Re: [Qemu-devel] [PATCH 04/16] aio: only call aio_poll_internal from iothread

[Eric Blake]

On 02/08/2016 09:14 AM, Paolo Bonzini wrote:
> aio_poll is not thread safe; it can report progress incorrectly when
> called from the main thread.  The bug remains latent as long as
> all of it is called within aio_context_acquire/aio_context_release,
> but this will change soon.
> 
> The details of the bug are pretty simple, but fixing it in an
> efficient way is thorny.  There are plenty of comments and formal
> models in the patch, so I will refer to it.
> 
> Signed-off-by: Paolo Bonzini <address@hidden>
> ---

> +++ b/async.c
> @@ -300,12 +300,224 @@ void aio_notify_accept(AioContext *ctx)
>      }
>  }
>  
> +/* aio_poll_internal is not thread-safe; it only reports progress
> + * correctly when called from one thread, because it has no
> + * history of what happened in different threads.  When called
> + * from two threads, there is a race:
> + *
> + *      main thread                       I/O thread
> + *      -----------------------           --------------------------
> + *      blk_drain
> + *        bdrv_requests_pending -> true
> + *                                        aio_poll_internal
> + *                                          process last request
> + *        aio_poll_internal
> + *
> + * Now aio_poll_internal will never exit, because there is no pending
> + * I/O on the AioContext.
> + *
> + * Therefore, aio_poll is a wrapper around aio_poll_internal that allows
> + * usage from _two_ threads: the I/O thread of course, and the main thread.
> + * When called from the main thread, aio_poll just asks the I/O thread
> + * for a nudge as soon as the next call to aio_poll is complete.
> + * Because we use QemuEvent, and QemuEvent supports a single consumer
> + * only, this only works when the calling thread holds the big QEMU lock.
> + *
> + * Because aio_poll is used in a loop, spurious wakeups are okay.
> + * Therefore, the I/O thread calls qemu_event_set very liberally
> + * (it helps that qemu_event_set is cheap on an already-set event).
> + * generally used in a loop, it's okay to have spurious wakeups.

Incomplete sentence due to bad rebase leftovers?

> + * Similarly it is okay to return true when no progress was made
> + * (as long as this doesn't happen forever, or you get livelock).
> + *
> +


-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature