| Message ID | pull.1069.git.git.1629576007891.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Give support for hooks based on platform | expand |
On 2021-08-21 at 20:00:07, Rafael Santiago via GitGitGadget wrote: > From: rafael-santiago <voidbrainvoid@tutanota.com> > > The idea behind this commit can be useful for teams > that share git-hooks into a custom directory and > dealing with projects that must be developed, > built, maintained on several different platforms. > > This commit allows the execution of git hooks > based on the current operating system. > A "native hook" is defined in the form: > hooks/hook-name_platform > Where platform must be equivalent to the > content returned in sysname field in utsname > struct when calling uname() [but all normalized > in lowercase]. > > On Windows, independent of version, flavor, SP, > whatever it is simply "windows". I'm not sure that this is going to work out very well. The MINGW environment used by Git for Windows and Cygwin are quite different. I would fully expect to write shell scripts and Unix tooling in Cygwin, whereas users using Git for Windows might not want that at all. That also doesn't take into effect using Git for Windows in WSL, which introduces some interesting logistical challenges. In addition, I have a few concerns about the grouping of Linux altogether. While in many cases it is possible to write tooling that works natively across Linux distros, most binaries will not. Therefore, binary hooks that might run fine on Debian would fail on a Fedora or Red Hat system, especially if those binaries link to any of a number of different shared libraries (e.g., OpenSSL). There's also work to move hooks into the config and out of the hooks directory, and I don't think this will mesh well with it. > The main motivation of this extension is to > reduce dependency of scripting languages, > logical trinkets etc just to execute minor > tasks during scm events that could be done > natively but differently from a platform > to another. Less dependencies, cleaner > repos: a small step for a better world > for any software developer. Is there a reason that the proper hooks couldn't be copied or symlinked into place with a script? I think that would resolve this concern with a lot less work.
In my opinion "binary hooks" (hooks that execute specific binaries not present in the system as a default tool) should be versioned and built as a support tool into the repository or in the worst case downloaded from somewhere, even because versioning binaries into git repos is considered a bad practice that could make bloated repos. The point is that in many cases a dependency with a script language is created only to make the hook actions portable from a platform to other, but what this script in its essence does is a thing that could be done with basic tools delivered with the current operating system. There is no problem on using cygwin on windows, you should use standard hook and do all the effort to make it unique for cygwin environments and true unix boxes (in other words: you would continue doing what you are doing, because it attends yours requirements). Notice that everything that have been working will stay working as before. Anyway, if cygwin becomes a point of incompatibility at some point, you could use the "_windows" version by coding your "cygwin script" there. Maybe this would add more possibilities in git-hook tooling by making them less plastered, I think. Rafael Santiago
On 2021-08-21 at 23:11:27, Rafael Santiago wrote: > In my opinion "binary hooks" (hooks that execute specific binaries not > present in the system as a default tool) should be versioned and built > as a support tool into the repository or in the worst case downloaded > from somewhere, even because versioning binaries into git repos is > considered a bad practice that could make bloated repos. Yes, I agree binary hooks should not be checked into the repository. > The point is that in many cases a dependency with a script language is > created only to make the hook actions portable from a platform to > other, but what this script in its essence does is a thing that could > be done with basic tools delivered with the current operating system. Then, in general, it can be done in a shell script containing an if-then statement per platform using the native tools, so I'm not seeing the particular reason that this series is necessary if the hooks being executed aren't binaries. All systems on which Git runs must contain a POSIX-compatible shell. Can you explain the rationale for your proposal in more detail so that we can understand why this change is necessary? Typically this is done in the commit message, but I don't think I understand why you want to do this. > There is no problem on using cygwin on windows, you should use > standard hook and do all the effort to make it unique for cygwin > environments and true unix boxes (in other words: you would continue > doing what you are doing, because it attends yours requirements). > Notice that everything that have been working will stay working as > before. Anyway, if cygwin becomes a point of incompatibility at some > point, you could use the "_windows" version by coding your "cygwin > script" there. Right, my point is that your commit message proposes using "windows" for Cygwin. The patch doesn't, but your commit message says that every version of Windows is considered "windows".
Well, you are taking into consideration that every git user that needs to automate some stuff during a scm event will do it from a shell (I meant true shell or a mocked up one such as cygwin, msys, etc) and it is true, by design. However, not all git users (out unix) uses git from "bash-like" environments. I know people that prefers using it from a well-cooked IDE by clicking and dragging things or even from command prompt when on Windows. Those people are not able to handle some scm event automation as unix users because git hook in its essence presupposes the existence/necessity of a powerful shell (okay, it is possible to put a shebang and call a batch file on windows, maybe, but it is a little clumsy, in my opinion). On Windows, users can do a bunch of stuff just by using the ready to go powershell, but open an if/else on a bash script to run a cygwin instance by calling powershell from there is not a good and clean solution for this type of user. Presupposing shell for git, limitates the idea behind the scm event handling with hooks, because currently it is strongly dependent from shell to work on every git supported platform. The idea of having hooks being executed by platform would be the first step to give support to execute commands on scm events without obligating users out of a unix have a shell interpreter to access native stuff. Currently, this commit does not implement it but would be possible to do and in a less noisy way for all unix-like stuff. I am not sure but currently a _windows hook out from cygwin would result on a spawn error, would not? Git hooks are useful features but would be more useful if it breaked up the shell jail. It could make git much more integrated with the current platform. Being possible to make it powerful as it is on a unix even on a total different platform as Windows, let's say. For sure, this commit are not a "panacea" but intends to start making git-hooks more independent from 3rd party software to work on as expected, on every platform that a git-repo is expected to be handled. I hope I was clearer from this time. Rafael Santiago -- 22 de ago. de 2021 19:07 por sandals@crustytoothpaste.net: > On 2021-08-21 at 23:11:27, Rafael Santiago wrote: > >> In my opinion "binary hooks" (hooks that execute specific binaries not >> present in the system as a default tool) should be versioned and built >> as a support tool into the repository or in the worst case downloaded >> from somewhere, even because versioning binaries into git repos is >> considered a bad practice that could make bloated repos. >> > > Yes, I agree binary hooks should not be checked into the repository. > >> The point is that in many cases a dependency with a script language is >> created only to make the hook actions portable from a platform to >> other, but what this script in its essence does is a thing that could >> be done with basic tools delivered with the current operating system. >> > > Then, in general, it can be done in a shell script containing an if-then > statement per platform using the native tools, so I'm not seeing the > particular reason that this series is necessary if the hooks being > executed aren't binaries. All systems on which Git runs must contain a > POSIX-compatible shell. > > Can you explain the rationale for your proposal in more detail so that > we can understand why this change is necessary? Typically this is done > in the commit message, but I don't think I understand why you want to do > this. > >> There is no problem on using cygwin on windows, you should use >> standard hook and do all the effort to make it unique for cygwin >> environments and true unix boxes (in other words: you would continue >> doing what you are doing, because it attends yours requirements). >> Notice that everything that have been working will stay working as >> before. Anyway, if cygwin becomes a point of incompatibility at some >> point, you could use the "_windows" version by coding your "cygwin >> script" there. >> > > Right, my point is that your commit message proposes using "windows" for > Cygwin. The patch doesn't, but your commit message says that every > version of Windows is considered "windows". > -- > brian m. carlson (he/him or they/them) > Toronto, Ontario, CA >
On Sun, Aug 22, 2021 at 10:07:41PM +0000, brian m. carlson wrote: > > The point is that in many cases a dependency with a script language is > > created only to make the hook actions portable from a platform to > > other, but what this script in its essence does is a thing that could > > be done with basic tools delivered with the current operating system. > > Then, in general, it can be done in a shell script containing an if-then > statement per platform using the native tools, so I'm not seeing the > particular reason that this series is necessary if the hooks being > executed aren't binaries. All systems on which Git runs must contain a > POSIX-compatible shell. This is my gut feeling, too (whether users know it or not, even on Windows most programs specified by config are being run by the shell). However, I do think there is room for Git to make this case a bit easier: conditional config includes. Once we are able to specify hooks via config (which is being worked on elsewhere), then we ought to be able to implement an includeIf like: [includeIf "uname_s:linux"] path = linux-hooks.config [includeIf "uname_s:windows"] path = windows-hooks.config The advantage being that this could apply to _all_ config, and not just hooks. It may still require fighting over which values should match on cygwin. ;) -Peff
"brian m. carlson" <sandals@crustytoothpaste.net> writes: > Then, in general, it can be done in a shell script containing an if-then > statement per platform using the native tools, so I'm not seeing the > particular reason that this series is necessary if the hooks being > executed aren't binaries. All systems on which Git runs must contain a > POSIX-compatible shell. When we start defining our hooks in the configuration, this may fit with the conditional inclusion of configuration files. Current conditions only can depend on where the repository is, but it is easy to imagine that a conditional inclusion based on the value of the configuration variable, so [includeIf "var:dev.host=mac"] path = ... [includeIf "var:dev.host=win"] path = ... might be a way to say "if the dev.host configuration (presumably set in somewhere like /etc/gitconfig or ~/.gitconfig) is set to this value, take the configuration from the specified path. It is up to the project to define the variable they use to switch on; some project may ship with a set of hooks that can be used on both windows and cygwin at the same time, in which case they do not need the distinction between the two, and some other project may care about the distinction. Git does not have to impose or enforce any policy about the granularity of what "the platform" is, with such a scheme.
Jeff King <peff@peff.net> writes: > On Sun, Aug 22, 2021 at 10:07:41PM +0000, brian m. carlson wrote: > >> > The point is that in many cases a dependency with a script language is >> > created only to make the hook actions portable from a platform to >> > other, but what this script in its essence does is a thing that could >> > be done with basic tools delivered with the current operating system. >> >> Then, in general, it can be done in a shell script containing an if-then >> statement per platform using the native tools, so I'm not seeing the >> particular reason that this series is necessary if the hooks being >> executed aren't binaries. All systems on which Git runs must contain a >> POSIX-compatible shell. > > This is my gut feeling, too (whether users know it or not, even on > Windows most programs specified by config are being run by the shell). > > However, I do think there is room for Git to make this case a bit > easier: conditional config includes. Once we are able to specify hooks > via config (which is being worked on elsewhere), then we ought to be > able to implement an includeIf like: > > [includeIf "uname_s:linux"] > path = linux-hooks.config > [includeIf "uname_s:windows"] > path = windows-hooks.config > > The advantage being that this could apply to _all_ config, and not just > hooks. Heh, it seems great minds think alike.
On Mon, Aug 23, 2021 at 10:59:28AM -0700, Junio C Hamano wrote: > Jeff King <peff@peff.net> writes: > > > On Sun, Aug 22, 2021 at 10:07:41PM +0000, brian m. carlson wrote: > > > >> > The point is that in many cases a dependency with a script language is > >> > created only to make the hook actions portable from a platform to > >> > other, but what this script in its essence does is a thing that could > >> > be done with basic tools delivered with the current operating system. > >> > >> Then, in general, it can be done in a shell script containing an if-then > >> statement per platform using the native tools, so I'm not seeing the > >> particular reason that this series is necessary if the hooks being > >> executed aren't binaries. All systems on which Git runs must contain a > >> POSIX-compatible shell. > > > > This is my gut feeling, too (whether users know it or not, even on > > Windows most programs specified by config are being run by the shell). > > > > However, I do think there is room for Git to make this case a bit > > easier: conditional config includes. Once we are able to specify hooks > > via config (which is being worked on elsewhere), then we ought to be > > able to implement an includeIf like: > > > > [includeIf "uname_s:linux"] > > path = linux-hooks.config > > [includeIf "uname_s:windows"] > > path = windows-hooks.config > > > > The advantage being that this could apply to _all_ config, and not just > > hooks. > > Heh, it seems great minds think alike. One important distinction in what you wrote is that you're expecting the user to set dev.host once. That nicely sidesteps any question of "how does Git label each platform?", but it does mean the user has to do that setup manually (which maybe is amortized across many repos, but in practice for many people I suspect is no better than them setting up the correct "include" in the first place). I hoped that by calling it "uname_s", it would be clear it was the same as "uname -s", and then we could blame any naming confusion on the OS. :) But even if it is not used for this particular application, I think the [includeIf "var:..."] you proposed might be a reasonable thing to support. -Peff
diff --git a/run-command.c b/run-command.c index f72e72cce73..973c1a3434b 100644 --- a/run-command.c +++ b/run-command.c @@ -1319,9 +1319,50 @@ int async_with_fork(void) #endif } +static inline const char *platform_name(void) +{ + static const char *platform = NULL; +#ifndef GIT_WINDOWS_NATIVE + static struct utsname un = { 0 }; +#endif + if (platform != NULL) + return platform; + +#ifndef GIT_WINDOWS_NATIVE + if (uname(&un) == 0) { + for (size_t s = 0; un.sysname[s] != 0; s++) + un.sysname[s] = tolower(un.sysname[s]); + platform = un.sysname; + } +#else + platform = "windows"; +#endif + + return platform; +} + +static const char *find_native_hook(const char *name) +{ + char native_name[64] = ""; + const char *platform = NULL; + if (name == NULL || strstr(name, "_") != NULL) + return NULL; + + platform = platform_name(); + if (platform == NULL) + return NULL; + + if (snprintf(native_name, sizeof(native_name) - 1, "%s_%s", name, platform) >= sizeof(native_name) - 1) + return NULL; + return find_hook(native_name); +} + const char *find_hook(const char *name) { + const char *native_hook = find_native_hook(name); static struct strbuf path = STRBUF_INIT; + if (native_hook != NULL) + return native_hook; strbuf_reset(&path); strbuf_git_path(&path, "hooks/%s", name); diff --git a/t/t7527-pre-commit-native-hook.sh b/t/t7527-pre-commit-native-hook.sh new file mode 100755 index 00000000000..f3835f943af --- /dev/null +++ b/t/t7527-pre-commit-native-hook.sh @@ -0,0 +1,63 @@ +#!/bin/sh + +test_description='Test native hooks extension' + +. ./test-lib.sh + +expected_platform=$(uname -s | tr A-Z a-z) + +if [ $(expr substr $(uname -s | tr A-Z a-z) 1 5) == "mingw" ] ; then + expected_platform="windows" +fi + +test_expect_success 'set standard and native pre-commit hooks' ' + mkdir -p test-repo && + cd test-repo && + git init && + mkdir -p .git/hooks && + echo \#!/bin/sh > .git/hooks/pre-commit && + echo echo Hello generic. >> .git/hooks/pre-commit && + chmod u+x .git/hooks/pre-commit && + echo \#!/bin/sh > .git/hooks/pre-commit_${expected_platform} && + echo echo Hello ${expected_platform} >> .git/hooks/pre-commit_${expected_platform} && + chmod u+x .git/hooks/pre-commit_${expected_platform} && + echo test > README && + git add README && + git commit -am "1-2-3 this is a test." 2>out.txt && + cat out.txt | grep Hello\ ${expected_platform} +' + +if [ ${expected_platform} != "windows" ] ; then + # chmod does not work well on Windows. + test_expect_success 'set standard and native pre-commit hooks but let the native one not executable' ' + mkdir -p test-repo && + cd test-repo && + git init && + mkdir -p .git/hooks && + echo \#!/bin/sh > .git/hooks/pre-commit && + echo echo Hello generic. >> .git/hooks/pre-commit && + chmod u+x .git/hooks/pre-commit && + echo \#!/bin/sh > .git/hooks/pre-commit_${expected_platform} && + echo echo Hello ${expected_platform} >> .git/hooks/pre-commit_${expected_platform} && + echo test > README && + git add README && + git commit -am "1-2-3 this is a test." 2>out.txt && + cat out.txt | grep Hello\ generic + ' + + test_expect_success 'set standard pre-commit hook only' ' + mkdir -p test-repo && + cd test-repo && + git init && + mkdir -p .git/hooks && + echo \#!/bin/sh > .git/hooks/pre-commit && + echo echo Hello standard hook. >> .git/hooks/pre-commit && + chmod u+x .git/hooks/pre-commit && + echo test > README && + git add README && + git commit -am "1-2-3 this is a test." 2>out.txt && + cat out.txt | grep Hello\ standard\ hook + ' +fi + +test_done