Skip to contentSkip to conversation

powerpc/mm: Use refcount_t for refcount

PATCH In review Chuhong Yuan 3 participants 4 messages Aug 8, 2019 – Nov 26, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc/mm: Use refcount_t for refcount

Reference counters are preferred to use refcount_t instead of atomic_t.
This is because the implementation of refcount_t can prevent overflows and detect possible use-after-free.
So convert atomic_t ref counters to refcount_t.

---
 arch/powerpc/mm/book3s64/mmu_context.c | 2 +-
 arch/powerpc/mm/book3s64/pgtable.c     | 7 +++----
 arch/powerpc/mm/pgtable-frag.c         | 9 ++++-----
 include/linux/mm_types.h               | 3 ++-
 4 files changed, 10 insertions(+), 11 deletions(-)
arch/powerpc/mm/book3s64/mmu_context.c +1 -1
index 2d0cb5ba9a47..f836fd5a6abc 100644
@@ -231,7 +231,7 @@ static void pmd_frag_destroy(void *pmd_frag)
 	/* drop all the pending references */
 	count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
 	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
 		pgtable_pmd_page_dtor(page);
 		__free_page(page);
 	}
arch/powerpc/mm/book3s64/pgtable.c +3 -4
index 7d0e0d0d22c4..40056896ce4e 100644
@@ -277,7 +277,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
 		return NULL;
 	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
 	ret = page_address(page);
 	/*
@@ -294,7 +294,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
 	 * count.
 	 */
 	if (likely(!mm->context.pmd_frag)) {
-		atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PMD_FRAG_NR);
 		mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
 	}
 	spin_unlock(&mm->page_table_lock);
@@ -317,8 +317,7 @@ void pmd_fragment_free(unsigned long *pmd)
 {
 	struct page *page = virt_to_page(pmd);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
 		pgtable_pmd_page_dtor(page);
 		__free_page(page);
 	}
arch/powerpc/mm/pgtable-frag.c +4 -5
index a7b05214760c..4ef8231b677f 100644
@@ -24,7 +24,7 @@ void pte_frag_destroy(void *pte_frag)
 	/* drop all the pending references */
 	count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
 	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
 		pgtable_page_dtor(page);
 		__free_page(page);
 	}
@@ -71,7 +71,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 			return NULL;
 	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
 	ret = page_address(page);
 	/*
@@ -87,7 +87,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 	 * count.
 	 */
 	if (likely(!pte_frag_get(&mm->context))) {
-		atomic_set(&page->pt_frag_refcount, PTE_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PTE_FRAG_NR);
 		pte_frag_set(&mm->context, ret + PTE_FRAG_SIZE);
 	}
 	spin_unlock(&mm->page_table_lock);
@@ -110,8 +110,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
 {
 	struct page *page = virt_to_page(table);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
 		if (!kernel)
 			pgtable_page_dtor(page);
 		__free_page(page);
include/linux/mm_types.h +2 -1
index 3a37a89eb7a7..7fe23a3faf95 100644
@@ -14,6 +14,7 @@
 #include <linux/uprobes.h>
 #include <linux/page-flags-layout.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 
 #include <asm/mmu.h>
 
@@ -147,7 +148,7 @@ struct page {
 			unsigned long _pt_pad_2;	/* mapping */
 			union {
 				struct mm_struct *pt_mm; /* x86 pgds only */
-				atomic_t pt_frag_refcount; /* powerpc */
+				refcount_t pt_frag_refcount; /* powerpc */
 			};
 #if ALLOC_SPLIT_PTLOCKS
 			spinlock_t *ptl;
Signature · 1 line
2.20.1
#1 lore raw
Quoted 7 lines · Chuhong Yuan

Reference counters are preferred to use refcount_t instead of atomic_t.
This is because the implementation of refcount_t can prevent overflows and detect possible use-after-free.
So convert atomic_t ref counters to refcount_t.

Thanks.

We don't have a fast implementation of refcount_t, so I'm worried this could cause a measurable performance regression.

Did you benchmark it at all?

cheers

Quoted 109 lines
arch/powerpc/mm/book3s64/mmu_context.c +1 -1
index 2d0cb5ba9a47..f836fd5a6abc 100644
@@ -231,7 +231,7 @@ static void pmd_frag_destroy(void *pmd_frag)
 	/* drop all the pending references */
 	count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
 	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
 		pgtable_pmd_page_dtor(page);
 		__free_page(page);
 	}
arch/powerpc/mm/book3s64/pgtable.c +3 -4
index 7d0e0d0d22c4..40056896ce4e 100644
@@ -277,7 +277,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
 		return NULL;
 	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
 	ret = page_address(page);
 	/*
@@ -294,7 +294,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
 	 * count.
 	 */
 	if (likely(!mm->context.pmd_frag)) {
-		atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PMD_FRAG_NR);
 		mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
 	}
 	spin_unlock(&mm->page_table_lock);
@@ -317,8 +317,7 @@ void pmd_fragment_free(unsigned long *pmd)
 {
 	struct page *page = virt_to_page(pmd);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
 		pgtable_pmd_page_dtor(page);
 		__free_page(page);
 	}
arch/powerpc/mm/pgtable-frag.c +4 -5
index a7b05214760c..4ef8231b677f 100644
@@ -24,7 +24,7 @@ void pte_frag_destroy(void *pte_frag)
 	/* drop all the pending references */
 	count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
 	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
 		pgtable_page_dtor(page);
 		__free_page(page);
 	}
@@ -71,7 +71,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 			return NULL;
 	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
 	ret = page_address(page);
 	/*
@@ -87,7 +87,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 	 * count.
 	 */
 	if (likely(!pte_frag_get(&mm->context))) {
-		atomic_set(&page->pt_frag_refcount, PTE_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PTE_FRAG_NR);
 		pte_frag_set(&mm->context, ret + PTE_FRAG_SIZE);
 	}
 	spin_unlock(&mm->page_table_lock);
@@ -110,8 +110,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
 {
 	struct page *page = virt_to_page(table);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
 		if (!kernel)
 			pgtable_page_dtor(page);
 		__free_page(page);
include/linux/mm_types.h +2 -1
index 3a37a89eb7a7..7fe23a3faf95 100644
@@ -14,6 +14,7 @@
 #include <linux/uprobes.h>
 #include <linux/page-flags-layout.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 
 #include <asm/mmu.h>
 
@@ -147,7 +148,7 @@ struct page {
 			unsigned long _pt_pad_2;	/* mapping */
 			union {
 				struct mm_struct *pt_mm; /* x86 pgds only */
-				atomic_t pt_frag_refcount; /* powerpc */
+				refcount_t pt_frag_refcount; /* powerpc */
 			};
 #if ALLOC_SPLIT_PTLOCKS
 			spinlock_t *ptl;
Signature · 1 line
2.20.1
#2 lore raw
Quoted 8 lines · Michael Ellerman
Quoted 6 lines · Chuhong Yuan

Reference counters are preferred to use refcount_t instead of atomic_t.
This is because the implementation of refcount_t can prevent overflows and detect possible use-after-free.
So convert atomic_t ref counters to refcount_t.

Quoted 9 lines
Quoted 1 line

Thanks.

We don't have a fast implementation of refcount_t, so I'm worried this could cause a measurable performance regression.

Did you benchmark it at all?

I did not benchmark it and I don't have the testing environment...

Quoted 111 lines

cheers

Quoted 109 lines
arch/powerpc/mm/book3s64/mmu_context.c +1 -1
index 2d0cb5ba9a47..f836fd5a6abc 100644
@@ -231,7 +231,7 @@ static void pmd_frag_destroy(void *pmd_frag)
      /* drop all the pending references */
      count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
      /* We allow PTE_FRAG_NR fragments from a PTE page */
-     if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
+     if (refcount_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
              pgtable_pmd_page_dtor(page);
              __free_page(page);
      }
arch/powerpc/mm/book3s64/pgtable.c +3 -4
index 7d0e0d0d22c4..40056896ce4e 100644
@@ -277,7 +277,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
              return NULL;
      }
 
-     atomic_set(&page->pt_frag_refcount, 1);
+     refcount_set(&page->pt_frag_refcount, 1);
 
      ret = page_address(page);
      /*
@@ -294,7 +294,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
       * count.
       */
      if (likely(!mm->context.pmd_frag)) {
-             atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
+             refcount_set(&page->pt_frag_refcount, PMD_FRAG_NR);
              mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
      }
      spin_unlock(&mm->page_table_lock);
@@ -317,8 +317,7 @@ void pmd_fragment_free(unsigned long *pmd)
 {
      struct page *page = virt_to_page(pmd);
 
-     BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-     if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+     if (refcount_dec_and_test(&page->pt_frag_refcount)) {
              pgtable_pmd_page_dtor(page);
              __free_page(page);
      }
arch/powerpc/mm/pgtable-frag.c +4 -5
index a7b05214760c..4ef8231b677f 100644
@@ -24,7 +24,7 @@ void pte_frag_destroy(void *pte_frag)
      /* drop all the pending references */
      count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
      /* We allow PTE_FRAG_NR fragments from a PTE page */
-     if (atomic_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
+     if (refcount_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
              pgtable_page_dtor(page);
              __free_page(page);
      }
@@ -71,7 +71,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
                      return NULL;
      }
 
-     atomic_set(&page->pt_frag_refcount, 1);
+     refcount_set(&page->pt_frag_refcount, 1);
 
      ret = page_address(page);
      /*
@@ -87,7 +87,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
       * count.
       */
      if (likely(!pte_frag_get(&mm->context))) {
-             atomic_set(&page->pt_frag_refcount, PTE_FRAG_NR);
+             refcount_set(&page->pt_frag_refcount, PTE_FRAG_NR);
              pte_frag_set(&mm->context, ret + PTE_FRAG_SIZE);
      }
      spin_unlock(&mm->page_table_lock);
@@ -110,8 +110,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
 {
      struct page *page = virt_to_page(table);
 
-     BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-     if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+     if (refcount_dec_and_test(&page->pt_frag_refcount)) {
              if (!kernel)
                      pgtable_page_dtor(page);
              __free_page(page);
include/linux/mm_types.h +2 -2
index 3a37a89eb7a7..7fe23a3faf95 100644
@@ -14,6 +14,7 @@
 #include <linux/uprobes.h>
 #include <linux/page-flags-layout.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 
 #include <asm/mmu.h>
 
@@ -147,7 +148,7 @@ struct page {
                      unsigned long _pt_pad_2;        /* mapping */
                      union {
                              struct mm_struct *pt_mm; /* x86 pgds only */
-                             atomic_t pt_frag_refcount; /* powerpc */
+                             refcount_t pt_frag_refcount; /* powerpc */
                      };
 #if ALLOC_SPLIT_PTLOCKS
                      spinlock_t *ptl;
--

2.20.1

#3 lore raw
In reply to Michael Ellerman

Le 09/08/2019 à 14:36, Michael Ellerman a écrit :

Quoted 7 lines
Quoted 6 lines · Chuhong Yuan

Reference counters are preferred to use refcount_t instead of atomic_t.
This is because the implementation of refcount_t can prevent overflows and detect possible use-after-free.
So convert atomic_t ref counters to refcount_t.

Quoted 6 lines
Quoted 1 line

Thanks.

We don't have a fast implementation of refcount_t, so I'm worried this could cause a measurable performance regression.

Fast implementations have been removed by commit https://github.com/linuxppc/linux/commit/fb041bb7c0a918b95c6889fc965cdc4a75b4c0ca

It's now considered that the generic implementation is good enough for everybody.

However, this series doesn't apply anymore and needs rebase:

Applying: powerpc/mm: Use refcount_t for refcount
Using index info to reconstruct a base tree...
M arch/powerpc/mm/book3s64/mmu_context.c
M arch/powerpc/mm/book3s64/pgtable.c
M arch/powerpc/mm/pgtable-frag.c
M include/linux/mm_types.h
Falling back to patching base and 3-way merge... Auto-merging include/linux/mm_types.h
CONFLICT (content): Merge conflict in include/linux/mm_types.h Auto-merging arch/powerpc/mm/pgtable-frag.c
CONFLICT (content): Merge conflict in arch/powerpc/mm/pgtable-frag.c Auto-merging arch/powerpc/mm/book3s64/pgtable.c
CONFLICT (content): Merge conflict in arch/powerpc/mm/book3s64/pgtable.c Auto-merging arch/powerpc/mm/book3s64/mmu_context.c
Patch failed at 0001 powerpc/mm: Use refcount_t for refcount When you have resolved this problem, run "git am --continue". If you prefer to skip this patch, run "git am --skip" instead. To restore the original branch and stop patching, run "git am --abort".

Thanks
Christophe

Quoted 114 lines

Did you benchmark it at all?

cheers

Quoted 109 lines
arch/powerpc/mm/book3s64/mmu_context.c +1 -1
index 2d0cb5ba9a47..f836fd5a6abc 100644
@@ -231,7 +231,7 @@ static void pmd_frag_destroy(void *pmd_frag)
  	/* drop all the pending references */
  	count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
  	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
  		pgtable_pmd_page_dtor(page);
  		__free_page(page);
  	}
arch/powerpc/mm/book3s64/pgtable.c +3 -4
index 7d0e0d0d22c4..40056896ce4e 100644
@@ -277,7 +277,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
  		return NULL;
  	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
  	ret = page_address(page);
  	/*
@@ -294,7 +294,7 @@ static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
  	 * count.
  	 */
  	if (likely(!mm->context.pmd_frag)) {
-		atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PMD_FRAG_NR);
  		mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
  	}
  	spin_unlock(&mm->page_table_lock);
@@ -317,8 +317,7 @@ void pmd_fragment_free(unsigned long *pmd)
  {
  	struct page *page = virt_to_page(pmd);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
  		pgtable_pmd_page_dtor(page);
  		__free_page(page);
  	}
arch/powerpc/mm/pgtable-frag.c +4 -5
index a7b05214760c..4ef8231b677f 100644
@@ -24,7 +24,7 @@ void pte_frag_destroy(void *pte_frag)
  	/* drop all the pending references */
  	count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
  	/* We allow PTE_FRAG_NR fragments from a PTE page */
-	if (atomic_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
+	if (refcount_sub_and_test(PTE_FRAG_NR - count, &page->pt_frag_refcount)) {
  		pgtable_page_dtor(page);
  		__free_page(page);
  	}
@@ -71,7 +71,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
  			return NULL;
  	}
 
-	atomic_set(&page->pt_frag_refcount, 1);
+	refcount_set(&page->pt_frag_refcount, 1);
 
  	ret = page_address(page);
  	/*
@@ -87,7 +87,7 @@ static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
  	 * count.
  	 */
  	if (likely(!pte_frag_get(&mm->context))) {
-		atomic_set(&page->pt_frag_refcount, PTE_FRAG_NR);
+		refcount_set(&page->pt_frag_refcount, PTE_FRAG_NR);
  		pte_frag_set(&mm->context, ret + PTE_FRAG_SIZE);
  	}
  	spin_unlock(&mm->page_table_lock);
@@ -110,8 +110,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
  {
  	struct page *page = virt_to_page(table);
 
-	BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
-	if (atomic_dec_and_test(&page->pt_frag_refcount)) {
+	if (refcount_dec_and_test(&page->pt_frag_refcount)) {
  		if (!kernel)
  			pgtable_page_dtor(page);
  		__free_page(page);
include/linux/mm_types.h +2 -1
index 3a37a89eb7a7..7fe23a3faf95 100644
@@ -14,6 +14,7 @@
  #include <linux/uprobes.h>
  #include <linux/page-flags-layout.h>
  #include <linux/workqueue.h>
+#include <linux/refcount.h>
 
  #include <asm/mmu.h>
 
@@ -147,7 +148,7 @@ struct page {
  			unsigned long _pt_pad_2;	/* mapping */
  			union {
  				struct mm_struct *pt_mm; /* x86 pgds only */
-				atomic_t pt_frag_refcount; /* powerpc */
+				refcount_t pt_frag_refcount; /* powerpc */
  			};
  #if ALLOC_SPLIT_PTLOCKS
  			spinlock_t *ptl;
Signature · 1 line
2.20.1