Re: [Qemu-devel] [PATCH v2 2/2] QEMUBH: make AioContext's bh re-entrant

[Paolo Bonzini]

Il 17/06/2013 17:28, Stefan Hajnoczi ha scritto:
>> > +    qemu_mutex_lock(&ctx->bh_lock);
>> >      bh->next = ctx->first_bh;
>> > +    /* Make sure the memebers ready before putting bh into list */
> s/memebers/members/
> 
>> > +    smp_wmb();
> Why lock bh_lock before assigning bh->next?  Could you lock the mutex
> here and then drop the smp_wmb() since the pthread function already does
> that?
> 
> http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_11

Not sure I understand, ctx->first_bh is read here and that's what the
lock protects.

    thread 1                           thread 2
    ------------------------------------------------------------------
    bh->next = ctx->first_bh;
                                       bh->next = ctx->first_bh;
                                       lock
                                       ctx->first_bh = bh;
                                       unlock
    lock
    ctx->first_bh = bh;
    unlock

and thread 2's bottom half is gone.  There is also a similar race that
leaves a dangling pointer if aio_bh_new races against aio_bh_poll.

Paolo