libc: Implement userland implementations of exit functions

[pussuw]

Summary

Implement atexit(), on_exit(), exit() to libc. Using the call gate versions is problematic as they basically inject userland code into the kernel, which is then run with kernel privileges. In any case the functions belong to libc.

Impact

Exit functions are executed with user privileges

Testing

mpfs icicle:knsh

[Ouss4]

@pussuw there seem to be some warnings:

stdlib/lib_cxa_atexit.c:83:13: error: '__cxa_callback' defined but not used [-Werror=unused-function]

CI turns warnings into errors, I don't know if you do that too, but you can with: make EXTRAFLAGS="-Werror".

(there are some make[3]: arm-nuttx-eabi-gcc: Command not found as well, there is an issue with the script apparently, please ignore it for now.)

[Ouss4]

The other build error which I listed above is way beyond my understanding. The make recipe works locally for the targets that fail in CI.

It could be a parallel make issue as it seemed that the dependencies were messed up. I took a quick look at libs/Makefile, nothing stands out. I hope (sorry :) ) that it can be reproduced once the warning is fixed.

[Ouss4]

@pussuw Apparently we have a name collision with https://github.com/apache/incubator-nuttx/blob/master/libs/libc/stdlib/lib_Exit.c and your newly created file lib_exit.c. I won't pretend that I understand the case sensitivity of all file systems, but I least am not able to create lib_exit.c under MacOS when lib_Exit.c exists.

[pussuw]

@pussuw Apparently we have a name collision with https://github.com/apache/incubator-nuttx/blob/master/libs/libc/stdlib/lib_Exit.c and your newly created file lib_exit.c. I won't pretend that I understand the case sensitivity of all file systems, but I least am not able to create lib_exit.c under MacOS when lib_Exit.c exists.

@Ouss4 thanks for doing the investigation! That makes sense then, since I don't get the error but I'm on Linux in which file names are case sensitive. I'll merge the two files, makes no sense to have two.

[pkarashchenko]

Seems like something on the ARM side still pulls __dso_handle and since cxa_atexit.c is not compiled the weak reference is not there. Need to move the weak symbol declaration somewhere else.

arm-none-eabi-ld: /github/workspace/sources/nuttx/staging/libapps.a(symtab.c.github.workspace.sources.apps.examples.elf.o):(.rodata.g_elf_exports+0x1c): undefined reference to `__dso_handle'

I still think that we are missing extern

[pkarashchenko]

LGTM and great work!

[pussuw]

Thanks for the review, turned into quite a magnus opus in the end.

[masayuki2009]

@Ouss4

I noticed that the ostest with esp32-devkitc:smp (real board not qemu) causes a deadlock.

user_main: waitpid test

Test waitpid()
waitpid_main: PID 28 Started
waitpid_start_child: Started waitpid_main at PID=28
waitpid_main: PID 29 Started
waitpid_start_child: Started waitpid_main at PID=29
waitpid_main: PID 30 Started
waitpid_start_child: Started waitpid_main at PID=30
waitpid_test: Waiting for PID=28 with waitpid()
waitpid_main: PID 28 exitting with result=14
waitpid_main: PID 29 exitting with result=14
waitpid_test: PID 28 waitpid succeeded with stat_loc=0e00
waitpid_last: Waiting for PID=30 with waitpid()
waitpid_main: PID 30 exitting with result=14
waitpid_last: PASS: PID 30 waitpid succeeded with stat_loc=0e00

End of test memory usage:
VARIABLE  BEFORE   AFTER
======== ======== ========
arena       4cda0    4cda0
ordblks         5       12
mxordblk    2b730    2b730
uordblks     b960     bce0
fordblks    41440    410c0

user_main: mutex test
Initializing mutex
Starting thread 1
Starting thread 2

[Ouss4]

@masayuki2009 did you enable any other option?
I just tried quickly running ostest 2 times and both passed.

Actually, I've been using the latest master the whole day for some tests.

[masayuki2009]

@masayuki2009 did you enable any other option?

No

I just tried quickly running ostest 2 times and both passed.
Actually, I've been using the latest master the whole day for some tests.

Hmm, I confirmed that the latest upstream works.
It seems that #6286 fixed the issue.
However, I'm still not understand why the PR fixed the issue.

[xiaoxiang781216]

Yes, it's strange. #6286 could improve the performance which may reduce the race condition.