From: Al Viro <viro@...>
Subject: [PATCH] fix adbhid mismerge
 [1]Date: Oct 16, 8:02 pm 2007

Something really odd has happened: the last couple of changesets
have
-       int up_flag;
+       int keycode, up_flag;
and
-       int up_flag;
+       int up_flag, key;
in another, both in adb_input_keycode().  Even with -m passed to
git-whatchanged there's no sign of anything in that area.

Aside of trivial conflict resolution (see below), what's the right
way to trace the things like that?  Linus?

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
---
diff --git a/drivers/macintosh/adbhid.c b/drivers/macintosh/adbhid.c
index 8cce016..2766e4f 100644
--- a/drivers/macintosh/adbhid.c
+++ b/drivers/macintosh/adbhid.c
@@ -282,7 +282,7 @@ static void
 adbhid_input_keycode(int id, int scancode, int repeat)
 {
 	struct adbhid *ahid = adbhid[id];
-	int keycode, up_flag;
+	int keycode, up_flag, key;
 
 	keycode = scancode & 0x7f;
 	up_flag = scancode & 0x80;
-


From: Linus Torvalds <torvalds@...>
Subject: Re: [PATCH] fix adbhid mismerge
 [1]Date: Oct 16, 9:21 pm 2007

On Wed, 17 Oct 2007, Al Viro wrote:
>
> 	Something really odd has happened: the last couple of changesets
> have
> -       int up_flag;
> +       int keycode, up_flag;
> and
> -       int up_flag;
> +       int up_flag, key;
> in another, both in adb_input_keycode().  Even with -m passed to
> git-whatchanged there's no sign of anything in that area.

I don't think you did anything wrong. You used both --full-history 
(implicitly: git-whatchanged) and you made sure to see the diffs for both 
sides of any merge (-m), and that means that you should see every single 
diff involved.

Looking into it, the "key" variable was declared in the commit that 
introduced the new line

	int up_flag, key;
	..
	key = adbhid[id]->keycode[keycode];

which is commit 555ddbb4e2191c8823df2d61525218ac39481385. But then that 
declaration of "int key" goes away at some later time.

And doing a

	git whatchanged -p -m 555ddbb4e2191c8823df2d61525218ac39481385.. drivers/macintosh/adbhid.c

does actually show the culprit. It's just that the "-p -m" format is so 
damn unreadable that it's almost impossible to see.

Anyway, it's b981d8b3f5e008ff10d993be633ad00564fc22cd, which had a 
conflict in that file, and Dmitry apparently mis-merged it and edited the 
result down so that it didn't have 'key' declared any more.

So the way I found it was to just search for the line in the diffs that 
makes that thing go away, ie just look for the line in the diffs that says

	-	int up_flag, key;

and then you need to look at which of those are totally bogus and are just 
because it shows the diff against one of the earlier trees that also don't 
have that "key = adbhid[...]" line!

(And that is actually *much* less obvious than it should be, since a lot 
of the case of those lines going away is becuase I had merged Dmitry's 
tree in the first place)

You can make git help you narrow it down a bit more by using -S, ie some 
horrible command line from hell like this:

	git whatchanged -S'int up_flag, key;' -m -p 		\
		555ddbb4e2191c8823df2d61525218ac39481385..	\
		drivers/macintosh/adbhid.c

will actually show only those commits that add/remove a line like the one 
you are wondering where it went. That can cut down on the noise a bit, but 
you'll get all the same false alarms, so no, it's probably not worth it.

In general, I'm afraid that merge errors are simply not very easy to find. 
The problem in this case is in that b981d8b3f5 merge, but if you actually 
then do a "git show b981d8b3f5" you won't even see the problem spot in the 
default "--cc" format output, because Dmitry had resolved that problmatic 
place to be the same as one of the parent branches, and that makes git not 
show the diff for it (since it's "uninteresting")

			Linus
-


From: Linus Torvalds <torvalds@...>
Subject: Re: [PATCH] fix adbhid mismerge
 [1]Date: Oct 16, 10:21 pm 2007

On Tue, 16 Oct 2007, Linus Torvalds wrote:
> 
> I don't think you did anything wrong. You used both --full-history 
> (implicitly: git-whatchanged) and you made sure to see the diffs for both 
> sides of any merge (-m), and that means that you should see every single 
> diff involved.

Btw, if anybody can come up with a better way to find these kinds of 
mis-merges, I'd love to hear about it.

In *this* particular case, the -c flag ("combined" merge diff) probably 
comes closest, and is certainly a lot better than passing in -m (which 
shows each merge against both parents separately), and in fact, I think 
you would have found the mis-merge immediately if you had used

	git whatchanged -p -c v2.6.23.. drivers/macintosh/adbhid.c

but I'm not going to guarantee that -c always gives you what you want.

In general, the rules are:

 - the default for merge diffs is to show "condensed combined" merge, ie 
   the diff of only those parts where the result actively differs from 
   *both* parents.

   This is very terse, and it has the wonderful property of showing merges 
   where you actually ended up doing "real work" and not just picking one 
   side or the other, but in this case the very fact that the mis-merge 
   had picked one side (and it really would have _needed_ a correct manual 
   merge) also meant that the default "--cc" format didn't show anything 
   at all.

 - "-c" is for regular combined merges: any file that was modified in both 
   parents will show up as a combination of the diffs of both sides, while 
   a file that was taken in its *entirety* is ignored.

   In this case that's exactly what you wanted. It's just too noisy to 
   necessarily be the default, and you can still have a silent mis-merge 
   if the merger picked *only* one side.

   But in general, I suspect that "-c" is often a good thing to try if you 
   cannot find the cause of some change in a regular commit, and suspect a 
   merge error.

 - "-m" shows each side totally independently. Quite frankly, I've never 
   found it useful. It is essentially guaranteed to show all changes, 
   since it shows the patches against all parents individually, so even if 
   we took only one side, we'll still show the patch against the *other* 
   side, but quite frankly, while it's thus useful in theory, in practice 
   the end result is just too noisy to likely ever really be useful as 
   anything but a "yes, the information is there (..but it's practically 
   impossible to find for all the other noise that is also there)"

The main reason "-m" exists is historical: before Junio implemented the 
combined formats, -m was the easy way to show *any* information. I bet -m 
can be useful in some case where you have some pattern you can search for 
(ie I used -m in this case to find the mis-merge, but realized only later 
that I would have been better off using -c), but it's not something I'd 
recommend unless you were really desperate.

What I'd actually really like would be something that shows the original 
conflict, but that's really expensive to compute (it basically involves 
re-doing the merge from scratch - finding the proper base commit(s) etc). 
So we never did that.

		Linus
-