diff mbox series

[v3,2/2] attr: add attr.tree for setting the treeish to read attributes from

Message ID dadb822da99772cd277417f564cf672f65d1cc24.1696967380.git.gitgitgadget@gmail.com (mailing list archive)
State Superseded
Headers show
Series attr: add attr.tree config | expand

Commit Message

John Cai Oct. 10, 2023, 7:49 p.m. UTC
From: John Cai <johncai86@gmail.com>

44451a2 (attr: teach "--attr-source=<tree>" global option to "git",
2023-05-06) provided the ability to pass in a treeish as the attr
source. In the context of serving Git repositories as bare repos like we
do at GitLab however, it would be easier to point --attr-source to HEAD
for all commands by setting it once.

Add a new config attr.tree that allows this.

Signed-off-by: John Cai <johncai86@gmail.com>
---
 Documentation/config.txt      |  2 ++
 Documentation/config/attr.txt |  5 +++
 attr.c                        |  7 ++++
 attr.h                        |  2 ++
 config.c                      | 14 ++++++++
 t/t0003-attributes.sh         | 62 +++++++++++++++++++++++++++++++++++
 6 files changed, 92 insertions(+)
 create mode 100644 Documentation/config/attr.txt

Comments

Junio C Hamano Oct. 10, 2023, 10:14 p.m. UTC | #1
"John Cai via GitGitGadget" <gitgitgadget@gmail.com> writes:

> From: John Cai <johncai86@gmail.com>
>
> 44451a2 (attr: teach "--attr-source=<tree>" global option to "git",
> 2023-05-06) provided the ability to pass in a treeish as the attr
> source. In the context of serving Git repositories as bare repos like we
> do at GitLab however, it would be easier to point --attr-source to HEAD
> for all commands by setting it once.
>
> Add a new config attr.tree that allows this.
>
> Signed-off-by: John Cai <johncai86@gmail.com>
> ---
>  Documentation/config.txt      |  2 ++
>  Documentation/config/attr.txt |  5 +++
>  attr.c                        |  7 ++++
>  attr.h                        |  2 ++
>  config.c                      | 14 ++++++++
>  t/t0003-attributes.sh         | 62 +++++++++++++++++++++++++++++++++++
>  6 files changed, 92 insertions(+)
>  create mode 100644 Documentation/config/attr.txt
>
> diff --git a/Documentation/config.txt b/Documentation/config.txt
> index 229b63a454c..b1891c2b5af 100644
> --- a/Documentation/config.txt
> +++ b/Documentation/config.txt
> @@ -371,6 +371,8 @@ other popular tools, and describe them in your documentation.
>  
>  include::config/advice.txt[]
>  
> +include::config/attr.txt[]
> +
>  include::config/core.txt[]
>  
>  include::config/add.txt[]
> diff --git a/Documentation/config/attr.txt b/Documentation/config/attr.txt
> new file mode 100644
> index 00000000000..be882523f8b
> --- /dev/null
> +++ b/Documentation/config/attr.txt
> @@ -0,0 +1,5 @@
> +attr.tree:
> +	A <tree-ish> to read gitattributes from instead of the worktree. See
> +	linkgit:gitattributes[5]. If `attr.tree` does not resolve to a valid tree,
> +	treat it as an empty tree. --attr-source and GIT_ATTR_SOURCE take
> +	precedence over attr.tree.

Properly typeset `--attr-source` and `GIT_ATTR_SOURCE`.

A quick "git grep" in Documentation/config/*.txt tells me that
nobody refers to an object type like <tree-ish>.  Imitate what this
was modeled after, namely Documentation/config/mailmap.txt, which
says just

	... a reference to a blob in the repository.

without any half mark-up.

More importantly, the description makes one wonder what the
precedence rule between these two (the general rule would be for
command line parameter to override environment, if I recall
correctly).

I think the enumeration header usually is followed by double-colons
among Documentation/config/*.txt files.  Let's be consistent.

In the context of this expression, "worktree" is a wrong noun to
use---the term of art refers to an instance of "working tree",
together with some "per worktree" administrative files inside .git/
directory.  On the other hand, "working tree" refers to the "files
meant to be visible to build tools and editors part of the non-bare
repository", which is what you want to use here.

attr.tree::
	A tree object to read the attributes from, instead of the
	`.gitattributes` file in the working tree.  In a bare
	repository, this defaults to HEAD:.gitattributes".  If a
	given value does not resolve to a valid tree object, an
	empty tree is used instead.  When `GIT_ATTR_SOURCE`
	environment variable or `--attr-source` command line option
	is used, this configuration variable has no effect.

or something along that line, perhaps.

> diff --git a/attr.c b/attr.c
> index bf2ea1626a6..0ae6852d12b 100644
> --- a/attr.c
> +++ b/attr.c
> @@ -24,6 +24,8 @@
>  #include "tree-walk.h"
>  #include "object-name.h"
>  
> +const char *git_attr_tree;
> +
>  const char git_attr__true[] = "(builtin)true";
>  const char git_attr__false[] = "\0(builtin)false";
>  static const char git_attr__unknown[] = "(builtin)unknown";
> @@ -1206,6 +1208,11 @@ static void compute_default_attr_source(struct object_id *attr_source)
>  	if (!default_attr_source_tree_object_name)
>  		default_attr_source_tree_object_name = getenv(GIT_ATTR_SOURCE_ENVIRONMENT);
>  
> +	if (!default_attr_source_tree_object_name) {
> +		default_attr_source_tree_object_name = git_attr_tree;
> +		ignore_bad_attr_tree = 1;
> +	}

As long as "attr.tree" was read by calling git_default_attr_config()
before we come here, git_attr_tree is not NULL and we allow bad attr
tree in default_attr_source_tree_object_name.  But stepping back a
bit, even if "attr.tree" is unspecified, i.e., git_attr_tree is
NULL, we set ignore_bad_attr_tree to true here.

What it means is that after the above if() statement, if
default_attr_source_tree_object_name is still NULL, we know that
ignore_bad_attr_tree is already set to true. 

>  	if (!default_attr_source_tree_object_name &&
>  	    startup_info->have_repository &&
>  	    is_bare_repository()) {

So would it make more sense to remove the assignment to the same
variable we made in [1/2] around here (not seen in the post
context)?

Alternatively, even though it makes the code a bit more verbose, the
logic might become clearer if you wrote the "assign from the config"
part like so:

	if (!default_attr_source_tree_object_name && git_attr_tree) {
        	default_attr_source_tree_object_name = git_attr_tree;
		ignore_bad_attr_tree = 1;
	}

It would leave more flexibility to the code around here.  You could
for example add code that assigns a different value, a tree object
that is required to exist, to default_attr_source_tree_object_name
after this point, for example, without having to wonder what the
"current" value of ignore_bad_attr_tree is.

> +static int git_default_attr_config(const char *var, const char *value)
> +{
> +	if (!strcmp(var, "attr.tree"))
> +		return git_config_string(&git_attr_tree, var, value);
> +
> +	/* Add other attribute related config variables here and to
> +	   Documentation/config/attr.txt. */

	/*
	 * Our multi-line comments should look
         * more like this; opening slash-asterisk
	 * and closing asterisk-slash sit on a line
	 * on its own.
	 */

> @@ -342,6 +346,46 @@ test_expect_success 'bare repository: check that .gitattribute is ignored' '
>  	)
>  '
>  
> +bad_attr_source_err="fatal: bad --attr-source or GIT_ATTR_SOURCE"

Not a fault of this two-patch series, but we probably should refine
this error reporting so that the reader can tell which one is being
complained about, and optionally what the offending value was.


> +test_expect_success 'attr.tree when HEAD is unborn' '
> +	test_when_finished rm -rf empty &&
> +	git init empty &&
> +	(
> +		cd empty &&
> +		echo $bad_attr_source_err >expect_err &&

Let's not rely on words in the error message split with exactly one
whitespace each and instead quote the variable properly.  I.e.,

		echo "$bad_attr_source_err" >expect_err &&

But this is not even used, as we do not expect it to fail.  Perhaps
remove it altogether?

> +		echo "f/path: test: unspecified" >expect &&
> +		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
> +		test_must_be_empty err &&
> +		test_cmp expect actual
> +	)
> +'
> +
> +test_expect_success 'attr.tree points to non-existing ref' '
> +	test_when_finished rm -rf empty &&
> +	git init empty &&
> +	(
> +		cd empty &&
> +		echo $bad_attr_source_err >expect_err &&

Ditto.

> +		echo "f/path: test: unspecified" >expect &&
> +		git -c attr.tree=refs/does/not/exist check-attr test -- f/path >actual 2>err &&
> +		test_must_be_empty err &&
> +		test_cmp expect actual
> +	)
> +'
> +
> +test_expect_success 'bad attr source defaults to reading .gitattributes file' '
> +	test_when_finished rm -rf empty &&
> +	git init empty &&
> +	(
> +		cd empty &&
> +		echo "f/path test=val" >.gitattributes &&
> +		echo "f/path: test: val" >expect &&
> +		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
> +		test_must_be_empty err &&
> +		test_cmp expect actual
> +	)
> +'

In other words, with the additional tests, we do not check error
cases (which may be perfectly OK, if they are covered by existing
tests).  A bit curious.

> @@ -356,6 +400,24 @@ test_expect_success 'bare repo defaults to reading .gitattributes from HEAD' '
>  	test_cmp expect actual
>  '
>  
> +test_expect_success '--attr-source and GIT_ATTR_SOURCE take precedence over attr.tree' '

Do we want to ensure which one takes precedence between the command
line option and the environment?  It's not like the one that is
given the last takes effect.

> +	test_when_finished rm -rf empty &&
> +	git init empty &&
> +	(
> +		cd empty &&
> +		git checkout -b attr-source &&
> +		test_commit "val1" .gitattributes "f/path test=val1" &&
> +		git checkout -b attr-tree &&
> +		test_commit "val2" .gitattributes "f/path test=val2" &&
> +		git checkout attr-source &&
> +		echo "f/path: test: val1" >expect &&
> +		git -c attr.tree=attr-tree --attr-source=attr-source check-attr test -- f/path >actual &&
> +		test_cmp expect actual &&
> +		GIT_ATTR_SOURCE=attr-source git -c attr.tree=attr-tree check-attr test -- f/path >actual &&
> +		test_cmp expect actual
> +	)
> +'
> +
>  test_expect_success 'bare repository: with --source' '
>  	(
>  		cd bare.git &&

Other than that, looking great.  Thanks.
John Cai Oct. 11, 2023, 2:19 a.m. UTC | #2
Hi Junio,

On 10 Oct 2023, at 18:14, Junio C Hamano wrote:

> "John Cai via GitGitGadget" <gitgitgadget@gmail.com> writes:
>
>> From: John Cai <johncai86@gmail.com>
>>
>> 44451a2 (attr: teach "--attr-source=<tree>" global option to "git",
>> 2023-05-06) provided the ability to pass in a treeish as the attr
>> source. In the context of serving Git repositories as bare repos like we
>> do at GitLab however, it would be easier to point --attr-source to HEAD
>> for all commands by setting it once.
>>
>> Add a new config attr.tree that allows this.
>>
>> Signed-off-by: John Cai <johncai86@gmail.com>
>> ---
>>  Documentation/config.txt      |  2 ++
>>  Documentation/config/attr.txt |  5 +++
>>  attr.c                        |  7 ++++
>>  attr.h                        |  2 ++
>>  config.c                      | 14 ++++++++
>>  t/t0003-attributes.sh         | 62 +++++++++++++++++++++++++++++++++++
>>  6 files changed, 92 insertions(+)
>>  create mode 100644 Documentation/config/attr.txt
>>
>> diff --git a/Documentation/config.txt b/Documentation/config.txt
>> index 229b63a454c..b1891c2b5af 100644
>> --- a/Documentation/config.txt
>> +++ b/Documentation/config.txt
>> @@ -371,6 +371,8 @@ other popular tools, and describe them in your documentation.
>>
>>  include::config/advice.txt[]
>>
>> +include::config/attr.txt[]
>> +
>>  include::config/core.txt[]
>>
>>  include::config/add.txt[]
>> diff --git a/Documentation/config/attr.txt b/Documentation/config/attr.txt
>> new file mode 100644
>> index 00000000000..be882523f8b
>> --- /dev/null
>> +++ b/Documentation/config/attr.txt
>> @@ -0,0 +1,5 @@
>> +attr.tree:
>> +	A <tree-ish> to read gitattributes from instead of the worktree. See
>> +	linkgit:gitattributes[5]. If `attr.tree` does not resolve to a valid tree,
>> +	treat it as an empty tree. --attr-source and GIT_ATTR_SOURCE take
>> +	precedence over attr.tree.
>
> Properly typeset `--attr-source` and `GIT_ATTR_SOURCE`.
>
> A quick "git grep" in Documentation/config/*.txt tells me that
> nobody refers to an object type like <tree-ish>.  Imitate what this
> was modeled after, namely Documentation/config/mailmap.txt, which
> says just
>
> 	... a reference to a blob in the repository.
>
> without any half mark-up.

sounds good

>
> More importantly, the description makes one wonder what the
> precedence rule between these two (the general rule would be for
> command line parameter to override environment, if I recall
> correctly).

yes, that should be the case.

>
> I think the enumeration header usually is followed by double-colons
> among Documentation/config/*.txt files.  Let's be consistent.
>
> In the context of this expression, "worktree" is a wrong noun to
> use---the term of art refers to an instance of "working tree",
> together with some "per worktree" administrative files inside .git/
> directory.  On the other hand, "working tree" refers to the "files
> meant to be visible to build tools and editors part of the non-bare
> repository", which is what you want to use here.
>
> attr.tree::
> 	A tree object to read the attributes from, instead of the
> 	`.gitattributes` file in the working tree.  In a bare
> 	repository, this defaults to HEAD:.gitattributes".  If a
> 	given value does not resolve to a valid tree object, an
> 	empty tree is used instead.  When `GIT_ATTR_SOURCE`
> 	environment variable or `--attr-source` command line option
> 	is used, this configuration variable has no effect.
>
> or something along that line, perhaps.

sounds good to me, thanks

>
>> diff --git a/attr.c b/attr.c
>> index bf2ea1626a6..0ae6852d12b 100644
>> --- a/attr.c
>> +++ b/attr.c
>> @@ -24,6 +24,8 @@
>>  #include "tree-walk.h"
>>  #include "object-name.h"
>>
>> +const char *git_attr_tree;
>> +
>>  const char git_attr__true[] = "(builtin)true";
>>  const char git_attr__false[] = "\0(builtin)false";
>>  static const char git_attr__unknown[] = "(builtin)unknown";
>> @@ -1206,6 +1208,11 @@ static void compute_default_attr_source(struct object_id *attr_source)
>>  	if (!default_attr_source_tree_object_name)
>>  		default_attr_source_tree_object_name = getenv(GIT_ATTR_SOURCE_ENVIRONMENT);
>>
>> +	if (!default_attr_source_tree_object_name) {
>> +		default_attr_source_tree_object_name = git_attr_tree;
>> +		ignore_bad_attr_tree = 1;
>> +	}
>
> As long as "attr.tree" was read by calling git_default_attr_config()
> before we come here, git_attr_tree is not NULL and we allow bad attr
> tree in default_attr_source_tree_object_name.  But stepping back a
> bit, even if "attr.tree" is unspecified, i.e., git_attr_tree is
> NULL, we set ignore_bad_attr_tree to true here.
>
> What it means is that after the above if() statement, if
> default_attr_source_tree_object_name is still NULL, we know that
> ignore_bad_attr_tree is already set to true.
>
>>  	if (!default_attr_source_tree_object_name &&
>>  	    startup_info->have_repository &&
>>  	    is_bare_repository()) {
>
> So would it make more sense to remove the assignment to the same
> variable we made in [1/2] around here (not seen in the post
> context)?
>
> Alternatively, even though it makes the code a bit more verbose, the
> logic might become clearer if you wrote the "assign from the config"
> part like so:
>
> 	if (!default_attr_source_tree_object_name && git_attr_tree) {
>         	default_attr_source_tree_object_name = git_attr_tree;
> 		ignore_bad_attr_tree = 1;
> 	}
>
> It would leave more flexibility to the code around here.  You could
> for example add code that assigns a different value, a tree object
> that is required to exist, to default_attr_source_tree_object_name
> after this point, for example, without having to wonder what the
> "current" value of ignore_bad_attr_tree is.

good point--I think this logic is easier to follow.

>
>> +static int git_default_attr_config(const char *var, const char *value)
>> +{
>> +	if (!strcmp(var, "attr.tree"))
>> +		return git_config_string(&git_attr_tree, var, value);
>> +
>> +	/* Add other attribute related config variables here and to
>> +	   Documentation/config/attr.txt. */
>
> 	/*
> 	 * Our multi-line comments should look
>          * more like this; opening slash-asterisk
> 	 * and closing asterisk-slash sit on a line
> 	 * on its own.
> 	 */
>
>> @@ -342,6 +346,46 @@ test_expect_success 'bare repository: check that .gitattribute is ignored' '
>>  	)
>>  '
>>
>> +bad_attr_source_err="fatal: bad --attr-source or GIT_ATTR_SOURCE"
>
> Not a fault of this two-patch series, but we probably should refine
> this error reporting so that the reader can tell which one is being
> complained about, and optionally what the offending value was.
>
>
>> +test_expect_success 'attr.tree when HEAD is unborn' '
>> +	test_when_finished rm -rf empty &&
>> +	git init empty &&
>> +	(
>> +		cd empty &&
>> +		echo $bad_attr_source_err >expect_err &&
>
> Let's not rely on words in the error message split with exactly one
> whitespace each and instead quote the variable properly.  I.e.,
>
> 		echo "$bad_attr_source_err" >expect_err &&
>
> But this is not even used, as we do not expect it to fail.  Perhaps
> remove it altogether?
>
>> +		echo "f/path: test: unspecified" >expect &&
>> +		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
>> +		test_must_be_empty err &&
>> +		test_cmp expect actual
>> +	)
>> +'
>> +
>> +test_expect_success 'attr.tree points to non-existing ref' '
>> +	test_when_finished rm -rf empty &&
>> +	git init empty &&
>> +	(
>> +		cd empty &&
>> +		echo $bad_attr_source_err >expect_err &&
>
> Ditto.
>
>> +		echo "f/path: test: unspecified" >expect &&
>> +		git -c attr.tree=refs/does/not/exist check-attr test -- f/path >actual 2>err &&
>> +		test_must_be_empty err &&
>> +		test_cmp expect actual
>> +	)
>> +'
>> +
>> +test_expect_success 'bad attr source defaults to reading .gitattributes file' '
>> +	test_when_finished rm -rf empty &&
>> +	git init empty &&
>> +	(
>> +		cd empty &&
>> +		echo "f/path test=val" >.gitattributes &&
>> +		echo "f/path: test: val" >expect &&
>> +		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
>> +		test_must_be_empty err &&
>> +		test_cmp expect actual
>> +	)
>> +'
>
> In other words, with the additional tests, we do not check error
> cases (which may be perfectly OK, if they are covered by existing
> tests).  A bit curious.

Just checked and it does seem like we don't have a test case that covers the
error case. Will add in the next version.

>
>> @@ -356,6 +400,24 @@ test_expect_success 'bare repo defaults to reading .gitattributes from HEAD' '
>>  	test_cmp expect actual
>>  '
>>
>> +test_expect_success '--attr-source and GIT_ATTR_SOURCE take precedence over attr.tree' '
>
> Do we want to ensure which one takes precedence between the command
> line option and the environment?  It's not like the one that is
> given the last takes effect.

yeah, might as well.

>
>> +	test_when_finished rm -rf empty &&
>> +	git init empty &&
>> +	(
>> +		cd empty &&
>> +		git checkout -b attr-source &&
>> +		test_commit "val1" .gitattributes "f/path test=val1" &&
>> +		git checkout -b attr-tree &&
>> +		test_commit "val2" .gitattributes "f/path test=val2" &&
>> +		git checkout attr-source &&
>> +		echo "f/path: test: val1" >expect &&
>> +		git -c attr.tree=attr-tree --attr-source=attr-source check-attr test -- f/path >actual &&
>> +		test_cmp expect actual &&
>> +		GIT_ATTR_SOURCE=attr-source git -c attr.tree=attr-tree check-attr test -- f/path >actual &&
>> +		test_cmp expect actual
>> +	)
>> +'
>> +
>>  test_expect_success 'bare repository: with --source' '
>>  	(
>>  		cd bare.git &&
>
> Other than that, looking great.  Thanks.

thanks!
John
diff mbox series

Patch

diff --git a/Documentation/config.txt b/Documentation/config.txt
index 229b63a454c..b1891c2b5af 100644
--- a/Documentation/config.txt
+++ b/Documentation/config.txt
@@ -371,6 +371,8 @@  other popular tools, and describe them in your documentation.
 
 include::config/advice.txt[]
 
+include::config/attr.txt[]
+
 include::config/core.txt[]
 
 include::config/add.txt[]
diff --git a/Documentation/config/attr.txt b/Documentation/config/attr.txt
new file mode 100644
index 00000000000..be882523f8b
--- /dev/null
+++ b/Documentation/config/attr.txt
@@ -0,0 +1,5 @@ 
+attr.tree:
+	A <tree-ish> to read gitattributes from instead of the worktree. See
+	linkgit:gitattributes[5]. If `attr.tree` does not resolve to a valid tree,
+	treat it as an empty tree. --attr-source and GIT_ATTR_SOURCE take
+	precedence over attr.tree.
diff --git a/attr.c b/attr.c
index bf2ea1626a6..0ae6852d12b 100644
--- a/attr.c
+++ b/attr.c
@@ -24,6 +24,8 @@ 
 #include "tree-walk.h"
 #include "object-name.h"
 
+const char *git_attr_tree;
+
 const char git_attr__true[] = "(builtin)true";
 const char git_attr__false[] = "\0(builtin)false";
 static const char git_attr__unknown[] = "(builtin)unknown";
@@ -1206,6 +1208,11 @@  static void compute_default_attr_source(struct object_id *attr_source)
 	if (!default_attr_source_tree_object_name)
 		default_attr_source_tree_object_name = getenv(GIT_ATTR_SOURCE_ENVIRONMENT);
 
+	if (!default_attr_source_tree_object_name) {
+		default_attr_source_tree_object_name = git_attr_tree;
+		ignore_bad_attr_tree = 1;
+	}
+
 	if (!default_attr_source_tree_object_name &&
 	    startup_info->have_repository &&
 	    is_bare_repository()) {
diff --git a/attr.h b/attr.h
index 2b745df4054..127998ae013 100644
--- a/attr.h
+++ b/attr.h
@@ -236,4 +236,6 @@  const char *git_attr_global_file(void);
 /* Return whether the system gitattributes file is enabled and should be used. */
 int git_attr_system_is_enabled(void);
 
+extern const char *git_attr_tree;
+
 #endif /* ATTR_H */
diff --git a/config.c b/config.c
index 3846a37be97..21a1590b505 100644
--- a/config.c
+++ b/config.c
@@ -18,6 +18,7 @@ 
 #include "repository.h"
 #include "lockfile.h"
 #include "mailmap.h"
+#include "attr.h"
 #include "exec-cmd.h"
 #include "strbuf.h"
 #include "quote.h"
@@ -1904,6 +1905,16 @@  static int git_default_mailmap_config(const char *var, const char *value)
 	return 0;
 }
 
+static int git_default_attr_config(const char *var, const char *value)
+{
+	if (!strcmp(var, "attr.tree"))
+		return git_config_string(&git_attr_tree, var, value);
+
+	/* Add other attribute related config variables here and to
+	   Documentation/config/attr.txt. */
+	return 0;
+}
+
 int git_default_config(const char *var, const char *value,
 		       const struct config_context *ctx, void *cb)
 {
@@ -1927,6 +1938,9 @@  int git_default_config(const char *var, const char *value,
 	if (starts_with(var, "mailmap."))
 		return git_default_mailmap_config(var, value);
 
+	if (starts_with(var, "attr."))
+		return git_default_attr_config(var, value);
+
 	if (starts_with(var, "advice.") || starts_with(var, "color.advice"))
 		return git_default_advice_config(var, value);
 
diff --git a/t/t0003-attributes.sh b/t/t0003-attributes.sh
index e6b1a117228..b0949125f26 100755
--- a/t/t0003-attributes.sh
+++ b/t/t0003-attributes.sh
@@ -40,6 +40,10 @@  attr_check_source () {
 	test_cmp expect actual &&
 	test_must_be_empty err
 
+	git $git_opts -c "attr.tree=$source" check-attr test -- "$path" >actual 2>err &&
+	test_cmp expect actual &&
+	test_must_be_empty err
+
 	GIT_ATTR_SOURCE="$source" git $git_opts check-attr test -- "$path" >actual 2>err &&
 	test_cmp expect actual &&
 	test_must_be_empty err
@@ -342,6 +346,46 @@  test_expect_success 'bare repository: check that .gitattribute is ignored' '
 	)
 '
 
+bad_attr_source_err="fatal: bad --attr-source or GIT_ATTR_SOURCE"
+
+test_expect_success 'attr.tree when HEAD is unborn' '
+	test_when_finished rm -rf empty &&
+	git init empty &&
+	(
+		cd empty &&
+		echo $bad_attr_source_err >expect_err &&
+		echo "f/path: test: unspecified" >expect &&
+		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
+		test_must_be_empty err &&
+		test_cmp expect actual
+	)
+'
+
+test_expect_success 'attr.tree points to non-existing ref' '
+	test_when_finished rm -rf empty &&
+	git init empty &&
+	(
+		cd empty &&
+		echo $bad_attr_source_err >expect_err &&
+		echo "f/path: test: unspecified" >expect &&
+		git -c attr.tree=refs/does/not/exist check-attr test -- f/path >actual 2>err &&
+		test_must_be_empty err &&
+		test_cmp expect actual
+	)
+'
+
+test_expect_success 'bad attr source defaults to reading .gitattributes file' '
+	test_when_finished rm -rf empty &&
+	git init empty &&
+	(
+		cd empty &&
+		echo "f/path test=val" >.gitattributes &&
+		echo "f/path: test: val" >expect &&
+		git -c attr.tree=HEAD check-attr test -- f/path >actual 2>err &&
+		test_must_be_empty err &&
+		test_cmp expect actual
+	)
+'
 
 test_expect_success 'bare repo defaults to reading .gitattributes from HEAD' '
 	test_when_finished rm -rf test bare_with_gitattribute &&
@@ -356,6 +400,24 @@  test_expect_success 'bare repo defaults to reading .gitattributes from HEAD' '
 	test_cmp expect actual
 '
 
+test_expect_success '--attr-source and GIT_ATTR_SOURCE take precedence over attr.tree' '
+	test_when_finished rm -rf empty &&
+	git init empty &&
+	(
+		cd empty &&
+		git checkout -b attr-source &&
+		test_commit "val1" .gitattributes "f/path test=val1" &&
+		git checkout -b attr-tree &&
+		test_commit "val2" .gitattributes "f/path test=val2" &&
+		git checkout attr-source &&
+		echo "f/path: test: val1" >expect &&
+		git -c attr.tree=attr-tree --attr-source=attr-source check-attr test -- f/path >actual &&
+		test_cmp expect actual &&
+		GIT_ATTR_SOURCE=attr-source git -c attr.tree=attr-tree check-attr test -- f/path >actual &&
+		test_cmp expect actual
+	)
+'
+
 test_expect_success 'bare repository: with --source' '
 	(
 		cd bare.git &&