I did a bit of digging on this one, I found out that the permissions are
indeed changed by make - see test/TestCommon.gmk:
# Prep for output
# Change execute permissions on shared library files.
# Files in repositories should never have execute permissions, but
# there are some tests that have pre-built shared libraries, and these
# windows dll files must have execute permission. Adding execute
# permission may happen automatically on windows when using certain
# versions of mercurial but it cannot be guaranteed. And blindly
# adding execute permission might be seen as a mercurial 'change', so
# we avoid adding execute permission to repository files. But testing
# from a plain source tree needs the chmod a+rx. Applying the chmod to
# all shared libraries not just dll files. And with CYGWIN and sshd
# service, you may need CYGWIN=ntsec for this to work.
prep:
@$(MKDIR) -p $(ABS_TEST_OUTPUT_DIR)
@$(MKDIR) -p `$(DIRNAME) $(ARCHIVE_BUNDLE)`
@if [ ! -d $(TEST_ROOT)/../.hg ] ;
then \
$(FIND) $(TEST_ROOT) \( -name \*.dll -o -name \*.DLL -o -name
\*.so \) \
-exec $(CHMOD) a+rx {} \;
; \
fi
This part seems to be responsible for changing the permissions on the
said lib files. That said, TestCommon.gmk is essentially the same as it
was in jdk10 unconsolidated, but yet the problem cannot be reporoduced
in the latest unconsolidated jdk10 repo. So, I believe something is up.
More specifically, this test:
if [ ! -d $(TEST_ROOT)/../.hg ] ; then \
was probably failing before, causing the permission changing branch not
to be taken?
Maurizio
On 26/09/17 12:10, Maurizio Cimadamore wrote:
On 26/09/17 07:13, Alan Bateman wrote:
On 25/09/2017 21:35, Maurizio Cimadamore wrote:
On 25/09/17 20:58, Alan Bateman wrote:
On 25/09/2017 20:11, Jonathan Gibbons wrote:
Erik,
It could be a feature of the build (i.e. test makefiles) to verify
that no source-controlled files were modified in the course of a
test run.
Something fishy in this thread as these tests have historically not
changed the permissions of these libraries. So I'm curious how they
were run, I wonder in case they were changed by a make file instead.
make test TEST=jdk_lang
AFAIK, this should be the correct way of invoking tests from make?
I assume this make file must be running chmod, not the tests
themselves as otherwise this issue would have been noticed a long
time ago. I suspect many people run jtreg directly in their local
environment rather than the make file.
Note that I've been running tests this way over the last 2-3 years,
and this only started to happen after the move to consolidated repo.
That's why I correlated the two things.
Maurizio
In any case, these older tests predate build test support for
compiling native code and probably should be migrated anyway.
-Alan
