Re: [Qemu-devel] [FIX V2] block: Fix device snapshots broken by the block filter snapshots patchset.

["Benoît Canet" <benoit.canet@irqsave.net>]

The Wednesday 12 Feb 2014 à 20:49:18 (+0100), Benoît Canet wrote :
> The Tuesday 11 Feb 2014 à 16:12:17 (+0800), Fam Zheng wrote :
> > On Mon, 02/10 22:49, Benoît Canet wrote:
> > > Take into account the fact that a block filter like quorum will be in bs->file
> > > while a regular block driver device is really on the top level.
> > > 
> > > Signed-off-by: Benoit Canet <benoit@irqsave.net>
> > > ---
> > >  block.c | 14 ++++++++------
> > >  1 file changed, 8 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/block.c b/block.c
> > > index 07ac50a..d04f535 100644
> > > --- a/block.c
> > > +++ b/block.c
> > > @@ -5400,14 +5400,16 @@ bool bdrv_is_first_non_filter(BlockDriverState *candidate)
> > >  
> > >      /* walk down the bs forest recursively */
> > >      QTAILQ_FOREACH(bs, &bdrv_states, device_list) {
> > > -        bool perm;
> > > -
> > > -        if (!bs->file) {
> > > -            continue;
> > > +        bool perm = false;
> > > +
> > > +        if (bs->file &&
> > > +            bs->file->drv &&
> > > +            bs->file->drv->authorizations[BS_IS_A_FILTER]) {
> > > +            perm = bdrv_recurse_is_first_non_filter(bs->file, candidate);
> > > +        } else if (bs == candidate) {
> > > +            perm = true;
> > >          }
> > >  
> > > -        perm = bdrv_recurse_is_first_non_filter(bs->file, candidate);
> > > -
> > >          /* candidate is the first non filter */
> > >          if (perm) {
> > >              return true;
> > 
> > With this change, if the top level driver has ->file, its implementation of
> > .bdrv_recurse_is_first_non_filter() is skipped and bs->file is the start point.
> > 
> > So we have an implication that single child block drivers (that has ->file)
> > doesn't need to, and shouldn't implement this operation, as commentted above
> > bdrv_generic_is_first_non_filter.
> > 
> > Tested that this patch fixes the external snapshot problem, but didn't test
> > the "quorum as bs->file case".
> > 
> > Thanks,
> > 
> > Reviewed-by: Fam Zheng <famz@redhat.com>
> 
> After extensive testing of all type of quorum instanciations and snapshots I
> discovered that we are not done yet with this issue.
> 
> When instantiating quorum from the command line the quorum driver is in
> bs->file->drv.
> When using QMP's blockdev_add at once or by using references the quorum driver
> is in bs->drv.
> 
> As a result this patch work in half the cases: regular file and command line
> quorum but not when quorum is instantiated via QMP.
> 
> I don't understand why the block layer have this irregular behavior regarding
> quorum instantiation.
> 
> What would be the best way to fix this ?

I have a pending resping of this patch which tests for the two cases as a
workaround.

Should I post it ?

Best regards

Benoît

> 
> Best regards
> 
> Benoît
>