AW: AW: [PATCH v3 3/3] i386: Port to RTEMS

[Sebastian Huber]

On 2020-03-27 17:19, Jan.Sommer at dlr.de wrote:

> Von: Sebastian Huber [mailto:sebastian.huber at embedded-brains.de]
> Gesendet: Montag, 23. März 2020 14:20
> An: Sommer, Jan; devel at rtems.org
> Betreff: Re: AW: [PATCH v3 3/3] i386: Port to RTEMS
>>
>>> -----Ursprüngliche Nachricht-----
>>> Von: devel [mailto:devel-bounces at rtems.org] Im Auftrag von
>>> Jan.Sommer at dlr.de
>>> Gesendet: Montag, 23. März 2020 13:50
>>> An: sebastian.huber at embedded-brains.de; devel at rtems.org
>>> Betreff: AW: [PATCH v3 3/3] i386: Port to RTEMS
>>>
>>>
>>>
>>>> -----Ursprüngliche Nachricht-----
>>>> Von: Sebastian Huber [mailto:sebastian.huber at embedded-brains.de]
>>>> Gesendet: Montag, 23. März 2020 10:59
>>>> An: Sommer, Jan; devel at rtems.org
>>>> Betreff: Re: [PATCH v3 3/3] i386: Port to RTEMS
>>>>
>>>> On 23/03/2020 09:06, Jan Sommer wrote:
>>>>
>>>>> @@ -4597,7 +4599,12 @@ iflib_device_register(device_t dev, void *sc,
>>>> if_shared_ctx_t sctx, if_ctx_t *ct
>>>>>    	/* Set up cpu set.  If it fails, use the set of all CPUs. */
>>>>>    	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx-
>>>>> ifc_cpus) != 0) {
>>>>>    		device_printf(dev, "Unable to fetch CPU list\n");
>>>>> +#ifndef __rtems__
>>>>>    		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
>>>>> +#else /* __rtems__ */
>>>>> +		cpuset_t cpus = {all_cpus};
>>>>> +		CPU_COPY(&cpus, &ctx->ifc_cpus);
>>>>> +#endif /* __rtems__ */
>>>>>    	}
>>>>>    	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
>>>> What is the reason for this change?
>>>>
>>> For RTEMS the all_cpus variable has been replaced in freebsd/sys/sys/smp.h
>>> to " #define all_cpus 1U".
>>> The additional variable is there because CPU_COPY requires an address.
>>>
>>>> Could you please move the changes in non-x86 specific files to a
>>>> separate commit. Especially the <sys/callout.h> is an important header
>>>> file used across all architectures.
>> Ok, problems like this are an indication that you try to use a feature which is
>> not really supported. I would first try to understand what the purpose of the
>> ifc_cpus is. Maybe this stuff could be disabled completely.
> Ok, I investigated the issue further.
> If I understand the RTEMS specific settings in sys/smp.h correctly, rtems-libbsd always assumes to run on a single core?
Yes, most of the time. There are some exceptions, for example the UMA 
and the Epoch Based Reclamation.
>
> ifc_cpus is the map of active CPUs, i.e. for RTEMS it should be 1, and is used to determine the queue sizes.
> It is only initialized once from all_cpus, but with CPU_COPY, which requires an address as source.
> In FreeBSD all_cpus is a variable, for RTEMS it is only a define producing a compile error.
>
> As far as I see it, we could either use a temporary variable as above or maybe an int for the all_cpus in sys/smp.h instead of the define.
> However the latter would move the assignment to a c-file, thus making it harder to find. Do you have any other idea?
>
I would prefer to either support this per CPU stuff properly or disable 
it via the preprocessor. I normally remove the CPU members from the 
structs to catch all use cases. For example in struct callout:

#ifndef __rtems__
     volatile int c_cpu;            /* CPU we're scheduled on */
#endif /* __rtems__ */

This tinkering makes it more difficult to figure out what is going on in 
one year or so.