display: pace the frame clock by the panel's EDID refresh rate

Both EDID moments the system has are now captured and carried to the
compositor's frame clock:

- EFI: the loader derives refresh from the preferred detailed timing
  (pixel clock / total pixels) while GOP is still alive — the only moment
  it is readable — and hands it through the boot handoff into the
  display0 node's DisplayInfo (new refresh_hz field, 0 = unknown).
- GPU: the virtio-gpu driver derives the same figure from its own EDID
  read and carries it in the attach_scanout announce (request.y).

updateFrameClock() re-derives the interval from the active backend's
info at bring-up and again on every backend change — the boot
framebuffer's clock dies with the GOP floor at upgrade, replaced by the
GPU's rate. Unknown rate defaults to 60 Hz; the result is clamped to
[30, 120] Hz so a mis-parsed EDID can neither starve nor flood the
compositor. Rate only, never phase: without vblank, presents still
free-run (docs/display-v2.md, 'Fenced is not vsync').

Observed in QEMU: OVMF exposes no EDID for the VGA adapter, so the GOP
floor logs 'frame clock 62 Hz (default)' (real firmware does expose it);
the virtio-gpu EDID advertises 75 Hz and the upgrade logs 'frame clock
76 Hz (panel EDID)'. The display kernel test asserts refresh_hz rides
the seeded node; all 7 display QEMU cases pass.

boot/efi.zig

@@ -97,7 +97,7 @@ fn boot() !noreturn {
 }
 
 /// A display resolution in pixels.
-const Resolution = struct { width: u32, height: u32 };
+const Resolution = struct { width: u32, height: u32, refresh_hz: u32 };
 
 /// Switch the GPU to the monitor's native resolution (when we can determine it)
 /// and read the resulting graphics mode into our own framebuffer description.
@@ -128,6 +128,10 @@ fn queryFramebuffer(bs: *uefi.tables.BootServices) !boot_handoff.Framebuffer {
         // Each pixel is 32 bits, so the byte pitch is 4 * pixels-per-row.
         .pitch = info.pixels_per_scan_line * 4,
         .format = try pixelFormat(info.pixel_format),
+        // The refresh rate rides the EDID preferred timing. If the firmware kept a
+        // non-native mode it may not describe that mode exactly — but it is the panel's
+        // own clock, a far better frame-clock seed than a hardcoded 60 Hz.
+        .refresh_hz = if (native) |n| n.refresh_hz else 0,
     };
 }
 
@@ -176,10 +180,12 @@ fn nativeResolution(bs: *uefi.tables.BootServices, handles: []uefi.Handle) ?Reso
     return null;
 }
 
-/// Parse the native resolution from a raw EDID block. The first Detailed Timing
-/// Descriptor (at byte 54) is the preferred — i.e. native — mode by convention;
-/// its active pixel counts are split across low bytes and the high nibbles of
-/// later bytes.
+/// Parse the native resolution and refresh rate from a raw EDID block. The first
+/// Detailed Timing Descriptor (at byte 54) is the preferred — i.e. native — mode by
+/// convention; its active pixel counts are split across low bytes and the high nibbles
+/// of later bytes. The refresh rate is derived, not stored: the descriptor carries the
+/// pixel clock (10 kHz units) and the active+blanking extents, and
+/// refresh = clock / (horizontal total × vertical total).
 fn edidNative(edid: []const u8) ?Resolution {
     if (edid.len < 128) return null;
     // Every EDID begins with this fixed 8-byte header.
@@ -193,7 +199,13 @@ fn edidNative(edid: []const u8) ?Resolution {
     const w = @as(u32, dtd[2]) | (@as(u32, dtd[4] & 0xf0) << 4);
     const h = @as(u32, dtd[5]) | (@as(u32, dtd[7] & 0xf0) << 4);
     if (w == 0 or h == 0) return null;
-    return .{ .width = w, .height = h };
+
+    const clock_hz = (@as(u64, dtd[0]) | (@as(u64, dtd[1]) << 8)) * 10_000;
+    const h_blank = @as(u64, dtd[3]) | (@as(u64, dtd[4] & 0x0f) << 8);
+    const v_blank = @as(u64, dtd[6]) | (@as(u64, dtd[7] & 0x0f) << 8);
+    const total = (@as(u64, w) + h_blank) * (@as(u64, h) + v_blank);
+    const refresh: u32 = if (total == 0) 0 else @intCast((clock_hz + total / 2) / total);
+    return .{ .width = w, .height = h, .refresh_hz = refresh };
 }
 
 /// Open the kernel on the volume we booted from, read it into a pool buffer,

build.zig

@@ -850,6 +850,53 @@ pub fn build(b: *std.Build) void {
     const run_efi_step = b.step("run-x86-64", "Boot the x86-64 kernel in QEMU (UEFI/OVMF); serial0 is logged to zig-out/qemu-test/run-x86-64-serial0-<timestamp>.log");
     run_efi_step.dependOn(&run_efi.step);
 
+    // --- run-x86-64-gpu: the same boot plus a virtio-gpu adapter ---
+    // The VGA device still supplies the boot (GOP) framebuffer the compositor starts
+    // on; the virtio-gpu function is discovered by the device-manager stack, its
+    // driver announces a shared scanout, and the compositor upgrades off the GOP
+    // floor to fenced, tear-free native presents (docs/display-v2.md).
+    // This is the interactive twin of the `display-native` test case, and 512M
+    // matches it (the whole driver stack + the compositor's surfaces at once).
+    // QEMU shows one head per adapter: pick the virtio-gpu head in the View menu
+    // to watch the native output.
+    const run_gpu = b.addSystemCommand(&.{
+        "qemu-system-x86_64",
+        "-device",
+        "qemu-xhci,id=xhci",
+        "-device",
+        "usb-mouse,bus=xhci.0",
+        "-device",
+        "usb-kbd,bus=xhci.0",
+        "-machine",
+        "q35",
+        "-m",
+        "512M",
+        "-drive",
+        b.fmt("if=pflash,format=raw,readonly=on,file={s}", .{ovmf_code}),
+    });
+    run_gpu.addArg("-drive");
+    run_gpu.addPrefixedFileArg("if=pflash,format=raw,file=", vars_out);
+    run_gpu.addArg("-drive");
+    run_gpu.addPrefixedFileArg("if=none,id=bootusb,format=raw,file=", fat_image_serial);
+    run_gpu.addArgs(&.{
+        "-device",
+        "usb-storage,bus=xhci.0,drive=bootusb,removable=on,bootindex=0",
+        "-net",
+        "none",
+        "-vga",
+        "none",
+        "-device",
+        "VGA,edid=on,xres=1280,yres=720",
+        "-device",
+        "virtio-gpu-pci",
+    });
+    const gpu_serial_log = b.fmt("{s}/run-x86-64-gpu-serial0-{s}.log", .{ log_dir, timestamp(b) });
+    run_gpu.addArgs(&.{ "-serial", b.fmt("file:{s}", .{gpu_serial_log}) });
+    run_gpu.step.dependOn(&make_log_dir.step);
+
+    const run_gpu_step = b.step("run-x86-64-gpu", "Boot in QEMU with a virtio-gpu adapter: the compositor upgrades to fenced (tear-free) native presents; watch the virtio-gpu head in QEMU's View menu");
+    run_gpu_step.dependOn(&run_gpu.step);
+
     // const run_cmd = b.addRunArtifact(exe);
     // const run_step = b.step("run", "Run the app");
     // run_step.dependOn(&run_cmd.step);

docs/display-v2-plan.md

@@ -6,7 +6,7 @@ lands on its own and ends in a **verifiable gate** — shaped for a `/loop` run,
 
 ## Locked decisions (do not relitigate)
 
-- **First native backend = virtio-gpu** (VM standard: mode-set + present/flush + vsync).
+- **First native backend = virtio-gpu** (VM standard: mode-set + fenced present/flush).
 - **Dynamic hot-attach**: boot on GOP, upgrade to native when the driver **announces**
   (push, not polling); re-attach across driver restarts; GOP is the floor for "no driver
   ever," not a live fall-back after a reprogram.
@@ -46,7 +46,7 @@ Extract scanout from the compositor so today's path becomes one backend among fu
 
 - [x] `system/services/display/backend.zig`: a `Backend` tagged union with `info()`,
       `surface()` (the cacheable compose target), `present(damage)`, and capability flags
-      (`canModeSet`/`hasVsync`, both false for GOP).
+      (`canModeSet`/`hasFencedPresent`, both false for GOP).
 - [x] The v1 GOP path is now `backend.Gop` (claims the `display` node, WC-maps the LFB,
       keeps the cacheable back buffer, `present` = the damage-rect WC copy). display.zig
       composes into `backend.surface()` and calls `backend.present(damage)` — no LFB or
@@ -123,7 +123,7 @@ confirm the composited frame landed (`display: native present verified`), while
 ok` still fires — checked order-independently. Without `-device virtio-gpu-pci` nothing is
 announced and it stays on GOP: the v1 `display-service`/`display-demo` gates pass unchanged.
 
-## V5 — Mode-setting, EDID, and vsync ✅
+## V5 — Mode-setting, EDID, and fenced presents ✅
 
 - [x] The driver negotiates `VIRTIO_GPU_F_EDID` (when offered) and reads the monitor's EDID,
       logging its preferred mode; it offers a small mode list over `.scanout` `get_modes`. The
@@ -131,14 +131,16 @@ announced and it stays on GOP: the v1 `display-service`/`display-demo` gates pas
       scanout rectangle (no resource/surface churn) — a runtime resolution change. `runtime.display`
       gains `modes()` / `setMode()` (display-protocol `get_modes`/`set_mode`, forwarded to the backend).
 - [x] Every `resource_flush` is issued fenced (`VIRTIO_GPU_FLAG_FENCE`); the device signals the
-      fence when the frame is on screen, which the used-ring ack the synchronous present waits on
-      already gates — a tear-free present.
-- [x] `backend.VirtioGpu` reports `canModeSet` / `hasVsync` = true.
+      fence when it has consumed the frame, which the used-ring ack the synchronous present waits
+      on already gates — a tear-free present. (Completion feedback, **not vblank**: base
+      virtio-gpu 2D has no display-refresh event, so nothing paces presents to the monitor —
+      see the "Fenced is not vsync" note in [display-v2.md](display-v2.md).)
+- [x] `backend.VirtioGpu` reports `canModeSet` / `hasFencedPresent` = true.
 
 **Gate (met):** the `display-modeset` case (reusing the display-native boot) upgrades to
 virtio-gpu, queries the driver's modes, `setMode`s to a different resolution, and confirms the
 change by reading the backend's geometry back (`display: mode set to {w}x{h}, verified`); the
-fenced present path is exercised and confirmed (`display: vsync present ok`) — both from serial,
+fenced present path is exercised and confirmed (`display: fenced present ok`) — both from serial,
 passing 3/3. The driver also logs the EDID preferred mode (`virtio-gpu: EDID preferred mode …`).
 
 ## V6 — Resilience (restart + re-attach) + tests + docs ✅

docs/display-v2.md

@@ -2,7 +2,7 @@
 
 **Status: complete (V1–V6).** The compositor boots on the GOP framebuffer and, when a
 virtio-gpu driver announces itself, hot-attaches a native backend over the shared `shm`
-scanout surface — with runtime mode-setting, EDID, and fenced (vsync) presents, and it
+scanout surface — with runtime mode-setting, EDID, and fenced presents, and it
 re-attaches across driver restarts. All serial-gated (see [display-v2-plan.md](display-v2-plan.md)).
 
 v1 ([display.md](display.md)) is a compositor that owns the **GOP framebuffer** — it
@@ -25,10 +25,10 @@ The compositor itself (layers, back buffer, damage) does not change. Only the la
    scanout backend  (selected at runtime — GOP by default, native when it appears)
         │
         ├─ GopBackend        the v1 path: WC copy back→front to the firmware LFB.
-        │                    Always available. No mode-set, no vsync. THE FLOOR.
+        │                    Always available. No mode-set, no present fence. THE FLOOR.
         │
         └─ VirtioGpuBackend  talks to a virtio-gpu driver process over a `scanout`
-                             service: present via a shared resource + flush (real vsync),
+                             service: present via a shared resource + fenced flush,
                              EDID mode list, runtime mode-set.
 ```
 
@@ -37,8 +37,8 @@ A **backend** is a small interface the compositor calls:
 - `surface()` → the pixels to compose into and their geometry `{ptr, pitch, format, w, h}`
   (the LFB for GOP; a shared scanout resource for virtio-gpu),
 - `present(damage: Rect)` → make the damaged region visible (a no-op-ish WC copy for GOP;
-  a virtio flush, optionally vsync-fenced, for the native path),
-- capability queries — `canModeSet`, `hasVsync` — and, when supported, `modes()` /
+  a fenced virtio flush for the native path),
+- capability queries — `canModeSet`, `hasFencedPresent` — and, when supported, `modes()` /
   `setMode(m)`.
 
 The compositor composes into `surface()` and calls `present(damage)` exactly as it does
@@ -98,23 +98,31 @@ compositor when a second backend arrives"). It claims the virtio-gpu PCI functio
   at a chosen mode for **runtime mode-setting**,
 - registers a `scanout` service and announces to the display service.
 
-Its `resource_flush` is the real **present** — and gives a genuine **vsync/tear-free**
-path a dumb GOP framebuffer can't.
+Its `resource_flush` is the real **present** — and gives a **fenced, tear-free** path a
+dumb GOP framebuffer can't.
+
+**Fenced is not vsync.** The fence completes when the device has *consumed* the frame:
+real completion feedback, and tear-freedom by snapshot semantics (the host displays
+discrete transferred frames, never a half-written surface). It is **not** a vblank —
+base virtio-gpu 2D has no display-refresh event at all (Linux's driver for this device
+fakes one with a software timer), so nothing paces presents to the monitor's refresh.
+Refresh-paced presents need either a native driver's vblank interrupt (delivered over
+the existing IRQ-as-IPC path) or the compositor's own frame clock.
 
 ## What v2 unlocks — and its honest scope
 
 Behind the abstraction, a native backend gives runtime **mode-setting** (resolution /
-refresh / bpp), **EDID** enumeration, and **vsync**. But only on devices we have a driver
+refresh / bpp), **EDID** enumeration, and **fenced presents**. But only on devices we have a driver
 for — realistically **VMs** (virtio-gpu, and later maybe Bochs DISPI). Real discrete GPUs
 need per-vendor KMS-class drivers that aren't getting written, so they **stay on GOP** —
 which is genuinely fine (v1 on the NVIDIA box is smooth). So v2's real value is twofold:
-the **pluggable architecture** (a driver slots in when one exists) and a **rich, vsync'd
+the **pluggable architecture** (a driver slots in when one exists) and a **rich, fenced
 path in VMs**, where danos development happens. The framebuffer floor never goes away.
 
 ## Locked decisions
 
 - **First native backend: virtio-gpu** — the VM standard; gives mode-set + a real
-  present/flush (and vsync), and exercises the whole pluggable design. Tested with QEMU
+  present/flush (fenced), and exercises the whole pluggable design. Tested with QEMU
   `-device virtio-gpu`.
 - **Dynamic hot-attach** — boot on GOP, upgrade to native on the driver's announce,
   re-attach across driver restarts; GOP is the floor for "no driver ever," not a live

docs/display.md

@@ -196,13 +196,21 @@ shell, a terminal, a cursor, and a wallpaper:
 | `fill_rect`        | fill a rectangle of a layer with a colour                     |
 | `blit_tile`        | copy a small client-supplied pixel tile into a layer (inline) |
 | `damage`           | mark a region of a layer dirty                                |
-| `present`          | composite dirty layers and flush to the screen                |
+| `present`          | request a repaint: composited at the next frame-clock tick    |
 
 Text is intentionally *not* an operation — a client renders glyphs by blitting tiles
 (the [PSF font](../system/kernel/font.psf) path the console already uses can move into a
 client). Keeping the protocol to rectangles and tiles keeps the compositor small and the
 policy in the client.
 
+`present` is a *request*, not an immediate flush: the compositor runs a ~60 Hz **frame
+clock** (a one-shot kernel timer re-armed on demand), and each tick composites all the
+damage accumulated since the last one. Any number of client presents and cursor moves
+inside one interval coalesce into a single repaint — the software stand-in for vblank
+pacing on backends that have none (all of them today; see
+[display-v2.md](display-v2.md), "Fenced is not vsync"). Bring-up paths that must put
+pixels on screen synchronously (initialisation, the self-checks) bypass the clock.
+
 ## `runtime.display`
 
 Clients speak the protocol through a new [`library/runtime/display.zig`](../library/runtime/runtime.zig),

system/boot-handoff.zig

@@ -37,6 +37,11 @@ pub const Framebuffer = extern struct {
     height: u32, // visible rows (e.g. 1080)
     pitch: u32, // bytes from the start of one row to the start of the next
     format: PixelFormat,
+    /// The panel's refresh rate in Hz, computed from its EDID preferred timing (pixel
+    /// clock / total pixels per frame) while GOP was still alive — the one moment it is
+    /// readable (docs/gop.md). 0 = unknown (no EDID). The display service paces its
+    /// frame clock by it; without vblank this fixes the *rate*, never the *phase*.
+    refresh_hz: u32 = 0,
 
     /// Whether a usable framebuffer was handed over.
     pub fn present(self: Framebuffer) bool {

system/devices/device-abi.zig

@@ -97,6 +97,7 @@ pub const DisplayInfo = extern struct {
     height: u32 = 0, // visible rows
     pitch: u32 = 0, // bytes from one row's start to the next
     format: u32 = 0, // a DisplayFormat value
+    refresh_hz: u32 = 0, // panel refresh rate from EDID (0 = unknown); see boot-handoff
 };
 
 /// `DeviceDescriptor.parent` for a device with no parent — a root of the device tree.

system/drivers/virtio-gpu/virtio-gpu.zig

@@ -53,13 +53,18 @@ const offered_modes = [_]Mode{ .{ .width = 640, .height = 480 }, .{ .width = 800
 var current_width: u32 = offered_modes[0].width;
 var current_height: u32 = offered_modes[0].height;
 
-/// Monotonic fence id for fenced (vsync) flushes; the device signals the fence when the flush
-/// is complete, which its used-ring ack already gates our synchronous present on.
+/// Monotonic fence id for fenced flushes; the device signals the fence when the flush is
+/// complete, which its used-ring ack already gates our synchronous present on. Completion
+/// feedback, not vblank — nothing here is paced to the display's refresh.
 var fence_next: u64 = 1;
 
 /// Whether the device offered VIRTIO_GPU_F_EDID, so `get_edid` is worth issuing.
 var edid_available = false;
 
+/// The panel refresh rate parsed from the EDID preferred timing (0 = unknown). Carried to
+/// the compositor in the announce so its frame clock paces to the panel, not a guess.
+var edid_refresh_hz: u32 = 0;
+
 /// The control virtqueue. We drive it synchronously — one command, notify, poll the used
 /// ring — so a depth of 16 is ample; we ask the device to shrink to it (virtio 1.0 lets the
 /// driver reduce queue_size), keeping the whole ring inside one page.
@@ -467,10 +472,18 @@ fn readEdid() void {
     }
     // The first detailed timing descriptor (EDID base-block offset 54) is the preferred mode:
     // active pixels are 12-bit, low byte + high nibble (bytes 2/4 horizontal, 5/7 vertical).
+    // The refresh rate is derived from the same descriptor: pixel clock (bytes 0-1, 10 kHz
+    // units) over total (active + blanking) pixels per frame — the loader does the identical
+    // computation for the boot framebuffer (boot/efi.zig edidNative).
     const e = &response.edid;
     const h_active = @as(u32, e[56]) | (@as(u32, e[58] & 0xF0) << 4);
     const v_active = @as(u32, e[59]) | (@as(u32, e[61] & 0xF0) << 4);
-    log("virtio-gpu: EDID preferred mode {d}x{d}\n", .{ h_active, v_active });
+    const clock_hz = (@as(u64, e[54]) | (@as(u64, e[55]) << 8)) * 10_000;
+    const h_blank = @as(u64, e[57]) | (@as(u64, e[58] & 0x0F) << 8);
+    const v_blank = @as(u64, e[60]) | (@as(u64, e[61] & 0x0F) << 8);
+    const total = (@as(u64, h_active) + h_blank) * (@as(u64, v_active) + v_blank);
+    if (total != 0) edid_refresh_hz = @intCast((clock_hz + total / 2) / total);
+    log("virtio-gpu: EDID preferred mode {d}x{d} @ {d} Hz\n", .{ h_active, v_active, edid_refresh_hz });
 }
 
 /// Present the whole surface: copy the guest backing into the host resource, then flush it to
@@ -492,8 +505,9 @@ fn presentFull() bool {
         if (command_nodata(@sizeOf(vg.TransferToHost2d)) != ok_nodata) return false;
     }
     {
-        // A fenced flush (vsync): the device signals the fence when the frame is actually on
-        // screen — which its used-ring ack, what our synchronous submit waits on, already gates.
+        // A fenced flush: the device signals the fence once it has consumed the frame — which
+        // its used-ring ack, what our synchronous submit waits on, already gates. Completion
+        // feedback and a tear-free snapshot, not vblank pacing.
         const request = requestAt(vg.ResourceFlush);
         request.* = .{
             .hdr = .{ .type = @intFromEnum(vg.CmdType.resource_flush), .flags = vg.flag_fence, .fence_id = fence_next },
@@ -548,6 +562,7 @@ fn announce() void {
     var request = dp.Request{
         .operation = @intFromEnum(dp.Operation.attach_scanout),
         .x = max_width, // the shared surface's row stride in pixels (it is sized to the max mode)
+        .y = edid_refresh_hz, // the panel refresh from EDID (0 = unknown) — the frame-clock seed
         .width = current_width,
         .height = current_height,
         .colour = display_format_bgrx,

system/kernel/devices-broker.zig

@@ -61,7 +61,7 @@ pub fn init(device_tree: *const platform.DeviceTree) void {
 /// [[boot-handoff]], not the device tree), so it is seeded explicitly, after `init`.
 /// Returns the new device id, or null when there is no framebuffer (headless) or the
 /// table is full. Idempotent-ish: only ever call once per boot.
-pub fn seedDisplay(base: u64, width: u32, height: u32, pitch: u32, format: u32) ?u64 {
+pub fn seedDisplay(base: u64, width: u32, height: u32, pitch: u32, format: u32, refresh_hz: u32) ?u64 {
     if (base == 0 or width == 0 or height == 0) return null; // headless
     if (count >= maximum_devices) {
         dropped += 1;
@@ -79,7 +79,7 @@ pub fn seedDisplay(base: u64, width: u32, height: u32, pitch: u32, format: u32)
         .len = @as(u64, height) * pitch,
         .flags = device_abi.resource_flag_write_combining,
     };
-    d.display = .{ .width = width, .height = height, .pitch = pitch, .format = format };
+    d.display = .{ .width = width, .height = height, .pitch = pitch, .format = format, .refresh_hz = refresh_hz };
     devices[count] = d;
     display_device = d.id;
     count += 1;

system/kernel/kernel.zig

@@ -200,8 +200,8 @@ fn kmain(boot_information: *const BootInformation) noreturn {
         // Publish the loader's framebuffer as a claimable `display` device, so a
         // user-space display service can take it over the same claim + mmio_map path as
         // any other hardware (it is not firmware-discovered; it rides the boot handoff).
-        if (devices_broker.seedDisplay(fb.base, fb.width, fb.height, fb.pitch, @intFromEnum(fb.format))) |display_id| {
-            log.print("/system/kernel: framebuffer device {d} seeded ({d}x{d}, pitch {d}, write-combining)\n", .{ display_id, fb.width, fb.height, fb.pitch });
+        if (devices_broker.seedDisplay(fb.base, fb.width, fb.height, fb.pitch, @intFromEnum(fb.format), fb.refresh_hz)) |display_id| {
+            log.print("/system/kernel: framebuffer device {d} seeded ({d}x{d}, pitch {d}, {d} Hz, write-combining)\n", .{ display_id, fb.width, fb.height, fb.pitch, fb.refresh_hz });
         }
 
         // Install the device-IRQ trampolines, so a driver's irq_bind has vectors to

system/kernel/tests.zig

@@ -3404,6 +3404,7 @@ fn displayTest(boot_information: *const BootInformation) void {
 
     check("the node is class display", d.class == @intFromEnum(device_abi.DeviceClass.display));
     check("it carries the framebuffer geometry", d.display.width == fb.width and d.display.height == fb.height and d.display.pitch == fb.pitch);
+    check("it carries the panel refresh rate", d.display.refresh_hz == fb.refresh_hz);
     check("it has exactly one resource", d.resource_count == 1);
     const r = d.resources[0];
     check("that resource is a memory window", r.kind == @intFromEnum(device_abi.ResourceKind.memory));

system/services/display/backend.zig

@@ -16,8 +16,10 @@ const scanout_protocol = runtime.scanout_protocol;
 const Rect = compositor.Rect;
 const Surface = compositor.Surface;
 
-/// The current display mode, as a backend reports it.
-pub const Info = struct { width: u32, height: u32, pitch: u32, format: u32 };
+/// The current display mode, as a backend reports it. `refresh_hz` is the panel's
+/// refresh rate from EDID (0 = unknown) — the frame clock's pacing seed; without vblank
+/// it fixes the rate, never the phase (docs/display-v2.md, "Fenced is not vsync").
+pub const Info = struct { width: u32, height: u32, pitch: u32, format: u32, refresh_hz: u32 };
 
 /// Enumeration scratch — a `DeviceDescriptor` is large, and only one scan is ever needed.
 var device_table: [64]device.DeviceDescriptor = undefined;
@@ -26,7 +28,7 @@ var device_table: [64]device.DeviceDescriptor = undefined;
 /// framebuffer write-combining as the front buffer, and keeps a cacheable back buffer of
 /// the same geometry as the compose target. `present` streams the damaged rectangle from
 /// the back buffer to the LFB (sequential WC writes; the LFB is never read). No mode-set,
-/// no vsync — the portable floor (docs/display-v2.md).
+/// no present fence — the portable floor (docs/display-v2.md).
 pub const Gop = struct {
     device_id: u64,
     front: [*]volatile u8, // the LFB (write-combining)
@@ -35,13 +37,14 @@ pub const Gop = struct {
     height: u32,
     pitch: u32,
     format: u32,
+    refresh_hz: u32, // from the boot EDID via the display0 node (0 = unknown)
 
     /// The framebuffer's id and geometry, captured together. `findDisplay` reads these out of
     /// the enumeration table and returns them by value, so the caller never re-reads the table
     /// across later syscalls (`device_enumerate` writes the whole table straight into this
     /// process's memory; reading a descriptor's tail again after other syscalls have run is a
     /// window we simply avoid by copying the few fields we need up front).
-    const Found = struct { id: u64, width: u32, height: u32, pitch: u32, format: u32 };
+    const Found = struct { id: u64, width: u32, height: u32, pitch: u32, format: u32, refresh_hz: u32 };
 
     /// The first `display`-class device with a *valid* (non-zero) geometry, or null. A zero
     /// geometry is treated as "not ready yet" so the caller retries — a real framebuffer always
@@ -52,7 +55,7 @@ pub const Gop = struct {
         for (device_table[0..n]) |*d| {
             if (d.class != @intFromEnum(device.DeviceClass.display)) continue;
             if (d.display.width == 0 or d.display.height == 0 or d.display.pitch == 0) continue;
-            return .{ .id = d.id, .width = d.display.width, .height = d.display.height, .pitch = d.display.pitch, .format = d.display.format };
+            return .{ .id = d.id, .width = d.display.width, .height = d.display.height, .pitch = d.display.pitch, .format = d.display.format, .refresh_hz = d.display.refresh_hz };
         }
         return null;
     }
@@ -93,11 +96,12 @@ pub const Gop = struct {
             .height = found.height,
             .pitch = found.pitch,
             .format = found.format,
+            .refresh_hz = found.refresh_hz,
         };
     }
 
     pub fn info(self: *const Gop) Info {
-        return .{ .width = self.width, .height = self.height, .pitch = self.pitch, .format = self.format };
+        return .{ .width = self.width, .height = self.height, .pitch = self.pitch, .format = self.format, .refresh_hz = self.refresh_hz };
     }
 
     /// The cacheable compose target (the back buffer).
@@ -110,22 +114,48 @@ pub const Gop = struct {
         };
     }
 
-    /// Stream the damaged rectangle from the back buffer to the write-combining LFB, row by
-    /// row (sequential writes — what WC memory wants; the LFB is never read).
-    pub fn present(self: *const Gop, damage: Rect) void {
-        const c = damage.intersect(.{ .x = 0, .y = 0, .w = @intCast(self.width), .h = @intCast(self.height) });
-        if (c.isEmpty()) return;
+    /// Stream each damaged rectangle from the back buffer to the write-combining LFB, row
+    /// by row (sequential writes — what WC memory wants; the LFB is never read). The rows
+    /// are copied by `presentSpan` below, which widens the stores by hand: `volatile`
+    /// keeps the compiler from eliding or reordering framebuffer writes, but it also
+    /// forbids it from merging them, so a naive per-pixel loop is stuck at one 4-byte
+    /// store per iteration. Keeping each copy small (the damage list) and each store wide
+    /// shrinks the window in which scanout can sample a half-written frame.
+    pub fn present(self: *const Gop, damage: []const Rect) void {
+        const bounds = Rect{ .x = 0, .y = 0, .w = @intCast(self.width), .h = @intCast(self.height) };
+        for (damage) |rect| {
+            const c = rect.intersect(bounds);
+            if (c.isEmpty()) continue;
+            const span: usize = @intCast(c.w);
             var y: i32 = c.y;
             while (y < c.bottom()) : (y += 1) {
-            const off = @as(usize, @intCast(y)) * self.pitch;
-            const src: [*]const u32 = @ptrCast(@alignCast(self.back + off));
-            const dst: [*]volatile u32 = @ptrCast(@alignCast(self.front + off));
-            var x: i32 = c.x;
-            while (x < c.right()) : (x += 1) dst[@intCast(x)] = src[@intCast(x)];
+                const offset = @as(usize, @intCast(y)) * self.pitch + @as(usize, @intCast(c.x)) * 4;
+                const source: [*]const u32 = @ptrCast(@alignCast(self.back + offset));
+                const front_row: [*]volatile u32 = @ptrCast(@alignCast(self.front + offset));
+                presentSpan(front_row, source, span);
+            }
         }
     }
 };
 
+/// Copy `count` pixels into the write-combining front buffer with 8-byte volatile stores
+/// (plus a 4-byte head/tail where the span isn't 8-aligned — pixel spans are always
+/// 4-aligned). The loads come from the cacheable back buffer and are assembled into a
+/// `u64` in registers, so nothing here reads the front buffer.
+fn presentSpan(destination: [*]volatile u32, source: [*]const u32, count: usize) void {
+    var i: usize = 0;
+    if (i < count and (@intFromPtr(destination) & 7) != 0) {
+        destination[0] = source[0];
+        i = 1;
+    }
+    while (i + 2 <= count) : (i += 2) {
+        const pair = @as(u64, source[i]) | (@as(u64, source[i + 1]) << 32);
+        const wide: *volatile u64 = @ptrCast(@alignCast(destination + i));
+        wide.* = pair;
+    }
+    if (i < count) destination[i] = source[i];
+}
+
 /// A display mode the native backend can switch to.
 pub const Mode = scanout_protocol.Mode;
 
@@ -143,17 +173,19 @@ pub const VirtioGpu = struct {
     width: u32, // the active mode
     height: u32,
     format: u32,
+    refresh_hz: u32, // from the driver's EDID read, carried in the announce (0 = unknown)
     scanout: ipc.Handle, // the driver's present + mode channel (looked up on `.scanout`)
 
     pub fn info(self: *const VirtioGpu) Info {
-        return .{ .width = self.width, .height = self.height, .pitch = self.stride * 4, .format = self.format };
+        return .{ .width = self.width, .height = self.height, .pitch = self.stride * 4, .format = self.format, .refresh_hz = self.refresh_hz };
     }
     pub fn surface(self: *const VirtioGpu) Surface {
         return .{ .pixels = self.pixels, .stride = self.stride, .width = self.width, .height = self.height };
     }
     /// Ask the driver to present. The composited pixels are already in the shared surface, so
-    /// this is a single request over `.scanout`; the driver transfers + fenced-flushes.
-    pub fn present(self: *const VirtioGpu, damage: Rect) void {
+    /// this is a single request over `.scanout` regardless of how many damage rectangles
+    /// accumulated; the driver transfers + fenced-flushes the whole frame.
+    pub fn present(self: *const VirtioGpu, damage: []const Rect) void {
         _ = damage;
         var request = scanout_protocol.Request{
             .operation = @intFromEnum(scanout_protocol.Operation.present),
@@ -210,7 +242,7 @@ pub const Backend = union(enum) {
             inline else => |*b| b.surface(),
         };
     }
-    pub fn present(self: *const Backend, damage: Rect) void {
+    pub fn present(self: *const Backend, damage: []const Rect) void {
         switch (self.*) {
             inline else => |*b| b.present(damage),
         }
@@ -236,9 +268,13 @@ pub const Backend = union(enum) {
             .virtio => true,
         };
     }
-    /// Whether this backend has a vblank/fence for tear-free present (virtio-gpu: yes, V5 — every
-    /// flush is fenced, so the device signals completion when the frame is actually on screen).
-    pub fn hasVsync(self: *const Backend) bool {
+    /// Whether this backend's present is **fenced** — it completes only once the device has
+    /// consumed the frame (virtio-gpu: every flush carries a fence the used-ring ack waits on).
+    /// A fence gives completion feedback and tear-free snapshot presents; it is *not* vblank —
+    /// nothing paces presents to the display's refresh (base virtio-gpu 2D has no vblank event
+    /// at all). True vsync needs a native driver's vblank interrupt. See docs/display-v2.md,
+    /// "Fenced is not vsync".
+    pub fn hasFencedPresent(self: *const Backend) bool {
         return switch (self.*) {
             .gop => false,
             .virtio => true,

system/services/display/compositor.zig

@@ -57,6 +57,177 @@ pub const Rect = struct {
     }
 };
 
+/// The dirty screen regions accumulated between presents. Kept as a *list* of rectangles,
+/// not one bounding box: when two small things move far apart — the cursor on one side of
+/// the screen, an animating layer on the other — a single bounding box unites them into a
+/// huge region, and presenting it streams megabytes to the framebuffer for a few thousand
+/// changed pixels. The long copy widens the window in which scanout (or QEMU's display
+/// refresh) samples a half-written frame — visible as tearing and cursor trails. Small
+/// separate rectangles keep each copy, and that window, tight.
+///
+/// A new rectangle that overlaps an existing entry is united into it (repainting a modest
+/// superset is harmless — compositing is idempotent); the grown entry is *not* re-merged
+/// against the rest, so entries may overlap, which costs only a duplicate repaint. When
+/// the table is full the newcomer folds into the last entry — degrading toward the old
+/// bounding-box behaviour instead of dropping damage.
+pub const DamageList = struct {
+    pub const capacity = 16;
+
+    rects: [capacity]Rect = [_]Rect{Rect.empty} ** capacity,
+    count: usize = 0,
+
+    pub fn add(self: *DamageList, r: Rect) void {
+        if (r.isEmpty()) return;
+        for (self.rects[0..self.count]) |*existing| {
+            if (!existing.intersect(r).isEmpty()) {
+                existing.* = existing.unite(r);
+                return;
+            }
+        }
+        if (self.count < capacity) {
+            self.rects[self.count] = r;
+            self.count += 1;
+            return;
+        }
+        self.rects[capacity - 1] = self.rects[capacity - 1].unite(r);
+    }
+
+    pub fn isEmpty(self: *const DamageList) bool {
+        return self.count == 0;
+    }
+
+    pub fn slice(self: *const DamageList) []const Rect {
+        return self.rects[0..self.count];
+    }
+
+    pub fn clear(self: *DamageList) void {
+        self.count = 0;
+    }
+};
+
+/// The alternative damage tracker: a **fixed tile grid**, the scheme browser compositors
+/// and tile-based GPUs use. The screen is divided into `tile_size`-pixel tiles up front;
+/// `add` marks the tiles a rectangle touches (a bit per tile — merging is free and exact,
+/// no heuristics), and `collect` walks the grid turning runs of adjacent dirty tiles into
+/// repaint rectangles (horizontal runs, then equal-span rows merged vertically, so
+/// full-screen damage collapses back to a single rectangle).
+///
+/// Trade-off against `DamageList`: tracking is O(1) with a strictly bounded worst case
+/// (never more than the dirty tiles), but repaints are quantized — a 1-pixel change
+/// repaints a whole tile. Which wins depends on the workload; the display service has a
+/// compile-time switch (`damage_mode`) to compare them.
+pub const TileGrid = struct {
+    pub const tile_size = 64;
+    pub const maximum_columns = 128; // supports screens up to 8192 px wide…
+    pub const maximum_rows = 128; // …and 8192 px tall (beyond that, edge tiles stretch)
+    pub const maximum_tiles = maximum_columns * maximum_rows;
+    /// The most rectangles `collect` produces; extras fold into the last (never dropped).
+    pub const maximum_rects = 64;
+
+    width: u32 = 0,
+    height: u32 = 0,
+    columns: u32 = 0,
+    rows: u32 = 0,
+    dirty_count: u32 = 0,
+    dirty: [maximum_tiles]bool = [_]bool{false} ** maximum_tiles,
+
+    /// Size the grid for a screen. Also clears it — callers reset on a geometry change,
+    /// where the mode-set paths damage the whole new screen anyway.
+    pub fn reset(self: *TileGrid, width: u32, height: u32) void {
+        self.width = width;
+        self.height = height;
+        self.columns = @min((width + tile_size - 1) / tile_size, maximum_columns);
+        self.rows = @min((height + tile_size - 1) / tile_size, maximum_rows);
+        self.clear();
+    }
+
+    pub fn matches(self: *const TileGrid, width: u32, height: u32) bool {
+        return self.width == width and self.height == height;
+    }
+
+    pub fn isEmpty(self: *const TileGrid) bool {
+        return self.dirty_count == 0;
+    }
+
+    pub fn clear(self: *TileGrid) void {
+        @memset(&self.dirty, false);
+        self.dirty_count = 0;
+    }
+
+    /// Mark every tile `r` touches. Clips to the screen first, so out-of-range
+    /// rectangles are harmless.
+    pub fn add(self: *TileGrid, r: Rect) void {
+        const screen = Rect{ .x = 0, .y = 0, .w = @intCast(self.width), .h = @intCast(self.height) };
+        const c = r.intersect(screen);
+        if (c.isEmpty()) return;
+        const column_first: u32 = @intCast(@divTrunc(c.x, tile_size));
+        const row_first: u32 = @intCast(@divTrunc(c.y, tile_size));
+        const column_last: u32 = @min(@as(u32, @intCast(@divTrunc(c.right() - 1, tile_size))), self.columns - 1);
+        const row_last: u32 = @min(@as(u32, @intCast(@divTrunc(c.bottom() - 1, tile_size))), self.rows - 1);
+        var row = row_first;
+        while (row <= row_last) : (row += 1) {
+            var column = column_first;
+            while (column <= column_last) : (column += 1) {
+                const index = row * self.columns + column;
+                if (!self.dirty[index]) {
+                    self.dirty[index] = true;
+                    self.dirty_count += 1;
+                }
+            }
+        }
+    }
+
+    /// The screen rectangle covered by tiles [column_first, column_end) of `row`. Edge
+    /// tiles clamp to the true screen size (the last column/row may be partial — or, on a
+    /// screen wider than the grid supports, stretched to cover the remainder).
+    fn tileSpanRect(self: *const TileGrid, column_first: u32, column_end: u32, row: u32) Rect {
+        const x: i32 = @intCast(column_first * tile_size);
+        const y: i32 = @intCast(row * tile_size);
+        const right: i32 = if (column_end >= self.columns) @intCast(self.width) else @intCast(column_end * tile_size);
+        const bottom: i32 = if (row + 1 >= self.rows) @intCast(self.height) else @intCast((row + 1) * tile_size);
+        return .{ .x = x, .y = y, .w = right - x, .h = bottom - y };
+    }
+
+    /// Turn the dirty tiles into repaint rectangles in `out`: coalesce each row's runs of
+    /// adjacent dirty tiles, then merge a run into the rectangle directly above it when
+    /// the spans match — so a dirty block of tiles becomes one rectangle. Returns the
+    /// filled prefix of `out`.
+    pub fn collect(self: *const TileGrid, out: []Rect) []Rect {
+        var count: usize = 0;
+        var row: u32 = 0;
+        while (row < self.rows) : (row += 1) {
+            var column: u32 = 0;
+            while (column < self.columns) {
+                if (!self.dirty[row * self.columns + column]) {
+                    column += 1;
+                    continue;
+                }
+                var run_end = column + 1;
+                while (run_end < self.columns and self.dirty[row * self.columns + run_end]) run_end += 1;
+                const rect = self.tileSpanRect(column, run_end, row);
+                column = run_end;
+
+                var merged = false;
+                for (out[0..count]) |*existing| {
+                    if (existing.x == rect.x and existing.w == rect.w and existing.bottom() == rect.y) {
+                        existing.h += rect.h;
+                        merged = true;
+                        break;
+                    }
+                }
+                if (merged) continue;
+                if (count < out.len) {
+                    out[count] = rect;
+                    count += 1;
+                } else {
+                    out[count - 1] = out[count - 1].unite(rect);
+                }
+            }
+        }
+        return out[0..count];
+    }
+};
+
 /// A block of 32-bit pixels: `pixels` addressed row-major with `stride` pixels between
 /// row starts (≥ width — the framebuffer's stride is pitch/4, a layer's is its width).
 pub const Surface = struct {
@@ -74,15 +245,17 @@ pub const Surface = struct {
     }
 };
 
-/// Fill `rect` of `s` with the native pixel `colour`, clipped to `s`'s bounds.
+/// Fill `rect` of `s` with the native pixel `colour`, clipped to `s`'s bounds. Each row is
+/// one `@memset` over the clipped span, so the compiler vectorizes it and the bounds check
+/// runs once per row, not once per pixel.
 pub fn fillRect(s: Surface, rect: Rect, colour: u32) void {
     const c = rect.intersect(s.bounds());
     if (c.isEmpty()) return;
+    const x0: usize = @intCast(c.x);
+    const span: usize = @intCast(c.w);
     var y: i32 = c.y;
     while (y < c.bottom()) : (y += 1) {
-        const r = s.row(@intCast(y));
-        var x: i32 = c.x;
-        while (x < c.right()) : (x += 1) r[@intCast(x)] = colour;
+        @memset((s.row(@intCast(y)) + x0)[0..span], colour);
     }
 }
 
@@ -94,35 +267,36 @@ pub fn composite(dst: Surface, dx: i32, dy: i32, layer: Surface, clip: Rect) voi
     const on_screen = Rect{ .x = dx, .y = dy, .w = @intCast(layer.width), .h = @intCast(layer.height) };
     const region = on_screen.intersect(clip).intersect(dst.bounds());
     if (region.isEmpty()) return;
+    const span: usize = @intCast(region.w);
+    const dst_x: usize = @intCast(region.x);
+    const src_x: usize = @intCast(region.x - dx);
     var y: i32 = region.y;
     while (y < region.bottom()) : (y += 1) {
-        const src = layer.row(@intCast(y - dy));
-        const d = dst.row(@intCast(y));
-        var x: i32 = region.x;
-        while (x < region.right()) : (x += 1) {
-            d[@intCast(x)] = src[@intCast(x - dx)];
-        }
+        const source_row = layer.row(@intCast(y - dy)) + src_x;
+        const destination_row = dst.row(@intCast(y)) + dst_x;
+        @memcpy(destination_row[0..span], source_row[0..span]);
     }
 }
 
 /// Copy a `w`×`h` tile of native pixels from `src` (raw little-endian bytes, row-major,
-/// tightly packed) into `dst` at (`dx`, `dy`), clipped to `dst`'s bounds. `src` is read
-/// with `readInt` because it comes straight out of an IPC message buffer and carries no
-/// alignment guarantee. Returns without touching anything if `src` is short.
+/// tightly packed) into `dst` at (`dx`, `dy`), clipped to `dst`'s bounds. `src` comes
+/// straight out of an IPC message buffer and carries no alignment guarantee, so each
+/// clipped row is a byte-wise `@memcpy` — which equals the old per-pixel little-endian
+/// `readInt` on every danos target (all little-endian) without the alignment concern.
+/// Returns without touching anything if `src` is short.
 pub fn blitTile(dst: Surface, dx: i32, dy: i32, src: []const u8, w: u32, h: u32) void {
     if (src.len < @as(usize, w) * h * 4) return;
-    var ty: u32 = 0;
-    while (ty < h) : (ty += 1) {
-        const yy = dy + @as(i32, @intCast(ty));
-        if (yy < 0 or yy >= dst.height) continue;
-        const drow = dst.row(@intCast(yy));
-        var tx: u32 = 0;
-        while (tx < w) : (tx += 1) {
-            const xx = dx + @as(i32, @intCast(tx));
-            if (xx < 0 or xx >= dst.width) continue;
-            const off = (@as(usize, ty) * w + tx) * 4;
-            drow[@intCast(xx)] = std.mem.readInt(u32, src[off..][0..4], .little);
-        }
+    const region = Rect.init(dx, dy, @intCast(w), @intCast(h)).intersect(dst.bounds());
+    if (region.isEmpty()) return;
+    const span: usize = @intCast(region.w);
+    const tile_x: usize = @intCast(region.x - dx);
+    const dst_x: usize = @intCast(region.x);
+    var y: i32 = region.y;
+    while (y < region.bottom()) : (y += 1) {
+        const tile_y: usize = @intCast(y - dy);
+        const offset = (tile_y * w + tile_x) * 4;
+        const destination_row = dst.row(@intCast(y)) + dst_x;
+        @memcpy(std.mem.sliceAsBytes(destination_row[0..span]), src[offset..][0 .. span * 4]);
     }
 }
 
@@ -179,6 +353,92 @@ test "composite honours the damage rectangle" {
     try std.testing.expectEqual(@as(u32, 0), back[4 * 8 + 4]); // outside damage
 }
 
+test "damage list keeps disjoint rectangles separate and merges overlap" {
+    var list = DamageList{};
+    list.add(Rect.init(0, 0, 10, 10));
+    list.add(Rect.init(100, 100, 10, 10)); // far away: its own entry
+    try std.testing.expectEqual(@as(usize, 2), list.slice().len);
+    list.add(Rect.init(5, 5, 10, 10)); // overlaps the first: united into it
+    try std.testing.expectEqual(@as(usize, 2), list.slice().len);
+    try std.testing.expectEqual(Rect.init(0, 0, 15, 15), list.slice()[0]);
+    try std.testing.expect(!list.isEmpty());
+    list.clear();
+    try std.testing.expect(list.isEmpty());
+}
+
+test "damage list folds overflow into the last entry instead of dropping it" {
+    var list = DamageList{};
+    var i: i32 = 0;
+    while (i < DamageList.capacity) : (i += 1) {
+        list.add(Rect.init(i * 100, 0, 10, 10)); // disjoint: fills every slot
+    }
+    try std.testing.expectEqual(@as(usize, DamageList.capacity), list.slice().len);
+    const overflow = Rect.init(0, 5000, 10, 10);
+    list.add(overflow);
+    try std.testing.expectEqual(@as(usize, DamageList.capacity), list.slice().len);
+    const last = list.slice()[DamageList.capacity - 1];
+    try std.testing.expect(!last.intersect(overflow).isEmpty()); // still covered
+}
+
+test "damage list ignores empty rectangles" {
+    var list = DamageList{};
+    list.add(Rect.empty);
+    try std.testing.expect(list.isEmpty());
+}
+
+test "tile grid coalesces a run of adjacent tiles into one rectangle" {
+    var grid = TileGrid{};
+    grid.reset(256, 128); // 4×2 tiles of 64 px
+    grid.add(Rect.init(10, 10, 100, 10)); // spans tiles (0,0) and (1,0)
+    var scratch: [TileGrid.maximum_rects]Rect = undefined;
+    const rects = grid.collect(&scratch);
+    try std.testing.expectEqual(@as(usize, 1), rects.len);
+    try std.testing.expectEqual(Rect.init(0, 0, 128, 64), rects[0]);
+}
+
+test "tile grid: full-screen damage collapses back to a single rectangle" {
+    var grid = TileGrid{};
+    grid.reset(1280, 720); // 20×12 tiles; the bottom row is partial (720 = 11*64 + 16)
+    grid.add(Rect.init(0, 0, 1280, 720));
+    var scratch: [TileGrid.maximum_rects]Rect = undefined;
+    const rects = grid.collect(&scratch);
+    try std.testing.expectEqual(@as(usize, 1), rects.len);
+    try std.testing.expectEqual(Rect.init(0, 0, 1280, 720), rects[0]);
+}
+
+test "tile grid keeps far-apart damage as separate rectangles" {
+    var grid = TileGrid{};
+    grid.reset(1280, 720);
+    grid.add(Rect.init(0, 0, 10, 10)); // top-left tile
+    grid.add(Rect.init(1000, 600, 10, 10)); // a far-away tile
+    var scratch: [TileGrid.maximum_rects]Rect = undefined;
+    const rects = grid.collect(&scratch);
+    try std.testing.expectEqual(@as(usize, 2), rects.len);
+}
+
+test "tile grid clamps edge tiles to the true screen size" {
+    var grid = TileGrid{};
+    grid.reset(100, 100); // 2×2 tiles, both partial in each axis
+    grid.add(Rect.init(0, 0, 100, 100));
+    var scratch: [TileGrid.maximum_rects]Rect = undefined;
+    const rects = grid.collect(&scratch);
+    try std.testing.expectEqual(@as(usize, 1), rects.len);
+    try std.testing.expectEqual(Rect.init(0, 0, 100, 100), rects[0]);
+}
+
+test "tile grid clear empties it and reset resizes it" {
+    var grid = TileGrid{};
+    grid.reset(256, 256);
+    grid.add(Rect.init(0, 0, 256, 256));
+    try std.testing.expect(!grid.isEmpty());
+    grid.clear();
+    try std.testing.expect(grid.isEmpty());
+    try std.testing.expect(grid.matches(256, 256));
+    grid.reset(512, 512);
+    try std.testing.expect(!grid.matches(256, 256));
+    try std.testing.expect(grid.isEmpty());
+}
+
 test "blitTile copies a packed tile, clipping and reading unaligned bytes" {
     var back = [_]u32{0} ** (4 * 4);
     const dst = Surface{ .pixels = &back, .stride = 4, .width = 4, .height = 4 };

system/services/display/display.zig

@@ -10,7 +10,9 @@
 //! z-order, and visibility. Clients create layers, draw into them by command (`fill_rect`,
 //! `blit_tile`), mark `damage`, and ask for a `present`; the compositor repaints only the
 //! damaged region — clear it, paint the visible layers bottom-to-top into the backend's
-//! surface, then `backend.present(damage)`. Shared-memory client surfaces are later
+//! surface, then `backend.present(damage)`. Presents are paced by a ~60 Hz **frame clock**
+//! (see `schedulePresent`), so any number of client presents and cursor moves inside one
+//! interval coalesce into a single frame. Shared-memory client surfaces are later
 //! (docs/display-v2.md).
 
 const std = @import("std");
@@ -50,8 +52,10 @@ var pending_modeset_check: bool = false;
 var background: u32 = 0;
 
 /// The layer stack. A fixed table (a compositor has few top-level surfaces during
-/// bring-up); each used slot owns an mmap'd surface. `damage` accumulates the dirty
-/// screen region since the last `present`, so a present touches only what changed.
+/// bring-up); each used slot owns an mmap'd surface. `damage_list` accumulates the dirty
+/// screen rectangles since the last `present`, so a present touches only what changed —
+/// and keeps far-apart changes (the cursor here, an animating layer there) as *separate*
+/// small copies rather than one huge bounding box (see compositor.DamageList).
 const maximum_layers = 16;
 
 const Layer = struct {
@@ -65,7 +69,67 @@ const Layer = struct {
 };
 
 var layers: [maximum_layers]Layer = [_]Layer{.{}} ** maximum_layers;
-var damage: Rect = Rect.empty;
+
+/// Which damage tracker drives `present` — a compile-time A/B switch (both are in
+/// compositor.zig with the trade-off discussion):
+///   .list — free-form dirty rectangles (tight bounds, heuristic merging)
+///   .grid — a fixed 64-px tile grid (exact O(1) merging, tile-quantized repaints)
+const DamageMode = enum { list, grid };
+const damage_mode: DamageMode = .grid;
+
+var damage_list: compositor.DamageList = .{};
+var damage_grid: compositor.TileGrid = .{};
+
+/// The **frame clock**: client `present` requests and cursor motion don't repaint
+/// immediately — they accumulate damage and arm a one-shot timer, and the tick composites
+/// everything pending as one frame. That paces presents to ~60 Hz no matter how fast
+/// clients draw or the mouse moves (previously every mouse event became a full present).
+/// No backend has a real vblank to pace by (docs/display-v2.md, "Fenced is not vsync");
+/// this is the software stand-in, the same strategy Linux uses atop virtio-gpu. Bring-up
+/// paths that need pixels on screen *now* (initialise, the self-checks) still call
+/// `present()` directly.
+///
+/// The interval comes from the *active backend's* panel refresh rate (EDID: the loader
+/// captures it for the GOP floor while firmware still runs; the native driver reads its
+/// own and carries it in the announce). `updateFrameClock` re-derives it whenever the
+/// backend changes — the boot framebuffer's clock dies with the GOP floor at upgrade.
+/// Without a rate the clock defaults to 60 Hz, and it is clamped to [30, 120] Hz so a
+/// mis-parsed EDID can neither starve nor flood the compositor.
+var frame_interval_milliseconds: u64 = 16;
+var frame_timer_armed = false;
+
+/// Derive the frame-clock interval from the active backend's refresh rate and log what
+/// the clock is now pacing to. Called at bring-up and again on every backend change.
+fn updateFrameClock() void {
+    const reported = backend.info().refresh_hz;
+    const rate: u64 = if (reported == 0) 60 else @min(@max(reported, 30), 120);
+    frame_interval_milliseconds = @max(1000 / rate, 1);
+    var line: [96]u8 = undefined;
+    _ = system.write(std.fmt.bufPrint(&line, "display: frame clock {d} Hz ({s})\n", .{
+        1000 / frame_interval_milliseconds,
+        if (reported == 0) "default" else "panel EDID",
+    }) catch return);
+}
+
+/// Arm the frame clock unless a tick is already pending: any number of requests inside
+/// one interval coalesce into that single tick's present.
+fn schedulePresent() void {
+    if (frame_timer_armed) return;
+    frame_timer_armed = true;
+    _ = system.timerOnce(service_endpoint, frame_interval_milliseconds);
+}
+
+/// A timer landing — the frame clock, or the deferred first native present armed by
+/// `attach_scanout`: present the accumulated damage, then run the one-shot mode-set
+/// self-check if the native upgrade queued it.
+fn frameTick() void {
+    frame_timer_armed = false;
+    present();
+    if (pending_modeset_check) {
+        pending_modeset_check = false;
+        modesetSelfCheck();
+    }
+}
 
 // --- geometry helpers -------------------------------------------------------
 
@@ -78,9 +142,20 @@ fn layerScreenRect(l: *const Layer) Rect {
     return .{ .x = l.x, .y = l.y, .w = @intCast(l.surface.width), .h = @intCast(l.surface.height) };
 }
 
-/// Add `r` (screen coordinates) to the pending damage, clipped to the screen.
+/// Add `r` (screen coordinates) to the pending damage, clipped to the screen. In grid
+/// mode the grid re-sizes itself lazily when the screen geometry changes — every
+/// geometry-changing path (`attach_scanout`, `set_mode`) damages the whole new screen
+/// right after, so damage pending from the old geometry is safely superseded.
 fn addDamage(r: Rect) void {
-    damage = damage.unite(r.intersect(screenRect()));
+    const clipped = r.intersect(screenRect());
+    switch (damage_mode) {
+        .list => damage_list.add(clipped),
+        .grid => {
+            const mode = backend.info();
+            if (!damage_grid.matches(mode.width, mode.height)) damage_grid.reset(mode.width, mode.height);
+            damage_grid.add(clipped);
+        },
+    }
 }
 
 // --- layer operations (called from onMessage and the self-check) ------------
@@ -183,21 +258,29 @@ fn compositeInto(clip: Rect) void {
     }
 }
 
-/// Composite the accumulated damage into the backend's surface, hand it to the backend to
-/// put on screen, then clear the damage. A no-op when nothing is dirty. The frame counter
-/// advances regardless, so callers can name frames.
+/// Composite each accumulated damage rectangle into the backend's surface, hand the list
+/// to the backend to put on screen, then clear the damage. A no-op when nothing is dirty.
+/// The frame counter advances regardless, so callers can name frames.
 fn present() void {
-    const dirty = damage.intersect(screenRect());
-    if (!dirty.isEmpty()) {
-        compositeInto(dirty);
+    var scratch: [compositor.TileGrid.maximum_rects]Rect = undefined;
+    const dirty: []const Rect = switch (damage_mode) {
+        .list => damage_list.slice(),
+        .grid => damage_grid.collect(&scratch),
+    };
+    const had_damage = dirty.len != 0;
+    if (had_damage) {
+        for (dirty) |region| compositeInto(region);
         backend.present(dirty);
     }
-    damage = Rect.empty;
+    switch (damage_mode) {
+        .list => damage_list.clear(),
+        .grid => damage_grid.clear(),
+    }
     frames += 1;
 
     // The first present after a native upgrade confirms the composited frame actually reached
     // the shared scanout surface (the automated stand-in for "it's on screen").
-    if (pending_native_verify and !dirty.isEmpty()) {
+    if (pending_native_verify and had_damage) {
         pending_native_verify = false;
         verifyNativePresent();
     }
@@ -221,7 +304,7 @@ fn verifyNativePresent() void {
 /// present channel, switch the backend to virtio-gpu, and queue a full-screen repaint. The
 /// present is deferred to a timer (see `service_endpoint`) so it happens after this reply
 /// unblocks the driver and it starts serving `.scanout`.
-fn attachScanout(stride: u32, width: u32, height: u32, format: u32, capability: ?ipc.Handle, reply: []u8) usize {
+fn attachScanout(stride: u32, width: u32, height: u32, format: u32, refresh_hz: u32, capability: ?ipc.Handle, reply: []u8) usize {
     const cap = capability orelse return fail(reply);
     if (width == 0 or height == 0 or stride < width) return fail(reply);
     const mapped = runtime.shm.map(cap) orelse return fail(reply);
@@ -240,9 +323,11 @@ fn attachScanout(stride: u32, width: u32, height: u32, format: u32, capability:
         .width = width,
         .height = height,
         .format = format,
+        .refresh_hz = refresh_hz,
         .scanout = scanout,
     } };
     background = protocol.pack(format, 0x20, 0x30, 0x48); // re-pack the wallpaper for the mode
+    updateFrameClock(); // the GOP floor's clock dies here — pace by the GPU's EDID now
     addDamage(screenRect()); // the whole new surface must be painted
     pending_native_verify = true;
     if (!reattach) pending_modeset_check = true; // the mode-set self-check runs once, on first upgrade
@@ -257,7 +342,8 @@ fn attachScanout(stride: u32, width: u32, height: u32, format: u32, capability:
 /// After the native upgrade is verified, prove the runtime-resolution-change and fenced-present
 /// paths: query the driver's modes, switch to one that differs from the current, re-composite
 /// the whole screen at the new size, and confirm the backend now reports that geometry. The
-/// present goes through the driver's fenced flush, so a clean present is a vsync present.
+/// present goes through the driver's fenced flush, so a clean present is a *fenced* present —
+/// completion-acknowledged and tear-free, not vblank-paced (docs/display-v2.md).
 fn modesetSelfCheck() void {
     if (!backend.canModeSet()) return;
     var mode_list: [4]backend_mod.Mode = undefined;
@@ -289,7 +375,7 @@ fn modesetSelfCheck() void {
     if (now.width == wanted.width and now.height == wanted.height) {
         var line: [80]u8 = undefined;
         _ = system.write(std.fmt.bufPrint(&line, "display: mode set to {d}x{d}, verified\n", .{ now.width, now.height }) catch "display: mode set, verified\n");
-        if (backend.hasVsync()) _ = system.write("display: vsync present ok\n");
+        if (backend.hasFencedPresent()) _ = system.write("display: fenced present ok\n");
     } else {
         _ = system.write("display: mode set FAILED (geometry unchanged)\n");
     }
@@ -443,15 +529,15 @@ fn mouseListener(width: u32, height: u32) void {
     }
 }
 
-/// Consume the latest cursor position from the channel and repaint the cursor layer at
-/// it. Runs on the main loop (the compositor owner) in response to a listener poke.
-/// `configureLayer` damages both the old and new footprints, so a plain `present`
-/// repaints exactly the two rectangles that changed.
+/// Consume the latest cursor position from the channel and move the cursor layer to it.
+/// Runs on the main loop (the compositor owner) in response to a listener poke.
+/// `configureLayer` damages both the old and new footprints; the frame clock presents
+/// them at the next tick, so a fast mouse coalesces to at most ~60 repaints a second.
 fn renderCursor() void {
     const snapshot = cursor_channel.take() orelse return;
     const id = cursor_layer orelse return;
     _ = configureLayer(id, snapshot.x, snapshot.y, cursor_z, true);
-    present();
+    schedulePresent();
     if (!cursor_tracking_reported and
         @abs(snapshot.x - cursor_origin_x) >= cursor_report_threshold and
         @abs(snapshot.y - cursor_origin_y) >= cursor_report_threshold)
@@ -500,6 +586,7 @@ fn initialise(endpoint: ipc.Handle) bool {
     _ = system.write(std.fmt.bufPrint(&line, "display: online {d}x{d} pitch {d} format {d}\n", .{
         mode.width, mode.height, mode.pitch, mode.format,
     }) catch "display: online\n");
+    updateFrameClock();
     _ = system.write("display: presented frame 0\n");
 
     selfCheck();
@@ -561,11 +648,13 @@ fn onMessage(message: []const u8, reply: []u8, sender: u32, capability: ?ipc.Han
             return ok(reply);
         },
         @intFromEnum(protocol.Operation.present) => {
-            present();
+            // Scheduled, not immediate: the frame clock composites the accumulated damage
+            // at the next tick, so back-to-back client presents coalesce into one frame.
+            schedulePresent();
             return ok(reply);
         },
         @intFromEnum(protocol.Operation.attach_scanout) => {
-            return attachScanout(request.x, request.width, request.height, request.colour, capability, reply);
+            return attachScanout(request.x, request.width, request.height, request.colour, request.y, capability, reply);
         },
         @intFromEnum(protocol.Operation.set_mode) => {
             if (!backend.setMode(request.width, request.height)) return fail(reply);
@@ -591,21 +680,15 @@ fn onMessage(message: []const u8, reply: []u8, sender: u32, capability: ?ipc.Han
     }
 }
 
-/// Two notification sources reach the compositor. A **message-notification** is a poke
-/// from the mouse-listener thread (a buffered self-`ipc.send`, `notify_message_bit`):
-/// repaint the cursor at its latest channel position. Anything else is the post-attach
-/// present **timer**: repaint into the freshly attached native surface, verify the frame
-/// landed, then run the one-shot mode-set self-check (V5).
+/// Two notification sources reach the compositor, and one coalesced badge can carry
+/// both, so each bit is handled independently. A **message-notification** is a poke from
+/// the mouse-listener thread (a buffered self-`ipc.send`, `notify_message_bit`): fold the
+/// newest cursor position into the scene. A **timer** (`notify_timer_bit`) is the frame
+/// clock — or the deferred first native present after `attach_scanout` — either way,
+/// present the accumulated damage.
 fn onNotification(badge: u64) void {
-    if (badge & ipc.notify_message_bit != 0) {
-        renderCursor();
-        return;
-    }
-    present(); // native present + verify (first timer fire after the upgrade)
-    if (pending_modeset_check) {
-        pending_modeset_check = false;
-        modesetSelfCheck();
-    }
+    if (badge & ipc.notify_message_bit != 0) renderCursor();
+    if (badge & ipc.notify_timer_bit != 0) frameTick();
 }
 
 pub fn main() void {

system/services/display/protocol.zig

@@ -24,11 +24,13 @@ pub const Operation = enum(u32) {
     damage = 6,
     /// present(): composite the dirty layers and flush to the screen.
     present = 7,
-    /// attach_scanout(x=stride, width, height, colour=format) + <surface capability>: a native
-    /// scanout driver announces itself, handing over the shared scanout surface as an `ipc_call`
-    /// send_cap. The compositor maps it, looks up the driver's `.scanout` present channel, and
-    /// upgrades off the GOP floor (docs/display-v2.md V4). `x` is the surface's row stride in
-    /// pixels, `colour` the DisplayFormat.
+    /// attach_scanout(x=stride, y=refresh_hz, width, height, colour=format) + <surface
+    /// capability>: a native scanout driver announces itself, handing over the shared scanout
+    /// surface as an `ipc_call` send_cap. The compositor maps it, looks up the driver's
+    /// `.scanout` present channel, and upgrades off the GOP floor (docs/display-v2.md V4).
+    /// `x` is the surface's row stride in pixels, `y` the panel refresh rate from the
+    /// driver's EDID read (0 = unknown; paces the compositor's frame clock), `colour` the
+    /// DisplayFormat.
     attach_scanout = 8,
     /// set_mode(width, height): change the display resolution — only a native backend that
     /// reports `canModeSet` honours it; on the GOP floor it fails (docs/display-v2.md V5).

test/qemu_test.py

@@ -221,15 +221,16 @@ CASES = [
      # require all three markers to appear somewhere rather than in a fixed order.
      "expect": r"(?s)(?=.*display: scanout upgraded to virtio-gpu)(?=.*display: native present verified)(?=.*display-demo: ok)",
      "fail":   r"display: native present FAILED|display: could not|display-demo: (no display|create failed)|CPU EXCEPTION|KERNEL PANIC"},
-    # Mode-set + EDID + vsync (v2 V5): same boot as display-native. After upgrading, the
-    # compositor queries the driver's modes, switches to a different resolution, and confirms the
-    # backend now reports it; the fenced present path makes it a vsync present. (The driver also
-    # logs the EDID preferred mode during bring-up.) Reuses the display-native kernel scenario.
+    # Mode-set + EDID + fenced presents (v2 V5): same boot as display-native. After upgrading,
+    # the compositor queries the driver's modes, switches to a different resolution, and confirms
+    # the backend now reports it; each present is fenced — completion-acknowledged and tear-free,
+    # not vblank-paced (docs/display-v2.md, "Fenced is not vsync"). (The driver also logs the
+    # EDID preferred mode during bring-up.) Reuses the display-native kernel scenario.
     {"name": "display-modeset",
      "build_case": "display-native",
      "qemu_extra": ["-device", "virtio-gpu-pci"],
      "mem": "512M",
-     "expect": r"(?s)(?=.*display: mode set to \d+x\d+, verified)(?=.*display: vsync present ok)",
+     "expect": r"(?s)(?=.*display: mode set to \d+x\d+, verified)(?=.*display: fenced present ok)",
      "fail":   r"display: mode set FAILED|display: mode-set self-check: |display: native present FAILED|CPU EXCEPTION|KERNEL PANIC"},
     # Resilience: driver restart + re-attach (v2 V6). device-manager (in test-scanout-restart
     # mode) kills the virtio-gpu driver once after it hellos; the restart policy respawns it, it