On 6/20/26 8:32 AM, Jose Fernandez (Anthropic) wrote:

reqsk_queue_hash_req() publishes a TCP_NEW_SYN_RECV request_sock onto
the ehash chain (via inet_ehash_insert(), which drops the bucket lock on
return) and only afterwards refcount_set()s rsk_refcnt to 3.

Lockless readers such as __inet_lookup_established() account for this by
using refcount_inc_not_zero(), but bpf_iter_tcp_established_batch() uses
plain sock_hold() while holding the bucket lock, on the assumption that
the lock guarantees sk_refcnt > 0. That assumption does not hold for
request_sock:

   CPU 0                                CPU 1
   -----                                -----
   tcp_conn_request()
    reqsk_queue_hash_req()
     inet_ehash_insert(req)
      spin_lock(bucket)
      __sk_nulls_add_node_rcu(req)      // rsk_refcnt == 0
      spin_unlock(bucket)
                                        bpf_iter_tcp_established_batch()
                                         spin_lock(bucket)
                                         sock_hold(req)   <-- addition on 0
                                         spin_unlock(bucket)
     refcount_set(&req->rsk_refcnt, 3)  // clobbers saturated value

which surfaces as:

   refcount_t: addition on 0; use-after-free.
   WARNING: lib/refcount.c:25 at refcount_warn_saturate+0x48/0x90, CPU#1
   Call Trace:
    bpf_iter_tcp_established_batch+0x14e/0x170
    bpf_iter_tcp_batch+0x53/0x200
    bpf_iter_tcp_seq_next+0x27/0x70
    bpf_seq_read+0x107/0x410
    vfs_read+0xb9/0x380

refcount_warn_saturate() then saturates the count, the publishing CPU's
refcount_set() clobbers it, and the socket is left one reference short.
When the last legitimate owner drops its reference the reqsk is freed
while still reachable, leading to use-after-free panics in e.g.
inet_csk_accept() or inet_csk_listen_stop().

This reproduces in seconds with tcp_syncookies=0, a handful of threads
doing connect()/close() to a local listener while others read an
iter/tcp link in a tight loop.

Use refcount_inc_not_zero() and skip the socket on failure, the same way
every other ehash walker does. The listening hash is unaffected as
listeners are always inserted into lhash2 with sk_refcnt >= 1, so
bpf_iter_tcp_listening_batch() is left as-is.

If every matching socket in a bucket is mid-init, end_sk can stay at 0;
advance to the next bucket in that case rather than terminating the
whole iteration on a stale batch[0].

Fixes: 04c7820b776f ("bpf: tcp: Bpf iter batching and lock_sock")
Reviewed-by: Ben Cressey <redacted>
Assisted-by: Claude:unspecified
Signed-off-by: Jose Fernandez (Anthropic) <redacted>

LGTM.

Reviewed-by: Jiayuan Chen <jiayuan.chen@linux.dev>

Folding the open-coded first *start_sk into the loop is a good
cleanup — it was the one socket that bypassed the refcnt check.


The double-put on the realloc-failure path reported by ai is a separate,
pre-existing issue and can be addressed in a follow-up.