]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
Bluetooth: fix UAF in l2cap_sock_cleanup_listen() vs l2cap_conn_del()
authorSafa Karakuş <safa.karakus@secunnix.com>
Sat, 16 May 2026 18:15:04 +0000 (21:15 +0300)
committerLuiz Augusto von Dentz <luiz.von.dentz@intel.com>
Wed, 20 May 2026 20:35:47 +0000 (16:35 -0400)
bt_accept_dequeue() unlinks a not-yet-accepted child from the parent
accept queue and release_sock()s it before returning, so the returned
sk has no caller reference and is unlocked.

l2cap_sock_cleanup_listen() walks these children on listening-socket
close.  A concurrent HCI disconnect drives hci_rx_work ->
l2cap_conn_del() which runs l2cap_chan_del() + l2cap_sock_kill() and
frees the child sk and its l2cap_chan; cleanup_listen() then uses both:

  BUG: KASAN: slab-use-after-free in l2cap_sock_kill
    l2cap_sock_kill / l2cap_sock_cleanup_listen / __x64_sys_close
  Freed by: l2cap_conn_del -> l2cap_sock_close_cb -> l2cap_sock_kill

This is distinct from the two fixes already in this area: commit
e83f5e24da741 ("Bluetooth: serialize accept_q access") serialises the
accept_q list/poll and takes temporary refs inside bt_accept_dequeue(),
and CVE-2025-39860 serialises the userspace close()/accept() race by
calling cleanup_listen() under lock_sock() in l2cap_sock_release().
Neither covers l2cap_conn_del() running from hci_rx_work, so this UAF
still reproduces on current bluetooth/master.

Take the reference at the source: bt_accept_dequeue() does sock_hold()
while sk is still locked, before release_sock(); callers sock_put().
cleanup_listen() pins the chan with l2cap_chan_hold_unless_zero() under
a brief child sk lock (serialising vs l2cap_sock_teardown_cb()), drops
it before l2cap_chan_lock(), and skips a duplicate l2cap_sock_kill() on
SOCK_DEAD.  conn->lock is not taken here: cleanup_listen() runs under
the parent sk lock and that would invert
conn->lock -> chan->lock -> sk_lock (lockdep).

KASAN/SMP: an unprivileged listen/close vs HCI-disconnect race produced
12 use-after-free reports per run before this change; 0, and no lockdep
report, over 1600+ raced iterations after it on bluetooth/master.

Fixes: 15f02b910562 ("Bluetooth: L2CAP: Add initial code for Enhanced Credit Based Mode")
Cc: stable@vger.kernel.org
Reported-by: Siwei Zhang <oss@fourdim.xyz>
Reviewed-by: Siwei Zhang <oss@fourdim.xyz>
Signed-off-by: Safa Karakuş <safa.karakus@secunnix.com>
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>
net/bluetooth/af_bluetooth.c
net/bluetooth/iso.c
net/bluetooth/l2cap_sock.c
net/bluetooth/rfcomm/sock.c
net/bluetooth/sco.c

index 9d68dd86023c9d1d19bc27644d0d4e8dfdcac536..1a6aa3f8d4d610f7c0ba47bf563e422326a822a0 100644 (file)
@@ -340,6 +340,16 @@ restart:
                        if (newsock)
                                sock_graft(sk, newsock);
 
+                       /* Hand the caller a reference taken while sk is
+                        * still locked.  bt_accept_unlink() just dropped
+                        * the accept-queue reference; without this hold a
+                        * concurrent teardown (e.g. l2cap_conn_del() ->
+                        * l2cap_sock_kill()) could free sk between
+                        * release_sock() and the caller using it.  Every
+                        * caller drops this with sock_put() when done.
+                        */
+                       sock_hold(sk);
+
                        release_sock(sk);
                        if (next)
                                sock_put(next);
index b971281f0a2b9bca27dc90a091e83526db5b461e..d7af617cda45e4581e11de60ba398712692e8316 100644 (file)
@@ -751,6 +751,8 @@ static void iso_sock_cleanup_listen(struct sock *parent)
        while ((sk = bt_accept_dequeue(parent, NULL))) {
                iso_sock_close(sk);
                iso_sock_kill(sk);
+               /* Drop the reference handed back by bt_accept_dequeue(). */
+               sock_put(sk);
        }
 
        /* If listening socket has a hcon, properly disconnect it */
@@ -1356,8 +1358,13 @@ static int iso_sock_accept(struct socket *sock, struct socket *newsock,
                }
 
                ch = bt_accept_dequeue(sk, newsock);
-               if (ch)
+               if (ch) {
+                       /* Drop the bridging ref from bt_accept_dequeue();
+                        * the grafted socket keeps ch alive from here.
+                        */
+                       sock_put(ch);
                        break;
+               }
 
                if (!timeo) {
                        err = -EAGAIN;
index cf590a67d3641c77f0eb25118ddb9593692fb6f6..b34e7da8d90672150859ea2035f47fd26d98a329 100644 (file)
@@ -349,8 +349,13 @@ static int l2cap_sock_accept(struct socket *sock, struct socket *newsock,
                }
 
                nsk = bt_accept_dequeue(sk, newsock);
-               if (nsk)
+               if (nsk) {
+                       /* Drop the bridging ref from bt_accept_dequeue();
+                        * the grafted socket keeps nsk alive from here.
+                        */
+                       sock_put(nsk);
                        break;
+               }
 
                if (!timeo) {
                        err = -EAGAIN;
@@ -1475,22 +1480,54 @@ static void l2cap_sock_cleanup_listen(struct sock *parent)
        BT_DBG("parent %p state %s", parent,
               state_to_string(parent->sk_state));
 
-       /* Close not yet accepted channels */
+       /* Close not yet accepted channels.
+        *
+        * bt_accept_dequeue() now returns sk with an extra reference held
+        * (taken while sk was still locked) so a concurrent l2cap_conn_del()
+        * -> l2cap_sock_kill() cannot free sk under us.
+        *
+        * cleanup_listen() runs under the parent sk lock, so unlike
+        * l2cap_sock_shutdown() we must NOT take conn->lock here: that would
+        * establish sk_lock -> conn->lock and invert the established
+        * conn->lock -> chan->lock -> sk_lock order (lockdep deadlock).
+        *
+        * Instead, briefly take the child sk lock to fetch and pin its chan.
+        * l2cap_conn_del() reaches the chan free only via
+        * l2cap_chan_del() -> l2cap_sock_teardown_cb(), which itself takes
+        * the child sk lock; holding it across l2cap_chan_hold_unless_zero()
+        * therefore guarantees the chan cannot be freed while we read and
+        * pin it (hold_unless_zero() additionally skips a chan already past
+        * its last reference).  We then drop the sk lock before taking
+        * chan->lock, so sk and chan locks are never held together.
+        */
        while ((sk = bt_accept_dequeue(parent, NULL))) {
-               struct l2cap_chan *chan = l2cap_pi(sk)->chan;
+               struct l2cap_chan *chan;
+
+               lock_sock_nested(sk, L2CAP_NESTING_NORMAL);
+               chan = l2cap_chan_hold_unless_zero(l2cap_pi(sk)->chan);
+               release_sock(sk);
+               if (!chan) {
+                       /* l2cap_conn_del() already tearing this child down */
+                       sock_put(sk);
+                       continue;
+               }
 
                BT_DBG("child chan %p state %s", chan,
                       state_to_string(chan->state));
 
-               l2cap_chan_hold(chan);
                l2cap_chan_lock(chan);
-
                __clear_chan_timer(chan);
                l2cap_chan_close(chan, ECONNRESET);
-               l2cap_sock_kill(sk);
-
+               /* l2cap_conn_del() may already have killed this socket
+                * (it sets SOCK_DEAD); skip the duplicate to avoid a
+                * double sock_put()/l2cap_chan_put().
+                */
+               if (!sock_flag(sk, SOCK_DEAD))
+                       l2cap_sock_kill(sk);
                l2cap_chan_unlock(chan);
+
                l2cap_chan_put(chan);
+               sock_put(sk);
        }
 }
 
index be6639cd6f59053fc1bc55c88a62dc8ac3d28fe8..bd7d959c6e9eb883358c49f48c0a89116304dcea 100644 (file)
@@ -180,6 +180,8 @@ static void rfcomm_sock_cleanup_listen(struct sock *parent)
        while ((sk = bt_accept_dequeue(parent, NULL))) {
                rfcomm_sock_close(sk);
                rfcomm_sock_kill(sk);
+               /* Drop the reference handed back by bt_accept_dequeue(). */
+               sock_put(sk);
        }
 
        parent->sk_state  = BT_CLOSED;
@@ -497,8 +499,13 @@ static int rfcomm_sock_accept(struct socket *sock, struct socket *newsock,
                }
 
                nsk = bt_accept_dequeue(sk, newsock);
-               if (nsk)
+               if (nsk) {
+                       /* Drop the bridging ref from bt_accept_dequeue();
+                        * the grafted socket keeps nsk alive from here.
+                        */
+                       sock_put(nsk);
                        break;
+               }
 
                if (!timeo) {
                        err = -EAGAIN;
index eba44525d41d9880fb76773540dd1a990f3f167f..f1799c6a6f87510060d38f6f1cefa955304df94d 100644 (file)
@@ -502,6 +502,8 @@ static void sco_sock_cleanup_listen(struct sock *parent)
        while ((sk = bt_accept_dequeue(parent, NULL))) {
                sco_sock_close(sk);
                sco_sock_kill(sk);
+               /* Drop the reference handed back by bt_accept_dequeue(). */
+               sock_put(sk);
        }
 
        parent->sk_state  = BT_CLOSED;
@@ -765,8 +767,13 @@ static int sco_sock_accept(struct socket *sock, struct socket *newsock,
                }
 
                ch = bt_accept_dequeue(sk, newsock);
-               if (ch)
+               if (ch) {
+                       /* Drop the bridging ref from bt_accept_dequeue();
+                        * the grafted socket keeps ch alive from here.
+                        */
+                       sock_put(ch);
                        break;
+               }
 
                if (!timeo) {
                        err = -EAGAIN;