| Message ID | 21ea9ca-a1f5-8b90-5e88-95fb1c49bbfa@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | mm/thp: fix THP splitting unmap BUGs and related (fwd) | expand |
On Tue, Jun 08, 2021 at 09:08:09PM -0700, Hugh Dickins wrote: > Most callers of is_huge_zero_pmd() supply a pmd already verified present; > but a few (notably zap_huge_pmd()) do not - it might be a pmd migration > entry, in which the pfn is encoded differently from a present pmd: which > might pass the is_huge_zero_pmd() test (though not on x86, since L1TF > forced us to protect against that); or perhaps even crash in pmd_page() > applied to a swap-like entry. > > Make it safe by adding pmd_present() check into is_huge_zero_pmd() itself; > and make it quicker by saving huge_zero_pfn, so that is_huge_zero_pmd() > will not need to do that pmd_page() lookup each time. > > __split_huge_pmd_locked() checked pmd_trans_huge() before: that worked, > but is unnecessary now that is_huge_zero_pmd() checks present. > > Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp") > Signed-off-by: Hugh Dickins <hughd@google.com> > Cc: <stable@vger.kernel.org> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
On Tue, Jun 8, 2021 at 9:08 PM Hugh Dickins <hughd@google.com> wrote: > > Most callers of is_huge_zero_pmd() supply a pmd already verified present; > but a few (notably zap_huge_pmd()) do not - it might be a pmd migration > entry, in which the pfn is encoded differently from a present pmd: which > might pass the is_huge_zero_pmd() test (though not on x86, since L1TF > forced us to protect against that); or perhaps even crash in pmd_page() > applied to a swap-like entry. > > Make it safe by adding pmd_present() check into is_huge_zero_pmd() itself; > and make it quicker by saving huge_zero_pfn, so that is_huge_zero_pmd() > will not need to do that pmd_page() lookup each time. > > __split_huge_pmd_locked() checked pmd_trans_huge() before: that worked, > but is unnecessary now that is_huge_zero_pmd() checks present. > > Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp") > Signed-off-by: Hugh Dickins <hughd@google.com> > Cc: <stable@vger.kernel.org> Reviewed-by: Yang Shi <shy828301@gmail.com> > --- > Patch added (replacing part of first) since the v1 series was posted. > > include/linux/huge_mm.h | 8 +++++++- > mm/huge_memory.c | 5 ++++- > 2 files changed, 11 insertions(+), 2 deletions(-) > > diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h > index 9626fda5efce..2a8ebe6c222e 100644 > --- a/include/linux/huge_mm.h > +++ b/include/linux/huge_mm.h > @@ -286,6 +286,7 @@ struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, > vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd); > > extern struct page *huge_zero_page; > +extern unsigned long huge_zero_pfn; > > static inline bool is_huge_zero_page(struct page *page) > { > @@ -294,7 +295,7 @@ static inline bool is_huge_zero_page(struct page *page) > > static inline bool is_huge_zero_pmd(pmd_t pmd) > { > - return is_huge_zero_page(pmd_page(pmd)); > + return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd); > } > > static inline bool is_huge_zero_pud(pud_t pud) > @@ -440,6 +441,11 @@ static inline bool is_huge_zero_page(struct page *page) > return false; > } > > +static inline bool is_huge_zero_pmd(pmd_t pmd) > +{ > + return false; > +} > + > static inline bool is_huge_zero_pud(pud_t pud) > { > return false; > diff --git a/mm/huge_memory.c b/mm/huge_memory.c > index 42cfefc6e66e..5885c5f5836f 100644 > --- a/mm/huge_memory.c > +++ b/mm/huge_memory.c > @@ -62,6 +62,7 @@ static struct shrinker deferred_split_shrinker; > > static atomic_t huge_zero_refcount; > struct page *huge_zero_page __read_mostly; > +unsigned long huge_zero_pfn __read_mostly = ~0UL; > > bool transparent_hugepage_enabled(struct vm_area_struct *vma) > { > @@ -98,6 +99,7 @@ static bool get_huge_zero_page(void) > __free_pages(zero_page, compound_order(zero_page)); > goto retry; > } > + WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); > > /* We take additional reference here. It will be put back by shrinker */ > atomic_set(&huge_zero_refcount, 2); > @@ -147,6 +149,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, > if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { > struct page *zero_page = xchg(&huge_zero_page, NULL); > BUG_ON(zero_page == NULL); > + WRITE_ONCE(huge_zero_pfn, ~0UL); > __free_pages(zero_page, compound_order(zero_page)); > return HPAGE_PMD_NR; > } > @@ -2071,7 +2074,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, > return; > } > > - if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) { > + if (is_huge_zero_pmd(*pmd)) { > /* > * FIXME: Do we want to invalidate secondary mmu by calling > * mmu_notifier_invalidate_range() see comments below inside > -- > 2.26.2 >
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h index 9626fda5efce..2a8ebe6c222e 100644 --- a/include/linux/huge_mm.h +++ b/include/linux/huge_mm.h @@ -286,6 +286,7 @@ struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd); extern struct page *huge_zero_page; +extern unsigned long huge_zero_pfn; static inline bool is_huge_zero_page(struct page *page) { @@ -294,7 +295,7 @@ static inline bool is_huge_zero_page(struct page *page) static inline bool is_huge_zero_pmd(pmd_t pmd) { - return is_huge_zero_page(pmd_page(pmd)); + return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd); } static inline bool is_huge_zero_pud(pud_t pud) @@ -440,6 +441,11 @@ static inline bool is_huge_zero_page(struct page *page) return false; } +static inline bool is_huge_zero_pmd(pmd_t pmd) +{ + return false; +} + static inline bool is_huge_zero_pud(pud_t pud) { return false; diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 42cfefc6e66e..5885c5f5836f 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -62,6 +62,7 @@ static struct shrinker deferred_split_shrinker; static atomic_t huge_zero_refcount; struct page *huge_zero_page __read_mostly; +unsigned long huge_zero_pfn __read_mostly = ~0UL; bool transparent_hugepage_enabled(struct vm_area_struct *vma) { @@ -98,6 +99,7 @@ static bool get_huge_zero_page(void) __free_pages(zero_page, compound_order(zero_page)); goto retry; } + WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); /* We take additional reference here. It will be put back by shrinker */ atomic_set(&huge_zero_refcount, 2); @@ -147,6 +149,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { struct page *zero_page = xchg(&huge_zero_page, NULL); BUG_ON(zero_page == NULL); + WRITE_ONCE(huge_zero_pfn, ~0UL); __free_pages(zero_page, compound_order(zero_page)); return HPAGE_PMD_NR; } @@ -2071,7 +2074,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, return; } - if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) { + if (is_huge_zero_pmd(*pmd)) { /* * FIXME: Do we want to invalidate secondary mmu by calling * mmu_notifier_invalidate_range() see comments below inside
Most callers of is_huge_zero_pmd() supply a pmd already verified present; but a few (notably zap_huge_pmd()) do not - it might be a pmd migration entry, in which the pfn is encoded differently from a present pmd: which might pass the is_huge_zero_pmd() test (though not on x86, since L1TF forced us to protect against that); or perhaps even crash in pmd_page() applied to a swap-like entry. Make it safe by adding pmd_present() check into is_huge_zero_pmd() itself; and make it quicker by saving huge_zero_pfn, so that is_huge_zero_pmd() will not need to do that pmd_page() lookup each time. __split_huge_pmd_locked() checked pmd_trans_huge() before: that worked, but is unnecessary now that is_huge_zero_pmd() checks present. Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp") Signed-off-by: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> --- Patch added (replacing part of first) since the v1 series was posted. include/linux/huge_mm.h | 8 +++++++- mm/huge_memory.c | 5 ++++- 2 files changed, 11 insertions(+), 2 deletions(-)