Judge loss by arrival order and a deadline, one tx ring per stream

system.go

@@ -722,6 +722,45 @@ func checkRings(fd int, ifname string, rxWant, txWant uint32) checkResult {
 	return res
 }
 
+// With PACKET_QDISC_BYPASS the hardware ring comes from XPS, which maps the
+// sending cpu to a ring, so an unpinned sender migrating across cpus spreads
+// one stream over rings that drain concurrently and reorder it on the wire.
+// Each stream's pinned feeder cpu maps to the stream's own ring instead. The
+// rings sit past the flow-rule queues because ixgbe pairs rx and tx of one
+// index on one interrupt vector, and tx cleanup does not belong on the
+// vectors delivering the streams.
+func checkTxQueues(ifname string, iface int) checkResult {
+	res := checkResult{item: ifname + " tx queues"}
+	queues, err := filepath.Glob("/sys/class/net/" + ifname + "/queues/tx-*/xps_cpus")
+	if err != nil || len(queues) == 0 {
+		res.err = fmt.Errorf("no tx queues found: %v", err)
+		return res
+	}
+	masks := make(map[string]string, numStreams)
+	for s := 0; s < numStreams; s++ {
+		masks[fmt.Sprintf("tx-%d", numStreams+s)] = fmt.Sprintf("%x", uint64(1)<<txCPU(iface, s))
+	}
+	pinned := 0
+	for _, q := range queues {
+		want := "0"
+		if m, ok := masks[filepath.Base(filepath.Dir(q))]; ok {
+			want = m
+			pinned++
+		}
+		if err := os.WriteFile(q, []byte(want), 0); err != nil {
+			res.err = fmt.Errorf("%s: %w", q, err)
+			return res
+		}
+	}
+	if pinned != numStreams {
+		res.err = fmt.Errorf("pinned %d of %d streams across %d tx queues", pinned, numStreams, len(queues))
+		return res
+	}
+	res.state = fmt.Sprintf("streams on tx-%d..%d, feeder cpus %d/%d",
+		numStreams, 2*numStreams-1, txCPU(iface, 0), txCPU(iface, numStreams-1))
+	return res
+}
+
 func withIoctlSocket(fn func(fd int) []checkResult) []checkResult {
 	fd, err := unix.Socket(unix.AF_INET, unix.SOCK_DGRAM, 0)
 	if err != nil {
@@ -737,16 +776,17 @@ func configureSystem(ifnames []string, ethertypes []uint16) []checkResult {
 	return withIoctlSocket(func(fd int) []checkResult {
 		out := []checkResult{checkGovernor(wantGovernor), checkMaxFreq(wantMaxFreqKHz),
 			checkIdleLatency(wantIdleLatency)}
-		for _, ifname := range ifnames {
+		for i, ifname := range ifnames {
 			out = append(out, checkLinkUp(fd, ifname))
 			out = append(out, checkMTU(fd, ifname, wantMTU))
 			out = append(out, clearFlowRules(fd, ifname))
 
-			// Ring changes reprogram the queues, so flow rules pointing at those
-			// queues have to be installed afterwards.
+			// Ring changes reprogram the queues, so flow rules and the xps pin,
+			// which point at those queues, have to be installed afterwards.
 			out = append(out, checkRings(fd, ifname, wantRxRing, wantTxRing))
 			out = append(out, checkCoalesce(fd, ifname, wantCoalesceUsecs))
 			out = append(out, checkFlowRules(fd, ifname, ethertypes))
+			out = append(out, checkTxQueues(ifname, i))
 		}
 		return out
 	})