From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:51440) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1er87P-0001QA-UF for qemu-devel@nongnu.org; Wed, 28 Feb 2018 15:14:30 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1er87L-0008BL-UT for qemu-devel@nongnu.org; Wed, 28 Feb 2018 15:14:27 -0500 Received: from mx3-rdu2.redhat.com ([66.187.233.73]:52940 helo=mx1.redhat.com) by eggs.gnu.org with esmtps (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.71) (envelope-from ) id 1er87L-0008Ab-JI for qemu-devel@nongnu.org; Wed, 28 Feb 2018 15:14:23 -0500 Date: Wed, 28 Feb 2018 20:14:19 +0000 From: "Dr. David Alan Gilbert" Message-ID: <20180228201418.GQ2981@work-vm> References: <20180208103132.28452-1-peterx@redhat.com> <20180208103132.28452-28-peterx@redhat.com> <20180213201059.GU2378@work-vm> <20180214045639.GI4472@xz-mi> <20180214185658.GF2507@work-vm> <20180222074346.GG18962@xz-mi> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20180222074346.GG18962@xz-mi> Subject: Re: [Qemu-devel] [PATCH v6 27/28] migration/qmp: add command migrate-pause List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Peter Xu Cc: qemu-devel@nongnu.org, Alexey Perevalov , "Daniel P . Berrange" , Juan Quintela , Andrea Arcangeli * 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 > > > > > --- > > > > > 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. > > > > > 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. Dave > > That's one reason why the to_dst_file is closed in migrate_fd_cleanup > > which is normally closed in the back-half run on the main thread. > > > > One way would be to make the state go to POSTCOPY_PAUSED here; > > note that migrate_set_state already uses an atomic_cmpxchg to do the > > update. > > Thanks, > > -- > Peter Xu -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK