Skip to contentSkip to conversation

af_unix: mark MSG_SPLICE_PAGES frags shared

PATCH In review Yiming Qian 2 participants 4 messages Jun 30 – Jul 2

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 af_unix: mark MSG_SPLICE_PAGES frags shared

unix_stream_sendmsg() splices external pages directly into skb frags when MSG_SPLICE_PAGES is set, but it does not propagate SKBFL_SHARED_FRAG afterward.

That leaves later writers without the shared-frag marker even though the skb still references externally owned pages.

Set the marker after a successful skb_splice_from_iter() call.

Fixes a0dbf5f818f90 ("af_unix: Support MSG_SPLICE_PAGES")
---
 net/unix/af_unix.c | 1 +
 1 file changed, 1 insertion(+)
net/unix/af_unix.c +1 -0
index f7a9d55eee8a1..f2cd0f8ec0914 100644
@@ -2458,6 +2458,7 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg,
 				goto out_free;
 
 			size = err;
+			skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
 			refcount_add(size, &sk->sk_wmem_alloc);
 		} else {
 			skb_put(skb, size - data_len);
Signature · 1 line
2.34.1
#1 lore raw
Quoted 4 lines · Yiming Qian

unix_stream_sendmsg() splices external pages directly into skb frags when MSG_SPLICE_PAGES is set, but it does not propagate SKBFL_SHARED_FRAG afterward.

I think it doesn't matter with the plain AF_UNIX.

Please elaborate on the scenario where this could be a problem. e.g. sockmap ?

Quoted 9 lines

That leaves later writers without the shared-frag marker even though the skb still references externally owned pages.

Set the marker after a successful skb_splice_from_iter() call.

Fixes a0dbf5f818f90 ("af_unix: Support MSG_SPLICE_PAGES")

Reported-by is not needed when it's identical to SOB tag.

Quoted 20 lines
---
 net/unix/af_unix.c | 1 +
 1 file changed, 1 insertion(+)
net/unix/af_unix.c +1 -1
index f7a9d55eee8a1..f2cd0f8ec0914 100644
@@ -2458,6 +2458,7 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg,
                                goto out_free;
 
                        size = err;
+                       skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
                        refcount_add(size, &sk->sk_wmem_alloc);
                } else {
                        skb_put(skb, size - data_len);
--

2.34.1

author
#2 lore raw

Thanks, I agree plain AF_UNIX does not by itself provide the writer side.

The concern was that MSG_SPLICE_PAGES imports externally owned pages into skb frags, while unlike TCP/UDP/KCM this path does not mark them with SKBFL_SHARED_FRAG. I checked the AF_UNIX sockmap path as well, but I do not currently have a concrete in-tree chain where those frags reach a writer that skips COW based on skb_has_shared_frag().

So this should be treated as a defensive consistency cleanup rather than a security fix. I will drop the Fixes tag and the duplicate Reported-by tags in v2.

Quoted 46 lines · Kuniyuki Iwashima <[email protected]> 于2026年6月30日周二 23:51写道:
Quoted 4 lines · Yiming Qian

unix_stream_sendmsg() splices external pages directly into skb frags when MSG_SPLICE_PAGES is set, but it does not propagate SKBFL_SHARED_FRAG afterward.

I think it doesn't matter with the plain AF_UNIX.

Please elaborate on the scenario where this could be a problem. e.g. sockmap ?

Quoted 9 lines

That leaves later writers without the shared-frag marker even though the skb still references externally owned pages.

Set the marker after a successful skb_splice_from_iter() call.

Fixes a0dbf5f818f90 ("af_unix: Support MSG_SPLICE_PAGES")

Reported-by is not needed when it's identical to SOB tag.

Quoted 20 lines
---
 net/unix/af_unix.c | 1 +
 1 file changed, 1 insertion(+)
net/unix/af_unix.c +1 -1
index f7a9d55eee8a1..f2cd0f8ec0914 100644
@@ -2458,6 +2458,7 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg,
                                goto out_free;
 
                        size = err;
+                       skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
                        refcount_add(size, &sk->sk_wmem_alloc);
                } else {
                        skb_put(skb, size - data_len);
--

2.34.1

#3 lore raw
Quoted 11 lines · 钱一铭

Thanks, I agree plain AF_UNIX does not by itself provide the writer side.

The concern was that MSG_SPLICE_PAGES imports externally owned pages into skb frags, while unlike TCP/UDP/KCM this path does not mark them with SKBFL_SHARED_FRAG. I checked the AF_UNIX sockmap path as well, but I do not currently have a concrete in-tree chain where those frags reach a writer that skips COW based on skb_has_shared_frag().

So this should be treated as a defensive consistency cleanup rather than a security fix.

Hmm, if it's not exploitable, let's not add that. It's rather confusing to future readers.

Quoted 50 lines

I will drop the Fixes tag and the duplicate Reported-by
tags in v2.

Quoted 46 lines · Kuniyuki Iwashima <[email protected]> 于2026年6月30日周二 23:51写道:
Quoted 4 lines · Yiming Qian

unix_stream_sendmsg() splices external pages directly into skb frags when MSG_SPLICE_PAGES is set, but it does not propagate SKBFL_SHARED_FRAG afterward.

I think it doesn't matter with the plain AF_UNIX.

Please elaborate on the scenario where this could be a problem. e.g. sockmap ?

Quoted 9 lines

That leaves later writers without the shared-frag marker even though the skb still references externally owned pages.

Set the marker after a successful skb_splice_from_iter() call.

Fixes a0dbf5f818f90 ("af_unix: Support MSG_SPLICE_PAGES")

Reported-by is not needed when it's identical to SOB tag.

Quoted 20 lines
---
 net/unix/af_unix.c | 1 +
 1 file changed, 1 insertion(+)
net/unix/af_unix.c +1 -1
index f7a9d55eee8a1..f2cd0f8ec0914 100644
@@ -2458,6 +2458,7 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg,
                                goto out_free;
 
                        size = err;
+                       skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
                        refcount_add(size, &sk->sk_wmem_alloc);
                } else {
                        skb_put(skb, size - data_len);
--

2.34.1