Complete sync performance (#1741)

* Parallelise PDU stream fetching for complete sync

* Fixes

* Fixes

* Worker queue

* Workers

* Don't populate device list changes on complete sync

* Don't fast-forward typing notifications either on complete sync

* Revert "Don't fast-forward typing notifications either on complete sync"
This reverts commit 01471f78431cdd840915111f71bd2b5176e584a8.

* Comments

2 files changed, 61 insertions, 12 deletions

diff --git a/syncapi/streams/stream_devicelist.go b/syncapi/streams/stream_devicelist.go
index c43d50a4..9ea9d088 100644
--- a/syncapi/streams/stream_devicelist.go
+++ b/syncapi/streams/stream_devicelist.go
@@ -19,7 +19,7 @@ func (p *DeviceListStreamProvider) CompleteSync(
 	ctx context.Context,
 	req *types.SyncRequest,
 ) types.LogPosition {
-	return p.IncrementalSync(ctx, req, types.LogPosition{}, p.LatestPosition(ctx))
+	return p.LatestPosition(ctx)
 }
 
 func (p *DeviceListStreamProvider) IncrementalSync(
diff --git a/syncapi/streams/stream_pdu.go b/syncapi/streams/stream_pdu.go
index d6d7ff44..ae38dc30 100644
--- a/syncapi/streams/stream_pdu.go
+++ b/syncapi/streams/stream_pdu.go
@@ -2,18 +2,54 @@ package streams
 
 import (
 	"context"
+	"sync"
+	"time"
 
 	"github.com/matrix-org/dendrite/syncapi/types"
 	userapi "github.com/matrix-org/dendrite/userapi/api"
 	"github.com/matrix-org/gomatrixserverlib"
+	"go.uber.org/atomic"
 )
 
+// The max number of per-room goroutines to have running.
+// Too high and this will consume lots of CPU, too low and complete
+// sync responses will take longer to process.
+const PDU_STREAM_WORKERS = 256
+
+// The maximum number of tasks that can be queued in total before
+// backpressure will build up and the rests will start to block.
+const PDU_STREAM_QUEUESIZE = PDU_STREAM_WORKERS * 8
+
 type PDUStreamProvider struct {
 	StreamProvider
+
+	tasks   chan func()
+	workers atomic.Int32
+}
+
+func (p *PDUStreamProvider) worker() {
+	defer p.workers.Dec()
+	for {
+		select {
+		case f := <-p.tasks:
+			f()
+		case <-time.After(time.Second * 10):
+			return
+		}
+	}
+}
+
+func (p *PDUStreamProvider) queue(f func()) {
+	if p.workers.Load() < PDU_STREAM_WORKERS {
+		p.workers.Inc()
+		go p.worker()
+	}
+	p.tasks <- f
 }
 
 func (p *PDUStreamProvider) Setup() {
 	p.StreamProvider.Setup()
+	p.tasks = make(chan func(), PDU_STREAM_QUEUESIZE)
 
 	p.latestMutex.Lock()
 	defer p.latestMutex.Unlock()
@@ -52,19 +88,32 @@ func (p *PDUStreamProvider) CompleteSync(
 	eventFilter := req.Filter.Room.Timeline
 
 	// Build up a /sync response. Add joined rooms.
-	for _, roomID := range joinedRoomIDs {
-		var jr *types.JoinResponse
-		jr, err = p.getJoinResponseForCompleteSync(
-			ctx, roomID, r, &stateFilter, &eventFilter, req.Device,
-		)
-		if err != nil {
-			req.Log.WithError(err).Error("p.getJoinResponseForCompleteSync failed")
-			return from
-		}
-		req.Response.Rooms.Join[roomID] = *jr
-		req.Rooms[roomID] = gomatrixserverlib.Join
+	var reqMutex sync.Mutex
+	var reqWaitGroup sync.WaitGroup
+	reqWaitGroup.Add(len(joinedRoomIDs))
+	for _, room := range joinedRoomIDs {
+		roomID := room
+		p.queue(func() {
+			defer reqWaitGroup.Done()
+
+			var jr *types.JoinResponse
+			jr, err = p.getJoinResponseForCompleteSync(
+				ctx, roomID, r, &stateFilter, &eventFilter, req.Device,
+			)
+			if err != nil {
+				req.Log.WithError(err).Error("p.getJoinResponseForCompleteSync failed")
+				return
+			}
+
+			reqMutex.Lock()
+			defer reqMutex.Unlock()
+			req.Response.Rooms.Join[roomID] = *jr
+			req.Rooms[roomID] = gomatrixserverlib.Join
+		})
 	}
 
+	reqWaitGroup.Wait()
+
 	// Add peeked rooms.
 	peeks, err := p.DB.PeeksInRange(ctx, req.Device.UserID, req.Device.ID, r)
 	if err != nil {