[PATCH 0/2] Fix sys_sync() bug and slightly cleanup the code

[PATCH 0/2] Fix sys_sync() bug and slightly cleanup the code

[Jan Kara]

  Hi,

  this series of two patches fixes a bug in sys_sync() which could result in
not all data being safely on disk after this function returns. It also fixes
the same issue in fsync_super() which leads to assertion failure in JBD with
Jens's per-BDI writeback patches.
  The first patch fixes those two bugs, the second patch just simplifies the
code afterwards.

									Honza

[PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Jan Kara]

So far, do_sync() called:
  sync_inodes(0);
  sync_supers();
  sync_filesystems(0);
  sync_filesystems(1);
  sync_inodes(1);

This ordering makes it kind of hard for filesystems as sync_inodes(0) need not
submit all the IO (for example it skips inodes with I_SYNC set) so e.g. forcing
transaction to disk in ->sync_fs() is not really enough. Therefore sys_sync has
not been completely reliable on some filesystems (ext3, ext4, reiserfs, ocfs2
and others are hit by this) when racing e.g. with background writeback. A
similar problem hits also other filesystems (e.g. ext2) because of
write_supers() being called before the sync_inodes(1).

Change the ordering of calls in do_sync() - this requires a new function
sync_blkdevs() to preserve the property that block devices are always synced
after write_super() / sync_fs() call.

The same issue is fixed in __fsync_super() function used on umount /
remount read-only.

Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/super.c         |   25 ++++++++++++++++++++++++-
 fs/sync.c          |    3 ++-
 include/linux/fs.h |    2 ++
 3 files changed, 28 insertions(+), 2 deletions(-)

diff --git a/fs/super.c b/fs/super.c
index 786fe7d..4f56333 100644
--- a/fs/super.c
+++ b/fs/super.c
@@ -267,6 +267,7 @@ void __fsync_super(struct super_block *sb)
 {
 	sync_inodes_sb(sb, 0);
 	vfs_dq_sync(sb);
+	sync_inodes_sb(sb, 1);
 	lock_super(sb);
 	if (sb->s_dirt && sb->s_op->write_super)
 		sb->s_op->write_super(sb);
@@ -274,7 +275,6 @@ void __fsync_super(struct super_block *sb)
 	if (sb->s_op->sync_fs)
 		sb->s_op->sync_fs(sb, 1);
 	sync_blockdev(sb->s_bdev);
-	sync_inodes_sb(sb, 1);
 }
 
 /*
@@ -502,6 +502,29 @@ restart:
 	mutex_unlock(&mutex);
 }
 
+/*
+ *  Sync all block devices underlying some superblock
+ */
+void sync_blockdevs(void)
+{
+	struct super_block *sb;
+
+	spin_lock(&sb_lock);
+restart:
+	list_for_each_entry(sb, &super_blocks, s_list) {
+		sb->s_count++;
+		spin_unlock(&sb_lock);
+		down_read(&sb->s_umount);
+		if (sb->s_root)
+			sync_blockdev(sb->s_bdev);
+		up_read(&sb->s_umount);
+		spin_lock(&sb_lock);
+		if (__put_super_and_need_restart(sb))
+			goto restart;
+	}
+	spin_unlock(&sb_lock);
+}
+
 /**
  *	get_super - get the superblock of a device
  *	@bdev: device to get the superblock for
diff --git a/fs/sync.c b/fs/sync.c
index 7abc65f..fa14e42 100644
--- a/fs/sync.c
+++ b/fs/sync.c
@@ -26,10 +26,11 @@ static void do_sync(unsigned long wait)
 	wakeup_pdflush(0);
 	sync_inodes(0);		/* All mappings, inodes and their blockdevs */
 	vfs_dq_sync(NULL);
+	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
 	sync_supers();		/* Write the superblocks */
 	sync_filesystems(0);	/* Start syncing the filesystems */
 	sync_filesystems(wait);	/* Waitingly sync the filesystems */
-	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
+	sync_blockdevs();
 	if (!wait)
 		printk("Emergency Sync complete\n");
 	if (unlikely(laptop_mode))
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 5bed436..4bad02e 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -1942,6 +1942,7 @@ extern void bdput(struct block_device *);
 extern struct block_device *open_by_devnum(dev_t, fmode_t);
 extern void invalidate_bdev(struct block_device *);
 extern int sync_blockdev(struct block_device *bdev);
+extern void sync_blockdevs(void);
 extern struct super_block *freeze_bdev(struct block_device *);
 extern void emergency_thaw_all(void);
 extern int thaw_bdev(struct block_device *bdev, struct super_block *sb);
@@ -1951,6 +1952,7 @@ extern int fsync_no_super(struct block_device *);
 #else
 static inline void bd_forget(struct inode *inode) {}
 static inline int sync_blockdev(struct block_device *bdev) { return 0; }
+static inline void sync_blockdevs(void) { }
 static inline void invalidate_bdev(struct block_device *bdev) {}
 
 static inline struct super_block *freeze_bdev(struct block_device *sb)
-- 
1.6.0.2

Re: [PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Trond Myklebust]

On Wed, 2009-04-22 at 17:56 +0200, Jan Kara wrote:
> So far, do_sync() called:
>   sync_inodes(0);
>   sync_supers();
>   sync_filesystems(0);
>   sync_filesystems(1);
>   sync_inodes(1);
> 
> This ordering makes it kind of hard for filesystems as sync_inodes(0) need not
> submit all the IO (for example it skips inodes with I_SYNC set) so e.g. forcing
> transaction to disk in ->sync_fs() is not really enough. Therefore sys_sync has
> not been completely reliable on some filesystems (ext3, ext4, reiserfs, ocfs2
> and others are hit by this) when racing e.g. with background writeback. A
> similar problem hits also other filesystems (e.g. ext2) because of
> write_supers() being called before the sync_inodes(1).
> 
> Change the ordering of calls in do_sync() - this requires a new function
> sync_blkdevs() to preserve the property that block devices are always synced
> after write_super() / sync_fs() call.
> 
> The same issue is fixed in __fsync_super() function used on umount /
> remount read-only.
> 
> Signed-off-by: Jan Kara <jack@suse.cz>
> ---
>  fs/super.c         |   25 ++++++++++++++++++++++++-
>  fs/sync.c          |    3 ++-
>  include/linux/fs.h |    2 ++
>  3 files changed, 28 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/super.c b/fs/super.c
> index 786fe7d..4f56333 100644
> --- a/fs/super.c
> +++ b/fs/super.c
> @@ -267,6 +267,7 @@ void __fsync_super(struct super_block *sb)
>  {
>  	sync_inodes_sb(sb, 0);
>  	vfs_dq_sync(sb);
> +	sync_inodes_sb(sb, 1);
>  	lock_super(sb);
>  	if (sb->s_dirt && sb->s_op->write_super)
>  		sb->s_op->write_super(sb);
> @@ -274,7 +275,6 @@ void __fsync_super(struct super_block *sb)
>  	if (sb->s_op->sync_fs)
>  		sb->s_op->sync_fs(sb, 1);
>  	sync_blockdev(sb->s_bdev);
> -	sync_inodes_sb(sb, 1);
>  }
>  
>  /*
> @@ -502,6 +502,29 @@ restart:
>  	mutex_unlock(&mutex);
>  }
>  
> +/*
> + *  Sync all block devices underlying some superblock
> + */
> +void sync_blockdevs(void)
> +{
> +	struct super_block *sb;
> +
> +	spin_lock(&sb_lock);
> +restart:
> +	list_for_each_entry(sb, &super_blocks, s_list) {
> +		sb->s_count++;
> +		spin_unlock(&sb_lock);
> +		down_read(&sb->s_umount);
> +		if (sb->s_root)
> +			sync_blockdev(sb->s_bdev);
> +		up_read(&sb->s_umount);

What's the point of going to all this trouble of upping the sb->s_count,
and grabbing sb->s_umount if there is no sb->s_bdev to sync in the first
place?

> +		spin_lock(&sb_lock);
> +		if (__put_super_and_need_restart(sb))
> +			goto restart;
> +	}
> +	spin_unlock(&sb_lock);
> +}
> +
>  /**
>   *	get_super - get the superblock of a device
>   *	@bdev: device to get the superblock for
> diff --git a/fs/sync.c b/fs/sync.c
> index 7abc65f..fa14e42 100644
> --- a/fs/sync.c
> +++ b/fs/sync.c
> @@ -26,10 +26,11 @@ static void do_sync(unsigned long wait)
>  	wakeup_pdflush(0);
>  	sync_inodes(0);		/* All mappings, inodes and their blockdevs */
>  	vfs_dq_sync(NULL);
> +	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
>  	sync_supers();		/* Write the superblocks */
>  	sync_filesystems(0);	/* Start syncing the filesystems */
>  	sync_filesystems(wait);	/* Waitingly sync the filesystems */
> -	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
> +	sync_blockdevs();
>  	if (!wait)
>  		printk("Emergency Sync complete\n");
>  	if (unlikely(laptop_mode))
> diff --git a/include/linux/fs.h b/include/linux/fs.h
> index 5bed436..4bad02e 100644
> --- a/include/linux/fs.h
> +++ b/include/linux/fs.h
> @@ -1942,6 +1942,7 @@ extern void bdput(struct block_device *);
>  extern struct block_device *open_by_devnum(dev_t, fmode_t);
>  extern void invalidate_bdev(struct block_device *);
>  extern int sync_blockdev(struct block_device *bdev);
> +extern void sync_blockdevs(void);
>  extern struct super_block *freeze_bdev(struct block_device *);
>  extern void emergency_thaw_all(void);
>  extern int thaw_bdev(struct block_device *bdev, struct super_block *sb);
> @@ -1951,6 +1952,7 @@ extern int fsync_no_super(struct block_device *);
>  #else
>  static inline void bd_forget(struct inode *inode) {}
>  static inline int sync_blockdev(struct block_device *bdev) { return 0; }
> +static inline void sync_blockdevs(void) { }
>  static inline void invalidate_bdev(struct block_device *bdev) {}
>  
>  static inline struct super_block *freeze_bdev(struct block_device *sb)
> -- 
> 1.6.0.2
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Jan Kara]

On Wed 22-04-09 13:16:16, Trond Myklebust wrote:
> On Wed, 2009-04-22 at 17:56 +0200, Jan Kara wrote:
> > So far, do_sync() called:
> >   sync_inodes(0);
> >   sync_supers();
> >   sync_filesystems(0);
> >   sync_filesystems(1);
> >   sync_inodes(1);
> > 
> > This ordering makes it kind of hard for filesystems as sync_inodes(0) need not
> > submit all the IO (for example it skips inodes with I_SYNC set) so e.g. forcing
> > transaction to disk in ->sync_fs() is not really enough. Therefore sys_sync has
> > not been completely reliable on some filesystems (ext3, ext4, reiserfs, ocfs2
> > and others are hit by this) when racing e.g. with background writeback. A
> > similar problem hits also other filesystems (e.g. ext2) because of
> > write_supers() being called before the sync_inodes(1).
> > 
> > Change the ordering of calls in do_sync() - this requires a new function
> > sync_blkdevs() to preserve the property that block devices are always synced
> > after write_super() / sync_fs() call.
> > 
> > The same issue is fixed in __fsync_super() function used on umount /
> > remount read-only.
> > 
> > Signed-off-by: Jan Kara <jack@suse.cz>
> > ---
> >  fs/super.c         |   25 ++++++++++++++++++++++++-
> >  fs/sync.c          |    3 ++-
> >  include/linux/fs.h |    2 ++
> >  3 files changed, 28 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/super.c b/fs/super.c
> > index 786fe7d..4f56333 100644
> > --- a/fs/super.c
> > +++ b/fs/super.c
> > @@ -267,6 +267,7 @@ void __fsync_super(struct super_block *sb)
> >  {
> >  	sync_inodes_sb(sb, 0);
> >  	vfs_dq_sync(sb);
> > +	sync_inodes_sb(sb, 1);
> >  	lock_super(sb);
> >  	if (sb->s_dirt && sb->s_op->write_super)
> >  		sb->s_op->write_super(sb);
> > @@ -274,7 +275,6 @@ void __fsync_super(struct super_block *sb)
> >  	if (sb->s_op->sync_fs)
> >  		sb->s_op->sync_fs(sb, 1);
> >  	sync_blockdev(sb->s_bdev);
> > -	sync_inodes_sb(sb, 1);
> >  }
> >  
> >  /*
> > @@ -502,6 +502,29 @@ restart:
> >  	mutex_unlock(&mutex);
> >  }
> >  
> > +/*
> > + *  Sync all block devices underlying some superblock
> > + */
> > +void sync_blockdevs(void)
> > +{
> > +	struct super_block *sb;
> > +
> > +	spin_lock(&sb_lock);
> > +restart:
> > +	list_for_each_entry(sb, &super_blocks, s_list) {
> > +		sb->s_count++;
> > +		spin_unlock(&sb_lock);
> > +		down_read(&sb->s_umount);
> > +		if (sb->s_root)
> > +			sync_blockdev(sb->s_bdev);
> > +		up_read(&sb->s_umount);
> 
> What's the point of going to all this trouble of upping the sb->s_count,
> and grabbing sb->s_umount if there is no sb->s_bdev to sync in the first
> place?
  Thanks for the comment. Yes, we could save something here. I can improve
this and similarly my next patch. I'll just wait till tomorrow whether
someone else does not have other comments...

> 
> > +		spin_lock(&sb_lock);
> > +		if (__put_super_and_need_restart(sb))
> > +			goto restart;
> > +	}
> > +	spin_unlock(&sb_lock);
> > +}
> > +
> >  /**
> >   *	get_super - get the superblock of a device
> >   *	@bdev: device to get the superblock for
> > diff --git a/fs/sync.c b/fs/sync.c
> > index 7abc65f..fa14e42 100644
> > --- a/fs/sync.c
> > +++ b/fs/sync.c
> > @@ -26,10 +26,11 @@ static void do_sync(unsigned long wait)
> >  	wakeup_pdflush(0);
> >  	sync_inodes(0);		/* All mappings, inodes and their blockdevs */
> >  	vfs_dq_sync(NULL);
> > +	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
> >  	sync_supers();		/* Write the superblocks */
> >  	sync_filesystems(0);	/* Start syncing the filesystems */
> >  	sync_filesystems(wait);	/* Waitingly sync the filesystems */
> > -	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
> > +	sync_blockdevs();
> >  	if (!wait)
> >  		printk("Emergency Sync complete\n");
> >  	if (unlikely(laptop_mode))
> > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > index 5bed436..4bad02e 100644
> > --- a/include/linux/fs.h
> > +++ b/include/linux/fs.h
> > @@ -1942,6 +1942,7 @@ extern void bdput(struct block_device *);
> >  extern struct block_device *open_by_devnum(dev_t, fmode_t);
> >  extern void invalidate_bdev(struct block_device *);
> >  extern int sync_blockdev(struct block_device *bdev);
> > +extern void sync_blockdevs(void);
> >  extern struct super_block *freeze_bdev(struct block_device *);
> >  extern void emergency_thaw_all(void);
> >  extern int thaw_bdev(struct block_device *bdev, struct super_block *sb);
> > @@ -1951,6 +1952,7 @@ extern int fsync_no_super(struct block_device *);
> >  #else
> >  static inline void bd_forget(struct inode *inode) {}
> >  static inline int sync_blockdev(struct block_device *bdev) { return 0; }
> > +static inline void sync_blockdevs(void) { }
> >  static inline void invalidate_bdev(struct block_device *bdev) {}
> >  
> >  static inline struct super_block *freeze_bdev(struct block_device *sb)
> > -- 
> > 1.6.0.2

									Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: [PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Christoph Hellwig]

On Wed, Apr 22, 2009 at 05:56:20PM +0200, Jan Kara wrote:
> So far, do_sync() called:
>   sync_inodes(0);
>   sync_supers();
>   sync_filesystems(0);
>   sync_filesystems(1);
>   sync_inodes(1);
> 
> This ordering makes it kind of hard for filesystems as sync_inodes(0) need not
> submit all the IO (for example it skips inodes with I_SYNC set) so e.g. forcing
> transaction to disk in ->sync_fs() is not really enough. Therefore sys_sync has
> not been completely reliable on some filesystems (ext3, ext4, reiserfs, ocfs2
> and others are hit by this) when racing e.g. with background writeback. A
> similar problem hits also other filesystems (e.g. ext2) because of
> write_supers() being called before the sync_inodes(1).
> 
> Change the ordering of calls in do_sync() - this requires a new function
> sync_blkdevs() to preserve the property that block devices are always synced
> after write_super() / sync_fs() call.
> 
> The same issue is fixed in __fsync_super() function used on umount /
> remount read-only.

This looks reasonable, but I always fear we break something when
touching this path. It would be really nice if we could rewrite do_sync
to sit ontop of __fsync_super.  E.g. do a 

	for_each_sb()
		__fsync_super(sb, ASYNC);
	for_each_sb()
		__fsync_super(sb, SYNC);

so that we have one central place that makes sure a filesystem is
properly synced.

Another thing I want to do in this area is sort out the meaning of
write_super.  I'd really prefer to have every filesystem implement
->sync_fs for actual data-integerity syncs, and only leave ->write_super
for the periodic writeouts, as the current implementation is extrenly
confusing and causes a lot of trouble for filesystems doing their own
periodic sb writeback.

Re: [PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Jan Kara]

On Thu 23-04-09 07:10:40, Christoph Hellwig wrote:
> On Wed, Apr 22, 2009 at 05:56:20PM +0200, Jan Kara wrote:
> > So far, do_sync() called:
> >   sync_inodes(0);
> >   sync_supers();
> >   sync_filesystems(0);
> >   sync_filesystems(1);
> >   sync_inodes(1);
> > 
> > This ordering makes it kind of hard for filesystems as sync_inodes(0) need not
> > submit all the IO (for example it skips inodes with I_SYNC set) so e.g. forcing
> > transaction to disk in ->sync_fs() is not really enough. Therefore sys_sync has
> > not been completely reliable on some filesystems (ext3, ext4, reiserfs, ocfs2
> > and others are hit by this) when racing e.g. with background writeback. A
> > similar problem hits also other filesystems (e.g. ext2) because of
> > write_supers() being called before the sync_inodes(1).
> > 
> > Change the ordering of calls in do_sync() - this requires a new function
> > sync_blkdevs() to preserve the property that block devices are always synced
> > after write_super() / sync_fs() call.
> > 
> > The same issue is fixed in __fsync_super() function used on umount /
> > remount read-only.
> 
> This looks reasonable, but I always fear we break something when
> touching this path. It would be really nice if we could rewrite do_sync
> to sit ontop of __fsync_super.  E.g. do a 
> 
> 	for_each_sb()
> 		__fsync_super(sb, ASYNC);
> 	for_each_sb()
> 		__fsync_super(sb, SYNC);
> 
> so that we have one central place that makes sure a filesystem is
> properly synced.
  OK, makes sence. Will do.

> Another thing I want to do in this area is sort out the meaning of
> write_super.  I'd really prefer to have every filesystem implement
> ->sync_fs for actual data-integerity syncs, and only leave ->write_super
> for the periodic writeouts, as the current implementation is extrenly
> confusing and causes a lot of trouble for filesystems doing their own
> periodic sb writeback.
  Yes, that would be nice but I guess it's a partly a separate issue
(and has to touch a lot of filesystems). I'll keep write_supers() call in
the next version of the patch so that this split isn't made harded by it.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: [PATCH] vfs: Fix sys_sync() and fsync_super() reliability

[Christoph Hellwig]

On Thu, Apr 23, 2009 at 03:07:24PM +0200, Jan Kara wrote:
> > Another thing I want to do in this area is sort out the meaning of
> > write_super.  I'd really prefer to have every filesystem implement
> > ->sync_fs for actual data-integerity syncs, and only leave ->write_super
> > for the periodic writeouts, as the current implementation is extrenly
> > confusing and causes a lot of trouble for filesystems doing their own
> > periodic sb writeback.
>   Yes, that would be nice but I guess it's a partly a separate issue
> (and has to touch a lot of filesystems). I'll keep write_supers() call in
> the next version of the patch so that this split isn't made harded by it.

Yeah, separate issue.  We need to do this sync massaging in a couple of
small steps..

[PATCH] vfs: Merge sync_supers(), sync_filesystems() and sync_blockdevs()

[Jan Kara]

These three functions are quite similar so merge them to save superblock list
traversal code. As a bonus we get livelock avoidance for all these superblock
traversals. Also remove the condition that if wait == 0 and sb->s_dirt is
not set, then ->sync_fs() is not called. This does not really make much sence
since s_dirt is generally used by filesystem to mean that ->write_super() needs
to be called. But ->sync_fs() does different things. I even suspect that some
filesystems (btrfs?) sets s_dirt just to fool this logic.

Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/super.c          |  104 ++++++++++++++------------------------------------
 fs/sync.c           |   12 ++++--
 include/linux/fs.h  |   11 +++--
 mm/page-writeback.c |    2 +-
 4 files changed, 44 insertions(+), 85 deletions(-)

diff --git a/fs/super.c b/fs/super.c
index 4f56333..6d6209e 100644
--- a/fs/super.c
+++ b/fs/super.c
@@ -413,58 +413,24 @@ void drop_super(struct super_block *sb)
 
 EXPORT_SYMBOL(drop_super);
 
-static inline void write_super(struct super_block *sb)
-{
-	lock_super(sb);
-	if (sb->s_root && sb->s_dirt)
-		if (sb->s_op->write_super)
-			sb->s_op->write_super(sb);
-	unlock_super(sb);
-}
-
-/*
- * 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;
-
-	spin_lock(&sb_lock);
-restart:
-	list_for_each_entry(sb, &super_blocks, s_list) {
-		if (sb->s_dirt) {
-			sb->s_count++;
-			spin_unlock(&sb_lock);
-			down_read(&sb->s_umount);
-			write_super(sb);
-			up_read(&sb->s_umount);
-			spin_lock(&sb_lock);
-			if (__put_super_and_need_restart(sb))
-				goto restart;
-		}
-	}
-	spin_unlock(&sb_lock);
-}
-
 /*
- * Call the ->sync_fs super_op against all filesystems which are r/w and
- * which implement it.
+ * Call the ->write_super, ->sync_fs, sync_blockdev() against all filesystems
+ * which are r/w and which implement it.
  *
- * This operation is careful to avoid the livelock which could easily happen
- * if two or more filesystems are being continuously dirtied.  s_need_sync_fs
- * is used only here.  We set it against all filesystems and then clear it as
- * we sync them.  So redirtied filesystems are skipped.
+ * This operation is careful to avoid the livelock. Currently it's not so easy
+ * to trigger as it used to be - we restart scanning the superblock list only
+ * if the superblock got unmounted under us - but it's better to be safe and it
+ * doesn't cost us much. s_need_sync is used only here.  We set it against all
+ * filesystems and then clear it as we sync them.  So redirtied filesystems are
+ * skipped.
  *
  * But if process A is currently running sync_filesystems and then process B
- * calls sync_filesystems as well, process B will set all the s_need_sync_fs
- * flags again, which will cause process A to resync everything.  Fix that with
- * a local mutex.
+ * calls sync_filesystems as well, process B will set all the s_need_sync flags
+ * again, which will cause process A to resync everything.  Fix that with a
+ * local mutex.
  *
- * (Fabian) Avoid sync_fs with clean fs & wait mode 0
  */
-void sync_filesystems(int wait)
+void sync_filesystems(int what, int wait)
 {
 	struct super_block *sb;
 	static DEFINE_MUTEX(mutex);
@@ -472,26 +438,37 @@ void sync_filesystems(int wait)
 	mutex_lock(&mutex);		/* Could be down_interruptible */
 	spin_lock(&sb_lock);
 	list_for_each_entry(sb, &super_blocks, s_list) {
-		if (!sb->s_op->sync_fs)
+		if ((!sb->s_op->sync_fs || !(what & FSSYNC_FS)) &&
+		    (!sb->s_op->write_super || !(what & FSSYNC_SUPER)) &&
+		    !(what & FSSYNC_BDEV))
 			continue;
 		if (sb->s_flags & MS_RDONLY)
 			continue;
-		sb->s_need_sync_fs = 1;
+		sb->s_need_sync = 1;
 	}
 
 restart:
 	list_for_each_entry(sb, &super_blocks, s_list) {
-		if (!sb->s_need_sync_fs)
+		if (!sb->s_need_sync)
 			continue;
-		sb->s_need_sync_fs = 0;
+		sb->s_need_sync = 0;
 		if (sb->s_flags & MS_RDONLY)
 			continue;	/* hm.  Was remounted r/o meanwhile */
 		sb->s_count++;
 		spin_unlock(&sb_lock);
 		down_read(&sb->s_umount);
 		async_synchronize_full_domain(&sb->s_async_list);
-		if (sb->s_root && (wait || sb->s_dirt))
-			sb->s_op->sync_fs(sb, wait);
+		if (sb->s_root) {
+			lock_super(sb);
+			if (what & FSSYNC_SUPER && sb->s_dirt &&
+			    sb->s_op->write_super)
+				sb->s_op->write_super(sb);
+			unlock_super(sb);
+			if (what & FSSYNC_FS && sb->s_op->sync_fs)
+				sb->s_op->sync_fs(sb, wait);
+			if (what & FSSYNC_BDEV)
+				sync_blockdev(sb->s_bdev);
+		}
 		up_read(&sb->s_umount);
 		/* restart only when sb is no longer on the list */
 		spin_lock(&sb_lock);
@@ -502,29 +479,6 @@ restart:
 	mutex_unlock(&mutex);
 }
 
-/*
- *  Sync all block devices underlying some superblock
- */
-void sync_blockdevs(void)
-{
-	struct super_block *sb;
-
-	spin_lock(&sb_lock);
-restart:
-	list_for_each_entry(sb, &super_blocks, s_list) {
-		sb->s_count++;
-		spin_unlock(&sb_lock);
-		down_read(&sb->s_umount);
-		if (sb->s_root)
-			sync_blockdev(sb->s_bdev);
-		up_read(&sb->s_umount);
-		spin_lock(&sb_lock);
-		if (__put_super_and_need_restart(sb))
-			goto restart;
-	}
-	spin_unlock(&sb_lock);
-}
-
 /**
  *	get_super - get the superblock of a device
  *	@bdev: device to get the superblock for
diff --git a/fs/sync.c b/fs/sync.c
index fa14e42..3c677ee 100644
--- a/fs/sync.c
+++ b/fs/sync.c
@@ -26,11 +26,15 @@ static void do_sync(unsigned long wait)
 	wakeup_pdflush(0);
 	sync_inodes(0);		/* All mappings, inodes and their blockdevs */
 	vfs_dq_sync(NULL);
+	sync_filesystems(FSSYNC_FS, 0);	/* Start syncing the filesystems */
+	/*
+	 * We have to reliably submit IO for all the inodes before writing
+	 * super blocks and calling sync_fs(). Otherwise superblock could miss
+	 * some updates or journal could still have uncommitted data.
+	 */
 	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
-	sync_supers();		/* Write the superblocks */
-	sync_filesystems(0);	/* Start syncing the filesystems */
-	sync_filesystems(wait);	/* Waitingly sync the filesystems */
-	sync_blockdevs();
+	/* Waitingly sync the filesystems */
+	sync_filesystems(FSSYNC_SUPER | FSSYNC_FS | FSSYNC_BDEV, wait);
 	if (!wait)
 		printk("Emergency Sync complete\n");
 	if (unlikely(laptop_mode))
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 4bad02e..2758e75 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -1321,7 +1321,7 @@ struct super_block {
 	struct rw_semaphore	s_umount;
 	struct mutex		s_lock;
 	int			s_count;
-	int			s_need_sync_fs;
+	int			s_need_sync;
 	atomic_t		s_active;
 #ifdef CONFIG_SECURITY
 	void                    *s_security;
@@ -1942,7 +1942,6 @@ extern void bdput(struct block_device *);
 extern struct block_device *open_by_devnum(dev_t, fmode_t);
 extern void invalidate_bdev(struct block_device *);
 extern int sync_blockdev(struct block_device *bdev);
-extern void sync_blockdevs(void);
 extern struct super_block *freeze_bdev(struct block_device *);
 extern void emergency_thaw_all(void);
 extern int thaw_bdev(struct block_device *bdev, struct super_block *sb);
@@ -1952,7 +1951,6 @@ extern int fsync_no_super(struct block_device *);
 #else
 static inline void bd_forget(struct inode *inode) {}
 static inline int sync_blockdev(struct block_device *bdev) { return 0; }
-static inline void sync_blockdevs(void) { }
 static inline void invalidate_bdev(struct block_device *bdev) {}
 
 static inline struct super_block *freeze_bdev(struct block_device *sb)
@@ -2082,8 +2080,11 @@ extern int filemap_fdatawrite_range(struct address_space *mapping,
 				loff_t start, loff_t end);
 
 extern int vfs_fsync(struct file *file, struct dentry *dentry, int datasync);
-extern void sync_supers(void);
-extern void sync_filesystems(int wait);
+/* Flags telling what should be synced by sync_filesystems() */
+#define FSSYNC_FS 1
+#define FSSYNC_SUPER 2
+#define FSSYNC_BDEV 4
+extern void sync_filesystems(int what, int wait);
 extern void __fsync_super(struct super_block *sb);
 extern void emergency_sync(void);
 extern void emergency_remount(void);
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 30351f0..05a2948 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -768,7 +768,7 @@ static void wb_kupdate(unsigned long arg)
 		.range_cyclic	= 1,
 	};
 
-	sync_supers();
+	sync_filesystems(FSSYNC_SUPER, 0);
 
 	oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval);
 	start_jif = jiffies;
-- 
1.6.0.2

Re: [PATCH] vfs: Merge sync_supers(), sync_filesystems() and sync_blockdevs()

[Trond Myklebust]

On Wed, 2009-04-22 at 17:56 +0200, Jan Kara wrote:
> These three functions are quite similar so merge them to save superblock list
> traversal code. As a bonus we get livelock avoidance for all these superblock
> traversals. Also remove the condition that if wait == 0 and sb->s_dirt is
> not set, then ->sync_fs() is not called. This does not really make much sence
> since s_dirt is generally used by filesystem to mean that ->write_super() needs
> to be called. But ->sync_fs() does different things. I even suspect that some
> filesystems (btrfs?) sets s_dirt just to fool this logic.
> 
> Signed-off-by: Jan Kara <jack@suse.cz>
> ---
>  fs/super.c          |  104 ++++++++++++++------------------------------------
>  fs/sync.c           |   12 ++++--
>  include/linux/fs.h  |   11 +++--
>  mm/page-writeback.c |    2 +-
>  4 files changed, 44 insertions(+), 85 deletions(-)
> 
> diff --git a/fs/super.c b/fs/super.c
> index 4f56333..6d6209e 100644
> --- a/fs/super.c
> +++ b/fs/super.c
> @@ -413,58 +413,24 @@ void drop_super(struct super_block *sb)
>  
>  EXPORT_SYMBOL(drop_super);
>  
> -static inline void write_super(struct super_block *sb)
> -{
> -	lock_super(sb);
> -	if (sb->s_root && sb->s_dirt)
> -		if (sb->s_op->write_super)
> -			sb->s_op->write_super(sb);
> -	unlock_super(sb);
> -}
> -
> -/*
> - * 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;
> -
> -	spin_lock(&sb_lock);
> -restart:
> -	list_for_each_entry(sb, &super_blocks, s_list) {
> -		if (sb->s_dirt) {
> -			sb->s_count++;
> -			spin_unlock(&sb_lock);
> -			down_read(&sb->s_umount);
> -			write_super(sb);
> -			up_read(&sb->s_umount);
> -			spin_lock(&sb_lock);
> -			if (__put_super_and_need_restart(sb))
> -				goto restart;
> -		}
> -	}
> -	spin_unlock(&sb_lock);
> -}
> -
>  /*
> - * Call the ->sync_fs super_op against all filesystems which are r/w and
> - * which implement it.
> + * Call the ->write_super, ->sync_fs, sync_blockdev() against all filesystems
> + * which are r/w and which implement it.
>   *
> - * This operation is careful to avoid the livelock which could easily happen
> - * if two or more filesystems are being continuously dirtied.  s_need_sync_fs
> - * is used only here.  We set it against all filesystems and then clear it as
> - * we sync them.  So redirtied filesystems are skipped.
> + * This operation is careful to avoid the livelock. Currently it's not so easy
> + * to trigger as it used to be - we restart scanning the superblock list only
> + * if the superblock got unmounted under us - but it's better to be safe and it
> + * doesn't cost us much. s_need_sync is used only here.  We set it against all
> + * filesystems and then clear it as we sync them.  So redirtied filesystems are
> + * skipped.
>   *
>   * But if process A is currently running sync_filesystems and then process B
> - * calls sync_filesystems as well, process B will set all the s_need_sync_fs
> - * flags again, which will cause process A to resync everything.  Fix that with
> - * a local mutex.
> + * calls sync_filesystems as well, process B will set all the s_need_sync flags
> + * again, which will cause process A to resync everything.  Fix that with a
> + * local mutex.
>   *
> - * (Fabian) Avoid sync_fs with clean fs & wait mode 0
>   */
> -void sync_filesystems(int wait)
> +void sync_filesystems(int what, int wait)
>  {
>  	struct super_block *sb;
>  	static DEFINE_MUTEX(mutex);
> @@ -472,26 +438,37 @@ void sync_filesystems(int wait)
>  	mutex_lock(&mutex);		/* Could be down_interruptible */
>  	spin_lock(&sb_lock);
>  	list_for_each_entry(sb, &super_blocks, s_list) {
> -		if (!sb->s_op->sync_fs)
> +		if ((!sb->s_op->sync_fs || !(what & FSSYNC_FS)) &&
> +		    (!sb->s_op->write_super || !(what & FSSYNC_SUPER)) &&
> +		    !(what & FSSYNC_BDEV))

Ditto. There should be a check for sb->s_bdev here too...

>  			continue;
>  		if (sb->s_flags & MS_RDONLY)
>  			continue;
> -		sb->s_need_sync_fs = 1;
> +		sb->s_need_sync = 1;
>  	}
>  
>  restart:
>  	list_for_each_entry(sb, &super_blocks, s_list) {
> -		if (!sb->s_need_sync_fs)
> +		if (!sb->s_need_sync)
>  			continue;
> -		sb->s_need_sync_fs = 0;
> +		sb->s_need_sync = 0;
>  		if (sb->s_flags & MS_RDONLY)
>  			continue;	/* hm.  Was remounted r/o meanwhile */
>  		sb->s_count++;
>  		spin_unlock(&sb_lock);
>  		down_read(&sb->s_umount);
>  		async_synchronize_full_domain(&sb->s_async_list);
> -		if (sb->s_root && (wait || sb->s_dirt))
> -			sb->s_op->sync_fs(sb, wait);
> +		if (sb->s_root) {
> +			lock_super(sb);
> +			if (what & FSSYNC_SUPER && sb->s_dirt &&
> +			    sb->s_op->write_super)
> +				sb->s_op->write_super(sb);
> +			unlock_super(sb);
> +			if (what & FSSYNC_FS && sb->s_op->sync_fs)
> +				sb->s_op->sync_fs(sb, wait);
> +			if (what & FSSYNC_BDEV)
> +				sync_blockdev(sb->s_bdev);
> +		}
>  		up_read(&sb->s_umount);
>  		/* restart only when sb is no longer on the list */
>  		spin_lock(&sb_lock);
> @@ -502,29 +479,6 @@ restart:
>  	mutex_unlock(&mutex);
>  }
>  
> -/*
> - *  Sync all block devices underlying some superblock
> - */
> -void sync_blockdevs(void)
> -{
> -	struct super_block *sb;
> -
> -	spin_lock(&sb_lock);
> -restart:
> -	list_for_each_entry(sb, &super_blocks, s_list) {
> -		sb->s_count++;
> -		spin_unlock(&sb_lock);
> -		down_read(&sb->s_umount);
> -		if (sb->s_root)
> -			sync_blockdev(sb->s_bdev);
> -		up_read(&sb->s_umount);
> -		spin_lock(&sb_lock);
> -		if (__put_super_and_need_restart(sb))
> -			goto restart;
> -	}
> -	spin_unlock(&sb_lock);
> -}
> -
>  /**
>   *	get_super - get the superblock of a device
>   *	@bdev: device to get the superblock for
> diff --git a/fs/sync.c b/fs/sync.c
> index fa14e42..3c677ee 100644
> --- a/fs/sync.c
> +++ b/fs/sync.c
> @@ -26,11 +26,15 @@ static void do_sync(unsigned long wait)
>  	wakeup_pdflush(0);
>  	sync_inodes(0);		/* All mappings, inodes and their blockdevs */
>  	vfs_dq_sync(NULL);
> +	sync_filesystems(FSSYNC_FS, 0);	/* Start syncing the filesystems */
> +	/*
> +	 * We have to reliably submit IO for all the inodes before writing
> +	 * super blocks and calling sync_fs(). Otherwise superblock could miss
> +	 * some updates or journal could still have uncommitted data.
> +	 */
>  	sync_inodes(wait);	/* Mappings, inodes and blockdevs, again. */
> -	sync_supers();		/* Write the superblocks */
> -	sync_filesystems(0);	/* Start syncing the filesystems */
> -	sync_filesystems(wait);	/* Waitingly sync the filesystems */
> -	sync_blockdevs();
> +	/* Waitingly sync the filesystems */
> +	sync_filesystems(FSSYNC_SUPER | FSSYNC_FS | FSSYNC_BDEV, wait);
>  	if (!wait)
>  		printk("Emergency Sync complete\n");
>  	if (unlikely(laptop_mode))
> diff --git a/include/linux/fs.h b/include/linux/fs.h
> index 4bad02e..2758e75 100644
> --- a/include/linux/fs.h
> +++ b/include/linux/fs.h
> @@ -1321,7 +1321,7 @@ struct super_block {
>  	struct rw_semaphore	s_umount;
>  	struct mutex		s_lock;
>  	int			s_count;
> -	int			s_need_sync_fs;
> +	int			s_need_sync;
>  	atomic_t		s_active;
>  #ifdef CONFIG_SECURITY
>  	void                    *s_security;
> @@ -1942,7 +1942,6 @@ extern void bdput(struct block_device *);
>  extern struct block_device *open_by_devnum(dev_t, fmode_t);
>  extern void invalidate_bdev(struct block_device *);
>  extern int sync_blockdev(struct block_device *bdev);
> -extern void sync_blockdevs(void);
>  extern struct super_block *freeze_bdev(struct block_device *);
>  extern void emergency_thaw_all(void);
>  extern int thaw_bdev(struct block_device *bdev, struct super_block *sb);
> @@ -1952,7 +1951,6 @@ extern int fsync_no_super(struct block_device *);
>  #else
>  static inline void bd_forget(struct inode *inode) {}
>  static inline int sync_blockdev(struct block_device *bdev) { return 0; }
> -static inline void sync_blockdevs(void) { }
>  static inline void invalidate_bdev(struct block_device *bdev) {}
>  
>  static inline struct super_block *freeze_bdev(struct block_device *sb)
> @@ -2082,8 +2080,11 @@ extern int filemap_fdatawrite_range(struct address_space *mapping,
>  				loff_t start, loff_t end);
>  
>  extern int vfs_fsync(struct file *file, struct dentry *dentry, int datasync);
> -extern void sync_supers(void);
> -extern void sync_filesystems(int wait);
> +/* Flags telling what should be synced by sync_filesystems() */
> +#define FSSYNC_FS 1
> +#define FSSYNC_SUPER 2
> +#define FSSYNC_BDEV 4
> +extern void sync_filesystems(int what, int wait);
>  extern void __fsync_super(struct super_block *sb);
>  extern void emergency_sync(void);
>  extern void emergency_remount(void);
> diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> index 30351f0..05a2948 100644
> --- a/mm/page-writeback.c
> +++ b/mm/page-writeback.c
> @@ -768,7 +768,7 @@ static void wb_kupdate(unsigned long arg)
>  		.range_cyclic	= 1,
>  	};
>  
> -	sync_supers();
> +	sync_filesystems(FSSYNC_SUPER, 0);
>  
>  	oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval);
>  	start_jif = jiffies;

Re: [PATCH] vfs: Merge sync_supers(), sync_filesystems() and sync_blockdevs()

[Christoph Hellwig]

On Wed, Apr 22, 2009 at 05:56:21PM +0200, Jan Kara wrote:
> These three functions are quite similar so merge them to save superblock list
> traversal code. As a bonus we get livelock avoidance for all these superblock
> traversals. Also remove the condition that if wait == 0 and sb->s_dirt is
> not set, then ->sync_fs() is not called. This does not really make much sence
> since s_dirt is generally used by filesystem to mean that ->write_super() needs
> to be called. But ->sync_fs() does different things. I even suspect that some
> filesystems (btrfs?) sets s_dirt just to fool this logic.

Please separate the s_dirt bit out into a small patch, that's something
we probably want for -stable, too.

Otherwise the patch looks reasonable.

Re: [PATCH] vfs: Merge sync_supers(), sync_filesystems() and sync_blockdevs()

[Christoph Hellwig]

On Wed, Apr 22, 2009 at 05:56:21PM +0200, Jan Kara wrote:
> These three functions are quite similar so merge them to save superblock list
> traversal code. As a bonus we get livelock avoidance for all these superblock
> traversals. Also remove the condition that if wait == 0 and sb->s_dirt is
> not set, then ->sync_fs() is not called. This does not really make much sence
> since s_dirt is generally used by filesystem to mean that ->write_super() needs
> to be called. But ->sync_fs() does different things. I even suspect that some
> filesystems (btrfs?) sets s_dirt just to fool this logic.

Some more comments after looking at it in more details:

 - the FSSYNC_SUPER case really needs to do a trylock on the mutex,
   otherwise any in-progress sync would block pdflush for a long time.
   And as any real sync should write out the superblock it's not needed
   anyway during that time.  (Need to double-check the filesystems,
   though)
 - sync_filesystems really should move to fs/sync.c
 - I get more and more inclined to make sync just case of looping over
   the superblocks and do an fsync_super.  A plain sync fsync_super
   might be too slow so we can try to do an async one first and then
   a sync one as a second pass
 - that wakeup_pdflush in do_sync looks extremly fishy, we need to do
   all page writeback via sync_inodes_(sb) anyway, and doing this in
   parallel from pdflush just introduced tons of potential race
   opportunities
 - now if sync_filesystems just ends up calling __fsync_super for
   the normal sync path I wonder if there really is a point unifying
   it with the periodic write_super case.