Re: [PATCH v6 1/4] block/replication.c: Ignore requests after failover

[Lukas Straub <lukasstraub2@web.de>]

On Wed, 23 Oct 2019 14:49:29 +0200
Max Reitz <mreitz@redhat.com> wrote:

> On 05.10.19 15:05, Lukas Straub wrote:
> > After failover the Secondary side of replication shouldn't change state, because
> > it now functions as our primary disk.
> > 
> > In replication_start, replication_do_checkpoint, replication_stop, ignore
> > the request if current state is BLOCK_REPLICATION_DONE (sucessful failover) or
> > BLOCK_REPLICATION_FAILOVER (failover in progres i.e. currently merging active
> > and hidden images into the base image).
> > 
> > Signed-off-by: Lukas Straub <lukasstraub2@web.de>
> > Reviewed-by: Zhang Chen <chen.zhang@intel.com>
> > ---
> >  block/replication.c | 38 +++++++++++++++++++++++++++++++++++---
> >  1 file changed, 35 insertions(+), 3 deletions(-)  
> 
> Disclaimer: I don’t know anything about the replication block driver.
> 
> > diff --git a/block/replication.c b/block/replication.c
> > index 3d4dedddfc..97cc65c0cf 100644
> > --- a/block/replication.c
> > +++ b/block/replication.c  
> 
> [...]
> 
> > @@ -529,8 +540,7 @@ static void replication_start(ReplicationState *rs, ReplicationMode mode,
> >                     "Block device is in use by internal backup job");
> > 
> >          top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL);
> > -        if (!top_bs || !bdrv_is_root_node(top_bs) ||
> > -            !check_top_bs(top_bs, bs)) {
> > +        if (!top_bs || !check_top_bs(top_bs, bs)) {  
> 
> It appears to me that top_bs is only used to install op blockers.  It
> seems reasonable to require a root node to be able to do so (because op
> blockers are really only checked on a root node).
> (And the commit message doesn’t tell why we’d want to drop the
> is_root_node check here.)
> 
> Now OTOH I don’t know whether the replication driver needs an op blocker
> at all or whether the permission system suffices...

Hi,
Now that I look at again, it actually works without this change, by passing
a correct top-id= parameter to the driver (I somehow overlooked that parameter). 
So I will revert this change in the next version.

> 
> I suppose the rest of this patch is not really about the block layer, so
> I can’t really comment on it.  (It looks OK to me from a generic and
> naïve standpoint, though.)
> 
> >              error_setg(errp, "No top_bs or it is invalid");
> >              reopen_backing_file(bs, false, NULL);
> >              aio_context_release(aio_context);  
> 
> [...]
> 
> > @@ -593,7 +614,7 @@ static void replication_get_error(ReplicationState *rs, Error **errp)
> >      aio_context_acquire(aio_context);
> >      s = bs->opaque;
> > 
> > -    if (s->stage != BLOCK_REPLICATION_RUNNING) {
> > +    if (s->stage == BLOCK_REPLICATION_NONE) {  
> 
> Just one question out of curiosity, though: Is this a bug fix?

No, It only applies to continuous replication, because colo will
check all replication nodes for errors before checkpointing. So
a secondary continuing replication would error out here, because
it is either in state BLOCK_REPLICATION_DONE or 
BLOCK_REPLICATION_FAILOVER.

> Max
> 
> >          error_setg(errp, "Block replication is not running");
> >          aio_context_release(aio_context);
> >          return;  
>