Re: [PATCH V7 02/24] coresight: associating path with session rather than tracer

From: Mathieu Poirier
Date: Wed Jan 13 2016 - 16:47:21 EST

On 10 January 2016 at 08:43, Rabin Vincent <rabin@xxxxxx> wrote:
> On Tue, Jan 05, 2016 at 03:23:20PM -0700, Mathieu Poirier wrote:
>> On 5 January 2016 at 06:15, Rabin Vincent <rabin@xxxxxx> wrote:
>> > The command sequence is exactly as mentioned above. The kernel is
>> > v4.4rc-8 + all your patches from this thread + the OMAP coresight patch
>> > I mentioned earlier.
>> What patch was that? Would mind providing me with more information?
>> I fixed the kmalloc bug - it stemmed from a rebase conflict in the
>> etm3x driver. On the flip side I haven't been able to reproduce the
>> lockdep problem. But looking at the log you provided I can understand
>> why the kernel is complaining.
>> I produced a patch [1] and added you as a reporter. It would be great
>> if you could give it a quick spin to see if the problem still persist.
>> [1].
> It does silence the warning but simply using mutex_trylock() instead of
> mutex_lock() does not seem like the correct fix. The lock dependency
> problem lockdep is complaining about can be fixed for example by not
> holding the coresight_mutex around the device_unregister()
> coresight_unregister(). I don't quite see what the mutex is protecting
> there?

You are correct, locking at that point doesn't help anything.

> BTW, there seem to be several problems with the unregistration:
> - csdev->refcnt which is allocated with kcalloc() in
> coresight_register() is never freed.
> - csdev->conns is freed before device_unregister() rather than in the
> release function.

I agree.

> - put_device() is not done on the csdev' conn->child_devs when it is
> unregistered, so the reference taken by the bus_find_device() in
> coresight_fixup_device_conns() is never released.

Right, I'll fix that.

> - There is no reference taken for the csdev when it is assigned as
> some other device's child_dev in coresight_orphan_match()

The only time a device shouldn't be removed is when it is part of an
active trace session, something that shouldn't happen due to the
static nature of the coresight devices. When coresight_unregister()
is called the framework should go through the topology and remove
references to the device that is being taken away, something that I
will fix promptly.

That way in 'coresight_fixup_device_conns()', put_device() can be
called when a match is found and references don't have to be taken in
'coresight_orphan_match()' when a device is assigned as a child.

Thanks for the review,