On 2016/7/21 20:55, Michal Hocko wrote: > On Thu 21-07-16 20:45:15, zhong jiang wrote: >> On 2016/7/21 20:30, Michal Hocko wrote: >>> On Thu 21-07-16 20:14:41, zhong jiang wrote: >>>> On 2016/7/21 19:27, Michal Hocko wrote: >>>>> On Thu 21-07-16 18:54:09, zhong jiang wrote: >>>>>> On 2016/7/21 15:43, Michal Hocko wrote: >>>>>>> We have further discussed the patch and I believe it is not correct. See [1]. >>>>>>> I am proposing the following alternative. >>>>>>> >>>>>>> [1] http://lkml.kernel.org/r/20160720132431.GM11249@xxxxxxxxxxxxxx >>>>>>> --- >>>>>>> >From b1e9b3214f1859fdf7d134cdcb56f5871933539c Mon Sep 17 00:00:00 2001 >>>>>>> From: Michal Hocko <mhocko@xxxxxxxx> >>>>>>> Date: Thu, 21 Jul 2016 09:28:13 +0200 >>>>>>> Subject: [PATCH] mm, hugetlb: fix huge_pte_alloc BUG_ON >>>>>>> >>>>>>> Zhong Jiang has reported a BUG_ON from huge_pte_alloc hitting when he >>>>>>> runs his database load with memory online and offline running in >>>>>>> parallel. The reason is that huge_pmd_share might detect a shared pmd >>>>>>> which is currently migrated and so it has migration pte which is >>>>>>> !pte_huge. >>>>>>> >>>>>>> There doesn't seem to be any easy way to prevent from the race and in >>>>>>> fact seeing the migration swap entry is not harmful. Both callers of >>>>>>> huge_pte_alloc are prepared to handle them. copy_hugetlb_page_range >>>>>>> will copy the swap entry and make it COW if needed. hugetlb_fault will >>>>>>> back off and so the page fault is retries if the page is still under >>>>>>> migration and waits for its completion in hugetlb_fault. >>>>>>> >>>>>>> That means that the BUG_ON is wrong and we should update it. Let's >>>>>>> simply check that all present ptes are pte_huge instead. >>>>>>> >>>>>>> Reported-by: zhongjiang <zhongjiang@xxxxxxxxxx> >>>>>>> Signed-off-by: Michal Hocko <mhocko@xxxxxxxx> >>>>>>> --- >>>>>>> mm/hugetlb.c | 2 +- >>>>>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>>>>> >>>>>>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c >>>>>>> index 34379d653aa3..31dd2b8b86b3 100644 >>>>>>> --- a/mm/hugetlb.c >>>>>>> +++ b/mm/hugetlb.c >>>>>>> @@ -4303,7 +4303,7 @@ pte_t *huge_pte_alloc(struct mm_struct *mm, >>>>>>> pte = (pte_t *)pmd_alloc(mm, pud, addr); >>>>>>> } >>>>>>> } >>>>>>> - BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); >>>>>>> + BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); >>>>>>> >>>>>>> return pte; >>>>>>> } >>>>>> I don't think that the patch can fix the question. The explain is as follow. >>>>>> >>>>>> cpu0 cpu1 >>>>>> copy_hugetlb_page_range try_to_unmap_one >>>>>> huge_pte_alloc #pmd may be shared >>>>>> lock dst_pte #dst_pte may be migrate >>>>>> lock src_pte #src_pte may be normal pt1 >>>>>> set_huge_pte_at #dst_pte points to normal >>>>>> spin_unlock (src_pt1) >>>>>> lock src_pte >>>>>> spin_unlock(dst_pt1) set src_pte migrate entry >>>>>> spin_unlock(src_pte) >>>>>> * dst_pte is a normal pte, but corresponding to the >>>>>> pfn is under migrate. it is dangerous. >>>>>> >>>>>> The race may occur. is right ? if the scenario exist. we should think about more. >>>>> Can this happen at all? copy_hugetlb_page_range does the following to >>>>> rule out shared page table entries. At least that is my understanding of >>>>> c5c99429fa57 ("fix hugepages leak due to pagetable page sharing") >>>>> >>>>> /* If the pagetables are shared don't copy or take references */ >>>>> if (dst_pte == src_pte) >>>>> continue; >>>> vm_file points to mapping should be shared, I am not sure, if it is >>>> so, the possibility is exist. of course, src_pte is the same as the >>>> dst_pte. >>> I am not sure I understand. This is a fork path where the ptes are >>> copied over from the parent to the child. So how would vm_file differ? >> I think you can misunderstand my meaning. A file refers to the >> mapping field can be shared by other process, parent process have the >> mapping , but is not only. This is only my viewpoint. is right ?? > OK, now I understand what you mean. So you mean that a different process > initiates the migration while this path copies to pte. That is certainly > possible but I still fail to see what is the problem about that. > huge_pte_alloc will return the identical pte whether it is regular or > migration one. So what exactly is the problem? > copy_hugetlb_page_range obtain the shared dst_pte, it may be not equal to the src_pte. The dst_pte can come from other process sharing the mapping. /* If the pagetables are shared don't copy or take references */ if (dst_pte == src_pte) continue; Even it do the fork path, we scan the i_mmap to find same pte. I think that dst_pte may come from other process. It is not the parent. it will lead to the dst_pte is not equal to the src_pte from the parent. vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { is right ? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>