Re: Lockdep report for hci_conn_get_phy()

[Jiri Kosina]

On Thu, 4 Mar 2021, Jiri Kosina wrote:

>  ======================================================
>  WARNING: possible circular locking dependency detected
>  5.12.0-rc1-00026-g73d464503354 #10 Not tainted
>  ------------------------------------------------------
>  bluetoothd/1118 is trying to acquire lock:
>  ffff8f078383c078 (&hdev->lock){+.+.}-{3:3}, at: hci_conn_get_phy+0x1c/0x150 [bluetooth]
>  
>  but task is already holding lock:
>  ffff8f07e831d920 (sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP){+.+.}-{0:0}, at: l2cap_sock_getsockopt+0x8b/0x610 
> 
>  
>  which lock already depends on the new lock.
> 
>  
>  the existing dependency chain (in reverse order) is:
>  
>  -> #3 (sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP){+.+.}-{0:0}:
>         lock_sock_nested+0x72/0xa0
>         l2cap_sock_ready_cb+0x18/0x70 [bluetooth]
>         l2cap_config_rsp+0x27a/0x520 [bluetooth]
>         l2cap_sig_channel+0x658/0x1330 [bluetooth]
>         l2cap_recv_frame+0x1ba/0x310 [bluetooth]
>         hci_rx_work+0x1cc/0x640 [bluetooth]
>         process_one_work+0x244/0x5f0
>         worker_thread+0x3c/0x380
>         kthread+0x13e/0x160
>         ret_from_fork+0x22/0x30
>  
>  -> #2 (&chan->lock#2/1){+.+.}-{3:3}:
>         __mutex_lock+0xa3/0xa10
>         l2cap_chan_connect+0x33a/0x940 [bluetooth]
>         l2cap_sock_connect+0x141/0x2a0 [bluetooth]
>         __sys_connect+0x9b/0xc0
>         __x64_sys_connect+0x16/0x20
>         do_syscall_64+0x33/0x80
>         entry_SYSCALL_64_after_hwframe+0x44/0xae
>  
>  -> #1 (&conn->chan_lock){+.+.}-{3:3}:
>         __mutex_lock+0xa3/0xa10
>         l2cap_chan_connect+0x322/0x940 [bluetooth]
>         l2cap_sock_connect+0x141/0x2a0 [bluetooth]
>         __sys_connect+0x9b/0xc0
>         __x64_sys_connect+0x16/0x20
>         do_syscall_64+0x33/0x80
>         entry_SYSCALL_64_after_hwframe+0x44/0xae
>  
>  -> #0 (&hdev->lock){+.+.}-{3:3}:
>         __lock_acquire+0x147a/0x1a50
>         lock_acquire+0x277/0x3d0
>         __mutex_lock+0xa3/0xa10
>         hci_conn_get_phy+0x1c/0x150 [bluetooth]
>         l2cap_sock_getsockopt+0x5a9/0x610 [bluetooth]
>         __sys_getsockopt+0xcc/0x200
>         __x64_sys_getsockopt+0x20/0x30
>         do_syscall_64+0x33/0x80
>         entry_SYSCALL_64_after_hwframe+0x44/0xae

So looking at the code and digging a bit in the history, it seems like the 
above dependency chain has been there since ever ...

>  other info that might help us debug this:
> 
>  Chain exists of:
>    &hdev->lock --> &chan->lock#2/1 --> sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP
>   Possible unsafe locking scenario:
> 
>         CPU0                    CPU1
>         ----                    ----
>    lock(sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP);
>                                 lock(&chan->lock#2/1);
>                                 lock(sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP);
>    lock(&hdev->lock);
>  
>   *** DEADLOCK ***
> 
>  1 lock held by bluetoothd/1118:
>   #0: ffff8f07e831d920 (sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP){+.+.}-{0:0}, at: l2cap_sock_getsockopt+0x8b/0x610 [bluetooth]
>  
>  stack backtrace:
>  CPU: 3 PID: 1118 Comm: bluetoothd Not tainted 5.12.0-rc1-00026-g73d464503354 #10
>  Hardware name: LENOVO 20K5S22R00/20K5S22R00, BIOS R0IET38W (1.16 ) 05/31/2017
>  Call Trace:
>   dump_stack+0x7f/0xa1
>   check_noncircular+0x105/0x120
>   ? __lock_acquire+0x147a/0x1a50
>   __lock_acquire+0x147a/0x1a50
>   lock_acquire+0x277/0x3d0
>   ? hci_conn_get_phy+0x1c/0x150 [bluetooth]
>   ? __lock_acquire+0x2e1/0x1a50
>   ? lock_is_held_type+0xb4/0x120
>   ? hci_conn_get_phy+0x1c/0x150 [bluetooth]
>   __mutex_lock+0xa3/0xa10
>   ? hci_conn_get_phy+0x1c/0x150 [bluetooth]
>   ? lock_acquire+0x277/0x3d0
>   ? mark_held_locks+0x49/0x70
>   ? mark_held_locks+0x49/0x70
>   ? hci_conn_get_phy+0x1c/0x150 [bluetooth]
>   hci_conn_get_phy+0x1c/0x150 [bluetooth]
>   l2cap_sock_getsockopt+0x5a9/0x610 [bluetooth]
>   __sys_getsockopt+0xcc/0x200
>   __x64_sys_getsockopt+0x20/0x30
>   do_syscall_64+0x33/0x80
>   entry_SYSCALL_64_after_hwframe+0x44/0xae

... but the sk_lock-AF_BLUETOOTH-BTPROTO_L2CAP -> conn->hdev dependency 
has been added only in eab2404ba798 ("Bluetooth: Add BT_PHY socket 
option") and I've started to see this splat only now as I've probably 
recently acquired userspace that excercises this getsockopt(BT_PHY).

-- 
Jiri Kosina
SUSE Labs