Re: [Qemu-devel] [RFC PATCH 0/4] aio: Don't poll ioeventfd in nested aio_poll()

[Fam Zheng]

On Thu, 05/28 14:01, Paolo Bonzini wrote:
> 
> 
> On 28/05/2015 13:49, Fam Zheng wrote:
> > On Thu, 05/28 13:19, Paolo Bonzini wrote:
> >>
> >>
> >> On 28/05/2015 13:16, Fam Zheng wrote:
> >>>> On 28/05/2015 03:46, Fam Zheng wrote:
> >>>>> The main context uses iohandler and aio_dispatch, neither calls
> >>>>> aio_set_dispatching(). However, if we have [2], they can be changed to
> >>>>> aio_poll(), then would this idea work?
> >>>>
> >>>> I think it's a bad idea to handle aio_poll for context B in a different
> >>>> way, just because you have an outer aio_poll for context A...
> >>>
> >>> But we already do something similar: ignoring slirp, main_loop_wait() is 
> >>> like
> >>> an iothread aio_poll() without the "outermost differentiation", while the
> >>> current aio_poll() in bdrv_drain() is roughly "main_loop_wait() minus
> >>> iohandlers".
> >>
> >> Right, but the two sets of iohandlers are stored in different places, so
> >> it's obvious that you don't execute all of them.  On the other hand,
> >> examining global state in aio_poll is really bad.
> >>
> > 
> > OK.
> > 
> > Would moving the ioeventfds to a new top level aio_loop_wait() be any 
> > better?
> > That way no global state is needed.
> 
> If we need pause/resume anyway due to block/mirror.c's use of
> block_job_defer_to_main_loop, I think this is not a problem anymore?
> 

I also hope to dedup the iohandler code with async.c, on top of [2]; and in the
longer term, convert slirp to use AioContext API too, so that all
*_pollfds_fill() will not be necessary - the whole event loop goes epoll style.

Fam