Pavel Begunkov [Wed, 27 Jul 2022 09:30:41 +0000 (10:30 +0100)]
io_uring: notification completion optimisation
We want to use all optimisations that we have for io_uring requests like
completion batching, memory caching and more but for zc notifications.
Fortunately, notification perfectly fit the request model so we can
overlay them onto struct io_kiocb and use all the infratructure.
Most of the fields of struct io_notif natively fits into io_kiocb, so we
replace struct io_notif with struct io_kiocb carrying struct
io_notif_data in the cmd cache line. Then we adapt io_alloc_notif() to
use io_alloc_req()/io_alloc_req_refill(), and kill leftovers of hand
coded caching. __io_notif_complete_tw() is converted to use io_uring's
tw infra.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:51 +0000 (21:52 +0100)]
selftests/io_uring: test zerocopy send
Add selftests for io_uring zerocopy sends and io_uring's notification
infrastructure. It's largely influenced by msg_zerocopy and uses it on
the receive side.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:50 +0000 (21:52 +0100)]
io_uring: enable managed frags with register buffers
io_uring's registered buffers infra has a good performant way of pinning
pages, so let's use SKBFL_MANAGED_FRAG_REFS when our requests are purely
register buffer backed.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:49 +0000 (21:52 +0100)]
io_uring: add zc notification flush requests
Overlay notification control onto IORING_OP_RSRC_UPDATE (former
IORING_OP_FILES_UPDATE). It allows to flush a range of zc notifications
from slots with indexes [sqe->off, sqe->off+sqe->len). If sqe->arg is
not zero, it also copies sqe->arg as a new tag for all flushed
notifications.
Note, it doesn't flush a notification of a slot if there was no requests
attached to it (since last flush or registration).
Pavel Begunkov [Tue, 12 Jul 2022 20:52:47 +0000 (21:52 +0100)]
io_uring: flush notifiers after sendzc
Allow to flush notifiers as a part of sendzc request by setting
IORING_SENDZC_FLUSH flag. When the sendzc request succeedes it will
flush the used [active] notifier.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:46 +0000 (21:52 +0100)]
io_uring: sendzc with fixed buffers
Allow zerocopy sends to use fixed buffers. There is an optimisation for
this case, the network layer don't need to reference the pages, see
SKBFL_MANAGED_FRAG_REFS, so io_uring have to ensure validity of fixed
buffers until the notifier is released.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:43 +0000 (21:52 +0100)]
io_uring: wire send zc request type
Add a new io_uring opcode IORING_OP_SENDZC. The main distinction from
IORING_OP_SEND is that the user should specify a notification slot
index in sqe::notification_idx and the buffers are safe to reuse only
when the used notification is flushed and completes.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:41 +0000 (21:52 +0100)]
io_uring: add rsrc referencing for notifiers
In preparation to zerocopy sends with fixed buffers make notifiers to
reference the rsrc node to protect the used fixed buffers. We can't just
grab it for a send request as notifiers can likely outlive requests that
used it.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:39 +0000 (21:52 +0100)]
io_uring: cache struct io_notif
kmalloc'ing struct io_notif is too expensive when done frequently, cache
them as many other resources in io_uring. Keep two list, the first one
is from where we're getting notifiers, it's protected by ->uring_lock.
The second is protected by ->completion_lock, to which we queue released
notifiers. Then we splice one list into another when needed.
Pavel Begunkov [Tue, 12 Jul 2022 20:52:38 +0000 (21:52 +0100)]
io_uring: add zc notification infrastructure
Add internal part of send zerocopy notifications. There are two main
structures, the first one is struct io_notif, which carries inside
struct ubuf_info and maps 1:1 to it. io_uring will be binding a number
of zerocopy send requests to it and ask to complete (aka flush) it. When
flushed and all attached requests and skbs complete, it'll generate one
and only one CQE. There are intended to be passed into the network layer
as struct msghdr::msg_ubuf.
The second concept is notification slots. The userspace will be able to
register an array of slots and subsequently addressing them by the index
in the array. Slots are independent of each other. Each slot can have
only one notifier at a time (called active notifier) but many notifiers
during the lifetime. When active, a notifier not going to post any
completion but the userspace can attach requests to it by specifying
the corresponding slot while issueing send zc requests. Eventually, the
userspace will want to "flush" the notifier losing any way to attach
new requests to it, however it can use the next atomatically added
notifier of this slot or of any other slot.
When the network layer is done with all enqueued skbs attached to a
notifier and doesn't need the specified in them user data, the flushed
notifier will post a CQE.
Merge branch 'for-5.20/io_uring' into for-5.20/io_uring-zerocopy-send
* for-5.20/io_uring: (716 commits)
io_uring: ensure REQ_F_ISREG is set async offload
net: fix compat pointer in get_compat_msghdr()
io_uring: Don't require reinitable percpu_ref
io_uring: fix types in io_recvmsg_multishot_overflow
io_uring: Use atomic_long_try_cmpxchg in __io_account_mem
io_uring: support multishot in recvmsg
net: copy from user before calling __get_compat_msghdr
net: copy from user before calling __copy_msghdr
io_uring: support 0 length iov in buffer select in compat
io_uring: fix multishot ending when not polled
io_uring: add netmsg cache
io_uring: impose max limit on apoll cache
io_uring: add abstraction around apoll cache
io_uring: move apoll cache to poll.c
io_uring: consolidate hash_locked io-wq handling
io_uring: clear REQ_F_HASH_LOCKED on hash removal
io_uring: don't race double poll setting REQ_F_ASYNC_DATA
io_uring: don't miss setting REQ_F_DOUBLE_POLL
io_uring: disable multishot recvmsg
io_uring: only trace one of complete or overflow
...
If we're offloading requests directly to io-wq because IOSQE_ASYNC was
set in the sqe, we can miss hashing writes appropriately because we
haven't set REQ_F_ISREG yet. This can cause a performance regression
with buffered writes, as io-wq then no longer correctly serializes writes
to that file.
Ensure that we set the flags in io_prep_async_work(), which will cause
the io-wq work item to be hashed appropriately.
Fixes: 2b76ab77d446 ("io_uring: move read/write file prep state into actual opcode handler") Link: https://lore.kernel.org/io-uring/20220608080054.GB22428@xsang-OptiPlex-9020/ Reported-by: kernel test robot <oliver.sang@intel.com> Tested-by: Yin Fengwei <fengwei.yin@intel.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
A previous change enabled external users to copy the data before
calling __get_compat_msghdr(), but didn't modify get_compat_msghdr() or
__io_compat_recvmsg_copy_hdr() to take that into account. They are both
stil passing in the __user pointer rather than the copied version.
Ensure we pass in the kernel struct, not the pointer to the user data.
Michal Koutný [Fri, 15 Jul 2022 17:45:01 +0000 (19:45 +0200)]
io_uring: Don't require reinitable percpu_ref
The commit 7bf09ad31555 ("io_uring: remove ring quiesce for
io_uring_register") removed the worklow relying on reinit/resurrection
of the percpu_ref, hence, initialization with that requested is a relic.
This is based on code review, this causes no real bug (and theoretically
can't). Technically it's a revert of commit 99b355628d97 ("io_uring:
initialize percpu refcounters using PERCU_REF_ALLOW_REINIT") but since
the flag omission is now justified, I'm not making this a revert.
Fixes: 7bf09ad31555 ("io_uring: remove ring quiesce for io_uring_register") Signed-off-by: Michal Koutný <mkoutny@suse.com> Acked-by: Roman Gushchin <roman.gushchin@linux.dev> Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring: fix types in io_recvmsg_multishot_overflow
io_recvmsg_multishot_overflow had incorrect types on non x64 system.
But also it had an unnecessary INT_MAX check, which could just be done
by changing the type of the accumulator to int (also simplifying the
casts).
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au> Fixes: a8b38c4ce724 ("io_uring: support multishot in recvmsg") Signed-off-by: Dylan Yudaken <dylany@fb.com> Link: https://lore.kernel.org/r/20220715130252.610639-1-dylany@fb.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring: Use atomic_long_try_cmpxchg in __io_account_mem
Use atomic_long_try_cmpxchg instead of
atomic_long_cmpxchg (*ptr, old, new) == old in __io_account_mem.
x86 CMPXCHG instruction returns success in ZF flag, so this
change saves a compare after cmpxchg (and related move
instruction in front of cmpxchg).
Also, atomic_long_try_cmpxchg implicitly assigns old *ptr value
to "old" when cmpxchg fails, enabling further code simplifications.
Similar to multishot recv, this will require provided buffers to be
used. However recvmsg is much more complex than recv as it has multiple
outputs. Specifically flags, name, and control messages.
Support this by introducing a new struct io_uring_recvmsg_out with 4
fields. namelen, controllen and flags match the similar out fields in
msghdr from standard recvmsg(2), payloadlen is the length of the payload
following the header.
This struct is placed at the start of the returned buffer. Based on what
the user specifies in struct msghdr, the next bytes of the buffer will be
name (the next msg_namelen bytes), and then control (the next
msg_controllen bytes). The payload will come at the end. The return value
in the CQE is the total used size of the provided buffer.
If multishot is not actually polling then return IOU_OK rather than the
result.
If the result was > 0 this will confuse things further up the callstack
which expect a return <= 0.
For recvmsg/sendmsg, if they don't complete inline, we currently need
to allocate a struct io_async_msghdr for each request. This is a
somewhat large struct.
Hook up sendmsg/recvmsg to use the io_alloc_cache. This reduces the
alloc + free overhead considerably, yielding 4-5% of extra performance
running netbench.
Caches like this tend to grow to the peak size, and then never get any
smaller. Impose a max limit on the size, to prevent it from growing too
big.
A somewhat randomly chosen 512 is the max size we'll allow the cache
to get. If a batch of frees come in and would bring it over that, we
simply start kfree'ing the surplus.
Just as with io_poll_double_prepare() setting REQ_F_DOUBLE_POLL, we can
race with the first poll entry when setting REQ_F_ASYNC_DATA. Move it
under io_poll_double_prepare().
Pavel Begunkov [Thu, 7 Jul 2022 14:13:14 +0000 (15:13 +0100)]
io_uring: don't miss setting REQ_F_DOUBLE_POLL
When adding a second poll entry we should set REQ_F_DOUBLE_POLL
unconditionally. We might race with the first entry removal but that
doesn't change the rule.
recvmsg has semantics that do not make it trivial to extend to
multishot. Specifically it has user pointers and returns data in the
original parameter. In order to make this API useful these will need to be
somehow included with the provided buffers.
For now remove multishot for recvmsg as it is not useful.
Dylan Yudaken [Thu, 30 Jun 2022 09:12:29 +0000 (02:12 -0700)]
io_uring: multishot recv
Support multishot receive for io_uring.
Typical server applications will run a loop where for each recv CQE it
requeues another recv/recvmsg.
This can be simplified by using the existing multishot functionality
combined with io_uring's provided buffers.
The API is to add the IORING_RECV_MULTISHOT flag to the SQE. CQEs will
then be posted (with IORING_CQE_F_MORE flag set) when data is available
and is read. Once an error occurs or the socket ends, the multishot will
be removed and a completion without IORING_CQE_F_MORE will be posted.
The benefit to this is that the recv is much more performant.
* Subsequent receives are queued up straight away without requiring the
application to finish a processing loop.
* If there are more data in the socket (sat the provided buffer size is
smaller than the socket buffer) then the data is immediately
returned, improving batching.
* Poll is only armed once and reused, saving CPU cycles
Dylan Yudaken [Thu, 30 Jun 2022 09:12:27 +0000 (02:12 -0700)]
io_uring: fix multishot poll on overflow
On overflow, multishot poll can still complete with the IORING_CQE_F_MORE
flag set.
If in the meantime the user clears a CQE and a the poll was cancelled then
the poll will post a CQE without the IORING_CQE_F_MORE (and likely result
-ECANCELED).
However when processing the application will encounter the non-overflow
CQE which indicates that there will be no more events posted. Typical
userspace applications would free memory associated with the poll in this
case.
It will then subsequently receive the earlier CQE which has overflowed,
which breaks the contract given by the IORING_CQE_F_MORE flag.
Dylan Yudaken [Thu, 30 Jun 2022 09:12:26 +0000 (02:12 -0700)]
io_uring: add allow_overflow to io_post_aux_cqe
Some use cases of io_post_aux_cqe would not want to overflow as is, but
might want to change the flags/result. For example multishot receive
requires in order CQE, and so if there is an overflow it would need to
stop receiving until the overflow is taken care of.
Pavel Begunkov [Sat, 25 Jun 2022 10:55:38 +0000 (11:55 +0100)]
io_uring: let to set a range for file slot allocation
From recently io_uring provides an option to allocate a file index for
operation registering fixed files. However, it's utterly unusable with
mixed approaches when for a part of files the userspace knows better
where to place it, as it may race and users don't have any sane way to
pick a slot and hoping it will not be taken.
Let the userspace to register a range of fixed file slots in which the
auto-allocation happens. The use case is splittting the fixed table in
two parts, where on of them is used for auto-allocation and another for
slot-specified operations.
Jens Axboe [Mon, 13 Jun 2022 10:47:02 +0000 (04:47 -0600)]
io_uring: add support for passing fixed file descriptors
With IORING_OP_MSG_RING, one ring can send a message to another ring.
Extend that support to also allow sending a fixed file descriptor to
that ring, enabling one ring to pass a registered descriptor to another
one.
Arguments are extended to pass in:
sqe->addr3 fixed file slot in source ring
sqe->file_index fixed file slot in destination ring
IORING_OP_MSG_RING is extended to take a command argument in sqe->addr.
If set to zero (or IORING_MSG_DATA), it sends just a message like before.
If set to IORING_MSG_SEND_FD, a fixed file descriptor is sent according
to the above arguments.
Two common use cases for this are:
1) Server needs to be shutdown or restarted, pass file descriptors to
another onei
2) Backend is split, and one accepts connections, while others then get
the fd passed and handle the actual connection.
Both of those are classic SCM_RIGHTS use cases, and it's not possible to
support them with direct descriptors today.
By default, this will post a CQE to the target ring, similarly to how
IORING_MSG_DATA does it. If IORING_MSG_RING_CQE_SKIP is set, no message
is posted to the target ring. The issuer is expected to notify the
receiver side separately.
Jens Axboe [Mon, 13 Jun 2022 10:42:56 +0000 (04:42 -0600)]
io_uring: split out fixed file installation and removal
Put it with the filetable code, which is where it belongs. While doing
so, have the helpers take a ctx rather than an io_kiocb. It doesn't make
sense to use a request, as it's not an operation on the request itself.
It applies to the ring itself.
io_uring: replace zero-length array with flexible-array member
There is a regular need in the kernel to provide a way to declare
having a dynamically sized set of trailing elements in a structure.
Kernel code should always use “flexible array members”[1] for these
cases. The older style of one-element or zero-length arrays should
no longer be used[2].
Pavel Begunkov [Sat, 25 Jun 2022 10:53:02 +0000 (11:53 +0100)]
io_uring: remove ctx->refs pinning on enter
io_uring_enter() takes ctx->refs, which was previously preventing racing
with register quiesce. However, as register now doesn't touch the refs,
we can freely kill extra ctx pinning and rely on the fact that we're
holding a file reference preventing the ring from being destroyed.
Pavel Begunkov [Sat, 25 Jun 2022 10:52:58 +0000 (11:52 +0100)]
io_uring: improve io_fail_links()
io_fail_links() is called with ->completion_lock held and for that
reason we'd want to keep it as small as we can. Instead of doing
__io_req_complete_post() for each linked request under the lock, fail
them in a task_work handler under ->uring_lock.
Jens Axboe [Tue, 21 Jun 2022 20:34:15 +0000 (14:34 -0600)]
io_uring: move POLLFREE handling to separate function
We really don't care about this at all in terms of performance. Outside
of having it already be marked unlikely(), shove it into a separate
__cold function.
Pavel Begunkov [Thu, 23 Jun 2022 13:24:49 +0000 (14:24 +0100)]
io_uring: optimise submission side poll_refs
The final poll_refs put in __io_arm_poll_handler() takes quite some
cycles. When we're arming from the original task context task_work won't
be run, so in this case we can assume that we won't race with task_works
and so not take the initial ownership ref.
One caveat is that after arming a poll we may race with it, so we have
to add a bunch of io_poll_get_ownership() hidden inside of
io_poll_can_finish_inline() whenever we want to complete arming inline.
For the same reason we can't just set REQ_F_DOUBLE_POLL in
__io_queue_proc() and so need to sync with the first poll entry by
taking its wq head lock.
Pavel Begunkov [Thu, 23 Jun 2022 13:24:48 +0000 (14:24 +0100)]
io_uring: refactor poll arm error handling
__io_arm_poll_handler() errors parsing is a horror, in case it failed it
returns 0 and the caller is expected to look at ipt.error, which already
led us to a number of problems before.
When it returns a valid mask, leave it as it's not, i.e. return 1 and
store the mask in ipt.result_mask. In case of a failure that can be
handled inline return an error code (negative value), and return 0 if
__io_arm_poll_handler() took ownership of the request and will complete
it.
Pavel Begunkov [Thu, 23 Jun 2022 13:24:47 +0000 (14:24 +0100)]
io_uring: change arm poll return values
The rules for __io_arm_poll_handler()'s result parsing are complicated,
as the first step don't pass return a mask but pass back a positive
return code and fill ipt->result_mask.
Pavel Begunkov [Thu, 23 Jun 2022 13:24:44 +0000 (14:24 +0100)]
io_uring: clean poll ->private flagging
We store a req pointer in wqe->private but also take one bit to mark
double poll entries. Replace macro helpers with inline functions for
better type checking and also name the double flag.
Jens Axboe [Sat, 18 Jun 2022 16:00:50 +0000 (10:00 -0600)]
io_uring: add sync cancelation API through io_uring_register()
The io_uring cancelation API is async, like any other API that we expose
there. For the case of finding a request to cancel, or not finding one,
it is fully sync in that when submission returns, the CQE for both the
cancelation request and the targeted request have been posted to the
CQ ring.
However, if the targeted work is being executed by io-wq, the API can
only start the act of canceling it. This makes it difficult to use in
some circumstances, as the caller then has to wait for the CQEs to come
in and match on the same cancelation data there.
Provide a IORING_REGISTER_SYNC_CANCEL command for io_uring_register()
that does sync cancelations, always. For the io-wq case, it'll wait
for the cancelation to come in before returning. The only expected
returns from this API is:
0 Request found and canceled fine.
> 0 Requests found and canceled. Only happens if asked to
cancel multiple requests, and if the work wasn't in
progress.
-ENOENT Request not found.
-ETIME A timeout on the operation was requested, but the timeout
expired before we could cancel.
and we won't get -EALREADY via this API.
If the timeout value passed in is -1 (tv_sec and tv_nsec), then that
means that no timeout is requested. Otherwise, the timespec passed in
is the amount of time the sync cancel will wait for a successful
cancelation.
Jens Axboe [Sat, 18 Jun 2022 15:47:04 +0000 (09:47 -0600)]
io_uring: add IORING_ASYNC_CANCEL_FD_FIXED cancel flag
In preparation for not having a request to pass in that carries this
state, add a separate cancelation flag that allows the caller to ask
for a fixed file for cancelation.
Dylan Yudaken [Wed, 22 Jun 2022 13:40:25 +0000 (06:40 -0700)]
io_uring: batch task_work
Batching task work up is an important performance optimisation, as
task_work_add is expensive.
In order to keep the semantics replace the task_list with a fake node
while processing the old list, and then do a cmpxchg at the end to see if
there is more work.
Dylan Yudaken [Wed, 22 Jun 2022 13:40:23 +0000 (06:40 -0700)]
io_uring: lockless task list
With networking use cases we see contention on the spinlock used to
protect the task_list when multiple threads try and add completions at once.
Instead we can use a lockless list, and assume that the first caller to
add to the list is responsible for kicking off task work.
Pavel Begunkov [Tue, 21 Jun 2022 09:09:00 +0000 (10:09 +0100)]
io_uring: improve io_run_task_work()
Since SQPOLL now uses TWA_SIGNAL_NO_IPI, there won't be task work items
without TIF_NOTIFY_SIGNAL. Simplify io_run_task_work() by removing
task->task_works check. Even though looks it doesn't cause extra cache
bouncing, it's still nice to not touch it an extra time when it might be
not cached.
Pavel Begunkov [Mon, 20 Jun 2022 14:27:35 +0000 (15:27 +0100)]
io_uring: optimize io_uring_task layout
task_work bits of io_uring_task are split into two cache lines causing
extra cache bouncing, place them into a separate cache line. Also move
the most used submission path fields closer together, so there are hot.
Pavel Begunkov [Mon, 20 Jun 2022 00:26:01 +0000 (01:26 +0100)]
io_uring: add a warn_once for poll_find
io_poll_remove() expects poll_find() to search only for poll requests and
passes a flag for this. Just be a little bit extra cautious considering
lots of recent poll/cancellation changes and add a WARN_ON_ONCE checking
that we don't get an apoll'ed request.
Pavel Begunkov [Mon, 20 Jun 2022 00:26:00 +0000 (01:26 +0100)]
io_uring: consistent naming for inline completion
Improve naming of the inline/deferred completion helper so it's
consistent with it's *_post counterpart. Add some comments and extra
lockdeps to ensure the locking is done right.
Pavel Begunkov [Mon, 20 Jun 2022 00:25:57 +0000 (01:25 +0100)]
io_uring: add io_commit_cqring_flush()
Since __io_commit_cqring_flush users moved to different files, introduce
io_commit_cqring_flush() helper and encapsulate all flags testing details
inside.
Pavel Begunkov [Mon, 20 Jun 2022 00:25:56 +0000 (01:25 +0100)]
io_uring: introduce locking helpers for CQE posting
spin_lock(&ctx->completion_lock);
/* post CQEs */
io_commit_cqring(ctx);
spin_unlock(&ctx->completion_lock);
io_cqring_ev_posted(ctx);
We have many places repeating this sequence, and the three function
unlock section is not perfect from the maintainance perspective and also
makes it harder to add new locking/sync trick.
Introduce two helpers. io_cq_lock(), which is simple and only grabs
->completion_lock, and io_cq_unlock_post() encapsulating the three call
section.
Pavel Begunkov [Mon, 20 Jun 2022 00:25:55 +0000 (01:25 +0100)]
io_uring: hide eventfd assumptions in eventfd paths
Some io_uring-eventfd users assume that there won't be spurious wakeups.
That assumption has to be honoured by all io_cqring_ev_posted() callers,
which is inconvenient and from time to time leads to problems but should
be maintained to not break the userspace.
Instead of making the callers track whether a CQE was posted or not, hide
it inside io_eventfd_signal(). It saves ->cached_cq_tail it saw last time
and triggers the eventfd only when ->cached_cq_tail changed since then.
Pavel Begunkov [Mon, 20 Jun 2022 00:25:54 +0000 (01:25 +0100)]
io_uring: fix io_poll_remove_all clang warnings
clang complains on bitwise operations with bools, add a bit more
verbosity to better show that we want to call io_poll_remove_all_table()
twice but with different arguments.
Pavel Begunkov [Mon, 20 Jun 2022 00:25:52 +0000 (01:25 +0100)]
io_uring: fix multi ctx cancellation
io_uring_try_cancel_requests() loops until there is nothing left to do
with the ring, however there might be several rings and they might have
dependencies between them, e.g. via poll requests.
Instead of cancelling rings one by one, try to cancel them all and only
then loop over if we still potenially some work to do.
Pavel Begunkov [Sun, 19 Jun 2022 11:26:08 +0000 (12:26 +0100)]
io_uring: remove ->flush_cqes optimisation
It's not clear how widely used IOSQE_CQE_SKIP_SUCCESS is, and how often
->flush_cqes flag prevents from completion being flushed. Sometimes it's
high level of concurrency that enables it at least for one CQE, but
sometimes it doesn't save much because nobody waiting on the CQ.
Remove ->flush_cqes flag and the optimisation, it should benefit the
normal use case. Note, that there is no spurious eventfd problem with
that as checks for spuriousness were incorporated into
io_eventfd_signal().
projects / kernel.git / log