Capture the panel by reading the displayed drm buffer instead of fbdev

.gitignore

@@ -1 +1,2 @@
 /cabletest
+/shots/

fb.go

@@ -1,53 +1,149 @@
 package main
 
 import (
+	"encoding/binary"
 	"fmt"
-	"os"
+	"path/filepath"
 	"unsafe"
 
 	"golang.org/x/sys/unix"
 )
 
+// Drawing lands in a plain memory buffer and is blitted to a scanout buffer the
+// display is not reading, which is then swapped in whole at a vertical blank.
+// Writing into the live scanout buffer instead, as the fbdev interface invites,
+// races the beam: the blit takes a few hundred microseconds, and whatever the
+// display reads during it is part of one frame and part of the next.
 const (
-	fbioGetVScreenInfo = 0x4600
+	drmIoctlBase = 0x64
-	fbioBlank          = 0x4611
-	fbBlankUnblank     = 0
 
-	kdSetMode   = 0x4b3a
+	drmModeConnected     = 1
-	kdText      = 0x00
+	drmModeTypePreferred = 1 << 3
-	kdGraphics  = 0x01
+	drmModePageFlipEvent = 0x01
-	consolePath = "/dev/tty0"
+	drmEventFlipComplete = 0x02
-	fbPath      = "/dev/fb0"
+
+	// What ADDFB means by 32 bits per pixel and 24 bits of colour.
+	xrgbRedShift   = 16
+	xrgbGreenShift = 8
+	xrgbBlueShift  = 0
+
+	// Two would be enough to stop the panel tearing, since one buffer being
+	// displayed while the other is drawn into is all double buffering means.
+	// More than two is for anything reading a frame back out: they are cycled
+	// in order, so a buffer is left alone for the three frames between going on
+	// screen and coming round again, and reading one out of uncached scanout
+	// memory takes a good fraction of a frame.
+	scanoutBuffers = 4
 )
 
-// The variable screen info is a flat run of 40 u32s, so it is read as an array
+func drmIO(nr uintptr) uintptr         { return drmIoctlBase<<8 | nr }
-// rather than a struct to sidestep any question of padding.
+func drmIOWR(nr, size uintptr) uintptr { return 3<<30 | size<<16 | drmIoctlBase<<8 | nr }
-const (
+
-	viXres          = 0
+type drmModeInfo struct {
-	viYres          = 1
+	clock                                         uint32
-	viBitsPerPixel  = 6
+	hdisplay, hsyncStart, hsyncEnd, htotal, hskew uint16
-	viRedOffset     = 8
+	vdisplay, vsyncStart, vsyncEnd, vtotal, vscan uint16
-	viGreenOffset   = 11
+	vrefresh, flags, typ                          uint32
-	viBlueOffset    = 14
+	name                                          [32]byte
-	viScreenInfoLen = 40
+}
+
+type drmModeCardRes struct {
+	fbIDPtr, crtcIDPtr, connIDPtr, encIDPtr     uint64
+	countFBs, countCRTCs, countConns, countEncs uint32
+	minWidth, maxWidth, minHeight, maxHeight    uint32
+}
+
+type drmModeGetConnector struct {
+	encodersPtr, modesPtr, propsPtr, propValuesPtr uint64
+	countModes, countProps, countEncoders          uint32
+	encoderID, connectorID, connectorType          uint32
+	connectorTypeID, connection                    uint32
+	mmWidth, mmHeight, subpixel, pad               uint32
+}
+
+type drmModeGetEncoder struct {
+	encoderID, encoderType        uint32
+	crtcID                        uint32
+	possibleCRTCs, possibleClones uint32
+}
+
+type drmModeCrtc struct {
+	setConnectorsPtr uint64
+	countConnectors  uint32
+	crtcID           uint32
+	fbID             uint32
+	x, y             uint32
+	gammaSize        uint32
+	modeValid        uint32
+	mode             drmModeInfo
+}
+
+type drmModeFBCmd struct {
+	fbID, width, height, pitch, bpp, depth uint32
+	handle                                 uint32
+}
+
+type drmModeCreateDumb struct {
+	height, width, bpp, flags uint32
+	handle, pitch             uint32
+	size                      uint64
+}
+
+type drmModeMapDumb struct {
+	handle, pad uint32
+	offset      uint64
+}
+
+type drmModeCrtcPageFlip struct {
+	crtcID, fbID, flags, reserved uint32
+	userData                      uint64
+}
+
+var (
+	drmSetMaster    = drmIO(0x1e)
+	drmDropMaster   = drmIO(0x1f)
+	drmGetResources = drmIOWR(0xa0, unsafe.Sizeof(drmModeCardRes{}))
+	drmSetCrtc      = drmIOWR(0xa2, unsafe.Sizeof(drmModeCrtc{}))
+	drmGetEncoder   = drmIOWR(0xa6, unsafe.Sizeof(drmModeGetEncoder{}))
+	drmGetConnector = drmIOWR(0xa7, unsafe.Sizeof(drmModeGetConnector{}))
+	drmAddFB        = drmIOWR(0xae, unsafe.Sizeof(drmModeFBCmd{}))
+	drmPageFlip     = drmIOWR(0xb0, unsafe.Sizeof(drmModeCrtcPageFlip{}))
+	drmCreateDumb   = drmIOWR(0xb2, unsafe.Sizeof(drmModeCreateDumb{}))
+	drmMapDumb      = drmIOWR(0xb3, unsafe.Sizeof(drmModeMapDumb{}))
 )
 
+func drmIoctl(fd int, req uintptr, arg unsafe.Pointer) error {
+	if _, _, errno := unix.Syscall(unix.SYS_IOCTL, uintptr(fd), req, uintptr(arg)); errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+type scanout struct {
+	fbID uint32
+	mem  []byte
+}
+
 // w and h are the logical canvas, which is portrait; pw and ph are the panel,
 // which is landscape. Every draw is turned a quarter turn on its way to memory,
 // so logical top lands on the panel's right edge.
 type framebuffer struct {
-	file   *os.File
+	fd     int
-	tty    *os.File
-	mem    []byte
 	back   []byte
 	w      int
 	h      int
 	pw     int
 	ph     int
 	stride int
-	rShift uint
+
-	gShift uint
+	bufs   [scanoutBuffers]scanout
-	bShift uint
+	front  int
+	crtcID uint32
+	connID uint32
+
+	// One token per completed flip. The render loop waits on this rather than on
+	// a timer, so drawing is paced by the panel instead of by a guess at its rate.
+	flips chan struct{}
 }
 
 func (fb *framebuffer) offset(x, y int) int {
@@ -58,88 +154,261 @@ func (fb *framebuffer) fromPanel(x, y int) (int, int) {
 	return y, fb.pw - 1 - x
 }
 
-// The kernel console draws into the same framebuffer, including a blinking
+func cardResources(fd int) (crtcs, conns []uint32, err error) {
-// cursor and any keyboard echo, which fights every frame we write. Putting the
+	var res drmModeCardRes
-// active console into graphics mode stops it touching the framebuffer at all.
+	if err := drmIoctl(fd, drmGetResources, unsafe.Pointer(&res)); err != nil {
-func takeConsole() (*os.File, error) {
+		return nil, nil, fmt.Errorf("get resources: %w", err)
-	tty, err := os.OpenFile(consolePath, os.O_RDWR, 0)
-	if err != nil {
-		return nil, err
 	}
-	if _, _, errno := unix.Syscall(unix.SYS_IOCTL, tty.Fd(),
+	if res.countCRTCs == 0 || res.countConns == 0 {
-		kdSetMode, kdGraphics); errno != 0 {
+		return nil, nil, fmt.Errorf("card has no crtcs or connectors")
-		tty.Close()
-		return nil, fmt.Errorf("console to graphics mode: %w", errno)
 	}
-	return tty, nil
+	crtcs = make([]uint32, res.countCRTCs)
+	conns = make([]uint32, res.countConns)
+	res.countFBs, res.countEncs = 0, 0
+	res.fbIDPtr, res.encIDPtr = 0, 0
+	res.crtcIDPtr = uint64(uintptr(unsafe.Pointer(&crtcs[0])))
+	res.connIDPtr = uint64(uintptr(unsafe.Pointer(&conns[0])))
+	if err := drmIoctl(fd, drmGetResources, unsafe.Pointer(&res)); err != nil {
+		return nil, nil, fmt.Errorf("get resources: %w", err)
+	}
+	return crtcs, conns, nil
 }
 
-func openFramebuffer(path string) (*framebuffer, error) {
+// The preferred mode is the panel's native one; anything else would be the
-	f, err := os.OpenFile(path, os.O_RDWR, 0)
+// driver scaling a wrong-sized image onto it.
+func preferredMode(fd int, connID uint32) (drmModeInfo, error) {
+	c := drmModeGetConnector{connectorID: connID}
+	if err := drmIoctl(fd, drmGetConnector, unsafe.Pointer(&c)); err != nil {
+		return drmModeInfo{}, err
+	}
+	if c.countModes == 0 {
+		return drmModeInfo{}, fmt.Errorf("connector %d reported no modes", connID)
+	}
+	modes := make([]drmModeInfo, c.countModes)
+	q := drmModeGetConnector{
+		connectorID: connID,
+		countModes:  c.countModes,
+		modesPtr:    uint64(uintptr(unsafe.Pointer(&modes[0]))),
+	}
+	if err := drmIoctl(fd, drmGetConnector, unsafe.Pointer(&q)); err != nil {
+		return drmModeInfo{}, err
+	}
+	if q.countModes == 0 {
+		return drmModeInfo{}, fmt.Errorf("connector %d reported no modes", connID)
+	}
+	for _, m := range modes[:q.countModes] {
+		if m.typ&drmModeTypePreferred != 0 {
+			return m, nil
+		}
+	}
+	return modes[0], nil
+}
+
+func crtcFor(fd int, c drmModeGetConnector, crtcs []uint32) (uint32, error) {
+	encoders := []uint32{c.encoderID}
+	if c.countEncoders > 0 {
+		list := make([]uint32, c.countEncoders)
+		q := drmModeGetConnector{
+			connectorID:   c.connectorID,
+			countEncoders: c.countEncoders,
+			encodersPtr:   uint64(uintptr(unsafe.Pointer(&list[0]))),
+		}
+		if err := drmIoctl(fd, drmGetConnector, unsafe.Pointer(&q)); err == nil {
+			encoders = append(encoders, list[:q.countEncoders]...)
+		}
+	}
+	for _, id := range encoders {
+		if id == 0 {
+			continue
+		}
+		e := drmModeGetEncoder{encoderID: id}
+		if err := drmIoctl(fd, drmGetEncoder, unsafe.Pointer(&e)); err != nil {
+			continue
+		}
+		// Already driving this connector, otherwise anything it can be wired to.
+		if e.crtcID != 0 {
+			return e.crtcID, nil
+		}
+		for i, crtc := range crtcs {
+			if e.possibleCRTCs&(1<<uint(i)) != 0 {
+				return crtc, nil
+			}
+		}
+	}
+	return 0, fmt.Errorf("connector %d has no usable crtc", c.connectorID)
+}
+
+// The connector to drive and a crtc that can drive it.
+func findDisplay(fd int) (connID, crtcID uint32, err error) {
+	crtcs, conns, err := cardResources(fd)
+	if err != nil {
+		return 0, 0, err
+	}
+	for _, id := range conns {
+		c := drmModeGetConnector{connectorID: id}
+		if err := drmIoctl(fd, drmGetConnector, unsafe.Pointer(&c)); err != nil {
+			continue
+		}
+		if c.connection != drmModeConnected || c.countModes == 0 {
+			continue
+		}
+		crtc, err := crtcFor(fd, c, crtcs)
+		if err != nil {
+			continue
+		}
+		return id, crtc, nil
+	}
+	return 0, 0, fmt.Errorf("no connected connector with a mode")
+}
+
+// The card number is not stable across machines, so the card driving a
+// connected display is the one we want.
+func openCard() (int, error) {
+	paths, err := filepath.Glob("/dev/dri/card*")
+	if err != nil {
+		return -1, err
+	}
+	for _, p := range paths {
+		fd, err := unix.Open(p, unix.O_RDWR|unix.O_CLOEXEC, 0)
+		if err != nil {
+			continue
+		}
+		if _, _, err := findDisplay(fd); err == nil {
+			return fd, nil
+		}
+		unix.Close(fd)
+	}
+	return -1, fmt.Errorf("no drm device with a connected display")
+}
+
+func (fb *framebuffer) addScanout(i int) error {
+	create := drmModeCreateDumb{width: uint32(fb.pw), height: uint32(fb.ph), bpp: 32}
+	if err := drmIoctl(fb.fd, drmCreateDumb, unsafe.Pointer(&create)); err != nil {
+		return fmt.Errorf("create dumb buffer: %w", err)
+	}
+	fb.stride = int(create.pitch)
+
+	add := drmModeFBCmd{
+		width:  uint32(fb.pw),
+		height: uint32(fb.ph),
+		pitch:  create.pitch,
+		bpp:    32,
+		depth:  24,
+		handle: create.handle,
+	}
+	if err := drmIoctl(fb.fd, drmAddFB, unsafe.Pointer(&add)); err != nil {
+		return fmt.Errorf("add fb: %w", err)
+	}
+	fb.bufs[i].fbID = add.fbID
+
+	m := drmModeMapDumb{handle: create.handle}
+	if err := drmIoctl(fb.fd, drmMapDumb, unsafe.Pointer(&m)); err != nil {
+		return fmt.Errorf("map dumb buffer: %w", err)
+	}
+	mem, err := unix.Mmap(fb.fd, int64(m.offset), int(create.size),
+		unix.PROT_READ|unix.PROT_WRITE, unix.MAP_SHARED)
+	if err != nil {
+		return fmt.Errorf("mmap scanout: %w", err)
+	}
+	fb.bufs[i].mem = mem
+	return nil
+}
+
+func openFramebuffer() (*framebuffer, error) {
+	fd, err := openCard()
 	if err != nil {
 		return nil, err
 	}
+	fb := &framebuffer{fd: fd, flips: make(chan struct{}, 1)}
 
-	var vi [viScreenInfoLen]uint32
+	// Without master the modeset below is refused, and taking it is also what
-	if _, _, errno := unix.Syscall(unix.SYS_IOCTL, f.Fd(),
+	// stops the kernel console drawing into the display behind us.
-		fbioGetVScreenInfo, uintptr(unsafe.Pointer(&vi[0]))); errno != 0 {
+	if err := drmIoctl(fd, drmSetMaster, nil); err != nil {
-		f.Close()
+		unix.Close(fd)
-		return nil, fmt.Errorf("get screen info: %w", errno)
+		return nil, fmt.Errorf("take drm master: %w", err)
-	}
-	if vi[viBitsPerPixel] != 32 {
-		f.Close()
-		return nil, fmt.Errorf("only 32bpp supported, got %d", vi[viBitsPerPixel])
 	}
 
-	// Wake the panel; the console blanks it after a timeout.
+	connID, crtcID, err := findDisplay(fd)
-	unix.Syscall(unix.SYS_IOCTL, f.Fd(), fbioBlank, fbBlankUnblank)
-
-	stride, ok := readUint("/sys/class/graphics/fb0/stride")
-	if !ok {
-		f.Close()
-		return nil, fmt.Errorf("cannot read framebuffer stride")
-	}
-
-	fb := &framebuffer{
-		file:   f,
-		pw:     int(vi[viXres]),
-		ph:     int(vi[viYres]),
-		w:      int(vi[viYres]),
-		h:      int(vi[viXres]),
-		stride: int(stride),
-		rShift: uint(vi[viRedOffset]),
-		gShift: uint(vi[viGreenOffset]),
-		bShift: uint(vi[viBlueOffset]),
-	}
-
-	fb.mem, err = unix.Mmap(int(f.Fd()), 0, fb.stride*fb.ph,
-		unix.PROT_READ|unix.PROT_WRITE, unix.MAP_SHARED)
 	if err != nil {
-		f.Close()
+		fb.close()
-		return nil, fmt.Errorf("mmap: %w", err)
+		return nil, err
+	}
+	mode, err := preferredMode(fd, connID)
+	if err != nil {
+		fb.close()
+		return nil, err
+	}
+	fb.connID, fb.crtcID = connID, crtcID
+	fb.pw, fb.ph = int(mode.hdisplay), int(mode.vdisplay)
+	fb.w, fb.h = fb.ph, fb.pw
+
+	for i := range fb.bufs {
+		if err := fb.addScanout(i); err != nil {
+			fb.close()
+			return nil, err
+		}
 	}
 	fb.back = make([]byte, fb.stride*fb.ph)
 
-	fb.tty, err = takeConsole()
+	set := drmModeCrtc{
-	if err != nil {
+		setConnectorsPtr: uint64(uintptr(unsafe.Pointer(&fb.connID))),
-		unix.Munmap(fb.mem)
+		countConnectors:  1,
-		f.Close()
+		crtcID:           crtcID,
-		return nil, err
+		fbID:             fb.bufs[0].fbID,
+		modeValid:        1,
+		mode:             mode,
 	}
+	if err := drmIoctl(fd, drmSetCrtc, unsafe.Pointer(&set)); err != nil {
+		fb.close()
+		return nil, fmt.Errorf("set crtc: %w", err)
+	}
+
+	go fb.readEvents()
+	// Nothing has been flipped yet, so the first frame is owed its turn.
+	fb.flips <- struct{}{}
 	return fb, nil
 }
 
+// Flip completions arrive on the drm fd as a stream of length-prefixed events.
+func (fb *framebuffer) readEvents() {
+	buf := make([]byte, 4096)
+	for {
+		n, err := unix.Read(fb.fd, buf)
+		if n <= 0 || err != nil {
+			return
+		}
+		for off := 0; off+8 <= n; {
+			typ := binary.LittleEndian.Uint32(buf[off:])
+			length := int(binary.LittleEndian.Uint32(buf[off+4:]))
+			if length < 8 || off+length > n {
+				return
+			}
+			if typ == drmEventFlipComplete {
+				select {
+				case fb.flips <- struct{}{}:
+				default:
+				}
+			}
+			off += length
+		}
+	}
+}
+
 func (fb *framebuffer) close() {
-	fb.fill(rgb{})
+	for i := range fb.bufs {
-	fb.flush()
+		if fb.bufs[i].mem != nil {
-	unix.Syscall(unix.SYS_IOCTL, fb.tty.Fd(), kdSetMode, kdText)
+			unix.Munmap(fb.bufs[i].mem)
-	fb.tty.Close()
+		}
-	unix.Munmap(fb.mem)
+	}
-	fb.file.Close()
+	// Dropping master hands the display back to the kernel console, which
+	// restores its own mode. The framebuffers and dumb buffers are reclaimed
+	// when the last reference to the fd goes.
+	drmIoctl(fb.fd, drmDropMaster, nil)
+	unix.Close(fb.fd)
 }
 
 func (fb *framebuffer) pixel(c rgb) uint32 {
-	return uint32(c.r)<<fb.rShift | uint32(c.g)<<fb.gShift | uint32(c.b)<<fb.bShift
+	return uint32(c.r)<<xrgbRedShift | uint32(c.g)<<xrgbGreenShift | uint32(c.b)<<xrgbBlueShift
 }
 
 type rgb struct {
@@ -197,9 +466,9 @@ func (fb *framebuffer) blend(x, y int, c rgb, cov uint8) {
 	a := uint32(cov)
 	old := uint32(fb.back[o+0]) | uint32(fb.back[o+1])<<8 |
 		uint32(fb.back[o+2])<<16 | uint32(fb.back[o+3])<<24
-	orr := uint8(old >> fb.rShift)
+	orr := uint8(old >> xrgbRedShift)
-	og := uint8(old >> fb.gShift)
+	og := uint8(old >> xrgbGreenShift)
-	ob := uint8(old >> fb.bShift)
+	ob := uint8(old >> xrgbBlueShift)
 	mix := rgb{
 		r: uint8((uint32(c.r)*a + uint32(orr)*(255-a)) / 255),
 		g: uint8((uint32(c.g)*a + uint32(og)*(255-a)) / 255),
@@ -212,6 +481,21 @@ func (fb *framebuffer) blend(x, y int, c rgb, cov uint8) {
 	fb.back[o+3] = byte(v >> 24)
 }
 
-func (fb *framebuffer) flush() {
+// Copies into the scanout buffer furthest from being displayed and asks for it
-	copy(fb.mem, fb.back)
+// at the next blank. The copy cannot tear because nothing is displaying that
+// buffer, and the swap cannot tear because the hardware does it between frames.
+func (fb *framebuffer) flush() error {
+	next := (fb.front + 1) % scanoutBuffers
+	copy(fb.bufs[next].mem, fb.back)
+
+	flip := drmModeCrtcPageFlip{
+		crtcID: fb.crtcID,
+		fbID:   fb.bufs[next].fbID,
+		flags:  drmModePageFlipEvent,
+	}
+	if err := drmIoctl(fb.fd, drmPageFlip, unsafe.Pointer(&flip)); err != nil {
+		return fmt.Errorf("page flip: %w", err)
+	}
+	fb.front = next
+	return nil
 }

harness/grab.go

@@ -0,0 +1,250 @@
+package main
+
+import (
+	"bytes"
+	"fmt"
+	"image"
+	"image/color"
+	"path/filepath"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// cabletest drives the panel through drm and page flips between two scanout
+// buffers of its own, so /dev/fb0 holds the kernel console and there is nothing
+// in it worth reading. The buffer being displayed is reachable from out here
+// though: GETCRTC names it, and root is allowed a handle to it without being
+// drm master, so it can be mapped and read without the program that owns it
+// having to cooperate or even notice.
+const (
+	drmCardGlob = "/dev/dri/card*"
+
+	drmVblankRelative     = 0x1
+	drmVblankHighCrtcMask = 0x3e
+	drmVblankHighCrtcShft = 1
+
+	xrgbRedShift   = 16
+	xrgbGreenShift = 8
+	xrgbBlueShift  = 0
+)
+
+func drmIOWR(nr, size uintptr) uintptr { return 3<<30 | size<<16 | 0x64<<8 | nr }
+
+type drmModeInfo struct {
+	clock                                         uint32
+	hdisplay, hsyncStart, hsyncEnd, htotal, hskew uint16
+	vdisplay, vsyncStart, vsyncEnd, vtotal, vscan uint16
+	vrefresh, flags, typ                          uint32
+	name                                          [32]byte
+}
+
+type drmModeCardRes struct {
+	fbIDPtr, crtcIDPtr, connIDPtr, encIDPtr     uint64
+	countFBs, countCRTCs, countConns, countEncs uint32
+	minWidth, maxWidth, minHeight, maxHeight    uint32
+}
+
+type drmModeCrtc struct {
+	setConnectorsPtr uint64
+	countConnectors  uint32
+	crtcID           uint32
+	fbID             uint32
+	x, y             uint32
+	gammaSize        uint32
+	modeValid        uint32
+	mode             drmModeInfo
+}
+
+type drmModeFBCmd struct {
+	fbID, width, height, pitch, bpp, depth uint32
+	handle                                 uint32
+}
+
+type drmModeMapDumb struct {
+	handle, pad uint32
+	offset      uint64
+}
+
+type drmWaitVblank struct {
+	typ      uint32
+	sequence uint32
+	signal   uint64
+	tvSec    int64
+	tvUsec   int64
+}
+
+var (
+	drmGetResources = drmIOWR(0xa0, unsafe.Sizeof(drmModeCardRes{}))
+	drmGetCrtc      = drmIOWR(0xa1, unsafe.Sizeof(drmModeCrtc{}))
+	drmGetFB        = drmIOWR(0xad, unsafe.Sizeof(drmModeFBCmd{}))
+	drmMapDumb      = drmIOWR(0xb3, unsafe.Sizeof(drmModeMapDumb{}))
+	drmWaitVblankIO = drmIOWR(0x3a, 24)
+)
+
+func drmIoctl(fd int, req uintptr, arg unsafe.Pointer) error {
+	if _, _, errno := unix.Syscall(unix.SYS_IOCTL, uintptr(fd), req, uintptr(arg)); errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+type grabber struct {
+	fd    int
+	crtc  uint32
+	index int
+}
+
+// The card and crtc actually putting something on a display, which is whichever
+// one cabletest chose when it set its mode.
+func openGrabber() (*grabber, error) {
+	paths, err := filepath.Glob(drmCardGlob)
+	if err != nil {
+		return nil, err
+	}
+	for _, p := range paths {
+		fd, err := unix.Open(p, unix.O_RDWR|unix.O_CLOEXEC, 0)
+		if err != nil {
+			continue
+		}
+		crtc, index, err := activeCrtc(fd)
+		if err == nil {
+			return &grabber{fd: fd, crtc: crtc, index: index}, nil
+		}
+		unix.Close(fd)
+	}
+	return nil, fmt.Errorf("no drm crtc is scanning out a framebuffer")
+}
+
+func activeCrtc(fd int) (uint32, int, error) {
+	var res drmModeCardRes
+	if err := drmIoctl(fd, drmGetResources, unsafe.Pointer(&res)); err != nil {
+		return 0, 0, err
+	}
+	if res.countCRTCs == 0 {
+		return 0, 0, fmt.Errorf("card has no crtcs")
+	}
+	crtcs := make([]uint32, res.countCRTCs)
+	res.countFBs, res.countConns, res.countEncs = 0, 0, 0
+	res.fbIDPtr, res.connIDPtr, res.encIDPtr = 0, 0, 0
+	res.crtcIDPtr = uint64(uintptr(unsafe.Pointer(&crtcs[0])))
+	if err := drmIoctl(fd, drmGetResources, unsafe.Pointer(&res)); err != nil {
+		return 0, 0, err
+	}
+	for i, id := range crtcs {
+		c := drmModeCrtc{crtcID: id}
+		if err := drmIoctl(fd, drmGetCrtc, unsafe.Pointer(&c)); err != nil {
+			continue
+		}
+		if c.fbID != 0 && c.modeValid != 0 {
+			return id, i, nil
+		}
+	}
+	return 0, 0, fmt.Errorf("no crtc is scanning out a framebuffer")
+}
+
+func (g *grabber) close() { unix.Close(g.fd) }
+
+// Reading the buffer takes about as long as a frame, so where the read starts
+// in the flip cycle is what decides whether it stays ahead of the writer. Woken
+// at a blank, the buffer named next has just gone on screen, which leaves a
+// whole frame plus however long cabletest spends drawing before anything
+// touches it again.
+func (g *grabber) waitVblank() error {
+	v := drmWaitVblank{
+		typ:      drmVblankRelative | uint32(g.index<<drmVblankHighCrtcShft)&drmVblankHighCrtcMask,
+		sequence: 1,
+	}
+	for {
+		err := drmIoctl(g.fd, drmWaitVblankIO, unsafe.Pointer(&v))
+		if err == unix.EINTR {
+			continue
+		}
+		return err
+	}
+}
+
+// Maps whichever buffer is on screen right now. The caller is expected to have
+// just woken at a blank, so this is the one with the longest life ahead of it.
+func (g *grabber) mapFront() (mem []byte, pitch, pw, ph int, err error) {
+	c := drmModeCrtc{crtcID: g.crtc}
+	if err := drmIoctl(g.fd, drmGetCrtc, unsafe.Pointer(&c)); err != nil {
+		return nil, 0, 0, 0, fmt.Errorf("get crtc: %w", err)
+	}
+	if c.fbID == 0 {
+		return nil, 0, 0, 0, fmt.Errorf("crtc %d is not scanning out anything", g.crtc)
+	}
+
+	// Only a drm master or root is given a handle to someone else's framebuffer.
+	fb := drmModeFBCmd{fbID: c.fbID}
+	if err := drmIoctl(g.fd, drmGetFB, unsafe.Pointer(&fb)); err != nil {
+		return nil, 0, 0, 0, fmt.Errorf("get fb %d: %w", c.fbID, err)
+	}
+	if fb.handle == 0 {
+		return nil, 0, 0, 0, fmt.Errorf("kernel gave no handle for fb %d", c.fbID)
+	}
+	if fb.bpp != 32 {
+		return nil, 0, 0, 0, fmt.Errorf("expected 32bpp, got %d", fb.bpp)
+	}
+
+	m := drmModeMapDumb{handle: fb.handle}
+	if err := drmIoctl(g.fd, drmMapDumb, unsafe.Pointer(&m)); err != nil {
+		return nil, 0, 0, 0, fmt.Errorf("map fb %d: %w", c.fbID, err)
+	}
+	pitch, pw, ph = int(fb.pitch), int(fb.width), int(fb.height)
+	mem, err = unix.Mmap(g.fd, int64(m.offset), pitch*ph,
+		unix.PROT_READ, unix.MAP_SHARED)
+	if err != nil {
+		return nil, 0, 0, 0, fmt.Errorf("mmap fb %d: %w", c.fbID, err)
+	}
+	return mem, pitch, pw, ph, nil
+}
+
+// Scanout memory is uncached, so reading a frame out of it costs a good
+// fraction of a frame's time and could in principle be overtaken by the next
+// redraw. Rather than trust that it was not, the buffer is read twice and the
+// pair only accepted if they agree: the writer cycles through its buffers in
+// order and leaves this one alone for several frames after putting it on
+// screen, which is comfortably long enough for both reads. Disagreement means
+// that reasoning is wrong somewhere, which is worth hearing about rather than
+// papering over with another go.
+func (g *grabber) frame() (*image.NRGBA, error) {
+	if err := g.waitVblank(); err != nil {
+		return nil, fmt.Errorf("wait for vblank: %w", err)
+	}
+	mem, pitch, pw, ph, err := g.mapFront()
+	if err != nil {
+		return nil, err
+	}
+	a := make([]byte, len(mem))
+	copy(a, mem)
+	b := make([]byte, len(mem))
+	copy(b, mem)
+	unix.Munmap(mem)
+
+	if !bytes.Equal(a, b) {
+		return nil, fmt.Errorf("scanout buffer was redrawn while being read")
+	}
+	return decode(a, pitch, pw, ph), nil
+}
+
+// The panel is landscape and every draw is turned a quarter turn on its way
+// into it, so the turn is undone here to get back what a person standing in
+// front of it sees.
+func decode(buf []byte, pitch, pw, ph int) *image.NRGBA {
+	img := image.NewNRGBA(image.Rect(0, 0, ph, pw))
+	for x := 0; x < ph; x++ {
+		for y := 0; y < pw; y++ {
+			o := x*pitch + (pw-1-y)*4
+			v := uint32(buf[o]) | uint32(buf[o+1])<<8 |
+				uint32(buf[o+2])<<16 | uint32(buf[o+3])<<24
+			img.SetNRGBA(x, y, color.NRGBA{
+				R: uint8(v >> xrgbRedShift),
+				G: uint8(v >> xrgbGreenShift),
+				B: uint8(v >> xrgbBlueShift),
+				A: 255,
+			})
+		}
+	}
+	return img
+}

harness/main.go

@@ -0,0 +1,168 @@
+// Runs cabletest for a fixed time and dumps the panel to png, since the machine
+// with the display on it is not the machine looking at it.
+//
+//	sudo go run ./harness -for 20s -at 3s,15s -- -a enp1s0f0np0 -b enp1s0f1np1
+package main
+
+import (
+	"flag"
+	"fmt"
+	"image/png"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"slices"
+	"strings"
+	"syscall"
+	"time"
+)
+
+const (
+	shotsDir = "shots"
+	grace    = 3 * time.Second
+)
+
+func main() {
+	runFor := flag.Duration("for", 15*time.Second, "how long to let cabletest run")
+	at := flag.String("at", "", "offsets to capture the panel at, comma separated, e.g. 3s,10s")
+	flag.Parse()
+
+	if err := run(*runFor, *at, flag.Args()); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+}
+
+func parseShots(s string, runFor time.Duration) ([]time.Duration, error) {
+	if s == "" {
+		return nil, nil
+	}
+	var out []time.Duration
+	for _, f := range strings.Split(s, ",") {
+		d, err := time.ParseDuration(strings.TrimSpace(f))
+		if err != nil {
+			return nil, fmt.Errorf("bad offset %q: %w", f, err)
+		}
+		if d < 0 || d >= runFor {
+			return nil, fmt.Errorf("offset %s falls outside the %s run", d, runFor)
+		}
+		out = append(out, d)
+	}
+	slices.Sort(out)
+	return out, nil
+}
+
+func run(runFor time.Duration, at string, args []string) error {
+	if os.Geteuid() != 0 {
+		return fmt.Errorf("needs root for the raw sockets and the framebuffer: sudo go run ./harness")
+	}
+	shots, err := parseShots(at, runFor)
+	if err != nil {
+		return err
+	}
+	if len(shots) > 0 {
+		if err := os.MkdirAll(shotsDir, 0o755); err != nil {
+			return err
+		}
+	}
+
+	cmd := exec.Command("go", append([]string{"run", "."}, args...)...)
+	cmd.Stdout, cmd.Stderr = os.Stdout, os.Stderr
+	// Its own process group: go run's compiled child is reparented rather than
+	// killed when go run dies, and an orphan holding the wire poisons every
+	// measurement after it.
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	pgid := cmd.Process.Pid
+
+	done := make(chan error, 1)
+	go func() { done <- cmd.Wait() }()
+
+	start := time.Now()
+	for _, d := range shots {
+		select {
+		case err := <-done:
+			return exited(start, err)
+		case <-time.After(time.Until(start.Add(d))):
+		}
+		name := filepath.Join(shotsDir, d.String()+".png")
+		if err := shoot(name); err != nil {
+			stop(pgid, done)
+			return err
+		}
+		fmt.Printf("captured %s\n", name)
+	}
+
+	select {
+	case err := <-done:
+		return exited(start, err)
+	case <-time.After(time.Until(start.Add(runFor))):
+	}
+	return stop(pgid, done)
+}
+
+func exited(start time.Time, err error) error {
+	ran := time.Since(start).Round(time.Millisecond)
+	if err != nil {
+		return fmt.Errorf("cabletest exited after %s: %w", ran, err)
+	}
+	return fmt.Errorf("cabletest exited after %s", ran)
+}
+
+func stop(pgid int, done chan error) error {
+	syscall.Kill(-pgid, syscall.SIGINT)
+	select {
+	case <-done:
+	case <-time.After(grace):
+		syscall.Kill(-pgid, syscall.SIGKILL)
+		<-done
+	}
+	// Reaping go run says nothing about what it left behind, so the group itself
+	// is what gets checked.
+	if groupGone(pgid) {
+		return nil
+	}
+	syscall.Kill(-pgid, syscall.SIGKILL)
+	if groupGone(pgid) {
+		return nil
+	}
+	return fmt.Errorf("process group %d survived SIGKILL; check with pgrep -a cabletest", pgid)
+}
+
+func groupGone(pgid int) bool {
+	deadline := time.Now().Add(grace)
+	for {
+		if syscall.Kill(-pgid, 0) == syscall.ESRCH {
+			return true
+		}
+		if time.Now().After(deadline) {
+			return false
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+}
+
+func shoot(name string) error {
+	g, err := openGrabber()
+	if err != nil {
+		return err
+	}
+	defer g.close()
+
+	img, err := g.frame()
+	if err != nil {
+		return err
+	}
+
+	out, err := os.Create(name)
+	if err != nil {
+		return err
+	}
+	if err := png.Encode(out, img); err != nil {
+		out.Close()
+		return err
+	}
+	return out.Close()
+}

main.go

@@ -50,12 +50,17 @@ type direction struct {
 	probeRxFD int
 	cable     *cableStats
 
+	// Guards everything the sampler touches. The counters are read on their own
+	// clock and drawn on another, and the two must not read them at once:
+	// sampleDrops consumes what it reads, so a second caller would see a gap.
+	mu          sync.Mutex
 	prevConsole counterSet
 	win         *rateWindow
-	heldFrames  heldValue
-	heldSent    heldValue
 	drops       uint64
 	base        counterSet
+
+	heldFrames heldValue
+	heldSent   heldValue
 	nic        atomic.Uint64
 	poller     *nicPoller
 }
@@ -266,7 +271,10 @@ func (d *direction) snapshot() sample {
 // everything is measured from. Rates are deliberately left running, since they
 // are instantaneous and would only blink to zero and back.
 func (d *direction) reset() {
+	d.mu.Lock()
 	d.base = d.capture(time.Now())
+	d.mu.Unlock()
+
 	d.heldFrames = heldValue{}
 	d.heldSent = heldValue{}
 	d.cable.reset()
@@ -364,6 +372,9 @@ func totalView(views []view) view {
 }
 
 func (d *direction) view(t time.Time) view {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
 	now := d.capture(t)
 	p := d.prevConsole
 	d.prevConsole = now
@@ -380,21 +391,36 @@ func (d *direction) view(t time.Time) view {
 	return v
 }
 
-func (d *direction) displayView(t time.Time) view {
+// The one place the counters are read for the ring. Runs on its own ticker, so
-	now := d.capture(t)
+// what the buckets measure does not move when the drawing does.
-	d.win.push(now)
+func (d *direction) sample(t time.Time) {
+	d.mu.Lock()
+	d.win.push(d.capture(t))
+	d.mu.Unlock()
+}
 
-	v := d.counters(now)
+// Draws whatever the sampler last put in the ring rather than reading the
-	v.rxFrames = d.heldFrames.get(t, v.rxFrames)
+// counters again, so the display is a consumer of the measurement and never a
-	v.rxGot = d.heldSent.get(t, v.rxGot)
+// participant in it.
-	if n := d.win.count(); n >= 2 {
+func (d *direction) displayView(t time.Time) view {
+	d.mu.Lock()
+	n := d.win.count()
+	if n == 0 {
+		d.mu.Unlock()
+		return view{cable: d.cable.view()}
+	}
+	v := d.counters(d.win.at(n - 1))
+	if n >= 2 {
 		v.window = errsBetween(d.win.at(0), d.win.at(n-1))
 	}
-
 	v.txPPS = d.win.median(txRatePPS)
 	v.rxPPS = d.win.median(rxRatePPS)
 	v.txGbps = d.win.median(txRateGbps)
 	v.rxGbps = d.win.median(rxRateGbps)
+	d.mu.Unlock()
+
+	v.rxFrames = d.heldFrames.get(t, v.rxFrames)
+	v.rxGot = d.heldSent.get(t, v.rxGot)
 	return v
 }
 
@@ -437,6 +463,7 @@ func buildDirection(label string, tx, rx endpoint, sizes []int, cfg config) (*di
 		cable:   newCableStats(),
 	}
 	d.poller = &nicPoller{txName: tx.name, rxName: rx.name, total: &d.nic}
+	d.win = newRateWindow(int(rateWindowSpan/sampleInterval) + 1)
 
 	for i := 0; i < cfg.streams; i++ {
 		et := uint16(etherBase + i)
@@ -575,15 +602,39 @@ func main() {
 
 const (
 	reportInterval = time.Second
-	// Redraw fast so the panel feels live, but measure rates over a much longer
+	// What a bucket covers. Nothing to do with how often the panel is redrawn:
-	// window than a frame, since a frame's worth of a bursty sender is noise.
+	// how often the counters are read is a property of the measurement, and
-	displayInterval = 16 * time.Millisecond
+	// letting the refresh set it would let a slow or blocked draw stretch the
+	// window it reports. Short enough that a step lands promptly, long enough
+	// that a bucket holds tens of thousands of frames at line rate and is not
+	// itself noise.
+	sampleInterval = 16 * time.Millisecond
 	// Both how far back the shown errors reach and how many buckets the median
 	// runs over, so a step in the rate lands half this late.
 	rateWindowSpan = time.Second
 	totalsHold     = 50 * time.Millisecond
 )
 
+// One sampler for both directions, so their buckets share an instant and the
+// cable length, which needs a figure from each, is never mixing two moments.
+type sampler struct {
+	dirs []*direction
+}
+
+func (s *sampler) run(done *atomic.Bool, startTx <-chan struct{}) {
+	<-startTx
+
+	tick := time.NewTicker(sampleInterval)
+	defer tick.Stop()
+
+	for !done.Load() {
+		now := <-tick.C
+		for _, d := range s.dirs {
+			d.sample(now)
+		}
+	}
+}
+
 func run(aName, bName, sizesArg string,
 	nStreams, batch int, nsPerM float64) error {
 
@@ -692,6 +743,12 @@ func run(aName, bName, sizesArg string,
 	for _, d := range dirs {
 		d.start(&wg, &doneTx, &doneRx, cfg, &rxReady, startTx)
 	}
+	samp := &sampler{dirs: dirs}
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
+		samp.run(&doneRx, startTx)
+	}()
 	rxReady.Wait()
 
 	sig := make(chan os.Signal, 1)
@@ -719,14 +776,9 @@ func run(aName, bName, sizesArg string,
 	close(startTx)
 	tick := time.NewTicker(reportInterval)
 	defer tick.Stop()
-	frame := time.NewTicker(displayInterval)
-	defer frame.Stop()
 
 	views := make([]view, len(dirs))
 	rows := make([]view, len(dirs))
-	for _, d := range dirs {
-		d.win = newRateWindow(int(rateWindowSpan/displayInterval) + 1)
-	}
 	stats := &streamTable{cols: intervalCols, headerEvery: 20}
 	for {
 		select {
@@ -737,7 +789,8 @@ func run(aName, bName, sizesArg string,
 			return nil
 		case <-space:
 			start = resetAll(dirs, stats)
-		case now := <-frame.C:
+		case <-disp.fb.flips:
+			now := time.Now()
 			px, py, down := touch.get()
 			x, y := disp.fb.fromPanel(px, py)
 			if disp.holdReset(x, y, down, now) {
@@ -750,8 +803,10 @@ func run(aName, bName, sizesArg string,
 			if m, ok := cfg.cableMetres(views); ok {
 				cable = fmt.Sprintf("%.1f m", m)
 			}
-			disp.render(totalView(views), now.Sub(start),
+			if err := disp.render(totalView(views), now.Sub(start),
-				target*float64(len(dirs)), cable)
+				target*float64(len(dirs)), cable); err != nil {
+				return err
+			}
 		case now := <-tick.C:
 			elapsed := now.Sub(start)
 			// Length needs both directions, so every row is sampled before any of

ui.go

@@ -58,7 +58,7 @@ type display struct {
 }
 
 func newDisplay() (*display, error) {
-	fb, err := openFramebuffer(fbPath)
+	fb, err := openFramebuffer()
 	if err != nil {
 		return nil, err
 	}
@@ -250,7 +250,7 @@ func rateColor(gb, target float64) rgb {
 	}
 }
 
-func (d *display) render(v view, elapsed time.Duration, target float64, cable string) {
+func (d *display) render(v view, elapsed time.Duration, target float64, cable string) error {
 	fb := d.fb
 	fb.fill(uiBg)
 
@@ -267,5 +267,5 @@ func (d *display) render(v view, elapsed time.Duration, target float64, cable st
 	d.errBlock(x, w, y+blockGap, v.since)
 
 	d.drawResetButton()
-	fb.flush()
+	return fb.flush()
 }