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

[Andrew Morton]

Mike Fedyk wrote:
> 
> Hello,
> 
> I am trying to track down a memory accounting problem I've seen ever since I
> tried a 2.4.13 based kernel.  Specifically I've noticed an overflow for the
> "Cached" entry in /proc/meminfo, but also the numbers don't add up to the
> total memory count.  Shouldn't they add up?  If they should, I haven't seen
> one that does....
> 
> I first noticed it on:
> 2.4.13freeswan-1.91+ac5+preempt+netdev_random+vm_freeswap
> 2.4.14-pre6+preempt+netdev_random+ext3_0.9.14-2414p5
> 
> I thought it may be preempt so I tried:
> 2.4.14-pre8+netdev_random-p7+xsched+ext3_0.9.14-2414p8+elevator
> 
> But I still get the same problem with "Cached".
> 
> Now I'm trying to see if it could be ext3 with:
> 2.4.14-ext3-2.4-0.9.14-2414p8
> 
> And I haven't noticed the problem after 16 hours uptime.  Sometimes it would
> show earlier, or later.
> 

Ah.  So are you saying that it does *not* occur on
ext3, but that it does occur on ext2?

Could you retest ext2 with this:

--- linux-2.4.14/mm/filemap.c	Mon Nov  5 21:01:12 2001
+++ linux-akpm/mm/filemap.c	Wed Nov  7 18:09:38 2001
@@ -223,6 +223,8 @@ static void truncate_complete_page(struc
 	/* Leave it on the LRU if it gets converted into anonymous buffers */
 	if (!page->buffers || block_flushpage(page, 0))
 		lru_cache_del(page);
+	else
+		atomic_inc(&buffermem_pages);
 
 	/*
 	 * We remove the page from the page cache _after_ we have

(This will break ext3's accounting)