lkml.org 
[lkml]   [2013]   [Jan]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2 2/3] mm: Update file times when inodes are written after mmaped writes
On Mon, Dec 31, 2012 at 8:11 AM, Jan Kara <jack@suse.cz> wrote:
> On Sat 22-12-12 00:43:30, Andy Lutomirski wrote:
>> On Sat, Dec 22, 2012 at 12:29 AM, Christoph Hellwig <hch@infradead.org> wrote:
>> > NAK, we went through great trouble to get rid of the nasty layering
>> > violation where the VM called file_update_time directly just a short
>> > while ago, reintroducing that is a massive step back.
>> >

[...]

>
>> The original version of this patch did the update in ->writepage and
>> ->writepages, but that may have had lock ordering issues. (I wasn't
>> able to confirm that there was any actual problem.)
> Well, your call of mapping_flush_cmtime() from do_writepages() is easy to
> move to generic_writepages(). Thus filesystem can easily implement it's own
> ->writepages() callback if time update doesn't suit it.

That sounds fine to me. Updating the handful of filesystems in there
isn't a big deal.

>With the call from
> remove_vma() it is more problematic (and the calling context there is
> harder as well because we hold mmap_sem). We could maybe leave the call
> upto filesystem's ->release callback (and provide generic ->release handler
> which just calls mapping_flush_cmtime()). It won't be perfect because that
> gets called only after the last file descriptor for that struct file is
> closed (i.e., if a process forks and child inherits mappings, ->release gets
> called only after both parent and the child unmap the file) but it should
> catch 99% of the real world cases. Christoph, would the be OK with
> you?

I'm not sure that 99% is good enough -- I'd be nervous about breaking
some build or versioning system.

vm_ops->close is almost a good place for this, except that it's called
on some failure paths and it will mess up is_mergeable_vma if lots of
filesystems suddenly have a ->close operation. What about adding
vm_ops->flush, which would be called in remove_vma and possibly
msync(MS_ASYNC)? I think that all real filesystems (i.e. things that
care about cmtime updates) have vm_operations.

--Andy


\
 
 \ /
  Last update: 2013-01-03 19:01    [W:0.389 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site