Merge branch 'for-linus' of git://git390.osdl.marist.edu/pub/scm/linux-2.6
[pandora-kernel.git] / lib / radix-tree.c
index 7097bb2..d69ddbe 100644 (file)
@@ -2,6 +2,7 @@
  * Copyright (C) 2001 Momchil Velikov
  * Portions Copyright (C) 2001 Christoph Hellwig
  * Copyright (C) 2005 SGI, Christoph Lameter <clameter@sgi.com>
+ * Copyright (C) 2006 Nick Piggin
  *
  * This program is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public License as
 #include <linux/gfp.h>
 #include <linux/string.h>
 #include <linux/bitops.h>
+#include <linux/rcupdate.h>
 
 
 #ifdef __KERNEL__
-#define RADIX_TREE_MAP_SHIFT   6
+#define RADIX_TREE_MAP_SHIFT   (CONFIG_BASE_SMALL ? 4 : 6)
 #else
 #define RADIX_TREE_MAP_SHIFT   3       /* For more stressful testing */
 #endif
@@ -45,7 +47,9 @@
        ((RADIX_TREE_MAP_SIZE + BITS_PER_LONG - 1) / BITS_PER_LONG)
 
 struct radix_tree_node {
+       unsigned int    height;         /* Height from the bottom */
        unsigned int    count;
+       struct rcu_head rcu_head;
        void            *slots[RADIX_TREE_MAP_SIZE];
        unsigned long   tags[RADIX_TREE_MAX_TAGS][RADIX_TREE_TAG_LONGS];
 };
@@ -63,7 +67,7 @@ static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH] __read_mostly;
 /*
  * Radix tree node cache.
  */
-static kmem_cache_t *radix_tree_node_cachep;
+static struct kmem_cache *radix_tree_node_cachep;
 
 /*
  * Per-cpu pool of preloaded nodes
@@ -74,6 +78,11 @@ struct radix_tree_preload {
 };
 DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
 
+static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
+{
+       return root->gfp_mask & __GFP_BITS_MASK;
+}
+
 /*
  * This assumes that the caller has performed appropriate preallocation, and
  * that the caller has pinned this thread of control to the current CPU.
@@ -82,9 +91,10 @@ static struct radix_tree_node *
 radix_tree_node_alloc(struct radix_tree_root *root)
 {
        struct radix_tree_node *ret;
+       gfp_t gfp_mask = root_gfp_mask(root);
 
-       ret = kmem_cache_alloc(radix_tree_node_cachep, root->gfp_mask);
-       if (ret == NULL && !(root->gfp_mask & __GFP_WAIT)) {
+       ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
+       if (ret == NULL && !(gfp_mask & __GFP_WAIT)) {
                struct radix_tree_preload *rtp;
 
                rtp = &__get_cpu_var(radix_tree_preloads);
@@ -94,13 +104,21 @@ radix_tree_node_alloc(struct radix_tree_root *root)
                        rtp->nr--;
                }
        }
+       BUG_ON(radix_tree_is_direct_ptr(ret));
        return ret;
 }
 
+static void radix_tree_node_rcu_free(struct rcu_head *head)
+{
+       struct radix_tree_node *node =
+                       container_of(head, struct radix_tree_node, rcu_head);
+       kmem_cache_free(radix_tree_node_cachep, node);
+}
+
 static inline void
 radix_tree_node_free(struct radix_tree_node *node)
 {
-       kmem_cache_free(radix_tree_node_cachep, node);
+       call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
 }
 
 /*
@@ -152,6 +170,27 @@ static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
        return test_bit(offset, node->tags[tag]);
 }
 
+static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
+{
+       root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
+}
+
+
+static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
+{
+       root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
+}
+
+static inline void root_tag_clear_all(struct radix_tree_root *root)
+{
+       root->gfp_mask &= __GFP_BITS_MASK;
+}
+
+static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
+{
+       return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
+}
+
 /*
  * Returns 1 if any slot in the node has this tag set.
  * Otherwise returns 0.
@@ -182,7 +221,6 @@ static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
 {
        struct radix_tree_node *node;
        unsigned int height;
-       char tags[RADIX_TREE_MAX_TAGS];
        int tag;
 
        /* Figure out what the height should be.  */
@@ -195,32 +233,25 @@ static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
                goto out;
        }
 
-       /*
-        * Prepare the tag status of the top-level node for propagation
-        * into the newly-pushed top-level node(s)
-        */
-       for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
-               tags[tag] = 0;
-               if (any_tag_set(root->rnode, tag))
-                       tags[tag] = 1;
-       }
-
        do {
+               unsigned int newheight;
                if (!(node = radix_tree_node_alloc(root)))
                        return -ENOMEM;
 
                /* Increase the height.  */
-               node->slots[0] = root->rnode;
+               node->slots[0] = radix_tree_direct_to_ptr(root->rnode);
 
                /* Propagate the aggregated tag info into the new root */
                for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
-                       if (tags[tag])
+                       if (root_tag_get(root, tag))
                                tag_set(node, tag, 0);
                }
 
+               newheight = root->height+1;
+               node->height = newheight;
                node->count = 1;
-               root->rnode = node;
-               root->height++;
+               rcu_assign_pointer(root->rnode, node);
+               root->height = newheight;
        } while (height > root->height);
 out:
        return 0;
@@ -242,9 +273,10 @@ int radix_tree_insert(struct radix_tree_root *root,
        int offset;
        int error;
 
+       BUG_ON(radix_tree_is_direct_ptr(item));
+
        /* Make sure the tree is high enough.  */
-       if ((!index && !root->rnode) ||
-                       index > radix_tree_maxindex(root->height)) {
+       if (index > radix_tree_maxindex(root->height)) {
                error = radix_tree_extend(root, index);
                if (error)
                        return error;
@@ -255,16 +287,17 @@ int radix_tree_insert(struct radix_tree_root *root,
        shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 
        offset = 0;                     /* uninitialised var warning */
-       do {
+       while (height > 0) {
                if (slot == NULL) {
                        /* Have to add a child node.  */
                        if (!(slot = radix_tree_node_alloc(root)))
                                return -ENOMEM;
+                       slot->height = height;
                        if (node) {
-                               node->slots[offset] = slot;
+                               rcu_assign_pointer(node->slots[offset], slot);
                                node->count++;
                        } else
-                               root->rnode = slot;
+                               rcu_assign_pointer(root->rnode, slot);
                }
 
                /* Go a level down */
@@ -273,60 +306,74 @@ int radix_tree_insert(struct radix_tree_root *root,
                slot = node->slots[offset];
                shift -= RADIX_TREE_MAP_SHIFT;
                height--;
-       } while (height > 0);
+       }
 
        if (slot != NULL)
                return -EEXIST;
 
-       BUG_ON(!node);
-       node->count++;
-       node->slots[offset] = item;
-       BUG_ON(tag_get(node, 0, offset));
-       BUG_ON(tag_get(node, 1, offset));
+       if (node) {
+               node->count++;
+               rcu_assign_pointer(node->slots[offset], item);
+               BUG_ON(tag_get(node, 0, offset));
+               BUG_ON(tag_get(node, 1, offset));
+       } else {
+               rcu_assign_pointer(root->rnode, radix_tree_ptr_to_direct(item));
+               BUG_ON(root_tag_get(root, 0));
+               BUG_ON(root_tag_get(root, 1));
+       }
 
        return 0;
 }
 EXPORT_SYMBOL(radix_tree_insert);
 
-static inline void **__lookup_slot(struct radix_tree_root *root,
-                                  unsigned long index)
+/**
+ *     radix_tree_lookup_slot    -    lookup a slot in a radix tree
+ *     @root:          radix tree root
+ *     @index:         index key
+ *
+ *     Returns:  the slot corresponding to the position @index in the
+ *     radix tree @root. This is useful for update-if-exists operations.
+ *
+ *     This function cannot be called under rcu_read_lock, it must be
+ *     excluded from writers, as must the returned slot for subsequent
+ *     use by radix_tree_deref_slot() and radix_tree_replace slot.
+ *     Caller must hold tree write locked across slot lookup and
+ *     replace.
+ */
+void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
 {
        unsigned int height, shift;
-       struct radix_tree_node **slot;
+       struct radix_tree_node *node, **slot;
 
-       height = root->height;
+       node = root->rnode;
+       if (node == NULL)
+               return NULL;
+
+       if (radix_tree_is_direct_ptr(node)) {
+               if (index > 0)
+                       return NULL;
+               return (void **)&root->rnode;
+       }
+
+       height = node->height;
        if (index > radix_tree_maxindex(height))
                return NULL;
 
        shift = (height-1) * RADIX_TREE_MAP_SHIFT;
-       slot = &root->rnode;
 
-       while (height > 0) {
-               if (*slot == NULL)
+       do {
+               slot = (struct radix_tree_node **)
+                       (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
+               node = *slot;
+               if (node == NULL)
                        return NULL;
 
-               slot = (struct radix_tree_node **)
-                       ((*slot)->slots +
-                               ((index >> shift) & RADIX_TREE_MAP_MASK));
                shift -= RADIX_TREE_MAP_SHIFT;
                height--;
-       }
+       } while (height > 0);
 
        return (void **)slot;
 }
-
-/**
- *     radix_tree_lookup_slot    -    lookup a slot in a radix tree
- *     @root:          radix tree root
- *     @index:         index key
- *
- *     Lookup the slot corresponding to the position @index in the radix tree
- *     @root. This is useful for update-if-exists operations.
- */
-void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
-{
-       return __lookup_slot(root, index);
-}
 EXPORT_SYMBOL(radix_tree_lookup_slot);
 
 /**
@@ -335,13 +382,45 @@ EXPORT_SYMBOL(radix_tree_lookup_slot);
  *     @index:         index key
  *
  *     Lookup the item at the position @index in the radix tree @root.
+ *
+ *     This function can be called under rcu_read_lock, however the caller
+ *     must manage lifetimes of leaf nodes (eg. RCU may also be used to free
+ *     them safely). No RCU barriers are required to access or modify the
+ *     returned item, however.
  */
 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
 {
-       void **slot;
+       unsigned int height, shift;
+       struct radix_tree_node *node, **slot;
+
+       node = rcu_dereference(root->rnode);
+       if (node == NULL)
+               return NULL;
 
-       slot = __lookup_slot(root, index);
-       return slot != NULL ? *slot : NULL;
+       if (radix_tree_is_direct_ptr(node)) {
+               if (index > 0)
+                       return NULL;
+               return radix_tree_direct_to_ptr(node);
+       }
+
+       height = node->height;
+       if (index > radix_tree_maxindex(height))
+               return NULL;
+
+       shift = (height-1) * RADIX_TREE_MAP_SHIFT;
+
+       do {
+               slot = (struct radix_tree_node **)
+                       (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
+               node = rcu_dereference(*slot);
+               if (node == NULL)
+                       return NULL;
+
+               shift -= RADIX_TREE_MAP_SHIFT;
+               height--;
+       } while (height > 0);
+
+       return node;
 }
 EXPORT_SYMBOL(radix_tree_lookup);
 
@@ -365,11 +444,10 @@ void *radix_tree_tag_set(struct radix_tree_root *root,
        struct radix_tree_node *slot;
 
        height = root->height;
-       if (index > radix_tree_maxindex(height))
-               return NULL;
+       BUG_ON(index > radix_tree_maxindex(height));
 
-       shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
        slot = root->rnode;
+       shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
 
        while (height > 0) {
                int offset;
@@ -383,6 +461,10 @@ void *radix_tree_tag_set(struct radix_tree_root *root,
                height--;
        }
 
+       /* set the root's tag bit */
+       if (slot && !root_tag_get(root, tag))
+               root_tag_set(root, tag);
+
        return slot;
 }
 EXPORT_SYMBOL(radix_tree_tag_set);
@@ -405,9 +487,8 @@ void *radix_tree_tag_clear(struct radix_tree_root *root,
                        unsigned long index, unsigned int tag)
 {
        struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
-       struct radix_tree_node *slot;
+       struct radix_tree_node *slot = NULL;
        unsigned int height, shift;
-       void *ret = NULL;
 
        height = root->height;
        if (index > radix_tree_maxindex(height))
@@ -432,20 +513,24 @@ void *radix_tree_tag_clear(struct radix_tree_root *root,
                height--;
        }
 
-       ret = slot;
-       if (ret == NULL)
+       if (slot == NULL)
                goto out;
 
-       do {
+       while (pathp->node) {
                if (!tag_get(pathp->node, tag, pathp->offset))
                        goto out;
                tag_clear(pathp->node, tag, pathp->offset);
                if (any_tag_set(pathp->node, tag))
                        goto out;
                pathp--;
-       } while (pathp->node);
+       }
+
+       /* clear the root's tag bit */
+       if (root_tag_get(root, tag))
+               root_tag_clear(root, tag);
+
 out:
-       return ret;
+       return slot;
 }
 EXPORT_SYMBOL(radix_tree_tag_clear);
 
@@ -458,28 +543,37 @@ EXPORT_SYMBOL(radix_tree_tag_clear);
  *
  * Return values:
  *
- *  0: tag not present
- *  1: tag present, set
- * -1: tag present, unset
+ *  0: tag not present or not set
+ *  1: tag set
  */
 int radix_tree_tag_get(struct radix_tree_root *root,
                        unsigned long index, unsigned int tag)
 {
        unsigned int height, shift;
-       struct radix_tree_node *slot;
+       struct radix_tree_node *node;
        int saw_unset_tag = 0;
 
-       height = root->height;
+       /* check the root's tag bit */
+       if (!root_tag_get(root, tag))
+               return 0;
+
+       node = rcu_dereference(root->rnode);
+       if (node == NULL)
+               return 0;
+
+       if (radix_tree_is_direct_ptr(node))
+               return (index == 0);
+
+       height = node->height;
        if (index > radix_tree_maxindex(height))
                return 0;
 
        shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
-       slot = root->rnode;
 
        for ( ; ; ) {
                int offset;
 
-               if (slot == NULL)
+               if (node == NULL)
                        return 0;
 
                offset = (index >> shift) & RADIX_TREE_MAP_MASK;
@@ -488,15 +582,15 @@ int radix_tree_tag_get(struct radix_tree_root *root,
                 * This is just a debug check.  Later, we can bale as soon as
                 * we see an unset tag.
                 */
-               if (!tag_get(slot, tag, offset))
+               if (!tag_get(node, tag, offset))
                        saw_unset_tag = 1;
                if (height == 1) {
-                       int ret = tag_get(slot, tag, offset);
+                       int ret = tag_get(node, tag, offset);
 
                        BUG_ON(ret && saw_unset_tag);
-                       return ret ? 1 : -1;
+                       return !!ret;
                }
-               slot = slot->slots[offset];
+               node = rcu_dereference(node->slots[offset]);
                shift -= RADIX_TREE_MAP_SHIFT;
                height--;
        }
@@ -505,44 +599,45 @@ EXPORT_SYMBOL(radix_tree_tag_get);
 #endif
 
 static unsigned int
-__lookup(struct radix_tree_root *root, void **results, unsigned long index,
+__lookup(struct radix_tree_node *slot, void **results, unsigned long index,
        unsigned int max_items, unsigned long *next_index)
 {
        unsigned int nr_found = 0;
        unsigned int shift, height;
-       struct radix_tree_node *slot;
        unsigned long i;
 
-       height = root->height;
+       height = slot->height;
        if (height == 0)
                goto out;
-
        shift = (height-1) * RADIX_TREE_MAP_SHIFT;
-       slot = root->rnode;
 
        for ( ; height > 1; height--) {
-
-               for (i = (index >> shift) & RADIX_TREE_MAP_MASK ;
-                               i < RADIX_TREE_MAP_SIZE; i++) {
+               i = (index >> shift) & RADIX_TREE_MAP_MASK;
+               for (;;) {
                        if (slot->slots[i] != NULL)
                                break;
                        index &= ~((1UL << shift) - 1);
                        index += 1UL << shift;
                        if (index == 0)
                                goto out;       /* 32-bit wraparound */
+                       i++;
+                       if (i == RADIX_TREE_MAP_SIZE)
+                               goto out;
                }
-               if (i == RADIX_TREE_MAP_SIZE)
-                       goto out;
 
                shift -= RADIX_TREE_MAP_SHIFT;
-               slot = slot->slots[i];
+               slot = rcu_dereference(slot->slots[i]);
+               if (slot == NULL)
+                       goto out;
        }
 
        /* Bottom level: grab some items */
        for (i = index & RADIX_TREE_MAP_MASK; i < RADIX_TREE_MAP_SIZE; i++) {
+               struct radix_tree_node *node;
                index++;
-               if (slot->slots[i]) {
-                       results[nr_found++] = slot->slots[i];
+               node = slot->slots[i];
+               if (node) {
+                       results[nr_found++] = rcu_dereference(node);
                        if (nr_found == max_items)
                                goto out;
                }
@@ -564,28 +659,51 @@ out:
  *     *@results.
  *
  *     The implementation is naive.
+ *
+ *     Like radix_tree_lookup, radix_tree_gang_lookup may be called under
+ *     rcu_read_lock. In this case, rather than the returned results being
+ *     an atomic snapshot of the tree at a single point in time, the semantics
+ *     of an RCU protected gang lookup are as though multiple radix_tree_lookups
+ *     have been issued in individual locks, and results stored in 'results'.
  */
 unsigned int
 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
                        unsigned long first_index, unsigned int max_items)
 {
-       const unsigned long max_index = radix_tree_maxindex(root->height);
+       unsigned long max_index;
+       struct radix_tree_node *node;
        unsigned long cur_index = first_index;
-       unsigned int ret = 0;
+       unsigned int ret;
+
+       node = rcu_dereference(root->rnode);
+       if (!node)
+               return 0;
+
+       if (radix_tree_is_direct_ptr(node)) {
+               if (first_index > 0)
+                       return 0;
+               node = radix_tree_direct_to_ptr(node);
+               results[0] = rcu_dereference(node);
+               return 1;
+       }
+
+       max_index = radix_tree_maxindex(node->height);
 
+       ret = 0;
        while (ret < max_items) {
                unsigned int nr_found;
                unsigned long next_index;       /* Index of next search */
 
                if (cur_index > max_index)
                        break;
-               nr_found = __lookup(root, results + ret, cur_index,
+               nr_found = __lookup(node, results + ret, cur_index,
                                        max_items - ret, &next_index);
                ret += nr_found;
                if (next_index == 0)
                        break;
                cur_index = next_index;
        }
+
        return ret;
 }
 EXPORT_SYMBOL(radix_tree_gang_lookup);
@@ -595,48 +713,63 @@ EXPORT_SYMBOL(radix_tree_gang_lookup);
  * open-coding the search.
  */
 static unsigned int
-__lookup_tag(struct radix_tree_root *root, void **results, unsigned long index,
+__lookup_tag(struct radix_tree_node *slot, void **results, unsigned long index,
        unsigned int max_items, unsigned long *next_index, unsigned int tag)
 {
        unsigned int nr_found = 0;
-       unsigned int shift;
-       unsigned int height = root->height;
-       struct radix_tree_node *slot;
+       unsigned int shift, height;
 
-       shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
-       slot = root->rnode;
+       height = slot->height;
+       if (height == 0)
+               goto out;
+       shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 
        while (height > 0) {
-               unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK;
+               unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
 
-               for ( ; i < RADIX_TREE_MAP_SIZE; i++) {
-                       if (tag_get(slot, tag, i)) {
-                               BUG_ON(slot->slots[i] == NULL);
+               for (;;) {
+                       if (tag_get(slot, tag, i))
                                break;
-                       }
                        index &= ~((1UL << shift) - 1);
                        index += 1UL << shift;
                        if (index == 0)
                                goto out;       /* 32-bit wraparound */
+                       i++;
+                       if (i == RADIX_TREE_MAP_SIZE)
+                               goto out;
                }
-               if (i == RADIX_TREE_MAP_SIZE)
-                       goto out;
                height--;
                if (height == 0) {      /* Bottom level: grab some items */
                        unsigned long j = index & RADIX_TREE_MAP_MASK;
 
                        for ( ; j < RADIX_TREE_MAP_SIZE; j++) {
+                               struct radix_tree_node *node;
                                index++;
-                               if (tag_get(slot, tag, j)) {
-                                       BUG_ON(slot->slots[j] == NULL);
-                                       results[nr_found++] = slot->slots[j];
+                               if (!tag_get(slot, tag, j))
+                                       continue;
+                               node = slot->slots[j];
+                               /*
+                                * Even though the tag was found set, we need to
+                                * recheck that we have a non-NULL node, because
+                                * if this lookup is lockless, it may have been
+                                * subsequently deleted.
+                                *
+                                * Similar care must be taken in any place that
+                                * lookup ->slots[x] without a lock (ie. can't
+                                * rely on its value remaining the same).
+                                */
+                               if (node) {
+                                       node = rcu_dereference(node);
+                                       results[nr_found++] = node;
                                        if (nr_found == max_items)
                                                goto out;
                                }
                        }
                }
                shift -= RADIX_TREE_MAP_SHIFT;
-               slot = slot->slots[i];
+               slot = rcu_dereference(slot->slots[i]);
+               if (slot == NULL)
+                       break;
        }
 out:
        *next_index = index;
@@ -661,23 +794,44 @@ radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
                unsigned long first_index, unsigned int max_items,
                unsigned int tag)
 {
-       const unsigned long max_index = radix_tree_maxindex(root->height);
+       struct radix_tree_node *node;
+       unsigned long max_index;
        unsigned long cur_index = first_index;
-       unsigned int ret = 0;
+       unsigned int ret;
+
+       /* check the root's tag bit */
+       if (!root_tag_get(root, tag))
+               return 0;
+
+       node = rcu_dereference(root->rnode);
+       if (!node)
+               return 0;
+
+       if (radix_tree_is_direct_ptr(node)) {
+               if (first_index > 0)
+                       return 0;
+               node = radix_tree_direct_to_ptr(node);
+               results[0] = rcu_dereference(node);
+               return 1;
+       }
+
+       max_index = radix_tree_maxindex(node->height);
 
+       ret = 0;
        while (ret < max_items) {
                unsigned int nr_found;
                unsigned long next_index;       /* Index of next search */
 
                if (cur_index > max_index)
                        break;
-               nr_found = __lookup_tag(root, results + ret, cur_index,
+               nr_found = __lookup_tag(node, results + ret, cur_index,
                                        max_items - ret, &next_index, tag);
                ret += nr_found;
                if (next_index == 0)
                        break;
                cur_index = next_index;
        }
+
        return ret;
 }
 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
@@ -689,12 +843,23 @@ EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
 static inline void radix_tree_shrink(struct radix_tree_root *root)
 {
        /* try to shrink tree height */
-       while (root->height > 1 &&
+       while (root->height > 0 &&
                        root->rnode->count == 1 &&
                        root->rnode->slots[0]) {
                struct radix_tree_node *to_free = root->rnode;
+               void *newptr;
 
-               root->rnode = to_free->slots[0];
+               /*
+                * We don't need rcu_assign_pointer(), since we are simply
+                * moving the node from one part of the tree to another. If
+                * it was safe to dereference the old pointer to it
+                * (to_free->slots[0]), it will be safe to dereference the new
+                * one (root->rnode).
+                */
+               newptr = to_free->slots[0];
+               if (root->height == 1)
+                       newptr = radix_tree_ptr_to_direct(newptr);
+               root->rnode = newptr;
                root->height--;
                /* must only free zeroed nodes into the slab */
                tag_clear(to_free, 0, 0);
@@ -717,12 +882,9 @@ static inline void radix_tree_shrink(struct radix_tree_root *root)
 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
 {
        struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
-       struct radix_tree_path *orig_pathp;
-       struct radix_tree_node *slot;
+       struct radix_tree_node *slot = NULL;
+       struct radix_tree_node *to_free;
        unsigned int height, shift;
-       void *ret = NULL;
-       char tags[RADIX_TREE_MAX_TAGS];
-       int nr_cleared_tags;
        int tag;
        int offset;
 
@@ -730,11 +892,18 @@ void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
        if (index > radix_tree_maxindex(height))
                goto out;
 
+       slot = root->rnode;
+       if (height == 0 && root->rnode) {
+               slot = radix_tree_direct_to_ptr(slot);
+               root_tag_clear_all(root);
+               root->rnode = NULL;
+               goto out;
+       }
+
        shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
        pathp->node = NULL;
-       slot = root->rnode;
 
-       for ( ; height > 0; height--) {
+       do {
                if (slot == NULL)
                        goto out;
 
@@ -744,46 +913,31 @@ void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
                pathp->node = slot;
                slot = slot->slots[offset];
                shift -= RADIX_TREE_MAP_SHIFT;
-       }
+               height--;
+       } while (height > 0);
 
-       ret = slot;
-       if (ret == NULL)
+       if (slot == NULL)
                goto out;
 
-       orig_pathp = pathp;
-
        /*
         * Clear all tags associated with the just-deleted item
         */
-       nr_cleared_tags = 0;
        for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
-               tags[tag] = 1;
-               if (tag_get(pathp->node, tag, pathp->offset)) {
-                       tag_clear(pathp->node, tag, pathp->offset);
-                       if (!any_tag_set(pathp->node, tag)) {
-                               tags[tag] = 0;
-                               nr_cleared_tags++;
-                       }
-               }
-       }
-
-       for (pathp--; nr_cleared_tags && pathp->node; pathp--) {
-               for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
-                       if (tags[tag])
-                               continue;
-
-                       tag_clear(pathp->node, tag, pathp->offset);
-                       if (any_tag_set(pathp->node, tag)) {
-                               tags[tag] = 1;
-                               nr_cleared_tags--;
-                       }
-               }
+               if (tag_get(pathp->node, tag, pathp->offset))
+                       radix_tree_tag_clear(root, index, tag);
        }
 
+       to_free = NULL;
        /* Now free the nodes we do not need anymore */
-       for (pathp = orig_pathp; pathp->node; pathp--) {
+       while (pathp->node) {
                pathp->node->slots[pathp->offset] = NULL;
                pathp->node->count--;
+               /*
+                * Queue the node for deferred freeing after the
+                * last reference to it disappears (set NULL, above).
+                */
+               if (to_free)
+                       radix_tree_node_free(to_free);
 
                if (pathp->node->count) {
                        if (pathp->node == root->rnode)
@@ -792,12 +946,18 @@ void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
                }
 
                /* Node with zero slots in use so free it */
-               radix_tree_node_free(pathp->node);
+               to_free = pathp->node;
+               pathp--;
+
        }
-       root->rnode = NULL;
+       root_tag_clear_all(root);
        root->height = 0;
+       root->rnode = NULL;
+       if (to_free)
+               radix_tree_node_free(to_free);
+
 out:
-       return ret;
+       return slot;
 }
 EXPORT_SYMBOL(radix_tree_delete);
 
@@ -808,16 +968,12 @@ EXPORT_SYMBOL(radix_tree_delete);
  */
 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
 {
-       struct radix_tree_node *rnode;
-       rnode = root->rnode;
-       if (!rnode)
-               return 0;
-       return any_tag_set(rnode, tag);
+       return root_tag_get(root, tag);
 }
 EXPORT_SYMBOL(radix_tree_tagged);
 
 static void
-radix_tree_node_ctor(void *node, kmem_cache_t *cachep, unsigned long flags)
+radix_tree_node_ctor(void *node, struct kmem_cache *cachep, unsigned long flags)
 {
        memset(node, 0, sizeof(struct radix_tree_node));
 }
@@ -840,7 +996,6 @@ static __init void radix_tree_init_maxindex(void)
                height_to_maxindex[i] = __maxindex(i);
 }
 
-#ifdef CONFIG_HOTPLUG_CPU
 static int radix_tree_callback(struct notifier_block *nfb,
                             unsigned long action,
                             void *hcpu)
@@ -860,7 +1015,6 @@ static int radix_tree_callback(struct notifier_block *nfb,
        }
        return NOTIFY_OK;
 }
-#endif /* CONFIG_HOTPLUG_CPU */
 
 void __init radix_tree_init(void)
 {