Skip to contentSkip to conversation

tcp: provide better locality for retransmit timer

PATCH Applied Eric Dumazet 4 participants 12 messages Nov 24, 2025 – Nov 26, 2025

Subsystems (9)
Networking Network drivers Documentation Kselftest +5 more subsystems
lore
Share & tools
lore

Eric Dumazet

author
#0 lore raw

cover tcp: provide better locality for retransmit timer

TCP stack uses three timers per flow, currently spread this way:

- sk->sk_timer : keepalive timer
- icsk->icsk_retransmit_timer : retransmit timer
- icsk->icsk_delack_timer : delayed ack timer

This series moves the retransmit timer to sk->sk_timer location, to increase data locality in TX paths.

keepalive timers are not often used, this change should be neutral for them.

After the series we have following fields:

- sk->tcp_retransmit_timer : retransmit timer, in sock_write_tx group
- icsk->icsk_delack_timer : delayed ack timer
- icsk->icsk_keepalive_timer : keepalive timer

Moving icsk_delack_timer in a beter location would also be welcomed.

Eric Dumazet (4):
tcp: rename icsk_timeout() to tcp_timeout_expires()
net: move sk_dst_pending_confirm and sk_pacing_status to sock_read_tx
group
tcp: introduce icsk->icsk_keepalive_timer
tcp: remove icsk->icsk_retransmit_timer

 .../net_cachelines/inet_connection_sock.rst   |  2 +-
 include/net/inet_connection_sock.h            | 20 ++++++++------
 include/net/sock.h                            | 13 ++++++---
 net/core/sock.c                               |  4 +--
 net/ipv4/inet_connection_sock.c               | 12 ++++-----
 net/ipv4/inet_diag.c                          |  8 +++---
 net/ipv4/tcp_ipv4.c                           |  8 +++---
 net/ipv4/tcp_timer.c                          | 23 ++++++++--------
 net/ipv6/tcp_ipv6.c                           |  8 +++---
 net/mptcp/protocol.c                          | 27 +++++++++----------
 net/mptcp/protocol.h                          |  2 +-
 .../selftests/bpf/progs/bpf_iter_tcp4.c       |  8 +++---
 .../selftests/bpf/progs/bpf_iter_tcp6.c       |  8 +++---
 13 files changed, 74 insertions(+), 69 deletions(-)
Signature · 1 line
2.52.0.460.gd25c4c69ec-goog
author
#1 lore raw

PATCH 1/4 tcp: rename icsk_timeout() to tcp_timeout_expires()

In preparation of sk->tcp_timeout_timer introduction, rename icsk_timeout() helper and change its argument to plain 'const struct sock *sk'.

Signed-off-by Eric Dumazet <[email protected]>
---
 include/net/inet_connection_sock.h | 5 ++---
 net/ipv4/inet_diag.c               | 4 ++--
 net/ipv4/tcp_ipv4.c                | 4 ++--
 net/ipv4/tcp_timer.c               | 6 +++---
 net/ipv6/tcp_ipv6.c                | 4 ++--
 net/mptcp/protocol.c               | 2 +-
 6 files changed, 12 insertions(+), 13 deletions(-)
include/net/inet_connection_sock.h +2 -3
index fd40af2221b99d74ce2baa79f0378686f86fe997..765c2149d6787ef1063e5f29d78547ec6ca79746 100644
@@ -184,10 +184,9 @@ static inline void inet_csk_delack_init(struct sock *sk)
 	memset(&inet_csk(sk)->icsk_ack, 0, sizeof(inet_csk(sk)->icsk_ack));
 }
 
-static inline unsigned long
-icsk_timeout(const struct inet_connection_sock *icsk)
+static inline unsigned long tcp_timeout_expires(const struct sock *sk)
 {
-	return READ_ONCE(icsk->icsk_retransmit_timer.expires);
+	return READ_ONCE(inet_csk(sk)->icsk_retransmit_timer.expires);
 }
 
 static inline unsigned long
net/ipv4/inet_diag.c +2 -2
index f0b6c5a411a2008e2a039ed37e262f3f132e58ac..9f63c09439a055550c49b659f23ff8a00ee80348 100644
@@ -287,12 +287,12 @@ int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
 		r->idiag_timer = 1;
 		r->idiag_retrans = READ_ONCE(icsk->icsk_retransmits);
 		r->idiag_expires =
-			jiffies_delta_to_msecs(icsk_timeout(icsk) - jiffies);
+			jiffies_delta_to_msecs(tcp_timeout_expires(sk) - jiffies);
 	} else if (icsk_pending == ICSK_TIME_PROBE0) {
 		r->idiag_timer = 4;
 		r->idiag_retrans = READ_ONCE(icsk->icsk_probes_out);
 		r->idiag_expires =
-			jiffies_delta_to_msecs(icsk_timeout(icsk) - jiffies);
+			jiffies_delta_to_msecs(tcp_timeout_expires(sk) - jiffies);
 	} else if (timer_pending(&sk->sk_timer)) {
 		r->idiag_timer = 2;
 		r->idiag_retrans = READ_ONCE(icsk->icsk_probes_out);
net/ipv4/tcp_ipv4.c +2 -2
index e0bb8d9e2d9c8c4a49519655d627e6e4d1b1cbac..7b8af2c8d03a4cf2c0d90029d2725c0f9dc1a071 100644
@@ -2869,10 +2869,10 @@ static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
 	    icsk_pending == ICSK_TIME_REO_TIMEOUT ||
 	    icsk_pending == ICSK_TIME_LOSS_PROBE) {
 		timer_active	= 1;
-		timer_expires	= icsk_timeout(icsk);
+		timer_expires	= tcp_timeout_expires(sk);
 	} else if (icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active	= 4;
-		timer_expires	= icsk_timeout(icsk);
+		timer_expires	= tcp_timeout_expires(sk);
 	} else if (timer_pending(&sk->sk_timer)) {
 		timer_active	= 2;
 		timer_expires	= sk->sk_timer.expires;
net/ipv4/tcp_timer.c +3 -3
index 0672c3d8f4f104e4358629ecd4039b3689d08903..afbd901e610e24c88439d5c152531074d514533a 100644
@@ -510,7 +510,7 @@ static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
 	 * and tp->rcv_tstamp might very well have been written recently.
 	 * rcv_delta can thus be negative.
 	 */
-	rcv_delta = icsk_timeout(icsk) - tp->rcv_tstamp;
+	rcv_delta = tcp_timeout_expires(sk) - tp->rcv_tstamp;
 	if (rcv_delta <= timeout)
 		return false;
 

@@ -697,9 +697,9 @@ void tcp_write_timer_handler(struct sock *sk)
 	    !icsk->icsk_pending)
 		return;
 
-	if (time_after(icsk_timeout(icsk), jiffies)) {
+	if (time_after(tcp_timeout_expires(sk), jiffies)) {
 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer,
-			       icsk_timeout(icsk));
+			       tcp_timeout_expires(sk));
 		return;
 	}
 	tcp_mstamp_refresh(tcp_sk(sk));
net/ipv6/tcp_ipv6.c +2 -2
index 08113f43012494df04ef7c66a89169856e5c98df..33c76c3a6da7cb0a1a49344ffe9ae27f0e949388 100644
@@ -2163,10 +2163,10 @@ static void get_tcp6_sock(struct seq_file *seq, struct sock *sp, int i)
 	    icsk_pending == ICSK_TIME_REO_TIMEOUT ||
 	    icsk_pending == ICSK_TIME_LOSS_PROBE) {
 		timer_active	= 1;
-		timer_expires	= icsk_timeout(icsk);
+		timer_expires	= tcp_timeout_expires(sp);
 	} else if (icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active	= 4;
-		timer_expires	= icsk_timeout(icsk);
+		timer_expires	= tcp_timeout_expires(sp);
 	} else if (timer_pending(&sp->sk_timer)) {
 		timer_active	= 2;
 		timer_expires	= sp->sk_timer.expires;
net/mptcp/protocol.c +1 -1
index 75bb1199bed9ae293e978d6f96d71fb70e9e8e13..e3fc001ea74d224ad3974c214c8e9d2c8b2fcf85 100644
@@ -524,7 +524,7 @@ static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subfl
 	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 
 	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
-	       icsk_timeout(inet_csk(ssk)) - jiffies : 0;
+	       tcp_timeout_expires(ssk) - jiffies : 0;
 }
 
 static void mptcp_set_timeout(struct sock *sk)
Signature · 1 line
2.52.0.460.gd25c4c69ec-goog
author
#3 lore raw

PATCH 2/4 net: move sk_dst_pending_confirm and sk_pacing_status to sock_read_tx group

In reply to Eric Dumazet

These two fields are mostly read in TCP tx path, move them in an more appropriate group for better cache locality.

Signed-off-by Eric Dumazet <[email protected]>
---
 include/net/sock.h | 4 ++--
 net/core/sock.c    | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)
include/net/sock.h +2 -2
index a5f36ea9d46f05970e36e3dc6c10051e5b0ecd42..7c372a9d9b10c97f9d0c4d6faab267a04bd68e0b 100644
@@ -481,8 +481,6 @@ struct sock {
 		struct rb_root	tcp_rtx_queue;
 	};
 	struct sk_buff_head	sk_write_queue;
-	u32			sk_dst_pending_confirm;
-	u32			sk_pacing_status; /* see enum sk_pacing */
 	struct page_frag	sk_frag;
 	struct timer_list	sk_timer;
 

@@ -493,6 +491,8 @@ struct sock {
 	__cacheline_group_end(sock_write_tx);
 
 	__cacheline_group_begin(sock_read_tx);
+	u32			sk_dst_pending_confirm;
+	u32			sk_pacing_status; /* see enum sk_pacing */
 	unsigned long		sk_max_pacing_rate;
 	long			sk_sndtimeo;
 	u32			sk_priority;
net/core/sock.c +2 -2
index 3b74fc71f51c13547591c4e8524d0475fcb27c8e..43553771bd5c43e951e09e4b9eeef82071937267 100644
@@ -4501,14 +4501,14 @@ static int __init sock_struct_check(void)
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_send_head);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_write_queue);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_write_pending);
-	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_dst_pending_confirm);
-	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_pacing_status);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_frag);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_timer);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_pacing_rate);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_zckey);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_write_tx, sk_tskey);
 
+	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_read_tx, sk_dst_pending_confirm);
+	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_read_tx, sk_pacing_status);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_read_tx, sk_max_pacing_rate);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_read_tx, sk_sndtimeo);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct sock, sock_read_tx, sk_priority);
Signature · 1 line
2.52.0.460.gd25c4c69ec-goog
author
#5 lore raw

PATCH 3/4 tcp: introduce icsk->icsk_keepalive_timer

In reply to Eric Dumazet

sk->sk_timer has been used for TCP keepalives.

Keepalive timers are not in fast path, we want to use sk->sk_timer storage for retransmit timers, for better cache locality.

Create icsk->icsk_keepalive_timer and change keepalive code to no longer use sk->sk_timer.

Added space is reclaimed in the following patch.

This includes changes to MPTCP, which was also using sk_timer.

Alias icsk->mptcp_tout_timer and icsk->icsk_keepalive_timer for inet_sk_diag_fill() sake.

Signed-off-by Eric Dumazet <[email protected]>
---
 .../net_cachelines/inet_connection_sock.rst           |  1 +
 include/net/inet_connection_sock.h                    | 11 +++++++++--
 net/ipv4/inet_connection_sock.c                       |  6 +++---
 net/ipv4/inet_diag.c                                  |  4 ++--
 net/ipv4/tcp_ipv4.c                                   |  4 ++--
 net/ipv4/tcp_timer.c                                  |  9 +++++----
 net/ipv6/tcp_ipv6.c                                   |  4 ++--
 net/mptcp/protocol.c                                  | 10 ++++++----
 net/mptcp/protocol.h                                  |  2 +-
 tools/testing/selftests/bpf/progs/bpf_iter_tcp4.c     |  4 ++--
 tools/testing/selftests/bpf/progs/bpf_iter_tcp6.c     |  4 ++--
 11 files changed, 35 insertions(+), 24 deletions(-)
Documentation/networking/net_cachelines/inet_connection_sock.rst +1 -0
index 8fae85ebb773085b249c606ce37872e0566b70b4..4f65de2def8c9ccef1108f8f3a3de1d8c12b8497 100644
@@ -14,6 +14,7 @@ struct inet_bind_bucket             icsk_bind_hash         read_mostly
 struct inet_bind2_bucket            icsk_bind2_hash        read_mostly                             tcp_set_state,inet_put_port
 struct timer_list                   icsk_retransmit_timer  read_write                              inet_csk_reset_xmit_timer,tcp_connect
 struct timer_list                   icsk_delack_timer      read_mostly                             inet_csk_reset_xmit_timer,tcp_connect
+struct timer_list                   icsk_keepalive_timer
 u32                                 icsk_rto               read_write                              tcp_cwnd_validate,tcp_schedule_loss_probe,tcp_connect_init,tcp_connect,tcp_write_xmit,tcp_push_one
 u32                                 icsk_rto_min
 u32                                 icsk_rto_max           read_mostly                             tcp_reset_xmit_timer
include/net/inet_connection_sock.h +9 -2
index 765c2149d6787ef1063e5f29d78547ec6ca79746..e0d90b996348d895256191a5f10275d8f3f3a69a 100644
@@ -57,6 +57,9 @@ struct inet_connection_sock_af_ops {
  * @icsk_bind_hash:	   Bind node
  * @icsk_bind2_hash:	   Bind node in the bhash2 table
  * @icsk_retransmit_timer: Resend (no ack)
+ * @icsk_delack_timer:     Delayed ACK timer
+ * @icsk_keepalive_timer:  Keepalive timer
+ * @mptcp_tout_timer: mptcp timer
  * @icsk_rto:		   Retransmit timeout
  * @icsk_pmtu_cookie	   Last pmtu seen by socket
  * @icsk_ca_ops		   Pluggable congestion control hook

@@ -81,8 +84,12 @@ struct inet_connection_sock {
 	struct request_sock_queue icsk_accept_queue;
 	struct inet_bind_bucket	  *icsk_bind_hash;
 	struct inet_bind2_bucket  *icsk_bind2_hash;
- 	struct timer_list	  icsk_retransmit_timer;
- 	struct timer_list	  icsk_delack_timer;
+	struct timer_list	  icsk_retransmit_timer;
+	struct timer_list	  icsk_delack_timer;
+	union {
+		struct timer_list icsk_keepalive_timer;
+		struct timer_list mptcp_tout_timer;
+	};
 	__u32			  icsk_rto;
 	__u32                     icsk_rto_min;
 	u32			  icsk_rto_max;
net/ipv4/inet_connection_sock.c +3 -3
index b4eae731c9ba5693b38ee063decaa6fd776d9b8b..4fc09f9bf25d59e8155107eba391f5c566f290a0 100644
@@ -739,7 +739,7 @@ void inet_csk_init_xmit_timers(struct sock *sk,
 
 	timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 	timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
-	timer_setup(&sk->sk_timer, keepalive_handler, 0);
+	timer_setup(&icsk->icsk_keepalive_timer, keepalive_handler, 0);
 	icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 }
 

@@ -752,7 +752,7 @@ void inet_csk_clear_xmit_timers(struct sock *sk)
 
 	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 	sk_stop_timer(sk, &icsk->icsk_delack_timer);
-	sk_stop_timer(sk, &sk->sk_timer);
+	sk_stop_timer(sk, &icsk->icsk_keepalive_timer);
 }
 
 void inet_csk_clear_xmit_timers_sync(struct sock *sk)

@@ -767,7 +767,7 @@ void inet_csk_clear_xmit_timers_sync(struct sock *sk)
 
 	sk_stop_timer_sync(sk, &icsk->icsk_retransmit_timer);
 	sk_stop_timer_sync(sk, &icsk->icsk_delack_timer);
-	sk_stop_timer_sync(sk, &sk->sk_timer);
+	sk_stop_timer_sync(sk, &icsk->icsk_keepalive_timer);
 }
 
 struct dst_entry *inet_csk_route_req(const struct sock *sk,
net/ipv4/inet_diag.c +2 -2
index 9f63c09439a055550c49b659f23ff8a00ee80348..3f5b1418a6109bd4e398fb2a7d95013044e75f08 100644
@@ -293,11 +293,11 @@ int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
 		r->idiag_retrans = READ_ONCE(icsk->icsk_probes_out);
 		r->idiag_expires =
 			jiffies_delta_to_msecs(tcp_timeout_expires(sk) - jiffies);
-	} else if (timer_pending(&sk->sk_timer)) {
+	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		r->idiag_timer = 2;
 		r->idiag_retrans = READ_ONCE(icsk->icsk_probes_out);
 		r->idiag_expires =
-			jiffies_delta_to_msecs(sk->sk_timer.expires - jiffies);
+			jiffies_delta_to_msecs(icsk->icsk_keepalive_timer.expires - jiffies);
 	}
 
 	if ((ext & (1 << (INET_DIAG_INFO - 1))) && handler->idiag_info_size) {
net/ipv4/tcp_ipv4.c +2 -2
index 7b8af2c8d03a4cf2c0d90029d2725c0f9dc1a071..f8a9596e8f4d41563896f02329d20b731fe7961f 100644
@@ -2873,9 +2873,9 @@ static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
 	} else if (icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active	= 4;
 		timer_expires	= tcp_timeout_expires(sk);
-	} else if (timer_pending(&sk->sk_timer)) {
+	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active	= 2;
-		timer_expires	= sk->sk_timer.expires;
+		timer_expires	= icsk->icsk_keepalive_timer.expires;
 	} else {
 		timer_active	= 0;
 		timer_expires = jiffies;
net/ipv4/tcp_timer.c +5 -4
index afbd901e610e24c88439d5c152531074d514533a..d2678dfd811806840cb332d47750dd771b20d6af 100644
@@ -755,12 +755,12 @@ void tcp_syn_ack_timeout(const struct request_sock *req)
 
 void tcp_reset_keepalive_timer(struct sock *sk, unsigned long len)
 {
-	sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
+	sk_reset_timer(sk, &inet_csk(sk)->icsk_keepalive_timer, jiffies + len);
 }
 
 static void tcp_delete_keepalive_timer(struct sock *sk)
 {
-	sk_stop_timer(sk, &sk->sk_timer);
+	sk_stop_timer(sk, &inet_csk(sk)->icsk_keepalive_timer);
 }
 
 void tcp_set_keepalive(struct sock *sk, int val)

@@ -777,8 +777,9 @@ EXPORT_IPV6_MOD_GPL(tcp_set_keepalive);
 
 static void tcp_keepalive_timer(struct timer_list *t)
 {
-	struct sock *sk = timer_container_of(sk, t, sk_timer);
-	struct inet_connection_sock *icsk = inet_csk(sk);
+	struct inet_connection_sock *icsk =
+		timer_container_of(icsk, t, icsk_keepalive_timer);
+	struct sock *sk = &icsk->icsk_inet.sk;
 	struct tcp_sock *tp = tcp_sk(sk);
 	u32 elapsed;
net/ipv6/tcp_ipv6.c +2 -2
index 33c76c3a6da7cb0a1a49344ffe9ae27f0e949388..280fe59785598e269183bf90f962ea8d58632b9a 100644
@@ -2167,9 +2167,9 @@ static void get_tcp6_sock(struct seq_file *seq, struct sock *sp, int i)
 	} else if (icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active	= 4;
 		timer_expires	= tcp_timeout_expires(sp);
-	} else if (timer_pending(&sp->sk_timer)) {
+	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active	= 2;
-		timer_expires	= sp->sk_timer.expires;
+		timer_expires	= icsk->icsk_keepalive_timer.expires;
 	} else {
 		timer_active	= 0;
 		timer_expires = jiffies;
net/mptcp/protocol.c +6 -4
index e3fc001ea74d224ad3974c214c8e9d2c8b2fcf85..6a3175c922add6d47f3268cc4cc3c663d9509cee 100644
@@ -2326,7 +2326,9 @@ static void mptcp_retransmit_timer(struct timer_list *t)
 
 static void mptcp_tout_timer(struct timer_list *t)
 {
-	struct sock *sk = timer_container_of(sk, t, sk_timer);
+	struct inet_connection_sock *icsk =
+		timer_container_of(icsk, t, mptcp_tout_timer);
+	struct sock *sk = &icsk->icsk_inet.sk;
 
 	mptcp_schedule_work(sk);
 	sock_put(sk);

@@ -2750,7 +2752,7 @@ void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
 	 */
 	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
 
-	sk_reset_timer(sk, &sk->sk_timer, timeout);
+	sk_reset_timer(sk, &inet_csk(sk)->mptcp_tout_timer, timeout);
 }
 
 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)

@@ -2875,7 +2877,7 @@ static void __mptcp_init_sock(struct sock *sk)
 
 	/* re-use the csk retrans timer for MPTCP-level retrans */
 	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
-	timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
+	timer_setup(&msk->sk.mptcp_tout_timer, mptcp_tout_timer, 0);
 }
 
 static void mptcp_ca_reset(struct sock *sk)

@@ -3077,7 +3079,7 @@ static void __mptcp_destroy_sock(struct sock *sk)
 	might_sleep();
 
 	mptcp_stop_rtx_timer(sk);
-	sk_stop_timer(sk, &sk->sk_timer);
+	sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
 	msk->pm.status = 0;
 	mptcp_release_sched(msk);
net/mptcp/protocol.h +1 -1
index a23780ff670fb2a098bf1b8ef83efa38d69beff5..f38e66cedd7e17bbbdc7f92ea2340fc90fe4f836 100644
@@ -847,7 +847,7 @@ static inline void mptcp_stop_tout_timer(struct sock *sk)
 	if (!inet_csk(sk)->icsk_mtup.probe_timestamp)
 		return;
 
-	sk_stop_timer(sk, &sk->sk_timer);
+	sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
 	inet_csk(sk)->icsk_mtup.probe_timestamp = 0;
 }
tools/testing/selftests/bpf/progs/bpf_iter_tcp4.c +2 -2
index 164640db3a29cf720e193453cab79f4bc317917c..685811326a04126f411da2199cbb5dba576cdde7 100644
@@ -103,9 +103,9 @@ static int dump_tcp_sock(struct seq_file *seq, struct tcp_sock *tp,
 	} else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active = 4;
 		timer_expires = icsk->icsk_retransmit_timer.expires;
-	} else if (timer_pending(&sp->sk_timer)) {
+	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active = 2;
-		timer_expires = sp->sk_timer.expires;
+		timer_expires = icsk->icsk_keepalive_timer.expires;
 	} else {
 		timer_active = 0;
 		timer_expires = bpf_jiffies64();
tools/testing/selftests/bpf/progs/bpf_iter_tcp6.c +2 -2
index 591c703f5032f024e4b511a6af8d63d1233a042a..0f4a927127517ce3d156c718c3ddece0407c3137 100644
@@ -103,9 +103,9 @@ static int dump_tcp6_sock(struct seq_file *seq, struct tcp6_sock *tp,
 	} else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active = 4;
 		timer_expires = icsk->icsk_retransmit_timer.expires;
-	} else if (timer_pending(&sp->sk_timer)) {
+	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active = 2;
-		timer_expires = sp->sk_timer.expires;
+		timer_expires = icsk->icsk_keepalive_timer.expires;
 	} else {
 		timer_active = 0;
 		timer_expires = bpf_jiffies64();
Signature · 1 line
2.52.0.460.gd25c4c69ec-goog
maintainer Reviewed-by on v1 3/4
#6 lore raw

Re: [PATCH net-next 3/4] tcp: introduce icsk->icsk_keepalive_timer

Quoted 9 lines · Eric Dumazet

sk->sk_timer has been used for TCP keepalives.

Keepalive timers are not in fast path, we want to use sk->sk_timer storage for retransmit timers, for better cache locality.

Create icsk->icsk_keepalive_timer and change keepalive code to no longer use sk->sk_timer.

Quoted 8 lines

Added space is reclaimed in the following patch.

This includes changes to MPTCP, which was also using sk_timer.

Alias icsk->mptcp_tout_timer and icsk->icsk_keepalive_timer for inet_sk_diag_fill() sake.

Signed-off-by Eric Dumazet <[email protected]>
Reviewed-by Kuniyuki Iwashima <[email protected]>
author
#7 lore raw

PATCH 4/4 tcp: remove icsk->icsk_retransmit_timer

In reply to Eric Dumazet

Now sk->sk_timer is no longer used by TCP keepalive, we can use its storage for TCP and MPTCP retransmit timers for better cache locality.

Signed-off-by Eric Dumazet <[email protected]>
---
 .../net_cachelines/inet_connection_sock.rst       |  1 -
 include/net/inet_connection_sock.h                |  8 +++-----
 include/net/sock.h                                |  9 +++++++--
 net/ipv4/inet_connection_sock.c                   |  6 +++---
 net/ipv4/tcp_timer.c                              |  8 +++-----
 net/mptcp/protocol.c                              | 15 +++++----------
 tools/testing/selftests/bpf/progs/bpf_iter_tcp4.c |  4 ++--
 tools/testing/selftests/bpf/progs/bpf_iter_tcp6.c |  4 ++--
 8 files changed, 25 insertions(+), 30 deletions(-)
Documentation/networking/net_cachelines/inet_connection_sock.rst +0 -1
index 4f65de2def8c9ccef1108f8f3a3de1d8c12b8497..cc2000f55c29879a12c0e4d238242b01cee18091 100644
@@ -12,7 +12,6 @@ struct inet_sock                    icsk_inet              read_mostly         r
 struct request_sock_queue           icsk_accept_queue
 struct inet_bind_bucket             icsk_bind_hash         read_mostly                             tcp_set_state
 struct inet_bind2_bucket            icsk_bind2_hash        read_mostly                             tcp_set_state,inet_put_port
-struct timer_list                   icsk_retransmit_timer  read_write                              inet_csk_reset_xmit_timer,tcp_connect
 struct timer_list                   icsk_delack_timer      read_mostly                             inet_csk_reset_xmit_timer,tcp_connect
 struct timer_list                   icsk_keepalive_timer
 u32                                 icsk_rto               read_write                              tcp_cwnd_validate,tcp_schedule_loss_probe,tcp_connect_init,tcp_connect,tcp_write_xmit,tcp_push_one
include/net/inet_connection_sock.h +3 -5
index e0d90b996348d895256191a5f10275d8f3f3a69a..ecb362025c4e5183ec78aef4b45c249da87c19ea 100644
@@ -56,7 +56,6 @@ struct inet_connection_sock_af_ops {
  * @icsk_accept_queue:	   FIFO of established children
  * @icsk_bind_hash:	   Bind node
  * @icsk_bind2_hash:	   Bind node in the bhash2 table
- * @icsk_retransmit_timer: Resend (no ack)
  * @icsk_delack_timer:     Delayed ACK timer
  * @icsk_keepalive_timer:  Keepalive timer
  * @mptcp_tout_timer: mptcp timer

@@ -84,7 +83,6 @@ struct inet_connection_sock {
 	struct request_sock_queue icsk_accept_queue;
 	struct inet_bind_bucket	  *icsk_bind_hash;
 	struct inet_bind2_bucket  *icsk_bind2_hash;
-	struct timer_list	  icsk_retransmit_timer;
 	struct timer_list	  icsk_delack_timer;
 	union {
 		struct timer_list icsk_keepalive_timer;

@@ -193,7 +191,7 @@ static inline void inet_csk_delack_init(struct sock *sk)
 
 static inline unsigned long tcp_timeout_expires(const struct sock *sk)
 {
-	return READ_ONCE(inet_csk(sk)->icsk_retransmit_timer.expires);
+	return READ_ONCE(sk->tcp_retransmit_timer.expires);
 }
 
 static inline unsigned long

@@ -209,7 +207,7 @@ static inline void inet_csk_clear_xmit_timer(struct sock *sk, const int what)
 	if (what == ICSK_TIME_RETRANS || what == ICSK_TIME_PROBE0) {
 		smp_store_release(&icsk->icsk_pending, 0);
 #ifdef INET_CSK_CLEAR_TIMERS
-		sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
+		sk_stop_timer(sk, &sk->tcp_retransmit_timer);
 #endif
 	} else if (what == ICSK_TIME_DACK) {
 		smp_store_release(&icsk->icsk_ack.pending, 0);

@@ -241,7 +239,7 @@ static inline void inet_csk_reset_xmit_timer(struct sock *sk, const int what,
 	if (what == ICSK_TIME_RETRANS || what == ICSK_TIME_PROBE0 ||
 	    what == ICSK_TIME_LOSS_PROBE || what == ICSK_TIME_REO_TIMEOUT) {
 		smp_store_release(&icsk->icsk_pending, what);
-		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, when);
+		sk_reset_timer(sk, &sk->tcp_retransmit_timer, when);
 	} else if (what == ICSK_TIME_DACK) {
 		smp_store_release(&icsk->icsk_ack.pending,
 				  icsk->icsk_ack.pending | ICSK_ACK_TIMER);
include/net/sock.h +7 -2
index 7c372a9d9b10c97f9d0c4d6faab267a04bd68e0b..ebe3897eb02a74d8cda337b8ebea4f920c826858 100644
@@ -305,6 +305,8 @@ struct sk_filter;
   *	@sk_txrehash: enable TX hash rethink
   *	@sk_filter: socket filtering instructions
   *	@sk_timer: sock cleanup timer
+  *	@tcp_retransmit_timer: tcp retransmit timer
+  *	@mptcp_retransmit_timer: mptcp retransmit timer
   *	@sk_stamp: time stamp of last packet received
   *	@sk_stamp_seq: lock for accessing sk_stamp on 32 bit architectures only
   *	@sk_tsflags: SO_TIMESTAMPING flags

@@ -482,8 +484,11 @@ struct sock {
 	};
 	struct sk_buff_head	sk_write_queue;
 	struct page_frag	sk_frag;
-	struct timer_list	sk_timer;
-
+	union {
+		struct timer_list	sk_timer;
+		struct timer_list	tcp_retransmit_timer;
+		struct timer_list	mptcp_retransmit_timer;
+	};
 	unsigned long		sk_pacing_rate; /* bytes per second */
 	atomic_t		sk_zckey;
 	atomic_t		sk_tskey;
net/ipv4/inet_connection_sock.c +3 -3
index 4fc09f9bf25d59e8155107eba391f5c566f290a0..97d57c52b9ad953d7ec1ad679237b5d122f47470 100644
@@ -737,7 +737,7 @@ void inet_csk_init_xmit_timers(struct sock *sk,
 {
 	struct inet_connection_sock *icsk = inet_csk(sk);
 
-	timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
+	timer_setup(&sk->tcp_retransmit_timer, retransmit_handler, 0);
 	timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 	timer_setup(&icsk->icsk_keepalive_timer, keepalive_handler, 0);
 	icsk->icsk_pending = icsk->icsk_ack.pending = 0;

@@ -750,7 +750,7 @@ void inet_csk_clear_xmit_timers(struct sock *sk)
 	smp_store_release(&icsk->icsk_pending, 0);
 	smp_store_release(&icsk->icsk_ack.pending, 0);
 
-	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
+	sk_stop_timer(sk, &sk->tcp_retransmit_timer);
 	sk_stop_timer(sk, &icsk->icsk_delack_timer);
 	sk_stop_timer(sk, &icsk->icsk_keepalive_timer);
 }

@@ -765,7 +765,7 @@ void inet_csk_clear_xmit_timers_sync(struct sock *sk)
 	smp_store_release(&icsk->icsk_pending, 0);
 	smp_store_release(&icsk->icsk_ack.pending, 0);
 
-	sk_stop_timer_sync(sk, &icsk->icsk_retransmit_timer);
+	sk_stop_timer_sync(sk, &sk->tcp_retransmit_timer);
 	sk_stop_timer_sync(sk, &icsk->icsk_delack_timer);
 	sk_stop_timer_sync(sk, &icsk->icsk_keepalive_timer);
 }
net/ipv4/tcp_timer.c +3 -5
index d2678dfd811806840cb332d47750dd771b20d6af..160080c9021d0717605520af3f1a47d071e7bf4d 100644
@@ -698,7 +698,7 @@ void tcp_write_timer_handler(struct sock *sk)
 		return;
 
 	if (time_after(tcp_timeout_expires(sk), jiffies)) {
-		sk_reset_timer(sk, &icsk->icsk_retransmit_timer,
+		sk_reset_timer(sk, &sk->tcp_retransmit_timer,
 			       tcp_timeout_expires(sk));
 		return;
 	}

@@ -725,12 +725,10 @@ void tcp_write_timer_handler(struct sock *sk)
 
 static void tcp_write_timer(struct timer_list *t)
 {
-	struct inet_connection_sock *icsk =
-			timer_container_of(icsk, t, icsk_retransmit_timer);
-	struct sock *sk = &icsk->icsk_inet.sk;
+	struct sock *sk = timer_container_of(sk, t, tcp_retransmit_timer);
 
 	/* Avoid locking the socket when there is no pending event. */
-	if (!smp_load_acquire(&icsk->icsk_pending))
+	if (!smp_load_acquire(&inet_csk(sk)->icsk_pending))
 		goto out;
 
 	bh_lock_sock(sk);
net/mptcp/protocol.c +5 -10
index 6a3175c922add6d47f3268cc4cc3c663d9509cee..d48c7947df6ef267ae770c63eeb70085a311a4df 100644
@@ -416,9 +416,7 @@ static bool __mptcp_move_skb(struct sock *sk, struct sk_buff *skb)
 
 static void mptcp_stop_rtx_timer(struct sock *sk)
 {
-	struct inet_connection_sock *icsk = inet_csk(sk);
-
-	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
+	sk_stop_timer(sk, &sk->mptcp_retransmit_timer);
 	mptcp_sk(sk)->timer_ival = 0;
 }
 

@@ -926,12 +924,11 @@ static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list
 
 static bool mptcp_rtx_timer_pending(struct sock *sk)
 {
-	return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
+	return timer_pending(&sk->mptcp_retransmit_timer);
 }
 
 static void mptcp_reset_rtx_timer(struct sock *sk)
 {
-	struct inet_connection_sock *icsk = inet_csk(sk);
 	unsigned long tout;
 
 	/* prevent rescheduling on close */

@@ -939,7 +936,7 @@ static void mptcp_reset_rtx_timer(struct sock *sk)
 		return;
 
 	tout = mptcp_sk(sk)->timer_ival;
-	sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
+	sk_reset_timer(sk, &sk->mptcp_retransmit_timer, jiffies + tout);
 }
 
 bool mptcp_schedule_work(struct sock *sk)

@@ -2306,9 +2303,7 @@ static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
 
 static void mptcp_retransmit_timer(struct timer_list *t)
 {
-	struct inet_connection_sock *icsk = timer_container_of(icsk, t,
-							       icsk_retransmit_timer);
-	struct sock *sk = &icsk->icsk_inet.sk;
+	struct sock *sk = timer_container_of(sk, t, mptcp_retransmit_timer);
 	struct mptcp_sock *msk = mptcp_sk(sk);
 
 	bh_lock_sock(sk);

@@ -2876,7 +2871,7 @@ static void __mptcp_init_sock(struct sock *sk)
 	spin_lock_init(&msk->fallback_lock);
 
 	/* re-use the csk retrans timer for MPTCP-level retrans */
-	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
+	timer_setup(&sk->mptcp_retransmit_timer, mptcp_retransmit_timer, 0);
 	timer_setup(&msk->sk.mptcp_tout_timer, mptcp_tout_timer, 0);
 }
tools/testing/selftests/bpf/progs/bpf_iter_tcp4.c +2 -2
index 685811326a04126f411da2199cbb5dba576cdde7..b1e509b231cd97dcc863db84ffa9bfa5897ca4ce 100644
@@ -99,10 +99,10 @@ static int dump_tcp_sock(struct seq_file *seq, struct tcp_sock *tp,
 	    icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 		timer_active = 1;
-		timer_expires = icsk->icsk_retransmit_timer.expires;
+		timer_expires = sp->tcp_retransmit_timer.expires;
 	} else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active = 4;
-		timer_expires = icsk->icsk_retransmit_timer.expires;
+		timer_expires = sp->tcp_retransmit_timer.expires;
 	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active = 2;
 		timer_expires = icsk->icsk_keepalive_timer.expires;
tools/testing/selftests/bpf/progs/bpf_iter_tcp6.c +2 -2
index 0f4a927127517ce3d156c718c3ddece0407c3137..dbc7166aee91f5403d3e275522cb652f104ac9c5 100644
@@ -99,10 +99,10 @@ static int dump_tcp6_sock(struct seq_file *seq, struct tcp6_sock *tp,
 	    icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 		timer_active = 1;
-		timer_expires = icsk->icsk_retransmit_timer.expires;
+		timer_expires = sp->tcp_retransmit_timer.expires;
 	} else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 		timer_active = 4;
-		timer_expires = icsk->icsk_retransmit_timer.expires;
+		timer_expires = sp->tcp_retransmit_timer.expires;
 	} else if (timer_pending(&icsk->icsk_keepalive_timer)) {
 		timer_active = 2;
 		timer_expires = icsk->icsk_keepalive_timer.expires;
Signature · 1 line
2.52.0.460.gd25c4c69ec-goog
CI Applied
#9 lore raw
In reply to Eric Dumazet
This series was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Hello:

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

Quoted 10 lines · On Mon, 24 Nov 2025 17:50:09 +0000 you wrote:

TCP stack uses three timers per flow, currently spread this way:

- sk->sk_timer : keepalive timer
- icsk->icsk_retransmit_timer : retransmit timer
- icsk->icsk_delack_timer : delayed ack timer

This series moves the retransmit timer to sk->sk_timer location, to increase data locality in TX paths.

[...]

Here is the summary with links:
- [net-next,1/4] tcp: rename icsk_timeout() to tcp_timeout_expires()

    https://git.kernel.org/netdev/net-next/c/3a6e8fd0bf40
  - [net-next,2/4] net: move sk_dst_pending_confirm and sk_pacing_status to sock_read_tx group
    https://git.kernel.org/netdev/net-next/c/27e8257a8651
  - [net-next,3/4] tcp: introduce icsk->icsk_keepalive_timer
    https://git.kernel.org/netdev/net-next/c/08dfe370239e
  - [net-next,4/4] tcp: remove icsk->icsk_retransmit_timer
    https://git.kernel.org/netdev/net-next/c/9a5e5334adc0

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
maintainer Acked-by on the v1 series
#10 lore raw
In reply to Eric Dumazet

Hi Eric,

Quoted 11 lines · Eric Dumazet

TCP stack uses three timers per flow, currently spread this way:

- sk->sk_timer : keepalive timer
- icsk->icsk_retransmit_timer : retransmit timer
- icsk->icsk_delack_timer : delayed ack timer

This series moves the retransmit timer to sk->sk_timer location, to increase data locality in TX paths.

keepalive timers are not often used, this change should be neutral for them.

Quoted 7 lines

After the series we have following fields:

- sk->tcp_retransmit_timer : retransmit timer, in sock_write_tx group
- icsk->icsk_delack_timer : delayed ack timer
- icsk->icsk_keepalive_timer : keepalive timer

Moving icsk_delack_timer in a beter location would also be welcomed.

Sorry for the delay, I didn't manage to look at it yesterday. I just wanted to say thank you for this series, and for having adapted MPTCP as well!

Just in case, even if the series has already been applied:

Acked-by Matthieu Baerts (NGI0) <[email protected]>

Cheers,
Matt

Signature · 1 line
Sponsored by the NGI0 Core fund.
author
#11 lore raw
Quoted 23 lines · Matthieu Baerts

Hi Eric,

Quoted 11 lines · Eric Dumazet

TCP stack uses three timers per flow, currently spread this way:

- sk->sk_timer : keepalive timer
- icsk->icsk_retransmit_timer : retransmit timer
- icsk->icsk_delack_timer : delayed ack timer

This series moves the retransmit timer to sk->sk_timer location, to increase data locality in TX paths.

keepalive timers are not often used, this change should be neutral for them.

Quoted 7 lines

After the series we have following fields:

- sk->tcp_retransmit_timer : retransmit timer, in sock_write_tx group
- icsk->icsk_delack_timer : delayed ack timer
- icsk->icsk_keepalive_timer : keepalive timer

Moving icsk_delack_timer in a beter location would also be welcomed.

Quoted 7 lines

Sorry for the delay, I didn't manage to look at it yesterday. I just wanted to say thank you for this series, and for having adapted MPTCP as well!

Just in case, even if the series has already been applied:

Acked-by Matthieu Baerts (NGI0) <[email protected]>

Thanks for checking the series !