Early work-in-progress io_uring support #1356

dmantipov · 2022-10-05T08:28:47Z

This code aims to provide an initial and very basic support for io_uring(7), a Linux-specific API for asynchronous I/O. Currently this work aims to provide an io_uring-based I/O for socket-based bufferevents. The most important issue of the whole thing is that bufferevent interface assumes the synchronous callbacks, somewhat similar to the convenient read() and write() system calls, but the whole io_uring subsystem is designed to be (mostly) asynchronous. Current approach is designed in attempt to avoid changing an external libevent API and basically works as follows:

evbuffer_io_uring_submit_rwv() should get and SQE, prepare it to issue I/O request, submit to the ring, and call event_base_active_by_fd() to enforce the next iteration of (per-base) event processing loop.
Being woken up to serve an event processing loop, event_process_active_single_queue() issues io_uring_wait_cqe_timeout() to fetch CQE and finally calls event_process_cqe() to complete an I/O.

I would strongly suggests everyone interested in this feature to try this code, run sample/bufferevent-benchmark test, share the numbers and maybe even an ideas on further improvements of this code, whatever.

Signed-off-by: Dmitry Antipov dantipov@cloudlinux.com

Fixes: #1019

azat · 2022-10-06T10:03:49Z

Great!

An idea is that an event_base should auto-tune itself to match the workload.

Actually this is not that important, and maybe it is better to do it separately, since right now there is no such auto tunning.

That is, io_uring makes no sense if there are just a few in-flight evbuffers for the most of the time and so individual readv()/writev() calls are enough

Yes, but it will also not harm, right?
So maybe yo can go with a simpler implementation that we can merge first and then work in some auto-tuning algorithms.

dmantipov · 2022-10-07T06:45:00Z

Is it possible to install liburing-2.2 on a few Linux systems running CI jobs? Although current io_uring quirks are not ready to pass all of the make verify tests, this may change in the near future :-).

azat · 2022-10-07T20:29:34Z

Is it possible to install liburing-2.2 on a few Linux systems running CI jobs?

Sure, just modify:

libevent/.github/workflows/build.yml

Line 76 in b5b4c7f

sudo apt-get install -y libmbedtls-dev
libevent/.github/workflows/build.yml

Line 202 in b5b4c7f

sudo apt-get install -y libmbedtls-dev

dmantipov · 2022-10-10T12:43:42Z

I've relaxed liburing requirements to >= 2.0, it should work as well.

dmantipov · 2022-10-13T15:13:12Z

NOTE: in the last version of the patch, io_uring is turned off by default even if supported, and it's required to call event_base_new_with_config(cfg) with enabled event_config_set_flag(cfg, EVENT_BASE_FLAG_IO_URING) flag.

azat · 2022-10-23T17:39:02Z

@dmantipov do you have some workload that can benefit from this?

azat · 2022-10-23T17:40:44Z

.github/workflows/build.yml

@@ -73,7 +73,7 @@ jobs:
      - name: Install Depends
        run: |
          sudo apt-get update
-          sudo apt-get install -y libmbedtls-dev
+          sudo apt-get install -y libmbedtls-dev liburing-dev


Hm, there is no liburing-dev in ubuntu 18.04, let's switch to ubuntu-22.04

azat · 2022-10-23T20:11:46Z

event-internal.h

+#ifdef EVENT__HAVE_LIBURING
+	struct io_uring io_uring;
+	unsigned io_uring_sqe_count;
+	struct __kernel_timespec io_uring_cqe_timeout;


Consider using plain timespec instead of private __kernel_timespec

azat · 2022-10-23T20:13:00Z

event.c

+			base->io_uring_cqe_timeout.tv_nsec = 0;
+		}
+		if (io_uring_queue_init(queue_size, &base->io_uring, 0) == 0)
+			base->flags |= EVENT_BASE_FLAG_IO_URING;


An error if it fails?

azat · 2022-10-23T20:19:35Z

event.c

@@ -1769,6 +1841,14 @@ event_process_active_single_queue(struct event_base *base,
 		if (base->event_continue)
 			break;
 	}
+#ifdef EVENT__HAVE_LIBURING
+	/* FIXME: handle unfinished I/O requests properly. */
+	if (base->io_uring_sqe_count > 0) {


Why this is bad? It can be handled in another iteration

azat · 2022-10-23T20:21:30Z

event.c

@@ -1755,6 +1817,16 @@ event_process_active_single_queue(struct event_base *base,
 		}
 #endif

+#ifdef EVENT__HAVE_LIBURING
+		if (base->io_uring_sqe_count) {


Hm, interface looks clumsy:

checking io_uring_sqe_count after each event looks odd, it should be checked in the event_base_loop, or it should be binded to specific event and processed only when the callback for that particular event is called (one way is to add new callback for event in case of uring, however a) this will bloat the structure and b) it will break ABI compatibility - even though it is highly not recommended to use event structure directly)

io_uring_sqe_count is set from the buffer.c, and checked here, how about adding some wrappers for this?

io_uring_buffers modified directly from the bufferevent code, how about adding some wrappers for this?

azat · 2022-10-23T20:31:28Z

buffer.c

+		else {
+			/* Caller might want to check errno, which
+			   is never used by liburing itself. */
+			n = -1, errno = -buf->io_uring_result;


Suggested change

n = -1, errno = -buf->io_uring_result;

n = -1;

errno = -buf->io_uring_result;

azat · 2022-10-23T20:32:55Z

buffer.c

+		else {
+			/* Caller might want to check errno, which
+			   is never used by liburing itself. */
+			n = -1, errno = -buffer->io_uring_result;


Suggested change

n = -1, errno = -buffer->io_uring_result;

n = -1;

errno = -buffer->io_uring_result;

azat · 2022-10-23T20:35:35Z

buffer.c

+		buf->io_uring_status = IO_URING_NONE;
+		goto complete;
+	} else if (buf->io_uring_status == IO_URING_RUNNING) {
+		result = -1, errno = EINPROGRESS;


Suggested change

result = -1, errno = EINPROGRESS;

result = -1;

errno = EINPROGRESS;

azat · 2022-10-23T20:35:51Z

buffer.c

+		buffer->io_uring_status = IO_URING_NONE;
+		goto complete;
+	} else if (buffer->io_uring_status == IO_URING_RUNNING) {
+		n = -1, errno = EINPROGRESS;


Suggested change

n = -1, errno = EINPROGRESS;

n = -1;

errno = EINPROGRESS;

azat · 2022-10-23T20:38:44Z

I will mark it as draft for now.

azat · 2022-10-24T20:45:07Z

CI failures:

/usr/bin/ld: lib/libevent_core-2.2.so.1.0.0: undefined reference to `io_uring_queue_init'
/usr/bin/ld: lib/libevent_core-2.2.so.1.0.0: undefined reference to `io_uring_submit'
/usr/bin/ld: lib/libevent_core-2.2.so.1.0.0: undefined reference to `io_uring_queue_exit'
/usr/bin/ld: lib/libevent_core-2.2.so.1.0.0: undefined reference to `io_uring_get_sqe'
/usr/bin/ld: lib/libevent_core-2.2.so.1.0.0: undefined reference to `io_uring_wait_cqe_timeout'

But by some reason one build with clang passed - https://github.com/libevent/libevent/actions/runs/3308507890/jobs/5460943018
Any clue?

Signed-off-by: Dmitry Antipov <dantipov@cloudlinux.com>

dmantipov force-pushed the early-work-in-progress-io-uring-support branch 2 times, most recently from 6a7036d to 7644b74 Compare October 10, 2022 10:50

dmantipov force-pushed the early-work-in-progress-io-uring-support branch 3 times, most recently from 8facc5c to 626677e Compare October 13, 2022 15:10

dmantipov force-pushed the early-work-in-progress-io-uring-support branch 2 times, most recently from b1e531f to 0241dc3 Compare October 18, 2022 09:05

azat reviewed Oct 23, 2022

View reviewed changes

azat marked this pull request as draft October 23, 2022 20:39

azat force-pushed the early-work-in-progress-io-uring-support branch from 0241dc3 to a575de7 Compare October 23, 2022 20:40

eakraly mentioned this pull request Oct 24, 2022

question : any plans to implement recent io_uring support for turn relay ?? coturn/coturn#1034

Open

vi mentioned this pull request Nov 16, 2022

Are there plans to use io_uring on Linux? coturn/coturn#591

Closed

dmantipov closed this Dec 8, 2022

dmantipov force-pushed the early-work-in-progress-io-uring-support branch from a575de7 to d8ecb88 Compare December 8, 2022 16:27

dmantipov reopened this Dec 8, 2022

dmantipov force-pushed the early-work-in-progress-io-uring-support branch from 97b88ad to bfe3f23 Compare December 9, 2022 08:58

dmantipov force-pushed the early-work-in-progress-io-uring-support branch from bfe3f23 to ee65796 Compare December 20, 2022 09:43

dmantipov force-pushed the early-work-in-progress-io-uring-support branch from ee65796 to 68d51a3 Compare December 30, 2022 09:52

Early work-in-progress io_uring(7) support

b528661

Signed-off-by: Dmitry Antipov <dantipov@cloudlinux.com>

dmantipov force-pushed the early-work-in-progress-io-uring-support branch from 68d51a3 to b528661 Compare January 13, 2023 09:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Early work-in-progress io_uring support #1356

Early work-in-progress io_uring support #1356

dmantipov commented Oct 5, 2022 •

edited by azat

azat commented Oct 6, 2022

dmantipov commented Oct 7, 2022

azat commented Oct 7, 2022

dmantipov commented Oct 10, 2022

dmantipov commented Oct 13, 2022

azat commented Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat Oct 23, 2022

azat commented Oct 23, 2022

azat commented Oct 24, 2022

	n = -1, errno = -buf->io_uring_result;
	n = -1;
	errno = -buf->io_uring_result;

	n = -1, errno = -buffer->io_uring_result;
	n = -1;
	errno = -buffer->io_uring_result;

	result = -1, errno = EINPROGRESS;
	result = -1;
	errno = EINPROGRESS;

Early work-in-progress io_uring support #1356

Are you sure you want to change the base?

Early work-in-progress io_uring support #1356

Conversation

dmantipov commented Oct 5, 2022 • edited by azat

azat commented Oct 6, 2022

dmantipov commented Oct 7, 2022

azat commented Oct 7, 2022

dmantipov commented Oct 10, 2022

dmantipov commented Oct 13, 2022

azat commented Oct 23, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

azat commented Oct 23, 2022

azat commented Oct 24, 2022

dmantipov commented Oct 5, 2022 •

edited by azat