FIFO ordering of input events (#1386)

* Initial FIFOing of roomserver inputs

* Remove EventID response from api.InputRoomEventsResponse

* Don't send back event ID unnecessarily

* Fix ordering hopefully

* Reduce copies, use buffered task channel to reduce contention on other rooms

* Fix error handling

1 files changed, 74 insertions, 8 deletions

diff --git a/roomserver/internal/input/input.go b/roomserver/internal/input/input.go
index 87bdc5db..7a44ff42 100644
--- a/roomserver/internal/input/input.go
+++ b/roomserver/internal/input/input.go
@@ -19,12 +19,14 @@ import (
 	"context"
 	"encoding/json"
 	"sync"
+	"time"
 
 	"github.com/Shopify/sarama"
 	"github.com/matrix-org/dendrite/roomserver/api"
 	"github.com/matrix-org/dendrite/roomserver/storage"
 	"github.com/matrix-org/gomatrixserverlib"
 	log "github.com/sirupsen/logrus"
+	"go.uber.org/atomic"
 )
 
 type Inputer struct {
@@ -33,7 +35,36 @@ type Inputer struct {
 	ServerName           gomatrixserverlib.ServerName
 	OutputRoomEventTopic string
 
-	mutexes sync.Map // room ID -> *sync.Mutex, protects calls to processRoomEvent
+	workers sync.Map // room ID -> *inputWorker
+}
+
+type inputTask struct {
+	ctx   context.Context
+	event *api.InputRoomEvent
+	wg    *sync.WaitGroup
+	err   error // written back by worker, only safe to read when all tasks are done
+}
+
+type inputWorker struct {
+	r       *Inputer
+	running atomic.Bool
+	input   chan *inputTask
+}
+
+func (w *inputWorker) start() {
+	if !w.running.CAS(false, true) {
+		return
+	}
+	defer w.running.Store(false)
+	for {
+		select {
+		case task := <-w.input:
+			_, task.err = w.r.processRoomEvent(task.ctx, task.event)
+			task.wg.Done()
+		case <-time.After(time.Second * 5):
+			return
+		}
+	}
 }
 
 // WriteOutputEvents implements OutputRoomEventWriter
@@ -73,19 +104,54 @@ func (r *Inputer) InputRoomEvents(
 	ctx context.Context,
 	request *api.InputRoomEventsRequest,
 	response *api.InputRoomEventsResponse,
-) (err error) {
+) error {
+	// Create a wait group. Each task that we dispatch will call Done on
+	// this wait group so that we know when all of our events have been
+	// processed.
+	wg := &sync.WaitGroup{}
+	wg.Add(len(request.InputRoomEvents))
+	tasks := make([]*inputTask, len(request.InputRoomEvents))
+
 	for i, e := range request.InputRoomEvents {
+		// Work out if we are running per-room workers or if we're just doing
+		// it on a global basis (e.g. SQLite).
 		roomID := "global"
 		if r.DB.SupportsConcurrentRoomInputs() {
 			roomID = e.Event.RoomID()
 		}
-		mutex, _ := r.mutexes.LoadOrStore(roomID, &sync.Mutex{})
-		mutex.(*sync.Mutex).Lock()
-		if response.EventID, err = r.processRoomEvent(ctx, request.InputRoomEvents[i]); err != nil {
-			mutex.(*sync.Mutex).Unlock()
-			return err
+
+		// Look up the worker, or create it if it doesn't exist. This channel
+		// is buffered to reduce the chance that we'll be blocked by another
+		// room - the channel will be quite small as it's just pointer types.
+		w, _ := r.workers.LoadOrStore(roomID, &inputWorker{
+			r:     r,
+			input: make(chan *inputTask, 10),
+		})
+		worker := w.(*inputWorker)
+
+		// Create a task. This contains the input event and a reference to
+		// the wait group, so that the worker can notify us when this specific
+		// task has been finished.
+		tasks[i] = &inputTask{
+			ctx:   ctx,
+			event: &request.InputRoomEvents[i],
+			wg:    wg,
+		}
+
+		// Send the task to the worker.
+		go worker.start()
+		worker.input <- tasks[i]
+	}
+
+	// Wait for all of the workers to return results about our tasks.
+	wg.Wait()
+
+	// If any of the tasks returned an error, we should probably report
+	// that back to the caller.
+	for _, task := range tasks {
+		if task.err != nil {
+			return task.err
 		}
-		mutex.(*sync.Mutex).Unlock()
 	}
 	return nil
 }