mm: make compound_head() robust
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Sat, 7 Nov 2015 00:29:54 +0000 (16:29 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 7 Nov 2015 01:50:42 +0000 (17:50 -0800)
Hugh has pointed that compound_head() call can be unsafe in some
context. There's one example:

CPU0 CPU1

isolate_migratepages_block()
  page_count()
    compound_head()
      !!PageTail() == true
put_page()
  tail->first_page = NULL
      head = tail->first_page
alloc_pages(__GFP_COMP)
   prep_compound_page()
     tail->first_page = head
     __SetPageTail(p);
      !!PageTail() == true
    <head == NULL dereferencing>

The race is pure theoretical. I don't it's possible to trigger it in
practice. But who knows.

We can fix the race by changing how encode PageTail() and compound_head()
within struct page to be able to update them in one shot.

The patch introduces page->compound_head into third double word block in
front of compound_dtor and compound_order. Bit 0 encodes PageTail() and
the rest bits are pointer to head page if bit zero is set.

The patch moves page->pmd_huge_pte out of word, just in case if an
architecture defines pgtable_t into something what can have the bit 0
set.

hugetlb_cgroup uses page->lru.next in the second tail page to store
pointer struct hugetlb_cgroup. The patch switch it to use page->private
in the second tail page instead. The space is free since ->first_page is
removed from the union.

The patch also opens possibility to remove HUGETLB_CGROUP_MIN_ORDER
limitation, since there's now space in first tail page to store struct
hugetlb_cgroup pointer. But that's out of scope of the patch.

That means page->compound_head shares storage space with:

 - page->lru.next;
 - page->next;
 - page->rcu_head.next;

That's too long list to be absolutely sure, but looks like nobody uses
bit 0 of the word.

page->rcu_head.next guaranteed[1] to have bit 0 clean as long as we use
call_rcu(), call_rcu_bh(), call_rcu_sched(), or call_srcu(). But future
call_rcu_lazy() is not allowed as it makes use of the bit and we can
get false positive PageTail().

[1] http://lkml.kernel.org/g/20150827163634.GD4029@linux.vnet.ibm.com

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
15 files changed:
Documentation/vm/split_page_table_lock
arch/xtensa/configs/iss_defconfig
include/linux/hugetlb_cgroup.h
include/linux/mm.h
include/linux/mm_types.h
include/linux/page-flags.h
mm/Kconfig
mm/debug.c
mm/huge_memory.c
mm/hugetlb.c
mm/hugetlb_cgroup.c
mm/internal.h
mm/memory-failure.c
mm/page_alloc.c
mm/swap.c

index 6dea4fd5c96100d75f9152ffd8ac9be5909acf5f..62842a857dab32477f90ad84f010bacb883d0aa6 100644 (file)
@@ -54,8 +54,8 @@ everything required is done by pgtable_page_ctor() and pgtable_page_dtor(),
 which must be called on PTE table allocation / freeing.
 
 Make sure the architecture doesn't use slab allocator for page table
-allocation: slab uses page->slab_cache and page->first_page for its pages.
-These fields share storage with page->ptl.
+allocation: slab uses page->slab_cache for its pages.
+This field shares storage with page->ptl.
 
 PMD split lock only makes sense if you have more than two page table
 levels.
index f3dfe0d921c2759a2ff201e3b6b7425078338004..44c6764d9146347d713a8d377d612ae7970f9e96 100644 (file)
@@ -169,7 +169,6 @@ CONFIG_FLATMEM_MANUAL=y
 # CONFIG_SPARSEMEM_MANUAL is not set
 CONFIG_FLATMEM=y
 CONFIG_FLAT_NODE_MEM_MAP=y
-CONFIG_PAGEFLAGS_EXTENDED=y
 CONFIG_SPLIT_PTLOCK_CPUS=4
 # CONFIG_PHYS_ADDR_T_64BIT is not set
 CONFIG_ZONE_DMA_FLAG=1
index 7edd305152983af1ab6aee93f470dd99289046e8..24154c26d469c60984020b5a0441fcb6dde3fcb0 100644 (file)
@@ -32,7 +32,7 @@ static inline struct hugetlb_cgroup *hugetlb_cgroup_from_page(struct page *page)
 
        if (compound_order(page) < HUGETLB_CGROUP_MIN_ORDER)
                return NULL;
-       return (struct hugetlb_cgroup *)page[2].lru.next;
+       return (struct hugetlb_cgroup *)page[2].private;
 }
 
 static inline
@@ -42,7 +42,7 @@ int set_hugetlb_cgroup(struct page *page, struct hugetlb_cgroup *h_cg)
 
        if (compound_order(page) < HUGETLB_CGROUP_MIN_ORDER)
                return -1;
-       page[2].lru.next = (void *)h_cg;
+       page[2].private = (unsigned long)h_cg;
        return 0;
 }
 
index 6581c21320cbb67819a057553c50aec969c8a962..9671b6f23edac81ced458674e715a267e4c434f6 100644 (file)
@@ -430,46 +430,6 @@ static inline void compound_unlock_irqrestore(struct page *page,
 #endif
 }
 
-static inline struct page *compound_head_by_tail(struct page *tail)
-{
-       struct page *head = tail->first_page;
-
-       /*
-        * page->first_page may be a dangling pointer to an old
-        * compound page, so recheck that it is still a tail
-        * page before returning.
-        */
-       smp_rmb();
-       if (likely(PageTail(tail)))
-               return head;
-       return tail;
-}
-
-/*
- * Since either compound page could be dismantled asynchronously in THP
- * or we access asynchronously arbitrary positioned struct page, there
- * would be tail flag race. To handle this race, we should call
- * smp_rmb() before checking tail flag. compound_head_by_tail() did it.
- */
-static inline struct page *compound_head(struct page *page)
-{
-       if (unlikely(PageTail(page)))
-               return compound_head_by_tail(page);
-       return page;
-}
-
-/*
- * If we access compound page synchronously such as access to
- * allocated page, there is no need to handle tail flag race, so we can
- * check tail flag directly without any synchronization primitive.
- */
-static inline struct page *compound_head_fast(struct page *page)
-{
-       if (unlikely(PageTail(page)))
-               return page->first_page;
-       return page;
-}
-
 /*
  * The atomic page->_mapcount, starts from -1: so that transitions
  * both from it and to it can be tracked, using atomic_inc_and_test
@@ -518,7 +478,7 @@ static inline void get_huge_page_tail(struct page *page)
        VM_BUG_ON_PAGE(!PageTail(page), page);
        VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
        VM_BUG_ON_PAGE(atomic_read(&page->_count) != 0, page);
-       if (compound_tail_refcounted(page->first_page))
+       if (compound_tail_refcounted(compound_head(page)))
                atomic_inc(&page->_mapcount);
 }
 
@@ -541,13 +501,7 @@ static inline struct page *virt_to_head_page(const void *x)
 {
        struct page *page = virt_to_page(x);
 
-       /*
-        * We don't need to worry about synchronization of tail flag
-        * when we call virt_to_head_page() since it is only called for
-        * already allocated page and this page won't be freed until
-        * this virt_to_head_page() is finished. So use _fast variant.
-        */
-       return compound_head_fast(page);
+       return compound_head(page);
 }
 
 /*
@@ -1586,8 +1540,7 @@ static inline bool ptlock_init(struct page *page)
         * with 0. Make sure nobody took it in use in between.
         *
         * It can happen if arch try to use slab for page table allocation:
-        * slab code uses page->slab_cache and page->first_page (for tail
-        * pages), which share storage with page->ptl.
+        * slab code uses page->slab_cache, which share storage with page->ptl.
         */
        VM_BUG_ON_PAGE(*(unsigned long *)&page->ptl, page);
        if (!ptlock_alloc(page))
index e334ef79cb436dda88dc7bc409a60d28430744e2..bb91658c603f65be58e687280e1c0249ca63ffd7 100644 (file)
@@ -111,7 +111,13 @@ struct page {
                };
        };
 
-       /* Third double word block */
+       /*
+        * Third double word block
+        *
+        * WARNING: bit 0 of the first word encode PageTail(). That means
+        * the rest users of the storage space MUST NOT use the bit to
+        * avoid collision and false-positive PageTail().
+        */
        union {
                struct list_head lru;   /* Pageout list, eg. active_list
                                         * protected by zone->lru_lock !
@@ -132,14 +138,23 @@ struct page {
                struct rcu_head rcu_head;       /* Used by SLAB
                                                 * when destroying via RCU
                                                 */
-               /* First tail page of compound page */
+               /* Tail pages of compound page */
                struct {
+                       unsigned long compound_head; /* If bit zero is set */
+
+                       /* First tail page only */
                        unsigned short int compound_dtor;
                        unsigned short int compound_order;
                };
 
 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && USE_SPLIT_PMD_PTLOCKS
-               pgtable_t pmd_huge_pte; /* protected by page->ptl */
+               struct {
+                       unsigned long __pad;    /* do not overlay pmd_huge_pte
+                                                * with compound_head to avoid
+                                                * possible bit 0 collision.
+                                                */
+                       pgtable_t pmd_huge_pte; /* protected by page->ptl */
+               };
 #endif
        };
 
@@ -160,7 +175,6 @@ struct page {
 #endif
 #endif
                struct kmem_cache *slab_cache;  /* SL[AU]B: Pointer to slab */
-               struct page *first_page;        /* Compound tail pages */
        };
 
 #ifdef CONFIG_MEMCG
index a525e50674844b24a994fde2554bea17a2700e4f..bb53c7b863152468196a390eb269449a03949e4d 100644 (file)
@@ -86,12 +86,7 @@ enum pageflags {
        PG_private,             /* If pagecache, has fs-private data */
        PG_private_2,           /* If pagecache, has fs aux data */
        PG_writeback,           /* Page is under writeback */
-#ifdef CONFIG_PAGEFLAGS_EXTENDED
        PG_head,                /* A head page */
-       PG_tail,                /* A tail page */
-#else
-       PG_compound,            /* A compound page */
-#endif
        PG_swapcache,           /* Swap page: swp_entry_t in private */
        PG_mappedtodisk,        /* Has blocks allocated on-disk */
        PG_reclaim,             /* To be reclaimed asap */
@@ -398,85 +393,46 @@ static inline void set_page_writeback_keepwrite(struct page *page)
        test_set_page_writeback_keepwrite(page);
 }
 
-#ifdef CONFIG_PAGEFLAGS_EXTENDED
-/*
- * System with lots of page flags available. This allows separate
- * flags for PageHead() and PageTail() checks of compound pages so that bit
- * tests can be used in performance sensitive paths. PageCompound is
- * generally not used in hot code paths except arch/powerpc/mm/init_64.c
- * and arch/powerpc/kvm/book3s_64_vio_hv.c which use it to detect huge pages
- * and avoid handling those in real mode.
- */
 __PAGEFLAG(Head, head) CLEARPAGEFLAG(Head, head)
-__PAGEFLAG(Tail, tail)
 
-static inline int PageCompound(struct page *page)
-{
-       return page->flags & ((1L << PG_head) | (1L << PG_tail));
-
-}
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-static inline void ClearPageCompound(struct page *page)
+static inline int PageTail(struct page *page)
 {
-       BUG_ON(!PageHead(page));
-       ClearPageHead(page);
+       return READ_ONCE(page->compound_head) & 1;
 }
-#endif
-
-#define PG_head_mask ((1L << PG_head))
 
-#else
-/*
- * Reduce page flag use as much as possible by overlapping
- * compound page flags with the flags used for page cache pages. Possible
- * because PageCompound is always set for compound pages and not for
- * pages on the LRU and/or pagecache.
- */
-TESTPAGEFLAG(Compound, compound)
-__SETPAGEFLAG(Head, compound)  __CLEARPAGEFLAG(Head, compound)
-
-/*
- * PG_reclaim is used in combination with PG_compound to mark the
- * head and tail of a compound page. This saves one page flag
- * but makes it impossible to use compound pages for the page cache.
- * The PG_reclaim bit would have to be used for reclaim or readahead
- * if compound pages enter the page cache.
- *
- * PG_compound & PG_reclaim    => Tail page
- * PG_compound & ~PG_reclaim   => Head page
- */
-#define PG_head_mask ((1L << PG_compound))
-#define PG_head_tail_mask ((1L << PG_compound) | (1L << PG_reclaim))
-
-static inline int PageHead(struct page *page)
+static inline void set_compound_head(struct page *page, struct page *head)
 {
-       return ((page->flags & PG_head_tail_mask) == PG_head_mask);
+       WRITE_ONCE(page->compound_head, (unsigned long)head + 1);
 }
 
-static inline int PageTail(struct page *page)
+static inline void clear_compound_head(struct page *page)
 {
-       return ((page->flags & PG_head_tail_mask) == PG_head_tail_mask);
+       WRITE_ONCE(page->compound_head, 0);
 }
 
-static inline void __SetPageTail(struct page *page)
+static inline struct page *compound_head(struct page *page)
 {
-       page->flags |= PG_head_tail_mask;
+       unsigned long head = READ_ONCE(page->compound_head);
+
+       if (unlikely(head & 1))
+               return (struct page *) (head - 1);
+       return page;
 }
 
-static inline void __ClearPageTail(struct page *page)
+static inline int PageCompound(struct page *page)
 {
-       page->flags &= ~PG_head_tail_mask;
-}
+       return PageHead(page) || PageTail(page);
 
+}
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 static inline void ClearPageCompound(struct page *page)
 {
-       BUG_ON((page->flags & PG_head_tail_mask) != (1 << PG_compound));
-       clear_bit(PG_compound, &page->flags);
+       BUG_ON(!PageHead(page));
+       ClearPageHead(page);
 }
 #endif
 
-#endif /* !PAGEFLAGS_EXTENDED */
+#define PG_head_mask ((1L << PG_head))
 
 #ifdef CONFIG_HUGETLB_PAGE
 int PageHuge(struct page *page);
index 0d9fdcd01e479d87a45cb487db54e8c1fff27883..97a4e06b15c00dfd7f8b8bc0f1e1e4610b769938 100644 (file)
@@ -200,18 +200,6 @@ config MEMORY_HOTREMOVE
        depends on MEMORY_HOTPLUG && ARCH_ENABLE_MEMORY_HOTREMOVE
        depends on MIGRATION
 
-#
-# If we have space for more page flags then we can enable additional
-# optimizations and functionality.
-#
-# Regular Sparsemem takes page flag bits for the sectionid if it does not
-# use a virtual memmap. Disable extended page flags for 32 bit platforms
-# that require the use of a sectionid in the page flags.
-#
-config PAGEFLAGS_EXTENDED
-       def_bool y
-       depends on 64BIT || SPARSEMEM_VMEMMAP || !SPARSEMEM
-
 # Heavily threaded applications may benefit from splitting the mm-wide
 # page_table_lock, so that faults on different parts of the user address
 # space can be handled with less contention: split it at this NR_CPUS.
index e784110fb51d1056e28ada2bd5bd26b859e68778..668aa35191ca1243d9be055b5e03ed9b17e17976 100644 (file)
@@ -25,12 +25,7 @@ static const struct trace_print_flags pageflag_names[] = {
        {1UL << PG_private,             "private"       },
        {1UL << PG_private_2,           "private_2"     },
        {1UL << PG_writeback,           "writeback"     },
-#ifdef CONFIG_PAGEFLAGS_EXTENDED
        {1UL << PG_head,                "head"          },
-       {1UL << PG_tail,                "tail"          },
-#else
-       {1UL << PG_compound,            "compound"      },
-#endif
        {1UL << PG_swapcache,           "swapcache"     },
        {1UL << PG_mappedtodisk,        "mappedtodisk"  },
        {1UL << PG_reclaim,             "reclaim"       },
index 73266ee7274c650c06925456a18bc8fb6fb407c5..e1ccc83f73d3b3cdaf0f07fd6db346d44d1d7476 100644 (file)
@@ -1755,8 +1755,7 @@ static void __split_huge_page_refcount(struct page *page,
                                      (1L << PG_unevictable)));
                page_tail->flags |= (1L << PG_dirty);
 
-               /* clear PageTail before overwriting first_page */
-               smp_wmb();
+               clear_compound_head(page_tail);
 
                if (page_is_young(page))
                        set_page_young(page_tail);
index e90a29024c5c69c7d0145b5b013b67bdda66d534..4eb0f0964883bb5f428bb4f6c7c81c0d6c7ba6f9 100644 (file)
@@ -1001,9 +1001,8 @@ static void destroy_compound_gigantic_page(struct page *page,
        struct page *p = page + 1;
 
        for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
-               __ClearPageTail(p);
+               clear_compound_head(p);
                set_page_refcounted(p);
-               p->first_page = NULL;
        }
 
        set_compound_order(page, 0);
@@ -1276,10 +1275,7 @@ static void prep_compound_gigantic_page(struct page *page, unsigned long order)
                 */
                __ClearPageReserved(p);
                set_page_count(p, 0);
-               p->first_page = page;
-               /* Make sure p->first_page is always valid for PageTail() */
-               smp_wmb();
-               __SetPageTail(p);
+               set_compound_head(p, page);
        }
 }
 
index 33d59abe91f1a72a8e98ae5b25310ebdafb9bc17..d8fb10de0f1449f0cd912aaa65d6887c338b20aa 100644 (file)
@@ -385,7 +385,7 @@ void __init hugetlb_cgroup_file_init(void)
                /*
                 * Add cgroup control files only if the huge page consists
                 * of more than two normal pages. This is because we use
-                * page[2].lru.next for storing cgroup details.
+                * page[2].private for storing cgroup details.
                 */
                if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
                        __hugetlb_cgroup_file_init(hstate_index(h));
index 5b7841f6fa2797cd8b11835baaded9bc4210b989..a7f5670fea23803b4381b0556b6b256d93f69a11 100644 (file)
@@ -80,9 +80,9 @@ static inline void __get_page_tail_foll(struct page *page,
         * speculative page access (like in
         * page_cache_get_speculative()) on tail pages.
         */
-       VM_BUG_ON_PAGE(atomic_read(&page->first_page->_count) <= 0, page);
+       VM_BUG_ON_PAGE(atomic_read(&compound_head(page)->_count) <= 0, page);
        if (get_page_head)
-               atomic_inc(&page->first_page->_count);
+               atomic_inc(&compound_head(page)->_count);
        get_huge_page_tail(page);
 }
 
index 16a0ec385320a06e09ddde61b804b51c63d952c6..8424b64711ac35955772078804b2e19f7cd99620 100644 (file)
@@ -776,8 +776,6 @@ static int me_huge_page(struct page *p, unsigned long pfn)
 #define lru            (1UL << PG_lru)
 #define swapbacked     (1UL << PG_swapbacked)
 #define head           (1UL << PG_head)
-#define tail           (1UL << PG_tail)
-#define compound       (1UL << PG_compound)
 #define slab           (1UL << PG_slab)
 #define reserved       (1UL << PG_reserved)
 
@@ -800,12 +798,7 @@ static struct page_state {
         */
        { slab,         slab,           MF_MSG_SLAB,    me_kernel },
 
-#ifdef CONFIG_PAGEFLAGS_EXTENDED
        { head,         head,           MF_MSG_HUGE,            me_huge_page },
-       { tail,         tail,           MF_MSG_HUGE,            me_huge_page },
-#else
-       { compound,     compound,       MF_MSG_HUGE,            me_huge_page },
-#endif
 
        { sc|dirty,     sc|dirty,       MF_MSG_DIRTY_SWAPCACHE, me_swapcache_dirty },
        { sc|dirty,     sc,             MF_MSG_CLEAN_SWAPCACHE, me_swapcache_clean },
index fae1bd6f9f37a808e782d8c193bdc4c8ae0c64a1..e361001519d3e44fcc50a9caf5af51a512ae4413 100644 (file)
@@ -445,15 +445,15 @@ out:
 /*
  * Higher-order pages are called "compound pages".  They are structured thusly:
  *
- * The first PAGE_SIZE page is called the "head page".
+ * The first PAGE_SIZE page is called the "head page" and have PG_head set.
  *
- * The remaining PAGE_SIZE pages are called "tail pages".
+ * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
+ * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
  *
- * All pages have PG_compound set.  All tail pages have their ->first_page
- * pointing at the head page.
+ * The first tail page's ->compound_dtor holds the offset in array of compound
+ * page destructors. See compound_page_dtors.
  *
- * The first tail page's ->lru.next holds the address of the compound page's
- * put_page() function.  Its ->lru.prev holds the order of allocation.
+ * The first tail page's ->compound_order holds the order of allocation.
  * This usage means that zero-order pages may not be compound.
  */
 
@@ -473,10 +473,7 @@ void prep_compound_page(struct page *page, unsigned long order)
        for (i = 1; i < nr_pages; i++) {
                struct page *p = page + i;
                set_page_count(p, 0);
-               p->first_page = page;
-               /* Make sure p->first_page is always valid for PageTail() */
-               smp_wmb();
-               __SetPageTail(p);
+               set_compound_head(p, page);
        }
 }
 
@@ -854,17 +851,30 @@ static void free_one_page(struct zone *zone,
 
 static int free_tail_pages_check(struct page *head_page, struct page *page)
 {
-       if (!IS_ENABLED(CONFIG_DEBUG_VM))
-               return 0;
+       int ret = 1;
+
+       /*
+        * We rely page->lru.next never has bit 0 set, unless the page
+        * is PageTail(). Let's make sure that's true even for poisoned ->lru.
+        */
+       BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
+
+       if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
+               ret = 0;
+               goto out;
+       }
        if (unlikely(!PageTail(page))) {
                bad_page(page, "PageTail not set", 0);
-               return 1;
+               goto out;
        }
-       if (unlikely(page->first_page != head_page)) {
-               bad_page(page, "first_page not consistent", 0);
-               return 1;
+       if (unlikely(compound_head(page) != head_page)) {
+               bad_page(page, "compound_head not consistent", 0);
+               goto out;
        }
-       return 0;
+       ret = 0;
+out:
+       clear_compound_head(page);
+       return ret;
 }
 
 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
@@ -931,6 +941,10 @@ void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
                        struct page *page = pfn_to_page(start_pfn);
 
                        init_reserved_page(start_pfn);
+
+                       /* Avoid false-positive PageTail() */
+                       INIT_LIST_HEAD(&page->lru);
+
                        SetPageReserved(page);
                }
        }
index 983f692a47fdfbb80505fa77f673b9af37d08739..39395fb549c013e20b34d5fe0be8afa6eee5afbf 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -201,7 +201,7 @@ out_put_single:
                                __put_single_page(page);
                        return;
                }
-               VM_BUG_ON_PAGE(page_head != page->first_page, page);
+               VM_BUG_ON_PAGE(page_head != compound_head(page), page);
                /*
                 * We can release the refcount taken by
                 * get_page_unless_zero() now that
@@ -262,7 +262,7 @@ static void put_compound_page(struct page *page)
         *  Case 3 is possible, as we may race with
         *  __split_huge_page_refcount tearing down a THP page.
         */
-       page_head = compound_head_by_tail(page);
+       page_head = compound_head(page);
        if (!__compound_tail_refcounted(page_head))
                put_unrefcounted_compound_page(page_head, page);
        else