Re: Memory accounting problem in 2.4.13, 2.4.14pre, and possibly 2.4.14

[Mike Fedyk]

On Wed, Nov 07, 2001 at 09:00:26PM -0800, Andrew Morton wrote:
> Mike Fedyk wrote:
> > 
> > >
> > 
> > I am running unpatched 2.4.14 now.
> > 
> > Do you still want me to try this patch now that you know I have been able to
> > see the problem with 2.2.14+ext3?
> > 
> 
> It's OK - I can reproduce it easily anyway.
> 
> There are two things here.  Recent -ac kernels had a merge
> bug down in the /proc code which caused `Cached:' to go
> negative.  It was recently fixed.
> 
> And quite independently, current ext3 for Linus kernels now has a
> bug which causes the `buffermem_pages' number to get too large.
> This has the exact same effect: `Cached:' goes negative. 
> 
> The buffermem_pages counter is purely for reporting - no VM decisions
> are based on its value.  But if it worries you, just remove line 1933 of fs/jbd/transaction.c.

Applied.

Here's the patch.

--- 2.4.14-ext3_0.9.15-2414/fs/jbd/transaction.c~	Wed Nov  7 22:41:13 2001
+++ 2.4.14-ext3_0.9.15-2414/fs/jbd/transaction.c	Wed Nov  7 22:43:14 2001
@@ -1930,7 +1930,6 @@
 
 	if (!offset) {
 		if (!may_free || !try_to_free_buffers(page, 0)) {
-			atomic_inc(&buffermem_pages);
 			return 0;
 		}
 		J_ASSERT(page->buffers == NULL);

I'll run it overnight, and test it tomorrow...

Mike