[PATCH bpf] bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

[PATCH bpf] bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

[Dudu Lu]

sk_psock_start_verdict() and sk_psock_stop_verdict() access
psock->saved_data_ready with plain loads and stores, while the sibling
sk_psock_stop_strp() uses WRITE_ONCE for the same field.  sock_map_link()
/ sock_map_unref() touch the same psock from another CPU without the
lock, so KCSAN reports the race:

  BUG: KCSAN: data-race in sock_map_link / sock_map_unref
  BUG: KCSAN: data-race in sk_psock_start_verdict / sock_map_unref

Use READ_ONCE() for the reads and WRITE_ONCE() for the writes, matching
the strp sibling.

Fixes: 2ef2b20cf4e0 ("net: annotate data-races around sk->sk_{data_ready,write_space}")
Assisted-by: LLM
Signed-off-by: Dudu Lu <phx0fer@gmail.com>
---
Tested base
-----------
Detected on Linux 7.3-rc4 (commit 93f51579e7df), arm64 QEMU guest.
Re-checked against mainline f47d223627f8 (v7.3-341): the affected code
is unchanged between the two, so the bug is still present there.
Checked against linux-next next-20261021 (commit 9a482184d910): the
affected code is unchanged there as well, so the bug is still present in
linux-next too.

Reproducer status
-----------------
A reproducer was written and executed on an arm64 QEMU guest (KCSAN); it
reported the race.  The reproducer source is not posted here per
Documentation/process/security-bugs.rst; it is available on request.

Triggering conditions
---------------------
CAP_BPF / the ability to update sockmap and attach verdict programs.

Notes
-----
Built and checkpatch-clean: this patch was applied on top of 7.3-rc4
(commit 93f51579e7df) and a full arm64 kernel (defconfig) was built with
it -- no new warnings or errors.  scripts/checkpatch.pl reports 0 errors,
0 warnings.

 net/core/skmsg.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/net/core/skmsg.c b/net/core/skmsg.c
index df385a5a961e..0bfb8c4e7f36 100644
--- a/net/core/skmsg.c
+++ b/net/core/skmsg.c
@@ -1260,10 +1260,10 @@ static void sk_psock_verdict_data_ready(struct sock *sk)
 
 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
 {
-	if (psock->saved_data_ready)
+	if (READ_ONCE(psock->saved_data_ready))
 		return;
 
-	psock->saved_data_ready = sk->sk_data_ready;
+	WRITE_ONCE(psock->saved_data_ready, sk->sk_data_ready);
 	WRITE_ONCE(sk->sk_data_ready, sk_psock_verdict_data_ready);
 	WRITE_ONCE(sk->sk_write_space, sk_psock_write_space);
 }
@@ -1273,9 +1273,9 @@ void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
 	psock_set_prog(&psock->progs.stream_verdict, NULL);
 	psock_set_prog(&psock->progs.skb_verdict, NULL);
 
-	if (!psock->saved_data_ready)
+	if (!READ_ONCE(psock->saved_data_ready))
 		return;
 
 	WRITE_ONCE(sk->sk_data_ready, psock->saved_data_ready);
-	psock->saved_data_ready = NULL;
+	WRITE_ONCE(psock->saved_data_ready, NULL);
 }

Re: [PATCH bpf] bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

[bot+bpf-ci]

[-- Attachment #1: Type: text/plain, Size: 2488 bytes --]

> bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

TL;DR: Commit message only: the Fixes: tag may point at the wrong commit;
5a685cd94b21 may be the better target (medium confidence), though 2ef2b20cf4e0
may be deliberate for stable backporting. The code change itself is not
criticized.

> bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path
>
>     sk_psock_start_verdict() and sk_psock_stop_verdict() access
>     psock->saved_data_ready with plain loads and stores, while the sibling
>     sk_psock_stop_strp() uses WRITE_ONCE for the same field.  sock_map_link()
>     / sock_map_unref() touch the same psock from another CPU without the
>     lock, so KCSAN reports the race:
>
>       BUG: KCSAN: data-race in sock_map_link / sock_map_unref
>       BUG: KCSAN: data-race in sk_psock_start_verdict / sock_map_unref
>
>     Use READ_ONCE() for the reads and WRITE_ONCE() for the writes, matching
>     the strp sibling.
>
>     Fixes: 2ef2b20cf4e0 ("net: annotate data-races around sk->sk_{data_ready,write_space}")
>     Assisted-by: LLM
>     Signed-off-by: Dudu Lu <phx0fer@gmail.com>

Is the Fixes: tag pointing at the right commit?

2ef2b20cf4e0 only annotated the sk->sk_data_ready and sk->sk_write_space
accesses, plus the saved_data_ready store in sk_psock_stop_strp().  Its
parent already had plain loads and stores of psock->saved_data_ready in
sk_psock_start_verdict() and sk_psock_stop_verdict(), so it left the verdict
path unannotated rather than introducing the accesses changed here.

The psock->saved_data_ready field, and the lockless reads of it that the
KCSAN reports are about, came from 5a685cd94b21 ("skmsg: Get rid of struct
sk_psock_parser"), which folded parser.saved_data_ready and parser.enabled
into psock->saved_data_ready.  Should the tag be:

  Fixes: 5a685cd94b21 ("skmsg: Get rid of struct sk_psock_parser")

The verdict start/stop code itself dates back to ef5659280eb1, but there the
unlocked reader tested parser.enabled rather than saved_data_ready, so this
attribution is a medium confidence suggestion.  It is also possible that
2ef2b20cf4e0 was chosen deliberately, since the context lines of this patch
depend on it for stable backporting.


---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/39631503837

Re: [PATCH bpf] bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

[Alexei Starovoitov]

On Wed, Oct 21, 2026 at 08:56 AM Dudu Lu <phx0fer@gmail.com> wrote:
> sk_psock_stop_strp() uses WRITE_ONCE for the same field.  sock_map_link()
> / sock_map_unref() touch the same psock from another CPU without the
> lock, so KCSAN reports the race:
>
>   BUG: KCSAN: data-race in sock_map_link / sock_map_unref
>   BUG: KCSAN: data-race in sk_psock_start_verdict / sock_map_unref

sashiko is right. KCSAN will print both with this patch applied.

pw-bot: cr

Re: [PATCH bpf] bpf, sockmap: use READ_ONCE/WRITE_ONCE for saved_data_ready in the verdict path

[Jakub Sitnicki]

On Wed, Oct 21, 2026 at 11:56 PM +08, Dudu Lu wrote:
> sk_psock_start_verdict() and sk_psock_stop_verdict() access
> psock->saved_data_ready with plain loads and stores, while the sibling
> sk_psock_stop_strp() uses WRITE_ONCE for the same field.  sock_map_link()
> / sock_map_unref() touch the same psock from another CPU without the
> lock, so KCSAN reports the race:
>
>   BUG: KCSAN: data-race in sock_map_link / sock_map_unref
>   BUG: KCSAN: data-race in sk_psock_start_verdict / sock_map_unref
>
> Use READ_ONCE() for the reads and WRITE_ONCE() for the writes, matching
> the strp sibling.
>
> Fixes: 2ef2b20cf4e0 ("net: annotate data-races around sk->sk_{data_ready,write_space}")
> Assisted-by: LLM
> Signed-off-by: Dudu Lu <phx0fer@gmail.com>
> ---
> Tested base
> -----------
> Detected on Linux 7.3-rc4 (commit 93f51579e7df), arm64 QEMU guest.
> Re-checked against mainline f47d223627f8 (v7.3-341): the affected code
> is unchanged between the two, so the bug is still present there.
> Checked against linux-next next-20261021 (commit 9a482184d910): the
> affected code is unchanged there as well, so the bug is still present in
> linux-next too.
>
> Reproducer status
> -----------------
> A reproducer was written and executed on an arm64 QEMU guest (KCSAN); it
> reported the race.  The reproducer source is not posted here per
> Documentation/process/security-bugs.rst; it is available on request.
>
> Triggering conditions
> ---------------------
> CAP_BPF / the ability to update sockmap and attach verdict programs.
>
> Notes
> -----
> Built and checkpatch-clean: this patch was applied on top of 7.3-rc4
> (commit 93f51579e7df) and a full arm64 kernel (defconfig) was built with
> it -- no new warnings or errors.  scripts/checkpatch.pl reports 0 errors,
> 0 warnings.
>
>  net/core/skmsg.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/net/core/skmsg.c b/net/core/skmsg.c
> index df385a5a961e..0bfb8c4e7f36 100644
> --- a/net/core/skmsg.c
> +++ b/net/core/skmsg.c
> @@ -1260,10 +1260,10 @@ static void sk_psock_verdict_data_ready(struct sock *sk)
>  
>  void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
>  {
> -	if (psock->saved_data_ready)
> +	if (READ_ONCE(psock->saved_data_ready))
>  		return;

Okay, but if we're going down that route, now sk_psock_stop_strp() is
missing a READ_ONCE around psock->saved_data_ready, or rather we should
do it once and use it through the whole function:

void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
{
	psock_set_prog(&psock->progs.stream_parser, NULL);

	if (!psock->saved_data_ready)
		return;

	WRITE_ONCE(sk->sk_data_ready, psock->saved_data_ready);
	WRITE_ONCE(psock->saved_data_ready, NULL);
	strp_stop(&psock->strp);
}

>  
> -	psock->saved_data_ready = sk->sk_data_ready;
> +	WRITE_ONCE(psock->saved_data_ready, sk->sk_data_ready);

I belive you also need to READ_ONCE(sk->sk_data_ready).

>  	WRITE_ONCE(sk->sk_data_ready, sk_psock_verdict_data_ready);
>  	WRITE_ONCE(sk->sk_write_space, sk_psock_write_space);
>  }
> @@ -1273,9 +1273,9 @@ void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
>  	psock_set_prog(&psock->progs.stream_verdict, NULL);
>  	psock_set_prog(&psock->progs.skb_verdict, NULL);
>  
> -	if (!psock->saved_data_ready)
> +	if (!READ_ONCE(psock->saved_data_ready))
>  		return;
>  
>  	WRITE_ONCE(sk->sk_data_ready, psock->saved_data_ready);
> -	psock->saved_data_ready = NULL;
> +	WRITE_ONCE(psock->saved_data_ready, NULL);
>  }

[PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Dudu Lu]

The sockmap link/unlink paths touch three psock fields from one CPU while
another CPU touches them under a different lock.  A KCSAN run with two
SOCKMAPs sharing one socket reports all three:

  BUG: KCSAN: data-race in sock_map_link / sock_map_unref
  BUG: KCSAN: data-race in sock_map_link / unix_dgram_bpf_update_proto
  BUG: KCSAN: data-race in sk_psock_start_verdict / sock_map_unref

  - psock->psock_update_sk_prot: written in sock_map_init_proto() without
    sk->sk_callback_lock, read in sock_map_del_link() under it.
  - psock->saved_data_ready: written by sk_psock_start_verdict() under
    sk->sk_callback_lock, read in sock_map_del_link() under
    psock->link_lock -- a different lock instance.
  - sk->sk_prot: replaced by sock_replace_proto() (via the protocol's
    psock_update_sk_prot()) under sk->sk_callback_lock, read by
    sock_map_link() / sock_map_sk_is_suitable() outside it.

Annotate every one of these accesses with READ_ONCE()/WRITE_ONCE(),
matching the existing convention for the same fields: sk->sk_data_ready
and sk->sk_write_space were already annotated this way, as were the
sk_psock_stop_strp() accesses to saved_data_ready.  The indirect calls
read the pointer once into a local, so the guard and the call cannot see
different values.

No Fixes: tag: the three fields were introduced by three different
commits, so no single one is the origin.  psock->saved_data_ready was
folded into struct sk_psock by commit 5a685cd94b21 ("skmsg: Get rid of
struct sk_psock_parser"); the psock->psock_update_sk_prot read in
sock_map_del_link() came with commit c0d95d3380ee ("bpf, sockmap:
Re-evaluate proto ops when psock is removed from sockmap"); the lockless
sk->sk_prot reads are older still.

Assisted-by: LLM
Signed-off-by: Dudu Lu <phx0fer@gmail.com>
---
v2:
 - Drop the Fixes: tag; the three fields came from three different
   commits, so explain the provenance in the message instead (bpf-ci).
 - Read the function pointer once into a local before the guard and the
   call, so the two cannot see different values: psock->saved_data_ready
   in sk_psock_stop_strp() and sk_psock_stop_verdict(), and
   psock->psock_update_sk_prot in sk_psock_restore_proto(),
   sock_map_del_link() and sock_map_init_proto() (Jakub).
 - READ_ONCE() the source sk->sk_data_ready in sk_psock_start_strp()
   and sk_psock_start_verdict() (Jakub).
 - Annotate the sock_map.c accesses that KCSAN actually reports:
   sock_map_init_proto(), sock_map_del_link(), sock_map_link(),
   sock_map_psock_get_checked() and sock_map_sk_is_suitable()
   (Alexei).

Tested base
-----------
Detected on Linux 7.3-rc4 (commit 93f51579e7df), arm64 QEMU guest.
Re-checked against mainline f47d223627f8 (v7.3-341): the affected code
is unchanged between the two, so the bug is still present there.
Checked against linux-next next-20261021 (commit 9a482184d910): the
affected code is unchanged there as well, so the bug is still present in
linux-next too.

Reproducer status
-----------------
A reproducer was written and executed on an arm64 QEMU guest (KCSAN); it
reported the races above.  The reproducer source is not posted here per
Documentation/process/security-bugs.rst; it is available on request.

Post-fix verification
---------------------
Tested: the patch was built for arm64 and booted in a QEMU guest with
KCSAN; the reproducer passes.  Concretely, the reproducer was re-run on
the guest, once on the base tree and once with this patch applied.  On
the base tree the sockmap-path races are five -- the three listed above,
plus

  BUG: KCSAN: data-race in sock_map_update_elem_sys / unix_dgram_bpf_update_proto
  BUG: KCSAN: data-race in sk_psock_stop_verdict / sock_map_unref

With this patch all five are gone and no new race is introduced.  The map
operations complete identically in both runs (same op counts, no failures,
no KASAN memory errors), so the annotation does not change behaviour.

The reproducer also reports four unrelated races that this patch does not
touch and that fire the same way with or without it: three are the
socket-lock fast path (the 8-byte try_cmpxchg on sk_lock.combined against
the slock-protected sk_lock.owned write, by design), and one is a plain
sk->sk_prot read in generic release_sock(), which is a different scope
from the sockmap paths annotated here.

Triggering conditions
---------------------
CAP_BPF / the ability to update sockmap and attach verdict programs.
The two writers run on separate CPUs (the guest has 2), one cycling a
map with a verdict program, the other a bare map, on the same socket.

Notes
-----
Built and checkpatch-clean: this patch was applied on top of 7.3-rc4
(commit 93f51579e7df) and the affected objects were rebuilt for arm64
with the KCSAN config -- no new warnings or errors.  scripts/checkpatch.pl
reports 0 errors, 1 warning (its 'BUG: KCSAN:' commit-message
heuristic, i.e. it is asking for a Fixes: tag, which we deliberately omit).

 include/linux/skmsg.h |  7 +++++--
 net/core/skmsg.c      | 24 +++++++++++++++---------
 net/core/sock_map.c   | 34 +++++++++++++++++++++-------------
 3 files changed, 41 insertions(+), 24 deletions(-)

diff --git a/include/linux/skmsg.h b/include/linux/skmsg.h
index d5e35f24738d..cb17426f63be 100644
--- a/include/linux/skmsg.h
+++ b/include/linux/skmsg.h
@@ -488,8 +488,11 @@ static inline void sk_psock_cork_free(struct sk_psock *psock)
 static inline void sk_psock_restore_proto(struct sock *sk,
 					  struct sk_psock *psock)
 {
-	if (psock->psock_update_sk_prot)
-		psock->psock_update_sk_prot(sk, psock, true);
+	int (*update_proto)(struct sock *sk, struct sk_psock *psock, bool restore);
+
+	update_proto = READ_ONCE(psock->psock_update_sk_prot);
+	if (update_proto)
+		update_proto(sk, psock, true);
 }
 
 static inline struct sk_psock *sk_psock_get(struct sock *sk)
diff --git a/net/core/skmsg.c b/net/core/skmsg.c
index df385a5a961e..54d832039f09 100644
--- a/net/core/skmsg.c
+++ b/net/core/skmsg.c
@@ -1167,22 +1167,25 @@ int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
 
 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
 {
-	if (psock->saved_data_ready)
+	if (READ_ONCE(psock->saved_data_ready))
 		return;
 
-	psock->saved_data_ready = sk->sk_data_ready;
+	WRITE_ONCE(psock->saved_data_ready, READ_ONCE(sk->sk_data_ready));
 	WRITE_ONCE(sk->sk_data_ready, sk_psock_strp_data_ready);
 	WRITE_ONCE(sk->sk_write_space, sk_psock_write_space);
 }
 
 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
 {
+	void (*saved_data_ready)(struct sock *sk);
+
 	psock_set_prog(&psock->progs.stream_parser, NULL);
 
-	if (!psock->saved_data_ready)
+	saved_data_ready = READ_ONCE(psock->saved_data_ready);
+	if (!saved_data_ready)
 		return;
 
-	WRITE_ONCE(sk->sk_data_ready, psock->saved_data_ready);
+	WRITE_ONCE(sk->sk_data_ready, saved_data_ready);
 	WRITE_ONCE(psock->saved_data_ready, NULL);
 	strp_stop(&psock->strp);
 }
@@ -1260,22 +1263,25 @@ static void sk_psock_verdict_data_ready(struct sock *sk)
 
 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock)
 {
-	if (psock->saved_data_ready)
+	if (READ_ONCE(psock->saved_data_ready))
 		return;
 
-	psock->saved_data_ready = sk->sk_data_ready;
+	WRITE_ONCE(psock->saved_data_ready, READ_ONCE(sk->sk_data_ready));
 	WRITE_ONCE(sk->sk_data_ready, sk_psock_verdict_data_ready);
 	WRITE_ONCE(sk->sk_write_space, sk_psock_write_space);
 }
 
 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock)
 {
+	void (*saved_data_ready)(struct sock *sk);
+
 	psock_set_prog(&psock->progs.stream_verdict, NULL);
 	psock_set_prog(&psock->progs.skb_verdict, NULL);
 
-	if (!psock->saved_data_ready)
+	saved_data_ready = READ_ONCE(psock->saved_data_ready);
+	if (!saved_data_ready)
 		return;
 
-	WRITE_ONCE(sk->sk_data_ready, psock->saved_data_ready);
-	psock->saved_data_ready = NULL;
+	WRITE_ONCE(sk->sk_data_ready, saved_data_ready);
+	WRITE_ONCE(psock->saved_data_ready, NULL);
 }
diff --git a/net/core/sock_map.c b/net/core/sock_map.c
index 38df84284328..4dc69316ac17 100644
--- a/net/core/sock_map.c
+++ b/net/core/sock_map.c
@@ -145,6 +145,7 @@ static void sock_map_del_link(struct sock *sk,
 {
 	bool strp_stop = false, verdict_stop = false;
 	struct sk_psock_link *link, *tmp;
+	int (*update_proto)(struct sock *sk, struct sk_psock *psock, bool restore);
 
 	spin_lock_bh(&psock->link_lock);
 	list_for_each_entry_safe(link, tmp, &psock->link, list) {
@@ -152,11 +153,11 @@ static void sock_map_del_link(struct sock *sk,
 			struct bpf_map *map = link->map;
 			struct sk_psock_progs *progs = sock_map_progs(map);
 
-			if (psock->saved_data_ready && progs->stream_parser)
+			if (READ_ONCE(psock->saved_data_ready) && progs->stream_parser)
 				strp_stop = true;
-			if (psock->saved_data_ready && progs->stream_verdict)
+			if (READ_ONCE(psock->saved_data_ready) && progs->stream_verdict)
 				verdict_stop = true;
-			if (psock->saved_data_ready && progs->skb_verdict)
+			if (READ_ONCE(psock->saved_data_ready) && progs->skb_verdict)
 				verdict_stop = true;
 			list_del(&link->list);
 			sk_psock_free_link(link);
@@ -171,8 +172,9 @@ static void sock_map_del_link(struct sock *sk,
 		if (verdict_stop)
 			sk_psock_stop_verdict(sk, psock);
 
-		if (psock->psock_update_sk_prot)
-			psock->psock_update_sk_prot(sk, psock, false);
+		update_proto = READ_ONCE(psock->psock_update_sk_prot);
+		if (update_proto)
+			update_proto(sk, psock, false);
 		write_unlock_bh(&sk->sk_callback_lock);
 	}
 }
@@ -189,10 +191,13 @@ static void sock_map_unref(struct sock *sk, void *link_raw)
 
 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
 {
-	if (!sk->sk_prot->psock_update_sk_prot)
+	int (*update_proto)(struct sock *sk, struct sk_psock *psock, bool restore);
+
+	update_proto = READ_ONCE(sk->sk_prot)->psock_update_sk_prot;
+	if (!update_proto)
 		return -EINVAL;
-	psock->psock_update_sk_prot = sk->sk_prot->psock_update_sk_prot;
-	return sk->sk_prot->psock_update_sk_prot(sk, psock, false);
+	WRITE_ONCE(psock->psock_update_sk_prot, update_proto);
+	return update_proto(sk, psock, false);
 }
 
 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
@@ -202,7 +207,7 @@ static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
 	rcu_read_lock();
 	psock = sk_psock(sk);
 	if (psock) {
-		if (sk->sk_prot->close != sock_map_close) {
+		if (READ_ONCE(sk->sk_prot)->close != sock_map_close) {
 			psock = ERR_PTR(-EBUSY);
 			goto out;
 		}
@@ -303,7 +308,8 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
 	}
 
 	write_lock_bh(&sk->sk_callback_lock);
-	if (stream_parser && stream_verdict && !psock->saved_data_ready) {
+	if (stream_parser && stream_verdict &&
+	    !READ_ONCE(psock->saved_data_ready)) {
 		if (sk_is_tcp(sk))
 			ret = sk_psock_init_strp(sk, psock);
 		else
@@ -314,9 +320,11 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
 			goto out;
 		}
 		sk_psock_start_strp(sk, psock);
-	} else if (!stream_parser && stream_verdict && !psock->saved_data_ready) {
+	} else if (!stream_parser && stream_verdict &&
+		   !READ_ONCE(psock->saved_data_ready)) {
 		sk_psock_start_verdict(sk,psock);
-	} else if (!stream_verdict && skb_verdict && !psock->saved_data_ready) {
+	} else if (!stream_verdict && skb_verdict &&
+		   !READ_ONCE(psock->saved_data_ready)) {
 		sk_psock_start_verdict(sk, psock);
 	}
 	write_unlock_bh(&sk->sk_callback_lock);
@@ -536,7 +544,7 @@ static bool sock_map_redirect_allowed(const struct sock *sk)
 
 static bool sock_map_sk_is_suitable(const struct sock *sk)
 {
-	return !!sk->sk_prot->psock_update_sk_prot;
+	return !!READ_ONCE(sk->sk_prot)->psock_update_sk_prot;
 }
 
 static bool sock_map_sk_state_allowed(const struct sock *sk)

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Alexei Starovoitov]

On Wed, Oct 21, 2026 at 10:48 PM Dudu Lu <phx0fer@gmail.com> wrote:
> Annotate every one of these accesses with READ_ONCE()/WRITE_ONCE(),
> matching the existing convention for the same fields: sk->sk_data_ready

This makes KCSAN quiet and leaves the race where it is.

sock_map_link() runs under the sock lock. sock_map_delete_elem() doesn't
take it. sk_callback_lock is the only lock they share and sock_map_link()
takes it after psock_set_prog() and sock_map_init_proto().
So two tcp_bpf_update_proto(sk, psock, false) can run in parallel:
  cpu0: sock_map_link() for a map with msg_parser sees
        psock->progs.stream_verdict and picks TCP_BPF_TXRX
  cpu1: sock_map_del_link() stops the verdict and stores TCP_BPF_TX
  cpu0: stores TCP_BPF_TXRX
The socket is left with tcp_bpf_recvmsg_parser() and no verdict prog.
That's what commit c0d95d3380ee fixed.
Take the lock earlier. Then psock_update_sk_prot needs no annotation.

All accesses to saved_data_ready are under sk_callback_lock except the
three reads in sock_map_del_link(). Move that check under the lock instead
of annotating it. No need to touch skmsg.c and skmsg.h.

Only the two lockless reads of sk->sk_prot need READ_ONCE().

Something like this (compile tested only):

diff --git a/net/core/sock_map.c b/net/core/sock_map.c
index 38df842..20fb605 100644
--- a/net/core/sock_map.c
+++ b/net/core/sock_map.c
@@ -152,11 +152,11 @@ static void sock_map_del_link(struct sock *sk,
 			struct bpf_map *map = link->map;
 			struct sk_psock_progs *progs = sock_map_progs(map);

-			if (psock->saved_data_ready && progs->stream_parser)
+			if (progs->stream_parser)
 				strp_stop = true;
-			if (psock->saved_data_ready && progs->stream_verdict)
+			if (progs->stream_verdict)
 				verdict_stop = true;
-			if (psock->saved_data_ready && progs->skb_verdict)
+			if (progs->skb_verdict)
 				verdict_stop = true;
 			list_del(&link->list);
 			sk_psock_free_link(link);
@@ -164,8 +164,11 @@ static void sock_map_del_link(struct sock *sk,
 		}
 	}
 	spin_unlock_bh(&psock->link_lock);
-	if (strp_stop || verdict_stop) {
-		write_lock_bh(&sk->sk_callback_lock);
+	if (!strp_stop && !verdict_stop)
+		return;
+
+	write_lock_bh(&sk->sk_callback_lock);
+	if (psock->saved_data_ready) {
 		if (strp_stop)
 			sk_psock_stop_strp(sk, psock);
 		if (verdict_stop)
@@ -173,8 +176,8 @@ static void sock_map_del_link(struct sock *sk,

 		if (psock->psock_update_sk_prot)
 			psock->psock_update_sk_prot(sk, psock, false);
-		write_unlock_bh(&sk->sk_callback_lock);
 	}
+	write_unlock_bh(&sk->sk_callback_lock);
 }

 static void sock_map_unref(struct sock *sk, void *link_raw)
@@ -202,7 +205,7 @@ static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
 	rcu_read_lock();
 	psock = sk_psock(sk);
 	if (psock) {
-		if (sk->sk_prot->close != sock_map_close) {
+		if (READ_ONCE(sk->sk_prot)->close != sock_map_close) {
 			psock = ERR_PTR(-EBUSY);
 			goto out;
 		}
@@ -284,6 +287,7 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
 		}
 	}

+	write_lock_bh(&sk->sk_callback_lock);
 	if (msg_parser)
 		psock_set_prog(&psock->progs.msg_parser, msg_parser);
 	if (stream_parser)
@@ -298,11 +302,11 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
 	 */
 	ret = sock_map_init_proto(sk, psock);
 	if (ret < 0) {
+		write_unlock_bh(&sk->sk_callback_lock);
 		sk_psock_put(sk, psock);
 		goto out;
 	}

-	write_lock_bh(&sk->sk_callback_lock);
 	if (stream_parser && stream_verdict && !psock->saved_data_ready) {
 		if (sk_is_tcp(sk))
 			ret = sk_psock_init_strp(sk, psock);
@@ -536,7 +540,7 @@ static bool sock_map_redirect_allowed(const struct sock *sk)

 static bool sock_map_sk_is_suitable(const struct sock *sk)
 {
-	return !!sk->sk_prot->psock_update_sk_prot;
+	return !!READ_ONCE(sk->sk_prot)->psock_update_sk_prot;
 }

 static bool sock_map_sk_state_allowed(const struct sock *sk)

Pls run your reproducer on it, post the reproducer with the respin
and target bpf-next.

pw-bot: cr

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Jakub Sitnicki]

On Thu, Oct 22, 2026 at 03:40 PM GMT, Alexei Starovoitov wrote:
> On Wed, Oct 21, 2026 at 10:48 PM Dudu Lu <phx0fer@gmail.com> wrote:
>> Annotate every one of these accesses with READ_ONCE()/WRITE_ONCE(),
>> matching the existing convention for the same fields: sk->sk_data_ready
>
> This makes KCSAN quiet and leaves the race where it is.
>
> sock_map_link() runs under the sock lock. sock_map_delete_elem() doesn't
> take it. sk_callback_lock is the only lock they share and sock_map_link()
> takes it after psock_set_prog() and sock_map_init_proto().
> So two tcp_bpf_update_proto(sk, psock, false) can run in parallel:
>   cpu0: sock_map_link() for a map with msg_parser sees
>         psock->progs.stream_verdict and picks TCP_BPF_TXRX
>   cpu1: sock_map_del_link() stops the verdict and stores TCP_BPF_TX
>   cpu0: stores TCP_BPF_TXRX

Right, so:
- map_update_elem syscall vs close(fd) - OK (lock_sock protected),
- map_update_elem syscall vs map_delete_elem syscall - RACY,
- map_update_elem syscall vs TCP RST (destroy) - RACY,
- map_update_elem from BPF vs close(fd) - RACY.

> The socket is left with tcp_bpf_recvmsg_parser() and no verdict prog.
> That's what commit c0d95d3380ee fixed.
> Take the lock earlier. Then psock_update_sk_prot needs no annotation.
>
> All accesses to saved_data_ready are under sk_callback_lock except the
> three reads in sock_map_del_link(). Move that check under the lock instead
> of annotating it. No need to touch skmsg.c and skmsg.h.
>
> Only the two lockless reads of sk->sk_prot need READ_ONCE().
>
> Something like this (compile tested only):
>
> diff --git a/net/core/sock_map.c b/net/core/sock_map.c
> index 38df842..20fb605 100644
> --- a/net/core/sock_map.c
> +++ b/net/core/sock_map.c
> @@ -152,11 +152,11 @@ static void sock_map_del_link(struct sock *sk,
>  			struct bpf_map *map = link->map;
>  			struct sk_psock_progs *progs = sock_map_progs(map);
>
> -			if (psock->saved_data_ready && progs->stream_parser)
> +			if (progs->stream_parser)
>  				strp_stop = true;
> -			if (psock->saved_data_ready && progs->stream_verdict)
> +			if (progs->stream_verdict)
>  				verdict_stop = true;
> -			if (psock->saved_data_ready && progs->skb_verdict)
> +			if (progs->skb_verdict)
>  				verdict_stop = true;
>  			list_del(&link->list);
>  			sk_psock_free_link(link);
> @@ -164,8 +164,11 @@ static void sock_map_del_link(struct sock *sk,
>  		}
>  	}
>  	spin_unlock_bh(&psock->link_lock);
> -	if (strp_stop || verdict_stop) {
> -		write_lock_bh(&sk->sk_callback_lock);
> +	if (!strp_stop && !verdict_stop)
> +		return;
> +
> +	write_lock_bh(&sk->sk_callback_lock);
> +	if (psock->saved_data_ready) {
>  		if (strp_stop)
>  			sk_psock_stop_strp(sk, psock);
>  		if (verdict_stop)
> @@ -173,8 +176,8 @@ static void sock_map_del_link(struct sock *sk,
>
>  		if (psock->psock_update_sk_prot)
>  			psock->psock_update_sk_prot(sk, psock, false);
> -		write_unlock_bh(&sk->sk_callback_lock);
>  	}
> +	write_unlock_bh(&sk->sk_callback_lock);
>  }
>
>  static void sock_map_unref(struct sock *sk, void *link_raw)
> @@ -202,7 +205,7 @@ static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
>  	rcu_read_lock();
>  	psock = sk_psock(sk);
>  	if (psock) {
> -		if (sk->sk_prot->close != sock_map_close) {
> +		if (READ_ONCE(sk->sk_prot)->close != sock_map_close) {
>  			psock = ERR_PTR(-EBUSY);
>  			goto out;
>  		}
> @@ -284,6 +287,7 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
>  		}
>  	}
>
> +	write_lock_bh(&sk->sk_callback_lock);
>  	if (msg_parser)
>  		psock_set_prog(&psock->progs.msg_parser, msg_parser);
>  	if (stream_parser)
> @@ -298,11 +302,11 @@ static int sock_map_link(struct bpf_map *map, struct sock *sk)
>  	 */
>  	ret = sock_map_init_proto(sk, psock);
>  	if (ret < 0) {
> +		write_unlock_bh(&sk->sk_callback_lock);
>  		sk_psock_put(sk, psock);
>  		goto out;
>  	}
>
> -	write_lock_bh(&sk->sk_callback_lock);
>  	if (stream_parser && stream_verdict && !psock->saved_data_ready) {
>  		if (sk_is_tcp(sk))
>  			ret = sk_psock_init_strp(sk, psock);
> @@ -536,7 +540,7 @@ static bool sock_map_redirect_allowed(const struct sock *sk)
>
>  static bool sock_map_sk_is_suitable(const struct sock *sk)
>  {
> -	return !!sk->sk_prot->psock_update_sk_prot;
> +	return !!READ_ONCE(sk->sk_prot)->psock_update_sk_prot;
>  }
>
>  static bool sock_map_sk_state_allowed(const struct sock *sk)

Code LGTM

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Alexei Starovoitov]

On Fri, Oct 23, 2026 at 11:49 AM Jakub Sitnicki <jakub@cloudflare.com> wrote:
>
> >
> >  static bool sock_map_sk_state_allowed(const struct sock *sk)
>
> Code LGTM

Jakub,
Could you commandeer it ?
Dudu is clearly a bot and doesn't care.

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Jakub Sitnicki]

On Fri, Oct 23, 2026 at 11:53 AM -07, Alexei Starovoitov wrote:
> On Fri, Oct 23, 2026 at 11:49 AM Jakub Sitnicki <jakub@cloudflare.com> wrote:
>>
>> >
>> >  static bool sock_map_sk_state_allowed(const struct sock *sk)
>>
>> Code LGTM
>
> Jakub,
> Could you commandeer it ?
> Dudu is clearly a bot and doesn't care.

Ack. Will drive this.

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[phx]

Hi Jakub, Alexei,

Sorry for the long silence — my laptop died right after v2 and took the
local mail and the working tree with it. Agreed that the ordering fix is
the right one; please go ahead with the commandeer, Jakub.

I do still have the reproducer (a KCSAN selftest for this race, verified
in QEMU). Say the word and I will post it, or send it to you to fold into
your respin.

Thanks,
Dudu


On Mon, Oct 26, 2026 at 8:33 PM Jakub Sitnicki <jakub@cloudflare.com> wrote:
>
> On Fri, Oct 23, 2026 at 11:53 AM -07, Alexei Starovoitov wrote:
> > On Fri, Oct 23, 2026 at 11:49 AM Jakub Sitnicki <jakub@cloudflare.com> wrote:
> >>
> >> >
> >> >  static bool sock_map_sk_state_allowed(const struct sock *sk)
> >>
> >> Code LGTM
> >
> > Jakub,
> > Could you commandeer it ?
> > Dudu is clearly a bot and doesn't care.
>
> Ack. Will drive this.

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Jakub Sitnicki]

On Tue, Oct 27, 2026 at 01:10 PM +08, phx wrote:
> Sorry for the long silence — my laptop died right after v2 and took the
> local mail and the working tree with it. Agreed that the ordering fix is
> the right one; please go ahead with the commandeer, Jakub.
>
> I do still have the reproducer (a KCSAN selftest for this race, verified
> in QEMU). Say the word and I will post it, or send it to you to fold into
> your respin.

Feel free to iterate on the patch once you get your dev env restored.

We would also need selftest coverage for the race, if it reproduces
relatively fast. You can take a look at test_race_insert_listen for an
example.

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Alexei Starovoitov]

On Tue Oct 27, 2026 at 4:33 PM UTC, Jakub Sitnicki wrote:
> On Tue, Oct 27, 2026 at 01:10 PM +08, phx wrote:
>> Sorry for the long silence — my laptop died right after v2 and took the
>> local mail and the working tree with it. Agreed that the ordering fix is
>> the right one; please go ahead with the commandeer, Jakub.
>>
>> I do still have the reproducer (a KCSAN selftest for this race, verified
>> in QEMU). Say the word and I will post it, or send it to you to fold into
>> your respin.
>
> Feel free to iterate on the patch once you get your dev env restored.

I very much prefer if you Jakub will take it over.
'phx' is a clueless bot that sends spam across all subsystem.

Re: [PATCH bpf v2] bpf, sockmap: annotate the psock teardown data races

[Jakub Sitnicki]

On Tue, Oct 27, 2026 at 04:46 PM GMT, Alexei Starovoitov wrote:
> On Tue Oct 27, 2026 at 4:33 PM UTC, Jakub Sitnicki wrote:
>> On Tue, Oct 27, 2026 at 01:10 PM +08, phx wrote:
>>> Sorry for the long silence — my laptop died right after v2 and took the
>>> local mail and the working tree with it. Agreed that the ordering fix is
>>> the right one; please go ahead with the commandeer, Jakub.
>>>
>>> I do still have the reproducer (a KCSAN selftest for this race, verified
>>> in QEMU). Say the word and I will post it, or send it to you to fold into
>>> your respin.
>>
>> Feel free to iterate on the patch once you get your dev env restored.
>
> I very much prefer if you Jakub will take it over.
> 'phx' is a clueless bot that sends spam across all subsystem.

Got me fooled. Thanks. I will respin it myself.