Re: [Qemu-devel] [PATCH 3/3] linux-aio: process completions from ioq_submit()

[Paolo Bonzini <pbonzini@redhat.com>]

On 19/07/2016 13:18, Roman Penyaev wrote:
>> > No need for this new field.  You can just do nothing here and check
>> > laiocb.ret == -EINPROGRESS here in laio_co_submit.
> 
> I have thought but did not like it, because we depend on the value,
> which kernel writes there.

(The kernel actually writes to ev->res).

> If kernel by chance writes -EINPROGRESS
> (whatever that means, bug in some ll driver?) we are screwed up.
> But probably that is my paranoia.

Understood.  However, QEMU relies elsewhere on EINPROGRESS not being
returned for file I/O, so I think it's safe.

> Also, I hope (I do not know how to reproduce this, virtio_blk does not nest),
> that we are allowed to nest (calling aio_poll() and all this machinery) from
> co-routine.

Hmm, good question.  The nesting scenario originally happened exactly
from a coroutine (commit_run), but I suspect that it cannot happen
anymore since we've introduced block_job_defer_to_main_loop and
bdrv_co_drain.  In any case your patch wouldn't change that.

Paolo

> Are we?  We can lead to deep nesting inside the following sequence:
> ioq_submit() -> complete() -> aio_poll() -> ioq_submit() -> complete() ...
> but that can happen of course even without this patch.  I would say this nesting
> is a clumsy stuff.

Paolo