[Devel] Re: [PATCH 8/8] Per-container pages reclamation

Balbir Singh balbir at linux.vnet.ibm.com
Thu May 24 00:59:55 PDT 2007


Pavel Emelianov wrote:
>> Index: linux-2.6.20/mm/rss_container.c
>> ===================================================================
>> --- linux-2.6.20.orig/mm/rss_container.c	2007-05-15 05:13:46.000000000 -0700
>> +++ linux-2.6.20/mm/rss_container.c	2007-05-16 20:45:45.000000000 -0700
>> @@ -212,6 +212,7 @@ void container_rss_del(struct page_conta
>>  
>>  	css_put(&rss->css);
>>  	kfree(pc);
>> +	init_page_container(page);
> 
> This hunk is bad.
> See, when the page drops its mapcount to 0 it may be reused right
> after this if it belongs to a file map - another CPU can touch it.
> Thus you're risking to reset the wrong container.
> 
> The main idea if the accounting is that you cannot trust the
> page_container(page) value after the page's mapcount became 0.
> 

Good catch, I'll move the initialization to free_hot_cold_page().
I'm attaching a new patch. I've also gotten rid of the unused
variable page in container_rss_del().

I've compile and boot tested the fix

-- 
	Thanks,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL
-------------- next part --------------
A non-text attachment was scrubbed...
Name: rss-fix-lru-race.patch
Type: text/x-patch
Size: 2617 bytes
Desc: not available
URL: <http://lists.openvz.org/pipermail/devel/attachments/20070524/3bd53df0/attachment-0001.bin>


More information about the Devel mailing list