renegade-project/linux-next
2
0
Fork 0
mirror of https://github.com/edk2-porting/linux-next.git synced 2024-12-21 19:53:59 +08:00
Code
e90886291c
linux-next/fs/io-wq.c

1425 lines
34 KiB
C
Raw Normal View History

io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
// SPDX-License-Identifier: GPL-2.0
/*
* Basic worker thread pool for io_uring
*
* Copyright (C) 2019 Jens Axboe
*
*/
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/errno.h>
#include <linux/sched/signal.h>
#include <linux/percpu.h>
#include <linux/slab.h>
#include <linux/rculist_nulls.h>
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
#include <linux/cpu.h>
tracehook: Remove tracehook.h Now that all of the definitions have moved out of tracehook.h into ptrace.h, sched/signal.h, resume_user_mode.h there is nothing left in tracehook.h so remove it. Update the few files that were depending upon tracehook.h to bring in definitions to use the headers they need directly. Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lkml.kernel.org/r/20220309162454.123006-13-ebiederm@xmission.com Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2022-02-10 02:47:08 +08:00
#include <linux/task_work.h>
audit,io_uring,io-wq: add some basic audit support to io_uring This patch adds basic auditing to io_uring operations, regardless of their context. This is accomplished by allocating audit_context structures for the io-wq worker and io_uring SQPOLL kernel threads as well as explicitly auditing the io_uring operations in io_issue_sqe(). Individual io_uring operations can bypass auditing through the "audit_skip" field in the struct io_op_def definition for the operation; although great care must be taken so that security relevant io_uring operations do not bypass auditing; please contact the audit mailing list (see the MAINTAINERS file) with any questions. The io_uring operations are audited using a new AUDIT_URINGOP record, an example is shown below: type=UNKNOWN[1336] msg=audit(1631800225.981:37289): uring_op=19 success=yes exit=0 items=0 ppid=15454 pid=15681 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=(null) Thanks to Richard Guy Briggs for review and feedback. Signed-off-by: Paul Moore <paul@paul-moore.com>
2021-02-17 08:46:48 +08:00
#include <linux/audit.h>
io-wq: provide IO_WQ_* constants for IORING_REGISTER_IOWQ_MAX_WORKERS arg items The items passed in the array pointed by the arg parameter of IORING_REGISTER_IOWQ_MAX_WORKERS io_uring_register operation carry certain semantics: they refer to different io-wq worker categories; provide IO_WQ_* constants in the UAPI, so these categories can be referenced in the user space code. Suggested-by: Jens Axboe <axboe@kernel.dk> Complements: 2e480058ddc21ec5 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com> Link: https://lore.kernel.org/r/20210913154415.GA12890@asgard.redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-13 23:44:15 +08:00
#include <uapi/linux/io_uring.h>
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
#include "io-wq.h"
#define WORKER_IDLE_TIMEOUT (5 * HZ)
enum {
IO_WORKER_F_UP = 1, /* up and active */
IO_WORKER_F_RUNNING = 2, /* account as running */
IO_WORKER_F_FREE = 4, /* worker on free list */
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
IO_WORKER_F_BOUND = 8, /* is doing bounded work */
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
};
enum {
IO_WQ_BIT_EXIT = 0, /* wq exiting */
};
enum {
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
};
/*
* One for each thread in a wqe pool
*/
struct io_worker {
refcount_t ref;
unsigned flags;
struct hlist_nulls_node nulls_node;
io-wq: ensure free/busy list browsing see all items We have two lists for workers in io-wq, a busy and a free list. For certain operations we want to browse all workers, and we currently do that by browsing the two separate lists. But since these lists are RCU protected, we can potentially miss workers if they move between the two lists while we're browsing them. Add a third list, all_list, that simply holds all workers. A worker is added to that list when it starts, and removed when it exits. This makes the worker iteration cleaner, too. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 04:54:49 +08:00
struct list_head all_list;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct task_struct *task;
struct io_wqe *wqe;
io-wq: ensure we have a stable view of ->cur_work for cancellations worker->cur_work is currently protected by the lock of the wqe that the worker belongs to. When we send a signal to a worker, we need a stable view of ->cur_work, so we need to hold that lock. But this doesn't work so well, since we have the opposite order potentially on queueing work. If POLL_ADD is used with a signalfd, then io_poll_wake() is called with the signal lock, and that sometimes needs to insert work items. Add a specific worker lock that protects the current work item. Then we can guarantee that the task we're sending a signal is currently processing the exact work we think it is. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 00:43:34 +08:00
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wq_work *cur_work;
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
struct io_wq_work *next_work;
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spinlock_t lock;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: fix race around io_worker grabbing There's a small window between lookup dropping the reference to the worker and calling wake_up_process() on the worker task, where the worker itself could have exited. We ensure that the worker struct itself is valid, but worker->task may very well be gone by the time we issue the wakeup. Fix the race by using a completion triggered by the reference going to zero, and having exit wait for that completion before proceeding. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-24 10:59:06 +08:00
struct completion ref_done;
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
unsigned long create_state;
struct callback_head create_work;
int create_index;
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
union {
struct rcu_head rcu;
struct work_struct work;
};
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
};
#if BITS_PER_LONG == 64
#define IO_WQ_HASH_ORDER 6
#else
#define IO_WQ_HASH_ORDER 5
#endif
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
#define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
struct io_wqe_acct {
unsigned nr_workers;
unsigned max_workers;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
int index;
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
atomic_t nr_running;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spinlock_t lock;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wq_work_list work_list;
unsigned long flags;
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
};
enum {
IO_WQ_ACCT_BOUND,
IO_WQ_ACCT_UNBOUND,
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
IO_WQ_ACCT_NR,
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
};
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
/*
* Per-node worker thread pool
*/
struct io_wqe {
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
raw_spinlock_t lock;
io-wq: use IO_WQ_ACCT_NR rather than hardcoded number It's better to use the defined enum stuff not the hardcoded number to define array. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-4-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:41 +08:00
struct io_wqe_acct acct[IO_WQ_ACCT_NR];
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
int node;
io-wq: remove now redundant struct io_wq_nulls_list Since we don't iterate these lists anymore after commit: e61df66c69b1 ("io-wq: ensure free/busy list browsing see all items") we don't need to retain the nulls value we use for them. That means it's pretty pointless to wrap the hlist_nulls_head in a structure, so get rid of it. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 23:00:41 +08:00
struct hlist_nulls_head free_list;
io-wq: ensure free/busy list browsing see all items We have two lists for workers in io-wq, a busy and a free list. For certain operations we want to browse all workers, and we currently do that by browsing the two separate lists. But since these lists are RCU protected, we can potentially miss workers if they move between the two lists while we're browsing them. Add a third list, all_list, that simply holds all workers. A worker is added to that list when it starts, and removed when it exits. This makes the worker iteration cleaner, too. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 04:54:49 +08:00
struct list_head all_list;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
struct wait_queue_entry wait;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wq *wq;
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
cpumask_var_t cpu_mask;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
};
/*
* Per io_wq state
*/
struct io_wq {
unsigned long state;
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
free_work_fn *free_work;
io_wq: add per-wq work handler instead of per work io_uring is the only user of io-wq, and now it uses only io-wq callback for all its requests, namely io_wq_submit_work(). Instead of storing work->runner callback in each instance of io_wq_work, keep it in io-wq itself. pros: - reduces io_wq_work size - more robust -- ->func won't be invalidated with mem{cpy,set}(req) - helps other work Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-09 02:08:20 +08:00
io_wq_work_fn *do_work;
io_wq: add get/put_work handlers to io_wq_create() For cancellation, we need to ensure that the work item stays valid for as long as ->cur_work is valid. Right now we can't safely dereference the work item even under the wqe->lock, because while the ->cur_work pointer will remain valid, the work could be completing and be freed in parallel. Only invoke ->get/put_work() on items we know that the caller queued themselves. Add IO_WQ_WORK_INTERNAL for io-wq to use, which is needed when we're queueing a flush item, for instance. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-13 13:31:31 +08:00
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
struct io_wq_hash *hash;
io-wq: have manager wait for all workers to exit Instead of having to wait separately on workers and manager, just have the manager wait on the workers. We use an atomic_t for the reference here, as we need to start at 0 and allow increment from that. Since the number of workers is naturally capped by the allowed nr of processes, and that uses an int, there is no risk of overflow. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 00:47:20 +08:00
atomic_t worker_refs;
struct completion worker_done;
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
struct hlist_node cpuhp_node;
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
struct task_struct *task;
io-wq: embed wqe ptr array into struct io_wq io-wq keeps an array of pointers to struct io_wqe, allocate this array as a part of struct io-wq, it's easier to code and saves an extra indirection for nearly each io-wq call. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/1482c6a001923bbed662dc38a8a580fb08b1ed8c.1623634181.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-14 09:36:12 +08:00
struct io_wqe *wqes[];
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
};
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
static enum cpuhp_state io_wq_online;
io-wq: ensure all pending work is canceled on exit If we race on shutting down the io-wq, then we should ensure that any work that was queued after workers shutdown is canceled. Harden the add work check a bit too, checking for IO_WQ_BIT_EXIT and cancel if it's set. Add a WARN_ON() for having any work before we kill the io-wq context. Reported-by: syzbot+91b4b56ead187d35c9d3@syzkaller.appspotmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-04 06:47:04 +08:00
struct io_cb_cancel_data {
work_cancel_fn *fn;
void *data;
int nr_running;
int nr_pending;
bool cancel_all;
};
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
static void io_wqe_dec_running(struct io_worker *worker);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
struct io_wqe_acct *acct,
struct io_cb_cancel_data *match);
io-wq: remove worker to owner tw dependency INFO: task iou-wrk-6609:6612 blocked for more than 143 seconds. Not tainted 5.15.0-rc5-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:iou-wrk-6609 state:D stack:27944 pid: 6612 ppid: 6526 flags:0x00004006 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xb44/0x5960 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_worker_exit fs/io-wq.c:183 [inline] io_wqe_worker+0x66d/0xc40 fs/io-wq.c:597 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 io-wq worker may submit a task_work to the master task and upon io_worker_exit() wait for the tw to get executed. The problem appears when the master task is waiting in coredump.c: 468 freezer_do_not_count(); 469 wait_for_completion(&core_state->startup); 470 freezer_count(); Apparently having some dependency on children threads getting everything stuck. Workaround it by cancelling the taks_work callback that causes it before going into io_worker_exit() waiting. p.s. probably a better option is to not submit tw elevating the refcount in the first place, but let's leave this excercise for the future. Cc: stable@vger.kernel.org Reported-and-tested-by: syzbot+27d62ee6f256b186883e@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/142a716f4ed936feae868959059154362bfa8c19.1635509451.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-29 20:11:33 +08:00
static void create_worker_cb(struct callback_head *cb);
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
static void io_wq_cancel_tw_create(struct io_wq *wq);
io-wq: ensure all pending work is canceled on exit If we race on shutting down the io-wq, then we should ensure that any work that was queued after workers shutdown is canceled. Harden the add work check a bit too, checking for IO_WQ_BIT_EXIT and cancel if it's set. Add a WARN_ON() for having any work before we kill the io-wq context. Reported-by: syzbot+91b4b56ead187d35c9d3@syzkaller.appspotmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-04 06:47:04 +08:00
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
static bool io_worker_get(struct io_worker *worker)
{
return refcount_inc_not_zero(&worker->ref);
}
static void io_worker_release(struct io_worker *worker)
{
if (refcount_dec_and_test(&worker->ref))
io-wq: fix race around io_worker grabbing There's a small window between lookup dropping the reference to the worker and calling wake_up_process() on the worker task, where the worker itself could have exited. We ensure that the worker struct itself is valid, but worker->task may very well be gone by the time we issue the wakeup. Fix the race by using a completion triggered by the reference going to zero, and having exit wait for that completion before proceeding. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-24 10:59:06 +08:00
complete(&worker->ref_done);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: refactor *_get_acct() Extract a helper for io_work_get_acct() and io_wqe_get_acct() to avoid duplication. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-22 09:58:28 +08:00
static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
{
return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
}
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
struct io_wq_work *work)
{
io-wq: refactor *_get_acct() Extract a helper for io_work_get_acct() and io_wqe_get_acct() to avoid duplication. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-22 09:58:28 +08:00
return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
}
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
{
io-wq: refactor *_get_acct() Extract a helper for io_work_get_acct() and io_wqe_get_acct() to avoid duplication. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-22 09:58:28 +08:00
return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
}
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
static void io_worker_ref_put(struct io_wq *wq)
{
if (atomic_dec_and_test(&wq->worker_refs))
complete(&wq->worker_done);
}
io-wq: remove worker to owner tw dependency INFO: task iou-wrk-6609:6612 blocked for more than 143 seconds. Not tainted 5.15.0-rc5-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:iou-wrk-6609 state:D stack:27944 pid: 6612 ppid: 6526 flags:0x00004006 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xb44/0x5960 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_worker_exit fs/io-wq.c:183 [inline] io_wqe_worker+0x66d/0xc40 fs/io-wq.c:597 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 io-wq worker may submit a task_work to the master task and upon io_worker_exit() wait for the tw to get executed. The problem appears when the master task is waiting in coredump.c: 468 freezer_do_not_count(); 469 wait_for_completion(&core_state->startup); 470 freezer_count(); Apparently having some dependency on children threads getting everything stuck. Workaround it by cancelling the taks_work callback that causes it before going into io_worker_exit() waiting. p.s. probably a better option is to not submit tw elevating the refcount in the first place, but let's leave this excercise for the future. Cc: stable@vger.kernel.org Reported-and-tested-by: syzbot+27d62ee6f256b186883e@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/142a716f4ed936feae868959059154362bfa8c19.1635509451.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-29 20:11:33 +08:00
static void io_worker_cancel_cb(struct io_worker *worker)
{
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
struct io_wqe *wqe = worker->wqe;
struct io_wq *wq = wqe->wq;
atomic_dec(&acct->nr_running);
raw_spin_lock(&worker->wqe->lock);
acct->nr_workers--;
raw_spin_unlock(&worker->wqe->lock);
io_worker_ref_put(wq);
clear_bit_unlock(0, &worker->create_state);
io_worker_release(worker);
}
static bool io_task_worker_match(struct callback_head *cb, void *data)
{
struct io_worker *worker;
if (cb->func != create_worker_cb)
return false;
worker = container_of(cb, struct io_worker, create_work);
return worker == data;
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
static void io_worker_exit(struct io_worker *worker)
{
struct io_wqe *wqe = worker->wqe;
io-wq: remove worker to owner tw dependency INFO: task iou-wrk-6609:6612 blocked for more than 143 seconds. Not tainted 5.15.0-rc5-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:iou-wrk-6609 state:D stack:27944 pid: 6612 ppid: 6526 flags:0x00004006 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xb44/0x5960 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_worker_exit fs/io-wq.c:183 [inline] io_wqe_worker+0x66d/0xc40 fs/io-wq.c:597 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 io-wq worker may submit a task_work to the master task and upon io_worker_exit() wait for the tw to get executed. The problem appears when the master task is waiting in coredump.c: 468 freezer_do_not_count(); 469 wait_for_completion(&core_state->startup); 470 freezer_count(); Apparently having some dependency on children threads getting everything stuck. Workaround it by cancelling the taks_work callback that causes it before going into io_worker_exit() waiting. p.s. probably a better option is to not submit tw elevating the refcount in the first place, but let's leave this excercise for the future. Cc: stable@vger.kernel.org Reported-and-tested-by: syzbot+27d62ee6f256b186883e@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/142a716f4ed936feae868959059154362bfa8c19.1635509451.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-29 20:11:33 +08:00
struct io_wq *wq = wqe->wq;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: remove worker to owner tw dependency INFO: task iou-wrk-6609:6612 blocked for more than 143 seconds. Not tainted 5.15.0-rc5-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:iou-wrk-6609 state:D stack:27944 pid: 6612 ppid: 6526 flags:0x00004006 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xb44/0x5960 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_worker_exit fs/io-wq.c:183 [inline] io_wqe_worker+0x66d/0xc40 fs/io-wq.c:597 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 io-wq worker may submit a task_work to the master task and upon io_worker_exit() wait for the tw to get executed. The problem appears when the master task is waiting in coredump.c: 468 freezer_do_not_count(); 469 wait_for_completion(&core_state->startup); 470 freezer_count(); Apparently having some dependency on children threads getting everything stuck. Workaround it by cancelling the taks_work callback that causes it before going into io_worker_exit() waiting. p.s. probably a better option is to not submit tw elevating the refcount in the first place, but let's leave this excercise for the future. Cc: stable@vger.kernel.org Reported-and-tested-by: syzbot+27d62ee6f256b186883e@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/142a716f4ed936feae868959059154362bfa8c19.1635509451.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-29 20:11:33 +08:00
while (1) {
struct callback_head *cb = task_work_cancel_match(wq->task,
io_task_worker_match, worker);
if (!cb)
break;
io_worker_cancel_cb(worker);
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: use helper for worker refcounting Use io_worker_release() instead of hand coding it in io_worker_exit(). Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/6f95f09d2cdbafcbb2e22ad0d1a2bc4d3962bf65.1634987320.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-23 19:13:55 +08:00
io_worker_release(worker);
io-wq: fix race around io_worker grabbing There's a small window between lookup dropping the reference to the worker and calling wake_up_process() on the worker task, where the worker itself could have exited. We ensure that the worker struct itself is valid, but worker->task may very well be gone by the time we issue the wakeup. Fix the race by using a completion triggered by the reference going to zero, and having exit wait for that completion before proceeding. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-24 10:59:06 +08:00
wait_for_completion(&worker->ref_done);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_lock(&wqe->lock);
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
if (worker->flags & IO_WORKER_F_FREE)
io-wq: only remove worker from free_list, if it was there If the worker isn't on the free_list, don't attempt to delete it. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 09:00:55 +08:00
hlist_nulls_del_rcu(&worker->nulls_node);
io-wq: ensure free/busy list browsing see all items We have two lists for workers in io-wq, a busy and a free list. For certain operations we want to browse all workers, and we currently do that by browsing the two separate lists. But since these lists are RCU protected, we can potentially miss workers if they move between the two lists while we're browsing them. Add a third list, all_list, that simply holds all workers. A worker is added to that list when it starts, and removed when it exits. This makes the worker iteration cleaner, too. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 04:54:49 +08:00
list_del_rcu(&worker->all_list);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
io_wqe_dec_running(worker);
worker->flags = 0;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
preempt_disable();
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
current->flags &= ~PF_IO_WORKER;
preempt_enable();
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: use kfree_rcu() to simplify the code The callback function of call_rcu() just calls kfree(), so we can use kfree_rcu() instead of call_rcu() + callback function. Signed-off-by: YueHaibing <yuehaibing@huawei.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-02 15:55:01 +08:00
kfree_rcu(worker, rcu);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io_worker_ref_put(wqe->wq);
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
do_exit(0);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
{
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
bool ret = false;
raw_spin_lock(&acct->lock);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
if (!wq_list_empty(&acct->work_list) &&
!test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
ret = true;
raw_spin_unlock(&acct->lock);
return ret;
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
}
/*
* Check head of free list for an available worker. If one isn't available,
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
* caller must create one.
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
*/
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
struct io_wqe_acct *acct)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
__must_hold(RCU)
{
struct hlist_nulls_node *n;
struct io_worker *worker;
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
/*
* Iterate free_list and see if we can find an idle worker to
* activate. If a given worker is on the free_list but in the process
* of exiting, keep trying.
*/
hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
if (!io_worker_get(worker))
continue;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
if (io_wqe_get_acct(worker) != acct) {
io_worker_release(worker);
continue;
}
io-wq: fix race between worker exiting and activating free worker Nadav correctly reports that we have a race between a worker exiting, and new work being queued. This can lead to work being queued behind an existing worker that could be sleeping on an event before it can run to completion, and hence introducing potential big latency gaps if we hit this race condition: cpu0 cpu1 ---- ---- io_wqe_worker() schedule_timeout() // timed out io_wqe_enqueue() io_wqe_wake_worker() // work_flags & IO_WQ_WORK_CONCURRENT io_wqe_activate_free_worker() io_worker_exit() Fix this by having the exiting worker go through the normal decrement of a running worker, which will spawn a new one if needed. The free worker activation is modified to only return success if we were able to find a sleeping worker - if not, we keep looking through the list. If we fail, we create a new worker as per usual. Cc: stable@vger.kernel.org Link: https://lore.kernel.org/io-uring/BFF746C0-FEDE-4646-A253-3021C57C26C9@gmail.com/ Reported-by: Nadav Amit <nadav.amit@gmail.com> Tested-by: Nadav Amit <nadav.amit@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-03 23:14:35 +08:00
if (wake_up_process(worker->task)) {
io_worker_release(worker);
return true;
}
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
io_worker_release(worker);
}
return false;
}
/*
* We need a worker. If we find a free one, we're good. If not, and we're
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
* below the max number of workers, create one.
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
*/
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
{
/*
* Most likely an attempt to queue unbounded work on an io_wq that
* wasn't setup with any unbounded workers.
*/
io_uring: fix false WARN_ONCE WARNING: CPU: 1 PID: 11749 at fs/io-wq.c:244 io_wqe_wake_worker fs/io-wq.c:244 [inline] WARNING: CPU: 1 PID: 11749 at fs/io-wq.c:244 io_wqe_enqueue+0x7f6/0x910 fs/io-wq.c:751 A WARN_ON_ONCE() in io_wqe_wake_worker() can be triggered by a valid userspace setup. Replace it with pr_warn. Reported-by: syzbot+ea2f1484cffe5109dc10@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/f7ede342c3342c4c26668f5168e2993e38bbd99c.1623949695.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 01:13:59 +08:00
if (unlikely(!acct->max_workers))
pr_warn_once("io-wq is not configured for unbound workers");
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
raw_spin_lock(&wqe->lock);
io-wq: max_worker fixes First, fix nr_workers checks against max_workers, with max_worker registration, it may pretty easily happen that nr_workers > max_workers. Also, synchronise writing to acct->max_worker with wqe->lock. It's not an actual problem, but as we don't care about io_wqe_create_worker(), it's better than WRITE_ONCE()/READ_ONCE(). Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Reported-by: Beld Zhang <beldzhang@gmail.com> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/11f90e6b49410b7d1a88f5d04fb8d95bb86b8cf3.1634671835.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-20 03:31:26 +08:00
if (acct->nr_workers >= acct->max_workers) {
io-wq: code clean of io_wqe_create_worker() Remove do_create to save a local variable. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-12 03:40:50 +08:00
raw_spin_unlock(&wqe->lock);
return true;
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
}
io-wq: code clean of io_wqe_create_worker() Remove do_create to save a local variable. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-12 03:40:50 +08:00
acct->nr_workers++;
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: code clean of io_wqe_create_worker() Remove do_create to save a local variable. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-12 03:40:50 +08:00
atomic_inc(&acct->nr_running);
atomic_inc(&wqe->wq->worker_refs);
return create_io_worker(wqe->wq, wqe, acct->index);
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
}
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
static void io_wqe_inc_running(struct io_worker *worker)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
{
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
atomic_inc(&acct->nr_running);
}
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
static void create_worker_cb(struct callback_head *cb)
{
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
struct io_worker *worker;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
struct io_wq *wq;
io-wq: fix lack of acct->nr_workers < acct->max_workers judgement There should be this judgement before we create an io-worker Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-05 18:05:38 +08:00
struct io_wqe *wqe;
struct io_wqe_acct *acct;
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
bool do_create = false;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
worker = container_of(cb, struct io_worker, create_work);
wqe = worker->wqe;
io-wq: fix lack of acct->nr_workers < acct->max_workers judgement There should be this judgement before we create an io-worker Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-05 18:05:38 +08:00
wq = wqe->wq;
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
acct = &wqe->acct[worker->create_index];
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_lock(&wqe->lock);
io-wq: fix bug of creating io-wokers unconditionally The former patch to add check between nr_workers and max_workers has a bug, which will cause unconditionally creating io-workers. That's because the result of the check doesn't affect the call of create_io_worker(), fix it by bringing in a boolean value for it. Fixes: 21698274da5b ("io-wq: fix lack of acct->nr_workers < acct->max_workers judgement") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20210808135434.68667-2-haoxu@linux.alibaba.com [axboe: drop hunk that isn't strictly needed] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-08 21:54:33 +08:00
if (acct->nr_workers < acct->max_workers) {
io-wq: fix lack of acct->nr_workers < acct->max_workers judgement There should be this judgement before we create an io-worker Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-05 18:05:38 +08:00
acct->nr_workers++;
io-wq: fix bug of creating io-wokers unconditionally The former patch to add check between nr_workers and max_workers has a bug, which will cause unconditionally creating io-workers. That's because the result of the check doesn't affect the call of create_io_worker(), fix it by bringing in a boolean value for it. Fixes: 21698274da5b ("io-wq: fix lack of acct->nr_workers < acct->max_workers judgement") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20210808135434.68667-2-haoxu@linux.alibaba.com [axboe: drop hunk that isn't strictly needed] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-08 21:54:33 +08:00
do_create = true;
}
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: fix bug of creating io-wokers unconditionally The former patch to add check between nr_workers and max_workers has a bug, which will cause unconditionally creating io-workers. That's because the result of the check doesn't affect the call of create_io_worker(), fix it by bringing in a boolean value for it. Fixes: 21698274da5b ("io-wq: fix lack of acct->nr_workers < acct->max_workers judgement") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20210808135434.68667-2-haoxu@linux.alibaba.com [axboe: drop hunk that isn't strictly needed] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-08 21:54:33 +08:00
if (do_create) {
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
create_io_worker(wq, wqe, worker->create_index);
io-wq: fix bug of creating io-wokers unconditionally The former patch to add check between nr_workers and max_workers has a bug, which will cause unconditionally creating io-workers. That's because the result of the check doesn't affect the call of create_io_worker(), fix it by bringing in a boolean value for it. Fixes: 21698274da5b ("io-wq: fix lack of acct->nr_workers < acct->max_workers judgement") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20210808135434.68667-2-haoxu@linux.alibaba.com [axboe: drop hunk that isn't strictly needed] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-08 21:54:33 +08:00
} else {
atomic_dec(&acct->nr_running);
io_worker_ref_put(wq);
}
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
clear_bit_unlock(0, &worker->create_state);
io_worker_release(worker);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
}
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static bool io_queue_worker_create(struct io_worker *worker,
struct io_wqe_acct *acct,
task_work_func_t func)
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
{
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
struct io_wqe *wqe = worker->wqe;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
struct io_wq *wq = wqe->wq;
/* raced with exit, just ignore create call */
if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
goto fail;
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
if (!io_worker_get(worker))
goto fail;
/*
* create_state manages ownership of create_work/index. We should
* only need one entry per worker, as the worker going to sleep
* will trigger the condition, and waking will clear it once it
* runs the task_work.
*/
if (test_bit(0, &worker->create_state) ||
test_and_set_bit_lock(0, &worker->create_state))
goto fail_release;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
atomic_inc(&wq->worker_refs);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
init_task_work(&worker->create_work, func);
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
worker->create_index = acct->index;
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
/*
* EXIT may have been set after checking it above, check after
* adding the task_work and remove any creation item if it is
* now set. wq exit does that too, but we can have added this
* work item after we canceled in io_wq_exit_workers().
*/
if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
io_wq_cancel_tw_create(wq);
io_worker_ref_put(wq);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
return true;
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
}
io_worker_ref_put(wq);
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
clear_bit_unlock(0, &worker->create_state);
fail_release:
io_worker_release(worker);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
fail:
atomic_dec(&acct->nr_running);
io_worker_ref_put(wq);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
return false;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
}
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
static void io_wqe_dec_running(struct io_worker *worker)
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
{
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
struct io_wqe *wqe = worker->wqe;
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
if (!(worker->flags & IO_WORKER_F_UP))
return;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
if (!atomic_dec_and_test(&acct->nr_running))
return;
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
if (!io_acct_run_queue(acct))
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
return;
atomic_inc(&acct->nr_running);
atomic_inc(&wqe->wq->worker_refs);
io_queue_worker_create(worker, acct, create_worker_cb);
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
/*
* Worker will start processing some work. Move it to the busy list, if
* it's currently on the freelist
*/
io-wq: remove useless 'work' argument to __io_worker_busy() We don't use 'work' anymore in the busy logic, remove the dead argument. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:10:11 +08:00
static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
if (worker->flags & IO_WORKER_F_FREE) {
worker->flags &= ~IO_WORKER_F_FREE;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&wqe->lock);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
hlist_nulls_del_init_rcu(&worker->nulls_node);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
}
/*
* No work, worker going to sleep. Move to freelist, and unuse mm if we
* have one attached. Dropping the mm may potentially sleep, so we drop
* the lock in that case and return success. Since the caller has to
* retry the loop in that case (we changed task state), we don't regrab
* the lock if we return success.
*/
io-wq: worker idling always returns false Remove the bool return, and the checking for it in the caller. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-16 04:26:34 +08:00
static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
__must_hold(wqe->lock)
{
if (!(worker->flags & IO_WORKER_F_FREE)) {
worker->flags |= IO_WORKER_F_FREE;
io-wq: remove now redundant struct io_wq_nulls_list Since we don't iterate these lists anymore after commit: e61df66c69b1 ("io-wq: ensure free/busy list browsing see all items") we don't need to retain the nulls value we use for them. That means it's pretty pointless to wrap the hlist_nulls_head in a structure, so get rid of it. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 23:00:41 +08:00
hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
}
io-wq: hash dependent work Enable io-wq hashing stuff for dependent works simply by re-enqueueing such requests. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:05 +08:00
static inline unsigned int io_get_work_hash(struct io_wq_work *work)
{
return work->flags >> IO_WQ_HASH_SHIFT;
}
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
{
struct io_wq *wq = wqe->wq;
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
bool ret = false;
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
io-wq: ensure that hash wait lock is IRQ disabling A previous commit removed the IRQ safety of the worker and wqe locks, but that left one spot of the hash wait lock now being done without already having IRQs disabled. Ensure that we use the right locking variant for the hashed waitqueue lock. Fixes: a9a4aa9fbfc5 ("io-wq: wqe and worker locks no longer need to be IRQ safe") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 20:57:25 +08:00
spin_lock_irq(&wq->hash->wait.lock);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
if (list_empty(&wqe->wait.entry)) {
__add_wait_queue(&wq->hash->wait, &wqe->wait);
if (!test_bit(hash, &wq->hash->map)) {
__set_current_state(TASK_RUNNING);
list_del_init(&wqe->wait.entry);
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
ret = true;
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
}
}
io-wq: ensure that hash wait lock is IRQ disabling A previous commit removed the IRQ safety of the worker and wqe locks, but that left one spot of the hash wait lock now being done without already having IRQs disabled. Ensure that we use the right locking variant for the hashed waitqueue lock. Fixes: a9a4aa9fbfc5 ("io-wq: wqe and worker locks no longer need to be IRQ safe") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 20:57:25 +08:00
spin_unlock_irq(&wq->hash->wait.lock);
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
return ret;
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
}
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
io-wq: fix queue stalling race We need to set the stalled bit early, before we drop the lock for adding us to the stall hash queue. If not, then we can race with new work being queued between adding us to the stall hash and io_worker_handle_work() marking us stalled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:53:00 +08:00
struct io_worker *worker)
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
__must_hold(acct->lock)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: shrink io_wq_work a bit Currently we're using 40 bytes for the io_wq_work structure, and 16 of those is the doubly link list node. We don't need doubly linked lists, we always add to tail to keep things ordered, and any other use case is list traversal with deletion. For the deletion case, we can easily support any node deletion by keeping track of the previous entry. This shrinks io_wq_work to 32 bytes, and subsequently io_kiock from io_uring to 216 to 208 bytes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:59:32 +08:00
struct io_wq_work_node *node, *prev;
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
struct io_wq_work *work, *tail;
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
unsigned int stall_hash = -1U;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wqe *wqe = worker->wqe;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_for_each(node, prev, &acct->work_list) {
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
unsigned int hash;
io-wq: shrink io_wq_work a bit Currently we're using 40 bytes for the io_wq_work structure, and 16 of those is the doubly link list node. We don't need doubly linked lists, we always add to tail to keep things ordered, and any other use case is list traversal with deletion. For the deletion case, we can easily support any node deletion by keeping track of the previous entry. This shrinks io_wq_work to 32 bytes, and subsequently io_kiock from io_uring to 216 to 208 bytes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:59:32 +08:00
work = container_of(node, struct io_wq_work, list);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
/* not hashed, can run anytime */
io-wq: split hashing and enqueueing It's a preparation patch removing io_wq_enqueue_hashed(), which now should be done by io_wq_hash_work() + io_wq_enqueue(). Also, set hash value for dependant works, and do it as late as possible, because req->file can be unavailable before. This hash will be ignored by io-wq. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:04 +08:00
if (!io_wq_is_hashed(work)) {
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_del(&acct->work_list, node, prev);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
return work;
}
io-wq: hash dependent work Enable io-wq hashing stuff for dependent works simply by re-enqueueing such requests. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:05 +08:00
hash = io_get_work_hash(work);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
/* all items with this hash lie in [work, tail] */
tail = wqe->hash_tail[hash];
/* hashed, can run if not already running */
if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
wqe->hash_tail[hash] = NULL;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_cut(&acct->work_list, &tail->list, prev);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
return work;
}
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
if (stall_hash == -1U)
stall_hash = hash;
/* fast forward to a next hash, for-each will fix up @prev */
node = &tail->list;
}
if (stall_hash != -1U) {
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
bool unstalled;
io-wq: fix queue stalling race We need to set the stalled bit early, before we drop the lock for adding us to the stall hash queue. If not, then we can race with new work being queued between adding us to the stall hash and io_worker_handle_work() marking us stalled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:53:00 +08:00
/*
* Set this before dropping the lock to avoid racing with new
* work being added and clearing the stalled bit.
*/
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&acct->lock);
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
unstalled = io_wait_on_hash(wqe, stall_hash);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&acct->lock);
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
if (unstalled) {
clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
if (wq_has_sleeper(&wqe->wq->hash->wait))
wake_up(&wqe->wq->hash->wait);
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
return NULL;
}
io-wq: ensure task is running before processing task_work Mark the current task as running if we need to run task_work from the io-wq threads as part of work handling. If that is the case, then return as such so that the caller can appropriately loop back and reset if it was part of a going-to-sleep flush. Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-21 21:06:56 +08:00
static bool io_flush_signals(void)
io_uring/io-wq: don't use static creds/mm assignments We currently setup the io_wq with a static set of mm and creds. Even for a single-use io-wq per io_uring, this is suboptimal as we have may have multiple enters of the ring. For sharing the io-wq backend, it doesn't work at all. Switch to passing in the creds and mm when the work item is setup. This means that async work is no longer deferred to the io_uring mm and creds, it is done with the current mm and creds. Flag this behavior with IORING_FEAT_CUR_PERSONALITY, so applications know they can rely on the current personality (mm and creds) being the same for direct issue and async issue. Reviewed-by: Stefan Metzmacher <metze@samba.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-01-28 07:34:48 +08:00
{
io_uring: don't use {test,clear}_tsk_thread_flag() for current Linus correctly points out that this is both unnecessary and generates much worse code on some archs as going from current to thread_info is actually backwards - and obviously just wasteful, since the thread_info is what we care about. Since io_uring only operates on current for these operations, just use test_thread_flag() instead. For io-wq, we can further simplify and use tracehook_notify_signal() to handle the TIF_NOTIFY_SIGNAL work and clear the flag. The latter isn't an actual bug right now, but it may very well be in the future if we place other work items under TIF_NOTIFY_SIGNAL. Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://lore.kernel.org/io-uring/CAHk-=wgYhNck33YHKZ14mFB5MzTTk8gqXHcfj=RWTAXKwgQJgg@mail.gmail.com/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-22 04:16:08 +08:00
if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
io-wq: ensure task is running before processing task_work Mark the current task as running if we need to run task_work from the io-wq threads as part of work handling. If that is the case, then return as such so that the caller can appropriately loop back and reset if it was part of a going-to-sleep flush. Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-21 21:06:56 +08:00
__set_current_state(TASK_RUNNING);
task_work: Decouple TIF_NOTIFY_SIGNAL and task_work There are a small handful of reasons besides pending signals that the kernel might want to break out of interruptible sleeps. The flag TIF_NOTIFY_SIGNAL and the helpers that set and clear TIF_NOTIFY_SIGNAL provide that the infrastructure for breaking out of interruptible sleeps and entering the return to user space slow path for those cases. Expand tracehook_notify_signal inline in it's callers and remove it, which makes clear that TIF_NOTIFY_SIGNAL and task_work are separate concepts. Update the comment on set_notify_signal to more accurately describe it's purpose. Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lkml.kernel.org/r/20220309162454.123006-9-ebiederm@xmission.com Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2022-02-10 01:18:54 +08:00
clear_notify_signal();
if (task_work_pending(current))
task_work_run();
io-wq: ensure task is running before processing task_work Mark the current task as running if we need to run task_work from the io-wq threads as part of work handling. If that is the case, then return as such so that the caller can appropriately loop back and reset if it was part of a going-to-sleep flush. Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-21 21:06:56 +08:00
return true;
io_uring/io-wq: don't use static creds/mm assignments We currently setup the io_wq with a static set of mm and creds. Even for a single-use io-wq per io_uring, this is suboptimal as we have may have multiple enters of the ring. For sharing the io-wq backend, it doesn't work at all. Switch to passing in the creds and mm when the work item is setup. This means that async work is no longer deferred to the io_uring mm and creds, it is done with the current mm and creds. Flag this behavior with IORING_FEAT_CUR_PERSONALITY, so applications know they can rely on the current personality (mm and creds) being the same for direct issue and async issue. Reviewed-by: Stefan Metzmacher <metze@samba.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-01-28 07:34:48 +08:00
}
io-wq: ensure task is running before processing task_work Mark the current task as running if we need to run task_work from the io-wq threads as part of work handling. If that is the case, then return as such so that the caller can appropriately loop back and reset if it was part of a going-to-sleep flush. Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-21 21:06:56 +08:00
return false;
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
}
static void io_assign_current_work(struct io_worker *worker,
struct io_wq_work *work)
{
io-wq: don't resched if there is no work This little tweak restores the behaviour that was before the recent io_worker_handle_work() optimisation patches. It makes the function do cond_resched() and flush_signals() only if there is an actual work to execute. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:03 +08:00
if (work) {
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
io_flush_signals();
io-wq: don't resched if there is no work This little tweak restores the behaviour that was before the recent io_worker_handle_work() optimisation patches. It makes the function do cond_resched() and flush_signals() only if there is an actual work to execute. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:03 +08:00
cond_resched();
}
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spin_lock(&worker->lock);
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
worker->cur_work = work;
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
worker->next_work = NULL;
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spin_unlock(&worker->lock);
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
}
io-wq: hash dependent work Enable io-wq hashing stuff for dependent works simply by re-enqueueing such requests. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:05 +08:00
static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
static void io_worker_handle_work(struct io_worker *worker)
{
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wqe *wqe = worker->wqe;
struct io_wq *wq = wqe->wq;
io-wq: cancel unbounded works on io-wq destroy WARNING: CPU: 5 PID: 227 at fs/io_uring.c:8578 io_ring_exit_work+0xe6/0x470 RIP: 0010:io_ring_exit_work+0xe6/0x470 Call Trace: process_one_work+0x206/0x400 worker_thread+0x4a/0x3d0 kthread+0x129/0x170 ret_from_fork+0x22/0x30 INFO: task lfs-openat:2359 blocked for more than 245 seconds. task:lfs-openat state:D stack: 0 pid: 2359 ppid: 1 flags:0x00000004 Call Trace: ... wait_for_completion+0x8b/0xf0 io_wq_destroy_manager+0x24/0x60 io_wq_put_and_exit+0x18/0x30 io_uring_clean_tctx+0x76/0xa0 __io_uring_files_cancel+0x1b9/0x2e0 do_exit+0xc0/0xb40 ... Even after io-wq destroy has been issued io-wq worker threads will continue executing all left work items as usual, and may hang waiting for I/O that won't ever complete (aka unbounded). [<0>] pipe_read+0x306/0x450 [<0>] io_iter_do_read+0x1e/0x40 [<0>] io_read+0xd5/0x330 [<0>] io_issue_sqe+0xd21/0x18a0 [<0>] io_wq_submit_work+0x6c/0x140 [<0>] io_worker_handle_work+0x17d/0x400 [<0>] io_wqe_worker+0x2c0/0x330 [<0>] ret_from_fork+0x22/0x30 Cancel all unbounded I/O instead of executing them. This changes the user visible behaviour, but that's inevitable as io-wq is not per task. Suggested-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/cd4b543154154cba055cf86f351441c2174d7f71.1617842918.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-08 08:54:42 +08:00
bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
do {
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
struct io_wq_work *work;
io-wq: delete dead lock shuffling code We used to have more code around the work loop, but now the goto and lock juggling just makes it less readable than it should. Get rid of it. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-20 04:11:58 +08:00
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
/*
* If we got some work, mark us as busy. If we didn't, but
* the list isn't empty, it means we stalled on hashed work.
* Mark us stalled so we don't keep looking for work when we
* can't make progress, any work completion or insertion will
* clear the stalled flag.
*/
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
raw_spin_lock(&acct->lock);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
work = io_get_next_work(acct, worker);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&acct->lock);
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
if (work) {
io-wq: remove useless 'work' argument to __io_worker_busy() We don't use 'work' anymore in the busy logic, remove the dead argument. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:10:11 +08:00
__io_worker_busy(wqe, worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
/*
* Make sure cancelation can find this, even before
* it becomes the active work. That avoids a window
* where the work has been removed from our general
* work list, but isn't yet discoverable as the
* current work item for this worker.
*/
raw_spin_lock(&worker->lock);
worker->next_work = work;
raw_spin_unlock(&worker->lock);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
} else {
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
break;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
}
io-wq: optimise locking in io_worker_handle_work() There are 2 optimisations: - Now, io_worker_handler_work() do io_assign_current_work() twice per request, and each one adds lock/unlock(worker->lock) pair. The first is to reset worker->cur_work to NULL, and the second to set a real work shortly after. If there is a dependant work, set it immediately, that effectively removes the extra NULL'ing. - And there is no use in taking wqe->lock for linked works, as they are not hashed now. Optimise it out. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:10 +08:00
io_assign_current_work(worker, work);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
__set_current_state(TASK_RUNNING);
io-wq: ensure we have a stable view of ->cur_work for cancellations worker->cur_work is currently protected by the lock of the wqe that the worker belongs to. When we send a signal to a worker, we need a stable view of ->cur_work, so we need to hold that lock. But this doesn't work so well, since we have the opposite order potentially on queueing work. If POLL_ADD is used with a signalfd, then io_poll_wake() is called with the signal lock, and that sometimes needs to insert work items. Add a specific worker lock that protects the current work item. Then we can guarantee that the task we're sending a signal is currently processing the exact work we think it is. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 00:43:34 +08:00
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
/* handle a whole dependent link */
do {
io_uring/io-wq: return 2-step work swap scheme Saving one lock/unlock for io-wq is not super important, but adds some ugliness in the code. More important, atomic decs not turning it to zero for some archs won't give the right ordering/barriers so the io_steal_work() may pretty easily get subtly and completely broken. Return back 2-step io-wq work exchange and clean it up. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-04 21:52:08 +08:00
struct io_wq_work *next_hashed, *linked;
io-wq: update hash bits Linked requests are hashed, remove a comment stating otherwise. Also move hash bits to emphasise that we don't carry it through loop iteration and set it every time. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-07-25 19:42:00 +08:00
unsigned int hash = io_get_work_hash(work);
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
next_hashed = wq_next_work(work);
io-wq: cancel unbounded works on io-wq destroy WARNING: CPU: 5 PID: 227 at fs/io_uring.c:8578 io_ring_exit_work+0xe6/0x470 RIP: 0010:io_ring_exit_work+0xe6/0x470 Call Trace: process_one_work+0x206/0x400 worker_thread+0x4a/0x3d0 kthread+0x129/0x170 ret_from_fork+0x22/0x30 INFO: task lfs-openat:2359 blocked for more than 245 seconds. task:lfs-openat state:D stack: 0 pid: 2359 ppid: 1 flags:0x00000004 Call Trace: ... wait_for_completion+0x8b/0xf0 io_wq_destroy_manager+0x24/0x60 io_wq_put_and_exit+0x18/0x30 io_uring_clean_tctx+0x76/0xa0 __io_uring_files_cancel+0x1b9/0x2e0 do_exit+0xc0/0xb40 ... Even after io-wq destroy has been issued io-wq worker threads will continue executing all left work items as usual, and may hang waiting for I/O that won't ever complete (aka unbounded). [<0>] pipe_read+0x306/0x450 [<0>] io_iter_do_read+0x1e/0x40 [<0>] io_read+0xd5/0x330 [<0>] io_issue_sqe+0xd21/0x18a0 [<0>] io_wq_submit_work+0x6c/0x140 [<0>] io_worker_handle_work+0x17d/0x400 [<0>] io_wqe_worker+0x2c0/0x330 [<0>] ret_from_fork+0x22/0x30 Cancel all unbounded I/O instead of executing them. This changes the user visible behaviour, but that's inevitable as io-wq is not per task. Suggested-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/cd4b543154154cba055cf86f351441c2174d7f71.1617842918.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-08 08:54:42 +08:00
if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
work->flags |= IO_WQ_WORK_CANCEL;
io_uring/io-wq: return 2-step work swap scheme Saving one lock/unlock for io-wq is not super important, but adds some ugliness in the code. More important, atomic decs not turning it to zero for some archs won't give the right ordering/barriers so the io_steal_work() may pretty easily get subtly and completely broken. Return back 2-step io-wq work exchange and clean it up. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-04 21:52:08 +08:00
wq->do_work(work);
io_assign_current_work(worker, NULL);
io-wq: shuffle io_worker_handle_work() code This is a preparation patch, it adds some helpers and makes the next patches cleaner. - extract io_impersonate_work() and io_assign_current_work() - replace @next label with nested do-while - move put_work() right after NULL'ing cur_work. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:09 +08:00
io_uring/io-wq: return 2-step work swap scheme Saving one lock/unlock for io-wq is not super important, but adds some ugliness in the code. More important, atomic decs not turning it to zero for some archs won't give the right ordering/barriers so the io_steal_work() may pretty easily get subtly and completely broken. Return back 2-step io-wq work exchange and clean it up. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-04 21:52:08 +08:00
linked = wq->free_work(work);
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
work = next_hashed;
if (!work && linked && !io_wq_is_hashed(linked)) {
work = linked;
linked = NULL;
}
io_assign_current_work(worker, work);
if (linked)
io_wqe_enqueue(wqe, linked);
if (hash != -1U && !next_hashed) {
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
/* serialize hash clear with wake_up() */
spin_lock_irq(&wq->hash->wait.lock);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
clear_bit(hash, &wq->hash->map);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
io-wq: serialize hash clear with wakeup We need to ensure that we serialize the stalled and hash bits with the wait_queue wait handler, or we could be racing with someone modifying the hashed state after we find it busy, but before we then give up and wait for it to be cleared. This can cause random delays or stalls when handling buffered writes for many files, where some of these files cause hash collisions between the worker threads. Cc: stable@vger.kernel.org Reported-by: Daniel Black <daniel@mariadb.org> Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-12 08:32:53 +08:00
spin_unlock_irq(&wq->hash->wait.lock);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
if (wq_has_sleeper(&wq->hash->wait))
wake_up(&wq->hash->wait);
io_wq: add get/put_work handlers to io_wq_create() For cancellation, we need to ensure that the work item stays valid for as long as ->cur_work is valid. Right now we can't safely dereference the work item even under the wqe->lock, because while the ->cur_work pointer will remain valid, the work could be completing and be freed in parallel. Only invoke ->get/put_work() on items we know that the caller queued themselves. Add IO_WQ_WORK_INTERNAL for io-wq to use, which is needed when we're queueing a flush item, for instance. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-13 13:31:31 +08:00
}
io-wq: optimise locking in io_worker_handle_work() There are 2 optimisations: - Now, io_worker_handler_work() do io_assign_current_work() twice per request, and each one adds lock/unlock(worker->lock) pair. The first is to reset worker->cur_work to NULL, and the second to set a real work shortly after. If there is a dependant work, set it immediately, that effectively removes the extra NULL'ing. - And there is no use in taking wqe->lock for linked works, as they are not hashed now. Optimise it out. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:10 +08:00
} while (work);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
} while (1);
}
static int io_wqe_worker(void *data)
{
struct io_worker *worker = data;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wqe *wqe = worker->wqe;
struct io_wq *wq = wqe->wq;
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
bool last_timeout = false;
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
char buf[TASK_COMM_LEN];
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
set_task_comm(current, buf);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
audit,io_uring,io-wq: add some basic audit support to io_uring This patch adds basic auditing to io_uring operations, regardless of their context. This is accomplished by allocating audit_context structures for the io-wq worker and io_uring SQPOLL kernel threads as well as explicitly auditing the io_uring operations in io_issue_sqe(). Individual io_uring operations can bypass auditing through the "audit_skip" field in the struct io_op_def definition for the operation; although great care must be taken so that security relevant io_uring operations do not bypass auditing; please contact the audit mailing list (see the MAINTAINERS file) with any questions. The io_uring operations are audited using a new AUDIT_URINGOP record, an example is shown below: type=UNKNOWN[1336] msg=audit(1631800225.981:37289): uring_op=19 success=yes exit=0 items=0 ppid=15454 pid=15681 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=(null) Thanks to Richard Guy Briggs for review and feedback. Signed-off-by: Paul Moore <paul@paul-moore.com>
2021-02-17 08:46:48 +08:00
audit_alloc_kernel(current);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
io_uring: allow IO worker threads to be frozen With the freezer using the proper signaling to notify us of when it's time to freeze a thread, we can re-enable normal freezer usage for the IO threads. Ensure that SQPOLL, io-wq, and the io-wq manager call try_to_freeze() appropriately, and remove the default setting of PF_NOFREEZE from create_io_thread(). Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-13 11:26:13 +08:00
long ret;
io-wq: remove worker->wait waitqueue We only have one cases of using the waitqueue to wake the worker, the rest are using wake_up_process(). Since we can save some cycles not fiddling with the waitqueue io_wqe_worker(), switch the work activation to task wakeup and get rid of the now unused wait_queue_head_t in struct io_worker. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-12-08 12:03:59 +08:00
set_current_state(TASK_INTERRUPTIBLE);
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
while (io_acct_run_queue(acct))
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io_worker_handle_work(worker);
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&wqe->lock);
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
/* timed out, exit unless we're the last worker */
if (last_timeout && acct->nr_workers > 1) {
io-wq: fix potential race of acct->nr_workers Given max_worker is 1, and we currently have 1 running and it is exiting. There may be race like: io_wqe_enqueue worker1 no work there and timeout unlock(wqe->lock) ->insert work -->io_worker_exit lock(wqe->lock) ->if(!nr_workers) //it's still 1 unlock(wqe->lock) goto run_cancel lock(wqe->lock) nr_workers-- ->dec_running ->worker creation fails unlock(wqe->lock) We enqueued one work but there is no workers, causes hung. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-12 03:40:52 +08:00
acct->nr_workers--;
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
raw_spin_unlock(&wqe->lock);
__set_current_state(TASK_RUNNING);
break;
}
last_timeout = false;
io-wq: worker idling always returns false Remove the bool return, and the checking for it in the caller. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-16 04:26:34 +08:00
__io_worker_idle(wqe, worker);
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: ensure task is running before processing task_work Mark the current task as running if we need to run task_work from the io-wq threads as part of work handling. If that is the case, then return as such so that the caller can appropriately loop back and reset if it was part of a going-to-sleep flush. Fixes: 3bfe6106693b ("io-wq: fork worker threads from original task") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-21 21:06:56 +08:00
if (io_flush_signals())
continue;
io_uring: allow IO worker threads to be frozen With the freezer using the proper signaling to notify us of when it's time to freeze a thread, we can re-enable normal freezer usage for the IO threads. Ensure that SQPOLL, io-wq, and the io-wq manager call try_to_freeze() appropriately, and remove the default setting of PF_NOFREEZE from create_io_thread(). Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-13 11:26:13 +08:00
ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
io_uring: handle signals for IO threads like a normal thread We go through various hoops to disallow signals for the IO threads, but there's really no reason why we cannot just allow them. The IO threads never return to userspace like a normal thread, and hence don't go through normal signal processing. Instead, just check for a pending signal as part of the work loop, and call get_signal() to handle it for us if anything is pending. With that, we can support receiving signals, including special ones like SIGSTOP. Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-26 08:16:06 +08:00
if (signal_pending(current)) {
struct ksignal ksig;
if (!get_signal(&ksig))
continue;
io-wq: exclusively gate signal based exit on get_signal() return io-wq threads block all signals, except SIGKILL and SIGSTOP. We should not need any extra checking of signal_pending or fatal_signal_pending, rely exclusively on whether or not get_signal() tells us to exit. The original debugging of this issue led to the false positive that we were exiting on non-fatal signals, but that is not the case. The issue was around races with nr_workers accounting. Fixes: 87c169665578 ("io-wq: ensure we exit if thread group is exiting") Fixes: 15e20db2e0ce ("io-wq: only exit on fatal signals") Reported-by: Eric W. Biederman <ebiederm@xmission.com> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-28 00:04:10 +08:00
break;
io_uring: handle signals for IO threads like a normal thread We go through various hoops to disallow signals for the IO threads, but there's really no reason why we cannot just allow them. The IO threads never return to userspace like a normal thread, and hence don't go through normal signal processing. Instead, just check for a pending signal as part of the work loop, and call get_signal() to handle it for us if anything is pending. With that, we can support receiving signals, including special ones like SIGSTOP. Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-26 08:16:06 +08:00
}
io-wq: get rid of FIXED worker flag It makes the logic easier to follow if we just get rid of the fixed worker flag, and simply ensure that we never exit the last worker in the group. This also means that no particular worker is special. Just track the last timeout state, and if we have hit it and no work is pending, check if there are other workers. If yes, then we can exit this one safely. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 03:01:17 +08:00
last_timeout = !ret;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: reduce acct->lock crossing functions lock/unlock reduce acct->lock lock and unlock in different functions to make the code clearer. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-3-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:40 +08:00
if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
io-wq: simplify worker exiting io_worker_handle_work() already takes care of the empty list case and releases spinlock, so get rid of ugly conditional unlocking and unconditionally call handle_work() Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/7521e485677f381036676943e876a0afecc23017.1623634181.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-14 09:36:17 +08:00
io_worker_handle_work(worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
audit,io_uring,io-wq: add some basic audit support to io_uring This patch adds basic auditing to io_uring operations, regardless of their context. This is accomplished by allocating audit_context structures for the io-wq worker and io_uring SQPOLL kernel threads as well as explicitly auditing the io_uring operations in io_issue_sqe(). Individual io_uring operations can bypass auditing through the "audit_skip" field in the struct io_op_def definition for the operation; although great care must be taken so that security relevant io_uring operations do not bypass auditing; please contact the audit mailing list (see the MAINTAINERS file) with any questions. The io_uring operations are audited using a new AUDIT_URINGOP record, an example is shown below: type=UNKNOWN[1336] msg=audit(1631800225.981:37289): uring_op=19 success=yes exit=0 items=0 ppid=15454 pid=15681 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=(null) Thanks to Richard Guy Briggs for review and feedback. Signed-off-by: Paul Moore <paul@paul-moore.com>
2021-02-17 08:46:48 +08:00
audit_free(current);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io_worker_exit(worker);
return 0;
}
/*
* Called when a worker is scheduled in. Mark us as currently running.
*/
void io_wq_worker_running(struct task_struct *tsk)
{
kthread: Generalize pf_io_worker so it can point to struct kthread The point of using set_child_tid to hold the kthread pointer was that it already did what is necessary. There are now restrictions on when set_child_tid can be initialized and when set_child_tid can be used in schedule_tail. Which indicates that continuing to use set_child_tid to hold the kthread pointer is a bad idea. Instead of continuing to use the set_child_tid field of task_struct generalize the pf_io_worker field of task_struct and use it to hold the kthread pointer. Rename pf_io_worker (which is a void * pointer) to worker_private so it can be used to store kthreads struct kthread pointer. Update the kthread code to store the kthread pointer in the worker_private field. Remove the places where set_child_tid had to be dealt with carefully because kthreads also used it. Link: https://lkml.kernel.org/r/CAHk-=wgtFAA9SbVYg0gR1tqPMC17-NYcs0GQkaYg1bGhh1uJQQ@mail.gmail.com Link: https://lkml.kernel.org/r/87a6grvqy8.fsf_-_@email.froward.int.ebiederm.org Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2021-12-23 12:10:09 +08:00
struct io_worker *worker = tsk->worker_private;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
if (!worker)
return;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
if (!(worker->flags & IO_WORKER_F_UP))
return;
if (worker->flags & IO_WORKER_F_RUNNING)
return;
worker->flags |= IO_WORKER_F_RUNNING;
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
io_wqe_inc_running(worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
/*
* Called when worker is going to sleep. If there are no workers currently
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
* running and we have work pending, wake up a free one or create a new one.
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
*/
void io_wq_worker_sleeping(struct task_struct *tsk)
{
kthread: Generalize pf_io_worker so it can point to struct kthread The point of using set_child_tid to hold the kthread pointer was that it already did what is necessary. There are now restrictions on when set_child_tid can be initialized and when set_child_tid can be used in schedule_tail. Which indicates that continuing to use set_child_tid to hold the kthread pointer is a bad idea. Instead of continuing to use the set_child_tid field of task_struct generalize the pf_io_worker field of task_struct and use it to hold the kthread pointer. Rename pf_io_worker (which is a void * pointer) to worker_private so it can be used to store kthreads struct kthread pointer. Update the kthread code to store the kthread pointer in the worker_private field. Remove the places where set_child_tid had to be dealt with carefully because kthreads also used it. Link: https://lkml.kernel.org/r/CAHk-=wgtFAA9SbVYg0gR1tqPMC17-NYcs0GQkaYg1bGhh1uJQQ@mail.gmail.com Link: https://lkml.kernel.org/r/87a6grvqy8.fsf_-_@email.froward.int.ebiederm.org Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2021-12-23 12:10:09 +08:00
struct io_worker *worker = tsk->worker_private;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
if (!worker)
return;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
if (!(worker->flags & IO_WORKER_F_UP))
return;
if (!(worker->flags & IO_WORKER_F_RUNNING))
return;
worker->flags &= ~IO_WORKER_F_RUNNING;
io-wq: don't pass 'wqe' needlessly around Just grab it from the worker itself, which we're already passing in. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-18 00:00:57 +08:00
io_wqe_dec_running(worker);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
struct task_struct *tsk)
{
kthread: Generalize pf_io_worker so it can point to struct kthread The point of using set_child_tid to hold the kthread pointer was that it already did what is necessary. There are now restrictions on when set_child_tid can be initialized and when set_child_tid can be used in schedule_tail. Which indicates that continuing to use set_child_tid to hold the kthread pointer is a bad idea. Instead of continuing to use the set_child_tid field of task_struct generalize the pf_io_worker field of task_struct and use it to hold the kthread pointer. Rename pf_io_worker (which is a void * pointer) to worker_private so it can be used to store kthreads struct kthread pointer. Update the kthread code to store the kthread pointer in the worker_private field. Remove the places where set_child_tid had to be dealt with carefully because kthreads also used it. Link: https://lkml.kernel.org/r/CAHk-=wgtFAA9SbVYg0gR1tqPMC17-NYcs0GQkaYg1bGhh1uJQQ@mail.gmail.com Link: https://lkml.kernel.org/r/87a6grvqy8.fsf_-_@email.froward.int.ebiederm.org Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2021-12-23 12:10:09 +08:00
tsk->worker_private = worker;
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
worker->task = tsk;
set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
tsk->flags |= PF_NO_SETAFFINITY;
raw_spin_lock(&wqe->lock);
hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
list_add_tail_rcu(&worker->all_list, &wqe->all_list);
worker->flags |= IO_WORKER_F_FREE;
raw_spin_unlock(&wqe->lock);
wake_up_new_task(tsk);
}
static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
{
return true;
}
static inline bool io_should_retry_thread(long err)
{
io-wq: don't retry task_work creation failure on fatal conditions We don't want to be retrying task_work creation failure if there's an actual signal pending for the parent task. If we do, then we can enter an infinite loop of perpetually retrying and each retry failing with -ERESTARTNOINTR because a signal is pending. Fixes: 3146cba99aa2 ("io-wq: make worker creation resilient against signals") Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20211202165606.mqryio4yzubl7ms5@pasture/ Tested-by: Florian Fischer <florian.fl.fischer@fau.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-03 10:40:15 +08:00
/*
* Prevent perpetual task_work retry, if the task (or its group) is
* exiting.
*/
if (fatal_signal_pending(current))
return false;
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
switch (err) {
case -EAGAIN:
case -ERESTARTSYS:
case -ERESTARTNOINTR:
case -ERESTARTNOHAND:
return true;
default:
return false;
}
}
static void create_worker_cont(struct callback_head *cb)
{
struct io_worker *worker;
struct task_struct *tsk;
struct io_wqe *wqe;
worker = container_of(cb, struct io_worker, create_work);
clear_bit_unlock(0, &worker->create_state);
wqe = worker->wqe;
tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
if (!IS_ERR(tsk)) {
io_init_new_worker(wqe, worker, tsk);
io_worker_release(worker);
return;
} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
atomic_dec(&acct->nr_running);
raw_spin_lock(&wqe->lock);
acct->nr_workers--;
if (!acct->nr_workers) {
struct io_cb_cancel_data match = {
.fn = io_wq_work_match_all,
.cancel_all = true,
};
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
while (io_acct_cancel_pending_work(wqe, acct, &match))
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
;
} else {
raw_spin_unlock(&wqe->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
}
io_worker_ref_put(wqe->wq);
io-wq: fix memory leak in create_io_worker() BUG: memory leak unreferenced object 0xffff888126fcd6c0 (size 192): comm "syz-executor.1", pid 11934, jiffies 4294983026 (age 15.690s) backtrace: [<ffffffff81632c91>] kmalloc_node include/linux/slab.h:609 [inline] [<ffffffff81632c91>] kzalloc_node include/linux/slab.h:732 [inline] [<ffffffff81632c91>] create_io_worker+0x41/0x1e0 fs/io-wq.c:739 [<ffffffff8163311e>] io_wqe_create_worker fs/io-wq.c:267 [inline] [<ffffffff8163311e>] io_wqe_enqueue+0x1fe/0x330 fs/io-wq.c:866 [<ffffffff81620b64>] io_queue_async_work+0xc4/0x200 fs/io_uring.c:1473 [<ffffffff8162c59c>] __io_queue_sqe+0x34c/0x510 fs/io_uring.c:6933 [<ffffffff8162c7ab>] io_req_task_submit+0x4b/0xa0 fs/io_uring.c:2233 [<ffffffff8162cb48>] io_async_task_func+0x108/0x1c0 fs/io_uring.c:5462 [<ffffffff816259e3>] tctx_task_work+0x1b3/0x3a0 fs/io_uring.c:2158 [<ffffffff81269b43>] task_work_run+0x73/0xb0 kernel/task_work.c:164 [<ffffffff812dcdd1>] tracehook_notify_signal include/linux/tracehook.h:212 [inline] [<ffffffff812dcdd1>] handle_signal_work kernel/entry/common.c:146 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_prepare+0x151/0x180 kernel/entry/common.c:209 [<ffffffff843ff25d>] __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] [<ffffffff843ff25d>] syscall_exit_to_user_mode+0x1d/0x40 kernel/entry/common.c:302 [<ffffffff843fa4a2>] do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 [<ffffffff84600068>] entry_SYSCALL_64_after_hwframe+0x44/0xae when create_io_thread() return error, and not retry, the worker object need to be freed. Reported-by: syzbot+65454c239241d3d647da@syzkaller.appspotmail.com Signed-off-by: Qiang.zhang <qiang.zhang@windriver.com> Link: https://lore.kernel.org/r/20210909115822.181188-1-qiang.zhang@windriver.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-09 19:58:22 +08:00
kfree(worker);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
return;
}
/* re-create attempts grab a new worker ref, drop the existing one */
io_worker_release(worker);
schedule_work(&worker->work);
}
static void io_workqueue_create(struct work_struct *work)
{
struct io_worker *worker = container_of(work, struct io_worker, work);
struct io_wqe_acct *acct = io_wqe_get_acct(worker);
io-wq: Remove duplicate code in io_workqueue_create() While task_work_add() in io_workqueue_create() is true, then duplicate code is executed: -> clear_bit_unlock(0, &worker->create_state); -> io_worker_release(worker); -> atomic_dec(&acct->nr_running); -> io_worker_ref_put(wq); -> return false; -> clear_bit_unlock(0, &worker->create_state); // back to io_workqueue_create() -> io_worker_release(worker); -> kfree(worker); The io_worker_release() and clear_bit_unlock() are executed twice. Fixes: 3146cba99aa2 ("io-wq: make worker creation resilient against signals") Signed-off-by: Bixuan Cui <cuibixuan@huawei.com> Link: https://lore.kernel.org/r/20210911085847.34849-1-cuibixuan@huawei.com Reviwed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-11 16:58:47 +08:00
if (!io_queue_worker_create(worker, acct, create_worker_cont))
io-wq: fix memory leak in create_io_worker() BUG: memory leak unreferenced object 0xffff888126fcd6c0 (size 192): comm "syz-executor.1", pid 11934, jiffies 4294983026 (age 15.690s) backtrace: [<ffffffff81632c91>] kmalloc_node include/linux/slab.h:609 [inline] [<ffffffff81632c91>] kzalloc_node include/linux/slab.h:732 [inline] [<ffffffff81632c91>] create_io_worker+0x41/0x1e0 fs/io-wq.c:739 [<ffffffff8163311e>] io_wqe_create_worker fs/io-wq.c:267 [inline] [<ffffffff8163311e>] io_wqe_enqueue+0x1fe/0x330 fs/io-wq.c:866 [<ffffffff81620b64>] io_queue_async_work+0xc4/0x200 fs/io_uring.c:1473 [<ffffffff8162c59c>] __io_queue_sqe+0x34c/0x510 fs/io_uring.c:6933 [<ffffffff8162c7ab>] io_req_task_submit+0x4b/0xa0 fs/io_uring.c:2233 [<ffffffff8162cb48>] io_async_task_func+0x108/0x1c0 fs/io_uring.c:5462 [<ffffffff816259e3>] tctx_task_work+0x1b3/0x3a0 fs/io_uring.c:2158 [<ffffffff81269b43>] task_work_run+0x73/0xb0 kernel/task_work.c:164 [<ffffffff812dcdd1>] tracehook_notify_signal include/linux/tracehook.h:212 [inline] [<ffffffff812dcdd1>] handle_signal_work kernel/entry/common.c:146 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_prepare+0x151/0x180 kernel/entry/common.c:209 [<ffffffff843ff25d>] __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] [<ffffffff843ff25d>] syscall_exit_to_user_mode+0x1d/0x40 kernel/entry/common.c:302 [<ffffffff843fa4a2>] do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 [<ffffffff84600068>] entry_SYSCALL_64_after_hwframe+0x44/0xae when create_io_thread() return error, and not retry, the worker object need to be freed. Reported-by: syzbot+65454c239241d3d647da@syzkaller.appspotmail.com Signed-off-by: Qiang.zhang <qiang.zhang@windriver.com> Link: https://lore.kernel.org/r/20210909115822.181188-1-qiang.zhang@windriver.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-09 19:58:22 +08:00
kfree(worker);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
}
static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
{
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
struct io_wqe_acct *acct = &wqe->acct[index];
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
struct io_worker *worker;
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
struct task_struct *tsk;
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
io-wq: fix races around manager/worker creation and task exit These races have always been there, they are just more apparent now that we do early cancel of io-wq when the task exits. Ensure that the io-wq manager sets task state correctly to not miss wakeups for task creation. This is important if we get a wakeup after having marked ourselves as TASK_INTERRUPTIBLE. If we do end up creating workers, then we flip the state back to running, making the subsequent schedule() a no-op. Also increment the wq ref count before forking the thread, to avoid a use-after-free. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-24 06:34:06 +08:00
__set_current_state(TASK_RUNNING);
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
if (!worker) {
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
fail:
atomic_dec(&acct->nr_running);
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_lock(&wqe->lock);
io-wq: fix no lock protection of acct->nr_worker There is an acct->nr_worker visit without lock protection. Think about the case: two callers call io_wqe_wake_worker(), one is the original context and the other one is an io-worker(by calling io_wqe_enqueue(wqe, linked)), on two cpus paralelly, this may cause nr_worker to be larger than max_worker. Let's fix it by adding lock for it, and let's do nr_workers++ before create_io_worker. There may be a edge cause that the first caller fails to create an io-worker, but the second caller doesn't know it and then quit creating io-worker as well: say nr_worker = max_worker - 1 cpu 0 cpu 1 io_wqe_wake_worker() io_wqe_wake_worker() nr_worker < max_worker nr_worker++ create_io_worker() nr_worker == max_worker failed return return But the chance of this case is very slim. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> [axboe: fix unconditional create_io_worker() call] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-05 18:05:37 +08:00
acct->nr_workers--;
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io_worker_ref_put(wq);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
return false;
io-wq: fork worker threads from original task Instead of using regular kthread kernel threads, create kernel threads that are like a real thread that the task would create. This ensures that we get all the context that we need, without having to carry that state around. This greatly reduces the code complexity, and the risk of missing state for a given request type. With the move away from kthread, we can also dump everything related to assigned state to the new threads. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-17 05:15:30 +08:00
}
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
refcount_set(&worker->ref, 1);
worker->wqe = wqe;
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spin_lock_init(&worker->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
init_completion(&worker->ref_done);
io_uring: move to using create_io_thread() This allows us to do task creation and setup without needing to use completions to try and synchronize with the starting thread. Get rid of the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup completion events - we can now do setup before the task is running. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-05 03:39:36 +08:00
if (index == IO_WQ_ACCT_BOUND)
worker->flags |= IO_WORKER_F_BOUND;
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
if (!IS_ERR(tsk)) {
io_init_new_worker(wqe, worker, tsk);
} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
io-wq: fix memory leak in create_io_worker() BUG: memory leak unreferenced object 0xffff888126fcd6c0 (size 192): comm "syz-executor.1", pid 11934, jiffies 4294983026 (age 15.690s) backtrace: [<ffffffff81632c91>] kmalloc_node include/linux/slab.h:609 [inline] [<ffffffff81632c91>] kzalloc_node include/linux/slab.h:732 [inline] [<ffffffff81632c91>] create_io_worker+0x41/0x1e0 fs/io-wq.c:739 [<ffffffff8163311e>] io_wqe_create_worker fs/io-wq.c:267 [inline] [<ffffffff8163311e>] io_wqe_enqueue+0x1fe/0x330 fs/io-wq.c:866 [<ffffffff81620b64>] io_queue_async_work+0xc4/0x200 fs/io_uring.c:1473 [<ffffffff8162c59c>] __io_queue_sqe+0x34c/0x510 fs/io_uring.c:6933 [<ffffffff8162c7ab>] io_req_task_submit+0x4b/0xa0 fs/io_uring.c:2233 [<ffffffff8162cb48>] io_async_task_func+0x108/0x1c0 fs/io_uring.c:5462 [<ffffffff816259e3>] tctx_task_work+0x1b3/0x3a0 fs/io_uring.c:2158 [<ffffffff81269b43>] task_work_run+0x73/0xb0 kernel/task_work.c:164 [<ffffffff812dcdd1>] tracehook_notify_signal include/linux/tracehook.h:212 [inline] [<ffffffff812dcdd1>] handle_signal_work kernel/entry/common.c:146 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] [<ffffffff812dcdd1>] exit_to_user_mode_prepare+0x151/0x180 kernel/entry/common.c:209 [<ffffffff843ff25d>] __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] [<ffffffff843ff25d>] syscall_exit_to_user_mode+0x1d/0x40 kernel/entry/common.c:302 [<ffffffff843fa4a2>] do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 [<ffffffff84600068>] entry_SYSCALL_64_after_hwframe+0x44/0xae when create_io_thread() return error, and not retry, the worker object need to be freed. Reported-by: syzbot+65454c239241d3d647da@syzkaller.appspotmail.com Signed-off-by: Qiang.zhang <qiang.zhang@windriver.com> Link: https://lore.kernel.org/r/20210909115822.181188-1-qiang.zhang@windriver.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-09 19:58:22 +08:00
kfree(worker);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
goto fail;
} else {
INIT_WORK(&worker->work, io_workqueue_create);
schedule_work(&worker->work);
}
return true;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: fix use-after-free in io_wq_worker_running The smart syzbot has found a reproducer for the following issue: ================================================================== BUG: KASAN: use-after-free in instrument_atomic_write include/linux/instrumented.h:71 [inline] BUG: KASAN: use-after-free in atomic_inc include/asm-generic/atomic-instrumented.h:240 [inline] BUG: KASAN: use-after-free in io_wqe_inc_running fs/io-wq.c:301 [inline] BUG: KASAN: use-after-free in io_wq_worker_running+0xde/0x110 fs/io-wq.c:613 Write of size 4 at addr ffff8882183db08c by task io_wqe_worker-0/7771 CPU: 0 PID: 7771 Comm: io_wqe_worker-0 Not tainted 5.9.0-rc4-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x198/0x1fd lib/dump_stack.c:118 print_address_description.constprop.0.cold+0xae/0x497 mm/kasan/report.c:383 __kasan_report mm/kasan/report.c:513 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:530 check_memory_region_inline mm/kasan/generic.c:186 [inline] check_memory_region+0x13d/0x180 mm/kasan/generic.c:192 instrument_atomic_write include/linux/instrumented.h:71 [inline] atomic_inc include/asm-generic/atomic-instrumented.h:240 [inline] io_wqe_inc_running fs/io-wq.c:301 [inline] io_wq_worker_running+0xde/0x110 fs/io-wq.c:613 schedule_timeout+0x148/0x250 kernel/time/timer.c:1879 io_wqe_worker+0x517/0x10e0 fs/io-wq.c:580 kthread+0x3b5/0x4a0 kernel/kthread.c:292 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294 Allocated by task 7768: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track mm/kasan/common.c:56 [inline] __kasan_kmalloc.constprop.0+0xbf/0xd0 mm/kasan/common.c:461 kmem_cache_alloc_node_trace+0x17b/0x3f0 mm/slab.c:3594 kmalloc_node include/linux/slab.h:572 [inline] kzalloc_node include/linux/slab.h:677 [inline] io_wq_create+0x57b/0xa10 fs/io-wq.c:1064 io_init_wq_offload fs/io_uring.c:7432 [inline] io_sq_offload_start fs/io_uring.c:7504 [inline] io_uring_create fs/io_uring.c:8625 [inline] io_uring_setup+0x1836/0x28e0 fs/io_uring.c:8694 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 21: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track+0x1c/0x30 mm/kasan/common.c:56 kasan_set_free_info+0x1b/0x30 mm/kasan/generic.c:355 __kasan_slab_free+0xd8/0x120 mm/kasan/common.c:422 __cache_free mm/slab.c:3418 [inline] kfree+0x10e/0x2b0 mm/slab.c:3756 __io_wq_destroy fs/io-wq.c:1138 [inline] io_wq_destroy+0x2af/0x460 fs/io-wq.c:1146 io_finish_async fs/io_uring.c:6836 [inline] io_ring_ctx_free fs/io_uring.c:7870 [inline] io_ring_exit_work+0x1e4/0x6d0 fs/io_uring.c:7954 process_one_work+0x94c/0x1670 kernel/workqueue.c:2269 worker_thread+0x64c/0x1120 kernel/workqueue.c:2415 kthread+0x3b5/0x4a0 kernel/kthread.c:292 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294 The buggy address belongs to the object at ffff8882183db000 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 140 bytes inside of 1024-byte region [ffff8882183db000, ffff8882183db400) The buggy address belongs to the page: page:000000009bada22b refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x2183db flags: 0x57ffe0000000200(slab) raw: 057ffe0000000200 ffffea0008604c48 ffffea00086a8648 ffff8880aa040700 raw: 0000000000000000 ffff8882183db000 0000000100000002 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8882183daf80: ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ffff8882183db000: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb >ffff8882183db080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff8882183db100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8882183db180: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ================================================================== which is down to the comment below, /* all workers gone, wq exit can proceed */ if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) complete(&wqe->wq->done); because there might be multiple cases of wqe in a wq and we would wait for every worker in every wqe to go home before releasing wq's resources on destroying. To that end, rework wq's refcount by making it independent of the tracking of workers because after all they are two different things, and keeping it balanced when workers come and go. Note the manager kthread, like other workers, now holds a grab to wq during its lifetime. Finally to help destroy wq, check IO_WQ_BIT_EXIT upon creating worker and do nothing for exiting wq. Cc: stable@vger.kernel.org # v5.5+ Reported-by: syzbot+45fa0a195b941764e0f0@syzkaller.appspotmail.com Reported-by: syzbot+9af99580130003da82b1@syzkaller.appspotmail.com Cc: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Hillf Danton <hdanton@sina.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-09-26 21:26:55 +08:00
/*
* Iterate the passed in list and call the specific function for each
* worker that isn't exiting
*/
static bool io_wq_for_each_worker(struct io_wqe *wqe,
bool (*func)(struct io_worker *, void *),
void *data)
{
struct io_worker *worker;
bool ret = false;
list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
if (io_worker_get(worker)) {
/* no task if node is/was offline */
if (worker->task)
ret = func(worker, data);
io_worker_release(worker);
if (ret)
break;
}
}
return ret;
}
static bool io_wq_worker_wake(struct io_worker *worker, void *data)
{
io-wq: use __set_notify_signal() to wake workers The only difference between set_notify_signal() and __set_notify_signal() is that the former checks if it needs to deliver an IPI to force a reschedule. As the io-wq workers never leave the kernel, and IPI is never needed, they simply need a wakeup. Reviewed-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/20220426014904.60384-4-axboe@kernel.dk Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-04-26 09:49:01 +08:00
__set_notify_signal(worker->task);
io-wq: fix use-after-free in io_wq_worker_running The smart syzbot has found a reproducer for the following issue: ================================================================== BUG: KASAN: use-after-free in instrument_atomic_write include/linux/instrumented.h:71 [inline] BUG: KASAN: use-after-free in atomic_inc include/asm-generic/atomic-instrumented.h:240 [inline] BUG: KASAN: use-after-free in io_wqe_inc_running fs/io-wq.c:301 [inline] BUG: KASAN: use-after-free in io_wq_worker_running+0xde/0x110 fs/io-wq.c:613 Write of size 4 at addr ffff8882183db08c by task io_wqe_worker-0/7771 CPU: 0 PID: 7771 Comm: io_wqe_worker-0 Not tainted 5.9.0-rc4-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x198/0x1fd lib/dump_stack.c:118 print_address_description.constprop.0.cold+0xae/0x497 mm/kasan/report.c:383 __kasan_report mm/kasan/report.c:513 [inline] kasan_report.cold+0x1f/0x37 mm/kasan/report.c:530 check_memory_region_inline mm/kasan/generic.c:186 [inline] check_memory_region+0x13d/0x180 mm/kasan/generic.c:192 instrument_atomic_write include/linux/instrumented.h:71 [inline] atomic_inc include/asm-generic/atomic-instrumented.h:240 [inline] io_wqe_inc_running fs/io-wq.c:301 [inline] io_wq_worker_running+0xde/0x110 fs/io-wq.c:613 schedule_timeout+0x148/0x250 kernel/time/timer.c:1879 io_wqe_worker+0x517/0x10e0 fs/io-wq.c:580 kthread+0x3b5/0x4a0 kernel/kthread.c:292 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294 Allocated by task 7768: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track mm/kasan/common.c:56 [inline] __kasan_kmalloc.constprop.0+0xbf/0xd0 mm/kasan/common.c:461 kmem_cache_alloc_node_trace+0x17b/0x3f0 mm/slab.c:3594 kmalloc_node include/linux/slab.h:572 [inline] kzalloc_node include/linux/slab.h:677 [inline] io_wq_create+0x57b/0xa10 fs/io-wq.c:1064 io_init_wq_offload fs/io_uring.c:7432 [inline] io_sq_offload_start fs/io_uring.c:7504 [inline] io_uring_create fs/io_uring.c:8625 [inline] io_uring_setup+0x1836/0x28e0 fs/io_uring.c:8694 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 21: kasan_save_stack+0x1b/0x40 mm/kasan/common.c:48 kasan_set_track+0x1c/0x30 mm/kasan/common.c:56 kasan_set_free_info+0x1b/0x30 mm/kasan/generic.c:355 __kasan_slab_free+0xd8/0x120 mm/kasan/common.c:422 __cache_free mm/slab.c:3418 [inline] kfree+0x10e/0x2b0 mm/slab.c:3756 __io_wq_destroy fs/io-wq.c:1138 [inline] io_wq_destroy+0x2af/0x460 fs/io-wq.c:1146 io_finish_async fs/io_uring.c:6836 [inline] io_ring_ctx_free fs/io_uring.c:7870 [inline] io_ring_exit_work+0x1e4/0x6d0 fs/io_uring.c:7954 process_one_work+0x94c/0x1670 kernel/workqueue.c:2269 worker_thread+0x64c/0x1120 kernel/workqueue.c:2415 kthread+0x3b5/0x4a0 kernel/kthread.c:292 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294 The buggy address belongs to the object at ffff8882183db000 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 140 bytes inside of 1024-byte region [ffff8882183db000, ffff8882183db400) The buggy address belongs to the page: page:000000009bada22b refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x2183db flags: 0x57ffe0000000200(slab) raw: 057ffe0000000200 ffffea0008604c48 ffffea00086a8648 ffff8880aa040700 raw: 0000000000000000 ffff8882183db000 0000000100000002 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8882183daf80: ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ffff8882183db000: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb >ffff8882183db080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ ffff8882183db100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8882183db180: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ================================================================== which is down to the comment below, /* all workers gone, wq exit can proceed */ if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs)) complete(&wqe->wq->done); because there might be multiple cases of wqe in a wq and we would wait for every worker in every wqe to go home before releasing wq's resources on destroying. To that end, rework wq's refcount by making it independent of the tracking of workers because after all they are two different things, and keeping it balanced when workers come and go. Note the manager kthread, like other workers, now holds a grab to wq during its lifetime. Finally to help destroy wq, check IO_WQ_BIT_EXIT upon creating worker and do nothing for exiting wq. Cc: stable@vger.kernel.org # v5.5+ Reported-by: syzbot+45fa0a195b941764e0f0@syzkaller.appspotmail.com Reported-by: syzbot+9af99580130003da82b1@syzkaller.appspotmail.com Cc: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Hillf Danton <hdanton@sina.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-09-26 21:26:55 +08:00
wake_up_process(worker->task);
return false;
}
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
io-wq: fix IO_WQ_WORK_NO_CANCEL cancellation To cancel a work, io-wq sets IO_WQ_WORK_CANCEL and executes the callback. However, IO_WQ_WORK_NO_CANCEL works will just execute and may return next work, which will be ignored and lost. Cancel the whole link. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-02 00:18:19 +08:00
{
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
struct io_wq *wq = wqe->wq;
io-wq: fix IO_WQ_WORK_NO_CANCEL cancellation To cancel a work, io-wq sets IO_WQ_WORK_CANCEL and executes the callback. However, IO_WQ_WORK_NO_CANCEL works will just execute and may return next work, which will be ignored and lost. Cancel the whole link. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-02 00:18:19 +08:00
do {
work->flags |= IO_WQ_WORK_CANCEL;
io_uring/io-wq: return 2-step work swap scheme Saving one lock/unlock for io-wq is not super important, but adds some ugliness in the code. More important, atomic decs not turning it to zero for some archs won't give the right ordering/barriers so the io_steal_work() may pretty easily get subtly and completely broken. Return back 2-step io-wq work exchange and clean it up. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-04 21:52:08 +08:00
wq->do_work(work);
work = wq->free_work(work);
io-wq: fix IO_WQ_WORK_NO_CANCEL cancellation To cancel a work, io-wq sets IO_WQ_WORK_CANCEL and executes the callback. However, IO_WQ_WORK_NO_CANCEL works will just execute and may return next work, which will be ignored and lost. Cancel the whole link. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-02 00:18:19 +08:00
} while (work);
}
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
{
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
unsigned int hash;
struct io_wq_work *tail;
if (!io_wq_is_hashed(work)) {
append:
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_add_tail(&work->list, &acct->work_list);
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
return;
}
hash = io_get_work_hash(work);
tail = wqe->hash_tail[hash];
wqe->hash_tail[hash] = work;
if (!tail)
goto append;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_add_after(&work->list, &tail->list, &acct->work_list);
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
}
io-wq: fix cancellation on create-worker failure WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1151 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1146 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 io_req_complete_post+0xf5b/0x1190 fs/io_uring.c:1794 Modules linked in: Call Trace: tctx_task_work+0x1e5/0x570 fs/io_uring.c:2158 task_work_run+0xe0/0x1a0 kernel/task_work.c:164 tracehook_notify_signal include/linux/tracehook.h:212 [inline] handle_signal_work kernel/entry/common.c:146 [inline] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] exit_to_user_mode_prepare+0x232/0x2a0 kernel/entry/common.c:209 __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:302 do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 entry_SYSCALL_64_after_hwframe+0x44/0xae When io_wqe_enqueue() -> io_wqe_create_worker() fails, we can't just call io_run_cancel() to clean up the request, it's already enqueued via io_wqe_insert_work() and will be executed either by some other worker during cancellation (e.g. in io_wq_put_and_exit()). Reported-by: Hao Sun <sunhao.th@gmail.com> Fixes: 3146cba99aa28 ("io-wq: make worker creation resilient against signals") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/93b9de0fcf657affab0acfd675d4abcd273ee863.1631092071.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-08 17:09:29 +08:00
static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
{
return work == data;
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
{
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
struct io_cb_cancel_data match;
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
unsigned work_flags = work->flags;
bool do_create;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io_uring: explicitly catch any illegal async queue attempt Catch an illegal case to queue async from an unrelated task that got the ring fd passed to it. This should not be possible to hit, but better be proactive and catch it explicitly. io-wq is extended to check for early IO_WQ_WORK_CANCEL being set on a work item as well, so it can run the request through the normal cancelation path. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-07-24 01:53:54 +08:00
/*
* If io-wq is exiting for this task, or if the request has explicitly
* been marked as one that should not get executed, cancel it here.
*/
if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
(work->flags & IO_WQ_WORK_CANCEL)) {
io-wq: fix ref leak for req in case of exit cancelations do_work such as io_wq_submit_work that cancel the work may leave a ref of req as 1 if we have links. Fix it by call io_run_cancel. Fixes: 4fb6ac326204 ("io-wq: improve manager/worker handling over exec") Signed-off-by: yangerkun <yangerkun@huawei.com> Link: https://lore.kernel.org/r/20210309030410.3294078-1-yangerkun@huawei.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 11:04:10 +08:00
io_run_cancel(work, wqe);
io-wq: improve manager/worker handling over exec exec will cancel any threads, including the ones that io-wq is using. This isn't a problem, in fact we'd prefer it to be that way since it means we know that any async work cancels naturally without having to handle it proactively. But it does mean that we need to setup a new manager, as the manager and workers are gone. Handle this at queue time, and cancel work if we fail. Since the manager can go away without us noticing, ensure that the manager itself holds a reference to the 'wq' as well. Rename io_wq_destroy() to io_wq_put() to reflect that. In the future we can now simplify exec cancelation handling, for now just leave it the same. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-26 01:17:09 +08:00
return;
}
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&acct->lock);
io-wq: handle hashed writes in chains We always punt async buffered writes to an io-wq helper, as the core kernel does not have IOCB_NOWAIT support for that. Most buffered async writes complete very quickly, as it's just a copy operation. This means that doing multiple locking roundtrips on the shared wqe lock for each buffered write is wasteful. Additionally, buffered writes are hashed work items, which means that any buffered write to a given file is serialized. Keep identicaly hashed work items contiguously in @wqe->work_list, and track a tail for each hash bucket. On dequeue of a hashed item, splice all of the same hash in one go using the tracked tail. Until the batch is done, the caller doesn't have to synchronize with the wqe or worker locks again. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-24 03:57:22 +08:00
io_wqe_insert_work(wqe, work);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&acct->lock);
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&wqe->lock);
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
rcu_read_lock();
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
do_create = !io_wqe_activate_free_worker(wqe, acct);
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
rcu_read_unlock();
io-wq: wqe and worker locks no longer need to be IRQ safe io_uring no longer queues async work off completion handlers that run in hard or soft interrupt context, and that use case was the only reason that io-wq had to use IRQ safe locks for wqe and worker locks. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-30 20:33:08 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: fix race between adding work and activating a free worker The attempt to find and activate a free worker for new work is currently combined with creating a new one if we don't find one, but that opens io-wq up to a race where the worker that is found and activated can put itself to sleep without knowing that it has been selected to perform this new work. Fix this by moving the activation into where we add the new work item, then we can retain it within the wqe->lock scope and elimiate the race with the worker itself checking inside the lock, but sleeping outside of it. Cc: stable@vger.kernel.org Reported-by: Andres Freund <andres@anarazel.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-31 01:55:22 +08:00
if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
!atomic_read(&acct->nr_running))) {
bool did_create;
did_create = io_wqe_create_worker(wqe, acct);
io-wq: fix cancellation on create-worker failure WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1151 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1146 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 io_req_complete_post+0xf5b/0x1190 fs/io_uring.c:1794 Modules linked in: Call Trace: tctx_task_work+0x1e5/0x570 fs/io_uring.c:2158 task_work_run+0xe0/0x1a0 kernel/task_work.c:164 tracehook_notify_signal include/linux/tracehook.h:212 [inline] handle_signal_work kernel/entry/common.c:146 [inline] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] exit_to_user_mode_prepare+0x232/0x2a0 kernel/entry/common.c:209 __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:302 do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 entry_SYSCALL_64_after_hwframe+0x44/0xae When io_wqe_enqueue() -> io_wqe_create_worker() fails, we can't just call io_run_cancel() to clean up the request, it's already enqueued via io_wqe_insert_work() and will be executed either by some other worker during cancellation (e.g. in io_wq_put_and_exit()). Reported-by: Hao Sun <sunhao.th@gmail.com> Fixes: 3146cba99aa28 ("io-wq: make worker creation resilient against signals") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/93b9de0fcf657affab0acfd675d4abcd273ee863.1631092071.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-08 17:09:29 +08:00
if (likely(did_create))
return;
raw_spin_lock(&wqe->lock);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
if (acct->nr_workers) {
raw_spin_unlock(&wqe->lock);
return;
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
}
io-wq: fix cancellation on create-worker failure WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1151 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 req_ref_put_and_test fs/io_uring.c:1146 [inline] WARNING: CPU: 0 PID: 10392 at fs/io_uring.c:1151 io_req_complete_post+0xf5b/0x1190 fs/io_uring.c:1794 Modules linked in: Call Trace: tctx_task_work+0x1e5/0x570 fs/io_uring.c:2158 task_work_run+0xe0/0x1a0 kernel/task_work.c:164 tracehook_notify_signal include/linux/tracehook.h:212 [inline] handle_signal_work kernel/entry/common.c:146 [inline] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] exit_to_user_mode_prepare+0x232/0x2a0 kernel/entry/common.c:209 __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:302 do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 entry_SYSCALL_64_after_hwframe+0x44/0xae When io_wqe_enqueue() -> io_wqe_create_worker() fails, we can't just call io_run_cancel() to clean up the request, it's already enqueued via io_wqe_insert_work() and will be executed either by some other worker during cancellation (e.g. in io_wq_put_and_exit()). Reported-by: Hao Sun <sunhao.th@gmail.com> Fixes: 3146cba99aa28 ("io-wq: make worker creation resilient against signals") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/93b9de0fcf657affab0acfd675d4abcd273ee863.1631092071.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-08 17:09:29 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
/* fatal condition, failed to create the first worker */
match.fn = io_wq_work_match_item,
match.data = work,
match.cancel_all = false,
io_acct_cancel_pending_work(wqe, acct, &match);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
{
struct io_wqe *wqe = wq->wqes[numa_node_id()];
io_wqe_enqueue(wqe, work);
}
/*
io-wq: split hashing and enqueueing It's a preparation patch removing io_wq_enqueue_hashed(), which now should be done by io_wq_hash_work() + io_wq_enqueue(). Also, set hash value for dependant works, and do it as late as possible, because req->file can be unavailable before. This hash will be ignored by io-wq. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:04 +08:00
* Work items that hash to the same value will not be done in parallel.
* Used to limit concurrent writes, generally hashed by inode.
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
*/
io-wq: split hashing and enqueueing It's a preparation patch removing io_wq_enqueue_hashed(), which now should be done by io_wq_hash_work() + io_wq_enqueue(). Also, set hash value for dependant works, and do it as late as possible, because req->file can be unavailable before. This hash will be ignored by io-wq. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:04 +08:00
void io_wq_hash_work(struct io_wq_work *work, void *val)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: split hashing and enqueueing It's a preparation patch removing io_wq_enqueue_hashed(), which now should be done by io_wq_hash_work() + io_wq_enqueue(). Also, set hash value for dependant works, and do it as late as possible, because req->file can be unavailable before. This hash will be ignored by io-wq. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-14 05:31:04 +08:00
unsigned int bit;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
bit = hash_ptr(val, IO_WQ_HASH_ORDER);
work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
}
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
static bool __io_wq_worker_cancel(struct io_worker *worker,
struct io_cb_cancel_data *match,
struct io_wq_work *work)
{
if (work && match->fn(work, match->data)) {
work->flags |= IO_WQ_WORK_CANCEL;
io-wq: use __set_notify_signal() to wake workers The only difference between set_notify_signal() and __set_notify_signal() is that the former checks if it needs to deliver an IPI to force a reschedule. As the io-wq workers never leave the kernel, and IPI is never needed, they simply need a wakeup. Reviewed-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/20220426014904.60384-4-axboe@kernel.dk Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-04-26 09:49:01 +08:00
__set_notify_signal(worker->task);
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
return true;
}
return false;
}
io-wq: remove duplicated cancel code Deduplicate cancellation parts, as many of them looks the same, as do e.g. - io_wqe_cancel_cb_work() and io_wqe_cancel_work() - io_wq_worker_cancel() and io_work_cancel() Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-07 06:15:39 +08:00
static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
io_uring: support for generic async request cancel This adds support for IORING_OP_ASYNC_CANCEL, which will attempt to cancel requests that have been punted to async context and are now in-flight. This works for regular read/write requests to files, as long as they haven't been started yet. For socket based IO (or things like accept4(2)), we can cancel work that is already running as well. To cancel a request, the sqe must have ->addr set to the user_data of the request it wishes to cancel. If the request is cancelled successfully, the original request is completed with -ECANCELED and the cancel request is completed with a result of 0. If the request was already running, the original may or may not complete in error. The cancel request will complete with -EALREADY for that case. And finally, if the request to cancel wasn't found, the cancel request is completed with -ENOENT. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-29 11:49:21 +08:00
{
io-wq: remove duplicated cancel code Deduplicate cancellation parts, as many of them looks the same, as do e.g. - io_wqe_cancel_cb_work() and io_wqe_cancel_work() - io_wq_worker_cancel() and io_work_cancel() Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-07 06:15:39 +08:00
struct io_cb_cancel_data *match = data;
io_uring: support for generic async request cancel This adds support for IORING_OP_ASYNC_CANCEL, which will attempt to cancel requests that have been punted to async context and are now in-flight. This works for regular read/write requests to files, as long as they haven't been started yet. For socket based IO (or things like accept4(2)), we can cancel work that is already running as well. To cancel a request, the sqe must have ->addr set to the user_data of the request it wishes to cancel. If the request is cancelled successfully, the original request is completed with -ECANCELED and the cancel request is completed with a result of 0. If the request was already running, the original may or may not complete in error. The cancel request will complete with -EALREADY for that case. And finally, if the request to cancel wasn't found, the cancel request is completed with -ENOENT. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-29 11:49:21 +08:00
/*
* Hold the lock to avoid ->cur_work going out of scope, caller
io-wq: ensure we have a stable view of ->cur_work for cancellations worker->cur_work is currently protected by the lock of the wqe that the worker belongs to. When we send a signal to a worker, we need a stable view of ->cur_work, so we need to hold that lock. But this doesn't work so well, since we have the opposite order potentially on queueing work. If POLL_ADD is used with a signalfd, then io_poll_wake() is called with the signal lock, and that sometimes needs to insert work items. Add a specific worker lock that protects the current work item. Then we can guarantee that the task we're sending a signal is currently processing the exact work we think it is. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 00:43:34 +08:00
* may dereference the passed in work.
io_uring: support for generic async request cancel This adds support for IORING_OP_ASYNC_CANCEL, which will attempt to cancel requests that have been punted to async context and are now in-flight. This works for regular read/write requests to files, as long as they haven't been started yet. For socket based IO (or things like accept4(2)), we can cancel work that is already running as well. To cancel a request, the sqe must have ->addr set to the user_data of the request it wishes to cancel. If the request is cancelled successfully, the original request is completed with -ECANCELED and the cancel request is completed with a result of 0. If the request was already running, the original may or may not complete in error. The cancel request will complete with -EALREADY for that case. And finally, if the request to cancel wasn't found, the cancel request is completed with -ENOENT. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-29 11:49:21 +08:00
*/
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spin_lock(&worker->lock);
io-wq: add intermediate work step between pending list and active work We have a gap where a worker removes an item from the work list and to when it gets added as the workers active work. In this state, the work item cannot be found by cancelations. This is a small window, but it does exist. Add a temporary pointer to a work item that isn't on the pending work list anymore, but also not the active work. This is needed as we need to drop the wqe lock in between grabbing the work item and marking it as active, to ensure that signal based cancelations are properly ordered. Reported-by: Florian Fischer <florian.fl.fischer@fau.de> Link: https://lore.kernel.org/io-uring/20220118151337.fac6cthvbnu7icoc@pasture/ Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:23:51 +08:00
if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
__io_wq_worker_cancel(worker, match, worker->next_work))
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
match->nr_running++;
io-wq: make io_worker lock a raw spinlock In preparation to nesting it under the wqe lock (which is raw due to being acquired from the scheduler side), change the io_worker lock from a normal spinlock to a raw spinlock. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:13:43 +08:00
raw_spin_unlock(&worker->lock);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
return match->nr_running && !match->cancel_all;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: fix hang after cancelling pending hashed work Don't forget to update wqe->hash_tail after cancelling a pending work item, if it was hashed. Cc: stable@vger.kernel.org # 5.7+ Reported-by: Dmitry Shulyak <yashulyak@gmail.com> Fixes: 86f3cd1b589a1 ("io-wq: handle hashed writes in chains") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-24 01:33:10 +08:00
static inline void io_wqe_remove_pending(struct io_wqe *wqe,
struct io_wq_work *work,
struct io_wq_work_node *prev)
{
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
io-wq: fix hang after cancelling pending hashed work Don't forget to update wqe->hash_tail after cancelling a pending work item, if it was hashed. Cc: stable@vger.kernel.org # 5.7+ Reported-by: Dmitry Shulyak <yashulyak@gmail.com> Fixes: 86f3cd1b589a1 ("io-wq: handle hashed writes in chains") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-24 01:33:10 +08:00
unsigned int hash = io_get_work_hash(work);
struct io_wq_work *prev_work = NULL;
if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
if (prev)
prev_work = container_of(prev, struct io_wq_work, list);
if (prev_work && io_get_work_hash(prev_work) == hash)
wqe->hash_tail[hash] = prev_work;
else
wqe->hash_tail[hash] = NULL;
}
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wq_list_del(&acct->work_list, &work->list, prev);
io-wq: fix hang after cancelling pending hashed work Don't forget to update wqe->hash_tail after cancelling a pending work item, if it was hashed. Cc: stable@vger.kernel.org # 5.7+ Reported-by: Dmitry Shulyak <yashulyak@gmail.com> Fixes: 86f3cd1b589a1 ("io-wq: handle hashed writes in chains") Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-24 01:33:10 +08:00
}
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
struct io_wqe_acct *acct,
struct io_cb_cancel_data *match)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: shrink io_wq_work a bit Currently we're using 40 bytes for the io_wq_work structure, and 16 of those is the doubly link list node. We don't need doubly linked lists, we always add to tail to keep things ordered, and any other use case is list traversal with deletion. For the deletion case, we can easily support any node deletion by keeping track of the previous entry. This shrinks io_wq_work to 32 bytes, and subsequently io_kiock from io_uring to 216 to 208 bytes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:59:32 +08:00
struct io_wq_work_node *node, *prev;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wq_work *work;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&acct->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
wq_list_for_each(node, prev, &acct->work_list) {
work = container_of(node, struct io_wq_work, list);
if (!match->fn(work, match->data))
continue;
io_wqe_remove_pending(wqe, work, prev);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&acct->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
io_run_cancel(work, wqe);
match->nr_pending++;
/* not safe to continue after unlock */
return true;
}
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_unlock(&acct->lock);
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
return false;
}
static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
struct io_cb_cancel_data *match)
{
int i;
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
retry:
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++) {
struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
io-wq: make worker creation resilient against signals If a task is queueing async work and also handling signals, then we can run into the case where create_io_thread() is interrupted and returns failure because of that. If this happens for creating the first worker in a group, then that worker will never get created and we can hang the ring. If we do get a fork failure, retry from task_work. With signals we have to be a bit careful as we cannot simply queue as task_work, as we'll still have signals pending at that point. Punt over a normal workqueue first and then create from task_work after that. Lastly, ensure that we handle fatal worker creations. Worker creation failures are normally not fatal, only if we fail to create one in an empty worker group can we not make progress. Right now that is ignored, ensure that we handle that and run cancel on the work item. There are two paths that create new workers - one is the "existing worker going to sleep", and the other is "no workers found for this work, create one". The former is never fatal, as workers do exist in the group. Only the latter needs to be carefully handled. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-02 01:20:10 +08:00
if (io_acct_cancel_pending_work(wqe, acct, match)) {
if (match->cancel_all)
goto retry;
io-wq: invoke work cancelation with wqe->lock held io_wqe_cancel_pending_work() grabs it internally, grab it upfront instead. For the running work cancelation, grab the lock around it as well. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:18:20 +08:00
break;
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
}
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
static void io_wqe_cancel_running_work(struct io_wqe *wqe,
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
struct io_cb_cancel_data *match)
{
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
rcu_read_lock();
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
rcu_read_unlock();
}
io-wq: remove duplicated cancel code Deduplicate cancellation parts, as many of them looks the same, as do e.g. - io_wqe_cancel_cb_work() and io_wqe_cancel_work() - io_wq_worker_cancel() and io_work_cancel() Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-07 06:15:39 +08:00
enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
void *data, bool cancel_all)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: remove duplicated cancel code Deduplicate cancellation parts, as many of them looks the same, as do e.g. - io_wqe_cancel_cb_work() and io_wqe_cancel_work() - io_wq_worker_cancel() and io_work_cancel() Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-07 06:15:39 +08:00
struct io_cb_cancel_data match = {
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
.fn = cancel,
.data = data,
.cancel_all = cancel_all,
io-wq: make io_wqe_cancel_work() take a match handler We want to use the cancel functionality for canceling based on not just the work itself. Instead of matching on the work address manually, allow a match handler to tell us if we found the right work item or not. No functional changes in this patch. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-09 10:13:32 +08:00
};
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
int node;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
/*
* First check pending list, if we're lucky we can just remove it
* from there. CANCEL_OK means that the work is returned as-new,
* no completion will be posted for it.
io-wq: perform both unstarted and started work cancelations in one go Rather than split these into two separate lookups and matches, combine them into one loop. This will become important when we can guarantee that we don't have a window where a pending work item isn't discoverable in either state. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:22:32 +08:00
*
* Then check if a free (going busy) or busy worker has the work
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
* currently running. If we find it there, we'll return CANCEL_RUNNING
* as an indication that we attempt to signal cancellation. The
* completion will run normally in this case.
io-wq: perform both unstarted and started work cancelations in one go Rather than split these into two separate lookups and matches, combine them into one loop. This will become important when we can guarantee that we don't have a window where a pending work item isn't discoverable in either state. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:22:32 +08:00
*
* Do both of these while holding the wqe->lock, to ensure that
* we'll find a work item regardless of state.
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
*/
for_each_node(node) {
struct io_wqe *wqe = wq->wqes[node];
io-wq: perform both unstarted and started work cancelations in one go Rather than split these into two separate lookups and matches, combine them into one loop. This will become important when we can guarantee that we don't have a window where a pending work item isn't discoverable in either state. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:22:32 +08:00
io_wqe_cancel_pending_work(wqe, &match);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
if (match.nr_pending && !match.cancel_all)
io-wq: perform both unstarted and started work cancelations in one go Rather than split these into two separate lookups and matches, combine them into one loop. This will become important when we can guarantee that we don't have a window where a pending work item isn't discoverable in either state. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:22:32 +08:00
return IO_WQ_CANCEL_OK;
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock(&wqe->lock);
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
io_wqe_cancel_running_work(wqe, &match);
io-wq: invoke work cancelation with wqe->lock held io_wqe_cancel_pending_work() grabs it internally, grab it upfront instead. For the running work cancelation, grab the lock around it as well. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-01-19 10:18:20 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
if (match.nr_running && !match.cancel_all)
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
return IO_WQ_CANCEL_RUNNING;
}
io-wq: add an option to cancel all matched reqs This adds support for cancelling all io-wq works matching a predicate. It isn't used yet, so no change in observable behaviour. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:03 +08:00
if (match.nr_running)
return IO_WQ_CANCEL_RUNNING;
if (match.nr_pending)
return IO_WQ_CANCEL_OK;
io-wq: reorder cancellation pending -> running Go all over all pending lists and cancel works there, and only then try to match running requests. No functional changes here, just a preparation for bulk cancellation. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-15 15:24:02 +08:00
return IO_WQ_CANCEL_NOTFOUND;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
int sync, void *key)
{
struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
int i;
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
list_del_init(&wait->entry);
rcu_read_lock();
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++) {
struct io_wqe_acct *acct = &wqe->acct[i];
if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
io_wqe_activate_free_worker(wqe, acct);
}
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
rcu_read_unlock();
return 1;
}
io-wq: have io_wq_create() take a 'data' argument We currently pass in 4 arguments outside of the bounded size. In preparation for adding one more argument, let's bundle them up in a struct to make it more readable. No functional changes in this patch. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-25 23:49:20 +08:00
struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
int ret, node, i;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wq *wq;
io_wq: add per-wq work handler instead of per work io_uring is the only user of io-wq, and now it uses only io-wq callback for all its requests, namely io_wq_submit_work(). Instead of storing work->runner callback in each instance of io_wq_work, keep it in io-wq itself. pros: - reduces io_wq_work size - more robust -- ->func won't be invalidated with mem{cpy,set}(req) - helps other work Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-09 02:08:20 +08:00
if (WARN_ON_ONCE(!data->free_work || !data->do_work))
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
return ERR_PTR(-EINVAL);
io_uring: fix false WARN_ONCE WARNING: CPU: 1 PID: 11749 at fs/io-wq.c:244 io_wqe_wake_worker fs/io-wq.c:244 [inline] WARNING: CPU: 1 PID: 11749 at fs/io-wq.c:244 io_wqe_enqueue+0x7f6/0x910 fs/io-wq.c:751 A WARN_ON_ONCE() in io_wqe_wake_worker() can be triggered by a valid userspace setup. Replace it with pr_warn. Reported-by: syzbot+ea2f1484cffe5109dc10@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/f7ede342c3342c4c26668f5168e2993e38bbd99c.1623949695.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 01:13:59 +08:00
if (WARN_ON_ONCE(!bounded))
return ERR_PTR(-EINVAL);
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
io-wq: embed wqe ptr array into struct io_wq io-wq keeps an array of pointers to struct io_wqe, allocate this array as a part of struct io-wq, it's easier to code and saves an extra indirection for nearly each io-wq call. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/1482c6a001923bbed662dc38a8a580fb08b1ed8c.1623634181.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-14 09:36:12 +08:00
wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
if (!wq)
return ERR_PTR(-ENOMEM);
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
if (ret)
io-wq: embed wqe ptr array into struct io_wq io-wq keeps an array of pointers to struct io_wqe, allocate this array as a part of struct io-wq, it's easier to code and saves an extra indirection for nearly each io-wq call. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/1482c6a001923bbed662dc38a8a580fb08b1ed8c.1623634181.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-14 09:36:12 +08:00
goto err_wq;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
refcount_inc(&data->hash->refs);
wq->hash = data->hash;
io_uring/io-wq: forward submission ref to async First it changes io-wq interfaces. It replaces {get,put}_work() with free_work(), which guaranteed to be called exactly once. It also enforces free_work() callback to be non-NULL. io_uring follows the changes and instead of putting a submission reference in io_put_req_async_completion(), it will be done in io_free_work(). As removes io_get_work() with corresponding refcount_inc(), the ref balance is maintained. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-03-04 21:14:12 +08:00
wq->free_work = data->free_work;
io_wq: add per-wq work handler instead of per work io_uring is the only user of io-wq, and now it uses only io-wq callback for all its requests, namely io_wq_submit_work(). Instead of storing work->runner callback in each instance of io_wq_work, keep it in io-wq itself. pros: - reduces io_wq_work size - more robust -- ->func won't be invalidated with mem{cpy,set}(req) - helps other work Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-06-09 02:08:20 +08:00
wq->do_work = data->do_work;
io_wq: add get/put_work handlers to io_wq_create() For cancellation, we need to ensure that the work item stays valid for as long as ->cur_work is valid. Right now we can't safely dereference the work item even under the wqe->lock, because while the ->cur_work pointer will remain valid, the work could be completing and be freed in parallel. Only invoke ->get/put_work() on items we know that the caller queued themselves. Add IO_WQ_WORK_INTERNAL for io-wq to use, which is needed when we're queueing a flush item, for instance. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-13 13:31:31 +08:00
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
ret = -ENOMEM;
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
for_each_node(node) {
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
struct io_wqe *wqe;
io-wq: don't call kXalloc_node() with non-online node Glauber reports a crash on init on a box he has: RIP: 0010:__alloc_pages_nodemask+0x132/0x340 Code: 18 01 75 04 41 80 ce 80 89 e8 48 8b 54 24 08 8b 74 24 1c c1 e8 0c 48 8b 3c 24 83 e0 01 88 44 24 20 48 85 d2 0f 85 74 01 00 00 <3b> 77 08 0f 82 6b 01 00 00 48 89 7c 24 10 89 ea 48 8b 07 b9 00 02 RSP: 0018:ffffb8be4d0b7c28 EFLAGS: 00010246 RAX: 0000000000000000 RBX: 0000000000000000 RCX: 000000000000e8e8 RDX: 0000000000000000 RSI: 0000000000000002 RDI: 0000000000002080 RBP: 0000000000012cc0 R08: 0000000000000000 R09: 0000000000000002 R10: 0000000000000dc0 R11: ffff995c60400100 R12: 0000000000000000 R13: 0000000000012cc0 R14: 0000000000000001 R15: ffff995c60db00f0 FS: 00007f4d115ca900(0000) GS:ffff995c60d80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000002088 CR3: 00000017cca66002 CR4: 00000000007606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 PKRU: 55555554 Call Trace: alloc_slab_page+0x46/0x320 new_slab+0x9d/0x4e0 ___slab_alloc+0x507/0x6a0 ? io_wq_create+0xb4/0x2a0 __slab_alloc+0x1c/0x30 kmem_cache_alloc_node_trace+0xa6/0x260 io_wq_create+0xb4/0x2a0 io_uring_setup+0x97f/0xaa0 ? io_remove_personalities+0x30/0x30 ? io_poll_trigger_evfd+0x30/0x30 do_syscall_64+0x5b/0x1c0 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f4d116cb1ed which is due to the 'wqe' and 'worker' allocation being node affine. But it isn't valid to call the node affine allocation if the node isn't online. Setup structures for even offline nodes, as usual, but skip them in terms of thread setup to not waste resources. If the node isn't online, just alloc memory with NUMA_NO_NODE. Reported-by: Glauber Costa <glauber@scylladb.com> Tested-by: Glauber Costa <glauber@scylladb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-11 21:30:06 +08:00
int alloc_node = node;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: don't call kXalloc_node() with non-online node Glauber reports a crash on init on a box he has: RIP: 0010:__alloc_pages_nodemask+0x132/0x340 Code: 18 01 75 04 41 80 ce 80 89 e8 48 8b 54 24 08 8b 74 24 1c c1 e8 0c 48 8b 3c 24 83 e0 01 88 44 24 20 48 85 d2 0f 85 74 01 00 00 <3b> 77 08 0f 82 6b 01 00 00 48 89 7c 24 10 89 ea 48 8b 07 b9 00 02 RSP: 0018:ffffb8be4d0b7c28 EFLAGS: 00010246 RAX: 0000000000000000 RBX: 0000000000000000 RCX: 000000000000e8e8 RDX: 0000000000000000 RSI: 0000000000000002 RDI: 0000000000002080 RBP: 0000000000012cc0 R08: 0000000000000000 R09: 0000000000000002 R10: 0000000000000dc0 R11: ffff995c60400100 R12: 0000000000000000 R13: 0000000000012cc0 R14: 0000000000000001 R15: ffff995c60db00f0 FS: 00007f4d115ca900(0000) GS:ffff995c60d80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000002088 CR3: 00000017cca66002 CR4: 00000000007606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 PKRU: 55555554 Call Trace: alloc_slab_page+0x46/0x320 new_slab+0x9d/0x4e0 ___slab_alloc+0x507/0x6a0 ? io_wq_create+0xb4/0x2a0 __slab_alloc+0x1c/0x30 kmem_cache_alloc_node_trace+0xa6/0x260 io_wq_create+0xb4/0x2a0 io_uring_setup+0x97f/0xaa0 ? io_remove_personalities+0x30/0x30 ? io_poll_trigger_evfd+0x30/0x30 do_syscall_64+0x5b/0x1c0 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f4d116cb1ed which is due to the 'wqe' and 'worker' allocation being node affine. But it isn't valid to call the node affine allocation if the node isn't online. Setup structures for even offline nodes, as usual, but skip them in terms of thread setup to not waste resources. If the node isn't online, just alloc memory with NUMA_NO_NODE. Reported-by: Glauber Costa <glauber@scylladb.com> Tested-by: Glauber Costa <glauber@scylladb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-11 21:30:06 +08:00
if (!node_online(alloc_node))
alloc_node = NUMA_NO_NODE;
wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
if (!wqe)
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
goto err;
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
goto err;
cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
wq->wqes[node] = wqe;
io-wq: don't call kXalloc_node() with non-online node Glauber reports a crash on init on a box he has: RIP: 0010:__alloc_pages_nodemask+0x132/0x340 Code: 18 01 75 04 41 80 ce 80 89 e8 48 8b 54 24 08 8b 74 24 1c c1 e8 0c 48 8b 3c 24 83 e0 01 88 44 24 20 48 85 d2 0f 85 74 01 00 00 <3b> 77 08 0f 82 6b 01 00 00 48 89 7c 24 10 89 ea 48 8b 07 b9 00 02 RSP: 0018:ffffb8be4d0b7c28 EFLAGS: 00010246 RAX: 0000000000000000 RBX: 0000000000000000 RCX: 000000000000e8e8 RDX: 0000000000000000 RSI: 0000000000000002 RDI: 0000000000002080 RBP: 0000000000012cc0 R08: 0000000000000000 R09: 0000000000000002 R10: 0000000000000dc0 R11: ffff995c60400100 R12: 0000000000000000 R13: 0000000000012cc0 R14: 0000000000000001 R15: ffff995c60db00f0 FS: 00007f4d115ca900(0000) GS:ffff995c60d80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000002088 CR3: 00000017cca66002 CR4: 00000000007606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 PKRU: 55555554 Call Trace: alloc_slab_page+0x46/0x320 new_slab+0x9d/0x4e0 ___slab_alloc+0x507/0x6a0 ? io_wq_create+0xb4/0x2a0 __slab_alloc+0x1c/0x30 kmem_cache_alloc_node_trace+0xa6/0x260 io_wq_create+0xb4/0x2a0 io_uring_setup+0x97f/0xaa0 ? io_remove_personalities+0x30/0x30 ? io_poll_trigger_evfd+0x30/0x30 do_syscall_64+0x5b/0x1c0 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f4d116cb1ed which is due to the 'wqe' and 'worker' allocation being node affine. But it isn't valid to call the node affine allocation if the node isn't online. Setup structures for even offline nodes, as usual, but skip them in terms of thread setup to not waste resources. If the node isn't online, just alloc memory with NUMA_NO_NODE. Reported-by: Glauber Costa <glauber@scylladb.com> Tested-by: Glauber Costa <glauber@scylladb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-11 21:30:06 +08:00
wqe->node = alloc_node;
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
io-wq: remove nr_process accounting We're now just using fork like we would from userspace, so there's no need to try and impose extra restrictions or accounting on the user side of things. That's already being done for us. That also means we don't have to pass in the user_struct anymore, that's correctly inherited through ->creds on fork. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-22 07:02:53 +08:00
wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
io-wq: add support for bounded vs unbunded work io_uring supports request types that basically have two different lifetimes: 1) Bounded completion time. These are requests like disk reads or writes, which we know will finish in a finite amount of time. 2) Unbounded completion time. These are generally networked IO, where we have no idea how long they will take to complete. Another example is POLL commands. This patch provides support for io-wq to handle these differently, so we don't starve bounded requests by tying up workers for too long. By default all work is bounded, unless otherwise specified in the work item. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-08 02:41:16 +08:00
task_rlimit(current, RLIMIT_NPROC);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
INIT_LIST_HEAD(&wqe->wait.entry);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
wqe->wait.func = io_wqe_hash_wake;
for (i = 0; i < IO_WQ_ACCT_NR; i++) {
struct io_wqe_acct *acct = &wqe->acct[i];
acct->index = i;
atomic_set(&acct->nr_running, 0);
INIT_WQ_LIST(&acct->work_list);
io-wq: decouple work_list protection from the big wqe->lock wqe->lock is abused, it now protects acct->work_list, hash stuff, nr_workers, wqe->free_list and so on. Lets first get the work_list out of the wqe-lock mess by introduce a specific lock for work list. This is the first step to solve the huge contension between work insertion and work consumption. good thing: - split locking for bound and unbound work list - reduce contension between work_list visit and (worker's)free_list. For the hash stuff, since there won't be a work with same file in both bound and unbound work list, thus they won't visit same hash entry. it works well to use the new lock to protect hash stuff. Results: set max_unbound_worker = 4, test with echo-server: nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16 (-n connection, -l workload) before this patch: Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074 Overhead Command Shared Object Symbol 28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath 6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock 2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work 1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue 1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker 1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule 0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked with this patch: Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943 Overhead Command Shared Object Symbol 16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock 4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat 2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work 2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave 2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work 1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock 1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker 1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39% TPS is similar, while cpu usage is from almost 400% to 350% Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:39 +08:00
raw_spin_lock_init(&acct->lock);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
}
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
wqe->wq = wq;
io_wq: Make io_wqe::lock a raw_spinlock_t During a context switch the scheduler invokes wq_worker_sleeping() with disabled preemption. Disabling preemption is needed because it protects access to `worker->sleeping'. As an optimisation it avoids invoking schedule() within the schedule path as part of possible wake up (thus preempt_enable_no_resched() afterwards). The io-wq has been added to the mix in the same section with disabled preemption. This breaks on PREEMPT_RT because io_wq_worker_sleeping() acquires a spinlock_t. Also within the schedule() the spinlock_t must be acquired after tsk_is_pi_blocked() otherwise it will block on the sleeping lock again while scheduling out. While playing with `io_uring-bench' I didn't notice a significant latency spike after converting io_wqe::lock to a raw_spinlock_t. The latency was more or less the same. In order to keep the spinlock_t it would have to be moved after the tsk_is_pi_blocked() check which would introduce a branch instruction into the hot path. The lock is used to maintain the `work_list' and wakes one task up at most. Should io_wqe_cancel_pending_work() cause latency spikes, while searching for a specific item, then it would need to drop the lock during iterations. revert_creds() is also invoked under the lock. According to debug cred::non_rcu is 0. Otherwise it should be moved outside of the locked section because put_cred_rcu()->free_uid() acquires a sleeping lock. Convert io_wqe::lock to a raw_spinlock_t.c Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-09-01 16:41:46 +08:00
raw_spin_lock_init(&wqe->lock);
io-wq: remove now redundant struct io_wq_nulls_list Since we don't iterate these lists anymore after commit: e61df66c69b1 ("io-wq: ensure free/busy list browsing see all items") we don't need to retain the nulls value we use for them. That means it's pretty pointless to wrap the hlist_nulls_head in a structure, so get rid of it. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 23:00:41 +08:00
INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
io-wq: ensure free/busy list browsing see all items We have two lists for workers in io-wq, a busy and a free list. For certain operations we want to browse all workers, and we currently do that by browsing the two separate lists. But since these lists are RCU protected, we can potentially miss workers if they move between the two lists while we're browsing them. Add a third list, all_list, that simply holds all workers. A worker is added to that list when it starts, and removed when it exits. This makes the worker iteration cleaner, too. Reported-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-14 04:54:49 +08:00
INIT_LIST_HEAD(&wqe->all_list);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
}
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
wq->task = get_task_struct(data->task);
atomic_set(&wq->worker_refs, 1);
init_completion(&wq->worker_done);
return wq;
io-wq: wait for io_wq_create() to setup necessary workers We currently have a race where if setup is really slow, we can be calling io_wq_destroy() before we're done setting up. This will cause the caller to get stuck waiting for the manager to set things up, but the manager already exited. Fix this by doing a sync setup of the manager. This also fixes the case where if we failed creating workers, we'd also get stuck. In practice this race window was really small, as we already wait for the manager to start. Hence someone would have to call io_wq_destroy() after the task has started, but before it started the first loop. The reported test case forked tons of these, which is why it became an issue. Reported-by: syzbot+0f1cc17f85154f400465@syzkaller.appspotmail.com Fixes: 771b53d033e8 ("io-wq: small threadpool implementation for io_uring") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-19 23:37:07 +08:00
err:
io-wq: fix error path leak of buffered write hash map The 'err' path should include the hash put, we already grabbed a reference once we get that far. Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-02 00:09:56 +08:00
io_wq_put_hash(data->hash);
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
for_each_node(node) {
if (!wq->wqes[node])
continue;
free_cpumask_var(wq->wqes[node]->cpu_mask);
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
kfree(wq->wqes[node]);
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
}
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
err_wq:
io-wq: wait for io_wq_create() to setup necessary workers We currently have a race where if setup is really slow, we can be calling io_wq_destroy() before we're done setting up. This will cause the caller to get stuck waiting for the manager to set things up, but the manager already exited. Fix this by doing a sync setup of the manager. This also fixes the case where if we failed creating workers, we'd also get stuck. In practice this race window was really small, as we already wait for the manager to start. Hence someone would have to call io_wq_destroy() after the task has started, but before it started the first loop. The reported test case forked tons of these, which is why it became an issue. Reported-by: syzbot+0f1cc17f85154f400465@syzkaller.appspotmail.com Fixes: 771b53d033e8 ("io-wq: small threadpool implementation for io_uring") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-19 23:37:07 +08:00
kfree(wq);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
return ERR_PTR(ret);
}
io-wq: cancel task_work on exit only targeting the current 'wq' With using task_work_cancel(), we're potentially canceling task_work that isn't related to this specific io_wq. Use the newly added task_work_cancel_match() to ensure that we only remove and cancel work items that are specific to this io_wq. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-02 09:57:07 +08:00
static bool io_task_work_match(struct callback_head *cb, void *data)
{
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
struct io_worker *worker;
io-wq: cancel task_work on exit only targeting the current 'wq' With using task_work_cancel(), we're potentially canceling task_work that isn't related to this specific io_wq. Use the newly added task_work_cancel_match() to ensure that we only remove and cancel work items that are specific to this io_wq. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-02 09:57:07 +08:00
io-wq: fix silly logic error in io_task_work_match() We check for the func with an OR condition, which means it always ends up being false and we never match the task_work we want to cancel. In the unexpected case that we do exit with that pending, we can trigger a hang waiting for a worker to exit, but it was never created. syzbot reports that as such: INFO: task syz-executor687:8514 blocked for more than 143 seconds. Not tainted 5.14.0-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:syz-executor687 state:D stack:27296 pid: 8514 ppid: 8479 flags:0x00024004 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0x940/0x26f0 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_wq_exit_workers fs/io-wq.c:1162 [inline] io_wq_put_and_exit+0x40c/0xc70 fs/io-wq.c:1197 io_uring_clean_tctx fs/io_uring.c:9607 [inline] io_uring_cancel_generic+0x5fe/0x740 fs/io_uring.c:9687 io_uring_files_cancel include/linux/io_uring.h:16 [inline] do_exit+0x265/0x2a30 kernel/exit.c:780 do_group_exit+0x125/0x310 kernel/exit.c:922 get_signal+0x47f/0x2160 kernel/signal.c:2868 arch_do_signal_or_restart+0x2a9/0x1c40 arch/x86/kernel/signal.c:865 handle_signal_work kernel/entry/common.c:148 [inline] exit_to_user_mode_loop kernel/entry/common.c:172 [inline] exit_to_user_mode_prepare+0x17d/0x290 kernel/entry/common.c:209 __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline] syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:302 do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x445cd9 RSP: 002b:00007fc657f4b308 EFLAGS: 00000246 ORIG_RAX: 00000000000000ca RAX: 0000000000000001 RBX: 00000000004cb448 RCX: 0000000000445cd9 RDX: 00000000000f4240 RSI: 0000000000000081 RDI: 00000000004cb44c RBP: 00000000004cb440 R08: 000000000000000e R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000246 R12: 000000000049b154 R13: 0000000000000003 R14: 00007fc657f4b400 R15: 0000000000022000 While in there, also decrement accr->nr_workers. This isn't strictly needed as we're exiting, but let's make sure the accounting matches up. Fixes: 3146cba99aa2 ("io-wq: make worker creation resilient against signals") Reported-by: syzbot+f62d3e0a4ea4f38f5326@syzkaller.appspotmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-09 09:57:26 +08:00
if (cb->func != create_worker_cb && cb->func != create_worker_cont)
io-wq: cancel task_work on exit only targeting the current 'wq' With using task_work_cancel(), we're potentially canceling task_work that isn't related to this specific io_wq. Use the newly added task_work_cancel_match() to ensure that we only remove and cancel work items that are specific to this io_wq. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-02 09:57:07 +08:00
return false;
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
worker = container_of(cb, struct io_worker, create_work);
return worker->wqe->wq == data;
io-wq: cancel task_work on exit only targeting the current 'wq' With using task_work_cancel(), we're potentially canceling task_work that isn't related to this specific io_wq. Use the newly added task_work_cancel_match() to ensure that we only remove and cancel work items that are specific to this io_wq. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-02 09:57:07 +08:00
}
io_uring/io-wq: close io-wq full-stop gap There is an old problem with io-wq cancellation where requests should be killed and are in io-wq but are not discoverable, e.g. in @next_hashed or @linked vars of io_worker_handle_work(). It adds some unreliability to individual request canellation, but also may potentially get __io_uring_cancel() stuck. For instance: 1) An __io_uring_cancel()'s cancellation round have not found any request but there are some as desribed. 2) __io_uring_cancel() goes to sleep 3) Then workers wake up and try to execute those hidden requests that happen to be unbound. As we already cancel all requests of io-wq there, set IO_WQ_BIT_EXIT in advance, so preventing 3) from executing unbound requests. The workers will initially break looping because of getting a signal as they are threads of the dying/exec()'ing user task. Cc: stable@vger.kernel.org Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/abfcf8c54cb9e8f7bfbad7e9a0cc5433cc70bdc2.1621781238.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-05-23 22:48:39 +08:00
void io_wq_exit_start(struct io_wq *wq)
{
set_bit(IO_WQ_BIT_EXIT, &wq->state);
}
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
static void io_wq_cancel_tw_create(struct io_wq *wq)
io-wq: provide an io_wq_put_and_exit() helper If we put the io-wq from io_uring, we really want it to exit. Provide a helper that does that for us. Couple that with not having the manager hold a reference to the 'wq' and the normal SQPOLL exit will tear down the io-wq context appropriate. On the io-wq side, our wq context is per task, so only the task itself is manipulating ->manager and hence it's safe to check and clear without any extra locking. We just need to ensure that the manager task stays around, in case it exits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 04:48:19 +08:00
{
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
struct callback_head *cb;
io-wq: cancel task_work on exit only targeting the current 'wq' With using task_work_cancel(), we're potentially canceling task_work that isn't related to this specific io_wq. Use the newly added task_work_cancel_match() to ensure that we only remove and cancel work items that are specific to this io_wq. Fixes: 685fe7feedb9 ("io-wq: eliminate the need for a manager thread") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-02 09:57:07 +08:00
while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
struct io_worker *worker;
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io-wq: remove GFP_ATOMIC allocation off schedule out path Daniel reports that the v5.14-rc4-rt4 kernel throws a BUG when running stress-ng: | [ 90.202543] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35 | [ 90.202549] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 2047, name: iou-wrk-2041 | [ 90.202555] CPU: 5 PID: 2047 Comm: iou-wrk-2041 Tainted: G W 5.14.0-rc4-rt4+ #89 | [ 90.202559] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 | [ 90.202561] Call Trace: | [ 90.202577] dump_stack_lvl+0x34/0x44 | [ 90.202584] ___might_sleep.cold+0x87/0x94 | [ 90.202588] rt_spin_lock+0x19/0x70 | [ 90.202593] ___slab_alloc+0xcb/0x7d0 | [ 90.202598] ? newidle_balance.constprop.0+0xf5/0x3b0 | [ 90.202603] ? dequeue_entity+0xc3/0x290 | [ 90.202605] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202610] ? pick_next_task_fair+0xb9/0x330 | [ 90.202612] ? __schedule+0x670/0x1410 | [ 90.202615] ? io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202618] kmem_cache_alloc_trace+0x79/0x1f0 | [ 90.202621] io_wqe_dec_running.isra.0+0x98/0xe0 | [ 90.202625] io_wq_worker_sleeping+0x37/0x50 | [ 90.202628] schedule+0x30/0xd0 | [ 90.202630] schedule_timeout+0x8f/0x1a0 | [ 90.202634] ? __bpf_trace_tick_stop+0x10/0x10 | [ 90.202637] io_wqe_worker+0xfd/0x320 | [ 90.202641] ? finish_task_switch.isra.0+0xd3/0x290 | [ 90.202644] ? io_worker_handle_work+0x670/0x670 | [ 90.202646] ? io_worker_handle_work+0x670/0x670 | [ 90.202649] ret_from_fork+0x22/0x30 which is due to the RT kernel not liking a GFP_ATOMIC allocation inside a raw spinlock. Besides that not working on RT, doing any kind of allocation from inside schedule() is kind of nasty and should be avoided if at all possible. This particular path happens when an io-wq worker goes to sleep, and we need a new worker to handle pending work. We currently allocate a small data item to hold the information we need to create a new worker, but we can instead include this data in the io_worker struct itself and just protect it with a single bit lock. We only really need one per worker anyway, as we will have run pending work between to sleep cycles. https://lore.kernel.org/lkml/20210804082418.fbibprcwtzyt5qax@beryllium.lan/ Reported-by: Daniel Wagner <dwagner@suse.de> Tested-by: Daniel Wagner <dwagner@suse.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-04 22:37:25 +08:00
worker = container_of(cb, struct io_worker, create_work);
io-wq: remove worker to owner tw dependency INFO: task iou-wrk-6609:6612 blocked for more than 143 seconds. Not tainted 5.15.0-rc5-syzkaller #0 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:iou-wrk-6609 state:D stack:27944 pid: 6612 ppid: 6526 flags:0x00004006 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xb44/0x5960 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_timeout+0x1db/0x2a0 kernel/time/timer.c:1857 do_wait_for_common kernel/sched/completion.c:85 [inline] __wait_for_common kernel/sched/completion.c:106 [inline] wait_for_common kernel/sched/completion.c:117 [inline] wait_for_completion+0x176/0x280 kernel/sched/completion.c:138 io_worker_exit fs/io-wq.c:183 [inline] io_wqe_worker+0x66d/0xc40 fs/io-wq.c:597 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 io-wq worker may submit a task_work to the master task and upon io_worker_exit() wait for the tw to get executed. The problem appears when the master task is waiting in coredump.c: 468 freezer_do_not_count(); 469 wait_for_completion(&core_state->startup); 470 freezer_count(); Apparently having some dependency on children threads getting everything stuck. Workaround it by cancelling the taks_work callback that causes it before going into io_worker_exit() waiting. p.s. probably a better option is to not submit tw elevating the refcount in the first place, but let's leave this excercise for the future. Cc: stable@vger.kernel.org Reported-and-tested-by: syzbot+27d62ee6f256b186883e@syzkaller.appspotmail.com Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/142a716f4ed936feae868959059154362bfa8c19.1635509451.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-29 20:11:33 +08:00
io_worker_cancel_cb(worker);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
}
io-wq: check for wq exit after adding new worker task_work We check IO_WQ_BIT_EXIT before attempting to create a new worker, and wq exit cancels pending work if we have any. But it's possible to have a race between the two, where creation checks exit finding it not set, but we're in the process of exiting. The exit side will cancel pending creation task_work, but there's a gap where we add task_work after we've canceled existing creations at exit time. Fix this by checking the EXIT bit post adding the creation task_work. If it's set, run the same cancelation that exit does. Reported-and-tested-by: syzbot+b60c982cb0efc5e05a47@syzkaller.appspotmail.com Reviewed-by: Hao Xu <haoxu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-12-10 23:29:30 +08:00
}
static void io_wq_exit_workers(struct io_wq *wq)
{
int node;
if (!wq->task)
return;
io_wq_cancel_tw_create(wq);
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
rcu_read_lock();
for_each_node(node) {
struct io_wqe *wqe = wq->wqes[node];
io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
io-wq: provide an io_wq_put_and_exit() helper If we put the io-wq from io_uring, we really want it to exit. Provide a helper that does that for us. Couple that with not having the manager hold a reference to the 'wq' and the normal SQPOLL exit will tear down the io-wq context appropriate. On the io-wq side, our wq context is per task, so only the task itself is manipulating ->manager and hence it's safe to check and clear without any extra locking. We just need to ensure that the manager task stays around, in case it exits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 04:48:19 +08:00
}
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
rcu_read_unlock();
io_worker_ref_put(wq);
wait_for_completion(&wq->worker_done);
io-wq: Fix UAF when wakeup wqe in hash waitqueue BUG: KASAN: use-after-free in __wake_up_common+0x637/0x650 Read of size 8 at addr ffff8880304250d8 by task iou-wrk-28796/28802 Call Trace: __dump_stack [inline] dump_stack+0x141/0x1d7 print_address_description.constprop.0.cold+0x5b/0x2c6 __kasan_report [inline] kasan_report.cold+0x7c/0xd8 __wake_up_common+0x637/0x650 __wake_up_common_lock+0xd0/0x130 io_worker_handle_work+0x9dd/0x1790 io_wqe_worker+0xb2a/0xd40 ret_from_fork+0x1f/0x30 Allocated by task 28798: kzalloc_node [inline] io_wq_create+0x3c4/0xdd0 io_init_wq_offload [inline] io_uring_alloc_task_context+0x1bf/0x6b0 __io_uring_add_task_file+0x29a/0x3c0 io_uring_add_task_file [inline] io_uring_install_fd [inline] io_uring_create [inline] io_uring_setup+0x209a/0x2bd0 do_syscall_64+0x3a/0xb0 entry_SYSCALL_64_after_hwframe+0x44/0xae Freed by task 28798: kfree+0x106/0x2c0 io_wq_destroy+0x182/0x380 io_wq_put [inline] io_wq_put_and_exit+0x7a/0xa0 io_uring_clean_tctx [inline] __io_uring_cancel+0x428/0x530 io_uring_files_cancel do_exit+0x299/0x2a60 do_group_exit+0x125/0x310 get_signal+0x47f/0x2150 arch_do_signal_or_restart+0x2a8/0x1eb0 handle_signal_work[inline] exit_to_user_mode_loop [inline] exit_to_user_mode_prepare+0x171/0x280 __syscall_exit_to_user_mode_work [inline] syscall_exit_to_user_mode+0x19/0x60 do_syscall_64+0x47/0xb0 entry_SYSCALL_64_after_hwframe There are the following scenarios, hash waitqueue is shared by io-wq1 and io-wq2. (note: wqe is worker) io-wq1:worker2 | locks bit1 io-wq2:worker1 | waits bit1 io-wq1:worker3 | waits bit1 io-wq1:worker2 | completes all wqe bit1 work items io-wq1:worker2 | drop bit1, exit io-wq2:worker1 | locks bit1 io-wq1:worker3 | can not locks bit1, waits bit1 and exit io-wq1 | exit and free io-wq1 io-wq2:worker1 | drops bit1 io-wq1:worker3 | be waked up, even though wqe is freed After all iou-wrk belonging to io-wq1 have exited, remove wqe form hash waitqueue, it is guaranteed that there will be no more wqe belonging to io-wq1 in the hash waitqueue. Reported-by: syzbot+6cb11ade52aa17095297@syzkaller.appspotmail.com Signed-off-by: Zqiang <qiang.zhang@windriver.com> Link: https://lore.kernel.org/r/20210526050826.30500-1-qiang.zhang@windriver.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-05-26 13:08:26 +08:00
for_each_node(node) {
spin_lock_irq(&wq->hash->wait.lock);
list_del_init(&wq->wqes[node]->wait.entry);
spin_unlock_irq(&wq->hash->wait.lock);
}
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
put_task_struct(wq->task);
wq->task = NULL;
io-wq: provide an io_wq_put_and_exit() helper If we put the io-wq from io_uring, we really want it to exit. Provide a helper that does that for us. Couple that with not having the manager hold a reference to the 'wq' and the normal SQPOLL exit will tear down the io-wq context appropriate. On the io-wq side, our wq context is per task, so only the task itself is manipulating ->manager and hence it's safe to check and clear without any extra locking. We just need to ensure that the manager task stays around, in case it exits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 04:48:19 +08:00
}
io-wq: improve manager/worker handling over exec exec will cancel any threads, including the ones that io-wq is using. This isn't a problem, in fact we'd prefer it to be that way since it means we know that any async work cancels naturally without having to handle it proactively. But it does mean that we need to setup a new manager, as the manager and workers are gone. Handle this at queue time, and cancel work if we fail. Since the manager can go away without us noticing, ensure that the manager itself holds a reference to the 'wq' as well. Rename io_wq_destroy() to io_wq_put() to reflect that. In the future we can now simplify exec cancelation handling, for now just leave it the same. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-26 01:17:09 +08:00
static void io_wq_destroy(struct io_wq *wq)
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
{
io-wq: fix handling of NUMA node IDs There are several things that can go wrong in the current code on NUMA systems, especially if not all nodes are online all the time: - If the identifiers of the online nodes do not form a single contiguous block starting at zero, wq->wqes will be too small, and OOB memory accesses will occur e.g. in the loop in io_wq_create(). - If a node comes online between the call to num_online_nodes() and the for_each_node() loop in io_wq_create(), an OOB write will occur. - If a node comes online between io_wq_create() and io_wq_enqueue(), a lookup is performed for an element that doesn't exist, and an OOB read will probably occur. Fix it by: - using nr_node_ids instead of num_online_nodes() for the allocation size; nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the highest node ID that could possibly come online at some point, even if those nodes' identifiers are not a contiguous block - creating workers for all possible CPUs, not just all online ones This is basically what the normal workqueue code also does, as far as I can tell. Signed-off-by: Jann Horn <jannh@google.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-11-27 02:10:20 +08:00
int node;
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
for_each_node(node) {
struct io_wqe *wqe = wq->wqes[node];
io-wq: fix race around pending work on teardown syzbot reports that it's triggering the warning condition on having pending work on shutdown: WARNING: CPU: 1 PID: 12346 at fs/io-wq.c:1061 io_wq_destroy fs/io-wq.c:1061 [inline] WARNING: CPU: 1 PID: 12346 at fs/io-wq.c:1061 io_wq_put+0x153/0x260 fs/io-wq.c:1072 Modules linked in: CPU: 1 PID: 12346 Comm: syz-executor.5 Not tainted 5.12.0-rc2-syzkaller #0 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:io_wq_destroy fs/io-wq.c:1061 [inline] RIP: 0010:io_wq_put+0x153/0x260 fs/io-wq.c:1072 Code: 8d e8 71 90 ea 01 49 89 c4 41 83 fc 40 7d 4f e8 33 4d 97 ff 42 80 7c 2d 00 00 0f 85 77 ff ff ff e9 7a ff ff ff e8 1d 4d 97 ff <0f> 0b eb b9 8d 6b ff 89 ee 09 de bf ff ff ff ff e8 18 51 97 ff 09 RSP: 0018:ffffc90001ebfb08 EFLAGS: 00010293 RAX: ffffffff81e16083 RBX: ffff888019038040 RCX: ffff88801e86b780 RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000040 RBP: 1ffff1100b2f8a80 R08: ffffffff81e15fce R09: ffffed100b2f8a82 R10: ffffed100b2f8a82 R11: 0000000000000000 R12: 0000000000000000 R13: dffffc0000000000 R14: ffff8880597c5400 R15: ffff888019038000 FS: 00007f8dcd89c700(0000) GS:ffff8880b9c00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 000055e9a054e160 CR3: 000000001dfb8000 CR4: 00000000001506f0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: io_uring_clean_tctx+0x1b7/0x210 fs/io_uring.c:8802 __io_uring_files_cancel+0x13c/0x170 fs/io_uring.c:8820 io_uring_files_cancel include/linux/io_uring.h:47 [inline] do_exit+0x258/0x2340 kernel/exit.c:780 do_group_exit+0x168/0x2d0 kernel/exit.c:922 get_signal+0x1734/0x1ef0 kernel/signal.c:2773 arch_do_signal_or_restart+0x3c/0x610 arch/x86/kernel/signal.c:811 handle_signal_work kernel/entry/common.c:147 [inline] exit_to_user_mode_loop kernel/entry/common.c:171 [inline] exit_to_user_mode_prepare+0xac/0x1e0 kernel/entry/common.c:208 __syscall_exit_to_user_mode_work kernel/entry/common.c:290 [inline] syscall_exit_to_user_mode+0x48/0x180 kernel/entry/common.c:301 entry_SYSCALL_64_after_hwframe+0x44/0xae RIP: 0033:0x465f69 which shouldn't happen, but seems to be possible due to a race on whether or not the io-wq manager sees a fatal signal first, or whether the io-wq workers do. If we race with queueing work and then send a fatal signal to the owning task, and the io-wq worker sees that before the manager sets IO_WQ_BIT_EXIT, then it's possible to have the worker exit and leave work behind. Just turn the WARN_ON_ONCE() into a cancelation condition instead. Reported-by: syzbot+77a738a6bc947bf639ca@syzkaller.appspotmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-26 00:16:12 +08:00
struct io_cb_cancel_data match = {
.fn = io_wq_work_match_all,
.cancel_all = true,
};
io_wqe_cancel_pending_work(wqe, &match);
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
free_cpumask_var(wqe->cpu_mask);
io-wq: make buffered file write hashed work map per-ctx Before the io-wq thread change, we maintained a hash work map and lock per-node per-ring. That wasn't ideal, as we really wanted it to be per ring. But now that we have per-task workers, the hash map ends up being just per-task. That'll work just fine for the normal case of having one task use a ring, but if you share the ring between tasks, then it's considerably worse than it was before. Make the hash map per ctx instead, which provides full per-ctx buffered write serialization on hashed writes. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-20 03:33:30 +08:00
kfree(wqe);
}
io_wq_put_hash(wq->hash);
io-wq: small threadpool implementation for io_uring This adds support for io-wq, a smaller and specialized thread pool implementation. This is meant to replace workqueues for io_uring. Among the reasons for this addition are: - We can assign memory context smarter and more persistently if we manage the life time of threads. - We can drop various work-arounds we have in io_uring, like the async_list. - We can implement hashed work insertion, to manage concurrency of buffered writes without needing a) an extra workqueue, or b) needlessly making the concurrency of said workqueue very low which hurts performance of multiple buffered file writers. - We can implement cancel through signals, for cancelling interruptible work like read/write (or send/recv) to/from sockets. - We need the above cancel for being able to assign and use file tables from a process. - We can implement a more thorough cancel operation in general. - We need it to move towards a syslet/threadlet model for even faster async execution. For that we need to take ownership of the used threads. This list is just off the top of my head. Performance should be the same, or better, at least that's what I've seen in my testing. io-wq supports basic NUMA functionality, setting up a pool per node. io-wq hooks up to the scheduler schedule in/out just like workqueue and uses that to drive the need for more/less workers. Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-10-23 00:25:58 +08:00
kfree(wq);
io-wq: improve manager/worker handling over exec exec will cancel any threads, including the ones that io-wq is using. This isn't a problem, in fact we'd prefer it to be that way since it means we know that any async work cancels naturally without having to handle it proactively. But it does mean that we need to setup a new manager, as the manager and workers are gone. Handle this at queue time, and cancel work if we fail. Since the manager can go away without us noticing, ensure that the manager itself holds a reference to the 'wq' as well. Rename io_wq_destroy() to io_wq_put() to reflect that. In the future we can now simplify exec cancelation handling, for now just leave it the same. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-26 01:17:09 +08:00
}
io-wq: provide an io_wq_put_and_exit() helper If we put the io-wq from io_uring, we really want it to exit. Provide a helper that does that for us. Couple that with not having the manager hold a reference to the 'wq' and the normal SQPOLL exit will tear down the io-wq context appropriate. On the io-wq side, our wq context is per task, so only the task itself is manipulating ->manager and hence it's safe to check and clear without any extra locking. We just need to ensure that the manager task stays around, in case it exits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 04:48:19 +08:00
void io_wq_put_and_exit(struct io_wq *wq)
{
io_uring/io-wq: close io-wq full-stop gap There is an old problem with io-wq cancellation where requests should be killed and are in io-wq but are not discoverable, e.g. in @next_hashed or @linked vars of io_worker_handle_work(). It adds some unreliability to individual request canellation, but also may potentially get __io_uring_cancel() stuck. For instance: 1) An __io_uring_cancel()'s cancellation round have not found any request but there are some as desribed. 2) __io_uring_cancel() goes to sleep 3) Then workers wake up and try to execute those hidden requests that happen to be unbound. As we already cancel all requests of io-wq there, set IO_WQ_BIT_EXIT in advance, so preventing 3) from executing unbound requests. The workers will initially break looping because of getting a signal as they are threads of the dying/exec()'ing user task. Cc: stable@vger.kernel.org Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/abfcf8c54cb9e8f7bfbad7e9a0cc5433cc70bdc2.1621781238.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-05-23 22:48:39 +08:00
WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
io-wq: eliminate the need for a manager thread io-wq relies on a manager thread to create/fork new workers, as needed. But there's really no strong need for it anymore. We have the following cases that fork a new worker: 1) Work queue. This is done from the task itself always, and it's trivial to create a worker off that path, if needed. 2) All workers have gone to sleep, and we have more work. This is called off the sched out path. For this case, use a task_work items to queue a fork-worker operation. 3) Hashed work completion. Don't think we need to do anything off this case. If need be, it could just use approach 2 as well. Part of this change is incrementing the running worker count before the fork, to avoid cases where we observe we need a worker and then queue creation of one. Then new work comes in, we fork a new one. That last queue operation should have waited for the previous worker to come up, it's quite possible we don't even need it. Hence move the worker running from before we fork it off to more efficiently handle that case. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-09 00:37:51 +08:00
io_wq_exit_workers(wq);
io-wq: remove unused io-wq refcounting iowq->refs is initialised to one and killed on exit, so it's not used and we can kill it. Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/401007393528ea7c102360e69a29b64498e15db2.1623634181.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-14 09:36:13 +08:00
io_wq_destroy(wq);
io-wq: provide an io_wq_put_and_exit() helper If we put the io-wq from io_uring, we really want it to exit. Provide a helper that does that for us. Couple that with not having the manager hold a reference to the 'wq' and the normal SQPOLL exit will tear down the io-wq context appropriate. On the io-wq side, our wq context is per task, so only the task itself is manipulating ->manager and hence it's safe to check and clear without any extra locking. We just need to ensure that the manager task stays around, in case it exits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-02-27 04:48:19 +08:00
}
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
struct online_data {
unsigned int cpu;
bool online;
};
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
{
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
struct online_data *od = data;
io-wq: Fix io_wq_worker_affinity() Do not include private headers and do not frob in internals. On top of that, while the previous code restores the affinity, it doesn't ensure the task actually moves there if it was running, leading to the fun situation that it can be observed running outside of its allowed mask for potentially significant time. Use the proper API instead. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lore.kernel.org/r/YG7QkiUzlEbW85TU@hirez.programming.kicks-ass.net Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-04-08 17:44:50 +08:00
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
if (od->online)
cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
else
cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
return false;
}
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
{
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
struct online_data od = {
.cpu = cpu,
.online = online
};
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
int i;
rcu_read_lock();
for_each_node(i)
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
rcu_read_unlock();
return 0;
}
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
{
struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
return __io_wq_cpu_online(wq, cpu, true);
}
static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
{
struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
return __io_wq_cpu_online(wq, cpu, false);
}
io_uring: allow user configurable IO thread CPU affinity io-wq defaults to per-node masks for IO workers. This works fine by default, but isn't particularly handy for workloads that prefer more specific affinities, for either performance or isolation reasons. This adds IORING_REGISTER_IOWQ_AFF that allows the user to pass in a CPU mask that is then applied to IO thread workers, and an IORING_UNREGISTER_IOWQ_AFF that simply resets the masks back to the default of per-node. Note that no care is given to existing IO threads, they will need to go through a reschedule before the affinity is correct if they are already running or sleeping. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:19:54 +08:00
int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
{
int i;
rcu_read_lock();
for_each_node(i) {
struct io_wqe *wqe = wq->wqes[i];
if (mask)
cpumask_copy(wqe->cpu_mask, mask);
else
cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
}
rcu_read_unlock();
return 0;
}
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
/*
* Set max number of unbounded workers, returns old value. If new_count is 0,
* then just return the old value.
*/
int io_wq_max_workers(struct io_wq *wq, int *new_count)
{
io-wq: fix max-workers not correctly set on multi-node system In io-wq.c:io_wq_max_workers(), new_count[] was changed right after each node's value was set. This caused the following node getting the setting of the previous one. Returned values are copied from node 0. Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Beld Zhang <beldzhang@gmail.com> [axboe: minor fixups] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-03 02:32:08 +08:00
int prev[IO_WQ_ACCT_NR];
bool first_node = true;
int i, node;
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
io-wq: provide IO_WQ_* constants for IORING_REGISTER_IOWQ_MAX_WORKERS arg items The items passed in the array pointed by the arg parameter of IORING_REGISTER_IOWQ_MAX_WORKERS io_uring_register operation carry certain semantics: they refer to different io-wq worker categories; provide IO_WQ_* constants in the UAPI, so these categories can be referenced in the user space code. Suggested-by: Jens Axboe <axboe@kernel.dk> Complements: 2e480058ddc21ec5 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com> Link: https://lore.kernel.org/r/20210913154415.GA12890@asgard.redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-13 23:44:15 +08:00
BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
io-wq: use IO_WQ_ACCT_NR rather than hardcoded number It's better to use the defined enum stuff not the hardcoded number to define array. Signed-off-by: Hao Xu <haoxu@linux.alibaba.com> Link: https://lore.kernel.org/r/20220206095241.121485-4-haoxu@linux.alibaba.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-06 17:52:41 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++) {
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
new_count[i] = task_rlimit(current, RLIMIT_NPROC);
}
io-wq: fix max-workers not correctly set on multi-node system In io-wq.c:io_wq_max_workers(), new_count[] was changed right after each node's value was set. This caused the following node getting the setting of the previous one. Returned values are copied from node 0. Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Beld Zhang <beldzhang@gmail.com> [axboe: minor fixups] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-03 02:32:08 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++)
prev[i] = 0;
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
rcu_read_lock();
for_each_node(node) {
io-wq: max_worker fixes First, fix nr_workers checks against max_workers, with max_worker registration, it may pretty easily happen that nr_workers > max_workers. Also, synchronise writing to acct->max_worker with wqe->lock. It's not an actual problem, but as we don't care about io_wqe_create_worker(), it's better than WRITE_ONCE()/READ_ONCE(). Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Reported-by: Beld Zhang <beldzhang@gmail.com> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/11f90e6b49410b7d1a88f5d04fb8d95bb86b8cf3.1634671835.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-20 03:31:26 +08:00
struct io_wqe *wqe = wq->wqes[node];
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
struct io_wqe_acct *acct;
io-wq: max_worker fixes First, fix nr_workers checks against max_workers, with max_worker registration, it may pretty easily happen that nr_workers > max_workers. Also, synchronise writing to acct->max_worker with wqe->lock. It's not an actual problem, but as we don't care about io_wqe_create_worker(), it's better than WRITE_ONCE()/READ_ONCE(). Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Reported-by: Beld Zhang <beldzhang@gmail.com> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/11f90e6b49410b7d1a88f5d04fb8d95bb86b8cf3.1634671835.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-20 03:31:26 +08:00
raw_spin_lock(&wqe->lock);
io-wq: split bounded and unbounded work into separate lists We've got a few issues that all boil down to the fact that we have one list of pending work items, yet two different types of workers to serve them. This causes some oddities around workers switching type and even hashed work vs regular work on the same bounded list. Just separate them out cleanly, similarly to how we already do accounting of what is running. That provides a clean separation and removes some corner cases that can cause stalls when handling IO that is punted to io-wq. Fixes: ecc53c48c13d ("io-wq: check max_worker limits if a worker transitions bound state") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-09-01 03:57:32 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++) {
io-wq: max_worker fixes First, fix nr_workers checks against max_workers, with max_worker registration, it may pretty easily happen that nr_workers > max_workers. Also, synchronise writing to acct->max_worker with wqe->lock. It's not an actual problem, but as we don't care about io_wqe_create_worker(), it's better than WRITE_ONCE()/READ_ONCE(). Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Reported-by: Beld Zhang <beldzhang@gmail.com> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/11f90e6b49410b7d1a88f5d04fb8d95bb86b8cf3.1634671835.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-20 03:31:26 +08:00
acct = &wqe->acct[i];
io-wq: fix max-workers not correctly set on multi-node system In io-wq.c:io_wq_max_workers(), new_count[] was changed right after each node's value was set. This caused the following node getting the setting of the previous one. Returned values are copied from node 0. Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Beld Zhang <beldzhang@gmail.com> [axboe: minor fixups] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-03 02:32:08 +08:00
if (first_node)
prev[i] = max_t(int, acct->max_workers, prev[i]);
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
if (new_count[i])
acct->max_workers = new_count[i];
}
io-wq: max_worker fixes First, fix nr_workers checks against max_workers, with max_worker registration, it may pretty easily happen that nr_workers > max_workers. Also, synchronise writing to acct->max_worker with wqe->lock. It's not an actual problem, but as we don't care about io_wqe_create_worker(), it's better than WRITE_ONCE()/READ_ONCE(). Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Reported-by: Beld Zhang <beldzhang@gmail.com> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Link: https://lore.kernel.org/r/11f90e6b49410b7d1a88f5d04fb8d95bb86b8cf3.1634671835.git.asml.silence@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-10-20 03:31:26 +08:00
raw_spin_unlock(&wqe->lock);
io-wq: fix max-workers not correctly set on multi-node system In io-wq.c:io_wq_max_workers(), new_count[] was changed right after each node's value was set. This caused the following node getting the setting of the previous one. Returned values are copied from node 0. Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Beld Zhang <beldzhang@gmail.com> [axboe: minor fixups] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-03 02:32:08 +08:00
first_node = false;
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
}
rcu_read_unlock();
io-wq: fix max-workers not correctly set on multi-node system In io-wq.c:io_wq_max_workers(), new_count[] was changed right after each node's value was set. This caused the following node getting the setting of the previous one. Returned values are copied from node 0. Fixes: 2e480058ddc2 ("io-wq: provide a way to limit max number of workers") Signed-off-by: Beld Zhang <beldzhang@gmail.com> [axboe: minor fixups] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-03 02:32:08 +08:00
for (i = 0; i < IO_WQ_ACCT_NR; i++)
new_count[i] = prev[i];
io-wq: provide a way to limit max number of workers io-wq divides work into two categories: 1) Work that completes in a bounded time, like reading from a regular file or a block device. This type of work is limited based on the size of the SQ ring. 2) Work that may never complete, we call this unbounded work. The amount of workers here is just limited by RLIMIT_NPROC. For various uses cases, it's handy to have the kernel limit the maximum amount of pending workers for both categories. Provide a way to do with with a new IORING_REGISTER_IOWQ_MAX_WORKERS operation. IORING_REGISTER_IOWQ_MAX_WORKERS takes an array of two integers and sets the max worker count to what is being passed in for each category. The old values are returned into that same array. If 0 is being passed in for either category, it simply returns the current value. The value is capped at RLIMIT_NPROC. This actually isn't that important as it's more of a hint, if we're exceeding the value then our attempt to fork a new worker will fail. This happens naturally already if more than one node is in the system, as these values are per-node internally for io-wq. Reported-by: Johannes Lundberg <johalun0@gmail.com> Link: https://github.com/axboe/liburing/issues/420 Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-08-28 01:33:19 +08:00
return 0;
}
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
static __init int io_wq_init(void)
{
int ret;
ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
io-wq: use private CPU mask In preparation for allowing user specific CPU masks for IO thread creation, switch to using a mask embedded in the per-node wqe structure. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-06-18 00:08:11 +08:00
io_wq_cpu_online, io_wq_cpu_offline);
io-wq: re-set NUMA node affinities if CPUs come online We correctly set io-wq NUMA node affinities when the io-wq context is setup, but if an entire node CPU set is offlined and then brought back online, the per node affinities are broken. Ensure that we set them again whenever a CPU comes online. This ensures that we always track the right node affinity. The usual cpuhp notifiers are used to drive it. Reported-by: Zhang Qiang <qiang.zhang@windriver.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-10-22 23:02:50 +08:00
if (ret < 0)
return ret;
io_wq_online = ret;
return 0;
}
subsys_initcall(io_wq_init);
Copy Permalink