560aaa9 perf: remove FFI reflection and per-frame allocations from the hot path (#31) · neferwl

M Makefile +1 -1


@@ -50,7 +50,7 @@ arch:
50	50		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...
51	51		# Fast, mandatory guards for tiled content publications and SHM copies.
52	52		perf-check:
53		-	CGO_ENABLED=0 go test ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
	53	+	CGO_ENABLED=0 go test ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestReportSeenAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
54	54		check: vet test arch fakes-check perf-check
55	55
56	56		# Arch package of the committed HEAD (packaging/arch/PKGBUILD). Go modules


@@ -50,7 +50,7 @@ arch:
50		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...	50		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...
51		# Fast, mandatory guards for tiled content publications and SHM copies.	51		# Fast, mandatory guards for tiled content publications and SHM copies.
52		perf-check:	52		perf-check:
53	-	CGO_ENABLED=0 go test ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1	53	+	CGO_ENABLED=0 go test ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestReportSeenAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
54		check: vet test arch fakes-check perf-check	54		check: vet test arch fakes-check perf-check
55			55
56		# Arch package of the committed HEAD (packaging/arch/PKGBUILD). Go modules	56		# Arch package of the committed HEAD (packaging/arch/PKGBUILD). Go modules

M README.md +1 -0


@@ -98,6 +98,7 @@ The file has one `key = value` per line, and `#` starts a comment. A missing fil
98	98		- [Configuration](docs/config.md): every key, action and default bind, and HDR.
99	99		- [Desktop integration](docs/desktop.md): running commands, clipboard, X11 apps, idle and screen off, bars, and the state file for scripts.
100	100		- [Headless mode](docs/headless.md): run without a screen, take screenshots and play input scripts.
	101	+	- [Performance](docs/performance.md): how to profile, the allocation guards, and reference numbers.
101	102
102	103		## Why not Rust?
103	104


@@ -98,6 +98,7 @@ The file has one `key = value` per line, and `#` starts a comment. A missing fil
98	- [Configuration](docs/config.md): every key, action and default bind, and HDR.	98		- [Configuration](docs/config.md): every key, action and default bind, and HDR.
99	- [Desktop integration](docs/desktop.md): running commands, clipboard, X11 apps, idle and screen off, bars, and the state file for scripts.	99		- [Desktop integration](docs/desktop.md): running commands, clipboard, X11 apps, idle and screen off, bars, and the state file for scripts.
100	- [Headless mode](docs/headless.md): run without a screen, take screenshots and play input scripts.	100		- [Headless mode](docs/headless.md): run without a screen, take screenshots and play input scripts.
		101	+	- [Performance](docs/performance.md): how to profile, the allocation guards, and reference numbers.
101		102
102	## Why not Rust?	103		## Why not Rust?
103		104

M cmd/neferwl/main.go +40 -8


@@ -6,6 +6,9 @@ import (
6	6		"errors"
7	7		"flag"
8	8		"fmt"
	9	+	"net"
	10	+	"net/http"
	11	+	"net/http/pprof"
9	12		"os"
10	13		"os/signal"
11	14		"runtime/debug"

@@ -31,19 +34,40 @@ func main() {
31	34		os.Exit(runCode())
32	35		}
33	36
34		-	// tuneGC keeps the heap close to what neferwl uses: a compositor runs for
35		-	// days and its per-frame garbage is small, so the default GOGC=100 (heap
36		-	// grows to twice the live data) mostly holds freed memory. A soft limit
37		-	// makes the collector work harder before the process gets large. GOGC and
38		-	// GOMEMLIMIT set in the environment win.
	37	+	// tuneGC bounds the heap of a process that runs for days. The Go live heap
	38	+	// is small (tens of MB; buffers live in GPU and shared memory), so the
	39	+	// default GOGC=100 is kept: on the tiled playback test, GOGC=50 ran 13
	40	+	// collections for 91 ms of GC CPU where GOGC=100 ran 2 for 3 ms, for 2 MB
	41	+	// more heap. A soft limit makes the collector work harder before the
	42	+	// process gets large. GOGC and GOMEMLIMIT set in the environment win.
39	43		func tuneGC() {
40		-	if os.Getenv("GOGC") == "" {
41		-	debug.SetGCPercent(50)
42		-	}
43	44		if os.Getenv("GOMEMLIMIT") == "" {
44	45		debug.SetMemoryLimit(256 << 20)
45	46		}
46	47		}
	48	+
	49	+	// servePprof exposes the runtime profiles under /debug/pprof/ until ctx is
	50	+	// done. It binds before returning so a bad address fails at startup.
	51	+	func servePprof(ctx context.Context, addr string) error {
	52	+	ln, err := net.Listen("tcp", addr)
	53	+	if err != nil {
	54	+	return fmt.Errorf("--pprof: %w", err)
	55	+	}
	56	+	mux := http.NewServeMux()
	57	+	mux.HandleFunc("/debug/pprof/", pprof.Index)
	58	+	mux.HandleFunc("/debug/pprof/cmdline", pprof.Cmdline)
	59	+	mux.HandleFunc("/debug/pprof/profile", pprof.Profile)
	60	+	mux.HandleFunc("/debug/pprof/symbol", pprof.Symbol)
	61	+	mux.HandleFunc("/debug/pprof/trace", pprof.Trace)
	62	+	srv := &http.Server{Handler: mux, ReadHeaderTimeout: 5 * time.Second}
	63	+	go func() { _ = srv.Serve(ln) }()
	64	+	go func() {
	65	+	<-ctx.Done()
	66	+	_ = srv.Close()
	67	+	}()
	68	+	return nil
	69	+	}
	70	+
47	71		func runCode() int {
48	72		err := run()
49	73		if err == nil \|\| errors.Is(err, flag.ErrHelp) \|\| errors.Is(err, context.Canceled) {

@@ -123,6 +147,7 @@ func run() error {
123	147		debugFlag := flags.String("debug", "", "debug components (comma-separated or all; input-motion logs every pointer motion)")
124	148		configFlag := flags.String("config", "", "config path (empty uses XDG default)")
125	149		session := flags.Bool("session", false, "run as the login session: share WAYLAND_DISPLAY and DISPLAY with systemd and D-Bus user services; notify systemd when ready")
	150	+	pprofAddr := flags.String("pprof", "", "serve Go runtime profiles (heap, CPU, trace) on this address, e.g. localhost:6060")
126	151		if err := flags.Parse(os.Args[1:]); err != nil {
127	152		if errors.Is(err, flag.ErrHelp) {
128	153		return err

@@ -224,6 +249,13 @@ func run() error {
224	249		panic(p)
225	250		}
226	251		}()
	252	+	if *pprofAddr != "" {
	253	+	if err := servePprof(ctx, *pprofAddr); err != nil {
	254	+	fmt.Fprintln(os.Stderr, err)
	255	+	return err
	256	+	}
	257	+	log.Info().Str("addr", *pprofAddr).Msg("pprof")
	258	+	}
227	259		start := time.Now()
228	260		log.Info().Strs("args", os.Args[1:]).Str("backend", *backend).Str("tty", os.Getenv("XDG_VTNR")).Str("session_type", os.Getenv("XDG_SESSION_TYPE")).Msg("run")
229	261		configLog := logging.For(ctx, "config")

 	"errors"
 	"flag"
 	"fmt"
+	"net"
+	"net/http"
+	"net/http/pprof"
 	"os"
 	"os/signal"
 	"runtime/debug"
 	os.Exit(runCode())
 }
-// tuneGC keeps the heap close to what neferwl uses: a compositor runs for
+// tuneGC bounds the heap of a process that runs for days. The Go live heap
-// days and its per-frame garbage is small, so the default GOGC=100 (heap
+// is small (tens of MB; buffers live in GPU and shared memory), so the
-// grows to twice the live data) mostly holds freed memory. A soft limit
+// default GOGC=100 is kept: on the tiled playback test, GOGC=50 ran 13
-// makes the collector work harder before the process gets large. GOGC and
+// collections for 91 ms of GC CPU where GOGC=100 ran 2 for 3 ms, for 2 MB
-// GOMEMLIMIT set in the environment win.
+// more heap. A soft limit makes the collector work harder before the
+// process gets large. GOGC and GOMEMLIMIT set in the environment win.
 func tuneGC() {
-	if os.Getenv("GOGC") == "" {
-		debug.SetGCPercent(50)
-	}
 	if os.Getenv("GOMEMLIMIT") == "" {
 		debug.SetMemoryLimit(256 << 20)
 	}
 }
+// servePprof exposes the runtime profiles under /debug/pprof/ until ctx is
+// done. It binds before returning so a bad address fails at startup.
+func servePprof(ctx context.Context, addr string) error {
+	ln, err := net.Listen("tcp", addr)
+	if err != nil {
+		return fmt.Errorf("--pprof: %w", err)
+	}
+	mux := http.NewServeMux()
+	mux.HandleFunc("/debug/pprof/", pprof.Index)
+	mux.HandleFunc("/debug/pprof/cmdline", pprof.Cmdline)
+	mux.HandleFunc("/debug/pprof/profile", pprof.Profile)
+	mux.HandleFunc("/debug/pprof/symbol", pprof.Symbol)
+	mux.HandleFunc("/debug/pprof/trace", pprof.Trace)
+	srv := &http.Server{Handler: mux, ReadHeaderTimeout: 5 * time.Second}
+	go func() { _ = srv.Serve(ln) }()
+	go func() {
+		<-ctx.Done()
+		_ = srv.Close()
+	}()
+	return nil
+}
 func runCode() int {
 	err := run()
 	if err == nil || errors.Is(err, flag.ErrHelp) || errors.Is(err, context.Canceled) {
 	debugFlag := flags.String("debug", "", "debug components (comma-separated or all; input-motion logs every pointer motion)")
 	configFlag := flags.String("config", "", "config path (empty uses XDG default)")
 	session := flags.Bool("session", false, "run as the login session: share WAYLAND_DISPLAY and DISPLAY with systemd and D-Bus user services; notify systemd when ready")
+	pprofAddr := flags.String("pprof", "", "serve Go runtime profiles (heap, CPU, trace) on this address, e.g. localhost:6060")
 	if err := flags.Parse(os.Args[1:]); err != nil {
 		if errors.Is(err, flag.ErrHelp) {
 			return err
 			panic(p)
 		}
 	}()
+	if *pprofAddr != "" {
+		if err := servePprof(ctx, *pprofAddr); err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			return err
+		}
+		log.Info().Str("addr", *pprofAddr).Msg("pprof")
+	}
 	start := time.Now()
 	log.Info().Strs("args", os.Args[1:]).Str("backend", *backend).Str("tty", os.Getenv("XDG_VTNR")).Str("session_type", os.Getenv("XDG_SESSION_TYPE")).Msg("run")
 	configLog := logging.For(ctx, "config")

A docs/performance.md +62 -0

@@ -0,0 +1,62 @@
	1	+	# Performance
	2	+
	3	+	How NeferWL keeps its CPU and memory use low, how to measure it, and the reference numbers to compare against.
	4	+
	5	+	## Design
	6	+
	7	+	- No reflection on the FFI boundary. libwayland and Vulkan are called through `purego.Syscall6/Syscall15` (fixed arity, zero allocation) and C calls back into Go through `purego.NewCallbackInts`. The Vulkan dispatch tables are generated methods; the libwayland wrapper caches resource id, version and client, and precomputes request argument counts.
	8	+	- No allocation per frame or per request on the steady-state path. Frame callbacks, presentation reports, DRM flips, Vulkan submissions and surface updates reuse owner-goroutine scratch or immutable snapshots. Surface updates (one per `wl_surface.commit`) are recycled once no queued update references them.
	9	+	- Default GOGC. The Go live heap is a few MB (client buffers live in shared and GPU memory), so the runtime default `GOGC=100` with a 256 MiB soft limit gives 2 collections on the tiled playback test where `GOGC=50` gave 13. `GOGC` and `GOMEMLIMIT` from the environment win.
	10	+
	11	+	## Guards
	12	+
	13	+	`make perf-check` runs `testing.AllocsPerRun` guards that fail when a hot path starts allocating: tiled SHM scene walk, tiled commit publication, headless tiled playback, steady-state `Render`, DRM report and flip snapshots, pacer bookkeeping, surface update recycling and input-region traversal. Run it with `make check`.
	14	+
	15	+	## Measuring
	16	+
	17	+	Profile a running compositor on another TTY or over SSH.
	18	+
	19	+	```sh
	20	+	# CPU: where the cycles go (Go and C frames).
	21	+	perf record -g -p (pgrep -f bin/neferwl) -- sleep 8
	22	+	perf report --children --percent-limit 1
	23	+
	24	+	# Allocations: which Go line allocates how many objects.
	25	+	neferwl --backend=drm --pprof=localhost:6060
	26	+	go tool pprof -sample_index=alloc_objects -top http://localhost:6060/debug/pprof/heap
	27	+
	28	+	# GC: one line per collection (count, heap before/after, CPU).
	29	+	GODEBUG=gctrace=1 neferwl --backend=drm
	30	+	```
	31	+
	32	+	`--pprof` also serves `/debug/pprof/profile` (CPU) and `/debug/pprof/trace` (scheduler and GC timeline for `go tool trace`). Escape analysis (`go build -gcflags=-m ./internal/adapters/vulkan`) shows why a value reaches the heap.
	33	+
	34	+	## Reference: 4K HDR tiled video playback
	35	+
	36	+	Workload: a browser playing 4K HDR video with 65 tiled subsurfaces committing every frame, DRM backend, `perf record -g` for 8 s. Percentages are the share of the process's own samples (inclusive).
	37	+
	38	+	\| Metric \| Before (issue #19) \| 2026-09-27 \|
	39	+	\|---\|---\|---\|
	40	+	\| CPU (top) \| 20–30 % \| 13–14 % \|
	41	+	\| GC (`gcBgMarkWorker`) \| 20 % \| 8.7 % \|
	42	+	\| `runtime.mallocgc` \| — \| 3.2 % \|
	43	+	\| Go → C reflection (`purego.RegisterFunc`) \| 46.7 % \| 0 \|
	44	+	\| C → Go reflection (`reflect.Value.Call`) \| ~30 % \| 0 \|
	45	+	\| Request handlers (`server.dispatch`) \| 25.9 % \| 27.6 % (real work) \|
	46	+	\| Vulkan `Render` \| — \| 15.3 % \|
	47	+
	48	+	Component numbers (linux/amd64, Go 1.27, `CGO_ENABLED=0`):
	49	+
	50	+	\| Path \| Before \| Now \|
	51	+	\|---\|---\|---\|
	52	+	\| C → Go callback, 5 args (purego) \| 428 ns, 3 allocs \| 101 ns, 0 \|
	53	+	\| Go → C, 6 args (purego) \| 71 ns, 1 alloc \| 34 ns, 0 \|
	54	+	\| Wayland request dispatch (purego-libwayland) \| 27 ns, 1 alloc \| 5 ns, 0 \|
	55	+	\| Wayland `PostEvent`, 3 args \| 344 ns, 5 allocs \| 41 ns, 0 \|
	56	+	\| Vulkan Reset+Begin+Barrier+End (purego-vulkan) \| 6.4 µs, 22 allocs \| 1.2 µs, 0 \|
	57	+	\| `Render`, tiled scene, per frame \| 1321 allocs \| 10 allocs \|
	58	+	\| Surface update per commit \| 2 allocs \| 0 (recycled) \|
	59	+	\| DRM report / flip `Shows` per flip \| 2 maps \| 0 (immutable snapshots) \|
	60	+	\| GC on tiled playback test \| 13 cycles, 91 ms \| 2 cycles, 3 ms \|
	61	+
	62	+	What remains is compositor work: `surface.Commit` and the fifo logic (~23 %), real C calls (~5 %), the Go scheduler (~8 %).

@@ -0,0 +1,62 @@
	1	+	# Performance
	2	+
	3	+	How NeferWL keeps its CPU and memory use low, how to measure it, and the reference numbers to compare against.
	4	+
	5	+	## Design
	6	+
	7	+	- No reflection on the FFI boundary. libwayland and Vulkan are called through `purego.Syscall6/Syscall15` (fixed arity, zero allocation) and C calls back into Go through `purego.NewCallbackInts`. The Vulkan dispatch tables are generated methods; the libwayland wrapper caches resource id, version and client, and precomputes request argument counts.
	8	+	- No allocation per frame or per request on the steady-state path. Frame callbacks, presentation reports, DRM flips, Vulkan submissions and surface updates reuse owner-goroutine scratch or immutable snapshots. Surface updates (one per `wl_surface.commit`) are recycled once no queued update references them.
	9	+	- Default GOGC. The Go live heap is a few MB (client buffers live in shared and GPU memory), so the runtime default `GOGC=100` with a 256 MiB soft limit gives 2 collections on the tiled playback test where `GOGC=50` gave 13. `GOGC` and `GOMEMLIMIT` from the environment win.
	10	+
	11	+	## Guards
	12	+
	13	+	`make perf-check` runs `testing.AllocsPerRun` guards that fail when a hot path starts allocating: tiled SHM scene walk, tiled commit publication, headless tiled playback, steady-state `Render`, DRM report and flip snapshots, pacer bookkeeping, surface update recycling and input-region traversal. Run it with `make check`.
	14	+
	15	+	## Measuring
	16	+
	17	+	Profile a running compositor on another TTY or over SSH.
	18	+
	19	+	```sh
	20	+	# CPU: where the cycles go (Go and C frames).
	21	+	perf record -g -p (pgrep -f bin/neferwl) -- sleep 8
	22	+	perf report --children --percent-limit 1
	23	+
	24	+	# Allocations: which Go line allocates how many objects.
	25	+	neferwl --backend=drm --pprof=localhost:6060
	26	+	go tool pprof -sample_index=alloc_objects -top http://localhost:6060/debug/pprof/heap
	27	+
	28	+	# GC: one line per collection (count, heap before/after, CPU).
	29	+	GODEBUG=gctrace=1 neferwl --backend=drm
	30	+	```
	31	+
	32	+	`--pprof` also serves `/debug/pprof/profile` (CPU) and `/debug/pprof/trace` (scheduler and GC timeline for `go tool trace`). Escape analysis (`go build -gcflags=-m ./internal/adapters/vulkan`) shows why a value reaches the heap.
	33	+
	34	+	## Reference: 4K HDR tiled video playback
	35	+
	36	+	Workload: a browser playing 4K HDR video with 65 tiled subsurfaces committing every frame, DRM backend, `perf record -g` for 8 s. Percentages are the share of the process's own samples (inclusive).
	37	+
	38	+	\| Metric \| Before (issue #19) \| 2026-09-27 \|
	39	+	\|---\|---\|---\|
	40	+	\| CPU (top) \| 20–30 % \| 13–14 % \|
	41	+	\| GC (`gcBgMarkWorker`) \| 20 % \| 8.7 % \|
	42	+	\| `runtime.mallocgc` \| — \| 3.2 % \|
	43	+	\| Go → C reflection (`purego.RegisterFunc`) \| 46.7 % \| 0 \|
	44	+	\| C → Go reflection (`reflect.Value.Call`) \| ~30 % \| 0 \|
	45	+	\| Request handlers (`server.dispatch`) \| 25.9 % \| 27.6 % (real work) \|
	46	+	\| Vulkan `Render` \| — \| 15.3 % \|
	47	+
	48	+	Component numbers (linux/amd64, Go 1.27, `CGO_ENABLED=0`):
	49	+
	50	+	\| Path \| Before \| Now \|
	51	+	\|---\|---\|---\|
	52	+	\| C → Go callback, 5 args (purego) \| 428 ns, 3 allocs \| 101 ns, 0 \|
	53	+	\| Go → C, 6 args (purego) \| 71 ns, 1 alloc \| 34 ns, 0 \|
	54	+	\| Wayland request dispatch (purego-libwayland) \| 27 ns, 1 alloc \| 5 ns, 0 \|
	55	+	\| Wayland `PostEvent`, 3 args \| 344 ns, 5 allocs \| 41 ns, 0 \|
	56	+	\| Vulkan Reset+Begin+Barrier+End (purego-vulkan) \| 6.4 µs, 22 allocs \| 1.2 µs, 0 \|
	57	+	\| `Render`, tiled scene, per frame \| 1321 allocs \| 10 allocs \|
	58	+	\| Surface update per commit \| 2 allocs \| 0 (recycled) \|
	59	+	\| DRM report / flip `Shows` per flip \| 2 maps \| 0 (immutable snapshots) \|
	60	+	\| GC on tiled playback test \| 13 cycles, 91 ms \| 2 cycles, 3 ms \|
	61	+
	62	+	What remains is compositor work: `surface.Commit` and the fifo logic (~23 %), real C calls (~5 %), the Go scheduler (~8 %).

M go.mod +3 -3


@@ -3,9 +3,9 @@ module github.com/bnema/neferwl
3	3		go 1.27
4	4
5	5		require (
6		-	github.com/bnema/purego v0.12.0-bnema.1
7		-	github.com/bnema/purego-libwayland v0.4.1
8		-	github.com/bnema/purego-vulkan v0.4.2
	6	+	github.com/bnema/purego v0.13.0-bnema.1
	7	+	github.com/bnema/purego-libwayland v0.5.0
	8	+	github.com/bnema/purego-vulkan v0.5.0
9	9		github.com/bnema/wlturbo v0.2.0
10	10		github.com/bnema/zerowrap v1.4.1
11	11		github.com/rs/zerolog v1.35.1


@@ -3,9 +3,9 @@ module github.com/bnema/neferwl
3		go 1.27	3		go 1.27
4			4
5		require (	5		require (
6	-	github.com/bnema/purego v0.12.0-bnema.1	6	+	github.com/bnema/purego v0.13.0-bnema.1
7	-	github.com/bnema/purego-libwayland v0.4.1	7	+	github.com/bnema/purego-libwayland v0.5.0
8	-	github.com/bnema/purego-vulkan v0.4.2	8	+	github.com/bnema/purego-vulkan v0.5.0
9		github.com/bnema/wlturbo v0.2.0	9		github.com/bnema/wlturbo v0.2.0
10		github.com/bnema/zerowrap v1.4.1	10		github.com/bnema/zerowrap v1.4.1
11		github.com/rs/zerolog v1.35.1	11		github.com/rs/zerolog v1.35.1

M go.sum +6 -6

@@ -1,9 +1,9 @@
1		-	github.com/bnema/purego v0.12.0-bnema.1 h1:/jQhDj74geYzFg57t6qDvTc2PbU5uEsM93YgzVXJenI=
2		-	github.com/bnema/purego v0.12.0-bnema.1/go.mod h1:xS1iX3noypxnbOUEUAqOvkfnUzgHuu19VtL4+dfUMJM=
3		-	github.com/bnema/purego-libwayland v0.4.1 h1:D7N0Gu7M5+iL/RPC9ijGBVMKLRYNR+w/X/n/H6iG/cA=
4		-	github.com/bnema/purego-libwayland v0.4.1/go.mod h1:6p9eNaWKaKcnk6lH22sJ9ueMhzv2kirq5P3KY9XGEvc=
5		-	github.com/bnema/purego-vulkan v0.4.2 h1:lQKXcSSnrHtU6GouKjVnDsUs1oituLrBEYpJ5y4NR/E=
6		-	github.com/bnema/purego-vulkan v0.4.2/go.mod h1:jJOFPhqZANTLVxGbJncUQeiXiv/KPtbvLzDwXR/cTic=
	1	+	github.com/bnema/purego v0.13.0-bnema.1 h1:fKRdztnvTmRqAjANMoN+utEIH73dvPqVPJh2iZCCAO8=
	2	+	github.com/bnema/purego v0.13.0-bnema.1/go.mod h1:xS1iX3noypxnbOUEUAqOvkfnUzgHuu19VtL4+dfUMJM=
	3	+	github.com/bnema/purego-libwayland v0.5.0 h1:+mrp3grCkl1URZSna5nC2ru1g3wixpjjD5Wg0HfKl0o=
	4	+	github.com/bnema/purego-libwayland v0.5.0/go.mod h1:+2zIEBX+iNlRB27xrm5Z8Dnd+cmqX+1AYvuwguxK2dg=
	5	+	github.com/bnema/purego-vulkan v0.5.0 h1:n6FzSEo8Ypmc98EFqLXGxQtSAWc4CtsuoCxV4vB6TWA=
	6	+	github.com/bnema/purego-vulkan v0.5.0/go.mod h1:7kFoR1p01JbknuZlfA20oeImryBe8Fx+Zd9hm2uSUM4=
7	7		github.com/bnema/wlturbo v0.2.0 h1:gZPe81TrrrYW8xrbZpEvZnDI7VQBm60+dgAapbXJyRQ=
8	8		github.com/bnema/wlturbo v0.2.0/go.mod h1:4i1k1wyw7FDWOkh2hg1ZwPfrPRn3+E6WrYg9Qrt74CA=
9	9		github.com/bnema/zerowrap v1.4.1 h1:v70w8LDU2+FU+EX/etY9DjVwOAWN0fUKShwZyZHnyLU=

@@ -1,9 +1,9 @@
1	-	github.com/bnema/purego v0.12.0-bnema.1 h1:/jQhDj74geYzFg57t6qDvTc2PbU5uEsM93YgzVXJenI=	1	+	github.com/bnema/purego v0.13.0-bnema.1 h1:fKRdztnvTmRqAjANMoN+utEIH73dvPqVPJh2iZCCAO8=
2	-	github.com/bnema/purego v0.12.0-bnema.1/go.mod h1:xS1iX3noypxnbOUEUAqOvkfnUzgHuu19VtL4+dfUMJM=	2	+	github.com/bnema/purego v0.13.0-bnema.1/go.mod h1:xS1iX3noypxnbOUEUAqOvkfnUzgHuu19VtL4+dfUMJM=
3	-	github.com/bnema/purego-libwayland v0.4.1 h1:D7N0Gu7M5+iL/RPC9ijGBVMKLRYNR+w/X/n/H6iG/cA=	3	+	github.com/bnema/purego-libwayland v0.5.0 h1:+mrp3grCkl1URZSna5nC2ru1g3wixpjjD5Wg0HfKl0o=
4	-	github.com/bnema/purego-libwayland v0.4.1/go.mod h1:6p9eNaWKaKcnk6lH22sJ9ueMhzv2kirq5P3KY9XGEvc=	4	+	github.com/bnema/purego-libwayland v0.5.0/go.mod h1:+2zIEBX+iNlRB27xrm5Z8Dnd+cmqX+1AYvuwguxK2dg=
5	-	github.com/bnema/purego-vulkan v0.4.2 h1:lQKXcSSnrHtU6GouKjVnDsUs1oituLrBEYpJ5y4NR/E=	5	+	github.com/bnema/purego-vulkan v0.5.0 h1:n6FzSEo8Ypmc98EFqLXGxQtSAWc4CtsuoCxV4vB6TWA=
6	-	github.com/bnema/purego-vulkan v0.4.2/go.mod h1:jJOFPhqZANTLVxGbJncUQeiXiv/KPtbvLzDwXR/cTic=	6	+	github.com/bnema/purego-vulkan v0.5.0/go.mod h1:7kFoR1p01JbknuZlfA20oeImryBe8Fx+Zd9hm2uSUM4=
7		github.com/bnema/wlturbo v0.2.0 h1:gZPe81TrrrYW8xrbZpEvZnDI7VQBm60+dgAapbXJyRQ=	7		github.com/bnema/wlturbo v0.2.0 h1:gZPe81TrrrYW8xrbZpEvZnDI7VQBm60+dgAapbXJyRQ=
8		github.com/bnema/wlturbo v0.2.0/go.mod h1:4i1k1wyw7FDWOkh2hg1ZwPfrPRn3+E6WrYg9Qrt74CA=	8		github.com/bnema/wlturbo v0.2.0/go.mod h1:4i1k1wyw7FDWOkh2hg1ZwPfrPRn3+E6WrYg9Qrt74CA=
9		github.com/bnema/zerowrap v1.4.1 h1:v70w8LDU2+FU+EX/etY9DjVwOAWN0fUKShwZyZHnyLU=	9		github.com/bnema/zerowrap v1.4.1 h1:v70w8LDU2+FU+EX/etY9DjVwOAWN0fUKShwZyZHnyLU=

M internal/adapters/drm/output.go +65 -11


@@ -93,6 +93,12 @@ type Output struct {
93	93		shown, queued uint64
94	94		reason string // why the last frame was composed ("" = scanout)
95	95		unsent []ports.OutputPresented
	96	+	// Last immutable snapshot sent in a report. Never mutate it: wayland may
	97	+	// still be reading a prior report on another goroutine.
	98	+	seenSnapshot map[ports.WindowID]uint64
	99	+	showsSnapshot map[ports.WindowID]uint64
	100	+	directShowsSnapshot map[ports.WindowID]uint64
	101	+	showsScratch map[ports.WindowID]uint64
96	102		// Tearing (ADR 006): async commits of scanned-out buffers that ask for
97	103		// them, when the card has DRM_CAP_ATOMIC_ASYNC_PAGE_FLIP. asyncFence
98	104		// is set when an async commit may carry IN_FENCE_FD (probed once).

@@ -967,6 +973,17 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
967	973		}()
968	974		stats := time.NewTicker(10 * time.Second)
969	975		defer stats.Stop()
	976	+	// Timer channels are enabled only when their condition holds. Go 1.27
	977	+	// guarantees a stopped/reset timer cannot deliver its old value.
	978	+	retryTimer := time.NewTimer(time.Hour)
	979	+	vrrTimer := time.NewTimer(time.Hour)
	980	+	stuckTimer := time.NewTimer(time.Hour)
	981	+	retryTimer.Stop()
	982	+	vrrTimer.Stop()
	983	+	stuckTimer.Stop()
	984	+	defer retryTimer.Stop()
	985	+	defer vrrTimer.Stop()
	986	+	defer stuckTimer.Stop()
970	987		// seen is the latest content Seq per window. It is reported only while
971	988		// no frame is in flight: then the GPU has finished every frame that
972	989		// read older buffers (the kernel waited for their fence to flip).

@@ -989,17 +1006,26 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
989	1006		// so is one that waits for an uncommitted frame's fence.
990	1007		var retry <-chan time.Time
991	1008		if len(o.unsent) > 0 \|\| readWait {
992		-	retry = time.After(time.Millisecond)
	1009	+	retryTimer.Reset(time.Millisecond)
	1010	+	retry = retryTimer.C
	1011	+	} else {
	1012	+	retryTimer.Stop()
993	1013		}
994	1014		// A composed screen that stopped changing still drops VRR on time.
995	1015		var vrrOff <-chan time.Time
996	1016		if enabled && o.vrrOn && !o.pending && !o.composedSince.IsZero() {
997		-	vrrOff = time.After(max(0, vrrHold-time.Since(o.composedSince)))
	1017	+	vrrTimer.Reset(max(0, vrrHold-time.Since(o.composedSince)))
	1018	+	vrrOff = vrrTimer.C
	1019	+	} else {
	1020	+	vrrTimer.Stop()
998	1021		}
999	1022		// A commit whose event never comes must not stop the output.
1000	1023		var stuck <-chan time.Time
1001	1024		if enabled && o.pending {
1002		-	stuck = time.After(max(0, time.Until(o.stuckAt)))
	1025	+	stuckTimer.Reset(max(0, time.Until(o.stuckAt)))
	1026	+	stuck = stuckTimer.C
	1027	+	} else {
	1028	+	stuckTimer.Stop()
1003	1029		}
1004	1030		stateDirty := false
1005	1031		select {

@@ -1144,7 +1170,7 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1144	1170		start := time.Now()
1145	1171		// shows is what the frame puts on screen: presentation feedback
1146	1172		// of windows it does not draw is discarded, not presented.
1147		-	f := pendingFrame{shows: shownBy(scene, seen)}
	1173	+	f := pendingFrame{shows: o.shownBy(scene, seen)}
1148	1174		var fb uint32
1149	1175		var c ports.SurfaceContent
1150	1176		forceCompose := len(requests) > 0

@@ -1153,7 +1179,7 @@ func (o *Output) Run(ctx context.Context, newRenderer func(w, h int) (ports.Rend
1153	1179		}
1154	1180		direct := false
1155	1181		if fb != 0 {
1156		-	direct, err = o.commitScanout(fb, c, pendingFrame{shows: map[ports.WindowID]uint64{c.ID: seen[c.ID]}})
	1182	+	direct, err = o.commitScanout(fb, c, pendingFrame{shows: o.directShownBy(c.ID, seen[c.ID])})
1157	1183		}
1158	1184		if !direct {
1159	1185		var ov overlayWin

@@ -1312,7 +1338,10 @@ const maxUnsent = 16
1312	1338		// buffers can be released. A flip-less report merges into the newest
1313	1339		// queued one; flips are never merged unless maxUnsent is reached.
1314	1340		func (o Output) report(flip ports.FlipInfo, seen map[ports.WindowID]uint64) {
1315		-	r := ports.OutputPresented{Output: o.conn.name, Flip: flip, Shown: o.shown, Queued: o.queued, Seen: maps.Clone(seen)}
	1341	+	if !maps.Equal(o.seenSnapshot, seen) {
	1342	+	o.seenSnapshot = maps.Clone(seen)
	1343	+	}
	1344	+	r := ports.OutputPresented{Output: o.conn.name, Flip: flip, Shown: o.shown, Queued: o.queued, Seen: o.seenSnapshot}
1316	1345		if n := len(o.unsent); n > 0 && (flip == nil \|\| o.unsent[n-1].Flip == nil) {
1317	1346		if flip == nil {
1318	1347		r.Flip = o.unsent[n-1].Flip

@@ -1325,8 +1354,18 @@ func (o Output) report(flip ports.FlipInfo, seen map[ports.WindowID]uint64) {
1325	1354		old, next := o.unsent[0].Flip, o.unsent[1].Flip
1326	1355		if next != nil && old != nil {
1327	1356		next.Merged += 1 + old.Merged
	1357	+	// A snapshot can be shared with previously sent flips. Clone only
	1358	+	// when a merge actually needs to add an absent window.
	1359	+	cloned := false
1328	1360		for id, seq := range old.Shows {
1329	1361		if _, ok := next.Shows[id]; !ok {
	1362	+	if !cloned {
	1363	+	next.Shows = maps.Clone(next.Shows)
	1364	+	cloned = true
	1365	+	}
	1366	+	if next.Shows == nil {
	1367	+	next.Shows = make(map[ports.WindowID]uint64)
	1368	+	}
1330	1369		next.Shows[id] = seq
1331	1370		}
1332	1371		}

@@ -1529,13 +1568,28 @@ func (o *Output) freeImages() {
1529	1568		}
1530	1569		}
1531	1570
1532		-	// shownBy is the content Seq of each window the scene draws.
1533		-	func shownBy(s ports.Scene, seen map[ports.WindowID]uint64) map[ports.WindowID]uint64 {
1534		-	out := map[ports.WindowID]uint64{}
	1571	+	// shownBy returns an immutable snapshot of the content Seq of drawn windows.
	1572	+	// The scratch map is private to the output; a snapshot may still be read by
	1573	+	// wayland after a subsequent flip, so never refill a previously sent map.
	1574	+	func (o *Output) shownBy(s ports.Scene, seen map[ports.WindowID]uint64) map[ports.WindowID]uint64 {
	1575	+	if o.showsScratch == nil {
	1576	+	o.showsScratch = make(map[ports.WindowID]uint64)
	1577	+	}
	1578	+	clear(o.showsScratch)
1535	1579		for id, seq := range seen {
1536	1580		if s.Shows(id) {
1537		-	out[id] = seq
	1581	+	o.showsScratch[id] = seq
1538	1582		}
1539	1583		}
1540		-	return out
	1584	+	if !maps.Equal(o.showsSnapshot, o.showsScratch) {
	1585	+	o.showsSnapshot = maps.Clone(o.showsScratch)
	1586	+	}
	1587	+	return o.showsSnapshot
	1588	+	}
	1589	+
	1590	+	func (o *Output) directShownBy(id ports.WindowID, seq uint64) map[ports.WindowID]uint64 {
	1591	+	if len(o.directShowsSnapshot) != 1 \|\| o.directShowsSnapshot[id] != seq {
	1592	+	o.directShowsSnapshot = map[ports.WindowID]uint64{id: seq}
	1593	+	}
	1594	+	return o.directShowsSnapshot
1541	1595		}

 	shown, queued uint64
 	reason        string // why the last frame was composed ("" = scanout)
 	unsent        []ports.OutputPresented
+	// Last immutable snapshot sent in a report. Never mutate it: wayland may
+	// still be reading a prior report on another goroutine.
+	seenSnapshot        map[ports.WindowID]uint64
+	showsSnapshot       map[ports.WindowID]uint64
+	directShowsSnapshot map[ports.WindowID]uint64
+	showsScratch        map[ports.WindowID]uint64
 	// Tearing (ADR 006): async commits of scanned-out buffers that ask for
 	// them, when the card has DRM_CAP_ATOMIC_ASYNC_PAGE_FLIP. asyncFence
 	// is set when an async commit may carry IN_FENCE_FD (probed once).
 	}()
 	stats := time.NewTicker(10 * time.Second)
 	defer stats.Stop()
+	// Timer channels are enabled only when their condition holds. Go 1.27
+	// guarantees a stopped/reset timer cannot deliver its old value.
+	retryTimer := time.NewTimer(time.Hour)
+	vrrTimer := time.NewTimer(time.Hour)
+	stuckTimer := time.NewTimer(time.Hour)
+	retryTimer.Stop()
+	vrrTimer.Stop()
+	stuckTimer.Stop()
+	defer retryTimer.Stop()
+	defer vrrTimer.Stop()
+	defer stuckTimer.Stop()
 	// seen is the latest content Seq per window. It is reported only while
 	// no frame is in flight: then the GPU has finished every frame that
 	// read older buffers (the kernel waited for their fence to flip).
 		// so is one that waits for an uncommitted frame's fence.
 		var retry <-chan time.Time
 		if len(o.unsent) > 0 || readWait {
-			retry = time.After(time.Millisecond)
+			retryTimer.Reset(time.Millisecond)
+			retry = retryTimer.C
+		} else {
+			retryTimer.Stop()
 		}
 		// A composed screen that stopped changing still drops VRR on time.
 		var vrrOff <-chan time.Time
 		if enabled && o.vrrOn && !o.pending && !o.composedSince.IsZero() {
-			vrrOff = time.After(max(0, vrrHold-time.Since(o.composedSince)))
+			vrrTimer.Reset(max(0, vrrHold-time.Since(o.composedSince)))
+			vrrOff = vrrTimer.C
+		} else {
+			vrrTimer.Stop()
 		}
 		// A commit whose event never comes must not stop the output.
 		var stuck <-chan time.Time
 		if enabled && o.pending {
-			stuck = time.After(max(0, time.Until(o.stuckAt)))
+			stuckTimer.Reset(max(0, time.Until(o.stuckAt)))
+			stuck = stuckTimer.C
+		} else {
+			stuckTimer.Stop()
 		}
 		stateDirty := false
 		select {
 		start := time.Now()
 		// shows is what the frame puts on screen: presentation feedback
 		// of windows it does not draw is discarded, not presented.
-		f := pendingFrame{shows: shownBy(scene, seen)}
+		f := pendingFrame{shows: o.shownBy(scene, seen)}
 		var fb uint32
 		var c ports.SurfaceContent
 		forceCompose := len(requests) > 0
 		}
 		direct := false
 		if fb != 0 {
-			direct, err = o.commitScanout(fb, c, pendingFrame{shows: map[ports.WindowID]uint64{c.ID: seen[c.ID]}})
+			direct, err = o.commitScanout(fb, c, pendingFrame{shows: o.directShownBy(c.ID, seen[c.ID])})
 		}
 		if !direct {
 			var ov overlayWin
 // buffers can be released. A flip-less report merges into the newest
 // queued one; flips are never merged unless maxUnsent is reached.
 func (o *Output) report(flip *ports.FlipInfo, seen map[ports.WindowID]uint64) {
-	r := ports.OutputPresented{Output: o.conn.name, Flip: flip, Shown: o.shown, Queued: o.queued, Seen: maps.Clone(seen)}
+	if !maps.Equal(o.seenSnapshot, seen) {
+		o.seenSnapshot = maps.Clone(seen)
+	}
+	r := ports.OutputPresented{Output: o.conn.name, Flip: flip, Shown: o.shown, Queued: o.queued, Seen: o.seenSnapshot}
 	if n := len(o.unsent); n > 0 && (flip == nil || o.unsent[n-1].Flip == nil) {
 		if flip == nil {
 			r.Flip = o.unsent[n-1].Flip
 		old, next := o.unsent[0].Flip, o.unsent[1].Flip
 		if next != nil && old != nil {
 			next.Merged += 1 + old.Merged
+			// A snapshot can be shared with previously sent flips. Clone only
+			// when a merge actually needs to add an absent window.
+			cloned := false
 			for id, seq := range old.Shows {
 				if _, ok := next.Shows[id]; !ok {
+					if !cloned {
+						next.Shows = maps.Clone(next.Shows)
+						cloned = true
+					}
+					if next.Shows == nil {
+						next.Shows = make(map[ports.WindowID]uint64)
+					}
 					next.Shows[id] = seq
 				}
 			}
 	}
 }
-// shownBy is the content Seq of each window the scene draws.
+// shownBy returns an immutable snapshot of the content Seq of drawn windows.
-func shownBy(s ports.Scene, seen map[ports.WindowID]uint64) map[ports.WindowID]uint64 {
+// The scratch map is private to the output; a snapshot may still be read by
-	out := map[ports.WindowID]uint64{}
+// wayland after a subsequent flip, so never refill a previously sent map.
+func (o *Output) shownBy(s ports.Scene, seen map[ports.WindowID]uint64) map[ports.WindowID]uint64 {
+	if o.showsScratch == nil {
+		o.showsScratch = make(map[ports.WindowID]uint64)
+	}
+	clear(o.showsScratch)
 	for id, seq := range seen {
 		if s.Shows(id) {
-			out[id] = seq
+			o.showsScratch[id] = seq
 		}
 	}
-	return out
+	if !maps.Equal(o.showsSnapshot, o.showsScratch) {
+		o.showsSnapshot = maps.Clone(o.showsScratch)
+	}
+	return o.showsSnapshot
+}
+func (o *Output) directShownBy(id ports.WindowID, seq uint64) map[ports.WindowID]uint64 {
+	if len(o.directShowsSnapshot) != 1 || o.directShowsSnapshot[id] != seq {
+		o.directShowsSnapshot = map[ports.WindowID]uint64{id: seq}
+	}
+	return o.directShowsSnapshot
 }

M internal/adapters/drm/output_test.go +43 -0


@@ -218,6 +218,49 @@ func TestFlipReportsTimestampAndShownFrame(t *testing.T) {
218	218		}
219	219		}
220	220
	221	+	// Unchanged reports share an immutable snapshot; a later content update
	222	+	// must not change an earlier report, even after it has been sent.
	223	+	func TestReportSeenAllocations(t *testing.T) {
	224	+	o, _, _ := testOutput(t)
	225	+	seen := map[ports.WindowID]uint64{1: 1, 2: 2}
	226	+	o.report(nil, seen)
	227	+	// Keep the queue bounded without measuring the queue's growth.
	228	+	if allocs := testing.AllocsPerRun(100, func() { o.report(nil, seen) }); allocs != 0 {
	229	+	t.Errorf("unchanged report: %.1f allocs, want 0", allocs)
	230	+	} else {
	231	+	t.Logf("unchanged report: %.1f allocs", allocs)
	232	+	}
	233	+	previous := o.unsent[0].Seen
	234	+	seen[1] = 3
	235	+	o.report(nil, seen)
	236	+	if previous[1] != 1 \|\| o.unsent[0].Seen[1] != 3 {
	237	+	t.Fatalf("snapshot mutated: old %v new %v", previous, o.unsent[0].Seen)
	238	+	}
	239	+	}
	240	+
	241	+	func TestShownBySnapshotAllocations(t *testing.T) {
	242	+	o := &Output{}
	243	+	s := ports.Scene{Windows: []ports.SceneWindow{{ID: 1}}}
	244	+	seen := map[ports.WindowID]uint64{1: 1, 2: 2}
	245	+	previous := o.shownBy(s, seen)
	246	+	if allocs := testing.AllocsPerRun(100, func() { o.shownBy(s, seen) }); allocs != 0 {
	247	+	t.Errorf("unchanged shows: %.1f allocs, want 0", allocs)
	248	+	}
	249	+	seen[1] = 3
	250	+	current := o.shownBy(s, seen)
	251	+	if previous[1] != 1 \|\| current[1] != 3 {
	252	+	t.Fatalf("snapshot mutated: old %v new %v", previous, current)
	253	+	}
	254	+	direct := o.directShownBy(1, 3)
	255	+	if allocs := testing.AllocsPerRun(100, func() { o.directShownBy(1, 3) }); allocs != 0 {
	256	+	t.Errorf("unchanged direct shows: %.1f allocs, want 0", allocs)
	257	+	}
	258	+	o.directShownBy(1, 4)
	259	+	if direct[1] != 3 {
	260	+	t.Fatalf("direct snapshot mutated: %v", direct)
	261	+	}
	262	+	}
	263	+
221	264		func TestReportsKeepEveryFlip(t *testing.T) {
222	265		o, _, _ := testOutput(t)
223	266		seen := map[ports.WindowID]uint64{1: 1}


@@ -218,6 +218,49 @@ func TestFlipReportsTimestampAndShownFrame(t *testing.T) {
218	}	218		}
219	}	219		}
220		220
		221	+	// Unchanged reports share an immutable snapshot; a later content update
		222	+	// must not change an earlier report, even after it has been sent.
		223	+	func TestReportSeenAllocations(t *testing.T) {
		224	+	o, _, _ := testOutput(t)
		225	+	seen := map[ports.WindowID]uint64{1: 1, 2: 2}
		226	+	o.report(nil, seen)
		227	+	// Keep the queue bounded without measuring the queue's growth.
		228	+	if allocs := testing.AllocsPerRun(100, func() { o.report(nil, seen) }); allocs != 0 {
		229	+	t.Errorf("unchanged report: %.1f allocs, want 0", allocs)
		230	+	} else {
		231	+	t.Logf("unchanged report: %.1f allocs", allocs)
		232	+	}
		233	+	previous := o.unsent[0].Seen
		234	+	seen[1] = 3
		235	+	o.report(nil, seen)
		236	+	if previous[1] != 1 \|\| o.unsent[0].Seen[1] != 3 {
		237	+	t.Fatalf("snapshot mutated: old %v new %v", previous, o.unsent[0].Seen)
		238	+	}
		239	+	}
		240	+
		241	+	func TestShownBySnapshotAllocations(t *testing.T) {
		242	+	o := &Output{}
		243	+	s := ports.Scene{Windows: []ports.SceneWindow{{ID: 1}}}
		244	+	seen := map[ports.WindowID]uint64{1: 1, 2: 2}
		245	+	previous := o.shownBy(s, seen)
		246	+	if allocs := testing.AllocsPerRun(100, func() { o.shownBy(s, seen) }); allocs != 0 {
		247	+	t.Errorf("unchanged shows: %.1f allocs, want 0", allocs)
		248	+	}
		249	+	seen[1] = 3
		250	+	current := o.shownBy(s, seen)
		251	+	if previous[1] != 1 \|\| current[1] != 3 {
		252	+	t.Fatalf("snapshot mutated: old %v new %v", previous, current)
		253	+	}
		254	+	direct := o.directShownBy(1, 3)
		255	+	if allocs := testing.AllocsPerRun(100, func() { o.directShownBy(1, 3) }); allocs != 0 {
		256	+	t.Errorf("unchanged direct shows: %.1f allocs, want 0", allocs)
		257	+	}
		258	+	o.directShownBy(1, 4)
		259	+	if direct[1] != 3 {
		260	+	t.Fatalf("direct snapshot mutated: %v", direct)
		261	+	}
		262	+	}
		263	+
221	func TestReportsKeepEveryFlip(t *testing.T) {	264		func TestReportsKeepEveryFlip(t *testing.T) {
222	o, _, _ := testOutput(t)	265		o, _, _ := testOutput(t)
223	seen := map[ports.WindowID]uint64{1: 1}	266		seen := map[ports.WindowID]uint64{1: 1}

M internal/adapters/drm/scanout_test.go +1 -1


@@ -224,7 +224,7 @@ func TestPlaneWithoutInFormatsIsLinear(t *testing.T) {
224	224		// a window on another workspace or hidden is discarded, not presented.
225	225		func TestShownByScene(t *testing.T) {
226	226		s := ports.Scene{Windows: []ports.SceneWindow{{ID: 1}, {ID: 2, Hidden: true}}, Layers: []ports.SceneLayer{{ID: 5}}}
227		-	got := shownBy(s, map[ports.WindowID]uint64{1: 3, 2: 4, 5: 6, 9: 7})
	227	+	got := (&Output{}).shownBy(s, map[ports.WindowID]uint64{1: 3, 2: 4, 5: 6, 9: 7})
228	228		if len(got) != 2 \|\| got[1] != 3 \|\| got[5] != 6 {
229	229		t.Fatalf("shows %v", got)
230	230		}


@@ -224,7 +224,7 @@ func TestPlaneWithoutInFormatsIsLinear(t *testing.T) {
224		// a window on another workspace or hidden is discarded, not presented.	224		// a window on another workspace or hidden is discarded, not presented.
225		func TestShownByScene(t *testing.T) {	225		func TestShownByScene(t *testing.T) {
226		s := ports.Scene{Windows: []ports.SceneWindow{{ID: 1}, {ID: 2, Hidden: true}}, Layers: []ports.SceneLayer{{ID: 5}}}	226		s := ports.Scene{Windows: []ports.SceneWindow{{ID: 1}, {ID: 2, Hidden: true}}, Layers: []ports.SceneLayer{{ID: 5}}}
227	-	got := shownBy(s, map[ports.WindowID]uint64{1: 3, 2: 4, 5: 6, 9: 7})	227	+	got := (&Output{}).shownBy(s, map[ports.WindowID]uint64{1: 3, 2: 4, 5: 6, 9: 7})
228		if len(got) != 2 \|\| got[1] != 3 \|\| got[5] != 6 {	228		if len(got) != 2 \|\| got[1] != 3 \|\| got[5] != 6 {
229		t.Fatalf("shows %v", got)	229		t.Fatalf("shows %v", got)
230		}	230		}

M internal/adapters/vulkan/cursor.go +1 -1


@@ -32,7 +32,7 @@ const fourccARGB = 'A' \| 'R'<<8 \| '2'<<16 \| '4'<<24
32	32		func (r *Renderer) CursorBuffers(size int) ([2]ports.DMABuf, error) {
33	33		var out [2]ports.DMABuf
34	34		r.dropCursors()
35		-	if r.dd.GetMemoryFdKHR == nil \|\| len(r.dmabuf.Formats) == 0 \|\| size <= 0 \|\| size > 512 {
	35	+	if !r.dd.HasGetMemoryFdKHR() \|\| len(r.dmabuf.Formats) == 0 \|\| size <= 0 \|\| size > 512 {
36	36		return out, errors.New("device cannot export cursor images")
37	37		}
38	38		for i := range r.cursors {


@@ -32,7 +32,7 @@ const fourccARGB = 'A' \| 'R'<<8 \| '2'<<16 \| '4'<<24
32		func (r *Renderer) CursorBuffers(size int) ([2]ports.DMABuf, error) {	32		func (r *Renderer) CursorBuffers(size int) ([2]ports.DMABuf, error) {
33		var out [2]ports.DMABuf	33		var out [2]ports.DMABuf
34		r.dropCursors()	34		r.dropCursors()
35	-	if r.dd.GetMemoryFdKHR == nil \|\| len(r.dmabuf.Formats) == 0 \|\| size <= 0 \|\| size > 512 {	35	+	if !r.dd.HasGetMemoryFdKHR() \|\| len(r.dmabuf.Formats) == 0 \|\| size <= 0 \|\| size > 512 {
36		return out, errors.New("device cannot export cursor images")	36		return out, errors.New("device cannot export cursor images")
37		}	37		}
38		for i := range r.cursors {	38		for i := range r.cursors {

M internal/adapters/vulkan/damage.go +18 -0


@@ -39,6 +39,24 @@ func newDamage(tg target, s ports.Scene, bounds image.Rectangle) damageRegion
39	39		return d
40	40		}
41	41
	42	+	// frameDamage reuses the bookkeeping maps. drawn is swapped with the
	43	+	// target's previous map after hold, so a target never observes mutations.
	44	+	func (r Renderer) frameDamage(tg target, s ports.Scene, bounds image.Rectangle) *damageRegion {
	45	+	if r.damageSeen == nil {
	46	+	r.damageSeen = make(map[ports.WindowID]bool)
	47	+	}
	48	+	clear(r.damageSeen)
	49	+	if r.damageDrawn == nil {
	50	+	r.damageDrawn = make(map[ports.WindowID]heldWindow)
	51	+	}
	52	+	clear(r.damageDrawn)
	53	+	r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds}
	54	+	if tg.valid && s.Seq != 0 && tg.sceneSeq == s.Seq {
	55	+	r.damage.held = tg
	56	+	}
	57	+	return &r.damage
	58	+	}
	59	+
42	60		// all reports whether the whole target is redrawn.
43	61		func (d *damageRegion) all() bool { return d.held == nil }
44	62


@@ -39,6 +39,24 @@ func newDamage(tg target, s ports.Scene, bounds image.Rectangle) damageRegion
39	return d	39		return d
40	}	40		}
41		41
		42	+	// frameDamage reuses the bookkeeping maps. drawn is swapped with the
		43	+	// target's previous map after hold, so a target never observes mutations.
		44	+	func (r Renderer) frameDamage(tg target, s ports.Scene, bounds image.Rectangle) *damageRegion {
		45	+	if r.damageSeen == nil {
		46	+	r.damageSeen = make(map[ports.WindowID]bool)
		47	+	}
		48	+	clear(r.damageSeen)
		49	+	if r.damageDrawn == nil {
		50	+	r.damageDrawn = make(map[ports.WindowID]heldWindow)
		51	+	}
		52	+	clear(r.damageDrawn)
		53	+	r.damage = damageRegion{seen: r.damageSeen, drawn: r.damageDrawn, bound: bounds}
		54	+	if tg.valid && s.Seq != 0 && tg.sceneSeq == s.Seq {
		55	+	r.damage.held = tg
		56	+	}
		57	+	return &r.damage
		58	+	}
		59	+
42	// all reports whether the whole target is redrawn.	60		// all reports whether the whole target is redrawn.
43	func (d *damageRegion) all() bool { return d.held == nil }	61		func (d *damageRegion) all() bool { return d.held == nil }
44		62

M internal/adapters/vulkan/frame.go +23 -6


@@ -176,7 +176,7 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
176	176		if hdr {
177	177		tg = &r.hdrOwn
178	178		}
179		-	dmg := newDamage(tg, s, image.Rect(0, 0, r.width, r.height))
	179	+	dmg := r.frameDamage(tg, s, image.Rect(0, 0, r.width, r.height))
180	180		ds := r.draws(s, contents, dmg)
181	181		dmg.finish()
182	182		r.dropPools()

@@ -184,7 +184,12 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
184	184		if partial {
185	185		// The pass keeps the image: repaint the background under the
186	186		// region, then everything clipped to it.
187		-	ds = dmg.clip(append([]draw{r.fillDraw(dmg.area, parseColor(s.Background))}, ds...))
	187	+	// Keep the background first without allocating a second draw slice.
	188	+	ds = append(ds, draw{})
	189	+	copy(ds[1:], ds[:len(ds)-1])
	190	+	ds[0] = r.fillDraw(dmg.area, parseColor(s.Background))
	191	+	ds = dmg.clip(ds)
	192	+	r.scratchDraws = ds
188	193		r.redrawn += dmg.area.Dx() * dmg.area.Dy()
189	194		} else {
190	195		r.redrawn += r.width * r.height

@@ -212,8 +217,12 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
212	217		d.CmdPipelineBarrier(cmd, vk.PipelineStageTransferBit\|vk.PipelineStageColorAttachmentOutputBit, vk.PipelineStageColorAttachmentOutputBit, 0, 0, nil, 0, nil, 1, &b)
213	218		// Client buffers come from the foreign queue family (their driver)
214	219		// and go back to it after the frame, so the next frame acquires them.
215		-	var dmas []*imported
216		-	acquires := map[imported]os.File{}
	220	+	dmas := r.frameDMAs[:0]
	221	+	if r.frameAcquires == nil {
	222	+	r.frameAcquires = make(map[imported]os.File)
	223	+	}
	224	+	clear(r.frameAcquires)
	225	+	acquires := r.frameAcquires
217	226		for _, dr := range ds {
218	227		if dr.im != nil && !slices.Contains(dmas, dr.im) {
219	228		dmas = append(dmas, dr.im)

@@ -223,6 +232,7 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
223	232		acquires[dr.im] = dr.acquire
224	233		}
225	234		}
	235	+	r.frameDMAs = dmas
226	236		r.ownership(cmd, dmas, true)
227	237		r.recordDraws(cmd, tg.view, parseColor(s.Background), ds, partial)
228	238		r.ownership(cmd, dmas, false)

@@ -247,7 +257,7 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
247	257		}
248	258		// Wait for the client's GPU writes to each buffer: its explicit
249	259		// acquire fence, else the buffer's implicit fences.
250		-	var waits []vk.Semaphore
	260	+	waits := r.frameWaits[:0]
251	261		for _, im := range dmas {
252	262		if f := acquires[im]; f != nil {
253	263		if sem := r.importFence(f); sem != 0 {

@@ -257,12 +267,17 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
257	267		waits = r.appendReadFences(waits, im)
258	268		}
259	269		}
	270	+	r.frameWaits = waits
260	271		frame := r.frame
261	272		for _, sem := range waits {
	273	+	sem := sem
262	274		r.retire(frame, func() { d.DestroySemaphore(r.device, sem, nil) })
263	275		}
264	276		submit := vk.SubmitInfo{SType: vk.StructureTypeSubmitInfo, CommandBufferCount: 1, CommandBuffers: &cmd}
265		-	stages := make([]vk.PipelineStageFlags, len(waits))
	277	+	if cap(r.frameStages) < len(waits) {
	278	+	r.frameStages = make([]vk.PipelineStageFlags, len(waits))
	279	+	}
	280	+	stages := r.frameStages[:len(waits)]
266	281		if len(waits) > 0 {
267	282		for i := range stages {
268	283		stages[i] = vk.PipelineStageFragmentShaderBit

@@ -285,7 +300,9 @@ func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.Surfa
285	300		} else {
286	301		tg.layout = b.NewLayout
287	302		}
	303	+	oldWindows := tg.windows
288	304		tg.hold(s, dmg)
	305	+	r.damageDrawn = oldWindows
289	306		r.last, r.readBack = tg, false
290	307		r.dropUnused()
291	308		r.dropShm()

 	if hdr {
 		tg = &r.hdrOwn
 	}
-	dmg := newDamage(tg, s, image.Rect(0, 0, r.width, r.height))
+	dmg := r.frameDamage(tg, s, image.Rect(0, 0, r.width, r.height))
 	ds := r.draws(s, contents, dmg)
 	dmg.finish()
 	r.dropPools()
 	if partial {
 		// The pass keeps the image: repaint the background under the
 		// region, then everything clipped to it.
-		ds = dmg.clip(append([]draw{r.fillDraw(dmg.area, parseColor(s.Background))}, ds...))
+		// Keep the background first without allocating a second draw slice.
+		ds = append(ds, draw{})
+		copy(ds[1:], ds[:len(ds)-1])
+		ds[0] = r.fillDraw(dmg.area, parseColor(s.Background))
+		ds = dmg.clip(ds)
+		r.scratchDraws = ds
 		r.redrawn += dmg.area.Dx() * dmg.area.Dy()
 	} else {
 		r.redrawn += r.width * r.height
 	d.CmdPipelineBarrier(cmd, vk.PipelineStageTransferBit|vk.PipelineStageColorAttachmentOutputBit, vk.PipelineStageColorAttachmentOutputBit, 0, 0, nil, 0, nil, 1, &b)
 	// Client buffers come from the foreign queue family (their driver)
 	// and go back to it after the frame, so the next frame acquires them.
-	var dmas []*imported
+	dmas := r.frameDMAs[:0]
-	acquires := map[*imported]*os.File{}
+	if r.frameAcquires == nil {
+		r.frameAcquires = make(map[*imported]*os.File)
+	}
+	clear(r.frameAcquires)
+	acquires := r.frameAcquires
 	for _, dr := range ds {
 		if dr.im != nil && !slices.Contains(dmas, dr.im) {
 			dmas = append(dmas, dr.im)
 			acquires[dr.im] = dr.acquire
 		}
 	}
+	r.frameDMAs = dmas
 	r.ownership(cmd, dmas, true)
 	r.recordDraws(cmd, tg.view, parseColor(s.Background), ds, partial)
 	r.ownership(cmd, dmas, false)
 	}
 	// Wait for the client's GPU writes to each buffer: its explicit
 	// acquire fence, else the buffer's implicit fences.
-	var waits []vk.Semaphore
+	waits := r.frameWaits[:0]
 	for _, im := range dmas {
 		if f := acquires[im]; f != nil {
 			if sem := r.importFence(f); sem != 0 {
 			waits = r.appendReadFences(waits, im)
 		}
 	}
+	r.frameWaits = waits
 	frame := r.frame
 	for _, sem := range waits {
+		sem := sem
 		r.retire(frame, func() { d.DestroySemaphore(r.device, sem, nil) })
 	}
 	submit := vk.SubmitInfo{SType: vk.StructureTypeSubmitInfo, CommandBufferCount: 1, CommandBuffers: &cmd}
-	stages := make([]vk.PipelineStageFlags, len(waits))
+	if cap(r.frameStages) < len(waits) {
+		r.frameStages = make([]vk.PipelineStageFlags, len(waits))
+	}
+	stages := r.frameStages[:len(waits)]
 	if len(waits) > 0 {
 		for i := range stages {
 			stages[i] = vk.PipelineStageFragmentShaderBit
 	} else {
 		tg.layout = b.NewLayout
 	}
+	oldWindows := tg.windows
 	tg.hold(s, dmg)
+	r.damageDrawn = oldWindows
 	r.last, r.readBack = tg, false
 	r.dropUnused()
 	r.dropShm()

M internal/adapters/vulkan/renderer.go +11 -1


@@ -75,6 +75,16 @@ type Renderer struct {
75	75		// storage buffer binding.
76	76		compose composer
77	77		maxRange int
	78	+	// Scratch belongs to the renderer goroutine; neither slice is published.
	79	+	scratchDraws []draw
	80	+	scratchCovers []image.Rectangle
	81	+	damage damageRegion
	82	+	damageSeen map[ports.WindowID]bool
	83	+	damageDrawn map[ports.WindowID]heldWindow
	84	+	frameDMAs []*imported
	85	+	frameAcquires map[imported]os.File
	86	+	frameWaits []vk.Semaphore
	87	+	frameStages []vk.PipelineStageFlags
78	88		}
79	89
80	90		func checked(name string, result vk.Result) error {

@@ -200,7 +210,7 @@ func New(width, height int) (r *Renderer, err error) {
200	210		}
201	211		// VK_KHR_external_semaphore_fd (enabled above): frames export
202	212		// their fence as a sync file.
203		-	r.syncFD = r.dd.GetSemaphoreFdKHR != nil
	213	+	r.syncFD = r.dd.HasGetSemaphoreFdKHR()
204	214		}
205	215		r.dd.GetDeviceQueue(r.device, family, 0, &r.queue)
206	216		extent := vk.Extent3D{Width: uint32(width), Height: uint32(height), Depth: 1}

 	// storage buffer binding.
 	compose  composer
 	maxRange int
+	// Scratch belongs to the renderer goroutine; neither slice is published.
+	scratchDraws  []draw
+	scratchCovers []image.Rectangle
+	damage        damageRegion
+	damageSeen    map[ports.WindowID]bool
+	damageDrawn   map[ports.WindowID]heldWindow
+	frameDMAs     []*imported
+	frameAcquires map[*imported]*os.File
+	frameWaits    []vk.Semaphore
+	frameStages   []vk.PipelineStageFlags
 }
 func checked(name string, result vk.Result) error {
 		}
 		// VK_KHR_external_semaphore_fd (enabled above): frames export
 		// their fence as a sync file.
-		r.syncFD = r.dd.GetSemaphoreFdKHR != nil
+		r.syncFD = r.dd.HasGetSemaphoreFdKHR()
 	}
 	r.dd.GetDeviceQueue(r.device, family, 0, &r.queue)
 	extent := vk.Extent3D{Width: uint32(width), Height: uint32(height), Depth: 1}

M internal/adapters/vulkan/scene.go +6 -2


@@ -24,7 +24,9 @@ type sceneWalk struct {
24	24
25	25		// draws walks the scene into quads in paint order.
26	26		func (r Renderer) draws(s ports.Scene, contents map[ports.WindowID]ports.SurfaceContent, dmg damageRegion) []draw {
27		-	w := &sceneWalk{r: r, s: s, contents: contents, dmg: dmg, scale: s.Scale, bounds: image.Rect(0, 0, r.width, r.height)}
	27	+	clear(r.scratchDraws) // release references from longer earlier scenes
	28	+	w := &sceneWalk{r: r, s: s, contents: contents, dmg: dmg, scale: s.Scale, bounds: image.Rect(0, 0, r.width, r.height), draws: r.scratchDraws[:0]}
	29	+	r.scratchCovers = r.scratchCovers[:0]
28	30		if w.scale <= 0 {
29	31		w.scale = 1
30	32		}

@@ -36,6 +38,7 @@ func (r *Renderer) draws(s ports.Scene, contents map[ports.WindowID]ports.Surfac
36	38		w.popups(false)
37	39		w.layers(true)
38	40		w.popups(true)
	41	+	r.scratchDraws = w.draws
39	42		return w.draws
40	43		}
41	44

@@ -160,7 +163,7 @@ func (w sceneWalk) place(id ports.WindowID, content ports.SurfaceContent, x, y
160	163		// left out, so the surface below it still shows: a dmabuf that fails to
161	164		// import (opaqueChildren pre-imports it) or an unreadable shm buffer.
162	165		func (w sceneWalk) opaqueChildren(content ports.SurfaceContent, ox, oy int, clip image.Rectangle) []image.Rectangle {
163		-	var covers []image.Rectangle
	166	+	covers := w.r.scratchCovers[:0]
164	167		for i := range content.Children {
165	168		ch := &content.Children[i]
166	169		if ch.Below \|\| !ch.Opaque {

@@ -172,6 +175,7 @@ func (w sceneWalk) opaqueChildren(content ports.SurfaceContent, ox, oy int, cl
172	175		}
173	176		covers = append(covers, dst)
174	177		}
	178	+	w.r.scratchCovers = covers
175	179		return covers
176	180		}
177	181

 // draws walks the scene into quads in paint order.
 func (r *Renderer) draws(s ports.Scene, contents map[ports.WindowID]ports.SurfaceContent, dmg *damageRegion) []draw {
-	w := &sceneWalk{r: r, s: s, contents: contents, dmg: dmg, scale: s.Scale, bounds: image.Rect(0, 0, r.width, r.height)}
+	clear(r.scratchDraws) // release references from longer earlier scenes
+	w := &sceneWalk{r: r, s: s, contents: contents, dmg: dmg, scale: s.Scale, bounds: image.Rect(0, 0, r.width, r.height), draws: r.scratchDraws[:0]}
+	r.scratchCovers = r.scratchCovers[:0]
 	if w.scale <= 0 {
 		w.scale = 1
 	}
 	w.popups(false)
 	w.layers(true)
 	w.popups(true)
+	r.scratchDraws = w.draws
 	return w.draws
 }
 // left out, so the surface below it still shows: a dmabuf that fails to
 // import (opaqueChildren pre-imports it) or an unreadable shm buffer.
 func (w *sceneWalk) opaqueChildren(content *ports.SurfaceContent, ox, oy int, clip image.Rectangle) []image.Rectangle {
-	var covers []image.Rectangle
+	covers := w.r.scratchCovers[:0]
 	for i := range content.Children {
 		ch := &content.Children[i]
 		if ch.Below || !ch.Opaque {
 		}
 		covers = append(covers, dst)
 	}
+	w.r.scratchCovers = covers
 	return covers
 }

M internal/adapters/vulkan/scene_perf_test.go +38 -2


@@ -7,6 +7,30 @@ import (
7	7		"github.com/bnema/neferwl/internal/ports"
8	8		)
9	9
	10	+	// TestRenderSteadyStateAllocations measures the cost of a small stable scene.
	11	+	func TestRenderSteadyStateAllocations(t *testing.T) {
	12	+	r, err := New(32, 32)
	13	+	if err != nil {
	14	+	t.Skipf("Vulkan unavailable: %v", err)
	15	+	}
	16	+	defer r.Close()
	17	+	scene := ports.Scene{Seq: 1, Windows: []ports.SceneWindow{{ID: 1, Rect: ports.Rect{W: 32, H: 32}}}}
	18	+	content := solidContent(t, 8, 8, color.RGBA{R: 255, A: 255})
	19	+	content.ID, content.Surface, content.Seq, content.Version = 1, 1, 1, 1
	20	+	contents := map[ports.WindowID]ports.SurfaceContent{1: content}
	21	+	frame := func() {
	22	+	if err := render(r, scene, contents); err != nil {
	23	+	t.Fatal(err)
	24	+	}
	25	+	}
	26	+	frame()
	27	+	if allocs := testing.AllocsPerRun(20, frame); allocs > 13 {
	28	+	t.Errorf("small steady-state Render: %.1f allocs/frame, want <=13", allocs)
	29	+	} else {
	30	+	t.Logf("small steady-state Render: %.1f allocs/frame", allocs)
	31	+	}
	32	+	}
	33	+
10	34		// TestSceneWalkUnchangedTiledSHM guards the GPU copy budget of an unchanged
11	35		// subsurface tree. Each child has its own real wl_shm pool.
12	36		func TestSceneWalkUnchangedTiledSHM(t *testing.T) {

@@ -37,13 +61,25 @@ func TestSceneWalkUnchangedTiledSHM(t *testing.T) {
37	61		_ = r.draws(scene, contents, dmg)
38	62		}
39	63		walk()
40		-	if allocs := testing.AllocsPerRun(20, walk); allocs > 16 {
41		-	t.Errorf("%d tiles: %.1f allocations/walk, want <=16", count, allocs)
	64	+	if allocs := testing.AllocsPerRun(20, walk); allocs > 8 {
	65	+	t.Errorf("%d tiles: %.1f allocations/walk, want <=8", count, allocs)
42	66		} else {
43	67		t.Logf("%d tiles: %.1f allocations/walk", count, allocs)
44	68		}
45	69		}
46	70		contents[1] = root
	71	+	// The same renderer and scene are reused; measure only steady-state frame work.
	72	+	frame := func() {
	73	+	if err := render(r, scene, contents); err != nil {
	74	+	t.Fatal(err)
	75	+	}
	76	+	}
	77	+	frame()
	78	+	if allocs := testing.AllocsPerRun(20, frame); allocs > 16 {
	79	+	t.Errorf("steady-state Render: %.1f allocs/frame, want <=16", allocs)
	80	+	} else {
	81	+	t.Logf("steady-state Render: %.1f allocs/frame", allocs)
	82	+	}
47	83		before := r.copied
48	84		for i := range 100 {
49	85		root.Seq++


@@ -7,6 +7,30 @@ import (
7		"github.com/bnema/neferwl/internal/ports"	7		"github.com/bnema/neferwl/internal/ports"
8		)	8		)
9			9
			10	+	// TestRenderSteadyStateAllocations measures the cost of a small stable scene.
			11	+	func TestRenderSteadyStateAllocations(t *testing.T) {
			12	+	r, err := New(32, 32)
			13	+	if err != nil {
			14	+	t.Skipf("Vulkan unavailable: %v", err)
			15	+	}
			16	+	defer r.Close()
			17	+	scene := ports.Scene{Seq: 1, Windows: []ports.SceneWindow{{ID: 1, Rect: ports.Rect{W: 32, H: 32}}}}
			18	+	content := solidContent(t, 8, 8, color.RGBA{R: 255, A: 255})
			19	+	content.ID, content.Surface, content.Seq, content.Version = 1, 1, 1, 1
			20	+	contents := map[ports.WindowID]ports.SurfaceContent{1: content}
			21	+	frame := func() {
			22	+	if err := render(r, scene, contents); err != nil {
			23	+	t.Fatal(err)
			24	+	}
			25	+	}
			26	+	frame()
			27	+	if allocs := testing.AllocsPerRun(20, frame); allocs > 13 {
			28	+	t.Errorf("small steady-state Render: %.1f allocs/frame, want <=13", allocs)
			29	+	} else {
			30	+	t.Logf("small steady-state Render: %.1f allocs/frame", allocs)
			31	+	}
			32	+	}
			33	+
10		// TestSceneWalkUnchangedTiledSHM guards the GPU copy budget of an unchanged	34		// TestSceneWalkUnchangedTiledSHM guards the GPU copy budget of an unchanged
11		// subsurface tree. Each child has its own real wl_shm pool.	35		// subsurface tree. Each child has its own real wl_shm pool.
12		func TestSceneWalkUnchangedTiledSHM(t *testing.T) {	36		func TestSceneWalkUnchangedTiledSHM(t *testing.T) {

@@ -37,13 +61,25 @@ func TestSceneWalkUnchangedTiledSHM(t *testing.T) {
37		_ = r.draws(scene, contents, dmg)	61		_ = r.draws(scene, contents, dmg)
38		}	62		}
39		walk()	63		walk()
40	-	if allocs := testing.AllocsPerRun(20, walk); allocs > 16 {	64	+	if allocs := testing.AllocsPerRun(20, walk); allocs > 8 {
41	-	t.Errorf("%d tiles: %.1f allocations/walk, want <=16", count, allocs)	65	+	t.Errorf("%d tiles: %.1f allocations/walk, want <=8", count, allocs)
42		} else {	66		} else {
43		t.Logf("%d tiles: %.1f allocations/walk", count, allocs)	67		t.Logf("%d tiles: %.1f allocations/walk", count, allocs)
44		}	68		}
45		}	69		}
46		contents[1] = root	70		contents[1] = root
			71	+	// The same renderer and scene are reused; measure only steady-state frame work.
			72	+	frame := func() {
			73	+	if err := render(r, scene, contents); err != nil {
			74	+	t.Fatal(err)
			75	+	}
			76	+	}
			77	+	frame()
			78	+	if allocs := testing.AllocsPerRun(20, frame); allocs > 16 {
			79	+	t.Errorf("steady-state Render: %.1f allocs/frame, want <=16", allocs)
			80	+	} else {
			81	+	t.Logf("steady-state Render: %.1f allocs/frame", allocs)
			82	+	}
47		before := r.copied	83		before := r.copied
48		for i := range 100 {	84		for i := range 100 {
49		root.Seq++	85		root.Seq++

M internal/adapters/vulkan/targets.go +1 -1


@@ -56,7 +56,7 @@ func (r *Renderer) ExportTargets(n int, modifiers []uint64) ([]ports.DMABuf, err
56	56		if n == 0 {
57	57		return nil, nil
58	58		}
59		-	if r.dd.GetMemoryFdKHR == nil \|\| r.dd.GetImageDrmFormatModifierPropertiesEXT == nil \|\| len(r.dmabuf.Formats) == 0 {
	59	+	if !r.dd.HasGetMemoryFdKHR() \|\| !r.dd.HasGetImageDrmFormatModifierPropertiesEXT() \|\| len(r.dmabuf.Formats) == 0 {
60	60		return nil, errors.New("device cannot export dmabufs")
61	61		}
62	62		mods := r.exportModifiers(modifiers)


@@ -56,7 +56,7 @@ func (r *Renderer) ExportTargets(n int, modifiers []uint64) ([]ports.DMABuf, err
56		if n == 0 {	56		if n == 0 {
57		return nil, nil	57		return nil, nil
58		}	58		}
59	-	if r.dd.GetMemoryFdKHR == nil \|\| r.dd.GetImageDrmFormatModifierPropertiesEXT == nil \|\| len(r.dmabuf.Formats) == 0 {	59	+	if !r.dd.HasGetMemoryFdKHR() \|\| !r.dd.HasGetImageDrmFormatModifierPropertiesEXT() \|\| len(r.dmabuf.Formats) == 0 {
60		return nil, errors.New("device cannot export dmabufs")	60		return nil, errors.New("device cannot export dmabufs")
61		}	61		}
62		mods := r.exportModifiers(modifiers)	62		mods := r.exportModifiers(modifiers)

M internal/adapters/wayland/fifo.go +71 -5

@@ -1,6 +1,7 @@
1	1		package wayland
2	2
3	3		import (
	4	+	"sync"
4	5		"time"
5	6
6	7		"github.com/bnema/neferwl/internal/ports"

@@ -233,6 +234,45 @@ type update struct {
233	234		bufDamage []ports.Rect
234	235		readyGeneration uint64
235	236		readyResult bool
	237	+	// refs counts incoming deps and prev links. Retired updates remain readable
	238	+	// until their last incoming link goes away, even after dropQueue.
	239	+	refs int
	240	+	retired bool
	241	+	}
	242	+
	243	+	// updatePool contains only retired updates with no incoming graph links. A
	244	+	// popped update is NOT reusable just because it left its owner's queue: a
	245	+	// parent's queued deps can still point to it (including after role teardown).
	246	+	// The display owner goroutine changes graph links; sync.Pool only shares dead,
	247	+	// zeroed storage between independent displays.
	248	+	var updatePool sync.Pool
	249	+
	250	+	func (u *update) releaseRef() {
	251	+	u.refs--
	252	+	if u.refs == 0 && u.retired {
	253	+	deps := u.deps[:0]
	254	+	*u = update{deps: deps}
	255	+	updatePool.Put(u)
	256	+	}
	257	+	}
	258	+
	259	+	// retireUpdate severs outgoing links only after the update is no longer used
	260	+	// by applyUpdate/dropQueue. Incoming links keep its identity until consumed.
	261	+	func (u *update) retireUpdate() {
	262	+	u.retired = true
	263	+	prev, deps := u.prev, u.deps
	264	+	u.prev, u.deps = nil, deps[:0]
	265	+	if prev != nil {
	266	+	prev.releaseRef()
	267	+	}
	268	+	for i, dep := range deps {
	269	+	deps[i] = nil // the pooled backing array must not retain old dependencies
	270	+	dep.releaseRef()
	271	+	}
	272	+	if u.refs == 0 {
	273	+	*u = update{deps: u.deps}
	274	+	updatePool.Put(u)
	275	+	}
236	276		}
237	277
238	278		type childLayout struct {

@@ -347,7 +387,13 @@ func (s *Server) nextRefresh(name string, now time.Time) time.Time {
347	387
348	388		// queueUpdate queues the pending commit until it can apply.
349	389		func (s *surface) queueUpdate() {
350		-	u := s.takePending()
	390	+	u, _ := updatePool.Get().(*update)
	391	+	if u == nil {
	392	+	u = new(update)
	393	+	}
	394	+	deps := u.deps[:0]
	395	+	*u = s.takePending()
	396	+	u.deps = deps
351	397		u.owner, u.synced = s, s.effectivelySynced()
352	398		for _, ch := range s.sub.children {
353	399		for i := len(ch.queue) - 1; i >= 0; i-- {

@@ -355,14 +401,16 @@ func (s *surface) queueUpdate() {
355	401		if candidate.synced && !candidate.bound {
356	402		candidate.bound = true
357	403		u.deps = append(u.deps, candidate)
	404	+	candidate.refs++
358	405		break
359	406		}
360	407		}
361	408		}
362	409		if len(s.queue) > 0 {
363	410		u.prev = s.queue[len(s.queue)-1]
	411	+	u.prev.refs++
364	412		}
365		-	s.queue = append(s.queue, &u)
	413	+	s.queue = append(s.queue, u)
366	414		s.server.fifoSurfaces[s] = struct{}{}
367	415		s.server.wakePacer()
368	416		}

@@ -431,7 +479,11 @@ func (s *surface) dropQueue() {
431	479		u.buffer.SendRelease()
432	480		}
433	481		}
434		-	s.queue = nil
	482	+	for i, u := range s.queue {
	483	+	s.queue[i] = nil // do not retain retired objects in the queue backing array
	484	+	u.retireUpdate()
	485	+	}
	486	+	s.queue = s.queue[:0]
435	487		delete(s.server.fifoSurfaces, s)
436	488		}
437	489

@@ -574,6 +626,9 @@ func (s *surface) flushDesync() {
574	626		}
575	627
576	628		func (u *update) graphReady(now time.Time) bool {
	629	+	if u.retired {
	630	+	return true // applied or dropped; a queued parent cannot replay it
	631	+	}
577	632		s := u.owner
578	633		gen := s.server.readinessGeneration
579	634		if gen != 0 && u.readyGeneration == gen {

@@ -598,6 +653,9 @@ func (u *update) graphReady(now time.Time) bool {
598	653		}
599	654
600	655		func (u *update) applyGraph() {
	656	+	if u.retired {
	657	+	return
	658	+	}
601	659		s := u.owner
602	660		for len(s.queue) > 0 && s.queue[0] != u {
603	661		s.queue[0].applyGraph()

@@ -610,11 +668,19 @@ func (u *update) applyGraph() {
610	668		if len(s.queue) == 0 \|\| s.queue[0] != u {
611	669		return
612	670		}
613		-	s.queue = s.queue[1:]
	671	+	s.queue[0] = nil
	672	+	if len(s.queue) == 1 {
	673	+	s.queue = s.queue[:0] // retain capacity for the next commit
	674	+	} else {
	675	+	s.queue = s.queue[1:]
	676	+	}
614	677		if len(s.queue) > 0 {
615		-	s.queue[0].prev = nil
	678	+	next := s.queue[0]
	679	+	next.prev = nil
	680	+	u.releaseRef() // next's prev no longer points to u
616	681		}
617	682		s.applyUpdate(u)
	683	+	u.retireUpdate()
618	684		}
619	685
620	686		// sameBuffer reports whether two wrappers are the same wl_buffer.

@@ -1,6 +1,7 @@
1		package wayland	1		package wayland
2			2
3		import (	3		import (
			4	+	"sync"
4		"time"	5		"time"
5			6
6		"github.com/bnema/neferwl/internal/ports"	7		"github.com/bnema/neferwl/internal/ports"

@@ -233,6 +234,45 @@ type update struct {
233		bufDamage []ports.Rect	234		bufDamage []ports.Rect
234		readyGeneration uint64	235		readyGeneration uint64
235		readyResult bool	236		readyResult bool
			237	+	// refs counts incoming deps and prev links. Retired updates remain readable
			238	+	// until their last incoming link goes away, even after dropQueue.
			239	+	refs int
			240	+	retired bool
			241	+	}
			242	+
			243	+	// updatePool contains only retired updates with no incoming graph links. A
			244	+	// popped update is NOT reusable just because it left its owner's queue: a
			245	+	// parent's queued deps can still point to it (including after role teardown).
			246	+	// The display owner goroutine changes graph links; sync.Pool only shares dead,
			247	+	// zeroed storage between independent displays.
			248	+	var updatePool sync.Pool
			249	+
			250	+	func (u *update) releaseRef() {
			251	+	u.refs--
			252	+	if u.refs == 0 && u.retired {
			253	+	deps := u.deps[:0]
			254	+	*u = update{deps: deps}
			255	+	updatePool.Put(u)
			256	+	}
			257	+	}
			258	+
			259	+	// retireUpdate severs outgoing links only after the update is no longer used
			260	+	// by applyUpdate/dropQueue. Incoming links keep its identity until consumed.
			261	+	func (u *update) retireUpdate() {
			262	+	u.retired = true
			263	+	prev, deps := u.prev, u.deps
			264	+	u.prev, u.deps = nil, deps[:0]
			265	+	if prev != nil {
			266	+	prev.releaseRef()
			267	+	}
			268	+	for i, dep := range deps {
			269	+	deps[i] = nil // the pooled backing array must not retain old dependencies
			270	+	dep.releaseRef()
			271	+	}
			272	+	if u.refs == 0 {
			273	+	*u = update{deps: u.deps}
			274	+	updatePool.Put(u)
			275	+	}
236		}	276		}
237			277
238		type childLayout struct {	278		type childLayout struct {

@@ -347,7 +387,13 @@ func (s *Server) nextRefresh(name string, now time.Time) time.Time {
347			387
348		// queueUpdate queues the pending commit until it can apply.	388		// queueUpdate queues the pending commit until it can apply.
349		func (s *surface) queueUpdate() {	389		func (s *surface) queueUpdate() {
350	-	u := s.takePending()	390	+	u, _ := updatePool.Get().(*update)
			391	+	if u == nil {
			392	+	u = new(update)
			393	+	}
			394	+	deps := u.deps[:0]
			395	+	*u = s.takePending()
			396	+	u.deps = deps
351		u.owner, u.synced = s, s.effectivelySynced()	397		u.owner, u.synced = s, s.effectivelySynced()
352		for _, ch := range s.sub.children {	398		for _, ch := range s.sub.children {
353		for i := len(ch.queue) - 1; i >= 0; i-- {	399		for i := len(ch.queue) - 1; i >= 0; i-- {

@@ -355,14 +401,16 @@ func (s *surface) queueUpdate() {
355		if candidate.synced && !candidate.bound {	401		if candidate.synced && !candidate.bound {
356		candidate.bound = true	402		candidate.bound = true
357		u.deps = append(u.deps, candidate)	403		u.deps = append(u.deps, candidate)
			404	+	candidate.refs++
358		break	405		break
359		}	406		}
360		}	407		}
361		}	408		}
362		if len(s.queue) > 0 {	409		if len(s.queue) > 0 {
363		u.prev = s.queue[len(s.queue)-1]	410		u.prev = s.queue[len(s.queue)-1]
			411	+	u.prev.refs++
364		}	412		}
365	-	s.queue = append(s.queue, &u)	413	+	s.queue = append(s.queue, u)
366		s.server.fifoSurfaces[s] = struct{}{}	414		s.server.fifoSurfaces[s] = struct{}{}
367		s.server.wakePacer()	415		s.server.wakePacer()
368		}	416		}

@@ -431,7 +479,11 @@ func (s *surface) dropQueue() {
431		u.buffer.SendRelease()	479		u.buffer.SendRelease()
432		}	480		}
433		}	481		}
434	-	s.queue = nil	482	+	for i, u := range s.queue {
			483	+	s.queue[i] = nil // do not retain retired objects in the queue backing array
			484	+	u.retireUpdate()
			485	+	}
			486	+	s.queue = s.queue[:0]
435		delete(s.server.fifoSurfaces, s)	487		delete(s.server.fifoSurfaces, s)
436		}	488		}
437			489

@@ -574,6 +626,9 @@ func (s *surface) flushDesync() {
574		}	626		}
575			627
576		func (u *update) graphReady(now time.Time) bool {	628		func (u *update) graphReady(now time.Time) bool {
			629	+	if u.retired {
			630	+	return true // applied or dropped; a queued parent cannot replay it
			631	+	}
577		s := u.owner	632		s := u.owner
578		gen := s.server.readinessGeneration	633		gen := s.server.readinessGeneration
579		if gen != 0 && u.readyGeneration == gen {	634		if gen != 0 && u.readyGeneration == gen {

@@ -598,6 +653,9 @@ func (u *update) graphReady(now time.Time) bool {
598		}	653		}
599			654
600		func (u *update) applyGraph() {	655		func (u *update) applyGraph() {
			656	+	if u.retired {
			657	+	return
			658	+	}
601		s := u.owner	659		s := u.owner
602		for len(s.queue) > 0 && s.queue[0] != u {	660		for len(s.queue) > 0 && s.queue[0] != u {
603		s.queue[0].applyGraph()	661		s.queue[0].applyGraph()

@@ -610,11 +668,19 @@ func (u *update) applyGraph() {
610		if len(s.queue) == 0 \|\| s.queue[0] != u {	668		if len(s.queue) == 0 \|\| s.queue[0] != u {
611		return	669		return
612		}	670		}
613	-	s.queue = s.queue[1:]	671	+	s.queue[0] = nil
			672	+	if len(s.queue) == 1 {
			673	+	s.queue = s.queue[:0] // retain capacity for the next commit
			674	+	} else {
			675	+	s.queue = s.queue[1:]
			676	+	}
614		if len(s.queue) > 0 {	677		if len(s.queue) > 0 {
615	-	s.queue[0].prev = nil	678	+	next := s.queue[0]
			679	+	next.prev = nil
			680	+	u.releaseRef() // next's prev no longer points to u
616		}	681		}
617		s.applyUpdate(u)	682		s.applyUpdate(u)
			683	+	u.retireUpdate()
618		}	684		}
619			685
620		// sameBuffer reports whether two wrappers are the same wl_buffer.	686		// sameBuffer reports whether two wrappers are the same wl_buffer.

M internal/adapters/wayland/fifo_test.go +49 -0


@@ -207,6 +207,55 @@ func TestFifoQueuedBufferDestroyed(t *testing.T) {
207	207		}
208	208		}
209	209
	210	+	// A parent can keep a dependency after the child's queue is dropped or its
	211	+	// update is applied independently. Neither path may recycle that dependency
	212	+	// into a new commit while the parent still holds it.
	213	+	func TestQueuedParentRetainsRetiredChild(t *testing.T) {
	214	+	for _, dropped := range []bool{true, false} {
	215	+	name := "applied"
	216	+	if dropped {
	217	+	name = "dropped"
	218	+	}
	219	+	t.Run(name, func(t *testing.T) {
	220	+	srv := &Server{fifoSurfaces: make(map[*surface]struct{}), frameReady: make(chan struct{}, 1)}
	221	+	parent := &surface{server: srv}
	222	+	child := &surface{server: srv}
	223	+	child.sub.parent, child.sub.synced = parent, true
	224	+	parent.sub.children = []*surface{child}
	225	+	child.pendingScale = 2
	226	+	child.queueUpdate()
	227	+	old := child.queue[0]
	228	+	parent.pendingScale = 3
	229	+	parent.queueUpdate()
	230	+	pu := parent.queue[0]
	231	+	if old.refs != 1 \|\| len(pu.deps) != 1 \|\| pu.deps[0] != old {
	232	+	t.Fatal("parent did not capture child's update")
	233	+	}
	234	+	if dropped {
	235	+	child.dropQueue()
	236	+	} else {
	237	+	old.applyGraph()
	238	+	}
	239	+	if !old.retired \|\| old.refs != 1 {
	240	+	t.Fatalf("retired child ref count: retired=%v refs=%d", old.retired, old.refs)
	241	+	}
	242	+	child.pendingScale = 7
	243	+	child.queueUpdate()
	244	+	if child.queue[0] == old \|\| old.scale != 2 \|\| pu.deps[0] != old {
	245	+	t.Fatal("live dependency was recycled as a new child commit")
	246	+	}
	247	+	if !pu.graphReady(time.Now()) {
	248	+	t.Fatal("retired dependency kept parent waiting")
	249	+	}
	250	+	pu.applyGraph()
	251	+	if parent.bufferScale != 3 \|\| len(parent.queue) != 0 \|\| len(child.queue) != 1 \|\| child.queue[0].scale != 7 {
	252	+	t.Fatal("parent replayed old dependency or changed new child commit")
	253	+	}
	254	+	child.dropQueue()
	255	+	})
	256	+	}
	257	+	}
	258	+
210	259		// Rechecking a shared predecessor in a traversal must reuse its readiness.
211	260		func TestGraphReadyMemoizesPredecessors(t *testing.T) {
212	261		s := &Server{readinessGeneration: 1}


@@ -207,6 +207,55 @@ func TestFifoQueuedBufferDestroyed(t *testing.T) {
207	}	207		}
208	}	208		}
209		209
		210	+	// A parent can keep a dependency after the child's queue is dropped or its
		211	+	// update is applied independently. Neither path may recycle that dependency
		212	+	// into a new commit while the parent still holds it.
		213	+	func TestQueuedParentRetainsRetiredChild(t *testing.T) {
		214	+	for _, dropped := range []bool{true, false} {
		215	+	name := "applied"
		216	+	if dropped {
		217	+	name = "dropped"
		218	+	}
		219	+	t.Run(name, func(t *testing.T) {
		220	+	srv := &Server{fifoSurfaces: make(map[*surface]struct{}), frameReady: make(chan struct{}, 1)}
		221	+	parent := &surface{server: srv}
		222	+	child := &surface{server: srv}
		223	+	child.sub.parent, child.sub.synced = parent, true
		224	+	parent.sub.children = []*surface{child}
		225	+	child.pendingScale = 2
		226	+	child.queueUpdate()
		227	+	old := child.queue[0]
		228	+	parent.pendingScale = 3
		229	+	parent.queueUpdate()
		230	+	pu := parent.queue[0]
		231	+	if old.refs != 1 \|\| len(pu.deps) != 1 \|\| pu.deps[0] != old {
		232	+	t.Fatal("parent did not capture child's update")
		233	+	}
		234	+	if dropped {
		235	+	child.dropQueue()
		236	+	} else {
		237	+	old.applyGraph()
		238	+	}
		239	+	if !old.retired \|\| old.refs != 1 {
		240	+	t.Fatalf("retired child ref count: retired=%v refs=%d", old.retired, old.refs)
		241	+	}
		242	+	child.pendingScale = 7
		243	+	child.queueUpdate()
		244	+	if child.queue[0] == old \|\| old.scale != 2 \|\| pu.deps[0] != old {
		245	+	t.Fatal("live dependency was recycled as a new child commit")
		246	+	}
		247	+	if !pu.graphReady(time.Now()) {
		248	+	t.Fatal("retired dependency kept parent waiting")
		249	+	}
		250	+	pu.applyGraph()
		251	+	if parent.bufferScale != 3 \|\| len(parent.queue) != 0 \|\| len(child.queue) != 1 \|\| child.queue[0].scale != 7 {
		252	+	t.Fatal("parent replayed old dependency or changed new child commit")
		253	+	}
		254	+	child.dropQueue()
		255	+	})
		256	+	}
		257	+	}
		258	+
210	// Rechecking a shared predecessor in a traversal must reuse its readiness.	259		// Rechecking a shared predecessor in a traversal must reuse its readiness.
211	func TestGraphReadyMemoizesPredecessors(t *testing.T) {	260		func TestGraphReadyMemoizesPredecessors(t *testing.T) {
212	s := &Server{readinessGeneration: 1}	261		s := &Server{readinessGeneration: 1}

M internal/adapters/wayland/frames.go +13 -3


@@ -26,9 +26,14 @@ func (s *Server) pace(ctx context.Context) {
26	26		timer := time.NewTimer(defaultFramePeriod)
27	27		defer timer.Stop()
28	28		presented := s.channels.Presented
	29	+	if s.flipped == nil {
	30	+	s.flipped = make(map[string]bool)
	31	+	}
29	32		for {
30		-	flipped := map[string]bool{}
31		-	var reports []ports.OutputPresented
	33	+	clear(s.flipped)
	34	+	flipped := s.flipped
	35	+	s.frameReports = s.frameReports[:0]
	36	+	reports := s.frameReports
32	37		select {
33	38		case <-ctx.Done():
34	39		return

@@ -52,6 +57,7 @@ func (s *Server) pace(ctx context.Context) {
52	57		break drain
53	58		}
54	59		}
	60	+	s.frameReports = reports
55	61		wait, idle := defaultFramePeriod, false
56	62		if !s.display.Do(func() {
57	63		if ctx.Err() == nil {

@@ -130,7 +136,11 @@ func (s Server) queueFrames(name string, callbacks []wayland.Callback) {
130	136		// their deadline. wait is the time to the next deadline; idle is set when
131	137		// no callback waits after firing.
132	138		func (s *Server) dueFrames(now time.Time, flipped map[string]bool) (fire []string, wait time.Duration, idle bool) {
133		-	periods := make(map[string]time.Duration, len(s.outputs)+1)
	139	+	if s.framePeriods == nil {
	140	+	s.framePeriods = make(map[string]time.Duration, len(s.outputs)+1)
	141	+	}
	142	+	periods := s.framePeriods
	143	+	clear(periods)
134	144		for _, o := range s.outputs {
135	145		periods[o.name()] = framePeriod(o.place.Info.RefreshMilli)
136	146		}

 	timer := time.NewTimer(defaultFramePeriod)
 	defer timer.Stop()
 	presented := s.channels.Presented
+	if s.flipped == nil {
+		s.flipped = make(map[string]bool)
+	}
 	for {
-		flipped := map[string]bool{}
+		clear(s.flipped)
-		var reports []ports.OutputPresented
+		flipped := s.flipped
+		s.frameReports = s.frameReports[:0]
+		reports := s.frameReports
 		select {
 		case <-ctx.Done():
 			return
 				break drain
 			}
 		}
+		s.frameReports = reports
 		wait, idle := defaultFramePeriod, false
 		if !s.display.Do(func() {
 			if ctx.Err() == nil {
 // their deadline. wait is the time to the next deadline; idle is set when
 // no callback waits after firing.
 func (s *Server) dueFrames(now time.Time, flipped map[string]bool) (fire []string, wait time.Duration, idle bool) {
-	periods := make(map[string]time.Duration, len(s.outputs)+1)
+	if s.framePeriods == nil {
+		s.framePeriods = make(map[string]time.Duration, len(s.outputs)+1)
+	}
+	periods := s.framePeriods
+	clear(periods)
 	for _, o := range s.outputs {
 		periods[o.name()] = framePeriod(o.place.Info.RefreshMilli)
 	}

M internal/adapters/wayland/frames_test.go +11 -0


@@ -9,6 +9,17 @@ import (
9	9		"github.com/bnema/purego-libwayland/protocol/wayland"
10	10		)
11	11
	12	+	func TestDueFramesAllocations(t *testing.T) {
	13	+	s := &Server{frameDue: map[string]time.Time{}, awaiting: map[string][]*wayland.Callback{}}
	14	+	now := time.Unix(100, 0)
	15	+	s.dueFrames(now, nil)
	16	+	if allocs := testing.AllocsPerRun(100, func() { s.dueFrames(now, nil) }); allocs != 0 {
	17	+	t.Errorf("idle dueFrames: %.1f allocs, want 0", allocs)
	18	+	} else {
	19	+	t.Logf("idle dueFrames: %.1f allocs", allocs)
	20	+	}
	21	+	}
	22	+
12	23		func TestFramePeriod(t *testing.T) {
13	24		for _, tc := range []struct {
14	25		milli int


@@ -9,6 +9,17 @@ import (
9	"github.com/bnema/purego-libwayland/protocol/wayland"	9		"github.com/bnema/purego-libwayland/protocol/wayland"
10	)	10		)
11		11
		12	+	func TestDueFramesAllocations(t *testing.T) {
		13	+	s := &Server{frameDue: map[string]time.Time{}, awaiting: map[string][]*wayland.Callback{}}
		14	+	now := time.Unix(100, 0)
		15	+	s.dueFrames(now, nil)
		16	+	if allocs := testing.AllocsPerRun(100, func() { s.dueFrames(now, nil) }); allocs != 0 {
		17	+	t.Errorf("idle dueFrames: %.1f allocs, want 0", allocs)
		18	+	} else {
		19	+	t.Logf("idle dueFrames: %.1f allocs", allocs)
		20	+	}
		21	+	}
		22	+
12	func TestFramePeriod(t *testing.T) {	23		func TestFramePeriod(t *testing.T) {
13	for _, tc := range []struct {	24		for _, tc := range []struct {
14	milli int	25		milli int

M internal/adapters/wayland/input_region.go +26 -24


@@ -34,33 +34,10 @@ func (g *region) boxRect() ports.Rect {
34	34
35	35		// effectiveInput includes all committed, bounded subsurface regions.
36	36		func (s *surface) effectiveInput() (bool, []ports.Rect) {
37		-	var rects []ports.Rect
38		-	var visit func(*surface, int, int)
39		-	visit = func(v *surface, x, y int) {
40		-	if v.has {
41		-	bounds := ports.Rect{W: v.content.LogicalW, H: v.content.LogicalH}
42		-	if v.inputAll {
43		-	if bounds.W > 0 && bounds.H > 0 {
44		-	rects = append(rects, ports.Rect{X: x, Y: y, W: bounds.W, H: bounds.H})
45		-	}
46		-	} else {
47		-	for _, r := range v.inputRects {
48		-	if r = intersectRect(r, bounds); r.W > 0 {
49		-	r.X += x
50		-	r.Y += y
51		-	rects = append(rects, r)
52		-	}
53		-	}
54		-	}
55		-	}
56		-	for _, item := range v.sub.layout {
57		-	visit(item.child, x+item.x, y+item.y)
58		-	}
59		-	}
60	37		if s.inputAll && len(s.sub.children) == 0 {
61	38		return true, nil
62	39		}
63		-	visit(s, 0, 0)
	40	+	rects := s.appendEffectiveInput(nil, 0, 0)
64	41		if s.xdg != nil {
65	42		for i := range rects {
66	43		rects[i].X -= s.xdg.geometry.X

@@ -70,6 +47,31 @@ func (s *surface) effectiveInput() (bool, []ports.Rect) {
70	47		return false, rects
71	48		}
72	49
	50	+	// appendEffectiveInput walks the committed tree without a per-call recursive
	51	+	// closure. The returned slice belongs to the caller (emitInput publishes it).
	52	+	func (s *surface) appendEffectiveInput(rects []ports.Rect, x, y int) []ports.Rect {
	53	+	if s.has {
	54	+	bounds := ports.Rect{W: s.content.LogicalW, H: s.content.LogicalH}
	55	+	if s.inputAll {
	56	+	if bounds.W > 0 && bounds.H > 0 {
	57	+	rects = append(rects, ports.Rect{X: x, Y: y, W: bounds.W, H: bounds.H})
	58	+	}
	59	+	} else {
	60	+	for _, r := range s.inputRects {
	61	+	if r = intersectRect(r, bounds); r.W > 0 {
	62	+	r.X += x
	63	+	r.Y += y
	64	+	rects = append(rects, r)
	65	+	}
	66	+	}
	67	+	}
	68	+	}
	69	+	for _, item := range s.sub.layout {
	70	+	rects = item.child.appendEffectiveInput(rects, x+item.x, y+item.y)
	71	+	}
	72	+	return rects
	73	+	}
	74	+
73	75		// resetInputEmission makes the next mapping publish its current region again.
74	76		func (s *surface) resetInputEmission() {
75	77		s.sentInput = false

 // effectiveInput includes all committed, bounded subsurface regions.
 func (s *surface) effectiveInput() (bool, []ports.Rect) {
-	var rects []ports.Rect
-	var visit func(*surface, int, int)
-	visit = func(v *surface, x, y int) {
-		if v.has {
-			bounds := ports.Rect{W: v.content.LogicalW, H: v.content.LogicalH}
-			if v.inputAll {
-				if bounds.W > 0 && bounds.H > 0 {
-					rects = append(rects, ports.Rect{X: x, Y: y, W: bounds.W, H: bounds.H})
-				}
-			} else {
-				for _, r := range v.inputRects {
-					if r = intersectRect(r, bounds); r.W > 0 {
-						r.X += x
-						r.Y += y
-						rects = append(rects, r)
-					}
-				}
-			}
-		}
-		for _, item := range v.sub.layout {
-			visit(item.child, x+item.x, y+item.y)
-		}
-	}
 	if s.inputAll && len(s.sub.children) == 0 {
 		return true, nil
 	}
-	visit(s, 0, 0)
+	rects := s.appendEffectiveInput(nil, 0, 0)
 	if s.xdg != nil {
 		for i := range rects {
 			rects[i].X -= s.xdg.geometry.X
 	return false, rects
 }
+// appendEffectiveInput walks the committed tree without a per-call recursive
+// closure. The returned slice belongs to the caller (emitInput publishes it).
+func (s *surface) appendEffectiveInput(rects []ports.Rect, x, y int) []ports.Rect {
+	if s.has {
+		bounds := ports.Rect{W: s.content.LogicalW, H: s.content.LogicalH}
+		if s.inputAll {
+			if bounds.W > 0 && bounds.H > 0 {
+				rects = append(rects, ports.Rect{X: x, Y: y, W: bounds.W, H: bounds.H})
+			}
+		} else {
+			for _, r := range s.inputRects {
+				if r = intersectRect(r, bounds); r.W > 0 {
+					r.X += x
+					r.Y += y
+					rects = append(rects, r)
+				}
+			}
+		}
+	}
+	for _, item := range s.sub.layout {
+		rects = item.child.appendEffectiveInput(rects, x+item.x, y+item.y)
+	}
+	return rects
+}
 // resetInputEmission makes the next mapping publish its current region again.
 func (s *surface) resetInputEmission() {
 	s.sentInput = false

M internal/adapters/wayland/server.go +5 -0


@@ -71,6 +71,11 @@ type Server struct {
71	71		awaiting map[string][]*wayland.Callback
72	72		frameDue map[string]time.Time
73	73		frameReady chan struct{}
	74	+	// Pacer scratch: periods is used on the display goroutine, while
	75	+	// flipped and frameReports belong to pace and its synchronous Do call.
	76	+	framePeriods map[string]time.Duration
	77	+	flipped map[string]bool
	78	+	frameReports []ports.OutputPresented
74	79		// held buffers wait until no output reads them (release.go); reports
75	80		// are the outputs' latest ports.OutputPresented.
76	81		held []heldBuffer


@@ -71,6 +71,11 @@ type Server struct {
71	awaiting map[string][]*wayland.Callback	71		awaiting map[string][]*wayland.Callback
72	frameDue map[string]time.Time	72		frameDue map[string]time.Time
73	frameReady chan struct{}	73		frameReady chan struct{}
		74	+	// Pacer scratch: periods is used on the display goroutine, while
		75	+	// flipped and frameReports belong to pace and its synchronous Do call.
		76	+	framePeriods map[string]time.Duration
		77	+	flipped map[string]bool
		78	+	frameReports []ports.OutputPresented
74	// held buffers wait until no output reads them (release.go); reports	79		// held buffers wait until no output reads them (release.go); reports
75	// are the outputs' latest ports.OutputPresented.	80		// are the outputs' latest ports.OutputPresented.
76	held []heldBuffer	81		held []heldBuffer

M internal/adapters/wayland/surface_perf_test.go +51 -6


@@ -9,9 +9,43 @@ import (
9	9		"github.com/bnema/neferwl/internal/ports"
10	10		)
11	11
	12	+	// TestEffectiveInputEmptyTreeAllocations guards against a heap-allocated
	13	+	// recursive closure for every input-region traversal of a tiled tree.
	14	+	func TestEffectiveInputEmptyTreeAllocations(t *testing.T) {
	15	+	root := &surface{}
	16	+	for range 64 {
	17	+	child := &surface{}
	18	+	root.sub.children = append(root.sub.children, child)
	19	+	root.sub.layout = append(root.sub.layout, childLayout{child: child})
	20	+	}
	21	+	walk := func() {
	22	+	all, rects := root.effectiveInput()
	23	+	if all \|\| len(rects) != 0 {
	24	+	t.Fatal("empty tree has input")
	25	+	}
	26	+	}
	27	+	if allocs := testing.AllocsPerRun(100, walk); allocs != 0 {
	28	+	t.Errorf("empty tiled input traversal: %.1f allocs, want 0", allocs)
	29	+	}
	30	+	}
	31	+
12	32		// TestTiledCommitPublishAllocations exercises real content publication without
13	33		// a client connection, holding old snapshots while subsequent commits publish.
14	34		func TestTiledCommitPublishAllocations(t *testing.T) {
	35	+	// The top-level test is listed in Makefile perf-check; this subtest guards
	36	+	// the per-commit update storage after its first retirement.
	37	+	t.Run("queueUpdateSteadyState", func(t *testing.T) {
	38	+	srv := &Server{fifoSurfaces: make(map[*surface]struct{}), frameReady: make(chan struct{}, 1)}
	39	+	surf := &surface{server: srv}
	40	+	commit := func() {
	41	+	surf.queueUpdate()
	42	+	surf.dropQueue()
	43	+	}
	44	+	commit()
	45	+	if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
	46	+	t.Errorf("steady-state queueUpdate/dropQueue: %.1f allocs/commit; want 0", allocs)
	47	+	}
	48	+	})
15	49		for _, count := range []int{16, 64} {
16	50		t.Run(fmt.Sprintf("%d", count), func(t *testing.T) {
17	51		channels := Channels{Contents: make(chan ports.SurfaceContent, 1)}

@@ -35,8 +69,8 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
35	69		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 {
36	70		t.Fatalf("previous publication changed: %+v", first)
37	71		}
38		-	if allocs := testing.AllocsPerRun(100, publish); allocs > 4 {
39		-	t.Errorf("%d children: %.1f allocs/publication; want <=4", count, allocs)
	72	+	if allocs := testing.AllocsPerRun(100, publish); allocs > 2 {
	73	+	t.Errorf("%d children: %.1f allocs/publication; want <=2", count, allocs)
40	74		} else {
41	75		t.Logf("%d children: %.1f allocs/unchanged publication", count, allocs)
42	76		}

@@ -44,8 +78,19 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
44	78		// proportionally to the number of children.
45	79		root.sub.pendingLayout = root.sub.layout
46	80		queue := func() { _ = root.takePending() }
47		-	if allocs := testing.AllocsPerRun(100, queue); allocs > 2 {
48		-	t.Errorf("%d children: %.1f allocs/queued commit; want <=2", count, allocs)
	81	+	if allocs := testing.AllocsPerRun(100, queue); allocs > 0 {
	82	+	t.Errorf("%d children: %.1f allocs/queued commit; want 0", count, allocs)
	83	+	} else {
	84	+	t.Logf("%d children: %.1f allocs/takePending", count, allocs)
	85	+	}
	86	+	commit := func() {
	87	+	root.queueUpdate()
	88	+	root.dropQueue()
	89	+	}
	90	+	if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
	91	+	t.Errorf("%d children: %.1f allocs/queueUpdate; want 0", count, allocs)
	92	+	} else {
	93	+	t.Logf("%d children: %.1f allocs/queueUpdate", count, allocs)
49	94		}
50	95		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 \|\| first.DamageHistory[0].Seq != 1 {
51	96		t.Fatal("previous publication was mutated")

@@ -68,8 +113,8 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
68	113		root.treeDirty = true
69	114		root.redraw()
70	115		}
71		-	if allocs := testing.AllocsPerRun(100, changed); allocs > 4 {
72		-	t.Errorf("%d children: %.1f allocs/changed publication; want <=4", count, allocs)
	116	+	if allocs := testing.AllocsPerRun(100, changed); allocs > 3 {
	117	+	t.Errorf("%d children: %.1f allocs/changed publication; want <=3", count, allocs)
73	118		} else {
74	119		t.Logf("%d children: %.1f allocs/changed publication", count, allocs)
75	120		}


@@ -9,9 +9,43 @@ import (
9		"github.com/bnema/neferwl/internal/ports"	9		"github.com/bnema/neferwl/internal/ports"
10		)	10		)
11			11
			12	+	// TestEffectiveInputEmptyTreeAllocations guards against a heap-allocated
			13	+	// recursive closure for every input-region traversal of a tiled tree.
			14	+	func TestEffectiveInputEmptyTreeAllocations(t *testing.T) {
			15	+	root := &surface{}
			16	+	for range 64 {
			17	+	child := &surface{}
			18	+	root.sub.children = append(root.sub.children, child)
			19	+	root.sub.layout = append(root.sub.layout, childLayout{child: child})
			20	+	}
			21	+	walk := func() {
			22	+	all, rects := root.effectiveInput()
			23	+	if all \|\| len(rects) != 0 {
			24	+	t.Fatal("empty tree has input")
			25	+	}
			26	+	}
			27	+	if allocs := testing.AllocsPerRun(100, walk); allocs != 0 {
			28	+	t.Errorf("empty tiled input traversal: %.1f allocs, want 0", allocs)
			29	+	}
			30	+	}
			31	+
12		// TestTiledCommitPublishAllocations exercises real content publication without	32		// TestTiledCommitPublishAllocations exercises real content publication without
13		// a client connection, holding old snapshots while subsequent commits publish.	33		// a client connection, holding old snapshots while subsequent commits publish.
14		func TestTiledCommitPublishAllocations(t *testing.T) {	34		func TestTiledCommitPublishAllocations(t *testing.T) {
			35	+	// The top-level test is listed in Makefile perf-check; this subtest guards
			36	+	// the per-commit update storage after its first retirement.
			37	+	t.Run("queueUpdateSteadyState", func(t *testing.T) {
			38	+	srv := &Server{fifoSurfaces: make(map[*surface]struct{}), frameReady: make(chan struct{}, 1)}
			39	+	surf := &surface{server: srv}
			40	+	commit := func() {
			41	+	surf.queueUpdate()
			42	+	surf.dropQueue()
			43	+	}
			44	+	commit()
			45	+	if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
			46	+	t.Errorf("steady-state queueUpdate/dropQueue: %.1f allocs/commit; want 0", allocs)
			47	+	}
			48	+	})
15		for _, count := range []int{16, 64} {	49		for _, count := range []int{16, 64} {
16		t.Run(fmt.Sprintf("%d", count), func(t *testing.T) {	50		t.Run(fmt.Sprintf("%d", count), func(t *testing.T) {
17		channels := Channels{Contents: make(chan ports.SurfaceContent, 1)}	51		channels := Channels{Contents: make(chan ports.SurfaceContent, 1)}

@@ -35,8 +69,8 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
35		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 {	69		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 {
36		t.Fatalf("previous publication changed: %+v", first)	70		t.Fatalf("previous publication changed: %+v", first)
37		}	71		}
38	-	if allocs := testing.AllocsPerRun(100, publish); allocs > 4 {	72	+	if allocs := testing.AllocsPerRun(100, publish); allocs > 2 {
39	-	t.Errorf("%d children: %.1f allocs/publication; want <=4", count, allocs)	73	+	t.Errorf("%d children: %.1f allocs/publication; want <=2", count, allocs)
40		} else {	74		} else {
41		t.Logf("%d children: %.1f allocs/unchanged publication", count, allocs)	75		t.Logf("%d children: %.1f allocs/unchanged publication", count, allocs)
42		}	76		}

@@ -44,8 +78,19 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
44		// proportionally to the number of children.	78		// proportionally to the number of children.
45		root.sub.pendingLayout = root.sub.layout	79		root.sub.pendingLayout = root.sub.layout
46		queue := func() { _ = root.takePending() }	80		queue := func() { _ = root.takePending() }
47	-	if allocs := testing.AllocsPerRun(100, queue); allocs > 2 {	81	+	if allocs := testing.AllocsPerRun(100, queue); allocs > 0 {
48	-	t.Errorf("%d children: %.1f allocs/queued commit; want <=2", count, allocs)	82	+	t.Errorf("%d children: %.1f allocs/queued commit; want 0", count, allocs)
			83	+	} else {
			84	+	t.Logf("%d children: %.1f allocs/takePending", count, allocs)
			85	+	}
			86	+	commit := func() {
			87	+	root.queueUpdate()
			88	+	root.dropQueue()
			89	+	}
			90	+	if allocs := testing.AllocsPerRun(100, commit); allocs != 0 {
			91	+	t.Errorf("%d children: %.1f allocs/queueUpdate; want 0", count, allocs)
			92	+	} else {
			93	+	t.Logf("%d children: %.1f allocs/queueUpdate", count, allocs)
49		}	94		}
50		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 \|\| first.DamageHistory[0].Seq != 1 {	95		if first.Seq != 1 \|\| len(first.Children) != count \|\| len(first.DamageHistory) != 1 \|\| first.DamageHistory[0].Seq != 1 {
51		t.Fatal("previous publication was mutated")	96		t.Fatal("previous publication was mutated")

@@ -68,8 +113,8 @@ func TestTiledCommitPublishAllocations(t *testing.T) {
68		root.treeDirty = true	113		root.treeDirty = true
69		root.redraw()	114		root.redraw()
70		}	115		}
71	-	if allocs := testing.AllocsPerRun(100, changed); allocs > 4 {	116	+	if allocs := testing.AllocsPerRun(100, changed); allocs > 3 {
72	-	t.Errorf("%d children: %.1f allocs/changed publication; want <=4", count, allocs)	117	+	t.Errorf("%d children: %.1f allocs/changed publication; want <=3", count, allocs)
73		} else {	118		} else {
74		t.Logf("%d children: %.1f allocs/changed publication", count, allocs)	119		t.Logf("%d children: %.1f allocs/changed publication", count, allocs)
75		}	120		}