threads / patch / 45595

patchpathspec: always honor `PATHSPEC_PREFIX_ORIGIN` flag

Subject: [PATCH] pathspec: always honor `PATHSPEC_PREFIX_ORIGIN` flag

## tl;dr

4 messages between Apr 3, 2017 and Apr 5, 2017. Diffs are folded; open one to read it.

replies: 3people: 3as markdown or json

Patrick Steinhardt· Apr 3, 2017, 08:03 UTC · lore

Previous to commit 5d8f084a5 (pathspec: simpler logic to prefix original pathspec elements, 2017-01-04), we were always using the computed `match` variable to perform pathspec matching whenever `PATHSPEC_PREFIX_ORIGIN` is set. This is for example useful when passing the parsed pathspecs to other commands, as the computed `match` may contain a pathspec relative to the repository root. The commit changed this logic to only do so when we do have an actual prefix and when literal pathspecs are deactivated.

But this change may actually break some commands which expect passed pathspecs to be relative to the repository root. One such case is `git add --patch`, which now fails when using relative paths from a subdirectory. For example if executing "git add -p ../foo.c" in a subdirectory, the `git-add--interactive` command will directly pass "../foo.c" to `git-ls-files`. As ls-files is executed at the repository's root, the command will notice that "../foo.c" is outside the repository and fail.

Fix the issue by again using the computed `match` variable whenever `PATHSPEC_PREFIX_ORIGIN` is set. This restores behavior previous to 5d8f084a5 and fixes interactive add.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 pathspec.c                 |  6 +++---
 t/t3701-add-interactive.sh | 20 ++++++++++++++++++++
 2 files changed, 23 insertions(+), 3 deletions(-)
Show changes to 2 files +23 −3

pathspec.c, t/t3701-add-interactive.sh

diff --git a/pathspec.c b/pathspec.c
index 303efda83..3193e45a6 100644
--- a/pathspec.c
+++ b/pathspec.c
@@ -504,12 +504,12 @@ static void init_pathspec_item(struct pathspec_item *item, unsigned flags,
 	 * Prefix the pathspec (keep all magic) and assign to
 	 * original. Useful for passing to another command.
 	 */
-	if ((flags & PATHSPEC_PREFIX_ORIGIN) &&
-	    prefixlen && !get_literal_global()) {
+	if (flags & PATHSPEC_PREFIX_ORIGIN) {
 		struct strbuf sb = STRBUF_INIT;
 
 		/* Preserve the actual prefix length of each pattern */
-		prefix_magic(&sb, prefixlen, element_magic);
+		if (prefixlen && !get_literal_global())
+			prefix_magic(&sb, prefixlen, element_magic);
 
 		strbuf_addstr(&sb, match);
 		item->original = strbuf_detach(&sb, NULL);
diff --git a/t/t3701-add-interactive.sh b/t/t3701-add-interactive.sh
index f9528fa00..640088dd6 100755
--- a/t/t3701-add-interactive.sh
+++ b/t/t3701-add-interactive.sh
@@ -436,6 +436,26 @@ test_expect_success 'add -p handles globs' '
 	test_cmp expect actual
 '
 
+test_expect_success 'add -p handles relative paths' '
+	git reset --hard &&
+
+	echo base >root.c &&
+	git add "*.c" &&
+	git commit -m base &&
+
+	echo change >root.c &&
+	mkdir -p subdir &&
+	git -C subdir add -p "../root.c" <<-\EOF &&
+	y
+	EOF
+
+	cat >expect <<-\EOF &&
+	root.c
+	EOF
+	git diff --cached --name-only >actual &&
+	test_cmp expect actual
+'
+
 test_expect_success 'add -p does not expand argument lists' '
 	git reset --hard &&
 
-- 
2.12.2
Brandon Williams· Apr 3, 2017, 16:26 UTC · re: Patrick Steinhardt · lore

Re: [PATCH] pathspec: always honor `PATHSPEC_PREFIX_ORIGIN` flag

On 04/03, Patrick Steinhardt wrote:
Show 48 quoted lines
> Previous to commit 5d8f084a5 (pathspec: simpler logic to prefix original
> pathspec elements, 2017-01-04), we were always using the computed
> `match` variable to perform pathspec matching whenever
> `PATHSPEC_PREFIX_ORIGIN` is set. This is for example useful when passing
> the parsed pathspecs to other commands, as the computed `match` may
> contain a pathspec relative to the repository root. The commit changed
> this logic to only do so when we do have an actual prefix and when
> literal pathspecs are deactivated.
> 
> But this change may actually break some commands which expect passed
> pathspecs to be relative to the repository root. One such case is `git
> add --patch`, which now fails when using relative paths from a
> subdirectory. For example if executing "git add -p ../foo.c" in a
> subdirectory, the `git-add--interactive` command will directly pass
> "../foo.c" to `git-ls-files`. As ls-files is executed at the
> repository's root, the command will notice that "../foo.c" is outside
> the repository and fail.
> 
> Fix the issue by again using the computed `match` variable whenever
> `PATHSPEC_PREFIX_ORIGIN` is set. This restores behavior previous to
> 5d8f084a5 and fixes interactive add.
> 
> Signed-off-by: Patrick Steinhardt <ps@pks.im>
> ---
>  pathspec.c                 |  6 +++---
>  t/t3701-add-interactive.sh | 20 ++++++++++++++++++++
>  2 files changed, 23 insertions(+), 3 deletions(-)
> 
> diff --git a/pathspec.c b/pathspec.c
> index 303efda83..3193e45a6 100644
> --- a/pathspec.c
> +++ b/pathspec.c
> @@ -504,12 +504,12 @@ static void init_pathspec_item(struct pathspec_item *item, unsigned flags,
>  	 * Prefix the pathspec (keep all magic) and assign to
>  	 * original. Useful for passing to another command.
>  	 */
> -	if ((flags & PATHSPEC_PREFIX_ORIGIN) &&
> -	    prefixlen && !get_literal_global()) {
> +	if (flags & PATHSPEC_PREFIX_ORIGIN) {
>  		struct strbuf sb = STRBUF_INIT;
>  
>  		/* Preserve the actual prefix length of each pattern */
> -		prefix_magic(&sb, prefixlen, element_magic);
> +		if (prefixlen && !get_literal_global())
> +			prefix_magic(&sb, prefixlen, element_magic);
>  
>  		strbuf_addstr(&sb, match);
>  		item->original = strbuf_detach(&sb, NULL);

Would it just make sense to drop the requirement that prefixlen be non-zero? My problem with this change currently is the ability to get an original string with is empty (ie "\0") which would cause git to throw some warnings about not allowing empty strings as pathspecs if they were then passed on to other processes.

Show 34 quoted lines
> diff --git a/t/t3701-add-interactive.sh b/t/t3701-add-interactive.sh
> index f9528fa00..640088dd6 100755
> --- a/t/t3701-add-interactive.sh
> +++ b/t/t3701-add-interactive.sh
> @@ -436,6 +436,26 @@ test_expect_success 'add -p handles globs' '
>  	test_cmp expect actual
>  '
>  
> +test_expect_success 'add -p handles relative paths' '
> +	git reset --hard &&
> +
> +	echo base >root.c &&
> +	git add "*.c" &&
> +	git commit -m base &&
> +
> +	echo change >root.c &&
> +	mkdir -p subdir &&
> +	git -C subdir add -p "../root.c" <<-\EOF &&
> +	y
> +	EOF
> +
> +	cat >expect <<-\EOF &&
> +	root.c
> +	EOF
> +	git diff --cached --name-only >actual &&
> +	test_cmp expect actual
> +'
> +
>  test_expect_success 'add -p does not expand argument lists' '
>  	git reset --hard &&
>  
> -- 
> 2.12.2
> 
-- 
Brandon Williams
Patrick Steinhardt· Apr 4, 2017, 09:17 UTC · re: Brandon Williams · lore

Re: [PATCH] pathspec: always honor `PATHSPEC_PREFIX_ORIGIN` flag

On Mon, Apr 03, 2017 at 09:26:48AM -0700, Brandon Williams wrote:
Show 55 quoted lines
> On 04/03, Patrick Steinhardt wrote:
> > Previous to commit 5d8f084a5 (pathspec: simpler logic to prefix original
> > pathspec elements, 2017-01-04), we were always using the computed
> > `match` variable to perform pathspec matching whenever
> > `PATHSPEC_PREFIX_ORIGIN` is set. This is for example useful when passing
> > the parsed pathspecs to other commands, as the computed `match` may
> > contain a pathspec relative to the repository root. The commit changed
> > this logic to only do so when we do have an actual prefix and when
> > literal pathspecs are deactivated.
> > 
> > But this change may actually break some commands which expect passed
> > pathspecs to be relative to the repository root. One such case is `git
> > add --patch`, which now fails when using relative paths from a
> > subdirectory. For example if executing "git add -p ../foo.c" in a
> > subdirectory, the `git-add--interactive` command will directly pass
> > "../foo.c" to `git-ls-files`. As ls-files is executed at the
> > repository's root, the command will notice that "../foo.c" is outside
> > the repository and fail.
> > 
> > Fix the issue by again using the computed `match` variable whenever
> > `PATHSPEC_PREFIX_ORIGIN` is set. This restores behavior previous to
> > 5d8f084a5 and fixes interactive add.
> > 
> > Signed-off-by: Patrick Steinhardt <ps@pks.im>
> > ---
> >  pathspec.c                 |  6 +++---
> >  t/t3701-add-interactive.sh | 20 ++++++++++++++++++++
> >  2 files changed, 23 insertions(+), 3 deletions(-)
> > 
> > diff --git a/pathspec.c b/pathspec.c
> > index 303efda83..3193e45a6 100644
> > --- a/pathspec.c
> > +++ b/pathspec.c
> > @@ -504,12 +504,12 @@ static void init_pathspec_item(struct pathspec_item *item, unsigned flags,
> >  	 * Prefix the pathspec (keep all magic) and assign to
> >  	 * original. Useful for passing to another command.
> >  	 */
> > -	if ((flags & PATHSPEC_PREFIX_ORIGIN) &&
> > -	    prefixlen && !get_literal_global()) {
> > +	if (flags & PATHSPEC_PREFIX_ORIGIN) {
> >  		struct strbuf sb = STRBUF_INIT;
> >  
> >  		/* Preserve the actual prefix length of each pattern */
> > -		prefix_magic(&sb, prefixlen, element_magic);
> > +		if (prefixlen && !get_literal_global())
> > +			prefix_magic(&sb, prefixlen, element_magic);
> >  
> >  		strbuf_addstr(&sb, match);
> >  		item->original = strbuf_detach(&sb, NULL);
> 
> Would it just make sense to drop the requirement that prefixlen be
> non-zero?  My problem with this change currently is the ability to get
> an original string with is empty (ie "\0") which would cause git to
> throw some warnings about not allowing empty strings as pathspecs if
> they were then passed on to other processes.
You're right. My patch results in:

$ git init repo $ cd repo $ touch file $ git add file $ git commit -mfile $ echo foo>file $ mkdir subdir $ cd subdir $ git add -p .. warning: empty strings as pathspecs will be made invalid... warning: empty strings as pathspecs will be made invalid...

Dropping only the `prefixlen` condition and then unconditionally calling `prefix_magic` suffices to fix the actual bug. I've improved the test to use a pathspec which resolves to "" and check stderr.

Thanks for catching this!

Regards Patrick

Duy Nguyen· Apr 5, 2017, 13:40 UTC · re: Brandon Williams · lore

Re: [PATCH] pathspec: always honor `PATHSPEC_PREFIX_ORIGIN` flag

On Mon, Apr 3, 2017 at 11:26 PM, Brandon Williams <bmwill@google.com> wrote:
Show 18 quoted lines
> On 04/03, Patrick Steinhardt wrote:
>> Previous to commit 5d8f084a5 (pathspec: simpler logic to prefix original
>> pathspec elements, 2017-01-04), we were always using the computed
>> `match` variable to perform pathspec matching whenever
>> `PATHSPEC_PREFIX_ORIGIN` is set. This is for example useful when passing
>> the parsed pathspecs to other commands, as the computed `match` may
>> contain a pathspec relative to the repository root. The commit changed
>> this logic to only do so when we do have an actual prefix and when
>> literal pathspecs are deactivated.
>>
>> But this change may actually break some commands which expect passed
>> pathspecs to be relative to the repository root. One such case is `git
>> add --patch`, which now fails when using relative paths from a
>> subdirectory. For example if executing "git add -p ../foo.c" in a
>> subdirectory, the `git-add--interactive` command will directly pass
>> "../foo.c" to `git-ls-files`. As ls-files is executed at the
>> repository's root, the command will notice that "../foo.c" is outside
>> the repository and fail.
Oops. Sorry I missed this.
Show 22 quoted lines
>> @@ -504,12 +504,12 @@ static void init_pathspec_item(struct pathspec_item *item, unsigned flags,
>>        * Prefix the pathspec (keep all magic) and assign to
>>        * original. Useful for passing to another command.
>>        */
>> -     if ((flags & PATHSPEC_PREFIX_ORIGIN) &&
>> -         prefixlen && !get_literal_global()) {
>> +     if (flags & PATHSPEC_PREFIX_ORIGIN) {
>>               struct strbuf sb = STRBUF_INIT;
>>
>>               /* Preserve the actual prefix length of each pattern */
>> -             prefix_magic(&sb, prefixlen, element_magic);
>> +             if (prefixlen && !get_literal_global())
>> +                     prefix_magic(&sb, prefixlen, element_magic);
>>
>>               strbuf_addstr(&sb, match);
>>               item->original = strbuf_detach(&sb, NULL);
>
> Would it just make sense to drop the requirement that prefixlen be
> non-zero?  My problem with this change currently is the ability to get
> an original string with is empty (ie "\0") which would cause git to
> throw some warnings about not allowing empty strings as pathspecs if
> they were then passed on to other processes.
Good catch.

I did wonder if it's a right thing, because the result pathspec is ':(prefix:0)'. After leaving out the magic path, you get an empty "path" of the pathspec, which probably should be warned because we wouldn't be able to handle it. For example, "git add ." at root will give you the path ".", not empty one. Maybe we can't handle the empty "path" part.

But we have an exception for this already. In "git add :/", the "path" of the pathspec is still empty ('/' is a magic, not "path" even if it looks like so) and we handle it just fine. So everything should be good here.

Patrick, please add a line or two about why you drop prefixlen when you re-roll (or even better, make it a separate patch; this sounds like an issue even before the code reorganization changes).

-- 
Duy

← back to recent threads