Making the newlib thread-safe

[haukepetersen]

When calling puts/printf after thread_create(), the CPU hangs. This seems to be connected to the mutex used in the uart driver for syncing the DMA transfer in uart_write(). When disabling DMA for a quick test, everything works fine. So can it be, that the mutex is somehow effected when a new thread is created?

[haukepetersen]

The problem is indeed related to f4's UART driver, it seems that the DMA enabled driver is not thread safe anymore...

[PeterKietzmann]

Sorry, I won't find the time to check this until friday :(

[DipSwitch]

I see that the atomic actions in mutex are different from what I always use. I have seen the volatile
At work I use these: https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html and/or https://gcc.gnu.org/onlinedocs/gcc-4.4.5/gcc/Atomic-Builtins.html for ARM.

But however, I'm porting an STM32F2xx now which I base of the STM32F4xx, the board and CPU port are ready so I could take a look tomorrow evening. It uses the DMA controller to since it's the same as the F4 :)

[haukepetersen]

Found the reason. The missbehavior is caused by the newlib not being threadsafe...

[DipSwitch]

For newlib to be made threadsafe you need to add the following to the task struct:

    /* Allocate a Newlib reent structure that is specific to this task.
    Note Newlib support has been included by popular demand, but is not
    used by the FreeRTOS maintainers themselves.  FreeRTOS is not
    responsible for resulting newlib operation.  User must be familiar with
    newlib and must provide system-wide implementations of the necessary
    stubs. Be warned that (at the time of writing) the current newlib design
    implements a system-wide malloc() that must be provided with locks. */
    struct  _reent xNewLib_reent;

And then on every task swap you need to update the impure pointer.

    #if ( configUSE_NEWLIB_REENTRANT == 1 )
    {
        /* Switch Newlib's _impure_ptr variable to point to the _reent
        structure specific to this task. */
        _impure_ptr = &( pxCurrentTCB->xNewLib_reent );
    }

To every struct and implement the following functions:

void __env_lock ( struct _reent *_r )
{
#if OS_THREAD_SAFE_NEWLIB
    if (xTaskGetSchedulerState() == taskSCHEDULER_NOT_STARTED)
        return;

    vTaskSuspendAll();
#endif /* OS_THREAD_SAFE_NEWLIB */
    return;
}

/* __env_unlock needs to provide recursive mutex unlocking */

void __env_unlock ( struct _reent *_r )
{
#if OS_THREAD_SAFE_NEWLIB
    if (!xTaskGetSchedulerState() == taskSCHEDULER_NOT_STARTED)
        return;

    xTaskResumeAll();
#endif /* OS_THREAD_SAFE_NEWLIB */
}

/* __malloc_lock needs to provide recursive mutex locking */

void __malloc_lock ( struct _reent *_r )
{
#if OS_THREAD_SAFE_NEWLIB
    if (xTaskGetSchedulerState() == taskSCHEDULER_NOT_STARTED)
        return;

    vTaskSuspendAll();
#endif /* OS_THREAD_SAFE_NEWLIB */
    return;
}

/* __malloc_unlock needs to provide recursive mutex unlocking */

void __malloc_unlock ( struct _reent *_r )
{
#if OS_THREAD_SAFE_NEWLIB
    if (xTaskGetSchedulerState() == taskSCHEDULER_NOT_STARTED)
        return;

    xTaskResumeAll();
#endif /* OS_THREAD_SAFE_NEWLIB */
}

This I copied from a port I made for FreeRTOS a while ago :)

[haukepetersen]

Interesting, I will try to look into this, but probably not before January... I renamed the issue to reflect this.

[PeterKietzmann]

(I changed the release tag to 2016.03)

[haukepetersen]

thx