Skip to contentSkip to conversation

mm, futex: Fix shared futex pgoff on shmem huge page

PATCH Superseded Hugh Dickins 3 participants 6 messages Jun 12, 2021 – Jun 15, 2021

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

PATCH 1/1 mm, futex: Fix shared futex pgoff on shmem huge page

If more than one futex is placed on a shmem huge page, it can happen that waking the second wakes the first instead, and leaves the second waiting:
the key's shared.pgoff is wrong.

When 3.11 commit 13d60f4b6ab5 ("futex: Take hugepages into account when generating futex_key"), the only shared huge pages came from hugetlbfs, and the code added to deal with its exceptional page->index was put into hugetlb source. Then that was missed when 4.8 added shmem huge pages.

page_to_pgoff() is what others use for this nowadays: except that, as currently written, it gives the right answer on hugetlbfs head, but nonsense on hugetlbfs tails. Fix that by calling hugetlbfs-specific hugetlb_basepage_index() on PageHuge tails as well as on head.

Yes, it's unconventional to declare hugetlb_basepage_index() there in pagemap.h, rather than in hugetlb.h; but I do not expect anything but page_to_pgoff() ever to need it.

Fixes 800d8c63b2e9 ("shmem: add huge pages support")

---

 include/linux/hugetlb.h |   16 ----------------
 include/linux/pagemap.h |   13 ++++++-------
 kernel/futex.c          |    3 +--
 mm/hugetlb.c            |    5 +----
 4 files changed, 8 insertions(+), 29 deletions(-)
include/linux/hugetlb.h +0 -16
@@ -733,17 +733,6 @@ static inline int hstate_index(struct hs
 	return h - hstates;
 }
 
-pgoff_t __basepage_index(struct page *page);
-
-/* Return page->index in PAGE_SIZE units */
-static inline pgoff_t basepage_index(struct page *page)
-{
-	if (!PageCompound(page))
-		return page->index;
-
-	return __basepage_index(page);
-}
-
 extern int dissolve_free_huge_page(struct page *page);
 extern int dissolve_free_huge_pages(unsigned long start_pfn,
 				    unsigned long end_pfn);
@@ -980,11 +969,6 @@ static inline int hstate_index(struct hs
 	return 0;
 }
 
-static inline pgoff_t basepage_index(struct page *page)
-{
-	return page->index;
-}
-
 static inline int dissolve_free_huge_page(struct page *page)
 {
 	return 0;
include/linux/pagemap.h +6 -7
@@ -516,8 +516,7 @@ static inline struct page *read_mapping_
 }
 
 /*
- * Get index of the page with in radix-tree
- * (TODO: remove once hugetlb pages will have ->index in PAGE_SIZE)
+ * Get index of the page within radix-tree (but not for hugetlb pages).
  */
 static inline pgoff_t page_to_index(struct page *page)
 {
@@ -536,14 +535,14 @@ static inline pgoff_t page_to_index(stru
 }
 
 /*
- * Get the offset in PAGE_SIZE.
- * (TODO: hugepage should have ->index in PAGE_SIZE)
+ * Get the offset in PAGE_SIZE (even for hugetlb pages).
  */
 static inline pgoff_t page_to_pgoff(struct page *page)
 {
-	if (unlikely(PageHeadHuge(page)))
-		return page->index << compound_order(page);
-
+	if (unlikely(PageHuge(page))) {
+		extern pgoff_t hugetlb_basepage_index(struct page *page);
+		return hugetlb_basepage_index(page);
+	}
 	return page_to_index(page);
 }
 
kernel/futex.c +1 -2
@@ -35,7 +35,6 @@
 #include <linux/jhash.h>
 #include <linux/pagemap.h>
 #include <linux/syscalls.h>
-#include <linux/hugetlb.h>
 #include <linux/freezer.h>
 #include <linux/memblock.h>
 #include <linux/fault-inject.h>
@@ -650,7 +649,7 @@ again:
 
 		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 		key->shared.i_seq = get_inode_sequence_number(inode);
-		key->shared.pgoff = basepage_index(tail);
+		key->shared.pgoff = page_to_pgoff(tail);
 		rcu_read_unlock();
 	}
 
mm/hugetlb.c +1 -4
@@ -1588,15 +1588,12 @@ struct address_space *hugetlb_page_mappi
 	return NULL;
 }
 
-pgoff_t __basepage_index(struct page *page)
+pgoff_t hugetlb_basepage_index(struct page *page)
 {
 	struct page *page_head = compound_head(page);
 	pgoff_t index = page_index(page_head);
 	unsigned long compound_idx;
 
-	if (!PageHuge(page_head))
-		return page_index(page);
-
 	if (compound_order(page_head) >= MAX_ORDER)
 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 	else
maintainer
#1 lore raw

include/linux/pagemap.h static inline struct page *read_mapping_ · line 520

Quoted 9 lines · Hugh Dickins

+++ linux/include/linux/pagemap.h 2021-06-11 17:30:28.726720252 -0700

include/linux/pagemap.h +1 -2
@@ -516,8 +516,7 @@ static inline struct page *read_mapping_
 }
 
 /*
- * Get index of the page with in radix-tree
- * (TODO: remove once hugetlb pages will have ->index in PAGE_SIZE)
+ * Get index of the page within radix-tree (but not for hugetlb pages).
  */

I think the TODO should be retained. It's still something that I intend to do.

include/linux/pagemap.h static inline pgoff_t page_to_index(stru · line 547

Quoted 11 lines
include/linux/pagemap.h +4 -3
 static inline pgoff_t page_to_pgoff(struct page *page)
 {
-	if (unlikely(PageHeadHuge(page)))
-		return page->index << compound_order(page);
-
+	if (unlikely(PageHuge(page))) {
+		extern pgoff_t hugetlb_basepage_index(struct page *page);
+		return hugetlb_basepage_index(page);
+	}
 	return page_to_index(page);
 }

Yes, this change absolutely makes sense. It's just a landmine if it returns the right answer for some tail pages but garbage for other kinds of tail pages.

Quoted 10 lines
mm/hugetlb.c +1 -1
@@ -1588,15 +1588,12 @@ struct address_space *hugetlb_page_mappi
 	return NULL;
 }
 
-pgoff_t __basepage_index(struct page *page)
+pgoff_t hugetlb_basepage_index(struct page *page)
 {

mm/hugetlb.c struct address_space *hugetlb_page_mappi · line 1599

Quoted 10 lines
mm/hugetlb.c +0 -3
 	struct page *page_head = compound_head(page);
 	pgoff_t index = page_index(page_head);
 	unsigned long compound_idx;
 
-	if (!PageHuge(page_head))
-		return page_index(page);
-
 	if (compound_order(page_head) >= MAX_ORDER)
 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 	else

urgh. this trailing bit should be:

#if defined(CONFIG_SPARSEMEM) && !defined(CONFIG_SPARSEMEM_VMEMMAP)
	compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
#else
	compound_idx = page - page_head;

#endif

#2 lore raw
Quoted 5 lines · Matthew Wilcox
Quoted 4 lines · Hugh Dickins

+++ linux/include/linux/pagemap.h 2021-06-11 17:30:28.726720252 -0700

(quoted diff) +0 -0
@@ -516,8 +516,7 @@ static inline struct page *read_mapping_
 }
 
Quoted 8 lines

include/linux/pagemap.h static inline struct page *read_mapping_ · line 520

Quoted 5 lines
include/linux/pagemap.h +1 -2
 /*
- * Get index of the page with in radix-tree
- * (TODO: remove once hugetlb pages will have ->index in PAGE_SIZE)
+ * Get index of the page within radix-tree (but not for hugetlb pages).
  */

I think the TODO should be retained. It's still something that I intend to do.

Okay. I did not mean to imply, by removing those TODOs, that they should not be done: just that they were a developer's notes to self, that I found distracting there.

I've restored both TODOs (but changed the second to say "hugetlb pages" explicitly, rather than the ambiguous "hugepage").

Quoted 23 lines
Quoted 10 lines
mm/hugetlb.c +1 -1
@@ -1588,15 +1588,12 @@ struct address_space *hugetlb_page_mappi
 	return NULL;
 }
 
-pgoff_t __basepage_index(struct page *page)
+pgoff_t hugetlb_basepage_index(struct page *page)
 {

mm/hugetlb.c struct address_space *hugetlb_page_mappi · line 1599

Quoted 10 lines
mm/hugetlb.c +0 -3
 	struct page *page_head = compound_head(page);
 	pgoff_t index = page_index(page_head);
 	unsigned long compound_idx;
 
-	if (!PageHuge(page_head))
-		return page_index(page);
-
 	if (compound_order(page_head) >= MAX_ORDER)
 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 	else

urgh. this trailing bit should be:

Quoted 5 lines
#if defined(CONFIG_SPARSEMEM) && !defined(CONFIG_SPARSEMEM_VMEMMAP)
	compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
#else
	compound_idx = page - page_head;

#endif

I don't see what's wrong with what's there, myself. Unfamiliar territory to me, but mem_map_next() appears to have the same MAX_ORDER expectation. Or perhaps you're just suggesting an optimization.

If it were obvious to me, I'd have gladly folded it in; but no, please send your own patch for that, running it by Mike Kravetz and Mike Rapoport and David Hildenbrand, I think.

Thanks, v2 follows,
Hugh

#3 lore raw

PATCH v2 1/1 mm, futex: Fix shared futex pgoff on shmem huge page

In reply to Matthew Wilcox

If more than one futex is placed on a shmem huge page, it can happen that waking the second wakes the first instead, and leaves the second waiting:
the key's shared.pgoff is wrong.

When 3.11 commit 13d60f4b6ab5 ("futex: Take hugepages into account when generating futex_key"), the only shared huge pages came from hugetlbfs, and the code added to deal with its exceptional page->index was put into hugetlb source. Then that was missed when 4.8 added shmem huge pages.

page_to_pgoff() is what others use for this nowadays: except that, as currently written, it gives the right answer on hugetlbfs head, but nonsense on hugetlbfs tails. Fix that by calling hugetlbfs-specific hugetlb_basepage_index() on PageHuge tails as well as on head.

Yes, it's unconventional to declare hugetlb_basepage_index() there in pagemap.h, rather than in hugetlb.h; but I do not expect anything but page_to_pgoff() ever to need it.

Fixes 800d8c63b2e9 ("shmem: add huge pages support")

---
v2: restored TODO comments, per Matthew; make "hugetlb pages" explicit.

 include/linux/hugetlb.h |   16 ----------------
 include/linux/pagemap.h |   13 +++++++------
 kernel/futex.c          |    3 +--
 mm/hugetlb.c            |    5 +----
 4 files changed, 9 insertions(+), 28 deletions(-)
include/linux/hugetlb.h +0 -16
@@ -733,17 +733,6 @@ static inline int hstate_index(struct hs
 	return h - hstates;
 }
 
-pgoff_t __basepage_index(struct page *page);
-
-/* Return page->index in PAGE_SIZE units */
-static inline pgoff_t basepage_index(struct page *page)
-{
-	if (!PageCompound(page))
-		return page->index;
-
-	return __basepage_index(page);
-}
-
 extern int dissolve_free_huge_page(struct page *page);
 extern int dissolve_free_huge_pages(unsigned long start_pfn,
 				    unsigned long end_pfn);
@@ -980,11 +969,6 @@ static inline int hstate_index(struct hs
 	return 0;
 }
 
-static inline pgoff_t basepage_index(struct page *page)
-{
-	return page->index;
-}
-
 static inline int dissolve_free_huge_page(struct page *page)
 {
 	return 0;
include/linux/pagemap.h +7 -6
@@ -516,7 +516,7 @@ static inline struct page *read_mapping_
 }
 
 /*
- * Get index of the page with in radix-tree
+ * Get index of the page within radix-tree (but not for hugetlb pages).
  * (TODO: remove once hugetlb pages will have ->index in PAGE_SIZE)
  */
 static inline pgoff_t page_to_index(struct page *page)
@@ -536,14 +536,15 @@ static inline pgoff_t page_to_index(stru
 }
 
 /*
- * Get the offset in PAGE_SIZE.
- * (TODO: hugepage should have ->index in PAGE_SIZE)
+ * Get the offset in PAGE_SIZE (even for hugetlb pages).
+ * (TODO: hugetlb pages should have ->index in PAGE_SIZE)
  */
 static inline pgoff_t page_to_pgoff(struct page *page)
 {
-	if (unlikely(PageHeadHuge(page)))
-		return page->index << compound_order(page);
-
+	if (unlikely(PageHuge(page))) {
+		extern pgoff_t hugetlb_basepage_index(struct page *page);
+		return hugetlb_basepage_index(page);
+	}
 	return page_to_index(page);
 }
 
kernel/futex.c +1 -2
@@ -35,7 +35,6 @@
 #include <linux/jhash.h>
 #include <linux/pagemap.h>
 #include <linux/syscalls.h>
-#include <linux/hugetlb.h>
 #include <linux/freezer.h>
 #include <linux/memblock.h>
 #include <linux/fault-inject.h>
@@ -650,7 +649,7 @@ again:
 
 		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 		key->shared.i_seq = get_inode_sequence_number(inode);
-		key->shared.pgoff = basepage_index(tail);
+		key->shared.pgoff = page_to_pgoff(tail);
 		rcu_read_unlock();
 	}
 
mm/hugetlb.c +1 -4
@@ -1588,15 +1588,12 @@ struct address_space *hugetlb_page_mappi
 	return NULL;
 }
 
-pgoff_t __basepage_index(struct page *page)
+pgoff_t hugetlb_basepage_index(struct page *page)
 {
 	struct page *page_head = compound_head(page);
 	pgoff_t index = page_index(page_head);
 	unsigned long compound_idx;
 
-	if (!PageHuge(page_head))
-		return page_index(page);
-
 	if (compound_order(page_head) >= MAX_ORDER)
 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 	else
maintainer Reviewed-by on v2 1/1
#4 lore raw
Quoted 14 lines · Hugh Dickins

If more than one futex is placed on a shmem huge page, it can happen that waking the second wakes the first instead, and leaves the second waiting:
the key's shared.pgoff is wrong.

When 3.11 commit 13d60f4b6ab5 ("futex: Take hugepages into account when generating futex_key"), the only shared huge pages came from hugetlbfs, and the code added to deal with its exceptional page->index was put into hugetlb source. Then that was missed when 4.8 added shmem huge pages.

page_to_pgoff() is what others use for this nowadays: except that, as currently written, it gives the right answer on hugetlbfs head, but nonsense on hugetlbfs tails. Fix that by calling hugetlbfs-specific hugetlb_basepage_index() on PageHuge tails as well as on head.

Quoted 8 lines

Yes, it's unconventional to declare hugetlb_basepage_index() there in pagemap.h, rather than in hugetlb.h; but I do not expect anything but page_to_pgoff() ever to need it.

Fixes 800d8c63b2e9 ("shmem: add huge pages support")
Acked-by on v2 1/1
#5 lore raw
In reply to Hugh Dickins
Quoted 14 lines · Hugh Dickins

If more than one futex is placed on a shmem huge page, it can happen that waking the second wakes the first instead, and leaves the second waiting:
the key's shared.pgoff is wrong.

When 3.11 commit 13d60f4b6ab5 ("futex: Take hugepages into account when generating futex_key"), the only shared huge pages came from hugetlbfs, and the code added to deal with its exceptional page->index was put into hugetlb source. Then that was missed when 4.8 added shmem huge pages.

page_to_pgoff() is what others use for this nowadays: except that, as currently written, it gives the right answer on hugetlbfs head, but nonsense on hugetlbfs tails. Fix that by calling hugetlbfs-specific hugetlb_basepage_index() on PageHuge tails as well as on head.

Quoted 8 lines

Yes, it's unconventional to declare hugetlb_basepage_index() there in pagemap.h, rather than in hugetlb.h; but I do not expect anything but page_to_pgoff() ever to need it.

Fixes 800d8c63b2e9 ("shmem: add huge pages support")

Assuming this goes through mm: