llbsolver: fix possible deadlock in history listen

The events for currently active builds were sent through
pubsub channel instead of directly to the current request,
like it was done for completed builds for example.

This meant that if there are more active builds running
than the pubsub channel buffer (32) the sends will block.
Because the history API mutex is held in this process it
will eventually block the requests for builds that try to
update their history records.

Signed-off-by: Tonis Tiigi <[email protected]>

solver/llbsolver/history.go

@@ -762,21 +762,30 @@ func (h *HistoryQueue) Listen(ctx context.Context, req *controlapi.BuildHistoryR
  }()
  }
 
+ // make a copy of events for active builds so we don't keep a lock during grpc send
+ actives := make([]*controlapi.BuildHistoryEvent, 0, len(h.active))
+
  for _, e := range h.active {
  if req.Ref != "" && e.Ref != req.Ref {
  continue
  }
  if _, ok := h.deleted[e.Ref]; ok {
  continue
  }
- sub.send(&controlapi.BuildHistoryEvent{
+ actives = append(actives, &controlapi.BuildHistoryEvent{
  Type: controlapi.BuildHistoryEventType_STARTED,
  Record: e,
  })
  }
 
  h.mu.Unlock()
 
+ for _, e := range actives {
+ if err := f(e); err != nil {
+ return err
+ }
+ }
+
  if !req.ActiveOnly {
  events := []*controlapi.BuildHistoryEvent{}
  if err := h.opt.DB.View(func(tx *bolt.Tx) error {