Fix data-race problems in git module (quick patch)

[wxiaoguang]

This patch is somewhat hacky, it might be the most simple one to resolve the data-race problem in git.Init.

Details:

Before git refactoring: git.Init was only called once in GlobalInitInstalled, which is incomplete, because the tests also need to init the git module.

With the last git refactoring #19732: git.Init was called in GlobalInitInstalled and many prepareTestEnv functions, but git.Init changes many global variables, then it leads to data-race problem.

This PR: introduce git.InitOnceWithSync which only changes the global variables once, just like what git.CheckLFSVersion does.

Although this PR is not ideal and hacky, (hopefully) it should work and be enough for 1.17 release. In the future, the setting related code should be refactored (eg: by using atomic.Value), including the git.CheckLFSVersion.

[zeripath]

By using a sync.Once we're limited to configuring this function once and once only per run (which is fine for Gitea as we are currently) but this also applies to the tests too.

Any attempt to call this function to reconfigure will fail but that means it could be possible that a different order of calling Tests could cause a different configuration to be set.

So ... the first function calling this Init should probably be being called from within the TestMain(s) to ensure that there is a consistent configuration.

[wxiaoguang]

By using a sync.Once we're limited to configuring this function once and once only per run (which is fine for Gitea as we are currently) but this also applies to the tests too.

Yes, but the same problem to git.CheckLFSVersion ......

Without a big refactoring, the problem can not be resolved clearly.

[wxiaoguang]

Why I think this PR is enough:

Before refactoring:

• Gitea web: call git.Init once
• Gitea serv: no call to git.Init
• Test code: no call to git.Init before GlobalInitInstalled (which calls git.Init)

After the refactoring and this PR:

• Gitea web: call git.InitOnceWithSync once, it can be guaranteed to be correct
• Gitea serv: call git.InitSimple once, not bad, no side effect
• Test code: call git.InitOnceWithSync in prepare with the same settings as GlobalInitInstalled, and at the moment everything works fine.

Indeed, the situation of refactoring the git module is more complex than I thought. IMO at the moment it's impossible to achieve the ideal status with a few PRs, especially there is a blocker in git.CheckLFSVersion, the changing to them will cause a lot of changed code.

And I am open to learn ideas about how to make the git module better.

[zeripath]

I guess we're somewhat forced to do the syncOnce here.

[codecov-commenter]

Codecov Report

Merging #19934 (ba99d59) into main (73382d2) will decrease coverage by 0.02%.
The diff coverage is 57.24%.

@@            Coverage Diff             @@
##             main   #19934      +/-   ##
==========================================
- Coverage   47.32%   47.30%   -0.03%     
==========================================
  Files         959      965       +6     
  Lines      133652   134013     +361     
==========================================
+ Hits        63252    63389     +137     
- Misses      62699    62904     +205     
- Partials     7701     7720      +19     

Impacted Files	Coverage Δ
cmd/admin.go	0.00% <0.00%> (ø)
cmd/serv.go	2.22% <0.00%> (-0.11%)	⬇️
models/db/engine.go	32.65% <ø> (-0.91%)	⬇️
models/repo/repo.go	64.66% <0.00%> (-0.49%)	⬇️
models/repo/repo_unit.go	77.14% <ø> (ø)
models/webhook/webhook.go	68.81% <0.00%> (-1.26%)	⬇️
modules/git/commit.go	56.92% <ø> (+0.38%)	⬆️
modules/git/lfs.go	66.66% <ø> (+7.84%)	⬆️
modules/git/remote.go	80.00% <ø> (+11.57%)	⬆️
modules/git/repo_attribute.go	69.17% <ø> (+0.65%)	⬆️
... and 114 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 23422f9...ba99d59. Read the comment docs.