commit ecf4ebf10054f70e51a0ce759b2ae91aa4febd1a
author Dmitri Shuralyov <dmitshur@golang.org> Sat Dec 05 15:03:27 2020 -0500
committer Dmitri Shuralyov <dmitshur@golang.org> Wed Jan 20 22:16:54 2021 +0000
tree a7e089d4ce6603af0135f96ec990c06b06721308
parent d2d155d1ae8c704a37f42fd3ebb1f3846f78e4d4

cmd/internal/moddeps: check content of all modules in GOROOT

Expand the scope of the TestAllDependenciesVendored test to check
that all modules in GOROOT are tidy, that packages are vendored,
the vendor content matches the upstream copy exactly, and that
bundled packages are re-generated (using x/tools/cmd/bundle at
the version selected in cmd module; this is deterministic and
guaranteed to be updated over time).

This is done in a conceptually simple way:

1.	Make a temporary copy of the entire GOROOT tree (except .git),
	one that is safe to modify.
2.	Run a list of high-level commands, the same commands we expect
	Go developers should be able to run in a normal complete GOROOT
	tree to make it clean and tidy.
3.	Diff the end result with the original GOROOT tree being tested
	to catch any unexpected differences.

The current set of commands that are run require the cmd/go command,
and a functional compiler itself (because re-generating the syscall
package involves a directive like //go:generate go run [...]). As a
result, copying a large majority of the GOROOT tree is a requirement.

Instead of looking for the few files or directories that can we can
get away not copying (e.g., the testdata directories aren't strictly
needed at this time), we opt not to optimize and just do the simple
copy. This is motivated by these reasons:

•	We end up having a complete, normal GOROOT tree, one that happens
	to be located at another path. There's a very high likelihood that
	module management/code generation commands, both the ones we run
	today and any additional ones that we might want to add in the
	future, will result in correct results even as the Go project
	evolves over time.

•	Having a completely stand-alone copy of the GOROOT tree without
	symlinks minimizes the risk of some of the module management/code
	generation commands, either now or in the future, from modifying
	the user's original GOROOT tree, something that should not happen
	during test execution. Overlays achieved with symlinks work well
	when we can guarantee only new files are added, but that isn't
	the case here.

•	Copying the entire GOROOT (without .git), takes around 5 seconds
	on a fairly modern computer with an SSD. The most we can save is
	a couple of seconds.

(We make some minor exceptions: the GOROOT/.git directory isn't copied,
and GOROOT/{bin,pkg} are deemed safe to share and thus symlink instead
of copying. If these optimizations cease to be viable to make, we'll
need to remove them.)

Since this functionality is fairly expensive to execute and requires
network access, it runs only when the test is executed without -short
flag. The previous behavior of the TestAllDependenciesVendored test is
kept in -short test mode. all.bash runs package tests with -short flag,
so its behavior is unchanged. The expectation is that the new test will
run on some of the longtest builders to catch problems. Users can invoke
the test manually 'go test cmd/internal/moddeps' (and it's run as part
of 'go test cmd', again, only when -short flag isn't provided).

On a 2017 MacBook Pro, a successful long test takes under 15 seconds,
which should be within scope of all long tests that are selected by
'go test std cmd'. We may further adjust when and where the test runs
by default based on our experience.

Fixes #36852.
Fixes #41409.
Fixes #43687.
Updates #43440.

Change-Id: I9eb85205fec7ec62e3f867831a0a82e3c767f618
Reviewed-on: https://go-review.googlesource.com/c/go/+/283643
Run-TryBot: Dmitri Shuralyov <dmitshur@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Bryan C. Mills <bcmills@google.com>
Trust: Dmitri Shuralyov <dmitshur@golang.org>