Bluetooth: Call drain_workqueue() before resetting state
authorJohan Hedberg <johan.hedberg@intel.com>
Tue, 18 Nov 2014 07:00:14 +0000 (09:00 +0200)
committerMarcel Holtmann <marcel@holtmann.org>
Tue, 18 Nov 2014 07:32:08 +0000 (08:32 +0100)
commit76727c02c1e14a2b561b806fa1d08acc1619ad27
treefb8c465d0f10a2f50a9840a73becb1ea7442807e
parent38da1703060a520e69b00405f9bdf765d1396cd0
Bluetooth: Call drain_workqueue() before resetting state

Doing things like hci_conn_hash_flush() while holding the hdev lock is
risky since its synchronous pending work cancellation could cause the
L2CAP layer to try to reacquire the hdev lock. Right now there doesn't
seem to be any obvious places where this would for certain happen but
it's already enough to cause lockdep to start warning against the hdev
and the work struct locks being taken in the "wrong" order:

[  +0.000373] mgmt-tester/1603 is trying to acquire lock:
[  +0.000292]  ((&conn->pending_rx_work)){+.+.+.}, at: [<c104266d>] flush_work+0x0/0x181
[  +0.000270]
but task is already holding lock:
[  +0.000000]  (&hdev->lock){+.+.+.}, at: [<c13b9a80>] hci_dev_do_close+0x166/0x359
[  +0.000000]
which lock already depends on the new lock.

[  +0.000000]
the existing dependency chain (in reverse order) is:
[  +0.000000]
-> #1 (&hdev->lock){+.+.+.}:
[  +0.000000]        [<c105ea8f>] lock_acquire+0xe3/0x156
[  +0.000000]        [<c140c663>] mutex_lock_nested+0x54/0x375
[  +0.000000]        [<c13d644b>] l2cap_recv_frame+0x293/0x1a9c
[  +0.000000]        [<c13d7ca4>] process_pending_rx+0x50/0x5e
[  +0.000000]        [<c1041a3f>] process_one_work+0x21c/0x436
[  +0.000000]        [<c1041e3d>] worker_thread+0x1be/0x251
[  +0.000000]        [<c1045a22>] kthread+0x94/0x99
[  +0.000000]        [<c140f801>] ret_from_kernel_thread+0x21/0x30
[  +0.000000]
-> #0 ((&conn->pending_rx_work)){+.+.+.}:
[  +0.000000]        [<c105e158>] __lock_acquire+0xa07/0xc89
[  +0.000000]        [<c105ea8f>] lock_acquire+0xe3/0x156
[  +0.000000]        [<c1042696>] flush_work+0x29/0x181
[  +0.000000]        [<c1042864>] __cancel_work_timer+0x76/0x8f
[  +0.000000]        [<c104288c>] cancel_work_sync+0xf/0x11
[  +0.000000]        [<c13d4c18>] l2cap_conn_del+0x72/0x183
[  +0.000000]        [<c13d8953>] l2cap_disconn_cfm+0x49/0x55
[  +0.000000]        [<c13be37a>] hci_conn_hash_flush+0x7a/0xc3
[  +0.000000]        [<c13b9af6>] hci_dev_do_close+0x1dc/0x359
[  +0.012038]        [<c13bbe38>] hci_unregister_dev+0x6e/0x1a3
[  +0.000000]        [<c12d33c1>] vhci_release+0x28/0x47
[  +0.000000]        [<c10dd6a9>] __fput+0xd6/0x154
[  +0.000000]        [<c10dd757>] ____fput+0xd/0xf
[  +0.000000]        [<c1044bb2>] task_work_run+0x6b/0x8d
[  +0.000000]        [<c1001bd2>] do_notify_resume+0x3c/0x3f
[  +0.000000]        [<c140fa70>] work_notifysig+0x29/0x31
[  +0.000000]
other info that might help us debug this:

[  +0.000000]  Possible unsafe locking scenario:

[  +0.000000]        CPU0                    CPU1
[  +0.000000]        ----                    ----
[  +0.000000]   lock(&hdev->lock);
[  +0.000000]                                lock((&conn->pending_rx_work));
[  +0.000000]                                lock(&hdev->lock);
[  +0.000000]   lock((&conn->pending_rx_work));
[  +0.000000]
 *** DEADLOCK ***

Fully fixing this would require some quite heavy refactoring to change
how the hdev lock and hci_conn instances are handled together. A simpler
solution for now which this patch takes is to try ensure that the hdev
workqueue is empty before proceeding with the various cleanup calls,
including hci_conn_hash_flush().

Signed-off-by: Johan Hedberg <johan.hedberg@intel.com>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
net/bluetooth/hci_core.c