Hi Xishi, On Mon, Jul 09, 2018 at 09:13:07PM +0800, 裘稀石(稀石) wrote: > Hi Naoya, > > - does the same issue happens on soft offlining of normal pages? > > I thinks yes. We can do anything during get_any_page and set hwpoison flag. > soft_offline_page > get_any_page > soft_offline_free_page > SetPageHWPoison > > I search the key word of PageHWPoison, and these two paths maybe have issue. > do_swap_page > if (PageHWPoison(page)) - ret = VM_FAULT_HWPOISON; > __do_fault > if (unlikely(PageHWPoison(vmf->page))) - return VM_FAULT_HWPOISON; > > It may cause mce kill later. I agree. I checked git history and found some related to this race. IIRC, the race windows seems to open at commit c6c919eb90e0 ("mm: use put_page() to free page instead of putback_lru_page()") which makes migration source pages freed by put_page(). Before that commit, we kept migration source pages pinned to avoid reusing, introduced by commit add05cecef80 ("mm: soft-offline: don't free target page in successful page migration".) And before this commit, migration source pages were freed via putback_lru_page() with changing migration type to MIGRATE_ISOLATE. The oldest behavior was buggy because of freeing with putback_lru_page(), but "changing migration type" part was still correct if there's a time window between freeing a page and setting the PageHWPoison flag. At c6c919eb90e0 we returned back to the behavior of releasing source pages (then we have such time window again,) so we should have restored to change the migration type. So I feel that this solves the above issues in the normal page's case. Does it make sense? This part of code has changed repeatedly and it's painful to track the whole story in git-log, so I'll clarify the latest (hopefully saner) behavior in inline comment. > > As you said allocation failure, I think we will get oom first. We get OOM if alloc_huge_page() returns ENOMEM, and that happens when - vma_needs_reservation() returns negative value, - hugetlb_cgroup_charge_cgroup() fails to charge. It seems to me that hwposion doesn't affect both of these. I presumed that dequeue_huge_page_vma() failed with NULL return due to dequeue_huge_page_node_exact()'s failure. And the result should be ENOSPC or allocation success by overcommitting. So the allocation failure might cause sigbus? > > - does hard offling of free (huge) page have the similar issue? > We can kill process in anytime, right? Yes, so as-is might be fine. Thanks, Naoya Horiguchi > > Thanks, > Xishi Qiu > On Mon, Jul 09, 2018 at 01:43:35PM +0800, 裘稀石(稀石) wrote: > > Hi Naoya, > > > > Shall we fix this path too? It also will set hwpoison before > > dissolve_free_huge_page(). > > > > soft_offline_huge_page > > migrate_pages > > unmap_and_move_huge_page > > if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison > (hpage)) > > dissolve_free_huge_page > > Thank you Xishi, I added it to the current (still draft) version below. > > I start feeling that current code is broken about behavior of PageHWPoison > (at least) in soft offline. And this patch might not cover all of the issues. > My current questions/concerns are: > > - does the same issue happens on soft offlining of normal pages? > - does hard offling of free (huge) page have the similar issue? > > I'll try to clarify these next and will update the patch if necessary. > I'm happy if I get some comment around these. > > Thanks, > Naoya Horiguchi > --- > From 9ce4df899f4c859001571958be6a281cdaf5a58f Mon Sep 17 00:00:00 2001 > From: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> > Date: Mon, 9 Jul 2018 13:07:46 +0900 > Subject: [PATCH] mm: fix race on soft-offlining free huge pages > > There's a race condition between soft offline and hugetlb_fault which > causes unexpected process killing and/or hugetlb allocation failure. > > The process killing is caused by the following flow: > > CPU 0 CPU 1 CPU 2 > > soft offline > get_any_page > // find the hugetlb is free > mmap a hugetlb file > page fault > ... > hugetlb_fault > hugetlb_no_page > alloc_huge_page > // succeed > soft_offline_free_page > // set hwpoison flag > mmap the hugetlb file > page fault > ... > hugetlb_fault > hugetlb_no_page > find_lock_page > return VM_FAULT_HWPOISON > mm_fault_error > do_sigbus > // kill the process > > > The hugetlb allocation failure comes from the following flow: > > CPU 0 CPU 1 > > mmap a hugetlb file > // reserve all free page but don't fault-in > soft offline > get_any_page > // find the hugetlb is free > soft_offline_free_page > // set hwpoison flag > dissolve_free_huge_page > // fail because all free hugepages are reserved > page fault > ... > hugetlb_fault > hugetlb_no_page > alloc_huge_page > ... > dequeue_huge_page_node_exact > // ignore hwpoisoned hugepage > // and finally fail due to no-mem > > The root cause of this is that current soft-offline code is written > based on an assumption that PageHWPoison flag should beset at first to > avoid accessing the corrupted data. This makes sense for memory_failure() > or hard offline, but does not for soft offline because soft offline is > not about corrected error and is safe from data lost. > This patch changes soft offline semantics where it sets PageHWPoison flag > only after containment of the error page completes succesfully. > > Reported-by: Xishi Qiu <xishi.qiuxishi@xxxxxxxxxxxxxxx> > Suggested-by: Xishi Qiu <xishi.qiuxishi@xxxxxxxxxxxxxxx> > Signed-off-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> > --- > mm/hugetlb.c | 11 +++++------ > mm/memory-failure.c | 13 +++++++------ > mm/migrate.c | 2 -- > 3 files changed, 12 insertions(+), 14 deletions(-) > > diff --git a/mm/hugetlb.c b/mm/hugetlb.c > index d34225c1cb5b..3c9ce4c05f1b 100644 > --- a/mm/hugetlb.c > +++ b/mm/hugetlb.c > @@ -1479,22 +1479,20 @@ static int free_pool_huge_page > (struct hstate *h, nodemask_t *nodes_allowed, > /* > * Dissolve a given free hugepage into free buddy pages. This function does > * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the > - * number of free hugepages would be reduced below the number of reserved > - * hugepages. > + * dissolution fails because a give page is not a free hugepage, or because > + * free hugepages are fully reserved. > */ > int dissolve_free_huge_page(struct page *page) > { > - int rc = 0; > + int rc = -EBUSY; > > spin_lock(&hugetlb_lock); > if (PageHuge(page) && !page_count(page)) { > struct page *head = compound_head(page); > struct hstate *h = page_hstate(head); > int nid = page_to_nid(head); > - if (h->free_huge_pages - h->resv_huge_pages == 0) { > - rc = -EBUSY; > + if (h->free_huge_pages - h->resv_huge_pages == 0) > goto out; > - } > /* > * Move PageHWPoison flag from head page to the raw error page, > * which makes any subpages rather than the error page reusable. > @@ -1508,6 +1506,7 @@ int dissolve_free_huge_page(struct page *page) > h->free_huge_pages_node[nid]--; > h->max_huge_pages--; > update_and_free_page(h, head); > + rc = 0; > } > out: > spin_unlock(&hugetlb_lock); > diff --git a/mm/memory-failure.c b/mm/memory-failure.c > index 9d142b9b86dc..7a519d947408 100644 > --- a/mm/memory-failure.c > +++ b/mm/memory-failure.c > @@ -1598,8 +1598,9 @@ static int soft_offline_huge_page > (struct page *page, int flags) > if (ret > 0) > ret = -EIO; > } else { > - if (PageHuge(page)) > - dissolve_free_huge_page(page); > + ret = dissolve_free_huge_page(page); > + if (!ret) > + num_poisoned_pages_inc(); > } > return ret; > } > @@ -1715,13 +1716,13 @@ static int soft_offline_in_use_page > (struct page *page, int flags) > > static void soft_offline_free_page(struct page *page) > { > + int rc = 0; > struct page *head = compound_head(page); > > - if (!TestSetPageHWPoison(head)) { > + if (PageHuge(head)) > + rc = dissolve_free_huge_page(page); > + if (!rc && !TestSetPageHWPoison(page)) > num_poisoned_pages_inc(); > - if (PageHuge(head)) > - dissolve_free_huge_page(page); > - } > } > > /** > diff --git a/mm/migrate.c b/mm/migrate.c > index 198af4289f9b..3ae213b799a1 100644 > --- a/mm/migrate.c > +++ b/mm/migrate.c > @@ -1318,8 +1318,6 @@ static int unmap_and_move_huge_page > (new_page_t get_new_page, > out: > if (rc != -EAGAIN) > putback_active_hugepage(hpage); > - if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage)) > - num_poisoned_pages_inc(); > > /* > * If migration was not successful and there's a freeing callback, use > -- > 2.7.4 >