While testing io_uring NAPI with DEFER_TASKRUN, I ran into slowdowns and
stalls in packet delivery. Turns out that while
io_napi_busy_loop_should_end() aborts appropriately on regular
task_work, it does not abort if we have local task_work pending.
Move io_has_work() into the private io_uring.h header, and gate whether
we should continue polling on that as well. This makes NAPI polling on
send/receive work as designed with IORING_SETUP_DEFER_TASKRUN as well.
Fixes: 8d0c12a80c ("io-uring: add napi busy poll support")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds the napi busy polling support in io_uring.c. It adds a new
napi_list to the io_ring_ctx structure. This list contains the list of
napi_id's that are currently enabled for busy polling. The list is
synchronized by the new napi_lock spin lock. The current default napi
busy polling time is stored in napi_busy_poll_to. If napi busy polling
is not enabled, the value is 0.
In addition there is also a hash table. The hash table store the napi
id and the pointer to the above list nodes. The hash table is used to
speed up the lookup to the list elements. The hash table is synchronized
with rcu.
The NAPI_TIMEOUT is stored as a timeout to make sure that the time a
napi entry is stored in the napi list is limited.
The busy poll timeout is also stored as part of the io_wait_queue. This
is necessary as for sq polling the poll interval needs to be adjusted
and the napi callback allows only to pass in one value.
This has been tested with two simple programs from the liburing library
repository: the napi client and the napi server program. The client
sends a request, which has a timestamp in its payload and the server
replies with the same payload. The client calculates the roundtrip time
and stores it to calculate the results.
The client is running on host1 and the server is running on host 2 (in
the same rack). The measured times below are roundtrip times. They are
average times over 5 runs each. Each run measures 1 million roundtrips.
no rx coal rx coal: frames=88,usecs=33
Default 57us 56us
client_poll=100us 47us 46us
server_poll=100us 51us 46us
client_poll=100us+ 40us 40us
server_poll=100us
client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on client
client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on server
client_poll=100us+ 41us 39us
server_poll=100us+
prefer napi busy poll on client + server
Signed-off-by: Stefan Roesch <shr@devkernel.io>
Suggested-by: Olivier Langlois <olivier@trillion01.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20230608163839.2891748-5-shr@devkernel.io
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This moves the definition of the io_wait_queue structure to the header
file so it can be also used from other files.
Signed-off-by: Stefan Roesch <shr@devkernel.io>
Link: https://lore.kernel.org/r/20230608163839.2891748-4-shr@devkernel.io
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Decouple from task_work running, and cap the number of entries we process
at the time. If we exceed that number, push remaining entries to a retry
list that we'll process first next time.
We cap the number of entries to process at 8, which is fairly random.
We just want to get enough per-ctx batching here, while not processing
endlessly.
Since we manually run PF_IO_WORKER related task_work anyway as the task
never exits to userspace, with this we no longer need to add an actual
task_work item to the per-process list.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Any of the fast paths will already have this locked, this helper only
exists to deal with io-wq invoking request issue where we do not have
the ctx->uring_lock held already. This means that any common or fast
path will already have this locked, mark it as such.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds a flag to avoid dipping dereferencing file and then f_op to
figure out if the file has a poll handler defined or not. We generally
call this at least twice for networked workloads, and if using ring
provided buffers, we do it on every buffer selection. Particularly the
latter is troublesome, as it's otherwise a very fast operation.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since our poll handling is edge triggered, multishot handlers retry
internally until they know that no more data is available. In
preparation for limiting these retries, add an internal return code,
IOU_REQUEUE, which can be used to inform the poll backend about the
handler wanting to retry, but that this should happen through a normal
task_work requeue rather than keep hammering on the issue side for this
one request.
No functional changes in this patch, nobody is using this return code
just yet.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since we no longer allow sending io_uring fds over SCM_RIGHTS, move to
using io_is_uring_fops() to detect whether this is a io_uring fd or not.
With that done, kill off io_uring_get_socket() as nobody calls it
anymore.
This is in preparation to yanking out the rest of the core related to
unix gc with io_uring.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Most of this code is basically self contained, move it out of the core
io_uring file to bring a bit more separation to the registration related
bits. This moves another ~10% of the code into register.c.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=okQ8
-----END PGP SIGNATURE-----
Merge tag 'for-6.7/io_uring-2023-10-30' of git://git.kernel.dk/linux
Pull io_uring updates from Jens Axboe:
"This contains the core io_uring updates, of which there are not many,
and adds support for using WAITID through io_uring and hence not
needing to block on these kinds of events.
Outside of that, tweaks to the legacy provided buffer handling and
some cleanups related to cancelations for uring_cmd support"
* tag 'for-6.7/io_uring-2023-10-30' of git://git.kernel.dk/linux:
io_uring/poll: use IOU_F_TWQ_LAZY_WAKE for wakeups
io_uring/kbuf: Use slab for struct io_buffer objects
io_uring/kbuf: Allow the full buffer id space for provided buffers
io_uring/kbuf: Fix check of BID wrapping in provided buffers
io_uring/rsrc: cleanup io_pin_pages()
io_uring: cancelable uring_cmd
io_uring: retain top 8bits of uring_cmd flags for kernel internal use
io_uring: add IORING_OP_WAITID support
exit: add internal include file with helpers
exit: add kernel_waitid_prepare() helper
exit: move core of do_wait() into helper
exit: abstract out should_wake helper for child_wait_callback()
io_uring/rw: add support for IORING_OP_READ_MULTISHOT
io_uring/rw: mark readv/writev as vectored in the opcode definition
io_uring/rw: split io_read() into a helper
The allocation of struct io_buffer for metadata of provided buffers is
done through a custom allocator that directly gets pages and
fragments them. But, slab would do just fine, as this is not a hot path
(in fact, it is a deprecated feature) and, by keeping a custom allocator
implementation we lose benefits like tracking, poisoning,
sanitizers. Finally, the custom code is more complex and requires
keeping the list of pages in struct ctx for no good reason. This patch
cleans this path up and just uses slab.
I microbenchmarked it by forcing the allocation of a large number of
objects with the least number of io_uring commands possible (keeping
nbufs=USHRT_MAX), with and without the patch. There is a slight
increase in time spent in the allocation with slab, of course, but even
when allocating to system resources exhaustion, which is not very
realistic and happened around 1/2 billion provided buffers for me, it
wasn't a significant hit in system time. Specially if we think of a
real-world scenario, an application doing register/unregister of
provided buffers will hit ctx->io_buffers_cache more often than actually
going to slab.
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20231005000531.30800-4-krisman@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_lockdep_assert_cq_locked() checks that locking is correctly done when
a CQE is posted. If the ring is setup in a disabled state with
IORING_SETUP_R_DISABLED, then ctx->submitter_task isn't assigned until
the ring is later enabled. We generally don't post CQEs in this state,
as no SQEs can be submitted. However it is possible to generate a CQE
if tagged resources are being updated. If this happens and PROVE_LOCKING
is enabled, then the locking check helper will dereference
ctx->submitter_task, which hasn't been set yet.
Fixup io_lockdep_assert_cq_locked() to handle this case correctly. While
at it, convert it to a static inline as well, so that generated line
offsets will actually reflect which condition failed, rather than just
the line offset for io_lockdep_assert_cq_locked() itself.
Reported-and-tested-by: syzbot+efc45d4e7ba6ab4ef1eb@syzkaller.appspotmail.com
Fixes: f26cc95935 ("io_uring: lockdep annotate CQ locking")
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_do_iopoll() and io_submit_flush_completions() are pretty similar,
both filling CQEs and then free a list of requests. Don't duplicate it
and make iopoll use __io_submit_flush_completions(), which also helps
with inlining and other optimisations.
For that, we need to first find all completed iopoll requests and splice
them from the iopoll list and then pass it down. This adds one extra
list traversal, which should be fine as requests will stay hot in cache.
CQ locking is already conditional, introduce ->lockless_cq and skip
locking for IOPOLL as it's protected by ->uring_lock.
We also add a wakeup optimisation for IOPOLL to __io_cq_unlock_post(),
so it works just like io_cqring_ev_posted_iopoll().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/3840473f5e8a960de35b77292026691880f6bdbc.1692916914.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the cached cqe check passes in io_get_cqe*() it already means that
the cqe we return is valid and non-zero, however the compiler is unable
to optimise null checks like in io_fill_cqe_req().
Do a bit of trickery, return success/fail boolean from io_get_cqe*()
and store cqe in the cqe parameter. That makes it do the right thing,
erasing the check together with the introduced indirection.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/322ea4d3377d3d4efd8ae90ab8ed28a99f518210.1692916914.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While looking at io_fill_cqe_req()'s asm I stumbled on our trace points
turning into the chunk below:
trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
req->cqe.res, req->cqe.flags,
req->extra1, req->extra2);
io_uring/io_uring.c:898: trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
movq 232(%rbx), %rdi # req_44(D)->big_cqe.extra2, _5
movq 224(%rbx), %rdx # req_44(D)->big_cqe.extra1, _6
movl 84(%rbx), %r9d # req_44(D)->cqe.D.81184.flags, _7
movl 80(%rbx), %r8d # req_44(D)->cqe.res, _8
movq 72(%rbx), %rcx # req_44(D)->cqe.user_data, _9
movq 88(%rbx), %rsi # req_44(D)->ctx, _10
./arch/x86/include/asm/jump_label.h:27: asm_volatile_goto("1:"
1:jmp .L1772 # objtool NOPs this #
...
It does a jump_label for actual tracing, but those 6 moves will stay
there in the hottest io_uring path. As an optimisation, add a
trace_io_uring_complete_enabled() check, which is also uses jump_labels,
it tricks the compiler into behaving. It removes the junk without
changing anything else int the hot path.
Note: apparently, it's not only me noticing it, and people are also
working it around. We should remove the check when it's solved
generically or rework tracing.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/555d8312644b3776f4be7e23f9b92943875c4bc7.1692916914.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now all callers of io_aux_cqe() set allow_overflow to false, remove the
parameter and not allow overflowing auxilary multishot cqes.
When CQ is full the function callers and all multishot requests in
general are expected to complete the request. That prevents indefinite
in-background grows of the overflow list and let's the userspace to
handle the backlog at its own pace.
Resubmitting a request should also be faster than accounting a bunch of
overflows, so it should be better for perf when it happens, but a well
behaving userspace should be trying to avoid overflows in any case.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/bb20d14d708ea174721e58bb53786b0521e4dd6d.1691757663.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_req_local_work_add() peeks into the work list, which can be executed
in the meanwhile. It's completely fine without KASAN as we're in an RCU
read section and it's SLAB_TYPESAFE_BY_RCU. With KASAN though it may
trigger a false positive warning because internal io_uring caches are
sanitised.
Remove sanitisation from the io_uring request cache for now.
Cc: stable@vger.kernel.org
Fixes: 8751d15426 ("io_uring: reduce scheduling due to tw")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c6fbf7a82a341e66a0007c76eefd9d57f2d3ba51.1691541473.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Extract a function for non-local task_work_add, and use it directly from
io_move_task_work_from_local(). Now we don't use IOU_F_TWQ_FORCE_NORMAL
and it can be killed.
As a small positive side effect we don't grab task->io_uring in
io_req_normal_work_add anymore, which is not needed for
io_req_local_work_add().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/2e55571e8ff2927ae3cc12da606d204e2485525b.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Everybody is passing in the request, so get rid of the io_ring_ctx and
explicit user_data pass-in. Both the ctx and user_data can be deduced
from the request at hand.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With IORING_REGISTER_USE_REGISTERED_RING, an application can register
the ring fd and use it via registered index rather than installed fd.
This allows using a registered ring for everything *except* the initial
mmap.
With IORING_SETUP_NO_MMAP, io_uring_setup uses buffers allocated by the
user, rather than requiring a subsequent mmap.
The combination of the two allows a user to operate *entirely* via a
registered ring fd, making it unnecessary to ever install the fd in the
first place. So, add a flag IORING_SETUP_REGISTERED_FD_ONLY to make
io_uring_setup register the fd and return a registered index, without
installing the fd.
This allows an application to avoid touching the fd table at all, and
allows a library to never even momentarily install a file descriptor.
This splits out an io_ring_add_registered_file helper from
io_ring_add_registered_fd, for use by io_uring_setup.
Signed-off-by: Josh Triplett <josh@joshtriplett.org>
Link: https://lore.kernel.org/r/bc8f431bada371c183b95a83399628b605e978a3.1682699803.git.josh@joshtriplett.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Create a simple helper that returns the size of the SQE. The SQE could
have two size, depending of the flags.
If IO_URING_SETUP_SQE128 flag is set, then return a double SQE,
otherwise returns the sizeof of io_uring_sqe (64 bytes).
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20230504121856.904491-2-leitao@debian.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Every task_work will try to wake the task to be executed, which causes
excessive scheduling and additional overhead. For some tw it's
justified, but others won't do much but post a single CQE.
When a task waits for multiple cqes, every such task_work will wake it
up. Instead, the task may give a hint about how many cqes it waits for,
io_req_local_work_add() will compare against it and skip wake ups
if #cqes + #tw is not enough to satisfy the waiting condition. Task_work
that uses the optimisation should be simple enough and never post more
than one CQE. It's also ignored for non DEFER_TASKRUN rings.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d2b77e99d1e86624d8a69f7037d764b739dcd225.1680782017.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We already do this manually for the !SQPOLL case, do it in general and
we can also dump the ugly min3() in io_submit_sqes().
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For task works we're passing around a bool pointer for whether the
current ring is locked or not, let's wrap it in a structure, that
will make it more opaque preventing abuse and will also help us
to pass more info in the future if needed.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1ecec9483d58696e248d1bfd52cf62b04442df1d.1679931367.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Just like for task_work, set the task mode to TASK_RUNNING before doing
any potential resume work. We're not holding any locks at this point,
but we may have already set the task state to TASK_INTERRUPTIBLE in
preparation for going to sleep waiting for events. Ensure that we set it
back to TASK_RUNNING if we have work to process, to avoid warnings on
calling blocking operations with !TASK_RUNNING.
Fixes: b5d3ae202f ("io_uring: handle TIF_NOTIFY_RESUME when checking for task_work")
Reported-by: kernel test robot <oliver.sang@intel.com>
Link: https://lore.kernel.org/oe-lkp/202302062208.24d3e563-oliver.sang@intel.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Every io_uring request is represented by struct io_kiocb, which is
cached locally by io_uring (not SLAB/SLUB) in the list called
submit_state.freelist. This patch simply enabled KASAN for this free
list.
This list is initially created by KMEM_CACHE, but later, managed by
io_uring. This patch basically poisons the objects that are not used
(i.e., they are the free list), and unpoisons it when the object is
allocated/removed from the list.
Touching these poisoned objects while in the freelist will cause a KASAN
warning.
Suggested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If TIF_NOTIFY_RESUME is set, then we need to call resume_user_mode_work()
for PF_IO_WORKER threads. They never return to usermode, hence never get
a chance to process any items that are marked by this flag. Most notably
this includes the final put of files, but also any throttling markers set
by block cgroups.
Cc: stable@vger.kernel.org # 5.10+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Even though io_poll_wq_wake()'s waitqueue_active reuses a barrier we do
for another waitqueue, it's not going to be the case in the future and
so we want to have a fast path for it when the ring has never been
polled.
Move poll_wq wake ups into __io_commit_cqring_flush() using a new flag
called ->poll_activated. The idea behind the flag is to set it when the
ring was polled for the first time. This requires additional sync to not
miss events, which is done here by using task_work for ->task_complete
rings, and by default enabling the flag for all other types of rings.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/060785e8e9137a920b232c0c7f575b131af19cac.1673274244.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Most places that want to run local tw explicitly and in advance check if
they are allowed to do so. Don't rely on a similar check in
__io_run_local_work(), leave it as a just-in-case warning and make sure
callers checks capabilities themselves.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/990fe0e8e70fd4d57e43625e5ce8fba584821d1a.1672916894.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
