The hci_conn objects don't have a dedicated lock themselves but rely
on the caller to hold the hci_dev lock for most types of access. The
hci_conn_timeout() function has so far sent certain HCI commands based
on the hci_conn state which has been possible without holding the
hci_dev lock.
The recent changes to do LE scanning before connect attempts added
even more operations to hci_conn and hci_dev from hci_conn_timeout,
thereby exposing potential race conditions with the hci_dev and
hci_conn states.
As an example of such a race, here there's a timeout but an
l2cap_sock_connect() call manages to race with the cleanup routine:
[Oct21 08:14] l2cap_chan_timeout: chan ee4b12c0 state BT_CONNECT
[ +0.000004] l2cap_chan_close: chan ee4b12c0 state BT_CONNECT
[ +0.000002] l2cap_chan_del: chan ee4b12c0, conn f3141580, err 111, state BT_CONNECT
[ +0.000002] l2cap_sock_teardown_cb: chan ee4b12c0 state BT_CONNECT
[ +0.000005] l2cap_chan_put: chan ee4b12c0 orig refcnt 4
[ +0.000010] hci_conn_drop: hcon f53d56e0 orig refcnt 1
[ +0.000013] l2cap_chan_put: chan ee4b12c0 orig refcnt 3
[ +0.000063] hci_conn_timeout: hcon f53d56e0 state BT_CONNECT
[ +0.000049] hci_conn_params_del: addr ee:0d:30:09:53:1f (type 1)
[ +0.000002] hci_chan_list_flush: hcon f53d56e0
[ +0.000001] hci_chan_del: hci0 hcon f53d56e0 chan f4e7ccc0
[ +0.004528] l2cap_sock_create: sock e708fc00
[ +0.000023] l2cap_chan_create: chan ee4b1770
[ +0.000001] l2cap_chan_hold: chan ee4b1770 orig refcnt 1
[ +0.000002] l2cap_sock_init: sk ee4b3390
[ +0.000029] l2cap_sock_bind: sk ee4b3390
[ +0.000010] l2cap_sock_setsockopt: sk ee4b3390
[ +0.000037] l2cap_sock_connect: sk ee4b3390
[ +0.000002] l2cap_chan_connect: 00:02:72:d9:e5:8b -> ee:0d:30:09:53:1f (type 2) psm 0x00
[ +0.000002] hci_get_route: 00:02:72:d9:e5:8b -> ee:0d:30:09:53:1f
[ +0.000001] hci_dev_hold: hci0 orig refcnt 8
[ +0.000003] hci_conn_hold: hcon f53d56e0 orig refcnt 0
Above the l2cap_chan_connect() shouldn't have been able to reach the
hci_conn f53d56e0 anymore but since hci_conn_timeout didn't do proper
locking that's not the case. The end result is a reference to hci_conn
that's not in the conn_hash list, resulting in list corruption when
trying to remove it later:
[Oct21 08:15] l2cap_chan_timeout: chan ee4b1770 state BT_CONNECT
[ +0.000004] l2cap_chan_close: chan ee4b1770 state BT_CONNECT
[ +0.000003] l2cap_chan_del: chan ee4b1770, conn f3141580, err 111, state BT_CONNECT
[ +0.000001] l2cap_sock_teardown_cb: chan ee4b1770 state BT_CONNECT
[ +0.000005] l2cap_chan_put: chan ee4b1770 orig refcnt 4
[ +0.000002] hci_conn_drop: hcon f53d56e0 orig refcnt 1
[ +0.000015] l2cap_chan_put: chan ee4b1770 orig refcnt 3
[ +0.000038] hci_conn_timeout: hcon f53d56e0 state BT_CONNECT
[ +0.000003] hci_chan_list_flush: hcon f53d56e0
[ +0.000002] hci_conn_hash_del: hci0 hcon f53d56e0
[ +0.000001] ------------[ cut here ]------------
[ +0.000461] WARNING: CPU: 0 PID: 1782 at lib/list_debug.c:56 __list_del_entry+0x3f/0x71()
[ +0.000839] list_del corruption, f53d56e0->prev is LIST_POISON2 (00000200)
The necessary fix is unfortunately more complicated than just adding
hci_dev_lock/unlock calls to the hci_conn_timeout() call path.
Particularly, the hci_conn_del() API, which expects the hci_dev lock to
be held, performs a cancel_delayed_work_sync(&hcon->disc_work) which
would lead to a deadlock if the hci_conn_timeout() call path tries to
acquire the same lock.
This patch solves the problem by deferring the cleanup work to a
separate work callback. To protect against the hci_dev or hci_conn
going away meanwhile temporary references are taken with the help of
hci_dev_hold() and hci_conn_get().
Signed-off-by: Johan Hedberg <johan.hedberg@intel.com>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
Cc: stable@vger.kernel.org # 4.3
8ce783dc5e