Skip to contentSkip to conversation

Ocfs2: Add a mount option "coherency=*" for O_DIRECT writes.

PATCH New Tristan Ye 3 participants 6 messages Oct 9, 2010 – Oct 10, 2010

Subsystems (2)
lore
Share & tools
lore

Tristan Ye

author
#0 lore raw

PATCH 1/2 Ocfs2: Add a mount option "coherency=*" for O_DIRECT writes.

Currently, default behavior of O_DIRECT writes was allowing concurrent writing among nodes, no cluster coherency guaranteed (no EX locks was taken), it hurts buffered reads on other nodes by reading stale data from cache.

The new mount option introduce a chance to choose two different behaviors for O_DIRECT writes:

* coherency=full, as the default value, will disallow concurrent O_DIRECT writes by taking EX locks.

* coherency=buffered, allow concurrent O_DIRECT writes without EX lock among nodes, which gains high performance at risk of getting stale data on other nodes.

---
 Documentation/filesystems/ocfs2.txt |    7 +++++++
 fs/ocfs2/ocfs2.h                    |    3 +++
 fs/ocfs2/super.c                    |   23 +++++++++++++++++++++++
 3 files changed, 33 insertions(+), 0 deletions(-)
Documentation/filesystems/ocfs2.txt +7 -0
index 1f7ae14..5393e66 100644
@@ -87,3 +87,10 @@ dir_resv_level=	(*)	By default, directory reservations will scale with file
 			reservations - users should rarely need to change this
 			value. If allocation reservations are turned off, this
 			option will have no effect.
+coherency=full  (*)	Disallow concurrent O_DIRECT writes, cluster inode
+			lock will be taken to force other nodes drop cache,
+			therefore full cluster coherency is guaranteed even
+			for O_DIRECT writes.
+coherency=buffered	Allow concurrent O_DIRECT writes without EX lock among
+			nodes, which gains high performance at risk of getting
+			stale data on other nodes.
fs/ocfs2/ocfs2.h +3 -0
index c67003b..2b987be 100644
@@ -256,6 +256,9 @@ enum ocfs2_mount_options
 						   control lists */
 	OCFS2_MOUNT_USRQUOTA = 1 << 10, /* We support user quotas */
 	OCFS2_MOUNT_GRPQUOTA = 1 << 11, /* We support group quotas */
+
+	OCFS2_MOUNT_COHERENCY_BUFFERED = 1 << 12 /* Allow concurrent O_DIRECT
+						    writes */
 };
 
 #define OCFS2_OSB_SOFT_RO			0x0001
fs/ocfs2/super.c +23 -0
index fa1be1b..36ee274 100644
@@ -177,6 +177,8 @@ enum {
 	Opt_noacl,
 	Opt_usrquota,
 	Opt_grpquota,
+	Opt_coherency_buffered,
+	Opt_coherency_full,
 	Opt_resv_level,
 	Opt_dir_resv_level,
 	Opt_err,
@@ -205,6 +207,8 @@ static const match_table_t tokens = {
 	{Opt_noacl, "noacl"},
 	{Opt_usrquota, "usrquota"},
 	{Opt_grpquota, "grpquota"},
+	{Opt_coherency_buffered, "coherency=buffered"},
+	{Opt_coherency_full, "coherency=full"},
 	{Opt_resv_level, "resv_level=%u"},
 	{Opt_dir_resv_level, "dir_resv_level=%u"},
 	{Opt_err, NULL}
@@ -631,6 +635,14 @@ static int ocfs2_remount(struct super_block *sb, int *flags, char *data)
 		goto out;
 	}
 
+	if ((osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED) !=
+	    (parsed_options.mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED)) {
+		ret = -EINVAL;
+		mlog(ML_ERROR, "Cannot change cluster coherency mode on "
+		     "remount\n");
+		goto out;
+	}
+
 	/* Probably don't want this on remount; it might
 	 * mess with other nodes */
 	if (!(osb->s_mount_opt & OCFS2_MOUNT_INODE64) &&
@@ -1438,6 +1450,12 @@ static int ocfs2_parse_options(struct super_block *sb,
 		case Opt_grpquota:
 			mopt->mount_opt |= OCFS2_MOUNT_GRPQUOTA;
 			break;
+		case Opt_coherency_buffered:
+			mopt->mount_opt |= OCFS2_MOUNT_COHERENCY_BUFFERED;
+			break;
+		case Opt_coherency_full:
+			mopt->mount_opt &= ~OCFS2_MOUNT_COHERENCY_BUFFERED;
+			break;
 		case Opt_acl:
 			mopt->mount_opt |= OCFS2_MOUNT_POSIX_ACL;
 			mopt->mount_opt &= ~OCFS2_MOUNT_NO_POSIX_ACL;
@@ -1536,6 +1554,11 @@ static int ocfs2_show_options(struct seq_file *s, struct vfsmount *mnt)
 	if (opts & OCFS2_MOUNT_GRPQUOTA)
 		seq_printf(s, ",grpquota");
 
+	if (opts & OCFS2_MOUNT_COHERENCY_BUFFERED)
+		seq_printf(s, ",coherency=buffered");
+	else
+		seq_printf(s, ",coherency=full");
+
 	if (opts & OCFS2_MOUNT_NOUSERXATTR)
 		seq_printf(s, ",nouser_xattr");
 	else
Signature · 1 line
1.5.5
author
#1 lore raw

PATCH 2/2 Ocfs2: Handle O_DIRECT writes with coherency option.

Change the default behavior to take the EX lock for all writes, both buffered and O_DIRECT, then also allow concurrent O_DIRECT writes by recognizing "coherency=buffered" option.

---
 fs/ocfs2/file.c |   30 ++++++++++++++++++++++++++++--
 1 files changed, 28 insertions(+), 2 deletions(-)
fs/ocfs2/file.c +28 -2
index 9a03c15..22710a0 100644
@@ -2255,14 +2255,40 @@ relock:
 		have_alloc_sem = 1;
 	}
 
-	/* concurrent O_DIRECT writes are allowed */
-	rw_level = !direct_io;
+	/* 
+	 * concurrent O_DIRECT writes are allowed with
+	 * mount_option "coherency=buffered".
+	 */
+	if (direct_io) {
+		rw_level = !(osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED);
+	} else
+		rw_level = !direct_io;
+
 	ret = ocfs2_rw_lock(inode, rw_level);
 	if (ret < 0) {
 		mlog_errno(ret);
 		goto out_sems;
 	}
 
+	/*
+	 * O_DIRECT writes with "coherency=full" need to take EX cluster
+	 * inode_lock to guarantee coherency.
+	 */
+	if ((direct_io) &&
+	    !(osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED)) {
+		ret = ocfs2_inode_lock(inode, NULL, 1);
+		if (ret < 0) {
+			mlog_errno(ret);
+			goto out_sems;
+		}
+		
+		/*
+		 * Safe to drop the inode_lock immediately since we're just
+		 * telling other nodes to flush their cache.
+		 */
+		ocfs2_inode_unlock(inode, 1);
+	}
+
 	can_do_direct = direct_io;
 	ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
 					    iocb->ki_left, appending,
Signature · 1 line
1.5.5
#2 lore raw

PATCH 2/2 Ocfs2: Handle O_DIRECT writes with coherency option.

fs/ocfs2/file.c relock: · line 2266

Quoted 11 lines · Tristan Ye
fs/ocfs2/file.c +9 -2
-	/* concurrent O_DIRECT writes are allowed */
-	rw_level = !direct_io;
+	/* 
+	 * concurrent O_DIRECT writes are allowed with
+	 * mount_option "coherency=buffered".
+	 */
+	if (direct_io) {
+		rw_level = !(osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED);
+	} else
+		rw_level = !direct_io;
+
	I think I'd like:

	if (direct_io && (osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED))
		rw_level = 0;
	else
		rw_level = 1;

It actually matches your comment much better. But since we're going to be using it again later, perhaps you should set 'int full_coherency = (osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED)' up at the top of the function and then do:

	rw_level = (!direct_io || full_coherency)

fs/ocfs2/file.c relock: · line 2278

Quoted 12 lines
fs/ocfs2/file.c +6 -0
 	ret = ocfs2_rw_lock(inode, rw_level);
 	if (ret < 0) {
 		mlog_errno(ret);
 		goto out_sems;
 	}
 
+	/*
+	 * O_DIRECT writes with "coherency=full" need to take EX cluster
+	 * inode_lock to guarantee coherency.
+	 */
+	if ((direct_io) &&
+	    !(osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED)) {

Then this check can be:

if (direct_io && full_coherency) {
/*
* We need to take and drop the inode lock to force
* other nodes to drop their caches. Buffered I/O
* already does this in write_begin(). */

fs/ocfs2/file.c relock: · line 2288

Quoted 10 lines
fs/ocfs2/file.c +10 -0
+		ret = ocfs2_inode_lock(inode, NULL, 1);
+		if (ret < 0) {
+			mlog_errno(ret);
+			goto out_sems;
+		}
+		
+		/*
+		 * Safe to drop the inode_lock immediately since we're just
+		 * telling other nodes to flush their cache.
+		 */

And you don't need this comment.

Quoted 2 lines
fs/ocfs2/file.c +2 -0
+		ocfs2_inode_unlock(inode, 1);
+	}
Signature · 11 lines
"The lawgiver, of all beings, most owes the law allegiance.  He of all
 men should behave as though the law compelled him.  But it is the
 universal weakness of mankind that what we are given to administer we
 presently imagine we own."
        - H.G. Wells

Joel Becker
Consulting Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127
#3 lore raw

PATCH 2/2 Ocfs2: Handle O_DIRECT writes with coherency option.

Quoted 19 lines · Joel Becker

fs/ocfs2/file.c relock: · line 2266

Quoted 11 lines · Tristan Ye
fs/ocfs2/file.c +9 -2
-	/* concurrent O_DIRECT writes are allowed */
-	rw_level = !direct_io;
+	/*
+	 * concurrent O_DIRECT writes are allowed with
+	 * mount_option "coherency=buffered".
+	 */
+	if (direct_io) {
+		rw_level = !(osb->s_mount_opt&  OCFS2_MOUNT_COHERENCY_BUFFERED);
+	} else
+		rw_level = !direct_io;
+
	I think I'd like:

	if (direct_io&&  (osb->s_mount_opt&  OCFS2_MOUNT_COHERENCY_BUFFERED))
		rw_level = 0;
	else
		rw_level = 1;
Quoted 6 lines

It actually matches your comment much better. But since we're going to be using it again later, perhaps you should set 'int full_coherency = (osb->s_mount_opt& OCFS2_MOUNT_COHERENCY_BUFFERED)' up at the top of the function and then do:

	rw_level = (!direct_io || full_coherency)

yeah, it looks more natural.

Quoted 21 lines

fs/ocfs2/file.c relock:

Quoted 12 lines
  	ret = ocfs2_rw_lock(inode, rw_level);
  	if (ret<  0) {
  		mlog_errno(ret);
  		goto out_sems;
  	}
fs/ocfs2/file.c +6 -0
+	/*
+	 * O_DIRECT writes with "coherency=full" need to take EX cluster
+	 * inode_lock to guarantee coherency.
+	 */
+	if ((direct_io)&&
+	    !(osb->s_mount_opt&  OCFS2_MOUNT_COHERENCY_BUFFERED)) {

Then this check can be:

if (direct_io&& full_coherency) {
/*
* We need to take and drop the inode lock to force
* other nodes to drop their caches. Buffered I/O
* already does this in write_begin(). */

Quoted 11 lines

fs/ocfs2/file.c relock: · line 2288

Quoted 10 lines
fs/ocfs2/file.c +10 -0
+		ret = ocfs2_inode_lock(inode, NULL, 1);
+		if (ret<  0) {
+			mlog_errno(ret);
+			goto out_sems;
+		}
+		
+		/*
+		 * Safe to drop the inode_lock immediately since we're just
+		 * telling other nodes to flush their cache.
+		 */

And you don't need this comment.

I also have another concern. Do we really need the exclusive lock? I think a PR lock should
let others to flush the cache for us.

Regards,
Tao

#4 lore raw

PATCH 2/2 Ocfs2: Handle O_DIRECT writes with coherency option.

Quoted 20 lines · Tao Ma
Quoted 8 lines · Joel Becker

On Sat, Oct 09, 2010 at 07:26:42PM +0800, Tristan Ye wrote:

	if (direct_io&&  full_coherency) {
		/*
		 * We need to take and drop the inode lock to force
		 * other nodes to drop their caches.  Buffered I/O
		 * already does this in write_begin().
		 */
Quoted 11 lines

fs/ocfs2/file.c relock: · line 2288

Quoted 10 lines
fs/ocfs2/file.c +10 -0
+		ret = ocfs2_inode_lock(inode, NULL, 1);
+		if (ret<  0) {
+			mlog_errno(ret);
+			goto out_sems;
+		}
+		
+		/*
+		 * Safe to drop the inode_lock immediately since we're just
+		 * telling other nodes to flush their cache.
+		 */

And you don't need this comment.

Quoted 3 lines

I also have another concern. Do we really need the exclusive lock? I think a PR lock should
let others to flush the cache for us.

If they already have a PR, they're not going to drop it just
because we take a PR. They're going to stay at PR and keep their cached data, which is the right thing for them to do from a performance standpoint.

Joel

Signature · 9 lines
"There is a country in Europe where multiple-choice tests are
 illegal."
        - Sigfried Hulzer

Joel Becker
Consulting Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127
#5 lore raw

PATCH 1/2 Ocfs2: Add a mount option "coherency=*" for O_DIRECT writes.

In reply to Tristan Ye
Quoted 12 lines · Tristan Ye

Currently, default behavior of O_DIRECT writes was allowing concurrent writing among nodes, no cluster coherency guaranteed (no EX locks was taken), it hurts buffered reads on other nodes by reading stale data from cache.

The new mount option introduce a chance to choose two different behaviors for O_DIRECT writes:

* coherency=full, as the default value, will disallow concurrent O_DIRECT writes by taking EX locks.

Quoted 6 lines

* coherency=buffered, allow concurrent O_DIRECT writes without EX lock among nodes, which gains high performance at risk of getting stale data on other nodes.

I think these two patches should be combined into one. It's not
a lot of change, and it makes sense as one unit.

fs/ocfs2/super.c static int ocfs2_remount(struct super_block *sb, int *flags, char *data) · line 644

Quoted 11 lines
fs/ocfs2/super.c +7 -0
@@ -631,6 +635,14 @@ static int ocfs2_remount(struct super_block *sb, int *flags, char *data)
 		goto out;
 	}
 
+	if ((osb->s_mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED) !=
+	    (parsed_options.mount_opt & OCFS2_MOUNT_COHERENCY_BUFFERED)) {
+		ret = -EINVAL;
+		mlog(ML_ERROR, "Cannot change cluster coherency mode on "
+		     "remount\n");
+		goto out;
+	}

Don't include this. I think it's a feature to allow coherency
mode to be changed on the fly. It doesn't hurt us any; it is a safe thing to change.

Joel

Signature · 9 lines
Life's Little Instruction Book #207

	"Swing for the fence."

Joel Becker
Consulting Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127