Re: ext3: call blkdev_issue_flush on fsync

["Fernando Luis Vázquez Cao" <fernando@oss.ntt.co.jp>]

On Mon, 2009-01-19 at 13:03 +0100, Jan Kara wrote:
> On Sat 17-01-09 19:00:49, Fernando Luis Vázquez Cao wrote:
> > On Sat, 2009-01-17 at 18:47 +0900, Fernando Luis Vázquez Cao wrote:
> > > On Fri, 2009-01-16 at 17:30 +0100, Jan Kara wrote:
> > > > On Fri 16-01-09 22:55:01, Fernando Luis Vázquez Cao wrote:
> > > > > To ensure that bits are truly on-disk after an fsync or fdatasync, we
> > > > > should force a disk flush explicitly when there is dirty data/metadata
> > > > > and the journal didn't emit a write barrier (either because metadata is
> > > > > not being synched or barriers are disabled).
> > > > > 
> > > > > Signed-off-by: Fernando Luis Vazquez Cao <fernando@oss.ntt.co.jp>
> > > > > ---
> > > >   Only two minor nits:
> > > > 
> > > > > --- linux-2.6.29-rc1-orig/fs/ext3/fsync.c	2008-12-25 08:26:37.000000000 +0900
> > > > > +++ linux-2.6.29-rc1/fs/ext3/fsync.c	2009-01-16 22:18:53.000000000 +0900
> > > > > @@ -27,6 +27,7 @@
> > > > >  #include <linux/sched.h>
> > > > >  #include <linux/writeback.h>
> > > > >  #include <linux/jbd.h>
> > > > > +#include <linux/blkdev.h>
> > > > >  #include <linux/ext3_fs.h>
> > > > >  #include <linux/ext3_jbd.h>
> > > > >  
> > > > > @@ -45,6 +46,8 @@
> > > > >  int ext3_sync_file(struct file * file, struct dentry *dentry, int datasync)
> > > > >  {
> > > > >  	struct inode *inode = dentry->d_inode;
> > > > > +	journal_t *journal = EXT3_SB(inode->i_sb)->s_journal;
> > > > > +	unsigned long i_state = inode->i_state;
> > > > >  	int ret = 0;
> > > > >  
> > > > >  	J_ASSERT(ext3_journal_current_handle() == NULL);
> > > > > @@ -69,23 +72,33 @@ int ext3_sync_file(struct file * file, s
> > > > >  	 */
> > > > >  	if (ext3_should_journal_data(inode)) {
> > > > >  		ret = ext3_force_commit(inode->i_sb);
> > > > > +		if (!(journal->j_flags & JFS_BARRIER))
> > > > > +			goto no_journal_barrier;
> > > > >  		goto out;
> > > > >  	}
> > > > >  
> > > > > -	if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
> > > > > -		goto out;
> > > > > +	if (datasync && !(i_state & I_DIRTY_DATASYNC))
> > > > > +		goto flush_blkdev;
> > > > >  
> > > > >  	/*
> > > > >  	 * The VFS has written the file data.  If the inode is unaltered
> > > > >  	 * then we need not start a commit.
> > > > >  	 */
> > > > > -	if (inode->i_state & (I_DIRTY_SYNC|I_DIRTY_DATASYNC)) {
> > > > > +	if (i_state & (I_DIRTY_SYNC|I_DIRTY_DATASYNC)) {
> > > > >  		struct writeback_control wbc = {
> > > > >  			.sync_mode = WB_SYNC_ALL,
> > > > >  			.nr_to_write = 0, /* sys_fsync did this */
> > > > >  		};
> > > > >  		ret = sync_inode(inode, &wbc);
> > > > > +		if (journal && !(journal->j_flags & JFS_BARRIER))
> > > > > +			goto no_journal_barrier;
> > > >   I cannot imagine "journal" will be NULL here.
> > > 
> > > I'll try to check whether that is always so just in case.
> > > 
> > > >   And we can also optimize here a bit and do "goto out" because here
> > > > we know the barrier has been issued.
> > > 
> > > Yep, I was considering the same optimization. By the way, I was
> > > wondering if we should honor ext3 and ext4's "barrier" mount option for
> > > sys_fsync()/sys_fdatasync() and do not force a flush when "barrier=1".
> > 
> > The last phrase should read " do not force a flush when "barrier=0" ".
>   I was hesitating about this a bit. But I don't think so. The reason is
> that POSIX (or any other reasonable specification) mandates that fsync()
> should return only after the data is safely on storage. So if we don't
> flush blockdevice caches, we effectively violate POSIX and we should never
> do that. With barriers the matter is a bit different - that is just a
> filesystem specific thing, no standard guarantees anything.

Hi Jan,

Sorry it's taken me so long to get back to you.

Thinking a lit bit more about this issue, it occurred to me that adding
a new mount option à la existing "barrier" is likely to be preferable.
As an example where such an option could make sense, let's consider a
system with battery-backup cache devices. Since the battery-backup
guarantees the data still not committed to the platter will not vanish
in the event of a power down, it should be possible to obtain a
performance gain by optimizing out the device flush on every
fsync()/fdatasync() call.

If there is consensus on the propriety of this approach, I will send
updated patches.

Regards,

Fernando