When ->poll() succeeds the AioHandler is placed on the ready list with
revents set to the magic value 0. This magic value causes
aio_dispatch_handler() to invoke ->poll_ready() instead of ->io_read()
for G_IO_IN or ->io_write() for G_IO_OUT.
This magic value 0 hack works for the IOThread where AioHandlers are
placed on ->ready_list and processed by aio_dispatch_ready_handlers().
It does not work for the main loop where all AioHandlers are processed
by aio_dispatch_handlers(), even those that are not ready and have a
revents value of 0.
As a result the main loop invokes ->poll_ready() on AioHandlers that are
not ready. These spurious ->poll_ready() calls waste CPU cycles and
could lead to crashes if the code assumes ->poll() must have succeeded
before ->poll_ready() is called (a reasonable asumption but I haven't
seen it in practice).
Stop using revents to track whether ->poll_ready() will be called on an
AioHandler. Introduce a separate AioHandler->poll_ready field instead.
This eliminates spurious ->poll_ready() calls in the main loop.
Fixes: 826cc32423 ("aio-posix: split poll check from ready handler")
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reported-by: Jason Wang <jasowang@redhat.com>
Tested-by: Jason Wang <jasowang@redhat.com>
Message-id: 20220223155703.136833-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Adaptive polling measures the execution time of the polling check plus
handlers called when a polled event becomes ready. Handlers can take a
significant amount of time, making it look like polling was running for
a long time when in fact the event handler was running for a long time.
For example, on Linux the io_submit(2) syscall invoked when a virtio-blk
device's virtqueue becomes ready can take 10s of microseconds. This
can exceed the default polling interval (32 microseconds) and cause
adaptive polling to stop polling.
By excluding the handler's execution time from the polling check we make
the adaptive polling calculation more accurate. As a result, the event
loop now stays in polling mode where previously it would have fallen
back to file descriptor monitoring.
The following data was collected with virtio-blk num-queues=2
event_idx=off using an IOThread. Before:
168k IOPS, IOThread syscalls:
9837.115 ( 0.020 ms): IO iothread1/620155 io_submit(ctx_id: 140512552468480, nr: 16, iocbpp: 0x7fcb9f937db0) = 16
9837.158 ( 0.002 ms): IO iothread1/620155 write(fd: 103, buf: 0x556a2ef71b88, count: 8) = 8
9837.161 ( 0.001 ms): IO iothread1/620155 write(fd: 104, buf: 0x556a2ef71b88, count: 8) = 8
9837.163 ( 0.001 ms): IO iothread1/620155 ppoll(ufds: 0x7fcb90002800, nfds: 4, tsp: 0x7fcb9f1342d0, sigsetsize: 8) = 3
9837.164 ( 0.001 ms): IO iothread1/620155 read(fd: 107, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.174 ( 0.001 ms): IO iothread1/620155 read(fd: 105, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.176 ( 0.001 ms): IO iothread1/620155 read(fd: 106, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.209 ( 0.035 ms): IO iothread1/620155 io_submit(ctx_id: 140512552468480, nr: 32, iocbpp: 0x7fca7d0cebe0) = 32
174k IOPS (+3.6%), IOThread syscalls:
9809.566 ( 0.036 ms): IO iothread1/623061 io_submit(ctx_id: 140539805028352, nr: 32, iocbpp: 0x7fd0cdd62be0) = 32
9809.625 ( 0.001 ms): IO iothread1/623061 write(fd: 103, buf: 0x5647cfba5f58, count: 8) = 8
9809.627 ( 0.002 ms): IO iothread1/623061 write(fd: 104, buf: 0x5647cfba5f58, count: 8) = 8
9809.663 ( 0.036 ms): IO iothread1/623061 io_submit(ctx_id: 140539805028352, nr: 32, iocbpp: 0x7fd0d0388b50) = 32
Notice that ppoll(2) and eventfd read(2) syscalls are eliminated because
the IOThread stays in polling mode instead of falling back to file
descriptor monitoring.
As usual, polling is not implemented on Windows so this patch ignores
the new io_poll_read() callback in aio-win32.c.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Message-id: 20211207132336.36627-2-stefanha@redhat.com
[Fixed up aio_set_event_notifier() calls in
tests/unit/test-fdmon-epoll.c added after this series was queued.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
When there are many poll handlers it's likely that some of them are idle
most of the time. Remove handlers that haven't had activity recently so
that the polling loop scales better for guests with a large number of
devices.
This feature only takes effect for the Linux io_uring fd monitoring
implementation because it is capable of combining fd monitoring with
userspace polling. The other implementations can't do that and risk
starving fds in favor of poll handlers, so don't try this optimization
when they are in use.
IOPS improves from 10k to 105k when the guest has 100
virtio-blk-pci,num-queues=32 devices and 1 virtio-blk-pci,num-queues=1
device for rw=randread,iodepth=1,bs=4k,ioengine=libaio on NVMe.
[Clarified aio_poll_handlers locking discipline explanation in comment
after discussion with Paolo Bonzini <pbonzini@redhat.com>.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-8-stefanha@redhat.com
Message-Id: <20200305170806.1313245-8-stefanha@redhat.com>
The recent Linux io_uring API has several advantages over ppoll(2) and
epoll(2). Details are given in the source code.
Add an io_uring implementation and make it the default on Linux.
Performance is the same as with epoll(7) but later patches add
optimizations that take advantage of io_uring.
It is necessary to change how aio_set_fd_handler() deals with deleting
AioHandlers since removing monitored file descriptors is asynchronous in
io_uring. fdmon_io_uring_remove() marks the AioHandler deleted and
aio_set_fd_handler() will let it handle deletion in that case.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-6-stefanha@redhat.com
Message-Id: <20200305170806.1313245-6-stefanha@redhat.com>
The ppoll(2) and epoll(7) file descriptor monitoring implementations are
mixed with the core util/aio-posix.c code. Before adding another
implementation for Linux io_uring, extract out the existing
ones so there is a clear interface and the core code is simpler.
The new interface is AioContext->fdmon_ops, a pointer to a FDMonOps
struct. See the patch for details.
Semantic changes:
1. ppoll(2) now reflects events from pollfds[] back into AioHandlers
while we're still on the clock for adaptive polling. This was
already happening for epoll(7), so if it's really an issue then we'll
need to fix both in the future.
2. epoll(7)'s fallback to ppoll(2) while external events are disabled
was broken when the number of fds exceeded the epoll(7) upgrade
threshold. I guess this code path simply wasn't tested and no one
noticed the bug. I didn't go out of my way to fix it but the correct
code is simpler than preserving the bug.
I also took some liberties in removing the unnecessary
AioContext->epoll_available (just check AioContext->epollfd != -1
instead) and AioContext->epoll_enabled (it's implicit if our
AioContext->fdmon_ops callbacks are being invoked) fields.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-4-stefanha@redhat.com
Message-Id: <20200305170806.1313245-4-stefanha@redhat.com>
