Re: [PATCH] memory mapped files not updating timestamps

[Andrew Morton <akpm@osdl.org>]

On Wed, 31 May 2006 13:53:16 -0400
Peter Staubach <staubach@redhat.com> wrote:

>

A few issues here.

> 
> I embarked on this work due to a bug reported by one of Red Hat's large
> customers.  They are finding that files, which should have been backed
> up, were not getting backed up.  This is due to the mtime on the files
> not changing and their backup software looking for mtime changes.  This
> is corruption and I need to get it fixed, sooner as opposed to later.
> 
> While I would like to get this fixed on top of Peter Zijlstra's changes,
> the process for those is looking long and complicated.  I am wondering
> if we could consider these changes and then add the requirement of
> maintaining these semantics to those that Peter's work is attempting to
> address.
> 

We seem to be having a problem getting a coherent changelog.  A changelog
should describe why the patch exists, what it does and how it does it. 
Please develop and maintain a changelog for each patch and reissue the
changelog with each reissuing of a patch, thanks.

This sequence:

+	if (test_and_clear_bit(AS_MCTIME, &mapping->flags))
+		inode_update_time(inode);

appears all over the place and should be implemented in a helper function.


The patch should work correctly for mmaps of block devices and I don't
think it does.  Sometimes it updates the timestamp on the kernel-internal
blockdev inode at mapping->host and sometimes it updates the inode of the
device node (/dev/hda1) at file->f_dentry->d_inode.  It should be updating
the /dev/hda1 inode.

The change to unlink_file_vma() is awkward, IMO.  It means that this helper
function "knows" things about what its caller is using it for.  I'd suggest
that this code should be moved up to a higher level where we have a more
sure semantic context, rather than being implemented in some low-level
helper function where it happens to be convenient.

Also, inode_update_time() can sleep. 
mark_inode_dirty_sync()->__mark_inode_dirty()->ext3_dirty_inode().  This is
despite Documentation/filesystems/Locking saying "must not sleep".  But
unlink_file_vma() (at least) is called from atomic context:

	unmap_region()
	->tlb_gather_mmu()
	  ->preempt_disable()
	->free_pgtables()
	  ->unlink_file_vma()
	->tlb_finish_mmu()
	  ->preempt_enable()

Which _should_ have triggered warnings if full kernel debugging was enabled
and sufficient testing was performed.  Perhaps a might_sleep() is needed in
_mark_inode_dirty().