On Wed, Feb 4, 2026 at 7:41 AM Michal Luczaj [off-list ref] wrote:

On 2/4/26 08:58, Kuniyuki Iwashima wrote:

quoted

On Tue, Feb 3, 2026 at 11:15 PM Martin KaFai Lau [off-list ref] wrote:

quoted

On 2/3/26 11:47 AM, Kuniyuki Iwashima wrote:

quoted

From: Michal Luczaj <redacted>
Date: Tue, 3 Feb 2026 10:57:46 +0100

quoted

On 2/3/26 04:53, Martin KaFai Lau wrote:

quoted

On 2/2/26 7:10 AM, Michal Luczaj wrote:

quoted

In related news, looks like bpf_iter_unix_seq_show() is missing
unix_state_lock(): lock_sock_fast() won't stop unix_release_sock(). E.g.
bpf iterator can grab unix_sock::peer as it is being released.

If the concern is the bpf iterator prog may use a released unix_peer(sk)
pointer, it should be fine. The unix_peer(sk) pointer is not a trusted
pointer to the bpf prog, so nothing bad will happen other than
potentially reading incorrect values.

But if the prog passes a released peer pointer to a bpf helper:

BUG: KASAN: slab-use-after-free in bpf_skc_to_unix_sock+0x95/0xb0
Read of size 1 at addr ffff888110654c92 by task test_progs/1936

hmm... bpf_skc_to_unix_sock is exposed to tracing. bpf_iter is a tracing
bpf prog.

quoted

Can you cook a patch for this ? probably like below

This can help the bpf_iter but not the other tracing prog such as fentry.

Oh well ... then bpf_skc_to_unix_sock() can be used even
with SEQ_START_TOKEN at fentry of bpf_iter_unix_seq_show() ??

How about adding notrace to all af_unix bpf iterator functions ?

The procfs iterator holds a spinlock of the hashtable from
->start/next() to ->stop() to prevent the race with unix_release_sock().

I think other (non-iterator) functions cannot do such racy
access with tracing prog.

But then there's SOCK_DGRAM where you can drop unix_peer(sk) without
releasing sk; see AF_UNSPEC in unix_dgram_connect(). I think Martin is
right, we can crash at many fentries.

Ah I was only considering SOCK_STREAM :p

Only solution that came to mind is breaking change enforcing
a release function to bpf_skc_to_unix_sock() so that it can
hold the peer's refcnt and save the pointer somewhere else
until the release function, but I think this is unacceptable.

Also, I guess this type of issue could be triggered with any
objects that are not refcounted by bpf tracing prog ?

For example, inet_sock(sk)->inet_opt could be freed by
setsockopt(IP_OPTIONS) even after fentry prog verifies
that it's not NULL.

I'm not sure if bpf_core_cast() etc allows such access, but
if it's allowed, I think there is no general solution.

Fortunately that's not null-deref nor oob-write, and it just reads
stale info as Martin mentioned... so probably this is WAI for
tracing prog ?

BUG: KASAN: slab-use-after-free in bpf_skc_to_unix_sock+0xa4/0xb0
Read of size 2 at addr ffff888147d38890 by task test_progs/2495
Call Trace:
 dump_stack_lvl+0x5d/0x80
 print_report+0x170/0x4f3
 kasan_report+0xe1/0x180
 bpf_skc_to_unix_sock+0xa4/0xb0
 bpf_prog_564a1c39c35d86a2_unix_shutdown_entry+0x8a/0x8e
 bpf_trampoline_6442564662+0x47/0xab
 unix_shutdown+0x9/0x880
 __sys_shutdown+0xe1/0x160
 __x64_sys_shutdown+0x52/0x90
 do_syscall_64+0x6b/0x3a0
 entry_SYSCALL_64_after_hwframe+0x76/0x7e

quoted

quoted

quoted

---8<---

diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 02ebad6afac7..9c7e9fbde362 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c

@@ -3740,8 +3740,9 @@ static int bpf_iter_unix_seq_show(struct seq_file *seq, void *v)
              return 0;

      slow = lock_sock_fast(sk);
+     unix_state_lock(sk);

-     if (unlikely(sk_unhashed(sk))) {
+     if (unlikely(sock_flag(other, SOCK_DEAD))) {
              ret = SEQ_SKIP;
              goto unlock;
      }

@@ -3751,6 +3752,7 @@ static int bpf_iter_unix_seq_show(struct seq_file *seq, void *v)
      prog = bpf_iter_get_info(&meta, false);
      ret = unix_prog_seq_show(prog, &meta, v, uid);
  unlock:
+     unix_staet_unlock(sk);
      unlock_sock_fast(sk, slow);
      return ret;
  }
---8<---

Thanks!