mm/rmap: allocate anon_vma_chain objects unlocked when possible

There is no reason to allocate the anon_vma_chain under the anon_vma write
lock when cloning - we can in fact assign these to the destination VMA
safely as we hold the exclusive mmap lock and therefore preclude anybody
else accessing these fields.

We only need take the anon_vma write lock when we link rbtree edges from
the anon_vma to the newly established AVCs.

This also allows us to eliminate the weird GFP_NOWAIT, GFP_KERNEL dance
introduced in commit dd34739c03f2 ("mm: avoid anon_vma_chain allocation
under anon_vma lock"), further simplifying this logic.

This should reduce lock anon_vma contention, and clarifies exactly where
the anon_vma lock is required.

We cannot adjust __anon_vma_prepare() in the same way as this is only
protected by VMA read lock, so we have to perform the allocation here
under the anon_vma write lock and page_table_lock (to protect against
racing threads), and we wish to retain the lock ordering.

With this change we can simplify cleanup_partial_anon_vmas() even further
- since we allocate AVC's without any lock taken and do not insert
anything into the interval tree until after the allocations are tried, we
can remove all logic pertaining to this and just free up AVC's only.

Link: https://lkml.kernel.org/r/624bf1ac0bde4871fcfca2c8c8e294b6d8f7ae7b.1768746221.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chriscli@google.com>
Cc: David Hildenbrand <david@kernel.org>
Cc: Harry Yoo <harry.yoo@oracle.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Pedro Falcato <pfalcato@suse.de>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

diff --git a/mm/rmap.c b/mm/rmap.c
index d4d2e7b9fe5f51ae3f7366370f8aea52508c4346..a5ce9163454a42952cc2c6153e7e5843f95bd78d 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -146,14 +146,13 @@ static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
        kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 }
 
-static void anon_vma_chain_link(struct vm_area_struct *vma,
-                               struct anon_vma_chain *avc,
-                               struct anon_vma *anon_vma)
+static void anon_vma_chain_assign(struct vm_area_struct *vma,
+                                 struct anon_vma_chain *avc,
+                                 struct anon_vma *anon_vma)
 {
        avc->vma = vma;
        avc->anon_vma = anon_vma;
        list_add(&avc->same_vma, &vma->anon_vma_chain);
-       anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 }
 
 /**
@@ -210,7 +209,8 @@ int __anon_vma_prepare(struct vm_area_struct *vma)
        spin_lock(&mm->page_table_lock);
        if (likely(!vma->anon_vma)) {
                vma->anon_vma = anon_vma;
-               anon_vma_chain_link(vma, avc, anon_vma);
+               anon_vma_chain_assign(vma, avc, anon_vma);
+               anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
                anon_vma->num_active_vmas++;
                allocated = NULL;
                avc = NULL;
@@ -291,21 +291,33 @@ int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 
        check_anon_vma_clone(dst, src);
 
-       /* All anon_vma's share the same root. */
+       /*
+        * Allocate AVCs. We don't need an anon_vma lock for this as we
+        * are not updating the anon_vma rbtree nor are we changing
+        * anon_vma statistics.
+        *
+        * Either src, dst have the same mm for which we hold an exclusive mmap
+        * write lock, or we are forking and we hold it on src->vm_mm and dst is
+        * not yet accessible to other threads so there's no possibliity of the
+        * unlinked AVC's being observed yet.
+        */
+       list_for_each_entry(pavc, &src->anon_vma_chain, same_vma) {
+               avc = anon_vma_chain_alloc(GFP_KERNEL);
+               if (!avc)
+                       goto enomem_failure;
+
+               anon_vma_chain_assign(dst, avc, pavc->anon_vma);
+       }
+
+       /*
+        * Now link the anon_vma's back to the newly inserted AVCs.
+        * Note that all anon_vma's share the same root.
+        */
        anon_vma_lock_write(src->anon_vma);
-       list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
-               struct anon_vma *anon_vma;
-
-               avc = anon_vma_chain_alloc(GFP_NOWAIT);
-               if (unlikely(!avc)) {
-                       anon_vma_unlock_write(src->anon_vma);
-                       avc = anon_vma_chain_alloc(GFP_KERNEL);
-                       if (!avc)
-                               goto enomem_failure;
-                       anon_vma_lock_write(src->anon_vma);
-               }
-               anon_vma = pavc->anon_vma;
-               anon_vma_chain_link(dst, avc, anon_vma);
+       list_for_each_entry_reverse(avc, &dst->anon_vma_chain, same_vma) {
+               struct anon_vma *anon_vma = avc->anon_vma;
+
+               anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 
                /*
                 * Reuse existing anon_vma if it has no vma and only one
@@ -321,7 +333,6 @@ int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
        }
        if (dst->anon_vma)
                dst->anon_vma->num_active_vmas++;
-
        anon_vma_unlock_write(src->anon_vma);
        return 0;
 
@@ -391,8 +402,10 @@ 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_chain_assign(vma, avc, anon_vma);
+       /* Now let rmap see it. */
        anon_vma_lock_write(anon_vma);
-       anon_vma_chain_link(vma, avc, anon_vma);
+       anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
        anon_vma->parent->num_children++;
        anon_vma_unlock_write(anon_vma);
 
@@ -403,40 +416,18 @@ int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
  * In the unfortunate case of anon_vma_clone() failing to allocate memory we
  * have to clean things up.
  *
- * On clone we hold the exclusive mmap write lock, so we can't race
- * unlink_anon_vmas(). Since we're cloning, we know we can't have empty
- * anon_vma's, since existing anon_vma's are what we're cloning from.
- *
- * So this function needs only traverse the anon_vma_chain and free each
- * allocated anon_vma_chain.
+ * Since we allocate anon_vma_chain's before we insert them into the interval
+ * trees, we simply have to free up the AVC's and remove the entries from the
+ * VMA's anon_vma_chain.
  */
 static void cleanup_partial_anon_vmas(struct vm_area_struct *vma)
 {
        struct anon_vma_chain *avc, *next;
-       struct anon_vma *root = NULL;
-
-       /*
-        * We exclude everybody else from being able to modify anon_vma's
-        * underneath us.
-        */
-       mmap_assert_locked(vma->vm_mm);
 
        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
-               struct anon_vma *anon_vma = avc->anon_vma;
-
-               /* All anon_vma's share the same root. */
-               if (!root) {
-                       root = anon_vma->root;
-                       anon_vma_lock_write(root);
-               }
-
-               anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
                list_del(&avc->same_vma);
                anon_vma_chain_free(avc);
        }
-
-       if (root)
-               anon_vma_unlock_write(root);
 }
 
 /**