| Message ID | 20221205234059.42971-2-jiaqiyan@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Memory poison recovery in khugepaged collapsing | expand |
On Mon, Dec 05, 2022 at 03:40:58PM -0800, Jiaqi Yan wrote: > Make __collapse_huge_page_copy return whether copying anonymous pages > succeeded, and make collapse_huge_page handle the return status. > > Break existing PTE scan loop into two for-loops. The first loop copies > source pages into target huge page, and can fail gracefully when running > into memory errors in source pages. If copying all pages succeeds, the > second loop releases and clears up these normal pages. Otherwise, the > second loop rolls back the page table and page states by: > - re-establishing the original PTEs-to-PMD connection. > - releasing source pages back to their LRU list. > > Tested manually: > 0. Enable khugepaged on system under test. > 1. Start a two-thread application. Each thread allocates a chunk of > non-huge anonymous memory buffer. > 2. Pick 4 random buffer locations (2 in each thread) and inject > uncorrectable memory errors at corresponding physical addresses. > 3. Signal both threads to make their memory buffer collapsible, i.e. > calling madvise(MADV_HUGEPAGE). > 4. Wait and check kernel log: khugepaged is able to recover from poisoned > pages and skips collapsing them. > 5. Signal both threads to inspect their buffer contents and make sure no > data corruption. > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > --- > include/trace/events/huge_memory.h | 3 +- > mm/khugepaged.c | 179 ++++++++++++++++++++++------- > 2 files changed, 139 insertions(+), 43 deletions(-) > > diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h > index 35d759d3b0104..5743ae970af31 100644 > --- a/include/trace/events/huge_memory.h > +++ b/include/trace/events/huge_memory.h > @@ -36,7 +36,8 @@ > EM( SCAN_ALLOC_HUGE_PAGE_FAIL, "alloc_huge_page_failed") \ > EM( SCAN_CGROUP_CHARGE_FAIL, "ccgroup_charge_failed") \ > EM( SCAN_TRUNCATED, "truncated") \ > - EMe(SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > + EM( SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > + EMe(SCAN_COPY_MC, "copy_poisoned_page") \ > > #undef EM > #undef EMe > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > index 5a7d2d5093f9c..0f1b9e05e17ec 100644 > --- a/mm/khugepaged.c > +++ b/mm/khugepaged.c > @@ -19,6 +19,7 @@ > #include <linux/page_table_check.h> > #include <linux/swapops.h> > #include <linux/shmem_fs.h> > +#include <linux/kmsan.h> > > #include <asm/tlb.h> > #include <asm/pgalloc.h> > @@ -55,6 +56,7 @@ enum scan_result { > SCAN_CGROUP_CHARGE_FAIL, > SCAN_TRUNCATED, > SCAN_PAGE_HAS_PRIVATE, > + SCAN_COPY_MC, > }; > > #define CREATE_TRACE_POINTS > @@ -530,6 +532,27 @@ static bool is_refcount_suitable(struct page *page) > return page_count(page) == expected_refcount; > } > > +/* > + * Copies memory with #MC in source page (@from) handled. Returns number > + * of bytes not copied if there was an exception; otherwise 0 for success. > + * Note handling #MC requires arch opt-in. > + */ > +static int copy_mc_page(struct page *to, struct page *from) > +{ > + char *vfrom, *vto; > + unsigned long ret; > + > + vfrom = kmap_local_page(from); > + vto = kmap_local_page(to); > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > + if (ret == 0) > + kmsan_copy_page_meta(to, from); > + kunmap_local(vto); > + kunmap_local(vfrom); > + > + return ret; > +} It is very similar to copy_mc_user_highpage(), but uses kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). Could you explain the difference? I don't quite get it. > + > static int __collapse_huge_page_isolate(struct vm_area_struct *vma, > unsigned long address, > pte_t *pte, > @@ -670,56 +693,124 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma, > return result; > } > > -static void __collapse_huge_page_copy(pte_t *pte, struct page *page, > - struct vm_area_struct *vma, > - unsigned long address, > - spinlock_t *ptl, > - struct list_head *compound_pagelist) > +/* > + * __collapse_huge_page_copy - attempts to copy memory contents from normal > + * pages to a hugepage. Cleans up the normal pages if copying succeeds; > + * otherwise restores the original page table and releases isolated normal pages. > + * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. > + * > + * @pte: starting of the PTEs to copy from > + * @page: the new hugepage to copy contents to > + * @pmd: pointer to the new hugepage's PMD > + * @rollback: the original normal pages' PMD > + * @vma: the original normal pages' virtual memory area > + * @address: starting address to copy > + * @pte_ptl: lock on normal pages' PTEs > + * @compound_pagelist: list that stores compound pages > + */ > +static int __collapse_huge_page_copy(pte_t *pte, > + struct page *page, > + pmd_t *pmd, > + pmd_t rollback, I think 'orig_pmd' is a better name. > + struct vm_area_struct *vma, > + unsigned long address, > + spinlock_t *pte_ptl, > + struct list_head *compound_pagelist) > { > struct page *src_page, *tmp; > pte_t *_pte; > - for (_pte = pte; _pte < pte + HPAGE_PMD_NR; > - _pte++, page++, address += PAGE_SIZE) { > - pte_t pteval = *_pte; > + pte_t pteval; > + unsigned long _address; > + spinlock_t *pmd_ptl; > + int result = SCAN_SUCCEED; > > - if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { > - clear_user_highpage(page, address); > - add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); > - if (is_zero_pfn(pte_pfn(pteval))) { > + /* > + * Copying pages' contents is subject to memory poison at any iteration. > + */ > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > + _pte++, page++, _address += PAGE_SIZE) { > + pteval = *_pte; > + > + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) > + clear_user_highpage(page, _address); > + else { > + src_page = pte_page(pteval); > + if (copy_mc_page(page, src_page) > 0) { > + result = SCAN_COPY_MC; > + break; > + } > + } > + } > + > + if (likely(result == SCAN_SUCCEED)) { > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > + _pte++, _address += PAGE_SIZE) { > + pteval = *_pte; > + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { > + add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); > + if (is_zero_pfn(pte_pfn(pteval))) { > + /* > + * pte_ptl mostly unnecessary. > + */ > + spin_lock(pte_ptl); > + pte_clear(vma->vm_mm, _address, _pte); > + spin_unlock(pte_ptl); > + } > + } else { > + src_page = pte_page(pteval); > + if (!PageCompound(src_page)) > + release_pte_page(src_page); > /* > - * ptl mostly unnecessary. > + * pte_ptl mostly unnecessary, but preempt has > + * to be disabled to update the per-cpu stats > + * inside page_remove_rmap(). > */ > - spin_lock(ptl); > - ptep_clear(vma->vm_mm, address, _pte); > - spin_unlock(ptl); > + spin_lock(pte_ptl); > + ptep_clear(vma->vm_mm, _address, _pte); > + page_remove_rmap(src_page, vma, false); > + spin_unlock(pte_ptl); > + free_page_and_swap_cache(src_page); > + } > + } > + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > + list_del(&src_page->lru); > + mod_node_page_state(page_pgdat(src_page), > + NR_ISOLATED_ANON + page_is_file_lru(src_page), > + -compound_nr(src_page)); > + unlock_page(src_page); > + free_swap_cache(src_page); > + putback_lru_page(src_page); > + } > + } else { > + /* > + * Re-establish the regular PMD that points to the regular > + * page table. Restoring PMD needs to be done prior to > + * releasing pages. Since pages are still isolated and > + * locked here, acquiring anon_vma_lock_write is unnecessary. > + */ > + pmd_ptl = pmd_lock(vma->vm_mm, pmd); > + pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback)); > + spin_unlock(pmd_ptl); > + /* > + * Release both raw and compound pages isolated > + * in __collapse_huge_page_isolate. > + */ > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > + _pte++, _address += PAGE_SIZE) { > + pteval = *_pte; > + if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) { > + src_page = pte_page(pteval); > + if (!PageCompound(src_page)) > + release_pte_page(src_page); Indentation levels get out of control. Maybe some code restructuring is required? > } > - } else { > - src_page = pte_page(pteval); > - copy_user_highpage(page, src_page, address, vma); > - if (!PageCompound(src_page)) > - release_pte_page(src_page); > - /* > - * ptl mostly unnecessary, but preempt has to > - * be disabled to update the per-cpu stats > - * inside page_remove_rmap(). > - */ > - spin_lock(ptl); > - ptep_clear(vma->vm_mm, address, _pte); > - page_remove_rmap(src_page, vma, false); > - spin_unlock(ptl); > - free_page_and_swap_cache(src_page); > + } > + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > + list_del(&src_page->lru); > + release_pte_page(src_page); > } > } > > - list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > - list_del(&src_page->lru); > - mod_node_page_state(page_pgdat(src_page), > - NR_ISOLATED_ANON + page_is_file_lru(src_page), > - -compound_nr(src_page)); > - unlock_page(src_page); > - free_swap_cache(src_page); > - putback_lru_page(src_page); > - } > + return result; > } > > static void khugepaged_alloc_sleep(void) > @@ -1079,9 +1170,13 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, > */ > anon_vma_unlock_write(vma->anon_vma); > > - __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, > - &compound_pagelist); > + result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, > + vma, address, pte_ptl, > + &compound_pagelist); > pte_unmap(pte); > + if (unlikely(result != SCAN_SUCCEED)) > + goto out_up_write; > + > /* > * spin_lock() below is not the equivalent of smp_wmb(), but > * the smp_wmb() inside __SetPageUptodate() can be reused to > -- > 2.39.0.rc0.267.gcb52ba06e7-goog >
On Thu, Jan 19, 2023 at 7:03 AM <kirill.shutemov@linux.intel.com> wrote: > > On Mon, Dec 05, 2022 at 03:40:58PM -0800, Jiaqi Yan wrote: > > Make __collapse_huge_page_copy return whether copying anonymous pages > > succeeded, and make collapse_huge_page handle the return status. > > > > Break existing PTE scan loop into two for-loops. The first loop copies > > source pages into target huge page, and can fail gracefully when running > > into memory errors in source pages. If copying all pages succeeds, the > > second loop releases and clears up these normal pages. Otherwise, the > > second loop rolls back the page table and page states by: > > - re-establishing the original PTEs-to-PMD connection. > > - releasing source pages back to their LRU list. > > > > Tested manually: > > 0. Enable khugepaged on system under test. > > 1. Start a two-thread application. Each thread allocates a chunk of > > non-huge anonymous memory buffer. > > 2. Pick 4 random buffer locations (2 in each thread) and inject > > uncorrectable memory errors at corresponding physical addresses. > > 3. Signal both threads to make their memory buffer collapsible, i.e. > > calling madvise(MADV_HUGEPAGE). > > 4. Wait and check kernel log: khugepaged is able to recover from poisoned > > pages and skips collapsing them. > > 5. Signal both threads to inspect their buffer contents and make sure no > > data corruption. > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > > --- > > include/trace/events/huge_memory.h | 3 +- > > mm/khugepaged.c | 179 ++++++++++++++++++++++------- > > 2 files changed, 139 insertions(+), 43 deletions(-) > > > > diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h > > index 35d759d3b0104..5743ae970af31 100644 > > --- a/include/trace/events/huge_memory.h > > +++ b/include/trace/events/huge_memory.h > > @@ -36,7 +36,8 @@ > > EM( SCAN_ALLOC_HUGE_PAGE_FAIL, "alloc_huge_page_failed") \ > > EM( SCAN_CGROUP_CHARGE_FAIL, "ccgroup_charge_failed") \ > > EM( SCAN_TRUNCATED, "truncated") \ > > - EMe(SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > + EM( SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > + EMe(SCAN_COPY_MC, "copy_poisoned_page") \ > > > > #undef EM > > #undef EMe > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > index 5a7d2d5093f9c..0f1b9e05e17ec 100644 > > --- a/mm/khugepaged.c > > +++ b/mm/khugepaged.c > > @@ -19,6 +19,7 @@ > > #include <linux/page_table_check.h> > > #include <linux/swapops.h> > > #include <linux/shmem_fs.h> > > +#include <linux/kmsan.h> > > > > #include <asm/tlb.h> > > #include <asm/pgalloc.h> > > @@ -55,6 +56,7 @@ enum scan_result { > > SCAN_CGROUP_CHARGE_FAIL, > > SCAN_TRUNCATED, > > SCAN_PAGE_HAS_PRIVATE, > > + SCAN_COPY_MC, > > }; > > > > #define CREATE_TRACE_POINTS > > @@ -530,6 +532,27 @@ static bool is_refcount_suitable(struct page *page) > > return page_count(page) == expected_refcount; > > } > > > > +/* > > + * Copies memory with #MC in source page (@from) handled. Returns number > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > + * Note handling #MC requires arch opt-in. > > + */ > > +static int copy_mc_page(struct page *to, struct page *from) > > +{ > > + char *vfrom, *vto; > > + unsigned long ret; > > + > > + vfrom = kmap_local_page(from); > > + vto = kmap_local_page(to); > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > + if (ret == 0) > > + kmsan_copy_page_meta(to, from); > > + kunmap_local(vto); > > + kunmap_local(vfrom); > > + > > + return ret; > > +} > > > It is very similar to copy_mc_user_highpage(), but uses > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > Could you explain the difference? I don't quite get it. copy_mc_page is actually the MC version of copy_highpage, which uses kmsan_copy_page_meta instead of kmsan_unpoison_memory. My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. When there is no metadata (kmsan_shadow or kmsan_origin), both kmsan_copy_page_meta and kmsan_unpoison_memory just do kmsan_internal_unpoison_memory to mark the memory range as initialized; when there is metadata in src page, kmsan_copy_page_meta will copy whatever metadata in src to dst. So I think kmsan_copy_page_meta is the right thing to do. > > > + > > static int __collapse_huge_page_isolate(struct vm_area_struct *vma, > > unsigned long address, > > pte_t *pte, > > @@ -670,56 +693,124 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma, > > return result; > > } > > > > -static void __collapse_huge_page_copy(pte_t *pte, struct page *page, > > - struct vm_area_struct *vma, > > - unsigned long address, > > - spinlock_t *ptl, > > - struct list_head *compound_pagelist) > > +/* > > + * __collapse_huge_page_copy - attempts to copy memory contents from normal > > + * pages to a hugepage. Cleans up the normal pages if copying succeeds; > > + * otherwise restores the original page table and releases isolated normal pages. > > + * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. > > + * > > + * @pte: starting of the PTEs to copy from > > + * @page: the new hugepage to copy contents to > > + * @pmd: pointer to the new hugepage's PMD > > + * @rollback: the original normal pages' PMD > > + * @vma: the original normal pages' virtual memory area > > + * @address: starting address to copy > > + * @pte_ptl: lock on normal pages' PTEs > > + * @compound_pagelist: list that stores compound pages > > + */ > > +static int __collapse_huge_page_copy(pte_t *pte, > > + struct page *page, > > + pmd_t *pmd, > > + pmd_t rollback, > > I think 'orig_pmd' is a better name. Will be renamed to orig_pmd in the next version v10. > > > + struct vm_area_struct *vma, > > + unsigned long address, > > + spinlock_t *pte_ptl, > > + struct list_head *compound_pagelist) > > { > > struct page *src_page, *tmp; > > pte_t *_pte; > > - for (_pte = pte; _pte < pte + HPAGE_PMD_NR; > > - _pte++, page++, address += PAGE_SIZE) { > > - pte_t pteval = *_pte; > > + pte_t pteval; > > + unsigned long _address; > > + spinlock_t *pmd_ptl; > > + int result = SCAN_SUCCEED; > > > > - if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { > > - clear_user_highpage(page, address); > > - add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); > > - if (is_zero_pfn(pte_pfn(pteval))) { > > + /* > > + * Copying pages' contents is subject to memory poison at any iteration. > > + */ > > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > > + _pte++, page++, _address += PAGE_SIZE) { > > + pteval = *_pte; > > + > > + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) > > + clear_user_highpage(page, _address); > > + else { > > + src_page = pte_page(pteval); > > + if (copy_mc_page(page, src_page) > 0) { > > + result = SCAN_COPY_MC; > > + break; > > + } > > + } > > + } > > + > > + if (likely(result == SCAN_SUCCEED)) { > > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > > + _pte++, _address += PAGE_SIZE) { > > + pteval = *_pte; > > + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { > > + add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); > > + if (is_zero_pfn(pte_pfn(pteval))) { > > + /* > > + * pte_ptl mostly unnecessary. > > + */ > > + spin_lock(pte_ptl); > > + pte_clear(vma->vm_mm, _address, _pte); > > + spin_unlock(pte_ptl); > > + } > > + } else { > > + src_page = pte_page(pteval); > > + if (!PageCompound(src_page)) > > + release_pte_page(src_page); > > /* > > - * ptl mostly unnecessary. > > + * pte_ptl mostly unnecessary, but preempt has > > + * to be disabled to update the per-cpu stats > > + * inside page_remove_rmap(). > > */ > > - spin_lock(ptl); > > - ptep_clear(vma->vm_mm, address, _pte); > > - spin_unlock(ptl); > > + spin_lock(pte_ptl); > > + ptep_clear(vma->vm_mm, _address, _pte); > > + page_remove_rmap(src_page, vma, false); > > + spin_unlock(pte_ptl); > > + free_page_and_swap_cache(src_page); > > + } > > + } > > + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > > + list_del(&src_page->lru); > > + mod_node_page_state(page_pgdat(src_page), > > + NR_ISOLATED_ANON + page_is_file_lru(src_page), > > + -compound_nr(src_page)); > > + unlock_page(src_page); > > + free_swap_cache(src_page); > > + putback_lru_page(src_page); > > + } > > + } else { > > + /* > > + * Re-establish the regular PMD that points to the regular > > + * page table. Restoring PMD needs to be done prior to > > + * releasing pages. Since pages are still isolated and > > + * locked here, acquiring anon_vma_lock_write is unnecessary. > > + */ > > + pmd_ptl = pmd_lock(vma->vm_mm, pmd); > > + pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback)); > > + spin_unlock(pmd_ptl); > > + /* > > + * Release both raw and compound pages isolated > > + * in __collapse_huge_page_isolate. > > + */ > > + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; > > + _pte++, _address += PAGE_SIZE) { > > + pteval = *_pte; > > + if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) { > > + src_page = pte_page(pteval); > > + if (!PageCompound(src_page)) > > + release_pte_page(src_page); > > Indentation levels get out of control. Maybe some code restructuring is > required? v10 will change to something like this to reduce 1 level of indentation: if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) continue; src_page = pte_page(pteval); if (!PageCompound(src_page)) release_pte_page(src_page); > > > } > > - } else { > > - src_page = pte_page(pteval); > > - copy_user_highpage(page, src_page, address, vma); > > - if (!PageCompound(src_page)) > > - release_pte_page(src_page); > > - /* > > - * ptl mostly unnecessary, but preempt has to > > - * be disabled to update the per-cpu stats > > - * inside page_remove_rmap(). > > - */ > > - spin_lock(ptl); > > - ptep_clear(vma->vm_mm, address, _pte); > > - page_remove_rmap(src_page, vma, false); > > - spin_unlock(ptl); > > - free_page_and_swap_cache(src_page); > > + } > > + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > > + list_del(&src_page->lru); > > + release_pte_page(src_page); > > } > > } > > > > - list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { > > - list_del(&src_page->lru); > > - mod_node_page_state(page_pgdat(src_page), > > - NR_ISOLATED_ANON + page_is_file_lru(src_page), > > - -compound_nr(src_page)); > > - unlock_page(src_page); > > - free_swap_cache(src_page); > > - putback_lru_page(src_page); > > - } > > + return result; > > } > > > > static void khugepaged_alloc_sleep(void) > > @@ -1079,9 +1170,13 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, > > */ > > anon_vma_unlock_write(vma->anon_vma); > > > > - __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, > > - &compound_pagelist); > > + result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, > > + vma, address, pte_ptl, > > + &compound_pagelist); > > pte_unmap(pte); > > + if (unlikely(result != SCAN_SUCCEED)) > > + goto out_up_write; > > + > > /* > > * spin_lock() below is not the equivalent of smp_wmb(), but > > * the smp_wmb() inside __SetPageUptodate() can be reused to > > -- > > 2.39.0.rc0.267.gcb52ba06e7-goog > > > > -- > Kiryl Shutsemau / Kirill A. Shutemov
On Fri, Jan 20, 2023 at 07:56:15AM -0800, Jiaqi Yan wrote: > On Thu, Jan 19, 2023 at 7:03 AM <kirill.shutemov@linux.intel.com> wrote: > > > > On Mon, Dec 05, 2022 at 03:40:58PM -0800, Jiaqi Yan wrote: > > > Make __collapse_huge_page_copy return whether copying anonymous pages > > > succeeded, and make collapse_huge_page handle the return status. > > > > > > Break existing PTE scan loop into two for-loops. The first loop copies > > > source pages into target huge page, and can fail gracefully when running > > > into memory errors in source pages. If copying all pages succeeds, the > > > second loop releases and clears up these normal pages. Otherwise, the > > > second loop rolls back the page table and page states by: > > > - re-establishing the original PTEs-to-PMD connection. > > > - releasing source pages back to their LRU list. > > > > > > Tested manually: > > > 0. Enable khugepaged on system under test. > > > 1. Start a two-thread application. Each thread allocates a chunk of > > > non-huge anonymous memory buffer. > > > 2. Pick 4 random buffer locations (2 in each thread) and inject > > > uncorrectable memory errors at corresponding physical addresses. > > > 3. Signal both threads to make their memory buffer collapsible, i.e. > > > calling madvise(MADV_HUGEPAGE). > > > 4. Wait and check kernel log: khugepaged is able to recover from poisoned > > > pages and skips collapsing them. > > > 5. Signal both threads to inspect their buffer contents and make sure no > > > data corruption. > > > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > > > --- > > > include/trace/events/huge_memory.h | 3 +- > > > mm/khugepaged.c | 179 ++++++++++++++++++++++------- > > > 2 files changed, 139 insertions(+), 43 deletions(-) > > > > > > diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h > > > index 35d759d3b0104..5743ae970af31 100644 > > > --- a/include/trace/events/huge_memory.h > > > +++ b/include/trace/events/huge_memory.h > > > @@ -36,7 +36,8 @@ > > > EM( SCAN_ALLOC_HUGE_PAGE_FAIL, "alloc_huge_page_failed") \ > > > EM( SCAN_CGROUP_CHARGE_FAIL, "ccgroup_charge_failed") \ > > > EM( SCAN_TRUNCATED, "truncated") \ > > > - EMe(SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > > + EM( SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > > + EMe(SCAN_COPY_MC, "copy_poisoned_page") \ > > > > > > #undef EM > > > #undef EMe > > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > > index 5a7d2d5093f9c..0f1b9e05e17ec 100644 > > > --- a/mm/khugepaged.c > > > +++ b/mm/khugepaged.c > > > @@ -19,6 +19,7 @@ > > > #include <linux/page_table_check.h> > > > #include <linux/swapops.h> > > > #include <linux/shmem_fs.h> > > > +#include <linux/kmsan.h> > > > > > > #include <asm/tlb.h> > > > #include <asm/pgalloc.h> > > > @@ -55,6 +56,7 @@ enum scan_result { > > > SCAN_CGROUP_CHARGE_FAIL, > > > SCAN_TRUNCATED, > > > SCAN_PAGE_HAS_PRIVATE, > > > + SCAN_COPY_MC, > > > }; > > > > > > #define CREATE_TRACE_POINTS > > > @@ -530,6 +532,27 @@ static bool is_refcount_suitable(struct page *page) > > > return page_count(page) == expected_refcount; > > > } > > > > > > +/* > > > + * Copies memory with #MC in source page (@from) handled. Returns number > > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > > + * Note handling #MC requires arch opt-in. > > > + */ > > > +static int copy_mc_page(struct page *to, struct page *from) > > > +{ > > > + char *vfrom, *vto; > > > + unsigned long ret; > > > + > > > + vfrom = kmap_local_page(from); > > > + vto = kmap_local_page(to); > > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > > + if (ret == 0) > > > + kmsan_copy_page_meta(to, from); > > > + kunmap_local(vto); > > > + kunmap_local(vfrom); > > > + > > > + return ret; > > > +} > > > > > > It is very similar to copy_mc_user_highpage(), but uses > > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > > > Could you explain the difference? I don't quite get it. > > copy_mc_page is actually the MC version of copy_highpage, which uses > kmsan_copy_page_meta instead of kmsan_unpoison_memory. > > My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. > When there is no metadata (kmsan_shadow or kmsan_origin), both > kmsan_copy_page_meta and kmsan_unpoison_memory just do > kmsan_internal_unpoison_memory to mark the memory range as > initialized; when there is metadata in src page, kmsan_copy_page_meta > will copy whatever metadata in src to dst. So I think > kmsan_copy_page_meta is the right thing to do. Should we fix copy_mc_user_highpage() then? > > Indentation levels get out of control. Maybe some code restructuring is > > required? > > v10 will change to something like this to reduce 1 level of indentation: > > if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) > continue; > src_page = pte_page(pteval); > if (!PageCompound(src_page)) > release_pte_page(src_page); I hoped for deeper rework. Maybe split the function into several functions and make overall structure more readable?
On Mon, Jan 23, 2023 at 4:34 PM <kirill.shutemov@linux.intel.com> wrote: > > On Fri, Jan 20, 2023 at 07:56:15AM -0800, Jiaqi Yan wrote: > > On Thu, Jan 19, 2023 at 7:03 AM <kirill.shutemov@linux.intel.com> wrote: > > > > > > On Mon, Dec 05, 2022 at 03:40:58PM -0800, Jiaqi Yan wrote: > > > > Make __collapse_huge_page_copy return whether copying anonymous pages > > > > succeeded, and make collapse_huge_page handle the return status. > > > > > > > > Break existing PTE scan loop into two for-loops. The first loop copies > > > > source pages into target huge page, and can fail gracefully when running > > > > into memory errors in source pages. If copying all pages succeeds, the > > > > second loop releases and clears up these normal pages. Otherwise, the > > > > second loop rolls back the page table and page states by: > > > > - re-establishing the original PTEs-to-PMD connection. > > > > - releasing source pages back to their LRU list. > > > > > > > > Tested manually: > > > > 0. Enable khugepaged on system under test. > > > > 1. Start a two-thread application. Each thread allocates a chunk of > > > > non-huge anonymous memory buffer. > > > > 2. Pick 4 random buffer locations (2 in each thread) and inject > > > > uncorrectable memory errors at corresponding physical addresses. > > > > 3. Signal both threads to make their memory buffer collapsible, i.e. > > > > calling madvise(MADV_HUGEPAGE). > > > > 4. Wait and check kernel log: khugepaged is able to recover from poisoned > > > > pages and skips collapsing them. > > > > 5. Signal both threads to inspect their buffer contents and make sure no > > > > data corruption. > > > > > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> > > > > --- > > > > include/trace/events/huge_memory.h | 3 +- > > > > mm/khugepaged.c | 179 ++++++++++++++++++++++------- > > > > 2 files changed, 139 insertions(+), 43 deletions(-) > > > > > > > > diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h > > > > index 35d759d3b0104..5743ae970af31 100644 > > > > --- a/include/trace/events/huge_memory.h > > > > +++ b/include/trace/events/huge_memory.h > > > > @@ -36,7 +36,8 @@ > > > > EM( SCAN_ALLOC_HUGE_PAGE_FAIL, "alloc_huge_page_failed") \ > > > > EM( SCAN_CGROUP_CHARGE_FAIL, "ccgroup_charge_failed") \ > > > > EM( SCAN_TRUNCATED, "truncated") \ > > > > - EMe(SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > > > + EM( SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ > > > > + EMe(SCAN_COPY_MC, "copy_poisoned_page") \ > > > > > > > > #undef EM > > > > #undef EMe > > > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > > > index 5a7d2d5093f9c..0f1b9e05e17ec 100644 > > > > --- a/mm/khugepaged.c > > > > +++ b/mm/khugepaged.c > > > > @@ -19,6 +19,7 @@ > > > > #include <linux/page_table_check.h> > > > > #include <linux/swapops.h> > > > > #include <linux/shmem_fs.h> > > > > +#include <linux/kmsan.h> > > > > > > > > #include <asm/tlb.h> > > > > #include <asm/pgalloc.h> > > > > @@ -55,6 +56,7 @@ enum scan_result { > > > > SCAN_CGROUP_CHARGE_FAIL, > > > > SCAN_TRUNCATED, > > > > SCAN_PAGE_HAS_PRIVATE, > > > > + SCAN_COPY_MC, > > > > }; > > > > > > > > #define CREATE_TRACE_POINTS > > > > @@ -530,6 +532,27 @@ static bool is_refcount_suitable(struct page *page) > > > > return page_count(page) == expected_refcount; > > > > } > > > > > > > > +/* > > > > + * Copies memory with #MC in source page (@from) handled. Returns number > > > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > > > + * Note handling #MC requires arch opt-in. > > > > + */ > > > > +static int copy_mc_page(struct page *to, struct page *from) > > > > +{ > > > > + char *vfrom, *vto; > > > > + unsigned long ret; > > > > + > > > > + vfrom = kmap_local_page(from); > > > > + vto = kmap_local_page(to); > > > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > > > + if (ret == 0) > > > > + kmsan_copy_page_meta(to, from); > > > > + kunmap_local(vto); > > > > + kunmap_local(vfrom); > > > > + > > > > + return ret; > > > > +} > > > > > > > > > It is very similar to copy_mc_user_highpage(), but uses > > > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > > > > > Could you explain the difference? I don't quite get it. > > > > copy_mc_page is actually the MC version of copy_highpage, which uses > > kmsan_copy_page_meta instead of kmsan_unpoison_memory. > > > > My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. > > When there is no metadata (kmsan_shadow or kmsan_origin), both > > kmsan_copy_page_meta and kmsan_unpoison_memory just do > > kmsan_internal_unpoison_memory to mark the memory range as > > initialized; when there is metadata in src page, kmsan_copy_page_meta > > will copy whatever metadata in src to dst. So I think > > kmsan_copy_page_meta is the right thing to do. > > Should we fix copy_mc_user_highpage() then? I think it depends on what copy_user_highpage() (the original of copy_mc_user_highpage) is used for. copy_mc_user_highpage is currently only used by __wp_page_copy_user, is it possible that here we don't want to (or don't need to) copy page metadata for userspace pages? > > > > Indentation levels get out of control. Maybe some code restructuring is > > > required? > > > > v10 will change to something like this to reduce 1 level of indentation: > > > > if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) > > continue; > > src_page = pte_page(pteval); > > if (!PageCompound(src_page)) > > release_pte_page(src_page); > > I hoped for deeper rework. Maybe split the function into several functions > and make overall structure more readable? How about turning the 2nd loop into __collapse_huge_page_copy_succeeded and __collapse_huge_page_copy_failed, one for the case copy succeeded, and one for failed? Like this: if (likely(result == SCAN_SUCCEED)) __collapse_huge_page_copy_succeeded(...); else __collapse_huge_page_copy_failed(...); My prototype shows it could reduce the level indents. > > -- > Kiryl Shutsemau / Kirill A. Shutemov
On Tue, Jan 31, 2023 at 09:16:45PM -0800, Jiaqi Yan wrote: > > > > > +/* > > > > > + * Copies memory with #MC in source page (@from) handled. Returns number > > > > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > > > > + * Note handling #MC requires arch opt-in. > > > > > + */ > > > > > +static int copy_mc_page(struct page *to, struct page *from) > > > > > +{ > > > > > + char *vfrom, *vto; > > > > > + unsigned long ret; > > > > > + > > > > > + vfrom = kmap_local_page(from); > > > > > + vto = kmap_local_page(to); > > > > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > > > > + if (ret == 0) > > > > > + kmsan_copy_page_meta(to, from); > > > > > + kunmap_local(vto); > > > > > + kunmap_local(vfrom); > > > > > + > > > > > + return ret; > > > > > +} > > > > > > > > > > > > It is very similar to copy_mc_user_highpage(), but uses > > > > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > > > > > > > Could you explain the difference? I don't quite get it. > > > > > > copy_mc_page is actually the MC version of copy_highpage, which uses > > > kmsan_copy_page_meta instead of kmsan_unpoison_memory. > > > > > > My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. > > > When there is no metadata (kmsan_shadow or kmsan_origin), both > > > kmsan_copy_page_meta and kmsan_unpoison_memory just do > > > kmsan_internal_unpoison_memory to mark the memory range as > > > initialized; when there is metadata in src page, kmsan_copy_page_meta > > > will copy whatever metadata in src to dst. So I think > > > kmsan_copy_page_meta is the right thing to do. > > > > Should we fix copy_mc_user_highpage() then? > > I think it depends on what copy_user_highpage() (the original of > copy_mc_user_highpage) is used for. copy_mc_user_highpage is currently > only used by __wp_page_copy_user, is it possible that here we don't > want to (or don't need to) copy page metadata for userspace pages? Tony, could chime in on this? Can we modify copy_mc_user_highpage() to also use kmsan_copy_page_meta()? I don't really understand KMSAN here. > > > > > > > Indentation levels get out of control. Maybe some code restructuring is > > > > required? > > > > > > v10 will change to something like this to reduce 1 level of indentation: > > > > > > if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) > > > continue; > > > src_page = pte_page(pteval); > > > if (!PageCompound(src_page)) > > > release_pte_page(src_page); > > > > I hoped for deeper rework. Maybe split the function into several functions > > and make overall structure more readable? > > How about turning the 2nd loop into > __collapse_huge_page_copy_succeeded and > __collapse_huge_page_copy_failed, one for the case copy succeeded, and > one for failed? Like this: > if (likely(result == SCAN_SUCCEED)) > __collapse_huge_page_copy_succeeded(...); > else > __collapse_huge_page_copy_failed(...); > > My prototype shows it could reduce the level indents. Give it a try and try to get into reader shoes. Get it easily digestible for someone who reads the code for the first time.
On Thu, Feb 02, 2023 at 03:01:02AM +0300, kirill.shutemov@linux.intel.com wrote: > On Tue, Jan 31, 2023 at 09:16:45PM -0800, Jiaqi Yan wrote: > > > > > > +/* > > > > > > + * Copies memory with #MC in source page (@from) handled. Returns number > > > > > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > > > > > + * Note handling #MC requires arch opt-in. > > > > > > + */ > > > > > > +static int copy_mc_page(struct page *to, struct page *from) > > > > > > +{ > > > > > > + char *vfrom, *vto; > > > > > > + unsigned long ret; > > > > > > + > > > > > > + vfrom = kmap_local_page(from); > > > > > > + vto = kmap_local_page(to); > > > > > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > > > > > + if (ret == 0) > > > > > > + kmsan_copy_page_meta(to, from); > > > > > > + kunmap_local(vto); > > > > > > + kunmap_local(vfrom); > > > > > > + > > > > > > + return ret; > > > > > > +} > > > > > > > > > > > > > > > It is very similar to copy_mc_user_highpage(), but uses > > > > > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > > > > > > > > > Could you explain the difference? I don't quite get it. > > > > > > > > copy_mc_page is actually the MC version of copy_highpage, which uses > > > > kmsan_copy_page_meta instead of kmsan_unpoison_memory. > > > > > > > > My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. > > > > When there is no metadata (kmsan_shadow or kmsan_origin), both > > > > kmsan_copy_page_meta and kmsan_unpoison_memory just do > > > > kmsan_internal_unpoison_memory to mark the memory range as > > > > initialized; when there is metadata in src page, kmsan_copy_page_meta > > > > will copy whatever metadata in src to dst. So I think > > > > kmsan_copy_page_meta is the right thing to do. > > > > > > Should we fix copy_mc_user_highpage() then? > > > > I think it depends on what copy_user_highpage() (the original of > > copy_mc_user_highpage) is used for. copy_mc_user_highpage is currently > > only used by __wp_page_copy_user, is it possible that here we don't > > want to (or don't need to) copy page metadata for userspace pages? > > Tony, could chime in on this? Can we modify copy_mc_user_highpage() to > also use kmsan_copy_page_meta()? I don't really understand KMSAN here. +KMSAN folks.
Pinging KMSAN experts, for the general guidance of kmsan_copy_page_meta vs kmsan_unpoison_memory On Wed, Feb 1, 2023 at 4:30 PM <kirill@shutemov.name> wrote: > > On Thu, Feb 02, 2023 at 03:01:02AM +0300, kirill.shutemov@linux.intel.com wrote: > > On Tue, Jan 31, 2023 at 09:16:45PM -0800, Jiaqi Yan wrote: > > > > > > > +/* > > > > > > > + * Copies memory with #MC in source page (@from) handled. Returns number > > > > > > > + * of bytes not copied if there was an exception; otherwise 0 for success. > > > > > > > + * Note handling #MC requires arch opt-in. > > > > > > > + */ > > > > > > > +static int copy_mc_page(struct page *to, struct page *from) > > > > > > > +{ > > > > > > > + char *vfrom, *vto; > > > > > > > + unsigned long ret; > > > > > > > + > > > > > > > + vfrom = kmap_local_page(from); > > > > > > > + vto = kmap_local_page(to); > > > > > > > + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); > > > > > > > + if (ret == 0) > > > > > > > + kmsan_copy_page_meta(to, from); > > > > > > > + kunmap_local(vto); > > > > > > > + kunmap_local(vfrom); > > > > > > > + > > > > > > > + return ret; > > > > > > > +} > > > > > > > > > > > > > > > > > > It is very similar to copy_mc_user_highpage(), but uses > > > > > > kmsan_copy_page_meta() instead of kmsan_unpoison_memory(). > > > > > > > > > > > > Could you explain the difference? I don't quite get it. > > > > > > > > > > copy_mc_page is actually the MC version of copy_highpage, which uses > > > > > kmsan_copy_page_meta instead of kmsan_unpoison_memory. > > > > > > > > > > My understanding is kmsan_copy_page_meta covers kmsan_unpoison_memory. > > > > > When there is no metadata (kmsan_shadow or kmsan_origin), both > > > > > kmsan_copy_page_meta and kmsan_unpoison_memory just do > > > > > kmsan_internal_unpoison_memory to mark the memory range as > > > > > initialized; when there is metadata in src page, kmsan_copy_page_meta > > > > > will copy whatever metadata in src to dst. So I think > > > > > kmsan_copy_page_meta is the right thing to do. > > > > > > > > Should we fix copy_mc_user_highpage() then? > > > > > > I think it depends on what copy_user_highpage() (the original of > > > copy_mc_user_highpage) is used for. copy_mc_user_highpage is currently > > > only used by __wp_page_copy_user, is it possible that here we don't > > > want to (or don't need to) copy page metadata for userspace pages? > > > > Tony, could chime in on this? Can we modify copy_mc_user_highpage() to > > also use kmsan_copy_page_meta()? I don't really understand KMSAN here. > > +KMSAN folks. > > -- > Kiryl Shutsemau / Kirill A. Shutemov
On Tue, Feb 7, 2023 at 7:20 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > Pinging KMSAN experts, for the general guidance of > kmsan_copy_page_meta vs kmsan_unpoison_memory Oh, sorry, I've missed the previous email. copy_mc_user_highpage() is expected to copy data from the user page, for which no metadata is ever allocated. Therefore we just initialize the destination shadow with zeros instead of copying anything. kmsan_copy_page_meta() is used when the metadata is copied between two kernel pages, therefore it handles the cases when page->kmsan_shadow is NULL for the source and destination pages. It might be a good idea to use kmsan_copy_page_meta() in both cases, but to do that I want to better understand what happens when kmap_local_page(from) is called in copy_mc_user_highpage(). Where does the corresponding struct page come from?
On Wed, Feb 8, 2023 at 3:45 AM Alexander Potapenko <glider@google.com> wrote: > > On Tue, Feb 7, 2023 at 7:20 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > > > Pinging KMSAN experts, for the general guidance of > > kmsan_copy_page_meta vs kmsan_unpoison_memory > > Oh, sorry, I've missed the previous email. NP, thanks for jumping in :) > > copy_mc_user_highpage() is expected to copy data from the user page, > for which no metadata is ever allocated. > Therefore we just initialize the destination shadow with zeros instead > of copying anything. > > kmsan_copy_page_meta() is used when the metadata is copied between two > kernel pages, therefore it handles the cases when page->kmsan_shadow > is NULL for the source and destination pages. > > It might be a good idea to use kmsan_copy_page_meta() in both cases, > but to do that I want to better understand what happens when > kmap_local_page(from) is called in copy_mc_user_highpage(). > Where does the corresponding struct page come from? Kirill can correct me, but I think khugepaged always copies user pages because it is trying to convert raw pages to THP for better userspace application performance. Therefore khugepaged should only need copy_mc_user_highpage(), for both file-backed and anonymous memory pages. However, copy_mc_user_highpage() needs both vaddr and vma, so it is a little bit hard to use it in collapse_file (i.e. in the file-backed case): 1. vma is not carried over to collapse_file from khugepaged_scan_mm_slot 2. collapse_file is not directly iterating with vaddrs of pages to be copied (Although both vaddr and vma are unused auguments in copy_mc_user_highpage, I think for cleanness, the caller e.g. khugepaged should feed valid values). So my patchset uses copy_mc_page(and kmsan_copy_page_meta) for both file-backed and anon memory pages. I guess as long as kmsan_copy_page_meta doesn't do anything unexpected for user pages (at least from my reading), we are good?
Hi Kirill, Given the discussion, are you ok with `copy_mc_page` introduced in this patch? If so I am going to send out v10 that has a major refactor on __collapse_huge_page_copy. Thanks, Jiaqi On Wed, Feb 8, 2023 at 3:00 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > On Wed, Feb 8, 2023 at 3:45 AM Alexander Potapenko <glider@google.com> wrote: > > > > On Tue, Feb 7, 2023 at 7:20 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > > > > > Pinging KMSAN experts, for the general guidance of > > > kmsan_copy_page_meta vs kmsan_unpoison_memory > > > > Oh, sorry, I've missed the previous email. > > NP, thanks for jumping in :) > > > > > copy_mc_user_highpage() is expected to copy data from the user page, > > for which no metadata is ever allocated. > > Therefore we just initialize the destination shadow with zeros instead > > of copying anything. > > > > kmsan_copy_page_meta() is used when the metadata is copied between two > > kernel pages, therefore it handles the cases when page->kmsan_shadow > > is NULL for the source and destination pages. > > > > It might be a good idea to use kmsan_copy_page_meta() in both cases, > > but to do that I want to better understand what happens when > > kmap_local_page(from) is called in copy_mc_user_highpage(). > > Where does the corresponding struct page come from? > > Kirill can correct me, but I think khugepaged always copies user pages > because it is trying to convert raw pages to THP for better userspace > application performance. Therefore khugepaged should only need > copy_mc_user_highpage(), for both file-backed and anonymous memory > pages. > > However, copy_mc_user_highpage() needs both vaddr and vma, so it is a > little bit hard to use it in collapse_file (i.e. in the file-backed > case): > 1. vma is not carried over to collapse_file from khugepaged_scan_mm_slot > 2. collapse_file is not directly iterating with vaddrs of pages to be copied > > (Although both vaddr and vma are unused auguments in > copy_mc_user_highpage, I think for cleanness, the caller e.g. > khugepaged should feed valid values). > > So my patchset uses copy_mc_page(and kmsan_copy_page_meta) for both > file-backed and anon memory pages. I guess as long as > kmsan_copy_page_meta doesn't do anything unexpected for user pages (at > least from my reading), we are good?
On Wed, Feb 08, 2023 at 03:00:59PM -0800, Jiaqi Yan wrote: > On Wed, Feb 8, 2023 at 3:45 AM Alexander Potapenko <glider@google.com> wrote: > > > > On Tue, Feb 7, 2023 at 7:20 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > > > > > Pinging KMSAN experts, for the general guidance of > > > kmsan_copy_page_meta vs kmsan_unpoison_memory > > > > Oh, sorry, I've missed the previous email. > > NP, thanks for jumping in :) > > > > > copy_mc_user_highpage() is expected to copy data from the user page, > > for which no metadata is ever allocated. > > Therefore we just initialize the destination shadow with zeros instead > > of copying anything. > > > > kmsan_copy_page_meta() is used when the metadata is copied between two > > kernel pages, therefore it handles the cases when page->kmsan_shadow > > is NULL for the source and destination pages. > > > > It might be a good idea to use kmsan_copy_page_meta() in both cases, > > but to do that I want to better understand what happens when > > kmap_local_page(from) is called in copy_mc_user_highpage(). > > Where does the corresponding struct page come from? > > Kirill can correct me, but I think khugepaged always copies user pages > because it is trying to convert raw pages to THP for better userspace > application performance. Therefore khugepaged should only need > copy_mc_user_highpage(), for both file-backed and anonymous memory > pages. > > However, copy_mc_user_highpage() needs both vaddr and vma, so it is a > little bit hard to use it in collapse_file (i.e. in the file-backed > case): > 1. vma is not carried over to collapse_file from khugepaged_scan_mm_slot > 2. collapse_file is not directly iterating with vaddrs of pages to be copied > > (Although both vaddr and vma are unused auguments in > copy_mc_user_highpage, I think for cleanness, the caller e.g. > khugepaged should feed valid values). It is not unused for !copy_mc_to_kernel case (basically everything but x86 and power). And it is used to flush caches. The fact that you don't have it in your implementation *may* indicate a problem. > So my patchset uses copy_mc_page(and kmsan_copy_page_meta) for both > file-backed and anon memory pages. I guess as long as > kmsan_copy_page_meta doesn't do anything unexpected for user pages (at > least from my reading), we are good? >
On Tue, Feb 28, 2023 at 5:40 AM <kirill@shutemov.name> wrote: > > On Wed, Feb 08, 2023 at 03:00:59PM -0800, Jiaqi Yan wrote: > > On Wed, Feb 8, 2023 at 3:45 AM Alexander Potapenko <glider@google.com> wrote: > > > > > > On Tue, Feb 7, 2023 at 7:20 PM Jiaqi Yan <jiaqiyan@google.com> wrote: > > > > > > > > Pinging KMSAN experts, for the general guidance of > > > > kmsan_copy_page_meta vs kmsan_unpoison_memory > > > > > > Oh, sorry, I've missed the previous email. > > > > NP, thanks for jumping in :) > > > > > > > > copy_mc_user_highpage() is expected to copy data from the user page, > > > for which no metadata is ever allocated. > > > Therefore we just initialize the destination shadow with zeros instead > > > of copying anything. > > > > > > kmsan_copy_page_meta() is used when the metadata is copied between two > > > kernel pages, therefore it handles the cases when page->kmsan_shadow > > > is NULL for the source and destination pages. > > > > > > It might be a good idea to use kmsan_copy_page_meta() in both cases, > > > but to do that I want to better understand what happens when > > > kmap_local_page(from) is called in copy_mc_user_highpage(). > > > Where does the corresponding struct page come from? > > > > Kirill can correct me, but I think khugepaged always copies user pages > > because it is trying to convert raw pages to THP for better userspace > > application performance. Therefore khugepaged should only need > > copy_mc_user_highpage(), for both file-backed and anonymous memory > > pages. > > > > However, copy_mc_user_highpage() needs both vaddr and vma, so it is a > > little bit hard to use it in collapse_file (i.e. in the file-backed > > case): > > 1. vma is not carried over to collapse_file from khugepaged_scan_mm_slot > > 2. collapse_file is not directly iterating with vaddrs of pages to be copied > > > > (Although both vaddr and vma are unused auguments in > > copy_mc_user_highpage, I think for cleanness, the caller e.g. > > khugepaged should feed valid values). > > It is not unused for !copy_mc_to_kernel case (basically everything but x86 > and power). And it is used to flush caches. The fact that you don't have > it in your implementation *may* indicate a problem. ah, I agree this looks problematic. Maybe I should give up on unifying the copying routine used by anon and file memory: * for anon, switch to copy_mc_user_highpage, which "works" for all architectures, but #MC is only recoverable on x86 and powerpc. * for shmem, extend copy_highpage to copy_mc_highpage, like how Tony makes copy_mc_user_highpage. > > > So my patchset uses copy_mc_page(and kmsan_copy_page_meta) for both > > file-backed and anon memory pages. I guess as long as > > kmsan_copy_page_meta doesn't do anything unexpected for user pages (at > > least from my reading), we are good? > > > > -- > Kiryl Shutsemau / Kirill A. Shutemov
diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h index 35d759d3b0104..5743ae970af31 100644 --- a/include/trace/events/huge_memory.h +++ b/include/trace/events/huge_memory.h @@ -36,7 +36,8 @@ EM( SCAN_ALLOC_HUGE_PAGE_FAIL, "alloc_huge_page_failed") \ EM( SCAN_CGROUP_CHARGE_FAIL, "ccgroup_charge_failed") \ EM( SCAN_TRUNCATED, "truncated") \ - EMe(SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ + EM( SCAN_PAGE_HAS_PRIVATE, "page_has_private") \ + EMe(SCAN_COPY_MC, "copy_poisoned_page") \ #undef EM #undef EMe diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 5a7d2d5093f9c..0f1b9e05e17ec 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c @@ -19,6 +19,7 @@ #include <linux/page_table_check.h> #include <linux/swapops.h> #include <linux/shmem_fs.h> +#include <linux/kmsan.h> #include <asm/tlb.h> #include <asm/pgalloc.h> @@ -55,6 +56,7 @@ enum scan_result { SCAN_CGROUP_CHARGE_FAIL, SCAN_TRUNCATED, SCAN_PAGE_HAS_PRIVATE, + SCAN_COPY_MC, }; #define CREATE_TRACE_POINTS @@ -530,6 +532,27 @@ static bool is_refcount_suitable(struct page *page) return page_count(page) == expected_refcount; } +/* + * Copies memory with #MC in source page (@from) handled. Returns number + * of bytes not copied if there was an exception; otherwise 0 for success. + * Note handling #MC requires arch opt-in. + */ +static int copy_mc_page(struct page *to, struct page *from) +{ + char *vfrom, *vto; + unsigned long ret; + + vfrom = kmap_local_page(from); + vto = kmap_local_page(to); + ret = copy_mc_to_kernel(vto, vfrom, PAGE_SIZE); + if (ret == 0) + kmsan_copy_page_meta(to, from); + kunmap_local(vto); + kunmap_local(vfrom); + + return ret; +} + static int __collapse_huge_page_isolate(struct vm_area_struct *vma, unsigned long address, pte_t *pte, @@ -670,56 +693,124 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma, return result; } -static void __collapse_huge_page_copy(pte_t *pte, struct page *page, - struct vm_area_struct *vma, - unsigned long address, - spinlock_t *ptl, - struct list_head *compound_pagelist) +/* + * __collapse_huge_page_copy - attempts to copy memory contents from normal + * pages to a hugepage. Cleans up the normal pages if copying succeeds; + * otherwise restores the original page table and releases isolated normal pages. + * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. + * + * @pte: starting of the PTEs to copy from + * @page: the new hugepage to copy contents to + * @pmd: pointer to the new hugepage's PMD + * @rollback: the original normal pages' PMD + * @vma: the original normal pages' virtual memory area + * @address: starting address to copy + * @pte_ptl: lock on normal pages' PTEs + * @compound_pagelist: list that stores compound pages + */ +static int __collapse_huge_page_copy(pte_t *pte, + struct page *page, + pmd_t *pmd, + pmd_t rollback, + struct vm_area_struct *vma, + unsigned long address, + spinlock_t *pte_ptl, + struct list_head *compound_pagelist) { struct page *src_page, *tmp; pte_t *_pte; - for (_pte = pte; _pte < pte + HPAGE_PMD_NR; - _pte++, page++, address += PAGE_SIZE) { - pte_t pteval = *_pte; + pte_t pteval; + unsigned long _address; + spinlock_t *pmd_ptl; + int result = SCAN_SUCCEED; - if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { - clear_user_highpage(page, address); - add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); - if (is_zero_pfn(pte_pfn(pteval))) { + /* + * Copying pages' contents is subject to memory poison at any iteration. + */ + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; + _pte++, page++, _address += PAGE_SIZE) { + pteval = *_pte; + + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) + clear_user_highpage(page, _address); + else { + src_page = pte_page(pteval); + if (copy_mc_page(page, src_page) > 0) { + result = SCAN_COPY_MC; + break; + } + } + } + + if (likely(result == SCAN_SUCCEED)) { + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; + _pte++, _address += PAGE_SIZE) { + pteval = *_pte; + if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { + add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); + if (is_zero_pfn(pte_pfn(pteval))) { + /* + * pte_ptl mostly unnecessary. + */ + spin_lock(pte_ptl); + pte_clear(vma->vm_mm, _address, _pte); + spin_unlock(pte_ptl); + } + } else { + src_page = pte_page(pteval); + if (!PageCompound(src_page)) + release_pte_page(src_page); /* - * ptl mostly unnecessary. + * pte_ptl mostly unnecessary, but preempt has + * to be disabled to update the per-cpu stats + * inside page_remove_rmap(). */ - spin_lock(ptl); - ptep_clear(vma->vm_mm, address, _pte); - spin_unlock(ptl); + spin_lock(pte_ptl); + ptep_clear(vma->vm_mm, _address, _pte); + page_remove_rmap(src_page, vma, false); + spin_unlock(pte_ptl); + free_page_and_swap_cache(src_page); + } + } + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { + list_del(&src_page->lru); + mod_node_page_state(page_pgdat(src_page), + NR_ISOLATED_ANON + page_is_file_lru(src_page), + -compound_nr(src_page)); + unlock_page(src_page); + free_swap_cache(src_page); + putback_lru_page(src_page); + } + } else { + /* + * Re-establish the regular PMD that points to the regular + * page table. Restoring PMD needs to be done prior to + * releasing pages. Since pages are still isolated and + * locked here, acquiring anon_vma_lock_write is unnecessary. + */ + pmd_ptl = pmd_lock(vma->vm_mm, pmd); + pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback)); + spin_unlock(pmd_ptl); + /* + * Release both raw and compound pages isolated + * in __collapse_huge_page_isolate. + */ + for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; + _pte++, _address += PAGE_SIZE) { + pteval = *_pte; + if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) { + src_page = pte_page(pteval); + if (!PageCompound(src_page)) + release_pte_page(src_page); } - } else { - src_page = pte_page(pteval); - copy_user_highpage(page, src_page, address, vma); - if (!PageCompound(src_page)) - release_pte_page(src_page); - /* - * ptl mostly unnecessary, but preempt has to - * be disabled to update the per-cpu stats - * inside page_remove_rmap(). - */ - spin_lock(ptl); - ptep_clear(vma->vm_mm, address, _pte); - page_remove_rmap(src_page, vma, false); - spin_unlock(ptl); - free_page_and_swap_cache(src_page); + } + list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { + list_del(&src_page->lru); + release_pte_page(src_page); } } - list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { - list_del(&src_page->lru); - mod_node_page_state(page_pgdat(src_page), - NR_ISOLATED_ANON + page_is_file_lru(src_page), - -compound_nr(src_page)); - unlock_page(src_page); - free_swap_cache(src_page); - putback_lru_page(src_page); - } + return result; } static void khugepaged_alloc_sleep(void) @@ -1079,9 +1170,13 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address, */ anon_vma_unlock_write(vma->anon_vma); - __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, - &compound_pagelist); + result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, + vma, address, pte_ptl, + &compound_pagelist); pte_unmap(pte); + if (unlikely(result != SCAN_SUCCEED)) + goto out_up_write; + /* * spin_lock() below is not the equivalent of smp_wmb(), but * the smp_wmb() inside __SetPageUptodate() can be reused to
Make __collapse_huge_page_copy return whether copying anonymous pages succeeded, and make collapse_huge_page handle the return status. Break existing PTE scan loop into two for-loops. The first loop copies source pages into target huge page, and can fail gracefully when running into memory errors in source pages. If copying all pages succeeds, the second loop releases and clears up these normal pages. Otherwise, the second loop rolls back the page table and page states by: - re-establishing the original PTEs-to-PMD connection. - releasing source pages back to their LRU list. Tested manually: 0. Enable khugepaged on system under test. 1. Start a two-thread application. Each thread allocates a chunk of non-huge anonymous memory buffer. 2. Pick 4 random buffer locations (2 in each thread) and inject uncorrectable memory errors at corresponding physical addresses. 3. Signal both threads to make their memory buffer collapsible, i.e. calling madvise(MADV_HUGEPAGE). 4. Wait and check kernel log: khugepaged is able to recover from poisoned pages and skips collapsing them. 5. Signal both threads to inspect their buffer contents and make sure no data corruption. Signed-off-by: Jiaqi Yan <jiaqiyan@google.com> --- include/trace/events/huge_memory.h | 3 +- mm/khugepaged.c | 179 ++++++++++++++++++++++------- 2 files changed, 139 insertions(+), 43 deletions(-)