Re: [PATCH v2] mm/hugetlb: fix race when migrate pages

[zhong jiang <zhongjiang@huawei.com>]

On 2016/7/20 21:24, Michal Hocko wrote:
> [Sorry for spammin, I though my headache would force me to not stare
>  into the code more but I couldn't resist ]
> On Wed 20-07-16 15:00:55, Michal Hocko wrote:
>> On Wed 20-07-16 14:45:01, Michal Hocko wrote:
>> [...]
>>> I was talking to Mel (CCed) and he has raised a good question. So if you
>>> encounter a page under migration and fail to share the pmd with it how
>>> can you have a proper content of the target page in the end?
>> Hmm, I was staring into the code some more and it seems this would be OK
>> because we should hit hugetlb_no_page with the newel instantiated pmd
>> and associate it with a page from the radix tree. So unless I am missing
>> something the corruption shouldn't be possible.
>>
>> I believe the post pmd_populate race is still there, though, so I
>> believe the approach should be rethought.
> So I think the swap entry is OK in fact. huge_pte_alloc would return a
> shared pmd which would just happen to be swap entry. hugetlb_fault would
> then recognize that by:
> 	/*
> 	 * entry could be a migration/hwpoison entry at this point, so this
> 	 * check prevents the kernel from going below assuming that we have
> 	 * a active hugepage in pagecache. This goto expects the 2nd page fault,
> 	 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
> 	 * handle it.
> 	 */
> 	if (!pte_present(entry))
> 		goto out_mutex;
>
> We would get back from the page fault and if the page was still under
> migration on the retry we would end up waiting for the migration entry
> on the next fault.
>
> 	ptep = huge_pte_offset(mm, address);
> 	if (ptep) {
> 		entry = huge_ptep_get(ptep);
> 		if (unlikely(is_hugetlb_entry_migration(entry))) {
> 			migration_entry_wait_huge(vma, mm, ptep);
> 			return 0;
>
> Or am I missing something? If not we should just reconsider the BUG_ON
> but I still have to think wether it is useful/needed at all.
  you are right. I agree.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>