Cycle every payload pattern per packet and always run both directions

This commit is contained in:
flamingcow
2026-07-25 23:03:31 -07:00
parent 0f16edfe61
commit e77574712d
8 changed files with 75 additions and 104 deletions
+3 -4
View File
@@ -6,10 +6,9 @@ import (
"strings"
)
// Everything the driver counts against a frame the link failed to carry, split
// by direction of travel so each counter belongs to exactly one direction: the
// transmitting interface owns the tx fields and the receiving interface the rx
// ones. Reading both sets off both interfaces would report every drop twice.
// Split by direction of travel so each counter belongs to exactly one
// direction: the transmitting interface owns the tx fields and the receiving one
// the rx fields. Reading both sets off both interfaces double counts every drop.
var (
nicTxFields = []string{
"tx_errors", "tx_dropped", "tx_fifo_errors",
+23 -24
View File
@@ -2,7 +2,6 @@ package main
import (
"encoding/binary"
"fmt"
"hash/crc32"
"math/rand/v2"
)
@@ -61,60 +60,60 @@ var patterns = []pattern{
}},
}
func patternIndex(name string) (int, error) {
for i, p := range patterns {
if p.name == name {
return i, nil
}
}
func patternNames() []string {
names := make([]string, len(patterns))
for i, p := range patterns {
names[i] = p.name
}
return 0, fmt.Errorf("unknown pattern %q (have %v)", name, names)
return names
}
// Every pattern is laid out at full length up front with its checksum at each
// size, so a sender picks a pattern by choosing a buffer, never by writing one.
type frameSpec struct {
patIdx int
ref []byte
crcFor map[int]uint32
refs [][]byte
crcFor []map[int]uint32
dstMAC [6]byte
srcMAC [6]byte
etherType uint16
sizes []int
maxSize int
maxPay int
}
func newFrameSpec(patIdx int, dst, src [6]byte, etherType uint16, sizes []int) *frameSpec {
func newFrameSpec(dst, src [6]byte, etherType uint16, sizes []int) *frameSpec {
maxSize := 0
for _, s := range sizes {
if s > maxSize {
maxSize = s
}
}
ref := make([]byte, maxSize-minFrame)
patterns[patIdx].fill(ref)
crcFor := make(map[int]uint32, len(sizes))
for _, s := range sizes {
crcFor[s] = crc32.Checksum(ref[:s-minFrame], crcTable)
}
return &frameSpec{
patIdx: patIdx,
ref: ref,
crcFor: crcFor,
f := &frameSpec{
dstMAC: dst,
srcMAC: src,
etherType: etherType,
sizes: sizes,
maxSize: maxSize,
maxPay: maxSize - minFrame,
}
for _, p := range patterns {
ref := make([]byte, f.maxPay)
p.fill(ref)
crcFor := make(map[int]uint32, len(sizes))
for _, s := range sizes {
crcFor[s] = crc32.Checksum(ref[:s-minFrame], crcTable)
}
f.refs = append(f.refs, ref)
f.crcFor = append(f.crcFor, crcFor)
}
return f
}
func (f *frameSpec) prefill(buf []byte) {
func (f *frameSpec) prefill(buf []byte, patIdx int) {
copy(buf[0:6], f.dstMAC[:])
copy(buf[6:12], f.srcMAC[:])
binary.BigEndian.PutUint16(buf[12:14], f.etherType)
copy(buf[minFrame:], f.ref)
copy(buf[minFrame:], f.refs[patIdx])
}
func putHeader(buf []byte, patIdx int, stream uint16, seq uint64, payLen int, crc uint32) {
+21 -37
View File
@@ -414,15 +414,15 @@ func (d *direction) row(elapsed time.Duration, v view, target float64, length st
}
}
// Read once a second rather than per frame, since these are sysfs files; the
// display uses whatever the last sample left behind.
// Sampled once a second, since these are sysfs reads; the display uses whatever
// the last sample left behind.
func (d *direction) sampleNIC() {
tx := readNIC(d.tx.name)
rx := readNIC(d.rx.name)
d.nicNow = tx.tx + rx.rx + tx.carrierDown
}
func buildDirection(label string, tx, rx endpoint, patIdx int, sizes []int, cfg config) (*direction, error) {
func buildDirection(label string, tx, rx endpoint, sizes []int, cfg config) (*direction, error) {
d := &direction{
label: label,
short: tx.tag + "→" + rx.tag,
@@ -434,7 +434,7 @@ func buildDirection(label string, tx, rx endpoint, patIdx int, sizes []int, cfg
for i := 0; i < cfg.streams; i++ {
et := uint16(etherBase + i)
d.specs = append(d.specs, newFrameSpec(patIdx, rx.mac, tx.mac, et, sizes))
d.specs = append(d.specs, newFrameSpec(rx.mac, tx.mac, et, sizes))
fd, err := openTxSocket(tx.idx)
if err != nil {
@@ -451,10 +451,9 @@ func buildDirection(label string, tx, rx endpoint, patIdx int, sizes []int, cfg
d.rxStats = append(d.rxStats, &rxStats{})
}
// The probe carries its own ethertype so it lands on a socket of its own, but
// it is left unsteered: it is a few frames a second and does not need a queue
// to itself, and the stamps are taken at the wire either way.
d.probeSpec = newFrameSpec(patIdx, rx.mac, tx.mac, cfg.probeEther, []int{probeSize})
// Deliberately given no flow rule: a few frames a second does not need a
// queue of its own, and the stamps are taken at the wire either way.
d.probeSpec = newFrameSpec(rx.mac, tx.mac, cfg.probeEther, []int{probeSize})
fd, err := openTxSocket(tx.idx)
if err != nil {
return nil, fmt.Errorf("%s probe tx socket: %w", label, err)
@@ -549,17 +548,15 @@ func main() {
aName = flag.String("a", "", "first interface")
bName = flag.String("b", "", "second interface")
sizesArg = flag.String("sizes", "64,128,256,512,1024,1280,1514", "frame sizes in bytes, excluding FCS, cycled per packet")
patArg = flag.String("pattern", "prbs", "payload pattern")
streams = flag.Int("streams", 7, "independent streams per direction, capped by rx rings; each gets its own ethertype, steered by a flow rule to its own rx queue")
batch = flag.Int("batch", 64, "frames per sendmmsg/recvmmsg call")
duplex = flag.Bool("duplex", true, "run both directions simultaneously")
zeroNS = flag.Float64("zero-ns", 4327.5, "both directions summed at zero cable length; belongs to the media adapters, recalibrate when they change")
nsPerM = flag.Float64("ns-per-m", 10.909, "both directions summed, per metre of cable")
zeroNS = flag.Float64("zero-ns", 2163.77, "mean of both directions at zero cable length; belongs to the media adapters, recalibrate when they change")
nsPerM = flag.Float64("ns-per-m", 5.4545, "mean of both directions, per metre of cable")
)
flag.Parse()
if err := run(*aName, *bName, *sizesArg, *patArg,
*streams, *batch, *duplex, *zeroNS, *nsPerM); err != nil {
if err := run(*aName, *bName, *sizesArg,
*streams, *batch, *zeroNS, *nsPerM); err != nil {
fmt.Fprintln(os.Stderr, "error:", err)
os.Exit(1)
}
@@ -576,8 +573,8 @@ const (
totalsHold = 50 * time.Millisecond
)
func run(aName, bName, sizesArg, patArg string,
nStreams, batch int, duplex bool, zeroNS, nsPerM float64) error {
func run(aName, bName, sizesArg string,
nStreams, batch int, zeroNS, nsPerM float64) error {
if aName == "" || bName == "" {
return fmt.Errorf("both -a and -b are required")
@@ -586,10 +583,6 @@ func run(aName, bName, sizesArg, patArg string,
if err != nil {
return err
}
patIdx, err := patternIndex(patArg)
if err != nil {
return err
}
a, err := lookupEndpoint(aName)
if err != nil {
return err
@@ -641,17 +634,12 @@ func run(aName, bName, sizesArg, patArg string,
}
var dirs []*direction
d0, err := buildDirection(a.name+"->"+b.name, a, b, patIdx, sizes, cfg)
if err != nil {
return err
}
dirs = append(dirs, d0)
if duplex {
d1, err := buildDirection(b.name+"->"+a.name, b, a, patIdx, sizes, cfg)
for _, p := range [][2]endpoint{{a, b}, {b, a}} {
d, err := buildDirection(p[0].name+"->"+p[1].name, p[0], p[1], sizes, cfg)
if err != nil {
return err
}
dirs = append(dirs, d1)
dirs = append(dirs, d)
}
defer func() {
for _, d := range dirs {
@@ -678,23 +666,19 @@ func run(aName, bName, sizesArg, patArg string,
for i, s := range sizes {
sizeStrs[i] = fmt.Sprintf("%d", s)
}
fmt.Println(renderBox("TEST",
fmt.Println(renderBox("CONFIG",
[]string{"SETTING", "VALUE"},
[]bool{false, false}, [][]string{
{"pattern", patterns[patIdx].name},
{"frame sizes", strings.Join(sizeStrs, " ")},
{"streams", fmt.Sprintf("%d per direction, ethertypes 0x%04x-0x%04x, one rx queue each",
{"streams", fmt.Sprintf("%d per direction, ethertypes 0x%04x-0x%04x",
nStreams, ethertypes[0], ethertypes[len(ethertypes)-1])},
{"probe", fmt.Sprintf("ethertype 0x%04x every %s", cfg.probeEther, probeInterval)},
{"batch", fmt.Sprintf("%d frames per syscall", batch)},
{"payload verify", "crc32c on every frame"},
{"cable probe", fmt.Sprintf("%d-byte frame on ethertype 0x%04x every %s, hardware stamped at both macs",
probeSize, cfg.probeEther, probeInterval)},
{"duplex", fmt.Sprintf("%v", duplex)},
{"socket buffers", fmt.Sprintf("sndbuf %s, rcvbuf %s (granted)",
{"calibration", fmt.Sprintf("%g ns at zero length, %g ns/m", zeroNS, nsPerM)},
{"buffers", fmt.Sprintf("sndbuf %s, rcvbuf %s",
humanBytes(uint64(sockBufSize(dirs[0].txFDs[0], unix.SO_SNDBUF))),
humanBytes(uint64(sockBufSize(dirs[0].rxFDs[0], unix.SO_RCVBUF))))},
}))
fmt.Println(paint("rates are per interval; error counts are cumulative, press space to reset them", cDim))
fmt.Println()
var doneTx, doneRx atomic.Bool
+16 -23
View File
@@ -10,10 +10,10 @@ import (
)
const (
// A stream number no data stream can take, so a probe is never mistaken for
// payload if one lands on the wrong socket.
probeStream = 0xffff
probeSize = 64
// Above any real stream number, so a probe is never taken for payload.
probeStream = 0xffff
probeSize = 64
probePattern = 0
// The mac has only a handful of transmit stamp slots. Asking faster than it
// can drain them gets slots recycled while a stamp is still outstanding, and
@@ -27,10 +27,9 @@ const (
)
// Both ports hang off one PTP clock, so a transmit stamp from one and a receive
// stamp from the other subtract directly. Both are taken at the mac, so the
// difference is the two phys plus the cable and nothing else: all host time and
// all queueing falls outside the stamped interval, which is why load does not
// move it.
// stamp from the other subtract directly. Both are taken at the mac, so all host
// time and all queueing falls outside the stamped interval, which is why load
// does not move it.
type cableStats struct {
mu sync.Mutex
min int64
@@ -52,10 +51,10 @@ func (v cableView) minText() string {
return commasInt(v.min)
}
// Summing both directions cancels the phy asymmetry between them, which is
// Averaging the two directions cancels the phy asymmetry between them, which is
// about 790ns and swamps any cable, so one direction alone cannot give a length.
func (c config) cableMetres(views []view) (float64, bool) {
if len(views) != 2 {
if len(views) == 0 {
return 0, false
}
var sum float64
@@ -65,7 +64,8 @@ func (c config) cableMetres(views []view) (float64, bool) {
}
sum += float64(v.cable.min)
}
return (sum - c.zeroNS) / c.nsPerM, true
mean := sum / float64(len(views))
return (mean - c.zeroNS) / c.nsPerM, true
}
func (c config) cableText(views []view) string {
@@ -83,8 +83,6 @@ func newCableStats() *cableStats {
}
}
// The two halves are produced by different goroutines in either order, so each
// deposits its stamp and whichever lands second completes the pair.
func (c *cableStats) put(seq uint64, ts int64, tx bool) {
c.mu.Lock()
defer c.mu.Unlock()
@@ -133,8 +131,6 @@ func (c *cableStats) reset() {
c.mu.Unlock()
}
// Sends one small frame at a time and collects its transmit stamp from the
// socket's error queue before sending the next.
type probeSender struct {
fd int
spec *frameSpec
@@ -143,7 +139,7 @@ type probeSender struct {
func (p *probeSender) run(done *atomic.Bool, startTx <-chan struct{}) {
buf := make([]byte, probeSize)
p.spec.prefill(buf)
p.spec.prefill(buf, probePattern)
oob := make([]byte, 512)
scratch := make([]byte, 1)
@@ -156,9 +152,8 @@ func (p *probeSender) run(done *atomic.Bool, startTx <-chan struct{}) {
for !done.Load() {
<-tick.C
// Stamps are matched to sends by position in the queue, so a stamp that
// arrived after its probe gave up would be handed to this one. Discard
// anything left over before sending.
// Stamps are matched to sends by position in the queue, so one that
// arrived after its probe gave up would be handed to this probe.
for {
if _, _, _, _, err := unix.Recvmsg(p.fd, scratch, oob,
unix.MSG_ERRQUEUE|unix.MSG_DONTWAIT); err != nil {
@@ -166,8 +161,8 @@ func (p *probeSender) run(done *atomic.Bool, startTx <-chan struct{}) {
}
}
putHeader(buf, p.spec.patIdx, probeStream, seq, probeSize-minFrame,
p.spec.crcFor[probeSize])
putHeader(buf, probePattern, probeStream, seq, probeSize-minFrame,
p.spec.crcFor[probePattern][probeSize])
err := unix.Send(p.fd, buf, 0)
// The sequence advances even when a probe fails, so a stale receive half
// can never be paired with a later probe that reused its number.
@@ -211,8 +206,6 @@ func (p *probeSender) awaitTx(scratch, oob []byte) (int64, bool) {
}
}
// Reads probes on the far interface, where every frame carries a receive stamp
// from the MAC.
type probeReceiver struct {
fd int
stats *cableStats
+1 -1
View File
@@ -61,7 +61,7 @@ func (w *rxWorker) run(done *atomic.Bool) {
w.streams[p.stream].observe(p.seq)
}
if p.payLen > len(w.spec.ref) {
if p.payLen > w.spec.maxPay {
w.stats.badLen.Add(1)
continue
}
+4 -11
View File
@@ -12,17 +12,14 @@ const (
hwtstampFilterAll = 1
)
// Read and written through the ifreq data pointer.
type hwtstampConfig struct {
flags int32
txType int32
rxFilter int32
}
// Hardware timestamping is a property of the interface, not the socket: the MAC
// has to be told to stamp on transmit and on receive before any socket can ask
// for the values. Receive stamping is filtered by protocol and ours is not PTP,
// so nothing narrower than "all" will see our frames.
// Receive stamping is filtered by protocol and ours is not PTP, so nothing
// narrower than "all" will see our frames.
func checkTimestamping(fd int, ifname string) checkResult {
res := checkResult{item: ifname + " hw timestamps"}
desc := func(c hwtstampConfig) string {
@@ -70,9 +67,6 @@ func checkTimestamping(fd int, ifname string) checkResult {
return res
}
// The socket asks for the values the MAC is now producing. Only the raw
// hardware clock is wanted; the software stamps would just be more control
// message to copy.
func enableTxTimestamps(fd int) error {
return unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_TIMESTAMPING,
unix.SOF_TIMESTAMPING_TX_HARDWARE|
@@ -85,9 +79,8 @@ func enableRxTimestamps(fd int) error {
unix.SOF_TIMESTAMPING_RX_HARDWARE|unix.SOF_TIMESTAMPING_RAW_HARDWARE)
}
// SCM_TIMESTAMPING carries three timespecs and the third is the raw hardware
// clock; the first two are software clocks we did not ask for and which arrive
// zeroed. A zero hardware stamp means the MAC did not produce one.
// Three timespecs, of which the third is the raw hardware clock. A zero there
// means the mac did not produce a stamp.
const scmTimestampingLen = 3 * int(unsafe.Sizeof(unix.Timespec{}))
func hwTimestamp(oob []byte) (int64, bool) {
+7 -2
View File
@@ -24,10 +24,14 @@ type txWorker struct {
}
func (w *txWorker) run(done *atomic.Bool) {
// The batch is not a multiple of the pattern count, so the pairing of pattern
// to size rotates every pass rather than settling into a fixed one.
bufs := make([][]byte, w.batch)
pats := make([]int, w.batch)
for i := range bufs {
bufs[i] = make([]byte, w.spec.maxSize)
w.spec.prefill(bufs[i])
pats[i] = i % len(patterns)
w.spec.prefill(bufs[i], pats[i])
}
hdrs, iovs := newMmsghdrs(bufs)
sizes := make([]int, w.batch)
@@ -44,7 +48,8 @@ func (w *txWorker) run(done *atomic.Bool) {
si = 0
}
sizes[i] = size
putHeader(bufs[i], w.spec.patIdx, w.stream, seq+uint64(i), size-minFrame, w.spec.crcFor[size])
putHeader(bufs[i], pats[i], w.stream, seq+uint64(i), size-minFrame,
w.spec.crcFor[pats[i]][size])
iovs[i].Len = uint64(size)
}
-2
View File
@@ -289,8 +289,6 @@ func (d *display) render(dirs []*direction, views []view, elapsed time.Duration,
fb.flush()
}
// The right-hand annotation sits above the table's last column so it lines up
// with the numbers under it rather than with the rule.
func (d *display) section(y int, title, right string) int {
d.small.draw(d.fb, uiMargin, y, title, uiFg)
if right != "" {