From 5d067ee82d7c0446efdd1ae2cffa40e208ec2843 Mon Sep 17 00:00:00 2001
From: Hermes <piotr@piksar.eu>
Date: Fri, 6 Nov 2026 16:37:39 +0100
Subject: [PATCH] fix: bound the event drain so the OSD timer punch survives
 storms (#3)

---
 internal/display/x11.go      |  17 +++++-
 internal/display/x11_test.go | 100 +++++++++++++++++++++++++++++++++++
 2 files changed, 116 insertions(+), 1 deletion(-)

diff --git a/internal/display/x11.go b/internal/display/x11.go
index 7cee954..251fa84 100644
--- a/internal/display/x11.go
+++ b/internal/display/x11.go
@@ -206,6 +206,16 @@ func createWindowValues(backPixel uint32, eventMask uint32, cursor xproto.Cursor
 // only as a downstream BadWindow.
 const cursorBitmapDepth = 1
 
+// eventDrainBatch bounds how many pending events one wait cycle drains. A
+// busy server (video fullscreen repaints) can serve unrelated non-Expose
+// events faster than they arrive; an unbounded drain would never reach the
+// timer checks, so the OSD overlay's initially empty bounding shape would
+// never be punched and its telemetry would never appear. Thirty-two events
+// per 100 ms poll cycle sustains well over three thousand events per
+// second — orders of magnitude above any plausible root-event rate — while
+// bounding liveness and cancellation latency to a single batch.
+const eventDrainBatch = 32
+
 // blankCursorX is the checked X-request surface of the blank cursor setup,
 // so the sequence can be driven by tests that script failures at exact
 // operations and pin request order and resource reclamation.
@@ -956,7 +966,12 @@ func (s *x11Session) WaitEvent(ctx context.Context) (bool, error) {
 	defer ticker.Stop()
 	lastRefresh := time.Now()
 	for {
-		for {
+		// Drain a bounded batch of pending events: a busy server (video
+		// fullscreen repaints) can serve unrelated non-Expose events faster
+		// than they arrive, and an unbounded drain would never reach the
+		// timer checks — the OSD overlay's initially empty bounding shape
+		// would never be punched and its telemetry would never appear.
+		for drained := 0; drained < eventDrainBatch; drained++ {
 			event, err := s.pollEvent()
 			if err != nil {
 				return false, err
diff --git a/internal/display/x11_test.go b/internal/display/x11_test.go
index 58fd132..bd0b778 100644
--- a/internal/display/x11_test.go
+++ b/internal/display/x11_test.go
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"strconv"
 	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -512,6 +513,105 @@ func TestTransparentShapesEmptyInputAndOutput(t *testing.T) {
 	}
 }
 
+// stormSession builds an OSD layer session whose event source never runs
+// dry: it serves an endless stream of non-Expose root events (the video
+// fullscreen repaint storm), so the drain loop can only exit through its
+// bound — never through an empty queue.
+func stormSession(refresh time.Duration) *x11Session {
+	return &x11Session{
+		poll:    time.Millisecond,
+		refresh: refresh,
+		pollEvent: func() (xgb.Event, xgb.Error) {
+			return xproto.ConfigureNotifyEvent{}, nil
+		},
+		probe: func() error { return nil },
+	}
+}
+
+func TestWaitEventRefreshesDuringContinuousEventStorm(t *testing.T) {
+	// The initially empty bounding shape is punched only by redraws: an
+	// endless stream of unrelated events must never starve the timer that
+	// drives the OSD refresh cadence.
+	session := stormSession(3 * time.Millisecond)
+	session.topVideo = 0
+
+	done := make(chan eventResult, 1)
+	go func() {
+		redraw, err := session.WaitEvent(context.Background())
+		done <- eventResult{redraw: redraw, err: err}
+	}()
+	select {
+	case result := <-done:
+		if result.err != nil {
+			t.Fatalf("WaitEvent=%v", result.err)
+		}
+		if !result.redraw {
+			t.Fatal("the refresh must request a redraw during an event storm")
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("WaitEvent starved: the timer redraw never fired during the storm")
+	}
+}
+
+func TestWaitEventHonorsCancellationDuringEventStorm(t *testing.T) {
+	session := stormSession(time.Hour)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	done := make(chan eventResult, 1)
+	go func() {
+		redraw, err := session.WaitEvent(ctx)
+		done <- eventResult{redraw: redraw, err: err}
+	}()
+	select {
+	case result := <-done:
+		if result.redraw {
+			t.Fatal("cancellation must not be reported as a redraw")
+		}
+		if !errors.Is(result.err, context.Canceled) {
+			t.Fatalf("err=%v want clean cancellation during the storm", result.err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("WaitEvent did not terminate on cancellation during the storm")
+	}
+}
+
+func TestWaitEventStillRedrawsOnExposeDuringEventStorm(t *testing.T) {
+	var served atomic.Bool
+	session := stormSession(0)
+	session.pollEvent = func() (xgb.Event, xgb.Error) {
+		if served.CompareAndSwap(false, true) {
+			return xproto.ExposeEvent{}, nil
+		}
+		return xproto.ConfigureNotifyEvent{}, nil
+	}
+
+	redraw, err := session.WaitEvent(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !redraw {
+		t.Fatal("an expose event must still trigger a redraw during a storm")
+	}
+}
+
+func TestRoleBelowCanvasStaysEventDrivenDuringEventStorm(t *testing.T) {
+	// The fallback canvas has no refresh interval: an endless unrelated
+	// event stream must never produce refresh redraws, and only
+	// cancellation ends the wait.
+	session := stormSession(0)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 50*time.Millisecond)
+	defer cancel()
+	redraw, err := session.WaitEvent(ctx)
+	if !errors.Is(err, context.DeadlineExceeded) {
+		t.Fatalf("err=%v want clean cancellation without a refresh interval", err)
+	}
+	if redraw {
+		t.Fatal("the canvas must stay event-driven with no refresh redraws")
+	}
+}
+
 func TestOSDTextRectCoversTheGlyphCellAroundTheBaseline(t *testing.T) {
 	rect := osdTextRect(32, 32, 9, 6, 11, 3)
 	if rect.X != 30 || rect.Y != 19 {