Re: [PATCH] block: temporarily hold the new AioContext of bs_top in bdrv_append()

[Kevin Wolf <kwolf@redhat.com>]

Am 14.02.2023 um 13:22 hat Stefano Garzarella geschrieben:
> On Tue, Feb 14, 2023 at 12:56 PM Kevin Wolf <kwolf@redhat.com> wrote:
> >
> > Am 14.02.2023 um 11:51 hat Stefano Garzarella geschrieben:
> > > bdrv_append() is called with bs_top AioContext held, but
> > > bdrv_attach_child_noperm() could change the AioContext of bs_top.
> > >
> > > bdrv_replace_node_noperm() calls bdrv_drained_begin() starting from
> > > commit 2398747128 ("block: Don't poll in bdrv_replace_child_noperm()").
> > > bdrv_drained_begin() can call BDRV_POLL_WHILE that assumes the new lock
> > > is taken, so let's temporarily hold the new AioContext to prevent QEMU
> > > from failing in BDRV_POLL_WHILE when it tries to release the wrong
> > > AioContext.
> > >
> > > Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=2168209
> > > Reported-by: Aihua Liang <aliang@redhat.com>
> > > Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> > > ---
> > > I'm not sure whether to use the following Fixes tag. That commit added the
> > > calls to bdrv_drained_begin() in bdrv_replace_node_noperm(), but maybe the
> > > problem was pre-existing.
> > >
> > > Fixes: 2398747128 ("block: Don't poll in bdrv_replace_child_noperm()")
> > >
> > > Note: a local reproducer is attached in the BZ, it is based on the Aihua Liang
> > > report and it hits the issue with a 20% ratio.
> > > ---
> > >  block.c | 23 +++++++++++++++++++++++
> > >  1 file changed, 23 insertions(+)
> > >
> > > diff --git a/block.c b/block.c
> > > index aa9062f2c1..0e2bc11e0b 100644
> > > --- a/block.c
> > > +++ b/block.c
> > > @@ -5266,6 +5266,8 @@ int bdrv_drop_filter(BlockDriverState *bs, Error **errp)
> > >   * child.
> > >   *
> > >   * This function does not create any image files.
> > > + *
> > > + * The caller must hold the AioContext lock for @bs_top.
> > >   */
> > >  int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top,
> > >                  Error **errp)
> > > @@ -5273,11 +5275,14 @@ int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top,
> > >      int ret;
> > >      BdrvChild *child;
> > >      Transaction *tran = tran_new();
> > > +    AioContext *old_context, *new_context;
> > >
> > >      GLOBAL_STATE_CODE();
> > >
> > >      assert(!bs_new->backing);
> > >
> > > +    old_context = bdrv_get_aio_context(bs_top);
> > > +
> > >      child = bdrv_attach_child_noperm(bs_new, bs_top, "backing",
> > >                                       &child_of_bds, bdrv_backing_role(bs_new),
> > >                                       tran, errp);
> > > @@ -5286,11 +5291,29 @@ int bdrv_append(BlockDriverState *bs_new, BlockDriverState *bs_top,
> > >          goto out;
> > >      }
> > >
> > > +    /*
> > > +     * bdrv_attach_child_noperm could change the AioContext of bs_top.
> > > +     * bdrv_replace_node_noperm calls bdrv_drained_begin, so let's temporarily
> > > +     * hold the new AioContext, since bdrv_drained_begin calls BDRV_POLL_WHILE
> > > +     * that assumes the new lock is taken.
> > > +     */
> > > +    new_context = bdrv_get_aio_context(bs_top);
> > > +
> > > +    if (old_context != new_context) {
> > > +        aio_context_release(old_context);
> > > +        aio_context_acquire(new_context);
> > > +    }
> > > +
> > >      ret = bdrv_replace_node_noperm(bs_top, bs_new, true, tran, errp);
> > >      if (ret < 0) {
> > >          goto out;
> >
> > If we take the error path, we return with new_context locked instead of
> > old_context now.
> 
> Grr, I'm blind...
> 
> >
> > >      }
> > >
> > > +    if (old_context != new_context) {
> > > +        aio_context_release(new_context);
> > > +        aio_context_acquire(old_context);
> > > +    }
> > > +
> > >      ret = bdrv_refresh_perms(bs_new, tran, errp);
> > >  out:
> > >      tran_finalize(tran, ret);
> >
> > Strictly speaking, don't we need to hold the lock across
> > tran_finalize(), too? It completes the bdrv_replace_node_noperm() call
> > you covered above.
> 
> Right!
> 
> >
> > Maybe bdrv_refresh_perms() and bdrv_refresh_limits(), too, in fact. We
> > never clearly defined which functions need the lock and which don't, so
> > hard to tell.
> 
> Okay, so to be on the safe side, I'll switch them back just before return.
> 
> > It's really time to get rid of it.
> 
> How could one disagree? :-)
> 
> What about the Fixes tag? Should I include it?

I'm not sure. Before the patch, bdrv_replace_child_noperm() had a drain
which could have caused the same kind of problems. But we're now
draining two BDSes, maybe that is the relevant difference. I guess we've
always had a bug, but that commit made it more likely to actually
trigger?

Kevin