Allow OS_THREAD_LIBSPACE_NUM as a macro

[AriParkkila]

Description

OS_THREAD_LIBSPACE_NUM is defined to 4, which is not a good default in all cases. This change makes it possible to give OS_THREAD_LIBSPACE_NUM also as a macro.

Mbed netsocket tests over cellular require OS_THREAD_LIBSPACE_NUM to be 5.

Pull request type

[x] Fix
[ ] Refactor
[ ] Target update
[ ] Functionality change
[ ] Docs update
[ ] Test update
[ ] Breaking change

Reviewers

@kivaisan

Release Notes

[0xc0170]

Changes to RTX should be also sent upstream, see https://github.com/ARM-software/CMSIS_5/blob/develop/CMSIS/RTOS2/RTX/Config/RTX_Config.h - this file should get the update there. Please send a patch for review there.

All the rest of config there contain a check if not already defined but not this one, so should be accepted as fix.

[kivaisan]

• I think @kjbracey-arm should review this from connectivity/core point-of-view.

[kjbracey]

Hmm, why isn't this configurable in RTX? Seems like it should be a setting that can be set, whether or not OS_THREAD_OBJ_MEM is set. Even if it is set, OS_THREAD_MEM only limits system-allocated threads, not total, so you might need more.

What actually triggers the failure? Presumably it's not just number of threads - is it number of threads that access errno?

I'm actually kind of happy to know that errno is actually thread-local on at least 1 toolchain - I thought it never was...

Agree that it should be submitted for upstreaming.

Although having to configure it at all is a bit poor - personally rather than having a separate pool of these things, and needing to search that pool, I'd put them in with the other thread info. They could live inside the os_thread_t itself, or be placed in the stack block. I guess that does mean more RAM though - guaranteeing it on every thread, rather than just threads that need it.

In the Mbed OS part, I'd make it a proper config option in the RTOS mbed_lib.json - some people might want even more. (Or less).

[AriParkkila]

The root cause might be that __user_perthread_libspace reserves TLS, which is never released (not even when a thread is terminated). As @kjbracey-arm suggested TLS should probably be part of thread's structs to follow its lifecycle.

@0xc0170 Thanks for pointing out that upstream commit, I'll do it.

[kjbracey]

The root cause might be that __user_perthread_libspace reserves TLS, which is never released (not even when a thread is terminated).

Ah, that's significant. I'd definitely suggest trying an RTX patch to add a free there too.

[0xc0170]

Once done upstream, let us know , should be aligned with changes here

[AriParkkila]

@0xc0170 Please find upstream PR at ARM-software/CMSIS_5#691

[AriParkkila]

@0xc0170 Upstream PR ARM-software/CMSIS_5#691 was merged. This should be ready for merge.

[0xc0170]

CI started

[mbed-ci]

Test run: FAILED

Summary: 1 of 4 test jobs failed
Build number : 1
Build artifacts

Failed test jobs:

• jenkins-ci/mbed-os-ci_unittests

[0xc0170]

Unittests failed [2019-10-18T11:28:44.880Z] 31: mbed assertation failed: 0, file: /builds/ws/mbed-os-ci_unittests@8/unitTest-4201/mbed-os/features/lorawan/LoRaWANStack.cpp, line 884 , please review

[mbed-ci]

Test run: SUCCESS

Summary: 11 of 11 test jobs passed
Build number : 2
Build artifacts

[0xc0170]

cc @ARMmbed/mbed-os-test This might need CI update?