[lkml]   [2008]   [May]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: GIT bisection range errors
On Fri, 9 May 2008, Ken Moffat wrote:

> On Thu, May 08, 2008 at 04:33:04PM -0700, Roland Dreier wrote:
> > > > $ git checkout -b rc v2.6.26-rc1
> > > > $ git bisect start
> > > > $ git bisect bad
> > > > $ git bisect good v2.6.25
> > > >
> > > > Yet, during this I'm finding myself at 2.6.25-rc6 and 2.6.25-rc8
> > > > as the last two results (both good...).
> >
> > > I reported a similar thing at the beginning of April
> > > - 2.6.25-rc1 bad, 2.6.24 good, git
> > > dropped me back at 2.6.24-rc4 (again, according to the top level
> > > Makefile).
> >
> > This is normal and expected, due to the distributed nature of git and
> > the fact that git-bisect operates on the full topology of history and
> > not just a linear sequence of commits.
> >
> > Imagine history like:
> >
> > A---B---C---D
> > \ /
> > \ /
> > \ /
> > E---F
> >
> > where B is good and D is bad. Now, when you bisect, there is no way to
> > know whether, say, E is good or bad and hence the bisect process may
> > present E as a tree to try.
> >
> > Now, if B is the 2.6.25 release, then since E branched off before B, it
> > will have a Makefile that says 2.6.25-rcX. Which is exactly the
> > behavior you are seeing.
> >
> > In short, everything looks fine and is behaving as expected.
> >
> > - R.
> But, surely those of us who bisect against linus' tree only
> care about the commits which made it into his tree, and in the
> context of whatever else was in _his_ tree at the time ?
> Maybe I'm under a misapprehension about changesets and merges. I
> thought a merge was just pulling in a series of changesets, and that
> each changeset only contains related items (comment, changed lines,
> added files, deleted files).

No, git tracks states and where they came from, not changes per se. That
is, when you look at a commit by David Miller, you're looking at exactly
the file contents that David Miller had when making the commit. Some other
systems linearize history such that what you'd see in Linus's tree is what
David Miller would have had if he'd made his changes to the tree Linus had
before merging David's branch, but that's not the normal thing to do with
git in this case.

> Whatever else may be in tree E, I don't expect it to have a commit
> which changes $EXTRAVERSION, purely because tree E is not Linus'
> tree. To me, that field is somewhat special - it indicates where I am
> (e.g. if bisecting across multiple rcs, or even across multiple
> releases) and it determines where the modules will go.

Tree E doesn't change versions; it's got the same version as A. But C and
D pick up the version change from B, which means that C and E have
different versions. You could also look at it like this: going back from D
to F changes the version, not because anybody on the lower path changed
it, but because Linus included the version change from his own side when
doing the merge.

> I see from Linus' reply to the original mail that this is indeed
> normal. That certainly isn't the word I would choose to use : we
> give things names to describe them and in this case the EXTRAVERSION
> appears to inhabit a parrallel universe to the pre-existing usage
> of "2.6.24 good, 2.6.25-rc1 bad". Colour me more confused than ever.

2.6.26-rc1 is bad, 2.6.25 is good, vanilla 2.6.25-rc1 is good, but some
modified version of 2.6.25-rc1 was bad. It's like if you take 2.6.25, and
you leave the Makefile the same but change some driver. You can find that
your 2.6.25 is now broken, while the original 2.6.25 is not. What's going
on in this bisect run is that Rene's seeing this same situation, but from
the perspective of looking back from the future and looking at somebody
else's state.

*This .sig left intentionally blank*

 \ /
  Last update: 2008-05-09 23:47    [W:0.115 / U:2.588 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site