Linux 3.1-rc6
[pandora-kernel.git] / fs / super.c
index ab3d672..3f56a26 100644 (file)
 LIST_HEAD(super_blocks);
 DEFINE_SPINLOCK(sb_lock);
 
+/*
+ * One thing we have to be careful of with a per-sb shrinker is that we don't
+ * drop the last active reference to the superblock from within the shrinker.
+ * If that happens we could trigger unregistering the shrinker from within the
+ * shrinker path and that leads to deadlock on the shrinker_rwsem. Hence we
+ * take a passive reference to the superblock to avoid this from occurring.
+ */
+static int prune_super(struct shrinker *shrink, struct shrink_control *sc)
+{
+       struct super_block *sb;
+       int     fs_objects = 0;
+       int     total_objects;
+
+       sb = container_of(shrink, struct super_block, s_shrink);
+
+       /*
+        * Deadlock avoidance.  We may hold various FS locks, and we don't want
+        * to recurse into the FS that called us in clear_inode() and friends..
+        */
+       if (sc->nr_to_scan && !(sc->gfp_mask & __GFP_FS))
+               return -1;
+
+       if (!grab_super_passive(sb))
+               return -1;
+
+       if (sb->s_op && sb->s_op->nr_cached_objects)
+               fs_objects = sb->s_op->nr_cached_objects(sb);
+
+       total_objects = sb->s_nr_dentry_unused +
+                       sb->s_nr_inodes_unused + fs_objects + 1;
+
+       if (sc->nr_to_scan) {
+               int     dentries;
+               int     inodes;
+
+               /* proportion the scan between the caches */
+               dentries = (sc->nr_to_scan * sb->s_nr_dentry_unused) /
+                                                       total_objects;
+               inodes = (sc->nr_to_scan * sb->s_nr_inodes_unused) /
+                                                       total_objects;
+               if (fs_objects)
+                       fs_objects = (sc->nr_to_scan * fs_objects) /
+                                                       total_objects;
+               /*
+                * prune the dcache first as the icache is pinned by it, then
+                * prune the icache, followed by the filesystem specific caches
+                */
+               prune_dcache_sb(sb, dentries);
+               prune_icache_sb(sb, inodes);
+
+               if (fs_objects && sb->s_op->free_cached_objects) {
+                       sb->s_op->free_cached_objects(sb, fs_objects);
+                       fs_objects = sb->s_op->nr_cached_objects(sb);
+               }
+               total_objects = sb->s_nr_dentry_unused +
+                               sb->s_nr_inodes_unused + fs_objects;
+       }
+
+       total_objects = (total_objects / 100) * sysctl_vfs_cache_pressure;
+       drop_super(sb);
+       return total_objects;
+}
+
 /**
  *     alloc_super     -       create new superblock
  *     @type:  filesystem type superblock should belong to
@@ -77,6 +140,8 @@ static struct super_block *alloc_super(struct file_system_type *type)
                INIT_HLIST_BL_HEAD(&s->s_anon);
                INIT_LIST_HEAD(&s->s_inodes);
                INIT_LIST_HEAD(&s->s_dentry_lru);
+               INIT_LIST_HEAD(&s->s_inode_lru);
+               spin_lock_init(&s->s_inode_lru_lock);
                init_rwsem(&s->s_umount);
                mutex_init(&s->s_lock);
                lockdep_set_class(&s->s_umount, &type->s_umount_key);
@@ -114,6 +179,10 @@ static struct super_block *alloc_super(struct file_system_type *type)
                s->s_op = &default_op;
                s->s_time_gran = 1000000000;
                s->cleancache_poolid = -1;
+
+               s->s_shrink.seeks = DEFAULT_SEEKS;
+               s->s_shrink.shrink = prune_super;
+               s->s_shrink.batch = 1024;
        }
 out:
        return s;
@@ -181,6 +250,10 @@ void deactivate_locked_super(struct super_block *s)
        if (atomic_dec_and_test(&s->s_active)) {
                cleancache_flush_fs(s);
                fs->kill_sb(s);
+
+               /* caches are now gone, we can safely kill the shrinker now */
+               unregister_shrinker(&s->s_shrink);
+
                /*
                 * We need to call rcu_barrier so all the delayed rcu free
                 * inodes are flushed before we release the fs module.
@@ -240,18 +313,49 @@ static int grab_super(struct super_block *s) __releases(sb_lock)
        return 0;
 }
 
+/*
+ *     grab_super_passive - acquire a passive reference
+ *     @s: reference we are trying to grab
+ *
+ *     Tries to acquire a passive reference. This is used in places where we
+ *     cannot take an active reference but we need to ensure that the
+ *     superblock does not go away while we are working on it. It returns
+ *     false if a reference was not gained, and returns true with the s_umount
+ *     lock held in read mode if a reference is gained. On successful return,
+ *     the caller must drop the s_umount lock and the passive reference when
+ *     done.
+ */
+bool grab_super_passive(struct super_block *sb)
+{
+       spin_lock(&sb_lock);
+       if (list_empty(&sb->s_instances)) {
+               spin_unlock(&sb_lock);
+               return false;
+       }
+
+       sb->s_count++;
+       spin_unlock(&sb_lock);
+
+       if (down_read_trylock(&sb->s_umount)) {
+               if (sb->s_root)
+                       return true;
+               up_read(&sb->s_umount);
+       }
+
+       put_super(sb);
+       return false;
+}
+
 /*
  * Superblock locking.  We really ought to get rid of these two.
  */
 void lock_super(struct super_block * sb)
 {
-       get_fs_excl();
        mutex_lock(&sb->s_lock);
 }
 
 void unlock_super(struct super_block * sb)
 {
-       put_fs_excl();
        mutex_unlock(&sb->s_lock);
 }
 
@@ -276,11 +380,9 @@ void generic_shutdown_super(struct super_block *sb)
 {
        const struct super_operations *sop = sb->s_op;
 
-
        if (sb->s_root) {
                shrink_dcache_for_umount(sb);
                sync_filesystem(sb);
-               get_fs_excl();
                sb->s_flags &= ~MS_ACTIVE;
 
                fsnotify_unmount_inodes(&sb->s_inodes);
@@ -295,7 +397,6 @@ void generic_shutdown_super(struct super_block *sb)
                           "Self-destruct in 5 seconds.  Have a nice day...\n",
                           sb->s_id);
                }
-               put_fs_excl();
        }
        spin_lock(&sb_lock);
        /* should be initialized for __put_super_and_need_restart() */
@@ -364,6 +465,7 @@ retry:
        list_add(&s->s_instances, &type->fs_supers);
        spin_unlock(&sb_lock);
        get_filesystem(type);
+       register_shrinker(&s->s_shrink);
        return s;
 }
 
@@ -451,6 +553,42 @@ void iterate_supers(void (*f)(struct super_block *, void *), void *arg)
        spin_unlock(&sb_lock);
 }
 
+/**
+ *     iterate_supers_type - call function for superblocks of given type
+ *     @type: fs type
+ *     @f: function to call
+ *     @arg: argument to pass to it
+ *
+ *     Scans the superblock list and calls given function, passing it
+ *     locked superblock and given argument.
+ */
+void iterate_supers_type(struct file_system_type *type,
+       void (*f)(struct super_block *, void *), void *arg)
+{
+       struct super_block *sb, *p = NULL;
+
+       spin_lock(&sb_lock);
+       list_for_each_entry(sb, &type->fs_supers, s_instances) {
+               sb->s_count++;
+               spin_unlock(&sb_lock);
+
+               down_read(&sb->s_umount);
+               if (sb->s_root)
+                       f(sb, arg);
+               up_read(&sb->s_umount);
+
+               spin_lock(&sb_lock);
+               if (p)
+                       __put_super(p);
+               p = sb;
+       }
+       if (p)
+               __put_super(p);
+       spin_unlock(&sb_lock);
+}
+
+EXPORT_SYMBOL(iterate_supers_type);
+
 /**
  *     get_super - get the superblock of a device
  *     @bdev: device to get the superblock for
@@ -657,7 +795,7 @@ static DEFINE_IDA(unnamed_dev_ida);
 static DEFINE_SPINLOCK(unnamed_dev_lock);/* protects the above */
 static int unnamed_dev_start = 0; /* don't bother trying below it */
 
-int set_anon_super(struct super_block *s, void *data)
+int get_anon_bdev(dev_t *p)
 {
        int dev;
        int error;
@@ -684,24 +822,38 @@ int set_anon_super(struct super_block *s, void *data)
                spin_unlock(&unnamed_dev_lock);
                return -EMFILE;
        }
-       s->s_dev = MKDEV(0, dev & MINORMASK);
-       s->s_bdi = &noop_backing_dev_info;
+       *p = MKDEV(0, dev & MINORMASK);
        return 0;
 }
+EXPORT_SYMBOL(get_anon_bdev);
 
-EXPORT_SYMBOL(set_anon_super);
-
-void kill_anon_super(struct super_block *sb)
+void free_anon_bdev(dev_t dev)
 {
-       int slot = MINOR(sb->s_dev);
-
-       generic_shutdown_super(sb);
+       int slot = MINOR(dev);
        spin_lock(&unnamed_dev_lock);
        ida_remove(&unnamed_dev_ida, slot);
        if (slot < unnamed_dev_start)
                unnamed_dev_start = slot;
        spin_unlock(&unnamed_dev_lock);
 }
+EXPORT_SYMBOL(free_anon_bdev);
+
+int set_anon_super(struct super_block *s, void *data)
+{
+       int error = get_anon_bdev(&s->s_dev);
+       if (!error)
+               s->s_bdi = &noop_backing_dev_info;
+       return error;
+}
+
+EXPORT_SYMBOL(set_anon_super);
+
+void kill_anon_super(struct super_block *sb)
+{
+       dev_t dev = sb->s_dev;
+       generic_shutdown_super(sb);
+       free_anon_bdev(dev);
+}
 
 EXPORT_SYMBOL(kill_anon_super);