Re: [PATCH] aio-posix: split poll check from ready handler

[Stefano Garzarella]

On Thu, Dec 02, 2021 at 01:09:23PM +0000, Stefan Hajnoczi wrote:
> On Wed, Dec 01, 2021 at 12:55:08PM +0100, Stefano Garzarella wrote:
> > On Tue, Nov 30, 2021 at 11:20:57AM +0000, Stefan Hajnoczi wrote:
> > > @@ -657,10 +704,7 @@ bool aio_poll(AioContext *ctx, bool blocking)
> > >     }
> > >
> > >     progress |= aio_bh_poll(ctx);
> > > -
> > > -    if (ret > 0) {
> > > -        progress |= aio_dispatch_ready_handlers(ctx, &ready_list);
> > > -    }
> > > +    progress |= aio_dispatch_ready_handlers(ctx, &ready_list);
> >
> > I honestly don't know this code well, so this change isn't clear to me.
> >
> > IIUC this is necessary because we now use REVENTS_POLL_READY, so even if
> > `wait` is interrupted, we want to run the handlers ready so far.
> > Or maybe is it an unrelated optimization?
>
> Previously we didn't need to run aio_dispatch_ready_handlers() in
> polling mode (ret would be 0). Now we do, so the conditional needs to be
> removed.
>
> There is no change for the ctx->fdmon_ops->wait() -EINTR case. fdmon
> won't add handlers to ready_list when returning -EINTR and previously
> polling would run handlers immediately. In order to preserve that
> behavior we now need to call aio_dispatch_ready_handlers()
> unconditionally because polling may have added handlers.

Yeah, now I see.
Aside from not knowing the code, I had only considered the ret < 0 case.

Thanks for clarifying,
Stefano