[PATCH] Add explicit --src/dst-prefix to git-formt-patch in git-rebase.sh for the case of "diff.noprefix" in git-config

---
 git-rebase.sh |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/git-rebase.sh b/git-rebase.sh
index 7508463..e83a0cf 100755
--- a/git-rebase.sh
+++ b/git-rebase.sh

@@ -565,7 +565,7 @@ fi
 
 if test -z "$do_merge"
 then
-	git format-patch -k --stdout --full-index --ignore-if-in-upstream \
+	git format-patch -k --stdout --full-index --ignore-if-in-upstream --src-prefix=a/ --dst-prefix=b/ \
 		--no-renames $root_flag "$revisions" |
 	git am $git_am_opt --rebasing --resolvemsg="$RESOLVEMSG" &&
 	move_to_original_branch

-- 
1.6.4.4

On Wed, Sep 08, 2010 at 04:54:05PM +0300, Oded Shimon wrote:

Hi.

I may have rudely just sent a patch without explaining anything, I thought 
the patch was self explanitory. I'm not sure on the expected procedure 
here for sending patches.

When using "git rebase", when "diff.noprefix" is set to true in 
git-config, then all the rebases mess up because they get the directory 
path wrong.

As far as I can tell, my patch fixes this, with no side-effects that I can 
think of.

Can anyone comment on the patch? Can it be pushed upstream?

- ods15

Oded Shimon [off-list ref] wrote:

---

A shorter summary (subject) and more of a commit message body might be a
good idea. People tend to like patch summaries that fit into one line
on a terminal, and shorter is even better for those tools that output
additional fields on the same line.

Signoff is missing, please see Documentation/SubmittingPatches for
details (including the reason why we need one in the first place).

I can't comment on the patch itself since I'm not familiar with the
whole diff machinery nor the innards of rebase.

-Jan

For the case of "diff.noprefix" in git-config, git-format-patch should
still output diff with standard prefixes for git-am

Signed-off-by: Oded Shimon <redacted>
---
 git-rebase.sh |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/git-rebase.sh b/git-rebase.sh
index 7508463..e83a0cf 100755
--- a/git-rebase.sh
+++ b/git-rebase.sh

@@ -565,7 +565,7 @@ fi
 
 if test -z "$do_merge"
 then
-	git format-patch -k --stdout --full-index --ignore-if-in-upstream \
+	git format-patch -k --stdout --full-index --ignore-if-in-upstream --src-prefix=a/ --dst-prefix=b/ \
 		--no-renames $root_flag "$revisions" |
 	git am $git_am_opt --rebasing --resolvemsg="$RESOLVEMSG" &&
 	move_to_original_branch

-- 
1.6.4.4

format-patch read and used the diff UI config, such as diff.renames,
diff.noprefix and diff.mnemnoicprefix.  These have a history of
breaking rebase and patch application in general; cf. 840b3ca (rebase:
protect against diff.renames configuration, 2008-11-10).

Instead of continually putting more options inside git-rebase to avoid
these issues, this patch takes the stance that output from
format-patch is intended primarily for git-am and only as a side
effect also for human consumption.  Hence, ignore the diff UI config
entirely when coming from format-patch.

Note that all existing calls to git_log_config except for the one in
git_format_config use a NULL callback.

Reported-by: Oded Shimon <redacted>
Signed-off-by: Thomas Rast <redacted>
---

This is a bolder approach that just outright ignores the backwards
compatibility complaints Junio had in 840b3ca.  Among the variables
parsed in git_diff_ui_config, namely

  color.diff (and its legacy alias diff.color)
  diff.renames
  diff.autorefreshindex
  diff.mnemonicprefix
  diff.noprefix
  diff.external
  diff.wordregex
  diff.ignoresubmodules

arguably only diff.renames (and perhaps diff.ignoresubmodules, I don't
use them) should affect format-patch.  Everything else undermines the
guarantee (by having a consistent format) that format-patch|am works.

So now I'm not so sure about diff.renames.  Perhaps it needs to be
retained, but that requires a special case since we cannot move it to
git_diff_basic_config() (which affects diff-* plumbing too).

In any case I also made a test.  If you decide to go for the original
patch, please feel free to "steal" it.



 builtin/log.c     |   20 ++++++++++++++++++--
 t/t3400-rebase.sh |    7 +++++++
 2 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/builtin/log.c b/builtin/log.c
index f2d9d61..a1079fe 100644
--- a/builtin/log.c
+++ b/builtin/log.c

@@ -385,8 +385,19 @@ static int cmd_log_walk(struct rev_info *rev)
 	return diff_result_code(&rev->diffopt, 0);
 }
 
+struct log_config_cb_data
+{
+	/*
+	 * If no_diff_ui_config is set, we use diff_basic_config
+	 * instead, ignoring the plumbing-specific UI settings.
+	 */
+	int no_diff_ui_config;
+};
+
 static int git_log_config(const char *var, const char *value, void *cb)
 {
+	struct log_config_cb_data *cb_data = cb;
+
 	if (!strcmp(var, "format.pretty"))
 		return git_config_string(&fmt_pretty, var, value);
 	if (!strcmp(var, "format.subjectprefix"))

@@ -406,7 +417,10 @@ static int git_log_config(const char *var, const char *value, void *cb)
 	if (!prefixcmp(var, "color.decorate."))
 		return parse_decorate_color_config(var, 15, value);
 
-	return git_diff_ui_config(var, value, cb);
+	if (!cb_data || !cb_data->no_diff_ui_config)
+		return git_diff_ui_config(var, value, cb);
+	else
+		return git_diff_basic_config(var, value, cb);
 }
 
 int cmd_whatchanged(int argc, const char **argv, const char *prefix)

@@ -1099,6 +1113,7 @@ int cmd_format_patch(int argc, const char **argv, const char *prefix)
 	char *add_signoff = NULL;
 	struct strbuf buf = STRBUF_INIT;
 	int use_patch_format = 0;
+	struct log_config_cb_data config_cb_data;
 	const struct option builtin_format_patch_options[] = {
 		{ OPTION_CALLBACK, 'n', "numbered", &numbered, NULL,
 			    "use [PATCH n/m] even with a single patch",

@@ -1160,7 +1175,8 @@ int cmd_format_patch(int argc, const char **argv, const char *prefix)
 	extra_hdr.strdup_strings = 1;
 	extra_to.strdup_strings = 1;
 	extra_cc.strdup_strings = 1;
-	git_config(git_format_config, NULL);
+	config_cb_data.no_diff_ui_config = 1;
+	git_config(git_format_config, &config_cb_data);
 	init_revisions(&rev, prefix);
 	rev.commit_format = CMIT_FMT_EMAIL;
 	rev.verbose_header = 1;

diff --git a/t/t3400-rebase.sh b/t/t3400-rebase.sh
index 349eebd..0e2fe71 100755
--- a/t/t3400-rebase.sh
+++ b/t/t3400-rebase.sh

@@ -144,6 +144,13 @@ test_expect_success 'rebase is not broken by diff.renames' '
 	GIT_TRACE=1 git rebase force-3way
 '
 
+test_expect_success 'rebase is not broken by diff.noprefix' '
+	git config diff.noprefix true &&
+	test_when_finished "git config --unset diff.noprefix" &&
+	git checkout -b noprefix side &&
+	GIT_TRACE=1 git rebase master
+'
+
 test_expect_success 'setup: recover' '
 	test_might_fail git rebase --abort &&
 	git reset --hard &&

-- 
1.7.3.rc0.289.gcd076

Heya,

On Thu, Sep 9, 2010 at 03:36, Thomas Rast [off-list ref] wrote:

+test_expect_success 'rebase is not broken by diff.noprefix' '
+       git config diff.noprefix true &&
+       test_when_finished "git config --unset diff.noprefix" &&
+       git checkout -b noprefix side &&
+       GIT_TRACE=1 git rebase master
+'
+

Can we have one for 'git config diff.color true' too?

-- 
Cheers,

Sverre Rabbelier

On Thu, Sep 09, 2010 at 10:36:54AM +0200, Thomas Rast wrote:

format-patch read and used the diff UI config, such as diff.renames,
diff.noprefix and diff.mnemnoicprefix.  These have a history of
breaking rebase and patch application in general; cf. 840b3ca (rebase:
protect against diff.renames configuration, 2008-11-10).

Instead of continually putting more options inside git-rebase to avoid
these issues, this patch takes the stance that output from
format-patch is intended primarily for git-am and only as a side
effect also for human consumption.  Hence, ignore the diff UI config
entirely when coming from format-patch.

Note that all existing calls to git_log_config except for the one in
git_format_config use a NULL callback.

This was my first thought upon reading Oded's patch, too. We would want
to cut out anything that will cause format-patch to create a patch that
could not be applied. So from your list:

This is a bolder approach that just outright ignores the backwards
compatibility complaints Junio had in 840b3ca.  Among the variables
parsed in git_diff_ui_config, namely

  color.diff (and its legacy alias diff.color)
  diff.renames
  diff.autorefreshindex
  diff.mnemonicprefix
  diff.noprefix
  diff.external
  diff.wordregex
  diff.ignoresubmodules

arguably only diff.renames (and perhaps diff.ignoresubmodules, I don't
use them) should affect format-patch.  Everything else undermines the
guarantee (by having a consistent format) that format-patch|am works.

I would agree that diff.renames should probably be the only thing we
want to allow (because it is not about making a broken diff, but because
the receiver may or may not support it, and we already know that
git-rebase will handle it).

diff.external is debatable. If your external diff is producing real,
applicable diffs, then it is fine to use it. I have to wonder why you
would use an external diff, then. I guess because it's faster, or maybe
has an algorithm that produces equivalent but easier-to-read results
(e.g., patience before we had --patience)?

So now I'm not so sure about diff.renames.  Perhaps it needs to be
retained, but that requires a special case since we cannot move it to
git_diff_basic_config() (which affects diff-* plumbing too).

I think it is reasonable to just move an explicit "diff.renames" check
into format_patch, and then set the diff_options appropriately. It
requires special case code because it _is_ a special case.

-Peff

Thomas Rast [off-list ref] writes:

...
arguably only diff.renames (and perhaps diff.ignoresubmodules, I don't
use them) should affect format-patch.  Everything else undermines the
guarantee (by having a consistent format) that format-patch|am works.

We need to be a bit careful here.

Each user must be able to find a combination of ($opts1, $opts2) to make
"format-patch $opts1 | am $opts2" run correctly with his funny settings
(e.g. diff.noprefix).  We must guarantee that [*1*].

I however don't think we need to guarantee that the pipeline always works
for empty opts1/2, and certainly we shouldn't insist what flows in that
pipe must be the bog-standard -p1 with a/ b/ prefix patch.  For example,
in circles under svn influence, people may prefer opts1=--no-prefix, and
as long as the recipient understands that is the community norm around
there, he can run his "am" with -p0 and everything should work.  It is not
unreasonable for the sender to have diff.noprefix in the repository config
in such a setup, don't you think?

There is no way to easily affect what options the "format-patch | am"
pipeline uses inside rebase.  It may make sense to introduce --rebasing
option to format-patch to cause it to ignore any funny setting the user
might have, so that we don't have to keep adding options to the command
invocation.  "am" has --rebasing already, and it may be beneficial to
teach the codepath to defeat some configuration variables in a similar
way.


[Footnote]

*1* ... within reason.  For example, I don't think there is no opts2 if
you had opts1="--src-prefix=a/ --dst-prefix=b/c/" that makes the pipeline
work reasonably.