mm/rmap, migration: Make rmap_walk_anon() and try_to_unmap_anon() more scalable
Ingo Molnar [Sun, 2 Dec 2012 19:56:50 +0000 (19:56 +0000)]
rmap_walk_anon() and try_to_unmap_anon() appears to be too
careful about locking the anon vma: while it needs protection
against anon vma list modifications, it does not need exclusive
access to the list itself.

Transforming this exclusive lock to a read-locked rwsem removes
a global lock from the hot path of page-migration intense
threaded workloads which can cause pathological performance like
this:

    96.43%        process 0  [kernel.kallsyms]  [k] perf_trace_sched_switch
                  |
                  --- perf_trace_sched_switch
                      __schedule
                      schedule
                      schedule_preempt_disabled
                      __mutex_lock_common.isra.6
                      __mutex_lock_slowpath
                      mutex_lock
                     |
                     |--50.61%-- rmap_walk
                     |          move_to_new_page
                     |          migrate_pages
                     |          migrate_misplaced_page
                     |          __do_numa_page.isra.69
                     |          handle_pte_fault
                     |          handle_mm_fault
                     |          __do_page_fault
                     |          do_page_fault
                     |          page_fault
                     |          __memset_sse2
                     |          |
                     |           --100.00%-- worker_thread
                     |                     |
                     |                      --100.00%-- start_thread
                     |
                      --49.39%-- page_lock_anon_vma
                                try_to_unmap_anon
                                try_to_unmap
                                migrate_pages
                                migrate_misplaced_page
                                __do_numa_page.isra.69
                                handle_pte_fault
                                handle_mm_fault
                                __do_page_fault
                                do_page_fault
                                page_fault
                                __memset_sse2
                                |
                                 --100.00%-- worker_thread
                                           start_thread

With this change applied the profile is now nicely flat
and there's no anon-vma related scheduling/blocking.

Rename anon_vma_[un]lock() => anon_vma_[un]lock_write(),
to make it clearer that it's an exclusive write-lock in
that case - suggested by Rik van Riel.

Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Turner <pjt@google.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>

include/linux/huge_mm.h
include/linux/rmap.h
mm/huge_memory.c
mm/ksm.c
mm/memory-failure.c
mm/migrate.c
mm/mmap.c
mm/mremap.c
mm/rmap.c

index 027ad04..0d1208c 100644 (file)
@@ -102,7 +102,7 @@ extern void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd);
 #define wait_split_huge_page(__anon_vma, __pmd)                                \
        do {                                                            \
                pmd_t *____pmd = (__pmd);                               \
-               anon_vma_lock(__anon_vma);                              \
+               anon_vma_lock_write(__anon_vma);                        \
                anon_vma_unlock(__anon_vma);                            \
                BUG_ON(pmd_trans_splitting(*____pmd) ||                 \
                       pmd_trans_huge(*____pmd));                       \
index f3f41d2..c20635c 100644 (file)
@@ -118,7 +118,7 @@ static inline void vma_unlock_anon_vma(struct vm_area_struct *vma)
                up_write(&anon_vma->root->rwsem);
 }
 
-static inline void anon_vma_lock(struct anon_vma *anon_vma)
+static inline void anon_vma_lock_write(struct anon_vma *anon_vma)
 {
        down_write(&anon_vma->root->rwsem);
 }
@@ -128,6 +128,17 @@ static inline void anon_vma_unlock(struct anon_vma *anon_vma)
        up_write(&anon_vma->root->rwsem);
 }
 
+static inline void anon_vma_lock_read(struct anon_vma *anon_vma)
+{
+       down_read(&anon_vma->root->rwsem);
+}
+
+static inline void anon_vma_unlock_read(struct anon_vma *anon_vma)
+{
+       up_read(&anon_vma->root->rwsem);
+}
+
+
 /*
  * anon_vma helper functions.
  */
@@ -220,8 +231,8 @@ int try_to_munlock(struct page *);
 /*
  * Called by memory-failure.c to kill processes.
  */
-struct anon_vma *page_lock_anon_vma(struct page *page);
-void page_unlock_anon_vma(struct anon_vma *anon_vma);
+struct anon_vma *page_lock_anon_vma_read(struct page *page);
+void page_unlock_anon_vma_read(struct anon_vma *anon_vma);
 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma);
 
 /*
index acd37fe..a24c9cb 100644 (file)
@@ -1549,7 +1549,7 @@ int split_huge_page(struct page *page)
        int ret = 1;
 
        BUG_ON(!PageAnon(page));
-       anon_vma = page_lock_anon_vma(page);
+       anon_vma = page_lock_anon_vma_read(page);
        if (!anon_vma)
                goto out;
        ret = 0;
@@ -1562,7 +1562,7 @@ int split_huge_page(struct page *page)
 
        BUG_ON(PageCompound(page));
 out_unlock:
-       page_unlock_anon_vma(anon_vma);
+       page_unlock_anon_vma_read(anon_vma);
 out:
        return ret;
 }
@@ -2074,7 +2074,7 @@ static void collapse_huge_page(struct mm_struct *mm,
        if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
                goto out;
 
-       anon_vma_lock(vma->anon_vma);
+       anon_vma_lock_write(vma->anon_vma);
 
        pte = pte_offset_map(pmd, address);
        ptl = pte_lockptr(mm, pmd);
index ae539f0..7fa37de 100644 (file)
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -1634,7 +1634,7 @@ again:
                struct anon_vma_chain *vmac;
                struct vm_area_struct *vma;
 
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
                                               0, ULONG_MAX) {
                        vma = vmac->vma;
@@ -1688,7 +1688,7 @@ again:
                struct anon_vma_chain *vmac;
                struct vm_area_struct *vma;
 
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
                                               0, ULONG_MAX) {
                        vma = vmac->vma;
@@ -1741,7 +1741,7 @@ again:
                struct anon_vma_chain *vmac;
                struct vm_area_struct *vma;
 
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
                                               0, ULONG_MAX) {
                        vma = vmac->vma;
index ddb68a1..f2cd830 100644 (file)
@@ -402,7 +402,7 @@ static void collect_procs_anon(struct page *page, struct list_head *to_kill,
        struct anon_vma *av;
        pgoff_t pgoff;
 
-       av = page_lock_anon_vma(page);
+       av = page_lock_anon_vma_read(page);
        if (av == NULL) /* Not actually mapped anymore */
                return;
 
@@ -423,7 +423,7 @@ static void collect_procs_anon(struct page *page, struct list_head *to_kill,
                }
        }
        read_unlock(&tasklist_lock);
-       page_unlock_anon_vma(av);
+       page_unlock_anon_vma_read(av);
 }
 
 /*
index f24e9cc..6e46485 100644 (file)
@@ -754,7 +754,7 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
         */
        if (PageAnon(page)) {
                /*
-                * Only page_lock_anon_vma() understands the subtleties of
+                * Only page_lock_anon_vma_read() understands the subtleties of
                 * getting a hold on an anon_vma from outside one of its mms.
                 */
                anon_vma = page_get_anon_vma(page);
index 8840863..68a16b4 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -602,7 +602,7 @@ again:                      remove_next = 1 + (end > next->vm_end);
        if (anon_vma) {
                VM_BUG_ON(adjust_next && next->anon_vma &&
                          anon_vma != next->anon_vma);
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                anon_vma_interval_tree_pre_update_vma(vma);
                if (adjust_next)
                        anon_vma_interval_tree_pre_update_vma(next);
index 1b61c2d..3dabd17 100644 (file)
@@ -104,7 +104,7 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
                }
                if (vma->anon_vma) {
                        anon_vma = vma->anon_vma;
-                       anon_vma_lock(anon_vma);
+                       anon_vma_lock_write(anon_vma);
                }
        }
 
index 6e3ee3b..b0f612d 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -87,24 +87,24 @@ static inline void anon_vma_free(struct anon_vma *anon_vma)
        VM_BUG_ON(atomic_read(&anon_vma->refcount));
 
        /*
-        * Synchronize against page_lock_anon_vma() such that
+        * Synchronize against page_lock_anon_vma_read() such that
         * we can safely hold the lock without the anon_vma getting
         * freed.
         *
         * Relies on the full mb implied by the atomic_dec_and_test() from
         * put_anon_vma() against the acquire barrier implied by
-        * mutex_trylock() from page_lock_anon_vma(). This orders:
+        * down_read_trylock() from page_lock_anon_vma_read(). This orders:
         *
-        * page_lock_anon_vma()         VS      put_anon_vma()
-        *   mutex_trylock()                      atomic_dec_and_test()
+        * page_lock_anon_vma_read()    VS      put_anon_vma()
+        *   down_read_trylock()                  atomic_dec_and_test()
         *   LOCK                                 MB
-        *   atomic_read()                        mutex_is_locked()
+        *   atomic_read()                        rwsem_is_locked()
         *
         * LOCK should suffice since the actual taking of the lock must
         * happen _before_ what follows.
         */
        if (rwsem_is_locked(&anon_vma->root->rwsem)) {
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                anon_vma_unlock(anon_vma);
        }
 
@@ -146,7 +146,7 @@ static void anon_vma_chain_link(struct vm_area_struct *vma,
  * allocate a new one.
  *
  * Anon-vma allocations are very subtle, because we may have
- * optimistically looked up an anon_vma in page_lock_anon_vma()
+ * optimistically looked up an anon_vma in page_lock_anon_vma_read()
  * and that may actually touch the spinlock even in the newly
  * allocated vma (it depends on RCU to make sure that the
  * anon_vma isn't actually destroyed).
@@ -181,7 +181,7 @@ int anon_vma_prepare(struct vm_area_struct *vma)
                        allocated = anon_vma;
                }
 
-               anon_vma_lock(anon_vma);
+               anon_vma_lock_write(anon_vma);
                /* page_table_lock to protect against threads */
                spin_lock(&mm->page_table_lock);
                if (likely(!vma->anon_vma)) {
@@ -306,7 +306,7 @@ int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
        get_anon_vma(anon_vma->root);
        /* Mark this anon_vma as the one where our new (COWed) pages go. */
        vma->anon_vma = anon_vma;
-       anon_vma_lock(anon_vma);
+       anon_vma_lock_write(anon_vma);
        anon_vma_chain_link(vma, avc, anon_vma);
        anon_vma_unlock(anon_vma);
 
@@ -442,7 +442,7 @@ out:
  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
  * reference like with page_get_anon_vma() and then block on the mutex.
  */
-struct anon_vma *page_lock_anon_vma(struct page *page)
+struct anon_vma *page_lock_anon_vma_read(struct page *page)
 {
        struct anon_vma *anon_vma = NULL;
        struct anon_vma *root_anon_vma;
@@ -457,14 +457,14 @@ struct anon_vma *page_lock_anon_vma(struct page *page)
 
        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
        root_anon_vma = ACCESS_ONCE(anon_vma->root);
-       if (down_write_trylock(&root_anon_vma->rwsem)) {
+       if (down_read_trylock(&root_anon_vma->rwsem)) {
                /*
                 * If the page is still mapped, then this anon_vma is still
                 * its anon_vma, and holding the mutex ensures that it will
                 * not go away, see anon_vma_free().
                 */
                if (!page_mapped(page)) {
-                       up_write(&root_anon_vma->rwsem);
+                       up_read(&root_anon_vma->rwsem);
                        anon_vma = NULL;
                }
                goto out;
@@ -484,15 +484,15 @@ struct anon_vma *page_lock_anon_vma(struct page *page)
 
        /* we pinned the anon_vma, its safe to sleep */
        rcu_read_unlock();
-       anon_vma_lock(anon_vma);
+       anon_vma_lock_read(anon_vma);
 
        if (atomic_dec_and_test(&anon_vma->refcount)) {
                /*
                 * Oops, we held the last refcount, release the lock
                 * and bail -- can't simply use put_anon_vma() because
-                * we'll deadlock on the anon_vma_lock() recursion.
+                * we'll deadlock on the anon_vma_lock_write() recursion.
                 */
-               anon_vma_unlock(anon_vma);
+               anon_vma_unlock_read(anon_vma);
                __put_anon_vma(anon_vma);
                anon_vma = NULL;
        }
@@ -504,9 +504,9 @@ out:
        return anon_vma;
 }
 
-void page_unlock_anon_vma(struct anon_vma *anon_vma)
+void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
 {
-       anon_vma_unlock(anon_vma);
+       anon_vma_unlock_read(anon_vma);
 }
 
 /*
@@ -732,7 +732,7 @@ static int page_referenced_anon(struct page *page,
        struct anon_vma_chain *avc;
        int referenced = 0;
 
-       anon_vma = page_lock_anon_vma(page);
+       anon_vma = page_lock_anon_vma_read(page);
        if (!anon_vma)
                return referenced;
 
@@ -754,7 +754,7 @@ static int page_referenced_anon(struct page *page,
                        break;
        }
 
-       page_unlock_anon_vma(anon_vma);
+       page_unlock_anon_vma_read(anon_vma);
        return referenced;
 }
 
@@ -1474,7 +1474,7 @@ static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
        struct anon_vma_chain *avc;
        int ret = SWAP_AGAIN;
 
-       anon_vma = page_lock_anon_vma(page);
+       anon_vma = page_lock_anon_vma_read(page);
        if (!anon_vma)
                return ret;
 
@@ -1501,7 +1501,7 @@ static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
                        break;
        }
 
-       page_unlock_anon_vma(anon_vma);
+       page_unlock_anon_vma_read(anon_vma);
        return ret;
 }
 
@@ -1696,7 +1696,7 @@ static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
        int ret = SWAP_AGAIN;
 
        /*
-        * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
+        * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
         * because that depends on page_mapped(); but not all its usages
         * are holding mmap_sem. Users without mmap_sem are required to
         * take a reference count to prevent the anon_vma disappearing
@@ -1704,7 +1704,7 @@ static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
        anon_vma = page_anon_vma(page);
        if (!anon_vma)
                return ret;
-       anon_vma_lock(anon_vma);
+       anon_vma_lock_read(anon_vma);
        anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
                struct vm_area_struct *vma = avc->vma;
                unsigned long address = vma_address(page, vma);
@@ -1712,7 +1712,7 @@ static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
                if (ret != SWAP_AGAIN)
                        break;
        }
-       anon_vma_unlock(anon_vma);
+       anon_vma_unlock_read(anon_vma);
        return ret;
 }