Re: [Bug 17752] 2.6.36-rc3: inconsistent lock state (iprune_sem, shrink_icache_memory)

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

bugzilla-daemon@bugzilla.kernel.org wrote:
> https://bugzilla.kernel.org/show_bug.cgi?id=17752
> 
> 
> Maciej Rutecki <maciej.rutecki@gmail.com> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>              Blocks|                            |16444

This is probably not a regression.  See this earlier discussion:
http://lkml.org/lkml/2010/1/15/76

And another report, coincidentally at the same time:
http://lkml.org/lkml/2010/1/18/108

As I understand Christoph's post on January 19, several code paths are
independently affected:
http://lkml.org/lkml/2010/1/19/267
(Maybe some of those have been fixed in the meantime.)

I don't see in my trace from September 1 which filesystem caused the
particular lockdep report.  I may have unmounted ext4, vfat, or fuse prior to
that.

Dave wrote on September 2:
>> Any memory allocation that enters reclaim in the unmount path will
>> generate this warning. The problem is that the normal memory reclaim
>> path is:
>> 
>> 	alloc -> reclaim -> shrink_slab -> shrink_icache_memory -> iprune_sem -> s_umount
>> 
>> while unmmount does:
>> 
>> 	unmount -> s_umount -> alloc -> lockdep goes boom!
>> or
>> 	unmount -> iprune_sem -> alloc -> lockdep goes boom!
>> 
>> I never got a straight answer on this, but it the warnings are
>> indicating that you must use GFP_NOFS allocations for every
>> allocation in the unmount path, which is kind of hard to know
>> about given the code is not unomunt path specific....

Is it feasible to add a gfp_flag argument to all call chains that could lead
to allocations in unmount?
-- 
Stefan Richter
-=====-==-=- =--= ---==
http://arcgraph.de/sr/