RE: Fatal error from git bisect

[Linus Torvalds <torvalds@osdl.org>]

On Thu, 9 Feb 2006, Mark E Mason wrote:
> 
> I thought of that, but even removing the workspace file doesn't get
> around this:

That doesn't necessarily help. Removing the workspace file potentially 
just makes the workspace even _more_ dirty, rather than less. It makes git 
think you've edited the file that it wants to check out another verion of.

(Of course, an extra file can _also_ be a sign of being dirty, so 
sometimes removing such a file _does_ generate a "more clean" tree).

Basically, "dirty" means that it doesn't match the head commit. 

> I've made no commits to this tree - it's just an up-to-date pull of the
> linux-mips.org tree.  I also have no local edits at this point.
> 
> [mason@hawaii linux.git]$ git diff
> [mason@hawaii linux.git]$
> 
> What's puzzling me is that (as I understand it) git bisect simply
> manipulates the workspace, not the repository.

Well, it doesn't _change_ the repository, but it does move around in it, 
using "git-read-tree -u -m". That can error out if the old tree and the 
new tree differs in a particular file, and the workspace doesn't match the 
old tree in that file.

Now, it could be a git bug, but before you go there, try to make sure tat 
your tree is really clean. Do a

	git checkout -f master

followed by "git status" to make sure that the workspace is clean (no 
unexpected untracked files, no diffs against HEAD, no nothing). The "git 
checkout -f master" should have cleaned everything up, but it won't 
actually touch extra files that it doesn't know about, so..

		Linus