On Tue, Jun 10, 2014 at 08:33:20AM +1000, Dave Chinner wrote:
> Hi everyone,
> Now that the 3.16 dev cycle has drawn to a close (one more
> linux-next build and I'll tag for-next and send a pull request),
> it's time to look ahead for the next couple of months. I think the
> current major pieces of work that are currently outstanding are
> - Jeff's bulkstat rework
> - Brian's EOF prealloc scanning
> - Namjae's FALLOC_FL_INSERT_RANGE work
> - Eric's XFS_ERROR() macro removal and return () cleanup.
I'm not tied to a particular kernel release by any means if there's
already a lot in the pipeline, but I'd like to include the basic sysfs
bits somewhere in the tail end of this.
> There's also two major pieces of infrastructure work I'd like to get
> - convert XFS to negative error returns
> - restructure code to have a fs/xfs/libxfs structure similar
> to userspace
> Because Eric's XFS_ERROR removal touches the entire codebase, as
> does the negative error return and the libxfs restructuring, I'd
> like to get these done first and base the rest of the dev cycle work
> on top of that. Eric's patches just need a minor rebase and the
> libxfs restructure needs some makefile rework and review and they
> should be good to go.
Sounds reasonable to me.
> The issue is the negative error number patchset, and how to handle
> review and testing. The patchset is already 62 patches long and it
> converts roughly half the code base. It'll take me another couple of
> days to convert the rest of the code, and that will probably take
> another 60 patches.
> I understand that reviewing 100+ patches is going to be a pain, but
> each patch currently averages about +/- 10 lines. The current
> diffstat is:
> 37 files changed, 723 insertions(+), 722 deletions(-)
> And that will probably double, so it's still going to be a fair
> amount of change.
Is there any sort of more coarse logical breakdown of this series, or do
we want/need to convert the entire codebase all at once? The individual
patches sound relatively small... is there a particular method at play
there? E.g., a patch per function? file? call chain?
> So the big question is how do we handle the review side of things.
> I think testing won't be a huge issue because of the time we have in
> the cycle (a couple of months to the 3.17 merge, and then a couple
> more months in the 3.17-rcX cycle) to find and catch regressions,
> but I'd like to know what people think about the best way to review
> this change will be.
> I'm happy for people to say "no, we need to review it patch by
> patch, so delay it for a cycle while we work through it", but I'm
> also happy for a "apply it all and look at error sources and
> inversion points for problems". The second is probably easier, as
> there will be very few remaining inversion points (only embedded
> errors in ioctl structures, I think) and all error sources should be
> negated at their definition and hence any error value (E* values)
> that are not defined as "-E*" is likely to be an mistake....
Personally, I generally prefer going through individual patches. Even if
we don't post the entire series and do a patch-by-patch reviewed-by
(which sounds like overkill in this case), that helps me do bits a time,
keep track of where I am, etc. I say that before I've seen any of these
patches of course ;), so I could certainly see running through some kind
of approach of doing it batches (i.e., "look at this sequence of
patches, identify the affected segments of code, make a direct pass
through that code, repeat"). I guess it's hard to say without just
digging in and finding the most effective approach to get through it.
Perhaps if we just make a branch available with the patches, put a
notification on the list, and we can use that as a review thread..?
> I'll be spending the next couple of days finishing this all off, so
> once it is done I can focus on review and bug fixes for the rest of
> the 3.17 dev cycle. That allows me to concentrate on a xfsprogs
> 3.2.1 release, subsequent userspace libxfs resync with the new
> kernel structure, and starting to work on some of the smart block
> device concepts I've been talking about recently....
> These are not concrete plans - just what I'd *like* to do in the
> next couple of months. Reality is bound to mess up any plans I have,
> so I figure I mays as well mess them up straight away....
> Dave Chinner
> xfs mailing list