2.6: marking individual directories as synchronous?

2.6: marking individual directories as synchronous?

[Dan Behman]

Hi,

I'm reading through Joseph Pranevich's great document "Wonderful World of 
Linux 2.6" and I came across something that I'd love to learn more about.  
In the "Block Device Support" -> "Filesystems" section, reference is made to 
"Individual directories can now be marked as synchronous so that all changes 
(additional files, etc.) will be atomic".  I searched through the update 
info at kernelnewbies but
couldn't find any more information on this - could someone please elaborate 
on this?  What is it and how does it work?  Is there any design 
documentation for this?

Thanks!

Dan Behman...

_________________________________________________________________
MSN 8 with e-mail virus protection service: 2 months FREE*  
http://join.msn.com/?page=features/virus

Re: 2.6: marking individual directories as synchronous?

[Theodore Ts'o]

On Thu, Jul 17, 2003 at 04:59:20PM -0400, Dan Behman wrote:
> Hi,
> 
> I'm reading through Joseph Pranevich's great document "Wonderful World of 
> Linux 2.6" and I came across something that I'd love to learn more about.  
> In the "Block Device Support" -> "Filesystems" section, reference is made 
> to "Individual directories can now be marked as synchronous so that all 
> changes (additional files, etc.) will be atomic".  I searched through the 
> update info at kernelnewbies but
> couldn't find any more information on this - could someone please elaborate 
> on this?  What is it and how does it work?  Is there any design 
> documentation for this?

He is is probably referring to "chattr +S".  See the man page for
chattr for more information.  Note that strictly speaking this does
not necessarily give you "atomic changes".  It does mean that changes
are scheduled to be immediately written to disk, but that does not
guarantee atomicity, at least not for all filesystems and for all
operations.  You *can* be guaranteed that system calls will not return
until the changes are on disk; note though that this does have has
some significant performance impacts.

						- Ted

Re: 2.6: marking individual directories as synchronous?

[Carl Spalletta]

Here is further info on the use of IS_DIRSYNC in ext2.

ext2_alloc_branch() is used only in the O_DIRECT path.
ext2_commit_chunk() is called indirectly in the following paths, which provide
functionality comparable to that of ext3:

[linux-2.6.0-test1]$ fscope -func=ext2_commit_chunk

ext2_commit_chunk ext2_add_link ext2_add_nondir ext2_create
ext2_commit_chunk ext2_add_link ext2_add_nondir ext2_link
ext2_commit_chunk ext2_add_link ext2_add_nondir ext2_mknod
ext2_commit_chunk ext2_add_link ext2_add_nondir ext2_symlink
ext2_commit_chunk ext2_add_link ext2_mkdir
ext2_commit_chunk ext2_add_link ext2_rename
ext2_commit_chunk ext2_delete_entry ext2_rename
ext2_commit_chunk ext2_delete_entry ext2_unlink ext2_rmdir
ext2_commit_chunk ext2_make_empty ext2_mkdir
ext2_commit_chunk ext2_set_link ext2_rename

FOR CLARITY:

items 1-4 rotated:
ext2_create        ext2_link          ext2_mknod         ext2_symlink
ext2_add_nondir    ext2_add_nondir    ext2_add_nondir    ext2_add_nondir
ext2_add_link      ext2_add_link      ext2_add_link      ext2_add_link
ext2_commit_chunk  ext2_commit_chunk  ext2_commit_chunk  ext2_commit_chunk

items 5-8 rotated:
(null)             (null)             (null)             ext2_rmdir
ext2_mkdir         ext2_rename        ext2_rename        ext2_unlink
ext2_add_link      ext2_add_link      ext2_delete_entry  ext2_delete_entry
ext2_commit_chunk  ext2_commit_chunk  ext2_commit_chunk  ext2_commit_chunk

items 9-10 rotated:
ext2_mkdir         ext2_rename
ext2_make_empty    ext2_set_link
ext2_commit_chunk  ext2_commit_chunk

(See this list for tool 'fscope'.)

_

Re: 2.6: marking individual directories as synchronous?

[Carl Spalletta]

Grepping around in 2.6.0-test1 src I found:

include/linux/fs.h:
105 #define MS_DIRSYNC 128 /* Directory modifications are synchronous */
138 #define S_DIRSYNC  128 /* Directory modifications are synchronous */

Therefore, study the definitions and uses of those flags as well as
IS_DIRSYNC(), EXT3_DIRSYNC_FL, ext3_ioctl() & ext3_set_inode_flags().

For example:
[linux-2.6.0-test1]$ cscope -d -L -3 IS_DIRSYNC
...
fs/ext2/dir.c     ext2_commit_chunk  71    if  (IS_DIRSYNC(dir))
fs/ext3/ialloc.c  ext3_new_inode     585   if  (IS_DIRSYNC(inode))
fs/ext3/namei.c   ext3_create        1638  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_mknod         1665  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_mkdir         1697  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_rmdir         1981  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_unlink        2033  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_symlink       2089  if  (IS_DIRSYNC(dir))
fs/ext3/namei.c   ext3_link          2139  if  (IS_DIRSYNC(dir))
fs/minix/dir.c    dir_commit_chunk   53    if  (IS_DIRSYNC(dir))
fs/sysv/dir.c     dir_commit_chunk   46    if  (IS_DIRSYNC(dir))
fs/ufs/dir.c      ufs_set_link       359   if  (IS_DIRSYNC(dir))
fs/ufs/dir.c      ufs_add_link       458   if  (IS_DIRSYNC(dir))
fs/ufs/dir.c      ufs_delete_entry   507   if  (IS_DIRSYNC(inode))
...

I haven't actually played with the application of this, but it would appear
to be some combination of ioctl's and/or mount flags.  Check the source for
chattr(1) to see if and how it uses the ioctl.