From: Andrea Arcangeli Date: Thu, 5 May 2016 23:22:20 +0000 (-0700) Subject: mm: thp: kvm: fix memory corruption in KVM with THP enabled X-Git-Url: https://git.stricted.de/?a=commitdiff_plain;h=127393fbe597dd85863a9bdccaa11007e7d4948f;p=GitHub%2Fmoto-9609%2Fandroid_kernel_motorola_exynos9610.git mm: thp: kvm: fix memory corruption in KVM with THP enabled After the THP refcounting change, obtaining a compound pages from get_user_pages() no longer allows us to assume the entire compound page is immediately mappable from a secondary MMU. A secondary MMU doesn't want to call get_user_pages() more than once for each compound page, in order to know if it can map the whole compound page. So a secondary MMU needs to know from a single get_user_pages() invocation when it can map immediately the entire compound page to avoid a flood of unnecessary secondary MMU faults and spurious atomic_inc()/atomic_dec() (pages don't have to be pinned by MMU notifier users). Ideally instead of the page->_mapcount < 1 check, get_user_pages() should return the granularity of the "page" mapping in the "mm" passed to get_user_pages(). However it's non trivial change to pass the "pmd" status belonging to the "mm" walked by get_user_pages up the stack (up to the caller of get_user_pages). So the fix just checks if there is not a single pte mapping on the page returned by get_user_pages, and in turn if the caller can assume that the whole compound page is mapped in the current "mm" (in a pmd_trans_huge()). In such case the entire compound page is safe to map into the secondary MMU without additional get_user_pages() calls on the surrounding tail/head pages. In addition of being faster, not having to run other get_user_pages() calls also reduces the memory footprint of the secondary MMU fault in case the pmd split happened as result of memory pressure. Without this fix after a MADV_DONTNEED (like invoked by QEMU during postcopy live migration or balloning) or after generic swapping (with a failure in split_huge_page() that would only result in pmd splitting and not a physical page split), KVM would map the whole compound page into the shadow pagetables, despite regular faults or userfaults (like UFFDIO_COPY) may map regular pages into the primary MMU as result of the pte faults, leading to the guest mode and userland mode going out of sync and not working on the same memory at all times. Any other secondary MMU notifier manager (KVM is just one of the many MMU notifier users) will need the same information if it doesn't want to run a flood of get_user_pages_fast and it can support multiple granularity in the secondary MMU mappings, so I think it is justified to be exposed not just to KVM. The other option would be to move transparent_hugepage_adjust to mm/huge_memory.c but that currently has all kind of KVM data structures in it, so it's definitely not a cut-and-paste work, so I couldn't do a fix as cleaner as this one for 4.6. Signed-off-by: Andrea Arcangeli Cc: "Dr. David Alan Gilbert" Cc: "Kirill A. Shutemov" Cc: "Li, Liang Z" Cc: Amit Shah Cc: Paolo Bonzini Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c index 58dbd5c439df..d6d4191e68f2 100644 --- a/arch/arm/kvm/mmu.c +++ b/arch/arm/kvm/mmu.c @@ -1004,7 +1004,7 @@ static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap) kvm_pfn_t pfn = *pfnp; gfn_t gfn = *ipap >> PAGE_SHIFT; - if (PageTransCompound(pfn_to_page(pfn))) { + if (PageTransCompoundMap(pfn_to_page(pfn))) { unsigned long mask; /* * The address we faulted on is backed by a transparent huge diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c index 1ff4dbb73fb7..b6f50e8b0a39 100644 --- a/arch/x86/kvm/mmu.c +++ b/arch/x86/kvm/mmu.c @@ -2823,7 +2823,7 @@ static void transparent_hugepage_adjust(struct kvm_vcpu *vcpu, */ if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn) && level == PT_PAGE_TABLE_LEVEL && - PageTransCompound(pfn_to_page(pfn)) && + PageTransCompoundMap(pfn_to_page(pfn)) && !mmu_gfn_lpage_is_disallowed(vcpu, gfn, PT_DIRECTORY_LEVEL)) { unsigned long mask; /* @@ -4785,7 +4785,7 @@ restart: */ if (sp->role.direct && !kvm_is_reserved_pfn(pfn) && - PageTransCompound(pfn_to_page(pfn))) { + PageTransCompoundMap(pfn_to_page(pfn))) { drop_spte(kvm, sptep); need_tlb_flush = 1; goto restart; diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h index f4ed4f1b0c77..6b052aa7b5b7 100644 --- a/include/linux/page-flags.h +++ b/include/linux/page-flags.h @@ -516,6 +516,27 @@ static inline int PageTransCompound(struct page *page) return PageCompound(page); } +/* + * PageTransCompoundMap is the same as PageTransCompound, but it also + * guarantees the primary MMU has the entire compound page mapped + * through pmd_trans_huge, which in turn guarantees the secondary MMUs + * can also map the entire compound page. This allows the secondary + * MMUs to call get_user_pages() only once for each compound page and + * to immediately map the entire compound page with a single secondary + * MMU fault. If there will be a pmd split later, the secondary MMUs + * will get an update through the MMU notifier invalidation through + * split_huge_pmd(). + * + * Unlike PageTransCompound, this is safe to be called only while + * split_huge_pmd() cannot run from under us, like if protected by the + * MMU notifier, otherwise it may result in page->_mapcount < 0 false + * positives. + */ +static inline int PageTransCompoundMap(struct page *page) +{ + return PageTransCompound(page) && atomic_read(&page->_mapcount) < 0; +} + /* * PageTransTail returns true for both transparent huge pages * and hugetlbfs pages, so it should only be called when it's known @@ -559,6 +580,7 @@ static inline int TestClearPageDoubleMap(struct page *page) #else TESTPAGEFLAG_FALSE(TransHuge) TESTPAGEFLAG_FALSE(TransCompound) +TESTPAGEFLAG_FALSE(TransCompoundMap) TESTPAGEFLAG_FALSE(TransTail) TESTPAGEFLAG_FALSE(DoubleMap) TESTSETFLAG_FALSE(DoubleMap)