[v2] config: introduce an Operating System-specific `includeIf` condition

Commit Message

Johannes Schindelin Nov. 21, 2022, 7:19 p.m. UTC

From: Johannes Schindelin <[e-mail]>

It is relatively common for users to maintain identical `~/.gitconfig`
files across all of their setups, using the `includeIf` construct
liberally to adjust the settings to the respective setup as needed.

In case of Operating System-specific adjustments, Git currently offers
no support to the users and they typically use a work-around like this:

	[includeIf "gitdir:/home/"]
		path = ~/.gitconfig-linux
	[includeIf "gitdir:/Users/"]
		path = ~/.gitconfig-mac
	[includeIf "gitdir:C:"]
		path = ~/.gitconfig-windows

However, this is fragile, as it would not even allow to discern between
Operating Systems that happen to host their home directories in
`/home/`, such as Linux and the BSDs.

Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
system name, i.e. the output of `uname -s`.

This addresses https://github.com/git-for-windows/git/issues/4125

Signed-off-by: Johannes Schindelin <[e-mail]>
---
    config: introduce an Operating System-specific includeIf condition
    
    I was about to write up guidelines how to write this patch, but it
    turned out that it was much faster to write the patch instead.
    
    Changes since v1:
    
     * The documentation now avoids mentioning uname -s and clarifies what
       it means by offering examples.
     * Replaced a double space in the test case with a single one.

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1429%2Fdscho%2Finclude-if-os-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1429/dscho/include-if-os-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/1429

Range-diff vs v1:

 1:  a7eb4a9d438 ! 1:  45231533883 config: introduce an Operating System-specific `includeIf` condition
     @@ Documentation/config.txt: As for the naming of this keyword, it is for forwards
       
      +`os`::
      +	The data that follows this keyword is taken as the name of an
     -+	Operating System; If it matches the output of `uname -s`, the
     -+	include condition is met.
     ++	Operating System, e.g. `Linux` or `Windows`; If it matches the
     ++	current Operating System, the include condition is met.
      +
       A few more notes on matching via `gitdir` and `gitdir/i`:
       
     @@ t/t1309-early-config.sh: test_expect_success 'onbranch config outside of git rep
      +		uname_s="$(uname -s)"
      +	fi &&
      +	test_config "includeIf.os:not-$uname_s.path" xyz &&
     -+	test 0 = "$(git  config x.y)" &&
     ++	test 0 = "$(git config x.y)" &&
      +	test_config "includeIf.os:$uname_s.path" xyz &&
      +	test z = "$(git config x.y)"
      +'


 Documentation/config.txt |  5 +++++
 config.c                 | 11 +++++++++++
 t/t1309-early-config.sh  | 16 ++++++++++++++++
 3 files changed, 32 insertions(+)


base-commit: e4a4b31577c7419497ac30cebe30d755b97752c5

Comments

Jeff King Nov. 21, 2022, 11:32 p.m. UTC | #1

On Mon, Nov 21, 2022 at 07:19:48PM +0000, Johannes Schindelin via GitGitGadget wrote:

> It is relatively common for users to maintain identical `~/.gitconfig`
> files across all of their setups, using the `includeIf` construct
> liberally to adjust the settings to the respective setup as needed.

This seems like a reasonable thing to have in general, but I wonder if
you have an example of how people use this. Mostly I am wondering:

  - is it sometimes a misuse, where users _think_ that the OS is
    correlated with some feature of Git.  And they would be better off
    with some flag like "does the current platform support fsmonitor".

  - for cases where it really  is "uname -s" the best differentiator? Or
    would they commonly want to lump FreeBSD and Linux into the same
    category, or to tell the difference between Debian versus Fedora?

-Peff

Ævar Arnfjörð Bjarmason Nov. 22, 2022, 2:01 p.m. UTC | #2

On Mon, Nov 21 2022, Johannes Schindelin via GitGitGadget wrote:

> From: Johannes Schindelin <[e-mail]>
>
> +`os`::
> +	The data that follows this keyword is taken as the name of an
> +	Operating System, e.g. `Linux` or `Windows`; If it matches the
> +	current Operating System, the include condition is met.
> +
>  A few more notes on matching via `gitdir` and `gitdir/i`:

The reste of the "includeif" use glob matching and "/i" for icase. IOW
this is how this new feature would fit in:
	
	|--------+--------+----------+----------+------------------+----|
	|        | gitdir | gitdir/i | onbranch | hasconfig:remote | os |
	|--------+--------+----------+----------+------------------+----|
	| icase? | N      | Y        | N        | N                | Y  |
	| glob?  | Y      | Y        | Y        | Y                | N  |
	| path?  | Y      | Y        | Y        | Y                | N  |
	|--------+--------+----------+----------+------------------+----|

I think at least flipping that "glob" to "Y" so you could match e.g.
"*BSD" would be useful, and easier to explain in context, rather than
why the rest use wildmatch() and this doesn't.

For matching the uname the case doesn't really matter, but for
consistency of the interface I think making it case-sensitive or adding
an "os/i" would make sense. I.e. let's consistently use "/i" if & when
something's case-insensitive.

> +test_expect_success '[includeIf "os:..."]' '
> +	test_config x.y 0 &&
> +	echo "[x] y = z" >.git/xyz &&
> +
> +	if test_have_prereq MINGW
> +	then
> +		uname_s=Windows
> +	else
> +		uname_s="$(uname -s)"
> +	fi &&
> +	test_config "includeIf.os:not-$uname_s.path" xyz &&
> +	test 0 = "$(git config x.y)" &&
> +	test_config "includeIf.os:$uname_s.path" xyz &&
> +	test z = "$(git config x.y)"
> +'

As I pointed out in the v1, this still:

 * Hides segfaults in "git config", let's check the exit code.
 * Doesn't test the "icase" semantics you're introducing. Let's do that
   if it's intentional.

Phillip Wood Nov. 22, 2022, 2:31 p.m. UTC | #3

On 22/11/2022 14:01, Ævar Arnfjörð Bjarmason wrote:
> 
> On Mon, Nov 21 2022, Johannes Schindelin via GitGitGadget wrote:
> 
>> From: Johannes Schindelin <[e-mail]>
>>
>> +`os`::
>> +	The data that follows this keyword is taken as the name of an
>> +	Operating System, e.g. `Linux` or `Windows`; If it matches the
>> +	current Operating System, the include condition is met.
>> +
>>   A few more notes on matching via `gitdir` and `gitdir/i`:
> 
> The reste of the "includeif" use glob matching and "/i" for icase. IOW
> this is how this new feature would fit in:
> 	
> 	|--------+--------+----------+----------+------------------+----|
> 	|        | gitdir | gitdir/i | onbranch | hasconfig:remote | os |
> 	|--------+--------+----------+----------+------------------+----|
> 	| icase? | N      | Y        | N        | N                | Y  |
> 	| glob?  | Y      | Y        | Y        | Y                | N  |
> 	| path?  | Y      | Y        | Y        | Y                | N  |
> 	|--------+--------+----------+----------+------------------+----|
> 
> I think at least flipping that "glob" to "Y" so you could match e.g.
> "*BSD" would be useful, and easier to explain in context, rather than
> why the rest use wildmatch() and this doesn't.

Globbing could be useful for the BSDs.

One other thing I thought of is will users know "Darwin" means MacOS?

> For matching the uname the case doesn't really matter, but for
> consistency of the interface I think making it case-sensitive or adding
> an "os/i" would make sense. I.e. let's consistently use "/i" if & when
> something's case-insensitive.

All the other items listed in your table such as branch names are case 
sensitive. The os name is not so it is of no benefit at all to the user 
to match it case sensitively. Let's consistently test case sensitive 
keys cases sensitively and case insensitive keys case insensitively.

Best Wishes

Phillip

>> +test_expect_success '[includeIf "os:..."]' '
>> +	test_config x.y 0 &&
>> +	echo "[x] y = z" >.git/xyz &&
>> +
>> +	if test_have_prereq MINGW
>> +	then
>> +		uname_s=Windows
>> +	else
>> +		uname_s="$(uname -s)"
>> +	fi &&
>> +	test_config "includeIf.os:not-$uname_s.path" xyz &&
>> +	test 0 = "$(git config x.y)" &&
>> +	test_config "includeIf.os:$uname_s.path" xyz &&
>> +	test z = "$(git config x.y)"
>> +'
> 
> As I pointed out in the v1, this still:
> 
>   * Hides segfaults in "git config", let's check the exit code.
>   * Doesn't test the "icase" semantics you're introducing. Let's do that
>     if it's intentional.

Philippe Blain Nov. 22, 2022, 6:40 p.m. UTC | #4

Hi Dscho,

Le 2022-11-21 à 14:19, Johannes Schindelin via GitGitGadget a écrit :
> From: Johannes Schindelin <[e-mail]>
> 
> It is relatively common for users to maintain identical `~/.gitconfig`
> files across all of their setups, using the `includeIf` construct
> liberally to adjust the settings to the respective setup as needed.
> 
> In case of Operating System-specific adjustments, 

Here, as well as in the commit title, the rest of the message, and the changes to 
the doc in the patch, I would downcase "operating system". 
It's a common word that I don't think should be capitalized (in contrast to
proper OS names like Windows and Linux).

Cheers,
Philippe.

Junio C Hamano Nov. 23, 2022, 12:16 a.m. UTC | #5

Phillip Wood <[e-mail]> writes:

> All the other items listed in your table such as branch names are case
> sensitive. The os name is not so it is of no benefit at all to the

You keep saying that you consider the OS name is case insensitive,
but I doubt that is the case, not in the sense that MacOS and macOS
are two different operating systems, but in the sense that OS
publishers have a single preferred way to spell their ware (which is
shown in "uname -s" output), and we should respect that.

Philip Oakley Nov. 23, 2022, 10:40 a.m. UTC | #6

Hi Dscho,

On 21/11/2022 19:19, Johannes Schindelin via GitGitGadget wrote:
> From: Johannes Schindelin <[e-mail]>
>
> It is relatively common for users to maintain identical `~/.gitconfig`
> files across all of their setups, using the `includeIf` construct
> liberally to adjust the settings to the respective setup as needed.
>
> In case of Operating System-specific adjustments, Git currently offers
> no support to the users and they typically use a work-around like this:
>
> 	[includeIf "gitdir:/home/"]
> 		path = ~/.gitconfig-linux
> 	[includeIf "gitdir:/Users/"]
> 		path = ~/.gitconfig-mac
> 	[includeIf "gitdir:C:"]
> 		path = ~/.gitconfig-windows
>
> However, this is fragile, as it would not even allow to discern between
> Operating Systems that happen to host their home directories in
> `/home/`, such as Linux and the BSDs.
>
> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
> system name, i.e. the output of `uname -s`.

This `uname -s` doesn't work well on Git-for-Windows (GfW) / Windows
because GfW has its own internal compatibility code to spoof the result.
Internally GfW sets it to "Windows" (see
https://github.com/git/git/blob/master/compat/mingw.c#L3084-L3095).

If I ask for `uname -s` on the GfW bash, for me it returns
`MINGW64_NT-10.0-19045`, both on the normal GfW install, and the GfW-SDK.

If I (as dumb user) try the CMD window prompt it's
    'uname' is not recognized as an internal or external command,
    operable program or batch file.

The Windows PowerShell also doesn't recognise the uname command.

My WSL reports `Linux`, though I haven't played with that for a while
(do all *nix variants report that?).

Thus we'll need some way of assisting the users in determining the
internal system name that their local Git will find. It maybe that the
man page just needs to be explicit about using "Windows" for the
Git-for-Windows implementation.

Or less helpful, maybe a `git-uname` command to disambiguate any other
systems with a compat::uname() variant.

Or just drop the mentions of "<uname-s>" in this commit message and
rename it 'sysname' to match the field of the struct utsname?

>
> This addresses https://github.com/git-for-windows/git/issues/4125
>
> Signed-off-by: Johannes Schindelin <[e-mail]>
> ---
>     config: introduce an Operating System-specific includeIf condition
>     
>     I was about to write up guidelines how to write this patch, but it
>     turned out that it was much faster to write the patch instead.
>     
>     Changes since v1:
>     
>      * The documentation now avoids mentioning uname -s and clarifies what
>        it means by offering examples.
>      * Replaced a double space in the test case with a single one.
>
> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1429%2Fdscho%2Finclude-if-os-v2
> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1429/dscho/include-if-os-v2
> Pull-Request: https://github.com/gitgitgadget/git/pull/1429
>
> Range-diff vs v1:
>
>  1:  a7eb4a9d438 ! 1:  45231533883 config: introduce an Operating System-specific `includeIf` condition
>      @@ Documentation/config.txt: As for the naming of this keyword, it is for forwards
>        
>       +`os`::
>       +	The data that follows this keyword is taken as the name of an
>      -+	Operating System; If it matches the output of `uname -s`, the
>      -+	include condition is met.
>      ++	Operating System, e.g. `Linux` or `Windows`; If it matches the
>      ++	current Operating System, the include condition is met.
>       +
>        A few more notes on matching via `gitdir` and `gitdir/i`:
>        
>      @@ t/t1309-early-config.sh: test_expect_success 'onbranch config outside of git rep
>       +		uname_s="$(uname -s)"
>       +	fi &&
>       +	test_config "includeIf.os:not-$uname_s.path" xyz &&
>      -+	test 0 = "$(git  config x.y)" &&
>      ++	test 0 = "$(git config x.y)" &&
>       +	test_config "includeIf.os:$uname_s.path" xyz &&
>       +	test z = "$(git config x.y)"
>       +'
>
>
>  Documentation/config.txt |  5 +++++
>  config.c                 | 11 +++++++++++
>  t/t1309-early-config.sh  | 16 ++++++++++++++++
>  3 files changed, 32 insertions(+)
>
> diff --git a/Documentation/config.txt b/Documentation/config.txt
> index e376d547ce0..b90bcd8ecfe 100644
> --- a/Documentation/config.txt
> +++ b/Documentation/config.txt
> @@ -186,6 +186,11 @@ As for the naming of this keyword, it is for forwards compatibiliy with
>  a naming scheme that supports more variable-based include conditions,
>  but currently Git only supports the exact keyword described above.
>  
> +`os`::
> +	The data that follows this keyword is taken as the name of an
> +	Operating System, e.g. `Linux` or `Windows`; If it matches the
> +	current Operating System

Is 'matches' the appropriate way to indicate that we compare just the
characters from the data string and ignore any trailing chars in the
uname.sysname field?

> , the include condition is met.
> +
>  A few more notes on matching via `gitdir` and `gitdir/i`:
>  
>   * Symlinks in `$GIT_DIR` are not resolved before matching.
> diff --git a/config.c b/config.c
> index 9b0e9c93285..9ab311ae99b 100644
> --- a/config.c
> +++ b/config.c
> @@ -394,6 +394,15 @@ static int include_by_remote_url(struct config_include_data *inc,
>  					     inc->remote_urls);
>  }
>  
> +static int include_by_os(const char *cond, size_t cond_len)
> +{
> +	struct utsname uname_info;
> +
> +	return !uname(&uname_info) &&
> +		!strncasecmp(uname_info.sysname, cond, cond_len) &&
> +		!uname_info.sysname[cond_len];
> +}
> +
>  static int include_condition_is_true(struct config_include_data *inc,
>  				     const char *cond, size_t cond_len)
>  {
> @@ -408,6 +417,8 @@ static int include_condition_is_true(struct config_include_data *inc,
>  	else if (skip_prefix_mem(cond, cond_len, "hasconfig:remote.*.url:", &cond,
>  				   &cond_len))
>  		return include_by_remote_url(inc, cond, cond_len);
> +	else if (skip_prefix_mem(cond, cond_len, "os:", &cond, &cond_len))
> +		return include_by_os(cond, cond_len);

(as above) We compare only on the basis of the few/many characters in
the config file so, IIUC, we could use `Win`, or `Lin` as the os:
string. Should this be noted in the man text? I'm thinking of users who
may be confused by having to change Win10 to Windows, but are happy to
shorten to `Win`.
>  
>  	/* unknown conditionals are always false */
>  	return 0;
> diff --git a/t/t1309-early-config.sh b/t/t1309-early-config.sh
> index 537435b90ae..b36afe1a528 100755
> --- a/t/t1309-early-config.sh
> +++ b/t/t1309-early-config.sh
> @@ -100,4 +100,20 @@ test_expect_success 'onbranch config outside of git repo' '
>  	nongit git help
>  '
>  
> +test_expect_success '[includeIf "os:..."]' '
> +	test_config x.y 0 &&
> +	echo "[x] y = z" >.git/xyz &&
> +
> +	if test_have_prereq MINGW
> +	then
> +		uname_s=Windows

This (correctly) copies/follows the compat code
(https://github.com/git/git/blob/master/compat/mingw.c#L3088), but isn't
what a GfW user sees if `uname-s` is run in their bash. Maybe change
uname_s to sysname, as noted above.
> +	else
> +		uname_s="$(uname -s)"
> +	fi &&
> +	test_config "includeIf.os:not-$uname_s.path" xyz &&
> +	test 0 = "$(git config x.y)" &&
> +	test_config "includeIf.os:$uname_s.path" xyz &&
> +	test z = "$(git config x.y)"
> +'
> +
>  test_done
>
> base-commit: e4a4b31577c7419497ac30cebe30d755b97752c5
--
Philip

Đoàn Trần Công Danh Nov. 23, 2022, 11:54 a.m. UTC | #7

On 2022-11-21 18:32:11-0500, Jeff King <[e-mail]> wrote:
> On Mon, Nov 21, 2022 at 07:19:48PM +0000, Johannes Schindelin via GitGitGadget wrote:
> 
> > It is relatively common for users to maintain identical `~/.gitconfig`
> > files across all of their setups, using the `includeIf` construct
> > liberally to adjust the settings to the respective setup as needed.
> 
> This seems like a reasonable thing to have in general, but I wonder if
> you have an example of how people use this. Mostly I am wondering:
> 
>   - is it sometimes a misuse, where users _think_ that the OS is
>     correlated with some feature of Git.  And they would be better off
>     with some flag like "does the current platform support fsmonitor".

A possible use-case is setting credential.helper based on OS, let's say
libsecret on Linux, and osxkeychain on macOS. Of course, users can
have their own helper on specific OS.

> 
>   - for cases where it really  is "uname -s" the best differentiator? Or
>     would they commonly want to lump FreeBSD and Linux into the same
>     category, or to tell the difference between Debian versus Fedora?
> 
> -Peff

Phillip Wood Nov. 23, 2022, 3:07 p.m. UTC | #8

Hi Junio

Welcome back, I hope you enjoyed your time away from the list.

On 23/11/2022 00:16, Junio C Hamano wrote:
> Phillip Wood <[e-mail]> writes:
> 
>> All the other items listed in your table such as branch names are case
>> sensitive. The os name is not so it is of no benefit at all to the
> 
> You keep saying that you consider the OS name is case insensitive,
> but I doubt that is the case, not in the sense that MacOS and macOS
> are two different operating systems, but in the sense that OS
> publishers have a single preferred way to spell their ware (which is
> shown in "uname -s" output), and we should respect that.

I can see that we would want to respect the preferred spelling in our 
documentation but it seems a bit mean to penalize users who write

	[IncludeIf "os:windows"]

rather than

	[IncludeIf "os:Windows"]

Best Wishes

Phillip

Junio C Hamano Nov. 23, 2022, 11:51 p.m. UTC | #9

Phillip Wood <[e-mail]> writes:

> I can see that we would want to respect the preferred spelling in our
> documentation but it seems a bit mean to penalize users who write
>
> 	[IncludeIf "os:windows"]
>
> rather than
>
> 	[IncludeIf "os:Windows"]

I do not think "uname-s/i:windows" is out of question.  I am saying
that the default should be case sensitive for consistency with
others.

Also, as I said already, I do not think we should squat on a good
name "os" with this "uname -s" feature that may not be a good
differenciator for two OSes for some cases (e.g. telling Debian and
Fedora apart was a good example raised upthread already).

Jeff King Nov. 24, 2022, 12:56 a.m. UTC | #10

On Wed, Nov 23, 2022 at 06:54:45PM +0700, Đoàn Trần Công Danh wrote:

> > This seems like a reasonable thing to have in general, but I wonder if
> > you have an example of how people use this. Mostly I am wondering:
> > 
> >   - is it sometimes a misuse, where users _think_ that the OS is
> >     correlated with some feature of Git.  And they would be better off
> >     with some flag like "does the current platform support fsmonitor".
> 
> A possible use-case is setting credential.helper based on OS, let's say
> libsecret on Linux, and osxkeychain on macOS. Of course, users can
> have their own helper on specific OS.

Thanks, that's a very nice concrete example. I agree that's a pretty
reasonable use of this feature, given the lack of other selection
mechanisms. I do wonder if folks might run into annoyances when they
want to use libsecret on Linux _and_ FreeBSD or similar. But this
feature is definitely better than the status quo, and is not very much
code to implement or support.

-Peff

Junio C Hamano Nov. 25, 2022, 7:31 a.m. UTC | #11

Philip Oakley <[e-mail]> writes:

>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
>> system name, i.e. the output of `uname -s`.
>
> This `uname -s` doesn't work well on Git-for-Windows (GfW) / Windows
> because GfW has its own internal compatibility code to spoof the result.
> ...
> Or just drop the mentions of "<uname-s>" in this commit message and
> rename it 'sysname' to match the field of the struct utsname?

FWIW I do not mind "sysname".  It is much better to say

	[includeIf "sysname:Linux"] path = ...

than "os:Linux", as "sysname" informs us the granularity used to
identify the system better than "os".

Samuel Ferencik April 17, 2023, 7:04 a.m. UTC | #12

>>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
>>> system name, i.e. the output of `uname -s`.

The discussion about https://github.com/gitgitgadget/git/pull/1429 seems to
have stalled on several points. I'll try to summarise; let's see if we can move
forward.

(I am the reporter of https://github.com/git-for-windows/git/issues/4125, which
led to this PR. I am vested in making progress here.)

1. name of the setting (`os` vs `uname-s` vs `sysname`)
   * dscho@ suggested `os`; Phillip and Philip suggested `uname-s` and
     `sysname`, respectively
   * I vote for `os`; I'm afraid perfect is the enemy of good here as
     * `man uname` says `-s` gives "the name of the operating system
       implementation"; no other `uname` switch comes closer to whatever
       concept "OS" represents
     * this is also correct on Windows (the "Windows" string) - see below
     * I find it extremely unlikely a future unforeseen git feature would have
       a better use for `includeIf os` (in parallel with `includeIf sysname`),
       i.e. I don't worry that we're squatting on a good name for a poor
       use-case
2. casing (use of `/i`)
   * dscho@ implemented case-insensitive comparison but without test coverage,
     documentation, and it's inconsistent with the other `includeIf` options
     that support the `/i` switch.
   * I propose that we compare case-sensitively because
     * no user can reasonably complain about this if the documentation is
       clear; the OS names are definitive and stable and it's not a big deal
       getting the case right for "Linux"
     * without the case insensitivity being documented, the users who [discover
       the insensitivity and] rely on it are risking breakage; plus the git
       maintainers are exposing themselves to the effects of Hyrum's law
       (https://xkcd.com/1172) - it's a disservice for both sides
     * this still allows us to add support for `/i` later (should a use-case
       emerge)
     * it is consistent with the other settings
     * it requires less code (incl. tests) and shorter documentation
3. handling Windows (MinGW, WSL)
   * As implemented currently, `includeIf "os:Windows"` would work in
     git-for-Windows. I think that's desirable, and no-one suggested otherwise.
   * In contrast, Philip points out `includeIf "os:Linux"` would be the way to
     match on WSL. Is that an issue? Do we want WSL to match "os:Windows" or
     "os:WSL"? As a Windows user, when I switch to WSL I do expect a "proper"
     Linux experience (unlike when I run "Git bash" on Windows, which is more
     like a port of utilities, but still Windows). I think this treating WLS as
     Linux is OK-ish, and we may get away with not discerning WSL. Thoughts?
4. documentation (w.r.t. the details in 1. - 3.)
   * We should document all of 1. - 3. I'm happy to give it a go if we can
     reach consensus.
   * Specifically, the documentation should mention that the OS string equals
     "Windows" in Git-for-Windows, and `$(uname -s)` otherwise; it should list
     examples, incl. "Linux" and "Darwin"; it should mention the case
     sensitivity.
5. tests (potential segfaults)
   * Johannes points out the tests hide segfaults. I haven't looked at this
     closely but hopefully Johannes's suggestion ("use test_cmp or
     test_cmp_config") is a clear enough pointer. I can try to fix this.
6. what's the use-case?
   * As the reporter of https://github.com/git-for-windows/git/issues/4125,
     here are my use-cases, i.e. settings that I currently set conditionally
     per OS (using `includeIf gitdir`):
     * different `difftool.path`, `mergetool.path` per OS (e.g. paths
       containing `C:\Program Files\...\...exe` vs Unix file paths)
     * different `merge.tool` per OS (I have a BeyondCompare license for Linux
       only)
     * different `core.autocrlf`: `true` on Windows, `input` otherwise
     * `core.whitespace` set to `cr-at-eol` on Windows
     * `core.editor` set to `gvim` on Windows
   * Note all of the use-cases above would cope with WSL reporting "Linux",
     except of `merge.tool`.

I hope this revives the discussion; I know it's been a while but I would
appreciate your input. If possible, please refer to the numbered points (1 - 6)
for clarity.

I'm happy to iterate on dscho@'s PR if we can reach consensus.


On Fri, 14 Apr 2023 at 01:44, Junio C Hamano <[e-mail]> wrote:
>
> Philip Oakley <[e-mail]> writes:
>
> >> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
> >> system name, i.e. the output of `uname -s`.
> >
> > This `uname -s` doesn't work well on Git-for-Windows (GfW) / Windows
> > because GfW has its own internal compatibility code to spoof the result.
> > ...
> > Or just drop the mentions of "<uname-s>" in this commit message and
> > rename it 'sysname' to match the field of the struct utsname?
>
> FWIW I do not mind "sysname".  It is much better to say
>
>         [includeIf "sysname:Linux"] path = ...
>
> than "os:Linux", as "sysname" informs us the granularity used to
> identify the system better than "os".
>
>
>

Junio C Hamano April 17, 2023, 6:46 p.m. UTC | #13

Samuel Ferencik <[e-mail]> writes:

>>>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
>>>> system name, i.e. the output of `uname -s`.
>
> The discussion about https://github.com/gitgitgadget/git/pull/1429 seems to
> have stalled on several points. I'll try to summarise; let's see if we can move
> forward.
>
> (I am the reporter of https://github.com/git-for-windows/git/issues/4125, which
> led to this PR. I am vested in making progress here.)
>
> 1. name of the setting (`os` vs `uname-s` vs `sysname`)

I do not think it is a good idea to squat on too generic a name like
'os', especially when there are multiple layers people will care
about.  But I think the original thread discussed this to death, and
I do not see a point bringing it up again as the first bullet point.

> 2. casing (use of `/i`)

My preference is to do this case sensitively (in other words, start
stupid) and if somebody wants to use "/i", add it later after the
dust settles.

> 3. handling Windows (MinGW, WSL)

This comes back to the reason why "os" is a horrible choice.  Is WSL
a Windows?  Is WSL a Linux?  The same question can be asked for Cygwin.

The answer depends on which layer you care about.  The underlying
kernel and system may be Windows, and some characteristics of the
underlying system may seep through the abstraction, but these
systems aim to give user experience of something like GNU/Linux.

And this is not limited to Windows.  There may be similar issue for
systems like PacBSD.  Is it a Linux?  Is it a BSD?

> 6. what's the use-case?

I think that this is the most important question to ask, and from
here, we'd see how #3 above should be resolved (I suspect that you
may want to have at least two layers to allow WSL to be grouped
together with MinGW and Cygwin at one level, and at the same time
allow it to be grouped together with Ubuntu at a different level).
And after we figure that out, we'll have a clear and intuitive
answer to #1.

Felipe Contreras April 18, 2023, 2:04 a.m. UTC | #14

Junio C Hamano wrote:
> Samuel Ferencik <[e-mail]> writes:
> 
> >>>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
> >>>> system name, i.e. the output of `uname -s`.
> >
> > The discussion about https://github.com/gitgitgadget/git/pull/1429 seems to
> > have stalled on several points. I'll try to summarise; let's see if we can move
> > forward.
> >
> > (I am the reporter of https://github.com/git-for-windows/git/issues/4125, which
> > led to this PR. I am vested in making progress here.)
> >
> > 1. name of the setting (`os` vs `uname-s` vs `sysname`)
> 
> I do not think it is a good idea to squat on too generic a name like
> 'os', especially when there are multiple layers people will care
> about.  But I think the original thread discussed this to death, and
> I do not see a point bringing it up again as the first bullet point.

"platform" is the term I've seen to be less OS-centric.

Johannes Schindelin April 19, 2023, 12:22 p.m. UTC | #15

Hi Junio,

On Mon, 17 Apr 2023, Junio C Hamano wrote:

> Samuel Ferencik <[e-mail]> writes:
>
> >>>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>` is the
> >>>> system name, i.e. the output of `uname -s`.
> >
> > The discussion about https://github.com/gitgitgadget/git/pull/1429 seems to
> > have stalled on several points. I'll try to summarise; let's see if we can move
> > forward.
> >
> > (I am the reporter of https://github.com/git-for-windows/git/issues/4125, which
> > led to this PR. I am vested in making progress here.)
> >
> > 1. name of the setting (`os` vs `uname-s` vs `sysname`)
>
> I do not think it is a good idea to squat on too generic a name like
> 'os', especially when there are multiple layers people will care
> about.  But I think the original thread discussed this to death, and
> I do not see a point bringing it up again as the first bullet point.

Given what you said about "Operating System", i.e. that both "Ubuntu" and
"Linux" should be able to match at the same time, I kind of concur, but in
the other direction: rather than forcing the current patch series to use a
less intuitive (i.e. user-unfriendlier) name than `os`, I would want to
modify the patch series so that it _can_ match "Ubuntu" and "Linux".

> > 2. casing (use of `/i`)
>
> My preference is to do this case sensitively (in other words, start
> stupid) and if somebody wants to use "/i", add it later after the
> dust settles.

I strongly disagree with this. This feature is meant for Git users, who I
must assume based on my experience would expect the value to be
case-insensitive. I.e. they would expect both `linux` and `Linux` to
match. Let's not make this feature harder to use than necessary!

> > 3. handling Windows (MinGW, WSL)
>
> This comes back to the reason why "os" is a horrible choice.  Is WSL
> a Windows?  Is WSL a Linux?  The same question can be asked for Cygwin.

These questions actually have pretty obvious answers, with the exception
of WSL1 (which nobody should use anymore): WSL is a Linux, Cygwin is not
an Operating System by itself but runs on Windows. But of course that's
not helpful to help configure Git specifically in a Cygwin setup.

> The answer depends on which layer you care about.

Yes, this is the crucial bit.

> The underlying kernel and system may be Windows, and some
> characteristics of the underlying system may seep through the
> abstraction, but these systems aim to give user experience of something
> like GNU/Linux.
>
> And this is not limited to Windows.  There may be similar issue for
> systems like PacBSD.  Is it a Linux?  Is it a BSD?
>
> > 6. what's the use-case?
>
> I think that this is the most important question to ask, and from
> here, we'd see how #3 above should be resolved (I suspect that you
> may want to have at least two layers to allow WSL to be grouped
> together with MinGW and Cygwin at one level, and at the same time
> allow it to be grouped together with Ubuntu at a different level).
> And after we figure that out, we'll have a clear and intuitive
> answer to #1.

This is probably the most valuable feedback in this entire thread: What is
the problem we're trying to solve here?

The original report (which this patch tries to address) asks for a way to
have a user-wide ("global") Git configuration that can be shared across
machines and that allows for adapting to the various environments. The
rationale is motivated well e.g. in
https://medium.com/doing-things-right/platform-specific-gitconfigs-and-the-wonderful-includeif-7376cd44994d
where platform-specific credential managers, editors, diff highlighters
that work only in certain setups, and work vs personal environments are
mentioned.

So as long as Git offers ways to discern between the mentioned
environments by including environment-specific configuration files, we
solve the problem.

Bonus points if we can do that without getting ever deeper into a pretty
contentious discussion about naming.

The strategy chosen in above-mentioned article uses the presence of
certain directories as tell-tales for the Operating System in which
Git is called: Linux, macOS or Windows. Concretely, it suggests this:

	[includeIf "gitdir:/Users"]
		path = ~/.gitconfig-macos
	[includeIf "gitdir:C:"]
		path = ~/.gitconfig-windows
	[includeIf "gitdir:/home"]
		path = ~/.gitconfig-linux

Now, the presence of directories like `/home/` might work well to discern
Linux from macOS, but this is not the correct way to identify Linux in
general (a `/home/` directory exists in many Unices, too). And it only
works when the _gitdir_ is in those directories, too. That's why I thought
that Git could do better.

In many cases, though, the presence of directories is probably "good
enough" to address the need described in above-mentioned article.

What triggered me to write this here patch was the report that those
`/home/` and `/Users` paths in the Git config ran into the warning that
Git for Windows no longer treats paths starting with a slash as relative
to the runtime prefix. This warning was introduced when the `%(prefix)/`
feature was upstreamed, superseding Git for Windows' original handling of
paths starting with a slash. The warning was introduced in November '21
(virtually together with Git for Windows v2.34.0):
https://github.com/git-for-windows/git/commit/28fdfd8a41836f49666c65e82d92de9befea2e69

So while I still think that something like `includeIf.os:<name>` would
make for a better way to address the concern in question, I realize that
the path of least resistance is simply to drop the now-deprecated feature
from Git for Windows (including the warning that was the reason for the
original report): https://github.com/git-for-windows/git/pull/4389

This still leaves Git in the somewhat unsatisfying state where there is no
better way to discern between Operating Systems than to work around by
detecting the presence of certain directories. But I do not see any viable
way to reach consensus about the `includeIf.os:<name>` patch, so I'll
leave it at that.

Ciao,
Johannes

Chris Torek April 19, 2023, 2:26 p.m. UTC | #16

Note: I'm going to mix two things here (maybe not the best idea)
and change the order.

On Wed, Apr 19, 2023 at 5:28 AM Johannes Schindelin
<[e-mail]> wrote:

> This is probably the most valuable feedback in this entire thread: What is
> the problem we're trying to solve here?
>
> The original report (which this patch tries to address) asks for a way to
> have a user-wide ("global") Git configuration that can be shared across
> machines and that allows for adapting to the various environments. The
> rationale is motivated well e.g. in
> https://medium.com/doing-things-right/platform-specific-gitconfigs-and-the-wonderful-includeif-7376cd44994d
> where platform-specific credential managers, editors, diff highlighters
> that work only in certain setups, and work vs personal environments are
> mentioned.

Why not allow use of environment variables?  Perhaps a simple:

git config include.env.USER_VAR

and/or:

git config include.ifexists.env.DIR

(feel free to invent better names!) or similar.  The user can then

export DIR=whatever

in some outside-of-Git setup.

On a separate note:

> ... This feature is meant for Git users, who I
> must assume based on my experience would expect the value to be
> case-insensitive. I.e. they would expect both `linux` and `Linux` to
> match.

I'm not at all sure about this: we already see a lot of confusion
from people who don't understand why, if Git is case-sensitive
about branch names (which it is), Git isn't case-sensitive about
branch names *sometimes* on Windows and macOS.

(In any case, using an env var sidesteps this question.)

Chris

Samuel Ferencik April 19, 2023, 2:32 p.m. UTC | #17

[resending my previous response, which I had sent only to the PR, not the
mailing list; composed before dscho@'s last email]

> > 2. casing (use of `/i`)
>
> My preference is to do this case sensitively (in other words, start
> stupid) and if somebody wants to use "/i", add it later after the
> dust settles.

Fantastic.

> > 6. what's the use-case?

> > * Note all of the use-cases above would cope with WSL reporting "Linux",
> >   except of `merge.tool`.
>
> I think that this is the most important question to ask, and from
> here, we'd see how #3 above should be resolved (I suspect that you
> may want to have at least two layers to allow WSL to be grouped
> together with MinGW and Cygwin at one level, and at the same time
> allow it to be grouped together with Ubuntu at a different level).

Actually, I take that back: I can install the Linux version of BeyondCompare
into WSL, use my license, and I'm fine with WSL reporting that it's Linux.
(This is just to correct my previous claim. More below.)

> > 3. handling Windows (MinGW, WSL)
>
> The answer depends on which layer you care about.  The underlying
> kernel and system may be Windows, and some characteristics of the
> underlying system may seep through the abstraction, but these
> systems aim to give user experience of something like GNU/Linux.

Fair point. I think we only care about the top-most layer (the one nearest to
the user).

The "seeping through" aspect means things often aren't clear-cut, but I also
think we can make practical decisions. The closer the top layer approximates an
OS, the more we can say it *is* that OS without compromising correctness. For
example, a Linux container on Windows should report (uname) it's a Linux, and
`includeIf` should go by this. So should WSL2, I think, because it provides a
Linux kernel. Maybe there is a use-case for git config (`includeIf`) to behave
differently in WSL because the host OS is Windows, but I'm not aware of such a
use-case. In contrast, "Git bash" and Cygwin report something other than "Linux"
(`MINGW64_NT-<version>`, `CYGWIN_NT-<version>`) because they are not Linuxes,
and that's also fine. Interestingly, though, they don't report the host OS
(Windows) either.

So I propose that we limit ourselves to the top-most layer. Can we come up with
a name (`os` or another) that would capture this concept?

- Felipe proposed `platform`. I'm afraid this may be confused with the hardware
  platform (as reported by `uname -m`).
- Phillip proposed `uname-s`. It's self-documenting in a way, but it's also
  wrong on Windows and describes the tool rather than the concept.
- Philip proposed `sysname`. It's slightly better but still very much a pointer
  to `utsname.sysname`.
- dscho@ proposed `os`. I like it for its obviousness. The issues with the name
  are marginal (not bigger than with any other name) and can be sufficiently
  addressed by good documentation.

If we can agree on the name, I can help fix up the documentation, the tests, and
remove the case insensitivity. Thoughts?

Randall S. Becker April 19, 2023, 3:21 p.m. UTC | #18

On Wednesday, April 19, 2023 8:23 AM, Johannes Schindelin wrote:
>On Mon, 17 Apr 2023, Junio C Hamano wrote:
>
>> Samuel Ferencik <[e-mail]> writes:
>>
>> >>>> Let's introduce a new condition: `os:<uname-s>` where `<uname-s>`
>> >>>> is the system name, i.e. the output of `uname -s`.
>> >
>> > The discussion about https://github.com/gitgitgadget/git/pull/1429
>> > seems to have stalled on several points. I'll try to summarise;
>> > let's see if we can move forward.
>> >
>> > (I am the reporter of
>> > https://github.com/git-for-windows/git/issues/4125, which led to
>> > this PR. I am vested in making progress here.)
>> >
>> > 1. name of the setting (`os` vs `uname-s` vs `sysname`)
>>
>> I do not think it is a good idea to squat on too generic a name like
>> 'os', especially when there are multiple layers people will care
>> about.  But I think the original thread discussed this to death, and I
>> do not see a point bringing it up again as the first bullet point.
>
>Given what you said about "Operating System", i.e. that both "Ubuntu" and
"Linux"
>should be able to match at the same time, I kind of concur, but in the
other direction:
>rather than forcing the current patch series to use a less intuitive (i.e.
user-
>unfriendlier) name than `os`, I would want to modify the patch series so
that it _can_
>match "Ubuntu" and "Linux".
>
>> > 2. casing (use of `/i`)
>>
>> My preference is to do this case sensitively (in other words, start
>> stupid) and if somebody wants to use "/i", add it later after the dust
>> settles.
>
>I strongly disagree with this. This feature is meant for Git users, who I
must assume
>based on my experience would expect the value to be case-insensitive. I.e.
they
>would expect both `linux` and `Linux` to match. Let's not make this feature
harder to
>use than necessary!
>
>> > 3. handling Windows (MinGW, WSL)
>>
>> This comes back to the reason why "os" is a horrible choice.  Is WSL a
>> Windows?  Is WSL a Linux?  The same question can be asked for Cygwin.
>
>These questions actually have pretty obvious answers, with the exception of
WSL1
>(which nobody should use anymore): WSL is a Linux, Cygwin is not an
Operating
>System by itself but runs on Windows. But of course that's not helpful to
help
>configure Git specifically in a Cygwin setup.
>
>> The answer depends on which layer you care about.
>
>Yes, this is the crucial bit.
>
>> The underlying kernel and system may be Windows, and some
>> characteristics of the underlying system may seep through the
>> abstraction, but these systems aim to give user experience of
>> something like GNU/Linux.
>>
>> And this is not limited to Windows.  There may be similar issue for
>> systems like PacBSD.  Is it a Linux?  Is it a BSD?
>>
>> > 6. what's the use-case?
>>
>> I think that this is the most important question to ask, and from
>> here, we'd see how #3 above should be resolved (I suspect that you may
>> want to have at least two layers to allow WSL to be grouped together
>> with MinGW and Cygwin at one level, and at the same time allow it to
>> be grouped together with Ubuntu at a different level).
>> And after we figure that out, we'll have a clear and intuitive answer
>> to #1.
>
>This is probably the most valuable feedback in this entire thread: What is
the problem
>we're trying to solve here?
>
>The original report (which this patch tries to address) asks for a way to
have a user-
>wide ("global") Git configuration that can be shared across machines and
that allows
>for adapting to the various environments. The rationale is motivated well
e.g. in
>https://medium.com/doing-things-right/platform-specific-gitconfigs-and-the-
>wonderful-includeif-7376cd44994d
>where platform-specific credential managers, editors, diff highlighters
that work only
>in certain setups, and work vs personal environments are mentioned.
>
>So as long as Git offers ways to discern between the mentioned environments
by
>including environment-specific configuration files, we solve the problem.

I would suggest using match of uname arguments. But there are variants. The
OS release should also be included here as something like NONSTOP_KERNEL is
not a sufficient answer. We should have at the very least, or encode the
includeif string accordingly:

uname -r = the release n (e.g., J06)
uname -n = the node name (e.g., hpitug)
uname -s = the OS (e.g., NONSTOP_KERNEL

We use these in config.mak.uname (except for -n).

Regards,
Randall

Junio C Hamano April 19, 2023, 4:07 p.m. UTC | #19

Johannes Schindelin <[e-mail]> writes:

>> > 2. casing (use of `/i`)
>>
>> My preference is to do this case sensitively (in other words, start
>> stupid) and if somebody wants to use "/i", add it later after the
>> dust settles.
>
> I strongly disagree with this. This feature is meant for Git users, who I
> must assume based on my experience would expect the value to be
> case-insensitive. I.e. they would expect both `linux` and `Linux` to
> match. Let's not make this feature harder to use than necessary!

[jc: Read below with 'os' replaced with 'uname-s' or any of your
favorite. This message does not take any position on that part of
the discussion.]

I am OK with the above position, if we make sure that all other
includeIf conditions are compared case insensitively, and if we make
sure there is no existing "/i" includeIf conditions.  Then those who
want to differentiate can later add "/casesensitive" option.

But I do not want to see a system where some of the conditions are
compared case insensitively while some other conditions are compared
case sensitively.  The end-users will then be forced to remember
"the condition 'os' can be spelled in any case permutation, but the
condition 'xyzzy' needs to be spelled in the right case".  It would
not be obvious to end users when they need to use "/i" variant in
such a system.

Unlike 'gitdir' whose value is arbitrarily chosen by the users and
the projects (where folks use both FooBar and foobar and want them
to mean different things), the vocabulary of "os" is limited and I
agree with you that it is very likely for any user to expect that
any case permutations of "linux" would mean the same thing.  But
spelling "windows" and have it match even when the official name of
the system may be "Windows" should be the choice made by the end
user, and the "/i" suffix in "os/i" is a way for them to express it.

You might be able to talk me into adding only "os/i" added without
adding "os", though.  I do not see much point in doing so, other
than that we can say "we do not take any position on what case
permutation is the right and official way to spell the name of each
system".

Thanks.

Junio C Hamano April 19, 2023, 4:14 p.m. UTC | #20

Johannes Schindelin <[e-mail]> writes:

> The original report (which this patch tries to address) asks for a way to
> have a user-wide ("global") Git configuration that can be shared across
> machines and that allows for adapting to the various environments.
> ...
> So as long as Git offers ways to discern between the mentioned
> environments by including environment-specific configuration files, we
> solve the problem.

Perhaps 

    [includeIf "ifExists:/path/to/foo-credential-manager.exe"]
	path = /path/to/config/needed/for/using/foo-credential-manager

is being called for?  "os" being "LiNuX" does not tell much about
the environment differences the end users would care about, like
what desktop environment is in use, which leads to the availability
and default choice of GUI tools, to help them auto-configure.

Patch

diff --git a/Documentation/config.txt b/Documentation/config.txt
index e376d547ce0..b90bcd8ecfe 100644
--- a/Documentation/config.txt
+++ b/Documentation/config.txt
@@ -186,6 +186,11 @@  As for the naming of this keyword, it is for forwards compatibiliy with
 a naming scheme that supports more variable-based include conditions,
 but currently Git only supports the exact keyword described above.
 
+`os`::
+	The data that follows this keyword is taken as the name of an
+	Operating System, e.g. `Linux` or `Windows`; If it matches the
+	current Operating System, the include condition is met.
+
 A few more notes on matching via `gitdir` and `gitdir/i`:
 
  * Symlinks in `$GIT_DIR` are not resolved before matching.
diff --git a/config.c b/config.c
index 9b0e9c93285..9ab311ae99b 100644
--- a/config.c
+++ b/config.c
@@ -394,6 +394,15 @@  static int include_by_remote_url(struct config_include_data *inc,
 					     inc->remote_urls);
 }
 
+static int include_by_os(const char *cond, size_t cond_len)
+{
+	struct utsname uname_info;
+
+	return !uname(&uname_info) &&
+		!strncasecmp(uname_info.sysname, cond, cond_len) &&
+		!uname_info.sysname[cond_len];
+}
+
 static int include_condition_is_true(struct config_include_data *inc,
 				     const char *cond, size_t cond_len)
 {
@@ -408,6 +417,8 @@  static int include_condition_is_true(struct config_include_data *inc,
 	else if (skip_prefix_mem(cond, cond_len, "hasconfig:remote.*.url:", &cond,
 				   &cond_len))
 		return include_by_remote_url(inc, cond, cond_len);
+	else if (skip_prefix_mem(cond, cond_len, "os:", &cond, &cond_len))
+		return include_by_os(cond, cond_len);
 
 	/* unknown conditionals are always false */
 	return 0;
diff --git a/t/t1309-early-config.sh b/t/t1309-early-config.sh
index 537435b90ae..b36afe1a528 100755
--- a/t/t1309-early-config.sh
+++ b/t/t1309-early-config.sh
@@ -100,4 +100,20 @@  test_expect_success 'onbranch config outside of git repo' '
 	nongit git help
 '
 
+test_expect_success '[includeIf "os:..."]' '
+	test_config x.y 0 &&
+	echo "[x] y = z" >.git/xyz &&
+
+	if test_have_prereq MINGW
+	then
+		uname_s=Windows
+	else
+		uname_s="$(uname -s)"
+	fi &&
+	test_config "includeIf.os:not-$uname_s.path" xyz &&
+	test 0 = "$(git config x.y)" &&
+	test_config "includeIf.os:$uname_s.path" xyz &&
+	test z = "$(git config x.y)"
+'
+
 test_done