Bluetooth: avoid circular locks in sco_sock_connect
authorDesmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
Tue, 10 Aug 2021 04:14:06 +0000 (12:14 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 18 Sep 2021 11:40:27 +0000 (13:40 +0200)
commit98d44b7be6f1bcfd4f824c5f8bc2b742f890879f
tree5795996d5594d7f420ee99301feb06227d785c91
parenta1073aad497d0d071a71f61b721966a176d50c08
Bluetooth: avoid circular locks in sco_sock_connect

[ Upstream commit 734bc5ff783115aa3164f4e9dd5967ae78e0a8ab ]

In a future patch, calls to bh_lock_sock in sco.c should be replaced
by lock_sock now that none of the functions are run in IRQ context.

However, doing so results in a circular locking dependency:

======================================================
WARNING: possible circular locking dependency detected
5.14.0-rc4-syzkaller #0 Not tainted
------------------------------------------------------
syz-executor.2/14867 is trying to acquire lock:
ffff88803e3c1120 (sk_lock-AF_BLUETOOTH-BTPROTO_SCO){+.+.}-{0:0}, at:
lock_sock include/net/sock.h:1613 [inline]
ffff88803e3c1120 (sk_lock-AF_BLUETOOTH-BTPROTO_SCO){+.+.}-{0:0}, at:
sco_conn_del+0x12a/0x2a0 net/bluetooth/sco.c:191

but task is already holding lock:
ffffffff8d2dc7c8 (hci_cb_list_lock){+.+.}-{3:3}, at:
hci_disconn_cfm include/net/bluetooth/hci_core.h:1497 [inline]
ffffffff8d2dc7c8 (hci_cb_list_lock){+.+.}-{3:3}, at:
hci_conn_hash_flush+0xda/0x260 net/bluetooth/hci_conn.c:1608

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (hci_cb_list_lock){+.+.}-{3:3}:
       __mutex_lock_common kernel/locking/mutex.c:959 [inline]
       __mutex_lock+0x12a/0x10a0 kernel/locking/mutex.c:1104
       hci_connect_cfm include/net/bluetooth/hci_core.h:1482 [inline]
       hci_remote_features_evt net/bluetooth/hci_event.c:3263 [inline]
       hci_event_packet+0x2f4d/0x7c50 net/bluetooth/hci_event.c:6240
       hci_rx_work+0x4f8/0xd30 net/bluetooth/hci_core.c:5122
       process_one_work+0x98d/0x1630 kernel/workqueue.c:2276
       worker_thread+0x658/0x11f0 kernel/workqueue.c:2422
       kthread+0x3e5/0x4d0 kernel/kthread.c:319
       ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295

-> #1 (&hdev->lock){+.+.}-{3:3}:
       __mutex_lock_common kernel/locking/mutex.c:959 [inline]
       __mutex_lock+0x12a/0x10a0 kernel/locking/mutex.c:1104
       sco_connect net/bluetooth/sco.c:245 [inline]
       sco_sock_connect+0x227/0xa10 net/bluetooth/sco.c:601
       __sys_connect_file+0x155/0x1a0 net/socket.c:1879
       __sys_connect+0x161/0x190 net/socket.c:1896
       __do_sys_connect net/socket.c:1906 [inline]
       __se_sys_connect net/socket.c:1903 [inline]
       __x64_sys_connect+0x6f/0xb0 net/socket.c:1903
       do_syscall_x64 arch/x86/entry/common.c:50 [inline]
       do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
       entry_SYSCALL_64_after_hwframe+0x44/0xae

-> #0 (sk_lock-AF_BLUETOOTH-BTPROTO_SCO){+.+.}-{0:0}:
       check_prev_add kernel/locking/lockdep.c:3051 [inline]
       check_prevs_add kernel/locking/lockdep.c:3174 [inline]
       validate_chain kernel/locking/lockdep.c:3789 [inline]
       __lock_acquire+0x2a07/0x54a0 kernel/locking/lockdep.c:5015
       lock_acquire kernel/locking/lockdep.c:5625 [inline]
       lock_acquire+0x1ab/0x510 kernel/locking/lockdep.c:5590
       lock_sock_nested+0xca/0x120 net/core/sock.c:3170
       lock_sock include/net/sock.h:1613 [inline]
       sco_conn_del+0x12a/0x2a0 net/bluetooth/sco.c:191
       sco_disconn_cfm+0x71/0xb0 net/bluetooth/sco.c:1202
       hci_disconn_cfm include/net/bluetooth/hci_core.h:1500 [inline]
       hci_conn_hash_flush+0x127/0x260 net/bluetooth/hci_conn.c:1608
       hci_dev_do_close+0x528/0x1130 net/bluetooth/hci_core.c:1778
       hci_unregister_dev+0x1c0/0x5a0 net/bluetooth/hci_core.c:4015
       vhci_release+0x70/0xe0 drivers/bluetooth/hci_vhci.c:340
       __fput+0x288/0x920 fs/file_table.c:280
       task_work_run+0xdd/0x1a0 kernel/task_work.c:164
       exit_task_work include/linux/task_work.h:32 [inline]
       do_exit+0xbd4/0x2a60 kernel/exit.c:825
       do_group_exit+0x125/0x310 kernel/exit.c:922
       get_signal+0x47f/0x2160 kernel/signal.c:2808
       arch_do_signal_or_restart+0x2a9/0x1c40 arch/x86/kernel/signal.c:865
       handle_signal_work kernel/entry/common.c:148 [inline]
       exit_to_user_mode_loop kernel/entry/common.c:172 [inline]
       exit_to_user_mode_prepare+0x17d/0x290 kernel/entry/common.c:209
       __syscall_exit_to_user_mode_work kernel/entry/common.c:291 [inline]
       syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:302
       ret_from_fork+0x15/0x30 arch/x86/entry/entry_64.S:288

other info that might help us debug this:

Chain exists of:
  sk_lock-AF_BLUETOOTH-BTPROTO_SCO --> &hdev->lock --> hci_cb_list_lock

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(hci_cb_list_lock);
                               lock(&hdev->lock);
                               lock(hci_cb_list_lock);
  lock(sk_lock-AF_BLUETOOTH-BTPROTO_SCO);

 *** DEADLOCK ***

The issue is that the lock hierarchy should go from &hdev->lock -->
hci_cb_list_lock --> sk_lock-AF_BLUETOOTH-BTPROTO_SCO. For example,
one such call trace is:

  hci_dev_do_close():
    hci_dev_lock();
    hci_conn_hash_flush():
      hci_disconn_cfm():
        mutex_lock(&hci_cb_list_lock);
        sco_disconn_cfm():
        sco_conn_del():
          lock_sock(sk);

However, in sco_sock_connect, we call lock_sock before calling
hci_dev_lock inside sco_connect, thus inverting the lock hierarchy.

We fix this by pulling the call to hci_dev_lock out from sco_connect.

Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
net/bluetooth/sco.c