On Wed, May 28, 2025 at 03:09:15PM -0400, Stefan Hajnoczi wrote:
> fdmon_ops->wait() is called with notify_me enabled. This makes it an
> expensive place to call qemu_bh_schedule() because aio_notify() invokes
> write(2) on the EventNotifier.
>
> Moving qemu_bh_schedule() after notify_me is reset improves IOPS from
> 270k to 300k IOPS with --blockdev file,aio=io_uring.
>
> I considered alternatives:
> 1. Introducing a variant of qemu_bh_schedule() that skips aio_notify().
> This only makes sense within the AioContext and fdmon implementation
> itself and is therefore a specialized internal API. I don't like
> that.
> 2. Changing fdmon_ops->wait() so implementors can reset notify_me
> themselves. This makes things complex and the other fdmon
> implementations don't need it, so it doesn't seem like a good
> solution.
>
> So in the end I moved the qemu_bh_schedule() call from fdmon-io_uring.c
> to aio-posix.c. It's ugly but straightforward.
>
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
> util/aio-posix.c | 11 +++++++++++
> util/fdmon-io_uring.c | 11 ++++++++++-
> 2 files changed, 21 insertions(+), 1 deletion(-)
Reviewed-by: Eric Blake <eblake@redhat.com>
>
> diff --git a/util/aio-posix.c b/util/aio-posix.c
> index 89bb215a2f..01428b141c 100644
> --- a/util/aio-posix.c
> +++ b/util/aio-posix.c
> @@ -693,6 +693,17 @@ bool aio_poll(AioContext *ctx, bool blocking)
> qatomic_read(&ctx->notify_me) - 2);
> }
>
> +#ifdef CONFIG_LINUX_IO_URING
> + /*
> + * This is part of fdmon-io_uring.c but it's more efficient to do it here
> + * after notify_me has been reset. That way qemu_bh_schedule() ->
> + * aio_notify() does not write the EventNotifier.
> + */
> + if (!QSIMPLEQ_EMPTY(&ctx->cqe_handler_ready_list)) {
> + qemu_bh_schedule(ctx->cqe_handler_bh);
> + }
> +#endif
> +
> aio_notify_accept(ctx);
>
> /* Calculate blocked time for adaptive polling */
> diff --git a/util/fdmon-io_uring.c b/util/fdmon-io_uring.c
> index 3a49d6a20a..03a07a4caf 100644
> --- a/util/fdmon-io_uring.c
> +++ b/util/fdmon-io_uring.c
> @@ -318,8 +318,12 @@ static bool process_cqe(AioContext *ctx,
> }
>
> cqe_handler->cqe = *cqe;
> +
> + /*
> + * aio_poll() and fdmon_io_uring_gsource_dispatch() schedule cqe_handler_bh
> + * when the list is non-empty.
> + */
> QSIMPLEQ_INSERT_TAIL(&ctx->cqe_handler_ready_list, cqe_handler, next);
> - qemu_bh_schedule(ctx->cqe_handler_bh);
> return false;
> }
>
> @@ -370,6 +374,11 @@ static void fdmon_io_uring_gsource_dispatch(AioContext *ctx,
> AioHandlerList *ready_list)
> {
> process_cq_ring(ctx, ready_list);
> +
> + /* Ensure CqeHandlers enqueued by process_cq_ring() will run */
> + if (!QSIMPLEQ_EMPTY(&ctx->cqe_handler_ready_list)) {
> + qemu_bh_schedule(ctx->cqe_handler_bh);
> + }
> }
>
> static int fdmon_io_uring_wait(AioContext *ctx, AioHandlerList *ready_list,
> --
> 2.49.0
>
>
--
Eric Blake, Principal Software Engineer
Red Hat, Inc.
Virtualization: qemu.org | libguestfs.org