Building mbedtls 2.2.0 on RHEL 5 leads to "No rule to make target `library/libmbedtls.so', needed by `tests/test_suite_aes.cbc.c'"

[robert-scheck]

Trying to build mbedtls 2.2.0 on RHEL 5 leads to the failure:

make[2]: *** No rule to make target `library/libmbedtls.so', needed by `tests/test_suite_aes.cbc.c'. Stop.

Running

sed -e 's/libmbedtls.so$/libmbedtls.so.2.2.0/' -i tests/CMakeFiles/test_suite_*.dir/build.make

after "cmake" and before "make" works around the issue. Interestingly, this only affects RHEL 5, not RHEL 6 or 7, thus this must be related to cmake somehow.

[ciarmcom]

ARM Internal Ref: IOTSSL-596

[mpg]

To be honest, I looks more like a CMake issue than an mbed TLS issue, if it works with RHEL 6 and 7 but not 5. Did you try asking CMake folks what they think about it?

Also, though I realize this might not be very satisfactory, if CMake is not working for you, we have good old Makefiles too.

[NWilson]

It could be that RHEL 5 has a very old version of CMake. You could try installing a recent CMake version, just to see if the problem then goes away.

Here, we use CMake on a variety of very old Unix and Linux platforms, and we actually installed the latest CMake from the source distribution when we moved from autoconf, since we couldn't find packages for all our build machines.

[robert-scheck]

RHEL 5 has cmake in Version 2.6.4 – which we can't upgrade in EPEL, because it is a RHEL package. Do you see chances to get the CMakefiles adapted accordingly so that they support this cmake version (like mbedtls 1.3.x did)?

[mpg]

I'm certainly not opposed in principle to supporting older CMake versions if we can, but right now it isn't obvious to me what we changed between 1.3 and now that broke the CMake build with old versions. If you can help us pinpoint what the culprit is, I'll be able to give you a better answer as to whether we can adapt our CMakelists.txt to solve your issue. If you're able to git-bisect the issue and point us to the commit that introduces it, that would probably be a good starting point.

[robert-scheck]

The problem is caused by your "make apidoc" target. Running "make -j4 all" seems to create the symlink, however "make apidoc" removes the symlink.

[mpg]

I read your patch and commit message, and while I agree that using cp -p in this case is desirable, I still think that:

• it should work without -p if you do make -j4 all && make apidoc instead of make -j4 all apidoc
• and you shouldn't do make -j4 all apidoc anyway because it's racy and you risk building the library with the wrong config.h modified by apidoc depending on the order things are built.

Anyway, I'm considering reverting the change to the apidoc target, because it's quite counter-intuitive and error-prone that make -j4 all apidoc is now racy.

[robert-scheck]

When I apply my -p patch from #392 on RHEL 5, make -j4 all apidoc fails like #390 is doing on RHEL 7 and 6. However, #392 together with make -j4 all && make apidoc works fine on RHEL 7, 6 and 5 for me.

Given I understand that make -j4 all apidoc is racy, I had a discussion with @mortenstevens already about maybe using make -j4 all && make apidoc for Fedora and RHEL (which makes our *.spec file hopefully more readable again).

[mpg]

Does make -j4 all && make apidoc also work without #392? I think it should work (expect that if you run make all again after that it will re-build instead of outputing "nothing to do" as expected, which I agree is bad). I'm asking because if it fails, that means there is another issue I didn't understand yet, and I'd rather understand everything that's going on, in order to avoid making similar mistakes in the future.

Again, sorry for breaking your build script, especially on a patch release, that was definitely not the intention.

[robert-scheck]

No. make -j4 all works as it should, however make apidoc kills the symlinks for library/libmbed*.so and this causes at latest a failure at make install then (already make apidoc fails if you drop the sed command from my initial posting).

This is because the header file timestamp changed, thus it needs to rebuild the library. As part of that, the symlinks library/libmbed*.so are partially removed and not created again. So the -p is essential for RHEL 5.

[mpg]

I'll see if I can reproduce the issue with cmake 2.8.9 (the oldest version I can easily access). I'm reverting the change anyway, but I'd really like to understand everything that's going on here.