In the 1.8 binary releases for linux, go test -i -race tries to re-build some packages from the standard library. If your go install is not writable, it will error trying to open the $pkg.a files for writing.

What version of Go are you using (go version)?

Today's release: go version go1.8 linux/amd64
Specifically, the binary release from: https://storage.googleapis.com/golang/go1.8.linux-amd64.tar.gz
Sha384sum: 5cf34c2335341febb707008ede937729c5cf09f66c98e11065ecc58f945fc7d95e1ed10c1d8de765591de5cf20c16457 go1.8.linux-amd64.tar.gz

What operating system and processor architecture are you using (go env)?

linux_amd64

What did you do?

Given a reproducer_test.go file:

package main

import "testing"

func Test(t *testing.T) {
	_ = 1
}

I run `go test in several different modes:

• go test ./... -- passes
• go test -race ./... -- passes
• go test -i ./... -- passes
• go test -i -race ./... -- FAILS

Go was installed by untar'ing the binary release. All of its uids are 0; all of its mtimes are as given by the tar headers.
The tarball is extracted in /installpath, and export GOROOT=/installpath/go applied to the environment.
(Identical steps have been used to install a go1.7 binary release, which does not exhibit the problem.)

What did you expect to see?

All four of these test commands should pass. (All four commands work correctly on go1.7.)

What did you see instead?

Test with -i and -race-mode gives the following error output:

go install internal/race: open /installpath/go/pkg/linux_amd64_race/internal/race.a: permission denied
go install errors: open /installpath/go/pkg/linux_amd64_race/errors.a: permission denied
go install unicode/utf8: open /installpath/go/pkg/linux_amd64_race/unicode/utf8.a: permission denied
go install unicode: open /installpath/go/pkg/linux_amd64_race/unicode.a: permission denied
go install math: open /installpath/go/pkg/linux_amd64_race/math.a: permission denied

and exits with a 1.

Other notes

When encountering this on less minimal code, I've seen several other packages also listed as offenders, such as crypto/subtle for example. The five listed above are what comes up with the minimal reproducer -- these five packages are dependencies of the testing package.

I thought perhaps this might be an issue from timestamps on the .a files being older than the relevant source files (I've made a mistake with an rsync command before that fubar'd that, and paid silly increases in compile time until I noticed and fixed it), but that doesn't appear to be the case, so I'm not sure what's going on here.