+
+// Lock the given container. Should be called in a new goroutine.
+func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
+ if !sch.uuidLock(uuid, "lock") {
+ return
+ }
+ defer sch.uuidUnlock(uuid)
+ if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
+ // This happens if the container has been cancelled or
+ // locked since runQueue called sch.queue.Entries(),
+ // possibly by a lockContainer() call from a previous
+ // runQueue iteration. In any case, we will respond
+ // appropriately on the next runQueue iteration, which
+ // will have already been triggered by the queue
+ // update.
+ logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
+ return
+ }
+ err := sch.queue.Lock(uuid)
+ if err != nil {
+ logger.WithError(err).Warn("error locking container")
+ return
+ }
+ logger.Debug("lock succeeded")
+ ctr, ok := sch.queue.Get(uuid)
+ if !ok {
+ logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
+ } else if ctr.State != arvados.ContainerStateLocked {
+ logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
+ }
+}
+
+// Acquire a non-blocking lock for specified UUID, returning true if
+// successful. The op argument is used only for debug logs.
+//
+// If the lock is not available, uuidLock arranges to wake up the
+// scheduler after a short delay, so it can retry whatever operation
+// is trying to get the lock (if that operation is still worth doing).
+//
+// This mechanism helps avoid spamming the controller/database with
+// concurrent updates for any single container, even when the
+// scheduler loop is running frequently.
+func (sch *Scheduler) uuidLock(uuid, op string) bool {
+ sch.mtx.Lock()
+ defer sch.mtx.Unlock()
+ logger := sch.logger.WithFields(logrus.Fields{
+ "ContainerUUID": uuid,
+ "Op": op,
+ })
+ if op, locked := sch.uuidOp[uuid]; locked {
+ logger.Debugf("uuidLock not available, Op=%s in progress", op)
+ // Make sure the scheduler loop wakes up to retry.
+ sch.wakeup.Reset(time.Second / 4)
+ return false
+ }
+ logger.Debug("uuidLock acquired")
+ sch.uuidOp[uuid] = op
+ return true
+}
+
+func (sch *Scheduler) uuidUnlock(uuid string) {
+ sch.mtx.Lock()
+ defer sch.mtx.Unlock()
+ delete(sch.uuidOp, uuid)
+}