xfs: stop advertising SB_I_VERSION
Subsystems (2)
PATCH 1/1 xfs: stop advertising SB_I_VERSION
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants inode->i_version to be updated is incomaptible with the XFS i_version mechanism. The VFS now wants inode->i_version to only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.
We still want XFS to update the inode change counter as it currently does, so convert all the code that checks SB_I_VERSION to check for v5 format support. Then we can remove the SB_I_VERSION flag from the VFS superblock to indicate that inode->i_version is not a valid change counter and should not be used as such.
--- fs/xfs/libxfs/xfs_trans_inode.c | 15 +++++---------- fs/xfs/xfs_iops.c | 16 +++------------- fs/xfs/xfs_super.c | 8 -------- 3 files changed, 8 insertions(+), 31 deletions(-)
fs/xfs/libxfs/xfs_trans_inode.c +5 -10
@@ -97,17 +97,12 @@ xfs_trans_log_inode(/* * First time we log the inode in a transaction, bump the inode change -* counter if it is configured for this to occur. While we have the -* inode locked exclusively for metadata modification, we can usually -* avoid setting XFS_ILOG_CORE if no one has queried the value since -* the last time it was incremented. If we have XFS_ILOG_CORE already -* set however, then go ahead and bump the i_version counter -* unconditionally. +* counter if it is configured for this to occur. */ -if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags)) { -if (IS_I_VERSION (inode) && -inode_maybe_inc_iversion (inode, flags &XFS_ILOG_CORE )) - flags |=XFS_ILOG_IVERSION ; +if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags) && +xfs_has_crc (ip->i_mount)) { + inode->i_version++; + flags |=XFS_ILOG_IVERSION ; } iip->ili_dirty_flags |= flags;
fs/xfs/xfs_iops.c +3 -13
@@ -584,11 +584,6 @@ xfs_vn_getattr( } } -if ((request_mask &STATX_CHANGE_COOKIE ) &&IS_I_VERSION (inode)) { - stat->change_cookie =inode_query_iversion (inode); - stat->result_mask |=STATX_CHANGE_COOKIE ; - } -/* * Note: If you add another clause to set an attribute flag, please * update attributes_mask below.
@@ -1044,16 +1039,11 @@ xfs_vn_update_time(struct timespec64 now;trace_xfs_update_time (ip); +ASSERT (!(flags &S_VERSION ));if (inode->i_sb->s_flags &SB_LAZYTIME ) { -if (!((flags &S_VERSION ) && -inode_maybe_inc_iversion (inode, false))) { -generic_update_time (inode, flags); -return 0 ; - } - -/* Capture the iversion update that just occurred */ - log_flags |=XFS_ILOG_CORE ; +generic_update_time (inode, flags); +return 0 ; } error =xfs_trans_alloc (mp, &M_RES (mp)->tr_fsyncts,0 ,0 ,0 , &tp);
fs/xfs/xfs_super.c +0 -8
@@ -1692,10 +1692,6 @@ xfs_fs_fill_super(set_posix_acl_flag (sb); -/* version 5 superblocks support inode version counters. */ -if (xfs_has_crc (mp)) - sb->s_flags |=SB_I_VERSION ; -if (xfs_has_dax_always (mp)) { error =xfs_setup_dax_always (mp);if (error)
@@ -1917,10 +1913,6 @@ xfs_fs_reconfigure(int flags = fc->sb_flags;int error; -/* version 5 superblocks always support version counters. */ -if (xfs_has_crc (mp)) - fc->sb_flags |=SB_I_VERSION ; - error =xfs_fs_validate_params (new_mp);if (error)return error;
Signature · 1 line
2.43.0
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.
Quoted 7 lines
We still want XFS to update the inode change counter as it currently does, so convert all the code that checks
SB_I_VERSIONto check for v5 format support. Then we can remove theSB_I_VERSIONflag from the VFS superblock to indicate thatinode->i_versionis not a valid change counter and should not be used as such.Signed-off-by Dave Chinner <[email protected]>
Seeing as NFS and XFS' definition of i_version have diverged, I suppose divorce is the only option. But please, let's get rid of all the *iversion() calls in the codebase.
With my paranoia hat on: let's add an i_changecounter to xfs_inode and completely stop using the inode.i_version to prevent the vfs from messing with us.
At some point we can rev the ondisk format to add a new field so that "di_version" can be whatever u64 cookie the vfs passes us through i_version.
--D
Quoted 99 lines
--- fs/xfs/libxfs/xfs_trans_inode.c | 15 +++++---------- fs/xfs/xfs_iops.c | 16 +++------------- fs/xfs/xfs_super.c | 8 -------- 3 files changed, 8 insertions(+), 31 deletions(-)fs/xfs/libxfs/xfs_trans_inode.c +5
-10@@ -97,17 +97,12 @@ xfs_trans_log_inode(/* * First time we log the inode in a transaction, bump the inode change -* counter if it is configured for this to occur. While we have the -* inode locked exclusively for metadata modification, we can usually -* avoid setting XFS_ILOG_CORE if no one has queried the value since -* the last time it was incremented. If we have XFS_ILOG_CORE already -* set however, then go ahead and bump the i_version counter -* unconditionally. +* counter if it is configured for this to occur. */ -if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags)) { -if (IS_I_VERSION (inode) && -inode_maybe_inc_iversion (inode, flags &XFS_ILOG_CORE )) - flags |=XFS_ILOG_IVERSION ; +if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags) && +xfs_has_crc (ip->i_mount)) { + inode->i_version++; + flags |=XFS_ILOG_IVERSION ; } iip->ili_dirty_flags |= flags;fs/xfs/xfs_iops.c +3
-13@@ -584,11 +584,6 @@ xfs_vn_getattr( } } -if ((request_mask &STATX_CHANGE_COOKIE ) &&IS_I_VERSION (inode)) { - stat->change_cookie =inode_query_iversion (inode); - stat->result_mask |=STATX_CHANGE_COOKIE ; - } -/* * Note: If you add another clause to set an attribute flag, please * update attributes_mask below. @@ -1044,16 +1039,11 @@ xfs_vn_update_time(struct timespec64 now;trace_xfs_update_time (ip); +ASSERT (!(flags &S_VERSION ));if (inode->i_sb->s_flags &SB_LAZYTIME ) { -if (!((flags &S_VERSION ) && -inode_maybe_inc_iversion (inode, false))) { -generic_update_time (inode, flags); -return 0 ; - } - -/* Capture the iversion update that just occurred */ - log_flags |=XFS_ILOG_CORE ; +generic_update_time (inode, flags); +return 0 ; } error =xfs_trans_alloc (mp, &M_RES (mp)->tr_fsyncts,0 ,0 ,0 , &tp);fs/xfs/xfs_super.c +0
-8@@ -1692,10 +1692,6 @@ xfs_fs_fill_super(set_posix_acl_flag (sb); -/* version 5 superblocks support inode version counters. */ -if (xfs_has_crc (mp)) - sb->s_flags |=SB_I_VERSION ; -if (xfs_has_dax_always (mp)) { error =xfs_setup_dax_always (mp);if (error)@@ -1917,10 +1913,6 @@ xfs_fs_reconfigure(int flags = fc->sb_flags;int error; -/* version 5 superblocks always support version counters. */ -if (xfs_has_crc (mp)) - fc->sb_flags |=SB_I_VERSION ; - error =xfs_fs_validate_params (new_mp);if (error)return error;Signature · 1 line
2.43.0
Quoted 18 lines · Darrick J. Wong
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.Quoted 7 lines
We still want XFS to update the inode change counter as it currently does, so convert all the code that checks
SB_I_VERSIONto check for v5 format support. Then we can remove theSB_I_VERSIONflag from the VFS superblock to indicate thatinode->i_versionis not a valid change counter and should not be used as such.Signed-off-by Dave Chinner <[email protected]>
Quoted 7 lines
Seeing as NFS and XFS' definition of
i_versionhave diverged, I suppose divorce is the only option. But please, let's get rid of all the*iversion()calls in the codebase.With my paranoia hat on: let's add an
i_changecountertoxfs_inodeand completely stop using the inode.i_version to prevent the vfs from messing with us.
Ok, I'll do that in a new patch rather than try to do everything in a single complicated patch.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 23 lines · Dave Chinner
Quoted 22 lines · Darrick J. Wong
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.Quoted 7 lines
We still want XFS to update the inode change counter as it currently does, so convert all the code that checks
SB_I_VERSIONto check for v5 format support. Then we can remove theSB_I_VERSIONflag from the VFS superblock to indicate thatinode->i_versionis not a valid change counter and should not be used as such.Signed-off-by Dave Chinner <[email protected]>Seeing as NFS and XFS' definition of
i_versionhave diverged, I suppose divorce is the only option. But please, let's get rid of all the*iversion()calls in the codebase.
Quoted 6 lines
Quoted 3 lines
With my paranoia hat on: let's add an
i_changecountertoxfs_inodeand completely stop using the inode.i_version to prevent the vfs from messing with us.Ok, I'll do that in a new patch rather than try to do everything in a single complicated patch.
Sounds good!
--D
Quoted 7 lines
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.
Are you saying that XFS has timestamp granularity finer than current_time() reports? I thought XFS used the same clocksource as everyone else.
At LPC, you mentioned you had some patches in progress to use the unused bits in the tv_nsec field as a change counter to track changes that occurred within the same timer tick.
Did that not pan out for some reason? I'd like to understand why if so. It sounded like a reasonable solution to the problem.
Quoted 93 lines
We still want XFS to update the inode change counter as it currently does, so convert all the code that checks
SB_I_VERSIONto check for v5 format support. Then we can remove theSB_I_VERSIONflag from the VFS superblock to indicate thatinode->i_versionis not a valid change counter and should not be used as such.Signed-off-by Dave Chinner <[email protected]>--- fs/xfs/libxfs/xfs_trans_inode.c | 15 +++++---------- fs/xfs/xfs_iops.c | 16 +++------------- fs/xfs/xfs_super.c | 8 -------- 3 files changed, 8 insertions(+), 31 deletions(-)fs/xfs/libxfs/xfs_trans_inode.c +5
-10@@ -97,17 +97,12 @@ xfs_trans_log_inode(/* * First time we log the inode in a transaction, bump the inode change -* counter if it is configured for this to occur. While we have the -* inode locked exclusively for metadata modification, we can usually -* avoid setting XFS_ILOG_CORE if no one has queried the value since -* the last time it was incremented. If we have XFS_ILOG_CORE already -* set however, then go ahead and bump the i_version counter -* unconditionally. +* counter if it is configured for this to occur. */ -if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags)) { -if (IS_I_VERSION (inode) && -inode_maybe_inc_iversion (inode, flags &XFS_ILOG_CORE )) - flags |=XFS_ILOG_IVERSION ; +if (!test_and_set_bit (XFS_LI_DIRTY , &iip->ili_item.li_flags) && +xfs_has_crc (ip->i_mount)) { + inode->i_version++; + flags |=XFS_ILOG_IVERSION ; } iip->ili_dirty_flags |= flags;fs/xfs/xfs_iops.c +3
-13@@ -584,11 +584,6 @@ xfs_vn_getattr( } } -if ((request_mask &STATX_CHANGE_COOKIE ) &&IS_I_VERSION (inode)) { - stat->change_cookie =inode_query_iversion (inode); - stat->result_mask |=STATX_CHANGE_COOKIE ; - } -/* * Note: If you add another clause to set an attribute flag, please * update attributes_mask below. @@ -1044,16 +1039,11 @@ xfs_vn_update_time(struct timespec64 now;trace_xfs_update_time (ip); +ASSERT (!(flags &S_VERSION ));if (inode->i_sb->s_flags &SB_LAZYTIME ) { -if (!((flags &S_VERSION ) && -inode_maybe_inc_iversion (inode, false))) { -generic_update_time (inode, flags); -return 0 ; - } - -/* Capture the iversion update that just occurred */ - log_flags |=XFS_ILOG_CORE ; +generic_update_time (inode, flags); +return 0 ; } error =xfs_trans_alloc (mp, &M_RES (mp)->tr_fsyncts,0 ,0 ,0 , &tp);fs/xfs/xfs_super.c +0
-4@@ -1692,10 +1692,6 @@ xfs_fs_fill_super(set_posix_acl_flag (sb); -/* version 5 superblocks support inode version counters. */ -if (xfs_has_crc (mp)) - sb->s_flags |=SB_I_VERSION ; -if (xfs_has_dax_always (mp)) { error =xfs_setup_dax_always (mp);if (error)
fs/xfs/xfs_super.c xfs_fs_reconfigure( · line 1918
Quoted 10 lines
fs/xfs/xfs_super.c +0
-4@@ -1917,10 +1913,6 @@ xfs_fs_reconfigure(int flags = fc->sb_flags;int error; -/* version 5 superblocks always support version counters. */ -if (xfs_has_crc (mp)) - fc->sb_flags |=SB_I_VERSION ; - error =xfs_fs_validate_params (new_mp);if (error)return error;
Quoted 11 lines · Jeff Layton
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.
Quoted 2 lines
Are you saying that XFS has timestamp granularity finer than
current_time()reports?
No.
Quoted 2 lines
I thought XFS used the same clocksource as
everyone else.
It does.
Quoted 3 lines
At LPC, you mentioned you had some patches in progress to use the unused bits in the
tv_nsecfield as a change counter to track changes that occurred within the same timer tick.
Still a possibility, but I wasn't going to do anything in that direction because it still seemed like you were still trying to make progress down the path of generic timestamp granularity improvements.
Quoted 2 lines
Did that not pan out for some reason? I'd like to understand why if so. It sounded like a reasonable solution to the problem.
Time. And the fact that ctime granularity isn't SB_I_VERSION at all, so whilst we might support statx change cookies in the future, that will not be via a SB_I_VERSION mechanism.
i.e. statx doesn't require us to support SB_I_VERSION for the change cookies, so until we are in a position to present a higher resolution change cookie via ctime we're just going to remove support for both.
Quoted 1 line
Acked-by Jeff Layton <[email protected]>
Thanks!
-Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 22 lines · Dave Chinner
Quoted 11 lines · Jeff Layton
Quoted 9 lines · Dave Chinner
From: Dave Chinner <[email protected]>
The redefinition of how NFS wants
inode->i_versionto be updated is incomaptible with the XFSi_versionmechanism. The VFS now wantsinode->i_versionto only change when ctime changes (i.e. it has become a ctime change counter, not an inode change counter). XFS has fine grained timestamps, so it can just use ctime for the NFS change cookie like it still does for V4 XFS filesystems.Quoted 2 lines
Are you saying that XFS has timestamp granularity finer than
current_time()reports?No.
Quoted 2 lines
I thought XFS used the same clocksource as
everyone else.It does.
Quoted 9 lines
Quoted 3 lines
At LPC, you mentioned you had some patches in progress to use the unused bits in the
tv_nsecfield as a change counter to track changes that occurred within the same timer tick.Still a possibility, but I wasn't going to do anything in that direction because it still seemed like you were still trying to make progress down the path of generic timestamp granularity improvements.
Ok, good. I'm glad to hear that that approach isn't unworkable (yet!).
As far as timestamp granularity goes, Linus made it clear that he was not on board with any approach that would even more complexity than the original multigrain ctime patches. That pretty much ended my work in that area.
Quoted 12 lines
Quoted 2 lines
Did that not pan out for some reason? I'd like to understand why if so. It sounded like a reasonable solution to the problem.
Time. And the fact that ctime granularity isn't
SB_I_VERSIONat all, so whilst we might support statx change cookies in the future, that will not be via aSB_I_VERSIONmechanism.i.e. statx doesn't require us to support
SB_I_VERSIONfor the change cookies, so until we are in a position to present a higher resolution change cookie via ctime we're just going to remove support for both.
Yeah, that makes sense.
Quoted 5 lines
Quoted 1 line
Acked-by Jeff Layton <[email protected]>Thanks!
-Dave.
Signature · 1 line
Jeff Layton <[email protected]>