JDK 20 update broke non-ascii argument encoding on macOS

[tetromino]

Description of the bug:

Non-ascii arguments to spawned tools (both in genrule and in ctx.actions.run) are encoded as mojibake garbage with Bazel at HEAD on macOS.

Everything works fine with Bazel 6.2.1 on macOS, and with any Bazel version (including at HEAD) on Linux.

bazelisk bisect identified ecf9b9a as the culprit.

What's the simplest, easiest way to reproduce this bug? Please provide a minimal example if possible.

BUILD file:

genrule(
    name = "genrule",
    cmd = "echo привет > \"$@\"",
    outs = ["genrule.out"],
)

With Bazel at HEAD on macOS, get garbage:

$ USE_BAZEL_VERSION=last_green bazelisk build //:genrule && cat bazel-bin/genrule.out
2023/06/27 14:55:01 Using unreleased version at commit 044a14cca2747aeff258fc71eaeb153c08cb34d5
INFO: Analyzed target //:genrule (0 packages loaded, 0 targets configured).
INFO: Found 1 target...
Target //:genrule up-to-date:
  bazel-bin/genrule.out
INFO: Elapsed time: 0.152s, Critical Path: 0.00s
INFO: 1 process: 1 internal.
INFO: Build completed successfully, 1 total action
пÑ�ивеÑ

With Bazel 6.2.1 on macOS, get expected result:

$ bazelisk build //:genrule && cat bazel-bin/genrule.out
INFO: Analyzed target //:genrule (5 packages loaded, 7 targets configured).
INFO: Found 1 target...
Target //:genrule up-to-date:
  bazel-bin/genrule.out
INFO: Elapsed time: 3.882s, Critical Path: 0.10s
INFO: 2 processes: 1 internal, 1 darwin-sandbox.
INFO: Build completed successfully, 2 total actions
привет

Which operating system are you running Bazel on?

macOS Ventura 13.4.1

What is the output of bazel info release?

development version

If bazel info release returns development version or (@non-git), tell us how you built Bazel.

Bazelisk with USE_BAZEL_VERSION=last_green

What's the output of git remote get-url origin; git rev-parse master; git rev-parse HEAD ?

No response

Is this a regression? If yes, please try to identify the Bazel commit where the bug was introduced.

ecf9b9a

Have you found anything relevant by searching the web?

No response

Any other information, logs, or outputs that you want to share?

No response

[aiuto]

/sub

[aiuto]

I think there are ways to make this fail on linux too.
If we spill args to a file instead of the command line I am seeing that a single 'é' will emit as the 4 octal bytes 303 203 302 251.

é = u00E9 =utf-8=> C3 A9 =(utf-8 again)=> C3 83 C2 A9 =%o=> 303 203 302 251 

That's a clear 8859-1 -> UTF-8 mis-encoding.

[meisterT]

cc @zhengwei143 @coeuvre

[aiuto]

This is a problem for linux also if you do it right.

bazelbuild/rules_pkg#712

• is a PR that should be good, but fails on a unicode problem only on macos with bazel at head.
• it creates an action by making an args list of a bunch of strings. When the command line gets built, the args are written just fine.

bazelbuild/rules_pkg#714

• takes the same code base, but creates the action by doing the proper thing - calling ctx.actions.args() and using arg.append for each arg.
• it we don't force spliling args to a file. It mostly works just like 712.
• if we do force writing the params file, it fails in many ways.

[tetromino]

I think that the CI failures in bazelbuild/rules_pkg#714 may be a red herring: it's bug (or more likely, several bugs) in the argparse module in certain versions of python, not a bug in Bazel :/

Look at the errors in https://buildkite.com/bazel/rules-pkg/builds/2466#01890405-7614-4d1f-a408-14e3bff80a92: there is no good reason for argparse to decode the param file using ascii given that you have set PYTHONIOENCODING, PYTHONUTF8, LANG, and LC_CTYPE.

I think argparse will finally be properly fixed in python 3.12 (currently in beta) by python/cpython#93277

[tetromino]

But bazelbuild/rules_pkg#714 does demonstrate a Bazel failure specific to macOS: the param file (bazel-out/darwin_x86_64-fastbuild/bin/tests/deb/fizzbuzz_4.5.6_all.deb-0.params) gets incorrectly encoded.

[aiuto]

I think that the CI failures in bazelbuild/rules_pkg#714

<bazelbuild/rules_pkg#714> may be a red herring: it's bug (or more likely, several bugs) in the argparse module in certain versions of python, not a bug in Bazel :/ The argparse thing is a red herring, but that is just an additional layer of problem. If I look at the .params files produced I can see the bad double encoding. So that problem is below argparse, but I don't have a good way to show that. I just won't have time to really make a good reproduction for the error cases until late next week.

…

On Wed, Jun 28, 2023 at 6:54 PM Alexandre Rostovtsev < ***@***.***> wrote: But bazelbuild/rules_pkg#714 <bazelbuild/rules_pkg#714> does demonstrate a Bazel failure specific to macOS: the param file (bazel-out/darwin_x86_64-fastbuild/bin/tests/deb/fizzbuzz_4.5.6_all.deb-0.params) gets incorrectly encoded. — Reply to this email directly, view it on GitHub <#18792 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAXHHHF6VATKD3KEA4CSY5LXNSY2XANCNFSM6AAAAAAZWBGMRQ> . You are receiving this because you commented.Message ID: ***@***.***>

[tetromino]

I think there are at least two separate problems in Bazel here, with separate code paths:

• argument vector encoding when spawning a process on macos - fixed by f00439d
• parameter file encoding on macos