[RFH] WTF did we change in git-fsck recently?

I noticed that "git-fsck --full" from 'master' takes forever to fsck the kernel repository (I left it running for 2 hours before killing it), while the one from 'maint' (or 1.5.1.1 which is installed on kernel.org) finishes within 2 or 3 minutes. There is some serious breakages there.

Rings a bell?

On Fri, 20 Apr 2007, Junio C Hamano wrote:

Hmm. Probably something broken in my "object decorator" thing then. 

Will check it out, I hadn't noticed myself.

		Linus

On Fri, 20 Apr 2007, Linus Torvalds wrote:

Duh.

When I did the object decorator thing, I made the "loop over the hash" function use the same logic for updating the hash, ie made them use

	if (++j >= size)
		j = 0;

for both the hash update for both "insert" and "lookup"

HOWEVER.

For some inexplicable reason I had an extraneous

	j++;

in the insert path (probably just from the fact that the old code there used

	j++;
	if (j >= size)
		j = 0;

and when I made them use the same logic I just didn't remove the old extraneous line properly.

This fixes it.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 decorate.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

Linus Torvalds <torvalds@linux-foundation.org> writes:

Thanks.

Learning from example by Steven Grimm, let me share a success story.

Earlier I noticed that "fsck --full" from 'master' took forever in linux-2.6 repository, but the one from 'maint' finished in 2 to 3 minutes.

We recently had a few enhancements by Christian Couder to git-bisect, and this was a perfect opportunity to see how well they worked:

(1) "git bisect start" now takes one bad and then one or more
    good commits, before suggesting the first revision to try.

Traditionally, immediately after you gave a bad and a good commit, it did a single bisection and then a checkout. This avoids repeated bisect computation and checkout when you know more than one good revisions before starting to bisect, and also let you bootstrap with a single command (you could instead give one good commit at a time and then finally a single bad commit to avoid the waste).

Not only I know 'maint' is good, I also know that the tips of "foreign projects" merged to git.git, that do not share any codepath the fsck takes, are irrelevant to the problem. So I want to mark tips of commit ancestry I merged from git-gui projects as good. Hence:

	$ git bisect start master maint remotes/git-gui/master

Mnemonic. Start takes a Bad before Goods, because B comes before G.

(2) "git bisect run <script>" takes a script to judge the
    goodness of the given revision.  Because I know each round
    of test takes around 3 minutes, I wrote a little script to
    automate the process and gave it to "git bisect run":

	$ git bisect run ./+run-script

This ran for a while (I do not know how long it took -- I was away from the machine and doing other things) and came back with the "object decoration" one Linus has fixed yesterday with his patch.

Here is the "+run-script". I have git.git repository and linux-2.6 repository next to each other.

-- >8 -- #!/bin/sh

# Build errors are not what I am interested in. make git-fsck && cd ../linux-2.6 || exit 255

# We are checking if it stops in a reasonable amount of time, so # let it run in the background...

../git.git/git-fsck --full >:log 2>&1 &

# ... and grab its process ID. fsck=$!

# ... and then wait for sufficiently long. sleep 240

# ... and then see if the process is still there.
if kill -0 $fsck
then
	# It is still running -- that is bad.
        # Three-kill is just a ritual and has no real meaning.
        # It is like "sync;sync;sync;reboot".
	kill $fsck; sleep 1; kill $fsck; sleep 1; kill $fsck;
	exit 1
else
	# It has already finished (the $fsck process was no more),
        # and we are happy.
	exit 0
fi