Re: [PATCH v2 5/6] block/linux-aio: convert to blk_io_plug_call() API

[Stefan Hajnoczi <stefanha@xxxxxxxxxx>]

On Mon, May 29, 2023 at 10:50:34AM +0200, Stefano Garzarella wrote:
> On Wed, May 24, 2023 at 03:36:34PM -0400, Stefan Hajnoczi wrote:
> > On Wed, May 24, 2023 at 10:52:03AM +0200, Stefano Garzarella wrote:
> > > On Tue, May 23, 2023 at 01:12:59PM -0400, Stefan Hajnoczi wrote:
> > > > Stop using the .bdrv_co_io_plug() API because it is not multi-queue
> > > > block layer friendly. Use the new blk_io_plug_call() API to batch I/O
> > > > submission instead.
> > > >
> > > > Signed-off-by: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
> > > > Reviewed-by: Eric Blake <eblake@xxxxxxxxxx>
> > > > ---
> > > > include/block/raw-aio.h |  7 -------
> > > > block/file-posix.c      | 28 ----------------------------
> > > > block/linux-aio.c       | 41 +++++++++++------------------------------
> > > > 3 files changed, 11 insertions(+), 65 deletions(-)
> > > >
> > > > diff --git a/include/block/raw-aio.h b/include/block/raw-aio.h
> > > > index da60ca13ef..0f63c2800c 100644
> > > > --- a/include/block/raw-aio.h
> > > > +++ b/include/block/raw-aio.h
> > > > @@ -62,13 +62,6 @@ int coroutine_fn laio_co_submit(int fd, uint64_t 
> > > > offset, QEMUIOVector *qiov,
> > > >
> > > > void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context);
> > > > void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context);
> > > > -
> > > > -/*
> > > > - * laio_io_plug/unplug work in the thread's current AioContext, 
> > > > therefore the
> > > > - * caller must ensure that they are paired in the same IOThread.
> > > > - */
> > > > -void laio_io_plug(void);
> > > > -void laio_io_unplug(uint64_t dev_max_batch);
> > > > #endif
> > > > /* io_uring.c - Linux io_uring implementation */
> > > > #ifdef CONFIG_LINUX_IO_URING
> > > > diff --git a/block/file-posix.c b/block/file-posix.c
> > > > index 7baa8491dd..ac1ed54811 100644
> > > > --- a/block/file-posix.c
> > > > +++ b/block/file-posix.c
> > > > @@ -2550,26 +2550,6 @@ static int coroutine_fn 
> > > > raw_co_pwritev(BlockDriverState *bs, int64_t offset,
> > > >     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_WRITE);
> > > > }
> > > >
> > > > -static void coroutine_fn raw_co_io_plug(BlockDriverState *bs)
> > > > -{
> > > > -    BDRVRawState __attribute__((unused)) *s = bs->opaque;
> > > > -#ifdef CONFIG_LINUX_AIO
> > > > -    if (s->use_linux_aio) {
> > > > -        laio_io_plug();
> > > > -    }
> > > > -#endif
> > > > -}
> > > > -
> > > > -static void coroutine_fn raw_co_io_unplug(BlockDriverState *bs)
> > > > -{
> > > > -    BDRVRawState __attribute__((unused)) *s = bs->opaque;
> > > > -#ifdef CONFIG_LINUX_AIO
> > > > -    if (s->use_linux_aio) {
> > > > -        laio_io_unplug(s->aio_max_batch);
> > > > -    }
> > > > -#endif
> > > > -}
> > > > -
> > > > static int coroutine_fn raw_co_flush_to_disk(BlockDriverState *bs)
> > > > {
> > > >     BDRVRawState *s = bs->opaque;
> > > > @@ -3914,8 +3894,6 @@ BlockDriver bdrv_file = {
> > > >     .bdrv_co_copy_range_from = raw_co_copy_range_from,
> > > >     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
> > > >     .bdrv_refresh_limits = raw_refresh_limits,
> > > > -    .bdrv_co_io_plug        = raw_co_io_plug,
> > > > -    .bdrv_co_io_unplug      = raw_co_io_unplug,
> > > >     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
> > > >
> > > >     .bdrv_co_truncate                   = raw_co_truncate,
> > > > @@ -4286,8 +4264,6 @@ static BlockDriver bdrv_host_device = {
> > > >     .bdrv_co_copy_range_from = raw_co_copy_range_from,
> > > >     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
> > > >     .bdrv_refresh_limits = raw_refresh_limits,
> > > > -    .bdrv_co_io_plug        = raw_co_io_plug,
> > > > -    .bdrv_co_io_unplug      = raw_co_io_unplug,
> > > >     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
> > > >
> > > >     .bdrv_co_truncate                   = raw_co_truncate,
> > > > @@ -4424,8 +4400,6 @@ static BlockDriver bdrv_host_cdrom = {
> > > >     .bdrv_co_pwritev        = raw_co_pwritev,
> > > >     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
> > > >     .bdrv_refresh_limits    = cdrom_refresh_limits,
> > > > -    .bdrv_co_io_plug        = raw_co_io_plug,
> > > > -    .bdrv_co_io_unplug      = raw_co_io_unplug,
> > > >     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
> > > >
> > > >     .bdrv_co_truncate                   = raw_co_truncate,
> > > > @@ -4552,8 +4526,6 @@ static BlockDriver bdrv_host_cdrom = {
> > > >     .bdrv_co_pwritev        = raw_co_pwritev,
> > > >     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
> > > >     .bdrv_refresh_limits    = cdrom_refresh_limits,
> > > > -    .bdrv_co_io_plug        = raw_co_io_plug,
> > > > -    .bdrv_co_io_unplug      = raw_co_io_unplug,
> > > >     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
> > > >
> > > >     .bdrv_co_truncate                   = raw_co_truncate,
> > > > diff --git a/block/linux-aio.c b/block/linux-aio.c
> > > > index 442c86209b..5021aed68f 100644
> > > > --- a/block/linux-aio.c
> > > > +++ b/block/linux-aio.c
> > > > @@ -15,6 +15,7 @@
> > > > #include "qemu/event_notifier.h"
> > > > #include "qemu/coroutine.h"
> > > > #include "qapi/error.h"
> > > > +#include "sysemu/block-backend.h"
> > > >
> > > > /* Only used for assertions.  */
> > > > #include "qemu/coroutine_int.h"
> > > > @@ -46,7 +47,6 @@ struct qemu_laiocb {
> > > > };
> > > >
> > > > typedef struct {
> > > > -    int plugged;
> > > >     unsigned int in_queue;
> > > >     unsigned int in_flight;
> > > >     bool blocked;
> > > > @@ -236,7 +236,7 @@ static void 
> > > > qemu_laio_process_completions_and_submit(LinuxAioState *s)
> > > > {
> > > >     qemu_laio_process_completions(s);
> > > >
> > > > -    if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
> > > > +    if (!QSIMPLEQ_EMPTY(&s->io_q.pending)) {
> > > >         ioq_submit(s);
> > > >     }
> > > > }
> > > > @@ -277,7 +277,6 @@ static void qemu_laio_poll_ready(EventNotifier 
> > > > *opaque)
> > > > static void ioq_init(LaioQueue *io_q)
> > > > {
> > > >     QSIMPLEQ_INIT(&io_q->pending);
> > > > -    io_q->plugged = 0;
> > > >     io_q->in_queue = 0;
> > > >     io_q->in_flight = 0;
> > > >     io_q->blocked = false;
> > > > @@ -354,31 +353,11 @@ static uint64_t laio_max_batch(LinuxAioState *s, 
> > > > uint64_t dev_max_batch)
> > > >     return max_batch;
> > > > }
> > > >
> > > > -void laio_io_plug(void)
> > > > +static void laio_unplug_fn(void *opaque)
> > > > {
> > > > -    AioContext *ctx = qemu_get_current_aio_context();
> > > > -    LinuxAioState *s = aio_get_linux_aio(ctx);
> > > > +    LinuxAioState *s = opaque;
> > > >
> > > > -    s->io_q.plugged++;
> > > > -}
> > > > -
> > > > -void laio_io_unplug(uint64_t dev_max_batch)
> > > > -{
> > > > -    AioContext *ctx = qemu_get_current_aio_context();
> > > > -    LinuxAioState *s = aio_get_linux_aio(ctx);
> > > > -
> > > > -    assert(s->io_q.plugged);
> > > > -    s->io_q.plugged--;
> > > > -
> > > > -    /*
> > > > -     * Why max batch checking is performed here:
> > > > -     * Another BDS may have queued requests with a higher 
> > > > dev_max_batch and
> > > > -     * therefore in_queue could now exceed our dev_max_batch. Re-check 
> > > > the max
> > > > -     * batch so we can honor our device's dev_max_batch.
> > > > -     */
> > > > -    if (s->io_q.in_queue >= laio_max_batch(s, dev_max_batch) ||
> > > 
> > > Why are we removing this condition?
> > > Could the same situation occur with the new API?
> > 
> > The semantics of unplug_fn() are different from .bdrv_co_unplug():
> > 1. unplug_fn() is only called when the last blk_io_unplug() call occurs,
> >   not every time blk_io_unplug() is called.
> > 2. unplug_fn() is per-thread, not per-BlockDriverState, so there is no
> >   way to get per-BlockDriverState fields like dev_max_batch.
> > 
> > Therefore this condition cannot be moved to laio_unplug_fn().
> 
> I see now.
> 
> > 
> > How important is this condition? I believe that dropping it does not
> > have much of an effect but maybe I missed something.
> 
> With Kevin we agreed to add it to avoid extra latency in some devices,
> but we didn't do much testing on this.
> 
> IIRC what solved the performance degradation was the check in
> laio_do_submit() that we still have after this changes.
> 
> So it may not have much effect, but maybe it's worth mentioning in
> the commit description.

I'll update the commit description.

> > 
> > Also, does it make sense to define per-BlockDriverState batching limits
> > when the AIO engine (Linux AIO or io_uring) is thread-local and shared
> > between all BlockDriverStates? I believe the fundamental reason (that we
> > discovered later) why dev_max_batch is effective is because the Linux
> > kernel processes 32 I/O request submissions at a time. Anything above 32
> > adds latency without a batching benefit.
> 
> This is a good point, maybe we should confirm it with some tests though.

Yes, I would benchmark it. Also, switching to per-thread max_batch
involves a command-line interface change and we still need to keep
aio-max-batch for compatibility for some time, so it's not urgent.

Stefan

Attachment: signature.asc
Description: PGP signature