Re: [Qemu-devel] [PATCH v6 27/28] migration/qmp: add command migrate-pause

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

* Peter Xu (peterx@redhat.com) wrote:
> On Wed, Feb 28, 2018 at 08:14:19PM +0000, Dr. David Alan Gilbert wrote:
> > * Peter Xu (peterx@redhat.com) wrote:
> > > On Wed, Feb 14, 2018 at 06:56:59PM +0000, Dr. David Alan Gilbert wrote:
> > > > * Peter Xu (peterx@redhat.com) wrote:
> > > > > On Tue, Feb 13, 2018 at 08:11:00PM +0000, Dr. David Alan Gilbert wrote:
> > > > > > * Peter Xu (peterx@redhat.com) wrote:
> > > > > > > It pauses an ongoing migration.  Currently it only supports postcopy.
> > > > > > > Note that this command will work on either side of the migration.
> > > > > > > Basically when we trigger this on one side, it'll interrupt the other
> > > > > > > side as well since the other side will get notified on the disconnect
> > > > > > > event.
> > > > > > > 
> > > > > > > However, it's still possible that the other side is not notified, for
> > > > > > > example, when the network is totally broken, or due to some firewall
> > > > > > > configuration changes.  In that case, we will also need to run the same
> > > > > > > command on the other side so both sides will go into the paused state.
> > > > > > > 
> > > > > > > Signed-off-by: Peter Xu <peterx@redhat.com>
> > > > > > > ---
> > > > > > >  migration/migration.c | 27 +++++++++++++++++++++++++++
> > > > > > >  qapi/migration.json   | 16 ++++++++++++++++
> > > > > > >  2 files changed, 43 insertions(+)
> > > > > > > 
> > > > > > > diff --git a/migration/migration.c b/migration/migration.c
> > > > > > > index bb57ed9ade..139abec0c3 100644
> > > > > > > --- a/migration/migration.c
> > > > > > > +++ b/migration/migration.c
> > > > > > > @@ -1448,6 +1448,33 @@ void qmp_migrate_recover(const char *uri, Error **errp)
> > > > > > >      qemu_start_incoming_migration(uri, errp);
> > > > > > >  }
> > > > > > >  
> > > > > > > +void qmp_migrate_pause(Error **errp)
> > > > > > > +{
> > > > > > > +    MigrationState *ms = migrate_get_current();
> > > > > > > +    MigrationIncomingState *mis = migration_incoming_get_current();
> > > > > > > +    int ret;
> > > > > > > +
> > > > > > > +    if (ms->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
> > > > > > > +        /* Source side, during postcopy */
> > > > > > > +        ret = qemu_file_shutdown(ms->to_dst_file);
> > > > > > 
> > > > > > This doesn't feel thread safe; although I'm not sure how to make it so.
> > > > > > If the migration finishes just after we check the state but before the
> > > > > > shutdown we end up using a bogus QEMUFile*
> > > > > > Making all the places that close a QEMUFile* set hte pointer Null before
> > > > > > they do the close doesn't help because you still race with that.
> > > > > > 
> > > > > > (The race is small, but still)
> > > > > 
> > > > > IMHO we can fix it by adding a migration lock for management code. If
> > > > > you see my previous migrate cleanup series, it's in my todo. ;)
> > > > > 
> > > > > The basic idea is that we take the lock for critical paths (but not
> > > > > during most of the migration process).  E.g., we may need the lock
> > > > > for:
> > > > > 
> > > > > - very beginning of migration, during setup
> > > > > - reaching the end of migration
> > > > > - every single migration QMP command (since HMP calls them so HMP will
> > > > >   also acquire the lock)
> > > > > - anywhere else I didn't mention that may necessary, e.g., when we
> > > > >   change migrate state, meanwhile we do something else - basically
> > > > >   that should be an "atomic operation", and we need the lock to make
> > > > >   sure of that.
> > > > 
> > > > But then we couldn't take that lock in an OOB command, you'd have to
> > > > audit all of those places that took it to make sure it didn't do any of
> > > > the things OOB commands aren't allowed to do.
> > > 
> > > Yeah OOB commands will be special - my plan is that they just never
> > > take the lock.  E.g., they only touches FDs, and FDs are naturally
> > > thread safe (like this command).
> > > 
> > > And some major migration commands (like "migrate" itself) should never
> > > be an OOB command.
> > 
> > OK; I'm not sure what makes FDs naturally thread safe though; but
> > lets see the code you have in mind.
> 
> I think I was wrong... it should need a lock.
> 
> > 
> > > > 
> > > > > For the recovery series, I would prefer that we ignore this issue for
> > > > > now - since this problem is there for quite a long time AFAICT in the
> > > > > whole migration code rather than this series only, and we need to
> > > > > solve it once and for all.
> > > > 
> > > > I don't think those problems happen (much) in the existing code, because
> > > > everything is done in the main thread.
> > > 
> > > But migration is running in its own thread (migration_thread)?
> > > 
> > > For example: What if we send migration commands during the end of
> > > migration or a failing migration?  Could there be risk in old code
> > > too since both main thread and migration thread may be manipulating
> > > MigrationState object?
> > 
> > Maybe; although migrate_set_state uses atomic_cmpxchg to ensure
> > that it's safe, and starting a migration can't happen unless there
> > isn't a migration in progress - and that's run under lock.
> 
> Yes I think if without OOB we should be fine since even the cleanup is
> running with the BQL.
> 
> Now I don't have good idea to solve this problem except introducing a
> lock.  How about I add a patch to introduce the mgmt_lock, which
> currently only protect the QEMUFile?  Like:
> 
> ----------------------------------
> diff --git a/migration/migration.c b/migration/migration.c
> index f31fcbb0d5..00c630326d 100644
> --- a/migration/migration.c
> +++ b/migration/migration.c
> @@ -1195,8 +1195,10 @@ static void migrate_fd_cleanup(void *opaque)
>          if (multifd_save_cleanup(&local_err) != 0) {
>              error_report_err(local_err);
>          }
> +        qemu_mutex_lock(&s->mgmt_lock);
>          qemu_fclose(s->to_dst_file);
>          s->to_dst_file = NULL;
> +        qemu_mutex_unlock(&s->mgmt_lock);
>      }
> 
>      assert((s->state != MIGRATION_STATUS_ACTIVE) &&
> @@ -2493,8 +2495,10 @@ static MigThrError postcopy_pause(MigrationState *s)
> 
>          /* Current channel is possibly broken. Release it. */
>          assert(s->to_dst_file);
> +        qemu_mutex_lock(&s->mgmt_lock);
>          qemu_file_shutdown(s->to_dst_file);
>          qemu_fclose(s->to_dst_file);
>          s->to_dst_file = NULL;
> +        qemu_mutex_unlock(&s->mgmt_lock);

That's only safe if we know qemu_fclose() can never block; otherwise
we're not allowed to take the same lock in the OOB command.

I think perhaps it's safer to always do something like:
  tmp = atomic_xchg(s->to_dst_file, NULL);
  qemu_file_shutdown(tmp);
  qemu_fclose(tmp);

then the OOB code can do the same?
Would that work - avoiding the lock?

Dave


>          error_report("Detected IO failure for postcopy. "
> @@ -2970,6 +2974,7 @@ static void migration_instance_finalize(Object *obj)
>      qemu_sem_destroy(&ms->postcopy_pause_sem);
>      qemu_sem_destroy(&ms->postcopy_pause_rp_sem);
>      qemu_sem_destroy(&ms->rp_state.rp_sem);
> +    qemu_mutex_destroy(&ms->mgmt_lock);
>  }
> 
>  static void migration_instance_init(Object *obj)
> @@ -3002,6 +3007,7 @@ static void migration_instance_init(Object *obj)
>      qemu_sem_init(&ms->postcopy_pause_sem, 0);
>      qemu_sem_init(&ms->postcopy_pause_rp_sem, 0);
>      qemu_sem_init(&ms->rp_state.rp_sem, 0);
> +    qemu_mutex_init(&ms->mgmt_lock);
>  }
> 
>  /*
> diff --git a/migration/migration.h b/migration/migration.h
> index c549859cc3..7fcb841978 100644
> --- a/migration/migration.h
> +++ b/migration/migration.h
> @@ -98,6 +98,11 @@ struct MigrationState
>      QemuThread thread;
>      QEMUBH *cleanup_bh;
>      QEMUFile *to_dst_file;
> +    /*
> +     * Currently it only protects to_dst_file.  We need to hold this
> +     * lock when we want to modify in/out QEMUFiles.
> +     */
> +    QemuMutex mgmt_lock;
> 
>      /* bytes already send at the beggining of current interation */
>      uint64_t iteration_initial_bytes;
> ----------------------------------
> 
> Then I take this lock in the OOB postcopy-pause handler.  Since the
> lock holding scenarios are always extremely fast, I assume it's pretty
> safe.  Would that be okay with you?
> 
> Thanks,
> 
> -- 
> Peter Xu
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK