mirror of
https://github.com/slackhq/nebula.git
synced 2026-08-15 15:57:01 +02:00
more fixes!
This commit is contained in:
@@ -93,8 +93,11 @@ func parseIPPrologue(pkt []byte, wantProto byte) (parsedIP, bool) {
|
||||
|
||||
// ipHeadersMatch compares the IP portion of two packet header prefixes for
|
||||
// byte-for-byte equality on every field that must be identical across
|
||||
// coalesced segments. Size/IPID/IPCsum and the 2-bit IP-level ECN field are
|
||||
// masked out — the appendPayload step merges CE into the seed.
|
||||
// coalesced segments. Size/IPID/IPCsum are masked out. The full DSCP/ECN
|
||||
// byte (IPv4 ToS / IPv6 traffic class) is compared, matching Linux kernel
|
||||
// GRO: segments with differing ECN codepoints must not coalesce, otherwise
|
||||
// ORing e.g. ECT(0) with ECT(1) would fabricate a false CE (congestion)
|
||||
// mark or mark a Not-ECT flow as ECN-capable.
|
||||
//
|
||||
// The transport (L4) portion of the header is checked separately by the
|
||||
// per-protocol matcher.
|
||||
@@ -102,11 +105,11 @@ func ipHeadersMatch(a, b []byte, isV6 bool) bool {
|
||||
if isV6 {
|
||||
// IPv6: byte 0 = version/TC[7:4], byte 1 = TC[3:0]/flow[19:16],
|
||||
// bytes [2:4] = flow[15:0], [6:8] = next_hdr/hop, [8:40] = src+dst.
|
||||
// ECN lives in TC[1:0] = byte 1 mask 0x30. Skip [4:6] payload_len.
|
||||
// Compare byte 1 fully so ECN (TC[1:0]) must match. Skip [4:6] payload_len.
|
||||
if a[0] != b[0] {
|
||||
return false
|
||||
}
|
||||
if a[1]&^0x30 != b[1]&^0x30 {
|
||||
if a[1] != b[1] {
|
||||
return false
|
||||
}
|
||||
if !bytes.Equal(a[2:4], b[2:4]) {
|
||||
@@ -119,11 +122,12 @@ func ipHeadersMatch(a, b []byte, isV6 bool) bool {
|
||||
}
|
||||
// IPv4: byte 0 = version/IHL, byte 1 = DSCP(6)|ECN(2),
|
||||
// [6:10] flags/fragoff/TTL/proto, [12:20] src+dst.
|
||||
// Compare byte 1 fully so ECN must match.
|
||||
// Skip [2:4] total len, [4:6] id, [10:12] csum.
|
||||
if a[0] != b[0] {
|
||||
return false
|
||||
}
|
||||
if a[1]&^0x03 != b[1]&^0x03 {
|
||||
if a[1] != b[1] {
|
||||
return false
|
||||
}
|
||||
if !bytes.Equal(a[6:10], b[6:10]) {
|
||||
@@ -135,19 +139,6 @@ func ipHeadersMatch(a, b []byte, isV6 bool) bool {
|
||||
return true
|
||||
}
|
||||
|
||||
// mergeECNIntoSeed ORs the 2-bit IP-level ECN field of pkt's IP header
|
||||
// onto the seed's IP header, so a CE mark on any coalesced segment
|
||||
// propagates to the final superpacket. (CE is 0b11; ORing yields CE if
|
||||
// any segment carried it.) Used by both TCP and UDP coalescers, so the
|
||||
// invariant lives in one place.
|
||||
func mergeECNIntoSeed(seedHdr, pktHdr []byte, isV6 bool) {
|
||||
if isV6 {
|
||||
seedHdr[1] |= pktHdr[1] & 0x30
|
||||
} else {
|
||||
seedHdr[1] |= pktHdr[1] & 0x03
|
||||
}
|
||||
}
|
||||
|
||||
// Arena is an injectable byte-slab that hands out non-overlapping borrowed
|
||||
// slices via Reserve and releases them in bulk via Reset. Coalescers take
|
||||
// an *Arena at construction so the caller controls the slab lifetime and
|
||||
|
||||
@@ -365,9 +365,6 @@ func (c *TCPCoalescer) appendPayload(s *coalesceSlot, pkt []byte, info parsedTCP
|
||||
// last segment. Without this the sender's push signal is dropped.
|
||||
s.hdrBuf[s.ipHdrLen+13] |= tcpFlagPsh
|
||||
}
|
||||
// Merge IP-level CE marks into the seed: headersMatch ignores ECN, so
|
||||
// this is the one place the signal is preserved.
|
||||
mergeECNIntoSeed(s.hdrBuf[:s.ipHdrLen], pkt[:s.ipHdrLen], s.isV6)
|
||||
if info.payLen < s.gsoSize || info.flags&tcpFlagPsh != 0 {
|
||||
s.psh = true
|
||||
}
|
||||
@@ -424,8 +421,9 @@ func (c *TCPCoalescer) flushSlot(s *coalesceSlot) error {
|
||||
|
||||
// headersMatch compares two IP+TCP header prefixes for byte-for-byte
|
||||
// equality on every field that must be identical across coalesced
|
||||
// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out, as is the
|
||||
// 2-bit IP-level ECN field — appendPayload merges CE into the seed.
|
||||
// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out. The IP-level
|
||||
// ECN codepoint is compared (via ipHeadersMatch) so segments with differing
|
||||
// ECN don't coalesce, matching kernel GRO.
|
||||
func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
|
||||
if len(a) != len(b) {
|
||||
return false
|
||||
@@ -632,6 +630,9 @@ func flowKeyCompare(a, b flowKey) int {
|
||||
// ECE state must agree across both slots: PSH is a semantic delimiter
|
||||
// (preserving the sender's push boundary) and ECE state must be uniform
|
||||
// across a window (the same rule canAppend enforces for in-flow appends).
|
||||
// The IP-level ECN codepoint must also match: this check calls headersMatch
|
||||
// → ipHeadersMatch, which compares the full DSCP/ECN byte, so two slots with
|
||||
// differing ECN marks stay separate superpackets, each keeping its own mark.
|
||||
//
|
||||
// Note: a slot sealed by reorder (canAppend returned false on seq
|
||||
// mismatch) keeps psh=false, so this restriction does not block the
|
||||
@@ -670,10 +671,9 @@ func canMergeSlots(prev, s *coalesceSlot) bool {
|
||||
}
|
||||
|
||||
// mergeSlots folds src into dst in place: payIovs concatenated, counters
|
||||
// and totals updated, PSH and IP-level CE bits OR'd into the seed header
|
||||
// so neither the push signal nor a CE mark is lost. The seed header's
|
||||
// seq, gsoSize, and fk are unchanged. Caller is responsible for releasing
|
||||
// src (it's no longer in c.slots after this call).
|
||||
// and totals updated, PSH OR'd into the seed header so the push signal is
|
||||
// not lost. The seed header's seq, gsoSize, and fk are unchanged. Caller
|
||||
// is responsible for releasing src (it's no longer in c.slots after this call).
|
||||
func mergeSlots(dst, src *coalesceSlot) {
|
||||
dst.payIovs = append(dst.payIovs, src.payIovs...)
|
||||
dst.numSeg += src.numSeg
|
||||
@@ -683,7 +683,6 @@ func mergeSlots(dst, src *coalesceSlot) {
|
||||
dst.psh = true
|
||||
dst.hdrBuf[dst.ipHdrLen+13] |= tcpFlagPsh
|
||||
}
|
||||
mergeECNIntoSeed(dst.hdrBuf[:dst.ipHdrLen], src.hdrBuf[:src.ipHdrLen], dst.isV6)
|
||||
}
|
||||
|
||||
// ipv4HdrChecksum computes the IPv4 header checksum over hdr (which must
|
||||
|
||||
@@ -762,39 +762,88 @@ func TestCoalescerEceMismatchReseeds(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
// TestCoalescerMergesCEMark confirms that an ECT(0) burst with a single
|
||||
// CE-marked packet still coalesces, and the merged superpacket carries CE.
|
||||
func TestCoalescerMergesCEMark(t *testing.T) {
|
||||
// TestCoalescerDifferingECNReseeds confirms that segments with differing IP
|
||||
// ECN codepoints do NOT coalesce: headersMatch compares the full ToS byte,
|
||||
// matching kernel GRO. Two ECT(0) segments merge; a CE stamp mid-run seals
|
||||
// the ECT(0) chain and starts a fresh superpacket that keeps CE; a trailing
|
||||
// ECT(0) starts yet another. Each superpacket keeps its own codepoint —
|
||||
// ORing the marks (the old buggy behavior) would have fabricated a false CE
|
||||
// across the whole burst.
|
||||
func TestCoalescerDifferingECNReseeds(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewTCPCoalescer(w, test.NewLogger(), NewArena(0))
|
||||
pay := make([]byte, 1200)
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT0, 1000, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
// Router along the path stamped CE on this one.
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnCE, 2200, tcpAck, pay)); err != nil {
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT0, 2200, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT0, 3400, tcpAck, pay)); err != nil {
|
||||
// Router along the path stamped CE on this one.
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnCE, 3400, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT0, 4600, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Flush(); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if len(w.gsoWrites) != 1 {
|
||||
t.Fatalf("want 1 merged gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
|
||||
if len(w.gsoWrites) != 3 {
|
||||
t.Fatalf("want 3 superpackets (ECN split), got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
|
||||
}
|
||||
g := w.gsoWrites[0]
|
||||
if len(g.pays) != 3 {
|
||||
t.Errorf("pay count=%d want 3", len(g.pays))
|
||||
// gso[0]: the two ECT(0) segments merged; gso[1]: CE alone; gso[2]:
|
||||
// trailing ECT(0) alone. Emitted in seq order.
|
||||
type want struct {
|
||||
pays int
|
||||
ecn byte
|
||||
}
|
||||
if got := g.hdr[1] & 0x03; got != ecnCE {
|
||||
t.Errorf("seed ECN=0x%02x want CE 0x%02x", got, ecnCE)
|
||||
wants := []want{{2, ecnECT0}, {1, ecnCE}, {1, ecnECT0}}
|
||||
for i, wnt := range wants {
|
||||
g := w.gsoWrites[i]
|
||||
if len(g.pays) != wnt.pays {
|
||||
t.Errorf("gso %d pay count=%d want %d", i, len(g.pays), wnt.pays)
|
||||
}
|
||||
if got := g.hdr[1] & 0x03; got != wnt.ecn {
|
||||
t.Errorf("gso %d ECN=0x%02x want 0x%02x", i, got, wnt.ecn)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// TestCoalescerDscpMismatchReseeds confirms that the new ECN-mask in
|
||||
// headersMatch did not also relax DSCP — different DSCP must still split.
|
||||
// TestCoalescerECT0ThenECT1NoCE is the core regression for the ECN merge
|
||||
// bug: ORing ECT(0)=0b10 with ECT(1)=0b01 fabricates CE=0b11. The two
|
||||
// segments must land in separate superpackets, each preserving its own
|
||||
// codepoint, and neither may end up CE-marked.
|
||||
func TestCoalescerECT0ThenECT1NoCE(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewTCPCoalescer(w, test.NewLogger(), NewArena(0))
|
||||
pay := make([]byte, 1200)
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT0, 1000, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv4WithToS(ecnECT1, 2200, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Flush(); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if len(w.gsoWrites) != 2 {
|
||||
t.Fatalf("want 2 separate superpackets (ECT0 vs ECT1), got %d", len(w.gsoWrites))
|
||||
}
|
||||
wantECN := []byte{ecnECT0, ecnECT1}
|
||||
for i, g := range w.gsoWrites {
|
||||
if got := g.hdr[1] & 0x03; got != wantECN[i] {
|
||||
t.Errorf("gso %d ECN=0x%02x want 0x%02x", i, got, wantECN[i])
|
||||
}
|
||||
if got := g.hdr[1] & 0x03; got == ecnCE {
|
||||
t.Errorf("gso %d fabricated CE from ECT merge", i)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// TestCoalescerDscpMismatchReseeds confirms that a DSCP difference (same
|
||||
// ECN) still splits — headersMatch compares the full ToS byte, so the upper
|
||||
// six DSCP bits must match too.
|
||||
func TestCoalescerDscpMismatchReseeds(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewTCPCoalescer(w, test.NewLogger(), NewArena(0))
|
||||
@@ -995,9 +1044,10 @@ func TestCoalescerSortKeepsPassthroughBarrier(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
// TestCoalescerIPv6MergesCEMark is the IPv6 analogue of
|
||||
// TestCoalescerMergesCEMark. ECN bits live in TC[1:0] = byte 1 mask 0x30.
|
||||
func TestCoalescerIPv6MergesCEMark(t *testing.T) {
|
||||
// TestCoalescerIPv6DifferingECNReseeds is the IPv6 analogue of
|
||||
// TestCoalescerDifferingECNReseeds. ECN bits live in TC[1:0] = byte 1 mask
|
||||
// 0x30, so ipHeadersMatch (comparing byte 1 fully) still splits them.
|
||||
func TestCoalescerIPv6DifferingECNReseeds(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewTCPCoalescer(w, test.NewLogger(), NewArena(0))
|
||||
pay := make([]byte, 1200)
|
||||
@@ -1005,20 +1055,36 @@ func TestCoalescerIPv6MergesCEMark(t *testing.T) {
|
||||
if err := c.Commit(buildTCPv6(ecnECT0, 1000, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv6(ecnCE, 2200, tcpAck, pay)); err != nil {
|
||||
if err := c.Commit(buildTCPv6(ecnECT0, 2200, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv6(ecnCE, 3400, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(buildTCPv6(ecnECT0, 4600, tcpAck, pay)); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Flush(); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if len(w.gsoWrites) != 1 {
|
||||
t.Fatalf("want 1 merged gso write, got %d", len(w.gsoWrites))
|
||||
if len(w.gsoWrites) != 3 {
|
||||
t.Fatalf("want 3 superpackets (ECN split), got %d", len(w.gsoWrites))
|
||||
}
|
||||
g := w.gsoWrites[0]
|
||||
// Byte 1 high nibble holds TC[3:0]; ECN is the low 2 bits of that nibble,
|
||||
// which appears in byte 1 mask 0x30 (>>4 to read the codepoint value).
|
||||
if got := (g.hdr[1] >> 4) & 0x03; got != ecnCE {
|
||||
t.Errorf("seed v6 ECN=0x%02x want CE 0x%02x", got, ecnCE)
|
||||
type want struct {
|
||||
pays int
|
||||
ecn byte
|
||||
}
|
||||
wants := []want{{2, ecnECT0}, {1, ecnCE}, {1, ecnECT0}}
|
||||
for i, wnt := range wants {
|
||||
g := w.gsoWrites[i]
|
||||
if len(g.pays) != wnt.pays {
|
||||
t.Errorf("gso %d pay count=%d want %d", i, len(g.pays), wnt.pays)
|
||||
}
|
||||
if got := (g.hdr[1] >> 4) & 0x03; got != wnt.ecn {
|
||||
t.Errorf("gso %d v6 ECN=0x%02x want 0x%02x", i, got, wnt.ecn)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -257,8 +257,6 @@ func (c *UDPCoalescer) appendPayload(s *udpSlot, pkt []byte, info parsedUDP) {
|
||||
s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
|
||||
s.numSeg++
|
||||
s.totalPay += info.payLen
|
||||
// Merge IP-level CE marks into the seed (same trick TCP coalescer uses).
|
||||
mergeECNIntoSeed(s.hdrBuf[:s.ipHdrLen], pkt[:s.ipHdrLen], s.isV6)
|
||||
if info.payLen < s.gsoSize {
|
||||
// Last-segment-can-be-shorter: this seals the chain.
|
||||
s.sealed = true
|
||||
@@ -329,8 +327,9 @@ func (c *UDPCoalescer) flushSlot(s *udpSlot) error {
|
||||
|
||||
// udpHeadersMatch compares two IP+UDP header prefixes for byte-equality on
|
||||
// every field that must be identical across coalesced segments. Length
|
||||
// fields and the ECN bits in IP TOS/TC are masked out — appendPayload
|
||||
// merges CE into the seed; flushSlot rewrites lengths.
|
||||
// fields are masked out (flushSlot rewrites them), but the IP-level ECN
|
||||
// codepoint is compared (via ipHeadersMatch) so segments with differing ECN
|
||||
// don't coalesce, matching kernel GRO.
|
||||
func udpHeadersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
|
||||
if len(a) != len(b) {
|
||||
return false
|
||||
|
||||
@@ -261,32 +261,37 @@ func TestUDPCoalescerCapsAtMaxSegs(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
// CE marks on appended segments must be merged into the seed's IP TOS.
|
||||
func TestUDPCoalescerMergesCEMark(t *testing.T) {
|
||||
// Differing IP ECN codepoints must not coalesce: udpHeadersMatch compares
|
||||
// the full ToS byte (matching kernel GRO). A CE-marked datagram mid-run
|
||||
// seals the Not-ECT chain and seeds a fresh superpacket that keeps CE; the
|
||||
// trailing Not-ECT datagram seeds another.
|
||||
func TestUDPCoalescerDifferingECNReseeds(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewUDPCoalescer(w, NewArena(0))
|
||||
pay := make([]byte, 800)
|
||||
pkt0 := buildUDPv4(1000, 53, pay) // ECN=00
|
||||
pkt0 := buildUDPv4(1000, 53, pay) // ECN=00 (Not-ECT)
|
||||
pkt1 := buildUDPv4(1000, 53, pay)
|
||||
pkt1[1] = 0x03 // CE
|
||||
pkt2 := buildUDPv4(1000, 53, pay)
|
||||
if err := c.Commit(pkt0); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(pkt1); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := c.Commit(pkt2); err != nil {
|
||||
t.Fatal(err)
|
||||
pkt1[1] = 0x03 // CE
|
||||
pkt2 := buildUDPv4(1000, 53, pay) // ECN=00 again
|
||||
for _, p := range [][]byte{pkt0, pkt1, pkt2} {
|
||||
if err := c.Commit(p); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
}
|
||||
if err := c.Flush(); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if len(w.gsoWrites) != 1 {
|
||||
t.Fatalf("want 1 merged gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
|
||||
if len(w.gsoWrites) != 3 {
|
||||
t.Fatalf("want 3 separate seeds (differing ECN), got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
|
||||
}
|
||||
if w.gsoWrites[0].hdr[1]&0x03 != 0x03 {
|
||||
t.Errorf("CE not merged into seed (tos=%#x)", w.gsoWrites[0].hdr[1])
|
||||
wantECN := []byte{0x00, 0x03, 0x00}
|
||||
for i, g := range w.gsoWrites {
|
||||
if len(g.pays) != 1 {
|
||||
t.Errorf("gso %d pay count=%d want 1", i, len(g.pays))
|
||||
}
|
||||
if got := g.hdr[1] & 0x03; got != wantECN[i] {
|
||||
t.Errorf("gso %d ECN=%#x want %#x", i, got, wantECN[i])
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -326,7 +331,7 @@ func TestUDPCoalescerIPv6Coalesces(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
// DSCP differences must reseed (headers don't match outside ECN).
|
||||
// DSCP differences must reseed: udpHeadersMatch compares the full ToS byte.
|
||||
func TestUDPCoalescerDSCPMismatchReseeds(t *testing.T) {
|
||||
w := &fakeTunWriter{gsoEnabled: true}
|
||||
c := NewUDPCoalescer(w, NewArena(0))
|
||||
|
||||
Reference in New Issue
Block a user