Perhaps somewhat relevant, the go command has a special behavior where if a nested module root doesn't have a file named "LICENSE", the LICENSE file is copied from the repository root if one exists there (see coderepo.go#L1027-L1032). Notably, it applies only to files named "LICENSE", not "COPYING".

CC @julieqiu, @jayconrod.

/cc @jba

pkg.go.dev works on modules, not repos. We fetch module zips from proxy.golang.org and process them. The zip for your nested module does not include a license file:

> curl -L -o /tmp/netstack.zip https://proxy.golang.org/golang.zx2c4.com/wireguard/tun/netstack/@v/v0.0.0-20210120232502-fcc8ad05df75.zip
> unzip -l /tmp/netstack.zip 
Archive:  /tmp/netstack.zip
  Length      Date    Time    Name
---------  ---------- -----   ----
     1047  1980-00-00 00:00   golang.zx2c4.com/wireguard/tun/netstack@v0.0.0-20210120232502-fcc8ad05df75/examples/http_client.go
     1223  1980-00-00 00:00   golang.zx2c4.com/wireguard/tun/netstack@v0.0.0-20210120232502-fcc8ad05df75/examples/http_server.go
      218  1980-00-00 00:00   golang.zx2c4.com/wireguard/tun/netstack@v0.0.0-20210120232502-fcc8ad05df75/go.mod
    40172  1980-00-00 00:00   golang.zx2c4.com/wireguard/tun/netstack@v0.0.0-20210120232502-fcc8ad05df75/go.sum
    20835  1980-00-00 00:00   golang.zx2c4.com/wireguard/tun/netstack@v0.0.0-20210120232502-fcc8ad05df75/tun.go
---------                     -------
    63495                     5 files

You should have a license file in each module rather than each repo, or you can take advantage of the special behavior that @dmitshur pointed out above.

You should have a license file in each module rather than each repo,

Thanks, but no thanks. The licensing concerns of my repository are per-repository, and I have no intention of changing this, or entering less-charted legal territory of a repository with differing licenses in different places. This isn't going to happen.

While I see the technical reason why this is challenging -- the proxy pre-makes zips for you -- I still believe this to be a bug in some part of the Go stack. Perhaps that proxy zipper should copy the license files in? Perhaps x/pkgsite should pursue a different way of license discovery.

But I can't imagine an acceptable solution being to change license-per-repo, which has lots of legal precedent, with license-per-go-module, whatever that might be in the mind of a judge.

Perhaps somewhat relevant, the go command has a special behavior where if a nested module root doesn't have a file named "LICENSE", the LICENSE file is copied from the repository root if one exists there (see coderepo.go#L1027-L1032). Notably, it applies only to files named "LICENSE", not "COPYING".

CC @julieqiu, @jayconrod.

Thanks. This seems like good behavior. I'll send a CL to extend that to COPYING.

Change https://golang.org/cl/285352 mentions this issue: modfetch: copy COPYING in addition to LICENSE

@dmitshur wrote on that CL:

A change like this may cause content of existing modules to change, invalidating their hashes. I don't know whether how viable it is to roll it out. Maybe it's possible by gating behind the Go language version stated in go.mod. All this is likely better discussed in an issue.

So, responding here.

This is a good point. I can update the CL to gate it on go.mod version. (Or, if you've got some way to dump the proxy server, you could just test that hypothesis on all known modules.) But I wonder if that's happening if there's anything else we should consider too. Should this actually take the full list of license file names that x/pkgsite recognizes?

Alternatively, if you feel that I should just rename COPYING to LICENSE, that would be perhaps an easier solution and just fine, but in that case, would it make sense to discuss changing x/pkgsite to only look at LICENSE and not the others, in order to keep the tooling semantics consistent across the ecosystem?

Alternatively, if you feel that I should just rename COPYING to LICENSE, that would be perhaps an easier solution and just fine, but in that case, would it make sense to discuss changing x/pkgsite to only look at LICENSE and not the others, in order to keep the tooling semantics consistent across the ecosystem?

I sent a CL in for this strategy. Feel free to -2 it if you hate that idea, but I figured I'd prepare it to give you the option. https://go-review.googlesource.com/c/pkgsite/+/285273

This issue is closely related to #39417. Copying my comments from there:

I won't say it's impossible, but it's very difficult to make any change to the set of files that get included in a module .zip file.

If different versions of Go choose different sets of files, they won't generate the same hashes, so when a module is downloaded by an old version of the go command it will look like someone has tampered with the module after it was released. That could be somewhat mitigated by gating the change on a new go version in go.mod. We'd have to choose a version far enough in the future (at least a year, probably more) that by the time it was current, versions of Go that don't support this are no longer supported.

It's also hard to decide what files to actually include: LICENSE, LICENSE.txt, sure. How about COPYING.html? copyright-license.pdf? pkg.go.dev has a list of names they use, and we'd probably use the same list. Not sure what they do if there are multiple matching files.

...

Just to close the loop on this: @bcmills, @matloob, and I discussed this last week, and we decided not to change anything here. Some mitigations are possible, but fundamentally, this would cause different versions of the go command to produce different sums for the same module, which would cause security errors. The cost of that outweighs the benefit.

Okay. I suppose I'll drop this, and rename COPYING->LICENSE.