[Xenomai-help] Masking / Disabling the local timer interrupt

[Xenomai-help] Masking / Disabling the local timer interrupt

[Matthew Fornero]

[-- Attachment #1: Type: text/plain, Size: 1126 bytes --]

Hello,

I have an application on x86 that uses an external event (either an
interrupt or a DMA write to a specific memory location) to trigger code
execution, and I believe the local timer interrupt is contributing a
significant amount of jitter. This is a bit of a unique case where the only
code that should be executing on that core is triggered by the event, so
there is no need for anything timer based.

I'm using a dual core processor with a quiet system, and have placed all of
the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
interrupt increment on CPU 1. When using the DMA write to trigger execution,
placing "sti"/"cli" around the polling routine improves the jitter
substantially.

Based on the above, is there some way to do one of the following:

1). Make a user space task higher priority than the timer interrupt

or

2). A proper way to disable the local timer (vs. disabling all interrupts)
on specific cores, as these cores would only need to run event based
routines.

Best Regards,

Matthew Fornero

[-- Attachment #2: Type: text/html, Size: 1216 bytes --]

Re: [Xenomai-help] Masking / Disabling the local timer interrupt

[Gilles Chanteperdrix]

Matthew Fornero wrote:
> Hello,
> 
> I have an application on x86 that uses an external event (either an
> interrupt or a DMA write to a specific memory location) to trigger code
> execution, and I believe the local timer interrupt is contributing a
> significant amount of jitter. This is a bit of a unique case where the only
> code that should be executing on that core is triggered by the event, so
> there is no need for anything timer based.
> 
> I'm using a dual core processor with a quiet system, and have placed all of
> the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
> Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
> interrupt increment on CPU 1. When using the DMA write to trigger execution,
> placing "sti"/"cli" around the polling routine improves the jitter
> substantially.
> 
> Based on the above, is there some way to do one of the following:
> 
> 1). Make a user space task higher priority than the timer interrupt
> 
> or
> 
> 2). A proper way to disable the local timer (vs. disabling all interrupts)
> on specific cores, as these cores would only need to run event based
> routines.

You can disable Xenomai timer interrupt by calling xnpod_stop_timer in
kernel-space. However, what you see in /proc/interrupt is not Xenomai
timer it is Linux', and it causes a bit of jitter (the time for the
I-pipe to mask, ack and return from interrupt). Disabling it would
require butchery in Linux code, but maybe simply masking the timer
interrupt could work.

-- 
					    Gilles.

Re: [Xenomai-help] Masking / Disabling the local timer interrupt

[Gilles Chanteperdrix]

Matthew Fornero wrote:
> On Tue, Nov 4, 2008 at 2:21 PM, Gilles Chanteperdrix
> <gilles.chanteperdrix@xenomai.org> wrote:
>> Matthew Fornero wrote:
>>> Hello,
>>>
>>> I have an application on x86 that uses an external event (either an
>>> interrupt or a DMA write to a specific memory location) to trigger code
>>> execution, and I believe the local timer interrupt is contributing a
>>> significant amount of jitter. This is a bit of a unique case where the only
>>> code that should be executing on that core is triggered by the event, so
>>> there is no need for anything timer based.
>>>
>>> I'm using a dual core processor with a quiet system, and have placed all of
>>> the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
>>> Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
>>> interrupt increment on CPU 1. When using the DMA write to trigger execution,
>>> placing "sti"/"cli" around the polling routine improves the jitter
>>> substantially.
>>>
>>> Based on the above, is there some way to do one of the following:
>>>
>>> 1). Make a user space task higher priority than the timer interrupt
>>>
>>> or
>>>
>>> 2). A proper way to disable the local timer (vs. disabling all interrupts)
>>> on specific cores, as these cores would only need to run event based
>>> routines.
>> You can disable Xenomai timer interrupt by calling xnpod_stop_timer in
>> kernel-space. However, what you see in /proc/interrupt is not Xenomai
>> timer it is Linux', and it causes a bit of jitter (the time for the
>> I-pipe to mask, ack and return from interrupt). Disabling it would
>> require butchery in Linux code, but maybe simply masking the timer
>> interrupt could work.
>>
>> --
>>                                            Gilles.
> 
> I could not find any reference to the function xnpod_stop_timer in the
> xenomai code (searched through the cross-reference).
> I tried using xnarch_stop_timer(cpu), but this seems to disable the
> timers across all CPUs, rather than just the one passed.
> What I'd really like to do is simply mask the interrupt temporarily
> while my critical section is executing (the DMA polling routine)-- is
> there any way to do this? I can go into the local APIC and change the
> mask bits, but using some sort of API seems like it'd be cleaner.

xnpod_stop_timer no longer exists ! The function to disable irqs is
rt_intr_disable if you use the native skin.

-- 
                                                 Gilles.

Re: [Xenomai-help] Masking / Disabling the local timer interrupt

[Matthew Fornero]

On Thu, Nov 6, 2008 at 8:51 AM, Gilles Chanteperdrix
<gilles.chanteperdrix@xenomai.org> wrote:
> Matthew Fornero wrote:
>> On Tue, Nov 4, 2008 at 2:21 PM, Gilles Chanteperdrix
>> <gilles.chanteperdrix@xenomai.org> wrote:
>>> Matthew Fornero wrote:
>>>> Hello,
>>>>
>>>> I have an application on x86 that uses an external event (either an
>>>> interrupt or a DMA write to a specific memory location) to trigger code
>>>> execution, and I believe the local timer interrupt is contributing a
>>>> significant amount of jitter. This is a bit of a unique case where the only
>>>> code that should be executing on that core is triggered by the event, so
>>>> there is no need for anything timer based.
>>>>
>>>> I'm using a dual core processor with a quiet system, and have placed all of
>>>> the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
>>>> Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
>>>> interrupt increment on CPU 1. When using the DMA write to trigger execution,
>>>> placing "sti"/"cli" around the polling routine improves the jitter
>>>> substantially.
>>>>
>>>> Based on the above, is there some way to do one of the following:
>>>>
>>>> 1). Make a user space task higher priority than the timer interrupt
>>>>
>>>> or
>>>>
>>>> 2). A proper way to disable the local timer (vs. disabling all interrupts)
>>>> on specific cores, as these cores would only need to run event based
>>>> routines.
>>> You can disable Xenomai timer interrupt by calling xnpod_stop_timer in
>>> kernel-space. However, what you see in /proc/interrupt is not Xenomai
>>> timer it is Linux', and it causes a bit of jitter (the time for the
>>> I-pipe to mask, ack and return from interrupt). Disabling it would
>>> require butchery in Linux code, but maybe simply masking the timer
>>> interrupt could work.
>>>
>>> --
>>>                                            Gilles.
>>
>> I could not find any reference to the function xnpod_stop_timer in the
>> xenomai code (searched through the cross-reference).
>> I tried using xnarch_stop_timer(cpu), but this seems to disable the
>> timers across all CPUs, rather than just the one passed.
>> What I'd really like to do is simply mask the interrupt temporarily
>> while my critical section is executing (the DMA polling routine)-- is
>> there any way to do this? I can go into the local APIC and change the
>> mask bits, but using some sort of API seems like it'd be cleaner.
>
> xnpod_stop_timer no longer exists ! The function to disable irqs is
> rt_intr_disable if you use the native skin.
>
> --
>                                                 Gilles.
>

On Thu, Nov 6, 2008 at 8:51 AM, Gilles Chanteperdrix
- Show quoted text -
<gilles.chanteperdrix@xenomai.org> wrote:
> Matthew Fornero wrote:
>> On Tue, Nov 4, 2008 at 2:21 PM, Gilles Chanteperdrix
>> <gilles.chanteperdrix@xenomai.org> wrote:
>>> Matthew Fornero wrote:
>>>> Hello,
>>>>
>>>> I have an application on x86 that uses an external event (either an
>>>> interrupt or a DMA write to a specific memory location) to trigger code
>>>> execution, and I believe the local timer interrupt is contributing a
>>>> significant amount of jitter. This is a bit of a unique case where the only
>>>> code that should be executing on that core is triggered by the event, so
>>>> there is no need for anything timer based.
>>>>
>>>> I'm using a dual core processor with a quiet system, and have placed all of
>>>> the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
>>>> Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
>>>> interrupt increment on CPU 1. When using the DMA write to trigger execution,
>>>> placing "sti"/"cli" around the polling routine improves the jitter
>>>> substantially.
>>>>
>>>> Based on the above, is there some way to do one of the following:
>>>>
>>>> 1). Make a user space task higher priority than the timer interrupt
>>>>
>>>> or
>>>>
>>>> 2). A proper way to disable the local timer (vs. disabling all interrupts)
>>>> on specific cores, as these cores would only need to run event based
>>>> routines.
>>> You can disable Xenomai timer interrupt by calling xnpod_stop_timer in
>>> kernel-space. However, what you see in /proc/interrupt is not Xenomai
>>> timer it is Linux', and it causes a bit of jitter (the time for the
>>> I-pipe to mask, ack and return from interrupt). Disabling it would
>>> require butchery in Linux code, but maybe simply masking the timer
>>> interrupt could work.
>>>
>>> --
>>>                                            Gilles.
>>
>> I could not find any reference to the function xnpod_stop_timer in the
>> xenomai code (searched through the cross-reference).
>> I tried using xnarch_stop_timer(cpu), but this seems to disable the
>> timers across all CPUs, rather than just the one passed.
>> What I'd really like to do is simply mask the interrupt temporarily
>> while my critical section is executing (the DMA polling routine)-- is
>> there any way to do this? I can go into the local APIC and change the
>> mask bits, but using some sort of API seems like it'd be cleaner.
>
> xnpod_stop_timer no longer exists ! The function to disable irqs is
> rt_intr_disable if you use the native skin.
>
> --
>                                                 Gilles.
>

A couple of follow ups:

1). I have not been able to find the RT_INTR pointer associated with
the local APIC timer interrupt-- is there some documentation on this?
Basically, I would need something to pass to rt_intr_disable

2). rt_intr_disable looks like it will disable the interrupt across
all CPUs. Is there any way to disable the Xenomai local APIC timer
handler on just a single CPU?

-- Matthew Fornero

Re: [Xenomai-help] Masking / Disabling the local timer interrupt

[Gilles Chanteperdrix]

Matthew Fornero wrote:
> A couple of follow ups:
> 
> 1). I have not been able to find the RT_INTR pointer associated with
> the local APIC timer interrupt-- is there some documentation on this?
> Basically, I would need something to pass to rt_intr_disable
> 
> 2). rt_intr_disable looks like it will disable the interrupt across
> all CPUs. Is there any way to disable the Xenomai local APIC timer
> handler on just a single CPU?

You do not need to send your mails twice, we receive your mails.

-- 
                                                 Gilles.

Re: [Xenomai-help] Masking / Disabling the local timer interrupt

[Gilles Chanteperdrix]

Matthew Fornero wrote:
> On Thu, Nov 6, 2008 at 8:51 AM, Gilles Chanteperdrix
> <gilles.chanteperdrix@xenomai.org> wrote:
>> Matthew Fornero wrote:
>>> On Tue, Nov 4, 2008 at 2:21 PM, Gilles Chanteperdrix
>>> <gilles.chanteperdrix@xenomai.org> wrote:
>>>> Matthew Fornero wrote:
>>>>> Hello,
>>>>>
>>>>> I have an application on x86 that uses an external event (either an
>>>>> interrupt or a DMA write to a specific memory location) to trigger code
>>>>> execution, and I believe the local timer interrupt is contributing a
>>>>> significant amount of jitter. This is a bit of a unique case where the only
>>>>> code that should be executing on that core is triggered by the event, so
>>>>> there is no need for anything timer based.
>>>>>
>>>>> I'm using a dual core processor with a quiet system, and have placed all of
>>>>> the "moveable" interrupts on CPU 0 (0-NR_IRQs using ipipe_set_irq_affinity).
>>>>> Looking at /proc/xenomai/irq and /proc/interrupts shows only the local timer
>>>>> interrupt increment on CPU 1. When using the DMA write to trigger execution,
>>>>> placing "sti"/"cli" around the polling routine improves the jitter
>>>>> substantially.
>>>>>
>>>>> Based on the above, is there some way to do one of the following:
>>>>>
>>>>> 1). Make a user space task higher priority than the timer interrupt
>>>>>
>>>>> or
>>>>>
>>>>> 2). A proper way to disable the local timer (vs. disabling all interrupts)
>>>>> on specific cores, as these cores would only need to run event based
>>>>> routines.
>>>> You can disable Xenomai timer interrupt by calling xnpod_stop_timer in
>>>> kernel-space. However, what you see in /proc/interrupt is not Xenomai
>>>> timer it is Linux', and it causes a bit of jitter (the time for the
>>>> I-pipe to mask, ack and return from interrupt). Disabling it would
>>>> require butchery in Linux code, but maybe simply masking the timer
>>>> interrupt could work.
>>>>
>>>> --
>>>>                                            Gilles.
>>> I could not find any reference to the function xnpod_stop_timer in the
>>> xenomai code (searched through the cross-reference).
>>> I tried using xnarch_stop_timer(cpu), but this seems to disable the
>>> timers across all CPUs, rather than just the one passed.
>>> What I'd really like to do is simply mask the interrupt temporarily
>>> while my critical section is executing (the DMA polling routine)-- is
>>> there any way to do this? I can go into the local APIC and change the
>>> mask bits, but using some sort of API seems like it'd be cleaner.
>> xnpod_stop_timer no longer exists ! The function to disable irqs is
>> rt_intr_disable if you use the native skin.
>>
>> --
>>                                                 Gilles.
>>
> 
> A couple of follow ups:
> 
> 1). I have not been able to find the RT_INTR pointer associated with
> the local APIC timer interrupt-- is there some documentation on this?
> Basically, I would need something to pass to rt_intr_disable

Right, I was going to suggest to use xnintr_disable, but it will not
work either since there is no xnintr object associated with the local
clock. You have to use xnarch_disable_irq and pass the irq number of the
local APIC timer interrupt.

> 
> 2). rt_intr_disable looks like it will disable the interrupt across
> all CPUs. Is there any way to disable the Xenomai local APIC timer
> handler on just a single CPU?

No, rt_intr_disable, xnintr_disable, xnarch_disable_irq only disable an
interrupt on the interrupt controller attached to the current cpu.

-- 
                                                 Gilles.