On Thu, 31 Aug 2017, Sebastian Andrzej Siewior wrote:
> On 2017-08-30 23:25:53 [+0200], Mart van de Wege wrote:
> > Hi,
> Hi,
>
> > The issue persists up to v4.11.12-rt10
> Does
> CONFIG_RWLOCK_RT_READER_BIASED=y
> make it go away?
>
> > Stacktrace:
> >
> > Aug 29 17:28:04 localhost kernel: [ 46.483810] ------------[ cut here ]------------
> > Aug 29 17:28:04 localhost kernel: [ 46.483812] kernel BUG at kernel/locking/rtmutex.c:1059!
>
> this is a deadlock on RT however !RT has a hidden problem which not
> yelled at by lockdep. We have the following call path:
>
> | hci_send_monitor_ctrl_event()
> | read_lock(&hci_sk_list.lock);
> |
> | hci_send_to_channel()
> | read_lock(&hci_sk_list.lock);
>
> So both functions acquire the same read_lock. If a write comes along
> between the first read_lock() and second read_lock() then we have a
> deadlock because the write_lock() will lock down further readers from
> acquiring the read-lock until the writer completed its task.
>
> This recursive locking was introduced in 38ceaa00d02d ("Bluetooth: Add
> support for sending MGMT commands and events to monitor").
Just for clarification: This is a mainline issue because qrlock based
rwlocks are writer fair, while the traditional rwlocks are reader
biased. Lockdep does not yell about it yet, because the wrapper function
use read-recursive mode.
Thanks,
tglx
--
To unsubscribe from this list: send the line "unsubscribe linux-bluetooth" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
