fs: clean up handling of i_version counter

This is a pretty big function for inlining. Move it to being non-inlined.

Reviewed-by NeilBrown <[email protected]>

Signed-off-by Jeff Layton <[email protected]>

---
fs/nfsd/nfsfh.c | 27 +++++++++++++++++++++++++++ fs/nfsd/nfsfh.h | 29 +----------------------------

2 files changed, 28 insertions(+), 28 deletions(-)

fs/nfsd/nfsfh.c +27 -0

index d73434200df9..7030d9209903 100644
@@ -748,3 +748,30 @@ enum fsid_source fsid_source(const struct svc_fh *fhp)
 		return FSIDSOURCE_UUID;
 	return FSIDSOURCE_DEV;
 }
+
+/*
+ * We could use i_version alone as the change attribute.  However,
+ * i_version can go backwards after a reboot.  On its own that doesn't
+ * necessarily cause a problem, but if i_version goes backwards and then
+ * is incremented again it could reuse a value that was previously used
+ * before boot, and a client who queried the two values might
+ * incorrectly assume nothing changed.
+ *
+ * By using both ctime and the i_version counter we guarantee that as
+ * long as time doesn't go backwards we never reuse an old value.
+ */
+u64 nfsd4_change_attribute(struct kstat *stat, struct inode *inode)
+{
+	if (inode->i_sb->s_export_op->fetch_iversion)
+		return inode->i_sb->s_export_op->fetch_iversion(inode);
+	else if (IS_I_VERSION(inode)) {
+		u64 chattr;
+
+		chattr =  stat->ctime.tv_sec;
+		chattr <<= 30;
+		chattr += stat->ctime.tv_nsec;
+		chattr += inode_query_iversion(inode);
+		return chattr;
+	} else
+		return time_to_chattr(&stat->ctime);
+}

fs/nfsd/nfsfh.h +1 -28

index c3ae6414fc5c..4c223a7a91d4 100644
@@ -291,34 +291,7 @@ static inline void fh_clear_pre_post_attrs(struct svc_fh *fhp)
 	fhp->fh_pre_saved = false;
 }
 
-/*
- * We could use i_version alone as the change attribute.  However,
- * i_version can go backwards after a reboot.  On its own that doesn't
- * necessarily cause a problem, but if i_version goes backwards and then
- * is incremented again it could reuse a value that was previously used
- * before boot, and a client who queried the two values might
- * incorrectly assume nothing changed.
- *
- * By using both ctime and the i_version counter we guarantee that as
- * long as time doesn't go backwards we never reuse an old value.
- */
-static inline u64 nfsd4_change_attribute(struct kstat *stat,
-					 struct inode *inode)
-{
-	if (inode->i_sb->s_export_op->fetch_iversion)
-		return inode->i_sb->s_export_op->fetch_iversion(inode);
-	else if (IS_I_VERSION(inode)) {
-		u64 chattr;
-
-		chattr =  stat->ctime.tv_sec;
-		chattr <<= 30;
-		chattr += stat->ctime.tv_nsec;
-		chattr += inode_query_iversion(inode);
-		return chattr;
-	} else
-		return time_to_chattr(&stat->ctime);
-}
-
+u64 nfsd4_change_attribute(struct kstat *stat, struct inode *inode);
 extern void fh_fill_pre_attrs(struct svc_fh *fhp);
 extern void fh_fill_post_attrs(struct svc_fh *fhp);
 extern void fh_fill_both_attrs(struct svc_fh *fhp);

Signature · 1 line

2.37.3

Now that we can call into vfs_getattr to get the i_version field, use that facility to fetch it instead of doing it in nfsd4_change_attribute.

Neil also pointed out recently that IS_I_VERSION directory operations are always logged, and so we only need to mitigate the rollback problem on regular files. Also, we don't need to factor in the ctime when reexporting NFS or Ceph.

Set the STATX_VERSION (and BTIME) bits in the request when we're dealing with a v4 request. Then, instead of looking at IS_I_VERSION when generating the change attr, look at the result mask and only use it if STATX_VERSION is set.

Change nfsd4_change_attribute to only factor in the ctime if it's a regular file and the fs doesn't advertise STATX_ATTR_VERSION_MONOTONIC.

Reviewed-by NeilBrown <[email protected]>

Signed-off-by Jeff Layton <[email protected]>

---

fs/nfsd/nfs4xdr.c |  4 +++-
fs/nfsd/nfsfh.c   | 53 +++++++++++++++++++++++++++++++----------------
fs/nfsd/vfs.h     |  7 ++++++-
3 files changed, 44 insertions(+), 20 deletions(-)

fs/nfsd/nfs4xdr.c +3 -1

index bcfeb1a922c0..c19b6b00b620 100644
@@ -2906,7 +2906,9 @@ nfsd4_encode_fattr(struct xdr_stream *xdr, struct svc_fh *fhp,
 			goto out;
 	}
 
-	err = vfs_getattr(&path, &stat, STATX_BASIC_STATS, AT_STATX_SYNC_AS_STAT);
+	err = vfs_getattr(&path, &stat,
+			  STATX_BASIC_STATS | STATX_BTIME | STATX_VERSION,
+			  AT_STATX_SYNC_AS_STAT);
 	if (err)
 		goto out_nfserr;
 	if (!(stat.result_mask & STATX_BTIME))

fs/nfsd/nfsfh.c +35 -18

index 7030d9209903..21b64ac97a06 100644
@@ -628,6 +628,10 @@ void fh_fill_pre_attrs(struct svc_fh *fhp)
 		stat.mtime = inode->i_mtime;
 		stat.ctime = inode->i_ctime;
 		stat.size  = inode->i_size;
+		if (v4 && IS_I_VERSION(inode)) {
+			stat.version = inode_query_iversion(inode);
+			stat.result_mask |= STATX_VERSION;
+		}
 	}
 	if (v4)
 		fhp->fh_pre_change = nfsd4_change_attribute(&stat, inode);

@@ -659,6 +663,10 @@ void fh_fill_post_attrs(struct svc_fh *fhp)
 	if (err) {
 		fhp->fh_post_saved = false;
 		fhp->fh_post_attr.ctime = inode->i_ctime;
+		if (v4 && IS_I_VERSION(inode)) {
+			fhp->fh_post_attr.version = inode_query_iversion(inode);
+			fhp->fh_post_attr.result_mask |= STATX_VERSION;
+		}
 	} else
 		fhp->fh_post_saved = true;
 	if (v4)

@@ -750,28 +758,37 @@ enum fsid_source fsid_source(const struct svc_fh *fhp)
 }
 
 /*
- * We could use i_version alone as the change attribute.  However,
- * i_version can go backwards after a reboot.  On its own that doesn't
- * necessarily cause a problem, but if i_version goes backwards and then
- * is incremented again it could reuse a value that was previously used
- * before boot, and a client who queried the two values might
- * incorrectly assume nothing changed.
+ * We could use i_version alone as the change attribute.  However, i_version
+ * can go backwards on a regular file after an unclean shutdown.  On its own
+ * that doesn't necessarily cause a problem, but if i_version goes backwards
+ * and then is incremented again it could reuse a value that was previously
+ * used before boot, and a client who queried the two values might incorrectly
+ * assume nothing changed.
+ *
+ * By using both ctime and the i_version counter we guarantee that as long as
+ * time doesn't go backwards we never reuse an old value. If the filesystem
+ * advertises STATX_ATTR_VERSION_MONOTONIC, then this mitigation is not needed.
 *
- * By using both ctime and the i_version counter we guarantee that as
- * long as time doesn't go backwards we never reuse an old value.
+ * We only need to do this for regular files as well. For directories, we
+ * assume that the new change attr is always logged to stable storage in some
+ * fashion before the results can be seen.
 */
 u64 nfsd4_change_attribute(struct kstat *stat, struct inode *inode)
 {
+	u64 chattr;
+
 	if (inode->i_sb->s_export_op->fetch_iversion)
 		return inode->i_sb->s_export_op->fetch_iversion(inode);
-	else if (IS_I_VERSION(inode)) {
-		u64 chattr;
-
-		chattr =  stat->ctime.tv_sec;
-		chattr <<= 30;
-		chattr += stat->ctime.tv_nsec;
-		chattr += inode_query_iversion(inode);
-		return chattr;
-	} else
-		return time_to_chattr(&stat->ctime);
+	if (stat->result_mask & STATX_VERSION) {
+		chattr = stat->version;
+
+		if (S_ISREG(inode->i_mode) &&
+		    !(stat->attributes & STATX_ATTR_VERSION_MONOTONIC)) {
+			chattr += (u64)stat->ctime.tv_sec << 30;
+			chattr += stat->ctime.tv_nsec;
+		}
+	} else {
+		chattr = time_to_chattr(&stat->ctime);
+	}
+	return chattr;
 }

fs/nfsd/vfs.h +6 -1

index 120521bc7b24..c98e13ec37b2 100644
@@ -168,9 +168,14 @@ static inline void fh_drop_write(struct svc_fh *fh)
 
 static inline __be32 fh_getattr(const struct svc_fh *fh, struct kstat *stat)
 {
+	u32 request_mask = STATX_BASIC_STATS;
 	struct path p = {.mnt = fh->fh_export->ex_path.mnt,
 			 .dentry = fh->fh_dentry};
-	return nfserrno(vfs_getattr(&p, stat, STATX_BASIC_STATS,
+
+	if (fh->fh_maxsize == NFS4_FHSIZE)
+		request_mask |= (STATX_BTIME | STATX_VERSION);
+
+	return nfserrno(vfs_getattr(&p, stat, request_mask,
 				    AT_STATX_SYNC_AS_STAT));
 }
 

Signature · 1 line

2.37.3

Now that the i_version counter is reported in struct kstat, there is no need for this export operation.

Reviewed-by NeilBrown <[email protected]>

Signed-off-by Jeff Layton <[email protected]>

---

fs/nfs/export.c          | 7 -------
fs/nfsd/nfsfh.c          | 2 --
include/linux/exportfs.h | 1 -
3 files changed, 10 deletions(-)

fs/nfs/export.c +0 -7

index 01596f2d0a1e..1a9d5aa51dfb 100644
@@ -145,17 +145,10 @@ nfs_get_parent(struct dentry *dentry)
 	return parent;
 }
 
-static u64 nfs_fetch_iversion(struct inode *inode)
-{
-	nfs_revalidate_inode(inode, NFS_INO_INVALID_CHANGE);
-	return inode_peek_iversion_raw(inode);
-}
-
 const struct export_operations nfs_export_ops = {
 	.encode_fh = nfs_encode_fh,
 	.fh_to_dentry = nfs_fh_to_dentry,
 	.get_parent = nfs_get_parent,
-	.fetch_iversion = nfs_fetch_iversion,
 	.flags = EXPORT_OP_NOWCC|EXPORT_OP_NOSUBTREECHK|
 		EXPORT_OP_CLOSE_BEFORE_UNLINK|EXPORT_OP_REMOTE_FS|
 		EXPORT_OP_NOATOMIC_ATTR,

fs/nfsd/nfsfh.c +0 -2

index 21b64ac97a06..9c1f697ffc72 100644
@@ -777,8 +777,6 @@ u64 nfsd4_change_attribute(struct kstat *stat, struct inode *inode)
 {
 	u64 chattr;
 
-	if (inode->i_sb->s_export_op->fetch_iversion)
-		return inode->i_sb->s_export_op->fetch_iversion(inode);
 	if (stat->result_mask & STATX_VERSION) {
 		chattr = stat->version;
 

include/linux/exportfs.h +0 -1

index fe848901fcc3..9f4d4bcbf251 100644
@@ -213,7 +213,6 @@ struct export_operations {
 			  bool write, u32 *device_generation);
 	int (*commit_blocks)(struct inode *inode, struct iomap *iomaps,
 			     int nr_iomaps, struct iattr *iattr);
-	u64 (*fetch_iversion)(struct inode *);
 #define	EXPORT_OP_NOWCC			(0x1) /* don't collect v3 wcc data */
 #define	EXPORT_OP_NOSUBTREECHK		(0x2) /* no subtree checking */
 #define	EXPORT_OP_CLOSE_BEFORE_UNLINK	(0x4) /* close files before unlink */

Signature · 1 line

2.37.3

From: Jeff Layton <[email protected]>

Claim one of the spare fields in struct statx to hold a 64-bit inode version attribute. When userland requests STATX_VERSION, copy the value from the kstat struct there, and stop masking off STATX_ATTR_VERSION_MONOTONIC.

Update the test-statx sample program to output the change attr and MountId.

Reviewed-by NeilBrown <[email protected]>

Signed-off-by Jeff Layton <[email protected]>

---
 fs/stat.c                 | 12 +++---------
 include/linux/stat.h      |  9 ---------
 include/uapi/linux/stat.h |  6 ++++--
 samples/vfs/test-statx.c  |  8 ++++++--
 4 files changed, 13 insertions(+), 22 deletions(-)

Posting this as an RFC as we're still trying to sort out what semantics we want to present to userland. In particular, this patch leaves the problem of crash resilience in to userland applications on filesystems that don't report as MONOTONIC.

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

Quoted 6 lines · Jeff Layton

From: Jeff Layton <[email protected]>

Claim one of the spare fields in struct statx to hold a 64-bit inode version attribute. When userland requests STATX_VERSION, copy the value from the kstat struct there, and stop masking off STATX_ATTR_VERSION_MONOTONIC.

Can we please make the name more sepcific than "version"? It's way too generic and - we already have userspace facing "version" fields for inodes that refer to the on-disk format version exposed in various UAPIs. It's common for UAPI structures used for file operations to have a "version" field that refers to the *UAPI structure version* rather than file metadata or data being retrieved from the file in question.

The need for an explanatory comment like this:

include/uapi/linux/stat.h struct statx { · line 130

Quoted 1 line

include/uapi/linux/stat.h +1 -0

+	__u64	stx_version; /* Inode change attribute */

demonstrates it is badly named. If you want it known as an inode change attribute, then don't name the variable "version". In reality, it really needs to be an opaque cookie, not something applications need to decode directly to make sense of.

Quoted 12 lines

Update the test-statx sample program to output the change attr and MountId.

Reviewed-by NeilBrown <[email protected]>

Signed-off-by Jeff Layton <[email protected]>

---
 fs/stat.c                 | 12 +++---------
 include/linux/stat.h      |  9 ---------
 include/uapi/linux/stat.h |  6 ++++--
 samples/vfs/test-statx.c  |  8 ++++++--
 4 files changed, 13 insertions(+), 22 deletions(-)

Quoted 4 lines

Posting this as an RFC as we're still trying to sort out what semantics we want to present to userland. In particular, this patch leaves the problem of crash resilience in to userland applications on filesystems that don't report as MONOTONIC.

Firstly, if userspace wants to use the change attribute, they are going to have to detect crashes themselves anyway because no fs in the kernel can set the MONOTONIC flag right now and it may be years before kernels/filesystems actually support it in production systems.

But more fundamentally, I think this monotonic increase guarantee is completely broken by the presence of snapshots and snapshot rollbacks. If you change something, then a while later decide it broke (e.g. a production system upgrade went awry) and you roll back the filesystem to the pre-upgrade snapshot, then all the change counters and m/ctimes are guaranteed to go backwards because they will revert to the snapshot values. Maybe the filesystem can bump some internal counter for the snapshot when the revert happens, but until that is implemented, filesystems that support snapshots and rollback can't assert MONOTONIC.

And that's worse for other filesystems, because if you put them on dm-thinp and roll them back, they are completely unaware of the fact that a rollback happened and there's *nothing* the filesystem can do about this. Indeed, snapshots are suppose to be done on clean filesystems so snapshot images don't require journal recovery, so any crash detection put in the filesystem recovery code to guarantee MONOTONIC behaviour will be soundly defeated by such block device snapshot rollbacks.

Hence I think MONOTONIC is completely unworkable for most existing filesystems because snapshots and rollbacks completely break the underlying assumption MONOTONIC relies on: that filesystem modifications always move forwards in both the time and modification order dimensions....

This means that monotonicity is probably not acheivable by any existing filesystem and so should not ever be mentioned in the UAPI. I think userspace semantics can be simplified down to "if the change cookie does not match exactly, caches are invalid" combined with "applications are responsible for detecting temporal discontiguities in filesystem presentation at start up (e.g. after a crash, unclean shutdown, restoration from backup, snapshot rollback, etc) for persistent cache invalidation purposes"....

Quoted 5 lines

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Quoted 75 lines · Jeff Layton

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 7 lines

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Note that I didn't propose to increment by offset on every reboot or mount. I have proposed that inode_maybe_inc_iversion() would not increment i_version by 1 (in fact 1 << I_VERSION_QUERIED_SHIFT) but rather by P (or P << I_VERSION_QUERIED_SHIFT) where P is a suitable number randomly selected on filesystem mount.

This will not cause cache invalidation after a clean unmount + remount. It will cause cache invalidation after a crash, snapshot rollback etc., only for inodes where i_version changed. If P is suitably selected (e.g. as being a prime), then the chances of collisions (even after a snapshot rollback) are very low (on the order of 2^(-50) if my piece of envelope calculations are right).

So this should nicely deal with all the problems we've spotted so far. But I may be missing something...

Quoted 97 lines · Jan Kara

Quoted 75 lines · Jeff Layton

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 7 lines

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Note that I didn't propose to increment by offset on every reboot or mount. I have proposed that inode_maybe_inc_iversion() would not increment i_version by 1 (in fact 1 << I_VERSION_QUERIED_SHIFT) but rather by P (or P << I_VERSION_QUERIED_SHIFT) where P is a suitable number randomly selected on filesystem mount.

This will not cause cache invalidation after a clean unmount + remount. It will cause cache invalidation after a crash, snapshot rollback etc., only for inodes where i_version changed. If P is suitably selected (e.g. as being a prime), then the chances of collisions (even after a snapshot rollback) are very low (on the order of 2^(-50) if my piece of envelope calculations are right).

Quoted 2 lines

So this should nicely deal with all the problems we've spotted so far. But I may be missing something...

Got it! That makes a lot more sense. Thinking about this some more...

What sort of range for P would be suitable?

Every increment would need to be by (shifted) P, so we can't choose too large a number. Queries are pretty rare vs. writes though, so that mitigates the issue somewhat.

Quoted 102 lines · Jeff Layton

Quoted 97 lines · Jan Kara

Quoted 75 lines · Jeff Layton

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 7 lines

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Note that I didn't propose to increment by offset on every reboot or mount. I have proposed that inode_maybe_inc_iversion() would not increment i_version by 1 (in fact 1 << I_VERSION_QUERIED_SHIFT) but rather by P (or P << I_VERSION_QUERIED_SHIFT) where P is a suitable number randomly selected on filesystem mount.

This will not cause cache invalidation after a clean unmount + remount. It will cause cache invalidation after a crash, snapshot rollback etc., only for inodes where i_version changed. If P is suitably selected (e.g. as being a prime), then the chances of collisions (even after a snapshot rollback) are very low (on the order of 2^(-50) if my piece of envelope calculations are right).

Quoted 2 lines

So this should nicely deal with all the problems we've spotted so far. But I may be missing something...

Quoted 7 lines

Got it! That makes a lot more sense. Thinking about this some more...

What sort of range for P would be suitable?

Every increment would need to be by (shifted) P, so we can't choose too large a number. Queries are pretty rare vs. writes though, so that mitigates the issue somewhat.

Well, I agree that for large P the counter would wrap earlier. But is that a problem? Note that if P is a prime (indivisible by 2 is enough), then the counter would get to already used value still only after 2^63 steps. Thus if we give up monotonicity and just treat the counter as an opaque cookie, we do not have to care about wrapping.

Sure given different P is selected for each mount the wrapping argument does not hold 100% but here comes the advantage of primes - if you have two different primes P and Q, then a collision means that k*P mod 2^63 = l*Q mod 2^63 and that holds for exactly one pair k,l from 1..2^63 range. So the chances of early collision even after selecting a different prime on each mount are *very* low.

So I think we should select from a relatively large set of primes so that the chance of randomly selecting the same prime (and thus reissuing the same change attr for different inode state sometime later) are small.

Quoted 117 lines · Jan Kara

Quoted 102 lines · Jeff Layton

Quoted 97 lines · Jan Kara

Quoted 75 lines · Jeff Layton

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 7 lines

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Note that I didn't propose to increment by offset on every reboot or mount. I have proposed that inode_maybe_inc_iversion() would not increment i_version by 1 (in fact 1 << I_VERSION_QUERIED_SHIFT) but rather by P (or P << I_VERSION_QUERIED_SHIFT) where P is a suitable number randomly selected on filesystem mount.

This will not cause cache invalidation after a clean unmount + remount. It will cause cache invalidation after a crash, snapshot rollback etc., only for inodes where i_version changed. If P is suitably selected (e.g. as being a prime), then the chances of collisions (even after a snapshot rollback) are very low (on the order of 2^(-50) if my piece of envelope calculations are right).

Quoted 2 lines

So this should nicely deal with all the problems we've spotted so far. But I may be missing something...

Quoted 7 lines

Got it! That makes a lot more sense. Thinking about this some more...

What sort of range for P would be suitable?

Every increment would need to be by (shifted) P, so we can't choose too large a number. Queries are pretty rare vs. writes though, so that mitigates the issue somewhat.

Well, I agree that for large P the counter would wrap earlier. But is that a problem? Note that if P is a prime (indivisible by 2 is enough), then the counter would get to already used value still only after 2^63 steps. Thus if we give up monotonicity and just treat the counter as an opaque cookie, we do not have to care about wrapping.

Quoted 6 lines

Sure given different P is selected for each mount the wrapping argument does not hold 100% but here comes the advantage of primes - if you have two different primes P and Q, then a collision means that k*P mod 2^63 = l*Q mod 2^63 and that holds for exactly one pair k,l from 1..2^63 range. So the chances of early collision even after selecting a different prime on each mount are *very* low.

I think we'll have to start avoiding 1 as a value for P if we do this, but the rest makes sense. I like this idea, Jan!

Quoted 4 lines

So I think we should select from a relatively large set of primes so that the chance of randomly selecting the same prime (and thus reissuing the same change attr for different inode state sometime later) are small.

Monotonicity allows you to discard "old" attr updates. For instance, sometimes a NFS GETATTR response may be delayed for various reasons. If the client sees a change attr that is provably older than one it has already seen, it can discard the update. So, there is value in servers advertising that property, and NFSv4.2 has a way to do that.

The Linux NFS client (at least) uses the same trick we do with jiffies to handle wrapping for MONOTONIC values. We should be able to advertise MONOTONIC as long as the client isn't comparing values that are more than ~2^62 apart.

Once we start talking about applications storing these values for incremental backups, then the time between checks could be very long.

So, I think we don't want _too_ large a value for P. The big question is how many individual change attr increments do we need to account for?

We have 64 bits total (it's an atomic64_t). We consume the lowest bit for the QUERIED flag. That leaves us 63 bits of counter (currently). When we increment by a larger value, we're effectively decreasing the size of the counter.

Let's assume a worst case of one increment per microsecond, interleaved by queries (so that they have to be real increments). 2^48 microseconds is close to 9 years.

That leaves 15 bits for the P, which is primes from 3..32749. Is that a large enough pool of prime numbers?

It looks like the kernel already has some infrastructure for handling primes in lib/math/prime_numbers.c. We could just select a global P value to use on every reboot, or just have filesystems set their own (maybe in a new field in the superblock?)

Quoted 125 lines · Jeff Layton

Quoted 117 lines · Jan Kara

Quoted 102 lines · Jeff Layton

Quoted 97 lines · Jan Kara

Quoted 75 lines · Jeff Layton

Quoted 53 lines · Jan Kara

Quoted 43 lines · Jeff Layton

Quoted 20 lines · Dave Chinner

Quoted 5 lines · Jeff Layton

Trond is of the opinion that monotonicity is a hard requirement, and that we should not allow filesystems that can't provide that quality to report STATX_VERSION at all. His rationale is that one of the main uses for this is for backup applications, and for those a counter that could go backward is worse than useless.

From the perspective of a backup program doing incremental backups, an inode with a change counter that has a different value to the current backup inventory means the file contains different information than what the current backup inventory holds. Again, snapshots, rollbacks, etc.

Therefore, regardless of whether the change counter has gone forwards or backwards, the backup program needs to back up this current version of the file in this backup because it is different to the inventory copy. Hence if the backup program fails to back it up, it will not be creating an exact backup of the user's data at the point in time the backup is run...

Quoted 3 lines

Hence I don't see that MONOTONIC is a requirement for backup programs - they really do have to be able to handle filesystems that have modifications that move backwards in time as well as forwards...

Rolling backward is not a problem in and of itself. The big issue is that after a crash, we can end up with a change attr seen before the crash that is now associated with a completely different inode state.

The scenario is something like:

- Change attr for an empty file starts at 1

- Write "A" to file, change attr goes to 2

- Read and statx happens (client sees "A" with change attr 2)

- Crash (before last change is logged to disk)

- Machine reboots, inode is empty, change attr back to 1

- Write "B" to file, change attr goes to 2

Quoted 8 lines

- Client stat's file, sees change attr 2 and assumes its cache is correct when it isn't (should be "B" not "A" now).

The real danger comes not from the thing going backward, but the fact that it can march forward again after going backward, and then the client can see two different inode states associated with the same change attr value. Jumping all the change attributes forward by a significant amount after a crash should avoid this issue.

Quoted 7 lines

As Dave pointed out, the problem with change attr having the same value for a different inode state (after going backwards) holds not only for the crashes but also for restore from backups, fs snapshots, device snapshots etc. So relying on change attr only looks a bit fragile. It works for the common case but the edge cases are awkward and there's no easy way to detect you are in the edge case.

This is true. In fact in the snapshot case you can't even rely on doing anything at reboot since you won't necessarily need to reboot to make it roll backward.

Whether that obviates the use of this value altogether, I'm not sure.

Quoted 6 lines

So I think any implementation caring about data integrity would have to include something like ctime into the picture anyway. Or we could just completely give up any idea of monotonicity and on each mount select random prime P < 2^64 and instead of doing inc when advancing the change attribute, we'd advance it by P. That makes collisions after restore / crash fairly unlikely.

Quoted 7 lines

Part of the goal (at least for NFS) is to avoid unnecessary cache invalidations.

If we just increment it by a particular offset on every reboot, then every time the server reboots, the clients will invalidate all of their cached inodes, and proceed to hammer the server with READ calls just as it's having to populate its own caches from disk.

Note that I didn't propose to increment by offset on every reboot or mount. I have proposed that inode_maybe_inc_iversion() would not increment i_version by 1 (in fact 1 << I_VERSION_QUERIED_SHIFT) but rather by P (or P << I_VERSION_QUERIED_SHIFT) where P is a suitable number randomly selected on filesystem mount.

This will not cause cache invalidation after a clean unmount + remount. It will cause cache invalidation after a crash, snapshot rollback etc., only for inodes where i_version changed. If P is suitably selected (e.g. as being a prime), then the chances of collisions (even after a snapshot rollback) are very low (on the order of 2^(-50) if my piece of envelope calculations are right).

Quoted 2 lines

So this should nicely deal with all the problems we've spotted so far. But I may be missing something...

Quoted 7 lines

Got it! That makes a lot more sense. Thinking about this some more...

What sort of range for P would be suitable?

Every increment would need to be by (shifted) P, so we can't choose too large a number. Queries are pretty rare vs. writes though, so that mitigates the issue somewhat.

Well, I agree that for large P the counter would wrap earlier. But is that a problem? Note that if P is a prime (indivisible by 2 is enough), then the counter would get to already used value still only after 2^63 steps. Thus if we give up monotonicity and just treat the counter as an opaque cookie, we do not have to care about wrapping.

Quoted 6 lines

Sure given different P is selected for each mount the wrapping argument does not hold 100% but here comes the advantage of primes - if you have two different primes P and Q, then a collision means that k*P mod 2^63 = l*Q mod 2^63 and that holds for exactly one pair k,l from 1..2^63 range. So the chances of early collision even after selecting a different prime on each mount are *very* low.

Quoted 2 lines

I think we'll have to start avoiding 1 as a value for P if we do this, but the rest makes sense. I like this idea, Jan!

Yes, 1 is kind of special so we should better avoid it in this scheme. Especially if we're going to select only smaller primes.

Quoted 19 lines

Quoted 4 lines

So I think we should select from a relatively large set of primes so that the chance of randomly selecting the same prime (and thus reissuing the same change attr for different inode state sometime later) are small.

Monotonicity allows you to discard "old" attr updates. For instance, sometimes a NFS GETATTR response may be delayed for various reasons. If the client sees a change attr that is provably older than one it has already seen, it can discard the update. So, there is value in servers advertising that property, and NFSv4.2 has a way to do that.

The Linux NFS client (at least) uses the same trick we do with jiffies to handle wrapping for MONOTONIC values. We should be able to advertise MONOTONIC as long as the client isn't comparing values that are more than ~2^62 apart.

Once we start talking about applications storing these values for incremental backups, then the time between checks could be very long.

Quoted 7 lines

So, I think we don't want _too_ large a value for P. The big question is how many individual change attr increments do we need to account for?

We have 64 bits total (it's an atomic64_t). We consume the lowest bit for the QUERIED flag. That leaves us 63 bits of counter (currently). When we increment by a larger value, we're effectively decreasing the size of the counter.

Yes, the larger value of P we take the sooner it will wrap which defeats comparisons attempting to establish any ordering of change cookie values.

Quoted 6 lines

Let's assume a worst case of one increment per microsecond, interleaved by queries (so that they have to be real increments). 2^48 microseconds is close to 9 years.

That leaves 15 bits for the P, which is primes from 3..32749. Is that a large enough pool of prime numbers?

Well, there are ~3000 primes in this range so that gives you a 1/3000 chance that after a crash, backup restore, snapshot rollback etc. you will pick the same prime which results in collisions of change cookies and thus possibility of data corruption. Is that low enough chance? The events I mention above should be relatively rare but given the number of machines running this code I would think the collision is bound to happen and the consequences could be ... unpleasant. That's why I would prefer to pick primes at least say upto 1m (there are ~78k of those). But that makes wrapping more frequent (~100 days with 1us update period). Probably still usable for NFS but not really for backup purposes. So I'm not sure we should be advertising the values have any ordering.

If the last used value would be persisted (e.g. in the filesystem's superblock), we could easily make sure the next selected P is different so in that case we could get away with a smaller set of primes but it would require filesystem on-disk format changes which has its own drawbacks. But that would be at least some path forward for providing change cookies that can be ordered on larger timescales.

Quoted 4 lines

It looks like the kernel already has some infrastructure for handling primes in lib/math/prime_numbers.c. We could just select a global P value to use on every reboot, or just have filesystems set their own (maybe in a new field in the superblock?)

IMO P needs to be selected on each mount to reliably solve the "restore from backup" and "snapshot rollback" scenarios. I agree it can be a new field in the VFS part of the superblock so that it is accessible by the iversion handling code.

Quoted 16 lines · Dave Chinner

Quoted 6 lines · Jeff Layton

From: Jeff Layton <[email protected]>

Claim one of the spare fields in struct statx to hold a 64-bit inode version attribute. When userland requests STATX_VERSION, copy the value from the kstat struct there, and stop masking off STATX_ATTR_VERSION_MONOTONIC.

Can we please make the name more sepcific than "version"? It's way too generic and - we already have userspace facing "version" fields for inodes that refer to the on-disk format version exposed in various UAPIs. It's common for UAPI structures used for file operations to have a "version" field that refers to the *UAPI structure version* rather than file metadata or data being retrieved from the file in question.

Quoted 9 lines

The need for an explanatory comment like this:

include/uapi/linux/stat.h struct statx { · line 130

Quoted 1 line

include/uapi/linux/stat.h +1 -0

+	__u64	stx_version; /* Inode change attribute */

demonstrates it is badly named. If you want it known as an inode change attribute, then don't name the variable "version". In reality, it really needs to be an opaque cookie, not something applications need to decode directly to make sense of.

Fair enough. I started with this being named stx_change_attr and other people objected. I then changed to stx_ino_version, but the "_ino" seemed redundant.

I'm open to suggestions here. Naming things like this is hard.

This patchset is intended to clean up the handling of the i_version counter by nfsd. Most of the changes are to internal interfaces.

This set is not intended to address crash resilience, or the fact that the counter is bumped before a change and not after. I intend to tackle those in follow-on patchsets.

My intention is to get this series included into linux-next soon, with an eye toward merging most of it during the v6.2 merge window. The last patch in the series is probably not suitable for merge as-is, at least until we sort out the semantics we want to present to userland for it.

Quoted 11 lines · Jeff Layton

This patchset is intended to clean up the handling of the i_version counter by nfsd. Most of the changes are to internal interfaces.

This set is not intended to address crash resilience, or the fact that the counter is bumped before a change and not after. I intend to tackle those in follow-on patchsets.

My intention is to get this series included into linux-next soon, with an eye toward merging most of it during the v6.2 merge window. The last patch in the series is probably not suitable for merge as-is, at least until we sort out the semantics we want to present to userland for it.

Over the course of the series I struggled a bit - and sorry for losing focus - with what i_version is supposed to represent for userspace. So I would support not exposing it to userspace before that. But that shouldn't affect your other changes iiuc.

Quoted 13 lines · Christian Brauner

Quoted 11 lines · Jeff Layton

This patchset is intended to clean up the handling of the i_version counter by nfsd. Most of the changes are to internal interfaces.

This set is not intended to address crash resilience, or the fact that the counter is bumped before a change and not after. I intend to tackle those in follow-on patchsets.

My intention is to get this series included into linux-next soon, with an eye toward merging most of it during the v6.2 merge window. The last patch in the series is probably not suitable for merge as-is, at least until we sort out the semantics we want to present to userland for it.

Quoted 4 lines

Over the course of the series I struggled a bit - and sorry for losing focus - with what i_version is supposed to represent for userspace. So I would support not exposing it to userspace before that. But that shouldn't affect your other changes iiuc.

Thanks Christian,

It has been a real struggle to nail this down, and yeah I too am not planning to expose this to userland until we have this much better defined. Patch #9 is just to give you an idea of what this would ultimately look like. I intend to re-post the first 8 patches with an eye toward merge in v6.2, once we've settled on the naming. On that note...

I believe you had mentioned that you didn't like STATX_CHANGE_ATTR for the name, and suggested STATX_I_VERSION (or something similar), which I later shortened to STATX_VERSION.

Dave C. objected to STATX_VERSION, as "version" fields in a struct usually refer to the version of the struct itself rather than the version of the thing it describes. It also sort of implies a monotonic counter, and I'm not ready to require that just yet.

What about STATX_CHANGE for the name (with corresponding names for the field and other flags)? That drops the redundant "_ATTR" postfix, while being sufficiently vague to allow for alternative implementations in the future.

Do you (or anyone else) have other suggestions for a name?

Signature · 1 line

Jeff Layton <[email protected]>

Quoted 42 lines · Jeff Layton

Quoted 13 lines · Christian Brauner

Quoted 11 lines · Jeff Layton

This patchset is intended to clean up the handling of the i_version counter by nfsd. Most of the changes are to internal interfaces.

This set is not intended to address crash resilience, or the fact that the counter is bumped before a change and not after. I intend to tackle those in follow-on patchsets.

My intention is to get this series included into linux-next soon, with an eye toward merging most of it during the v6.2 merge window. The last patch in the series is probably not suitable for merge as-is, at least until we sort out the semantics we want to present to userland for it.

Quoted 4 lines

Over the course of the series I struggled a bit - and sorry for losing focus - with what i_version is supposed to represent for userspace. So I would support not exposing it to userspace before that. But that shouldn't affect your other changes iiuc.

Thanks Christian,

It has been a real struggle to nail this down, and yeah I too am not planning to expose this to userland until we have this much better defined. Patch #9 is just to give you an idea of what this would ultimately look like. I intend to re-post the first 8 patches with an eye toward merge in v6.2, once we've settled on the naming. On that note...

I believe you had mentioned that you didn't like STATX_CHANGE_ATTR for the name, and suggested STATX_I_VERSION (or something similar), which I later shortened to STATX_VERSION.

Dave C. objected to STATX_VERSION, as "version" fields in a struct usually refer to the version of the struct itself rather than the version of the thing it describes. It also sort of implies a monotonic counter, and I'm not ready to require that just yet.

What about STATX_CHANGE for the name (with corresponding names for the field and other flags)? That drops the redundant "_ATTR" postfix, while being sufficiently vague to allow for alternative implementations in the future.

Quoted 1 line

Do you (or anyone else) have other suggestions for a name?

Welllll it's really a u32 whose value doesn't have any intrinsic meaning other than "if (value_now != value_before) flush_cache();" right? I think it really only tracks changes to file data, right?

STATX_CHANGE_COOKIE (wait, does this cookie augment i_ctime?)

STATX_MOD_COOKIE (...or just file modifications/i_mtime?)

STATX_MONITOR_COOKIE (...what are we monitoring??)

STATX_MON_COOKIE

STATX_COOKIE_MON

STATX_COOKIE_MONSTER

There we go. ;)

In seriousness, I'd probably go with one of the first two. I wouldn't be opposed to the last one, either, but others may disagree. ;)

--D

Quoted 43 lines · Darrick J. Wong

Quoted 42 lines · Jeff Layton

Quoted 13 lines · Christian Brauner

Quoted 11 lines · Jeff Layton

This patchset is intended to clean up the handling of the i_version counter by nfsd. Most of the changes are to internal interfaces.

This set is not intended to address crash resilience, or the fact that the counter is bumped before a change and not after. I intend to tackle those in follow-on patchsets.

My intention is to get this series included into linux-next soon, with an eye toward merging most of it during the v6.2 merge window. The last patch in the series is probably not suitable for merge as-is, at least until we sort out the semantics we want to present to userland for it.

Quoted 4 lines

Over the course of the series I struggled a bit - and sorry for losing focus - with what i_version is supposed to represent for userspace. So I would support not exposing it to userspace before that. But that shouldn't affect your other changes iiuc.

Thanks Christian,

It has been a real struggle to nail this down, and yeah I too am not planning to expose this to userland until we have this much better defined. Patch #9 is just to give you an idea of what this would ultimately look like. I intend to re-post the first 8 patches with an eye toward merge in v6.2, once we've settled on the naming. On that note...

I believe you had mentioned that you didn't like STATX_CHANGE_ATTR for the name, and suggested STATX_I_VERSION (or something similar), which I later shortened to STATX_VERSION.

Dave C. objected to STATX_VERSION, as "version" fields in a struct usually refer to the version of the struct itself rather than the version of the thing it describes. It also sort of implies a monotonic counter, and I'm not ready to require that just yet.

What about STATX_CHANGE for the name (with corresponding names for the field and other flags)? That drops the redundant "_ATTR" postfix, while being sufficiently vague to allow for alternative implementations in the future.

Quoted 6 lines

Quoted 1 line

Do you (or anyone else) have other suggestions for a name?

Welllll it's really a u32 whose value doesn't have any intrinsic meaning other than "if (value_now != value_before) flush_cache();" right? I think it really only tracks changes to file data, right?

It's a u64, but yeah, you're not supposed to assign any intrinsic meaning to the value itself.

Quoted 17 lines

STATX_CHANGE_COOKIE (wait, does this cookie augment i_ctime?)

STATX_MOD_COOKIE (...or just file modifications/i_mtime?)

STATX_MONITOR_COOKIE (...what are we monitoring??)

STATX_MON_COOKIE

STATX_COOKIE_MON

STATX_COOKIE_MONSTER

There we go. ;)

In seriousness, I'd probably go with one of the first two. I wouldn't be opposed to the last one, either, but others may disagree. ;)

Quoted 3 lines

--D

STATX_CHANGE_COOKIE is probably the best one. I'll plan to go with that unless someone has a better idea. Thanks for the suggestions!