Re: tasks getting stuck on mmap_sem?

[David Miller <davem@davemloft.net>]

From: Tony Luck <tony.luck@intel.com>
Date: Tue, 17 Aug 2010 11:24:19 -0700

> On Tue, Aug 17, 2010 at 9:14 AM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>> No. Looks like the rwsem changes broke sparc too. ia64 had some problems too.
> 
> I did have some similar mmap_sem issues - but the combination of
> fixing the types
> of the RWSEM_* defines to be unsigned, and the return value of
> ia64_atomic64_add()
> to be "long" rather than "int" looks to have cleared up the problems I
> was seeing. I
> could generally see the hung processes in less than 10 consecutive
> kernel builds, but
> ran a few thousand builds over the weekend with no issues.

You might be triggering it via threading as well, since make uses
vfork() for running sub-jobs.

> If git is multi-threaded, it may be hitting some different code path
> ... but it isn't trivial
> for me to try this out (my systems are on an isolated lab network segment).

Like you I tried fixing atomic64, but as I mentioned it turns out
sparc64 (like powerpc) always uses a 32-bit 'int' semaphore count.

I thought perhaps that for some reason the rwsem generic code had
a dependency on 'long' so I switched sparc64's rwsems over to 'long'
counters last night too, but I still get the problem.

Even reverting the rwsem commit that added the "<" test didn't fix
things, so now I'm simply going to bisect.