Re: [rfcomm_run] WARNING: CPU: 1 PID: 79 at kernel/sched/core.c:7156 __might_sleep()

[Peter Hurley <peter@hurleysoftware.com>]

On 10/02/2014 08:54 AM, Peter Zijlstra wrote:
> On Thu, Oct 02, 2014 at 08:38:46AM -0400, Peter Hurley wrote:
>> On 10/02/2014 08:31 AM, Peter Zijlstra wrote:
>>> On Thu, Oct 02, 2014 at 01:09:27PM +0200, Peter Zijlstra wrote:
>>>> On Tue, Sep 30, 2014 at 04:02:28PM +0800, Fengguang Wu wrote:
>>>>> Hi Peter,
>>>>>
>>>>> We possibly find a rfcomm bug (maintainers CCed) exposed by your debug patch
>>>>>
>>>>> [    1.861895] NET: Registered protocol family 5
>>>>> [    1.862978] Bluetooth: RFCOMM TTY layer initialized
>>>>> [    1.863099] ------------[ cut here ]------------
>>>>> [    1.863105] WARNING: CPU: 1 PID: 79 at kernel/sched/core.c:7156 __might_sleep+0x17d/0x1a1()
>>>>> [    1.863112] do not call blocking ops when !TASK_RUNNING; state=1 set at [<c14dc381>] rfcomm_run+0xdf/0x130e
>>>>> [    1.863591]  [<c1058b73>] ? kthread_stop+0x53/0x53
>>>>> [    1.864906]  [<c155a411>] dump_stack+0x48/0x60
>>>>> [    1.866298]  [<c14dc381>] ? rfcomm_run+0xdf/0x130e
>>>>
>>>> Ha yes, rfcomm_run is a complete buggy mess indeed. Lemme go see what I
>>>> can make of it.
>>>
>>> ---
>>> Subject: rfcomm: Fix broken wait construct
>>>
>>> rfcomm_run() is a tad broken in that is has a nested wait loop. One
>>> cannot rely on p->state for the outer wait because the inner wait will
>>> overwrite it.
>>>
>>> While at it, rename rfcomm_schedule to rfcomm_wake, since that is what
>>> it actually does.
>>
>> rfcomm_schedule() as in schedule_work(), which is how it's used.
> 
> Not really, all it does is wake the rfcomm_thread. The thread then does
> a linear walk of all known sessions looking for work -- which is clearly
> suboptimal as well, but I didn't feel like fixing that.
> 
> Also, the current implementation already disagrees with you, all it
> basically does it call wake_up_process() which is a big clue right
> there.

You're thinking of it from the point of view of the scheduler, so to you
it should be named what it does.

However, from the users' point of view, it's an abstraction of work
dispatching; the fact that a kthread (which needs waking) does the work
is irrelevant.

Consider if the kthread is converted to work_structs instead and your now-
renamed rfcomm_wake() is calling schedule_work().

Regards,
Peter Hurley