Skip to contentSkip to conversation

commit_metadata export operation v4

PATCH In review Ben Myers 4 participants 11 messages Feb 11, 2010 – Feb 12, 2010

Subsystems (4)
lore
Share & tools
lore

Ben Myers

author
#0 lore raw

cover commit_metadata export operation v4

Hey Bruce,

Here is the latest version of the knfsd sync changes. I have addressed the suggestions provided by Christoph, Trond, and Dave. This time rebased atop for-2.6.34 (I hope). Sorry for the extra noise.

Thanks,
Ben

---

Ben Myers (2):
commit_metadata export operation replacing nfsd_sync_dir xfs_export_operations.commit_metadata

 fs/nfsd/nfs3proc.c            |    2 -
 fs/nfsd/nfs4proc.c            |    2 -
 fs/nfsd/nfs4recover.c         |    4 -
 fs/nfsd/nfs4state.c           |    2 -
 fs/nfsd/nfsproc.c             |    4 +
 fs/nfsd/vfs.c                 |  116 +++++++++++++++++++----------------------
 fs/nfsd/vfs.h                 |    3 -
 fs/xfs/linux-2.6/xfs_export.c |   43 +++++++++++++++
 include/linux/exportfs.h      |    6 ++
 9 files changed, 111 insertions(+), 71 deletions(-)
Signature · 4 lines
_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs
author
#1 lore raw

PATCH 1/2 commit_metadata export operation replacing nfsd_sync_dir

- Add commit_metadata export_operation to allow the underlying filesystem to decide how to sync parent and child inodes most efficiently.

- Usage of nfsd_sync_dir and write_inode_now has been replaced with the commit_metadata function that takes a svc_fh and optional dentry for the child.

- The commit_metadata function calls the commit_metadata export_op if it's there, or else falls back to sync_inode instead of fsync and write_inode_now because only metadata need be synched here.

- nfsd4_sync_rec_dir now uses vfs_fsync so that commit_metadata can be static

- Add a 'delay_commit' arg to nfsd_setattr so that callers of nfsd_create_setattr can commit parent and child together avoiding an extra sync.

---
 fs/nfsd/nfs3proc.c       |    2 -
 fs/nfsd/nfs4proc.c       |    2 -
 fs/nfsd/nfs4recover.c    |    4 --
 fs/nfsd/nfs4state.c      |    2 -
 fs/nfsd/nfsproc.c        |    4 +-
 fs/nfsd/vfs.c            |  116 ++++++++++++++++++++++------------------------
 fs/nfsd/vfs.h            |    3 -
 include/linux/exportfs.h |    6 ++
 8 files changed, 68 insertions(+), 71 deletions(-)
fs/nfsd/nfs3proc.c +1 -1
index 3d68f45..fe3af23 100644
@@ -76,7 +76,7 @@ nfsd3_proc_setattr(struct svc_rqst *rqstp, struct nfsd3_sattrargs *argp,
 
 	fh_copy(&resp->fh, &argp->fh);
 	nfserr = nfsd_setattr(rqstp, &resp->fh, &argp->attrs,
-			      argp->check_guard, argp->guardtime);
+			      argp->check_guard, argp->guardtime, 0);
 	RETURN_STATUS(nfserr);
 }
 
fs/nfsd/nfs4proc.c +1 -1
index 37514c4..c6e3c7f 100644
@@ -803,7 +803,7 @@ nfsd4_setattr(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 	if (status)
 		goto out;
 	status = nfsd_setattr(rqstp, &cstate->current_fh, &setattr->sa_iattr,
-				0, (time_t)0);
+				0, 0, 0);
 out:
 	mnt_drop_write(cstate->current_fh.fh_export->ex_path.mnt);
 	return status;
fs/nfsd/nfs4recover.c +1 -3
index 5a754f7..98fb98e 100644
@@ -119,9 +119,7 @@ out_no_tfm:
 static void
 nfsd4_sync_rec_dir(void)
 {
-	mutex_lock(&rec_dir.dentry->d_inode->i_mutex);
-	nfsd_sync_dir(rec_dir.dentry);
-	mutex_unlock(&rec_dir.dentry->d_inode->i_mutex);
+	vfs_fsync(NULL, rec_dir.dentry, 0);
 }
 
 int
fs/nfsd/nfs4state.c +1 -1
index 3a20c09..f2fc8c8 100644
@@ -2290,7 +2290,7 @@ nfsd4_truncate(struct svc_rqst *rqstp, struct svc_fh *fh,
 		return 0;
 	if (!(open->op_share_access & NFS4_SHARE_ACCESS_WRITE))
 		return nfserr_inval;
-	return nfsd_setattr(rqstp, fh, &iattr, 0, (time_t)0);
+	return nfsd_setattr(rqstp, fh, &iattr, 0, 0, 0);
 }
 
 static __be32
fs/nfsd/nfsproc.c +2 -2
index a047ad6..f5e8280 100644
@@ -69,7 +69,7 @@ nfsd_proc_setattr(struct svc_rqst *rqstp, struct nfsd_sattrargs *argp,
 		argp->attrs.ia_valid, (long) argp->attrs.ia_size);
 
 	fh_copy(&resp->fh, &argp->fh);
-	nfserr = nfsd_setattr(rqstp, &resp->fh, &argp->attrs,0, (time_t)0);
+	nfserr = nfsd_setattr(rqstp, &resp->fh, &argp->attrs, 0, 0, 0);
 	return nfsd_return_attrs(nfserr, resp);
 }
 
@@ -326,7 +326,7 @@ nfsd_proc_create(struct svc_rqst *rqstp, struct nfsd_createargs *argp,
 		 */
 		attr->ia_valid &= ATTR_SIZE;
 		if (attr->ia_valid)
-			nfserr = nfsd_setattr(rqstp, newfhp, attr, 0, (time_t)0);
+			nfserr = nfsd_setattr(rqstp, newfhp, attr, 0, 0, 0);
 	}
 
 out_unlock:
fs/nfsd/vfs.c +55 -61

203 lines not on this page · raw message

fs/nfsd/vfs.h +1 -2
index 4b1de0a..5062afd 100644
@@ -41,7 +41,7 @@ __be32		 nfsd_lookup_dentry(struct svc_rqst *, struct svc_fh *,
 				const char *, unsigned int,
 				struct svc_export **, struct dentry **);
 __be32		nfsd_setattr(struct svc_rqst *, struct svc_fh *,
-				struct iattr *, int, time_t);
+				struct iattr *, int, time_t, int);
 int nfsd_mountpoint(struct dentry *, struct svc_export *);
 #ifdef CONFIG_NFSD_V4
 __be32          nfsd4_set_nfs4_acl(struct svc_rqst *, struct svc_fh *,
@@ -91,7 +91,6 @@ __be32		nfsd_statfs(struct svc_rqst *, struct svc_fh *,
 int		nfsd_notify_change(struct inode *, struct iattr *);
 __be32		nfsd_permission(struct svc_rqst *, struct svc_export *,
 				struct dentry *, int);
-int		nfsd_sync_dir(struct dentry *dp);
 
 #if defined(CONFIG_NFSD_V2_ACL) || defined(CONFIG_NFSD_V3_ACL)
 struct posix_acl *nfsd_get_posix_acl(struct svc_fh *, int);
include/linux/exportfs.h +6 -0
index dc12f41..9102ecf 100644
@@ -96,6 +96,7 @@ struct fid {
  * @fh_to_parent:   find the implied object's parent and get a dentry for it
  * @get_name:       find the name for a given inode in a given directory
  * @get_parent:     find the parent of a given directory
+ * @commit_metadata: commit metadata changes to stable storage
  *
  * See Documentation/filesystems/nfs/Exporting for details on how to use
  * this interface correctly.
@@ -137,6 +138,10 @@ struct fid {
  *    is also a directory.  In the event that it cannot be found, or storage
  *    space cannot be allocated, a %ERR_PTR should be returned.
  *
+ * commit_metadata:
+ *    @commit_metadata should commit metadata changes to stable storage.
+ *    Parent or child can be NULL.
+ *
  * Locking rules:
  *    get_parent is called with child->d_inode->i_mutex down
  *    get_name is not (which is possibly inconsistent)
@@ -152,6 +157,7 @@ struct export_operations {
 	int (*get_name)(struct dentry *parent, char *name,
 			struct dentry *child);
 	struct dentry * (*get_parent)(struct dentry *child);
+	int (*commit_metadata)(struct inode *parent, struct inode *child);
 };
 
 extern int exportfs_encode_fh(struct dentry *dentry, struct fid *fid,

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

author
#4 lore raw

PATCH 2/2 xfs_export_operations.commit_metadata

In reply to Ben Myers

This is the commit_metadata export operation for XFS, including the changes suggested by hch and dgc:

- Takes two two inodes instead of dentries and can assume the parent is always set.

- Uses xfs_lock_two_inodes for the ilock.

- Forces the log up to the larger lsn of parent and child.

- Uses XFS_LSN_CMP for lsn comparison.

- Doesn't force the log if nobody had a pincount.

---
 fs/xfs/linux-2.6/xfs_export.c |   43 +++++++++++++++++++++++++++++++++++++++++
 1 files changed, 43 insertions(+), 0 deletions(-)
fs/xfs/linux-2.6/xfs_export.c +43 -0
index 87b8cbd..d37ced1 100644
@@ -29,6 +29,7 @@
 #include "xfs_vnodeops.h"
 #include "xfs_bmap_btree.h"
 #include "xfs_inode.h"
+#include "xfs_inode_item.h"
 
 /*
  * Note that we only accept fileids which are long enough rather than allow
@@ -215,9 +216,51 @@ xfs_fs_get_parent(
 	return d_obtain_alias(VFS_I(cip));
 }
 
+STATIC int
+xfs_fs_nfs_commit_metadata(
+	struct inode		*parent,
+	struct inode		*child)
+{
+	struct xfs_inode	*dp = XFS_I(parent);
+	struct xfs_inode	*ip = NULL;
+	struct xfs_mount	*mp = dp->i_mount;
+	xfs_lsn_t		force_lsn = NULLCOMMITLSN;
+	int			error = 0;
+
+	if (child) {
+		ip = XFS_I(child);
+		xfs_lock_two_inodes(dp, ip, XFS_ILOCK_SHARED);
+	} else {
+		xfs_ilock(dp, XFS_ILOCK_SHARED);
+	}
+
+	if (xfs_ipincount(dp)) {
+		force_lsn = dp->i_itemp->ili_last_lsn;
+		if (ip && xfs_ipincount(ip) &&
+		    XFS_LSN_CMP(force_lsn, ip->i_itemp->ili_last_lsn) < 0) {
+			force_lsn = ip->i_itemp->ili_last_lsn;
+		}
+	} else if (ip && xfs_ipincount(ip)) {
+		force_lsn = ip->i_itemp->ili_last_lsn;
+	}
+
+	if (force_lsn != NULLCOMMITLSN) {
+		error = _xfs_log_force(mp, force_lsn,
+				XFS_LOG_FORCE | XFS_LOG_SYNC, NULL);
+	}
+
+	if (child)
+		xfs_iunlock(ip, XFS_ILOCK_SHARED);
+	if (parent)
+		xfs_iunlock(dp, XFS_ILOCK_SHARED);
+
+	return error;
+}
+
 const struct export_operations xfs_export_operations = {
 	.encode_fh		= xfs_fs_encode_fh,
 	.fh_to_dentry		= xfs_fs_fh_to_dentry,
 	.fh_to_parent		= xfs_fs_fh_to_parent,
 	.get_parent		= xfs_fs_get_parent,
+	.commit_metadata	= xfs_fs_nfs_commit_metadata,
 };

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#2 lore raw

Re: [PATCH 1/2] commit_metadata export operation replacing nfsd_sync_dir

In reply to Ben Myers

Generally this looks like a good change, but I do have some questions embedded below.

-Alex

Quoted 5 lines · Ben Myers

- Add commit_metadata export_operation to allow the underlying filesystem to decide how to sync parent and child inodes most efficiently.

- Usage of nfsd_sync_dir and write_inode_now has been replaced with the commit_metadata function that takes a svc_fh and optional dentry for the child.

In the cases you pass the child dentry, you will now be syncing things to disk in a slightly different way/order from before. I think this way is actually much better, but I thought I'd ask
anyway: are you sure that it's OK to sync them at the same time, rather than directory first, then child (or the other way around, depending on the case)?

Quoted 9 lines

- The commit_metadata function calls the commit_metadata export_op if it's there, or else falls back to sync_inode instead of fsync and write_inode_now because only metadata need be synched here.

- nfsd4_sync_rec_dir now uses vfs_fsync so that commit_metadata can be static

- Add a 'delay_commit' arg to nfsd_setattr so that callers of nfsd_create_setattr can commit parent and child together avoiding an extra sync.

I think this is a distinct enough change to warrant a separate patch from the commit_metadata feature.

Quoted 2 lines

. . .

fs/nfsd/nfs4recover.c out_no_tfm: · line 122

Quoted 11 lines

index 5a754f7..98fb98e 100644

fs/nfsd/nfs4recover.c +1 -3
@@ -119,9 +119,7 @@ out_no_tfm:
 static void
 nfsd4_sync_rec_dir(void)
 {
-	mutex_lock(&rec_dir.dentry->d_inode->i_mutex);
-	nfsd_sync_dir(rec_dir.dentry);
-	mutex_unlock(&rec_dir.dentry->d_inode->i_mutex);
+	vfs_fsync(NULL, rec_dir.dentry, 0);

Why do you no longer need to acquire the mutex here? I see it gets acquired during the ->fsync() call inside vfs_fsync_range(), but is there a reason that it needed to be held during the filemap_write_and_wait() (as is done in the nfsd_sync_dir() code) also?

. . .

fs/nfsd/vfs.c

Quoted 4 lines
fs/nfsd/vfs.c +0 -0
index ed024d3..97474fb 100644

. . .

fs/nfsd/vfs.c nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap, · line 453

Quoted 9 lines
fs/nfsd/vfs.c +2 -3
@@ -415,9 +449,8 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	}
 	if (size_change)
 		put_write_access(inode);
-	if (!err)
-		if (EX_ISSYNC(fhp->fh_export))
-			write_inode_now(inode, 1);
+	if (!err && !delay_commit)
+		err = nfserrno(commit_metadata(fhp, NULL));

Is this sufficient even if the size has changed?

fs/nfsd/vfs.c nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap, · line 456

Quoted 3 lines
fs/nfsd/vfs.c +0 -0
 out:
 	return err;
 

. . .

fs/nfsd/vfs.c nfsd_rename(struct svc_rqst *rqstp, struct svc_fh *ffhp, char *fname, int flen, · line 1769

Quoted 11 lines
fs/nfsd/vfs.c +3 -3
@@ -1766,10 +1763,10 @@ nfsd_rename(struct svc_rqst *rqstp, struct svc_fh *ffhp, char *fname, int flen,
 		goto out_dput_new;
 
 	host_err = vfs_rename(fdir, odentry, tdir, ndentry);
-	if (!host_err && EX_ISSYNC(tfhp->fh_export)) {
-		host_err = nfsd_sync_dir(tdentry);
+	if (!host_err) {
+		host_err = commit_metadata(tfhp, NULL);
 		if (!host_err)
-			host_err = nfsd_sync_dir(fdentry);
+			host_err = commit_metadata(ffhp, NULL);

It violates the spirit of the "parent" and "child" nature of its arguments, but it might be nice to commit both directories' metadata with the same call here.

. . .

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#5 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

In reply to Ben Myers
Quoted 5 lines · Ben Myers

This is the commit_metadata export operation for XFS, including the changes suggested by hch and dgc:

- Takes two two inodes instead of dentries and can assume the parent is always set.

I alluded to this in my review of the first patch. This could be changed considered in a more generic sense, "sync one or two inodes' metadata" rather than presupposing the two inodes have a parent/child relationship.

Quoted 7 lines

- Uses xfs_lock_two_inodes for the ilock.

- Forces the log up to the larger lsn of parent and child.

- Uses XFS_LSN_CMP for lsn comparison.

- Doesn't force the log if nobody had a pincount.

So if the log doesn't get forced, what causes the desired metadata sync expected as a result of this call? (Maybe this is a dumb question.)

Quoted 42 lines
---
 fs/xfs/linux-2.6/xfs_export.c |   43 +++++++++++++++++++++++++++++++++++++++++
 1 files changed, 43 insertions(+), 0 deletions(-)
fs/xfs/linux-2.6/xfs_export.c +22 -0
index 87b8cbd..d37ced1 100644
@@ -29,6 +29,7 @@
 #include "xfs_vnodeops.h"
 #include "xfs_bmap_btree.h"
 #include "xfs_inode.h"
+#include "xfs_inode_item.h"
 
 /*
  * Note that we only accept fileids which are long enough rather than allow
@@ -215,9 +216,51 @@ xfs_fs_get_parent(
 	return d_obtain_alias(VFS_I(cip));
 }
 
+STATIC int
+xfs_fs_nfs_commit_metadata(
+	struct inode		*parent,
+	struct inode		*child)
+{
+	struct xfs_inode	*dp = XFS_I(parent);
+	struct xfs_inode	*ip = NULL;
+	struct xfs_mount	*mp = dp->i_mount;
+	xfs_lsn_t		force_lsn = NULLCOMMITLSN;
+	int			error = 0;
+
+	if (child) {
+		ip = XFS_I(child);
+		xfs_lock_two_inodes(dp, ip, XFS_ILOCK_SHARED);
+	} else {
+		xfs_ilock(dp, XFS_ILOCK_SHARED);
+	}
+
+	if (xfs_ipincount(dp)) {
+		force_lsn = dp->i_itemp->ili_last_lsn;
+		if (ip && xfs_ipincount(ip) &&

fs/xfs/linux-2.6/xfs_export.c xfs_fs_get_parent( · line 249

Quoted 10 lines
fs/xfs/linux-2.6/xfs_export.c +10 -0
+		    XFS_LSN_CMP(force_lsn, ip->i_itemp->ili_last_lsn) < 0) {
+			force_lsn = ip->i_itemp->ili_last_lsn;
+		}
+	} else if (ip && xfs_ipincount(ip)) {
+		force_lsn = ip->i_itemp->ili_last_lsn;
+	}
+
+	if (force_lsn != NULLCOMMITLSN) {
+		error = _xfs_log_force(mp, force_lsn,
+				XFS_LOG_FORCE | XFS_LOG_SYNC, NULL);

You want this here:
error = xfs_log_force_lsn(mp, force_lsn, XFS_LOG_FORCE | XFS_LOG_SYNC);

Quoted 18 lines
fs/xfs/linux-2.6/xfs_export.c +11 -0
+	}
+
+	if (child)
+		xfs_iunlock(ip, XFS_ILOCK_SHARED);
+	if (parent)
+		xfs_iunlock(dp, XFS_ILOCK_SHARED);
+
+	return error;
+}
+
 const struct export_operations xfs_export_operations = {
 	.encode_fh		= xfs_fs_encode_fh,
 	.fh_to_dentry		= xfs_fs_fh_to_dentry,
 	.fh_to_parent		= xfs_fs_fh_to_parent,
 	.get_parent		= xfs_fs_get_parent,
+	.commit_metadata	= xfs_fs_nfs_commit_metadata,
 };
Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#3 lore raw

Re: [PATCH 1/2] commit_metadata export operation replacing nfsd_sync_dir

In reply to Alex Elder
Quoted 6 lines

In the cases you pass the child dentry, you will now be syncing things to disk in a slightly different way/order from before. I think this way is actually much better, but I thought I'd ask
anyway: are you sure that it's OK to sync them at the same time, rather than directory first, then child (or the other way around, depending on the case)?

The order is defined by when we commit the transactions, if we do the log force on the later one first we already write out the first transaction. Note that in any transaction filesystems the changes that create/mkdir/link/unlink do to parent and child will be in the same transaction anyway, which is kinda the point of adding the transactions to start with. Only for the case where we do a setattr in addition to the primary transaction we'll actually have another transaction to deal with and the above applies.

Quoted 4 lines
Quoted 4 lines

nfsd4_sync_rec_dir(void)
{

fs/nfsd/nfs4recover.c +0 -2
-	mutex_lock(&rec_dir.dentry->d_inode->i_mutex);
-	nfsd_sync_dir(rec_dir.dentry);
Quoted 8 lines

fs/nfsd/nfs4recover.c out_no_tfm: · line 122

Quoted 2 lines
fs/nfsd/nfs4recover.c +1 -1
-	mutex_unlock(&rec_dir.dentry->d_inode->i_mutex);
+	vfs_fsync(NULL, rec_dir.dentry, 0);

Why do you no longer need to acquire the mutex here? I see it gets acquired during the ->fsync() call inside vfs_fsync_range(), but is there a reason that it needed to be held during the filemap_write_and_wait() (as is done in the nfsd_sync_dir() code) also?

We don't need i_mutex for filemap_write_and_wait, it's just held because someone used the nfsd_sync_dir helper where it doesn't fit very well.

Quoted 11 lines

fs/nfsd/vfs.c nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap, · line 453

Quoted 9 lines
fs/nfsd/vfs.c +2 -3
@@ -415,9 +449,8 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	}
 	if (size_change)
 		put_write_access(inode);
-	if (!err)
-		if (EX_ISSYNC(fhp->fh_export))
-			write_inode_now(inode, 1);
+	if (!err && !delay_commit)
+		err = nfserrno(commit_metadata(fhp, NULL));

Is this sufficient even if the size has changed?

Yes.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#6 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

In reply to Alex Elder
Quoted 4 lines · Alex Elder
Quoted 3 lines · Ben Myers

This is the commit_metadata export operation for XFS, including the changes suggested by hch and dgc:

Quoted 8 lines
Quoted 2 lines

- Takes two two inodes instead of dentries and can assume the parent is always set.

I alluded to this in my review of the first patch. This could be changed considered in a more generic sense, "sync one or two inodes' metadata" rather than presupposing the two inodes have a parent/child relationship.

Or just implement my later suggestion to only pass one inode to the
method and cal in on both inodes. For the non-create case where
we have only one transaction to deal with fhr first call will take care of it and unpin the second inode by forcing the log buffer out. For the create case we need to make sure to call it on the child first so that we force out the setattr transaction which also forced out the
earlier one. This keeps the calling convention quite a bit simpler,
and also means we don't have to bother with locking two inodes or lsn comparisms.

Quoted 11 lines
Quoted 7 lines

- Uses xfs_lock_two_inodes for the ilock.

- Forces the log up to the larger lsn of parent and child.

- Uses XFS_LSN_CMP for lsn comparison.

- Doesn't force the log if nobody had a pincount.

So if the log doesn't get forced, what causes the desired metadata sync expected as a result of this call? (Maybe this is a dumb question.)

Inodes get pinned on transaction commit, and unpinned when the log I/O for that transaction completes. If the inode is not pinned this implies it has already been written to disk, e.g. because we're filling the log so fast that we need to write out more log buffers in that tiny window between the metada operation and the commit_metadata call.

Quoted 9 lines

fs/xfs/linux-2.6/xfs_export.c xfs_fs_get_parent( · line 249

Quoted 6 lines
fs/xfs/linux-2.6/xfs_export.c +6 -0
+		force_lsn = ip->i_itemp->ili_last_lsn;
+	}
+
+	if (force_lsn != NULLCOMMITLSN) {
+		error = _xfs_log_force(mp, force_lsn,
+				XFS_LOG_FORCE | XFS_LOG_SYNC, NULL);

You want this here:
error = xfs_log_force_lsn(mp, force_lsn, XFS_LOG_FORCE | XFS_LOG_SYNC);

In the XFS tree we do want either

	xfs_log_force_lsn(mp, force_lsn, XFS_LOG_SYNC);

or

	error = _xfs_log_force_lsn(mp, force_lsn, XFS_LOG_SYNC, NULL);

if we care enough about the returned error. But Ben is working against the NFS tree which doesn't have that change yet.

We can deal with that by either commiting the old variant to the nfs tree and then leaving sending Stephen a patch to fix it up in -next, or just not apply the xfs commit_metadata implementation yet, and wait for it until both the xfs and nfs trees have hit mainline.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#7 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

Quoted 14 lines · Christoph Hellwig
Quoted 4 lines · Alex Elder
Quoted 3 lines · Ben Myers

This is the commit_metadata export operation for XFS, including the changes suggested by hch and dgc:

Quoted 8 lines
Quoted 2 lines

- Takes two two inodes instead of dentries and can assume the parent is always set.

I alluded to this in my review of the first patch. This could be changed considered in a more generic sense, "sync one or two inodes' metadata" rather than presupposing the two inodes have a parent/child relationship.

Quoted 7 lines

Or just implement my later suggestion to only pass one inode to the
method and cal in on both inodes. For the non-create case where
we have only one transaction to deal with fhr first call will take care of it and unpin the second inode by forcing the log buffer out. For the create case we need to make sure to call it on the child first so that we force out the setattr transaction which also forced out the earlier one.

I chose not implement that suggestion because I prefer not to rely upon the coincidence that the child be modified last and synced first in knfsd. It is better that the intent of the patch be clear, and that filesystems other than xfs also have enough information to sort out how to sync more efficiently. knfsd shouldn't be forced to sync in a specific order just because that's what works best for xfs. Or, mebbe it should and I'm being thick. ;)

Alex's suggestion that ->commit_metadata be more generic about the relationship of the two inodes seems reasonable, not sure what that means for vfs.c:commit_metadata. This will help with the rename case.

Quoted 3 lines

This keeps the calling convention quite a bit simpler,
and also means we don't have to bother with locking two inodes or lsn comparisms.

Don't need the ilock to check pincount?

Quoted 13 lines
Quoted 5 lines

fs/xfs/linux-2.6/xfs_export.c xfs_fs_get_parent( · line 249

Quoted 2 lines
fs/xfs/linux-2.6/xfs_export.c +2 -0
+		error = _xfs_log_force(mp, force_lsn,
+				XFS_LOG_FORCE | XFS_LOG_SYNC, NULL);

You want this here:
error = xfs_log_force_lsn(mp, force_lsn, XFS_LOG_FORCE | XFS_LOG_SYNC);

In the XFS tree we do want either

	xfs_log_force_lsn(mp, force_lsn, XFS_LOG_SYNC);

or

	error = _xfs_log_force_lsn(mp, force_lsn, XFS_LOG_SYNC, NULL);
Quoted 7 lines

if we care enough about the returned error. But Ben is working against the NFS tree which doesn't have that change yet.

We can deal with that by either commiting the old variant to the nfs tree and then leaving sending Stephen a patch to fix it up in -next, or just not apply the xfs commit_metadata implementation yet, and wait for it until both the xfs and nfs trees have hit mainline.

Yeah. I don't know who Stephen is.

Thanks,
Ben

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#8 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

Quoted 8 lines · On Fri, Feb 12, 2010 at 01:56:47PM -0600, [email protected] wrote:
Quoted 7 lines · Christoph Hellwig

if we care enough about the returned error. But Ben is working against the NFS tree which doesn't have that change yet.

We can deal with that by either commiting the old variant to the nfs tree and then leaving sending Stephen a patch to fix it up in -next, or just not apply the xfs commit_metadata implementation yet, and wait for it until both the xfs and nfs trees have hit mainline.

Assuming the xfs tree's something stable, I could also just go ahead and pull it into the nfsd tree.

(Or we could do it the other way around--as long as when the merge window opens we try to send the pull requests in the same order.)

--b.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#10 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

In reply to [email protected]
Quoted 3 lines · On Fri, Feb 12, 2010 at 01:56:47PM -0600, [email protected] wrote:

I chose not implement that suggestion because I prefer not to rely upon the coincidence that the child be modified last and synced first in knfsd.

It does ot rely on that coincidence for correctness, just for a small performance optimization.

Quoted 5 lines

It is better that the intent of the patch be clear, and that filesystems other than xfs also have enough information to sort out how to sync more efficiently. knfsd shouldn't be forced to sync in a specific order just because that's what works best for xfs. Or, mebbe it should and I'm being thick. ;)

The order of parent and child only happens in the create case where NFS does a separate ->setattr call. For all the operations handled by the filesystem parent and child are in the same transaction for every transactional filesystem.

Quoted 5 lines
Quoted 3 lines

This keeps the calling convention quite a bit simpler,
and also means we don't have to bother with locking two inodes or lsn comparisms.

Don't need the ilock to check pincount?

Ah sorry, that should have read not bothering with locking two inodes at the same time, which always is a bit troublesome. We do need to lock the inode for looking at the pincount.

Quoted 6 lines
Quoted 4 lines

We can deal with that by either commiting the old variant to the nfs tree and then leaving sending Stephen a patch to fix it up in -next, or just not apply the xfs commit_metadata implementation yet, and wait for it until both the xfs and nfs trees have hit mainline.

Yeah. I don't know who Stephen is.

Stephen Rothwell is the maintainer of the linux-next tree.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#9 lore raw

Re: [PATCH 2/2] xfs_export_operations.commit_metadata

In reply to J. Bruce Fields
Quoted 5 lines · J. Bruce Fields

Assuming the xfs tree's something stable, I could also just go ahead and pull it into the nfsd tree.

(Or we could do it the other way around--as long as when the merge window opens we try to send the pull requests in the same order.)

The xfs tree is supposed to be stable and we only had very few incidents where a patch had to be pulled again. But it's also a quite large set of changes.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs