Skip to contentSkip to conversation

mm: hugetlb: fix a race between memory-failure/soft_offline and gather_surplus_pages

PATCH In review Muchun Song 6 participants 20 messages Apr 21, 2021 – May 7, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm: hugetlb: fix a race between memory-failure/soft_offline and gather_surplus_pages

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

---
 mm/hugetlb.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)
mm/hugetlb.c +4 -7
index 3476aa06da70..6c96332db34b 100644
@@ -2145,17 +2145,14 @@ static int gather_surplus_pages(struct hstate *h, long delta)
 
 	/* Free the needed pages to the hugetlb pool */
 	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
-		int zeroed;
-
 		if ((--needed) < 0)
 			break;
 		/*
-		 * This page is now managed by the hugetlb allocator and has
-		 * no users -- drop the buddy allocator's reference.
+		 * The refcount can possibly be increased by memory-failure or
+		 * soft_offline handlers.
 		 */
-		zeroed = put_page_testzero(page);
-		VM_BUG_ON_PAGE(!zeroed, page);
-		enqueue_huge_page(h, page);
+		if (likely(put_page_testzero(page)))
+			enqueue_huge_page(h, page);
 	}
 free:
 	spin_unlock_irq(&hugetlb_lock);
Signature · 1 line
2.11.0
#1 lore raw

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Quoted 34 lines
---
 mm/hugetlb.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)
mm/hugetlb.c +4 -7
index 3476aa06da70..6c96332db34b 100644
@@ -2145,17 +2145,14 @@ static int gather_surplus_pages(struct hstate *h, long delta)
 
 	/* Free the needed pages to the hugetlb pool */
 	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
-		int zeroed;
-
 		if ((--needed) < 0)
 			break;
 		/*
-		 * This page is now managed by the hugetlb allocator and has
-		 * no users -- drop the buddy allocator's reference.
+		 * The refcount can possibly be increased by memory-failure or
+		 * soft_offline handlers.
 		 */
-		zeroed = put_page_testzero(page);
-		VM_BUG_ON_PAGE(!zeroed, page);
-		enqueue_huge_page(h, page);
+		if (likely(put_page_testzero(page)))
+			enqueue_huge_page(h, page);
 	}
 free:
 	spin_unlock_irq(&hugetlb_lock);
Signature · 1 line
2.11.0
Signature · 2 lines
Michal Hocko
SUSE Labs
#2 lore raw
Quoted 23 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

Quoted 5 lines

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

If the page has been enqueued into the pool, then the page can be allocated to other users. The page reference count will be reset to 1 in the dequeue_huge_page_node_exact(). Then memory-failure will free the page because of put_page(). This is wrong. Because there is another user.

Quoted 39 lines
Quoted 33 lines
---
 mm/hugetlb.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)
mm/hugetlb.c +4 -8
index 3476aa06da70..6c96332db34b 100644
@@ -2145,17 +2145,14 @@ static int gather_surplus_pages(struct hstate *h, long delta)
 
      /* Free the needed pages to the hugetlb pool */
      list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
-             int zeroed;
-
              if ((--needed) < 0)
                      break;
              /*
-              * This page is now managed by the hugetlb allocator and has
-              * no users -- drop the buddy allocator's reference.
+              * The refcount can possibly be increased by memory-failure or
+              * soft_offline handlers.
               */
-             zeroed = put_page_testzero(page);
-             VM_BUG_ON_PAGE(!zeroed, page);
-             enqueue_huge_page(h, page);
+             if (likely(put_page_testzero(page)))
+                     enqueue_huge_page(h, page);
      }
 free:
      spin_unlock_irq(&hugetlb_lock);
--
Quoted 2 lines

2.11.0

--
Michal Hocko
SUSE Labs

maintainer
#3 lore raw
Quoted 11 lines · Muchun Song
Quoted 5 lines

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

If the page has been enqueued into the pool, then the page can be allocated to other users. The page reference count will be reset to 1 in the dequeue_huge_page_node_exact(). Then memory-failure will free the page because of put_page(). This is wrong. Because there is another user.

Note that dequeue_huge_page_node_exact() will not hand over any pages which are poisoned, so in this case it will not be allocated. But it is true that we might need hugetlb lock, this needs some more thought.

I will have a look.

Signature · 2 lines
Oscar Salvador
SUSE L3
#4 lore raw
Quoted 8 lines · Oscar Salvador
Quoted 6 lines · Muchun Song
Quoted 5 lines

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Quoted 8 lines
Quoted 5 lines

If the page has been enqueued into the pool, then the page can be allocated to other users. The page reference count will be reset to 1 in the dequeue_huge_page_node_exact(). Then memory-failure will free the page because of put_page(). This is wrong. Because there is another user.

Note that dequeue_huge_page_node_exact() will not hand over any pages which are poisoned, so in this case it will not be allocated.

But softoffline does not set page hwpoison before
__get_hwpoison_page(). So the page still can be
allocated. Right?

Quoted 8 lines

But it is true that we might need hugetlb lock, this needs some more thought.

I will have a look.

--
Oscar Salvador
SUSE L3

maintainer
#5 lore raw
Quoted 3 lines · Muchun Song

But softoffline does not set page hwpoison before
__get_hwpoison_page(). So the page still can be
allocated. Right?

Yep, soft_offline() only marks the page as hwpoison once the page has been fully contended and no other use is possible.
But yeah, hugetlb is a bit trickier in that regard.

This needs fixing in there.

Signature · 2 lines
Oscar Salvador
SUSE L3
#6 lore raw
Quoted 7 lines · Oscar Salvador
Quoted 3 lines · Muchun Song

But softoffline does not set page hwpoison before
__get_hwpoison_page(). So the page still can be
allocated. Right?

Quoted 5 lines

Yep, soft_offline() only marks the page as hwpoison once the page has been fully contended and no other use is possible.
But yeah, hugetlb is a bit trickier in that regard.

This needs fixing in there.

It is OK to fix it in softoffline/memory-failure.
I just want to expose the race. Thanks.

Quoted 5 lines

--
Oscar Salvador
SUSE L3

#7 lore raw
In reply to Oscar Salvador
Quoted 7 lines · Oscar Salvador
Quoted 6 lines · Muchun Song
Quoted 5 lines

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Quoted 8 lines
Quoted 5 lines

If the page has been enqueued into the pool, then the page can be allocated to other users. The page reference count will be reset to 1 in the dequeue_huge_page_node_exact(). Then memory-failure will free the page because of put_page(). This is wrong. Because there is another user.

Note that dequeue_huge_page_node_exact() will not hand over any pages which are poisoned, so in this case it will not be allocated.

I have to say I have missed the HWPoison check so the this particular scenario is not possible indeed.

Quoted 2 lines

But it is true that we might need hugetlb lock, this needs some more thought.

yes, nobody should be touching to the reference count of hugetlb pool pages out of the hugetlb proper.

Quoted 1 line

I will have a look.

Thanks!

Signature · 2 lines
Michal Hocko
SUSE Labs
#8 lore raw
In reply to Muchun Song
Quoted 30 lines · Muchun Song
Quoted 23 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

Quoted 5 lines

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held). What would happen if the reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Quoted 5 lines

If the page has been enqueued into the pool, then the page can be allocated to other users. The page reference count will be reset to 1 in the dequeue_huge_page_node_exact(). Then memory-failure will free the page because of put_page(). This is wrong. Because there is another user.

Yes that is one of the scenarios but I suspect there are more lurking there. That was my point that this should be addressed at the hwpoison side.

Signature · 2 lines
Michal Hocko
SUSE Labs
#9 lore raw
In reply to Michal Hocko
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

Quoted 3 lines

What would happen if the
reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Yes, this is another concern.

Thanks,
Naoya Horiguchi

#10 lore raw
Quoted 28 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

Quoted 6 lines

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

HPageFreed() can indicate whether a page is on the free pool list.

Quoted 9 lines
Quoted 4 lines

What would happen if the
reference count was increased after the page has been enqueed into the pool? This can just blow up later.

Yes, this is another concern.

Thanks,
Naoya Horiguchi

#11 lore raw
In reply to HORIGUCHI NAOYA(堀口 直也)
Quoted 27 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

Quoted 7 lines

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

It seems we can also race with the code that makes a compound page a hugetlb page. The memory failure code could be called after allocating pages from buddy and before setting compound page DTOR. So, the memory handling code will process it as a compound page.

Just thinking that this may not be limited to the hugetlb specific memory failure handling?

Signature · 1 line
Mike Kravetz
#12 lore raw
Quoted 36 lines · Mike Kravetz
Quoted 27 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

Quoted 7 lines

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

Quoted 4 lines

It seems we can also race with the code that makes a compound page a hugetlb page. The memory failure code could be called after allocating pages from buddy and before setting compound page DTOR. So, the memory handling code will process it as a compound page.

Yes, so get_hwpoison_page() has to call get_page_unless_zero() only when memory_failure() can surely handle the error.

Quoted 3 lines

Just thinking that this may not be limited to the hugetlb specific memory failure handling?

Currently hugetlb page is the only type of compound page supported by memory failure. But I agree with you that other types of compound pages have the same race window, and judging only with get_page_unless_zero() is dangerous. So I think that __get_hwpoison_page() should have the following structure:

  if (PageCompound) {
      if (PageHuge) {
          if (PageHugeFreed || PageHugeActive) {
              if (get_page_unless_zero)
                  return 0;   // path for in-use hugetlb page
              else
                  return 1;   // path for free hugetlb page
          } else {
              return -EBUSY;  // any transient hugetlb page
          }
      } else {
          ... // any other compound page (like thp, slab, ...)
      }
  } else {
      ...   // any non-compound page
  }

Thanks,
Naoya Horiguchi

#13 lore raw
Quoted 54 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 36 lines · Mike Kravetz
Quoted 27 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

Quoted 7 lines

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

Quoted 4 lines

It seems we can also race with the code that makes a compound page a hugetlb page. The memory failure code could be called after allocating pages from buddy and before setting compound page DTOR. So, the memory handling code will process it as a compound page.

Yes, so get_hwpoison_page() has to call get_page_unless_zero() only when memory_failure() can surely handle the error.

Quoted 3 lines

Just thinking that this may not be limited to the hugetlb specific memory failure handling?

Currently hugetlb page is the only type of compound page supported by memory failure. But I agree with you that other types of compound pages have the same race window, and judging only with get_page_unless_zero() is dangerous. So I think that __get_hwpoison_page() should have the following structure:

Quoted 16 lines
  if (PageCompound) {
      if (PageHuge) {
          if (PageHugeFreed || PageHugeActive) {
              if (get_page_unless_zero)
                  return 0;   // path for in-use hugetlb page
              else
                  return 1;   // path for free hugetlb page
          } else {
              return -EBUSY;  // any transient hugetlb page
          }
      } else {
          ... // any other compound page (like thp, slab, ...)
      }
  } else {
      ...   // any non-compound page
  }

The above pseudo code was wrong, so let me update my thought. I'm now trying to solve the reported issue by changing __get_hwpoison_page() like below:

  static int __get_hwpoison_page(struct page *page)
  {
          struct page *head = compound_head(page);
          if (PageCompound(page)) {
                  if (PageSlab(page)) {
                          return get_page_unless_zero(page);
                  } else if (PageHuge(head)) {
                          if (HPageFreed(head) || HPageMigratable(head))
                                  return get_page_unless_zero(head);
                  } else if (PageTransHuge(head)) {
                          /*
                           * Non anonymous thp exists only in allocation/free time. We
                           * can't handle such a case correctly, so let's give it up.
                           * This should be better than triggering BUG_ON when kernel
                           * tries to touch the "partially handled" page.
                           */
                          if (!PageAnon(head)) {
                                  pr_err("Memory failure: %#lx: non anonymous thp\n",
                                         page_to_pfn(page));
                                  return 0;
                          }
                          if (get_page_unless_zero(head)) {
                                  if (head == compound_head(page))
                                          return 1;
                                  pr_info("Memory failure: %#lx cannot catch tail\n",
                                          page_to_pfn(page));
                                  put_page(head);
                          }
                  }
                  return 0;
          }
  
          return get_page_unless_zero(page);

}

Some notes:

- in hugetlb path, new HPage* checks should avoid the reported race, but I still need more testing to confirm it,
- PageSlab check is added because otherwise I found that "non anonymous thp" path is chosen, that's obviously wrong,
- thp's branch has a known issue unrelated to the current issue, which will/should be improved later.

I'll send a patch next week.

Thanks,
Naoya Horiguchi

#14 lore raw

PATCH 1/1 mm,hwpoison: fix race with compound page allocation

Quoted 114 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 54 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 36 lines · Mike Kravetz
Quoted 27 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 18 lines · Michal Hocko

[Cc Naoya]

Quoted 15 lines · Muchun Song

The possible bad scenario:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines
Quoted 3 lines

The refcount can possibly be increased by memory-failure or soft_offline handlers, we can trigger VM_BUG_ON_PAGE and wrongly add the page to the hugetlb pool list.

The hwpoison side of this looks really suspicious to me. It shouldn't really touch the reference count of hugetlb pages without being very careful (and having hugetlb_lock held).

Quoted 7 lines

I have the same feeling, there is a window where a hugepage is refcounted during converting from buddy free pages into free hugepage, so refcount alone is not enough to prevent the race. hugetlb_lock is retaken after alloc_surplus_huge_page returns, so simply holding hugetlb_lock in get_hwpoison_page() seems not work. Is there any status bit to show that a hugepage is just being initialized (not in free hugepage pool or in use)?

Quoted 4 lines

It seems we can also race with the code that makes a compound page a hugetlb page. The memory failure code could be called after allocating pages from buddy and before setting compound page DTOR. So, the memory handling code will process it as a compound page.

Yes, so get_hwpoison_page() has to call get_page_unless_zero() only when memory_failure() can surely handle the error.

Quoted 3 lines

Just thinking that this may not be limited to the hugetlb specific memory failure handling?

Currently hugetlb page is the only type of compound page supported by memory failure. But I agree with you that other types of compound pages have the same race window, and judging only with get_page_unless_zero() is dangerous. So I think that __get_hwpoison_page() should have the following structure:

Quoted 16 lines
  if (PageCompound) {
      if (PageHuge) {
          if (PageHugeFreed || PageHugeActive) {
              if (get_page_unless_zero)
                  return 0;   // path for in-use hugetlb page
              else
                  return 1;   // path for free hugetlb page
          } else {
              return -EBUSY;  // any transient hugetlb page
          }
      } else {
          ... // any other compound page (like thp, slab, ...)
      }
  } else {
      ...   // any non-compound page
  }

The above pseudo code was wrong, so let me update my thought. I'm now trying to solve the reported issue by changing __get_hwpoison_page() like below:

  static int __get_hwpoison_page(struct page *page)
  {
          struct page *head = compound_head(page);
          if (PageCompound(page)) {
                  if (PageSlab(page)) {
                          return get_page_unless_zero(page);
                  } else if (PageHuge(head)) {
                          if (HPageFreed(head) || HPageMigratable(head))
                                  return get_page_unless_zero(head);
                  } else if (PageTransHuge(head)) {
                          /*
                           * Non anonymous thp exists only in allocation/free time. We
                           * can't handle such a case correctly, so let's give it up.
                           * This should be better than triggering BUG_ON when kernel
                           * tries to touch the "partially handled" page.
                           */
                          if (!PageAnon(head)) {
                                  pr_err("Memory failure: %#lx: non anonymous thp\n",
                                         page_to_pfn(page));
                                  return 0;
                          }
                          if (get_page_unless_zero(head)) {
                                  if (head == compound_head(page))
                                          return 1;
                                  pr_info("Memory failure: %#lx cannot catch tail\n",
                                          page_to_pfn(page));
                                  put_page(head);
                          }
                  }
                  return 0;
          }
  
          return get_page_unless_zero(page);

}

Some notes:

Quoted 8 lines

- in hugetlb path, new HPage* checks should avoid the reported race, but I still need more testing to confirm it,
- PageSlab check is added because otherwise I found that "non anonymous thp" path is chosen, that's obviously wrong,
- thp's branch has a known issue unrelated to the current issue, which will/should be improved later.

I'll send a patch next week.

I confirmed that the patch fixes the reported problem (in the testcase triggering VM_BUG_ON_PAGE() without this patch).
So let me suggest this as a fix on hwpoison side.

Thanks,
Naoya Horiguchi

---
From: Naoya Horiguchi <[email protected]>
Date: Wed, 28 Apr 2021 15:55:47 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen for slab pages.

---
 mm/memory-failure.c | 48 +++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 21 deletions(-)
mm/memory-failure.c +27 -21
index a3659619d293..61988e332712 100644
@@ -1095,30 +1095,36 @@ static int __get_hwpoison_page(struct page *page)
 {
 	struct page *head = compound_head(page);
 
-	if (!PageHuge(head) && PageTransHuge(head)) {
-		/*
-		 * Non anonymous thp exists only in allocation/free time. We
-		 * can't handle such a case correctly, so let's give it up.
-		 * This should be better than triggering BUG_ON when kernel
-		 * tries to touch the "partially handled" page.
-		 */
-		if (!PageAnon(head)) {
-			pr_err("Memory failure: %#lx: non anonymous thp\n",
-				page_to_pfn(page));
-			return 0;
+	if (PageCompound(page)) {
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			if (HPageFreed(head) || HPageMigratable(head))
+				return get_page_unless_zero(head);
+		} else if (PageTransHuge(head)) {
+			/*
+			 * Non anonymous thp exists only in allocation/free time. We
+			 * can't handle such a case correctly, so let's give it up.
+			 * This should be better than triggering BUG_ON when kernel
+			 * tries to touch the "partially handled" page.
+			 */
+			if (!PageAnon(head)) {
+				pr_err("Memory failure: %#lx: non anonymous thp\n",
+				       page_to_pfn(page));
+				return 0;
+			}
+			if (get_page_unless_zero(head)) {
+				if (head == compound_head(page))
+					return 1;
+				pr_info("Memory failure: %#lx cannot catch tail\n",
+					page_to_pfn(page));
+				put_page(head);
+			}
 		}
+		return 0;
 	}
 
-	if (get_page_unless_zero(head)) {
-		if (head == compound_head(page))
-			return 1;
-
-		pr_info("Memory failure: %#lx cannot catch tail\n",
-			page_to_pfn(page));
-		put_page(head);
-	}
-
-	return 0;
+	return get_page_unless_zero(page);
 }
 
 /*
Signature · 1 line
2.25.1
maintainer
#15 lore raw

Re: [PATCH] mm,hwpoison: fix race with compound page allocation

Quoted 21 lines · Naoya Horiguchi

---
From: Naoya Horiguchi <[email protected]>
Date: Wed, 28 Apr 2021 15:55:47 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

Quoted 7 lines

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen for slab pages.

Was it wrongly chosen even before? If so, maybe a Fix tag is warranted.

Quoted 12 lines
---
 mm/memory-failure.c | 48 +++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 21 deletions(-)
mm/memory-failure.c +0 -0
index a3659619d293..61988e332712 100644

mm/memory-failure.c static int __get_hwpoison_page(struct page *page) · line 1101

Quoted 7 lines
mm/memory-failure.c +6 -0
@@ -1095,30 +1095,36 @@ static int __get_hwpoison_page(struct page *page)
 
+	if (PageCompound(page)) {
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			if (HPageFreed(head) || HPageMigratable(head))
+				return get_page_unless_zero(head);

There were concerns raised wrt. memory-failure should not be fiddling with page's refcount without holding a hugetlb lock.
So, if we really want to make this more stable, we might want to hold the lock here.

The clearing and setting of HPageFreed happens under the lock, and for HPageMigratable that is also true for the clearing part, so I think it would be more sane to do this under the lock to close any possible race.

Does it make sense?

Signature · 2 lines
Oscar Salvador
SUSE L3
#16 lore raw

Re: [PATCH] mm,hwpoison: fix race with compound page allocation

Quoted 26 lines · Oscar Salvador
Quoted 25 lines · Naoya Horiguchi

---
From: Naoya Horiguchi <[email protected]>
Date: Wed, 28 Apr 2021 15:55:47 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

Quoted 5 lines
Quoted 3 lines

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen for slab pages.

Was it wrongly chosen even before? If so, maybe a Fix tag is warranted.

OK, I'll check when this was introduced.

Quoted 26 lines
Quoted 13 lines
---
 mm/memory-failure.c | 48 +++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 21 deletions(-)
mm/memory-failure.c +0 -0
index a3659619d293..61988e332712 100644

mm/memory-failure.c static int __get_hwpoison_page(struct page *page) · line 1101

Quoted 7 lines
mm/memory-failure.c +6 -0
@@ -1095,30 +1095,36 @@ static int __get_hwpoison_page(struct page *page)
 
+	if (PageCompound(page)) {
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			if (HPageFreed(head) || HPageMigratable(head))
+				return get_page_unless_zero(head);

There were concerns raised wrt. memory-failure should not be fiddling with page's refcount without holding a hugetlb lock.
So, if we really want to make this more stable, we might want to hold the lock here.

Quoted 5 lines

The clearing and setting of HPageFreed happens under the lock, and for HPageMigratable that is also true for the clearing part, so I think it would be more sane to do this under the lock to close any possible race.

Does it make sense?

Thanks, I'll update to do the check under hugetlb_lock.

- Naoya Horiguchi

#17 lore raw

PATCH v2 1/1 mm,hwpoison: fix race with compound page allocation

Quoted 61 lines · HORIGUCHI NAOYA(堀口 直也)
Quoted 26 lines · Oscar Salvador
Quoted 25 lines · Naoya Horiguchi

---
From: Naoya Horiguchi <[email protected]>
Date: Wed, 28 Apr 2021 15:55:47 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

Quoted 5 lines
Quoted 3 lines

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen for slab pages.

Was it wrongly chosen even before? If so, maybe a Fix tag is warranted.

OK, I'll check when this was introduced.

Quoted 26 lines
Quoted 13 lines
---
 mm/memory-failure.c | 48 +++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 21 deletions(-)
mm/memory-failure.c +0 -0
index a3659619d293..61988e332712 100644

mm/memory-failure.c static int __get_hwpoison_page(struct page *page) · line 1101

Quoted 7 lines
mm/memory-failure.c +6 -0
@@ -1095,30 +1095,36 @@ static int __get_hwpoison_page(struct page *page)
 
+	if (PageCompound(page)) {
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			if (HPageFreed(head) || HPageMigratable(head))
+				return get_page_unless_zero(head);

There were concerns raised wrt. memory-failure should not be fiddling with page's refcount without holding a hugetlb lock.
So, if we really want to make this more stable, we might want to hold the lock here.

Quoted 7 lines
Quoted 5 lines

The clearing and setting of HPageFreed happens under the lock, and for HPageMigratable that is also true for the clearing part, so I think it would be more sane to do this under the lock to close any possible race.

Does it make sense?

Thanks, I'll update to do the check under hugetlb_lock.

Hi,

Let me share the update below. Two changes:
- hold hugetlb_lock in hugetlb path,
- added Fixes tag and cc to stable. I limited the stable branch only to 5.12+ due to the dependency on HPage* pseudo flags.

- Naoya

---
From: Naoya Horiguchi <[email protected]>
Date: Thu, 6 May 2021 09:54:39 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen.

Fixes ead07f6a867b ("mm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling")
---
 mm/memory-failure.c | 53 +++++++++++++++++++++++++++------------------
 1 file changed, 32 insertions(+), 21 deletions(-)
mm/memory-failure.c +32 -21
index a3659619d293..966a1d6b0bc8 100644
@@ -1095,30 +1095,41 @@ static int __get_hwpoison_page(struct page *page)
 {
 	struct page *head = compound_head(page);
 
-	if (!PageHuge(head) && PageTransHuge(head)) {
-		/*
-		 * Non anonymous thp exists only in allocation/free time. We
-		 * can't handle such a case correctly, so let's give it up.
-		 * This should be better than triggering BUG_ON when kernel
-		 * tries to touch the "partially handled" page.
-		 */
-		if (!PageAnon(head)) {
-			pr_err("Memory failure: %#lx: non anonymous thp\n",
-				page_to_pfn(page));
-			return 0;
+	if (PageCompound(page)) {
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			int ret = 0;
+
+			spin_lock(&hugetlb_lock);
+			if (HPageFreed(head) || HPageMigratable(head))
+				ret = get_page_unless_zero(head);
+			spin_unlock(&hugetlb_lock);
+			return ret;
+		} else if (PageTransHuge(head)) {
+			/*
+			 * Non anonymous thp exists only in allocation/free time. We
+			 * can't handle such a case correctly, so let's give it up.
+			 * This should be better than triggering BUG_ON when kernel
+			 * tries to touch the "partially handled" page.
+			 */
+			if (!PageAnon(head)) {
+				pr_err("Memory failure: %#lx: non anonymous thp\n",
+				       page_to_pfn(page));
+				return 0;
+			}
+			if (get_page_unless_zero(head)) {
+				if (head == compound_head(page))
+					return 1;
+				pr_info("Memory failure: %#lx cannot catch tail\n",
+					page_to_pfn(page));
+				put_page(head);
+			}
 		}
+		return 0;
 	}
 
-	if (get_page_unless_zero(head)) {
-		if (head == compound_head(page))
-			return 1;
-
-		pr_info("Memory failure: %#lx cannot catch tail\n",
-			page_to_pfn(page));
-		put_page(head);
-	}
-
-	return 0;
+	return get_page_unless_zero(page);
 }
 
 /*
Signature · 1 line
2.25.1
maintainer
#18 lore raw

Re: [PATCH v2] mm,hwpoison: fix race with compound page allocation

Quoted 24 lines · Naoya Horiguchi

From: Naoya Horiguchi <[email protected]>
Date: Thu, 6 May 2021 09:54:39 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

Quoted 8 lines

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen.

Fixes ead07f6a867b ("mm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling")

Hi Naoya,

thanks for the patch.
I have some concerns though, more below:

Quoted 24 lines
---
 mm/memory-failure.c | 53 +++++++++++++++++++++++++++------------------
 1 file changed, 32 insertions(+), 21 deletions(-)
mm/memory-failure.c +1 -11
index a3659619d293..966a1d6b0bc8 100644
@@ -1095,30 +1095,41 @@ static int __get_hwpoison_page(struct page *page)
 {
 	struct page *head = compound_head(page);
 
-	if (!PageHuge(head) && PageTransHuge(head)) {
-		/*
-		 * Non anonymous thp exists only in allocation/free time. We
-		 * can't handle such a case correctly, so let's give it up.
-		 * This should be better than triggering BUG_ON when kernel
-		 * tries to touch the "partially handled" page.
-		 */
-		if (!PageAnon(head)) {
-			pr_err("Memory failure: %#lx: non anonymous thp\n",
-				page_to_pfn(page));
-			return 0;
+	if (PageCompound(page)) {

mm/memory-failure.c static int __get_hwpoison_page(struct page *page) · line 1108

Quoted 10 lines
mm/memory-failure.c +10 -0
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			int ret = 0;
+
+			spin_lock(&hugetlb_lock);
+			if (HPageFreed(head) || HPageMigratable(head))
+				ret = get_page_unless_zero(head);
+			spin_unlock(&hugetlb_lock);
+			return ret;

Ok, I am probably overthinking this but should we re-check under the lock wehther the page is a hugetlb page?
My concern is, what would happen if:

CPU0 CPU1
__get_hwpoison_page
PageHuge(head) == T
dissolve hugetlb page
hugetlb_lock

In that case, by the time we get to check hugetlb flags, those checks might return false, and we do not get a refcount.
So, I guess my question is: Should we re-check under the lock, and if it is not, do a "goto try_to_get_ref" that starts right at the beginning, or goes directly to the get_page_unless_zero at the end (the former probably better)?

As I said, I might be overthinking this, but well.

Signature · 2 lines
Oscar Salvador
SUSE L3
#19 lore raw

Re: [PATCH v2] mm,hwpoison: fix race with compound page allocation

Quoted 85 lines · Oscar Salvador
Quoted 24 lines · Naoya Horiguchi

From: Naoya Horiguchi <[email protected]>
Date: Thu, 6 May 2021 09:54:39 +0900
Subject: [PATCH] mm,hwpoison: fix race with compound page allocation

When hugetlb page fault (under overcommiting situation) and memory_failure() race, VM_BUG_ON_PAGE() is triggered by the following race:

CPU0: CPU1:

gather_surplus_pages()
page = alloc_surplus_huge_page()
memory_failure_hugetlb()
get_hwpoison_page(page)
__get_hwpoison_page(page)
get_page_unless_zero(page)
zero = put_page_testzero(page) VM_BUG_ON_PAGE(!zero, page) enqueue_huge_page(h, page)
put_page(page)

__get_hwpoison_page() only checks page refcount before taking additional one for memory error handling, which is wrong because there's time windows where compound pages have non-zero refcount during initialization.

Quoted 8 lines

So makes __get_hwpoison_page() check more page status for a few types of compound pages. PageSlab() check is added because otherwise "non anonymous thp" path is wrongly chosen.

Fixes ead07f6a867b ("mm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling")

Hi Naoya,

thanks for the patch.
I have some concerns though, more below:

Quoted 24 lines
---
 mm/memory-failure.c | 53 +++++++++++++++++++++++++++------------------
 1 file changed, 32 insertions(+), 21 deletions(-)
mm/memory-failure.c +1 -11
index a3659619d293..966a1d6b0bc8 100644
@@ -1095,30 +1095,41 @@ static int __get_hwpoison_page(struct page *page)
 {
 	struct page *head = compound_head(page);
 
-	if (!PageHuge(head) && PageTransHuge(head)) {
-		/*
-		 * Non anonymous thp exists only in allocation/free time. We
-		 * can't handle such a case correctly, so let's give it up.
-		 * This should be better than triggering BUG_ON when kernel
-		 * tries to touch the "partially handled" page.
-		 */
-		if (!PageAnon(head)) {
-			pr_err("Memory failure: %#lx: non anonymous thp\n",
-				page_to_pfn(page));
-			return 0;
+	if (PageCompound(page)) {

mm/memory-failure.c static int __get_hwpoison_page(struct page *page) · line 1108

Quoted 10 lines
mm/memory-failure.c +10 -0
+		if (PageSlab(page)) {
+			return get_page_unless_zero(page);
+		} else if (PageHuge(head)) {
+			int ret = 0;
+
+			spin_lock(&hugetlb_lock);
+			if (HPageFreed(head) || HPageMigratable(head))
+				ret = get_page_unless_zero(head);
+			spin_unlock(&hugetlb_lock);
+			return ret;

Ok, I am probably overthinking this but should we re-check under the lock wehther the page is a hugetlb page?
My concern is, what would happen if:

CPU0 CPU1
__get_hwpoison_page
PageHuge(head) == T
dissolve hugetlb page
hugetlb_lock

Quoted 2 lines

In that case, by the time we get to check hugetlb flags, those checks might return false, and we do not get a refcount.

Thanks, we had better add rechecking as we do in dissolve_free_huge_page().

Quoted 4 lines

So, I guess my question is: Should we re-check under the lock, and if it is not, do a "goto try_to_get_ref" that starts right at the beginning, or goes directly to the get_page_unless_zero at the end (the former probably better)?

Yes, retry could work in this case. Looking at existing code, get_any_page() provides "retry" layer, but it's not called now by get_hwpoison_page() when called from memory_failure(). So I think of trying to adjust code and make get_hwpoison_page call get_any_page() instead of calling __get_hwpoison_page(() directly.

Thanks,
Naoya Horiguchi