gfs2: Remove sb_start_write from gfs2_statfs_sync
PATCH 1/1 gfs2: Remove sb_start_write from gfs2_statfs_sync
In reply to a message not in the archiveHi,
Before this patch, function gfs2_statfs_sync called sb_start_write. This is a violation of the basic vfs rules that state that sb_start_write should always be taken before s_umount. See this document:
https://www.kernel.org/doc/htmldocs/filesystems/API-sb-start-write.html
"Since freeze protection behaves as a lock, users have to preserve ordering of freeze protection and other filesystem locks. Generally, freeze protection should be the outermost lock. In particular, we have:
sb_start_write -> i_mutex (write path, truncate, directory ops, ...) -> s_umount (freeze_super, thaw_super)"
deactivate_super
down_write(&s->s_umount); <------------------------------------ s_umount deactivate_locked_super
gfs2_kill_sb
kill_block_super
generic_shutdown_super
gfs2_put_super
gfs2_make_fs_ro
gfs2_statfs_sync(sdp->sd_vfs, 0);
sb_start_write <--------------------- sb_start_writeAs far as I can tell, gfs2_statfs_sync doesn't need to call sb_start_write any more than any other write to the file system, which are policed by glocks. None of the other functions in gfs2 lock sb_start_write so it only affects how vfs calls gfs2.
This patch simply removes the call to sb_start_write.
--- fs/gfs2/super.c | 2 -- 1 file changed, 2 deletions(-)
fs/gfs2/super.c +0 -2
@@ -353,7 +353,6 @@ int gfs2_statfs_sync(struct super_block *sb, int type)struct buffer_head *m_bh, *l_bh;int error; -sb_start_write (sb); error =gfs2_glock_nq_init (m_ip->i_gl,LM_ST_EXCLUSIVE ,GL_NOCACHE , &gh);if (error)
@@ -392,7 +391,6 @@ int gfs2_statfs_sync(struct super_block *sb, int type) out_unlock:gfs2_glock_dq_uninit (&gh); out: -sb_end_write (sb);return error; }
Andreas Gruenbacher
maintainerPATCH 1/1 gfs2: Remove sb_start_write from gfs2_statfs_sync
Hi Bob,
Quoted 24 lines · Bob Peterson
Before this patch, function
gfs2_statfs_synccalledsb_start_write. This is a violation of the basic vfs rules that state thatsb_start_writeshould always be taken befores_umount. See this document:https://www.kernel.org/doc/htmldocs/filesystems/API-sb-start-write.html
"Since freeze protection behaves as a lock, users have to preserve ordering of freeze protection and other filesystem locks. Generally, freeze protection should be the outermost lock. In particular, we have:
sb_start_write->i_mutex(write path, truncate, directory ops, ...) ->s_umount(freeze_super,thaw_super)"
deactivate_super
down_write(&s->s_umount); <------------------------------------s_umountdeactivate_locked_supergfs2_kill_sb kill_block_super generic_shutdown_super gfs2_put_super gfs2_make_fs_ro gfs2_statfs_sync(sdp->sd_vfs, 0); sb_start_write <--------------------- sb_start_write
Quoted 4 lines
As far as I can tell,
gfs2_statfs_syncdoesn't need to callsb_start_writeany more than any other write to the file system, which are policed by glocks. None of the other functions in gfs2 locksb_start_writeso it only affects how vfs calls gfs2.
you're quite right that the sb_start_write doesn't make sense in the above code path. That was equally true when the call was added in commit 2e60d7683c8d2 ("GFS2: update freeze code to use freeze/thaw_super on all nodes"), so I'm wondering what the intention may have been here. Are there any code paths not going through the vfs that need protection from filesystem freezes?
I'll leave this patch out for now, at least until it's more obvious what's going on exactly.
Quoted 20 lines
This patch simply removes the call to
sb_start_write.Signed-off-by Bob Peterson <[email protected]>--- fs/gfs2/super.c | 2 -- 1 file changed, 2 deletions(-)fs/gfs2/super.c +0
-1@@ -353,7 +353,6 @@ int gfs2_statfs_sync(struct super_block *sb, int type)struct buffer_head *m_bh, *l_bh;int error; -sb_start_write (sb); error =gfs2_glock_nq_init (m_ip->i_gl,LM_ST_EXCLUSIVE ,GL_NOCACHE , &gh);if (error)
fs/gfs2/super.c int gfs2_statfs_sync(struct super_block *sb, int type) · line 396
Quoted 7 lines
fs/gfs2/super.c +0
-1@@ -392,7 +391,6 @@ int gfs2_statfs_sync(struct super_block *sb, int type) out_unlock:gfs2_glock_dq_uninit (&gh); out: -sb_end_write (sb);return error; }
Thanks,
Andreas