lwq avoids using back pointers in lists, and uses less locking.
This introduces a new spinlock, but the other one will be removed in a
future patch.
For svc_clean_up_xprts(), we now dequeue the entire queue, walk it to
remove and process the xprts that need cleaning up, then re-enqueue the
remaining queue.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Currently if several items of work become available in quick succession,
that number of threads (if available) will be woken. By the time some
of them wake up another thread that was already cache-warm might have
come along and completed the work. Anecdotal evidence suggests as many
as 15% of wakes find nothing to do once they get to the point of
looking.
This patch changes svc_pool_wake_idle_thread() to wake the first thread
on the queue but NOT remove it. Subsequent calls will wake the same
thread. Once that thread starts it will dequeue itself and after
dequeueing some work to do, it will wake the next thread if there is more
work ready. This results in a more orderly increase in the number of
busy threads.
As a bonus, this allows us to reduce locking around the idle queue.
svc_pool_wake_idle_thread() no longer needs to take a lock (beyond
rcu_read_lock()) as it doesn't manipulate the queue, it just looks at
the first item.
The thread itself can avoid locking by using the new
llist_del_first_this() interface. This will safely remove the thread
itself if it is the head. If it isn't the head, it will do nothing.
If multiple threads call this concurrently only one will succeed. The
others will do nothing, so no corruption can result.
If a thread wakes up and finds that it cannot dequeue itself that means
either
- that it wasn't woken because it was the head of the queue. Maybe the
freezer woke it. In that case it can go back to sleep (after trying
to freeze of course).
- some other thread found there was nothing to do very recently, and
placed itself on the head of the queue in front of this thread.
It must check again after placing itself there, so it can be deemed to
be responsible for any pending work, and this thread can go back to
sleep until woken.
No code ever tests for busy threads any more. Only each thread itself
cares if it is busy. So svc_thread_busy() is no longer needed.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Functions which directly manipulate a 'struct rqst', such as
svc_rqst_alloc() or svc_rqst_release_pages(), can reasonably
have "rqst" in there name.
However functions that act on the running thread, such as
XX_should_sleep() or XX_wait_for_work() should seem more
natural with a "svc_thread_" prefix.
So make those changes.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
With an llist we don't need to take a lock to add a thread to the list,
though we still need a lock to remove it. That will go in the next
patch.
Unlike double-linked lists, a thread cannot reliably remove itself from
the list. Only the first thread can be removed, and that can change
asynchronously. So some care is needed.
We already check if there is pending work to do, so we are unlikely to
add ourselves to the idle list and then want to remove ourselves again.
If we DO find something needs to be done after adding ourselves to the
list, we simply wake up the first thread on the list. If that was us,
we successfully removed ourselves and can continue. If it was some
other thread, they will do the work that needs to be done. We can
safely sleep until woken.
We also remove the test on freezing() from rqst_should_sleep(). Instead
we set TASK_FREEZABLE before scheduling. This makes is safe to
schedule() when a freeze is pending. As we now loop waiting to be
removed from the idle queue, this is a cleaner way to handle freezing.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
We can tell if a pool is congested by checking if the idle list is
empty. We don't need a separate flag.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Rather than searching a list of threads to find an idle one, having a
list of idle threads allows an idle thread to be found immediately.
This adds some spin_lock calls which is not ideal, but as the hold-time
is tiny it is still faster than searching a list. A future patch will
remove them using llist.h. This involves some subtlety and so is left
to a separate patch.
This removes the need for the RQ_BUSY flag. The rqst is "busy"
precisely when it is not on the "idle" list.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc threads are currently stopped using kthread_stop(). This requires
identifying a specific thread. However we don't care which thread
stops, just as long as one does.
So instead, set a flag in the svc_pool to say that a thread needs to
die, and have each thread check this flag instead of calling
kthread_should_stop(). The first thread to find and clear this flag
then moves towards exiting.
This removes an explicit dependency on sp_all_threads which will make a
future patch simpler.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Using svc_recv() for (NFSv4.1) back-channel handling means we have just
one mechanism for waking threads.
Also change kthread_freezable_should_stop() in nfs4_callback_svc() to
kthread_should_stop() as used elsewhere.
kthread_freezable_should_stop() effectively adds a try_to_freeze() call,
and svc_recv() already contains that at an appropriate place.
Signed-off-by: NeilBrown <neilb@suse.de>
Cc: Trond Myklebust <trond.myklebust@hammerspace.com>
Cc: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_get_next_xprt() does a lot more than just get an xprt. It also
decides if it needs to sleep, depending not only on the availability of
xprts but also on the need to exit or handle external work.
So rename it to svc_rqst_wait_for_work() and only do the testing and
waiting. Move all the waiting-related code out of svc_recv() into the
new svc_rqst_wait_for_work().
Move the dequeueing code out of svc_get_next_xprt() into svc_recv().
Previously svc_xprt_dequeue() would be called twice, once before waiting
and possibly once after. Now instead rqst_should_sleep() is called
twice. Once to decide if waiting is needed, and once to check against
after setting the task state do see if we might have missed a wakeup.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_xprt_handle() does lots of things itself, but leaves some to the
caller - svc_recv(). This isn't elegant.
Move that code out of svc_recv() into svc_xprt_handle()
Move the calls to svc_xprt_release() from svc_send() and svc_drop()
(the two possible final steps in svc_process()) and from svc_recv() (in
the case where svc_process() wasn't called) into svc_xprt_handle().
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Based on its name you would think that rqst_should_sleep() would be
read-only, not changing anything. But in fact it will clear
SP_TASK_PENDING if that was set. This is surprising, and it blurs the
line between "check for work to do" and "dequeue work to do".
So change the "test_and_clear" to simple "test" and clear the bit once
the thread has decided to wake up and return to the caller.
With this, it makes sense to *always* set SP_TASK_PENDING when asked,
rather than to set it only if no thread could be woken up.
[ cel: Previously TASK_PENDING indicated there is work waiting but no
idle threads were found to pick up that work. After this patch, it acts
as an XPT_BUSY flag for wake-ups that have no associated xprt. ]
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_xprt_enqueue() can be costly, since it involves selecting and
waking up a process.
More than one enqueue is done per incoming RPC. For example,
svc_data_ready() enqueues, and so does svc_xprt_receive(). Also, if
an RPC message requires more than one call to ->recvfrom() to
receive it fully, each one of those calls does an enqueue.
To get a sense of the average number of transport enqueue operations
needed to process an incoming RPC message, re-use the "packets" pool
stat. Track the number of complete RPC messages processed by each
thread pool.
Reviewed-by: Jeff Layton <jlayton@redhat.com>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Refactor: Extract the loop that finds an idle service thread from
svc_xprt_enqueue() and svc_wake_up(). Both functions do just about
the same thing.
Note that svc_wake_up() currently does not hold the RCU read lock
while waking the target thread. It indeed should hold the lock, just
as svc_xprt_enqueue() does, to ensure the rqstp does not vanish
during the wake-up. This patch adds the RCU lock for svc_wake_up().
Note that shrinking the pool thread count is rare, and calls to
svc_wake_up() are also quite infrequent. In practice, this race is
very unlikely to be hit, so we are not marking the lock fix for
stable backport at this time.
Reviewed-by: Jeff Layton <jlayton@redhat.com>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The xpt_flags field frequently changes between the time that
svc_xprt_ready() grabs a copy and execution flow arrives at the
tracepoint at the tail of svc_xprt_enqueue(). In fact, there's
usually a sleep/wake-up in there, so those flags are almost
guaranteed to be different.
It would be more useful to record the exact flags that were used to
decide whether the transport is ready, so move the tracepoint.
Moving it means the tracepoint can't pick up the waker's pid. That
can be added to struct svc_rqst if it turns out that is important.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Most svc threads have no interest in a timeout.
nfsd sets it to 1 hour, but this is a wart of no significance.
lockd uses the timeout so that it can call nlmsvc_retry_blocked().
It also sometimes calls svc_wake_up() to ensure this is called.
So change lockd to be consistent and always use svc_wake_up() to trigger
nlmsvc_retry_blocked() - using a timer instead of a timeout to
svc_recv().
And change svc_recv() to not take a timeout arg.
This makes the sp_threads_timedout counter always zero.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_recv() currently returns a 0 on success or one of two errors:
- -EAGAIN means no message was successfully received
- -EINTR means the thread has been told to stop
Previously nfsd would stop as the result of a signal as well as
following kthread_stop(). In that case the difference was useful: EINTR
means stop unconditionally. EAGAIN means stop if kthread_should_stop(),
continue otherwise.
Now threads only exit when kthread_should_stop() so we don't need the
distinction.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
All callers of svc_recv() go on to call svc_process() on success.
Simplify callers by having svc_recv() do that for them.
This loses one call to validate_process_creds() in nfsd. That was
debugging code added 14 years ago. I don't think we need to keep it.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The original implementation of nfsd used signals to stop threads during
shutdown.
In Linux 2.3.46pre5 nfsd gained the ability to shutdown threads
internally it if was asked to run "0" threads. After this user-space
transitioned to using "rpc.nfsd 0" to stop nfsd and sending signals to
threads was no longer an important part of the API.
In commit 3ebdbe5203 ("SUNRPC: discard svo_setup and rename
svc_set_num_threads_sync()") (v5.17-rc1~75^2~41) we finally removed the
use of signals for stopping threads, using kthread_stop() instead.
This patch makes the "obvious" next step and removes the ability to
signal nfsd threads - or any svc threads. nfsd stops allowing signals
and we don't check for their delivery any more.
This will allow for some simplification in later patches.
A change worth noting is in nfsd4_ssc_setup_dul(). There was previously
a signal_pending() check which would only succeed when the thread was
being shut down. It should really have tested kthread_should_stop() as
well. Now it just does the latter, not the former.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Remove a couple of dprintk call sites that are of little value.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Acked-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The preceding block comment before svc_register_xprt_class() is
not related to that function.
While we're here, add proper documenting comments for these two
publicly-visible functions.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Acked-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Micro-optimization: Call ktime_get() only when ->xpo_recvfrom() has
given us a full RPC message to process. rq_stime isn't used
otherwise, so this avoids pointless work.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Acked-by: Tom Talpey <tom@talpey.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_init_buffer() is careful to allocate the initial set of server
thread buffer pages from memory on the local NUMA node.
svc_alloc_arg() should also be that careful.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
When enabled, this dprintk() fires for every incoming RPC, which is
an enormous amount of log traffic. These days, after the first few
hundred log messages, the system journald is just going to mute it,
along with all other NFSD debug output.
Let's rely on trace points for this high-traffic information
instead.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Since the ->xprt_ctxt pointer was added to svc_deferred_req, it has not
been sufficient to use kfree() to free a deferred request. We may need
to free the ctxt as well.
As freeing the ctxt is all that ->xpo_release_rqst() does, we repurpose
it to explicit do that even when the ctxt is not stored in an rqst.
So we now have ->xpo_release_ctxt() which is given an xprt and a ctxt,
which may have been taken either from an rqst or from a dreq. The
caller is now responsible for clearing that pointer after the call to
->xpo_release_ctxt.
We also clear dr->xprt_ctxt when the ctxt is moved into a new rqst when
revisiting a deferred request. This ensures there is only one pointer
to the ctxt, so the risk of double freeing in future is reduced. The
new code in svc_xprt_release which releases both the ctxt and any
rq_deferred depends on this.
Fixes: 773f91b2cf ("SUNRPC: Fix NFSD's request deferral on RDMA transports")
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
When an RPC request is deferred, the rq_xprt_ctxt pointer is moved out
of the svc_rqst into the svc_deferred_req.
When the deferred request is revisited, the pointer is copied into
the new svc_rqst - and also remains in the svc_deferred_req.
In the (rare?) case that the request is deferred a second time, the old
svc_deferred_req is reused - it still has all the correct content.
However in that case the rq_xprt_ctxt pointer is NOT cleared so that
when xpo_release_xprt is called, the ctxt is freed (UDP) or possible
added to a free list (RDMA).
When the deferred request is revisited for a second time, it will
reference this ctxt which may be invalid, and the free the object a
second time which is likely to oops.
So change svc_defer() to *always* clear rq_xprt_ctxt, and assert that
the value is now stored in the svc_deferred_req.
Fixes: 773f91b2cf ("SUNRPC: Fix NFSD's request deferral on RDMA transports")
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This patch adds opportunitistic RPC-with-TLS to the Linux in-kernel
NFS server. If the client requests RPC-with-TLS and the user space
handshake agent is running, the server will set up a TLS session.
There are no policy settings yet. For example, the server cannot
yet require the use of RPC-with-TLS to access its data.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
This is an eye-catcher for tracepoints that record the XID: it means
svc_rqst() has not received a full RPC Call with an XID yet.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
A single RPC transaction that touches only a couple of pages means
rq_pvec will not be even close to full in svc_xpt_release(). This is
a common case.
Instead, just leave the pages in rq_pvec until it is completely
full. This improves the efficiency of the batch release mechanism
on workloads that involve small RPC messages.
The rq_pvec is also fully emptied just before thread exit.
Reviewed-by: Calum Mackay <calum.mackay@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean-up: There doesn't seem to be a reason why this function is
stuck in a header. One thing it prevents is the convenient addition
of tracing. Moving it to a source file also makes the rq_respages
clean-up logic easier to find.
Reviewed-by: Calum Mackay <calum.mackay@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean up: All callers of svc_process() ignore its return value, so
svc_process() can safely be converted to return void. Ditto for
svc_send().
The return value of ->xpo_sendto() is now used only as part of a
trace event.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
There's no need for the cost of this extra virtual function call
during every RPC transaction: the RQ_SECURE bit can be set properly
in ->xpo_recvfrom() instead.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Eliminate the use of bus-locked operations in svc_xprt_enqueue(),
which is a hot path. Replace them with per-cpu variables to reduce
cross-CPU memory bus traffic.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Now the entire RPC Call header parsing path is handled via struct
xdr_stream-based decoders.
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
The premise that "Once an svc thread is scheduled and executing an
RPC, no other processes will touch svc_rqst::rq_flags" is false.
svc_xprt_enqueue() examines the RQ_BUSY flag in scheduled nfsd
threads when determining which thread to wake up next.
Found via KCSAN.
Fixes: 28df098881 ("SUNRPC: Use RMW bitops in single-threaded hot paths")
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Record not only the number of pages requested, but the number of
pages that were actually allocated, to get a measure of progress
(or lack thereof).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
I noticed CPU pipeline stalls while using perf.
Once an svc thread is scheduled and executing an RPC, no other
processes will touch svc_rqst::rq_flags. Thus bus-locked atomics are
not needed outside the svc thread scheduler.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Clean up: There is one caller. The @cpu argument can be made
implicit now that a get_cpu/put_cpu pair is no longer needed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
svc_xprt_enqueue() disables preemption via get_cpu() and then asks
for a pool of a specific CPU (current) via svc_pool_for_cpu().
While preemption is disabled, svc_xprt_enqueue() acquires
svc_pool::sp_lock with bottom-halfs disabled, which can sleep on
PREEMPT_RT.
Disabling preemption is not required here. The pool is protected with a
lock so the following list access is safe even cross-CPU. The following
iteration through svc_pool::sp_all_threads is under RCU-readlock and
remaining operations within the loop are atomic and do not rely on
disabled-preemption.
Use raw_smp_processor_id() as the argument for the requested CPU in
svc_pool_for_cpu().
Reported-by: Mike Galbraith <umgwanakikbuti@gmail.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Trond Myklebust reports an NFSD crash in svc_rdma_sendto(). Further
investigation shows that the crash occurred while NFSD was handling
a deferred request.
This patch addresses two inter-related issues that prevent request
deferral from working correctly for RPC/RDMA requests:
1. Prevent the crash by ensuring that the original
svc_rqst::rq_xprt_ctxt value is available when the request is
revisited. Otherwise svc_rdma_sendto() does not have a Receive
context available with which to construct its reply.
2. Possibly since before commit 71641d99ce ("svcrdma: Properly
compute .len and .buflen for received RPC Calls"),
svc_rdma_recvfrom() did not include the transport header in the
returned xdr_buf. There should have been no need for svc_defer()
and friends to save and restore that header, as of that commit.
This issue is addressed in a backport-friendly way by simply
having svc_rdma_recvfrom() set rq_xprt_hlen to zero
unconditionally, just as svc_tcp_recvfrom() does. This enables
svc_deferred_recv() to correctly reconstruct an RPC message
received via RPC/RDMA.
Reported-by: Trond Myklebust <trondmy@hammerspace.com>
Link: https://lore.kernel.org/linux-nfs/82662b7190f26fb304eb0ab1bb04279072439d4e.camel@hammerspace.com/
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Cc: <stable@vger.kernel.org>
Core
----
- Introduce XDP multi-buffer support, allowing the use of XDP with
jumbo frame MTUs and combination with Rx coalescing offloads (LRO).
- Speed up netns dismantling (5x) and lower the memory cost a little.
Remove unnecessary per-netns sockets. Scope some lists to a netns.
Cut down RCU syncing. Use batch methods. Allow netdev registration
to complete out of order.
- Support distinguishing timestamp types (ingress vs egress) and
maintaining them across packet scrubbing points (e.g. redirect).
- Continue the work of annotating packet drop reasons throughout
the stack.
- Switch netdev error counters from an atomic to dynamically
allocated per-CPU counters.
- Rework a few preempt_disable(), local_irq_save() and busy waiting
sections problematic on PREEMPT_RT.
- Extend the ref_tracker to allow catching use-after-free bugs.
BPF
---
- Introduce "packing allocator" for BPF JIT images. JITed code is
marked read only, and used to be allocated at page granularity.
Custom allocator allows for more efficient memory use, lower
iTLB pressure and prevents identity mapping huge pages from
getting split.
- Make use of BTF type annotations (e.g. __user, __percpu) to enforce
the correct probe read access method, add appropriate helpers.
- Convert the BPF preload to use light skeleton and drop
the user-mode-driver dependency.
- Allow XDP BPF_PROG_RUN test infra to send real packets, enabling
its use as a packet generator.
- Allow local storage memory to be allocated with GFP_KERNEL if called
from a hook allowed to sleep.
- Introduce fprobe (multi kprobe) to speed up mass attachment (arch
bits to come later).
- Add unstable conntrack lookup helpers for BPF by using the BPF
kfunc infra.
- Allow cgroup BPF progs to return custom errors to user space.
- Add support for AF_UNIX iterator batching.
- Allow iterator programs to use sleepable helpers.
- Support JIT of add, and, or, xor and xchg atomic ops on arm64.
- Add BTFGen support to bpftool which allows to use CO-RE in kernels
without BTF info.
- Large number of libbpf API improvements, cleanups and deprecations.
Protocols
---------
- Micro-optimize UDPv6 Tx, gaining up to 5% in test on dummy netdev.
- Adjust TSO packet sizes based on min_rtt, allowing very low latency
links (data centers) to always send full-sized TSO super-frames.
- Make IPv6 flow label changes (AKA hash rethink) more configurable,
via sysctl and setsockopt. Distinguish between server and client
behavior.
- VxLAN support to "collect metadata" devices to terminate only
configured VNIs. This is similar to VLAN filtering in the bridge.
- Support inserting IPv6 IOAM information to a fraction of frames.
- Add protocol attribute to IP addresses to allow identifying where
given address comes from (kernel-generated, DHCP etc.)
- Support setting socket and IPv6 options via cmsg on ping6 sockets.
- Reject mis-use of ECN bits in IP headers as part of DSCP/TOS.
Define dscp_t and stop taking ECN bits into account in fib-rules.
- Add support for locked bridge ports (for 802.1X).
- tun: support NAPI for packets received from batched XDP buffs,
doubling the performance in some scenarios.
- IPv6 extension header handling in Open vSwitch.
- Support IPv6 control message load balancing in bonding, prevent
neighbor solicitation and advertisement from using the wrong port.
Support NS/NA monitor selection similar to existing ARP monitor.
- SMC
- improve performance with TCP_CORK and sendfile()
- support auto-corking
- support TCP_NODELAY
- MCTP (Management Component Transport Protocol)
- add user space tag control interface
- I2C binding driver (as specified by DMTF DSP0237)
- Multi-BSSID beacon handling in AP mode for WiFi.
- Bluetooth:
- handle MSFT Monitor Device Event
- add MGMT Adv Monitor Device Found/Lost events
- Multi-Path TCP:
- add support for the SO_SNDTIMEO socket option
- lots of selftest cleanups and improvements
- Increase the max PDU size in CAN ISOTP to 64 kB.
Driver API
----------
- Add HW counters for SW netdevs, a mechanism for devices which
offload packet forwarding to report packet statistics back to
software interfaces such as tunnels.
- Select the default NIC queue count as a fraction of number of
physical CPU cores, instead of hard-coding to 8.
- Expose devlink instance locks to drivers. Allow device layer of
drivers to use that lock directly instead of creating their own
which always runs into ordering issues in devlink callbacks.
- Add header/data split indication to guide user space enabling
of TCP zero-copy Rx.
- Allow configuring completion queue event size.
- Refactor page_pool to enable fragmenting after allocation.
- Add allocation and page reuse statistics to page_pool.
- Improve Multiple Spanning Trees support in the bridge to allow
reuse of topologies across VLANs, saving HW resources in switches.
- DSA (Distributed Switch Architecture):
- replay and offload of host VLAN entries
- offload of static and local FDB entries on LAG interfaces
- FDB isolation and unicast filtering
New hardware / drivers
----------------------
- Ethernet:
- LAN937x T1 PHYs
- Davicom DM9051 SPI NIC driver
- Realtek RTL8367S, RTL8367RB-VB switch and MDIO
- Microchip ksz8563 switches
- Netronome NFP3800 SmartNICs
- Fungible SmartNICs
- MediaTek MT8195 switches
- WiFi:
- mt76: MediaTek mt7916
- mt76: MediaTek mt7921u USB adapters
- brcmfmac: Broadcom BCM43454/6
- Mobile:
- iosm: Intel M.2 7360 WWAN card
Drivers
-------
- Convert many drivers to the new phylink API built for split PCS
designs but also simplifying other cases.
- Intel Ethernet NICs:
- add TTY for GNSS module for E810T device
- improve AF_XDP performance
- GTP-C and GTP-U filter offload
- QinQ VLAN support
- Mellanox Ethernet NICs (mlx5):
- support xdp->data_meta
- multi-buffer XDP
- offload tc push_eth and pop_eth actions
- Netronome Ethernet NICs (nfp):
- flow-independent tc action hardware offload (police / meter)
- AF_XDP
- Other Ethernet NICs:
- at803x: fiber and SFP support
- xgmac: mdio: preamble suppression and custom MDC frequencies
- r8169: enable ASPM L1.2 if system vendor flags it as safe
- macb/gem: ZynqMP SGMII
- hns3: add TX push mode
- dpaa2-eth: software TSO
- lan743x: multi-queue, mdio, SGMII, PTP
- axienet: NAPI and GRO support
- Mellanox Ethernet switches (mlxsw):
- source and dest IP address rewrites
- RJ45 ports
- Marvell Ethernet switches (prestera):
- basic routing offload
- multi-chain TC ACL offload
- NXP embedded Ethernet switches (ocelot & felix):
- PTP over UDP with the ocelot-8021q DSA tagging protocol
- basic QoS classification on Felix DSA switch using dcbnl
- port mirroring for ocelot switches
- Microchip high-speed industrial Ethernet (sparx5):
- offloading of bridge port flooding flags
- PTP Hardware Clock
- Other embedded switches:
- lan966x: PTP Hardward Clock
- qca8k: mdio read/write operations via crafted Ethernet packets
- Qualcomm 802.11ax WiFi (ath11k):
- add LDPC FEC type and 802.11ax High Efficiency data in radiotap
- enable RX PPDU stats in monitor co-exist mode
- Intel WiFi (iwlwifi):
- UHB TAS enablement via BIOS
- band disablement via BIOS
- channel switch offload
- 32 Rx AMPDU sessions in newer devices
- MediaTek WiFi (mt76):
- background radar detection
- thermal management improvements on mt7915
- SAR support for more mt76 platforms
- MBSSID and 6 GHz band on mt7915
- RealTek WiFi:
- rtw89: AP mode
- rtw89: 160 MHz channels and 6 GHz band
- rtw89: hardware scan
- Bluetooth:
- mt7921s: wake on Bluetooth, SCO over I2S, wide-band-speed (WBS)
- Microchip CAN (mcp251xfd):
- multiple RX-FIFOs and runtime configurable RX/TX rings
- internal PLL, runtime PM handling simplification
- improve chip detection and error handling after wakeup
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
-----BEGIN PGP SIGNATURE-----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=ELpe
-----END PGP SIGNATURE-----
Merge tag 'net-next-5.18' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next
Pull networking updates from Jakub Kicinski:
"The sprinkling of SPI drivers is because we added a new one and Mark
sent us a SPI driver interface conversion pull request.
Core
----
- Introduce XDP multi-buffer support, allowing the use of XDP with
jumbo frame MTUs and combination with Rx coalescing offloads (LRO).
- Speed up netns dismantling (5x) and lower the memory cost a little.
Remove unnecessary per-netns sockets. Scope some lists to a netns.
Cut down RCU syncing. Use batch methods. Allow netdev registration
to complete out of order.
- Support distinguishing timestamp types (ingress vs egress) and
maintaining them across packet scrubbing points (e.g. redirect).
- Continue the work of annotating packet drop reasons throughout the
stack.
- Switch netdev error counters from an atomic to dynamically
allocated per-CPU counters.
- Rework a few preempt_disable(), local_irq_save() and busy waiting
sections problematic on PREEMPT_RT.
- Extend the ref_tracker to allow catching use-after-free bugs.
BPF
---
- Introduce "packing allocator" for BPF JIT images. JITed code is
marked read only, and used to be allocated at page granularity.
Custom allocator allows for more efficient memory use, lower iTLB
pressure and prevents identity mapping huge pages from getting
split.
- Make use of BTF type annotations (e.g. __user, __percpu) to enforce
the correct probe read access method, add appropriate helpers.
- Convert the BPF preload to use light skeleton and drop the
user-mode-driver dependency.
- Allow XDP BPF_PROG_RUN test infra to send real packets, enabling
its use as a packet generator.
- Allow local storage memory to be allocated with GFP_KERNEL if
called from a hook allowed to sleep.
- Introduce fprobe (multi kprobe) to speed up mass attachment (arch
bits to come later).
- Add unstable conntrack lookup helpers for BPF by using the BPF
kfunc infra.
- Allow cgroup BPF progs to return custom errors to user space.
- Add support for AF_UNIX iterator batching.
- Allow iterator programs to use sleepable helpers.
- Support JIT of add, and, or, xor and xchg atomic ops on arm64.
- Add BTFGen support to bpftool which allows to use CO-RE in kernels
without BTF info.
- Large number of libbpf API improvements, cleanups and deprecations.
Protocols
---------
- Micro-optimize UDPv6 Tx, gaining up to 5% in test on dummy netdev.
- Adjust TSO packet sizes based on min_rtt, allowing very low latency
links (data centers) to always send full-sized TSO super-frames.
- Make IPv6 flow label changes (AKA hash rethink) more configurable,
via sysctl and setsockopt. Distinguish between server and client
behavior.
- VxLAN support to "collect metadata" devices to terminate only
configured VNIs. This is similar to VLAN filtering in the bridge.
- Support inserting IPv6 IOAM information to a fraction of frames.
- Add protocol attribute to IP addresses to allow identifying where
given address comes from (kernel-generated, DHCP etc.)
- Support setting socket and IPv6 options via cmsg on ping6 sockets.
- Reject mis-use of ECN bits in IP headers as part of DSCP/TOS.
Define dscp_t and stop taking ECN bits into account in fib-rules.
- Add support for locked bridge ports (for 802.1X).
- tun: support NAPI for packets received from batched XDP buffs,
doubling the performance in some scenarios.
- IPv6 extension header handling in Open vSwitch.
- Support IPv6 control message load balancing in bonding, prevent
neighbor solicitation and advertisement from using the wrong port.
Support NS/NA monitor selection similar to existing ARP monitor.
- SMC
- improve performance with TCP_CORK and sendfile()
- support auto-corking
- support TCP_NODELAY
- MCTP (Management Component Transport Protocol)
- add user space tag control interface
- I2C binding driver (as specified by DMTF DSP0237)
- Multi-BSSID beacon handling in AP mode for WiFi.
- Bluetooth:
- handle MSFT Monitor Device Event
- add MGMT Adv Monitor Device Found/Lost events
- Multi-Path TCP:
- add support for the SO_SNDTIMEO socket option
- lots of selftest cleanups and improvements
- Increase the max PDU size in CAN ISOTP to 64 kB.
Driver API
----------
- Add HW counters for SW netdevs, a mechanism for devices which
offload packet forwarding to report packet statistics back to
software interfaces such as tunnels.
- Select the default NIC queue count as a fraction of number of
physical CPU cores, instead of hard-coding to 8.
- Expose devlink instance locks to drivers. Allow device layer of
drivers to use that lock directly instead of creating their own
which always runs into ordering issues in devlink callbacks.
- Add header/data split indication to guide user space enabling of
TCP zero-copy Rx.
- Allow configuring completion queue event size.
- Refactor page_pool to enable fragmenting after allocation.
- Add allocation and page reuse statistics to page_pool.
- Improve Multiple Spanning Trees support in the bridge to allow
reuse of topologies across VLANs, saving HW resources in switches.
- DSA (Distributed Switch Architecture):
- replay and offload of host VLAN entries
- offload of static and local FDB entries on LAG interfaces
- FDB isolation and unicast filtering
New hardware / drivers
----------------------
- Ethernet:
- LAN937x T1 PHYs
- Davicom DM9051 SPI NIC driver
- Realtek RTL8367S, RTL8367RB-VB switch and MDIO
- Microchip ksz8563 switches
- Netronome NFP3800 SmartNICs
- Fungible SmartNICs
- MediaTek MT8195 switches
- WiFi:
- mt76: MediaTek mt7916
- mt76: MediaTek mt7921u USB adapters
- brcmfmac: Broadcom BCM43454/6
- Mobile:
- iosm: Intel M.2 7360 WWAN card
Drivers
-------
- Convert many drivers to the new phylink API built for split PCS
designs but also simplifying other cases.
- Intel Ethernet NICs:
- add TTY for GNSS module for E810T device
- improve AF_XDP performance
- GTP-C and GTP-U filter offload
- QinQ VLAN support
- Mellanox Ethernet NICs (mlx5):
- support xdp->data_meta
- multi-buffer XDP
- offload tc push_eth and pop_eth actions
- Netronome Ethernet NICs (nfp):
- flow-independent tc action hardware offload (police / meter)
- AF_XDP
- Other Ethernet NICs:
- at803x: fiber and SFP support
- xgmac: mdio: preamble suppression and custom MDC frequencies
- r8169: enable ASPM L1.2 if system vendor flags it as safe
- macb/gem: ZynqMP SGMII
- hns3: add TX push mode
- dpaa2-eth: software TSO
- lan743x: multi-queue, mdio, SGMII, PTP
- axienet: NAPI and GRO support
- Mellanox Ethernet switches (mlxsw):
- source and dest IP address rewrites
- RJ45 ports
- Marvell Ethernet switches (prestera):
- basic routing offload
- multi-chain TC ACL offload
- NXP embedded Ethernet switches (ocelot & felix):
- PTP over UDP with the ocelot-8021q DSA tagging protocol
- basic QoS classification on Felix DSA switch using dcbnl
- port mirroring for ocelot switches
- Microchip high-speed industrial Ethernet (sparx5):
- offloading of bridge port flooding flags
- PTP Hardware Clock
- Other embedded switches:
- lan966x: PTP Hardward Clock
- qca8k: mdio read/write operations via crafted Ethernet packets
- Qualcomm 802.11ax WiFi (ath11k):
- add LDPC FEC type and 802.11ax High Efficiency data in radiotap
- enable RX PPDU stats in monitor co-exist mode
- Intel WiFi (iwlwifi):
- UHB TAS enablement via BIOS
- band disablement via BIOS
- channel switch offload
- 32 Rx AMPDU sessions in newer devices
- MediaTek WiFi (mt76):
- background radar detection
- thermal management improvements on mt7915
- SAR support for more mt76 platforms
- MBSSID and 6 GHz band on mt7915
- RealTek WiFi:
- rtw89: AP mode
- rtw89: 160 MHz channels and 6 GHz band
- rtw89: hardware scan
- Bluetooth:
- mt7921s: wake on Bluetooth, SCO over I2S, wide-band-speed (WBS)
- Microchip CAN (mcp251xfd):
- multiple RX-FIFOs and runtime configurable RX/TX rings
- internal PLL, runtime PM handling simplification
- improve chip detection and error handling after wakeup"
* tag 'net-next-5.18' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (2521 commits)
llc: fix netdevice reference leaks in llc_ui_bind()
drivers: ethernet: cpsw: fix panic when interrupt coaleceing is set via ethtool
ice: don't allow to run ice_send_event_to_aux() in atomic ctx
ice: fix 'scheduling while atomic' on aux critical err interrupt
net/sched: fix incorrect vlan_push_eth dest field
net: bridge: mst: Restrict info size queries to bridge ports
net: marvell: prestera: add missing destroy_workqueue() in prestera_module_init()
drivers: net: xgene: Fix regression in CRC stripping
net: geneve: add missing netlink policy and size for IFLA_GENEVE_INNER_PROTO_INHERIT
net: dsa: fix missing host-filtered multicast addresses
net/mlx5e: Fix build warning, detected write beyond size of field
iwlwifi: mvm: Don't fail if PPAG isn't supported
selftests/bpf: Fix kprobe_multi test.
Revert "rethook: x86: Add rethook x86 implementation"
Revert "arm64: rethook: Add arm64 rethook implementation"
Revert "powerpc: Add rethook support"
Revert "ARM: rethook: Add rethook arm implementation"
netdevice: add missing dm_private kdoc
net: bridge: mst: prevent NULL deref in br_mst_info_size()
selftests: forwarding: Use same VRF for port and VLAN upper
...
Clean up: svc_shutdown_net() now does nothing but call
svc_close_net(). Replace all external call sites.
svc_close_net() is renamed to be the inverse of svc_xprt_create().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Neil says:
"These functions were separated in commit 0971374e28 ("SUNRPC:
Reduce contention in svc_xprt_enqueue()") so that the XPT_BUSY check
happened before taking any spinlocks.
We have since moved or removed the spinlocks so the extra test is
fairly pointless."
I've made this a separate patch in case the XPT_BUSY change has
unexpected consequences and needs to be reverted.
Suggested-by: Neil Brown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
We have never been able to track down and address the underlying
cause of the performance issues with workqueue-based service
support. svo_enqueue_xprt is called multiple times per RPC, so
it adds instruction path length, but always ends up at the same
function: svc_xprt_do_enqueue(). We do not anticipate needing
this flexibility for dynamic nfsd thread management support.
As a micro-optimization, remove .svo_enqueue_xprt because
Spectre/Meltdown makes virtual function calls more costly.
This change essentially reverts commit b9e13cdfac ("nfsd/sunrpc:
turn enqueueing a svc_xprt into a svc_serv operation").
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
struct svc_xprt holds a long lived reference to a netns,
it is worth tracking it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- Bruce steps down as NFSD maintainer
- Prepare for dynamic nfsd thread management
- More work on supporting re-exporting NFS mounts
- One fs/locks patch on behalf of Jeff Layton
Notable bug fixes:
- Fix zero-length NFSv3 WRITEs
- Fix directory cinfo on FS's that do not support iversion
- Fix WRITE verifiers for stable writes
- Fix crash on COPY_NOTIFY with a special state ID
-----BEGIN PGP SIGNATURE-----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=/kX3
-----END PGP SIGNATURE-----
Merge tag 'nfsd-5.17' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux
Pull nfsd updates from Chuck Lever:
"Bruce has announced he is leaving Red Hat at the end of the month and
is stepping back from his role as NFSD co-maintainer. As a result,
this includes a patch removing him from the MAINTAINERS file.
There is one patch in here that Jeff Layton was carrying in the locks
tree. Since he had only one for this cycle, he asked us to send it to
you via the nfsd tree.
There continues to be 0-day reports from Robert Morris @MIT. This time
we include a fix for a crash in the COPY_NOTIFY operation.
Highlights:
- Bruce steps down as NFSD maintainer
- Prepare for dynamic nfsd thread management
- More work on supporting re-exporting NFS mounts
- One fs/locks patch on behalf of Jeff Layton
Notable bug fixes:
- Fix zero-length NFSv3 WRITEs
- Fix directory cinfo on FS's that do not support iversion
- Fix WRITE verifiers for stable writes
- Fix crash on COPY_NOTIFY with a special state ID"
* tag 'nfsd-5.17' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux: (51 commits)
SUNRPC: Fix sockaddr handling in svcsock_accept_class trace points
SUNRPC: Fix sockaddr handling in the svc_xprt_create_error trace point
fs/locks: fix fcntl_getlk64/fcntl_setlk64 stub prototypes
nfsd: fix crash on COPY_NOTIFY with special stateid
MAINTAINERS: remove bfields
NFSD: Move fill_pre_wcc() and fill_post_wcc()
Revert "nfsd: skip some unnecessary stats in the v4 case"
NFSD: Trace boot verifier resets
NFSD: Rename boot verifier functions
NFSD: Clean up the nfsd_net::nfssvc_boot field
NFSD: Write verifier might go backwards
nfsd: Add a tracepoint for errors in nfsd4_clone_file_range()
NFSD: De-duplicate net_generic(nf->nf_net, nfsd_net_id)
NFSD: De-duplicate net_generic(SVC_NET(rqstp), nfsd_net_id)
NFSD: Clean up nfsd_vfs_write()
nfsd: Replace use of rwsem with errseq_t
NFSD: Fix verifier returned in stable WRITEs
nfsd: Retry once in nfsd_open on an -EOPENSTALE return
nfsd: Add errno mapping for EREMOTEIO
nfsd: map EBADF
...
Various places in the kernel - largely in filesystems - respond to a
memory allocation failure by looping around and re-trying. Some of
these cannot conveniently use __GFP_NOFAIL, for reasons such as:
- a GFP_ATOMIC allocation, which __GFP_NOFAIL doesn't work on
- a need to check for the process being signalled between failures
- the possibility that other recovery actions could be performed
- the allocation is quite deep in support code, and passing down an
extra flag to say if __GFP_NOFAIL is wanted would be clumsy.
Many of these currently use congestion_wait() which (in almost all
cases) simply waits the given timeout - congestion isn't tracked for
most devices.
It isn't clear what the best delay is for loops, but it is clear that
the various filesystems shouldn't be responsible for choosing a timeout.
This patch introduces memalloc_retry_wait() with takes on that
responsibility. Code that wants to retry a memory allocation can call
this function passing the GFP flags that were used. It will wait
however is appropriate.
For now, it only considers __GFP_NORETRY and whatever
gfpflags_allow_blocking() tests. If blocking is allowed without
__GFP_NORETRY, then alloc_page either made some reclaim progress, or
waited for a while, before failing. So there is no need for much
further waiting. memalloc_retry_wait() will wait until the current
jiffie ends. If this condition is not met, then alloc_page() won't have
waited much if at all. In that case memalloc_retry_wait() waits about
200ms. This is the delay that most current loops uses.
linux/sched/mm.h needs to be included in some files now,
but linux/backing-dev.h does not.
Link: https://lkml.kernel.org/r/163754371968.13692.1277530886009912421@noble.neil.brown.name
Signed-off-by: NeilBrown <neilb@suse.de>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Cc: Chao Yu <chao@kernel.org>
Cc: Darrick J. Wong <djwong@kernel.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
