AZL-99633

In the Linux kernel, the following vulnerability has been resolved:

Bluetooth: RFCOMM: take rfcomm_mutex for the deferred setup accept

rfcommsockrecvmsg() completes a deferred setup by calling rfcommdlcaccept() without holding any RFCOMM lock:

if (test_and_clear_bit(RFCOMM_DEFER_SETUP, &d->flags)) {
    rfcomm_dlc_accept(d);
    return 0;
}

and rfcommdlcaccept() dereferences the session on its first line:

struct sock *sk = d->session->sock->sk;

Every other path that touches d->session runs under rfcommmutex: rfcommdlcopen(), rfcommdlcclose(), rfcommdlcexists(), rfcommdlcsendrpn(), and the RFCOMM thread through rfcommprocesssessions(). rfcommconnectind() is even documented as "called under rfcomm_lock()". This call site is the only one that skips it.

The RFCOMMDEFERSETUP bit looks like it serialises the accept against teardown, since __rfcommdlcclose() returns early when it wins the testandclear. But rfcommrecvdisc() forces the state first:

d->state = BT_CLOSED;
__rfcomm_dlc_close(d, err);

and the early return only covers BTCONNECT, BTCONFIG, BTOPEN and BTCONNECT2. With the state already BT_CLOSED that switch does not match, the bit is never consulted, and __rfcommdlcclose() falls through to rfcommdlcunlink(), which sets d->session = NULL.

So a remote DISC on a deferred dlc clears the session while leaving RFCOMMDEFERSETUP set. The next recvmsg() then passes the testandclear and dereferences a NULL session. No timing window is needed: once the DISC has been processed, the dereference is unconditional.

Give rfcommdlcaccept() the same shape as rfcommdlcopen() and rfcommdlcclose(): an exported wrapper that takes rfcomm_mutex and re-checks the session, around a __rfcommdlcaccept() that the two in-core callers, which already hold the mutex, keep using.

Reproduced on a KASAN + PROVELOCKING kernel with a BR/EDR peer emulated over /dev/vhci: the peer brings up an ACL link, opens L2CAP on the RFCOMM PSM, starts a session, opens a dlc on a channel bound with BTDEFER_SETUP, and sends DISC after the socket is accepted. recv() on the accepted socket then hits:

Oops: general protection fault KASAN: null-ptr-deref in range [0x0000000000000010-0x0000000000000017] RIP: 0010:rfcommdlcaccept+0x54/0x350 Call Trace: rfcommsockrecvmsg+0x1cd/0x230 sock_recvmsg+0x166/0x1c0 _sysrecvfrom+0x20d/0x300

0x10 is the offset of sock in struct rfcommsession. With this patch the same run completes with recv() returning 0 and no report, and lockdep stays quiet, confirming rfcommmutex is still taken before lock_sock on this path as it is on the thread side.