lkml.org 
[lkml]   [2014]   [Jan]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [Intel-gfx] [PATCH v2] drm/i915: fix dp/sdvo i2c cleanup
From
On Sun, Jan 26, 2014 at 1:51 AM, Imre Deak <imre.deak@intel.com> wrote:
> On Sat, 2014-01-25 at 21:37 +0100, Daniel Vetter wrote:
>> On Fri, Jan 24, 2014 at 02:47:33PM +0200, Imre Deak wrote:
>> > Atm we try to remove the connector's i2c sysfs entry too late in the
>> > encoder's destroy callback. By that time the kobject used as the parent
>> > for all connector sysfs entries is already removed when we do an early
>> > removal of all connector sysfs entries in intel_modeset_cleanup(). Fix
>> > this by adding an early_destroy encoder callback, where we remove the
>> > encoder's i2c adapter.
>> >
>> > v2:
>> > - add missing static to function, use existing sdvo cast helper,
>> > s/intel_sdvo/sdvo/ (Chris)
>> >
>> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70523
>> >
>> > Signed-off-by: Imre Deak <imre.deak@intel.com>
>>
>> This looks fishy ... sysfs should all be reference-counted, so I'm
>> confused what's going on here. Also I think this smells a bit like it's
>> going overall in the wrong direction - essentially with sysfs we can't
>> really force-remove stuff but have to wait until the refcount drops to 0.
>> Or at least that's how I think it works, I'd need to blow through a pile
>> of time to figure this all out.
>
> Hm, I haven't thought about refcounting :) Now, I agree that should
> normally allow for removing a parent and child device in both order.
>
> What happens and why we can't remove first the parent then the child:
>
> In
>
> intel_dp_init_connector()->
> intel_dp_i2c_init()
>
> we set the i2c adapter.dev.parent = intel_connector->base.kdev;
>
> device_register will then add the i2c sysfs entry to the connector sysfs
> dir. Refcounting here looks ok, both the parent connector kobject and
> its sysfs dir entry gets a reference from the child.
>
> During module cleanup, we call
>
> intel_modeset_cleanup()->
> drm_sysfs_connector_remove()
> device_unregister(connector->kdev)
>
> which is the i2c adapter's parent kdev. Then:
>
> device_del()->
> kobject_del()->
> sysfs_remove_dir()
>
> will remove all entries recursively in the connector's sysfs dir, along
> with all the i2c sysfs entries. Afterwards the intel encoder->destroy()
> callback calls i2c_del_adapter()->device_unregister()->device_del() and
> that will try to remove its own sysfs attributes, namely the power sysfs
> group, but won't find it since it was removed above and give a WARN.
>
> Note that the parent's recursive removal happens regardless of its
> kobject's or sysfs entry's refcount. The kobject itself will be put
> after device_del() in put_device() and only destroyed after the i2c
> adapter releases the refcount it holds on it. I admit it feels strange
> that the sysfs entries are removed before the last reference on the
> kobject is dropped, not sure if it's by design or an overlook..

I have no idea either how exactly this is supposed to work, and I
quick scan through Documentation/ didn't point me into a useful
direction either.

Adding Greg (and more mailing lists) for insight.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch


\
 
 \ /
  Last update: 2014-01-26 10:41    [W:0.043 / U:0.460 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site