vserver 1.9.3
[linux-2.6.git] / fs / dcache.c
index 613ff66..f938ad5 100644 (file)
 #include <linux/security.h>
 #include <linux/seqlock.h>
 #include <linux/swap.h>
+#include <linux/bootmem.h>
 
-#define DCACHE_PARANOIA 1
 /* #define DCACHE_DEBUG 1 */
 
+int sysctl_vfs_cache_pressure = 100;
+
 spinlock_t dcache_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
 seqlock_t rename_lock __cacheline_aligned_in_smp = SEQLOCK_UNLOCKED;
 
@@ -64,9 +66,9 @@ 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_name.name);
@@ -81,7 +83,7 @@ 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);
 }
 
 /*
@@ -142,6 +144,8 @@ 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;
 
@@ -284,12 +288,11 @@ struct dentry * dget_locked(struct dentry *dentry)
  * any other hashed alias over that one.
  */
 
-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) {
@@ -300,19 +303,26 @@ struct dentry * d_find_alias(struct inode *inode)
                if (!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.
@@ -611,7 +621,7 @@ 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.
  *
  */
@@ -626,7 +636,7 @@ 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);
                        }
 
                        /* 
@@ -663,7 +673,7 @@ static int shrink_dcache_memory(int nr, unsigned int gfp_mask)
                        return -1;
                prune_dcache(nr);
        }
-       return dentry_stat.nr_unused;
+       return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
 }
 
 /**
@@ -831,33 +841,27 @@ 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;
+               spin_lock(&res->d_lock);
+               res->d_sb = inode->i_sb;
+               res->d_parent = res;
+               res->d_inode = inode;
+
+               /*
+                * Set d_bucket to an "impossible" bucket address so
+                * that d_move() doesn't get a false positive
+                */
+               res->d_bucket = NULL;
+               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);
 
-                       /*
-                        * Set d_bucket to an "impossible" bucket address so
-                        * that d_move() doesn't get a false positive
-                        */
-                       res->d_bucket = NULL;
-                       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);
@@ -879,7 +883,7 @@ 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
@@ -890,11 +894,11 @@ 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);
@@ -968,11 +972,10 @@ 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; 
                struct qstr *qstr;
 
-               smp_read_barrier_depends();
                dentry = hlist_entry(node, struct dentry, d_hash);
 
                smp_rmb();
@@ -999,8 +1002,7 @@ struct dentry * __d_lookup(struct dentry * parent, struct qstr * name)
                if (dentry->d_parent != parent)
                        goto next;
 
-               qstr = &dentry->d_name;
-               smp_read_barrier_depends();
+               qstr = rcu_dereference(&dentry->d_name);
                if (parent->d_op && parent->d_op->d_compare) {
                        if (parent->d_op->d_compare(parent, qstr, name))
                                goto next;
@@ -1053,7 +1055,7 @@ 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)) {
@@ -1561,13 +1563,25 @@ static int __init set_dhash_entries(char *str)
 }
 __setup("dhash_entries=", set_dhash_entries);
 
-static void __init dcache_init(unsigned long mempages)
+static void __init dcache_init_early(void)
 {
-       struct hlist_head *d;
-       unsigned long order;
-       unsigned int nr_hash;
-       int i;
+       int loop;
+
+       dentry_hashtable =
+               alloc_large_system_hash("Dentry cache",
+                                       sizeof(struct hlist_head),
+                                       dhash_entries,
+                                       13,
+                                       0,
+                                       &d_hash_shift,
+                                       &d_hash_mask);
+
+       for (loop = 0; loop < (1 << d_hash_shift); loop++)
+               INIT_HLIST_HEAD(&dentry_hashtable[loop]);
+}
 
+static void __init dcache_init(unsigned long mempages)
+{
        /* 
         * A constructor could be added for stable state like the lists,
         * but it is probably not worth it because of the cache nature
@@ -1580,45 +1594,6 @@ static void __init dcache_init(unsigned long mempages)
                                         NULL, NULL);
        
        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");
-
-       d = dentry_hashtable;
-       i = nr_hash;
-       do {
-               INIT_HLIST_HEAD(d);
-               d++;
-               i--;
-       } while (i);
 }
 
 /* SLAB cache for __getname() consumers */
@@ -1632,6 +1607,12 @@ 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;