Re: [Qemu-devel] [PATCH] migration: re-active images while migration been canceled after inactive them

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Hailiang Zhang (zhang.zhanghailiang@huawei.com) wrote:
> On 2017/1/23 21:40, Dr. David Alan Gilbert wrote:
> > * zhanghailiang (zhang.zhanghailiang@huawei.com) wrote:
> > > commit fe904ea8242cbae2d7e69c052c754b8f5f1ba1d6 fixed a case
> > > which migration aborted QEMU because it didn't regain the control
> > > of images while some errors happened.
> > > 
> > > Actually, there are another two cases can trigger the same error reports:
> > > " bdrv_co_do_pwritev: Assertion `!(bs->open_flags & 0x0800)' failed",
> > > 
> > > Case 1, codes path:
> > > migration_thread()
> > >      migration_completion()
> > >          bdrv_inactivate_all() ----------------> inactivate images
> > >          qemu_savevm_state_complete_precopy()
> > >              socket_writev_buffer() --------> error because destination fails
> > >                  qemu_fflush() -------------------> set error on migration stream
> > > -> qmp_migrate_cancel() ---------------------> user cancelled migration concurrently
> > >      -> migrate_set_state() ------------------> set migrate CANCELLIN
> > >      migration_completion() -----------------> go on to fail_invalidate
> > > 	if (s->state == MIGRATION_STATUS_ACTIVE) -> Jump this branch
> > > 
> > > Case 2, codes path:
> > > migration_thread()
> > >      migration_completion()
> > >          bdrv_inactivate_all() ----------------> inactivate images
> > >      migreation_completion() finished
> > > -> qmp_migrate_cancel() ---------------------> user cancelled migration concurrently
> > >      qemu_mutex_lock_iothread();
> > >      qemu_bh_schedule (s->cleanup_bh);
> > > 
> > > As we can see from above, qmp_migrate_cancel can slip in whenever
> > > migration_thread does not hold the global lock. If this happens after
> > > bdrv_inactive_all() been called, the above error reports will appear.
> > > 
> > > To prevent this, we can call bdrv_invalidate_cache_all() in qmp_migrate_cancel()
> > > directly if we find images become inactive.
> > > 
> > > Signed-off-by: zhanghailiang <zhang.zhanghailiang@huawei.com>
> > > ---
> > > Hi,
> > > 
> > > I have sent another patch before to fix this problem, but didn't cover
> > > all the scenes, and there are some discussions about this problem,
> > > For more detail, please refer to
> > > https://lists.gnu.org/archive/html/qemu-block/2016-12/msg00003.html
> > > ---
> > >   include/migration/migration.h |  3 +++
> > >   migration/migration.c         | 13 +++++++++++++
> > >   2 files changed, 16 insertions(+)
> > > 
> > > diff --git a/include/migration/migration.h b/include/migration/migration.h
> > > index c309d23..2d5b724 100644
> > > --- a/include/migration/migration.h
> > > +++ b/include/migration/migration.h
> > > @@ -177,6 +177,9 @@ struct MigrationState
> > >       /* Flag set once the migration thread is running (and needs joining) */
> > >       bool migration_thread_running;
> > > 
> > > +    /* Flag set once the migration thread called bdrv_inactivate_all */
> > > +    bool block_inactive;
> > > +
> > >       /* Queue of outstanding page requests from the destination */
> > >       QemuMutex src_page_req_mutex;
> > >       QSIMPLEQ_HEAD(src_page_requests, MigrationSrcPageRequest) src_page_requests;
> > > diff --git a/migration/migration.c b/migration/migration.c
> > > index f498ab8..9defb3e 100644
> > > --- a/migration/migration.c
> > > +++ b/migration/migration.c
> > > @@ -1006,6 +1006,16 @@ static void migrate_fd_cancel(MigrationState *s)
> > >       if (s->state == MIGRATION_STATUS_CANCELLING && f) {
> > >           qemu_file_shutdown(f);
> > >       }
> > > +    if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
> > > +        Error *local_err = NULL;
> > > +
> > > +        bdrv_invalidate_cache_all(&local_err);
> > > +        if (local_err) {
> > > +            error_report_err(local_err);
> > > +        } else {
> > > +            s->block_inactive = false;
> > > +        }
> > > +    }
> > >   }
> > > 
> > >   void add_migration_state_change_notifier(Notifier *notify)
> > > @@ -1705,6 +1715,7 @@ static void migration_completion(MigrationState *s, int current_active_state,
> > >               if (ret >= 0) {
> > >                   qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
> > >                   qemu_savevm_state_complete_precopy(s->to_dst_file, false);
> > > +                s->block_inactive = true;
> > >               }
> > >           }
> > >           qemu_mutex_unlock_iothread();
> > > @@ -1758,6 +1769,8 @@ fail_invalidate:
> > >           bdrv_invalidate_cache_all(&local_err);
> > >           if (local_err) {
> > >               error_report_err(local_err);
> > > +        } else {
> > > +            s->block_inactive = false;
> > >           }
> > 
> > I think the fe904 commit also add the problem that this bdrv_invalidate_cache_all
> > is done outside the big lock (Stefan and Kevin tell me bdrv_* calls generally
> > need the lock).
> > 
> 
> Ha, you are right, I didn't noticed that all the time.

I started thinking about it because I wondered about the bdrv_invalidate_cache_all 
in fd_cancel racing against that one.

> So should I just add the big lock there ? Is that enough?

Yes, I think that should be fine.

Dave

> Thanks,
> Hailiang
> 
> > Dave
> > 
> > >       }
> > > 
> > > --
> > > 1.8.3.1
> > > 
> > > 
> > --
> > Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
> > 
> > .
> > 
> 
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK