Provided buffers allow an application to supply io_uring with buffers
that can then be grabbed for a read/receive request, when the data
source is ready to deliver data. The existing scheme relies on using
IORING_OP_PROVIDE_BUFFERS to do that, but it can be difficult to use
in real world applications. It's pretty efficient if the application
is able to supply back batches of provided buffers when they have been
consumed and the application is ready to recycle them, but if
fragmentation occurs in the buffer space, it can become difficult to
supply enough buffers at the time. This hurts efficiency.
Add a register op, IORING_REGISTER_PBUF_RING, which allows an application
to setup a shared queue for each buffer group of provided buffers. The
application can then supply buffers simply by adding them to this ring,
and the kernel can consume then just as easily. The ring shares the head
with the application, the tail remains private in the kernel.
Provided buffers setup with IORING_REGISTER_PBUF_RING cannot use
IORING_OP_{PROVIDE,REMOVE}_BUFFERS for adding or removing entries to the
ring, they must use the mapped ring. Mapped provided buffer rings can
co-exist with normal provided buffers, just not within the same group ID.
To gauge overhead of the existing scheme and evaluate the mapped ring
approach, a simple NOP benchmark was written. It uses a ring of 128
entries, and submits/completes 32 at the time. 'Replenish' is how
many buffers are provided back at the time after they have been
consumed:
Test Replenish NOPs/sec
================================================================
No provided buffers NA ~30M
Provided buffers 32 ~16M
Provided buffers 1 ~10M
Ring buffers 32 ~27M
Ring buffers 1 ~27M
The ring mapped buffers perform almost as well as not using provided
buffers at all, and they don't care if you provided 1 or more back at
the same time. This means application can just replenish as they go,
rather than need to batch and compact, further reducing overhead in the
application. The NOP benchmark above doesn't need to do any compaction,
so that overhead isn't even reflected in the above test.
Co-developed-by: Dylan Yudaken <dylany@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Abstract this out from io_sqe_buffer_register() so we can use it
elsewhere too without duplicating this code.
No intended functional changes in this patch.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Obviously not really useful since it's not transferring data, but it
is helpful in benchmarking overhead of provided buffers.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_provided_buffer_select() must drop the submit lock, if needed, even
in the error handling case. Failure to do so will leave us with the
ctx->uring_lock held, causing spew like:
====================================
WARNING: iou-wrk-366/368 still has locks held!
5.18.0-rc6-00294-gdf8dc7004331 #994 Not tainted
------------------------------------
1 lock held by iou-wrk-366/368:
#0: ffff0000c72598a8 (&ctx->uring_lock){+.+.}-{3:3}, at: io_ring_submit_lock+0x20/0x48
stack backtrace:
CPU: 4 PID: 368 Comm: iou-wrk-366 Not tainted 5.18.0-rc6-00294-gdf8dc7004331 #994
Hardware name: linux,dummy-virt (DT)
Call trace:
dump_backtrace.part.0+0xa4/0xd4
show_stack+0x14/0x5c
dump_stack_lvl+0x88/0xb0
dump_stack+0x14/0x2c
debug_check_no_locks_held+0x84/0x90
try_to_freeze.isra.0+0x18/0x44
get_signal+0x94/0x6ec
io_wqe_worker+0x1d8/0x2b4
ret_from_fork+0x10/0x20
and triggering later hangs off get_signal() because we attempt to
re-grab the lock.
Reported-by: syzbot+987d7bb19195ae45208c@syzkaller.appspotmail.com
Fixes: 149c69b04a ("io_uring: abstract out provided buffer list selection")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Refactor io_accept() to support multishot mode.
theoretical analysis:
1) when connections come in fast
- singleshot:
add accept sqe(userspace) --> accept inline
^ |
|-----------------|
- multishot:
add accept sqe(userspace) --> accept inline
^ |
|--*--|
we do accept repeatedly in * place until get EAGAIN
2) when connections come in at a low pressure
similar thing like 1), we reduce a lot of userspace-kernel context
switch and useless vfs_poll()
tests:
Did some tests, which goes in this way:
server client(multiple)
accept connect
read write
write read
close close
Basically, raise up a number of clients(on same machine with server) to
connect to the server, and then write some data to it, the server will
write those data back to the client after it receives them, and then
close the connection after write return. Then the client will read the
data and then close the connection. Here I test 10000 clients connect
one server, data size 128 bytes. And each client has a go routine for
it, so they come to the server in short time.
test 20 times before/after this patchset, time spent:(unit cycle, which
is the return value of clock())
before:
1930136+1940725+1907981+1947601+1923812+1928226+1911087+1905897+1941075
+1934374+1906614+1912504+1949110+1908790+1909951+1941672+1969525+1934984
+1934226+1914385)/20.0 = 1927633.75
after:
1858905+1917104+1895455+1963963+1892706+1889208+1874175+1904753+1874112
+1874985+1882706+1884642+1864694+1906508+1916150+1924250+1869060+1889506
+1871324+1940803)/20.0 = 1894750.45
(1927633.75 - 1894750.45) / 1927633.75 = 1.65%
Signed-off-by: Hao Xu <howeyxu@tencent.com>
Link: https://lore.kernel.org/r/20220514142046.58072-5-haoxu.linux@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For operations like accept, multishot is a useful feature, since we can
reduce a number of accept sqe. Let's integrate it to fast poll, it may
be good for other operations in the future.
Signed-off-by: Hao Xu <howeyxu@tencent.com>
Link: https://lore.kernel.org/r/20220514142046.58072-4-haoxu.linux@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add a flag to indicate multishot mode for fast poll. currently only
accept use it, but there may be more operations leveraging it in the
future. Also add a mask IO_APOLL_MULTI_POLLED which stands for
REQ_F_APOLL_MULTI | REQ_F_POLLED, to make the code short and cleaner.
Signed-off-by: Hao Xu <howeyxu@tencent.com>
Link: https://lore.kernel.org/r/20220514142046.58072-3-haoxu.linux@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The check for waking up a request compares the poll_t bits, however this
will always contain some common flags so this always wakes up.
For files with single wait queues such as sockets this can cause the
request to be sent to the async worker unnecesarily. Further if it is
non-blocking will complete the request with EAGAIN which is not desired.
Here exclude these common events, making sure to not exclude POLLERR which
might be important.
Fixes: d7718a9d25 ("io_uring: use poll driven retry for files that support it")
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220512091834.728610-3-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If an opcode handler semi-reliably returns -EAGAIN, io_wq_submit_work()
might continue busily hammer the same handler over and over again, which
is not ideal. The -EAGAIN handling in question was put there only for
IOPOLL, so restrict it to IOPOLL mode only where there is no other
recourse than to retry as we cannot wait.
Fixes: def596e955 ("io_uring: support for IO polling")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/f168b4f24181942f3614dd8ff648221736f572e6.1652433740.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently to setup a fully sparse descriptor space upfront, the app needs
to alloate an array of the full size and memset it to -1 and then pass
that in. Make this a bit easier by allowing a flag that simply does
this internally rather than needing to copy each slot separately.
This works with IORING_REGISTER_FILES2 as the flag is set in struct
io_uring_rsrc_register, and is only allow when the type is
IORING_RSRC_FILE as this doesn't make sense for registered buffers.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently limit these to 32K, but since we're now backing the table
space with vmalloc when needed, there's no reason why we can't make it
bigger. The total space is limited by RLIMIT_NOFILE as well.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the application passes in IORING_FILE_INDEX_ALLOC as the file_slot,
then that's a hint to allocate a fixed file descriptor rather than have
one be passed in directly.
This can be useful for having io_uring manage the direct descriptor space,
and also allows multi-shot support to work with fixed files.
Normal accept direct requests will complete with 0 for success, and < 0
in case of error. If io_uring is asked to allocated the direct descriptor,
then the direct descriptor is returned in case of success.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the application passes in IORING_FILE_INDEX_ALLOC as the file_slot,
then that's a hint to allocate a fixed file descriptor rather than have
one be passed in directly.
This can be useful for having io_uring manage the direct descriptor space.
Normal open direct requests will complete with 0 for success, and < 0
in case of error. If io_uring is asked to allocated the direct descriptor,
then the direct descriptor is returned in case of success.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Applications currently always pick where they want fixed files to go.
In preparation for allowing these types of commands with multishot
support, add a basic allocator in the fixed file table.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In preparation for adding a basic allocator for direct descriptors,
add helpers that set/clear whether a file slot is used.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's not needed as the REQ_F_BUFFER_SELECTED flag tracks the state of
whether or not kbuf is valid, so just drop it.
Suggested-by: Dylan Yudaken <dylany@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have io_kiocb->buf_index which is used for either fixed buffers, or
for provided buffers. For the latter, it's used to hold the buffer group
ID for buffer selection. Post selection, req->kbuf->bid is used to get
the buffer ID.
Store the buffer ID, when selected, in req->buf_index. If we do end up
recycling the buffer, reset it back to the buffer group ID.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
These are mutually exclusive - if you use provided buffers, then you
cannot use fixed buffers and vice versa. Move them into the same spot
in the io_kiocb, which is also advantageous for provided buffers as
they get near the submit side hot cacheline.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Callers already have room to store the addr and length information,
clean it up by having the caller just assign the previously provided
data.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use a plain array for any group ID that's less than 64, and punt
anything beyond that to an xarray. 64 fits in a page even for 4KB
page sizes and with the planned additions.
This makes the expected group usage faster by avoiding a hash and lookup
to find our list, and it uses less memory upfront by not allocating any
memory for provided buffers unless it's actually being used.
Suggested-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The read/write opcodes use it already, but the recv/recvmsg do not. If
we switch them over and read and validate this at init time while we're
checking if the opcode supports it anyway, then we can do it in one spot
and we don't have to pass in a separate group ID for io_buffer_select().
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There's no point in validity checking buf_index if the request doesn't
have REQ_F_BUFFER_SELECT set, as we will never use it for that case.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
After the recent changes, this is direct call to io_buffer_select()
anyway. With this change, there are no wrappers left for provided
buffer selection.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For all of send/sendmsg and recv/recvmsg we have the local 'sr' variable,
yet some cases still use req->sr_msg which sr points to. Use 'sr'
consistently.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If IORING_RECVSEND_POLL_FIRST is set for recv/recvmsg or send/sendmsg,
then we arm poll first rather than attempt a receive or send upfront.
This can be useful if we expect there to be no data (or space) available
for the request, as we can then avoid wasting time on the initial
issue attempt.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Don't punt this check to the op prep handlers, add the support to
io_op_defs and we can check them while setting up the request.
This reduces the text size by 500 bytes on aarch64, and makes this less
fragile by having the check in one spot and needing opcodes to opt in
to IOPOLL or ioprio support.
Reviewed-by: Hao Xu <howeyxu@tencent.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The IORING_SQ_NEED_WAKEUP flag is now set using atomic_or() which
implies a full barrier on some architectures but it is not required to
do so. Use the more appropriate smp_mb__after_atomic() which avoids the
extra barrier on those architectures.
Signed-off-by: Almog Khaikin <almogkh@gmail.com>
Link: https://lore.kernel.org/r/20220426163403.112692-1-almogkh@gmail.com
Fixes: 8018823e6987 ("io_uring: serialize ctx->rings->sq_flags with atomic_or/and")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If IORING_SETUP_COOP_TASKRUN is set to use cooperative scheduling for
running task_work, then IORING_SETUP_TASKRUN_FLAG can be set so the
application can tell if task_work is pending in the kernel for this
ring. This allows use cases like io_uring_peek_cqe() to still function
appropriately, or for the task to know when it would be useful to
call io_uring_wait_cqe() to run pending events.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20220426014904.60384-7-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If this is set, io_uring will never use an IPI to deliver a task_work
notification. This can be used in the common case where a single task or
thread communicates with the ring, and doesn't rely on
io_uring_cqe_peek().
This provides a noticeable win in performance, both from eliminating
the IPI itself, but also from avoiding interrupting the submitting
task unnecessarily.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20220426014904.60384-6-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While doing so, switch SQPOLL to TWA_SIGNAL_NO_IPI as well, as that
just does a task wakeup and then we can remove the special wakeup we
have in task_work_add.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20220426014904.60384-5-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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>
Rather than require ctx->completion_lock for ensuring that we don't
clobber the flags, use the atomic bitop helpers instead. This removes
the need to grab the completion_lock, in preparation for needing to set
or clear sq_flags when we don't know the status of this lock.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20220426014904.60384-3-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If IO_URING_SCM_ALL isn't set, as it would not be on 32-bit builds,
then we trigger a warning:
fs/io_uring.c: In function '__io_sqe_files_unregister':
fs/io_uring.c:8992:13: warning: unused variable 'i' [-Wunused-variable]
8992 | int i;
| ^
Move the ifdef up to include the 'i' variable declaration.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Fixes: 5e45690a1c ("io_uring: store SCM state in io_fixed_file->file_ptr")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Right now io_uring will not actively inform userspace if a CQE is
dropped. This is extremely rare, requiring a CQ ring overflow, as well as
a GFP_ATOMIC kmalloc failure. However the consequences could cause for
example applications to go into an undefined state, possibly waiting for a
CQE that never arrives.
Return an error code (EBADR) in these cases. Since this is expected to be
incredibly rare, try and avoid as much as possible affecting the hot code
paths, and so it only is returned lazily and when there is no other
available CQEs.
Once the error is returned, reset the error condition assuming the user is
either ok with it or will clean up appropriately.
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220421091345.2115755-6-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring_enter returns the count submitted preferrably over an error
code. In some code paths this check is not required, so reorganise the
code so that the check is only done as needed.
This is also a prep for returning error codes only in waiting scenarios.
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220421091345.2115755-4-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Trace cqe overflows in io_uring. Print ocqe before the check, so if it is
NULL it indicates that it has been dropped.
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220421091345.2115755-3-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently check REQ_F_POLLED before arming async poll for a
notification to retry. If it's set, then we don't allow poll and will
punt to io-wq instead. This is done to prevent a situation where a buggy
driver will repeatedly return that there's space/data available yet we
get -EAGAIN.
However, if we already transferred data, then it should be safe to rely
on poll again. Gate the check on whether or not REQ_F_PARTIAL_IO is
also set.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Like commit 7ba89d2af1 for recv/recvmsg, support MSG_WAITALL for the
send side. If this flag is set and we do a short send, retry for a
stream of seqpacket socket.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than match on a specific key, be it user_data or file, allow
canceling any request that we can lookup. Works like
IORING_ASYNC_CANCEL_ALL in that it cancels multiple requests, but it
doesn't key off user_data or the file.
Can't be set with IORING_ASYNC_CANCEL_FD, as that's a key selector.
Only one may be used at the time.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20220418164402.75259-6-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently sqe->addr must contain the user_data of the request being
canceled. Introduce the IORING_ASYNC_CANCEL_FD flag, which tells the
kernel that we're keying off the file fd instead for cancelation. This
allows canceling any request that a) uses a file, and b) was assigned the
file based on the value being passed in.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20220418164402.75259-5-axboe@kernel.dk
The current cancelation will lookup and cancel the first request it
finds based on the key passed in. Add a flag that allows to cancel any
request that matches they key. It completes with the number of requests
found and canceled, or res < 0 if an error occured.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20220418164402.75259-4-axboe@kernel.dk
In preparation for being able to not only key cancel off the user_data,
pass in the io_cancel_data struct for the various functions that deal
with request cancelation.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Link: https://lore.kernel.org/r/20220418164402.75259-3-axboe@kernel.dk
Move ->timeout_lock grabbing inside of io_timeout_cancel(), so
we can do io_req_task_queue_fail() outside of the lock. It's much nicer
than relying on triple nested locking.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/cde758c2897930d31e205ed8f476d4ec879a8849.1650458197.git.asml.silence@gmail.com
[axboe: drop now wrong timeout_lock annotation]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
