Re: downgrade_write replacement in remap_file_pages

[Andrew Morton <akpm@osdl.org>]

Andrea Arcangeli <andrea@suse.de> wrote:
>
> Apparently downgrade_write deadlocks the kernel in the mmap_sem
> under load. I suspect some rwsem bug. Anyways it's matter of time before
> in my tree I replace the rwsem implementation with my spinlock based
> common code implementation again that I can understand trivially (unlike
> the current code). I don't mind a microoptimization when the code is so
> complicated, so I don't mind too much to fix whatever current bug in
> downgrade_write.

I must say I agree with the sentiments - the current implementation doesn't
have a very attractive complexity/benefit ratio.  But I wrote a
spinlock-based version three years ago too, so I'm biased ;) Certainly it
is bog-simple and fixes up the overflow-at-32768-waiters bug.

I think a spinlock-based implementation would be OK if it was x86-specific,
because x86 spin_unlock is cheap.  But some architectures do atomic ops in
spin_unlock and won't like it.  Plus those architectures which can
implement atomic_add_return() can implement nice versions of rwsems such as
the ppc64 code.  Although ppc64 still seems to have an overflow bug.

So ho-hum.  As a first step, David, could you please take a look into
what's up with downgrade_write()?

(Then again, we need to have a serious think about the overflow bug.  It's
fatal.  Should we fix it?  If so, the current rwsem implementation is
probably unsalvageable).