Re: [Adeos-main] Reworking ipipe timer subsystem,

[Gilles Chanteperdrix <gilles.chanteperdrix@xenomai.org>]

On 04/02/2012 05:54 PM, Jan Kiszka wrote:
> On 2012-04-02 17:39, Gilles Chanteperdrix wrote:
>> On 02/24/2012 03:40 PM, Philippe Gerum wrote:
>>> On 02/24/2012 01:28 PM, Gilles Chanteperdrix wrote:
>>>> On 02/20/2012 05:46 PM, Philippe Gerum wrote:
>>>>> On 02/19/2012 08:43 PM, Gilles Chanteperdrix wrote:
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> restarting the ipipe-core is the good opportunity to look a bit at our
>>>>>> current state and think about what we could improve. On ARM, at least,
>>>>>> the thing we could improve is the timer subsystem. A long time ago,
>>>>>> linux has switched to a system allowing to select at run-time which
>>>>>> timer to use, and we do not really benefit from this, what timer we use
>>>>>> is selected at compile time, resulting in a massive ifdefery on arm, and
>>>>>> even on x86, we have to select at compile time whether using the 8254 or
>>>>>> APIC, whereas we could decide to use whatever linux is using, even say
>>>>>> HPET, without any compilation option. That is assuming we want to move
>>>>>> the x86 timer-specific code from xenomai to I-pipe.
>>>>>>
>>>>>> The idea to reach this goal would be to add some ipipe specific members
>>>>>> to the struct clock_event_device, the way we do for the interrupt
>>>>>> controller:
>>>>>> - a CLOCK_EVT_FEAT_IPIPE would signal that the clockevent device is
>>>>>> ipipe capable, meaning that it implements the following callback
>>>>>> - ipipe_steal would be called when stealing the timer, we could decide
>>>>>> to call this callback either as part of ipipe_request_timer, or with a
>>>>>> specific ipipe_steal_timer call, currently we simply set
>>>>>> __ipipe_mach_timerstolen to 1, but it is pure luck that we never needed
>>>>>> something more complicated, besides, we may want to start a timer that
>>>>>> was not started by linux so, we would put its initialization here;
>>>>>> - ipipe_stolen would record whether the timer is stolen
>>>>>> - ipipe_min_delta_ticks, ipipe_max_delta_ticks would be used by the
>>>>>> ipipe_set_next_event callback
>>>>>> - ipipe_ack would be called to ack the timer interrupt the way we
>>>>>> currently do it currently on arm with __ipipe_mach_acktimer
>>>>>> - ipipe_set_next_event would program the next timer shot, the way it is
>>>>>> currently done in __ipipe_mach_set_dec.
>>>>>>
>>>>>> All this is pretty straightforward, but there is still a question: how
>>>>>> does ipipe_request_timer select a timer. The idea is that on platform
>>>>>> without PIC muting, it is probably more efficient to use the same timer
>>>>>> for linux and xenomai. But on platforms with PIC muting, we could want
>>>>>> to select a different timer for linux and xenomai, but how do we find
>>>>>> it, what if linux has not selected a timer because it is unusable on
>>>>>> that platform?
>>>>>
>>>>> Checking the clock device mode for CLOCK_EVT_MODE_SHUTDOWN, and falling
>>>>> back to sharing the active kernel tick device + disabling PIC muting?
>>>>
>>>> OK. Another question is: do we want to move x86 timer handling from
>>>> xenomai to ipipe. If not, we have to find a way to support the two
>>>> possible configurations. What we could do is using the timer name in
>>>> ipipe_request_tickdev: if a timer name is supplied, we keep the old
>>>> implementation, if no timer name is supplied, we use the newho
>>>> implementation which looks for the best candidate with the
>>>> CLOCK_EVT_FEAT_IPIPE bit.
>>>>
>>>>
>>>
>>> Yes, makes sense. At any rate, handling the real-time timer the way we 
>>> want for Xenomai directly from the pipeline is the only sane option. We 
>>> only have to be careful about backward compatibility of the newest core 
>>> pipelines with 2.6.x, until we stop maintaining this branch in favor of 
>>> 3.x. We may also move ipipe_request_tckdev() to the compat module fully, 
>>> removing it from the current API if that makes sense.
>>>
>>
>> For those who would like to follow, the result, a bit different from what
>> was originally planned is the interface implemented by this file:
>> http://git.xenomai.org/?p=ipipe-gch.git;a=blob;f=kernel/ipipe/timer.c;h=b9936469652d8fe998157d155fda77df81f0425a;hb=52d36aa86d5c5d463d86d384ad717f26ec96ef8c
>>
>> And ARM and x86 architectures were re-factored over this interface. As 
>> an example, the implementation of x86 timers is in this patch:
>> http://git.xenomai.org/?p=ipipe-gch.git;a=commitdiff;h=52d36aa86d5c5d463d86d384ad717f26ec96ef8c;hp=675a8ed4365eb1f23b098f913caf40e4dc792b80
>>
>> 8254, local APIC and HPET in legacy mode were tested, even selected 
>> at run-time. Only per-cpu HPET remains to be tested (the hardware 
>> I have does not support it).
> 
> FWI, QEMU (w/ or w/o KVM) can emulated enough HPET timers, also with MSI
> support, but that was broken in I-pipe last time I checked. Use -global
> hpet.timers=4 (or more for more CPUs) and -global hpet.msi=on.

Ah thanks. How was it broken? MSI were broken? As far as I understood
MSI support for HPET are required for linux to select per-cpu HPET.

-- 
					    Gilles.