ARM: tegra: make device can run on UP
[linux-3.10.git] / fs / super.c
index 5101f05..12f1237 100644 (file)
@@ -20,7 +20,7 @@
  *  Heavily rewritten for 'one fs - one tree' dcache architecture. AV, Mar 2000
  */
 
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/slab.h>
 #include <linux/acct.h>
 #include <linux/blkdev.h>
 #include <linux/backing-dev.h>
 #include <linux/rculist_bl.h>
 #include <linux/cleancache.h>
+#include <linux/fsnotify.h>
+#include <linux/lockdep.h>
 #include "internal.h"
 
 
 LIST_HEAD(super_blocks);
 DEFINE_SPINLOCK(sb_lock);
 
+static char *sb_writers_name[SB_FREEZE_LEVELS] = {
+       "sb_writers",
+       "sb_pagefaults",
+       "sb_internal",
+};
+
 /*
  * 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.
@@ -101,32 +109,63 @@ static int prune_super(struct shrinker *shrink, struct shrink_control *sc)
        return total_objects;
 }
 
+static int init_sb_writers(struct super_block *s, struct file_system_type *type)
+{
+       int err;
+       int i;
+
+       for (i = 0; i < SB_FREEZE_LEVELS; i++) {
+               err = percpu_counter_init(&s->s_writers.counter[i], 0);
+               if (err < 0)
+                       goto err_out;
+               lockdep_init_map(&s->s_writers.lock_map[i], sb_writers_name[i],
+                                &type->s_writers_key[i], 0);
+       }
+       init_waitqueue_head(&s->s_writers.wait);
+       init_waitqueue_head(&s->s_writers.wait_unfrozen);
+       return 0;
+err_out:
+       while (--i >= 0)
+               percpu_counter_destroy(&s->s_writers.counter[i]);
+       return err;
+}
+
+static void destroy_sb_writers(struct super_block *s)
+{
+       int i;
+
+       for (i = 0; i < SB_FREEZE_LEVELS; i++)
+               percpu_counter_destroy(&s->s_writers.counter[i]);
+}
+
 /**
  *     alloc_super     -       create new superblock
  *     @type:  filesystem type superblock should belong to
+ *     @flags: the mount flags
  *
  *     Allocates and initializes a new &struct super_block.  alloc_super()
  *     returns a pointer new superblock or %NULL if allocation had failed.
  */
-static struct super_block *alloc_super(struct file_system_type *type)
+static struct super_block *alloc_super(struct file_system_type *type, int flags)
 {
        struct super_block *s = kzalloc(sizeof(struct super_block),  GFP_USER);
        static const struct super_operations default_op;
 
        if (s) {
                if (security_sb_alloc(s)) {
+                       /*
+                        * We cannot call security_sb_free() without
+                        * security_sb_alloc() succeeding. So bail out manually
+                        */
                        kfree(s);
                        s = NULL;
                        goto out;
                }
 #ifdef CONFIG_SMP
                s->s_files = alloc_percpu(struct list_head);
-               if (!s->s_files) {
-                       security_sb_free(s);
-                       kfree(s);
-                       s = NULL;
-                       goto out;
-               } else {
+               if (!s->s_files)
+                       goto err_out;
+               else {
                        int i;
 
                        for_each_possible_cpu(i)
@@ -135,23 +174,20 @@ static struct super_block *alloc_super(struct file_system_type *type)
 #else
                INIT_LIST_HEAD(&s->s_files);
 #endif
+               if (init_sb_writers(s, type))
+                       goto err_out;
+               s->s_flags = flags;
                s->s_bdi = &default_backing_dev_info;
-               INIT_LIST_HEAD(&s->s_instances);
+               INIT_HLIST_NODE(&s->s_instances);
                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_LIST_HEAD(&s->s_mounts);
                init_rwsem(&s->s_umount);
-               mutex_init(&s->s_lock);
                lockdep_set_class(&s->s_umount, &type->s_umount_key);
                /*
-                * The locking rules for s_lock are up to the
-                * filesystem. For example ext3fs has different
-                * lock ordering than usbfs:
-                */
-               lockdep_set_class(&s->s_lock, &type->s_lock_key);
-               /*
                 * sget() can have s_umount recursion.
                 *
                 * When it cannot find a suitable sb, it allocates a new
@@ -174,7 +210,6 @@ static struct super_block *alloc_super(struct file_system_type *type)
                mutex_init(&s->s_dquot.dqio_mutex);
                mutex_init(&s->s_dquot.dqonoff_mutex);
                init_rwsem(&s->s_dquot.dqptr_sem);
-               init_waitqueue_head(&s->s_wait_unfrozen);
                s->s_maxbytes = MAX_NON_LFS;
                s->s_op = &default_op;
                s->s_time_gran = 1000000000;
@@ -182,9 +217,20 @@ static struct super_block *alloc_super(struct file_system_type *type)
 
                s->s_shrink.seeks = DEFAULT_SEEKS;
                s->s_shrink.shrink = prune_super;
+               s->s_shrink.batch = 1024;
        }
 out:
        return s;
+err_out:
+       security_sb_free(s);
+#ifdef CONFIG_SMP
+       if (s->s_files)
+               free_percpu(s->s_files);
+#endif
+       destroy_sb_writers(s);
+       kfree(s);
+       s = NULL;
+       goto out;
 }
 
 /**
@@ -198,7 +244,9 @@ static inline void destroy_super(struct super_block *s)
 #ifdef CONFIG_SMP
        free_percpu(s->s_files);
 #endif
+       destroy_sb_writers(s);
        security_sb_free(s);
+       WARN_ON(!list_empty(&s->s_mounts));
        kfree(s->s_subtype);
        kfree(s->s_options);
        kfree(s);
@@ -209,7 +257,7 @@ static inline void destroy_super(struct super_block *s)
 /*
  * Drop a superblock's refcount.  The caller must hold sb_lock.
  */
-void __put_super(struct super_block *sb)
+static void __put_super(struct super_block *sb)
 {
        if (!--sb->s_count) {
                list_del_init(&sb->s_list);
@@ -224,7 +272,7 @@ void __put_super(struct super_block *sb)
  *     Drops a temporary reference, frees superblock if there's no
  *     references left.
  */
-void put_super(struct super_block *sb)
+static void put_super(struct super_block *sb)
 {
        spin_lock(&sb_lock);
        __put_super(sb);
@@ -247,17 +295,11 @@ void deactivate_locked_super(struct super_block *s)
 {
        struct file_system_type *fs = s->s_type;
        if (atomic_dec_and_test(&s->s_active)) {
-               cleancache_flush_fs(s);
+               cleancache_invalidate_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.
-                */
-               rcu_barrier();
                put_filesystem(fs);
                put_super(s);
        } else {
@@ -314,7 +356,7 @@ static int grab_super(struct super_block *s) __releases(sb_lock)
 
 /*
  *     grab_super_passive - acquire a passive reference
- *     @s: reference we are trying to grab
+ *     @sb: 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
@@ -327,7 +369,7 @@ static int grab_super(struct super_block *s) __releases(sb_lock)
 bool grab_super_passive(struct super_block *sb)
 {
        spin_lock(&sb_lock);
-       if (list_empty(&sb->s_instances)) {
+       if (hlist_unhashed(&sb->s_instances)) {
                spin_unlock(&sb_lock);
                return false;
        }
@@ -336,7 +378,7 @@ bool grab_super_passive(struct super_block *sb)
        spin_unlock(&sb_lock);
 
        if (down_read_trylock(&sb->s_umount)) {
-               if (sb->s_root)
+               if (sb->s_root && (sb->s_flags & MS_BORN))
                        return true;
                up_read(&sb->s_umount);
        }
@@ -345,24 +387,6 @@ bool grab_super_passive(struct super_block *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);
-}
-
-EXPORT_SYMBOL(lock_super);
-EXPORT_SYMBOL(unlock_super);
-
 /**
  *     generic_shutdown_super  -       common helper for ->kill_sb()
  *     @sb: superblock to kill
@@ -384,7 +408,6 @@ void generic_shutdown_super(struct super_block *sb)
        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);
@@ -399,11 +422,10 @@ 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() */
-       list_del_init(&sb->s_instances);
+       hlist_del_init(&sb->s_instances);
        spin_unlock(&sb_lock);
        up_write(&sb->s_umount);
 }
@@ -415,21 +437,24 @@ EXPORT_SYMBOL(generic_shutdown_super);
  *     @type:  filesystem type superblock should belong to
  *     @test:  comparison callback
  *     @set:   setup callback
+ *     @flags: mount flags
  *     @data:  argument to each of them
  */
 struct super_block *sget(struct file_system_type *type,
                        int (*test)(struct super_block *,void *),
                        int (*set)(struct super_block *,void *),
+                       int flags,
                        void *data)
 {
        struct super_block *s = NULL;
+       struct hlist_node *node;
        struct super_block *old;
        int err;
 
 retry:
        spin_lock(&sb_lock);
        if (test) {
-               list_for_each_entry(old, &type->fs_supers, s_instances) {
+               hlist_for_each_entry(old, node, &type->fs_supers, s_instances) {
                        if (!test(old, data))
                                continue;
                        if (!grab_super(old))
@@ -449,7 +474,7 @@ retry:
        }
        if (!s) {
                spin_unlock(&sb_lock);
-               s = alloc_super(type);
+               s = alloc_super(type, flags);
                if (!s)
                        return ERR_PTR(-ENOMEM);
                goto retry;
@@ -465,7 +490,7 @@ retry:
        s->s_type = type;
        strlcpy(s->s_id, type->name, sizeof(s->s_id));
        list_add_tail(&s->s_list, &super_blocks);
-       list_add(&s->s_instances, &type->fs_supers);
+       hlist_add_head(&s->s_instances, &type->fs_supers);
        spin_unlock(&sb_lock);
        get_filesystem(type);
        register_shrinker(&s->s_shrink);
@@ -483,46 +508,6 @@ void drop_super(struct super_block *sb)
 EXPORT_SYMBOL(drop_super);
 
 /**
- * sync_supers - helper for periodic superblock writeback
- *
- * Call the write_super method if present on all dirty superblocks in
- * the system.  This is for the periodic writeback used by most older
- * filesystems.  For data integrity superblock writeback use
- * sync_filesystems() instead.
- *
- * Note: check the dirty flag before waiting, so we don't
- * hold up the sync while mounting a device. (The newly
- * mounted device won't need syncing.)
- */
-void sync_supers(void)
-{
-       struct super_block *sb, *p = NULL;
-
-       spin_lock(&sb_lock);
-       list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
-                       continue;
-               if (sb->s_op->write_super && sb->s_dirt) {
-                       sb->s_count++;
-                       spin_unlock(&sb_lock);
-
-                       down_read(&sb->s_umount);
-                       if (sb->s_root && sb->s_dirt)
-                               sb->s_op->write_super(sb);
-                       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);
-}
-
-/**
  *     iterate_supers - call function for all active superblocks
  *     @f: function to call
  *     @arg: argument to pass to it
@@ -536,13 +521,13 @@ void iterate_supers(void (*f)(struct super_block *, void *), void *arg)
 
        spin_lock(&sb_lock);
        list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
+               if (hlist_unhashed(&sb->s_instances))
                        continue;
                sb->s_count++;
                spin_unlock(&sb_lock);
 
                down_read(&sb->s_umount);
-               if (sb->s_root)
+               if (sb->s_root && (sb->s_flags & MS_BORN))
                        f(sb, arg);
                up_read(&sb->s_umount);
 
@@ -569,14 +554,15 @@ void iterate_supers_type(struct file_system_type *type,
        void (*f)(struct super_block *, void *), void *arg)
 {
        struct super_block *sb, *p = NULL;
+       struct hlist_node *node;
 
        spin_lock(&sb_lock);
-       list_for_each_entry(sb, &type->fs_supers, s_instances) {
+       hlist_for_each_entry(sb, node, &type->fs_supers, s_instances) {
                sb->s_count++;
                spin_unlock(&sb_lock);
 
                down_read(&sb->s_umount);
-               if (sb->s_root)
+               if (sb->s_root && (sb->s_flags & MS_BORN))
                        f(sb, arg);
                up_read(&sb->s_umount);
 
@@ -610,14 +596,14 @@ struct super_block *get_super(struct block_device *bdev)
        spin_lock(&sb_lock);
 rescan:
        list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
+               if (hlist_unhashed(&sb->s_instances))
                        continue;
                if (sb->s_bdev == bdev) {
                        sb->s_count++;
                        spin_unlock(&sb_lock);
                        down_read(&sb->s_umount);
                        /* still alive? */
-                       if (sb->s_root)
+                       if (sb->s_root && (sb->s_flags & MS_BORN))
                                return sb;
                        up_read(&sb->s_umount);
                        /* nope, got unmounted */
@@ -633,6 +619,29 @@ rescan:
 EXPORT_SYMBOL(get_super);
 
 /**
+ *     get_super_thawed - get thawed superblock of a device
+ *     @bdev: device to get the superblock for
+ *
+ *     Scans the superblock list and finds the superblock of the file system
+ *     mounted on the device. The superblock is returned once it is thawed
+ *     (or immediately if it was not frozen). %NULL is returned if no match
+ *     is found.
+ */
+struct super_block *get_super_thawed(struct block_device *bdev)
+{
+       while (1) {
+               struct super_block *s = get_super(bdev);
+               if (!s || s->s_writers.frozen == SB_UNFROZEN)
+                       return s;
+               up_read(&s->s_umount);
+               wait_event(s->s_writers.wait_unfrozen,
+                          s->s_writers.frozen == SB_UNFROZEN);
+               put_super(s);
+       }
+}
+EXPORT_SYMBOL(get_super_thawed);
+
+/**
  * get_active_super - get an active reference to the superblock of a device
  * @bdev: device to get the superblock for
  *
@@ -650,7 +659,7 @@ struct super_block *get_active_super(struct block_device *bdev)
 restart:
        spin_lock(&sb_lock);
        list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
+               if (hlist_unhashed(&sb->s_instances))
                        continue;
                if (sb->s_bdev == bdev) {
                        if (grab_super(sb)) /* drops sb_lock */
@@ -670,14 +679,14 @@ struct super_block *user_get_super(dev_t dev)
        spin_lock(&sb_lock);
 rescan:
        list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
+               if (hlist_unhashed(&sb->s_instances))
                        continue;
                if (sb->s_dev ==  dev) {
                        sb->s_count++;
                        spin_unlock(&sb_lock);
                        down_read(&sb->s_umount);
                        /* still alive? */
-                       if (sb->s_root)
+                       if (sb->s_root && (sb->s_flags & MS_BORN))
                                return sb;
                        up_read(&sb->s_umount);
                        /* nope, got unmounted */
@@ -704,7 +713,7 @@ int do_remount_sb(struct super_block *sb, int flags, void *data, int force)
        int retval;
        int remount_ro;
 
-       if (sb->s_frozen != SB_UNFROZEN)
+       if (sb->s_writers.frozen != SB_UNFROZEN)
                return -EBUSY;
 
 #ifdef CONFIG_BLOCK
@@ -722,18 +731,29 @@ int do_remount_sb(struct super_block *sb, int flags, void *data, int force)
        /* If we are remounting RDONLY and current sb is read/write,
           make sure there are no rw files opened */
        if (remount_ro) {
-               if (force)
+               if (force) {
                        mark_files_ro(sb);
-               else if (!fs_may_remount_ro(sb))
-                       return -EBUSY;
+               } else {
+                       retval = sb_prepare_remount_readonly(sb);
+                       if (retval)
+                               return retval;
+               }
        }
 
        if (sb->s_op->remount_fs) {
                retval = sb->s_op->remount_fs(sb, &flags, data);
-               if (retval)
-                       return retval;
+               if (retval) {
+                       if (!force)
+                               goto cancel_readonly;
+                       /* If forced remount, go ahead despite any errors */
+                       WARN(1, "forced remount of a %s fs returned %i\n",
+                            sb->s_type->name, retval);
+               }
        }
        sb->s_flags = (sb->s_flags & ~MS_RMT_MASK) | (flags & MS_RMT_MASK);
+       /* Needs to be ordered wrt mnt_is_readonly() */
+       smp_wmb();
+       sb->s_readonly_remount = 0;
 
        /*
         * Some filesystems modify their metadata via some other path than the
@@ -746,6 +766,10 @@ int do_remount_sb(struct super_block *sb, int flags, void *data, int force)
        if (remount_ro && sb->s_bdev)
                invalidate_bdev(sb->s_bdev);
        return 0;
+
+cancel_readonly:
+       sb->s_readonly_remount = 0;
+       return retval;
 }
 
 static void do_emergency_remount(struct work_struct *work)
@@ -754,12 +778,13 @@ static void do_emergency_remount(struct work_struct *work)
 
        spin_lock(&sb_lock);
        list_for_each_entry(sb, &super_blocks, s_list) {
-               if (list_empty(&sb->s_instances))
+               if (hlist_unhashed(&sb->s_instances))
                        continue;
                sb->s_count++;
                spin_unlock(&sb_lock);
                down_write(&sb->s_umount);
-               if (sb->s_root && sb->s_bdev && !(sb->s_flags & MS_RDONLY)) {
+               if (sb->s_root && sb->s_bdev && (sb->s_flags & MS_BORN) &&
+                   !(sb->s_flags & MS_RDONLY)) {
                        /*
                         * What lock protects sb->s_flags??
                         */
@@ -817,7 +842,7 @@ int get_anon_bdev(dev_t *p)
        else if (error)
                return -EAGAIN;
 
-       if ((dev & MAX_ID_MASK) == (1 << MINORBITS)) {
+       if ((dev & MAX_IDR_MASK) == (1 << MINORBITS)) {
                spin_lock(&unnamed_dev_lock);
                ida_remove(&unnamed_dev_ida, dev);
                if (unnamed_dev_start > dev)
@@ -885,13 +910,12 @@ struct dentry *mount_ns(struct file_system_type *fs_type, int flags,
 {
        struct super_block *sb;
 
-       sb = sget(fs_type, ns_test_super, ns_set_super, data);
+       sb = sget(fs_type, ns_test_super, ns_set_super, flags, data);
        if (IS_ERR(sb))
                return ERR_CAST(sb);
 
        if (!sb->s_root) {
                int err;
-               sb->s_flags = flags;
                err = fill_super(sb, data, flags & MS_SILENT ? 1 : 0);
                if (err) {
                        deactivate_locked_super(sb);
@@ -952,7 +976,8 @@ struct dentry *mount_bdev(struct file_system_type *fs_type,
                error = -EBUSY;
                goto error_bdev;
        }
-       s = sget(fs_type, test_bdev_super, set_bdev_super, bdev);
+       s = sget(fs_type, test_bdev_super, set_bdev_super, flags | MS_NOSEC,
+                bdev);
        mutex_unlock(&bdev->bd_fsfreeze_mutex);
        if (IS_ERR(s))
                goto error_s;
@@ -977,7 +1002,6 @@ struct dentry *mount_bdev(struct file_system_type *fs_type,
        } else {
                char b[BDEVNAME_SIZE];
 
-               s->s_flags = flags | MS_NOSEC;
                s->s_mode = mode;
                strlcpy(s->s_id, bdevname(bdev, b), sizeof(s->s_id));
                sb_set_blocksize(s, block_size(bdev));
@@ -1022,13 +1046,11 @@ struct dentry *mount_nodev(struct file_system_type *fs_type,
        int (*fill_super)(struct super_block *, void *, int))
 {
        int error;
-       struct super_block *s = sget(fs_type, NULL, set_anon_super, NULL);
+       struct super_block *s = sget(fs_type, NULL, set_anon_super, flags, NULL);
 
        if (IS_ERR(s))
                return ERR_CAST(s);
 
-       s->s_flags = flags;
-
        error = fill_super(s, data, flags & MS_SILENT ? 1 : 0);
        if (error) {
                deactivate_locked_super(s);
@@ -1051,11 +1073,10 @@ struct dentry *mount_single(struct file_system_type *fs_type,
        struct super_block *s;
        int error;
 
-       s = sget(fs_type, compare_single, set_anon_super, NULL);
+       s = sget(fs_type, compare_single, set_anon_super, flags, NULL);
        if (IS_ERR(s))
                return ERR_CAST(s);
        if (!s->s_root) {
-               s->s_flags = flags;
                error = fill_super(s, data, flags & MS_SILENT ? 1 : 0);
                if (error) {
                        deactivate_locked_super(s);
@@ -1123,6 +1144,120 @@ out:
        return ERR_PTR(error);
 }
 
+/*
+ * This is an internal function, please use sb_end_{write,pagefault,intwrite}
+ * instead.
+ */
+void __sb_end_write(struct super_block *sb, int level)
+{
+       percpu_counter_dec(&sb->s_writers.counter[level-1]);
+       /*
+        * Make sure s_writers are updated before we wake up waiters in
+        * freeze_super().
+        */
+       smp_mb();
+       if (waitqueue_active(&sb->s_writers.wait))
+               wake_up(&sb->s_writers.wait);
+       rwsem_release(&sb->s_writers.lock_map[level-1], 1, _RET_IP_);
+}
+EXPORT_SYMBOL(__sb_end_write);
+
+#ifdef CONFIG_LOCKDEP
+/*
+ * We want lockdep to tell us about possible deadlocks with freezing but
+ * it's it bit tricky to properly instrument it. Getting a freeze protection
+ * works as getting a read lock but there are subtle problems. XFS for example
+ * gets freeze protection on internal level twice in some cases, which is OK
+ * only because we already hold a freeze protection also on higher level. Due
+ * to these cases we have to tell lockdep we are doing trylock when we
+ * already hold a freeze protection for a higher freeze level.
+ */
+static void acquire_freeze_lock(struct super_block *sb, int level, bool trylock,
+                               unsigned long ip)
+{
+       int i;
+
+       if (!trylock) {
+               for (i = 0; i < level - 1; i++)
+                       if (lock_is_held(&sb->s_writers.lock_map[i])) {
+                               trylock = true;
+                               break;
+                       }
+       }
+       rwsem_acquire_read(&sb->s_writers.lock_map[level-1], 0, trylock, ip);
+}
+#endif
+
+/*
+ * This is an internal function, please use sb_start_{write,pagefault,intwrite}
+ * instead.
+ */
+int __sb_start_write(struct super_block *sb, int level, bool wait)
+{
+retry:
+       if (unlikely(sb->s_writers.frozen >= level)) {
+               if (!wait)
+                       return 0;
+               wait_event(sb->s_writers.wait_unfrozen,
+                          sb->s_writers.frozen < level);
+       }
+
+#ifdef CONFIG_LOCKDEP
+       acquire_freeze_lock(sb, level, !wait, _RET_IP_);
+#endif
+       percpu_counter_inc(&sb->s_writers.counter[level-1]);
+       /*
+        * Make sure counter is updated before we check for frozen.
+        * freeze_super() first sets frozen and then checks the counter.
+        */
+       smp_mb();
+       if (unlikely(sb->s_writers.frozen >= level)) {
+               __sb_end_write(sb, level);
+               goto retry;
+       }
+       return 1;
+}
+EXPORT_SYMBOL(__sb_start_write);
+
+/**
+ * sb_wait_write - wait until all writers to given file system finish
+ * @sb: the super for which we wait
+ * @level: type of writers we wait for (normal vs page fault)
+ *
+ * This function waits until there are no writers of given type to given file
+ * system. Caller of this function should make sure there can be no new writers
+ * of type @level before calling this function. Otherwise this function can
+ * livelock.
+ */
+static void sb_wait_write(struct super_block *sb, int level)
+{
+       s64 writers;
+
+       /*
+        * We just cycle-through lockdep here so that it does not complain
+        * about returning with lock to userspace
+        */
+       rwsem_acquire(&sb->s_writers.lock_map[level-1], 0, 0, _THIS_IP_);
+       rwsem_release(&sb->s_writers.lock_map[level-1], 1, _THIS_IP_);
+
+       do {
+               DEFINE_WAIT(wait);
+
+               /*
+                * We use a barrier in prepare_to_wait() to separate setting
+                * of frozen and checking of the counter
+                */
+               prepare_to_wait(&sb->s_writers.wait, &wait,
+                               TASK_UNINTERRUPTIBLE);
+
+               writers = percpu_counter_sum(&sb->s_writers.counter[level-1]);
+               if (writers)
+                       schedule();
+
+               finish_wait(&sb->s_writers.wait, &wait);
+       } while (writers);
+}
+
 /**
  * freeze_super - lock the filesystem and force it into a consistent state
  * @sb: the super to lock
@@ -1130,6 +1265,31 @@ out:
  * Syncs the super to make sure the filesystem is consistent and calls the fs's
  * freeze_fs.  Subsequent calls to this without first thawing the fs will return
  * -EBUSY.
+ *
+ * During this function, sb->s_writers.frozen goes through these values:
+ *
+ * SB_UNFROZEN: File system is normal, all writes progress as usual.
+ *
+ * SB_FREEZE_WRITE: The file system is in the process of being frozen.  New
+ * writes should be blocked, though page faults are still allowed. We wait for
+ * all writes to complete and then proceed to the next stage.
+ *
+ * SB_FREEZE_PAGEFAULT: Freezing continues. Now also page faults are blocked
+ * but internal fs threads can still modify the filesystem (although they
+ * should not dirty new pages or inodes), writeback can run etc. After waiting
+ * for all running page faults we sync the filesystem which will clean all
+ * dirty pages and inodes (no new dirty pages or inodes can be created when
+ * sync is running).
+ *
+ * SB_FREEZE_FS: The file system is frozen. Now all internal sources of fs
+ * modification are blocked (e.g. XFS preallocation truncation on inode
+ * reclaim). This is usually implemented by blocking new transactions for
+ * filesystems that have them and need this additional guard. After all
+ * internal writers are finished we call ->freeze_fs() to finish filesystem
+ * freezing. Then we transition to SB_FREEZE_COMPLETE state. This state is
+ * mostly auxiliary for filesystems to verify they do not modify frozen fs.
+ *
+ * sb->s_writers.frozen is protected by sb->s_umount.
  */
 int freeze_super(struct super_block *sb)
 {
@@ -1137,37 +1297,64 @@ int freeze_super(struct super_block *sb)
 
        atomic_inc(&sb->s_active);
        down_write(&sb->s_umount);
-       if (sb->s_frozen) {
+       if (sb->s_writers.frozen != SB_UNFROZEN) {
                deactivate_locked_super(sb);
                return -EBUSY;
        }
 
+       if (!(sb->s_flags & MS_BORN)) {
+               up_write(&sb->s_umount);
+               return 0;       /* sic - it's "nothing to do" */
+       }
+
        if (sb->s_flags & MS_RDONLY) {
-               sb->s_frozen = SB_FREEZE_TRANS;
-               smp_wmb();
+               /* Nothing to do really... */
+               sb->s_writers.frozen = SB_FREEZE_COMPLETE;
                up_write(&sb->s_umount);
                return 0;
        }
 
-       sb->s_frozen = SB_FREEZE_WRITE;
+       /* From now on, no new normal writers can start */
+       sb->s_writers.frozen = SB_FREEZE_WRITE;
        smp_wmb();
 
+       /* Release s_umount to preserve sb_start_write -> s_umount ordering */
+       up_write(&sb->s_umount);
+
+       sb_wait_write(sb, SB_FREEZE_WRITE);
+
+       /* Now we go and block page faults... */
+       down_write(&sb->s_umount);
+       sb->s_writers.frozen = SB_FREEZE_PAGEFAULT;
+       smp_wmb();
+
+       sb_wait_write(sb, SB_FREEZE_PAGEFAULT);
+
+       /* All writers are done so after syncing there won't be dirty data */
        sync_filesystem(sb);
 
-       sb->s_frozen = SB_FREEZE_TRANS;
+       /* Now wait for internal filesystem counter */
+       sb->s_writers.frozen = SB_FREEZE_FS;
        smp_wmb();
+       sb_wait_write(sb, SB_FREEZE_FS);
 
-       sync_blockdev(sb->s_bdev);
        if (sb->s_op->freeze_fs) {
                ret = sb->s_op->freeze_fs(sb);
                if (ret) {
                        printk(KERN_ERR
                                "VFS:Filesystem freeze failed\n");
-                       sb->s_frozen = SB_UNFROZEN;
+                       sb->s_writers.frozen = SB_UNFROZEN;
+                       smp_wmb();
+                       wake_up(&sb->s_writers.wait_unfrozen);
                        deactivate_locked_super(sb);
                        return ret;
                }
        }
+       /*
+        * This is just for debugging purposes so that fs can warn if it
+        * sees write activity when frozen is set to SB_FREEZE_COMPLETE.
+        */
+       sb->s_writers.frozen = SB_FREEZE_COMPLETE;
        up_write(&sb->s_umount);
        return 0;
 }
@@ -1184,7 +1371,7 @@ int thaw_super(struct super_block *sb)
        int error;
 
        down_write(&sb->s_umount);
-       if (sb->s_frozen == SB_UNFROZEN) {
+       if (sb->s_writers.frozen == SB_UNFROZEN) {
                up_write(&sb->s_umount);
                return -EINVAL;
        }
@@ -1197,16 +1384,15 @@ int thaw_super(struct super_block *sb)
                if (error) {
                        printk(KERN_ERR
                                "VFS:Filesystem thaw failed\n");
-                       sb->s_frozen = SB_FREEZE_TRANS;
                        up_write(&sb->s_umount);
                        return error;
                }
        }
 
 out:
-       sb->s_frozen = SB_UNFROZEN;
+       sb->s_writers.frozen = SB_UNFROZEN;
        smp_wmb();
-       wake_up(&sb->s_wait_unfrozen);
+       wake_up(&sb->s_writers.wait_unfrozen);
        deactivate_locked_super(sb);
 
        return 0;