Re: [PATCH -mm] mm,numa,mprotect: always continue after finding a stable thp page

[Mel Gorman]

On Thu, Mar 06, 2014 at 05:52:47PM -0500, Rik van Riel wrote:
> On 03/06/2014 05:31 PM, Rik van Riel wrote:
> >On Thu, 06 Mar 2014 16:12:28 -0500
> >Sasha Levin <sasha.levin@xxxxxxxxxx> wrote:
> >
> >>While fuzzing with trinity inside a KVM tools guest running latest -next kernel I've hit the
> >>following spew. This seems to be introduced by your patch "mm,numa: reorganize change_pmd_range()".
> >
> >That patch should not introduce any functional changes, except for
> >the VM_BUG_ON that catches the fact that we fell through to the 4kB
> >pte handling code, despite having just handled a THP pmd...
> >
> >Does this patch fix the issue?
> >
> >Mel, am I overlooking anything obvious? :)
> >
> >---8<---
> >
> >Subject: mm,numa,mprotect: always continue after finding a stable thp page
> >
> >When turning a thp pmds into a NUMA one, change_huge_pmd will
> >return 0 when the pmd already is a NUMA pmd.
> 
> I did miss something obvious.  In this case, the code returns 1.
> 
> >However, change_pmd_range would fall through to the code that
> >handles 4kB pages, instead of continuing on to the next pmd.
> 
> Maybe the case that I missed is when khugepaged is in the
> process of collapsing pages into a transparent huge page?
> 
> If the virtual CPU gets de-scheduled by the host for long
> enough, it would be possible for khugepaged to run on
> another virtual CPU, and turn the pmd into a THP pmd,
> before that VM_BUG_ON test.
> 
> I see that khugepaged takes the mmap_sem for writing in the
> collapse code, and it looks like task_numa_work takes the
> mmap_sem for reading, so I guess that may not be possible?
> 

mmap_sem will prevent a parallel collapse but what prevents something
like the following?

							do_huge_pmd_wp_page
change_pmd_range
if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
	continue;
							pmdp_clear_flush(vma, haddr, pmd);
if (pmd_trans_huge(*pmd)) {
	.... path not taken ....
}
							page_add_new_anon_rmap(new_page, vma, haddr);
							set_pmd_at(mm, haddr, pmd, entry);
VM_BUG_ON(pmd_trans_huge(*pmd));

We do not hold the page table lock during the pmd_trans_huge check and we
do not recheck it under PTF lock in change_pte_range()

-- 
Mel Gorman
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/