Message ID | CAJnKYQ=Sj99wh_M1zVpX3V6TyDbZedF0kTLQ9phgRttz6e8Wrg@mail.gmail.com |
---|---|
State | New |
Headers | show |
Il 20/06/2013 11:41, liu ping fan ha scritto: > On Thu, Jun 20, 2013 at 4:16 PM, Paolo Bonzini <pbonzini@redhat.com> wrote: >> Il 20/06/2013 09:39, Stefan Hajnoczi ha scritto: >>> qemu_bh_cancel() and qemu_bh_delete() are not modified by this patch. >>> >>> It seems that calling them from a thread is a little risky because there >>> is no guarantee that the BH is no longer invoked after a thread calls >>> these functions. >>> >>> I think that's worth a comment or do you want them to take the lock so >>> they become safe? >> >> Taking the lock wouldn't help. The invoking loop of aio_bh_poll runs >> lockless. I think a comment is better. >> >> qemu_bh_cancel is inherently not thread-safe, there's not much you can >> do about it. >> >> qemu_bh_delete is safe as long as you wait for the bottom half to stop >> before deleting the containing object. Once we have RCU, deletion of >> QOM objects will be RCU-protected. Hence, a simple way could be to put >> the first part of aio_bh_poll() within rcu_read_lock/unlock. >> > In fact, I have some idea about this, introduce another member - > Object for QEMUBH which will be refereed in cb, then we leave anything > to refcnt mechanism. > For qemu_bh_cancel(), I do not figure out whether it is important or > not to sync with caller. This is a separate patch anyway... and a long discussion to have before too. :) Let's concentrate on one thing at a time. Paolo > diff --git a/async.c b/async.c > index 4b17eb7..60c35a1 100644 > --- a/async.c > +++ b/async.c > @@ -61,6 +61,7 @@ int aio_bh_poll(AioContext *ctx) > { > QEMUBH *bh, **bhp, *next; > int ret; > + int sched; > > { > QEMUBH *bh, **bhp, *next; > int ret; > + int sched; > > ctx->walking_bh++; > > @@ -69,8 +70,10 @@ int aio_bh_poll(AioContext *ctx) > /* Make sure fetching bh before accessing its members */ > smp_read_barrier_depends(); > next = bh->next; > - if (!bh->deleted && bh->scheduled) { > - bh->scheduled = 0; > + sched = 0; > + atomic_xchg(&bh->scheduled, sched); This is expensive. > + if (!bh->deleted && sched) { > + //bh->scheduled = 0; > if (!bh->idle) > ret = 1; > bh->idle = 0; > @@ -79,6 +82,9 @@ int aio_bh_poll(AioContext *ctx) > */ > smp_rmb(); > bh->cb(bh->opaque); > + if (bh->obj) { > + object_unref(bh->obj); > + } > } > } > > @@ -105,8 +111,12 @@ int aio_bh_poll(AioContext *ctx) > > void qemu_bh_schedule_idle(QEMUBH *bh) > { > - if (bh->scheduled) > + int sched = 1; > + > + atomic_xchg( &bh->scheduled, sched); > + if (sched) { > return; > + } > /* Make sure any writes that are needed by the callback are done > * before the locations are read in the aio_bh_poll. > */ > @@ -117,25 +127,46 @@ void qemu_bh_schedule_idle(QEMUBH *bh) > > void qemu_bh_schedule(QEMUBH *bh) > { > - if (bh->scheduled) > + int sched = 1; > + > + atomic_xchg( &bh->scheduled, sched); > + if (sched) { > return; > + } > /* Make sure any writes that are needed by the callback are done > * before the locations are read in the aio_bh_poll. > */ > smp_wmb(); > bh->scheduled = 1; > + if (bh->obj) { > + object_ref(bh->obj); > + } > bh->idle = 0; > aio_notify(bh->ctx); > } > > void qemu_bh_cancel(QEMUBH *bh) > { > - bh->scheduled = 0; > + int sched = 0; > + > + atomic_xchg( &bh->scheduled, sched); > + if (sched) { > + if (bh->obj) { > + object_ref(bh->obj); > + } > + } > } > > void qemu_bh_delete(QEMUBH *bh) > { > - bh->scheduled = 0; > + int sched = 0; > + > + atomic_xchg( &bh->scheduled, sched); > + if (sched) { > + if (bh->obj) { > + object_ref(bh->obj); > + } > + } > bh->deleted = 1; > } > > Regards, > Pingfan >>> The other thing I'm unclear on is the ->idle assignment followed >>> immediately by a ->scheduled assignment. Without memory barriers >>> aio_bh_poll() isn't guaranteed to get an ordered view of these updates: >>> it may see an idle BH as a regular scheduled BH because ->idle is still >>> 0. >> >> Right. You need to order ->idle writes before ->scheduled writes, and >> add memory barriers, or alternatively use two bits in ->scheduled so >> that you can assign both atomically. >> >> Paolo
[...] >>> >>> qemu_bh_delete is safe as long as you wait for the bottom half to stop >>> before deleting the containing object. Once we have RCU, deletion of >>> QOM objects will be RCU-protected. Hence, a simple way could be to put >>> the first part of aio_bh_poll() within rcu_read_lock/unlock. >>> >> In fact, I have some idea about this, introduce another member - >> Object for QEMUBH which will be refereed in cb, then we leave anything >> to refcnt mechanism. >> For qemu_bh_cancel(), I do not figure out whether it is important or >> not to sync with caller. > > This is a separate patch anyway... and a long discussion to have before > too. :) > > Let's concentrate on one thing at a time. > Yes, will do like this. Regards, Pingfan > Paolo > >> diff --git a/async.c b/async.c >> index 4b17eb7..60c35a1 100644 >> --- a/async.c >> +++ b/async.c >> @@ -61,6 +61,7 @@ int aio_bh_poll(AioContext *ctx) >> { >> QEMUBH *bh, **bhp, *next; >> int ret; >> + int sched; >> >> { >> QEMUBH *bh, **bhp, *next; >> int ret; >> + int sched; >> >> ctx->walking_bh++; >> >> @@ -69,8 +70,10 @@ int aio_bh_poll(AioContext *ctx) >> /* Make sure fetching bh before accessing its members */ >> smp_read_barrier_depends(); >> next = bh->next; >> - if (!bh->deleted && bh->scheduled) { >> - bh->scheduled = 0; >> + sched = 0; >> + atomic_xchg(&bh->scheduled, sched); > > This is expensive. > >> + if (!bh->deleted && sched) { >> + //bh->scheduled = 0; >> if (!bh->idle) >> ret = 1; >> bh->idle = 0; >> @@ -79,6 +82,9 @@ int aio_bh_poll(AioContext *ctx) >> */ >> smp_rmb(); >> bh->cb(bh->opaque); >> + if (bh->obj) { >> + object_unref(bh->obj); >> + } >> } >> } >> >> @@ -105,8 +111,12 @@ int aio_bh_poll(AioContext *ctx) >> >> void qemu_bh_schedule_idle(QEMUBH *bh) >> { >> - if (bh->scheduled) >> + int sched = 1; >> + >> + atomic_xchg( &bh->scheduled, sched); >> + if (sched) { >> return; >> + } >> /* Make sure any writes that are needed by the callback are done >> * before the locations are read in the aio_bh_poll. >> */ >> @@ -117,25 +127,46 @@ void qemu_bh_schedule_idle(QEMUBH *bh) >> >> void qemu_bh_schedule(QEMUBH *bh) >> { >> - if (bh->scheduled) >> + int sched = 1; >> + >> + atomic_xchg( &bh->scheduled, sched); >> + if (sched) { >> return; >> + } >> /* Make sure any writes that are needed by the callback are done >> * before the locations are read in the aio_bh_poll. >> */ >> smp_wmb(); >> bh->scheduled = 1; >> + if (bh->obj) { >> + object_ref(bh->obj); >> + } >> bh->idle = 0; >> aio_notify(bh->ctx); >> } >> >> void qemu_bh_cancel(QEMUBH *bh) >> { >> - bh->scheduled = 0; >> + int sched = 0; >> + >> + atomic_xchg( &bh->scheduled, sched); >> + if (sched) { >> + if (bh->obj) { >> + object_ref(bh->obj); >> + } >> + } >> } >> >> void qemu_bh_delete(QEMUBH *bh) >> { >> - bh->scheduled = 0; >> + int sched = 0; >> + >> + atomic_xchg( &bh->scheduled, sched); >> + if (sched) { >> + if (bh->obj) { >> + object_ref(bh->obj); >> + } >> + } >> bh->deleted = 1; >> } >> >> Regards, >> Pingfan >>>> The other thing I'm unclear on is the ->idle assignment followed >>>> immediately by a ->scheduled assignment. Without memory barriers >>>> aio_bh_poll() isn't guaranteed to get an ordered view of these updates: >>>> it may see an idle BH as a regular scheduled BH because ->idle is still >>>> 0. >>> >>> Right. You need to order ->idle writes before ->scheduled writes, and >>> add memory barriers, or alternatively use two bits in ->scheduled so >>> that you can assign both atomically. >>> >>> Paolo >
diff --git a/async.c b/async.c index 4b17eb7..60c35a1 100644 --- a/async.c +++ b/async.c @@ -61,6 +61,7 @@ int aio_bh_poll(AioContext *ctx) { QEMUBH *bh, **bhp, *next; int ret; + int sched; { QEMUBH *bh, **bhp, *next;