From 267bf3cf9a6f0ffb98b8afd983c1950e835f07c9 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:06 +0000
Subject: [PATCH 01/14] tcp: annotate data-races in tcp_get_info_chrono_stats()

tcp_get_timestamping_opt_stats() does not own the socket lock,
this is intentional.

It calls tcp_get_info_chrono_stats() while other threads could
change chrono fields in tcp_chrono_set().

I do not think we need coherent TCP socket state snapshot
in tcp_get_timestamping_opt_stats(), I chose to only
add annotations to keep KCSAN happy.

Fixes: 1c885808e456 ("tcp: SOF_TIMESTAMPING_OPT_STATS option for SO_TIMESTAMPING")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 include/net/tcp.h | 10 +++++++---
 net/ipv4/tcp.c    | 14 ++++++++++----
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/include/net/tcp.h b/include/net/tcp.h
index dfa52ceefd23..674af493882c 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -2208,10 +2208,14 @@ static inline void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new
 	const u32 now = tcp_jiffies32;
 	enum tcp_chrono old = tp->chrono_type;
 
+	/* Following WRITE_ONCE()s pair with READ_ONCE()s in
+	 * tcp_get_info_chrono_stats().
+	 */
 	if (old > TCP_CHRONO_UNSPEC)
-		tp->chrono_stat[old - 1] += now - tp->chrono_start;
-	tp->chrono_start = now;
-	tp->chrono_type = new;
+		WRITE_ONCE(tp->chrono_stat[old - 1],
+			   tp->chrono_stat[old - 1] + now - tp->chrono_start);
+	WRITE_ONCE(tp->chrono_start, now);
+	WRITE_ONCE(tp->chrono_type, new);
 }
 
 static inline void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 1a494d18c5fd..7b7812cb710f 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4191,12 +4191,18 @@ static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
 				      struct tcp_info *info)
 {
 	u64 stats[__TCP_CHRONO_MAX], total = 0;
-	enum tcp_chrono i;
+	enum tcp_chrono i, cur;
 
+	/* Following READ_ONCE()s pair with WRITE_ONCE()s in tcp_chrono_set().
+	 * This is because socket lock might not be owned by us at this point.
+	 * This is best effort, tcp_get_timestamping_opt_stats() can
+	 * see wrong values. A real fix would be too costly for TCP fast path.
+	 */
+	cur = READ_ONCE(tp->chrono_type);
 	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
-		stats[i] = tp->chrono_stat[i - 1];
-		if (i == tp->chrono_type)
-			stats[i] += tcp_jiffies32 - tp->chrono_start;
+		stats[i] = READ_ONCE(tp->chrono_stat[i - 1]);
+		if (i == cur)
+			stats[i] += tcp_jiffies32 - READ_ONCE(tp->chrono_start);
 		stats[i] *= USEC_PER_SEC / HZ;
 		total += stats[i];
 	}

From 21e92a38cfd891538598ba8f805e0165a820d532 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:07 +0000
Subject: [PATCH 02/14] tcp: add data-race annotations around tp->data_segs_out
 and tp->total_retrans

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 7e98102f4897 ("tcp: record pkts sent and retransmistted")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c        | 4 ++--
 net/ipv4/tcp_output.c | 8 +++++---
 2 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 7b7812cb710f..e39e0734d958 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4434,9 +4434,9 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
 			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
 	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
-			  tp->data_segs_out, TCP_NLA_PAD);
+			  READ_ONCE(tp->data_segs_out), TCP_NLA_PAD);
 	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
-			  tp->total_retrans, TCP_NLA_PAD);
+			  READ_ONCE(tp->total_retrans), TCP_NLA_PAD);
 
 	rate = READ_ONCE(sk->sk_pacing_rate);
 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 8e99687526a6..d8e8bba2d03a 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -1688,7 +1688,8 @@ static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
 
 	if (skb->len != tcp_header_size) {
 		tcp_event_data_sent(tp, sk);
-		tp->data_segs_out += tcp_skb_pcount(skb);
+		WRITE_ONCE(tp->data_segs_out,
+			   tp->data_segs_out + tcp_skb_pcount(skb));
 		tp->bytes_sent += skb->len - tcp_header_size;
 	}
 
@@ -3642,7 +3643,7 @@ int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
 	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
-	tp->total_retrans += segs;
+	WRITE_ONCE(tp->total_retrans, tp->total_retrans + segs);
 	tp->bytes_retrans += skb->len;
 
 	/* make sure skb->data is aligned on arches that require it
@@ -4646,7 +4647,8 @@ int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
 			 * However in this case, we are dealing with a passive fastopen
 			 * socket thus we can change total_retrans value.
 			 */
-			tcp_sk_rw(sk)->total_retrans++;
+			WRITE_ONCE(tcp_sk_rw(sk)->total_retrans,
+				   tcp_sk_rw(sk)->total_retrans + 1);
 		}
 		trace_tcp_retransmit_synack(sk, req);
 		WRITE_ONCE(req->num_retrans, req->num_retrans + 1);

From 829ba1f329cb7cbd56d599a6d225997fba66dc32 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:08 +0000
Subject: [PATCH 03/14] tcp: add data-races annotations around tp->reordering,
 tp->snd_cwnd

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE(), WRITE_ONCE() data_race() annotations to keep KCSAN happy.

Fixes: bb7c19f96012 ("tcp: add related fields into SCM_TIMESTAMPING_OPT_STATS")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 include/net/tcp.h      |  2 +-
 net/ipv4/tcp.c         |  8 ++++----
 net/ipv4/tcp_input.c   | 14 ++++++++------
 net/ipv4/tcp_metrics.c |  2 +-
 4 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/include/net/tcp.h b/include/net/tcp.h
index 674af493882c..ecbadcb3a744 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1513,7 +1513,7 @@ static inline u32 tcp_snd_cwnd(const struct tcp_sock *tp)
 static inline void tcp_snd_cwnd_set(struct tcp_sock *tp, u32 val)
 {
 	WARN_ON_ONCE((int)val <= 0);
-	tp->snd_cwnd = val;
+	WRITE_ONCE(tp->snd_cwnd, val);
 }
 
 static inline bool tcp_in_slow_start(const struct tcp_sock *tp)
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index e39e0734d958..24ba80d244b1 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4445,13 +4445,13 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	rate64 = tcp_compute_delivery_rate(tp);
 	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
 
-	nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
-	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
-	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
+	nla_put_u32(stats, TCP_NLA_SND_CWND, READ_ONCE(tp->snd_cwnd));
+	nla_put_u32(stats, TCP_NLA_REORDERING, READ_ONCE(tp->reordering));
+	nla_put_u32(stats, TCP_NLA_MIN_RTT, data_race(tcp_min_rtt(tp)));
 
 	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS,
 		   READ_ONCE(inet_csk(sk)->icsk_retransmits));
-	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
+	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, data_race(!!tp->rate_app_limited));
 	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
 	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
 	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 021f745747c5..6bb6bf049a35 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1293,8 +1293,9 @@ static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
 			 tp->sacked_out,
 			 tp->undo_marker ? tp->undo_retrans : 0);
 #endif
-		tp->reordering = min_t(u32, (metric + mss - 1) / mss,
-				       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
+		WRITE_ONCE(tp->reordering,
+			   min_t(u32, (metric + mss - 1) / mss,
+				 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)));
 	}
 
 	/* This exciting event is worth to be remembered. 8) */
@@ -2439,8 +2440,9 @@ static void tcp_check_reno_reordering(struct sock *sk, const int addend)
 	if (!tcp_limit_reno_sacked(tp))
 		return;
 
-	tp->reordering = min_t(u32, tp->packets_out + addend,
-			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
+	WRITE_ONCE(tp->reordering,
+		   min_t(u32, tp->packets_out + addend,
+			 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)));
 	tp->reord_seen++;
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
 }
@@ -2579,8 +2581,8 @@ void tcp_enter_loss(struct sock *sk)
 	reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
 	if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
 	    tp->sacked_out >= reordering)
-		tp->reordering = min_t(unsigned int, tp->reordering,
-				       reordering);
+		WRITE_ONCE(tp->reordering,
+			   min_t(unsigned int, tp->reordering, reordering));
 
 	tcp_set_ca_state(sk, TCP_CA_Loss);
 	tp->high_seq = tp->snd_nxt;
diff --git a/net/ipv4/tcp_metrics.c b/net/ipv4/tcp_metrics.c
index 06b1d5d3b6df..7a9d6d9006f6 100644
--- a/net/ipv4/tcp_metrics.c
+++ b/net/ipv4/tcp_metrics.c
@@ -496,7 +496,7 @@ void tcp_init_metrics(struct sock *sk)
 	}
 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 	if (val && tp->reordering != val)
-		tp->reordering = val;
+		WRITE_ONCE(tp->reordering, val);
 
 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 	rcu_read_unlock();

From fd571afb05ebaeac5d8f09460a0640d4cf6755f8 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:09 +0000
Subject: [PATCH 04/14] tcp: annotate data-races around tp->snd_ssthresh

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 7156d194a077 ("tcp: add snd_ssthresh stat in SCM_TIMESTAMPING_OPT_STATS")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/core/filter.c       | 2 +-
 net/ipv4/tcp.c          | 4 ++--
 net/ipv4/tcp_bbr.c      | 6 +++---
 net/ipv4/tcp_bic.c      | 2 +-
 net/ipv4/tcp_cdg.c      | 4 ++--
 net/ipv4/tcp_cubic.c    | 6 +++---
 net/ipv4/tcp_dctcp.c    | 2 +-
 net/ipv4/tcp_input.c    | 8 ++++----
 net/ipv4/tcp_metrics.c  | 4 ++--
 net/ipv4/tcp_nv.c       | 4 ++--
 net/ipv4/tcp_output.c   | 4 ++--
 net/ipv4/tcp_vegas.c    | 9 +++++----
 net/ipv4/tcp_westwood.c | 4 ++--
 net/ipv4/tcp_yeah.c     | 3 ++-
 14 files changed, 32 insertions(+), 30 deletions(-)

diff --git a/net/core/filter.c b/net/core/filter.c
index fcfcb72663ca..3b5609fb96de 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -5396,7 +5396,7 @@ static int bpf_sol_tcp_setsockopt(struct sock *sk, int optname,
 		if (val <= 0)
 			return -EINVAL;
 		tp->snd_cwnd_clamp = val;
-		tp->snd_ssthresh = val;
+		WRITE_ONCE(tp->snd_ssthresh, val);
 		break;
 	case TCP_BPF_DELACK_MAX:
 		timeout = usecs_to_jiffies(val);
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 24ba80d244b1..802a9ea05211 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -3425,7 +3425,7 @@ int tcp_disconnect(struct sock *sk, int flags)
 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
 	WRITE_ONCE(icsk->icsk_rto_min, TCP_RTO_MIN);
 	WRITE_ONCE(icsk->icsk_delack_max, TCP_DELACK_MAX);
-	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
+	WRITE_ONCE(tp->snd_ssthresh, TCP_INFINITE_SSTHRESH);
 	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
 	tp->snd_cwnd_cnt = 0;
 	tp->is_cwnd_limited = 0;
@@ -4452,7 +4452,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS,
 		   READ_ONCE(inet_csk(sk)->icsk_retransmits));
 	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, data_race(!!tp->rate_app_limited));
-	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
+	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, READ_ONCE(tp->snd_ssthresh));
 	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
 	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
 
diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
index 1ddc20a399b0..aec7805b1d37 100644
--- a/net/ipv4/tcp_bbr.c
+++ b/net/ipv4/tcp_bbr.c
@@ -897,8 +897,8 @@ static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
 
 	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
 		bbr->mode = BBR_DRAIN;	/* drain queue we created */
-		tcp_sk(sk)->snd_ssthresh =
-				bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT);
+		WRITE_ONCE(tcp_sk(sk)->snd_ssthresh,
+			   bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT));
 	}	/* fall through to check if in-flight is already small: */
 	if (bbr->mode == BBR_DRAIN &&
 	    bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <=
@@ -1043,7 +1043,7 @@ __bpf_kfunc static void bbr_init(struct sock *sk)
 	struct bbr *bbr = inet_csk_ca(sk);
 
 	bbr->prior_cwnd = 0;
-	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
+	WRITE_ONCE(tp->snd_ssthresh, TCP_INFINITE_SSTHRESH);
 	bbr->rtt_cnt = 0;
 	bbr->next_rtt_delivered = tp->delivered;
 	bbr->prev_ca_state = TCP_CA_Open;
diff --git a/net/ipv4/tcp_bic.c b/net/ipv4/tcp_bic.c
index 58358bf92e1b..65444ff14241 100644
--- a/net/ipv4/tcp_bic.c
+++ b/net/ipv4/tcp_bic.c
@@ -74,7 +74,7 @@ static void bictcp_init(struct sock *sk)
 	bictcp_reset(ca);
 
 	if (initial_ssthresh)
-		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
+		WRITE_ONCE(tcp_sk(sk)->snd_ssthresh, initial_ssthresh);
 }
 
 /*
diff --git a/net/ipv4/tcp_cdg.c b/net/ipv4/tcp_cdg.c
index ceabfd690a29..0812c390aee5 100644
--- a/net/ipv4/tcp_cdg.c
+++ b/net/ipv4/tcp_cdg.c
@@ -162,7 +162,7 @@ static void tcp_cdg_hystart_update(struct sock *sk)
 				NET_ADD_STATS(sock_net(sk),
 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
 					      tcp_snd_cwnd(tp));
-				tp->snd_ssthresh = tcp_snd_cwnd(tp);
+				WRITE_ONCE(tp->snd_ssthresh, tcp_snd_cwnd(tp));
 				return;
 			}
 		}
@@ -181,7 +181,7 @@ static void tcp_cdg_hystart_update(struct sock *sk)
 				NET_ADD_STATS(sock_net(sk),
 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
 					      tcp_snd_cwnd(tp));
-				tp->snd_ssthresh = tcp_snd_cwnd(tp);
+				WRITE_ONCE(tp->snd_ssthresh, tcp_snd_cwnd(tp));
 			}
 		}
 	}
diff --git a/net/ipv4/tcp_cubic.c b/net/ipv4/tcp_cubic.c
index ab78b5ae8d0e..119bf8cbb007 100644
--- a/net/ipv4/tcp_cubic.c
+++ b/net/ipv4/tcp_cubic.c
@@ -136,7 +136,7 @@ __bpf_kfunc static void cubictcp_init(struct sock *sk)
 		bictcp_hystart_reset(sk);
 
 	if (!hystart && initial_ssthresh)
-		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
+		WRITE_ONCE(tcp_sk(sk)->snd_ssthresh, initial_ssthresh);
 }
 
 __bpf_kfunc static void cubictcp_cwnd_event_tx_start(struct sock *sk)
@@ -420,7 +420,7 @@ static void hystart_update(struct sock *sk, u32 delay)
 				NET_ADD_STATS(sock_net(sk),
 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
 					      tcp_snd_cwnd(tp));
-				tp->snd_ssthresh = tcp_snd_cwnd(tp);
+				WRITE_ONCE(tp->snd_ssthresh, tcp_snd_cwnd(tp));
 			}
 		}
 	}
@@ -440,7 +440,7 @@ static void hystart_update(struct sock *sk, u32 delay)
 				NET_ADD_STATS(sock_net(sk),
 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
 					      tcp_snd_cwnd(tp));
-				tp->snd_ssthresh = tcp_snd_cwnd(tp);
+				WRITE_ONCE(tp->snd_ssthresh, tcp_snd_cwnd(tp));
 			}
 		}
 	}
diff --git a/net/ipv4/tcp_dctcp.c b/net/ipv4/tcp_dctcp.c
index 96c99999e09d..274e628e7cf8 100644
--- a/net/ipv4/tcp_dctcp.c
+++ b/net/ipv4/tcp_dctcp.c
@@ -177,7 +177,7 @@ static void dctcp_react_to_loss(struct sock *sk)
 	struct tcp_sock *tp = tcp_sk(sk);
 
 	ca->loss_cwnd = tcp_snd_cwnd(tp);
-	tp->snd_ssthresh = max(tcp_snd_cwnd(tp) >> 1U, 2U);
+	WRITE_ONCE(tp->snd_ssthresh, max(tcp_snd_cwnd(tp) >> 1U, 2U));
 }
 
 __bpf_kfunc static void dctcp_state(struct sock *sk, u8 new_state)
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 6bb6bf049a35..c6361447535f 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -2567,7 +2567,7 @@ void tcp_enter_loss(struct sock *sk)
 	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
 		tp->prior_cwnd = tcp_snd_cwnd(tp);
-		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
+		WRITE_ONCE(tp->snd_ssthresh, icsk->icsk_ca_ops->ssthresh(sk));
 		tcp_ca_event(sk, CA_EVENT_LOSS);
 		tcp_init_undo(tp);
 	}
@@ -2860,7 +2860,7 @@ static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
 		tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
 
 		if (tp->prior_ssthresh > tp->snd_ssthresh) {
-			tp->snd_ssthresh = tp->prior_ssthresh;
+			WRITE_ONCE(tp->snd_ssthresh, tp->prior_ssthresh);
 			tcp_ecn_withdraw_cwr(tp);
 		}
 	}
@@ -2978,7 +2978,7 @@ static void tcp_init_cwnd_reduction(struct sock *sk)
 	tp->prior_cwnd = tcp_snd_cwnd(tp);
 	tp->prr_delivered = 0;
 	tp->prr_out = 0;
-	tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
+	WRITE_ONCE(tp->snd_ssthresh, inet_csk(sk)->icsk_ca_ops->ssthresh(sk));
 	tcp_ecn_queue_cwr(tp);
 }
 
@@ -3120,7 +3120,7 @@ static void tcp_non_congestion_loss_retransmit(struct sock *sk)
 
 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
 		tp->high_seq = tp->snd_nxt;
-		tp->snd_ssthresh = tcp_current_ssthresh(sk);
+		WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk));
 		tp->prior_ssthresh = 0;
 		tp->undo_marker = 0;
 		tcp_set_ca_state(sk, TCP_CA_Loss);
diff --git a/net/ipv4/tcp_metrics.c b/net/ipv4/tcp_metrics.c
index 7a9d6d9006f6..dc0c081fc1f3 100644
--- a/net/ipv4/tcp_metrics.c
+++ b/net/ipv4/tcp_metrics.c
@@ -490,9 +490,9 @@ void tcp_init_metrics(struct sock *sk)
 	val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
 	      0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 	if (val) {
-		tp->snd_ssthresh = val;
+		WRITE_ONCE(tp->snd_ssthresh, val);
 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
-			tp->snd_ssthresh = tp->snd_cwnd_clamp;
+			WRITE_ONCE(tp->snd_ssthresh, tp->snd_cwnd_clamp);
 	}
 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 	if (val && tp->reordering != val)
diff --git a/net/ipv4/tcp_nv.c b/net/ipv4/tcp_nv.c
index a60662f4bdf9..f345897a68df 100644
--- a/net/ipv4/tcp_nv.c
+++ b/net/ipv4/tcp_nv.c
@@ -396,8 +396,8 @@ static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample)
 
 			/* We have enough data to determine we are congested */
 			ca->nv_allow_cwnd_growth = 0;
-			tp->snd_ssthresh =
-				(nv_ssthresh_factor * max_win) >> 3;
+			WRITE_ONCE(tp->snd_ssthresh,
+				   (nv_ssthresh_factor * max_win) >> 3);
 			if (tcp_snd_cwnd(tp) - max_win > 2) {
 				/* gap > 2, we do exponential cwnd decrease */
 				int dec;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index d8e8bba2d03a..2663505a0dd7 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -171,7 +171,7 @@ void tcp_cwnd_restart(struct sock *sk, s32 delta)
 
 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 
-	tp->snd_ssthresh = tcp_current_ssthresh(sk);
+	WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk));
 	restart_cwnd = min(restart_cwnd, cwnd);
 
 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
@@ -2143,7 +2143,7 @@ static void tcp_cwnd_application_limited(struct sock *sk)
 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
 		u32 win_used = max(tp->snd_cwnd_used, init_win);
 		if (win_used < tcp_snd_cwnd(tp)) {
-			tp->snd_ssthresh = tcp_current_ssthresh(sk);
+			WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk));
 			tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
 		}
 		tp->snd_cwnd_used = 0;
diff --git a/net/ipv4/tcp_vegas.c b/net/ipv4/tcp_vegas.c
index 950a66966059..574453af6bc0 100644
--- a/net/ipv4/tcp_vegas.c
+++ b/net/ipv4/tcp_vegas.c
@@ -245,7 +245,8 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 				 */
 				tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp),
 							 (u32)target_cwnd + 1));
-				tp->snd_ssthresh = tcp_vegas_ssthresh(tp);
+				WRITE_ONCE(tp->snd_ssthresh,
+					   tcp_vegas_ssthresh(tp));
 
 			} else if (tcp_in_slow_start(tp)) {
 				/* Slow start.  */
@@ -261,8 +262,8 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 					 * we slow down.
 					 */
 					tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
-					tp->snd_ssthresh
-						= tcp_vegas_ssthresh(tp);
+					WRITE_ONCE(tp->snd_ssthresh,
+						   tcp_vegas_ssthresh(tp));
 				} else if (diff < alpha) {
 					/* We don't have enough extra packets
 					 * in the network, so speed up.
@@ -280,7 +281,7 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 			else if (tcp_snd_cwnd(tp) > tp->snd_cwnd_clamp)
 				tcp_snd_cwnd_set(tp, tp->snd_cwnd_clamp);
 
-			tp->snd_ssthresh = tcp_current_ssthresh(sk);
+			WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk));
 		}
 
 		/* Wipe the slate clean for the next RTT. */
diff --git a/net/ipv4/tcp_westwood.c b/net/ipv4/tcp_westwood.c
index c6e97141eef2..b5a42adfd6ca 100644
--- a/net/ipv4/tcp_westwood.c
+++ b/net/ipv4/tcp_westwood.c
@@ -244,11 +244,11 @@ static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
 
 	switch (event) {
 	case CA_EVENT_COMPLETE_CWR:
-		tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
+		WRITE_ONCE(tp->snd_ssthresh, tcp_westwood_bw_rttmin(sk));
 		tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
 		break;
 	case CA_EVENT_LOSS:
-		tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
+		WRITE_ONCE(tp->snd_ssthresh, tcp_westwood_bw_rttmin(sk));
 		/* Update RTT_min when next ack arrives */
 		w->reset_rtt_min = 1;
 		break;
diff --git a/net/ipv4/tcp_yeah.c b/net/ipv4/tcp_yeah.c
index b22b3dccd05e..9e581154f18f 100644
--- a/net/ipv4/tcp_yeah.c
+++ b/net/ipv4/tcp_yeah.c
@@ -147,7 +147,8 @@ static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 					tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp),
 								 yeah->reno_count));
 
-					tp->snd_ssthresh = tcp_snd_cwnd(tp);
+					WRITE_ONCE(tp->snd_ssthresh,
+						   tcp_snd_cwnd(tp));
 				}
 
 				if (yeah->reno_count <= 2)

From faa886ad3ce5fc8f5156493491fe189b2b726bc9 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:10 +0000
Subject: [PATCH 05/14] tcp: annotate data-races around tp->delivered and
 tp->delivered_ce

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: feb5f2ec6464 ("tcp: export packets delivery info")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 include/net/tcp_ecn.h | 2 +-
 net/ipv4/tcp.c        | 4 ++--
 net/ipv4/tcp_input.c  | 8 ++++----
 3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/include/net/tcp_ecn.h b/include/net/tcp_ecn.h
index e9a933641636..865d5c5a7718 100644
--- a/include/net/tcp_ecn.h
+++ b/include/net/tcp_ecn.h
@@ -181,7 +181,7 @@ static inline void tcp_accecn_third_ack(struct sock *sk,
 		    tcp_accecn_validate_syn_feedback(sk, ace, sent_ect)) {
 			if ((tcp_accecn_extract_syn_ect(ace) == INET_ECN_CE) &&
 			    !tp->delivered_ce)
-				tp->delivered_ce++;
+				WRITE_ONCE(tp->delivered_ce, 1);
 		}
 		break;
 	}
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 802a9ea05211..0aabd02d4496 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4453,8 +4453,8 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 		   READ_ONCE(inet_csk(sk)->icsk_retransmits));
 	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, data_race(!!tp->rate_app_limited));
 	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, READ_ONCE(tp->snd_ssthresh));
-	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
-	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
+	nla_put_u32(stats, TCP_NLA_DELIVERED, READ_ONCE(tp->delivered));
+	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, READ_ONCE(tp->delivered_ce));
 
 	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
 	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index c6361447535f..63ff89210a72 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -476,14 +476,14 @@ static bool tcp_accecn_process_option(struct tcp_sock *tp,
 
 static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count)
 {
-	tp->delivered_ce += ecn_count;
+	WRITE_ONCE(tp->delivered_ce, tp->delivered_ce + ecn_count);
 }
 
 /* Updates the delivered and delivered_ce counts */
 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
 				bool ece_ack)
 {
-	tp->delivered += delivered;
+	WRITE_ONCE(tp->delivered, tp->delivered + delivered);
 	if (tcp_ecn_mode_rfc3168(tp) && ece_ack)
 		tcp_count_delivered_ce(tp, delivered);
 }
@@ -6779,7 +6779,7 @@ static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
 		/* SYN-data is counted as two separate packets in tcp_ack() */
 		if (tp->delivered > 1)
-			--tp->delivered;
+			WRITE_ONCE(tp->delivered, tp->delivered - 1);
 	}
 
 	tcp_fastopen_add_skb(sk, synack);
@@ -7212,7 +7212,7 @@ tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 	SKB_DR_SET(reason, NOT_SPECIFIED);
 	switch (sk->sk_state) {
 	case TCP_SYN_RECV:
-		tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
+		WRITE_ONCE(tp->delivered, tp->delivered + 1); /* SYN-ACK delivery isn't tracked in tcp_ack */
 		if (!tp->srtt_us)
 			tcp_synack_rtt_meas(sk, req);
 

From 124199444de467767175a9004e1574dc42523e62 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:11 +0000
Subject: [PATCH 06/14] tcp: add data-race annotations for TCP_NLA_SNDQ_SIZE

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 87ecc95d81d9 ("tcp: add send queue size stat in SCM_TIMESTAMPING_OPT_STATS")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-7-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c        | 4 +++-
 net/ipv4/tcp_input.c  | 4 ++--
 net/ipv4/tcp_output.c | 2 +-
 3 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 0aabd02d4496..729936d13a5c 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4456,7 +4456,9 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u32(stats, TCP_NLA_DELIVERED, READ_ONCE(tp->delivered));
 	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, READ_ONCE(tp->delivered_ce));
 
-	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
+	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE,
+		    max_t(int, 0,
+			  READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una)));
 	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
 
 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 63ff89210a72..edb5013873e0 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -3912,7 +3912,7 @@ static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
 	sock_owned_by_me((struct sock *)tp);
 	tp->bytes_acked += delta;
 	tcp_snd_sne_update(tp, ack);
-	tp->snd_una = ack;
+	WRITE_ONCE(tp->snd_una, ack);
 }
 
 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
@@ -7240,7 +7240,7 @@ tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 		if (sk->sk_socket)
 			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
 
-		tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
+		WRITE_ONCE(tp->snd_una, TCP_SKB_CB(skb)->ack_seq);
 		tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
 
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 2663505a0dd7..9f83c7e4acab 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -4153,7 +4153,7 @@ static void tcp_connect_init(struct sock *sk)
 	tp->snd_wnd = 0;
 	tcp_init_wl(tp, 0);
 	tcp_write_queue_purge(sk);
-	tp->snd_una = tp->write_seq;
+	WRITE_ONCE(tp->snd_una, tp->write_seq);
 	tp->snd_sml = tp->write_seq;
 	tp->snd_up = tp->write_seq;
 	WRITE_ONCE(tp->snd_nxt, tp->write_seq);

From ee43e957ce2ec77b2ec47fef28f3c0df6ab01a31 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:12 +0000
Subject: [PATCH 07/14] tcp: annotate data-races around tp->bytes_sent

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: ba113c3aa79a ("tcp: add data bytes sent stats")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-8-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c        | 2 +-
 net/ipv4/tcp_output.c | 3 ++-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 729936d13a5c..f999b86851cd 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4461,7 +4461,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 			  READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una)));
 	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
 
-	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
+	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, READ_ONCE(tp->bytes_sent),
 			  TCP_NLA_PAD);
 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
 			  TCP_NLA_PAD);
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 9f83c7e4acab..87af4731df87 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -1690,7 +1690,8 @@ static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
 		tcp_event_data_sent(tp, sk);
 		WRITE_ONCE(tp->data_segs_out,
 			   tp->data_segs_out + tcp_skb_pcount(skb));
-		tp->bytes_sent += skb->len - tcp_header_size;
+		WRITE_ONCE(tp->bytes_sent,
+			   tp->bytes_sent + skb->len - tcp_header_size);
 	}
 
 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)

From 5efc7b9f7cbd43401f1af81d3d7f2be00f93390d Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:13 +0000
Subject: [PATCH 08/14] tcp: annotate data-races around tp->bytes_retrans

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: fb31c9b9f6c8 ("tcp: add data bytes retransmitted stats")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-9-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c        | 4 ++--
 net/ipv4/tcp_output.c | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index f999b86851cd..8c84639dc54b 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4463,8 +4463,8 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 
 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, READ_ONCE(tp->bytes_sent),
 			  TCP_NLA_PAD);
-	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
-			  TCP_NLA_PAD);
+	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS,
+			  READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD);
 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
 	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 87af4731df87..f9d8755705f7 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -3645,7 +3645,7 @@ int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
 	WRITE_ONCE(tp->total_retrans, tp->total_retrans + segs);
-	tp->bytes_retrans += skb->len;
+	WRITE_ONCE(tp->bytes_retrans, tp->bytes_retrans + skb->len);
 
 	/* make sure skb->data is aligned on arches that require it
 	 * and check if ack-trimming & collapsing extended the headroom

From a984705ca88b976bf1087978fd98b7f3993da88c Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:14 +0000
Subject: [PATCH 09/14] tcp: annotate data-races around tp->dsack_dups

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 7e10b6554ff2 ("tcp: add dsack blocks received stats")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-10-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c       | 2 +-
 net/ipv4/tcp_input.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 8c84639dc54b..57c4dcc8bfe9 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4465,7 +4465,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 			  TCP_NLA_PAD);
 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS,
 			  READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD);
-	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
+	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups));
 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
 	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index edb5013873e0..65b3ecc6be4b 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1246,7 +1246,7 @@ static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
 	else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
 		state->flag |= FLAG_DSACK_TLP;
 
-	tp->dsack_dups += dup_segs;
+	WRITE_ONCE(tp->dsack_dups, tp->dsack_dups + dup_segs);
 	/* Skip the DSACK if dup segs weren't retransmitted by sender */
 	if (tp->dsack_dups > tp->total_retrans)
 		return 0;

From 62585690e6b2a112c408fe25f142b246ac833c42 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:15 +0000
Subject: [PATCH 10/14] tcp: annotate data-races around tp->reord_seen

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 7ec65372ca53 ("tcp: add stat of data packet reordering events")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-11-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c       | 2 +-
 net/ipv4/tcp_input.c | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 57c4dcc8bfe9..39a4b06e36bb 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4466,7 +4466,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS,
 			  READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD);
 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups));
-	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
+	nla_put_u32(stats, TCP_NLA_REORD_SEEN, READ_ONCE(tp->reord_seen));
 	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 65b3ecc6be4b..896a5a5a6b1a 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1299,7 +1299,7 @@ static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
 	}
 
 	/* This exciting event is worth to be remembered. 8) */
-	tp->reord_seen++;
+	WRITE_ONCE(tp->reord_seen, tp->reord_seen + 1);
 	NET_INC_STATS(sock_net(sk),
 		      ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
 }
@@ -2443,7 +2443,7 @@ static void tcp_check_reno_reordering(struct sock *sk, const int addend)
 	WRITE_ONCE(tp->reordering,
 		   min_t(u32, tp->packets_out + addend,
 			 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)));
-	tp->reord_seen++;
+	WRITE_ONCE(tp->reord_seen, tp->reord_seen + 1);
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
 }
 

From 290b693ce7c9d48588d88b15a782a3efc6fa036b Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:16 +0000
Subject: [PATCH 11/14] tcp: annotate data-races around tp->srtt_us

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: e8bd8fca6773 ("tcp: add SRTT to SCM_TIMESTAMPING_OPT_STATS")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-12-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c       | 5 +++--
 net/ipv4/tcp_input.c | 2 +-
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 39a4b06e36bb..541bd0d2d8c4 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -3623,7 +3623,8 @@ static void tcp_enable_tx_delay(struct sock *sk, int val)
 	if (delta && sk->sk_state == TCP_ESTABLISHED) {
 		s64 srtt = (s64)tp->srtt_us + delta;
 
-		tp->srtt_us = clamp_t(s64, srtt, 1, ~0U);
+		WRITE_ONCE(tp->srtt_us,
+			   clamp_t(s64, srtt, 1, ~0U));
 
 		/* Note: does not deal with non zero icsk_backoff */
 		tcp_set_rto(sk);
@@ -4467,7 +4468,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 			  READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD);
 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups));
 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, READ_ONCE(tp->reord_seen));
-	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
+	nla_put_u32(stats, TCP_NLA_SRTT, READ_ONCE(tp->srtt_us) >> 3);
 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
 		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 896a5a5a6b1a..e04ae105893c 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1132,7 +1132,7 @@ static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
 
 		tcp_bpf_rtt(sk, mrtt_us, srtt);
 	}
-	tp->srtt_us = max(1U, srtt);
+	WRITE_ONCE(tp->srtt_us, max(1U, srtt));
 }
 
 void tcp_update_pacing_rate(struct sock *sk)

From 71c675358b711bbfd8528949249419dc2dfa4ce1 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:17 +0000
Subject: [PATCH 12/14] tcp: annotate data-races around tp->timeout_rehash

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 32efcc06d2a1 ("tcp: export count for rehash attempts")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-13-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c       | 3 ++-
 net/ipv4/tcp_timer.c | 2 +-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 541bd0d2d8c4..192e95b71ce9 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4469,7 +4469,8 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups));
 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, READ_ONCE(tp->reord_seen));
 	nla_put_u32(stats, TCP_NLA_SRTT, READ_ONCE(tp->srtt_us) >> 3);
-	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
+	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH,
+		    READ_ONCE(tp->timeout_rehash));
 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
 		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
 	nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index ea99988795e7..8d791a954cd6 100644
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c
@@ -297,7 +297,7 @@ static int tcp_write_timeout(struct sock *sk)
 	}
 
 	if (sk_rethink_txhash(sk)) {
-		tp->timeout_rehash++;
+		WRITE_ONCE(tp->timeout_rehash, tp->timeout_rehash + 1);
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
 	}
 

From 3a63b3d160560ef51e43fb4c880a5cde8078053c Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:18 +0000
Subject: [PATCH 13/14] tcp: annotate data-races around (tp->write_seq -
 tp->snd_nxt)

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() annotations to keep KCSAN happy.

WRITE_ONCE() annotations are already present.

Fixes: e08ab0b377a1 ("tcp: add bytes not sent to SCM_TIMESTAMPING_OPT_STATS")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-14-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 192e95b71ce9..68894c03f262 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4472,7 +4472,8 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH,
 		    READ_ONCE(tp->timeout_rehash));
 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
-		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
+		    max_t(int, 0,
+			  READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt)));
 	nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
 			  TCP_NLA_PAD);
 	if (ack_skb)

From 9e89b9d03a2d2e30dcca166d5af52f9a8eceab25 Mon Sep 17 00:00:00 2001
From: Eric Dumazet <edumazet@google.com>
Date: Thu, 16 Apr 2026 20:03:19 +0000
Subject: [PATCH 14/14] tcp: annotate data-races around tp->plb_rehash

tcp_get_timestamping_opt_stats() intentionally runs lockless, we must
add READ_ONCE() and WRITE_ONCE() annotations to keep KCSAN happy.

Fixes: 29c1c44646ae ("tcp: add u32 counter in tcp_sock and an SNMP counter for PLB")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260416200319.3608680-15-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
 net/ipv4/tcp.c     | 3 ++-
 net/ipv4/tcp_plb.c | 2 +-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 68894c03f262..7fbf2fca5eb2 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4480,7 +4480,8 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
 		nla_put_u8(stats, TCP_NLA_TTL,
 			   tcp_skb_ttl_or_hop_limit(ack_skb));
 
-	nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
+	nla_put_u32(stats, TCP_NLA_REHASH,
+		    READ_ONCE(tp->plb_rehash) + READ_ONCE(tp->timeout_rehash));
 	return stats;
 }
 
diff --git a/net/ipv4/tcp_plb.c b/net/ipv4/tcp_plb.c
index 68ccdb9a5412..c11a0cd3f8fe 100644
--- a/net/ipv4/tcp_plb.c
+++ b/net/ipv4/tcp_plb.c
@@ -80,7 +80,7 @@ void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb)
 
 	sk_rethink_txhash(sk);
 	plb->consec_cong_rounds = 0;
-	tcp_sk(sk)->plb_rehash++;
+	WRITE_ONCE(tcp_sk(sk)->plb_rehash, tcp_sk(sk)->plb_rehash + 1);
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPLBREHASH);
 }
 EXPORT_SYMBOL_GPL(tcp_plb_check_rehash);