[Xenomai] FreeRTOS skin #2
ma30002000 at yahoo.de
Mon Jun 23 12:31:50 CEST 2014
----- Original Message -----
> From: Philippe Gerum <rpm at xenomai.org>
> To: Matthias Schneider <ma30002000 at yahoo.de>; "xenomai at xenomai.org" <xenomai at xenomai.org>
> Sent: Friday, June 20, 2014 9:55 AM
> Subject: Re: [Xenomai] FreeRTOS skin #2
> On 06/02/2014 08:46 PM, Matthias Schneider wrote:
>> Please find enclosed a new version of the FreeRTOS skin. It adapts to the
>> latest changes in forge/next and also removes the previously included
>> workarounds - thanks to the changes done based on the first review round
>> they were no longer necessary.
> Thanks. Here we go:
> * vTaskStartScheduler()
> Dropping the task_list lock then yielding is superfluous, the caller
> will wait on the condvar for the scheduler state to switch to SUSPENDING
> anyway, implicitly releasing this lock before going to sleep.
> * vTaskDelay()
> Cobalt wraps sched_yield(), so you likely want to call
> __RT(sched_yield()) to pick the proper implementation depending on the
> underlying real-time core.
> * __xTaskGenericCreate()
> - It looks like the mutex attribute initialized locally is a left over
> from a previous implementation.
> - Drawing unique identifier labels should be done using a name
> generator, e.g.
> static DEFINE_NAME_GENERATOR(task_namegen, struct freertos_task, name);
> generate_name(task->name, name, &task_namegen);
> The open-coded method also used by other emulators is racy.
Note that some freertos specific code remains to make duplicated task names
> * vPortEnter/ExitCritical()
> - wouldn't these calls better emulated by threadobj_lock_sched()
> assuming they stand for a global serialization mechanism?
Unfortunately, this seems cause some test cases in the demo to fail.
Since these tests are rather time consuming to investigate, I would like
to leave this item open until all other issues have been addressed. We
had also talked about other options concerning this subject as well
(removing the function or leaving it empty), so I still need some
time to think about it.
> * Task registry support
> The code looks outdated, using fsobstacks is recommended. The registry
> helpers in lib/alchemy/alarm.c illustrate a basic usage.
> * vTaskSuspendAll()
> Running check_processors() there may switch the caller to linux mode
> at each invocation, due to calling sysconf/sched_getaffinity indirectly.
> I would rather assume that the CPU affinity is not allowed to change
> once started, and determine whether a task is bound to more than a
> single CPU in the trampoline code. Then vTaskSuspendAll() may do a
> simple flag-based test for checking the consistency, still allowing
> different freertos tasks to be pinned on distinct CPUs though. At any
> rate, this would be an acceptable restriction, since freertos does not
> seem to be SMP-capable in its reference implementation anyway.
Ok, reading the affinity information once and caching it makes sense to
avoid the switches to secondary mode. However, vTaskSuspendAll() / vTaskResumeAll() only works as expected when all tasks are pinned to a
single cpu, storing this information on a by-task basis will not be sufficient.
Now if there is one place to retrieve this information, which would it be?
I can think of freertos_init() / freertos_scheduler_init() or
vTaskStartScheduler(). However, freertos_init() would probably not be
able to allow setting process cpu affinity programatically at the
beginning of main() if I am not mistaken.
> Next time, I'll start reviewing other parts of the emulator.
Great, thanks in advance for all the input..
More information about the Xenomai