]> git.kernelconcepts.de Git - karo-tx-linux.git/blobdiff - mm/ksm.c
mm/zswap.c: improve a size determination in zswap_frontswap_init()
[karo-tx-linux.git] / mm / ksm.c
index 5621840401c05ef439764bdf74f59fed8b5fbc3f..4dc92f138786988c4ef0f9d371ff8a48b2e6e905 100644 (file)
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -1313,14 +1313,14 @@ bool is_page_sharing_candidate(struct stable_node *stable_node)
        return __is_page_sharing_candidate(stable_node, 0);
 }
 
-static struct stable_node *stable_node_dup(struct stable_node **_stable_node,
-                                          struct page **tree_page,
-                                          struct rb_root *root,
-                                          bool prune_stale_stable_nodes)
+struct page *stable_node_dup(struct stable_node **_stable_node_dup,
+                            struct stable_node **_stable_node,
+                            struct rb_root *root,
+                            bool prune_stale_stable_nodes)
 {
        struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
        struct hlist_node *hlist_safe;
-       struct page *_tree_page;
+       struct page *_tree_page, *tree_page = NULL;
        int nr = 0;
        int found_rmap_hlist_len;
 
@@ -1353,27 +1353,28 @@ static struct stable_node *stable_node_dup(struct stable_node **_stable_node,
                        if (!found ||
                            dup->rmap_hlist_len > found_rmap_hlist_len) {
                                if (found)
-                                       put_page(*tree_page);
+                                       put_page(tree_page);
                                found = dup;
                                found_rmap_hlist_len = found->rmap_hlist_len;
-                               *tree_page = _tree_page;
+                               tree_page = _tree_page;
 
+                               /* skip put_page for found dup */
                                if (!prune_stale_stable_nodes)
                                        break;
-                               /* skip put_page */
                                continue;
                        }
                }
                put_page(_tree_page);
        }
 
-       /*
-        * nr is relevant only if prune_stale_stable_nodes is true,
-        * otherwise we may break the loop at nr == 1 even if there
-        * are multiple entries.
-        */
-       if (prune_stale_stable_nodes && found) {
-               if (nr == 1) {
+       if (found) {
+               /*
+                * nr is counting all dups in the chain only if
+                * prune_stale_stable_nodes is true, otherwise we may
+                * break the loop at nr == 1 even if there are
+                * multiple entries.
+                */
+               if (prune_stale_stable_nodes && nr == 1) {
                        /*
                         * If there's not just one entry it would
                         * corrupt memory, better BUG_ON. In KSM
@@ -1392,20 +1393,34 @@ static struct stable_node *stable_node_dup(struct stable_node **_stable_node,
                        ksm_stable_node_chains--;
                        ksm_stable_node_dups--;
                        /*
-                        * NOTE: the caller depends on the
-                        * *_stable_node to become NULL if the chain
-                        * was collapsed. Enforce that if anything
-                        * uses a stale (freed) stable_node chain a
-                        * visible crash will materialize (instead of
-                        * an use after free).
+                        * NOTE: the caller depends on the stable_node
+                        * to be equal to stable_node_dup if the chain
+                        * was collapsed.
                         */
-                       *_stable_node = stable_node = NULL;
-               } else if (__is_page_sharing_candidate(found, 1)) {
+                       *_stable_node = found;
                        /*
-                        * Refile our candidate at the head
-                        * after the prune if our candidate
-                        * can accept one more future sharing
-                        * in addition to the one underway.
+                        * Just for robustneess as stable_node is
+                        * otherwise left as a stable pointer, the
+                        * compiler shall optimize it away at build
+                        * time.
+                        */
+                       stable_node = NULL;
+               } else if (stable_node->hlist.first != &found->hlist_dup &&
+                          __is_page_sharing_candidate(found, 1)) {
+                       /*
+                        * If the found stable_node dup can accept one
+                        * more future merge (in addition to the one
+                        * that is underway) and is not at the head of
+                        * the chain, put it there so next search will
+                        * be quicker in the !prune_stale_stable_nodes
+                        * case.
+                        *
+                        * NOTE: it would be inaccurate to use nr > 1
+                        * instead of checking the hlist.first pointer
+                        * directly, because in the
+                        * prune_stale_stable_nodes case "nr" isn't
+                        * the position of the found dup in the chain,
+                        * but the total number of dups in the chain.
                         */
                        hlist_del(&found->hlist_dup);
                        hlist_add_head(&found->hlist_dup,
@@ -1413,7 +1428,8 @@ static struct stable_node *stable_node_dup(struct stable_node **_stable_node,
                }
        }
 
-       return found;
+       *_stable_node_dup = found;
+       return tree_page;
 }
 
 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
@@ -1429,35 +1445,60 @@ static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
                           typeof(*stable_node), hlist_dup);
 }
 
-static struct stable_node *__stable_node_chain(struct stable_node **_stable_node,
-                                              struct page **tree_page,
-                                              struct rb_root *root,
-                                              bool prune_stale_stable_nodes)
+/*
+ * Like for get_ksm_page, this function can free the *_stable_node and
+ * *_stable_node_dup if the returned tree_page is NULL.
+ *
+ * It can also free and overwrite *_stable_node with the found
+ * stable_node_dup if the chain is collapsed (in which case
+ * *_stable_node will be equal to *_stable_node_dup like if the chain
+ * never existed). It's up to the caller to verify tree_page is not
+ * NULL before dereferencing *_stable_node or *_stable_node_dup.
+ *
+ * *_stable_node_dup is really a second output parameter of this
+ * function and will be overwritten in all cases, the caller doesn't
+ * need to initialize it.
+ */
+static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
+                                       struct stable_node **_stable_node,
+                                       struct rb_root *root,
+                                       bool prune_stale_stable_nodes)
 {
        struct stable_node *stable_node = *_stable_node;
        if (!is_stable_node_chain(stable_node)) {
                if (is_page_sharing_candidate(stable_node)) {
-                       *tree_page = get_ksm_page(stable_node, false);
-                       return stable_node;
+                       *_stable_node_dup = stable_node;
+                       return get_ksm_page(stable_node, false);
                }
+               /*
+                * _stable_node_dup set to NULL means the stable_node
+                * reached the ksm_max_page_sharing limit.
+                */
+               *_stable_node_dup = NULL;
                return NULL;
        }
-       return stable_node_dup(_stable_node, tree_page, root,
+       return stable_node_dup(_stable_node_dup, _stable_node, root,
                               prune_stale_stable_nodes);
 }
 
-static __always_inline struct stable_node *chain_prune(struct stable_node **s_n,
-                                                      struct page **t_p,
-                                                      struct rb_root *root)
+static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
+                                               struct stable_node **s_n,
+                                               struct rb_root *root)
 {
-       return __stable_node_chain(s_n, t_p, root, true);
+       return __stable_node_chain(s_n_d, s_n, root, true);
 }
 
-static __always_inline struct stable_node *chain(struct stable_node *s_n,
-                                                struct page **t_p,
-                                                struct rb_root *root)
+static __always_inline struct page *chain(struct stable_node **s_n_d,
+                                         struct stable_node *s_n,
+                                         struct rb_root *root)
 {
-       return __stable_node_chain(&s_n, t_p, root, false);
+       struct stable_node *old_stable_node = s_n;
+       struct page *tree_page;
+
+       tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
+       /* not pruning dups so s_n cannot have changed */
+       VM_BUG_ON(s_n != old_stable_node);
+       return tree_page;
 }
 
 /*
@@ -1498,14 +1539,18 @@ again:
                cond_resched();
                stable_node = rb_entry(*new, struct stable_node, node);
                stable_node_any = NULL;
-               stable_node_dup = chain_prune(&stable_node, &tree_page, root);
+               tree_page = chain_prune(&stable_node_dup, &stable_node, root);
                /*
                 * NOTE: stable_node may have been freed by
                 * chain_prune() if the returned stable_node_dup is
                 * not NULL. stable_node_dup may have been inserted in
                 * the rbtree instead as a regular stable_node (in
                 * order to collapse the stable_node chain if a single
-                * stable_node dup was found in it).
+                * stable_node dup was found in it). In such case the
+                * stable_node is overwritten by the calleee to point
+                * to the stable_node_dup that was collapsed in the
+                * stable rbtree and stable_node will be equal to
+                * stable_node_dup like if the chain never existed.
                 */
                if (!stable_node_dup) {
                        /*
@@ -1623,15 +1668,13 @@ out:
 replace:
        /*
         * If stable_node was a chain and chain_prune collapsed it,
-        * stable_node will be NULL here. In that case the
-        * stable_node_dup is the regular stable_node that has
-        * replaced the chain. If stable_node is not NULL and equal to
-        * stable_node_dup there was no chain and stable_node_dup is
-        * the regular stable_node in the stable rbtree. Otherwise
-        * stable_node is the chain and stable_node_dup is the dup to
-        * replace.
+        * stable_node has been updated to be the new regular
+        * stable_node. A collapse of the chain is indistinguishable
+        * from the case there was no chain in the stable
+        * rbtree. Otherwise stable_node is the chain and
+        * stable_node_dup is the dup to replace.
         */
-       if (!stable_node || stable_node_dup == stable_node) {
+       if (stable_node_dup == stable_node) {
                VM_BUG_ON(is_stable_node_chain(stable_node_dup));
                VM_BUG_ON(is_stable_node_dup(stable_node_dup));
                /* there is no chain */
@@ -1676,13 +1719,13 @@ chain_append:
                stable_node_dup = stable_node_any;
        /*
         * If stable_node was a chain and chain_prune collapsed it,
-        * stable_node will be NULL here. In that case the
-        * stable_node_dup is the regular stable_node that has
-        * replaced the chain. If stable_node is not NULL and equal to
-        * stable_node_dup there was no chain and stable_node_dup is
-        * the regular stable_node in the stable rbtree.
+        * stable_node has been updated to be the new regular
+        * stable_node. A collapse of the chain is indistinguishable
+        * from the case there was no chain in the stable
+        * rbtree. Otherwise stable_node is the chain and
+        * stable_node_dup is the dup to replace.
         */
-       if (!stable_node || stable_node_dup == stable_node) {
+       if (stable_node_dup == stable_node) {
                VM_BUG_ON(is_stable_node_chain(stable_node_dup));
                VM_BUG_ON(is_stable_node_dup(stable_node_dup));
                /* chain is missing so create it */
@@ -1737,7 +1780,7 @@ again:
                cond_resched();
                stable_node = rb_entry(*new, struct stable_node, node);
                stable_node_any = NULL;
-               stable_node_dup = chain(stable_node, &tree_page, root);
+               tree_page = chain(&stable_node_dup, stable_node, root);
                if (!stable_node_dup) {
                        /*
                         * Either all stable_node dups were full in