[lkml]   [2014]   [Dec]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC PATCH v3 0/7] btrfs: implement swap file support
On Fri, Dec 12, 2014 at 11:32:13AM +0100, David Sterba wrote:
> On Tue, Dec 09, 2014 at 05:45:41PM -0800, Omar Sandoval wrote:
> > After some discussion on the mailing list, I decided that for simplicity and
> > reliability, it's best to simply disallow COW files and files with shared
> > extents (like files with extents shared with a snapshot). From a user's
> > perspective, this means that a snapshotted subvolume cannot be used for a swap
> > file, but keeping the swap file in a separate subvolume that is never
> > snapshotted seems entirely reasonable to me.
> Well, there are enough special cases how to do things on btrfs and I'd
> like to avoid introducing another one.
> > An alternative suggestion was to
> > allow swap files to be snapshotted and to do an implied COW on swap file
> > activation, which I was ready to implement until I realized that we can't permit
> > snapshotting a subvolume with an active swap file, so this creates a surprising
> > inconsistency for users (in my opinion).
> I still don't see why it's not possible to do the snapshot with an
> active swapfile.
Creating a snapshot of an active swapfile would create shared extents,
so the next time we have to swap out a page, we'd have to do a COW,
which we're already trying pretty hard to avoid. We could allow it, but
it might lead to some unreliable behavior and unhappy emails to the
mailing list. However, I do see your point about wanting to avoid
special cases, so I'd like to get some more input from others on this as

> > As with before, this functionality is tenuously tested in a virtual machine with
> > some artificial workloads, but it "works for me". I'm pretty happy with the
> > results on my end, so please comment away.
> The non-btrfs changes can go independently and do not have to wait until
> we resolve the swap vs snapshot problem.
> I did a simple test and it crashed instantly, lockep complains:
> memory: 2G
> swap file: 1G
> kernel: 3.17 + v3

That's my fault for not running with lockdep enabled. The problem here
is that swap-over-NFS is the only caller of nfs_direct_IO, so
nfs_direct_IO doesn't observe the normal direct_IO locking conventions
and neither does swap_writepage. I'll have to shuffle around some code
on the NFS side to fix that.

It looks like the non-btrfs parts of this might get a bit bigger, so
I'll look into getting that in separately.


 \ /
  Last update: 2014-12-12 21:41    [W:0.064 / U:16.724 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site