On Thu, Feb 08, 2018 at 10:02:17AM +1100, Dave Chinner wrote: > On Wed, Feb 07, 2018 at 09:47:12AM -0500, Brian Foster wrote: > > On Wed, Feb 07, 2018 at 12:34:41PM +1100, Dave Chinner wrote: > > > On Tue, Feb 06, 2018 at 08:02:47AM -0500, Brian Foster wrote: > > > > Extent swap is a low level mechanism exported by XFS to facilitate > > > > filesystem defragmentation. It is typically invoked by xfs_fsr under > > > > conditions that will atomically adjust inode extent state without > > > > loss of file data. > > > > > > > > xfs_fsr does not provide the necessary controls for low-level > > > > testing of the extent swap command, however. For example, xfs_fsr > > > > implements policies that dictate when to perform an extent swap and > > > > offers no control over the donor file characteristics. > > > > > > Hmmmm. I'm confused - we already use xfs_fsr for low level testing > > > of swapext functionality with carefully constructed donor file > > > characteristics in xfs/227. > > > > > > What am I missing here? > > > > > > > AFAICT, xfs/227 does so by formatting the filesystem in a particular way > > and running fsr over it, so it wasn't immediately clear to me what > > you're referring to. Staring at it some more, it appears to be using the > > -C debug option which looks like it forces an extent count..? If that's > > the case, the test[1] that was posted to use this feature could probably > > be modified to do a similar thing via xfs_fsr. > > > > Giving it a quick test, it doesn't appear to do exactly what I want with > > larger extent counts, however. It seems to work with a few extents or > > so, then doesn't ever create more than 4 or 5 if that's what I ask for > > with the donor file. Hmm, I wonder if prealloc or something is getting > > in the way of whatever it's attempting to do. If I try a similar test > > using falloc instead of pwrite, it doesn't do the swap at all. :/ > > Yeah, that's because of the way xfs_fsr preserves unwritten extents > across defragmentation. i.e. it assumes unwritten extents are a > result of application level preallocation and so it's there > intentionally and be preserved. > > When I wrote this yesterday, I hadn't see the test you'd posted to > the fstests list so I didn't have that context. Now that I've seen > it, and you've explained the problems, I can see why you wrote the > simpler xfs_io test command.... > > [....] > > > So unless there's any objections, I can try to reword the commit log for > > this patch to make the advantages/differences more clear. > > Yeah, I think that's all that is necessary - just expand on > "xfs_fsr does not provide the necessary controls" a bit so I don't > have to rely on my brain to remember this conversation. :P <shrug> I had figured xfs_fsr == "polished user program not to be cluttered up with weird xfstests warts" and xfs_io == "warty program we use to make raw calls for testing" so except for tests that check the behavior of xfs_fsr itself, there ought to be a raw interface via xfs_io that regression tests use directly. :) --D > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx > -- > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
