[Top] [All Lists]

Re: [ANNOUNCE, DISCUSS] xfsprogs: libxfs-4.1-update branch created

To: Brian Foster <bfoster@xxxxxxxxxx>
Subject: Re: [ANNOUNCE, DISCUSS] xfsprogs: libxfs-4.1-update branch created
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 11 May 2015 23:20:09 +1000
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20150511123917.GA43723@xxxxxxxxxxxxxxx>
References: <20150511000508.GD16689@dastard> <20150511123917.GA43723@xxxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Mon, May 11, 2015 at 08:39:18AM -0400, Brian Foster wrote:
> On Mon, May 11, 2015 at 10:05:08AM +1000, Dave Chinner wrote:
> > Hi folks,
> > 
> > I' ve just added a branch to the xfsprogs repository on kernel.org
> > that is up to date with the current 4.1-rc2 kernel code. It's smoke
> > tested, so I don't think there's any major problems with it. Can
> > everyone with development patches (i.e. things that change on disk
> > formats) for xfsprogs please rebase their work against this branch,
> > retest and repost? This branch will end up being the 3.3 release
> > tree...
> > 
> > FWIW, now that I've done this update, I'm considered how we keep
> > this up to date in future. Keeping the userspace code up to date and
> > working with all the internal kernel API changes takes quite a bit
> > of effort and that currently is all being done by the Maintainer.
> > i.e. me. This is the reason that userspace is not updated
> > frequently.
> > 
> > With this update, libxfs in the kernel and userspace are mostly
> > identical, so I'm thinking that a new rule for XFS developers needs
> > to be put in place: if you modify a libxfs kernel API, then you need
> > to make the same change to userspace. This means all the changes
> > needed for xfs_repair, xfs_db, etc, need to be done as well, because
> > sometimes these things are non-trivial.
> > 
> > A good example of this is Christoph's xfs_trans_commit() patchset.
> > There's 42 xfs_trans_commit() calls in userspace and it has a
> > completely different implementation in userspace. To put that in
> > context, there are 61 xfs_trans_commit() in the kernel code. If I
> > merge that changeset into the kernel code, then Christoph walks away
> > and it has become the Maintainer's problem to make userspace work
> > with the API change.
> > 
> > This *sucks* from a maintainer's POV - - it simply does not scale.
> > This is one of the reasons why my productivity as an XFS developer
> > has cratered over the last year. This userspace update work needs to
> > be distributed over the developers making the API and functionality
> > changes, so I'm throwing this out there to see what people think
> > about solving this problem.
> > 
> > I'm thinking that we need a userspace dev branch similar to the
> > for-next branch for the kernel code, where we aggregate topic
> > branches that contain each set of userspace patches that add/change
> > functionality that is shared with the kernel. If the developer
> > hasn't done the work to update userspace and test the changes there,
> > then the change is incomplete.
> > 
> > Note that I'm only suggesting this for the main XFS developers
> > making internal API or on-disk format changes to XFS - the typical
> > drive-by patches and bug fixes from random people we commit to the
> > kernel code will still need the maintainer to push them across to
> > userspace.
> > 
> > This will make my life easier from both an update and maintenance
> > point of view, and will make it easier for us as developers to get
> > the latest dev code faster. It will also make porting kernel changes
> > over to userspace faster, as the patches will mostly apply cleanly
> > to the userspace libxfs code base. And if it gets merged quicker,
> > tehn we'll get better test coverage of the dev tree because we'll be
> > testing it sooner. And as developers, we won't have to be working
> > out what changed and needs forward porting from the kernel to
> > userspace to make stuff work....
> > 
> > So, what do the people I'm asking to do more work so I don't have
> > to spend so much time on time consuming maintenance tasks think?
> > 
> This seems reasonable to me. As it is, I'd expect to do this anyways for
> anything that changes on-disk format. Such a change isn't really
> complete if mkfs/xfs_repair/xfs_db can't handle it appropriately.
> Provided we keep the development branch up to date, I don't see much
> harm in extending that to libxfs API level changes as well.

Right - the point of this is making sure we keep userspace up to
date, rather than doing an update once every blue moon and it being
a lot of work to do.

> What's the proposition with regard to submission/review process? I don't
> think we necessarily need the userspace bits until there is some review
> feedback on the kernel bits because that just increases development and
> review overhead (though nothing precludes posting both, of course). Also

Yes, that's fine - the userspace libxfs bits should be a straight
conversion of the kernel patches with a simple bit of path
stripping. e.g. the kernel patch path will be a/fs/xfs/libxfs/....
and the xfsprogs patch path is a/libxfs/...

Hence the same patch should apply, simply with a -p 5 (?) path strip
level rather than -p 1. i.e. `patch -p 5 < kernel.diff` should
simply apply the kernel patches to the xfsprogs libxfs.

> (and I think we discussed this briefly at LSF), I assume it is
> reasonable to condense a kernel patch series to a single userspace "sync
> XYZ feature to xfsprogs" patch for the bits that port directly over,
> since we have the kernel git log for finer grained history..?

Haven't really thought about it. Fine grained history is more
important for the kernel, but for userspace I can still see it being
useful for things like API changes where the same set of steps made
to the kernel code to make review easy should be made to make the
xfs_repair/db/... changes easy to review.

> Case in
> point: I could squash the sparse inode kernel patches into a single
> xfsprogs patch. The functional xfsprogs bits on top of that (e.g., mkfs,
> repair, etc.) would of course remain as independent patches that require
> indepenent review.

Right, for something that is all new code this can be done. It does
mean we'll have to be vigilant about merging bug fixes into the
userspace code, though. Those are really only picked up during
the large merges right now...

> Something to note here is that I think this might warrant a bit more
> positive feedback with regard to when things are considered reviewed on
> the kernel side (i.e., merge is imminent). Perhaps we could create the
> kernel topic branch but not merge it until the userspace bits are
> available..? Just a thought, whatever works best for you is probably
> fine provided the developer is made clear where things stand.

Well, for on-disk format changes we'll need the userspace code at
tha same time so that reviewers can test the changes, but for
everything else a port from the topic branch that gets merged into
the for-next tree is probably fine from a timing POV.

What I'd really like to is for all the changes merged in a kernel
-rc1 window are already merged into the userspace tree, or are at
least well on their way to being merged. That way we can even start
to think about a simple rolling release of xfsprogs that mirrors
the timing of the kernel release cycle. That way for new features
we get userspace support released at the same time that kernel
support is released....

The other benefit of doing this is that we get the changes committed
while it is still fresh in our minds - we don't have to look back to
code written 6-9 months ago and scratch our heads wondering why the
obvious merge changes aren't working because we've forgotten some
little detail about the change....


Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>