Skip to contentSkip to conversation

mm/hugetlb_vmemmap: fix memory loads ordering

PATCH Superseded Yu Zhao 4 participants 9 messages Jan 7, 2025 – Jan 10, 2025

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Yu Zhao

author
#0 lore raw

PATCH 1/1 mm/hugetlb_vmemmap: fix memory loads ordering

Using x86_64 as an example, for a 32KB struct page[] area describing a 2MB hugeTLB, HVO reduces the area to 4KB by the following steps:
1. Split the (r/w vmemmap) PMD mapping the area into 512 (r/w) PTEs;
2. For the 8 PTEs mapping the area, remap PTE 1-7 to the page mapped by PTE 0, and at the same time change the permission from r/w to r/o;
3. Free the pages PTE 1-7 used to map, hence the reduction from 32KB to 4KB.

However, the following race can happen due to improperly memory loads ordering:
CPU 1 (HVO) CPU 2 (speculative PFN walker)

page_ref_freeze()
synchronize_rcu()
rcu_read_lock() page_is_fake_head() is false
vmemmap_remap_pte()
XXX: struct page[] becomes r/o

page_ref_unfreeze()

                                  page_ref_count() is not zero

                                  atomic_add_unless(&page->_refcount)
                                  XXX: try to modify r/o struct page[]

Specifically, page_is_fake_head() must be ordered after page_ref_count() on CPU 2 so that it can only return true for this case, to avoid the later attempt to modify r/o struct page[].

This patch adds the missing memory barrier and makes the tests on page_is_fake_head() and page_ref_count() done in the proper order.

Fixes bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")
---
 include/linux/page-flags.h | 2 +-
 include/linux/page_ref.h   | 8 ++++++--
 2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/page-flags.h +1 -1
index 691506bdf2c5..6b8ecf86f1b6 100644
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 	 * cold cacheline in some cases.
 	 */
 	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {
 		/*
 		 * We can safely access the field of the @page[1] with PG_head
 		 * because the @page is a compound page composed with at least
include/linux/page_ref.h +6 -2
index 8c236c651d1d..5becea98bd79 100644
@@ -233,8 +233,12 @@ static inline bool page_ref_add_unless(struct page *page, int nr, int u)
 	bool ret = false;
 
 	rcu_read_lock();
-	/* avoid writing to the vmemmap area being remapped */
-	if (!page_is_fake_head(page) && page_ref_count(page) != u)
+	/*
+	 * To avoid writing to the vmemmap area remapped into r/o in parallel,
+	 * the page_ref_count() test must precede the page_is_fake_head() test
+	 * so that test_bit_acquire() in the latter is ordered after the former.
+	 */
+	if (page_ref_count(page) != u && !page_is_fake_head(page))
 		ret = atomic_add_unless(&page->_refcount, nr, u);
 	rcu_read_unlock();
 
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
#1 lore raw
Quoted 65 lines · Yu Zhao

Using x86_64 as an example, for a 32KB struct page[] area describing a 2MB hugeTLB, HVO reduces the area to 4KB by the following steps:
1. Split the (r/w vmemmap) PMD mapping the area into 512 (r/w) PTEs;
2. For the 8 PTEs mapping the area, remap PTE 1-7 to the page mapped by PTE 0, and at the same time change the permission from r/w to r/o;
3. Free the pages PTE 1-7 used to map, hence the reduction from 32KB to 4KB.

However, the following race can happen due to improperly memory loads ordering:
CPU 1 (HVO) CPU 2 (speculative PFN walker)

page_ref_freeze()
synchronize_rcu()
rcu_read_lock() page_is_fake_head() is false
vmemmap_remap_pte()
XXX: struct page[] becomes r/o

page_ref_unfreeze()

                                 page_ref_count() is not zero

                                 atomic_add_unless(&page->_refcount)
                                 XXX: try to modify r/o struct page[]

Specifically, page_is_fake_head() must be ordered after page_ref_count() on CPU 2 so that it can only return true for this case, to avoid the later attempt to modify r/o struct page[].

This patch adds the missing memory barrier and makes the tests on page_is_fake_head() and page_ref_count() done in the proper order.

Fixes bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")

---
include/linux/page-flags.h | 2 +-
include/linux/page_ref.h | 8 ++++++--

2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/page-flags.h +1 -1
index 691506bdf2c5..6b8ecf86f1b6 100644
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 * cold cacheline in some cases.
 */
 	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {
 		/*
 		 * We can safely access the field of the @page[1] with PG_head
 		 * because the @page is a compound page composed with at least
include/linux/page_ref.h +0 -0
index 8c236c651d1d..5becea98bd79 100644
@@ -233,8 +233,12 @@ static inline bool page_ref_add_unless(struct page *page, int nr, int u)
 	bool ret = false;

include/linux/page_ref.h static inline bool page_ref_add_unless(struct page *page, int nr, int u) · line 241

Quoted 10 lines
include/linux/page_ref.h +6 -2
 
 	rcu_read_lock();
- 	/* avoid writing to the vmemmap area being remapped */
- 	if (!page_is_fake_head(page) && page_ref_count(page) != u)
+ 	/*
+	 * To avoid writing to the vmemmap area remapped into r/o in parallel,
+	 * the page_ref_count() test must precede the page_is_fake_head() test
+	 * so that test_bit_acquire() in the latter is ordered after the former.
+	 */
+ 	if (page_ref_count(page) != u && !page_is_fake_head(page))

IIUC, we need to insert a memory barrier between page_ref_count() and page_is_fake_head(). Specifically, accessing between page->_refcount and page->flags. So we should insert a read memory barrier here, right? But I saw you added an acquire barrier in page_fixed_fake_head(), I don't understand why an acquire barrier could stop the CPU reordering the accessing between them. What am I missing here?

Muchun,
Thanks.

Quoted 6 lines
include/linux/page_ref.h +0 -0
 		ret = atomic_add_unless(&page->_refcount, nr, u);
 	rcu_read_unlock();
 
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
author
#2 lore raw
Quoted 67 lines · Muchun Song
Quoted 67 lines · Yu Zhao

Using x86_64 as an example, for a 32KB struct page[] area describing a 2MB hugeTLB, HVO reduces the area to 4KB by the following steps:
1. Split the (r/w vmemmap) PMD mapping the area into 512 (r/w) PTEs;
2. For the 8 PTEs mapping the area, remap PTE 1-7 to the page mapped by PTE 0, and at the same time change the permission from r/w to r/o;
3. Free the pages PTE 1-7 used to map, hence the reduction from 32KB to 4KB.

However, the following race can happen due to improperly memory loads ordering:
CPU 1 (HVO) CPU 2 (speculative PFN walker)

page_ref_freeze()
synchronize_rcu()
rcu_read_lock() page_is_fake_head() is false
vmemmap_remap_pte()
XXX: struct page[] becomes r/o

page_ref_unfreeze()

                                 page_ref_count() is not zero

                                 atomic_add_unless(&page->_refcount)
                                 XXX: try to modify r/o struct page[]

Specifically, page_is_fake_head() must be ordered after page_ref_count() on CPU 2 so that it can only return true for this case, to avoid the later attempt to modify r/o struct page[].

This patch adds the missing memory barrier and makes the tests on page_is_fake_head() and page_ref_count() done in the proper order.

Fixes bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")

---
include/linux/page-flags.h | 2 +-
include/linux/page_ref.h | 8 ++++++--

2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/page-flags.h +1 -1
index 691506bdf2c5..6b8ecf86f1b6 100644
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 * cold cacheline in some cases.
 */
      if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-         test_bit(PG_head, &page->flags)) {
+         test_bit_acquire(PG_head, &page->flags)) {
              /*
               * We can safely access the field of the @page[1] with PG_head
               * because the @page is a compound page composed with at least
include/linux/page_ref.h +0 -0
index 8c236c651d1d..5becea98bd79 100644
@@ -233,8 +233,12 @@ static inline bool page_ref_add_unless(struct page *page, int nr, int u)
      bool ret = false;
 
Quoted 13 lines

include/linux/page_ref.h static inline bool page_ref_add_unless(struct page *page, int nr, int u) · line 241

Quoted 9 lines
include/linux/page_ref.h +6 -2
      rcu_read_lock();
-     /* avoid writing to the vmemmap area being remapped */
-     if (!page_is_fake_head(page) && page_ref_count(page) != u)
+     /*
+      * To avoid writing to the vmemmap area remapped into r/o in parallel,
+      * the page_ref_count() test must precede the page_is_fake_head() test
+      * so that test_bit_acquire() in the latter is ordered after the former.
+      */
+     if (page_ref_count(page) != u && !page_is_fake_head(page))

IIUC, we need to insert a memory barrier between page_ref_count() and page_is_fake_head(). Specifically, accessing between page->_refcount and page->flags. So we should insert a read memory barrier here, right?

Correct, i.e., page_ref_count(page) != u; smp_rmb(); !page_is_fake_head(page).

Quoted 3 lines

But I saw you added an acquire barrier in page_fixed_fake_head(),
I don't understand why an acquire barrier could stop the CPU reordering the accessing between them. What am I missing here?

A load-acquire on page->_refcount would be equivalent to the smp_rmb() above. But apparently I used on page->flags because I misremembered whether a load-acquire inserts the equivalent smp_rmb() before or after (it's after, not before). Will fix this in v2.

#3 lore raw
In reply to Yu Zhao
Quoted 21 lines · Yu Zhao

Using x86_64 as an example, for a 32KB struct page[] area describing a 2MB hugeTLB, HVO reduces the area to 4KB by the following steps:
1. Split the (r/w vmemmap) PMD mapping the area into 512 (r/w) PTEs;
2. For the 8 PTEs mapping the area, remap PTE 1-7 to the page mapped by PTE 0, and at the same time change the permission from r/w to r/o;
3. Free the pages PTE 1-7 used to map, hence the reduction from 32KB

    to 4KB.

However, the following race can happen due to improperly memory loads ordering:
CPU 1 (HVO) CPU 2 (speculative PFN walker)

page_ref_freeze()
synchronize_rcu()
rcu_read_lock() page_is_fake_head() is false
vmemmap_remap_pte()
XXX: struct page[] becomes r/o

page_ref_unfreeze()

Quoted 8 lines
                                   page_ref_count() is not zero

                                   atomic_add_unless(&page->_refcount)
                                   XXX: try to modify r/o struct page[]

Specifically, page_is_fake_head() must be ordered after page_ref_count() on CPU 2 so that it can only return true for this case, to avoid the later attempt to modify r/o struct page[].

I *think* this is correct.

Quoted 18 lines

This patch adds the missing memory barrier and makes the tests on page_is_fake_head() and page_ref_count() done in the proper order.

Fixes bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")
---
  include/linux/page-flags.h | 2 +-
  include/linux/page_ref.h   | 8 ++++++--
  2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/page-flags.h +0 -0
index 691506bdf2c5..6b8ecf86f1b6 100644

include/linux/page-flags.h static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 5 lines
include/linux/page-flags.h +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
  	 * cold cacheline in some cases.
  	 */
  	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

I assume you want an explicit memory barrier in the single problematic caller instead.

Signature · 3 lines
Cheers,

David / dhildenb
#4 lore raw
Quoted 10 lines · David Hildenbrand

include/linux/page-flags.h static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 7 lines

+++ b/include/linux/page-flags.h

include/linux/page-flags.h +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
  	 * cold cacheline in some cases.
  	 */
  	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

I've been looking at the callers of PageTail() because it's going to be a bit of a weird thing to be checking in the separate-page-and-folio world. Obviously we can implement it, but there's a bit of a "But why would you want to ask that question" question.

Most current occurrences of PageTail() are in assertions of one form or another. Fair enough, not performance critical.

make_device_exclusive_range() is a little weird; looks like it's trying to make sure that each folio is only made exclusive once, and ignore any partial folios which overlap the start of the area.

damon_get_folio() wants to fail for tail pages. Fair enough.

split_huge_pages_all() is debug code.

page_idle_get_folio() is like damon.

That's it. We don't seem to have any PageTail() callers in critical code any more.

#5 lore raw
Quoted 17 lines · Matthew Wilcox
Quoted 10 lines · David Hildenbrand

Review static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 7 lines

+++ b/include/linux/page-flags.h

(quoted diff) +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
   	 * cold cacheline in some cases.
   	 */
   	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

I've been looking at the callers of PageTail() because it's going to be a bit of a weird thing to be checking in the separate-page-and-folio world. Obviously we can implement it, but there's a bit of a "But why would you want to ask that question" question.

Quoted 6 lines

Most current occurrences of PageTail() are in assertions of one form or another. Fair enough, not performance critical.

make_device_exclusive_range() is a little weird; looks like it's trying to make sure that each folio is only made exclusive once, and ignore any partial folios which overlap the start of the area.

I could have sworn we only support small folios here, but looks like we do support large folios.

IIUC, there is no way to identify reliably "this folio is device exclusive", the only hint is "no mappings". The following might do:

mm/rmap.c +7 -1
index c6c4d4ea29a7e..1424d0a351a86 100644
@@ -2543,7 +2543,13 @@ int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
 
        for (i = 0; i < npages; i++, start += PAGE_SIZE) {
                struct folio *folio = page_folio(pages[i]);
-               if (PageTail(pages[i]) || !folio_trylock(folio)) {
+
+               /*
+                * If there are no mappings, either the folio is actually
+                * unmapped or only device-exclusive swap entries point at
+                * this folio.
+                */
+               if (!folio_mapped(folio) || !folio_trylock(folio)) {
                        folio_put(folio);
                        pages[i] = NULL;
                        continue;
Quoted 9 lines

damon_get_folio() wants to fail for tail pages. Fair enough.

split_huge_pages_all() is debug code.

page_idle_get_folio() is like damon.

That's it. We don't seem to have any PageTail() callers in critical code any more.

Ah, you're right. Interestingly, PageTransTail() is even unused?

Signature · 3 lines
Cheers,

David / dhildenb
#6 lore raw
Quoted 40 lines · David Hildenbrand
Quoted 17 lines · Matthew Wilcox
Quoted 10 lines · David Hildenbrand

Review static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 7 lines

+++ b/include/linux/page-flags.h

(quoted diff) +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
    	 * cold cacheline in some cases.
    	 */
    	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

I've been looking at the callers of PageTail() because it's going to be a bit of a weird thing to be checking in the separate-page-and-folio world. Obviously we can implement it, but there's a bit of a "But why would you want to ask that question" question.

Quoted 6 lines

Most current occurrences of PageTail() are in assertions of one form or another. Fair enough, not performance critical.

make_device_exclusive_range() is a little weird; looks like it's trying to make sure that each folio is only made exclusive once, and ignore any partial folios which overlap the start of the area.

I could have sworn we only support small folios here, but looks like we do support large folios.

IIUC, there is no way to identify reliably "this folio is device exclusive", the only hint is "no mappings". The following might do:

mm/rmap.c +0 -1
index c6c4d4ea29a7e..1424d0a351a86 100644
@@ -2543,7 +2543,13 @@ int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
 
          for (i = 0; i < npages; i++, start += PAGE_SIZE) {
                  struct folio *folio = page_folio(pages[i]);
-               if (PageTail(pages[i]) || !folio_trylock(folio)) {

mm/rmap.c int make_device_exclusive_range(struct mm_struct *mm, unsigned long start, · line 2555

Quoted 10 lines
mm/rmap.c +7 -0
+
+               /*
+                * If there are no mappings, either the folio is actually
+                * unmapped or only device-exclusive swap entries point at
+                * this folio.
+                */
+               if (!folio_mapped(folio) || !folio_trylock(folio)) {
                          folio_put(folio);
                          pages[i] = NULL;
                          continue;

I stared longer at this, and not sure if that will work.

The PageTail() is in place because we return with the folio locked on success, so we won't trylock again on tail pages.

But staring at page_make_device_exclusive_one(), I am not sure if it does what we want in all cases ...

... and the hmm selftests just keeps failing upstream as well?! huh. :)

I'll try spending some time on this to see if I can grasp what needs to be done and how it could be handled ... better.

Signature · 3 lines
Cheers,

David / dhildenb
#7 lore raw
Quoted 57 lines · David Hildenbrand
Quoted 40 lines · David Hildenbrand
Quoted 17 lines · Matthew Wilcox
Quoted 10 lines · David Hildenbrand

Review static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 7 lines

+++ b/include/linux/page-flags.h

(quoted diff) +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
     	 * cold cacheline in some cases.
     	 */
     	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags)) {
+	    test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

I've been looking at the callers of PageTail() because it's going to be a bit of a weird thing to be checking in the separate-page-and-folio world. Obviously we can implement it, but there's a bit of a "But why would you want to ask that question" question.

Quoted 6 lines

Most current occurrences of PageTail() are in assertions of one form or another. Fair enough, not performance critical.

make_device_exclusive_range() is a little weird; looks like it's trying to make sure that each folio is only made exclusive once, and ignore any partial folios which overlap the start of the area.

I could have sworn we only support small folios here, but looks like we do support large folios.

IIUC, there is no way to identify reliably "this folio is device exclusive", the only hint is "no mappings". The following might do:

mm/rmap.c +0 -1
index c6c4d4ea29a7e..1424d0a351a86 100644
@@ -2543,7 +2543,13 @@ int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
 
           for (i = 0; i < npages; i++, start += PAGE_SIZE) {
                   struct folio *folio = page_folio(pages[i]);
-               if (PageTail(pages[i]) || !folio_trylock(folio)) {

mm/rmap.c int make_device_exclusive_range(struct mm_struct *mm, unsigned long start, · line 2555

Quoted 10 lines
mm/rmap.c +7 -0
+
+               /*
+                * If there are no mappings, either the folio is actually
+                * unmapped or only device-exclusive swap entries point at
+                * this folio.
+                */
+               if (!folio_mapped(folio) || !folio_trylock(folio)) {
                           folio_put(folio);
                           pages[i] = NULL;
                           continue;

I stared longer at this, and not sure if that will work.

The PageTail() is in place because we return with the folio locked on success, so we won't trylock again on tail pages.

Quoted 8 lines

But staring at page_make_device_exclusive_one(), I am not sure if it does what we want in all cases ...

... and the hmm selftests just keeps failing upstream as well?! huh. :)

I'll try spending some time on this to see if I can grasp what needs to be done and how it could be handled ... better.

As expected ...

# echo never > /sys/kernel/mm/transparent_hugepage/enabled
# ./hmm-tests
...
#  RUN           hmm.hmm_device_private.exclusive ...
#            OK  hmm.hmm_device_private.exclusive
ok 21 hmm.hmm_device_private.exclusive
#  RUN           hmm.hmm_device_private.exclusive_mprotect ...
#            OK  hmm.hmm_device_private.exclusive_mprotect
ok 22 hmm.hmm_device_private.exclusive_mprotect
#  RUN           hmm.hmm_device_private.exclusive_cow ...
#            OK  hmm.hmm_device_private.exclusive_cow
ok 23 hmm.hmm_device_private.exclusive_cow
#  RUN           hmm.hmm_device_private.hmm_gup_test ...
#            OK  hmm.hmm_device_private.hmm_gup_test
...
# echo always > /sys/kernel/mm/transparent_hugepage/enabled
...
#  RUN           hmm.hmm_device_private.exclusive ...
# hmm-tests.c:1751:exclusive:Expected ret (-16) == 0 (0)
# exclusive: Test terminated by assertion
#          FAIL  hmm.hmm_device_private.exclusive
not ok 21 hmm.hmm_device_private.exclusive
#  RUN           hmm.hmm_device_private.exclusive_mprotect ...
# hmm-tests.c:1805:exclusive_mprotect:Expected ret (-16) == 0 (0)
# exclusive_mprotect: Test terminated by assertion
#          FAIL  hmm.hmm_device_private.exclusive_mprotect
not ok 22 hmm.hmm_device_private.exclusive_mprotect
#  RUN           hmm.hmm_device_private.exclusive_cow ...
# hmm-tests.c:1858:exclusive_cow:Expected ret (-16) == 0 (0)
# exclusive_cow: Test terminated by assertion
#          FAIL  hmm.hmm_device_private.exclusive_cow
not ok 23 hmm.hmm_device_private.exclusive_cow

So rejecting folio_test_large() would likely achieve the same thing right now.

Signature · 3 lines
Cheers,

David / dhildenb
author
#8 lore raw
In reply to David Hildenbrand
Quoted 61 lines · David Hildenbrand
Quoted 21 lines · Yu Zhao

Using x86_64 as an example, for a 32KB struct page[] area describing a 2MB hugeTLB, HVO reduces the area to 4KB by the following steps:
1. Split the (r/w vmemmap) PMD mapping the area into 512 (r/w) PTEs;
2. For the 8 PTEs mapping the area, remap PTE 1-7 to the page mapped by PTE 0, and at the same time change the permission from r/w to r/o;
3. Free the pages PTE 1-7 used to map, hence the reduction from 32KB

    to 4KB.

However, the following race can happen due to improperly memory loads ordering:
CPU 1 (HVO) CPU 2 (speculative PFN walker)

page_ref_freeze()
synchronize_rcu()
rcu_read_lock() page_is_fake_head() is false
vmemmap_remap_pte()
XXX: struct page[] becomes r/o

page_ref_unfreeze()

Quoted 8 lines
                                   page_ref_count() is not zero

                                   atomic_add_unless(&page->_refcount)
                                   XXX: try to modify r/o struct page[]

Specifically, page_is_fake_head() must be ordered after page_ref_count() on CPU 2 so that it can only return true for this case, to avoid the later attempt to modify r/o struct page[].

I *think* this is correct.

Quoted 18 lines

This patch adds the missing memory barrier and makes the tests on page_is_fake_head() and page_ref_count() done in the proper order.

Fixes bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")
---
  include/linux/page-flags.h | 2 +-
  include/linux/page_ref.h   | 8 ++++++--
  2 files changed, 7 insertions(+), 3 deletions(-)
include/linux/page-flags.h +0 -0
index 691506bdf2c5..6b8ecf86f1b6 100644

include/linux/page-flags.h static __always_inline const struct page *page_fixed_fake_head(const struct page · line 215

Quoted 5 lines
include/linux/page-flags.h +1 -1
@@ -212,7 +212,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
       * cold cacheline in some cases.
       */
      if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-         test_bit(PG_head, &page->flags)) {
+         test_bit_acquire(PG_head, &page->flags)) {

This change will affect all page_fixed_fake_head() users, like ordinary PageTail even on !hugetlb.

Quoted 2 lines

I assume you want an explicit memory barrier in the single problematic caller instead.

Let me make it HVO specific in v2. It might look cleaner that way.