Re: [1/2] 2.6.22-rc7: known regressions

["Dmitry Torokhov" <dmitry.torokhov@gmail.com>]

On 7/5/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:
>
> Looks input-related..
>
> On Thu, 5 Jul 2007, David Woodhouse wrote:
> >
> > Hm, it's not something new. It's an oops I saw occasionally in 2.6.21-rc
> > too, whenever we had CONFIG_SYSFS_DEPRECATED set.
> >
> >  Unable to handle kernel paging request for data at address 0x6b6b6b6b
>
> Ok, that 0x6b is obviously the kfree() poisoning, ie it looks like a
> use-after-free problem with a pointer being loaded from a structure that
> had been free'd-
>
> And the trace seems to be (ignore the unreliable one):
>
> >  NIP [c001870c] strlen+0x4/0x18
> >  LR [c0134fec] kobject_get_path+0x34/0xc4
> >  Call Trace:
> >  [eed5be90] [c01d5124] class_uevent+0xac/0x1bc
> >  [eed5bed0] [c01357e4] kobject_uevent_env+0x23c/0x460
> >  [eed5bf20] [c01d485c] class_device_del+0x178/0x1a0
> >  [eed5bf40] [c01d489c] class_device_unregister+0x18/0x30
> >  [eed5bf60] [c021f820] input_unregister_device+0xf4/0x130
> >  [eed5bf70] [c0242f4c] hidinput_disconnect+0x2c/0x60
> >  [eed5bf90] [f27f2bac] hidp_session+0x550/0x584 [hidp]
> >  [eed5bff0] [c0013e28] kernel_thread+0x44/0x60
>
> Where we have a few missing functions due to inlining, ie the real
> sequence seems to be:
>
> class_device_del ->
>  kobject_uevent_env ->
>    class_uevent ->
>      kobject_get_path ->
>        get_kobj_path_length ->
>          parent = kobj;
>          do {
>            strlen(parent->k_name /* kobject_name(parent) */);
>            parent = parent->parent;
>          } while (parent);
>
> so either the kobj or one of it's parents had already been freed when it
> was unregistered due to the disconnect.
>
> I'm not seeing any reference counting or other protection for the device
> ("input") on "hid->inputs" list. But I don't know the code. Dmitry? Jiri?
>

In hidinput_connect we do:

          input_dev->dev.parent = hid->dev;

This should pin hid object untill all inputs are released. However
bluetooth does not use driver model and does not have hid->dev set up
and so it looks like we are simply trying to unregister an input
device that is already gone... I still don't quite get how we
unregister the same device twice - it is done from a per-hid-device
thread in hidp...

-- 
Dmitry