[RFC v1 0/2] Memory poison recovery in khugepaged

[RFC v1 0/2] Memory poison recovery in khugepaged

[Jiaqi Yan]

Problem
=======
Memory DIMMs are subject to multi-bit flips, i.e. memory errors.
As memory size and density increase, the chances of and number of
memory errors increase. The increasing size and density of server
RAM in the data center and cloud have shown increased uncorrectable
memory errors. There are already mechanisms in the kernel to recover
from uncorrectable memory errors. This series of patches provides
the recovery mechanism for the particular kernel agent khugepaged.

Impact
======
The main reason we chose to make khugepaged tolerant of memory failures
was its high possibility of accessing poisoned memory while performing
functionally optional compaction actions. Standard applications
typically don't have strict requirements on the size of its pages.
So they are given 4K pages by the kernel. The kernel is able to improve
application performance by either 1) giving application 2M pages
to begin with, or 2) collapsing 4K pages into 2M pages when possible.
This collapsing operation is done by khugepaged, a kernel agent that
is constantly scanning memory. When collapsing 4K pages into a 2M page,
it must copy the data from the 4K pages into a physically contiguous
2M page. Therefore, as long as there exists one poisoned cache line in
collapsible 4K pages, khugepaged will eventually access it. The current
impact to users is a machine check exception triggered kernel panic.
However, khugepaged’s compaction operations are not functionally required
kernel actions. Therefore making khugepaged tolerant to poisoned memory
will greatly improve user experience.

Solution
========
As stated before, it is less desirable to crash the system only because
khugepaged accesses poisoned pages while it is collapsing 4K pages.
The high level idea of this patch series is to skip the group of pages
(usually 512 4K-size pages) once khugepaged finds one of them is poisoned,
as these pages have become ineligible to be collapsed.

We are also careful to unwind operations khuagepaged has performed before
it detects memory failures. For example, before copying and collapsing
a group of anonymous pages into a huge page, the source pages will be
isolated and their page table is unlinked from their PMD. These operations
need to be undone in order to ensure these pages are not changed/lost from
the perspective of other threads (both user and kernel space). As for
file backed memory pages, there already exists a rollback case. This
patch just extends it so that khugepaged also correctly rolls back when
it fails to copy poisoned 4K pages.

Jiaqi Yan (2):
  mm: khugepaged: recover from poisoned anonymous memory
  mm: khugepaged: recover from poisoned file-backed memory

 include/linux/highmem.h |  37 +++++++
 mm/khugepaged.c         | 211 +++++++++++++++++++++++++++++-----------
 2 files changed, 189 insertions(+), 59 deletions(-)

-- 
2.35.1.894.gb6a874cedc-goog

[RFC v1 1/2] mm: khugepaged: recover from poisoned anonymous memory

[Jiaqi Yan]

Make __collapse_huge_page_copy return whether
collapsing/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. Roll back the page table and page states
in the 2nd loop when copy failed:
1) re-establish the PTEs-to-PMD connection.
2) release pages back to their LRU list.

Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
---
 include/linux/highmem.h |  19 ++++++
 mm/khugepaged.c         | 136 ++++++++++++++++++++++++++++++----------
 2 files changed, 122 insertions(+), 33 deletions(-)

diff --git a/include/linux/highmem.h b/include/linux/highmem.h
index 39bb9b47fa9c..15d0aa4d349c 100644
--- a/include/linux/highmem.h
+++ b/include/linux/highmem.h
@@ -281,6 +281,25 @@ static inline void copy_user_highpage(struct page *to, struct page *from,
 	kunmap_local(vfrom);
 }
 
+/*
+ * Machine check exception handled version of copy_user_highpage.
+ * Return true if copying page content failed; otherwise false.
+ */
+static inline bool copy_user_highpage_mc(struct page *to, struct page *from,
+	unsigned long vaddr, struct vm_area_struct *vma)
+{
+	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);
+	kunmap_local(vto);
+	kunmap_local(vfrom);
+
+	return ret > 0;
+}
+
 #endif
 
 #ifndef __HAVE_ARCH_COPY_HIGHPAGE
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 131492fd1148..84ed177f56ff 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -739,44 +739,97 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 	return 0;
 }
 
-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. Cleanup the normal pages if copy succeeds;
+ * otherwise restore the original pmd page table.
+ *
+ * @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 PTEs' PMD
+ * @address: starting address to copy
+ * @pte_ptl: lock on normal pages' PTEs
+ * @compound_pagelist: list that stores compound pages
+ */
+static bool __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;
+	bool copy_succeeded = true;
 
-		if (pte_none(pteval) || 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);
-			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
-			if (is_zero_pfn(pte_pfn(pteval))) {
-				/*
-				 * ptl mostly unnecessary.
-				 */
-				spin_lock(ptl);
-				ptep_clear(vma->vm_mm, address, _pte);
-				spin_unlock(ptl);
+		else {
+			src_page = pte_page(pteval);
+			if (copy_user_highpage_mc(page, src_page, address, vma)) {
+				copy_succeeded = false;
+				break;
+			}
+		}
+	}
+
+	if (!copy_succeeded) {
+		/*
+		 * Copying failed, re-establish the regular PMD that
+		 * points to regular page table. Since PTEs are still
+		 * isolated and locked, acquiring anon_vma_lock is unnecessary.
+		 */
+		pmd_ptl = pmd_lock(vma->vm_mm, pmd);
+		pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback));
+		spin_unlock(pmd_ptl);
+	}
+
+	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))) {
+			if (copy_succeeded) {
+				add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
+				if (is_zero_pfn(pte_pfn(pteval))) {
+					/*
+					 * ptl mostly unnecessary.
+					 */
+					spin_lock(pte_ptl);
+					pte_clear(vma->vm_mm, _address, _pte);
+					spin_unlock(pte_ptl);
+				}
 			}
 		} 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, false);
-			spin_unlock(ptl);
-			free_page_and_swap_cache(src_page);
+
+			if (copy_succeeded) {
+				/*
+				 * ptl mostly unnecessary, but preempt has to
+				 * be disabled to update the per-cpu stats
+				 * inside page_remove_rmap().
+				 */
+				spin_lock(pte_ptl);
+				pte_clear(vma->vm_mm, _address, _pte);
+				page_remove_rmap(src_page, false);
+				spin_unlock(pte_ptl);
+				free_page_and_swap_cache(src_page);
+			}
 		}
 	}
 
@@ -784,6 +837,8 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 		list_del(&src_page->lru);
 		release_pte_page(src_page);
 	}
+
+	return copy_succeeded;
 }
 
 static void khugepaged_alloc_sleep(void)
@@ -1066,6 +1121,7 @@ static void collapse_huge_page(struct mm_struct *mm,
 	struct vm_area_struct *vma;
 	struct mmu_notifier_range range;
 	gfp_t gfp;
+	bool copy_succeeded = false;
 
 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 
@@ -1177,9 +1233,13 @@ static void collapse_huge_page(struct mm_struct *mm,
 	 */
 	anon_vma_unlock_write(vma->anon_vma);
 
-	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
-			&compound_pagelist);
+	copy_succeeded = __collapse_huge_page_copy(pte, new_page, pmd, _pmd,
+			vma, address, pte_ptl, &compound_pagelist);
 	pte_unmap(pte);
+	if (!copy_succeeded) {
+		result = SCAN_FAIL;
+		goto out_up_write;
+	}
 	/*
 	 * spin_lock() below is not the equivalent of smp_wmb(), but
 	 * the smp_wmb() inside __SetPageUptodate() can be reused to
@@ -1364,9 +1424,14 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
 	pte_unmap_unlock(pte, ptl);
 	if (ret) {
 		node = khugepaged_find_target_node();
-		/* collapse_huge_page will return with the mmap_lock released */
-		collapse_huge_page(mm, address, hpage, node,
-				referenced, unmapped);
+		/*
+		 * collapse_huge_page will return with the mmap_r+w_lock released.
+		 * It is uncertain if *hpage is NULL or not when collapse_huge_page
+		 * returns, so keep ret=1 to jump to breakouterloop_mmap_lock
+		 * in khugepaged_scan_mm_slot, then *hpage will be freed
+		 * if collapse failed.
+		 */
+		collapse_huge_page(mm, address, hpage, node, referenced, unmapped);
 	}
 out:
 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
@@ -2168,6 +2233,11 @@ static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
 				khugepaged_scan_file(mm, file, pgoff, hpage);
 				fput(file);
 			} else {
+				/*
+				 * mmap_read_lock is
+				 * 1) released if both scan and collapse succeeded;
+				 * 2) still held if either scan or collapse failed.
+				 */
 				ret = khugepaged_scan_pmd(mm, vma,
 						khugepaged_scan.address,
 						hpage);
-- 
2.35.1.894.gb6a874cedc-goog

Re: [RFC v1 1/2] mm: khugepaged: recover from poisoned anonymous memory

[Yang Shi]

On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> Make __collapse_huge_page_copy return whether
> collapsing/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. Roll back the page table and page states
> in the 2nd loop when copy failed:
> 1) re-establish the PTEs-to-PMD connection.
> 2) release pages back to their LRU list.
>
> Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> ---
>  include/linux/highmem.h |  19 ++++++
>  mm/khugepaged.c         | 136 ++++++++++++++++++++++++++++++----------
>  2 files changed, 122 insertions(+), 33 deletions(-)
>
> diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> index 39bb9b47fa9c..15d0aa4d349c 100644
> --- a/include/linux/highmem.h
> +++ b/include/linux/highmem.h
> @@ -281,6 +281,25 @@ static inline void copy_user_highpage(struct page *to, struct page *from,
>         kunmap_local(vfrom);
>  }
>
> +/*
> + * Machine check exception handled version of copy_user_highpage.
> + * Return true if copying page content failed; otherwise false.
> + */
> +static inline bool copy_user_highpage_mc(struct page *to, struct page *from,
> +       unsigned long vaddr, struct vm_area_struct *vma)

Patch #2 defined copy_highpage_mc() which has the same implementation.
This function has two unused parameters: vaddr and vma. It seems we
just need to keep one.

> +{
> +       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);
> +       kunmap_local(vto);
> +       kunmap_local(vfrom);
> +
> +       return ret > 0;
> +}
> +
>  #endif
>
>  #ifndef __HAVE_ARCH_COPY_HIGHPAGE
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 131492fd1148..84ed177f56ff 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -739,44 +739,97 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
>         return 0;
>  }
>
> -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. Cleanup the normal pages if copy succeeds;
> + * otherwise restore the original pmd page table.
> + *
> + * @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 PTEs' PMD
> + * @address: starting address to copy
> + * @pte_ptl: lock on normal pages' PTEs
> + * @compound_pagelist: list that stores compound pages
> + */
> +static bool __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;
> +       bool copy_succeeded = true;
>
> -               if (pte_none(pteval) || 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);
> -                       add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> -                       if (is_zero_pfn(pte_pfn(pteval))) {
> -                               /*
> -                                * ptl mostly unnecessary.
> -                                */
> -                               spin_lock(ptl);
> -                               ptep_clear(vma->vm_mm, address, _pte);
> -                               spin_unlock(ptl);
> +               else {
> +                       src_page = pte_page(pteval);
> +                       if (copy_user_highpage_mc(page, src_page, address, vma)) {
> +                               copy_succeeded = false;
> +                               break;
> +                       }
> +               }
> +       }
> +
> +       if (!copy_succeeded) {
> +               /*
> +                * Copying failed, re-establish the regular PMD that
> +                * points to regular page table. Since PTEs are still
> +                * isolated and locked, acquiring anon_vma_lock is unnecessary.
> +                */
> +               pmd_ptl = pmd_lock(vma->vm_mm, pmd);
> +               pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback));
> +               spin_unlock(pmd_ptl);

I think we could just jump to list_for_each_entry_safe, right? But it
seems that list_for_each_entry_safe is not enough, please see the
later comments.

> +       }
> +
> +       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))) {
> +                       if (copy_succeeded) {

With the above jump, you don't need to check if copy is succeeded or
not since this is the succeeded path only.

> +                               add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> +                               if (is_zero_pfn(pte_pfn(pteval))) {
> +                                       /*
> +                                        * ptl mostly unnecessary.
> +                                        */
> +                                       spin_lock(pte_ptl);
> +                                       pte_clear(vma->vm_mm, _address, _pte);
> +                                       spin_unlock(pte_ptl);
> +                               }
>                         }
>                 } 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, false);
> -                       spin_unlock(ptl);
> -                       free_page_and_swap_cache(src_page);
> +
> +                       if (copy_succeeded) {
> +                               /*
> +                                * ptl mostly unnecessary, but preempt has to
> +                                * be disabled to update the per-cpu stats
> +                                * inside page_remove_rmap().
> +                                */
> +                               spin_lock(pte_ptl);
> +                               pte_clear(vma->vm_mm, _address, _pte);
> +                               page_remove_rmap(src_page, false);
> +                               spin_unlock(pte_ptl);
> +                               free_page_and_swap_cache(src_page);
> +                       }
>                 }
>         }
>
> @@ -784,6 +837,8 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
>                 list_del(&src_page->lru);
>                 release_pte_page(src_page);
>         }

If copy failed you need to unpin the isolated pages and put them back
to LRU IIUC.

> +
> +       return copy_succeeded;
>  }
>
>  static void khugepaged_alloc_sleep(void)
> @@ -1066,6 +1121,7 @@ static void collapse_huge_page(struct mm_struct *mm,
>         struct vm_area_struct *vma;
>         struct mmu_notifier_range range;
>         gfp_t gfp;
> +       bool copy_succeeded = false;
>
>         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
>
> @@ -1177,9 +1233,13 @@ static void collapse_huge_page(struct mm_struct *mm,
>          */
>         anon_vma_unlock_write(vma->anon_vma);
>
> -       __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
> -                       &compound_pagelist);
> +       copy_succeeded = __collapse_huge_page_copy(pte, new_page, pmd, _pmd,
> +                       vma, address, pte_ptl, &compound_pagelist);
>         pte_unmap(pte);
> +       if (!copy_succeeded) {

Shall the fail handling be moved before pte_unmap()?

> +               result = SCAN_FAIL;

I think a new result is preferred, for example, SCAN_COPY_MC, it would
be helpful for debug.

> +               goto out_up_write;
> +       }
>         /*
>          * spin_lock() below is not the equivalent of smp_wmb(), but
>          * the smp_wmb() inside __SetPageUptodate() can be reused to
> @@ -1364,9 +1424,14 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
>         pte_unmap_unlock(pte, ptl);
>         if (ret) {
>                 node = khugepaged_find_target_node();
> -               /* collapse_huge_page will return with the mmap_lock released */
> -               collapse_huge_page(mm, address, hpage, node,
> -                               referenced, unmapped);
> +               /*
> +                * collapse_huge_page will return with the mmap_r+w_lock released.
> +                * It is uncertain if *hpage is NULL or not when collapse_huge_page
> +                * returns, so keep ret=1 to jump to breakouterloop_mmap_lock
> +                * in khugepaged_scan_mm_slot, then *hpage will be freed
> +                * if collapse failed.

It may be not true for !NUMA case, the huge page may be reused, so you
need to make sure the huge page is cleared before return, I had a
patch that remove the special case for !NUMA
(https://lore.kernel.org/linux-mm/20210817202146.3218-1-shy828301@gmail.com/),
it could avoid clearing huge page, I should resurrect it.

> +                */
> +               collapse_huge_page(mm, address, hpage, node, referenced, unmapped);
>         }
>  out:
>         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
> @@ -2168,6 +2233,11 @@ static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
>                                 khugepaged_scan_file(mm, file, pgoff, hpage);
>                                 fput(file);
>                         } else {
> +                               /*
> +                                * mmap_read_lock is
> +                                * 1) released if both scan and collapse succeeded;
> +                                * 2) still held if either scan or collapse failed.
> +                                */
>                                 ret = khugepaged_scan_pmd(mm, vma,
>                                                 khugepaged_scan.address,
>                                                 hpage);
> --
> 2.35.1.894.gb6a874cedc-goog
>

Re: [RFC v1 1/2] mm: khugepaged: recover from poisoned anonymous memory

[Jiaqi Yan]

On Thu, Mar 24, 2022 at 7:20 PM Yang Shi <shy828301@gmail.com> wrote:
>
> On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> >
> > Make __collapse_huge_page_copy return whether
> > collapsing/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. Roll back the page table and page states
> > in the 2nd loop when copy failed:
> > 1) re-establish the PTEs-to-PMD connection.
> > 2) release pages back to their LRU list.
> >
> > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> > ---
> >  include/linux/highmem.h |  19 ++++++
> >  mm/khugepaged.c         | 136 ++++++++++++++++++++++++++++++----------
> >  2 files changed, 122 insertions(+), 33 deletions(-)
> >
> > diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> > index 39bb9b47fa9c..15d0aa4d349c 100644
> > --- a/include/linux/highmem.h
> > +++ b/include/linux/highmem.h
> > @@ -281,6 +281,25 @@ static inline void copy_user_highpage(struct page *to, struct page *from,
> >         kunmap_local(vfrom);
> >  }
> >
> > +/*
> > + * Machine check exception handled version of copy_user_highpage.
> > + * Return true if copying page content failed; otherwise false.
> > + */
> > +static inline bool copy_user_highpage_mc(struct page *to, struct page *from,
> > +       unsigned long vaddr, struct vm_area_struct *vma)
>
> Patch #2 defined copy_highpage_mc() which has the same implementation.
> This function has two unused parameters: vaddr and vma. It seems we
> just need to keep one.
>

Version 2 will delete copy_user_highpage_mc() from the patch series,
and just keep copy_highpage_mc().

> > +{
> > +       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);
> > +       kunmap_local(vto);
> > +       kunmap_local(vfrom);
> > +
> > +       return ret > 0;
> > +}
> > +
> >  #endif
> >
> >  #ifndef __HAVE_ARCH_COPY_HIGHPAGE
> > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > index 131492fd1148..84ed177f56ff 100644
> > --- a/mm/khugepaged.c
> > +++ b/mm/khugepaged.c
> > @@ -739,44 +739,97 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
> >         return 0;
> >  }
> >
> > -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. Cleanup the normal pages if copy succeeds;
> > + * otherwise restore the original pmd page table.
> > + *
> > + * @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 PTEs' PMD
> > + * @address: starting address to copy
> > + * @pte_ptl: lock on normal pages' PTEs
> > + * @compound_pagelist: list that stores compound pages
> > + */
> > +static bool __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;
> > +       bool copy_succeeded = true;
> >
> > -               if (pte_none(pteval) || 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);
> > -                       add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> > -                       if (is_zero_pfn(pte_pfn(pteval))) {
> > -                               /*
> > -                                * ptl mostly unnecessary.
> > -                                */
> > -                               spin_lock(ptl);
> > -                               ptep_clear(vma->vm_mm, address, _pte);
> > -                               spin_unlock(ptl);
> > +               else {
> > +                       src_page = pte_page(pteval);
> > +                       if (copy_user_highpage_mc(page, src_page, address, vma)) {
> > +                               copy_succeeded = false;
> > +                               break;
> > +                       }
> > +               }
> > +       }
> > +
> > +       if (!copy_succeeded) {
> > +               /*
> > +                * Copying failed, re-establish the regular PMD that
> > +                * points to regular page table. Since PTEs are still
> > +                * isolated and locked, acquiring anon_vma_lock is unnecessary.
> > +                */
> > +               pmd_ptl = pmd_lock(vma->vm_mm, pmd);
> > +               pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback));
> > +               spin_unlock(pmd_ptl);
>
> I think we could just jump to list_for_each_entry_safe, right? But it
> seems that list_for_each_entry_safe is not enough, please see the
> later comments.

Yes, it is not enough, we need to release_pte_page() for non-compound pages.

If we are going to use jump/goto, we need two for-loops over PTEs, one
for the copy
succeeded case, and one for the failure case. What I have right now is combining
the two for-loops into one.
I don't like to switch to jump because: 1) goto statements have
bad readability, 2) one for-loop is more concise than two for-loops.

>
> > +       }
> > +
> > +       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))) {
> > +                       if (copy_succeeded) {
>
> With the above jump, you don't need to check if copy is succeeded or
> not since this is the succeeded path only.
>
> > +                               add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> > +                               if (is_zero_pfn(pte_pfn(pteval))) {
> > +                                       /*
> > +                                        * ptl mostly unnecessary.
> > +                                        */
> > +                                       spin_lock(pte_ptl);
> > +                                       pte_clear(vma->vm_mm, _address, _pte);
> > +                                       spin_unlock(pte_ptl);
> > +                               }
> >                         }
> >                 } 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, false);
> > -                       spin_unlock(ptl);
> > -                       free_page_and_swap_cache(src_page);
> > +
> > +                       if (copy_succeeded) {
> > +                               /*
> > +                                * ptl mostly unnecessary, but preempt has to
> > +                                * be disabled to update the per-cpu stats
> > +                                * inside page_remove_rmap().
> > +                                */
> > +                               spin_lock(pte_ptl);
> > +                               pte_clear(vma->vm_mm, _address, _pte);
> > +                               page_remove_rmap(src_page, false);
> > +                               spin_unlock(pte_ptl);
> > +                               free_page_and_swap_cache(src_page);
> > +                       }
> >                 }
> >         }
> >
> > @@ -784,6 +837,8 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
> >                 list_del(&src_page->lru);
> >                 release_pte_page(src_page);
> >         }
>
> If copy failed you need to unpin the isolated pages and put them back
> to LRU IIUC.

No matter if the copy failed or succeeded, we need to use a for-loop
to unpin isolated pages
and put them back to LRU (i.e. release_pte_page()). In current code,
this is a common
operation in the above for-loop.

>
> > +
> > +       return copy_succeeded;
> >  }
> >
> >  static void khugepaged_alloc_sleep(void)
> > @@ -1066,6 +1121,7 @@ static void collapse_huge_page(struct mm_struct *mm,
> >         struct vm_area_struct *vma;
> >         struct mmu_notifier_range range;
> >         gfp_t gfp;
> > +       bool copy_succeeded = false;
> >
> >         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
> >
> > @@ -1177,9 +1233,13 @@ static void collapse_huge_page(struct mm_struct *mm,
> >          */
> >         anon_vma_unlock_write(vma->anon_vma);
> >
> > -       __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
> > -                       &compound_pagelist);
> > +       copy_succeeded = __collapse_huge_page_copy(pte, new_page, pmd, _pmd,
> > +                       vma, address, pte_ptl, &compound_pagelist);
> >         pte_unmap(pte);
> > +       if (!copy_succeeded) {
>
> Shall the fail handling be moved before pte_unmap()?

I don't think so. I think at this point we need to pte_unmap() no
matter if the copying
succeeded or not. It is obviously needed if __collapse_huge_page_copy succeeded
(because existing code already assumes __collapse_huge_page_copy
always succeeded).
For the failure case, my reasoning is based on __collapse_huge_page_isolate.
If __collapse_huge_page_isolate failed, we need to pte_unmap()
before goto out_up_write. So I would assume if __collapse_huge_page_copy failed,
pte_unmap() is also required.

>
> > +               result = SCAN_FAIL;
>
> I think a new result is preferred, for example, SCAN_COPY_MC, it would
> be helpful for debug.

Version 2 will add this new enum.


>
> > +               goto out_up_write;
> > +       }
> >         /*
> >          * spin_lock() below is not the equivalent of smp_wmb(), but
> >          * the smp_wmb() inside __SetPageUptodate() can be reused to
> > @@ -1364,9 +1424,14 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> >         pte_unmap_unlock(pte, ptl);
> >         if (ret) {
> >                 node = khugepaged_find_target_node();
> > -               /* collapse_huge_page will return with the mmap_lock released */
> > -               collapse_huge_page(mm, address, hpage, node,
> > -                               referenced, unmapped);
> > +               /*
> > +                * collapse_huge_page will return with the mmap_r+w_lock released.
> > +                * It is uncertain if *hpage is NULL or not when collapse_huge_page
> > +                * returns, so keep ret=1 to jump to breakouterloop_mmap_lock
> > +                * in khugepaged_scan_mm_slot, then *hpage will be freed
> > +                * if collapse failed.
>
> It may be not true for !NUMA case, the huge page may be reused, so you
> need to make sure the huge page is cleared before return, I had a
> patch that remove the special case for !NUMA
> (https://lore.kernel.org/linux-mm/20210817202146.3218-1-shy828301@gmail.com/),
> it could avoid clearing huge page, I should resurrect it.

Thanks for catching this! Clearing the hugepage before returning is
for security reasons
or to avoid memory corruption, right?
If my understanding is correct, once your patch is merged, this will
be good. Your patch
seems to be a very simple one, do you expect it to be merged soon?

>
> > +                */
> > +               collapse_huge_page(mm, address, hpage, node, referenced, unmapped);
> >         }
> >  out:
> >         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
> > @@ -2168,6 +2233,11 @@ static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
> >                                 khugepaged_scan_file(mm, file, pgoff, hpage);
> >                                 fput(file);
> >                         } else {
> > +                               /*
> > +                                * mmap_read_lock is
> > +                                * 1) released if both scan and collapse succeeded;
> > +                                * 2) still held if either scan or collapse failed.
> > +                                */
> >                                 ret = khugepaged_scan_pmd(mm, vma,
> >                                                 khugepaged_scan.address,
> >                                                 hpage);
> > --
> > 2.35.1.894.gb6a874cedc-goog
> >

Thanks for the quick and valuable feedback.
I will send out the v2 RFC, and please leave/reply there. Thanks!

Re: [RFC v1 1/2] mm: khugepaged: recover from poisoned anonymous memory

[Yang Shi]

On Tue, Apr 5, 2022 at 1:48 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> On Thu, Mar 24, 2022 at 7:20 PM Yang Shi <shy828301@gmail.com> wrote:
> >
> > On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> > >
> > > Make __collapse_huge_page_copy return whether
> > > collapsing/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. Roll back the page table and page states
> > > in the 2nd loop when copy failed:
> > > 1) re-establish the PTEs-to-PMD connection.
> > > 2) release pages back to their LRU list.
> > >
> > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> > > ---
> > >  include/linux/highmem.h |  19 ++++++
> > >  mm/khugepaged.c         | 136 ++++++++++++++++++++++++++++++----------
> > >  2 files changed, 122 insertions(+), 33 deletions(-)
> > >
> > > diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> > > index 39bb9b47fa9c..15d0aa4d349c 100644
> > > --- a/include/linux/highmem.h
> > > +++ b/include/linux/highmem.h
> > > @@ -281,6 +281,25 @@ static inline void copy_user_highpage(struct page *to, struct page *from,
> > >         kunmap_local(vfrom);
> > >  }
> > >
> > > +/*
> > > + * Machine check exception handled version of copy_user_highpage.
> > > + * Return true if copying page content failed; otherwise false.
> > > + */
> > > +static inline bool copy_user_highpage_mc(struct page *to, struct page *from,
> > > +       unsigned long vaddr, struct vm_area_struct *vma)
> >
> > Patch #2 defined copy_highpage_mc() which has the same implementation.
> > This function has two unused parameters: vaddr and vma. It seems we
> > just need to keep one.
> >
>
> Version 2 will delete copy_user_highpage_mc() from the patch series,
> and just keep copy_highpage_mc().
>
> > > +{
> > > +       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);
> > > +       kunmap_local(vto);
> > > +       kunmap_local(vfrom);
> > > +
> > > +       return ret > 0;
> > > +}
> > > +
> > >  #endif
> > >
> > >  #ifndef __HAVE_ARCH_COPY_HIGHPAGE
> > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > > index 131492fd1148..84ed177f56ff 100644
> > > --- a/mm/khugepaged.c
> > > +++ b/mm/khugepaged.c
> > > @@ -739,44 +739,97 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
> > >         return 0;
> > >  }
> > >
> > > -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. Cleanup the normal pages if copy succeeds;
> > > + * otherwise restore the original pmd page table.
> > > + *
> > > + * @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 PTEs' PMD
> > > + * @address: starting address to copy
> > > + * @pte_ptl: lock on normal pages' PTEs
> > > + * @compound_pagelist: list that stores compound pages
> > > + */
> > > +static bool __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;
> > > +       bool copy_succeeded = true;
> > >
> > > -               if (pte_none(pteval) || 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);
> > > -                       add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> > > -                       if (is_zero_pfn(pte_pfn(pteval))) {
> > > -                               /*
> > > -                                * ptl mostly unnecessary.
> > > -                                */
> > > -                               spin_lock(ptl);
> > > -                               ptep_clear(vma->vm_mm, address, _pte);
> > > -                               spin_unlock(ptl);
> > > +               else {
> > > +                       src_page = pte_page(pteval);
> > > +                       if (copy_user_highpage_mc(page, src_page, address, vma)) {
> > > +                               copy_succeeded = false;
> > > +                               break;
> > > +                       }
> > > +               }
> > > +       }
> > > +
> > > +       if (!copy_succeeded) {
> > > +               /*
> > > +                * Copying failed, re-establish the regular PMD that
> > > +                * points to regular page table. Since PTEs are still
> > > +                * isolated and locked, acquiring anon_vma_lock is unnecessary.
> > > +                */
> > > +               pmd_ptl = pmd_lock(vma->vm_mm, pmd);
> > > +               pmd_populate(vma->vm_mm, pmd, pmd_pgtable(rollback));
> > > +               spin_unlock(pmd_ptl);
> >
> > I think we could just jump to list_for_each_entry_safe, right? But it
> > seems that list_for_each_entry_safe is not enough, please see the
> > later comments.
>
> Yes, it is not enough, we need to release_pte_page() for non-compound pages.
>
> If we are going to use jump/goto, we need two for-loops over PTEs, one
> for the copy
> succeeded case, and one for the failure case. What I have right now is combining
> the two for-loops into one.
> I don't like to switch to jump because: 1) goto statements have
> bad readability, 2) one for-loop is more concise than two for-loops.

Makes sense to me, it is not worth two loops.

>
> >
> > > +       }
> > > +
> > > +       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))) {
> > > +                       if (copy_succeeded) {
> >
> > With the above jump, you don't need to check if copy is succeeded or
> > not since this is the succeeded path only.
> >
> > > +                               add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
> > > +                               if (is_zero_pfn(pte_pfn(pteval))) {
> > > +                                       /*
> > > +                                        * ptl mostly unnecessary.
> > > +                                        */
> > > +                                       spin_lock(pte_ptl);
> > > +                                       pte_clear(vma->vm_mm, _address, _pte);
> > > +                                       spin_unlock(pte_ptl);
> > > +                               }
> > >                         }
> > >                 } 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, false);
> > > -                       spin_unlock(ptl);
> > > -                       free_page_and_swap_cache(src_page);
> > > +
> > > +                       if (copy_succeeded) {
> > > +                               /*
> > > +                                * ptl mostly unnecessary, but preempt has to
> > > +                                * be disabled to update the per-cpu stats
> > > +                                * inside page_remove_rmap().
> > > +                                */
> > > +                               spin_lock(pte_ptl);
> > > +                               pte_clear(vma->vm_mm, _address, _pte);
> > > +                               page_remove_rmap(src_page, false);
> > > +                               spin_unlock(pte_ptl);
> > > +                               free_page_and_swap_cache(src_page);
> > > +                       }
> > >                 }
> > >         }
> > >
> > > @@ -784,6 +837,8 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
> > >                 list_del(&src_page->lru);
> > >                 release_pte_page(src_page);
> > >         }
> >
> > If copy failed you need to unpin the isolated pages and put them back
> > to LRU IIUC.
>
> No matter if the copy failed or succeeded, we need to use a for-loop
> to unpin isolated pages
> and put them back to LRU (i.e. release_pte_page()). In current code,
> this is a common
> operation in the above for-loop.
>
> >
> > > +
> > > +       return copy_succeeded;
> > >  }
> > >
> > >  static void khugepaged_alloc_sleep(void)
> > > @@ -1066,6 +1121,7 @@ static void collapse_huge_page(struct mm_struct *mm,
> > >         struct vm_area_struct *vma;
> > >         struct mmu_notifier_range range;
> > >         gfp_t gfp;
> > > +       bool copy_succeeded = false;
> > >
> > >         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
> > >
> > > @@ -1177,9 +1233,13 @@ static void collapse_huge_page(struct mm_struct *mm,
> > >          */
> > >         anon_vma_unlock_write(vma->anon_vma);
> > >
> > > -       __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
> > > -                       &compound_pagelist);
> > > +       copy_succeeded = __collapse_huge_page_copy(pte, new_page, pmd, _pmd,
> > > +                       vma, address, pte_ptl, &compound_pagelist);
> > >         pte_unmap(pte);
> > > +       if (!copy_succeeded) {
> >
> > Shall the fail handling be moved before pte_unmap()?
>
> I don't think so. I think at this point we need to pte_unmap() no
> matter if the copying
> succeeded or not. It is obviously needed if __collapse_huge_page_copy succeeded
> (because existing code already assumes __collapse_huge_page_copy
> always succeeded).
> For the failure case, my reasoning is based on __collapse_huge_page_isolate.
> If __collapse_huge_page_isolate failed, we need to pte_unmap()
> before goto out_up_write. So I would assume if __collapse_huge_page_copy failed,
> pte_unmap() is also required.

Yeah, make sense.

>
> >
> > > +               result = SCAN_FAIL;
> >
> > I think a new result is preferred, for example, SCAN_COPY_MC, it would
> > be helpful for debug.
>
> Version 2 will add this new enum.
>
>
> >
> > > +               goto out_up_write;
> > > +       }
> > >         /*
> > >          * spin_lock() below is not the equivalent of smp_wmb(), but
> > >          * the smp_wmb() inside __SetPageUptodate() can be reused to
> > > @@ -1364,9 +1424,14 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> > >         pte_unmap_unlock(pte, ptl);
> > >         if (ret) {
> > >                 node = khugepaged_find_target_node();
> > > -               /* collapse_huge_page will return with the mmap_lock released */
> > > -               collapse_huge_page(mm, address, hpage, node,
> > > -                               referenced, unmapped);
> > > +               /*
> > > +                * collapse_huge_page will return with the mmap_r+w_lock released.
> > > +                * It is uncertain if *hpage is NULL or not when collapse_huge_page
> > > +                * returns, so keep ret=1 to jump to breakouterloop_mmap_lock
> > > +                * in khugepaged_scan_mm_slot, then *hpage will be freed
> > > +                * if collapse failed.
> >
> > It may be not true for !NUMA case, the huge page may be reused, so you
> > need to make sure the huge page is cleared before return, I had a
> > patch that remove the special case for !NUMA
> > (https://lore.kernel.org/linux-mm/20210817202146.3218-1-shy828301@gmail.com/),
> > it could avoid clearing huge page, I should resurrect it.
>
> Thanks for catching this! Clearing the hugepage before returning is
> for security reasons
> or to avoid memory corruption, right?

Yes, but it should be fine since the users typically clear the page
before using it, for example, do_anonymous_page() does call
alloc_zeroed_user_highpage_movable() to allocate a page, which clears
the page before returning. So this should not block you, but remind me
to relook my patch.

> If my understanding is correct, once your patch is merged, this will
> be good. Your patch
> seems to be a very simple one, do you expect it to be merged soon?

It still has some unsolved review comments, and I didn't get time to
work on it yet.

>
> >
> > > +                */
> > > +               collapse_huge_page(mm, address, hpage, node, referenced, unmapped);
> > >         }
> > >  out:
> > >         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
> > > @@ -2168,6 +2233,11 @@ static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
> > >                                 khugepaged_scan_file(mm, file, pgoff, hpage);
> > >                                 fput(file);
> > >                         } else {
> > > +                               /*
> > > +                                * mmap_read_lock is
> > > +                                * 1) released if both scan and collapse succeeded;
> > > +                                * 2) still held if either scan or collapse failed.
> > > +                                */
> > >                                 ret = khugepaged_scan_pmd(mm, vma,
> > >                                                 khugepaged_scan.address,
> > >                                                 hpage);
> > > --
> > > 2.35.1.894.gb6a874cedc-goog
> > >
>
> Thanks for the quick and valuable feedback.
> I will send out the v2 RFC, and please leave/reply there. Thanks!

I may not be able to take a deep look at v2 soon, I will be off for
the following week. I will take a look at it once I'm back.

[RFC v1 2/2] mm: khugepaged: recover from poisoned file-backed memory

[Jiaqi Yan]

Make collapse_file roll back when copying pages failed.
More concretely:
* extract copy operations into a separate loop
* postpone the updates for nr_none until both scan and copy succeeded
* postpone joining small xarray entries until both scan and copy
  succeeded
* as for update operations to NR_XXX_THPS
    * for SHMEM file, postpone until both scan and copy succeeded
    * for other file, roll back if scan succeeded but copy failed

Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
---
 include/linux/highmem.h | 18 ++++++++++
 mm/khugepaged.c         | 75 +++++++++++++++++++++++++++--------------
 2 files changed, 67 insertions(+), 26 deletions(-)

diff --git a/include/linux/highmem.h b/include/linux/highmem.h
index 15d0aa4d349c..fc5aa221bdb5 100644
--- a/include/linux/highmem.h
+++ b/include/linux/highmem.h
@@ -315,6 +315,24 @@ static inline void copy_highpage(struct page *to, struct page *from)
 	kunmap_local(vfrom);
 }
 
+/*
+ * Machine check exception handled version of copy_highpage.
+ * Return true if copying page content failed; otherwise false.
+ */
+static inline bool copy_highpage_mc(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);
+	kunmap_local(vto);
+	kunmap_local(vfrom);
+
+	return ret > 0;
+}
+
 #endif
 
 static inline void memcpy_page(struct page *dst_page, size_t dst_off,
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 84ed177f56ff..ed2b1cd4bbc6 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1708,12 +1708,13 @@ static void collapse_file(struct mm_struct *mm,
 {
 	struct address_space *mapping = file->f_mapping;
 	gfp_t gfp;
-	struct page *new_page;
+	struct page *new_page, *page, *tmp;
 	pgoff_t index, end = start + HPAGE_PMD_NR;
 	LIST_HEAD(pagelist);
 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
 	int nr_none = 0, result = SCAN_SUCCEED;
 	bool is_shmem = shmem_file(file);
+	bool copy_failed = false;
 	int nr;
 
 	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
@@ -1936,9 +1937,7 @@ static void collapse_file(struct mm_struct *mm,
 	}
 	nr = thp_nr_pages(new_page);
 
-	if (is_shmem)
-		__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
-	else {
+	if (!is_shmem) {
 		__mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
 		filemap_nr_thps_inc(mapping);
 		/*
@@ -1956,34 +1955,39 @@ static void collapse_file(struct mm_struct *mm,
 		}
 	}
 
-	if (nr_none) {
-		__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
-		if (is_shmem)
-			__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
-	}
-
-	/* Join all the small entries into a single multi-index entry */
-	xas_set_order(&xas, start, HPAGE_PMD_ORDER);
-	xas_store(&xas, new_page);
 xa_locked:
 	xas_unlock_irq(&xas);
 xa_unlocked:
 
 	if (result == SCAN_SUCCEED) {
-		struct page *page, *tmp;
-
 		/*
 		 * Replacing old pages with new one has succeeded, now we
-		 * need to copy the content and free the old pages.
+		 * attempt to copy the contents.
 		 */
 		index = start;
-		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
+		list_for_each_entry(page, &pagelist, lru) {
 			while (index < page->index) {
 				clear_highpage(new_page + (index % HPAGE_PMD_NR));
 				index++;
 			}
-			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
-					page);
+			if (copy_highpage_mc(new_page + (page->index % HPAGE_PMD_NR), page)) {
+				copy_failed = true;
+				break;
+			}
+			index++;
+		}
+		while (!copy_failed && index < end) {
+			clear_highpage(new_page + (page->index % HPAGE_PMD_NR));
+			index++;
+		}
+	}
+
+	if (result == SCAN_SUCCEED && !copy_failed) {
+		/*
+		 * Copying old pages to huge one has succeeded, now we
+		 * need to free the old pages.
+		 */
+		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
 			list_del(&page->lru);
 			page->mapping = NULL;
 			page_ref_unfreeze(page, 1);
@@ -1991,12 +1995,20 @@ static void collapse_file(struct mm_struct *mm,
 			ClearPageUnevictable(page);
 			unlock_page(page);
 			put_page(page);
-			index++;
 		}
-		while (index < end) {
-			clear_highpage(new_page + (index % HPAGE_PMD_NR));
-			index++;
+
+		xas_lock_irq(&xas);
+		if (is_shmem)
+			__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
+		if (nr_none) {
+			__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
+			if (is_shmem)
+				__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
 		}
+		/* Join all the small entries into a single multi-index entry. */
+		xas_set_order(&xas, start, HPAGE_PMD_ORDER);
+		xas_store(&xas, new_page);
+		xas_unlock_irq(&xas);
 
 		SetPageUptodate(new_page);
 		page_ref_add(new_page, HPAGE_PMD_NR - 1);
@@ -2012,9 +2024,11 @@ static void collapse_file(struct mm_struct *mm,
 
 		khugepaged_pages_collapsed++;
 	} else {
-		struct page *page;
-
-		/* Something went wrong: roll back page cache changes */
+		/*
+		 * Something went wrong:
+		 * either result != SCAN_SUCCEED or copy_failed,
+		 * roll back page cache changes
+		 */
 		xas_lock_irq(&xas);
 		mapping->nrpages -= nr_none;
 
@@ -2047,6 +2061,15 @@ static void collapse_file(struct mm_struct *mm,
 			xas_lock_irq(&xas);
 		}
 		VM_BUG_ON(nr_none);
+		/*
+		 * Undo the updates of thp_nr_pages(new_page) for non-SHMEM file,
+		 * which is not updated yet for SHMEM file.
+		 * These undos are not needed if result is not SCAN_SUCCEED.
+		 */
+		if (!is_shmem && result == SCAN_SUCCEED) {
+			__mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
+			filemap_nr_thps_dec(mapping);
+		}
 		xas_unlock_irq(&xas);
 
 		new_page->mapping = NULL;
-- 
2.35.1.894.gb6a874cedc-goog

Re: [RFC v1 2/2] mm: khugepaged: recover from poisoned file-backed memory

[Yang Shi]

On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> Make collapse_file roll back when copying pages failed.
> More concretely:
> * extract copy operations into a separate loop
> * postpone the updates for nr_none until both scan and copy succeeded
> * postpone joining small xarray entries until both scan and copy
>   succeeded
> * as for update operations to NR_XXX_THPS
>     * for SHMEM file, postpone until both scan and copy succeeded
>     * for other file, roll back if scan succeeded but copy failed
>
> Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> ---
>  include/linux/highmem.h | 18 ++++++++++
>  mm/khugepaged.c         | 75 +++++++++++++++++++++++++++--------------
>  2 files changed, 67 insertions(+), 26 deletions(-)
>
> diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> index 15d0aa4d349c..fc5aa221bdb5 100644
> --- a/include/linux/highmem.h
> +++ b/include/linux/highmem.h
> @@ -315,6 +315,24 @@ static inline void copy_highpage(struct page *to, struct page *from)
>         kunmap_local(vfrom);
>  }
>
> +/*
> + * Machine check exception handled version of copy_highpage.
> + * Return true if copying page content failed; otherwise false.
> + */
> +static inline bool copy_highpage_mc(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);
> +       kunmap_local(vto);
> +       kunmap_local(vfrom);
> +
> +       return ret > 0;
> +}
> +
>  #endif
>
>  static inline void memcpy_page(struct page *dst_page, size_t dst_off,
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 84ed177f56ff..ed2b1cd4bbc6 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1708,12 +1708,13 @@ static void collapse_file(struct mm_struct *mm,
>  {
>         struct address_space *mapping = file->f_mapping;
>         gfp_t gfp;
> -       struct page *new_page;
> +       struct page *new_page, *page, *tmp;

It seems you removed the "struct page *page" from " if (result ==
SCAN_SUCCEED)", but keep the "struct page *page" under "for (index =
start; index < end; index++)". I think the "struct page *page" in the
for loop could be removed too.

>         pgoff_t index, end = start + HPAGE_PMD_NR;
>         LIST_HEAD(pagelist);
>         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
>         int nr_none = 0, result = SCAN_SUCCEED;
>         bool is_shmem = shmem_file(file);
> +       bool copy_failed = false;
>         int nr;
>
>         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
> @@ -1936,9 +1937,7 @@ static void collapse_file(struct mm_struct *mm,
>         }
>         nr = thp_nr_pages(new_page);
>
> -       if (is_shmem)
> -               __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> -       else {
> +       if (!is_shmem) {
>                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
>                 filemap_nr_thps_inc(mapping);
>                 /*
> @@ -1956,34 +1955,39 @@ static void collapse_file(struct mm_struct *mm,
>                 }
>         }
>
> -       if (nr_none) {
> -               __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> -               if (is_shmem)
> -                       __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> -       }
> -
> -       /* Join all the small entries into a single multi-index entry */
> -       xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> -       xas_store(&xas, new_page);
>  xa_locked:
>         xas_unlock_irq(&xas);
>  xa_unlocked:
>
>         if (result == SCAN_SUCCEED) {
> -               struct page *page, *tmp;
> -
>                 /*
>                  * Replacing old pages with new one has succeeded, now we
> -                * need to copy the content and free the old pages.
> +                * attempt to copy the contents.
>                  */
>                 index = start;
> -               list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> +               list_for_each_entry(page, &pagelist, lru) {
>                         while (index < page->index) {
>                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
>                                 index++;
>                         }
> -                       copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
> -                                       page);
> +                       if (copy_highpage_mc(new_page + (page->index % HPAGE_PMD_NR), page)) {
> +                               copy_failed = true;
> +                               break;
> +                       }
> +                       index++;
> +               }
> +               while (!copy_failed && index < end) {
> +                       clear_highpage(new_page + (page->index % HPAGE_PMD_NR));
> +                       index++;
> +               }
> +       }
> +
> +       if (result == SCAN_SUCCEED && !copy_failed) {

I think you could set "result" to SCAN_COPY_MC (as same as the
anonymous one), then you could drop !copy_failed and use "result"
alone afterwards.

> +               /*
> +                * Copying old pages to huge one has succeeded, now we
> +                * need to free the old pages.
> +                */
> +               list_for_each_entry_safe(page, tmp, &pagelist, lru) {
>                         list_del(&page->lru);
>                         page->mapping = NULL;
>                         page_ref_unfreeze(page, 1);
> @@ -1991,12 +1995,20 @@ static void collapse_file(struct mm_struct *mm,
>                         ClearPageUnevictable(page);
>                         unlock_page(page);
>                         put_page(page);
> -                       index++;
>                 }
> -               while (index < end) {
> -                       clear_highpage(new_page + (index % HPAGE_PMD_NR));
> -                       index++;
> +
> +               xas_lock_irq(&xas);
> +               if (is_shmem)
> +                       __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> +               if (nr_none) {
> +                       __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> +                       if (is_shmem)
> +                               __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
>                 }
> +               /* Join all the small entries into a single multi-index entry. */
> +               xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> +               xas_store(&xas, new_page);
> +               xas_unlock_irq(&xas);
>
>                 SetPageUptodate(new_page);
>                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
> @@ -2012,9 +2024,11 @@ static void collapse_file(struct mm_struct *mm,
>
>                 khugepaged_pages_collapsed++;
>         } else {
> -               struct page *page;
> -
> -               /* Something went wrong: roll back page cache changes */
> +               /*
> +                * Something went wrong:
> +                * either result != SCAN_SUCCEED or copy_failed,
> +                * roll back page cache changes
> +                */
>                 xas_lock_irq(&xas);
>                 mapping->nrpages -= nr_none;
>
> @@ -2047,6 +2061,15 @@ static void collapse_file(struct mm_struct *mm,
>                         xas_lock_irq(&xas);
>                 }
>                 VM_BUG_ON(nr_none);
> +               /*
> +                * Undo the updates of thp_nr_pages(new_page) for non-SHMEM file,
> +                * which is not updated yet for SHMEM file.
> +                * These undos are not needed if result is not SCAN_SUCCEED.
> +                */
> +               if (!is_shmem && result == SCAN_SUCCEED) {

Handling the error case when "result == SCAN_SUCCEED" looks awkward.
With the above fixed (set result to SCAN_COPY_MC) we could avoid the
awkwardness.

> +                       __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);

I'm wondering whether we may defer NR_FILE_THPS update just like SHMEM
because it has not to be updated in advance so that we have the user
visible counters update in the single place. Just filemap_nr_thps
needs to be updated in advance since it is used to sync with file open
path to truncate huge pages.

> +                       filemap_nr_thps_dec(mapping);
> +               }
>                 xas_unlock_irq(&xas);
>
>                 new_page->mapping = NULL;
> --
> 2.35.1.894.gb6a874cedc-goog
>

Re: [RFC v1 2/2] mm: khugepaged: recover from poisoned file-backed memory

[Jiaqi Yan]

On Mon, Mar 28, 2022 at 4:37 PM Yang Shi <shy828301@gmail.com> wrote:
>
> On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> >
> > Make collapse_file roll back when copying pages failed.
> > More concretely:
> > * extract copy operations into a separate loop
> > * postpone the updates for nr_none until both scan and copy succeeded
> > * postpone joining small xarray entries until both scan and copy
> >   succeeded
> > * as for update operations to NR_XXX_THPS
> >     * for SHMEM file, postpone until both scan and copy succeeded
> >     * for other file, roll back if scan succeeded but copy failed
> >
> > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> > ---
> >  include/linux/highmem.h | 18 ++++++++++
> >  mm/khugepaged.c         | 75 +++++++++++++++++++++++++++--------------
> >  2 files changed, 67 insertions(+), 26 deletions(-)
> >
> > diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> > index 15d0aa4d349c..fc5aa221bdb5 100644
> > --- a/include/linux/highmem.h
> > +++ b/include/linux/highmem.h
> > @@ -315,6 +315,24 @@ static inline void copy_highpage(struct page *to, struct page *from)
> >         kunmap_local(vfrom);
> >  }
> >
> > +/*
> > + * Machine check exception handled version of copy_highpage.
> > + * Return true if copying page content failed; otherwise false.
> > + */
> > +static inline bool copy_highpage_mc(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);
> > +       kunmap_local(vto);
> > +       kunmap_local(vfrom);
> > +
> > +       return ret > 0;
> > +}
> > +
> >  #endif
> >
> >  static inline void memcpy_page(struct page *dst_page, size_t dst_off,
> > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > index 84ed177f56ff..ed2b1cd4bbc6 100644
> > --- a/mm/khugepaged.c
> > +++ b/mm/khugepaged.c
> > @@ -1708,12 +1708,13 @@ static void collapse_file(struct mm_struct *mm,
> >  {
> >         struct address_space *mapping = file->f_mapping;
> >         gfp_t gfp;
> > -       struct page *new_page;
> > +       struct page *new_page, *page, *tmp;
>
> It seems you removed the "struct page *page" from " if (result ==
> SCAN_SUCCEED)", but keep the "struct page *page" under "for (index =
> start; index < end; index++)". I think the "struct page *page" in the
> for loop could be removed too.

I agree. Removed in the next version.

>
> >         pgoff_t index, end = start + HPAGE_PMD_NR;
> >         LIST_HEAD(pagelist);
> >         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
> >         int nr_none = 0, result = SCAN_SUCCEED;
> >         bool is_shmem = shmem_file(file);
> > +       bool copy_failed = false;
> >         int nr;
> >
> >         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
> > @@ -1936,9 +1937,7 @@ static void collapse_file(struct mm_struct *mm,
> >         }
> >         nr = thp_nr_pages(new_page);
> >
> > -       if (is_shmem)
> > -               __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> > -       else {
> > +       if (!is_shmem) {
> >                 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
> >                 filemap_nr_thps_inc(mapping);
> >                 /*
> > @@ -1956,34 +1955,39 @@ static void collapse_file(struct mm_struct *mm,
> >                 }
> >         }
> >
> > -       if (nr_none) {
> > -               __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> > -               if (is_shmem)
> > -                       __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> > -       }
> > -
> > -       /* Join all the small entries into a single multi-index entry */
> > -       xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> > -       xas_store(&xas, new_page);
> >  xa_locked:
> >         xas_unlock_irq(&xas);
> >  xa_unlocked:
> >
> >         if (result == SCAN_SUCCEED) {
> > -               struct page *page, *tmp;
> > -
> >                 /*
> >                  * Replacing old pages with new one has succeeded, now we
> > -                * need to copy the content and free the old pages.
> > +                * attempt to copy the contents.
> >                  */
> >                 index = start;
> > -               list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> > +               list_for_each_entry(page, &pagelist, lru) {
> >                         while (index < page->index) {
> >                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
> >                                 index++;
> >                         }
> > -                       copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
> > -                                       page);
> > +                       if (copy_highpage_mc(new_page + (page->index % HPAGE_PMD_NR), page)) {
> > +                               copy_failed = true;
> > +                               break;
> > +                       }
> > +                       index++;
> > +               }
> > +               while (!copy_failed && index < end) {
> > +                       clear_highpage(new_page + (page->index % HPAGE_PMD_NR));
> > +                       index++;
> > +               }
> > +       }
> > +
> > +       if (result == SCAN_SUCCEED && !copy_failed) {
>
> I think you could set "result" to SCAN_COPY_MC (as same as the
> anonymous one), then you could drop !copy_failed and use "result"
> alone afterwards.

Yes, copy_failed variable will be eliminated in the next version.

>
> > +               /*
> > +                * Copying old pages to huge one has succeeded, now we
> > +                * need to free the old pages.
> > +                */
> > +               list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> >                         list_del(&page->lru);
> >                         page->mapping = NULL;
> >                         page_ref_unfreeze(page, 1);
> > @@ -1991,12 +1995,20 @@ static void collapse_file(struct mm_struct *mm,
> >                         ClearPageUnevictable(page);
> >                         unlock_page(page);
> >                         put_page(page);
> > -                       index++;
> >                 }
> > -               while (index < end) {
> > -                       clear_highpage(new_page + (index % HPAGE_PMD_NR));
> > -                       index++;
> > +
> > +               xas_lock_irq(&xas);
> > +               if (is_shmem)
> > +                       __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> > +               if (nr_none) {
> > +                       __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> > +                       if (is_shmem)
> > +                               __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> >                 }
> > +               /* Join all the small entries into a single multi-index entry. */
> > +               xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> > +               xas_store(&xas, new_page);
> > +               xas_unlock_irq(&xas);
> >
> >                 SetPageUptodate(new_page);
> >                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
> > @@ -2012,9 +2024,11 @@ static void collapse_file(struct mm_struct *mm,
> >
> >                 khugepaged_pages_collapsed++;
> >         } else {
> > -               struct page *page;
> > -
> > -               /* Something went wrong: roll back page cache changes */
> > +               /*
> > +                * Something went wrong:
> > +                * either result != SCAN_SUCCEED or copy_failed,
> > +                * roll back page cache changes
> > +                */
> >                 xas_lock_irq(&xas);
> >                 mapping->nrpages -= nr_none;
> >
> > @@ -2047,6 +2061,15 @@ static void collapse_file(struct mm_struct *mm,
> >                         xas_lock_irq(&xas);
> >                 }
> >                 VM_BUG_ON(nr_none);
> > +               /*
> > +                * Undo the updates of thp_nr_pages(new_page) for non-SHMEM file,
> > +                * which is not updated yet for SHMEM file.
> > +                * These undos are not needed if result is not SCAN_SUCCEED.
> > +                */
> > +               if (!is_shmem && result == SCAN_SUCCEED) {
>
> Handling the error case when "result == SCAN_SUCCEED" looks awkward.
> With the above fixed (set result to SCAN_COPY_MC) we could avoid the
> awkwardness.
>

Will look less awkward in the next version :)

> > +                       __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
>
> I'm wondering whether we may defer NR_FILE_THPS update just like SHMEM
> because it has not to be updated in advance so that we have the user
> visible counters update in the single place. Just filemap_nr_thps
> needs to be updated in advance since it is used to sync with file open
> path to truncate huge pages.

In version 2, I will postpone __mod_lruvec_page_state() until copying succeeded,
and undo filemap_nr_thps_inc if (!is_shmem && result == SCAN_COPY_MC).

>
> > +                       filemap_nr_thps_dec(mapping);
> > +               }
> >                 xas_unlock_irq(&xas);
> >
> >                 new_page->mapping = NULL;
> > --
> > 2.35.1.894.gb6a874cedc-goog
> >

Thanks for the valuable feedback. I will send out the v2 RFC separately,
and please leave/reply there. Thanks!

Re: [RFC v1 2/2] mm: khugepaged: recover from poisoned file-backed memory

[Tong Tiangen]

在 2022/3/24 7:29, Jiaqi Yan 写道:
> Make collapse_file roll back when copying pages failed.
> More concretely:
> * extract copy operations into a separate loop
> * postpone the updates for nr_none until both scan and copy succeeded
> * postpone joining small xarray entries until both scan and copy
>    succeeded
> * as for update operations to NR_XXX_THPS
>      * for SHMEM file, postpone until both scan and copy succeeded
>      * for other file, roll back if scan succeeded but copy failed
> 
> Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> ---
>   include/linux/highmem.h | 18 ++++++++++
>   mm/khugepaged.c         | 75 +++++++++++++++++++++++++++--------------
>   2 files changed, 67 insertions(+), 26 deletions(-)
> 
> diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> index 15d0aa4d349c..fc5aa221bdb5 100644
> --- a/include/linux/highmem.h
> +++ b/include/linux/highmem.h
> @@ -315,6 +315,24 @@ static inline void copy_highpage(struct page *to, struct page *from)
>   	kunmap_local(vfrom);
>   }
>   
> +/*
> + * Machine check exception handled version of copy_highpage.
> + * Return true if copying page content failed; otherwise false.
> + */
> +static inline bool copy_highpage_mc(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);
> +	kunmap_local(vto);
> +	kunmap_local(vfrom);
> +
> +	return ret > 0;
> +}
> +
>   #endif
>   
>   static inline void memcpy_page(struct page *dst_page, size_t dst_off,
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 84ed177f56ff..ed2b1cd4bbc6 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1708,12 +1708,13 @@ static void collapse_file(struct mm_struct *mm,
>   {
>   	struct address_space *mapping = file->f_mapping;
>   	gfp_t gfp;
> -	struct page *new_page;
> +	struct page *new_page, *page, *tmp;
>   	pgoff_t index, end = start + HPAGE_PMD_NR;
>   	LIST_HEAD(pagelist);
>   	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
>   	int nr_none = 0, result = SCAN_SUCCEED;
>   	bool is_shmem = shmem_file(file);
> +	bool copy_failed = false;
>   	int nr;
>   
>   	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
> @@ -1936,9 +1937,7 @@ static void collapse_file(struct mm_struct *mm,
>   	}
>   	nr = thp_nr_pages(new_page);
>   
> -	if (is_shmem)
> -		__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> -	else {
> +	if (!is_shmem) {
>   		__mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
>   		filemap_nr_thps_inc(mapping);
>   		/*
> @@ -1956,34 +1955,39 @@ static void collapse_file(struct mm_struct *mm,
>   		}
>   	}
>   
> -	if (nr_none) {
> -		__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> -		if (is_shmem)
> -			__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> -	}
> -
> -	/* Join all the small entries into a single multi-index entry */
> -	xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> -	xas_store(&xas, new_page);
>   xa_locked:
>   	xas_unlock_irq(&xas);
>   xa_unlocked:
>   
>   	if (result == SCAN_SUCCEED) {
> -		struct page *page, *tmp;
> -
>   		/*
>   		 * Replacing old pages with new one has succeeded, now we
> -		 * need to copy the content and free the old pages.
> +		 * attempt to copy the contents.
>   		 */
>   		index = start;
> -		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> +		list_for_each_entry(page, &pagelist, lru) {
>   			while (index < page->index) {
>   				clear_highpage(new_page + (index % HPAGE_PMD_NR));
>   				index++;
>   			}
> -			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
> -					page);
> +			if (copy_highpage_mc(new_page + (page->index % HPAGE_PMD_NR), page)) {
> +				copy_failed = true;

The 1st patch here used "copy_succeed = false", It is best that the 
logic of the two positions can be unified.

> +				break;
> +			}
> +			index++;
> +		}
> +		while (!copy_failed && index < end) {
> +			clear_highpage(new_page + (page->index % HPAGE_PMD_NR));
> +			index++;
> +		}
> +	}
> +
> +	if (result == SCAN_SUCCEED && !copy_failed) {
> +		/*
> +		 * Copying old pages to huge one has succeeded, now we
> +		 * need to free the old pages.
> +		 */
> +		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
>   			list_del(&page->lru);
>   			page->mapping = NULL;
>   			page_ref_unfreeze(page, 1);
> @@ -1991,12 +1995,20 @@ static void collapse_file(struct mm_struct *mm,
>   			ClearPageUnevictable(page);
>   			unlock_page(page);
>   			put_page(page);
> -			index++;
>   		}
> -		while (index < end) {
> -			clear_highpage(new_page + (index % HPAGE_PMD_NR));
> -			index++;
> +
> +		xas_lock_irq(&xas);
> +		if (is_shmem)
> +			__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> +		if (nr_none) {
> +			__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> +			if (is_shmem)
> +				__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
>   		}
> +		/* Join all the small entries into a single multi-index entry. */
> +		xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> +		xas_store(&xas, new_page);
> +		xas_unlock_irq(&xas);
>   
>   		SetPageUptodate(new_page);
>   		page_ref_add(new_page, HPAGE_PMD_NR - 1);
> @@ -2012,9 +2024,11 @@ static void collapse_file(struct mm_struct *mm,
>   
>   		khugepaged_pages_collapsed++;
>   	} else {
> -		struct page *page;
> -
> -		/* Something went wrong: roll back page cache changes */
> +		/*
> +		 * Something went wrong:
> +		 * either result != SCAN_SUCCEED or copy_failed,
> +		 * roll back page cache changes
> +		 */
>   		xas_lock_irq(&xas);
>   		mapping->nrpages -= nr_none;
>   
> @@ -2047,6 +2061,15 @@ static void collapse_file(struct mm_struct *mm,
>   			xas_lock_irq(&xas);
>   		}
>   		VM_BUG_ON(nr_none);
> +		/*
> +		 * Undo the updates of thp_nr_pages(new_page) for non-SHMEM file,
> +		 * which is not updated yet for SHMEM file.
> +		 * These undos are not needed if result is not SCAN_SUCCEED.
> +		 */
> +		if (!is_shmem && result == SCAN_SUCCEED) {
> +			__mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
> +			filemap_nr_thps_dec(mapping);
> +		}
>   		xas_unlock_irq(&xas);
>   
>   		new_page->mapping = NULL;

Re: [RFC v1 2/2] mm: khugepaged: recover from poisoned file-backed memory

[Jiaqi Yan]

On Mon, Mar 28, 2022 at 9:02 PM Tong Tiangen <tongtiangen@huawei.com> wrote:
>
>
>
> 在 2022/3/24 7:29, Jiaqi Yan 写道:
> > Make collapse_file roll back when copying pages failed.
> > More concretely:
> > * extract copy operations into a separate loop
> > * postpone the updates for nr_none until both scan and copy succeeded
> > * postpone joining small xarray entries until both scan and copy
> >    succeeded
> > * as for update operations to NR_XXX_THPS
> >      * for SHMEM file, postpone until both scan and copy succeeded
> >      * for other file, roll back if scan succeeded but copy failed
> >
> > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> > ---
> >   include/linux/highmem.h | 18 ++++++++++
> >   mm/khugepaged.c         | 75 +++++++++++++++++++++++++++--------------
> >   2 files changed, 67 insertions(+), 26 deletions(-)
> >
> > diff --git a/include/linux/highmem.h b/include/linux/highmem.h
> > index 15d0aa4d349c..fc5aa221bdb5 100644
> > --- a/include/linux/highmem.h
> > +++ b/include/linux/highmem.h
> > @@ -315,6 +315,24 @@ static inline void copy_highpage(struct page *to, struct page *from)
> >       kunmap_local(vfrom);
> >   }
> >
> > +/*
> > + * Machine check exception handled version of copy_highpage.
> > + * Return true if copying page content failed; otherwise false.
> > + */
> > +static inline bool copy_highpage_mc(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);
> > +     kunmap_local(vto);
> > +     kunmap_local(vfrom);
> > +
> > +     return ret > 0;
> > +}
> > +
> >   #endif
> >
> >   static inline void memcpy_page(struct page *dst_page, size_t dst_off,
> > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > index 84ed177f56ff..ed2b1cd4bbc6 100644
> > --- a/mm/khugepaged.c
> > +++ b/mm/khugepaged.c
> > @@ -1708,12 +1708,13 @@ static void collapse_file(struct mm_struct *mm,
> >   {
> >       struct address_space *mapping = file->f_mapping;
> >       gfp_t gfp;
> > -     struct page *new_page;
> > +     struct page *new_page, *page, *tmp;
> >       pgoff_t index, end = start + HPAGE_PMD_NR;
> >       LIST_HEAD(pagelist);
> >       XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
> >       int nr_none = 0, result = SCAN_SUCCEED;
> >       bool is_shmem = shmem_file(file);
> > +     bool copy_failed = false;
> >       int nr;
> >
> >       VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
> > @@ -1936,9 +1937,7 @@ static void collapse_file(struct mm_struct *mm,
> >       }
> >       nr = thp_nr_pages(new_page);
> >
> > -     if (is_shmem)
> > -             __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> > -     else {
> > +     if (!is_shmem) {
> >               __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
> >               filemap_nr_thps_inc(mapping);
> >               /*
> > @@ -1956,34 +1955,39 @@ static void collapse_file(struct mm_struct *mm,
> >               }
> >       }
> >
> > -     if (nr_none) {
> > -             __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> > -             if (is_shmem)
> > -                     __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> > -     }
> > -
> > -     /* Join all the small entries into a single multi-index entry */
> > -     xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> > -     xas_store(&xas, new_page);
> >   xa_locked:
> >       xas_unlock_irq(&xas);
> >   xa_unlocked:
> >
> >       if (result == SCAN_SUCCEED) {
> > -             struct page *page, *tmp;
> > -
> >               /*
> >                * Replacing old pages with new one has succeeded, now we
> > -              * need to copy the content and free the old pages.
> > +              * attempt to copy the contents.
> >                */
> >               index = start;
> > -             list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> > +             list_for_each_entry(page, &pagelist, lru) {
> >                       while (index < page->index) {
> >                               clear_highpage(new_page + (index % HPAGE_PMD_NR));
> >                               index++;
> >                       }
> > -                     copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
> > -                                     page);
> > +                     if (copy_highpage_mc(new_page + (page->index % HPAGE_PMD_NR), page)) {
> > +                             copy_failed = true;
>
> The 1st patch here used "copy_succeed = false", It is best that the
> logic of the two positions can be unified.

copy_failed here will be eliminated once we have SCAN_COPY_MC
defined in version 2. Version 2 also renames "copy_succeeded"
in collapse_huge_page() to "copied", mimicking the "isolated" variable
for __collapse_huge_page_isolate().


>
> > +                             break;
> > +                     }
> > +                     index++;
> > +             }
> > +             while (!copy_failed && index < end) {
> > +                     clear_highpage(new_page + (page->index % HPAGE_PMD_NR));
> > +                     index++;
> > +             }
> > +     }
> > +
> > +     if (result == SCAN_SUCCEED && !copy_failed) {
> > +             /*
> > +              * Copying old pages to huge one has succeeded, now we
> > +              * need to free the old pages.
> > +              */
> > +             list_for_each_entry_safe(page, tmp, &pagelist, lru) {
> >                       list_del(&page->lru);
> >                       page->mapping = NULL;
> >                       page_ref_unfreeze(page, 1);
> > @@ -1991,12 +1995,20 @@ static void collapse_file(struct mm_struct *mm,
> >                       ClearPageUnevictable(page);
> >                       unlock_page(page);
> >                       put_page(page);
> > -                     index++;
> >               }
> > -             while (index < end) {
> > -                     clear_highpage(new_page + (index % HPAGE_PMD_NR));
> > -                     index++;
> > +
> > +             xas_lock_irq(&xas);
> > +             if (is_shmem)
> > +                     __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
> > +             if (nr_none) {
> > +                     __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
> > +                     if (is_shmem)
> > +                             __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
> >               }
> > +             /* Join all the small entries into a single multi-index entry. */
> > +             xas_set_order(&xas, start, HPAGE_PMD_ORDER);
> > +             xas_store(&xas, new_page);
> > +             xas_unlock_irq(&xas);
> >
> >               SetPageUptodate(new_page);
> >               page_ref_add(new_page, HPAGE_PMD_NR - 1);
> > @@ -2012,9 +2024,11 @@ static void collapse_file(struct mm_struct *mm,
> >
> >               khugepaged_pages_collapsed++;
> >       } else {
> > -             struct page *page;
> > -
> > -             /* Something went wrong: roll back page cache changes */
> > +             /*
> > +              * Something went wrong:
> > +              * either result != SCAN_SUCCEED or copy_failed,
> > +              * roll back page cache changes
> > +              */
> >               xas_lock_irq(&xas);
> >               mapping->nrpages -= nr_none;
> >
> > @@ -2047,6 +2061,15 @@ static void collapse_file(struct mm_struct *mm,
> >                       xas_lock_irq(&xas);
> >               }
> >               VM_BUG_ON(nr_none);
> > +             /*
> > +              * Undo the updates of thp_nr_pages(new_page) for non-SHMEM file,
> > +              * which is not updated yet for SHMEM file.
> > +              * These undos are not needed if result is not SCAN_SUCCEED.
> > +              */
> > +             if (!is_shmem && result == SCAN_SUCCEED) {
> > +                     __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
> > +                     filemap_nr_thps_dec(mapping);
> > +             }
> >               xas_unlock_irq(&xas);
> >
> >               new_page->mapping = NULL;

Re: [RFC v1 0/2] Memory poison recovery in khugepaged

[Yang Shi]

On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> Problem
> =======
> Memory DIMMs are subject to multi-bit flips, i.e. memory errors.
> As memory size and density increase, the chances of and number of
> memory errors increase. The increasing size and density of server
> RAM in the data center and cloud have shown increased uncorrectable
> memory errors. There are already mechanisms in the kernel to recover
> from uncorrectable memory errors. This series of patches provides
> the recovery mechanism for the particular kernel agent khugepaged.
>
> Impact
> ======
> The main reason we chose to make khugepaged tolerant of memory failures
> was its high possibility of accessing poisoned memory while performing
> functionally optional compaction actions. Standard applications
> typically don't have strict requirements on the size of its pages.
> So they are given 4K pages by the kernel. The kernel is able to improve
> application performance by either 1) giving application 2M pages
> to begin with, or 2) collapsing 4K pages into 2M pages when possible.
> This collapsing operation is done by khugepaged, a kernel agent that
> is constantly scanning memory. When collapsing 4K pages into a 2M page,
> it must copy the data from the 4K pages into a physically contiguous
> 2M page. Therefore, as long as there exists one poisoned cache line in
> collapsible 4K pages, khugepaged will eventually access it. The current
> impact to users is a machine check exception triggered kernel panic.
> However, khugepaged’s compaction operations are not functionally required
> kernel actions. Therefore making khugepaged tolerant to poisoned memory
> will greatly improve user experience.
>
> Solution
> ========
> As stated before, it is less desirable to crash the system only because
> khugepaged accesses poisoned pages while it is collapsing 4K pages.
> The high level idea of this patch series is to skip the group of pages
> (usually 512 4K-size pages) once khugepaged finds one of them is poisoned,
> as these pages have become ineligible to be collapsed.
>
> We are also careful to unwind operations khuagepaged has performed before
> it detects memory failures. For example, before copying and collapsing
> a group of anonymous pages into a huge page, the source pages will be
> isolated and their page table is unlinked from their PMD. These operations
> need to be undone in order to ensure these pages are not changed/lost from
> the perspective of other threads (both user and kernel space). As for
> file backed memory pages, there already exists a rollback case. This
> patch just extends it so that khugepaged also correctly rolls back when
> it fails to copy poisoned 4K pages.

Actually I should asked the question in the first place before diving
into the implementation details, if uncorrectable memory error
happens, kernel will pin the poisoned page and set hwpoison flag, the
bumped page refcount would prevent the page from being collapsed IIUC.
So I'm wondering why we need this?

>
> Jiaqi Yan (2):
>   mm: khugepaged: recover from poisoned anonymous memory
>   mm: khugepaged: recover from poisoned file-backed memory
>
>  include/linux/highmem.h |  37 +++++++
>  mm/khugepaged.c         | 211 +++++++++++++++++++++++++++++-----------
>  2 files changed, 189 insertions(+), 59 deletions(-)
>
> --
> 2.35.1.894.gb6a874cedc-goog
>

Re: [RFC v1 0/2] Memory poison recovery in khugepaged

[Jiaqi Yan]

On Thu, Mar 24, 2022 at 7:51 PM Yang Shi <shy828301@gmail.com> wrote:
>
> On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> >
> > Problem
> > =======
> > Memory DIMMs are subject to multi-bit flips, i.e. memory errors.
> > As memory size and density increase, the chances of and number of
> > memory errors increase. The increasing size and density of server
> > RAM in the data center and cloud have shown increased uncorrectable
> > memory errors. There are already mechanisms in the kernel to recover
> > from uncorrectable memory errors. This series of patches provides
> > the recovery mechanism for the particular kernel agent khugepaged.
> >
> > Impact
> > ======
> > The main reason we chose to make khugepaged tolerant of memory failures
> > was its high possibility of accessing poisoned memory while performing
> > functionally optional compaction actions. Standard applications
> > typically don't have strict requirements on the size of its pages.
> > So they are given 4K pages by the kernel. The kernel is able to improve
> > application performance by either 1) giving application 2M pages
> > to begin with, or 2) collapsing 4K pages into 2M pages when possible.
> > This collapsing operation is done by khugepaged, a kernel agent that
> > is constantly scanning memory. When collapsing 4K pages into a 2M page,
> > it must copy the data from the 4K pages into a physically contiguous
> > 2M page. Therefore, as long as there exists one poisoned cache line in
> > collapsible 4K pages, khugepaged will eventually access it. The current
> > impact to users is a machine check exception triggered kernel panic.
> > However, khugepaged’s compaction operations are not functionally required
> > kernel actions. Therefore making khugepaged tolerant to poisoned memory
> > will greatly improve user experience.
> >
> > Solution
> > ========
> > As stated before, it is less desirable to crash the system only because
> > khugepaged accesses poisoned pages while it is collapsing 4K pages.
> > The high level idea of this patch series is to skip the group of pages
> > (usually 512 4K-size pages) once khugepaged finds one of them is poisoned,
> > as these pages have become ineligible to be collapsed.
> >
> > We are also careful to unwind operations khuagepaged has performed before
> > it detects memory failures. For example, before copying and collapsing
> > a group of anonymous pages into a huge page, the source pages will be
> > isolated and their page table is unlinked from their PMD. These operations
> > need to be undone in order to ensure these pages are not changed/lost from
> > the perspective of other threads (both user and kernel space). As for
> > file backed memory pages, there already exists a rollback case. This
> > patch just extends it so that khugepaged also correctly rolls back when
> > it fails to copy poisoned 4K pages.
>
> Actually I should asked the question in the first place before diving
> into the implementation details, if uncorrectable memory error
> happens, kernel will pin the poisoned page and set hwpoison flag, the
> bumped page refcount would prevent the page from being collapsed IIUC.

This patch series is for cases where khugepaged is the first guy that detects
the memory errors on these poisoned pages. IOW, the pages are not known to
have memory errors when khugepaged collapsing gets to them.
In our observation, this happens frequently when the huge page ratio of
the system is relatively low, which is fairly common in cloud VMs.



> So I'm wondering why we need this?
>
> >
> > Jiaqi Yan (2):
> >   mm: khugepaged: recover from poisoned anonymous memory
> >   mm: khugepaged: recover from poisoned file-backed memory
> >
> >  include/linux/highmem.h |  37 +++++++
> >  mm/khugepaged.c         | 211 +++++++++++++++++++++++++++++-----------
> >  2 files changed, 189 insertions(+), 59 deletions(-)
> >
> > --
> > 2.35.1.894.gb6a874cedc-goog
> >

Re: [RFC v1 0/2] Memory poison recovery in khugepaged

[Yang Shi]

On Fri, Mar 25, 2022 at 2:11 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> On Thu, Mar 24, 2022 at 7:51 PM Yang Shi <shy828301@gmail.com> wrote:
> >
> > On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> > >
> > > Problem
> > > =======
> > > Memory DIMMs are subject to multi-bit flips, i.e. memory errors.
> > > As memory size and density increase, the chances of and number of
> > > memory errors increase. The increasing size and density of server
> > > RAM in the data center and cloud have shown increased uncorrectable
> > > memory errors. There are already mechanisms in the kernel to recover
> > > from uncorrectable memory errors. This series of patches provides
> > > the recovery mechanism for the particular kernel agent khugepaged.
> > >
> > > Impact
> > > ======
> > > The main reason we chose to make khugepaged tolerant of memory failures
> > > was its high possibility of accessing poisoned memory while performing
> > > functionally optional compaction actions. Standard applications
> > > typically don't have strict requirements on the size of its pages.
> > > So they are given 4K pages by the kernel. The kernel is able to improve
> > > application performance by either 1) giving application 2M pages
> > > to begin with, or 2) collapsing 4K pages into 2M pages when possible.
> > > This collapsing operation is done by khugepaged, a kernel agent that
> > > is constantly scanning memory. When collapsing 4K pages into a 2M page,
> > > it must copy the data from the 4K pages into a physically contiguous
> > > 2M page. Therefore, as long as there exists one poisoned cache line in
> > > collapsible 4K pages, khugepaged will eventually access it. The current
> > > impact to users is a machine check exception triggered kernel panic.
> > > However, khugepaged’s compaction operations are not functionally required
> > > kernel actions. Therefore making khugepaged tolerant to poisoned memory
> > > will greatly improve user experience.
> > >
> > > Solution
> > > ========
> > > As stated before, it is less desirable to crash the system only because
> > > khugepaged accesses poisoned pages while it is collapsing 4K pages.
> > > The high level idea of this patch series is to skip the group of pages
> > > (usually 512 4K-size pages) once khugepaged finds one of them is poisoned,
> > > as these pages have become ineligible to be collapsed.
> > >
> > > We are also careful to unwind operations khuagepaged has performed before
> > > it detects memory failures. For example, before copying and collapsing
> > > a group of anonymous pages into a huge page, the source pages will be
> > > isolated and their page table is unlinked from their PMD. These operations
> > > need to be undone in order to ensure these pages are not changed/lost from
> > > the perspective of other threads (both user and kernel space). As for
> > > file backed memory pages, there already exists a rollback case. This
> > > patch just extends it so that khugepaged also correctly rolls back when
> > > it fails to copy poisoned 4K pages.
> >
> > Actually I should asked the question in the first place before diving
> > into the implementation details, if uncorrectable memory error
> > happens, kernel will pin the poisoned page and set hwpoison flag, the
> > bumped page refcount would prevent the page from being collapsed IIUC.
>
> This patch series is for cases where khugepaged is the first guy that detects
> the memory errors on these poisoned pages. IOW, the pages are not known to
> have memory errors when khugepaged collapsing gets to them.
> In our observation, this happens frequently when the huge page ratio of
> the system is relatively low, which is fairly common in cloud VMs.

Thanks, this is the very important information that needs to be caught
in the 1st patch's commit log.

>
>
>
> > So I'm wondering why we need this?
> >
> > >
> > > Jiaqi Yan (2):
> > >   mm: khugepaged: recover from poisoned anonymous memory
> > >   mm: khugepaged: recover from poisoned file-backed memory
> > >
> > >  include/linux/highmem.h |  37 +++++++
> > >  mm/khugepaged.c         | 211 +++++++++++++++++++++++++++++-----------
> > >  2 files changed, 189 insertions(+), 59 deletions(-)
> > >
> > > --
> > > 2.35.1.894.gb6a874cedc-goog
> > >

Re: [RFC v1 0/2] Memory poison recovery in khugepaged

[Jiaqi Yan]

On Fri, Mar 25, 2022 at 2:42 PM Yang Shi <shy828301@gmail.com> wrote:
>
> On Fri, Mar 25, 2022 at 2:11 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> >
> > On Thu, Mar 24, 2022 at 7:51 PM Yang Shi <shy828301@gmail.com> wrote:
> > >
> > > On Wed, Mar 23, 2022 at 4:29 PM Jiaqi Yan <jiaqiyan@google.com> wrote:
> > > >
> > > > Problem
> > > > =======
> > > > Memory DIMMs are subject to multi-bit flips, i.e. memory errors.
> > > > As memory size and density increase, the chances of and number of
> > > > memory errors increase. The increasing size and density of server
> > > > RAM in the data center and cloud have shown increased uncorrectable
> > > > memory errors. There are already mechanisms in the kernel to recover
> > > > from uncorrectable memory errors. This series of patches provides
> > > > the recovery mechanism for the particular kernel agent khugepaged.
> > > >
> > > > Impact
> > > > ======
> > > > The main reason we chose to make khugepaged tolerant of memory failures
> > > > was its high possibility of accessing poisoned memory while performing
> > > > functionally optional compaction actions. Standard applications
> > > > typically don't have strict requirements on the size of its pages.
> > > > So they are given 4K pages by the kernel. The kernel is able to improve
> > > > application performance by either 1) giving application 2M pages
> > > > to begin with, or 2) collapsing 4K pages into 2M pages when possible.
> > > > This collapsing operation is done by khugepaged, a kernel agent that
> > > > is constantly scanning memory. When collapsing 4K pages into a 2M page,
> > > > it must copy the data from the 4K pages into a physically contiguous
> > > > 2M page. Therefore, as long as there exists one poisoned cache line in
> > > > collapsible 4K pages, khugepaged will eventually access it. The current
> > > > impact to users is a machine check exception triggered kernel panic.
> > > > However, khugepaged’s compaction operations are not functionally required
> > > > kernel actions. Therefore making khugepaged tolerant to poisoned memory
> > > > will greatly improve user experience.
> > > >
> > > > Solution
> > > > ========
> > > > As stated before, it is less desirable to crash the system only because
> > > > khugepaged accesses poisoned pages while it is collapsing 4K pages.
> > > > The high level idea of this patch series is to skip the group of pages
> > > > (usually 512 4K-size pages) once khugepaged finds one of them is poisoned,
> > > > as these pages have become ineligible to be collapsed.
> > > >
> > > > We are also careful to unwind operations khuagepaged has performed before
> > > > it detects memory failures. For example, before copying and collapsing
> > > > a group of anonymous pages into a huge page, the source pages will be
> > > > isolated and their page table is unlinked from their PMD. These operations
> > > > need to be undone in order to ensure these pages are not changed/lost from
> > > > the perspective of other threads (both user and kernel space). As for
> > > > file backed memory pages, there already exists a rollback case. This
> > > > patch just extends it so that khugepaged also correctly rolls back when
> > > > it fails to copy poisoned 4K pages.
> > >
> > > Actually I should asked the question in the first place before diving
> > > into the implementation details, if uncorrectable memory error
> > > happens, kernel will pin the poisoned page and set hwpoison flag, the
> > > bumped page refcount would prevent the page from being collapsed IIUC.
> >
> > This patch series is for cases where khugepaged is the first guy that detects
> > the memory errors on these poisoned pages. IOW, the pages are not known to
> > have memory errors when khugepaged collapsing gets to them.
> > In our observation, this happens frequently when the huge page ratio of
> > the system is relatively low, which is fairly common in cloud VMs.
>
> Thanks, this is the very important information that needs to be caught
> in the 1st patch's commit log.

Thanks for this valuable feedback. I will add this in the commit msg of v2,
but I will wait for your comments on patch 2/2 before sending out v2.

>
> >
> >
> >
> > > So I'm wondering why we need this?
> > >
> > > >
> > > > Jiaqi Yan (2):
> > > >   mm: khugepaged: recover from poisoned anonymous memory
> > > >   mm: khugepaged: recover from poisoned file-backed memory
> > > >
> > > >  include/linux/highmem.h |  37 +++++++
> > > >  mm/khugepaged.c         | 211 +++++++++++++++++++++++++++++-----------
> > > >  2 files changed, 189 insertions(+), 59 deletions(-)
> > > >
> > > > --
> > > > 2.35.1.894.gb6a874cedc-goog
> > > >