By default, io_uring will stop submitting a batch of requests if we run
into an error submitting a request. This isn't strictly necessary, as
the error result is passed out-of-band via a CQE anyway. And it can be
a bit confusing for some applications.
Provide a way to setup a ring that will continue submitting on error,
when the error CQE has been posted.
There's still one case that will break out of submission. If we fail
allocating a request, then we'll still return -ENOMEM. We could in theory
post a CQE for that condition too even if we never got a request. Leave
that for a potential followup.
Reported-by: Dylan Yudaken <dylany@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we are using provided buffers, it's less than useful to have a buffer
selected and pinned if a request needs to go async or arms poll for
notification trigger on when we can process it.
Recycle the buffer in those events, so we don't pin it for the duration
of the request.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Most of the logic in io_read() deals with regular files, and in some ways
it would make sense to split the handling into S_IFREG and others. But
at least for retry, we don't need to bother setting up a bunch of state
just to abort in the loop later. In particular, don't bother forcing
setup of async data for a normal non-vectored read when we don't need it.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The sqpoll thread can be used for performing the napi busy poll in a
similar way that it does io polling for file systems supporting direct
access bypassing the page cache.
The other way that io_uring can be used for napi busy poll is by
calling io_uring_enter() to get events.
If the user specify a timeout value, it is distributed between polling
and sleeping by using the systemwide setting
/proc/sys/net/core/busy_poll.
The changes have been tested with this program:
https://github.com/lano1106/io_uring_udp_ping
and the result is:
Without sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 40.631/42.050/58.667/1.547 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 30.619/31.753/61.433/1.456 us
With sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 42.087/44.438/59.508/1.533 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 35.779/37.347/52.201/0.924 us
Co-developed-by: Hao Xu <haoxu@linux.alibaba.com>
Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Link: https://lore.kernel.org/r/810bd9408ffc510ff08269e78dca9df4af0b9e4e.1646777484.git.olivier@trillion01.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds support for IORING_OP_MSG_RING, which allows an SQE to signal
another ring. That allows either waking up someone waiting on the ring,
or even passing a 64-bit value via the user_data field in the CQE.
sqe->fd must contain the fd of a ring that should receive the CQE.
sqe->off will be propagated to the cqe->user_data on the target ring,
and sqe->len will be propagated to cqe->res. The results CQE will have
IORING_CQE_F_MSG set in its flags, to indicate that this CQE was generated
from a messaging request rather than a SQE issued locally on that ring.
This effectively allows passing a 64-bit and a 32-bit quantify between
the two rings.
This request type has the following request specific error cases:
- -EBADFD. Set if the sqe->fd doesn't point to a file descriptor that is
of the io_uring type.
- -EOVERFLOW. Set if we were not able to deliver a request to the target
ring.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In testing high frequency workloads with provided buffers, we spend a
lot of time in allocating and freeing the buffer units themselves.
Rather than repeatedly free and alloc them, add a recycling cache
instead. There are two caches:
- ctx->io_buffers_cache. This is the one we grab from in the submission
path, and it's protected by ctx->uring_lock. For inline completions,
we can recycle straight back to this cache and not need any extra
locking.
- ctx->io_buffers_comp. If we're not under uring_lock, then we use this
list to recycle buffers. It's protected by the completion_lock.
On adding a new buffer, check io_buffers_cache. If it's empty, check if
we can splice entries from the io_buffers_comp_cache.
This reduces about 5-10% of overhead from provided buffers, bringing it
pretty close to the non-provided path.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Lots of workloads use multiple threads, in which case the file table is
shared between them. This makes getting and putting the ring file
descriptor for each io_uring_enter(2) system call more expensive, as it
involves an atomic get and put for each call.
Similarly to how we allow registering normal file descriptors to avoid
this overhead, add support for an io_uring_register(2) API that allows
to register the ring fds themselves:
1) IORING_REGISTER_RING_FDS - takes an array of io_uring_rsrc_update
structs, and registers them with the task.
2) IORING_UNREGISTER_RING_FDS - takes an array of io_uring_src_update
structs, and unregisters them.
When a ring fd is registered, it is internally represented by an offset.
This offset is returned to the application, and the application then
uses this offset and sets IORING_ENTER_REGISTERED_RING for the
io_uring_enter(2) system call. This works just like using a registered
file descriptor, rather than a real one, in an SQE, where
IOSQE_FIXED_FILE gets set to tell io_uring that we're using an internal
offset/descriptor rather than a real file descriptor.
In initial testing, this provides a nice bump in performance for
threaded applications in real world cases where the batch count (eg
number of requests submitted per io_uring_enter(2) invocation) is low.
In a microbenchmark, submitting NOP requests, we see the following
increases in performance:
Requests per syscall Baseline Registered Increase
----------------------------------------------------------------
1 ~7030K ~8080K +15%
2 ~13120K ~14800K +13%
4 ~22740K ~25300K +11%
Co-developed-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is a slight optimisation to be had by calculating the correct pos
pointer inside io_kiocb_update_pos and then using that later.
It seems code size drops by a bit:
000000000000a1b0 0000000000000400 t io_read
000000000000a5b0 0000000000000319 t io_write
vs
000000000000a1b0 00000000000003f6 t io_read
000000000000a5b0 0000000000000310 t io_write
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Update kiocb->ki_pos at execution time rather than in io_prep_rw().
io_prep_rw() happens before the job is enqueued to a worker and so the
offset might be read multiple times before being executed once.
Ensures that the file position in a set of _linked_ SQEs will be only
obtained after earlier SQEs have completed, and so will include their
incremented file position.
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_kiocb_ppos is called in both branches, and it seems that the compiler
does not fuse this. Fusing removes a few bytes from loop_rw_iter.
Before:
$ nm -S fs/io_uring.o | grep loop_rw_iter
0000000000002430 0000000000000124 t loop_rw_iter
After:
$ nm -S fs/io_uring.o | grep loop_rw_iter
0000000000002430 000000000000010d t loop_rw_iter
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
wqe->lock is abused, it now protects acct->work_list, hash stuff,
nr_workers, wqe->free_list and so on. Lets first get the work_list out
of the wqe-lock mess by introduce a specific lock for work list. This
is the first step to solve the huge contension between work insertion
and work consumption.
good thing:
- split locking for bound and unbound work list
- reduce contension between work_list visit and (worker's)free_list.
For the hash stuff, since there won't be a work with same file in both
bound and unbound work list, thus they won't visit same hash entry. it
works well to use the new lock to protect hash stuff.
Results:
set max_unbound_worker = 4, test with echo-server:
nice -n -15 ./io_uring_echo_server -p 8081 -f -n 1000 -l 16
(-n connection, -l workload)
before this patch:
Samples: 2M of event 'cycles:ppp', Event count (approx.): 1239982111074
Overhead Command Shared Object Symbol
28.59% iou-wrk-10021 [kernel.vmlinux] [k] native_queued_spin_lock_slowpath
8.89% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpath
6.20% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock
2.45% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work
2.36% iou-wrk-10021 [kernel.vmlinux] [k] _raw_spin_lock_irqsave
2.29% iou-wrk-10021 [kernel.vmlinux] [k] io_worker_handle_work
1.29% io_uring_echo_s [kernel.vmlinux] [k] io_wqe_enqueue
1.06% iou-wrk-10021 [kernel.vmlinux] [k] io_wqe_worker
1.06% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock
1.03% iou-wrk-10021 [kernel.vmlinux] [k] __schedule
0.99% iou-wrk-10021 [kernel.vmlinux] [k] tcp_sendmsg_locked
with this patch:
Samples: 1M of event 'cycles:ppp', Event count (approx.): 708446691943
Overhead Command Shared Object Symbol
16.86% iou-wrk-10893 [kernel.vmlinux] [k] native_queued_spin_lock_slowpat
9.10% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock
4.53% io_uring_echo_s [kernel.vmlinux] [k] native_queued_spin_lock_slowpat
2.87% iou-wrk-10893 [kernel.vmlinux] [k] io_worker_handle_work
2.57% iou-wrk-10893 [kernel.vmlinux] [k] _raw_spin_lock_irqsave
2.56% io_uring_echo_s [kernel.vmlinux] [k] io_prep_async_work
1.82% io_uring_echo_s [kernel.vmlinux] [k] _raw_spin_lock
1.33% iou-wrk-10893 [kernel.vmlinux] [k] io_wqe_worker
1.26% io_uring_echo_s [kernel.vmlinux] [k] try_to_wake_up
spin_lock failure from 25.59% + 8.89% = 34.48% to 16.86% + 4.53% = 21.39%
TPS is similar, while cpu usage is from almost 400% to 350%
Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
Link: https://lore.kernel.org/r/20220206095241.121485-2-haoxu@linux.alibaba.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Clang warns:
fs/io_uring.c:9396:9: warning: variable 'ret' is uninitialized when used here [-Wuninitialized]
return ret;
^~~
fs/io_uring.c:9373:13: note: initialize the variable 'ret' to silence this warning
int fd, ret;
^
= 0
1 warning generated.
Just return 0 directly and reduce the scope of ret to the if statement,
as that is the only place that it is used, which is how the function was
before the fixes commit.
Fixes: 1a75fac9a0f9 ("io_uring: avoid ring quiesce while registering/unregistering eventfd")
Link: https://github.com/ClangBuiltLinux/linux/issues/1579
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Link: https://lore.kernel.org/r/20220207162410.1013466-1-nathan@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is done using the RCU data structure (io_ev_fd). eventfd_async is
moved from io_ring_ctx to io_ev_fd which is RCU protected hence avoiding
ring quiesce which is much more expensive than an RCU lock. The place
where eventfd_async is read is already under rcu_read_lock so there is no
extra RCU read-side critical section needed.
Signed-off-by: Usama Arif <usama.arif@bytedance.com>
Link: https://lore.kernel.org/r/20220204145117.1186568-4-usama.arif@bytedance.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is done by creating a new RCU data structure (io_ev_fd) as part of
io_ring_ctx that holds the eventfd_ctx.
The function io_eventfd_signal is executed under rcu_read_lock with a
single rcu_dereference to io_ev_fd so that if another thread unregisters
the eventfd while io_eventfd_signal is still being executed, the
eventfd_signal for which io_eventfd_signal was called completes
successfully.
The process of registering/unregistering eventfd is already done under
uring_lock so multiple threads won't enter a race condition while
registering/unregistering eventfd.
With the above approach ring quiesce can be avoided which is much more
expensive then using RCU lock. On the system tested, io_uring_register
with IORING_REGISTER_EVENTFD takes less than 1ms with RCU lock, compared
to 15ms before with ring quiesce.
Signed-off-by: Usama Arif <usama.arif@bytedance.com>
Link: https://lore.kernel.org/r/20220204145117.1186568-3-usama.arif@bytedance.com
[axboe: long line fixups]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe
Olivier Langlois
Dylan Yudaken
Stefan Roesch
Hao Xu
Nathan Chancellor
Usama Arif