Go: cmd/go: TestScript/get_update_all changes mtimes in GOROOT/bin

[bcmills]

2020-02-28T20:15:38-964fac3/linux-amd64-longtest

test_race_install_cgo was converted to a script test in CL 214426. The old test was not run in parallel, and asserted that go test -race -i runtime/race did not reinstall cmd/cgo.

I suspect that this flake was caused by some _other_ test somehow overwriting cmd/cgo when it should not have.

[bcmills]

2020-03-01T05:45:39-1e9665d/windows-amd64-longtest

[dmitshur]

Another failure on windows-amd64-longtest at a7f918c1 from today.

[bcmills]

2020-03-16T22:31:39-2fbca94/windows-amd64-longtest
2020-03-15T00:17:52-dad94e7/windows-amd64-longtest
2020-03-13T19:06:00-be72e3c/linux-amd64-longtest
2020-03-11T19:59:39-f6a0d72/windows-amd64-longtest
2020-03-09T20:19:25-5fac45a/linux-amd64-longtest

I think this is a pairwise-incompatibility with some other test, but so far I haven't been able to track it down running each cmd/go test pairwise with test_race_install_cgo with a high -count.

[gopherbot]

Change https://golang.org/cl/223745 mentions this issue: cmd/go: replace TestCgoDependsOnSyscall with a simpler script test

[gopherbot]

Change https://golang.org/cl/223746 mentions this issue: cmd/go: add a "don't care about success" operator to script_test

[bcmills]

I'm hoping that CL 223745 does the trick, but there may be other interactions lurking.

[bcmills]

That CL didn't do it: https://build.golang.org/log/d9ad32ca35be46c7fac61bf609655f3e3100a803

At least the feedback was quick!

[gopherbot]

Change https://golang.org/cl/223755 mentions this issue: cmd/dist: run a checkNotStale on the builders before running tests

[bcmills]

FINALLY found it!

Doesn't seem to reproduce at all when running as a not-root user. That's why it was so [redacted] hard to track down.

For some reason, get_update_all adjusts file _attributes_ (but not content!) on ~everything in GOROOT/pkg/tool, but _only_ when running as root.

~/bisect/src$ sudo ../bin/go test cmd/go -run='TestScript/(get_update_all|test_race_install_cgo)'
go test proxy running at GOPROXY=http://127.0.0.1:41653/mod
--- FAIL: TestScript (0.02s)
    --- FAIL: TestScript/test_race_install_cgo (0.95s)
        script_test.go:202:
            # Tests Issue #10500 (0.938s)
            > [!race] skip
            > ! stale cmd/cgo
            > env GOBIN=$WORK/bin
            > go install mtime sametime
            > go tool -n cgo
            [stdout]
            /usr/local/google/home/bcmills/bisect/pkg/tool/linux_amd64/cgo
            > cp stdout cgopath.txt
            > exec $GOBIN/mtime cgopath.txt # get the mtime of the file whose name is in cgopath.txt
            [stdout]
            "2020-03-17T23:53:40.190152563-04:00"
            > cp stdout cgotime_before.txt
            > ? go test -race -i runtime/race
            > exec $GOBIN/mtime cgopath.txt # get the mtime of the file whose name is in cgopath.txt
            [stdout]
            "2020-03-17T23:54:37.402358823-04:00"
            > cp stdout cgotime_after.txt
            > exec $GOBIN/sametime cgotime_before.txt cgotime_after.txt
            [stderr]
            time in cgotime_before.txt (2020-03-17 23:53:40.190152563 -0400 EDT) is not the same as time in cgotime_after.txt (2020-03-17 23:54:37.402358823 -0400 EDT)[exit status 1]
            FAIL: testdata/script/test_race_install_cgo.txt:22: unexpected command failure

FAIL
FAIL    cmd/go  2.813s
FAIL

[bcmills]

I think I've figured out what's going on.

Our own test from CL 12174 (2015) was rendered unparallelizable by our own intentional breakage of the non-intrusive behavior that test was relying on in CL 76590 (2017).

https://github.com/golang/go/blob/938ad552cdfdb1707acedcc641c8f8ada642644d/src/cmd/go/internal/work/exec.go#L1465-L1476

[gopherbot]

Change https://golang.org/cl/223919 mentions this issue: cmd/go: make 'go get' and 'go install' avoid updating mtimes of binaries that are not stale

[bcmills]

Retitled for the diagnosed root cause.

[gopherbot]

Change https://golang.org/cl/223925 mentions this issue: cmd/go: rebuild cmd/cgo in test_race_install_cgo instead of asserting that it is not stale

[bcmills]

Retitled again. After discussion with @rsc, we decided that the mtime change is correct and should be kept, and instead we should prohibit the test version of the cmd/go binary from attempting to change metadata in GOROOT/bin.

[gopherbot]

Change https://golang.org/cl/225897 mentions this issue: cmd/go/internal/work: disallow testgo binary from installing to GOROOT