Is malloc() thread-safe in ARMC6?

[ccli8]

Description

• Type: Question
• Priority: Major

---

Question

Target
NUMAKER_PFM_NUC472

Toolchain:
ARM|ARMC6

Toolchain version:
ARM: 5.06
ARMC6: 6.9

mbed-cli version:
1.4.0

mbed-os sha:
f6bcbfe Merge pull request #6356 from hug-dev/asm-include-flags

How to?
According to Arm C and C++ Libraries and Floating-Point Support User Guide, malloc() can be made thread-safe by implementing the _mutex_*() functions. I want to check its validity in ARMC6.

I compile Greentea test code mbed-os-tests-mbedmicro-rtos-mbed-malloc with ARM toolchain and check generated MAP file malloc.map. I can find _mutex_*() functions (_mutex_initialize/acquire/release/free) are included, so I am convinced malloc() is thread-safe.

mbed test -m NUMAKER_PFM_NUC472 -t ARM -n mbed-os-tests-mbedmicro-rtos-mbed-malloc --compile

But then I compile the same code with ARMC6 toolchain and check generated MAP file malloc.map. I just find _mutex_initialize and doesn't find _mutex_acquuire/release/free. Is malloc thread-safe in ARMC6, or there's some other mechanism?

mbed test -m NUMAKER_PFM_NUC472 -t ARMC6 -n mbed-os-tests-mbedmicro-rtos-mbed-malloc --compile

[0xc0170]

https://github.com/ARMmbed/mbed-os/blob/master/rtos/TARGET_CORTEX/mbed_boot.c#L442

They are present for ARMCC, can you verify they are compiled for ARMC6 that you are using?

In the user guide for ARMCC 6, it is still defined same API for thread safe allocators (section 1.6.8 Management of locks in multithreaded applications there).

[ccli8]

In the user guide for ARMCC 6, it is still defined same API for thread safe allocators (section 1.6.8 Management of locks in multithreaded applications there).

@0xc0170 Yes. But per my compiling with ARMC6 (V6.9), I didn't find _mutex_acquuire/_mutex_release included in generated MAP file like ARMCC (V5.06). So I question thread-safety of malloc with ARMC6 in Mbed OS.

The test mbed-os-tests-mbedmicro-rtos-mbed-malloc is to test malloc in multiple threads. It is passed in NUMAKER_PFM_NUC472 with both ARM/ARMC6. But I still want to clarify it to avoid potential thread-safety issue.

[pan-]

@0xc0170 _mutex_acquire is defined in RTX I'm not sure to understand why the function is defined as weak; it doesn't seems right.

[0xc0170]

@0xc0170 _mutex_acquire is defined in RTX I'm not sure to understand why the function is defined as weak; it doesn't seems right.

I believe that is protected via macro that we define not to be enabled. Same is for some startup hooks that are defined there (we define our owns in our startup - mbed boot code file). I assume that is a reason why is not weakly defined.

cc @JonatanAntoni

[ccli8]

@0xc0170 @pan- Any update?

[JonatanAntoni]

Sorry, I don't understand your issue. The multi-thread library integration in RTX is defined weak to allow a redefinition if needed. In case of mbedOS I expect these function to be defined somewhere else to be compliant with your memory allocation style.

[ccli8]

@JonatanAntoni I know the link mechanism and expect it should behave in such way. But after I inspect generated MAP file, I find it doesn't go as expected.

I can find _mutex_* symbols in ARM toolchain generated MAP file:

    0x000068b0   0x00000030   Code   RO         9770    i._mutex_acquire    C:/ccli8/iot/mbed/mbed-os/nu-greentea/BUILD/tests/NUMAKER_PFM_NUC472/ARM/mbed-os/rtos/TARGET_CORTEX/rtx5/RTX/Source/rtx_lib.o
    0x000068e0   0x00000006   Code   RO         9771    i._mutex_free       C:/ccli8/iot/mbed/mbed-os/nu-greentea/BUILD/tests/NUMAKER_PFM_NUC472/ARM/mbed-os/rtos/TARGET_CORTEX/rtx5/RTX/Source/rtx_lib.o
    0x000068e6   0x00000002   PAD
    0x000068e8   0x000000b8   Code   RO         9200    i._mutex_initialize  C:/ccli8/iot/mbed/mbed-os/nu-greentea/BUILD/tests/NUMAKER_PFM_NUC472/ARM/mbed-os/rtos/TARGET_CORTEX/mbed_boot.o
    0x000069a0   0x00000020   Code   RO         9772    i._mutex_initialize  C:/ccli8/iot/mbed/mbed-os/nu-greentea/BUILD/tests/NUMAKER_PFM_NUC472/ARM/mbed-os/rtos/TARGET_CORTEX/rtx5/RTX/Source/rtx_lib.o
    0x000069c0   0x00000028   Code   RO         9773    i._mutex_release    C:/ccli8/iot/mbed/mbed-os/nu-greentea/BUILD/tests/NUMAKER_PFM_NUC472/ARM/mbed-os/rtos/TARGET_CORTEX/rtx5/RTX/Source/rtx_lib.o

But except _mutex_initialize, I cannot find other _mutex_* symbols in ARMC6 toolchain generated MAP file:

    0x000057f4   0x000000a8   Code   RO         6163    .text._mutex_initialize  mbed_boot.o
    0x0000589c   0x00000004   Code   RO         6682    .text._platform_post_stackheap_init  rtx_lib.o

Because _mutex_acquire/_mutex_free symbols are not linked in with ARMC6 (or I get wrong), I question the thread safety of malloc series functions with ARMC6.

@0xc0170 @pan-

[deepikabhavnani]

@JonatanAntoni - I believe the question here should be why __USED define is only set for ARMCC version < 6. https://github.com/ARMmbed/mbed-os/blob/master/rtos/TARGET_CORTEX/rtx5/RTX/Source/rtx_lib.c#L623

@ccli8 has submitted a PR for this fix (#6973), will it make sense to upstream the change in CMSIS Repo?

[deepikabhavnani]

@JonatanAntoni - Sorry my bad, change is done in mbed-os repo only, CMSIS is fine.

[geky]

Responding to @deepikabhavnani, I believe the __USED defines are behind the ifdef because I was only able to test ARMC5 at the time. (ARMC6 has frontend changes and that pr needed to go in quickly). If they work in ARMC6 we should adopt them there as well.

[ciarmcom]

ARM Internal Ref: MBOTRIAGE-215

[deepikabhavnani]

Fix merged to master #6973