detached 0e0d396

Merge pull request 'perf: quiet idle CPU, DRM master retry and stall watchdog' (#37) from perf/idle-cpu-watchdog into main

Reviewed-on: https://git.bnema.dev/bnema/neferwl/pulls/37
commit
0e0d396c2b460d9552be66bc61190226b08b4cca
author
bnema · · 2026-10-10 16:13 UTC
merge of
3305370 + 2382937 · diff against the first parent
signature
SSH UnverifiedSHA256:VOCA5lYARHcbKDZ7GhTV45VmcjWnJ6ZQIBazj17LivA The key is unknown
26 files · +1348 -161
Split
UnifiedSplit
M Makefile +1 -1
@@ -69,7 +69,7 @@ adapter-imports:
6969 [ -z "$$bad" ] || { echo "$$bad"; echo 'adapter imports another adapter: use ports or a shared library (Makefile ADAPTER_LIBS)' >&2; exit 1; }
7070 # Fast, mandatory guards for tiled content publications and SHM copies.
7171 perf-check:
72- CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations|TestPublishAllocations|TestPublishSubscribersAllocations|TestPublishStashSlideAllocations|TestPublishOverviewAllocations|TestPublishLeavingAllocations|TestOutputRoutingAllocations|TestReportSeenAllocations|TestFrameLifecycleTransitionsAllocations|TestFullscreenShownAllocations|TestFrameDecisionAllocations|TestCommitFrameAllocations|TestColorPipelineAllocations|TestShownBySnapshotAllocations|TestMRUNoteFocusDoesNotAllocateWhenSettled|TestDueFramesAllocations|TestRenderSteadyStateAllocations|TestSceneOpaqueRegionAllocations|TestSceneDeltaAllocations|TestAccountFlipAllocations|TestFlipDoneAllocations|TestSceneWalkUnchangedTiledSHM|TestTiledCommitPublishAllocations|TestCapturedCommitApplyAllocations|TestCapturedViewportApplyAllocations|TestEffectiveInputEmptyTreeAllocations|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
72+ CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations|TestPublishAllocations|TestPublishSubscribersAllocations|TestPublishStashSlideAllocations|TestPublishOverviewAllocations|TestPublishLeavingAllocations|TestOutputRoutingAllocations|TestReportSeenAllocations|TestFrameLifecycleTransitionsAllocations|TestFullscreenShownAllocations|TestFrameDecisionAllocations|TestCommitFrameAllocations|TestColorPipelineAllocations|TestShownBySnapshotAllocations|TestMRUNoteFocusDoesNotAllocateWhenSettled|TestDueFramesAllocations|TestRenderSteadyStateAllocations|TestSceneOpaqueRegionAllocations|TestSceneDeltaAllocations|TestAccountFlipAllocations|TestFlipDoneAllocations|TestSceneWalkUnchangedTiledSHM|TestTiledCommitPublishAllocations|TestCapturedCommitApplyAllocations|TestCapturedViewportApplyAllocations|TestEffectiveInputEmptyTreeAllocations|TestHeadlessTiledSHMCallbackAndCopyBudget|TestReleaseHeldAllocations|TestCommitDamageAllocations|TestEmitContentAllocations)$$' -count=1
7373 check: vet test arch fakes-check log-check perf-check
7474
7575 # On-demand SDR/HDR colour check with examples/testpattern on a headless
@@ -69,7 +69,7 @@ adapter-imports:
69 [ -z "$$bad" ] || { echo "$$bad"; echo 'adapter imports another adapter: use ports or a shared library (Makefile ADAPTER_LIBS)' >&2; exit 1; }69 [ -z "$$bad" ] || { echo "$$bad"; echo 'adapter imports another adapter: use ports or a shared library (Makefile ADAPTER_LIBS)' >&2; exit 1; }
70 # Fast, mandatory guards for tiled content publications and SHM copies.70 # Fast, mandatory guards for tiled content publications and SHM copies.
71 perf-check:71 perf-check:
72- CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations|TestPublishAllocations|TestPublishSubscribersAllocations|TestPublishStashSlideAllocations|TestPublishOverviewAllocations|TestPublishLeavingAllocations|TestOutputRoutingAllocations|TestReportSeenAllocations|TestFrameLifecycleTransitionsAllocations|TestFullscreenShownAllocations|TestFrameDecisionAllocations|TestCommitFrameAllocations|TestColorPipelineAllocations|TestShownBySnapshotAllocations|TestMRUNoteFocusDoesNotAllocateWhenSettled|TestDueFramesAllocations|TestRenderSteadyStateAllocations|TestSceneOpaqueRegionAllocations|TestSceneDeltaAllocations|TestAccountFlipAllocations|TestFlipDoneAllocations|TestSceneWalkUnchangedTiledSHM|TestTiledCommitPublishAllocations|TestCapturedCommitApplyAllocations|TestCapturedViewportApplyAllocations|TestEffectiveInputEmptyTreeAllocations|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=172+ CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations|TestPublishAllocations|TestPublishSubscribersAllocations|TestPublishStashSlideAllocations|TestPublishOverviewAllocations|TestPublishLeavingAllocations|TestOutputRoutingAllocations|TestReportSeenAllocations|TestFrameLifecycleTransitionsAllocations|TestFullscreenShownAllocations|TestFrameDecisionAllocations|TestCommitFrameAllocations|TestColorPipelineAllocations|TestShownBySnapshotAllocations|TestMRUNoteFocusDoesNotAllocateWhenSettled|TestDueFramesAllocations|TestRenderSteadyStateAllocations|TestSceneOpaqueRegionAllocations|TestSceneDeltaAllocations|TestAccountFlipAllocations|TestFlipDoneAllocations|TestSceneWalkUnchangedTiledSHM|TestTiledCommitPublishAllocations|TestCapturedCommitApplyAllocations|TestCapturedViewportApplyAllocations|TestEffectiveInputEmptyTreeAllocations|TestHeadlessTiledSHMCallbackAndCopyBudget|TestReleaseHeldAllocations|TestCommitDamageAllocations|TestEmitContentAllocations)$$' -count=1
73 check: vet test arch fakes-check log-check perf-check73 check: vet test arch fakes-check log-check perf-check
74 74
75 # On-demand SDR/HDR colour check with examples/testpattern on a headless75 # On-demand SDR/HDR colour check with examples/testpattern on a headless
M internal/adapters/drm/output.go +113 -37
@@ -78,6 +78,8 @@ type Output struct {
7878 // seatDisable carries a request to leave the planes on Bypass before
7979 // the seat is disabled (PrepareSeatDisable); Run answers it.
8080 seatDisable chan chan struct{}
81+ // ping carries watchdog probes (Ping); Run answers each one.
82+ ping chan chan struct{}
8183 // fbs are the two renderer images frames alternate between, back the
8284 // one the next frame draws into (ADR 014: zero copy).
8385 fbs [2]uint32
@@ -206,7 +208,7 @@ type CursorLoader func(c ports.CursorChange, scale float64, t ports.BufferTransf
206208 func newOutput(card *Card, c connector, mode modeInfo, crtc uint32) (*Output, error) {
207209 log := card.log
208210 pipe := slices.Index(card.crtcs, crtc)
209- o := &Output{k: card.k, flipped: card.flips[crtc], frame: frameLifecycle{serials: &card.serials}, formats: card.formats, sampled: card.want.Sampled, device: card.want.Device, crtc: crtc, conn: c, mode: mode, log: log, renderLog: card.renderLog, monitor: readMonitor(card.path, c.name), scanout: !card.want.NoScanout, clientFBs: map[uint64]*clientFB{}, reason: "start", ready: make(chan error, 1), seatDisable: make(chan chan struct{}), traceFlips: card.want.TraceFlips, vrrFlipGap: card.want.VRRFlipGap}
211+ o := &Output{k: card.k, flipped: card.flips[crtc], frame: frameLifecycle{serials: &card.serials}, formats: card.formats, sampled: card.want.Sampled, device: card.want.Device, crtc: crtc, conn: c, mode: mode, log: log, renderLog: card.renderLog, monitor: readMonitor(card.path, c.name), scanout: !card.want.NoScanout, clientFBs: map[uint64]*clientFB{}, reason: "start", ready: make(chan error, 1), seatDisable: make(chan chan struct{}), ping: make(chan chan struct{}), traceFlips: card.want.TraceFlips, vrrFlipGap: card.want.VRRFlipGap}
210212 var err error
211213 if o.saved, err = getCrtc(card.fd, crtc); err != nil {
212214 log.Warn().Err(err).Uint32("crtc", crtc).Msg("save crtc; it will not be restored on exit")
@@ -319,6 +321,21 @@ func cursorSize(fd int) int {
319321 // Cursor is the hardware cursor, or nil when the CRTC has no cursor plane.
320322 func (o *Output) Cursor() *Cursor { return o.cursor }
321323
324+// Ping returns once Run handled a probe, or once ctx ends: the watchdog
325+// uses it to detect a stuck output loop.
326+func (o *Output) Ping(ctx context.Context) {
327+ ack := make(chan struct{})
328+ select {
329+ case o.ping <- ack:
330+ case <-ctx.Done():
331+ return
332+ }
333+ select {
334+ case <-ack:
335+ case <-ctx.Done():
336+ }
337+}
338+
322339 // Info describes the output for core and clients. Unknown EDID fields are empty.
323340 func (o *Output) Info() ports.OutputInfo {
324341 known := func(v string) string {
@@ -1098,6 +1115,64 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
10981115 cursorWake = o.cursor.wake
10991116 }
11001117 securityChanges := o.SecurityChanges
1118+ reacquireTimer := time.NewTimer(time.Hour)
1119+ reacquireTimer.Stop()
1120+ defer reacquireTimer.Stop()
1121+ var reacquire masterRetry
1122+ // refuse records an enable refused for lack of DRM master.
1123+ refuseMaster := func(err error) {
1124+ if reacquire.refused(err, enabled, time.Now()) {
1125+ o.log.Warn().Err(err).Str("connector", o.conn.name).Int("attempts", masterRetryLimit).Msg("DRM master not back; output stays off until the next seat change")
1126+ }
1127+ }
1128+ // setActive applies a seat change. skip reports that the loop must
1129+ // go on without touching the frame state (refused or protected).
1130+ setActive := func(on bool) (skip bool, err error) {
1131+ o.observeSecurity()
1132+ if o.protected {
1133+ o.invalidateSecurity()
1134+ enabled = on
1135+ if on && o.fbs[0] == 0 {
1136+ if err := o.showImages(r, imagesDriver, nil); err != nil {
1137+ if !o.commitFailed(err, &enabled) {
1138+ return true, err
1139+ }
1140+ refuseMaster(err)
1141+ return true, nil
1142+ }
1143+ }
1144+ // With images, the protected modeset runs at the top of the
1145+ // loop (prepareSecurity): it ends or reschedules the retry.
1146+ return true, nil
1147+ }
1148+ // Modeset on every enable: a fast disable+enable can coalesce to one true.
1149+ if on {
1150+ enabled = true // this enable may supersede a previous master loss
1151+ var err error
1152+ if o.fbs[0] == 0 {
1153+ err = o.showImages(r, imagesDriver, nil)
1154+ } else {
1155+ err = o.modeset()
1156+ }
1157+ if err != nil {
1158+ if !o.commitFailed(err, &enabled) {
1159+ return true, err
1160+ }
1161+ refuseMaster(err)
1162+ return true, nil // epoch retry or wait disabled for master reacquisition
1163+ }
1164+ reacquire.done()
1165+ o.observeSecurity()
1166+ if o.protected {
1167+ return true, nil
1168+ }
1169+ o.probeAsync(r)
1170+ dirty = haveScene
1171+ }
1172+ enabled = on
1173+ o.log.Info().Bool("enabled", on).Msg("output")
1174+ return false, nil
1175+ }
11011176 for {
11021177 if o.observeSecurity() || haveScene && !o.sceneCurrent(scene) {
11031178 scene, haveScene, dirty = ports.Scene{}, false, false
@@ -1118,6 +1193,11 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
11181193 if !o.commitFailed(err, &enabled) {
11191194 return err
11201195 }
1196+ // A lock screen refused for lack of DRM master after a VT
1197+ // switch is retried like an ordinary enable.
1198+ refuseMaster(err)
1199+ } else {
1200+ reacquire.done()
11211201 }
11221202 // Recheck immediately: a takeover during the clear needs its own
11231203 // transaction, not another scene or wake to make progress.
@@ -1203,6 +1283,14 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
12031283 } else {
12041284 protectTimer.Stop()
12051285 }
1286+ // An enable refused for lack of DRM master is retried on time.
1287+ var reacquireDue <-chan time.Time
1288+ if d, ok := reacquire.next(time.Now()); ok {
1289+ reacquireTimer.Reset(d)
1290+ reacquireDue = reacquireTimer.C
1291+ } else {
1292+ reacquireTimer.Stop()
1293+ }
12061294 // A commit whose event never comes must not stop the output.
12071295 stuck := o.frame.wait(enabled)
12081296 stateDirty := false
@@ -1260,52 +1348,40 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
12601348 refuse = capture.ErrOutputOff
12611349 }
12621350 dirty = requests.Admit(q, refuse, scene, haveScene) || dirty
1351+ case ack := <-o.ping:
1352+ close(ack)
1353+ continue
12631354 case ack := <-o.seatDisable:
12641355 o.bypassForSeatDisable()
12651356 // The seat is about to be disabled: no frame may select the
1266- // pipeline again before the next enable (a modeset).
1357+ // pipeline again before the next enable (a modeset), not even a
1358+ // retried one.
1359+ reacquire.done()
12671360 enabled = false
12681361 ack <- struct{}{}
12691362 continue
12701363 case on := <-active:
1271- o.observeSecurity()
1272- if o.protected {
1273- o.invalidateSecurity()
1274- enabled = on
1275- if on && o.fbs[0] == 0 {
1276- if err := o.showImages(r, imagesDriver, nil); err != nil {
1277- if !o.commitFailed(err, &enabled) {
1278- return err
1279- }
1280- }
1281- }
1364+ reacquire.done()
1365+ skip, err := setActive(on)
1366+ if err != nil {
1367+ return err
1368+ }
1369+ if skip {
12821370 continue
12831371 }
1284- // Modeset on every enable: a fast disable+enable can coalesce to one true.
1285- if on {
1286- enabled = true // this enable may supersede a previous master loss
1287- var err error
1288- if o.fbs[0] == 0 {
1289- err = o.showImages(r, imagesDriver, nil)
1290- } else {
1291- err = o.modeset()
1292- }
1293- if err != nil {
1294- if !o.commitFailed(err, &enabled) {
1295- return err
1296- }
1297- continue // epoch retry or wait disabled for master reacquisition
1298- }
1299- o.observeSecurity()
1300- if o.protected {
1301- continue
1302- }
1303- o.probeAsync(r)
1304- dirty = haveScene
1305- stateDirty = true
1372+ stateDirty = on
1373+ case <-reacquireDue:
1374+ // The seat reported active before the kernel gave DRM master
1375+ // back: the enable modeset was refused. Try it again.
1376+ reacquire.fire()
1377+ skip, err := setActive(true)
1378+ if err != nil {
1379+ return err
13061380 }
1307- enabled = on
1308- o.log.Info().Bool("enabled", on).Msg("output")
1381+ if skip {
1382+ continue
1383+ }
1384+ stateDirty = true
13091385 case ev := <-o.flipped:
13101386 o.observeSecurity()
13111387 if !o.completed(ev, seen) {
@@ -78,6 +78,8 @@ type Output struct {
78 // seatDisable carries a request to leave the planes on Bypass before78 // seatDisable carries a request to leave the planes on Bypass before
79 // the seat is disabled (PrepareSeatDisable); Run answers it.79 // the seat is disabled (PrepareSeatDisable); Run answers it.
80 seatDisable chan chan struct{}80 seatDisable chan chan struct{}
81+ // ping carries watchdog probes (Ping); Run answers each one.
82+ ping chan chan struct{}
81 // fbs are the two renderer images frames alternate between, back the83 // fbs are the two renderer images frames alternate between, back the
82 // one the next frame draws into (ADR 014: zero copy).84 // one the next frame draws into (ADR 014: zero copy).
83 fbs [2]uint3285 fbs [2]uint32
@@ -206,7 +208,7 @@ type CursorLoader func(c ports.CursorChange, scale float64, t ports.BufferTransf
206 func newOutput(card *Card, c connector, mode modeInfo, crtc uint32) (*Output, error) {208 func newOutput(card *Card, c connector, mode modeInfo, crtc uint32) (*Output, error) {
207 log := card.log209 log := card.log
208 pipe := slices.Index(card.crtcs, crtc)210 pipe := slices.Index(card.crtcs, crtc)
209- o := &Output{k: card.k, flipped: card.flips[crtc], frame: frameLifecycle{serials: &card.serials}, formats: card.formats, sampled: card.want.Sampled, device: card.want.Device, crtc: crtc, conn: c, mode: mode, log: log, renderLog: card.renderLog, monitor: readMonitor(card.path, c.name), scanout: !card.want.NoScanout, clientFBs: map[uint64]*clientFB{}, reason: "start", ready: make(chan error, 1), seatDisable: make(chan chan struct{}), traceFlips: card.want.TraceFlips, vrrFlipGap: card.want.VRRFlipGap}211+ o := &Output{k: card.k, flipped: card.flips[crtc], frame: frameLifecycle{serials: &card.serials}, formats: card.formats, sampled: card.want.Sampled, device: card.want.Device, crtc: crtc, conn: c, mode: mode, log: log, renderLog: card.renderLog, monitor: readMonitor(card.path, c.name), scanout: !card.want.NoScanout, clientFBs: map[uint64]*clientFB{}, reason: "start", ready: make(chan error, 1), seatDisable: make(chan chan struct{}), ping: make(chan chan struct{}), traceFlips: card.want.TraceFlips, vrrFlipGap: card.want.VRRFlipGap}
210 var err error212 var err error
211 if o.saved, err = getCrtc(card.fd, crtc); err != nil {213 if o.saved, err = getCrtc(card.fd, crtc); err != nil {
212 log.Warn().Err(err).Uint32("crtc", crtc).Msg("save crtc; it will not be restored on exit")214 log.Warn().Err(err).Uint32("crtc", crtc).Msg("save crtc; it will not be restored on exit")
@@ -319,6 +321,21 @@ func cursorSize(fd int) int {
319 // Cursor is the hardware cursor, or nil when the CRTC has no cursor plane.321 // Cursor is the hardware cursor, or nil when the CRTC has no cursor plane.
320 func (o *Output) Cursor() *Cursor { return o.cursor }322 func (o *Output) Cursor() *Cursor { return o.cursor }
321 323
324+// Ping returns once Run handled a probe, or once ctx ends: the watchdog
325+// uses it to detect a stuck output loop.
326+func (o *Output) Ping(ctx context.Context) {
327+ ack := make(chan struct{})
328+ select {
329+ case o.ping <- ack:
330+ case <-ctx.Done():
331+ return
332+ }
333+ select {
334+ case <-ack:
335+ case <-ctx.Done():
336+ }
337+}
338+
322 // Info describes the output for core and clients. Unknown EDID fields are empty.339 // Info describes the output for core and clients. Unknown EDID fields are empty.
323 func (o *Output) Info() ports.OutputInfo {340 func (o *Output) Info() ports.OutputInfo {
324 known := func(v string) string {341 known := func(v string) string {
@@ -1098,6 +1115,64 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1098 cursorWake = o.cursor.wake1115 cursorWake = o.cursor.wake
1099 }1116 }
1100 securityChanges := o.SecurityChanges1117 securityChanges := o.SecurityChanges
1118+ reacquireTimer := time.NewTimer(time.Hour)
1119+ reacquireTimer.Stop()
1120+ defer reacquireTimer.Stop()
1121+ var reacquire masterRetry
1122+ // refuse records an enable refused for lack of DRM master.
1123+ refuseMaster := func(err error) {
1124+ if reacquire.refused(err, enabled, time.Now()) {
1125+ o.log.Warn().Err(err).Str("connector", o.conn.name).Int("attempts", masterRetryLimit).Msg("DRM master not back; output stays off until the next seat change")
1126+ }
1127+ }
1128+ // setActive applies a seat change. skip reports that the loop must
1129+ // go on without touching the frame state (refused or protected).
1130+ setActive := func(on bool) (skip bool, err error) {
1131+ o.observeSecurity()
1132+ if o.protected {
1133+ o.invalidateSecurity()
1134+ enabled = on
1135+ if on && o.fbs[0] == 0 {
1136+ if err := o.showImages(r, imagesDriver, nil); err != nil {
1137+ if !o.commitFailed(err, &enabled) {
1138+ return true, err
1139+ }
1140+ refuseMaster(err)
1141+ return true, nil
1142+ }
1143+ }
1144+ // With images, the protected modeset runs at the top of the
1145+ // loop (prepareSecurity): it ends or reschedules the retry.
1146+ return true, nil
1147+ }
1148+ // Modeset on every enable: a fast disable+enable can coalesce to one true.
1149+ if on {
1150+ enabled = true // this enable may supersede a previous master loss
1151+ var err error
1152+ if o.fbs[0] == 0 {
1153+ err = o.showImages(r, imagesDriver, nil)
1154+ } else {
1155+ err = o.modeset()
1156+ }
1157+ if err != nil {
1158+ if !o.commitFailed(err, &enabled) {
1159+ return true, err
1160+ }
1161+ refuseMaster(err)
1162+ return true, nil // epoch retry or wait disabled for master reacquisition
1163+ }
1164+ reacquire.done()
1165+ o.observeSecurity()
1166+ if o.protected {
1167+ return true, nil
1168+ }
1169+ o.probeAsync(r)
1170+ dirty = haveScene
1171+ }
1172+ enabled = on
1173+ o.log.Info().Bool("enabled", on).Msg("output")
1174+ return false, nil
1175+ }
1101 for {1176 for {
1102 if o.observeSecurity() || haveScene && !o.sceneCurrent(scene) {1177 if o.observeSecurity() || haveScene && !o.sceneCurrent(scene) {
1103 scene, haveScene, dirty = ports.Scene{}, false, false1178 scene, haveScene, dirty = ports.Scene{}, false, false
@@ -1118,6 +1193,11 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1118 if !o.commitFailed(err, &enabled) {1193 if !o.commitFailed(err, &enabled) {
1119 return err1194 return err
1120 }1195 }
1196+ // A lock screen refused for lack of DRM master after a VT
1197+ // switch is retried like an ordinary enable.
1198+ refuseMaster(err)
1199+ } else {
1200+ reacquire.done()
1121 }1201 }
1122 // Recheck immediately: a takeover during the clear needs its own1202 // Recheck immediately: a takeover during the clear needs its own
1123 // transaction, not another scene or wake to make progress.1203 // transaction, not another scene or wake to make progress.
@@ -1203,6 +1283,14 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1203 } else {1283 } else {
1204 protectTimer.Stop()1284 protectTimer.Stop()
1205 }1285 }
1286+ // An enable refused for lack of DRM master is retried on time.
1287+ var reacquireDue <-chan time.Time
1288+ if d, ok := reacquire.next(time.Now()); ok {
1289+ reacquireTimer.Reset(d)
1290+ reacquireDue = reacquireTimer.C
1291+ } else {
1292+ reacquireTimer.Stop()
1293+ }
1206 // A commit whose event never comes must not stop the output.1294 // A commit whose event never comes must not stop the output.
1207 stuck := o.frame.wait(enabled)1295 stuck := o.frame.wait(enabled)
1208 stateDirty := false1296 stateDirty := false
@@ -1260,52 +1348,40 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1260 refuse = capture.ErrOutputOff1348 refuse = capture.ErrOutputOff
1261 }1349 }
1262 dirty = requests.Admit(q, refuse, scene, haveScene) || dirty1350 dirty = requests.Admit(q, refuse, scene, haveScene) || dirty
1351+ case ack := <-o.ping:
1352+ close(ack)
1353+ continue
1263 case ack := <-o.seatDisable:1354 case ack := <-o.seatDisable:
1264 o.bypassForSeatDisable()1355 o.bypassForSeatDisable()
1265 // The seat is about to be disabled: no frame may select the1356 // The seat is about to be disabled: no frame may select the
1266- // pipeline again before the next enable (a modeset).1357+ // pipeline again before the next enable (a modeset), not even a
1358+ // retried one.
1359+ reacquire.done()
1267 enabled = false1360 enabled = false
1268 ack <- struct{}{}1361 ack <- struct{}{}
1269 continue1362 continue
1270 case on := <-active:1363 case on := <-active:
1271- o.observeSecurity()1364+ reacquire.done()
1272- if o.protected {1365+ skip, err := setActive(on)
1273- o.invalidateSecurity()1366+ if err != nil {
1274- enabled = on1367+ return err
1275- if on && o.fbs[0] == 0 {1368+ }
1276- if err := o.showImages(r, imagesDriver, nil); err != nil {1369+ if skip {
1277- if !o.commitFailed(err, &enabled) {
1278- return err
1279- }
1280- }
1281- }
1282 continue1370 continue
1283 }1371 }
1284- // Modeset on every enable: a fast disable+enable can coalesce to one true.1372+ stateDirty = on
1285- if on {1373+ case <-reacquireDue:
1286- enabled = true // this enable may supersede a previous master loss1374+ // The seat reported active before the kernel gave DRM master
1287- var err error1375+ // back: the enable modeset was refused. Try it again.
1288- if o.fbs[0] == 0 {1376+ reacquire.fire()
1289- err = o.showImages(r, imagesDriver, nil)1377+ skip, err := setActive(true)
1290- } else {1378+ if err != nil {
1291- err = o.modeset()1379+ return err
1292- }
1293- if err != nil {
1294- if !o.commitFailed(err, &enabled) {
1295- return err
1296- }
1297- continue // epoch retry or wait disabled for master reacquisition
1298- }
1299- o.observeSecurity()
1300- if o.protected {
1301- continue
1302- }
1303- o.probeAsync(r)
1304- dirty = haveScene
1305- stateDirty = true
1306 }1380 }
1307- enabled = on1381+ if skip {
1308- o.log.Info().Bool("enabled", on).Msg("output")1382+ continue
1383+ }
1384+ stateDirty = true
1309 case ev := <-o.flipped:1385 case ev := <-o.flipped:
1310 o.observeSecurity()1386 o.observeSecurity()
1311 if !o.completed(ev, seen) {1387 if !o.completed(ev, seen) {
A internal/adapters/drm/reacquire.go +56 -0
@@ -0,0 +1,56 @@
1+package drm
2+
3+import "time"
4+
5+// masterRetry retries an enable modeset that the kernel refused because
6+// DRM master was not back yet. The seat can report the session active
7+// before the kernel hands master back; without a retry the output stays
8+// dark until the next VT switch.
9+//
10+// The retry has a fixed deadline: the output loop wakes for many other
11+// events (pings, cursor, reports) and must not push it back on each one.
12+type masterRetry struct {
13+ pending bool
14+ due time.Time
15+ attempts int
16+}
17+
18+const (
19+ masterRetryFirst = 20 * time.Millisecond
20+ masterRetryMax = time.Second
21+ // masterRetryLimit bounds the retries of one enable (about 35 s).
22+ masterRetryLimit = 40
23+)
24+
25+// refused records an enable refused with err at now; only a lost master
26+// is retried, and enabled stays false until a retry succeeds. It reports
27+// that the retries are exhausted.
28+func (m *masterRetry) refused(err error, enabled bool, now time.Time) (gaveUp bool) {
29+ if !lostMaster(err) || enabled {
30+ return false
31+ }
32+ if m.attempts >= masterRetryLimit {
33+ m.pending = false
34+ return true
35+ }
36+ m.pending = true
37+ m.due = now.Add(min(masterRetryFirst<<min(m.attempts, 6), masterRetryMax))
38+ return false
39+}
40+
41+// fire starts an attempt: the retry is pending again only if it is refused.
42+func (m *masterRetry) fire() {
43+ m.pending = false
44+ m.attempts++
45+}
46+
47+// done ends the retry: the enable succeeded or the seat changed.
48+func (m *masterRetry) done() { *m = masterRetry{} }
49+
50+// next is the time left before the next attempt, if one is due.
51+func (m *masterRetry) next(now time.Time) (time.Duration, bool) {
52+ if !m.pending {
53+ return 0, false
54+ }
55+ return max(0, m.due.Sub(now)), true
56+}
@@ -0,0 +1,56 @@
1+package drm
2+
3+import "time"
4+
5+// masterRetry retries an enable modeset that the kernel refused because
6+// DRM master was not back yet. The seat can report the session active
7+// before the kernel hands master back; without a retry the output stays
8+// dark until the next VT switch.
9+//
10+// The retry has a fixed deadline: the output loop wakes for many other
11+// events (pings, cursor, reports) and must not push it back on each one.
12+type masterRetry struct {
13+ pending bool
14+ due time.Time
15+ attempts int
16+}
17+
18+const (
19+ masterRetryFirst = 20 * time.Millisecond
20+ masterRetryMax = time.Second
21+ // masterRetryLimit bounds the retries of one enable (about 35 s).
22+ masterRetryLimit = 40
23+)
24+
25+// refused records an enable refused with err at now; only a lost master
26+// is retried, and enabled stays false until a retry succeeds. It reports
27+// that the retries are exhausted.
28+func (m *masterRetry) refused(err error, enabled bool, now time.Time) (gaveUp bool) {
29+ if !lostMaster(err) || enabled {
30+ return false
31+ }
32+ if m.attempts >= masterRetryLimit {
33+ m.pending = false
34+ return true
35+ }
36+ m.pending = true
37+ m.due = now.Add(min(masterRetryFirst<<min(m.attempts, 6), masterRetryMax))
38+ return false
39+}
40+
41+// fire starts an attempt: the retry is pending again only if it is refused.
42+func (m *masterRetry) fire() {
43+ m.pending = false
44+ m.attempts++
45+}
46+
47+// done ends the retry: the enable succeeded or the seat changed.
48+func (m *masterRetry) done() { *m = masterRetry{} }
49+
50+// next is the time left before the next attempt, if one is due.
51+func (m *masterRetry) next(now time.Time) (time.Duration, bool) {
52+ if !m.pending {
53+ return 0, false
54+ }
55+ return max(0, m.due.Sub(now)), true
56+}
A internal/adapters/drm/reacquire_test.go +184 -0
@@ -0,0 +1,184 @@
1+package drm
2+
3+import (
4+ "context"
5+ "errors"
6+ "sync"
7+ "testing"
8+ "testing/synctest"
9+ "time"
10+
11+ portsmocks "github.com/bnema/neferwl/internal/mocks/ports"
12+ "github.com/bnema/neferwl/internal/ports"
13+ "github.com/stretchr/testify/mock"
14+ "golang.org/x/sys/unix"
15+)
16+
17+func TestMasterRetryBackoff(t *testing.T) {
18+ var m masterRetry
19+ now := time.Unix(0, 0)
20+ if _, ok := m.next(now); ok {
21+ t.Fatal("retry due before a refusal")
22+ }
23+ if m.refused(unix.EBUSY, false, now); m.pending {
24+ t.Fatal("a busy refusal is retried by the busy path, not here")
25+ }
26+ if m.refused(unix.EACCES, true, now); m.pending {
27+ t.Fatal("an enabled output is not retried")
28+ }
29+ m.refused(unix.EACCES, false, now)
30+ // The deadline is fixed: asking later leaves less, never more.
31+ if d, ok := m.next(now); !ok || d != masterRetryFirst {
32+ t.Fatalf("first retry %v %v", d, ok)
33+ }
34+ if d, _ := m.next(now.Add(masterRetryFirst / 2)); d != masterRetryFirst/2 {
35+ t.Fatalf("half-way %v", d)
36+ }
37+ if d, _ := m.next(now.Add(time.Hour)); d != 0 {
38+ t.Fatalf("overdue %v", d)
39+ }
40+ m.fire()
41+ if _, ok := m.next(now); ok {
42+ t.Fatal("pending while an attempt runs")
43+ }
44+ m.attempts = 30
45+ m.refused(unix.EPERM, false, now)
46+ if d, _ := m.next(now); d != masterRetryMax {
47+ t.Fatalf("capped retry %v", d)
48+ }
49+ m.attempts = masterRetryLimit
50+ if !m.refused(unix.EACCES, false, now) || m.pending {
51+ t.Fatal("retries past the limit")
52+ }
53+ m.done()
54+ if _, ok := m.next(now); ok || m.attempts != 0 {
55+ t.Fatal("retry after done")
56+ }
57+}
58+
59+// retryHarness runs an output whose enable modesets the kernel refuses
60+// (lost master) until allow is called.
61+type retryHarness struct {
62+ o *Output
63+ active chan bool
64+ modesets func() int
65+ allow func()
66+ stop func()
67+}
68+
69+func newRetryHarness(t *testing.T) retryHarness {
70+ t.Helper()
71+ var mu sync.Mutex
72+ refuse := true
73+ o, k, commits, commitMu := testOutputRules(t, func(_ *atomicReq, flags uint32) error {
74+ mu.Lock()
75+ defer mu.Unlock()
76+ if refuse && flags&atomicAllowModes != 0 && flags&atomicTestOnly == 0 {
77+ return unix.EACCES
78+ }
79+ return nil
80+ })
81+ o.cursor, o.tearing = nil, false
82+ r := portsmocks.NewMockRenderer(t)
83+ r.EXPECT().SetHDR(mock.Anything).Return().Maybe()
84+ r.EXPECT().UseTarget(mock.Anything).Return().Maybe()
85+ r.EXPECT().Render(mock.Anything, mock.Anything).Return(nil, nil).Maybe()
86+ r.EXPECT().Close().Return().Once()
87+ // The 10 s stats tick runs in long tests.
88+ r.EXPECT().Trim(mock.Anything).Return(nil).Maybe()
89+ r.EXPECT().TakeRedrawn().Return(0).Maybe()
90+ k.EXPECT().createBlob(mock.Anything).Return(99, nil).Maybe()
91+ k.EXPECT().destroyBlob(mock.Anything).Return(nil).Maybe()
92+ k.EXPECT().rmFB(mock.Anything).Return(nil).Maybe()
93+ o.ping = make(chan chan struct{})
94+ active := make(chan bool, 1)
95+ active <- false
96+ ctx, cancel := context.WithCancel(context.Background())
97+ done := make(chan error, 1)
98+ go func() {
99+ done <- o.Run(ctx, func(int, int) (ports.Renderer, error) { return r, nil }, nil, active, nil, nil, nil, make(chan ports.OutputPresented, 1), nil, nil)
100+ }()
101+ synctest.Wait()
102+ return retryHarness{
103+ o: o, active: active,
104+ modesets: func() int {
105+ commitMu.Lock()
106+ defer commitMu.Unlock()
107+ n := 0
108+ for _, c := range *commits {
109+ if c.flags&atomicAllowModes != 0 && c.flags&atomicTestOnly == 0 {
110+ n++
111+ }
112+ }
113+ return n
114+ },
115+ allow: func() { mu.Lock(); refuse = false; mu.Unlock() },
116+ stop: func() {
117+ cancel()
118+ if err := <-done; err != nil && !errors.Is(err, context.Canceled) {
119+ t.Fatal(err)
120+ }
121+ },
122+ }
123+}
124+
125+// The seat reports active before the kernel gives DRM master back: the
126+// enable modeset is refused, then retried until it succeeds, without a
127+// second seat change, even while other events (watchdog pings) keep
128+// waking the loop faster than the backoff.
129+func TestEnableRefusedForMasterIsRetried(t *testing.T) {
130+ synctest.Test(t, func(t *testing.T) {
131+ h := newRetryHarness(t)
132+ h.active <- true
133+ synctest.Wait()
134+ if n := h.modesets(); n != 1 {
135+ t.Fatalf("%d modesets after the refused enable", n)
136+ }
137+ // Let the backoff grow past the ping period.
138+ pingFor := func(d time.Duration) {
139+ for end := time.Now().Add(d); time.Now().Before(end); {
140+ time.Sleep(100 * time.Millisecond)
141+ h.o.Ping(context.Background())
142+ }
143+ synctest.Wait()
144+ }
145+ pingFor(3 * time.Second)
146+ before := h.modesets()
147+ if before < 6 {
148+ t.Fatalf("%d modesets in 3 s of pings: the retry was pushed back", before)
149+ }
150+ h.allow()
151+ pingFor(2 * masterRetryMax)
152+ after := h.modesets()
153+ if after != before+1 {
154+ t.Fatalf("%d modesets before success, %d after", before, after)
155+ }
156+ // Once enabled, nothing retries any more.
157+ pingFor(2 * masterRetryMax)
158+ if n := h.modesets(); n != after {
159+ t.Fatalf("%d modesets after success", n)
160+ }
161+ h.stop()
162+ })
163+}
164+
165+// A seat disable ends a pending retry: no modeset may follow the plane
166+// reset before the next enable.
167+func TestSeatDisableCancelsMasterRetry(t *testing.T) {
168+ synctest.Test(t, func(t *testing.T) {
169+ h := newRetryHarness(t)
170+ h.active <- true
171+ synctest.Wait()
172+ ack := make(chan struct{})
173+ h.o.seatDisable <- ack
174+ <-ack
175+ synctest.Wait()
176+ n := h.modesets()
177+ time.Sleep(10 * masterRetryMax)
178+ synctest.Wait()
179+ if got := h.modesets(); got != n {
180+ t.Fatalf("%d modesets after the seat disable, want %d", got, n)
181+ }
182+ h.stop()
183+ })
184+}
@@ -0,0 +1,184 @@
1+package drm
2+
3+import (
4+ "context"
5+ "errors"
6+ "sync"
7+ "testing"
8+ "testing/synctest"
9+ "time"
10+
11+ portsmocks "github.com/bnema/neferwl/internal/mocks/ports"
12+ "github.com/bnema/neferwl/internal/ports"
13+ "github.com/stretchr/testify/mock"
14+ "golang.org/x/sys/unix"
15+)
16+
17+func TestMasterRetryBackoff(t *testing.T) {
18+ var m masterRetry
19+ now := time.Unix(0, 0)
20+ if _, ok := m.next(now); ok {
21+ t.Fatal("retry due before a refusal")
22+ }
23+ if m.refused(unix.EBUSY, false, now); m.pending {
24+ t.Fatal("a busy refusal is retried by the busy path, not here")
25+ }
26+ if m.refused(unix.EACCES, true, now); m.pending {
27+ t.Fatal("an enabled output is not retried")
28+ }
29+ m.refused(unix.EACCES, false, now)
30+ // The deadline is fixed: asking later leaves less, never more.
31+ if d, ok := m.next(now); !ok || d != masterRetryFirst {
32+ t.Fatalf("first retry %v %v", d, ok)
33+ }
34+ if d, _ := m.next(now.Add(masterRetryFirst / 2)); d != masterRetryFirst/2 {
35+ t.Fatalf("half-way %v", d)
36+ }
37+ if d, _ := m.next(now.Add(time.Hour)); d != 0 {
38+ t.Fatalf("overdue %v", d)
39+ }
40+ m.fire()
41+ if _, ok := m.next(now); ok {
42+ t.Fatal("pending while an attempt runs")
43+ }
44+ m.attempts = 30
45+ m.refused(unix.EPERM, false, now)
46+ if d, _ := m.next(now); d != masterRetryMax {
47+ t.Fatalf("capped retry %v", d)
48+ }
49+ m.attempts = masterRetryLimit
50+ if !m.refused(unix.EACCES, false, now) || m.pending {
51+ t.Fatal("retries past the limit")
52+ }
53+ m.done()
54+ if _, ok := m.next(now); ok || m.attempts != 0 {
55+ t.Fatal("retry after done")
56+ }
57+}
58+
59+// retryHarness runs an output whose enable modesets the kernel refuses
60+// (lost master) until allow is called.
61+type retryHarness struct {
62+ o *Output
63+ active chan bool
64+ modesets func() int
65+ allow func()
66+ stop func()
67+}
68+
69+func newRetryHarness(t *testing.T) retryHarness {
70+ t.Helper()
71+ var mu sync.Mutex
72+ refuse := true
73+ o, k, commits, commitMu := testOutputRules(t, func(_ *atomicReq, flags uint32) error {
74+ mu.Lock()
75+ defer mu.Unlock()
76+ if refuse && flags&atomicAllowModes != 0 && flags&atomicTestOnly == 0 {
77+ return unix.EACCES
78+ }
79+ return nil
80+ })
81+ o.cursor, o.tearing = nil, false
82+ r := portsmocks.NewMockRenderer(t)
83+ r.EXPECT().SetHDR(mock.Anything).Return().Maybe()
84+ r.EXPECT().UseTarget(mock.Anything).Return().Maybe()
85+ r.EXPECT().Render(mock.Anything, mock.Anything).Return(nil, nil).Maybe()
86+ r.EXPECT().Close().Return().Once()
87+ // The 10 s stats tick runs in long tests.
88+ r.EXPECT().Trim(mock.Anything).Return(nil).Maybe()
89+ r.EXPECT().TakeRedrawn().Return(0).Maybe()
90+ k.EXPECT().createBlob(mock.Anything).Return(99, nil).Maybe()
91+ k.EXPECT().destroyBlob(mock.Anything).Return(nil).Maybe()
92+ k.EXPECT().rmFB(mock.Anything).Return(nil).Maybe()
93+ o.ping = make(chan chan struct{})
94+ active := make(chan bool, 1)
95+ active <- false
96+ ctx, cancel := context.WithCancel(context.Background())
97+ done := make(chan error, 1)
98+ go func() {
99+ done <- o.Run(ctx, func(int, int) (ports.Renderer, error) { return r, nil }, nil, active, nil, nil, nil, make(chan ports.OutputPresented, 1), nil, nil)
100+ }()
101+ synctest.Wait()
102+ return retryHarness{
103+ o: o, active: active,
104+ modesets: func() int {
105+ commitMu.Lock()
106+ defer commitMu.Unlock()
107+ n := 0
108+ for _, c := range *commits {
109+ if c.flags&atomicAllowModes != 0 && c.flags&atomicTestOnly == 0 {
110+ n++
111+ }
112+ }
113+ return n
114+ },
115+ allow: func() { mu.Lock(); refuse = false; mu.Unlock() },
116+ stop: func() {
117+ cancel()
118+ if err := <-done; err != nil && !errors.Is(err, context.Canceled) {
119+ t.Fatal(err)
120+ }
121+ },
122+ }
123+}
124+
125+// The seat reports active before the kernel gives DRM master back: the
126+// enable modeset is refused, then retried until it succeeds, without a
127+// second seat change, even while other events (watchdog pings) keep
128+// waking the loop faster than the backoff.
129+func TestEnableRefusedForMasterIsRetried(t *testing.T) {
130+ synctest.Test(t, func(t *testing.T) {
131+ h := newRetryHarness(t)
132+ h.active <- true
133+ synctest.Wait()
134+ if n := h.modesets(); n != 1 {
135+ t.Fatalf("%d modesets after the refused enable", n)
136+ }
137+ // Let the backoff grow past the ping period.
138+ pingFor := func(d time.Duration) {
139+ for end := time.Now().Add(d); time.Now().Before(end); {
140+ time.Sleep(100 * time.Millisecond)
141+ h.o.Ping(context.Background())
142+ }
143+ synctest.Wait()
144+ }
145+ pingFor(3 * time.Second)
146+ before := h.modesets()
147+ if before < 6 {
148+ t.Fatalf("%d modesets in 3 s of pings: the retry was pushed back", before)
149+ }
150+ h.allow()
151+ pingFor(2 * masterRetryMax)
152+ after := h.modesets()
153+ if after != before+1 {
154+ t.Fatalf("%d modesets before success, %d after", before, after)
155+ }
156+ // Once enabled, nothing retries any more.
157+ pingFor(2 * masterRetryMax)
158+ if n := h.modesets(); n != after {
159+ t.Fatalf("%d modesets after success", n)
160+ }
161+ h.stop()
162+ })
163+}
164+
165+// A seat disable ends a pending retry: no modeset may follow the plane
166+// reset before the next enable.
167+func TestSeatDisableCancelsMasterRetry(t *testing.T) {
168+ synctest.Test(t, func(t *testing.T) {
169+ h := newRetryHarness(t)
170+ h.active <- true
171+ synctest.Wait()
172+ ack := make(chan struct{})
173+ h.o.seatDisable <- ack
174+ <-ack
175+ synctest.Wait()
176+ n := h.modesets()
177+ time.Sleep(10 * masterRetryMax)
178+ synctest.Wait()
179+ if got := h.modesets(); got != n {
180+ t.Fatalf("%d modesets after the seat disable, want %d", got, n)
181+ }
182+ h.stop()
183+ })
184+}
M internal/adapters/drm/security_resume_test.go +26 -19
@@ -16,13 +16,23 @@ import (
1616 func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
1717 for _, stage := range []string{"engage-during-request", "master-loss"} {
1818 t.Run(stage, func(t *testing.T) {
19- var failures []error
19+ var state atomic.Uint64
20+ var rule func(*atomicReq, uint32) error
2021 if stage == "master-loss" {
21- failures = []error{unix.EACCES}
22+ // The lock engages as the kernel refuses the first enable for
23+ // lack of master: the retry must be a protected modeset.
24+ refused := false
25+ rule = func(*atomicReq, uint32) error {
26+ if refused {
27+ return nil
28+ }
29+ refused = true
30+ state.Store(3)
31+ return unix.EACCES
32+ }
2233 }
23- o, k, commits, mu := testOutputMu(t, failures...)
34+ o, k, commits, mu := testOutputRules(t, rule)
2435 o.cursor, o.tearing = nil, false
25- var state atomic.Uint64
2636 securityGate(t, o, &state)
2737 events := make(chan ports.SecurityBackendEvent, 4)
2838 o.SecurityEvents = events
@@ -55,22 +65,11 @@ func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
5565 }()
5666 active <- true
5767 if stage == "master-loss" {
58- waitFor(t, func() bool { mu.Lock(); defer mu.Unlock(); return len(*commits) == 1 })
59- state.Store(3)
68+ waitFor(t, func() bool { mu.Lock(); defer mu.Unlock(); return len(*commits) >= 1 })
6069 wake <- ports.SecurityState{}
61- select {
62- case ev := <-events:
63- if _, ok := ev.(ports.SecurityOutputProof); ok {
64- t.Fatal("master loss forged proof")
65- }
66- case <-time.After(20 * time.Millisecond):
67- }
68- select {
69- case err := <-done:
70- t.Fatalf("resume stopped owner: %v", err)
71- default:
72- }
73- active <- true // only real seat reacquisition permits preparation
70+ // The seat reported active before master came back: the owner
71+ // retries the protected modeset itself, with no second seat
72+ // event. The proof below must follow that successful retry.
7473 }
7574 for {
7675 select {
@@ -79,6 +78,14 @@ func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
7978 if p.Proof.Generation != 1 || p.Proof.Instance != 77 {
8079 t.Fatalf("proof %+v", p)
8180 }
81+ if stage == "master-loss" {
82+ mu.Lock()
83+ n := len(*commits)
84+ mu.Unlock()
85+ if n < 2 {
86+ t.Fatal("master loss forged proof: no successful retry")
87+ }
88+ }
8289 cancel()
8390 if err := <-done; err != nil {
8491 t.Fatal(err)
@@ -16,13 +16,23 @@ import (
16 func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {16 func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
17 for _, stage := range []string{"engage-during-request", "master-loss"} {17 for _, stage := range []string{"engage-during-request", "master-loss"} {
18 t.Run(stage, func(t *testing.T) {18 t.Run(stage, func(t *testing.T) {
19- var failures []error19+ var state atomic.Uint64
20+ var rule func(*atomicReq, uint32) error
20 if stage == "master-loss" {21 if stage == "master-loss" {
21- failures = []error{unix.EACCES}22+ // The lock engages as the kernel refuses the first enable for
23+ // lack of master: the retry must be a protected modeset.
24+ refused := false
25+ rule = func(*atomicReq, uint32) error {
26+ if refused {
27+ return nil
28+ }
29+ refused = true
30+ state.Store(3)
31+ return unix.EACCES
32+ }
22 }33 }
23- o, k, commits, mu := testOutputMu(t, failures...)34+ o, k, commits, mu := testOutputRules(t, rule)
24 o.cursor, o.tearing = nil, false35 o.cursor, o.tearing = nil, false
25- var state atomic.Uint64
26 securityGate(t, o, &state)36 securityGate(t, o, &state)
27 events := make(chan ports.SecurityBackendEvent, 4)37 events := make(chan ports.SecurityBackendEvent, 4)
28 o.SecurityEvents = events38 o.SecurityEvents = events
@@ -55,22 +65,11 @@ func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
55 }()65 }()
56 active <- true66 active <- true
57 if stage == "master-loss" {67 if stage == "master-loss" {
58- waitFor(t, func() bool { mu.Lock(); defer mu.Unlock(); return len(*commits) == 1 })68+ waitFor(t, func() bool { mu.Lock(); defer mu.Unlock(); return len(*commits) >= 1 })
59- state.Store(3)
60 wake <- ports.SecurityState{}69 wake <- ports.SecurityState{}
61- select {70+ // The seat reported active before master came back: the owner
62- case ev := <-events:71+ // retries the protected modeset itself, with no second seat
63- if _, ok := ev.(ports.SecurityOutputProof); ok {72+ // event. The proof below must follow that successful retry.
64- t.Fatal("master loss forged proof")
65- }
66- case <-time.After(20 * time.Millisecond):
67- }
68- select {
69- case err := <-done:
70- t.Fatalf("resume stopped owner: %v", err)
71- default:
72- }
73- active <- true // only real seat reacquisition permits preparation
74 }73 }
75 for {74 for {
76 select {75 select {
@@ -79,6 +78,14 @@ func TestSecurityUnlockedResumeUsesOwnerRetry(t *testing.T) {
79 if p.Proof.Generation != 1 || p.Proof.Instance != 77 {78 if p.Proof.Generation != 1 || p.Proof.Instance != 77 {
80 t.Fatalf("proof %+v", p)79 t.Fatalf("proof %+v", p)
81 }80 }
81+ if stage == "master-loss" {
82+ mu.Lock()
83+ n := len(*commits)
84+ mu.Unlock()
85+ if n < 2 {
86+ t.Fatal("master loss forged proof: no successful retry")
87+ }
88+ }
82 cancel()89 cancel()
83 if err := <-done; err != nil {90 if err := <-done; err != nil {
84 t.Fatal(err)91 t.Fatal(err)
M internal/adapters/logging/logging.go +88 -11
@@ -4,14 +4,18 @@ package logging
44 import (
55 "context"
66 "fmt"
7+ "io"
78 "os"
89 "path/filepath"
10+ runtimedebug "runtime/debug"
911 "sort"
1012 "strings"
13+ "sync"
1114 "time"
1215
1316 "github.com/bnema/zerowrap"
1417 "github.com/rs/zerolog"
18+ "github.com/rs/zerolog/diode"
1519 "golang.org/x/term"
1620 )
1721
@@ -29,6 +33,7 @@ var categories = map[string]string{"input-motion": "input", "input-keys": "input
2933
3034 type debugKey struct{}
3135 type levelKey struct{}
36+type dirKey struct{}
3237 type debugSet map[string]bool
3338
3439 func ParseDebug(value string) (map[string]bool, error) {
@@ -56,8 +61,9 @@ const keepRuns = 20
5661 // directory at it, prunes old runs of that backend and attaches the logger
5762 // to ctx. Each backend keeps its own runs: headless test runs never rotate
5863 // a DRM session's log away.
59-// The caller must invoke close after all logging is complete.
60-func Open(ctx context.Context, backend, level, debug string) (context.Context, func() error, error) {
64+// The caller must invoke close after all logging is complete; crashed
65+// keeps the crash file for a panic about to be re-raised.
66+func Open(ctx context.Context, backend, level, debug string) (context.Context, func(crashed bool) error, error) {
6167 selected, err := ParseDebug(debug)
6268 if err != nil {
6369 return nil, nil, err
@@ -85,23 +91,88 @@ func Open(ctx context.Context, backend, level, debug string) (context.Context, f
8591 _ = file.Close()
8692 return nil, nil, err
8793 }
94+ // Queued entries are lost on a fatal crash: the runtime's own report
95+ // (panic and every stack) goes to its own file, besides stderr. A clean
96+ // Close removes it while still empty. Created before pruneRuns so
97+ // crash files rotate like logs.
98+ crashPath := filepath.Join(dir, strings.TrimSuffix(name, ".log")+".crash")
99+ if crash, err := os.OpenFile(crashPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600); err == nil {
100+ _ = runtimedebug.SetCrashOutput(crash, runtimedebug.CrashOptions{})
101+ _ = crash.Close() // SetCrashOutput keeps its own duplicate
102+ } else {
103+ crashPath = ""
104+ }
88105 pruneRuns(dir)
89- var output = zerowrap.Config{Level: level, Format: "json", Output: file}
106+ // Hide Close from the diode: asyncWriter closes the file once.
107+ var sink io.Writer = struct{ io.Writer }{file}
90108 if term.IsTerminal(int(os.Stderr.Fd())) {
91- output.Output = &consoleAndFile{file: file}
109+ sink = &consoleAndFile{file: file}
92110 }
93- ctx = zerowrap.WithCtx(ctx, zerowrap.New(output))
111+ w := newAsyncWriter(sink, file, crashPath)
112+ ctx = zerowrap.WithCtx(ctx, zerowrap.New(zerowrap.Config{Level: level, Format: "json", Output: w}))
94113 ctx = context.WithValue(ctx, debugKey{}, debugSet(selected))
95114 ctx = context.WithValue(ctx, levelKey{}, level)
96- return ctx, file.Close, nil
115+ ctx = context.WithValue(ctx, dirKey{}, dir)
116+ return ctx, func(crashed bool) error {
117+ if !crashed {
118+ w.Clean()
119+ }
120+ return w.Close()
121+ }, nil
122+}
123+
124+// logQueue is how many entries may wait for the disk. Past it, entries are
125+// dropped (and counted on stderr) rather than blocking the logging
126+// goroutine: input, outputs and the display must never wait on a slow disk.
127+const logQueue = 8192
128+
129+// newAsyncWriter writes entries to w from its own goroutine. Close drains
130+// what is queued, then closes c.
131+func newAsyncWriter(w io.Writer, c io.Closer, crashPath string) *asyncWriter {
132+ d := diode.NewWriter(w, logQueue, 0, func(missed int) {
133+ fmt.Fprintf(os.Stderr, "neferwl: %d log entries dropped: disk too slow\n", missed)
134+ })
135+ return &asyncWriter{Writer: d, closer: c, crashPath: crashPath}
136+}
137+
138+type asyncWriter struct {
139+ diode.Writer
140+ closer io.Closer
141+ crashPath string
142+ once sync.Once
143+ err error
144+}
145+
146+// Close drains the queue and closes the file once.
147+func (a *asyncWriter) Close() error {
148+ a.once.Do(func() {
149+ _ = a.Writer.Close() // drains the queue
150+ a.err = a.closer.Close()
151+ })
152+ return a.err
153+}
154+
155+// Clean ends a run that did not crash: the crash output is released and
156+// the crash file removed while still empty.
157+func (a *asyncWriter) Clean() {
158+ if a.crashPath == "" {
159+ return
160+ }
161+ _ = runtimedebug.SetCrashOutput(nil, runtimedebug.CrashOptions{})
162+ if fi, err := os.Stat(a.crashPath); err == nil && fi.Size() == 0 {
163+ _ = os.Remove(a.crashPath)
164+ }
97165 }
98166
99167 func pruneRuns(dir string) {
100- runs, _ := filepath.Glob(filepath.Join(dir, "2*.log"))
101- sort.Strings(runs)
102- for len(runs) > keepRuns {
103- _ = os.Remove(runs[0])
104- runs = runs[1:]
168+ // Watchdog stall dumps (stall-<time>-<loop>.txt) rotate like runs.
169+ for _, pattern := range []string{"2*.log", "2*.crash", "stall-*.txt"} {
170+ runs, _ := filepath.Glob(filepath.Join(dir, pattern))
171+ sort.Strings(runs)
172+ for len(runs) > keepRuns {
173+ _ = os.Remove(runs[0])
174+ runs = runs[1:]
175+ }
105176 }
106177 }
107178
@@ -113,6 +184,12 @@ func (w *consoleAndFile) Write(p []byte) (int, error) {
113184 return w.file.Write(p)
114185 }
115186
187+// Dir is the directory of the run log Open created, or "" without one.
188+func Dir(ctx context.Context) string {
189+ dir, _ := ctx.Value(dirKey{}).(string)
190+ return dir
191+}
192+
116193 // Enabled reports whether a debug category was named in --debug.
117194 func Enabled(ctx context.Context, category string) bool {
118195 selected, _ := ctx.Value(debugKey{}).(debugSet)
@@ -4,14 +4,18 @@ package logging
4 import (4 import (
5 "context"5 "context"
6 "fmt"6 "fmt"
7+ "io"
7 "os"8 "os"
8 "path/filepath"9 "path/filepath"
10+ runtimedebug "runtime/debug"
9 "sort"11 "sort"
10 "strings"12 "strings"
13+ "sync"
11 "time"14 "time"
12 15
13 "github.com/bnema/zerowrap"16 "github.com/bnema/zerowrap"
14 "github.com/rs/zerolog"17 "github.com/rs/zerolog"
18+ "github.com/rs/zerolog/diode"
15 "golang.org/x/term"19 "golang.org/x/term"
16 )20 )
17 21
@@ -29,6 +33,7 @@ var categories = map[string]string{"input-motion": "input", "input-keys": "input
29 33
30 type debugKey struct{}34 type debugKey struct{}
31 type levelKey struct{}35 type levelKey struct{}
36+type dirKey struct{}
32 type debugSet map[string]bool37 type debugSet map[string]bool
33 38
34 func ParseDebug(value string) (map[string]bool, error) {39 func ParseDebug(value string) (map[string]bool, error) {
@@ -56,8 +61,9 @@ const keepRuns = 20
56 // directory at it, prunes old runs of that backend and attaches the logger61 // directory at it, prunes old runs of that backend and attaches the logger
57 // to ctx. Each backend keeps its own runs: headless test runs never rotate62 // to ctx. Each backend keeps its own runs: headless test runs never rotate
58 // a DRM session's log away.63 // a DRM session's log away.
59-// The caller must invoke close after all logging is complete.64+// The caller must invoke close after all logging is complete; crashed
60-func Open(ctx context.Context, backend, level, debug string) (context.Context, func() error, error) {65+// keeps the crash file for a panic about to be re-raised.
66+func Open(ctx context.Context, backend, level, debug string) (context.Context, func(crashed bool) error, error) {
61 selected, err := ParseDebug(debug)67 selected, err := ParseDebug(debug)
62 if err != nil {68 if err != nil {
63 return nil, nil, err69 return nil, nil, err
@@ -85,23 +91,88 @@ func Open(ctx context.Context, backend, level, debug string) (context.Context, f
85 _ = file.Close()91 _ = file.Close()
86 return nil, nil, err92 return nil, nil, err
87 }93 }
94+ // Queued entries are lost on a fatal crash: the runtime's own report
95+ // (panic and every stack) goes to its own file, besides stderr. A clean
96+ // Close removes it while still empty. Created before pruneRuns so
97+ // crash files rotate like logs.
98+ crashPath := filepath.Join(dir, strings.TrimSuffix(name, ".log")+".crash")
99+ if crash, err := os.OpenFile(crashPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600); err == nil {
100+ _ = runtimedebug.SetCrashOutput(crash, runtimedebug.CrashOptions{})
101+ _ = crash.Close() // SetCrashOutput keeps its own duplicate
102+ } else {
103+ crashPath = ""
104+ }
88 pruneRuns(dir)105 pruneRuns(dir)
89- var output = zerowrap.Config{Level: level, Format: "json", Output: file}106+ // Hide Close from the diode: asyncWriter closes the file once.
107+ var sink io.Writer = struct{ io.Writer }{file}
90 if term.IsTerminal(int(os.Stderr.Fd())) {108 if term.IsTerminal(int(os.Stderr.Fd())) {
91- output.Output = &consoleAndFile{file: file}109+ sink = &consoleAndFile{file: file}
92 }110 }
93- ctx = zerowrap.WithCtx(ctx, zerowrap.New(output))111+ w := newAsyncWriter(sink, file, crashPath)
112+ ctx = zerowrap.WithCtx(ctx, zerowrap.New(zerowrap.Config{Level: level, Format: "json", Output: w}))
94 ctx = context.WithValue(ctx, debugKey{}, debugSet(selected))113 ctx = context.WithValue(ctx, debugKey{}, debugSet(selected))
95 ctx = context.WithValue(ctx, levelKey{}, level)114 ctx = context.WithValue(ctx, levelKey{}, level)
96- return ctx, file.Close, nil115+ ctx = context.WithValue(ctx, dirKey{}, dir)
116+ return ctx, func(crashed bool) error {
117+ if !crashed {
118+ w.Clean()
119+ }
120+ return w.Close()
121+ }, nil
122+}
123+
124+// logQueue is how many entries may wait for the disk. Past it, entries are
125+// dropped (and counted on stderr) rather than blocking the logging
126+// goroutine: input, outputs and the display must never wait on a slow disk.
127+const logQueue = 8192
128+
129+// newAsyncWriter writes entries to w from its own goroutine. Close drains
130+// what is queued, then closes c.
131+func newAsyncWriter(w io.Writer, c io.Closer, crashPath string) *asyncWriter {
132+ d := diode.NewWriter(w, logQueue, 0, func(missed int) {
133+ fmt.Fprintf(os.Stderr, "neferwl: %d log entries dropped: disk too slow\n", missed)
134+ })
135+ return &asyncWriter{Writer: d, closer: c, crashPath: crashPath}
136+}
137+
138+type asyncWriter struct {
139+ diode.Writer
140+ closer io.Closer
141+ crashPath string
142+ once sync.Once
143+ err error
144+}
145+
146+// Close drains the queue and closes the file once.
147+func (a *asyncWriter) Close() error {
148+ a.once.Do(func() {
149+ _ = a.Writer.Close() // drains the queue
150+ a.err = a.closer.Close()
151+ })
152+ return a.err
153+}
154+
155+// Clean ends a run that did not crash: the crash output is released and
156+// the crash file removed while still empty.
157+func (a *asyncWriter) Clean() {
158+ if a.crashPath == "" {
159+ return
160+ }
161+ _ = runtimedebug.SetCrashOutput(nil, runtimedebug.CrashOptions{})
162+ if fi, err := os.Stat(a.crashPath); err == nil && fi.Size() == 0 {
163+ _ = os.Remove(a.crashPath)
164+ }
97 }165 }
98 166
99 func pruneRuns(dir string) {167 func pruneRuns(dir string) {
100- runs, _ := filepath.Glob(filepath.Join(dir, "2*.log"))168+ // Watchdog stall dumps (stall-<time>-<loop>.txt) rotate like runs.
101- sort.Strings(runs)169+ for _, pattern := range []string{"2*.log", "2*.crash", "stall-*.txt"} {
102- for len(runs) > keepRuns {170+ runs, _ := filepath.Glob(filepath.Join(dir, pattern))
103- _ = os.Remove(runs[0])171+ sort.Strings(runs)
104- runs = runs[1:]172+ for len(runs) > keepRuns {
173+ _ = os.Remove(runs[0])
174+ runs = runs[1:]
175+ }
105 }176 }
106 }177 }
107 178
@@ -113,6 +184,12 @@ func (w *consoleAndFile) Write(p []byte) (int, error) {
113 return w.file.Write(p)184 return w.file.Write(p)
114 }185 }
115 186
187+// Dir is the directory of the run log Open created, or "" without one.
188+func Dir(ctx context.Context) string {
189+ dir, _ := ctx.Value(dirKey{}).(string)
190+ return dir
191+}
192+
116 // Enabled reports whether a debug category was named in --debug.193 // Enabled reports whether a debug category was named in --debug.
117 func Enabled(ctx context.Context, category string) bool {194 func Enabled(ctx context.Context, category string) bool {
118 selected, _ := ctx.Value(debugKey{}).(debugSet)195 selected, _ := ctx.Value(debugKey{}).(debugSet)
M internal/adapters/logging/logging_test.go +34 -3
@@ -4,6 +4,7 @@ import (
44 "bytes"
55 "context"
66 "encoding/json"
7+ "io"
78 "os"
89 "path/filepath"
910 "testing"
@@ -47,7 +48,7 @@ func TestComponentLevelsAndRotation(t *testing.T) {
4748 core.Debug().Msg("visible")
4849 app.Debug().Msg("hidden")
4950 app.Info().Msg("startup")
50- if err := closeLog(); err != nil {
51+ if err := closeLog(false); err != nil {
5152 t.Fatal(err)
5253 }
5354 data, err := os.ReadFile(path)
@@ -72,6 +73,10 @@ func TestComponentLevelsAndRotation(t *testing.T) {
7273 if len(runs) != keepRuns {
7374 t.Fatalf("kept %d runs, want %d", len(runs), keepRuns)
7475 }
76+ // A clean run leaves no empty crash file behind.
77+ if crashes, _ := filepath.Glob(filepath.Join(dir, "2*.crash")); len(crashes) != 0 {
78+ t.Fatalf("empty crash files kept: %v", crashes)
79+ }
7580 }
7681
7782 func splitLines(data []byte) [][]byte {
@@ -94,7 +99,7 @@ func TestBackendsRotateSeparately(t *testing.T) {
9499 if err != nil {
95100 t.Fatal(err)
96101 }
97- if err := closeLog(); err != nil {
102+ if err := closeLog(false); err != nil {
98103 t.Fatal(err)
99104 }
100105 }
@@ -123,7 +128,7 @@ func TestCategoryEnablesComponentDebug(t *testing.T) {
123128 input, core := For(ctx, "input"), For(ctx, "core")
124129 input.Debug().Msg("motion")
125130 core.Debug().Msg("hidden")
126- if err := closeLog(); err != nil {
131+ if err := closeLog(false); err != nil {
127132 t.Fatal(err)
128133 }
129134 data, err := os.ReadFile(filepath.Join(os.Getenv("XDG_STATE_HOME"), "neferwl", "runs", "drm", "latest.log"))
@@ -135,6 +140,32 @@ func TestCategoryEnablesComponentDebug(t *testing.T) {
135140 }
136141 }
137142
143+// A disk that stops answering must not block the goroutine that logs:
144+// input and outputs log from their own loops.
145+func TestLogWriteNeverBlocksOnSlowDisk(t *testing.T) {
146+ r, pw, err := os.Pipe() // nobody reads: writes block once the pipe is full
147+ if err != nil {
148+ t.Fatal(err)
149+ }
150+ defer r.Close()
151+ w := newAsyncWriter(struct{ io.Writer }{pw}, pw, "")
152+ line := bytes.Repeat([]byte{'x'}, 1024)
153+ done := make(chan struct{})
154+ go func() {
155+ defer close(done)
156+ for range 4 * logQueue { // far past the pipe buffer and the queue
157+ _, _ = w.Write(line)
158+ }
159+ }()
160+ select {
161+ case <-done:
162+ case <-time.After(5 * time.Second):
163+ t.Fatal("log write blocked on a stuck disk")
164+ }
165+ r.Close() // unblock the writer so Close can drain and return
166+ _ = w.Close()
167+}
168+
138169 // Narrow categories are on only when named, never through "all".
139170 func TestEnabledCategories(t *testing.T) {
140171 for _, category := range []string{"input-motion", "input-keys", "drm-flip"} {
@@ -4,6 +4,7 @@ import (
4 "bytes"4 "bytes"
5 "context"5 "context"
6 "encoding/json"6 "encoding/json"
7+ "io"
7 "os"8 "os"
8 "path/filepath"9 "path/filepath"
9 "testing"10 "testing"
@@ -47,7 +48,7 @@ func TestComponentLevelsAndRotation(t *testing.T) {
47 core.Debug().Msg("visible")48 core.Debug().Msg("visible")
48 app.Debug().Msg("hidden")49 app.Debug().Msg("hidden")
49 app.Info().Msg("startup")50 app.Info().Msg("startup")
50- if err := closeLog(); err != nil {51+ if err := closeLog(false); err != nil {
51 t.Fatal(err)52 t.Fatal(err)
52 }53 }
53 data, err := os.ReadFile(path)54 data, err := os.ReadFile(path)
@@ -72,6 +73,10 @@ func TestComponentLevelsAndRotation(t *testing.T) {
72 if len(runs) != keepRuns {73 if len(runs) != keepRuns {
73 t.Fatalf("kept %d runs, want %d", len(runs), keepRuns)74 t.Fatalf("kept %d runs, want %d", len(runs), keepRuns)
74 }75 }
76+ // A clean run leaves no empty crash file behind.
77+ if crashes, _ := filepath.Glob(filepath.Join(dir, "2*.crash")); len(crashes) != 0 {
78+ t.Fatalf("empty crash files kept: %v", crashes)
79+ }
75 }80 }
76 81
77 func splitLines(data []byte) [][]byte {82 func splitLines(data []byte) [][]byte {
@@ -94,7 +99,7 @@ func TestBackendsRotateSeparately(t *testing.T) {
94 if err != nil {99 if err != nil {
95 t.Fatal(err)100 t.Fatal(err)
96 }101 }
97- if err := closeLog(); err != nil {102+ if err := closeLog(false); err != nil {
98 t.Fatal(err)103 t.Fatal(err)
99 }104 }
100 }105 }
@@ -123,7 +128,7 @@ func TestCategoryEnablesComponentDebug(t *testing.T) {
123 input, core := For(ctx, "input"), For(ctx, "core")128 input, core := For(ctx, "input"), For(ctx, "core")
124 input.Debug().Msg("motion")129 input.Debug().Msg("motion")
125 core.Debug().Msg("hidden")130 core.Debug().Msg("hidden")
126- if err := closeLog(); err != nil {131+ if err := closeLog(false); err != nil {
127 t.Fatal(err)132 t.Fatal(err)
128 }133 }
129 data, err := os.ReadFile(filepath.Join(os.Getenv("XDG_STATE_HOME"), "neferwl", "runs", "drm", "latest.log"))134 data, err := os.ReadFile(filepath.Join(os.Getenv("XDG_STATE_HOME"), "neferwl", "runs", "drm", "latest.log"))
@@ -135,6 +140,32 @@ func TestCategoryEnablesComponentDebug(t *testing.T) {
135 }140 }
136 }141 }
137 142
143+// A disk that stops answering must not block the goroutine that logs:
144+// input and outputs log from their own loops.
145+func TestLogWriteNeverBlocksOnSlowDisk(t *testing.T) {
146+ r, pw, err := os.Pipe() // nobody reads: writes block once the pipe is full
147+ if err != nil {
148+ t.Fatal(err)
149+ }
150+ defer r.Close()
151+ w := newAsyncWriter(struct{ io.Writer }{pw}, pw, "")
152+ line := bytes.Repeat([]byte{'x'}, 1024)
153+ done := make(chan struct{})
154+ go func() {
155+ defer close(done)
156+ for range 4 * logQueue { // far past the pipe buffer and the queue
157+ _, _ = w.Write(line)
158+ }
159+ }()
160+ select {
161+ case <-done:
162+ case <-time.After(5 * time.Second):
163+ t.Fatal("log write blocked on a stuck disk")
164+ }
165+ r.Close() // unblock the writer so Close can drain and return
166+ _ = w.Close()
167+}
168+
138 // Narrow categories are on only when named, never through "all".169 // Narrow categories are on only when named, never through "all".
139 func TestEnabledCategories(t *testing.T) {170 func TestEnabledCategories(t *testing.T) {
140 for _, category := range []string{"input-motion", "input-keys", "drm-flip"} {171 for _, category := range []string{"input-motion", "input-keys", "drm-flip"} {
M internal/adapters/vulkan/damage.go +49 -9
@@ -27,6 +27,8 @@ type damageRegion struct {
2727 drawn map[ports.WindowID]heldWindow
2828 // rects is scratch for a content's damage rects (reused across frames).
2929 rects []ports.Rect
30+ // why counts the reasons of full redraws (the renderer's; nil: none).
31+ why *fullRedraws
3032 }
3133
3234 // heldWindow is a window a target holds: its content Seq and rect.
@@ -46,6 +48,7 @@ func newDamage(tg *target, s ports.Scene, bounds image.Rectangle) *damageRegion
4648 // one keeps all but what sceneDelta finds changed.
4749 func (d *damageRegion) hold(tg *target, s ports.Scene) {
4850 if !tg.valid || s.Seq == 0 {
51+ d.why.count(fullTarget)
4952 return
5053 }
5154 if tg.sceneSeq == s.Seq {
@@ -64,6 +67,8 @@ func (d *damageRegion) hold(tg *target, s ports.Scene) {
6467 limit := d.bound.Dx() * d.bound.Dy() * 9 / 10
6568 if area, ok := sceneDelta(tg.scene, s, phys, limit); ok {
6669 d.held, d.area = tg, area
70+ } else {
71+ d.why.count(fullScene)
6772 }
6873 }
6974
@@ -78,7 +83,7 @@ func (r *Renderer) frameDamage(tg *target, s ports.Scene, bounds image.Rectangle
7883 r.damageDrawn = make(map[ports.WindowID]heldWindow)
7984 }
8085 clear(r.damageDrawn)
81- r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds, rects: r.damage.rects[:0]}
86+ r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds, rects: r.damage.rects[:0], why: &r.fullRedraws}
8287 r.damage.hold(tg, s)
8388 return &r.damage
8489 }
@@ -108,16 +113,51 @@ func (d *damageRegion) window(id ports.WindowID, c ports.SurfaceContent, rect im
108113 if ok && held == c.Seq {
109114 return
110115 }
111- if !ok || c.Empty() || len(c.Children) > 0 {
112- d.add(rect)
113- return
116+ switch {
117+ case !ok:
118+ d.why.count(fullNewWindow)
119+ case c.Empty():
120+ d.why.count(fullEmpty)
121+ case len(c.Children) > 0:
122+ d.why.count(fullChildren)
123+ default:
124+ var known bool
125+ d.rects, known = c.AppendDamageSince(d.rects[:0], held)
126+ if known {
127+ return // mapped by content, called next
128+ }
129+ d.why.count(fullHistory)
114130 }
115- var known bool
116- d.rects, known = c.AppendDamageSince(d.rects[:0], held)
117- if !known {
118- d.add(rect)
131+ d.add(rect)
132+}
133+
134+// fullReason is why a whole target or window rect is redrawn.
135+type fullReason int
136+
137+const (
138+ fullTarget fullReason = iota // target never drawn, or an unnumbered scene
139+ fullScene // the scene changed past sceneDelta
140+ fullNewWindow // the target never held the window
141+ fullEmpty // the window has no content
142+ fullChildren // the window has subsurfaces
143+ fullHistory // the damage history does not reach the held content
144+ fullReasons
145+)
146+
147+// fullRedraws counts full redraws by reason since the last take.
148+type fullRedraws [fullReasons]int
149+
150+func (f *fullRedraws) count(r fullReason) {
151+ if f != nil {
152+ f[r]++
119153 }
120- // Otherwise mapped by content, called next: nothing more here.
154+}
155+
156+// take returns the counts and resets them.
157+func (f *fullRedraws) take() fullRedraws {
158+ n := *f
159+ *f = fullRedraws{}
160+ return n
121161 }
122162
123163 // content maps a root surface's buffer damage onto full (where its w×h
@@ -27,6 +27,8 @@ type damageRegion struct {
27 drawn map[ports.WindowID]heldWindow27 drawn map[ports.WindowID]heldWindow
28 // rects is scratch for a content's damage rects (reused across frames).28 // rects is scratch for a content's damage rects (reused across frames).
29 rects []ports.Rect29 rects []ports.Rect
30+ // why counts the reasons of full redraws (the renderer's; nil: none).
31+ why *fullRedraws
30 }32 }
31 33
32 // heldWindow is a window a target holds: its content Seq and rect.34 // heldWindow is a window a target holds: its content Seq and rect.
@@ -46,6 +48,7 @@ func newDamage(tg *target, s ports.Scene, bounds image.Rectangle) *damageRegion
46 // one keeps all but what sceneDelta finds changed.48 // one keeps all but what sceneDelta finds changed.
47 func (d *damageRegion) hold(tg *target, s ports.Scene) {49 func (d *damageRegion) hold(tg *target, s ports.Scene) {
48 if !tg.valid || s.Seq == 0 {50 if !tg.valid || s.Seq == 0 {
51+ d.why.count(fullTarget)
49 return52 return
50 }53 }
51 if tg.sceneSeq == s.Seq {54 if tg.sceneSeq == s.Seq {
@@ -64,6 +67,8 @@ func (d *damageRegion) hold(tg *target, s ports.Scene) {
64 limit := d.bound.Dx() * d.bound.Dy() * 9 / 1067 limit := d.bound.Dx() * d.bound.Dy() * 9 / 10
65 if area, ok := sceneDelta(tg.scene, s, phys, limit); ok {68 if area, ok := sceneDelta(tg.scene, s, phys, limit); ok {
66 d.held, d.area = tg, area69 d.held, d.area = tg, area
70+ } else {
71+ d.why.count(fullScene)
67 }72 }
68 }73 }
69 74
@@ -78,7 +83,7 @@ func (r *Renderer) frameDamage(tg *target, s ports.Scene, bounds image.Rectangle
78 r.damageDrawn = make(map[ports.WindowID]heldWindow)83 r.damageDrawn = make(map[ports.WindowID]heldWindow)
79 }84 }
80 clear(r.damageDrawn)85 clear(r.damageDrawn)
81- r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds, rects: r.damage.rects[:0]}86+ r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds, rects: r.damage.rects[:0], why: &r.fullRedraws}
82 r.damage.hold(tg, s)87 r.damage.hold(tg, s)
83 return &r.damage88 return &r.damage
84 }89 }
@@ -108,16 +113,51 @@ func (d *damageRegion) window(id ports.WindowID, c ports.SurfaceContent, rect im
108 if ok && held == c.Seq {113 if ok && held == c.Seq {
109 return114 return
110 }115 }
111- if !ok || c.Empty() || len(c.Children) > 0 {116+ switch {
112- d.add(rect)117+ case !ok:
113- return118+ d.why.count(fullNewWindow)
119+ case c.Empty():
120+ d.why.count(fullEmpty)
121+ case len(c.Children) > 0:
122+ d.why.count(fullChildren)
123+ default:
124+ var known bool
125+ d.rects, known = c.AppendDamageSince(d.rects[:0], held)
126+ if known {
127+ return // mapped by content, called next
128+ }
129+ d.why.count(fullHistory)
114 }130 }
115- var known bool131+ d.add(rect)
116- d.rects, known = c.AppendDamageSince(d.rects[:0], held)132+}
117- if !known {133+
118- d.add(rect)134+// fullReason is why a whole target or window rect is redrawn.
135+type fullReason int
136+
137+const (
138+ fullTarget fullReason = iota // target never drawn, or an unnumbered scene
139+ fullScene // the scene changed past sceneDelta
140+ fullNewWindow // the target never held the window
141+ fullEmpty // the window has no content
142+ fullChildren // the window has subsurfaces
143+ fullHistory // the damage history does not reach the held content
144+ fullReasons
145+)
146+
147+// fullRedraws counts full redraws by reason since the last take.
148+type fullRedraws [fullReasons]int
149+
150+func (f *fullRedraws) count(r fullReason) {
151+ if f != nil {
152+ f[r]++
119 }153 }
120- // Otherwise mapped by content, called next: nothing more here.154+}
155+
156+// take returns the counts and resets them.
157+func (f *fullRedraws) take() fullRedraws {
158+ n := *f
159+ *f = fullRedraws{}
160+ return n
121 }161 }
122 162
123 // content maps a root surface's buffer damage onto full (where its w×h163 // content maps a root surface's buffer damage onto full (where its w×h
A internal/adapters/vulkan/fullredraw_test.go +39 -0
@@ -0,0 +1,39 @@
1+package vulkan
2+
3+import (
4+ "image"
5+ "testing"
6+
7+ "github.com/bnema/neferwl/internal/ports"
8+)
9+
10+// Each whole-window redraw is counted under its reason, so the stats tell
11+// a client claiming full damage from a history too short.
12+func TestFullRedrawReasons(t *testing.T) {
13+ rect := image.Rect(0, 0, 10, 10)
14+ held := &target{valid: true, windows: map[ports.WindowID]heldWindow{1: {seq: 1, rect: rect}}}
15+ var why fullRedraws
16+ d := &damageRegion{held: held, seen: map[ports.WindowID]bool{}, drawn: map[ports.WindowID]heldWindow{}, bound: rect, why: &why}
17+ pixels := ports.SurfaceContent{Width: 10, Height: 10, SHM: &ports.SHMBuffer{}}
18+
19+ known := pixels
20+ known.ID, known.Seq = 1, 2
21+ known.DamageHistory = []ports.SeqDamage{{Seq: 2, Rects: []ports.Rect{{W: 1, H: 1}}}}
22+ d.window(1, known, rect)
23+ if why != (fullRedraws{}) {
24+ t.Fatalf("partial damage counted as full: %v", why)
25+ }
26+
27+ d.seen = map[ports.WindowID]bool{}
28+ gap := pixels
29+ gap.ID, gap.Seq = 1, 9
30+ gap.DamageHistory = []ports.SeqDamage{{Seq: 9, Rects: []ports.Rect{{W: 1, H: 1}}}}
31+ d.window(1, gap, rect)
32+ d.window(2, gap, rect)
33+ if why[fullHistory] != 1 || why[fullNewWindow] != 1 {
34+ t.Fatalf("reasons %v", why)
35+ }
36+ if got := why.take(); got[fullHistory] != 1 || why != (fullRedraws{}) {
37+ t.Fatalf("take %v left %v", got, why)
38+ }
39+}
@@ -0,0 +1,39 @@
1+package vulkan
2+
3+import (
4+ "image"
5+ "testing"
6+
7+ "github.com/bnema/neferwl/internal/ports"
8+)
9+
10+// Each whole-window redraw is counted under its reason, so the stats tell
11+// a client claiming full damage from a history too short.
12+func TestFullRedrawReasons(t *testing.T) {
13+ rect := image.Rect(0, 0, 10, 10)
14+ held := &target{valid: true, windows: map[ports.WindowID]heldWindow{1: {seq: 1, rect: rect}}}
15+ var why fullRedraws
16+ d := &damageRegion{held: held, seen: map[ports.WindowID]bool{}, drawn: map[ports.WindowID]heldWindow{}, bound: rect, why: &why}
17+ pixels := ports.SurfaceContent{Width: 10, Height: 10, SHM: &ports.SHMBuffer{}}
18+
19+ known := pixels
20+ known.ID, known.Seq = 1, 2
21+ known.DamageHistory = []ports.SeqDamage{{Seq: 2, Rects: []ports.Rect{{W: 1, H: 1}}}}
22+ d.window(1, known, rect)
23+ if why != (fullRedraws{}) {
24+ t.Fatalf("partial damage counted as full: %v", why)
25+ }
26+
27+ d.seen = map[ports.WindowID]bool{}
28+ gap := pixels
29+ gap.ID, gap.Seq = 1, 9
30+ gap.DamageHistory = []ports.SeqDamage{{Seq: 9, Rects: []ports.Rect{{W: 1, H: 1}}}}
31+ d.window(1, gap, rect)
32+ d.window(2, gap, rect)
33+ if why[fullHistory] != 1 || why[fullNewWindow] != 1 {
34+ t.Fatalf("reasons %v", why)
35+ }
36+ if got := why.take(); got[fullHistory] != 1 || why != (fullRedraws{}) {
37+ t.Fatalf("take %v left %v", got, why)
38+ }
39+}
M internal/adapters/vulkan/renderer.go +7 -0
@@ -89,6 +89,9 @@ type Renderer struct {
8989 marks []trimMark
9090 // redrawn counts the target pixels drawn (TakeRedrawn).
9191 redrawn int
92+ // fullRedraws counts why whole targets or windows were redrawn; each
93+ // TakeRedrawn logs and resets it.
94+ fullRedraws fullRedraws
9295 // cursors are the exported cursor images (CursorBuffers).
9396 cursors [2]*cursorImage
9497 // compose draws every frame (compose.go); maxRange bounds one
@@ -345,6 +348,10 @@ func (r *Renderer) CopiedBytes() int { return r.copied }
345348 func (r *Renderer) TakeRedrawn() int {
346349 n := r.redrawn
347350 r.redrawn = 0
351+ if f := r.fullRedraws.take(); f != (fullRedraws{}) {
352+ r.log.Debug().Int("target", f[fullTarget]).Int("scene", f[fullScene]).Int("new_window", f[fullNewWindow]).
353+ Int("empty", f[fullEmpty]).Int("children", f[fullChildren]).Int("history", f[fullHistory]).Msg("full redraws")
354+ }
348355 return n
349356 }
350357
@@ -89,6 +89,9 @@ type Renderer struct {
89 marks []trimMark89 marks []trimMark
90 // redrawn counts the target pixels drawn (TakeRedrawn).90 // redrawn counts the target pixels drawn (TakeRedrawn).
91 redrawn int91 redrawn int
92+ // fullRedraws counts why whole targets or windows were redrawn; each
93+ // TakeRedrawn logs and resets it.
94+ fullRedraws fullRedraws
92 // cursors are the exported cursor images (CursorBuffers).95 // cursors are the exported cursor images (CursorBuffers).
93 cursors [2]*cursorImage96 cursors [2]*cursorImage
94 // compose draws every frame (compose.go); maxRange bounds one97 // compose draws every frame (compose.go); maxRange bounds one
@@ -345,6 +348,10 @@ func (r *Renderer) CopiedBytes() int { return r.copied }
345 func (r *Renderer) TakeRedrawn() int {348 func (r *Renderer) TakeRedrawn() int {
346 n := r.redrawn349 n := r.redrawn
347 r.redrawn = 0350 r.redrawn = 0
351+ if f := r.fullRedraws.take(); f != (fullRedraws{}) {
352+ r.log.Debug().Int("target", f[fullTarget]).Int("scene", f[fullScene]).Int("new_window", f[fullNewWindow]).
353+ Int("empty", f[fullEmpty]).Int("children", f[fullChildren]).Int("history", f[fullHistory]).Msg("full redraws")
354+ }
348 return n355 return n
349 }356 }
350 357
A internal/adapters/watchdog/watchdog.go +117 -0
@@ -0,0 +1,117 @@
1+// Package watchdog writes every goroutine stack to a file when a probed
2+// owner loop stops answering, so a frozen session leaves the exact place
3+// it is stuck on disk.
4+package watchdog
5+
6+import (
7+ "context"
8+ "fmt"
9+ "os"
10+ "path/filepath"
11+ "runtime/pprof"
12+ "sync"
13+ "time"
14+
15+ "github.com/bnema/zerowrap"
16+)
17+
18+// Probe asks an owner loop to run an empty task: Ping returns once the
19+// loop ran it. A loop that is stuck never returns.
20+type Probe struct {
21+ Name string
22+ Ping func()
23+}
24+
25+// Options set how often the probes run and how long a loop may take.
26+type Options struct {
27+ Dir string // where stall dumps go (the run log directory)
28+ Every time.Duration // between two probes of a loop
29+ Timeout time.Duration // a probe slower than this is a stall
30+ Log zerowrap.Logger
31+}
32+
33+// Run probes each loop probes returns until ctx ends; the set may change
34+// between ticks (outputs come and go). On a stall it writes one dump per
35+// stall: a loop must answer again before it is dumped again. Every Ping
36+// must return once ctx ends: Run waits for the probes in flight.
37+func Run(ctx context.Context, probes func() []Probe, o Options) {
38+ var inflight sync.WaitGroup
39+ defer inflight.Wait()
40+ type state struct {
41+ answer chan struct{} // the probe in flight, nil when none
42+ since time.Time
43+ stalled bool
44+ seen bool
45+ }
46+ states := map[string]*state{}
47+ tick := time.NewTicker(o.Every)
48+ defer tick.Stop()
49+ for {
50+ select {
51+ case <-ctx.Done():
52+ return
53+ case now := <-tick.C:
54+ for _, s := range states {
55+ s.seen = false
56+ }
57+ for _, p := range probes() {
58+ s := states[p.Name]
59+ if s == nil {
60+ s = &state{}
61+ states[p.Name] = s
62+ }
63+ s.seen = true
64+ if s.answer != nil {
65+ select {
66+ case <-s.answer:
67+ if s.stalled {
68+ o.Log.Warn().Str("loop", p.Name).Dur("after", now.Sub(s.since)).Msg("loop answers again")
69+ }
70+ s.answer, s.stalled = nil, false
71+ default:
72+ if !s.stalled && now.Sub(s.since) >= o.Timeout {
73+ s.stalled = true
74+ path, err := dump(o.Dir, p.Name, now)
75+ o.Log.Error().Err(err).Str("loop", p.Name).Dur("waited", now.Sub(s.since)).Str("dump", path).Msg("loop stalled")
76+ }
77+ continue // one probe in flight per loop
78+ }
79+ }
80+ answer := make(chan struct{})
81+ s.answer, s.since = answer, now
82+ inflight.Go(func() { p.Ping(); close(answer) })
83+ }
84+ // A loop that left (an unplugged output) is forgotten; its
85+ // probe goroutine ends with the loop's context.
86+ for name, s := range states {
87+ if !s.seen {
88+ delete(states, name)
89+ }
90+ }
91+ }
92+ }
93+}
94+
95+// dump writes every goroutine stack to dir and syncs it: a frozen machine
96+// is often force rebooted, which loses what the page cache still holds.
97+func dump(dir, name string, now time.Time) (string, error) {
98+ path := filepath.Join(dir, fmt.Sprintf("stall-%s-%s.txt", now.Format("20060102-150405.000"), name))
99+ f, err := os.OpenFile(path, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600)
100+ if err != nil {
101+ return path, err
102+ }
103+ fmt.Fprintf(f, "loop %q stopped answering at %s\n\n", name, now.Format(time.RFC3339Nano))
104+ werr := pprof.Lookup("goroutine").WriteTo(f, 2)
105+ serr := f.Sync()
106+ cerr := f.Close()
107+ if d, err := os.Open(dir); err == nil {
108+ _ = d.Sync()
109+ _ = d.Close()
110+ }
111+ for _, err := range []error{werr, serr, cerr} {
112+ if err != nil {
113+ return path, err
114+ }
115+ }
116+ return path, nil
117+}
@@ -0,0 +1,117 @@
1+// Package watchdog writes every goroutine stack to a file when a probed
2+// owner loop stops answering, so a frozen session leaves the exact place
3+// it is stuck on disk.
4+package watchdog
5+
6+import (
7+ "context"
8+ "fmt"
9+ "os"
10+ "path/filepath"
11+ "runtime/pprof"
12+ "sync"
13+ "time"
14+
15+ "github.com/bnema/zerowrap"
16+)
17+
18+// Probe asks an owner loop to run an empty task: Ping returns once the
19+// loop ran it. A loop that is stuck never returns.
20+type Probe struct {
21+ Name string
22+ Ping func()
23+}
24+
25+// Options set how often the probes run and how long a loop may take.
26+type Options struct {
27+ Dir string // where stall dumps go (the run log directory)
28+ Every time.Duration // between two probes of a loop
29+ Timeout time.Duration // a probe slower than this is a stall
30+ Log zerowrap.Logger
31+}
32+
33+// Run probes each loop probes returns until ctx ends; the set may change
34+// between ticks (outputs come and go). On a stall it writes one dump per
35+// stall: a loop must answer again before it is dumped again. Every Ping
36+// must return once ctx ends: Run waits for the probes in flight.
37+func Run(ctx context.Context, probes func() []Probe, o Options) {
38+ var inflight sync.WaitGroup
39+ defer inflight.Wait()
40+ type state struct {
41+ answer chan struct{} // the probe in flight, nil when none
42+ since time.Time
43+ stalled bool
44+ seen bool
45+ }
46+ states := map[string]*state{}
47+ tick := time.NewTicker(o.Every)
48+ defer tick.Stop()
49+ for {
50+ select {
51+ case <-ctx.Done():
52+ return
53+ case now := <-tick.C:
54+ for _, s := range states {
55+ s.seen = false
56+ }
57+ for _, p := range probes() {
58+ s := states[p.Name]
59+ if s == nil {
60+ s = &state{}
61+ states[p.Name] = s
62+ }
63+ s.seen = true
64+ if s.answer != nil {
65+ select {
66+ case <-s.answer:
67+ if s.stalled {
68+ o.Log.Warn().Str("loop", p.Name).Dur("after", now.Sub(s.since)).Msg("loop answers again")
69+ }
70+ s.answer, s.stalled = nil, false
71+ default:
72+ if !s.stalled && now.Sub(s.since) >= o.Timeout {
73+ s.stalled = true
74+ path, err := dump(o.Dir, p.Name, now)
75+ o.Log.Error().Err(err).Str("loop", p.Name).Dur("waited", now.Sub(s.since)).Str("dump", path).Msg("loop stalled")
76+ }
77+ continue // one probe in flight per loop
78+ }
79+ }
80+ answer := make(chan struct{})
81+ s.answer, s.since = answer, now
82+ inflight.Go(func() { p.Ping(); close(answer) })
83+ }
84+ // A loop that left (an unplugged output) is forgotten; its
85+ // probe goroutine ends with the loop's context.
86+ for name, s := range states {
87+ if !s.seen {
88+ delete(states, name)
89+ }
90+ }
91+ }
92+ }
93+}
94+
95+// dump writes every goroutine stack to dir and syncs it: a frozen machine
96+// is often force rebooted, which loses what the page cache still holds.
97+func dump(dir, name string, now time.Time) (string, error) {
98+ path := filepath.Join(dir, fmt.Sprintf("stall-%s-%s.txt", now.Format("20060102-150405.000"), name))
99+ f, err := os.OpenFile(path, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600)
100+ if err != nil {
101+ return path, err
102+ }
103+ fmt.Fprintf(f, "loop %q stopped answering at %s\n\n", name, now.Format(time.RFC3339Nano))
104+ werr := pprof.Lookup("goroutine").WriteTo(f, 2)
105+ serr := f.Sync()
106+ cerr := f.Close()
107+ if d, err := os.Open(dir); err == nil {
108+ _ = d.Sync()
109+ _ = d.Close()
110+ }
111+ for _, err := range []error{werr, serr, cerr} {
112+ if err != nil {
113+ return path, err
114+ }
115+ }
116+ return path, nil
117+}
A internal/adapters/watchdog/watchdog_test.go +77 -0
@@ -0,0 +1,77 @@
1+package watchdog
2+
3+import (
4+ "context"
5+ "os"
6+ "path/filepath"
7+ "strings"
8+ "sync/atomic"
9+ "testing"
10+ "testing/synctest"
11+ "time"
12+
13+ "github.com/bnema/zerowrap"
14+)
15+
16+// stuckOwnerLoop blocks like an owner goroutine that never reaches its
17+// select again; the dump must name it.
18+func stuckOwnerLoop(ctx context.Context, release <-chan struct{}) {
19+ select {
20+ case <-release:
21+ case <-ctx.Done():
22+ }
23+}
24+
25+func TestStalledLoopIsDumpedOnce(t *testing.T) {
26+ dir := t.TempDir()
27+ synctest.Test(t, func(t *testing.T) {
28+ ctx, cancel := context.WithCancel(context.Background())
29+ release := make(chan struct{})
30+ var answered atomic.Int32
31+ probes := func() []Probe {
32+ return []Probe{
33+ {Name: "stuck", Ping: func() { stuckOwnerLoop(ctx, release) }},
34+ {Name: "fine", Ping: func() { answered.Add(1) }},
35+ }
36+ }
37+ done := make(chan struct{})
38+ go func() {
39+ defer close(done)
40+ Run(ctx, probes, Options{Dir: dir, Every: 10 * time.Millisecond, Timeout: 50 * time.Millisecond, Log: zerowrap.Default()})
41+ }()
42+ dumps := func() []string {
43+ m, _ := filepath.Glob(filepath.Join(dir, "stall-*.txt"))
44+ return m
45+ }
46+ time.Sleep(40 * time.Millisecond)
47+ synctest.Wait()
48+ if d := dumps(); len(d) != 0 {
49+ t.Fatalf("dumped before the timeout: %v", d)
50+ }
51+ time.Sleep(30 * time.Millisecond)
52+ synctest.Wait()
53+ d := dumps()
54+ if len(d) != 1 || !strings.HasSuffix(d[0], "-stuck.txt") {
55+ t.Fatalf("dumps %v", d)
56+ }
57+ // Still stuck: no second dump.
58+ time.Sleep(time.Second)
59+ synctest.Wait()
60+ if again := dumps(); len(again) != 1 {
61+ t.Fatalf("dumps after the first %v", again)
62+ }
63+ data, err := os.ReadFile(d[0])
64+ if err != nil {
65+ t.Fatal(err)
66+ }
67+ if !strings.Contains(string(data), "stuckOwnerLoop") {
68+ t.Fatalf("dump does not name the stuck function:\n%s", data)
69+ }
70+ if n := answered.Load(); n < 50 {
71+ t.Fatalf("a healthy loop was probed %d times in 1 s", n)
72+ }
73+ // Run waits for its probes: a stuck one ends with ctx.
74+ cancel()
75+ <-done
76+ })
77+}
@@ -0,0 +1,77 @@
1+package watchdog
2+
3+import (
4+ "context"
5+ "os"
6+ "path/filepath"
7+ "strings"
8+ "sync/atomic"
9+ "testing"
10+ "testing/synctest"
11+ "time"
12+
13+ "github.com/bnema/zerowrap"
14+)
15+
16+// stuckOwnerLoop blocks like an owner goroutine that never reaches its
17+// select again; the dump must name it.
18+func stuckOwnerLoop(ctx context.Context, release <-chan struct{}) {
19+ select {
20+ case <-release:
21+ case <-ctx.Done():
22+ }
23+}
24+
25+func TestStalledLoopIsDumpedOnce(t *testing.T) {
26+ dir := t.TempDir()
27+ synctest.Test(t, func(t *testing.T) {
28+ ctx, cancel := context.WithCancel(context.Background())
29+ release := make(chan struct{})
30+ var answered atomic.Int32
31+ probes := func() []Probe {
32+ return []Probe{
33+ {Name: "stuck", Ping: func() { stuckOwnerLoop(ctx, release) }},
34+ {Name: "fine", Ping: func() { answered.Add(1) }},
35+ }
36+ }
37+ done := make(chan struct{})
38+ go func() {
39+ defer close(done)
40+ Run(ctx, probes, Options{Dir: dir, Every: 10 * time.Millisecond, Timeout: 50 * time.Millisecond, Log: zerowrap.Default()})
41+ }()
42+ dumps := func() []string {
43+ m, _ := filepath.Glob(filepath.Join(dir, "stall-*.txt"))
44+ return m
45+ }
46+ time.Sleep(40 * time.Millisecond)
47+ synctest.Wait()
48+ if d := dumps(); len(d) != 0 {
49+ t.Fatalf("dumped before the timeout: %v", d)
50+ }
51+ time.Sleep(30 * time.Millisecond)
52+ synctest.Wait()
53+ d := dumps()
54+ if len(d) != 1 || !strings.HasSuffix(d[0], "-stuck.txt") {
55+ t.Fatalf("dumps %v", d)
56+ }
57+ // Still stuck: no second dump.
58+ time.Sleep(time.Second)
59+ synctest.Wait()
60+ if again := dumps(); len(again) != 1 {
61+ t.Fatalf("dumps after the first %v", again)
62+ }
63+ data, err := os.ReadFile(d[0])
64+ if err != nil {
65+ t.Fatal(err)
66+ }
67+ if !strings.Contains(string(data), "stuckOwnerLoop") {
68+ t.Fatalf("dump does not name the stuck function:\n%s", data)
69+ }
70+ if n := answered.Load(); n < 50 {
71+ t.Fatalf("a healthy loop was probed %d times in 1 s", n)
72+ }
73+ // Run waits for its probes: a stuck one ends with ctx.
74+ cancel()
75+ <-done
76+ })
77+}
A internal/adapters/wayland/alloc_hotpath_test.go +153 -0
@@ -0,0 +1,153 @@
1+package wayland
2+
3+import (
4+ "context"
5+ "testing"
6+ "time"
7+
8+ "github.com/bnema/neferwl/internal/ports"
9+ "github.com/stretchr/testify/mock"
10+)
11+
12+// The steady-state hot paths that run for every frame or commit allocate
13+// nothing (or one pooled object per damage-history generation).
14+
15+func TestReleaseHeldAllocations(t *testing.T) {
16+ out := &output{place: ports.OutputPlacement{Info: ports.OutputInfo{Name: "DP-2", RefreshMilli: 60000}}}
17+ s := &Server{outputs: []*output{out}, reports: map[string]ports.OutputPresented{}}
18+ reports := []ports.OutputPresented{{Output: "DP-2", Seen: map[ports.WindowID]uint64{1: 1}}}
19+ now := time.Unix(100, 0)
20+ s.releaseHeld(now, reports)
21+ if allocs := testing.AllocsPerRun(100, func() { s.releaseHeld(now, reports) }); allocs != 0 {
22+ t.Errorf("releaseHeld: %.1f allocs, want 0", allocs)
23+ }
24+}
25+
26+func TestCommitDamageAllocations(t *testing.T) {
27+ srv := &Server{fifoSurfaces: map[*surface]struct{}{}, frameReady: make(chan struct{}, 1)}
28+ surf := &surface{server: srv}
29+ commit := func() {
30+ surf.Damage(nil, 0, 0, 4, 4)
31+ surf.DamageBuffer(nil, 1, 1, 2, 2)
32+ surf.queueUpdate()
33+ surf.queue[0].applyGraph()
34+ }
35+ commit()
36+ if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
37+ t.Errorf("damage requests + queueUpdate + apply: %.1f allocs, want 0", allocs)
38+ }
39+ // The queued update keeps its rects while the next commit collects its own.
40+ surf.Damage(nil, 0, 0, 1, 1)
41+ surf.queueUpdate()
42+ first := surf.queue[0]
43+ surf.Damage(nil, 5, 5, 2, 2)
44+ if len(first.damage) != 1 || first.damage[0].W != 1 || len(surf.next.damage) != 1 || surf.next.damage[0].W != 2 {
45+ t.Fatalf("damage shared between commits: queued %v pending %v", first.damage, surf.next.damage)
46+ }
47+ surf.queueUpdate()
48+ if second := surf.queue[1]; len(second.damage) != 1 || second.damage[0].W != 2 || first.damage[0].W != 1 {
49+ t.Fatalf("damage mixed up: %v %v", first.damage, second.damage)
50+ }
51+}
52+
53+// An explicit-sync commit moves its points into the update: the commit
54+// itself allocates no commitSync, and a pending one is not overwritten.
55+func TestCommitSyncStorageIsPerUpdate(t *testing.T) {
56+ dev := newMocksyncobjDevice(t)
57+ dev.EXPECT().signalled(mock.Anything, mock.Anything).Return(true, nil)
58+ srv := &Server{fifoSurfaces: map[*surface]struct{}{}, frameReady: make(chan struct{}, 1)}
59+ sw, err := newSyncWaiter(func() {})
60+ if err != nil {
61+ t.Fatal(err)
62+ }
63+ defer sw.pipeR.Close()
64+ defer sw.close()
65+ srv.syncWait = sw
66+ tl := &timeline{dev: dev, handle: 7, alive: true}
67+ surf := &surface{server: srv, sync: &syncState{}}
68+ take := func(point uint64) {
69+ surf.sync.acquire, surf.sync.release = syncPoint{tl: tl, point: point}, syncPoint{tl: tl, point: point + 1}
70+ surf.takeSyncPoints()
71+ surf.queueUpdate()
72+ }
73+ take(10)
74+ take(20)
75+ a, b := surf.queue[0], surf.queue[1]
76+ if a.sync == nil || b.sync == nil || a.sync == b.sync || a.sync.acquire.point != 10 || b.sync.acquire.point != 20 || surf.next.sync != nil {
77+ t.Fatalf("sync points mixed up: %+v %+v pending %+v", a.sync, b.sync, surf.next.sync)
78+ }
79+ // The commit's points live in the surface until queueUpdate: no heap
80+ // commitSync per commit (the mocked device's own allocations would drown
81+ // an AllocsPerRun count here).
82+ surf.sync.acquire, surf.sync.release = syncPoint{tl: tl, point: 30}, syncPoint{tl: tl, point: 31}
83+ surf.takeSyncPoints()
84+ if surf.next.sync != &surf.syncScratch {
85+ t.Fatal("commit points not held in the surface scratch")
86+ }
87+}
88+
89+func TestEmitContentAllocations(t *testing.T) {
90+ s := &Server{
91+ channels: Channels{Contents: make(chan ports.SurfaceContent)},
92+ contents: map[ports.WindowID]ports.SurfaceContent{},
93+ contentSeq: map[ports.WindowID]uint64{},
94+ damage: map[ports.WindowID][]ports.SeqDamage{},
95+ contentReady: make(chan struct{}, 1),
96+ }
97+ c := ports.SurfaceContent{ID: 1, Solid: &ports.SolidColor{}}
98+ var held []ports.SurfaceContent // readers retaining published histories
99+ var snapshots [][]ports.SeqDamage
100+ for range 3 * damageHistoryCap {
101+ s.emitContent(c, damage{full: true})
102+ got := s.contents[1]
103+ held = append(held, got)
104+ snapshots = append(snapshots, append([]ports.SeqDamage(nil), got.DamageHistory...))
105+ }
106+ for i, h := range held {
107+ // A published history never changes, nor can its owner append to it.
108+ if len(h.DamageHistory) != len(snapshots[i]) || cap(h.DamageHistory) != len(h.DamageHistory) {
109+ t.Fatalf("history %d: len %d cap %d", i, len(h.DamageHistory), cap(h.DamageHistory))
110+ }
111+ for j, d := range h.DamageHistory {
112+ if d.Seq != snapshots[i][j].Seq || d.Full != snapshots[i][j].Full {
113+ t.Fatalf("history %d changed at %d: %+v want %+v", i, j, d, snapshots[i][j])
114+ }
115+ }
116+ if n := len(h.DamageHistory); h.DamageHistory[n-1].Seq != h.Seq || n > damageHistory {
117+ t.Fatalf("history %d: %+v", i, h.DamageHistory)
118+ }
119+ }
120+ allocs := testing.AllocsPerRun(100, func() { s.emitContent(c, damage{full: true}) })
121+ if allocs != 0 {
122+ t.Errorf("emitContent: %.2f allocs, want 0 (amortized)", allocs)
123+ }
124+ t.Logf("emitContent: %.2f allocs", allocs)
125+}
126+
127+// A forwarder blocked on a stale frame is woken by the next publication
128+// and sends the newer frame instead.
129+func TestForwardContentsSendsNewerFrame(t *testing.T) {
130+ contents := make(chan ports.SurfaceContent)
131+ s := &Server{
132+ channels: Channels{Contents: contents},
133+ contents: map[ports.WindowID]ports.SurfaceContent{},
134+ contentSeq: map[ports.WindowID]uint64{},
135+ damage: map[ports.WindowID][]ports.SeqDamage{},
136+ contentReady: make(chan struct{}, 1),
137+ }
138+ ctx, cancel := context.WithCancel(context.Background())
139+ done := make(chan struct{})
140+ go func() { defer close(done); s.forwardContents(ctx) }()
141+ defer func() { cancel(); <-done }()
142+ s.emitContent(ports.SurfaceContent{ID: 1, Seq: 1, Solid: &ports.SolidColor{}}, damage{full: true})
143+ time.Sleep(10 * time.Millisecond) // let the forwarder block on seq 1
144+ s.emitContent(ports.SurfaceContent{ID: 1, Seq: 2, Solid: &ports.SolidColor{}}, damage{full: true})
145+ select {
146+ case c := <-contents:
147+ if c.Seq != 2 {
148+ t.Fatalf("stale frame %d forwarded, want 2", c.Seq)
149+ }
150+ case <-time.After(2 * time.Second):
151+ t.Fatal("no frame forwarded")
152+ }
153+}
@@ -0,0 +1,153 @@
1+package wayland
2+
3+import (
4+ "context"
5+ "testing"
6+ "time"
7+
8+ "github.com/bnema/neferwl/internal/ports"
9+ "github.com/stretchr/testify/mock"
10+)
11+
12+// The steady-state hot paths that run for every frame or commit allocate
13+// nothing (or one pooled object per damage-history generation).
14+
15+func TestReleaseHeldAllocations(t *testing.T) {
16+ out := &output{place: ports.OutputPlacement{Info: ports.OutputInfo{Name: "DP-2", RefreshMilli: 60000}}}
17+ s := &Server{outputs: []*output{out}, reports: map[string]ports.OutputPresented{}}
18+ reports := []ports.OutputPresented{{Output: "DP-2", Seen: map[ports.WindowID]uint64{1: 1}}}
19+ now := time.Unix(100, 0)
20+ s.releaseHeld(now, reports)
21+ if allocs := testing.AllocsPerRun(100, func() { s.releaseHeld(now, reports) }); allocs != 0 {
22+ t.Errorf("releaseHeld: %.1f allocs, want 0", allocs)
23+ }
24+}
25+
26+func TestCommitDamageAllocations(t *testing.T) {
27+ srv := &Server{fifoSurfaces: map[*surface]struct{}{}, frameReady: make(chan struct{}, 1)}
28+ surf := &surface{server: srv}
29+ commit := func() {
30+ surf.Damage(nil, 0, 0, 4, 4)
31+ surf.DamageBuffer(nil, 1, 1, 2, 2)
32+ surf.queueUpdate()
33+ surf.queue[0].applyGraph()
34+ }
35+ commit()
36+ if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
37+ t.Errorf("damage requests + queueUpdate + apply: %.1f allocs, want 0", allocs)
38+ }
39+ // The queued update keeps its rects while the next commit collects its own.
40+ surf.Damage(nil, 0, 0, 1, 1)
41+ surf.queueUpdate()
42+ first := surf.queue[0]
43+ surf.Damage(nil, 5, 5, 2, 2)
44+ if len(first.damage) != 1 || first.damage[0].W != 1 || len(surf.next.damage) != 1 || surf.next.damage[0].W != 2 {
45+ t.Fatalf("damage shared between commits: queued %v pending %v", first.damage, surf.next.damage)
46+ }
47+ surf.queueUpdate()
48+ if second := surf.queue[1]; len(second.damage) != 1 || second.damage[0].W != 2 || first.damage[0].W != 1 {
49+ t.Fatalf("damage mixed up: %v %v", first.damage, second.damage)
50+ }
51+}
52+
53+// An explicit-sync commit moves its points into the update: the commit
54+// itself allocates no commitSync, and a pending one is not overwritten.
55+func TestCommitSyncStorageIsPerUpdate(t *testing.T) {
56+ dev := newMocksyncobjDevice(t)
57+ dev.EXPECT().signalled(mock.Anything, mock.Anything).Return(true, nil)
58+ srv := &Server{fifoSurfaces: map[*surface]struct{}{}, frameReady: make(chan struct{}, 1)}
59+ sw, err := newSyncWaiter(func() {})
60+ if err != nil {
61+ t.Fatal(err)
62+ }
63+ defer sw.pipeR.Close()
64+ defer sw.close()
65+ srv.syncWait = sw
66+ tl := &timeline{dev: dev, handle: 7, alive: true}
67+ surf := &surface{server: srv, sync: &syncState{}}
68+ take := func(point uint64) {
69+ surf.sync.acquire, surf.sync.release = syncPoint{tl: tl, point: point}, syncPoint{tl: tl, point: point + 1}
70+ surf.takeSyncPoints()
71+ surf.queueUpdate()
72+ }
73+ take(10)
74+ take(20)
75+ a, b := surf.queue[0], surf.queue[1]
76+ if a.sync == nil || b.sync == nil || a.sync == b.sync || a.sync.acquire.point != 10 || b.sync.acquire.point != 20 || surf.next.sync != nil {
77+ t.Fatalf("sync points mixed up: %+v %+v pending %+v", a.sync, b.sync, surf.next.sync)
78+ }
79+ // The commit's points live in the surface until queueUpdate: no heap
80+ // commitSync per commit (the mocked device's own allocations would drown
81+ // an AllocsPerRun count here).
82+ surf.sync.acquire, surf.sync.release = syncPoint{tl: tl, point: 30}, syncPoint{tl: tl, point: 31}
83+ surf.takeSyncPoints()
84+ if surf.next.sync != &surf.syncScratch {
85+ t.Fatal("commit points not held in the surface scratch")
86+ }
87+}
88+
89+func TestEmitContentAllocations(t *testing.T) {
90+ s := &Server{
91+ channels: Channels{Contents: make(chan ports.SurfaceContent)},
92+ contents: map[ports.WindowID]ports.SurfaceContent{},
93+ contentSeq: map[ports.WindowID]uint64{},
94+ damage: map[ports.WindowID][]ports.SeqDamage{},
95+ contentReady: make(chan struct{}, 1),
96+ }
97+ c := ports.SurfaceContent{ID: 1, Solid: &ports.SolidColor{}}
98+ var held []ports.SurfaceContent // readers retaining published histories
99+ var snapshots [][]ports.SeqDamage
100+ for range 3 * damageHistoryCap {
101+ s.emitContent(c, damage{full: true})
102+ got := s.contents[1]
103+ held = append(held, got)
104+ snapshots = append(snapshots, append([]ports.SeqDamage(nil), got.DamageHistory...))
105+ }
106+ for i, h := range held {
107+ // A published history never changes, nor can its owner append to it.
108+ if len(h.DamageHistory) != len(snapshots[i]) || cap(h.DamageHistory) != len(h.DamageHistory) {
109+ t.Fatalf("history %d: len %d cap %d", i, len(h.DamageHistory), cap(h.DamageHistory))
110+ }
111+ for j, d := range h.DamageHistory {
112+ if d.Seq != snapshots[i][j].Seq || d.Full != snapshots[i][j].Full {
113+ t.Fatalf("history %d changed at %d: %+v want %+v", i, j, d, snapshots[i][j])
114+ }
115+ }
116+ if n := len(h.DamageHistory); h.DamageHistory[n-1].Seq != h.Seq || n > damageHistory {
117+ t.Fatalf("history %d: %+v", i, h.DamageHistory)
118+ }
119+ }
120+ allocs := testing.AllocsPerRun(100, func() { s.emitContent(c, damage{full: true}) })
121+ if allocs != 0 {
122+ t.Errorf("emitContent: %.2f allocs, want 0 (amortized)", allocs)
123+ }
124+ t.Logf("emitContent: %.2f allocs", allocs)
125+}
126+
127+// A forwarder blocked on a stale frame is woken by the next publication
128+// and sends the newer frame instead.
129+func TestForwardContentsSendsNewerFrame(t *testing.T) {
130+ contents := make(chan ports.SurfaceContent)
131+ s := &Server{
132+ channels: Channels{Contents: contents},
133+ contents: map[ports.WindowID]ports.SurfaceContent{},
134+ contentSeq: map[ports.WindowID]uint64{},
135+ damage: map[ports.WindowID][]ports.SeqDamage{},
136+ contentReady: make(chan struct{}, 1),
137+ }
138+ ctx, cancel := context.WithCancel(context.Background())
139+ done := make(chan struct{})
140+ go func() { defer close(done); s.forwardContents(ctx) }()
141+ defer func() { cancel(); <-done }()
142+ s.emitContent(ports.SurfaceContent{ID: 1, Seq: 1, Solid: &ports.SolidColor{}}, damage{full: true})
143+ time.Sleep(10 * time.Millisecond) // let the forwarder block on seq 1
144+ s.emitContent(ports.SurfaceContent{ID: 1, Seq: 2, Solid: &ports.SolidColor{}}, damage{full: true})
145+ select {
146+ case c := <-contents:
147+ if c.Seq != 2 {
148+ t.Fatalf("stale frame %d forwarded, want 2", c.Seq)
149+ }
150+ case <-time.After(2 * time.Second):
151+ t.Fatal("no frame forwarded")
152+ }
153+}
M internal/adapters/wayland/fifo.go +28 -6
@@ -209,6 +209,7 @@ type update struct {
209209 deps []*update
210210 owner *surface
211211 implicit [4]*syncWait
212+ syncStore commitSync // storage of sync when the commit had explicit-sync points
212213 prev *update
213214 synced, bound bool
214215 xdg *xdgSurface
@@ -237,8 +238,8 @@ var updatePool sync.Pool
237238 func (u *update) releaseRef() {
238239 u.refs--
239240 if u.refs == 0 && u.retired {
240- deps := u.deps[:0]
241- *u = update{deps: deps}
241+ deps, dmg, bufDmg := u.deps[:0], u.damage[:0], u.bufDamage[:0]
242+ *u = update{deps: deps, damage: dmg, bufDamage: bufDmg}
242243 updatePool.Put(u)
243244 }
244245 }
@@ -257,7 +258,7 @@ func (u *update) retireUpdate() {
257258 dep.releaseRef()
258259 }
259260 if u.refs == 0 {
260- *u = update{deps: u.deps}
261+ *u = update{deps: u.deps, damage: u.damage[:0], bufDamage: u.bufDamage[:0]}
261262 updatePool.Put(u)
262263 }
263264 }
@@ -346,9 +347,19 @@ func (s *surface) queueUpdate() {
346347 if u == nil {
347348 u = new(update)
348349 }
349- deps := u.deps[:0]
350+ deps, dmg, bufDmg := u.deps[:0], u.damage[:0], u.bufDamage[:0]
350351 *u = s.takePending()
351352 u.deps = deps
353+ // The pooled update brings the capacity a retired commit's damage had;
354+ // the surface collects the next commit's rects in it. The update keeps
355+ // the rects of this commit until it retires.
356+ s.next.damage, s.next.bufDamage = dmg, bufDmg
357+ if u.sync != nil && u.sync == &s.syncScratch {
358+ // The update owns its points from here: the surface's scratch is
359+ // free for the next commit while this one waits in the queue.
360+ u.syncStore, s.syncScratch = s.syncScratch, commitSync{}
361+ u.sync = &u.syncStore
362+ }
352363 u.owner, u.synced = s, s.effectivelySynced()
353364 if s.server.syncWait != nil && u.attached && u.buffer != nil && u.sync == nil {
354365 if b, ok := s.server.buffers[u.buffer.Resource].(*dmabufBuffer); ok {
@@ -377,7 +388,6 @@ func (s *surface) queueUpdate() {
377388 }
378389 s.queue = append(s.queue, u)
379390 s.server.fifoSurfaces[s] = struct{}{}
380- s.server.wakePacer()
381391 }
382392
383393 // applyUpdate applies only the captured commit; later requests remain pending.
@@ -745,8 +755,20 @@ func sameBuffer(a, b *wayland.Buffer) bool {
745755 return a != nil && b != nil && a.Resource == b.Resource
746756 }
747757
748-// wakePacer makes the pacer look at its queues now.
758+// wakePacer makes the pacer look at its queues. Display goroutine only.
759+// While a commit or a tick runs (paceDepth) it only records the request: the
760+// outermost caller then ticks inline (paceNow), saving the round trip of
761+// waking the pacer goroutine just so that it calls Do back.
749762 func (s *Server) wakePacer() {
763+ if s.paceDepth > 0 {
764+ s.paceDirty = true
765+ return
766+ }
767+ s.signalPacer()
768+}
769+
770+// signalPacer wakes the pacer goroutine; safe from any goroutine.
771+func (s *Server) signalPacer() {
750772 select {
751773 case s.frameReady <- struct{}{}:
752774 default:
@@ -209,6 +209,7 @@ type update struct {
209 deps []*update209 deps []*update
210 owner *surface210 owner *surface
211 implicit [4]*syncWait211 implicit [4]*syncWait
212+ syncStore commitSync // storage of sync when the commit had explicit-sync points
212 prev *update213 prev *update
213 synced, bound bool214 synced, bound bool
214 xdg *xdgSurface215 xdg *xdgSurface
@@ -237,8 +238,8 @@ var updatePool sync.Pool
237 func (u *update) releaseRef() {238 func (u *update) releaseRef() {
238 u.refs--239 u.refs--
239 if u.refs == 0 && u.retired {240 if u.refs == 0 && u.retired {
240- deps := u.deps[:0]241+ deps, dmg, bufDmg := u.deps[:0], u.damage[:0], u.bufDamage[:0]
241- *u = update{deps: deps}242+ *u = update{deps: deps, damage: dmg, bufDamage: bufDmg}
242 updatePool.Put(u)243 updatePool.Put(u)
243 }244 }
244 }245 }
@@ -257,7 +258,7 @@ func (u *update) retireUpdate() {
257 dep.releaseRef()258 dep.releaseRef()
258 }259 }
259 if u.refs == 0 {260 if u.refs == 0 {
260- *u = update{deps: u.deps}261+ *u = update{deps: u.deps, damage: u.damage[:0], bufDamage: u.bufDamage[:0]}
261 updatePool.Put(u)262 updatePool.Put(u)
262 }263 }
263 }264 }
@@ -346,9 +347,19 @@ func (s *surface) queueUpdate() {
346 if u == nil {347 if u == nil {
347 u = new(update)348 u = new(update)
348 }349 }
349- deps := u.deps[:0]350+ deps, dmg, bufDmg := u.deps[:0], u.damage[:0], u.bufDamage[:0]
350 *u = s.takePending()351 *u = s.takePending()
351 u.deps = deps352 u.deps = deps
353+ // The pooled update brings the capacity a retired commit's damage had;
354+ // the surface collects the next commit's rects in it. The update keeps
355+ // the rects of this commit until it retires.
356+ s.next.damage, s.next.bufDamage = dmg, bufDmg
357+ if u.sync != nil && u.sync == &s.syncScratch {
358+ // The update owns its points from here: the surface's scratch is
359+ // free for the next commit while this one waits in the queue.
360+ u.syncStore, s.syncScratch = s.syncScratch, commitSync{}
361+ u.sync = &u.syncStore
362+ }
352 u.owner, u.synced = s, s.effectivelySynced()363 u.owner, u.synced = s, s.effectivelySynced()
353 if s.server.syncWait != nil && u.attached && u.buffer != nil && u.sync == nil {364 if s.server.syncWait != nil && u.attached && u.buffer != nil && u.sync == nil {
354 if b, ok := s.server.buffers[u.buffer.Resource].(*dmabufBuffer); ok {365 if b, ok := s.server.buffers[u.buffer.Resource].(*dmabufBuffer); ok {
@@ -377,7 +388,6 @@ func (s *surface) queueUpdate() {
377 }388 }
378 s.queue = append(s.queue, u)389 s.queue = append(s.queue, u)
379 s.server.fifoSurfaces[s] = struct{}{}390 s.server.fifoSurfaces[s] = struct{}{}
380- s.server.wakePacer()
381 }391 }
382 392
383 // applyUpdate applies only the captured commit; later requests remain pending.393 // applyUpdate applies only the captured commit; later requests remain pending.
@@ -745,8 +755,20 @@ func sameBuffer(a, b *wayland.Buffer) bool {
745 return a != nil && b != nil && a.Resource == b.Resource755 return a != nil && b != nil && a.Resource == b.Resource
746 }756 }
747 757
748-// wakePacer makes the pacer look at its queues now.758+// wakePacer makes the pacer look at its queues. Display goroutine only.
759+// While a commit or a tick runs (paceDepth) it only records the request: the
760+// outermost caller then ticks inline (paceNow), saving the round trip of
761+// waking the pacer goroutine just so that it calls Do back.
749 func (s *Server) wakePacer() {762 func (s *Server) wakePacer() {
763+ if s.paceDepth > 0 {
764+ s.paceDirty = true
765+ return
766+ }
767+ s.signalPacer()
768+}
769+
770+// signalPacer wakes the pacer goroutine; safe from any goroutine.
771+func (s *Server) signalPacer() {
750 select {772 select {
751 case s.frameReady <- struct{}{}:773 case s.frameReady <- struct{}{}:
752 default:774 default:
M internal/adapters/wayland/frames.go +96 -44
@@ -24,14 +24,23 @@ const defaultFramePeriod = time.Second / 60
2424 const suspendedFramePeriod = time.Second
2525 const suspendedFrameQueue = "\x00suspended" // never a wl_output name
2626
27-// pace fires the due frame callbacks on page flips and deadlines.
27+// maxPacePasses bounds the passes one tick repeats while a pass queues
28+// more work; the rest goes back to the pacer goroutine.
29+const maxPacePasses = 3
30+
31+// pace sleeps until a flip, a wake or the deadline armed by the last tick,
32+// then runs a tick on the display goroutine. The timer is only armed from
33+// ticks, which all run there, so a commit that ticks inline (paceNow) never
34+// races the pacer for it.
2835 func (s *Server) pace(ctx context.Context) {
29- timer := time.NewTimer(defaultFramePeriod)
36+ timer := s.paceTimer // created by Run before the pacer starts
3037 defer timer.Stop()
3138 presented := s.channels.Presented
3239 if s.flipped == nil {
3340 s.flipped = make(map[string]bool)
3441 }
42+ // One method value for the whole loop: no closure per wake.
43+ tick := s.paceFromPacer
3544 for {
3645 clear(s.flipped)
3746 flipped := s.flipped
@@ -61,52 +70,95 @@ func (s *Server) pace(ctx context.Context) {
6170 }
6271 }
6372 s.frameReports = reports
64- wait, idle := defaultFramePeriod, false
65- if !s.display.Do(func() {
66- if ctx.Err() == nil {
67- // Each flip answers its own presentation feedbacks, in
68- // order: flips are never merged here.
69- for _, p := range reports {
70- if p.Flip != nil {
71- s.flips[p.Output] = *p.Flip
72- s.presentFlip(p.Output, p.Flip)
73- }
74- }
75- s.dropFeedbacks(time.Now())
76- if len(s.feedbacks) > 0 {
77- wait = feedbackTimeout
78- }
79- var due []string
80- var dueWait time.Duration
81- due, dueWait, idle = s.dueFrames(time.Now(), flipped)
82- if len(s.feedbacks) > 0 {
83- // Unanswered feedbacks time out: keep ticking.
84- dueWait, idle = min(dueWait, feedbackTimeout), false
85- }
86- wait = dueWait
87- for _, name := range due {
88- s.sendFrames(name, flipped[name])
89- }
90- if s.releaseHeld(time.Now(), reports) && idle {
91- // Held buffers wait for a flip or heldTimeout.
92- wait, idle = heldTimeout, false
93- }
94- // Fifo barriers and queued commits wait for a refresh.
95- if fw, waiting := s.tickFifo(time.Now(), flipped); waiting {
96- wait, idle = min(wait, fw), false
97- }
98- }
99- }) {
73+ if !s.display.Do(tick) {
10074 return
10175 }
102- if idle {
103- // Nothing waits: sleep until a commit queues a callback
104- // (frameReady) or a flip.
105- timer.Stop()
106- continue
76+ }
77+}
78+
79+// paceFromPacer runs the pacer's collected flips (flipped, frameReports)
80+// on the display goroutine.
81+func (s *Server) paceFromPacer() { s.paceTick(s.flipped, s.frameReports) }
82+
83+// paceNow runs a tick from code already on the display goroutine (a client
84+// commit): the pacer goroutine would only call Do to get back here. Without
85+// a running pacer (no timer) it falls back to waking it.
86+func (s *Server) paceNow() {
87+ if s.paceTimer == nil {
88+ s.signalPacer()
89+ return
90+ }
91+ s.paceTick(nil, nil)
92+}
93+
94+// paceTick answers the flips in reports, fires the due frame callbacks,
95+// releases buffers and applies the ready FIFO commits, then arms the timer
96+// for the next deadline (stopped when nothing waits). Display goroutine only.
97+func (s *Server) paceTick(flipped map[string]bool, reports []ports.OutputPresented) {
98+ if s.ctx != nil && s.ctx.Err() != nil {
99+ return
100+ }
101+ wait, idle := defaultFramePeriod, false
102+ for pass := 0; ; pass++ {
103+ s.paceDirty = false
104+ s.paceDepth++
105+ wait, idle = s.pacePass(flipped, reports)
106+ s.paceDepth--
107+ if !s.paceDirty {
108+ break
107109 }
108- timer.Reset(wait)
110+ // The pass queued work behind its own checks (applied commits
111+ // queue callbacks, released buffers): look again, as a flip-less
112+ // tick, instead of waking the pacer for it.
113+ flipped, reports = nil, nil
114+ if pass+1 == maxPacePasses {
115+ s.paceDirty = false
116+ s.signalPacer()
117+ break
118+ }
119+ }
120+ if s.paceTimer == nil {
121+ return
122+ }
123+ if idle {
124+ // Nothing waits: sleep until a commit queues a callback
125+ // (wakePacer) or a flip.
126+ s.paceTimer.Stop()
127+ return
128+ }
129+ s.paceTimer.Reset(wait)
130+}
131+
132+// pacePass is one pass of paceTick; it returns the time to the next
133+// deadline and whether nothing waits.
134+func (s *Server) pacePass(flipped map[string]bool, reports []ports.OutputPresented) (wait time.Duration, idle bool) {
135+ // Each flip answers its own presentation feedbacks, in order: flips
136+ // are never merged here.
137+ for _, p := range reports {
138+ if p.Flip != nil {
139+ s.flips[p.Output] = *p.Flip
140+ s.presentFlip(p.Output, p.Flip)
141+ }
142+ }
143+ s.dropFeedbacks(time.Now())
144+ var due []string
145+ due, wait, idle = s.dueFrames(time.Now(), flipped)
146+ if len(s.feedbacks) > 0 {
147+ // Unanswered feedbacks time out: keep ticking.
148+ wait, idle = min(wait, feedbackTimeout), false
149+ }
150+ for _, name := range due {
151+ s.sendFrames(name, flipped[name])
152+ }
153+ if s.releaseHeld(time.Now(), reports) && idle {
154+ // Held buffers wait for a flip or heldTimeout.
155+ wait, idle = heldTimeout, false
156+ }
157+ // Fifo barriers and queued commits wait for a refresh.
158+ if fw, waiting := s.tickFifo(time.Now(), flipped); waiting {
159+ wait, idle = min(wait, fw), false
109160 }
161+ return wait, idle
110162 }
111163
112164 // toplevelRoot resolves subsurfaces and popups to their parent toplevel.
@@ -24,14 +24,23 @@ const defaultFramePeriod = time.Second / 60
24 const suspendedFramePeriod = time.Second24 const suspendedFramePeriod = time.Second
25 const suspendedFrameQueue = "\x00suspended" // never a wl_output name25 const suspendedFrameQueue = "\x00suspended" // never a wl_output name
26 26
27-// pace fires the due frame callbacks on page flips and deadlines.27+// maxPacePasses bounds the passes one tick repeats while a pass queues
28+// more work; the rest goes back to the pacer goroutine.
29+const maxPacePasses = 3
30+
31+// pace sleeps until a flip, a wake or the deadline armed by the last tick,
32+// then runs a tick on the display goroutine. The timer is only armed from
33+// ticks, which all run there, so a commit that ticks inline (paceNow) never
34+// races the pacer for it.
28 func (s *Server) pace(ctx context.Context) {35 func (s *Server) pace(ctx context.Context) {
29- timer := time.NewTimer(defaultFramePeriod)36+ timer := s.paceTimer // created by Run before the pacer starts
30 defer timer.Stop()37 defer timer.Stop()
31 presented := s.channels.Presented38 presented := s.channels.Presented
32 if s.flipped == nil {39 if s.flipped == nil {
33 s.flipped = make(map[string]bool)40 s.flipped = make(map[string]bool)
34 }41 }
42+ // One method value for the whole loop: no closure per wake.
43+ tick := s.paceFromPacer
35 for {44 for {
36 clear(s.flipped)45 clear(s.flipped)
37 flipped := s.flipped46 flipped := s.flipped
@@ -61,52 +70,95 @@ func (s *Server) pace(ctx context.Context) {
61 }70 }
62 }71 }
63 s.frameReports = reports72 s.frameReports = reports
64- wait, idle := defaultFramePeriod, false73+ if !s.display.Do(tick) {
65- if !s.display.Do(func() {
66- if ctx.Err() == nil {
67- // Each flip answers its own presentation feedbacks, in
68- // order: flips are never merged here.
69- for _, p := range reports {
70- if p.Flip != nil {
71- s.flips[p.Output] = *p.Flip
72- s.presentFlip(p.Output, p.Flip)
73- }
74- }
75- s.dropFeedbacks(time.Now())
76- if len(s.feedbacks) > 0 {
77- wait = feedbackTimeout
78- }
79- var due []string
80- var dueWait time.Duration
81- due, dueWait, idle = s.dueFrames(time.Now(), flipped)
82- if len(s.feedbacks) > 0 {
83- // Unanswered feedbacks time out: keep ticking.
84- dueWait, idle = min(dueWait, feedbackTimeout), false
85- }
86- wait = dueWait
87- for _, name := range due {
88- s.sendFrames(name, flipped[name])
89- }
90- if s.releaseHeld(time.Now(), reports) && idle {
91- // Held buffers wait for a flip or heldTimeout.
92- wait, idle = heldTimeout, false
93- }
94- // Fifo barriers and queued commits wait for a refresh.
95- if fw, waiting := s.tickFifo(time.Now(), flipped); waiting {
96- wait, idle = min(wait, fw), false
97- }
98- }
99- }) {
100 return74 return
101 }75 }
102- if idle {76+ }
103- // Nothing waits: sleep until a commit queues a callback77+}
104- // (frameReady) or a flip.78+
105- timer.Stop()79+// paceFromPacer runs the pacer's collected flips (flipped, frameReports)
106- continue80+// on the display goroutine.
81+func (s *Server) paceFromPacer() { s.paceTick(s.flipped, s.frameReports) }
82+
83+// paceNow runs a tick from code already on the display goroutine (a client
84+// commit): the pacer goroutine would only call Do to get back here. Without
85+// a running pacer (no timer) it falls back to waking it.
86+func (s *Server) paceNow() {
87+ if s.paceTimer == nil {
88+ s.signalPacer()
89+ return
90+ }
91+ s.paceTick(nil, nil)
92+}
93+
94+// paceTick answers the flips in reports, fires the due frame callbacks,
95+// releases buffers and applies the ready FIFO commits, then arms the timer
96+// for the next deadline (stopped when nothing waits). Display goroutine only.
97+func (s *Server) paceTick(flipped map[string]bool, reports []ports.OutputPresented) {
98+ if s.ctx != nil && s.ctx.Err() != nil {
99+ return
100+ }
101+ wait, idle := defaultFramePeriod, false
102+ for pass := 0; ; pass++ {
103+ s.paceDirty = false
104+ s.paceDepth++
105+ wait, idle = s.pacePass(flipped, reports)
106+ s.paceDepth--
107+ if !s.paceDirty {
108+ break
107 }109 }
108- timer.Reset(wait)110+ // The pass queued work behind its own checks (applied commits
111+ // queue callbacks, released buffers): look again, as a flip-less
112+ // tick, instead of waking the pacer for it.
113+ flipped, reports = nil, nil
114+ if pass+1 == maxPacePasses {
115+ s.paceDirty = false
116+ s.signalPacer()
117+ break
118+ }
119+ }
120+ if s.paceTimer == nil {
121+ return
122+ }
123+ if idle {
124+ // Nothing waits: sleep until a commit queues a callback
125+ // (wakePacer) or a flip.
126+ s.paceTimer.Stop()
127+ return
128+ }
129+ s.paceTimer.Reset(wait)
130+}
131+
132+// pacePass is one pass of paceTick; it returns the time to the next
133+// deadline and whether nothing waits.
134+func (s *Server) pacePass(flipped map[string]bool, reports []ports.OutputPresented) (wait time.Duration, idle bool) {
135+ // Each flip answers its own presentation feedbacks, in order: flips
136+ // are never merged here.
137+ for _, p := range reports {
138+ if p.Flip != nil {
139+ s.flips[p.Output] = *p.Flip
140+ s.presentFlip(p.Output, p.Flip)
141+ }
142+ }
143+ s.dropFeedbacks(time.Now())
144+ var due []string
145+ due, wait, idle = s.dueFrames(time.Now(), flipped)
146+ if len(s.feedbacks) > 0 {
147+ // Unanswered feedbacks time out: keep ticking.
148+ wait, idle = min(wait, feedbackTimeout), false
149+ }
150+ for _, name := range due {
151+ s.sendFrames(name, flipped[name])
152+ }
153+ if s.releaseHeld(time.Now(), reports) && idle {
154+ // Held buffers wait for a flip or heldTimeout.
155+ wait, idle = heldTimeout, false
156+ }
157+ // Fifo barriers and queued commits wait for a refresh.
158+ if fw, waiting := s.tickFifo(time.Now(), flipped); waiting {
159+ wait, idle = min(wait, fw), false
109 }160 }
161+ return wait, idle
110 }162 }
111 163
112 // toplevelRoot resolves subsurfaces and popups to their parent toplevel.164 // toplevelRoot resolves subsurfaces and popups to their parent toplevel.
A internal/adapters/wayland/pace_inline_test.go +118 -0
@@ -0,0 +1,118 @@
1+package wayland
2+
3+import (
4+ "testing"
5+ "time"
6+
7+ "github.com/bnema/go-wayland-bindings/server/wayland"
8+ "github.com/bnema/neferwl/internal/ports"
9+)
10+
11+// paceFixture is a server whose surface waits on an output whose next
12+// frame is an hour away: a commit queues its callback and nothing fires.
13+func paceFixture(timer bool) (*Server, *surface) {
14+ out := &output{place: ports.OutputPlacement{Info: ports.OutputInfo{Name: "DP-2", RefreshMilli: 60000}}}
15+ s := &Server{
16+ outputs: []*output{out},
17+ awaiting: map[string][]*wayland.Callback{},
18+ frameDue: map[string]time.Time{"DP-2": time.Now().Add(time.Hour)},
19+ fifoSurfaces: map[*surface]struct{}{},
20+ lastFlip: map[string]time.Time{},
21+ flips: map[string]ports.FlipInfo{},
22+ reports: map[string]ports.OutputPresented{},
23+ frameReady: make(chan struct{}, 1),
24+ }
25+ if timer {
26+ s.paceTimer = time.NewTimer(time.Hour)
27+ s.paceTimer.Stop()
28+ }
29+ w := &window{hasLast: true, last: ports.ConfigureWindow{Output: "DP-2", Visible: true}}
30+ surf := &surface{server: s, xdg: &xdgSurface{window: w}}
31+ surf.next.callbacks = []*wayland.Callback{new(wayland.Callback)}
32+ return s, surf
33+}
34+
35+// A commit on the display goroutine ticks the pacer inline: it arms the
36+// pacer's timer itself and does not wake the pacer goroutine, which would
37+// only call display.Do to get back to the display goroutine.
38+func TestCommitTicksPacerInline(t *testing.T) {
39+ s, surf := paceFixture(true)
40+ surf.queueAndTick()
41+ if len(s.awaiting["DP-2"]) != 1 {
42+ t.Fatalf("callback not queued: %+v", s.awaiting)
43+ }
44+ if len(s.frameReady) != 0 {
45+ t.Fatal("commit woke the pacer goroutine")
46+ }
47+ waitArmed(t, s, "inline tick did not arm the pacer timer for the next deadline")
48+ if s.paceDepth != 0 || s.paceDirty {
49+ t.Fatalf("pace state leaked: depth %d dirty %v", s.paceDepth, s.paceDirty)
50+ }
51+}
52+
53+// Without a running pacer (no timer) the wake falls back to the channel.
54+func TestCommitWakesPacerWithoutTimer(t *testing.T) {
55+ s, surf := paceFixture(false)
56+ surf.queueAndTick()
57+ if len(s.frameReady) != 1 {
58+ t.Fatal("pacer not woken")
59+ }
60+}
61+
62+// Wakes inside a tick coalesce into one more pass, not one channel wake each.
63+func TestWakePacerDefersInsideTick(t *testing.T) {
64+ s, _ := paceFixture(true)
65+ s.paceDepth = 1
66+ s.wakePacer()
67+ s.wakePacer()
68+ if len(s.frameReady) != 0 || !s.paceDirty {
69+ t.Fatalf("wake inside a tick: queued %d dirty %v", len(s.frameReady), s.paceDirty)
70+ }
71+ s.paceDepth = 0
72+ s.wakePacer()
73+ if len(s.frameReady) != 1 {
74+ t.Fatal("wake outside a tick lost")
75+ }
76+}
77+
78+// A commit that waits (a fifo barrier) is timed by the inline tick.
79+func TestCommitWaitingOnBarrierArmsTimer(t *testing.T) {
80+ s, surf := paceFixture(true)
81+ surf.next.callbacks = nil
82+ surf.barrier, surf.barrierAt = true, time.Now()
83+ surf.next.wait = true
84+ surf.queueAndTick()
85+ if len(surf.queue) != 1 || len(s.frameReady) != 0 {
86+ t.Fatalf("queued %d, pacer wakes %d", len(surf.queue), len(s.frameReady))
87+ }
88+ waitArmed(t, s, "timer not armed for the barrier deadline")
89+}
90+
91+// waitArmed asserts the timer was armed: it is pending, or it already
92+// fired and holds its value.
93+func waitArmed(t *testing.T, s *Server, msg string) {
94+ t.Helper()
95+ if s.paceTimer.Stop() {
96+ return
97+ }
98+ select {
99+ case <-s.paceTimer.C:
100+ default:
101+ t.Fatal(msg)
102+ }
103+}
104+
105+// A synchronized subsurface commit waits for its parent's commit: it does
106+// not tick the pacer, so a burst of child commits costs no ticks.
107+func TestSyncedCommitDoesNotTick(t *testing.T) {
108+ s, surf := paceFixture(true)
109+ surf.sub.parent = &surface{server: s}
110+ surf.sub.synced = true
111+ surf.queueAndTick()
112+ if len(surf.queue) != 1 || !surf.queue[0].synced {
113+ t.Fatalf("synced commit not queued: %d", len(surf.queue))
114+ }
115+ if len(s.frameReady) != 0 || s.paceTimer.Stop() {
116+ t.Fatal("synced commit ticked the pacer")
117+ }
118+}
@@ -0,0 +1,118 @@
1+package wayland
2+
3+import (
4+ "testing"
5+ "time"
6+
7+ "github.com/bnema/go-wayland-bindings/server/wayland"
8+ "github.com/bnema/neferwl/internal/ports"
9+)
10+
11+// paceFixture is a server whose surface waits on an output whose next
12+// frame is an hour away: a commit queues its callback and nothing fires.
13+func paceFixture(timer bool) (*Server, *surface) {
14+ out := &output{place: ports.OutputPlacement{Info: ports.OutputInfo{Name: "DP-2", RefreshMilli: 60000}}}
15+ s := &Server{
16+ outputs: []*output{out},
17+ awaiting: map[string][]*wayland.Callback{},
18+ frameDue: map[string]time.Time{"DP-2": time.Now().Add(time.Hour)},
19+ fifoSurfaces: map[*surface]struct{}{},
20+ lastFlip: map[string]time.Time{},
21+ flips: map[string]ports.FlipInfo{},
22+ reports: map[string]ports.OutputPresented{},
23+ frameReady: make(chan struct{}, 1),
24+ }
25+ if timer {
26+ s.paceTimer = time.NewTimer(time.Hour)
27+ s.paceTimer.Stop()
28+ }
29+ w := &window{hasLast: true, last: ports.ConfigureWindow{Output: "DP-2", Visible: true}}
30+ surf := &surface{server: s, xdg: &xdgSurface{window: w}}
31+ surf.next.callbacks = []*wayland.Callback{new(wayland.Callback)}
32+ return s, surf
33+}
34+
35+// A commit on the display goroutine ticks the pacer inline: it arms the
36+// pacer's timer itself and does not wake the pacer goroutine, which would
37+// only call display.Do to get back to the display goroutine.
38+func TestCommitTicksPacerInline(t *testing.T) {
39+ s, surf := paceFixture(true)
40+ surf.queueAndTick()
41+ if len(s.awaiting["DP-2"]) != 1 {
42+ t.Fatalf("callback not queued: %+v", s.awaiting)
43+ }
44+ if len(s.frameReady) != 0 {
45+ t.Fatal("commit woke the pacer goroutine")
46+ }
47+ waitArmed(t, s, "inline tick did not arm the pacer timer for the next deadline")
48+ if s.paceDepth != 0 || s.paceDirty {
49+ t.Fatalf("pace state leaked: depth %d dirty %v", s.paceDepth, s.paceDirty)
50+ }
51+}
52+
53+// Without a running pacer (no timer) the wake falls back to the channel.
54+func TestCommitWakesPacerWithoutTimer(t *testing.T) {
55+ s, surf := paceFixture(false)
56+ surf.queueAndTick()
57+ if len(s.frameReady) != 1 {
58+ t.Fatal("pacer not woken")
59+ }
60+}
61+
62+// Wakes inside a tick coalesce into one more pass, not one channel wake each.
63+func TestWakePacerDefersInsideTick(t *testing.T) {
64+ s, _ := paceFixture(true)
65+ s.paceDepth = 1
66+ s.wakePacer()
67+ s.wakePacer()
68+ if len(s.frameReady) != 0 || !s.paceDirty {
69+ t.Fatalf("wake inside a tick: queued %d dirty %v", len(s.frameReady), s.paceDirty)
70+ }
71+ s.paceDepth = 0
72+ s.wakePacer()
73+ if len(s.frameReady) != 1 {
74+ t.Fatal("wake outside a tick lost")
75+ }
76+}
77+
78+// A commit that waits (a fifo barrier) is timed by the inline tick.
79+func TestCommitWaitingOnBarrierArmsTimer(t *testing.T) {
80+ s, surf := paceFixture(true)
81+ surf.next.callbacks = nil
82+ surf.barrier, surf.barrierAt = true, time.Now()
83+ surf.next.wait = true
84+ surf.queueAndTick()
85+ if len(surf.queue) != 1 || len(s.frameReady) != 0 {
86+ t.Fatalf("queued %d, pacer wakes %d", len(surf.queue), len(s.frameReady))
87+ }
88+ waitArmed(t, s, "timer not armed for the barrier deadline")
89+}
90+
91+// waitArmed asserts the timer was armed: it is pending, or it already
92+// fired and holds its value.
93+func waitArmed(t *testing.T, s *Server, msg string) {
94+ t.Helper()
95+ if s.paceTimer.Stop() {
96+ return
97+ }
98+ select {
99+ case <-s.paceTimer.C:
100+ default:
101+ t.Fatal(msg)
102+ }
103+}
104+
105+// A synchronized subsurface commit waits for its parent's commit: it does
106+// not tick the pacer, so a burst of child commits costs no ticks.
107+func TestSyncedCommitDoesNotTick(t *testing.T) {
108+ s, surf := paceFixture(true)
109+ surf.sub.parent = &surface{server: s}
110+ surf.sub.synced = true
111+ surf.queueAndTick()
112+ if len(surf.queue) != 1 || !surf.queue[0].synced {
113+ t.Fatalf("synced commit not queued: %d", len(surf.queue))
114+ }
115+ if len(s.frameReady) != 0 || s.paceTimer.Stop() {
116+ t.Fatal("synced commit ticked the pacer")
117+ }
118+}
M internal/adapters/wayland/presentation.go +1 -0
@@ -108,6 +108,7 @@ func (s *surface) commitFeedback(pending []*presentationtime.WpPresentationFeedb
108108 for _, fb := range pending {
109109 srv.feedbacks = append(srv.feedbacks, feedbackWait{fb: fb, surf: s, win: win, seq: seq, at: time.Now()})
110110 }
111+ srv.wakePacer() // an unanswered feedback times out: the pacer must keep ticking
111112 }
112113
113114 func discard(fb *presentationtime.WpPresentationFeedback) {
@@ -108,6 +108,7 @@ func (s *surface) commitFeedback(pending []*presentationtime.WpPresentationFeedb
108 for _, fb := range pending {108 for _, fb := range pending {
109 srv.feedbacks = append(srv.feedbacks, feedbackWait{fb: fb, surf: s, win: win, seq: seq, at: time.Now()})109 srv.feedbacks = append(srv.feedbacks, feedbackWait{fb: fb, surf: s, win: win, seq: seq, at: time.Now()})
110 }110 }
111+ srv.wakePacer() // an unanswered feedback times out: the pacer must keep ticking
111 }112 }
112 113
113 func discard(fb *presentationtime.WpPresentationFeedback) {114 func discard(fb *presentationtime.WpPresentationFeedback) {
A internal/adapters/wayland/probe.go +15 -0
@@ -0,0 +1,15 @@
1+package wayland
2+
3+import "context"
4+
5+// Ping returns once the display loop ran an empty task, the display
6+// stopped or ctx ended. The watchdog uses it to detect a stuck display
7+// loop; a stuck loop leaves only the inner Do goroutine behind.
8+func (s *Server) Ping(ctx context.Context) {
9+ done := make(chan struct{})
10+ go func() { s.display.Do(func() {}); close(done) }()
11+ select {
12+ case <-done:
13+ case <-ctx.Done():
14+ }
15+}
@@ -0,0 +1,15 @@
1+package wayland
2+
3+import "context"
4+
5+// Ping returns once the display loop ran an empty task, the display
6+// stopped or ctx ended. The watchdog uses it to detect a stuck display
7+// loop; a stuck loop leaves only the inner Do goroutine behind.
8+func (s *Server) Ping(ctx context.Context) {
9+ done := make(chan struct{})
10+ go func() { s.display.Do(func() {}); close(done) }()
11+ select {
12+ case <-done:
13+ case <-ctx.Done():
14+ }
15+}
M internal/adapters/wayland/release.go +5 -5
@@ -52,10 +52,7 @@ func (s *Server) releaseBuffer(surf *surface, b *wayland.Buffer, sync syncHold)
5252 s.contentMu.Unlock()
5353 s.held = append(s.held, h)
5454 // Wake the pacer: it releases the buffer on the next output report.
55- select {
56- case s.frameReady <- struct{}{}:
57- default:
58- }
55+ s.wakePacer()
5956 }
6057
6158 // release gives a buffer back: its release point is signalled when it
@@ -98,7 +95,7 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
9895 for _, r := range reports {
9996 s.reports[r.Output] = r
10097 }
101- live := make([]ports.OutputPresented, 0, len(s.outputs))
98+ live := s.heldLive[:0]
10299 for name := range s.reports {
103100 if s.outputByNameExact(name) == nil {
104101 delete(s.reports, name)
@@ -107,6 +104,8 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
107104 for _, r := range s.reports {
108105 live = append(live, r)
109106 }
107+ // keepHeld only reads live: the scratch holds nothing across calls.
108+ s.heldLive = live
110109 kept := s.held[:0]
111110 for _, h := range s.held {
112111 if s.bufferReferenced(h.res) || keepHeld(h, live, now) {
@@ -117,6 +116,7 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
117116 }
118117 clear(s.held[len(kept):])
119118 s.held = kept
119+ clear(live) // the reports carry maps: do not retain them
120120 return len(s.held) > 0
121121 }
122122
@@ -52,10 +52,7 @@ func (s *Server) releaseBuffer(surf *surface, b *wayland.Buffer, sync syncHold)
52 s.contentMu.Unlock()52 s.contentMu.Unlock()
53 s.held = append(s.held, h)53 s.held = append(s.held, h)
54 // Wake the pacer: it releases the buffer on the next output report.54 // Wake the pacer: it releases the buffer on the next output report.
55- select {55+ s.wakePacer()
56- case s.frameReady <- struct{}{}:
57- default:
58- }
59 }56 }
60 57
61 // release gives a buffer back: its release point is signalled when it58 // release gives a buffer back: its release point is signalled when it
@@ -98,7 +95,7 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
98 for _, r := range reports {95 for _, r := range reports {
99 s.reports[r.Output] = r96 s.reports[r.Output] = r
100 }97 }
101- live := make([]ports.OutputPresented, 0, len(s.outputs))98+ live := s.heldLive[:0]
102 for name := range s.reports {99 for name := range s.reports {
103 if s.outputByNameExact(name) == nil {100 if s.outputByNameExact(name) == nil {
104 delete(s.reports, name)101 delete(s.reports, name)
@@ -107,6 +104,8 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
107 for _, r := range s.reports {104 for _, r := range s.reports {
108 live = append(live, r)105 live = append(live, r)
109 }106 }
107+ // keepHeld only reads live: the scratch holds nothing across calls.
108+ s.heldLive = live
110 kept := s.held[:0]109 kept := s.held[:0]
111 for _, h := range s.held {110 for _, h := range s.held {
112 if s.bufferReferenced(h.res) || keepHeld(h, live, now) {111 if s.bufferReferenced(h.res) || keepHeld(h, live, now) {
@@ -117,6 +116,7 @@ func (s *Server) releaseHeld(now time.Time, reports []ports.OutputPresented) boo
117 }116 }
118 clear(s.held[len(kept):])117 clear(s.held[len(kept):])
119 s.held = kept118 s.held = kept
119+ clear(live) // the reports carry maps: do not retain them
120 return len(s.held) > 0120 return len(s.held) > 0
121 }121 }
122 122
M internal/adapters/wayland/server.go +36 -20
@@ -151,6 +151,12 @@ type Server struct {
151151 frameOwners map[*wayland.Callback]*surface
152152 frameDue map[string]time.Time
153153 frameReady chan struct{}
154+ // paceTimer is the pacer's deadline, armed by ticks on the display
155+ // goroutine; paceDepth/paceDirty defer wakePacer inside a commit or
156+ // tick (display goroutine only).
157+ paceTimer *time.Timer
158+ paceDepth int
159+ paceDirty bool
154160 // Pacer scratch: periods is used on the display goroutine, while
155161 // flipped and frameReports belong to pace and its synchronous Do call.
156162 framePeriods map[string]time.Duration
@@ -158,10 +164,12 @@ type Server struct {
158164 frameReports []ports.OutputPresented
159165 // held buffers wait until no output reads them (release.go); reports
160166 // are the outputs' latest ports.OutputPresented.
161- held []heldBuffer
162- reports map[string]ports.OutputPresented
163- frames uint64
164- started time.Time
167+ held []heldBuffer
168+ // heldLive is releaseHeld's scratch (display goroutine only).
169+ heldLive []ports.OutputPresented
170+ reports map[string]ports.OutputPresented
171+ frames uint64
172+ started time.Time
165173 // fifoSurfaces have a fifo barrier or queued commits (fifo.go);
166174 // lastFlip is each output's latest page flip.
167175 fifoSurfaces map[*surface]struct{}
@@ -227,7 +235,6 @@ type Server struct {
227235 clock ports.Clock
228236 powers []*outputPower
229237 outputsOff map[string]bool
230- contentNotify chan struct{}
231238 contentReady chan struct{}
232239 outputs []*output
233240 outputManagers []*outputManager
@@ -329,7 +336,6 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
329336 dataSources: map[*server.Resource]*clipSource{},
330337 primarySources: map[*server.Resource]*clipSource{},
331338 controlSources: map[*server.Resource]*clipSource{},
332- contentNotify: make(chan struct{}),
333339 regions: map[*server.Resource]*region{},
334340 relatives: map[server.Client][]*relativepointer.ZwpRelativePointerV1{},
335341 gestures: map[server.Client][]*gestureObject{},
@@ -369,7 +375,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
369375 var node *renderNode
370376 if opts.syncDev != nil {
371377 s.syncDev = opts.syncDev
372- if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {
378+ if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
373379 cleanup()
374380 d.Close()
375381 return nil, err
@@ -378,7 +384,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
378384 if node, err = openSyncobj(opts.SyncobjNode); err != nil {
379385 log.Info().Err(err).Msg("explicit sync off")
380386 node = nil
381- } else if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {
387+ } else if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
382388 node.close()
383389 node = nil
384390 } else {
@@ -387,7 +393,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
387393 }
388394 // Implicit dma-buf fences also need this waiter when explicit sync is off.
389395 if s.syncWait == nil {
390- if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {
396+ if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
391397 cleanup()
392398 d.Close()
393399 return nil, err
@@ -453,6 +459,7 @@ func (s *Server) Run(ctx context.Context) error {
453459 s.log.Info().Str("socket", s.name).Msg("starting wayland")
454460 s.started = time.Now()
455461 s.ctx = ctx
462+ s.paceTimer = time.NewTimer(defaultFramePeriod)
456463 var wg sync.WaitGroup
457464 wg.Add(15)
458465 go func() { defer wg.Done(); s.forwardBackendSecurity(ctx) }()
@@ -563,6 +570,10 @@ func (s *Server) forwardOutputApplied(ctx context.Context) {
563570 // enough for a renderer target a few frames old.
564571 const damageHistory = 4
565572
573+// damageHistoryCap is the capacity of a history's backing array: one
574+// allocation serves damageHistoryCap-damageHistory+1 publications.
575+const damageHistoryCap = 8 * damageHistory
576+
566577 // forwardOutputFormats applies the outputs' scanout formats on the
567578 // display goroutine.
568579 func (s *Server) forwardOutputFormats(ctx context.Context) {
@@ -632,23 +643,28 @@ func (s *Server) emitContent(c ports.SurfaceContent, d damage) {
632643 s.contentMu.Lock()
633644 s.contentSeq[c.ID]++
634645 c.Seq = s.contentSeq[c.ID]
635- // Copy-on-write: previous publications can still be in an output's hands.
646+ // Published histories are immutable: an output may hold one for as long
647+ // as it likes. The history grows inside one backing array, appended past
648+ // the end of every published slice (never over an element a reader can
649+ // see), and moves to a fresh array only when that one is full. Readers
650+ // get a capacity-clipped slice, so they cannot append into the spare.
636651 old := s.damage[c.ID]
637652 if len(old) >= damageHistory {
638653 old = old[len(old)-damageHistory+1:]
639654 }
640- h := make([]ports.SeqDamage, len(old)+1)
641- copy(h, old)
642- h[len(old)] = ports.SeqDamage{Seq: c.Seq, Full: d.full, Rects: d.rects}
655+ if len(old) == cap(old) {
656+ grown := make([]ports.SeqDamage, len(old), damageHistoryCap)
657+ copy(grown, old)
658+ old = grown
659+ }
660+ h := append(old, ports.SeqDamage{Seq: c.Seq, Full: d.full, Rects: d.rects})
643661 if c.Empty() {
644662 delete(s.damage, c.ID)
645663 } else {
646664 s.damage[c.ID] = h
647665 }
648- c.DamageHistory = h
666+ c.DamageHistory = h[:len(h):len(h)]
649667 s.contents[c.ID] = c
650- close(s.contentNotify)
651- s.contentNotify = make(chan struct{})
652668 s.contentMu.Unlock()
653669 select {
654670 case s.contentReady <- struct{}{}:
@@ -663,13 +679,11 @@ func (s *Server) forwardContents(ctx context.Context) {
663679 var c ports.SurfaceContent
664680 found := false
665681 var seq uint64
666- var notify <-chan struct{}
667682 for id, c = range s.contents {
668683 found = true
669684 seq = s.contentSeq[id]
670685 break
671686 }
672- notify = s.contentNotify
673687 s.contentMu.Unlock()
674688 if !found {
675689 select {
@@ -679,11 +693,13 @@ func (s *Server) forwardContents(ctx context.Context) {
679693 continue
680694 }
681695 }
682- // Do not hold a stale frame while the receiver is busy: wake on newer data.
696+ // Do not hold a stale frame while the receiver is busy: wake on newer
697+ // data. contentReady is signalled after every publication; a stale
698+ // signal only costs one more loop.
683699 select {
684700 case <-ctx.Done():
685701 return
686- case <-notify:
702+ case <-s.contentReady:
687703 continue
688704 case s.channels.Contents <- c:
689705 s.contentMu.Lock()
@@ -151,6 +151,12 @@ type Server struct {
151 frameOwners map[*wayland.Callback]*surface151 frameOwners map[*wayland.Callback]*surface
152 frameDue map[string]time.Time152 frameDue map[string]time.Time
153 frameReady chan struct{}153 frameReady chan struct{}
154+ // paceTimer is the pacer's deadline, armed by ticks on the display
155+ // goroutine; paceDepth/paceDirty defer wakePacer inside a commit or
156+ // tick (display goroutine only).
157+ paceTimer *time.Timer
158+ paceDepth int
159+ paceDirty bool
154 // Pacer scratch: periods is used on the display goroutine, while160 // Pacer scratch: periods is used on the display goroutine, while
155 // flipped and frameReports belong to pace and its synchronous Do call.161 // flipped and frameReports belong to pace and its synchronous Do call.
156 framePeriods map[string]time.Duration162 framePeriods map[string]time.Duration
@@ -158,10 +164,12 @@ type Server struct {
158 frameReports []ports.OutputPresented164 frameReports []ports.OutputPresented
159 // held buffers wait until no output reads them (release.go); reports165 // held buffers wait until no output reads them (release.go); reports
160 // are the outputs' latest ports.OutputPresented.166 // are the outputs' latest ports.OutputPresented.
161- held []heldBuffer167+ held []heldBuffer
162- reports map[string]ports.OutputPresented168+ // heldLive is releaseHeld's scratch (display goroutine only).
163- frames uint64169+ heldLive []ports.OutputPresented
164- started time.Time170+ reports map[string]ports.OutputPresented
171+ frames uint64
172+ started time.Time
165 // fifoSurfaces have a fifo barrier or queued commits (fifo.go);173 // fifoSurfaces have a fifo barrier or queued commits (fifo.go);
166 // lastFlip is each output's latest page flip.174 // lastFlip is each output's latest page flip.
167 fifoSurfaces map[*surface]struct{}175 fifoSurfaces map[*surface]struct{}
@@ -227,7 +235,6 @@ type Server struct {
227 clock ports.Clock235 clock ports.Clock
228 powers []*outputPower236 powers []*outputPower
229 outputsOff map[string]bool237 outputsOff map[string]bool
230- contentNotify chan struct{}
231 contentReady chan struct{}238 contentReady chan struct{}
232 outputs []*output239 outputs []*output
233 outputManagers []*outputManager240 outputManagers []*outputManager
@@ -329,7 +336,6 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
329 dataSources: map[*server.Resource]*clipSource{},336 dataSources: map[*server.Resource]*clipSource{},
330 primarySources: map[*server.Resource]*clipSource{},337 primarySources: map[*server.Resource]*clipSource{},
331 controlSources: map[*server.Resource]*clipSource{},338 controlSources: map[*server.Resource]*clipSource{},
332- contentNotify: make(chan struct{}),
333 regions: map[*server.Resource]*region{},339 regions: map[*server.Resource]*region{},
334 relatives: map[server.Client][]*relativepointer.ZwpRelativePointerV1{},340 relatives: map[server.Client][]*relativepointer.ZwpRelativePointerV1{},
335 gestures: map[server.Client][]*gestureObject{},341 gestures: map[server.Client][]*gestureObject{},
@@ -369,7 +375,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
369 var node *renderNode375 var node *renderNode
370 if opts.syncDev != nil {376 if opts.syncDev != nil {
371 s.syncDev = opts.syncDev377 s.syncDev = opts.syncDev
372- if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {378+ if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
373 cleanup()379 cleanup()
374 d.Close()380 d.Close()
375 return nil, err381 return nil, err
@@ -378,7 +384,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
378 if node, err = openSyncobj(opts.SyncobjNode); err != nil {384 if node, err = openSyncobj(opts.SyncobjNode); err != nil {
379 log.Info().Err(err).Msg("explicit sync off")385 log.Info().Err(err).Msg("explicit sync off")
380 node = nil386 node = nil
381- } else if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {387+ } else if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
382 node.close()388 node.close()
383 node = nil389 node = nil
384 } else {390 } else {
@@ -387,7 +393,7 @@ func New(opts Options, ch Channels, log zerowrap.Logger) (*Server, error) {
387 }393 }
388 // Implicit dma-buf fences also need this waiter when explicit sync is off.394 // Implicit dma-buf fences also need this waiter when explicit sync is off.
389 if s.syncWait == nil {395 if s.syncWait == nil {
390- if s.syncWait, err = newSyncWaiter(s.wakePacer); err != nil {396+ if s.syncWait, err = newSyncWaiter(s.signalPacer); err != nil {
391 cleanup()397 cleanup()
392 d.Close()398 d.Close()
393 return nil, err399 return nil, err
@@ -453,6 +459,7 @@ func (s *Server) Run(ctx context.Context) error {
453 s.log.Info().Str("socket", s.name).Msg("starting wayland")459 s.log.Info().Str("socket", s.name).Msg("starting wayland")
454 s.started = time.Now()460 s.started = time.Now()
455 s.ctx = ctx461 s.ctx = ctx
462+ s.paceTimer = time.NewTimer(defaultFramePeriod)
456 var wg sync.WaitGroup463 var wg sync.WaitGroup
457 wg.Add(15)464 wg.Add(15)
458 go func() { defer wg.Done(); s.forwardBackendSecurity(ctx) }()465 go func() { defer wg.Done(); s.forwardBackendSecurity(ctx) }()
@@ -563,6 +570,10 @@ func (s *Server) forwardOutputApplied(ctx context.Context) {
563 // enough for a renderer target a few frames old.570 // enough for a renderer target a few frames old.
564 const damageHistory = 4571 const damageHistory = 4
565 572
573+// damageHistoryCap is the capacity of a history's backing array: one
574+// allocation serves damageHistoryCap-damageHistory+1 publications.
575+const damageHistoryCap = 8 * damageHistory
576+
566 // forwardOutputFormats applies the outputs' scanout formats on the577 // forwardOutputFormats applies the outputs' scanout formats on the
567 // display goroutine.578 // display goroutine.
568 func (s *Server) forwardOutputFormats(ctx context.Context) {579 func (s *Server) forwardOutputFormats(ctx context.Context) {
@@ -632,23 +643,28 @@ func (s *Server) emitContent(c ports.SurfaceContent, d damage) {
632 s.contentMu.Lock()643 s.contentMu.Lock()
633 s.contentSeq[c.ID]++644 s.contentSeq[c.ID]++
634 c.Seq = s.contentSeq[c.ID]645 c.Seq = s.contentSeq[c.ID]
635- // Copy-on-write: previous publications can still be in an output's hands.646+ // Published histories are immutable: an output may hold one for as long
647+ // as it likes. The history grows inside one backing array, appended past
648+ // the end of every published slice (never over an element a reader can
649+ // see), and moves to a fresh array only when that one is full. Readers
650+ // get a capacity-clipped slice, so they cannot append into the spare.
636 old := s.damage[c.ID]651 old := s.damage[c.ID]
637 if len(old) >= damageHistory {652 if len(old) >= damageHistory {
638 old = old[len(old)-damageHistory+1:]653 old = old[len(old)-damageHistory+1:]
639 }654 }
640- h := make([]ports.SeqDamage, len(old)+1)655+ if len(old) == cap(old) {
641- copy(h, old)656+ grown := make([]ports.SeqDamage, len(old), damageHistoryCap)
642- h[len(old)] = ports.SeqDamage{Seq: c.Seq, Full: d.full, Rects: d.rects}657+ copy(grown, old)
658+ old = grown
659+ }
660+ h := append(old, ports.SeqDamage{Seq: c.Seq, Full: d.full, Rects: d.rects})
643 if c.Empty() {661 if c.Empty() {
644 delete(s.damage, c.ID)662 delete(s.damage, c.ID)
645 } else {663 } else {
646 s.damage[c.ID] = h664 s.damage[c.ID] = h
647 }665 }
648- c.DamageHistory = h666+ c.DamageHistory = h[:len(h):len(h)]
649 s.contents[c.ID] = c667 s.contents[c.ID] = c
650- close(s.contentNotify)
651- s.contentNotify = make(chan struct{})
652 s.contentMu.Unlock()668 s.contentMu.Unlock()
653 select {669 select {
654 case s.contentReady <- struct{}{}:670 case s.contentReady <- struct{}{}:
@@ -663,13 +679,11 @@ func (s *Server) forwardContents(ctx context.Context) {
663 var c ports.SurfaceContent679 var c ports.SurfaceContent
664 found := false680 found := false
665 var seq uint64681 var seq uint64
666- var notify <-chan struct{}
667 for id, c = range s.contents {682 for id, c = range s.contents {
668 found = true683 found = true
669 seq = s.contentSeq[id]684 seq = s.contentSeq[id]
670 break685 break
671 }686 }
672- notify = s.contentNotify
673 s.contentMu.Unlock()687 s.contentMu.Unlock()
674 if !found {688 if !found {
675 select {689 select {
@@ -679,11 +693,13 @@ func (s *Server) forwardContents(ctx context.Context) {
679 continue693 continue
680 }694 }
681 }695 }
682- // Do not hold a stale frame while the receiver is busy: wake on newer data.696+ // Do not hold a stale frame while the receiver is busy: wake on newer
697+ // data. contentReady is signalled after every publication; a stale
698+ // signal only costs one more loop.
683 select {699 select {
684 case <-ctx.Done():700 case <-ctx.Done():
685 return701 return
686- case <-notify:702+ case <-s.contentReady:
687 continue703 continue
688 case s.channels.Contents <- c:704 case s.channels.Contents <- c:
689 s.contentMu.Lock()705 s.contentMu.Lock()
M internal/adapters/wayland/surface.go +26 -1
@@ -33,6 +33,9 @@ type surface struct {
3333 committedViewport viewportState // value snapshot; no allocation on viewport commits
3434 // next is the state requested since the last commit (see pendingCommit).
3535 next pendingCommit
36+ // syncScratch holds next.sync between takeSyncPoints and queueUpdate,
37+ // which moves it into the update.
38+ syncScratch commitSync
3639 // bufferScale, transform and opaque are committed state. formatOpaque is
3740 // the current buffer's own opacity (x formats).
3841 bufferScale int
@@ -371,8 +374,30 @@ func (s *surface) Commit(*wayland.Surface) {
371374 s.takeSyncPoints()
372375 // Each request creates an ordered content update. A synchronized update
373376 // can only be reached through a parent's committed dependency graph.
377+ s.queueAndTick()
378+}
379+
380+// queueAndTick queues the commit and applies what is ready. The pacer is
381+// woken at most once, after the work, and by a tick on this goroutine
382+// rather than by a wake that makes the pacer call Do back.
383+func (s *surface) queueAndTick() {
384+ srv := s.server
385+ srv.paceDepth++
374386 s.queueUpdate()
375- s.server.tickFifo(time.Now(), nil)
387+ srv.tickFifo(time.Now(), nil)
388+ if n := len(s.queue); n > 0 && !s.queue[n-1].synced {
389+ // The commit waits for a barrier, a fence or a refresh: the pacer
390+ // must time it. A commit applied at once needs no tick of its own;
391+ // its callbacks, barrier and held buffers wake the pacer where
392+ // they are queued. A synchronized commit waits for its parent's
393+ // commit, which ticks for it.
394+ srv.wakePacer()
395+ }
396+ srv.paceDepth--
397+ if srv.paceDepth == 0 && srv.paceDirty {
398+ srv.paceDirty = false
399+ srv.paceNow()
400+ }
376401 }
377402
378403 // applyCommit makes only the captured update current. Live pending requests
@@ -33,6 +33,9 @@ type surface struct {
33 committedViewport viewportState // value snapshot; no allocation on viewport commits33 committedViewport viewportState // value snapshot; no allocation on viewport commits
34 // next is the state requested since the last commit (see pendingCommit).34 // next is the state requested since the last commit (see pendingCommit).
35 next pendingCommit35 next pendingCommit
36+ // syncScratch holds next.sync between takeSyncPoints and queueUpdate,
37+ // which moves it into the update.
38+ syncScratch commitSync
36 // bufferScale, transform and opaque are committed state. formatOpaque is39 // bufferScale, transform and opaque are committed state. formatOpaque is
37 // the current buffer's own opacity (x formats).40 // the current buffer's own opacity (x formats).
38 bufferScale int41 bufferScale int
@@ -371,8 +374,30 @@ func (s *surface) Commit(*wayland.Surface) {
371 s.takeSyncPoints()374 s.takeSyncPoints()
372 // Each request creates an ordered content update. A synchronized update375 // Each request creates an ordered content update. A synchronized update
373 // can only be reached through a parent's committed dependency graph.376 // can only be reached through a parent's committed dependency graph.
377+ s.queueAndTick()
378+}
379+
380+// queueAndTick queues the commit and applies what is ready. The pacer is
381+// woken at most once, after the work, and by a tick on this goroutine
382+// rather than by a wake that makes the pacer call Do back.
383+func (s *surface) queueAndTick() {
384+ srv := s.server
385+ srv.paceDepth++
374 s.queueUpdate()386 s.queueUpdate()
375- s.server.tickFifo(time.Now(), nil)387+ srv.tickFifo(time.Now(), nil)
388+ if n := len(s.queue); n > 0 && !s.queue[n-1].synced {
389+ // The commit waits for a barrier, a fence or a refresh: the pacer
390+ // must time it. A commit applied at once needs no tick of its own;
391+ // its callbacks, barrier and held buffers wake the pacer where
392+ // they are queued. A synchronized commit waits for its parent's
393+ // commit, which ticks for it.
394+ srv.wakePacer()
395+ }
396+ srv.paceDepth--
397+ if srv.paceDepth == 0 && srv.paceDirty {
398+ srv.paceDirty = false
399+ srv.paceNow()
400+ }
376 }401 }
377 402
378 // applyCommit makes only the captured update current. Live pending requests403 // applyCommit makes only the captured update current. Live pending requests
M internal/adapters/wayland/syncobj.go +4 -1
@@ -487,7 +487,10 @@ func (s *surface) takeSyncPoints() {
487487 if st == nil || !st.acquire.set() {
488488 return
489489 }
490- cs := &commitSync{acquire: st.acquire, release: st.release}
490+ // The commit's points wait in the surface until queueUpdate moves them
491+ // into the update: no allocation per commit.
492+ cs := &s.syncScratch
493+ *cs = commitSync{acquire: st.acquire, release: st.release}
491494 st.acquire, st.release = syncPoint{}, syncPoint{}
492495 cs.acquire.tl.uses++
493496 cs.release.tl.uses++
@@ -487,7 +487,10 @@ func (s *surface) takeSyncPoints() {
487 if st == nil || !st.acquire.set() {487 if st == nil || !st.acquire.set() {
488 return488 return
489 }489 }
490- cs := &commitSync{acquire: st.acquire, release: st.release}490+ // The commit's points wait in the surface until queueUpdate moves them
491+ // into the update: no allocation per commit.
492+ cs := &s.syncScratch
493+ *cs = commitSync{acquire: st.acquire, release: st.release}
491 st.acquire, st.release = syncPoint{}, syncPoint{}494 st.acquire, st.release = syncPoint{}, syncPoint{}
492 cs.acquire.tl.uses++495 cs.acquire.tl.uses++
493 cs.release.tl.uses++496 cs.release.tl.uses++
M internal/app/drm.go +34 -0
@@ -19,6 +19,7 @@ import (
1919 "github.com/bnema/neferwl/internal/adapters/logging"
2020 "github.com/bnema/neferwl/internal/adapters/sched"
2121 "github.com/bnema/neferwl/internal/adapters/seat"
22+ "github.com/bnema/neferwl/internal/adapters/watchdog"
2223 "github.com/bnema/neferwl/internal/ports"
2324 "github.com/bnema/zerowrap"
2425 )
@@ -119,6 +120,35 @@ type drmBackend struct {
119120 // reads them while the output loop adds and removes them.
120121 mu sync.Mutex
121122 preparers map[seatDisablePreparer]struct{}
123+ // probes are the running output loops the watchdog pings, by
124+ // connector, guarded by mu.
125+ probes map[string]func()
126+}
127+
128+func (b *drmBackend) addProbe(name string, ping func()) {
129+ b.mu.Lock()
130+ defer b.mu.Unlock()
131+ if b.probes == nil {
132+ b.probes = map[string]func(){}
133+ }
134+ b.probes[name] = ping
135+}
136+
137+func (b *drmBackend) removeProbe(name string) {
138+ b.mu.Lock()
139+ defer b.mu.Unlock()
140+ delete(b.probes, name)
141+}
142+
143+// watchdogProbes lists the running output loops for the watchdog.
144+func (b *drmBackend) watchdogProbes() []watchdog.Probe {
145+ b.mu.Lock()
146+ defer b.mu.Unlock()
147+ out := make([]watchdog.Probe, 0, len(b.probes))
148+ for name, ping := range b.probes {
149+ out = append(out, watchdog.Probe{Name: "output-" + name, Ping: ping})
150+ }
151+ return out
122152 }
123153
124154 // seatDisablePreparer is an output that can reset its planes before the
@@ -490,6 +520,10 @@ func (b *drmBackend) runOutputs(ctx context.Context, want func(ports.Config) drm
490520 defer o.Close()
491521 b.addPreparer(o)
492522 defer b.removePreparer(o) // before Close: it handles no request any more
523+ // octx derives from the session: the probe ends with this
524+ // output or the session.
525+ b.addProbe(name, func() { o.Ping(octx) })
526+ defer b.removeProbe(name)
493527 if realtime {
494528 runtime.LockOSThread()
495529 // Keep the thread locked until this goroutine exits: an RT
@@ -19,6 +19,7 @@ import (
19 "github.com/bnema/neferwl/internal/adapters/logging"19 "github.com/bnema/neferwl/internal/adapters/logging"
20 "github.com/bnema/neferwl/internal/adapters/sched"20 "github.com/bnema/neferwl/internal/adapters/sched"
21 "github.com/bnema/neferwl/internal/adapters/seat"21 "github.com/bnema/neferwl/internal/adapters/seat"
22+ "github.com/bnema/neferwl/internal/adapters/watchdog"
22 "github.com/bnema/neferwl/internal/ports"23 "github.com/bnema/neferwl/internal/ports"
23 "github.com/bnema/zerowrap"24 "github.com/bnema/zerowrap"
24 )25 )
@@ -119,6 +120,35 @@ type drmBackend struct {
119 // reads them while the output loop adds and removes them.120 // reads them while the output loop adds and removes them.
120 mu sync.Mutex121 mu sync.Mutex
121 preparers map[seatDisablePreparer]struct{}122 preparers map[seatDisablePreparer]struct{}
123+ // probes are the running output loops the watchdog pings, by
124+ // connector, guarded by mu.
125+ probes map[string]func()
126+}
127+
128+func (b *drmBackend) addProbe(name string, ping func()) {
129+ b.mu.Lock()
130+ defer b.mu.Unlock()
131+ if b.probes == nil {
132+ b.probes = map[string]func(){}
133+ }
134+ b.probes[name] = ping
135+}
136+
137+func (b *drmBackend) removeProbe(name string) {
138+ b.mu.Lock()
139+ defer b.mu.Unlock()
140+ delete(b.probes, name)
141+}
142+
143+// watchdogProbes lists the running output loops for the watchdog.
144+func (b *drmBackend) watchdogProbes() []watchdog.Probe {
145+ b.mu.Lock()
146+ defer b.mu.Unlock()
147+ out := make([]watchdog.Probe, 0, len(b.probes))
148+ for name, ping := range b.probes {
149+ out = append(out, watchdog.Probe{Name: "output-" + name, Ping: ping})
150+ }
151+ return out
122 }152 }
123 153
124 // seatDisablePreparer is an output that can reset its planes before the154 // seatDisablePreparer is an output that can reset its planes before the
@@ -490,6 +520,10 @@ func (b *drmBackend) runOutputs(ctx context.Context, want func(ports.Config) drm
490 defer o.Close()520 defer o.Close()
491 b.addPreparer(o)521 b.addPreparer(o)
492 defer b.removePreparer(o) // before Close: it handles no request any more522 defer b.removePreparer(o) // before Close: it handles no request any more
523+ // octx derives from the session: the probe ends with this
524+ // output or the session.
525+ b.addProbe(name, func() { o.Ping(octx) })
526+ defer b.removeProbe(name)
493 if realtime {527 if realtime {
494 runtime.LockOSThread()528 runtime.LockOSThread()
495 // Keep the thread locked until this goroutine exits: an RT529 // Keep the thread locked until this goroutine exits: an RT
M internal/app/session.go +17 -0
@@ -23,6 +23,7 @@ import (
2323 "github.com/bnema/neferwl/internal/adapters/sessionsecurity"
2424 "github.com/bnema/neferwl/internal/adapters/statefile"
2525 "github.com/bnema/neferwl/internal/adapters/vulkan"
26+ "github.com/bnema/neferwl/internal/adapters/watchdog"
2627 "github.com/bnema/neferwl/internal/adapters/wayland"
2728 "github.com/bnema/neferwl/internal/adapters/workspaceid"
2829 "github.com/bnema/neferwl/internal/adapters/xkb"
@@ -302,6 +303,22 @@ func (s *session) run(ctx context.Context, inject func(chan<- ports.InputEvent))
302303 close(script)
303304 }
304305 go func() { defer workers.Done(); done <- server.Run(ctx) }()
306+ if dir := logging.Dir(ctx); dir != "" {
307+ workers.Add(1)
308+ go func() {
309+ defer workers.Done()
310+ watchdog.Run(ctx, func() []watchdog.Probe {
311+ probes := []watchdog.Probe{
312+ {Name: "wayland", Ping: func() { server.Ping(ctx) }},
313+ {Name: "core", Ping: func() { c.Ping(ctx) }},
314+ }
315+ if hw != nil {
316+ probes = append(probes, hw.watchdogProbes()...)
317+ }
318+ return probes
319+ }, watchdog.Options{Dir: dir, Every: time.Second, Timeout: 3 * time.Second, Log: logging.For(ctx, "app")})
320+ }()
321+ }
305322 go func() {
306323 defer workers.Done()
307324 // Without a watch the allowlist stays as loaded: restart to apply.
@@ -23,6 +23,7 @@ import (
23 "github.com/bnema/neferwl/internal/adapters/sessionsecurity"23 "github.com/bnema/neferwl/internal/adapters/sessionsecurity"
24 "github.com/bnema/neferwl/internal/adapters/statefile"24 "github.com/bnema/neferwl/internal/adapters/statefile"
25 "github.com/bnema/neferwl/internal/adapters/vulkan"25 "github.com/bnema/neferwl/internal/adapters/vulkan"
26+ "github.com/bnema/neferwl/internal/adapters/watchdog"
26 "github.com/bnema/neferwl/internal/adapters/wayland"27 "github.com/bnema/neferwl/internal/adapters/wayland"
27 "github.com/bnema/neferwl/internal/adapters/workspaceid"28 "github.com/bnema/neferwl/internal/adapters/workspaceid"
28 "github.com/bnema/neferwl/internal/adapters/xkb"29 "github.com/bnema/neferwl/internal/adapters/xkb"
@@ -302,6 +303,22 @@ func (s *session) run(ctx context.Context, inject func(chan<- ports.InputEvent))
302 close(script)303 close(script)
303 }304 }
304 go func() { defer workers.Done(); done <- server.Run(ctx) }()305 go func() { defer workers.Done(); done <- server.Run(ctx) }()
306+ if dir := logging.Dir(ctx); dir != "" {
307+ workers.Add(1)
308+ go func() {
309+ defer workers.Done()
310+ watchdog.Run(ctx, func() []watchdog.Probe {
311+ probes := []watchdog.Probe{
312+ {Name: "wayland", Ping: func() { server.Ping(ctx) }},
313+ {Name: "core", Ping: func() { c.Ping(ctx) }},
314+ }
315+ if hw != nil {
316+ probes = append(probes, hw.watchdogProbes()...)
317+ }
318+ return probes
319+ }, watchdog.Options{Dir: dir, Every: time.Second, Timeout: 3 * time.Second, Log: logging.For(ctx, "app")})
320+ }()
321+ }
305 go func() {322 go func() {
306 defer workers.Done()323 defer workers.Done()
307 // Without a watch the allowlist stays as loaded: restart to apply.324 // Without a watch the allowlist stays as loaded: restart to apply.
M internal/cli/run.go +2 -2
@@ -123,12 +123,12 @@ func (f *runFlags) run(ctx context.Context, sizes [][2]int) error {
123123 if err != nil {
124124 return err
125125 }
126- defer closeLog()
126+ defer closeLog(false) // no-op after the panic path closed it
127127 log := logging.For(ctx, "app")
128128 defer func() {
129129 if p := recover(); p != nil {
130130 log.Error().Interface("panic", p).Str("stack", string(debug.Stack())).Msg("exit")
131- _ = closeLog()
131+ _ = closeLog(true) // keep the crash file for the re-panic
132132 panic(p)
133133 }
134134 }()
@@ -123,12 +123,12 @@ func (f *runFlags) run(ctx context.Context, sizes [][2]int) error {
123 if err != nil {123 if err != nil {
124 return err124 return err
125 }125 }
126- defer closeLog()126+ defer closeLog(false) // no-op after the panic path closed it
127 log := logging.For(ctx, "app")127 log := logging.For(ctx, "app")
128 defer func() {128 defer func() {
129 if p := recover(); p != nil {129 if p := recover(); p != nil {
130 log.Error().Interface("panic", p).Str("stack", string(debug.Stack())).Msg("exit")130 log.Error().Interface("panic", p).Str("stack", string(debug.Stack())).Msg("exit")
131- _ = closeLog()131+ _ = closeLog(true) // keep the crash file for the re-panic
132 panic(p)132 panic(p)
133 }133 }
134 }()134 }()
M internal/core/core.go +22 -2
@@ -138,7 +138,9 @@ type Core struct {
138138 swipe *swipeGesture
139139 // gesture is the touchpad gesture forwarded to a client
140140 // (clientgesture.go).
141- gesture clientGesture
141+ gesture clientGesture
142+ // ping carries watchdog probes (Ping); Run answers each one.
143+ ping chan chan struct{}
142144 frameC <-chan time.Time
143145 frameTimer ports.Timer
144146 // shots is the snapshot of what the screens show before an action
@@ -401,7 +403,7 @@ func New(cfg ports.Config, ch Channels, opts Options) (*Core, error) {
401403 return nil, fmt.Errorf("scenes, layouts, constraints, state and workspaces must have capacity 1")
402404 }
403405 // A placeholder screen holds windows until the first output arrives.
404- c := &Core{inputKeys: map[string]bool{}, slots: map[slotKey]*slotState{}, placement: newSpawnPlacement(), windows: newWindowRegistry(), popups: map[WindowID]*popupState{}, ch: ch, opts: opts, pressed: map[string]bool{}, buttons: map[uint32]bool{}, swallow: map[uint32]bool{}, configures: newConfigures()}
406+ c := &Core{inputKeys: map[string]bool{}, slots: map[slotKey]*slotState{}, placement: newSpawnPlacement(), windows: newWindowRegistry(), popups: map[WindowID]*popupState{}, ch: ch, opts: opts, pressed: map[string]bool{}, buttons: map[uint32]bool{}, swallow: map[uint32]bool{}, configures: newConfigures(), ping: make(chan chan struct{})}
405407 c.screens = []*screen{{mon: newMonitorWithIDs("", "", &c.nextWorkspaceID), scale: 1, cfgScale: 1}}
406408 if err := c.apply(cfg); err != nil {
407409 return nil, err
@@ -1023,6 +1025,21 @@ func (c *Core) hit(x, y float64) (WindowID, float64, float64) {
10231025 return id, sx, sy
10241026 }
10251027
1028+// Ping returns once Run handled a probe, or once ctx ends: the watchdog
1029+// uses it to detect a stuck owner loop.
1030+func (c *Core) Ping(ctx context.Context) {
1031+ ack := make(chan struct{})
1032+ select {
1033+ case c.ping <- ack:
1034+ case <-ctx.Done():
1035+ return
1036+ }
1037+ select {
1038+ case <-ack:
1039+ case <-ctx.Done():
1040+ }
1041+}
1042+
10261043 func (c *Core) Run(ctx context.Context) error {
10271044 c.syncSecurity()
10281045 defer c.stopFrame()
@@ -1084,6 +1101,9 @@ func (c *Core) Run(ctx context.Context) error {
10841101 return nil
10851102 }
10861103 continue
1104+ case ack := <-c.ping:
1105+ close(ack)
1106+ continue
10871107 case <-c.frameC:
10881108 if c.securityCheckpoint(ctx) != nil {
10891109 return nil
@@ -138,7 +138,9 @@ type Core struct {
138 swipe *swipeGesture138 swipe *swipeGesture
139 // gesture is the touchpad gesture forwarded to a client139 // gesture is the touchpad gesture forwarded to a client
140 // (clientgesture.go).140 // (clientgesture.go).
141- gesture clientGesture141+ gesture clientGesture
142+ // ping carries watchdog probes (Ping); Run answers each one.
143+ ping chan chan struct{}
142 frameC <-chan time.Time144 frameC <-chan time.Time
143 frameTimer ports.Timer145 frameTimer ports.Timer
144 // shots is the snapshot of what the screens show before an action146 // shots is the snapshot of what the screens show before an action
@@ -401,7 +403,7 @@ func New(cfg ports.Config, ch Channels, opts Options) (*Core, error) {
401 return nil, fmt.Errorf("scenes, layouts, constraints, state and workspaces must have capacity 1")403 return nil, fmt.Errorf("scenes, layouts, constraints, state and workspaces must have capacity 1")
402 }404 }
403 // A placeholder screen holds windows until the first output arrives.405 // A placeholder screen holds windows until the first output arrives.
404- c := &Core{inputKeys: map[string]bool{}, slots: map[slotKey]*slotState{}, placement: newSpawnPlacement(), windows: newWindowRegistry(), popups: map[WindowID]*popupState{}, ch: ch, opts: opts, pressed: map[string]bool{}, buttons: map[uint32]bool{}, swallow: map[uint32]bool{}, configures: newConfigures()}406+ c := &Core{inputKeys: map[string]bool{}, slots: map[slotKey]*slotState{}, placement: newSpawnPlacement(), windows: newWindowRegistry(), popups: map[WindowID]*popupState{}, ch: ch, opts: opts, pressed: map[string]bool{}, buttons: map[uint32]bool{}, swallow: map[uint32]bool{}, configures: newConfigures(), ping: make(chan chan struct{})}
405 c.screens = []*screen{{mon: newMonitorWithIDs("", "", &c.nextWorkspaceID), scale: 1, cfgScale: 1}}407 c.screens = []*screen{{mon: newMonitorWithIDs("", "", &c.nextWorkspaceID), scale: 1, cfgScale: 1}}
406 if err := c.apply(cfg); err != nil {408 if err := c.apply(cfg); err != nil {
407 return nil, err409 return nil, err
@@ -1023,6 +1025,21 @@ func (c *Core) hit(x, y float64) (WindowID, float64, float64) {
1023 return id, sx, sy1025 return id, sx, sy
1024 }1026 }
1025 1027
1028+// Ping returns once Run handled a probe, or once ctx ends: the watchdog
1029+// uses it to detect a stuck owner loop.
1030+func (c *Core) Ping(ctx context.Context) {
1031+ ack := make(chan struct{})
1032+ select {
1033+ case c.ping <- ack:
1034+ case <-ctx.Done():
1035+ return
1036+ }
1037+ select {
1038+ case <-ack:
1039+ case <-ctx.Done():
1040+ }
1041+}
1042+
1026 func (c *Core) Run(ctx context.Context) error {1043 func (c *Core) Run(ctx context.Context) error {
1027 c.syncSecurity()1044 c.syncSecurity()
1028 defer c.stopFrame()1045 defer c.stopFrame()
@@ -1084,6 +1101,9 @@ func (c *Core) Run(ctx context.Context) error {
1084 return nil1101 return nil
1085 }1102 }
1086 continue1103 continue
1104+ case ack := <-c.ping:
1105+ close(ack)
1106+ continue
1087 case <-c.frameC:1107 case <-c.frameC:
1088 if c.securityCheckpoint(ctx) != nil {1108 if c.securityCheckpoint(ctx) != nil {
1089 return nil1109 return nil

Keyboard shortcuts

Single-key shortcuts are on. ctrl k and esc work whatever this says.

ctrl k :
go to a section, branch, tag, file or repository
?
this list
j k
move the row cursor down / up
enter
open the row under the cursor; on Files, a directory opens or closes in place
↑ ↓ ← → home end
in the file tree (focus on a row): previous / next row, close / open a directory or go to its parent / first entry, first / last row. They never open a file
l h
on Files: open a directory or a file / close a directory or go up to its parent, then to the parent page. Elsewhere: into the side panel (on a merge request, its files) / back to the content; there j k move between its controls
g g G
first / last row
/
filter the list; on Files, find a file anywhere in the repository (esc clears, enter opens the best match)
b
switch branch or tag
g s g f g c g b g m g i g r g a
tabs: summary, files, commits, branches & tags, merge requests, issues, releases, actions (when available)
g h
home
w
wide mode: use the whole width of the screen, or not (remembered)
y
copy a permalink pinned to the commit
d u x
diff: split / unified / collapse or expand all
e E
diff: the whole file in view / every file, with the lines the hunks hide
n
merge request: next note in the diff
esc
close a dialog, leave a field

Branches 23

Tags 10

↑↓ moveenter openesc close