Re: lockdep splat from ioctl and mmap fops sharing lock

[Stefan Richter <stefanr@s5r6.in-berlin.de>]

Johannes Weiner wrote:
> Johannes Weiner <hannes@saeurebad.de> writes:
> 
>> Hi,
>>
>> fb_mmap() is called under mmap_sem and acquires fb_info->lock.
>>
>> Since 3e680aae4e53ab "fb: convert lock/unlock_kernel() into local fb
>> mutex", fb_ioctl() takes fb_info->lock and does copy_to/from_user()
>> under it, which in turn might acquire mmap_sem.
> 
> More affected drivers:
> 
> agp_ioctl() doing usercopy under agp_fe.agp_mutex
> agp_mmap() taking agp_fe.agp_mutex under mmap_sem
> 
> dv1394_write() doing usercopy under video->mtx
> dv1394_mmap() taking video->mtx under mmap_sem

OK, it doesn't surprise me that I didn't notice this one myself:  The
video->mtx usage is old, but dv1394 is not frequently used anymore; I
for one don't test anything of it beyond loading and unloading.

> raw1394_ioctl() doing usercopy under fi->state_mutex
> raw1394_mmap() taking fi->state_mutex under mmap_sem

The state_mutex in raw1394 however was introduced by me in patches
written against 2.6.26 and 2.6.27-rcs.  And I tested the ioctls and I'd
like to think that I also tested mmaps.  But maybe I didn't.  Of course
I ahve all sorts of lockdep options enabled.

So, was the usage of mmap_sem changed after 2.6.27 or were my tests
insufficient?
-- 
Stefan Richter
-=====-==--- =-=- =-===
http://arcgraph.de/sr/