Re: [PATCH] block/rbd: Do not use BDS's AioContext

[Stefan Hajnoczi <stefanha@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 5117 bytes --]

On Wed, Feb 12, 2025 at 03:27:26PM +0100, Hanna Czenczek wrote:
> On 12.02.25 14:26, Kevin Wolf wrote:
> > Am 12.02.2025 um 10:32 hat Hanna Czenczek geschrieben:
> > > RBD schedules the request completion code (qemu_rbd_finish_bh()) to run
> > > in the BDS's AioContext.  The intent seems to be to run it in the same
> > > context that the original request coroutine ran in, i.e. the thread on
> > > whose stack the RBDTask object exists (see qemu_rbd_start_co()).
> > > 
> > > However, with multiqueue, that thread is not necessarily the same as the
> > > BDS's AioContext.  Instead, we need to remember the actual AioContext
> > > and schedule the completion BH there.
> > > 
> > > Buglink:https://issues.redhat.com/browse/RHEL-67115
> > Please add a short summary of what actually happens to the commit
> > message. I had to check the link to remember what the symptoms are.
> 
> Sure.  The problem is, I don’t know exactly what’s going wrong (looked like
> a coroutine being rescheduled after it was already done), and I don’t know
> how this fixes it.
> 
> > > Reported-by: Junyao Zhao<junzhao@redhat.com>
> > > Signed-off-by: Hanna Czenczek<hreitz@redhat.com>
> > > ---
> > > I think I could also drop RBDTask.ctx and just use
> > > `qemu_coroutine_get_aio_context(RBDTask.co)` instead, but this is the
> > > version of the patch that was tested and confirmed to fix the issue (I
> > > don't have a local reproducer), so I thought I'll post this first.
> > Did  you figure out why it even makes a difference in which thread
> > qemu_rbd_finish_bh() runs? For context:
> > 
> >      static void qemu_rbd_finish_bh(void *opaque)
> >      {
> >          RBDTask *task = opaque;
> >          task->complete = true;
> >          aio_co_wake(task->co);
> >      }
> > 
> > This looks as if it should be working in any thread, except maybe for a
> > missing barrier after updating task->complete - but I think the failure
> > mode for that would be a hang in qemu_rbd_start_co().
> 
> Yes, I thought the same thing.  All I could imagine was that maybe reading
> task->co returns the wrong result, but given how long ago that must have
> been set, it seems quite unlikely (to say the least).  In addition,
> qemu_rbd_completion_cb() already reads the object from a different thread,
> and that seems to work fine.
> 
> Really, all I know is that the notion of a BDS’s AioContext no longer makes
> sense in a multiqueue I/O path, so this should be scheduled in the I/O’s
> AioContext (just conceptually speaking), and that this seems to fix the bug.

Scheduling the BH directly in the task's AioContext should help
performance, so I think this patch makes sense irrespective of the bug.
Performance results would be needed to justify that though.

With regards to the bug, I'm also not sure where the issue lies. The
root cause needs to be understood so that we can be sure that this patch
is really a fix and not just papering over a bug that is still reachable
in other ways.

> 
> > >   block/rbd.c | 10 ++++++----
> > >   1 file changed, 6 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/block/rbd.c b/block/rbd.c
> > > index af984fb7db..9d4e0817e0 100644
> > > --- a/block/rbd.c
> > > +++ b/block/rbd.c
> > > @@ -102,7 +102,7 @@ typedef struct BDRVRBDState {
> > >   } BDRVRBDState;
> > >   typedef struct RBDTask {
> > > -    BlockDriverState *bs;
> > > +    AioContext *ctx;
> > >       Coroutine *co;
> > >       bool complete;
> > >       int64_t ret;
> > > @@ -1269,8 +1269,7 @@ static void qemu_rbd_completion_cb(rbd_completion_t c, RBDTask *task)
> > >   {
> > >       task->ret = rbd_aio_get_return_value(c);
> > >       rbd_aio_release(c);
> > > -    aio_bh_schedule_oneshot(bdrv_get_aio_context(task->bs),
> > > -                            qemu_rbd_finish_bh, task);
> > > +    aio_bh_schedule_oneshot(task->ctx, qemu_rbd_finish_bh, task);
> > >   }
> > >   static int coroutine_fn qemu_rbd_start_co(BlockDriverState *bs,
> > > @@ -1281,7 +1280,10 @@ static int coroutine_fn qemu_rbd_start_co(BlockDriverState *bs,
> > >                                             RBDAIOCmd cmd)
> > >   {
> > >       BDRVRBDState *s = bs->opaque;
> > > -    RBDTask task = { .bs = bs, .co = qemu_coroutine_self() };
> > > +    RBDTask task = {
> > > +        .ctx = qemu_get_current_aio_context(),
> > > +        .co = qemu_coroutine_self(),
> > > +    };
> > >       rbd_completion_t c;
> > >       int r;
> > Nothing wrong I can see about the change, but I don't understand why it
> > fixes the problem.
> 
> Me neither.  But if this patch had been part of one of the original
> multiqueue series (without pointing out the linked bug), would there have
> been any argument against it?
> 
> Indeed it is a problem that I don’t understand what’s happening. But even
> more honestly, I’ll have to admit I can’t ever claim to understand what’s
> happening in a multi-threaded asynchronous C environment; even more so when
> the reproducer is installing Windows on RBD.
> 
> Hanna

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]