vserver 1.9.5.x5
[linux-2.6.git] / fs / dcache.c
index d4fa197..ed90b72 100644 (file)
  */
 
 #include <linux/config.h>
+#include <linux/syscalls.h>
 #include <linux/string.h>
 #include <linux/mm.h>
 #include <linux/fs.h>
 #include <linux/slab.h>
 #include <linux/init.h>
 #include <linux/smp_lock.h>
+#include <linux/hash.h>
 #include <linux/cache.h>
 #include <linux/module.h>
 #include <linux/mount.h>
 #include <linux/security.h>
 #include <linux/seqlock.h>
 #include <linux/swap.h>
+#include <linux/bootmem.h>
 
-#define DCACHE_PARANOIA 1
 /* #define DCACHE_DEBUG 1 */
 
-spinlock_t dcache_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
+int sysctl_vfs_cache_pressure = 100;
+
+ __cacheline_aligned_in_smp DEFINE_SPINLOCK(dcache_lock);
 seqlock_t rename_lock __cacheline_aligned_in_smp = SEQLOCK_UNLOCKED;
 
 EXPORT_SYMBOL(dcache_lock);
 
 static kmem_cache_t *dentry_cache; 
 
+#define DNAME_INLINE_LEN (sizeof(struct dentry)-offsetof(struct dentry,d_iname))
+
 /*
  * This is the single most critical data structure when it comes
  * to the dcache: the hashtable for lookups. Somebody should try
@@ -61,13 +67,12 @@ struct dentry_stat_t dentry_stat = {
        .age_limit = 45,
 };
 
-static void d_callback(void *arg)
+static void d_callback(struct rcu_head *head)
 {
-       struct dentry * dentry = (struct dentry *)arg;
+       struct dentry * dentry = container_of(head, struct dentry, d_rcu);
 
-       if (dname_external(dentry)) {
-               kfree(dentry->d_qstr);
-       }
+       if (dname_external(dentry))
+               kfree(dentry->d_name.name);
        kmem_cache_free(dentry_cache, dentry); 
 }
 
@@ -79,7 +84,7 @@ static void d_free(struct dentry *dentry)
 {
        if (dentry->d_op && dentry->d_op->d_release)
                dentry->d_op->d_release(dentry);
-       call_rcu(&dentry->d_rcu, d_callback, dentry);
+       call_rcu(&dentry->d_rcu, d_callback);
 }
 
 /*
@@ -140,6 +145,8 @@ void dput(struct dentry *dentry)
                return;
 
 repeat:
+       if (atomic_read(&dentry->d_count) == 1)
+               might_sleep();
        if (!atomic_dec_and_lock(&dentry->d_count, &dcache_lock))
                return;
 
@@ -149,7 +156,7 @@ repeat:
                spin_unlock(&dcache_lock);
                return;
        }
-                       
+
        /*
         * AV: ->d_delete() is _NOT_ allowed to block now.
         */
@@ -161,7 +168,7 @@ repeat:
        if (d_unhashed(dentry))
                goto kill_it;
        if (list_empty(&dentry->d_lru)) {
-               dentry->d_vfs_flags |= DCACHE_REFERENCED;
+               dentry->d_flags |= DCACHE_REFERENCED;
                list_add(&dentry->d_lru, &dentry_unused);
                dentry_stat.nr_unused++;
        }
@@ -257,7 +264,7 @@ int d_invalidate(struct dentry * dentry)
 static inline struct dentry * __dget_locked(struct dentry *dentry)
 {
        atomic_inc(&dentry->d_count);
-       if (atomic_read(&dentry->d_count) == 1) {
+       if (!list_empty(&dentry->d_lru)) {
                dentry_stat.nr_unused--;
                list_del_init(&dentry->d_lru);
        }
@@ -272,22 +279,25 @@ struct dentry * dget_locked(struct dentry *dentry)
 /**
  * d_find_alias - grab a hashed alias of inode
  * @inode: inode in question
+ * @want_discon:  flag, used by d_splice_alias, to request
+ *          that only a DISCONNECTED alias be returned.
  *
- * If inode has a hashed alias - acquire the reference to alias and
- * return it. Otherwise return NULL. Notice that if inode is a directory
- * there can be only one alias and it can be unhashed only if it has
- * no children.
+ * If inode has a hashed alias, or is a directory and has any alias,
+ * acquire the reference to alias and return it. Otherwise return NULL.
+ * Notice that if inode is a directory there can be only one alias and
+ * it can be unhashed only if it has no children, or if it is the root
+ * of a filesystem.
  *
  * If the inode has a DCACHE_DISCONNECTED alias, then prefer
- * any other hashed alias over that one.
+ * any other hashed alias over that one unless @want_discon is set,
+ * in which case only return a DCACHE_DISCONNECTED alias.
  */
 
-struct dentry * d_find_alias(struct inode *inode)
+static struct dentry * __d_find_alias(struct inode *inode, int want_discon)
 {
        struct list_head *head, *next, *tmp;
        struct dentry *alias, *discon_alias=NULL;
 
-       spin_lock(&dcache_lock);
        head = &inode->i_dentry;
        next = inode->i_dentry.next;
        while (next != head) {
@@ -295,22 +305,29 @@ struct dentry * d_find_alias(struct inode *inode)
                next = tmp->next;
                prefetch(next);
                alias = list_entry(tmp, struct dentry, d_alias);
-               if (!d_unhashed(alias)) {
+               if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
                        if (alias->d_flags & DCACHE_DISCONNECTED)
                                discon_alias = alias;
-                       else {
+                       else if (!want_discon) {
                                __dget_locked(alias);
-                               spin_unlock(&dcache_lock);
                                return alias;
                        }
                }
        }
        if (discon_alias)
                __dget_locked(discon_alias);
-       spin_unlock(&dcache_lock);
        return discon_alias;
 }
 
+struct dentry * d_find_alias(struct inode *inode)
+{
+       struct dentry *de;
+       spin_lock(&dcache_lock);
+       de = __d_find_alias(inode, 0);
+       spin_unlock(&dcache_lock);
+       return de;
+}
+
 /*
  *     Try to kill dentries associated with this inode.
  * WARNING: you must own a reference to inode.
@@ -375,6 +392,8 @@ static void prune_dcache(int count)
                struct dentry *dentry;
                struct list_head *tmp;
 
+               cond_resched_lock(&dcache_lock);
+
                tmp = dentry_unused.prev;
                if (tmp == &dentry_unused)
                        break;
@@ -394,8 +413,8 @@ static void prune_dcache(int count)
                        continue;
                }
                /* If the dentry was recently referenced, don't free it. */
-               if (dentry->d_vfs_flags & DCACHE_REFERENCED) {
-                       dentry->d_vfs_flags &= ~DCACHE_REFERENCED;
+               if (dentry->d_flags & DCACHE_REFERENCED) {
+                       dentry->d_flags &= ~DCACHE_REFERENCED;
                        list_add(&dentry->d_lru, &dentry_unused);
                        dentry_stat.nr_unused++;
                        spin_unlock(&dentry->d_lock);
@@ -531,6 +550,13 @@ positive:
  * list for prune_dcache(). We descend to the next level
  * whenever the d_subdirs list is non-empty and continue
  * searching.
+ *
+ * It returns zero iff there are no unused children,
+ * otherwise  it returns the number of children moved to
+ * the end of the unused list. This may not be the total
+ * number of unused children, because select_parent can
+ * drop the lock and return early due to latency
+ * constraints.
  */
 static int select_parent(struct dentry * parent)
 {
@@ -560,6 +586,15 @@ resume:
                        dentry_stat.nr_unused++;
                        found++;
                }
+
+               /*
+                * We can return to the caller if we have found some (this
+                * ensures forward progress). We'll be coming back to find
+                * the rest.
+                */
+               if (found && need_resched())
+                       goto out;
+
                /*
                 * Descend a level if the d_subdirs list is non-empty.
                 */
@@ -584,6 +619,7 @@ this_parent->d_parent->d_name.name, this_parent->d_name.name, found);
 #endif
                goto resume;
        }
+out:
        spin_unlock(&dcache_lock);
        return found;
 }
@@ -609,7 +645,7 @@ void shrink_dcache_parent(struct dentry * parent)
  *
  * Prune the dentries that are anonymous
  *
- * parsing d_hash list does not read_barrier_depends() as it
+ * parsing d_hash list does not hlist_for_each_rcu() as it
  * done under dcache_lock.
  *
  */
@@ -624,7 +660,7 @@ void shrink_dcache_anon(struct hlist_head *head)
                        struct dentry *this = hlist_entry(lp, struct dentry, d_hash);
                        if (!list_empty(&this->d_lru)) {
                                dentry_stat.nr_unused--;
-                               list_del(&this->d_lru);
+                               list_del_init(&this->d_lru);
                        }
 
                        /* 
@@ -643,30 +679,27 @@ void shrink_dcache_anon(struct hlist_head *head)
 }
 
 /*
- * This is called from kswapd when we think we need some more memory.
+ * Scan `nr' dentries and return the number which remain.
+ *
+ * We need to avoid reentering the filesystem if the caller is performing a
+ * GFP_NOFS allocation attempt.  One example deadlock is:
+ *
+ * ext2_new_block->getblk->GFP->shrink_dcache_memory->prune_dcache->
+ * prune_one_dentry->dput->dentry_iput->iput->inode->i_sb->s_op->put_inode->
+ * ext2_discard_prealloc->ext2_free_blocks->lock_super->DEADLOCK.
+ *
+ * In this case we return -1 to tell the caller that we baled.
  */
 static int shrink_dcache_memory(int nr, unsigned int gfp_mask)
 {
        if (nr) {
-               /*
-                * Nasty deadlock avoidance.
-                *
-                * ext2_new_block->getblk->GFP->shrink_dcache_memory->
-                * prune_dcache->prune_one_dentry->dput->dentry_iput->iput->
-                * inode->i_sb->s_op->put_inode->ext2_discard_prealloc->
-                * ext2_free_blocks->lock_super->DEADLOCK.
-                *
-                * We should make sure we don't hold the superblock lock over
-                * block allocations, but for now:
-                */
-               if (gfp_mask & __GFP_FS)
-                       prune_dcache(nr);
+               if (!(gfp_mask & __GFP_FS))
+                       return -1;
+               prune_dcache(nr);
        }
-       return dentry_stat.nr_unused;
+       return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
 }
 
-#define NAME_ALLOC_LEN(len)    ((len+16) & ~15)
-
 /**
  * d_alloc     -       allocate a dcache entry
  * @parent: parent of entry to allocate
@@ -677,52 +710,41 @@ static int shrink_dcache_memory(int nr, unsigned int gfp_mask)
  * copied and the copy passed in may be reused after this call.
  */
  
-struct dentry * d_alloc(struct dentry * parent, const struct qstr *name)
+struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
 {
-       char * str;
        struct dentry *dentry;
-       struct qstr * qstr;
+       char *dname;
 
        dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL); 
        if (!dentry)
                return NULL;
 
        if (name->len > DNAME_INLINE_LEN-1) {
-               qstr = kmalloc(sizeof(*qstr) + NAME_ALLOC_LEN(name->len), 
-                               GFP_KERNEL);  
-               if (!qstr) {
+               dname = kmalloc(name->len + 1, GFP_KERNEL);
+               if (!dname) {
                        kmem_cache_free(dentry_cache, dentry); 
                        return NULL;
                }
-               qstr->name = qstr->name_str;
-               qstr->len = name->len;
-               qstr->hash = name->hash;
-               dentry->d_qstr = qstr;
-               str = qstr->name_str;
        } else  {
-               dentry->d_qstr = &dentry->d_name;
-               str = dentry->d_iname;
+               dname = dentry->d_iname;
        }       
+       dentry->d_name.name = dname;
 
-       memcpy(str, name->name, name->len);
-       str[name->len] = 0;
+       dentry->d_name.len = name->len;
+       dentry->d_name.hash = name->hash;
+       memcpy(dname, name->name, name->len);
+       dname[name->len] = 0;
 
        atomic_set(&dentry->d_count, 1);
-       dentry->d_vfs_flags = DCACHE_UNHASHED;
-       dentry->d_lock = SPIN_LOCK_UNLOCKED;
-       dentry->d_flags = 0;
+       dentry->d_flags = DCACHE_UNHASHED;
+       spin_lock_init(&dentry->d_lock);
        dentry->d_inode = NULL;
        dentry->d_parent = NULL;
-       dentry->d_move_count = 0;
        dentry->d_sb = NULL;
-       dentry->d_name.name = str;
-       dentry->d_name.len = name->len;
-       dentry->d_name.hash = name->hash;
        dentry->d_op = NULL;
        dentry->d_fsdata = NULL;
        dentry->d_mounted = 0;
        dentry->d_cookie = NULL;
-       dentry->d_bucket = NULL;
        INIT_HLIST_NODE(&dentry->d_hash);
        INIT_LIST_HEAD(&dentry->d_lru);
        INIT_LIST_HEAD(&dentry->d_subdirs);
@@ -744,6 +766,16 @@ struct dentry * d_alloc(struct dentry * parent, const struct qstr *name)
        return dentry;
 }
 
+struct dentry *d_alloc_name(struct dentry *parent, const char *name)
+{
+       struct qstr q;
+
+       q.name = name;
+       q.len = strlen(name);
+       q.hash = full_name_hash(q.name, q.len);
+       return d_alloc(parent, &q);
+}
+
 /**
  * d_instantiate - fill in inode information for a dentry
  * @entry: dentry to complete
@@ -770,6 +802,54 @@ void d_instantiate(struct dentry *entry, struct inode * inode)
        security_d_instantiate(entry, inode);
 }
 
+/**
+ * d_instantiate_unique - instantiate a non-aliased dentry
+ * @entry: dentry to instantiate
+ * @inode: inode to attach to this dentry
+ *
+ * Fill in inode information in the entry. On success, it returns NULL.
+ * If an unhashed alias of "entry" already exists, then we return the
+ * aliased dentry instead.
+ *
+ * Note that in order to avoid conflicts with rename() etc, the caller
+ * had better be holding the parent directory semaphore.
+ */
+struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode)
+{
+       struct dentry *alias;
+       int len = entry->d_name.len;
+       const char *name = entry->d_name.name;
+       unsigned int hash = entry->d_name.hash;
+
+       BUG_ON(!list_empty(&entry->d_alias));
+       spin_lock(&dcache_lock);
+       if (!inode)
+               goto do_negative;
+       list_for_each_entry(alias, &inode->i_dentry, d_alias) {
+               struct qstr *qstr = &alias->d_name;
+
+               if (qstr->hash != hash)
+                       continue;
+               if (alias->d_parent != entry->d_parent)
+                       continue;
+               if (qstr->len != len)
+                       continue;
+               if (memcmp(qstr->name, name, len))
+                       continue;
+               dget_locked(alias);
+               spin_unlock(&dcache_lock);
+               BUG_ON(!d_unhashed(alias));
+               return alias;
+       }
+       list_add(&entry->d_alias, &inode->i_dentry);
+do_negative:
+       entry->d_inode = inode;
+       spin_unlock(&dcache_lock);
+       security_d_instantiate(entry, inode);
+       return NULL;
+}
+EXPORT_SYMBOL(d_instantiate_unique);
+
 /**
  * d_alloc_root - allocate root dentry
  * @root_inode: inode to allocate the root for
@@ -784,7 +864,8 @@ struct dentry * d_alloc_root(struct inode * root_inode)
        struct dentry *res = NULL;
 
        if (root_inode) {
-               static const struct qstr name = { .name = "/", .len = 1, .hash = 0 };
+               static const struct qstr name = { .name = "/", .len = 1 };
+
                res = d_alloc(NULL, &name);
                if (res) {
                        res->d_sb = root_inode->i_sb;
@@ -795,10 +876,11 @@ struct dentry * d_alloc_root(struct inode * root_inode)
        return res;
 }
 
-static inline struct hlist_head * d_hash(struct dentry * parent, unsigned long hash)
+static inline struct hlist_head *d_hash(struct dentry *parent,
+                                       unsigned long hash)
 {
-       hash += (unsigned long) parent / L1_CACHE_BYTES;
-       hash = hash ^ (hash >> D_HASHBITS);
+       hash += ((unsigned long) parent ^ GOLDEN_RATIO_PRIME) / L1_CACHE_BYTES;
+       hash = hash ^ ((hash ^ GOLDEN_RATIO_PRIME) >> D_HASHBITS);
        return dentry_hashtable + (hash & D_HASHMASK);
 }
 
@@ -824,7 +906,7 @@ static inline struct hlist_head * d_hash(struct dentry * parent, unsigned long h
 
 struct dentry * d_alloc_anon(struct inode *inode)
 {
-       static const struct qstr anonstring = { "", 0, 0};
+       static const struct qstr anonstring = { .name = "" };
        struct dentry *tmp;
        struct dentry *res;
 
@@ -840,28 +922,21 @@ struct dentry * d_alloc_anon(struct inode *inode)
        tmp->d_parent = tmp; /* make sure dput doesn't croak */
        
        spin_lock(&dcache_lock);
-       if (S_ISDIR(inode->i_mode) && !list_empty(&inode->i_dentry)) {
-               /* A directory can only have one dentry.
-                * This (now) has one, so use it.
-                */
-               res = list_entry(inode->i_dentry.next, struct dentry, d_alias);
-               __dget_locked(res);
-       } else {
+       res = __d_find_alias(inode, 0);
+       if (!res) {
                /* attach a disconnected dentry */
                res = tmp;
                tmp = NULL;
-               if (res) {
-                       spin_lock(&res->d_lock);
-                       res->d_sb = inode->i_sb;
-                       res->d_parent = res;
-                       res->d_inode = inode;
-                       res->d_bucket = d_hash(res, res->d_name.hash);
-                       res->d_flags |= DCACHE_DISCONNECTED;
-                       res->d_vfs_flags &= ~DCACHE_UNHASHED;
-                       list_add(&res->d_alias, &inode->i_dentry);
-                       hlist_add_head(&res->d_hash, &inode->i_sb->s_anon);
-                       spin_unlock(&res->d_lock);
-               }
+               spin_lock(&res->d_lock);
+               res->d_sb = inode->i_sb;
+               res->d_parent = res;
+               res->d_inode = inode;
+               res->d_flags |= DCACHE_DISCONNECTED;
+               res->d_flags &= ~DCACHE_UNHASHED;
+               list_add(&res->d_alias, &inode->i_dentry);
+               hlist_add_head(&res->d_hash, &inode->i_sb->s_anon);
+               spin_unlock(&res->d_lock);
+
                inode = NULL; /* don't drop reference */
        }
        spin_unlock(&dcache_lock);
@@ -883,7 +958,7 @@ struct dentry * d_alloc_anon(struct inode *inode)
  * DCACHE_DISCONNECTED), then d_move that in place of the given dentry
  * and return it, else simply d_add the inode to the dentry and return NULL.
  *
- * This is (will be) needed in the lookup routine of any filesystem that is exportable
+ * This is needed in the lookup routine of any filesystem that is exportable
  * (via knfsd) so that we can build dcache paths to directories effectively.
  *
  * If a dentry was found and moved, then it is returned.  Otherwise NULL
@@ -894,11 +969,11 @@ struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
 {
        struct dentry *new = NULL;
 
-       if (inode && S_ISDIR(inode->i_mode)) {
+       if (inode) {
                spin_lock(&dcache_lock);
-               if (!list_empty(&inode->i_dentry)) {
-                       new = list_entry(inode->i_dentry.next, struct dentry, d_alias);
-                       __dget_locked(new);
+               new = __d_find_alias(inode, 1);
+               if (new) {
+                       BUG_ON(!(new->d_flags & DCACHE_DISCONNECTED));
                        spin_unlock(&dcache_lock);
                        security_d_instantiate(new, inode);
                        d_rehash(dentry);
@@ -930,8 +1005,7 @@ struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
  *
  * __d_lookup is dcache_lock free. The hash list is protected using RCU.
  * Memory barriers are used while updating and doing lockless traversal. 
- * To avoid races with d_move while rename is happening, d_move_count is 
- * used. 
+ * To avoid races with d_move while rename is happening, d_lock is used.
  *
  * Overflows in memcmp(), while d_move, are avoided by keeping the length
  * and name pointer in one structure pointed by d_qstr.
@@ -940,8 +1014,9 @@ struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
  * lookup is going on.
  *
  * dentry_unused list is not updated even if lookup finds the required dentry
- * in there. It is updated in places such as prune_dcache, shrink_dcache_sb and
- * select_parent. This laziness saves lookup from dcache_lock acquisition.
+ * in there. It is updated in places such as prune_dcache, shrink_dcache_sb,
+ * select_parent and __dget_locked. This laziness saves lookup from dcache_lock
+ * acquisition.
  *
  * d_lookup() is protected against the concurrent renames in some unrelated
  * directory using the seqlockt_t rename_lock.
@@ -972,55 +1047,50 @@ struct dentry * __d_lookup(struct dentry * parent, struct qstr * name)
 
        rcu_read_lock();
        
-       hlist_for_each (node, head) { 
+       hlist_for_each_rcu(node, head) {
                struct dentry *dentry; 
-               unsigned long move_count;
-               struct qstr * qstr;
+               struct qstr *qstr;
 
-               smp_read_barrier_depends();
                dentry = hlist_entry(node, struct dentry, d_hash);
 
-               /* if lookup ends up in a different bucket 
-                * due to concurrent rename, fail it
-                */
-               if (unlikely(dentry->d_bucket != head))
-                       break;
-
-               /*
-                * We must take a snapshot of d_move_count followed by
-                * read memory barrier before any search key comparison 
-                */
-               move_count = dentry->d_move_count;
-               smp_rmb();
-
                if (dentry->d_name.hash != hash)
                        continue;
                if (dentry->d_parent != parent)
                        continue;
 
-               qstr = dentry->d_qstr;
-               smp_read_barrier_depends();
+               spin_lock(&dentry->d_lock);
+
+               /*
+                * Recheck the dentry after taking the lock - d_move may have
+                * changed things.  Don't bother checking the hash because we're
+                * about to compare the whole name anyway.
+                */
+               if (dentry->d_parent != parent)
+                       goto next;
+
+               /*
+                * It is safe to compare names since d_move() cannot
+                * change the qstr (protected by d_lock).
+                */
+               qstr = &dentry->d_name;
                if (parent->d_op && parent->d_op->d_compare) {
                        if (parent->d_op->d_compare(parent, qstr, name))
-                               continue;
+                               goto next;
                } else {
                        if (qstr->len != len)
-                               continue;
+                               goto next;
                        if (memcmp(qstr->name, str, len))
-                               continue;
+                               goto next;
                }
-               spin_lock(&dentry->d_lock);
-               /*
-                * If dentry is moved, fail the lookup
-                */ 
-               if (likely(move_count == dentry->d_move_count)) {
-                       if (!d_unhashed(dentry)) {
-                               atomic_inc(&dentry->d_count);
-                               found = dentry;
-                       }
+
+               if (!d_unhashed(dentry)) {
+                       atomic_inc(&dentry->d_count);
+                       found = dentry;
                }
                spin_unlock(&dentry->d_lock);
                break;
+next:
+               spin_unlock(&dentry->d_lock);
        }
        rcu_read_unlock();
 
@@ -1054,7 +1124,7 @@ int d_validate(struct dentry *dentry, struct dentry *dparent)
        spin_lock(&dcache_lock);
        base = d_hash(dparent, dentry->d_name.hash);
        hlist_for_each(lhp,base) { 
-               /* read_barrier_depends() not required for d_hash list
+               /* hlist_for_each_rcu() not required for d_hash list
                 * as it is parsed under dcache_lock
                 */
                if (dentry == hlist_entry(lhp, struct dentry, d_hash)) {
@@ -1108,6 +1178,13 @@ void d_delete(struct dentry * dentry)
        spin_unlock(&dcache_lock);
 }
 
+static void __d_rehash(struct dentry * entry, struct hlist_head *list)
+{
+
+       entry->d_flags &= ~DCACHE_UNHASHED;
+       hlist_add_head_rcu(&entry->d_hash, list);
+}
+
 /**
  * d_rehash    - add an entry back to the hash
  * @entry: dentry to add to the hash
@@ -1118,10 +1195,11 @@ void d_delete(struct dentry * dentry)
 void d_rehash(struct dentry * entry)
 {
        struct hlist_head *list = d_hash(entry->d_parent, entry->d_name.hash);
+
        spin_lock(&dcache_lock);
-       entry->d_vfs_flags &= ~DCACHE_UNHASHED;
-       entry->d_bucket = list;
-       hlist_add_head_rcu(&entry->d_hash, list);
+       spin_lock(&entry->d_lock);
+       __d_rehash(entry, list);
+       spin_unlock(&entry->d_lock);
        spin_unlock(&dcache_lock);
 }
 
@@ -1140,28 +1218,40 @@ void d_rehash(struct dentry * entry)
  * then no longer matches the actual (corrupted) string of the target.
  * The hash value has to match the hash queue that the dentry is on..
  */
-static inline void switch_names(struct dentry * dentry, struct dentry * target)
+static void switch_names(struct dentry *dentry, struct dentry *target)
 {
-       const unsigned char *old_name, *new_name;
-       struct qstr *old_qstr, *new_qstr;
-
-       memcpy(dentry->d_iname, target->d_iname, DNAME_INLINE_LEN); 
-       old_qstr = target->d_qstr;
-       old_name = target->d_name.name;
-       new_qstr = dentry->d_qstr;
-       new_name = dentry->d_name.name;
-       if (old_name == target->d_iname) {
-               old_name = dentry->d_iname;
-               old_qstr = &dentry->d_name;
-       }
-       if (new_name == dentry->d_iname) {
-               new_name = target->d_iname;
-               new_qstr = &target->d_name;
+       if (dname_external(target)) {
+               if (dname_external(dentry)) {
+                       /*
+                        * Both external: swap the pointers
+                        */
+                       do_switch(target->d_name.name, dentry->d_name.name);
+               } else {
+                       /*
+                        * dentry:internal, target:external.  Steal target's
+                        * storage and make target internal.
+                        */
+                       dentry->d_name.name = target->d_name.name;
+                       target->d_name.name = target->d_iname;
+               }
+       } else {
+               if (dname_external(dentry)) {
+                       /*
+                        * dentry:external, target:internal.  Give dentry's
+                        * storage to target and make dentry internal
+                        */
+                       memcpy(dentry->d_iname, target->d_name.name,
+                                       target->d_name.len + 1);
+                       target->d_name.name = dentry->d_name.name;
+                       dentry->d_name.name = dentry->d_iname;
+               } else {
+                       /*
+                        * Both are internal.  Just copy target to dentry
+                        */
+                       memcpy(dentry->d_iname, target->d_name.name,
+                                       target->d_name.len + 1);
+               }
        }
-       target->d_name.name = new_name;
-       dentry->d_name.name = old_name;
-       target->d_qstr = new_qstr;
-       dentry->d_qstr = old_qstr;
 }
 
 /*
@@ -1187,6 +1277,8 @@ static inline void switch_names(struct dentry * dentry, struct dentry * target)
 
 void d_move(struct dentry * dentry, struct dentry * target)
 {
+       struct hlist_head *list;
+
        if (!dentry->d_inode)
                printk(KERN_WARNING "VFS: moving negative dcache entry\n");
 
@@ -1204,15 +1296,14 @@ void d_move(struct dentry * dentry, struct dentry * target)
        }
 
        /* Move the dentry to the target hash queue, if on different bucket */
-       if (dentry->d_vfs_flags & DCACHE_UNHASHED)
+       if (dentry->d_flags & DCACHE_UNHASHED)
                goto already_unhashed;
-       if (dentry->d_bucket != target->d_bucket) {
-               hlist_del_rcu(&dentry->d_hash);
+
+       hlist_del_rcu(&dentry->d_hash);
+
 already_unhashed:
-               dentry->d_bucket = target->d_bucket;
-               hlist_add_head_rcu(&dentry->d_hash, target->d_bucket);
-               dentry->d_vfs_flags &= ~DCACHE_UNHASHED;
-       }
+       list = d_hash(target->d_parent, target->d_name.hash);
+       __d_rehash(dentry, list);
 
        /* Unhash the target: dput() will then get rid of it */
        __d_drop(target);
@@ -1222,7 +1313,6 @@ already_unhashed:
 
        /* Switch the names.. */
        switch_names(dentry, target);
-       smp_wmb();
        do_switch(dentry->d_name.len, target->d_name.len);
        do_switch(dentry->d_name.hash, target->d_name.hash);
 
@@ -1239,7 +1329,6 @@ already_unhashed:
        }
 
        list_add(&dentry->d_child, &dentry->d_parent->d_subdirs);
-       dentry->d_move_count++;
        spin_unlock(&target->d_lock);
        spin_unlock(&dentry->d_lock);
        write_sequnlock(&rename_lock);
@@ -1337,6 +1426,7 @@ char * d_path(struct dentry *dentry, struct vfsmount *vfsmnt,
        char *res;
        struct vfsmount *rootmnt;
        struct dentry *root;
+
        read_lock(&current->fs->lock);
        rootmnt = mntget(current->fs->rootmnt);
        root = dget(current->fs->root);
@@ -1547,12 +1637,33 @@ static int __init set_dhash_entries(char *str)
 }
 __setup("dhash_entries=", set_dhash_entries);
 
+static void __init dcache_init_early(void)
+{
+       int loop;
+
+       /* If hashes are distributed across NUMA nodes, defer
+        * hash allocation until vmalloc space is available.
+        */
+       if (hashdist)
+               return;
+
+       dentry_hashtable =
+               alloc_large_system_hash("Dentry cache",
+                                       sizeof(struct hlist_head),
+                                       dhash_entries,
+                                       13,
+                                       HASH_EARLY,
+                                       &d_hash_shift,
+                                       &d_hash_mask,
+                                       0);
+
+       for (loop = 0; loop < (1 << d_hash_shift); loop++)
+               INIT_HLIST_HEAD(&dentry_hashtable[loop]);
+}
+
 static void __init dcache_init(unsigned long mempages)
 {
-       struct hlist_head *d;
-       unsigned long order;
-       unsigned int nr_hash;
-       int i;
+       int loop;
 
        /* 
         * A constructor could be added for stable state like the lists,
@@ -1562,51 +1673,27 @@ static void __init dcache_init(unsigned long mempages)
        dentry_cache = kmem_cache_create("dentry_cache",
                                         sizeof(struct dentry),
                                         0,
-                                        SLAB_RECLAIM_ACCOUNT,
+                                        SLAB_RECLAIM_ACCOUNT|SLAB_PANIC,
                                         NULL, NULL);
-       if (!dentry_cache)
-               panic("Cannot create dentry cache");
        
        set_shrinker(DEFAULT_SEEKS, shrink_dcache_memory);
 
-       if (!dhash_entries)
-               dhash_entries = PAGE_SHIFT < 13 ?
-                               mempages >> (13 - PAGE_SHIFT) :
-                               mempages << (PAGE_SHIFT - 13);
-
-       dhash_entries *= sizeof(struct hlist_head);
-       for (order = 0; ((1UL << order) << PAGE_SHIFT) < dhash_entries; order++)
-               ;
-
-       do {
-               unsigned long tmp;
-
-               nr_hash = (1UL << order) * PAGE_SIZE /
-                       sizeof(struct hlist_head);
-               d_hash_mask = (nr_hash - 1);
-
-               tmp = nr_hash;
-               d_hash_shift = 0;
-               while ((tmp >>= 1UL) != 0UL)
-                       d_hash_shift++;
-
-               dentry_hashtable = (struct hlist_head *)
-                       __get_free_pages(GFP_ATOMIC, order);
-       } while (dentry_hashtable == NULL && --order >= 0);
-
-       printk(KERN_INFO "Dentry cache hash table entries: %d (order: %ld, %ld bytes)\n",
-                       nr_hash, order, (PAGE_SIZE << order));
-
-       if (!dentry_hashtable)
-               panic("Failed to allocate dcache hash table\n");
+       /* Hash may have been set up in dcache_init_early */
+       if (!hashdist)
+               return;
 
-       d = dentry_hashtable;
-       i = nr_hash;
-       do {
-               INIT_HLIST_HEAD(d);
-               d++;
-               i--;
-       } while (i);
+       dentry_hashtable =
+               alloc_large_system_hash("Dentry cache",
+                                       sizeof(struct hlist_head),
+                                       dhash_entries,
+                                       13,
+                                       0,
+                                       &d_hash_shift,
+                                       &d_hash_mask,
+                                       0);
+
+       for (loop = 0; loop < (1 << d_hash_shift); loop++)
+               INIT_HLIST_HEAD(&dentry_hashtable[loop]);
 }
 
 /* SLAB cache for __getname() consumers */
@@ -1620,6 +1707,12 @@ EXPORT_SYMBOL(d_genocide);
 extern void bdev_cache_init(void);
 extern void chrdev_init(void);
 
+void __init vfs_caches_init_early(void)
+{
+       dcache_init_early();
+       inode_init_early();
+}
+
 void __init vfs_caches_init(unsigned long mempages)
 {
        unsigned long reserve;
@@ -1627,20 +1720,14 @@ void __init vfs_caches_init(unsigned long mempages)
        /* Base hash sizes on available memory, with a reserve equal to
            150% of current kernel size */
 
-       reserve = (mempages - nr_free_pages()) * 3/2;
+       reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1);
        mempages -= reserve;
 
-       names_cachep = kmem_cache_create("names_cache",
-                       PATH_MAX, 0,
-                       SLAB_HWCACHE_ALIGN, NULL, NULL);
-       if (!names_cachep)
-               panic("Cannot create names SLAB cache");
+       names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
+                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
 
-       filp_cachep = kmem_cache_create("filp",
-                       sizeof(struct file), 0,
-                       SLAB_HWCACHE_ALIGN, filp_ctor, filp_dtor);
-       if(!filp_cachep)
-               panic("Cannot create filp SLAB cache");
+       filp_cachep = kmem_cache_create("filp", sizeof(struct file), 0,
+                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, filp_ctor, filp_dtor);
 
        dcache_init(mempages);
        inode_init(mempages);
@@ -1668,8 +1755,6 @@ EXPORT_SYMBOL(dget_locked);
 EXPORT_SYMBOL(dput);
 EXPORT_SYMBOL(find_inode_number);
 EXPORT_SYMBOL(have_submounts);
-EXPORT_SYMBOL(is_subdir);
 EXPORT_SYMBOL(names_cachep);
-EXPORT_SYMBOL(shrink_dcache_anon);
 EXPORT_SYMBOL(shrink_dcache_parent);
 EXPORT_SYMBOL(shrink_dcache_sb);