lkml.org 
[lkml]   [2011]   [May]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] [BUGFIX] mm: hugepages can cause negative commitlimit
On Thu, May 19, 2011 at 10:37:13AM -0300, Rafael Aquini wrote:
> On Thu, May 19, 2011 at 1:56 AM, Russ Anderson <rja@sgi.com> wrote:
> >
> > The way it was verified was putting a printk in to print totalram_pages
> > and hugetlb_total_pages. First the system was booted without any huge
> > pages. The next boot one huge page was allocated. The next boot more
> > hugepages allocated. Each time totalram_pages was reduced by the nuber
> > of huge pages allocated, with totalram_pages + hugetlb_total_pages
> > equaling the original number of pages.
> >
> > That behavior is also consistent with allocating over half of memory
> > resulting in CommitLimit going negative (as is shown in the above
> > output).
> >
> > Here is some data. Each represents a boot using 1G hugepages.
> > 0 hugepages : totalram_pages 16519867 hugetlb_total_pages 0
> > 1 hugepages : totalram_pages 16257723 hugetlb_total_pages 262144
> > 2 hugepages : totalram_pages 15995578 hugetlb_total_pages 524288
> > 31 hugepages : totalram_pages 8393403 hugetlb_total_pages 8126464
> > 32 hugepages : totalram_pages 8131258 hugetlb_total_pages 8388608
> >
> >
> > > hugepages are reserved, hugetlb_total_pages() has to be accounted and
> > > subtracted from totalram_pages in order to render an accurate number of
> > > remaining pages available to the general memory workload commitment.
> > >
> > > I've tried to reproduce your findings on my boxes, without
> > > success, unfortunately.
> >
> > Put a printk in meminfo_proc_show() to print totalram_pages and
> > hugetlb_total_pages(). Add "default_hugepagesz=1G hugepagesz=1G
> > hugepages=64"
> > to the boot line (varying the number of hugepages).
> >
> > > I'll keep chasing to hit this behaviour, though.
>
> I got what I was doing different, and you are partially right.
> Checking mm/hugetlb.c:
> 1811 static int __init hugetlb_nrpages_setup(char *s)
> 1812 {
> ....
> 1834 /*
> 1835 * Global state is always initialized later in hugetlb_init.
> 1836 * But we need to allocate >= MAX_ORDER hstates here early to
> still
> 1837 * use the bootmem allocator.
> 1838 */
> 1839 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
> 1840 hugetlb_hstate_alloc_pages(parsed_hstate);
> 1841
> 1842 last_mhp = mhp;
> 1843
> 1844 return 1;
> 1845 }
> 1846 __setup("hugepages=", hugetlb_nrpages_setup);
>
> I realize this issue you've reported only happens when you're using
> oversized hugepages. As their order are always >= MAX_ORDER, they got pages
> early allocated from bootmem allocator. So, these pages are not accounted
> for totalram_pages.
>
> Although your patch covers a fix for the proposed case, it only works for
> scenarios where oversized hugepages are allocated on boot. I think it will,
> unfortunately, cause a bug for the remaining scenarios.

OK, I see your point. The root problem is hugepages allocated at boot are
subtracted from totalram_pages but hugepages allocated at run time are not.
Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages".
That allocation will not change totalram_pages but will change
hugetlb_total_pages().
How best to fix this inconsistency? Should totalram_pages include or exclude
hugepages? What are the implications?

I have no strong preference as to which way to go as long as it is consistent.

> Cheers!
> --aquini

--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc rja@sgi.com


\
 
 \ /
  Last update: 2011-05-20 00:13    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site