Re: [Qemu-devel] [PATCH 07/17] nbd: convert to use qio_channel_yield

[Stefan Hajnoczi]

On Mon, Jan 30, 2017 at 04:18:10PM -0500, Paolo Bonzini wrote:
> On 30/01/2017 10:50, Stefan Hajnoczi wrote:
> > On Fri, Jan 20, 2017 at 05:43:12PM +0100, Paolo Bonzini wrote:
> >> +        aio_co_wake(s->recv_coroutine[i]);
> >>  
> >> -    qemu_coroutine_enter(nbd_get_client_session(bs)->send_coroutine);
> >> +        /* We're woken up by the recv_coroutine itself.  */
> >> +        qemu_coroutine_yield();
> > 
> > This relies on recv_coroutine() entering us only after we've yielded -
> > otherwise QEMU will crash.  The code and comments don't make it obvious
> > why this is guaranteed to be safe.
> 
> It doesn't rely on that (that's the magic hidden behind aio_co_wake),
> but you're right there is a documentation problem because I needed 10
> minutes to remind myself why it's correct.
>
> It works because neither coroutine is moving context:
> 
> - if the two coroutines are in the same context, aio_co_wake queues the
> coroutine for execution after the yield, which is obviously okay;
> 
> - if they are in different context, the recv_coroutine's aio_co_wake
> queues the read_reply_co with aio_co_schedule.  It is then woken up
> through a bottom half, which executes only after read_reply has yielded.
> 
> It is implied by the aio_co_wake and aio_co_schedule documentation; all
> requirements are satisfied:
> 
> 1) aio_co_wake's comment says:
> 
>    * aio_co_wake may be executed either in coroutine or non-coroutine
>    * context.  The coroutine must not be entered by anyone else while
>    * aio_co_wake() is active.
> 
>    read_reply_co is only woken by one recv_coroutine at a time
> 
> 2) And for the case where read_reply_co and recv_coroutine are in
>    different contexts, aio_co_schedule's comment says:
> 
>    * In addition the coroutine must have yielded unless ctx
>    * is the context in which the coroutine is running (i.e. the value of
>    * qemu_get_current_aio_context() from the coroutine itself).
> 
>    which is okay because aio_co_wake always uses "the context in
>    which the coroutine is running" as the argument to aio_co_schedule.
> 
> Any suggestions on how to document this properly?  I'm not sure a
> comment in the NBD driver is the best place, because similar logic will
> appear soon in other networked block drivers.

Maybe add a new function that just does these two calls.  I don't have a
good name for it.

Stefan

signature.asc
Description: PGP signature