Travis is okay.
https://travis-ci.org/ruby/ruby/builds/610461332

Rebsaed on the latest master branch.

I believe it's Travis-specific issue. Do you have any other ideas to fix the issue without touching the build system (common.mk, Makefile.in)?

I believe it's Travis-specific issue. Do you have any other ideas to fix the issue without touching the build system (common.mk, Makefile.in)?

Why do you think so?

I think the reason of the error is

$ make -jN
  => ext/foo/Makefile executed.

For example. In ext/json/Makefile, following part is executed.

$(TIMESTAMP_DIR)/.RUBYCOMMONDIR.time:
    $(Q) $(MAKEDIRS) $(@D) $(RUBYLIBDIR)
    $(Q) $(TOUCH) $@

Then

$ make install

In ext/json/Makefile, following part is executed wrongly, but the part should not be executed in make install in Travis.

$(RUBYLIBDIR)/json.rb: $(srcdir)/lib/json.rb $(TIMESTAMP_DIR)/.RUBYCOMMONDIR.time
    $(Q) $(COPY) $(srcdir)/lib/json.rb $(@D)

So, this issue can be happened when running make -jN in multi CPU cores environment.
I assume as it is -j33 in arm64, it can happen easily.

But if you want to fix the issue without touching the build system (common.mk, Makefile.in), maybe just adding following code fixes it.

$ git diff
diff --git a/.travis.yml b/.travis.yml
index 8b72afaaae..f30c270df7 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -453,6 +453,10 @@ before_script:
   - rm -rf ~/config_2nd
   - mv ../config_2nd ~
   - chmod u-w ..
+  - |-
+    if [ "$(uname -m)" = aarch64 ]; then
+      mkdir -p .ext/common .ext/.timestamp && touch .ext/.timestamp/.RUBYCOMMONDIR.time
+    fi
   - $SETARCH make -s $JOBS

Or run make COPY='cp -f' install instead of make install in case of aarch64 case to pass the make install step forcely.

Thank you for the report.
But this doesn't seem to fix the issue enough.

Thank you for your fixing.

Seeing your commit: fb02e31 .

Now I understand that there has already been the similar logic for .ext/$(arch) directory in common.mk.

$(TIMESTAMPDIR)/$(arch)/.time:
    $(Q)$(MAKEDIRS) $(@D) $(EXTOUT)/$(arch)
    @$(NULLCMD) > $@

And my modification was not at a proper place and was not the proper way in it.

Seeing https://travis-ci.org/ruby/ruby/builds after the commit, arm64-linux case is 100% success 4 times.

I think that it's time to remove arm64 from allow_failures again adapting this commit.
212f4d4

I think that it's time to remove arm64 from allow_failures again adapting this commit.

I sent PR for that.
#2672