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 well. > > 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 > [snip] 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. Thanks! -- Omar -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>