Skip to contentSkip to conversation

tcp: rehash onto different local ECMP path on retransmit timeout

PATCH In review Neil Spring 3 participants 6 messages Jun 15

Subsystems (8)
lore
Share & tools
lore
#0 lore raw

v14 cover tcp: rehash onto different local ECMP path on retransmit timeout

Currently sk_rethink_txhash() re-rolls the socket's txhash on RTO, PLB, and spurious-retransmission events, but the new hash is not propagated into the IPv6 ECMP path selection. The cached route is reused and fib6_select_path() is never re-invoked, so the connection uses the same local ECMP decision.

This series adds the two missing pieces:

1. __sk_dst_reset() alongside sk_rethink_txhash() so the cached dst is invalidated and the next transmit triggers a fresh route lookup.

2. fl6->mp_hash set from sk_txhash before each route lookup so fib6_select_path() picks a path from the (potentially re-rolled) hash.

The override applies only to fib_multipath_hash_policy 0 (the default L3 policy). Its hash includes the flow label, but that is 0 by default (np->flow_label is unset; auto_flowlabels computes the on-wire label later, per packet), so flows to the same peer share one local path. Keying it on sk_txhash makes that local path per-connection and lets a rehash re-select it; even when a flow label is present (reflected REPFLOW or explicitly set) only local path selection changes -- the on-wire flow label is unaffected. Policies 1-3 are left unchanged.

Patch 1 is the kernel change; patch 2 adds selftests covering rehash on SYN, SYN/ACK, midstream RTO, midstream spurious-retransmission, and PLB events, plus a policy 1 negative test, a flowlabel-leak regression test, a dst-rebuild consistency test, and a syncookie path-consistency test.

Changes since v13: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Set the syncookie txhash from the cookie unconditionally, dropping the
v11 IPv6-only guard; on IPv4 it only feeds TX-queue selection (Paolo Abeni)
Patch 2:
- Pin DCTCP for the PLB test with a route congctl attribute instead of
the host-global tcp_allowed_congestion_control (Sashiko AI review)
- Drop the redundant syncookie dst-rebuild test, already covered by the
dst rebuild and syncookie path consistency tests (Sashiko AI review)
- Give each test (and retry) its own port via an alloc_ports helper to
avoid overlap (Sashiko AI review)

Changes since v12: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Factor the repeated policy-0 IPv6 mp_hash assignment into a shared ip6_ecmp_set_mp_hash() helper (Paolo Abeni)
- Replace the open-coded txhash reroll + dst reset at the three rehash
sites with a __sk_rethink_txhash_reset_dst() helper, kept separate
from sk_rethink_txhash() so dst_negative_advice()'s dst op still runs (Paolo Abeni)
Patch 2:
- Check the first rehash attempt's exit status directly instead of via
$? (shellcheck SC2181), and drop the redundant fail_reason capture on
the tolerated first attempt (Paolo Abeni)
- Redirect the remaining slowwait stdout to /dev/null so loopy_wait's
counter output cannot leak into the captured failure message
(Sashiko AI review)

Changes since v11: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Fix the IPv6-only rule to exclude IPv4-mapped connections: key the
cookie txhash on skb->protocol, not sk->sk_family (Sashiko AI review)
- Set fl6->mp_hash in tcp_v6_send_response() so RSTs and time-wait
ACKs use the connection's ECMP path (Sashiko AI review)
- Remove the bpf_sk_assign_tcp_reqsk() txhash init added in v7; it is redundant, as cookie_tcp_reqsk_init() always sets txhash before the
request socket is routed (verified by poisoning txhash and running
the tcp_custom_syncookie BPF selftest: the route lookup never saw
the poison)
- Document that policy 0 IPv6 TCP ECMP selection follows txhash over a reflected/explicit flow label (on-wire flow label unchanged)
Patch 2:
- Drain TCP teardown between rounds so late FIN/RST packets do not
pollute the next round's tc filter counters (bot+bpf-ci)
- Skip the syncookie tests when CONFIG_SYN_COOKIES is unavailable;
select it in selftests/net/config

Changes since v10: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Fix build without CONFIG_SYN_COOKIES
- Leave IPv4 syncookie txhash unmodified (net_tx_rndhash())
- Document the IPv6 TCP policy 0 behavior change in ip-sysctl.rst
Patch 2:
- Correct runtime estimate from ~15s to ~60s
- Build DCTCP as =y instead of =m to avoid module load races
- Fix false failure of the midstream ACK test by limiting the send
buffer to avoid a closed receive window; window probes do not
cause rehash

Changes since v9: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Split cookie_init_sequence() into pure computation and a new cookie_record_sent() helper for the side effects; call
cookie_record_sent() after route_req() succeeds so the overflow
timestamp and SYNCOOKIESSENT counter are not bumped when no
SYN-ACK is sent
Patch 2:
- Make midstream ACK rehash test more reliable by blocking the unused
path first
- Fix port overlap when ECMP_REBUILD_ROUNDS exceeds the default

Changes since v8: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Fix REPFLOW flowlabel reflection for syncookie SYN-ACKs: pass 0 as
tw_isn to route_req() so tcp_v6_init_req() saves ireq->pktopts
Patch 2:
- Give midstream and ACK rehash attempt helpers distinct failure
messages (no TX activity vs no data on alternate path vs counter
not incrementing) instead of a single generic error
- Drop unused ns_server parameter from ecmp_dst_rebuild_check()
- Clean up server socat before break on setup failure in the dst
rebuild loop

Changes since v7: https://lore.kernel.org/netdev/[email protected]/
Patch 1:
- Remove #if IS_ENABLED(CONFIG_IPV6) guards around __sk_dst_reset()
in tcp_plb.c and tcp_timer.c (Eric Dumazet)
- Guard mp_hash in inet6_csk_route_socket() on sk_protocol == IPPROTO_TCP instead of txhash != 0, since non-TCP callers like L2TP set sk_txhash
in __ip6_datagram_connect() and should retain flow-key-based ECMP
- Use the syncookie (ISN) as txhash for both the SYN-ACK route lookup
and cookie_v6_check() socket creation, so the server's ECMP selection is consistent across the stateless SYN-ACK and the subsequent full socket. Move cookie_init_sequence() before route_req() in tcp_conn_request()
so the SYN-ACK dst is computed with the cookie-derived txhash; derive txhash from snt_isn in cookie_tcp_reqsk_init() to match
Patch 2:
- Invalidate dst via dummy route add/del instead of route replace to
avoid a transient single-nexthop state during multipath replacement
- Add syncookie server path consistency test verifying the SYN-ACK and post-cookie ACKs use the same ECMP path
- Strengthen policy 1 negative test to wait for multiple rehash attempts
and verify SYNs landed on exactly one interface

Changes since v6: https://lore.kernel.org/netdev/[email protected]/
- Guard mp_hash assignment so that non-TCP callers of inet6_csk_route_socket() fall through to rt6_multipath_hash()
(superseded in v8 by sk_protocol == IPPROTO_TCP guard)
- Initialize txhash in bpf_sk_assign_tcp_reqsk() to avoid reading uninitialized slab memory in inet6_csk_route_req() (reverted in v12
as redundant)
- Check post-rebuild busywait return status to avoid silent false pass

Changes since v5: https://lore.kernel.org/netdev/[email protected]/
- Improve selftest reliability: suppress __dst_negative_advice() via tcp_retries1=255 in dst rebuild tests so a real RTO cannot trigger
an unintended rehash; add internal retry to midstream and ACK
rehash tests to tolerate probabilistic ECMP path selection; fix
midstream baseline capture to account for packets that bypass tc
filters during the prio qdisc's TCQ_F_CAN_BYPASS window
- Increase ECMP_REBUILD_ROUNDS default to 10 for reliable regression detection with 2-way ECMP; replace sleep with busywait
- Use tcp_allowed_congestion_control instead of changing the host's
default congestion control for PLB test (superseded in v14 by route congctl)
- Use (txhash >> 1) ?: 1 to guarantee non-zero mp_hash, since zero
falls back to rt6_multipath_hash()

Changes since v4: https://lore.kernel.org/netdev/[email protected]/
- Condition fl6->mp_hash on fib_multipath_hash_policy == 0 to preserve deterministic hash policies 1-3 (e.g., symmetric 5-tuple for policy 1)
- Set fl6->mp_hash in tcp_v6_connect() and cookie_v6_check() for
initial route lookup consistency; move sk_set_txhash() earlier
(Jakub Kicinski)
- Add policy 1 negative test; improve sysctl save/restore
- Add flowlabel leak test confirming mp_hash does not alter the
on-wire IPv6 flow label
- Add dst rebuild consistency tests (normal and syncookie) verifying
that route table changes do not cause unintended ECMP path changes

Changes since v3: https://lore.kernel.org/netdev/[email protected]/
- Use __sk_dst_reset() instead of sk_dst_reset() since the socket lock
is held in all three call sites (Eric Dumazet)
- Guard __sk_dst_reset() with sk->sk_family == AF_INET6 since IPv4 ECMP
does not use sk_txhash for path selection
- Guard __sk_dst_reset() in tcp_plb_check_rehash() with the return value
of sk_rethink_txhash()
- Move tcp_rsk(req)->txhash initialization before route_req() in tcp_conn_request() to avoid reading uninitialized memory
- Add CONFIG_TCP_CONG_DCTCP=m to selftests/net/config for PLB test
- Skip PLB test gracefully if DCTCP is not available
- Save and restore original congestion control algorithm in PLB test
- Default get_netstat_counter() to 0 when counter is not found
- Skip all tests if tcp_syn_linear_timeouts is not available
- Replace bash/pipe data sources with socat OPEN:/dev/zero for
cleaner process cleanup
- Fix shellcheck warnings

Changes since v2: https://lore.kernel.org/netdev/[email protected]/
- Retitle "ECMP" to "local ECMP" to distinguish from remote ECMP
(Neal Cardwell)
- Add fl6->mp_hash propagation in inet6_sk_rebuild_header() (af_inet6.c), covering the dst rebuild path used on established sockets
- Remove incorrect ir_iif update from tcp_check_req() in tcp_minisocks.c;
the SYN/ACK rehash is already handled by tcp_rtx_synack() re-rolling
txhash which feeds into inet6_csk_route_req()'s mp_hash
(Eric Dumazet)
- Add ACK rehash and PLB rehash selftests
- Improve selftest reliability

Changes since v1: https://lore.kernel.org/netdev/[email protected]/
- Use tcp_rsk(req)->txhash instead of jhash_1word(req->num_retrans, ...)
for ECMP path selection in inet6_csk_route_req(), making the request
socket path consistent with the established socket path (Eric Dumazet)
- Add comments explaining the >> 1 shift for 31-bit mp_hash range
- Use socat -u (unidirectional) in selftest to avoid SIGPIPE race
- Increase tcp_syn_retries and tcp_syn_linear_timeouts to 25 for
better rehash coverage

Neil Spring (2):
tcp: rehash onto different local ECMP path on retransmit timeout
selftests: net: add local ECMP rehash test

 Documentation/networking/ip-sysctl.rst     |    6 +-
 include/net/ipv6.h                         |   12 +
 include/net/sock.h                         |   14 +
 include/net/tcp.h                          |   20 +-
 net/ipv4/syncookies.c                      |    6 +-
 net/ipv4/tcp_input.c                       |   17 +-
 net/ipv4/tcp_plb.c                         |    2 +-
 net/ipv4/tcp_timer.c                       |    2 +-
 net/ipv6/af_inet6.c                        |    2 +
 net/ipv6/inet6_connection_sock.c           |    5 +
 net/ipv6/syncookies.c                      |    2 +
 net/ipv6/tcp_ipv6.c                        |   19 +-
 tools/testing/selftests/net/Makefile       |    1 +
 tools/testing/selftests/net/config         |    2 +
 tools/testing/selftests/net/ecmp_rehash.sh | 1109 ++++++++++++++++++++
 15 files changed, 1203 insertions(+), 16 deletions(-)
 create mode 100755 tools/testing/selftests/net/ecmp_rehash.sh
Signature · 1 line
2.53.0-Meta
#1 lore raw

PATCH v14 1/2 tcp: rehash onto different local ECMP path on retransmit timeout

Currently sk_rethink_txhash() re-rolls the socket's txhash on RTO, PLB, and spurious-retransmission events, but the cached route is reused and the new hash is not propagated into the ECMP path selection logic. Two changes are needed to make rehash select a different local ECMP path:

1. Add __sk_dst_reset() alongside sk_rethink_txhash() in tcp_write_timeout(), tcp_rcv_spurious_retrans(), and tcp_plb_check_rehash() so the cached dst is invalidated and the next transmit triggers a fresh route lookup.

2. Set fl6->mp_hash from sk_txhash (or tcp_rsk(req)->txhash for SYN/ACK retransmits and syncookies) in tcp_v6_connect(), inet6_sk_rebuild_header(), inet6_csk_route_req(), inet6_csk_route_socket(), tcp_v6_send_response(), and cookie_v6_check() so fib6_select_path() picks a path based on the new hash.

The mp_hash override only applies to fib_multipath_hash_policy 0 (the default L3 policy). Its hash includes the flow label, but that is 0 by default -- np->flow_label is unset, and auto_flowlabels only computes the on-wire label later, per packet -- so flows to the same peer share one local path. Keying the hash on sk_txhash makes the local path per-connection and lets a rehash re-select it. Policies 1-3 are left unchanged.

The mp_hash assignment is factored into a small helper, ip6_ecmp_set_mp_hash(), shared by inet6_csk_route_req(), inet6_csk_route_socket(), tcp_v6_connect(), inet6_sk_rebuild_header(), tcp_v6_send_response(), and cookie_v6_check(). It applies (txhash >> 1) ?: 1 for policy 0 (the >> 1 keeps mp_hash in the 31-bit range; ?: 1 keeps it non-zero, since 0 would fall back to rt6_multipath_hash()). inet6_csk_route_socket() calls it only for sk_protocol == IPPROTO_TCP so that non-TCP callers (e.g., L2TP via inet6_csk_xmit) fall through to rt6_multipath_hash() and retain their existing flow-key-based ECMP behavior.

tcp_v6_send_response() also sets mp_hash from the response txhash so that a control packet (a RST from the full socket, or an ACK from a time-wait socket) selects the same local ECMP nexthop as the connection's txhash rather than falling back to the flow hash. The time-wait socket's tw_txhash is copied from sk_txhash when the connection enters TIME_WAIT, so it reflects any rehash that occurred.

Setting mp_hash explicitly is necessary because the default ECMP hash derives from fl6->flowlabel via np->flow_label, which is not updated from sk_txhash (REPFLOW is off by default). ip6_make_flowlabel() cannot help either, as it runs after the route lookup.

As a consequence, for policy 0 the local ECMP path of an IPv6 TCP flow follows sk_txhash even when fl6->flowlabel is non-zero, e.g. a reflected (REPFLOW) or explicitly set (IPV6_FLOWLABEL_MGR) flow label. This is intentional: only local path selection changes, so rehash can recover from a failed path; the on-wire flow label is unchanged.

sk_set_txhash() is moved before ip6_dst_lookup_flow() in tcp_v6_connect() so the initial ECMP path is selected by the same txhash that subsequent route rebuilds will use. This avoids unintended path changes when the cached dst is naturally invalidated (e.g., by PMTU discovery or route changes).

The rehash sites (tcp_write_timeout(), tcp_plb_check_rehash(), and tcp_rcv_spurious_retrans()) call __sk_rethink_txhash_reset_dst(), which re-rolls the txhash and, when it changed, drops the cached dst so the next transmit re-runs route selection. The dst reset is guarded by sk->sk_family == AF_INET6 since IPv4 ECMP does not currently use sk_txhash for path selection. For IPv4-mapped IPv6 sockets this produces a redundant dst reset on a cold path (RTO/PLB); the subsequent IPv4 route lookup returns the same result. The helper is deliberately separate from sk_rethink_txhash() itself:
dst_negative_advice() calls sk_rethink_txhash() before its own dst op, so resetting the dst inside sk_rethink_txhash() would skip that op (e.g. rt6_remove_exception_rt()).

For syncookies, cookie_init_sequence() computes the cookie value before route_req() and sets txhash so the SYN-ACK selects the same ECMP path that cookie_v6_check() will use when the full socket is created. cookie_tcp_reqsk_init() derives txhash from the cookie so the full socket's ECMP path matches the SYN-ACK. Both the SYN-ACK assignment in tcp_conn_request() and the full-socket assignment in cookie_tcp_reqsk_init() set txhash from the cookie for IPv4 and IPv6 alike. On IPv6 this drives ECMP path selection; on IPv4, which does not use sk_txhash for ECMP, it only affects TX-queue selection. That selection scales the hash by its high bits (reciprocal_scale()), which are uniform in the keyed secure_tcp_syn_cookie() output -- the MSS index only perturbs the low bits -- so the queue distribution matches net_tx_rndhash().

cookie_init_sequence() is split from the former version that also called tcp_synq_overflow() and incremented SYNCOOKIESSENT; those side effects are now in cookie_record_sent(), called after route_req() succeeds so they are not bumped when route_req() fails. cookie_record_sent() is guarded by CONFIG_SYN_COOKIES to match the guard on tcp_synq_overflow(). route_req() receives 0 as tw_isn for the syncookie path so that tcp_v6_init_req() still saves ireq->pktopts for REPFLOW flowlabel reflection and IPv6 cmsg options. The ecn_ok clear for syncookies without timestamps stays after tcp_ecn_create_request() so it takes precedence.

---
 Documentation/networking/ip-sysctl.rst |  6 +++++-
 include/net/ipv6.h                     | 12 ++++++++++++
 include/net/sock.h                     | 14 ++++++++++++++
 include/net/tcp.h                      | 20 ++++++++++++++------
 net/ipv4/syncookies.c                  |  6 +++++-
 net/ipv4/tcp_input.c                   | 17 +++++++++++++----
 net/ipv4/tcp_plb.c                     |  2 +-
 net/ipv4/tcp_timer.c                   |  2 +-
 net/ipv6/af_inet6.c                    |  2 ++
 net/ipv6/inet6_connection_sock.c       |  5 +++++
 net/ipv6/syncookies.c                  |  2 ++
 net/ipv6/tcp_ipv6.c                    | 19 +++++++++++++++++--
 12 files changed, 91 insertions(+), 16 deletions(-)
Documentation/networking/ip-sysctl.rst +5 -1
index 2e3a746fcc6d..9905f5aa2427 100644
@@ -2444,7 +2444,11 @@ fib_multipath_hash_policy - INTEGER
 
 	Possible values:
 
-	- 0 - Layer 3 (source and destination addresses plus flow label)
+	- 0 - Layer 3 (source and destination addresses plus flow label).
+	  For IPv6 TCP, the local ECMP path is selected from the socket
+	  txhash rather than the flow label, and may change after a TCP
+	  rehash event (such as a retransmission timeout) to recover from
+	  path failure.  The on-wire flow label is unaffected.
 	- 1 - Layer 4 (standard 5-tuple)
 	- 2 - Layer 3 or inner Layer 3 if present
 	- 3 - Custom multipath hash. Fields used for multipath hash calculation
include/net/ipv6.h +12 -0
index d042afe7a245..8a8eb30e2980 100644
@@ -952,6 +952,18 @@ static inline u32 ip6_multipath_hash_fields(const struct net *net)
 }
 #endif
 
+/* Derive the IPv6 ECMP hash from txhash so a rehash may pick a different path;
+ * policy 0 only, and only when txhash is set.  >> 1 clears the top bit
+ * (fib6_select_path() uses mp_hash as a signed 31-bit value); ?: 1 keeps the
+ * result non-zero, since mp_hash 0 falls back to rt6_multipath_hash().
+ */
+static inline void ip6_ecmp_set_mp_hash(const struct net *net,
+					struct flowi6 *fl6, u32 txhash)
+{
+	if (ip6_multipath_hash_policy(net) == 0 && txhash)
+		fl6->mp_hash = (txhash >> 1) ?: 1;
+}
+
 /*
  *	Header manipulation
  */
include/net/sock.h +14 -0
index dccd3738c368..6ea7daab7660 100644
@@ -2252,6 +2252,20 @@ sk_dst_reset(struct sock *sk)
 	sk_dst_set(sk, NULL);
 }
 
+/* Re-roll the socket txhash.  On a rehash, IPv6 also drops the cached route
+ * so the next transmit re-selects an ECMP path; IPv4 keeps its route, since
+ * IPv4 ECMP path selection does not use sk_txhash.
+ */
+static inline bool __sk_rethink_txhash_reset_dst(struct sock *sk)
+{
+	if (sk_rethink_txhash(sk)) {
+		if (sk->sk_family == AF_INET6)
+			__sk_dst_reset(sk);
+		return true;
+	}
+	return false;
+}
+
 struct dst_entry *__sk_dst_check(struct sock *sk, u32 cookie);
 
 struct dst_entry *sk_dst_check(struct sock *sk, u32 cookie);
include/net/tcp.h +14 -6
index 3c4e6adb0dbd..75d265d19bce 100644
@@ -2540,22 +2540,30 @@ extern const struct tcp_request_sock_ops tcp_request_sock_ipv6_ops;
 
 #ifdef CONFIG_SYN_COOKIES
 static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
-					 const struct sock *sk, struct sk_buff *skb,
-					 __u16 *mss)
+					 struct sk_buff *skb, __u16 *mss)
 {
-	tcp_synq_overflow(sk);
-	__NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
 	return ops->cookie_init_seq(skb, mss);
 }
 #else
 static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
-					 const struct sock *sk, struct sk_buff *skb,
-					 __u16 *mss)
+					 struct sk_buff *skb, __u16 *mss)
 {
 	return 0;
 }
 #endif
 
+#ifdef CONFIG_SYN_COOKIES
+static inline void cookie_record_sent(const struct sock *sk)
+{
+	tcp_synq_overflow(sk);
+	__NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
+}
+#else
+static inline void cookie_record_sent(const struct sock *sk)
+{
+}
+#endif
+
 struct tcp_key {
 	union {
 		struct {
net/ipv4/syncookies.c +5 -1
index df479277fb80..73e129768184 100644
@@ -280,9 +280,13 @@ static int cookie_tcp_reqsk_init(struct sock *sk, struct sk_buff *skb,
 	treq->snt_synack = 0;
 	treq->snt_tsval_first = 0;
 	treq->tfo_listener = false;
-	treq->txhash = net_tx_rndhash();
 	treq->rcv_isn = ntohl(th->seq) - 1;
 	treq->snt_isn = ntohl(th->ack_seq) - 1;
+	/* The request socket was freed after the SYN-ACK; use the cookie
+	 * (snt_isn) as txhash so the full socket and the SYN-ACK make the
+	 * same egress choice (IPv6 ECMP path; IPv4 TX queue).
+	 */
+	treq->txhash = treq->snt_isn;
 	treq->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
 
 #if IS_ENABLED(CONFIG_MPTCP)
net/ipv4/tcp_input.c +13 -4
index 7995a89bafc9..928a065a242b 100644
@@ -5020,8 +5020,9 @@ static void tcp_rcv_spurious_retrans(struct sock *sk,
 	    skb->protocol == htons(ETH_P_IPV6) &&
 	    (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
 	     ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
-	    sk_rethink_txhash(sk))
+	    __sk_rethink_txhash_reset_dst(sk)) {
 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
+	}
 
 	/* Save last flowlabel after a spurious retrans. */
 	tcp_save_lrcv_flowlabel(sk, skb);
@@ -7636,6 +7637,7 @@ int tcp_conn_request(struct request_sock_ops *rsk_ops,
 	tcp_rsk(req)->af_specific = af_ops;
 	tcp_rsk(req)->ts_off = 0;
 	tcp_rsk(req)->req_usec_ts = false;
+	tcp_rsk(req)->txhash = net_tx_rndhash();
 #if IS_ENABLED(CONFIG_MPTCP)
 	tcp_rsk(req)->is_mptcp = 0;
 #endif
@@ -7659,7 +7661,15 @@ int tcp_conn_request(struct request_sock_ops *rsk_ops,
 	/* Note: tcp_v6_init_req() might override ir_iif for link locals */
 	inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
 
-	dst = af_ops->route_req(sk, skb, &fl, req, isn);
+	if (want_cookie) {
+		isn = cookie_init_sequence(af_ops, skb, &req->mss);
+		/* Use the cookie as txhash so the SYN-ACK and the later full
+		 * socket make the same egress choice (IPv6 ECMP path; IPv4 TX queue).
+		 */
+		tcp_rsk(req)->txhash = isn;
+	}
+
+	dst = af_ops->route_req(sk, skb, &fl, req, want_cookie ? 0 : isn);
 	if (!dst)
 		goto drop_and_free;
 
@@ -7699,7 +7709,7 @@ int tcp_conn_request(struct request_sock_ops *rsk_ops,
 	tcp_ecn_create_request(req, skb, sk, dst);
 
 	if (want_cookie) {
-		isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
+		cookie_record_sent(sk);
 		if (!tmp_opt.tstamp_ok)
 			inet_rsk(req)->ecn_ok = 0;
 	}
@@ -7717,7 +7727,6 @@ int tcp_conn_request(struct request_sock_ops *rsk_ops,
 	}
 #endif
 	tcp_rsk(req)->snt_isn = isn;
-	tcp_rsk(req)->txhash = net_tx_rndhash();
 	tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
 	tcp_openreq_init_rwin(req, sk, dst);
 	sk_rx_queue_set(req_to_sk(req), skb);
net/ipv4/tcp_plb.c +1 -1
index c11a0cd3f8fe..bcc2f0add6af 100644
@@ -78,7 +78,7 @@ void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb)
 	if (plb->pause_until)
 		return;
 
-	sk_rethink_txhash(sk);
+	__sk_rethink_txhash_reset_dst(sk);
 	plb->consec_cong_rounds = 0;
 	WRITE_ONCE(tcp_sk(sk)->plb_rehash, tcp_sk(sk)->plb_rehash + 1);
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPLBREHASH);
net/ipv4/tcp_timer.c +1 -1
index 322db13333c7..bf171b5e1eb3 100644
@@ -297,7 +297,7 @@ static int tcp_write_timeout(struct sock *sk)
 		return 1;
 	}
 
-	if (sk_rethink_txhash(sk)) {
+	if (__sk_rethink_txhash_reset_dst(sk)) {
 		WRITE_ONCE(tp->timeout_rehash, tp->timeout_rehash + 1);
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
 	}
net/ipv6/af_inet6.c +2 -0
index 0a88b376141d..a5f3327d9f7d 100644
@@ -823,6 +823,8 @@ int inet6_sk_rebuild_header(struct sock *sk)
 	fl6->flowi6_uid = sk_uid(sk);
 	security_sk_classify_flow(sk, flowi6_to_flowi_common(fl6));
 
+	ip6_ecmp_set_mp_hash(sock_net(sk), fl6, sk->sk_txhash);
+
 	rcu_read_lock();
 	final_p = fl6_update_dst(fl6, rcu_dereference(np->opt), &np->final);
 	rcu_read_unlock();
net/ipv6/inet6_connection_sock.c +5 -0
index 37534e116899..fbdb8c8b9ba1 100644
@@ -48,6 +48,8 @@ struct dst_entry *inet6_csk_route_req(const struct sock *sk,
 	fl6->flowi6_uid = sk_uid(sk);
 	security_req_classify_flow(req, flowi6_to_flowi_common(fl6));
 
+	ip6_ecmp_set_mp_hash(sock_net(sk), fl6, tcp_rsk(req)->txhash);
+
 	if (!dst) {
 		dst = ip6_dst_lookup_flow(sock_net(sk), sk, fl6, final_p);
 		if (IS_ERR(dst))
@@ -70,6 +72,9 @@ struct dst_entry *inet6_csk_route_socket(struct sock *sk,
 	fl6->saddr = np->saddr;
 	fl6->flowlabel = np->flow_label;
 	IP6_ECN_flow_xmit(sk, fl6->flowlabel);
+
+	if (sk->sk_protocol == IPPROTO_TCP)
+		ip6_ecmp_set_mp_hash(sock_net(sk), fl6, sk->sk_txhash);
 	fl6->flowi6_oif = sk->sk_bound_dev_if;
 	fl6->flowi6_mark = sk->sk_mark;
 	fl6->fl6_sport = inet->inet_sport;
net/ipv6/syncookies.c +2 -0
index 4f6f0d751d6c..b581cb1ee2e8 100644
@@ -245,6 +245,8 @@ struct sock *cookie_v6_check(struct sock *sk, struct sk_buff *skb)
 		fl6.flowi6_uid = sk_uid(sk);
 		security_req_classify_flow(req, flowi6_to_flowi_common(&fl6));
 
+		ip6_ecmp_set_mp_hash(net, &fl6, tcp_rsk(req)->txhash);
+
 		dst = ip6_dst_lookup_flow(net, sk, &fl6, final_p);
 		if (IS_ERR(dst)) {
 			SKB_DR_SET(reason, IP_OUTNOROUTES);
net/ipv6/tcp_ipv6.c +17 -2
index 2c3f7a739709..e3a99f88cb6c 100644
@@ -258,6 +258,8 @@ static int tcp_v6_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
 	if (!ipv6_addr_any(&sk->sk_v6_rcv_saddr))
 		saddr = &sk->sk_v6_rcv_saddr;
 
+	sk_set_txhash(sk);
+
 	fl6->flowi6_proto = IPPROTO_TCP;
 	fl6->daddr = sk->sk_v6_daddr;
 	fl6->saddr = saddr ? *saddr : np->saddr;
@@ -275,6 +277,14 @@ static int tcp_v6_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
 
 	security_sk_classify_flow(sk, flowi6_to_flowi_common(fl6));
 
+	/* Non-zero mp_hash bypasses rt6_multipath_hash() in
+	 * fib6_select_path(), letting txhash control ECMP path
+	 * selection so that sk_rethink_txhash() rehashes onto a
+	 * different path.  Policies 1-3 derive a deterministic
+	 * hash from the flow keys and must not be overridden.
+	 */
+	ip6_ecmp_set_mp_hash(net, fl6, sk->sk_txhash);
+
 	dst = ip6_dst_lookup_flow(net, sk, fl6, final_p);
 	if (IS_ERR(dst)) {
 		err = PTR_ERR(dst);
@@ -313,8 +323,6 @@ static int tcp_v6_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
 	if (err)
 		goto late_failure;
 
-	sk_set_txhash(sk);
-
 	if (likely(!tp->repair)) {
 		union tcp_seq_and_ts_off st;
 
@@ -955,6 +963,13 @@ static void tcp_v6_send_response(const struct sock *sk, struct sk_buff *skb, u32
 	if (txhash) {
 		/* autoflowlabel/skb_get_hash_flowi6 rely on buff->hash */
 		skb_set_hash(buff, txhash, PKT_HASH_TYPE_L4);
+
+		/* Select the local ECMP path from the connection's txhash,
+		 * so a control packet (RST, or ACK from a time-wait socket)
+		 * uses the same nexthop as the data.  Only policy 0 uses
+		 * mp_hash; policies 1-3 derive a deterministic hash.
+		 */
+		ip6_ecmp_set_mp_hash(net, &fl6, txhash);
 	}
 	fl6.flowi6_mark = IP6_REPLY_MARK(net, skb->mark) ?: mark;
 	fl6.fl6_dport = t1->dest;
Signature · 1 line
2.53.0-Meta
#3 lore raw

PATCH v14 1/1 selftests: net: add local ECMP rehash test

In reply to Neil Spring

Add ecmp_rehash.sh with nine scenarios verifying that TCP rehash selects a different local ECMP path for IPv6:

- SYN retransmission (forward path blocked during setup)
- SYN/ACK retransmission (reverse path blocked during setup)
- Midstream RTO (forward path blocked on established connection)
- Midstream ACK rehash (reverse path blocked on established connection)
- PLB rehash (ECN-driven congestion on established connection)
- Hash policy 1 negative test (rehash attempted but path unchanged)
- No flowlabel leak (client mp_hash does not alter on-wire flowlabel)
- Dst rebuild consistency (dst invalidation does not change path)
- Syncookie server path consistency (SYN-ACK and post-cookie ACKs use the same ECMP path)

The policy 1 test verifies that fib_multipath_hash_policy=1 computes a deterministic 5-tuple hash, so txhash re-rolls do not change the ECMP path while TcpTimeoutRehash still increments.

The flowlabel leak test sets auto_flowlabels=0 on the client and installs tc filters on client egress that drop TCP packets with nonzero flowlabel, confirming that the client's fl6->mp_hash does not leak into the on-wire IPv6 flow label.

The PLB test needs DCTCP, a restricted congestion control. Rather than relax the host-global tcp_allowed_congestion_control (no per-netns equivalent), it pins dctcp on the test routes via the congctl route attribute, confined to the test namespaces.

The dst rebuild test streams data, invalidates the cached dst by adding and removing a dummy route (bumping the fib6_node sernum), and verifies that traffic stays on the same path. The sernum change causes ip6_dst_check() to fail on the next transmit, triggering a fresh route lookup via inet6_csk_route_socket(). ECMP_REBUILD_ROUNDS=10 repeats the check to reduce the probability of a buggy kernel passing by chance with 2-way ECMP.

The syncookie server path consistency test verifies that the server's SYN-ACK and subsequent ACKs use the same ECMP path. With syncookies, the request socket is freed after the SYN-ACK, so cookie_tcp_reqsk_init() must derive the same txhash (from the cookie) that was used for the SYN-ACK's route lookup.

The syncookie test forces tcp_syncookies=2; it skips when CONFIG_SYN_COOKIES is not available. selftests/net/config selects it (and CONFIG_TCP_CONG_DCTCP for the PLB test).

---
 tools/testing/selftests/net/Makefile       |    1 +
 tools/testing/selftests/net/config         |    2 +
 tools/testing/selftests/net/ecmp_rehash.sh | 1109 ++++++++++++++++++++
 3 files changed, 1112 insertions(+)
 create mode 100755 tools/testing/selftests/net/ecmp_rehash.sh
tools/testing/selftests/net/Makefile +1 -0
index baa30287cf22..6ec1b24218ad 100644
@@ -26,6 +26,7 @@ TEST_PROGS := \
 	cmsg_time.sh \
 	double_udp_encap.sh \
 	drop_monitor_tests.sh \
+	ecmp_rehash.sh \
 	fcnal-ipv4.sh \
 	fcnal-ipv6.sh \
 	fcnal-other.sh \
tools/testing/selftests/net/config +2 -0
index 94d722770420..31479bc7f0c4 100644
@@ -120,8 +120,10 @@ CONFIG_OPENVSWITCH_VXLAN=m
 CONFIG_PROC_SYSCTL=y
 CONFIG_PSAMPLE=m
 CONFIG_RPS=y
+CONFIG_SYN_COOKIES=y
 CONFIG_SYSFS=y
 CONFIG_TAP=m
+CONFIG_TCP_CONG_DCTCP=y
 CONFIG_TCP_MD5SIG=y
 CONFIG_TEST_BLACKHOLE_DEV=m
 CONFIG_TEST_BPF=m
tools/testing/selftests/net/ecmp_rehash.sh new file +1109 -0

1,109 lines not on this page · raw message

Signature · 1 line
2.53.0-Meta
Reviewed-by on v14 1/2
#2 lore raw
In reply to Neil Spring
Quoted 100 lines · Neil Spring

Currently sk_rethink_txhash() re-rolls the socket's txhash on RTO, PLB, and spurious-retransmission events, but the cached route is reused and the new hash is not propagated into the ECMP path selection logic. Two changes are needed to make rehash select a different local ECMP path:

1. Add __sk_dst_reset() alongside sk_rethink_txhash() in tcp_write_timeout(), tcp_rcv_spurious_retrans(), and tcp_plb_check_rehash() so the cached dst is invalidated and the next transmit triggers a fresh route lookup.

2. Set fl6->mp_hash from sk_txhash (or tcp_rsk(req)->txhash for SYN/ACK retransmits and syncookies) in tcp_v6_connect(), inet6_sk_rebuild_header(), inet6_csk_route_req(), inet6_csk_route_socket(), tcp_v6_send_response(), and cookie_v6_check() so fib6_select_path() picks a path based on the new hash.

The mp_hash override only applies to fib_multipath_hash_policy 0 (the default L3 policy). Its hash includes the flow label, but that is 0 by default -- np->flow_label is unset, and auto_flowlabels only computes the on-wire label later, per packet -- so flows to the same peer share one local path. Keying the hash on sk_txhash makes the local path per-connection and lets a rehash re-select it. Policies 1-3 are left unchanged.

The mp_hash assignment is factored into a small helper, ip6_ecmp_set_mp_hash(), shared by inet6_csk_route_req(), inet6_csk_route_socket(), tcp_v6_connect(), inet6_sk_rebuild_header(), tcp_v6_send_response(), and cookie_v6_check(). It applies (txhash >> 1) ?: 1 for policy 0 (the >> 1 keeps mp_hash in the 31-bit range; ?: 1 keeps it non-zero, since 0 would fall back to rt6_multipath_hash()). inet6_csk_route_socket() calls it only for sk_protocol == IPPROTO_TCP so that non-TCP callers (e.g., L2TP via inet6_csk_xmit) fall through to rt6_multipath_hash() and retain their existing flow-key-based ECMP behavior.

tcp_v6_send_response() also sets mp_hash from the response txhash so that a control packet (a RST from the full socket, or an ACK from a time-wait socket) selects the same local ECMP nexthop as the connection's txhash rather than falling back to the flow hash. The time-wait socket's tw_txhash is copied from sk_txhash when the connection enters TIME_WAIT, so it reflects any rehash that occurred.

Setting mp_hash explicitly is necessary because the default ECMP hash derives from fl6->flowlabel via np->flow_label, which is not updated from sk_txhash (REPFLOW is off by default). ip6_make_flowlabel() cannot help either, as it runs after the route lookup.

As a consequence, for policy 0 the local ECMP path of an IPv6 TCP flow follows sk_txhash even when fl6->flowlabel is non-zero, e.g. a reflected (REPFLOW) or explicitly set (IPV6_FLOWLABEL_MGR) flow label. This is intentional: only local path selection changes, so rehash can recover from a failed path; the on-wire flow label is unchanged.

sk_set_txhash() is moved before ip6_dst_lookup_flow() in tcp_v6_connect() so the initial ECMP path is selected by the same txhash that subsequent route rebuilds will use. This avoids unintended path changes when the cached dst is naturally invalidated (e.g., by PMTU discovery or route changes).

The rehash sites (tcp_write_timeout(), tcp_plb_check_rehash(), and tcp_rcv_spurious_retrans()) call __sk_rethink_txhash_reset_dst(), which re-rolls the txhash and, when it changed, drops the cached dst so the next transmit re-runs route selection. The dst reset is guarded by sk->sk_family == AF_INET6 since IPv4 ECMP does not currently use sk_txhash for path selection. For IPv4-mapped IPv6 sockets this produces a redundant dst reset on a cold path (RTO/PLB); the subsequent IPv4 route lookup returns the same result. The helper is deliberately separate from sk_rethink_txhash() itself:
dst_negative_advice() calls sk_rethink_txhash() before its own dst op, so resetting the dst inside sk_rethink_txhash() would skip that op (e.g. rt6_remove_exception_rt()).

For syncookies, cookie_init_sequence() computes the cookie value before route_req() and sets txhash so the SYN-ACK selects the same ECMP path that cookie_v6_check() will use when the full socket is created. cookie_tcp_reqsk_init() derives txhash from the cookie so the full socket's ECMP path matches the SYN-ACK. Both the SYN-ACK assignment in tcp_conn_request() and the full-socket assignment in cookie_tcp_reqsk_init() set txhash from the cookie for IPv4 and IPv6 alike. On IPv6 this drives ECMP path selection; on IPv4, which does not use sk_txhash for ECMP, it only affects TX-queue selection. That selection scales the hash by its high bits (reciprocal_scale()), which are uniform in the keyed secure_tcp_syn_cookie() output -- the MSS index only perturbs the low bits -- so the queue distribution matches net_tx_rndhash().

cookie_init_sequence() is split from the former version that also called tcp_synq_overflow() and incremented SYNCOOKIESSENT; those side effects are now in cookie_record_sent(), called after route_req() succeeds so they are not bumped when route_req() fails. cookie_record_sent() is guarded by CONFIG_SYN_COOKIES to match the guard on tcp_synq_overflow(). route_req() receives 0 as tw_isn for the syncookie path so that tcp_v6_init_req() still saves ireq->pktopts for REPFLOW flowlabel reflection and IPv6 cmsg options. The ecn_ok clear for syncookies without timestamps stays after tcp_ecn_create_request() so it takes precedence.

Quoted 1 line
#4 lore raw

Re: [PATCH v14] selftests: net: add local ECMP rehash test

In reply to Neil Spring
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Quoted 14 lines · On Sun, 14 Jun 2026 21:21:58 -0700 you wrote:

Add ecmp_rehash.sh with nine scenarios verifying that TCP rehash selects a different local ECMP path for IPv6:

- SYN retransmission (forward path blocked during setup)
- SYN/ACK retransmission (reverse path blocked during setup)
- Midstream RTO (forward path blocked on established connection)
- Midstream ACK rehash (reverse path blocked on established connection)
- PLB rehash (ECN-driven congestion on established connection)
- Hash policy 1 negative test (rehash attempted but path unchanged)
- No flowlabel leak (client mp_hash does not alter on-wire flowlabel)
- Dst rebuild consistency (dst invalidation does not change path)
- Syncookie server path consistency (SYN-ACK and post-cookie ACKs use the same ECMP path)

Quoted 1 line

[...]

Here is the summary with links:
- [v14] selftests: net: add local ECMP rehash test

    https://git.kernel.org/netdev/net-next/c/ae57a5533243

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
#5 lore raw
In reply to Neil Spring
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Quoted 9 lines · On Sun, 14 Jun 2026 21:21:56 -0700 you wrote:

Currently sk_rethink_txhash() re-rolls the socket's txhash on RTO, PLB, and spurious-retransmission events, but the new hash is not propagated into the IPv6 ECMP path selection. The cached route is reused and fib6_select_path() is never re-invoked, so the connection uses the same local ECMP decision.

This series adds the two missing pieces:

[...]

Here is the summary with links:
- [v14,1/2] tcp: rehash onto different local ECMP path on retransmit timeout

    https://git.kernel.org/netdev/net-next/c/658eb696544c

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html