[Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Hi Gilles,

how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
default settings? At least on x86-64, the default 32K is not enough.
Unless we talk about GB ;), maybe it makes sense to adjust the default
size accordingly.

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Gilles Chanteperdrix]

Jan Kiszka wrote:
> Hi Gilles,
> 
> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
> default settings? At least on x86-64, the default 32K is not enough.
> Unless we talk about GB ;), maybe it makes sense to adjust the default
> size accordingly.

It depends on the arguments you pass to switchtest.

-- 
                                                 Gilles.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Hi Gilles,
>>
>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>> default settings? At least on x86-64, the default 32K is not enough.
>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>> size accordingly.
> 
> It depends on the arguments you pass to switchtest.

None, ie. the default settings.

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Gilles Chanteperdrix]

Jan Kiszka wrote:
> Gilles Chanteperdrix wrote:
>> Jan Kiszka wrote:
>>> Hi Gilles,
>>>
>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>> default settings? At least on x86-64, the default 32K is not enough.
>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>> size accordingly.
>> It depends on the arguments you pass to switchtest.
> 
> None, ie. the default settings.

Then 6 kernel-space tasks are created. Since switchtest is not the
average use of Xenomai, and is supposed to be used only for validating
the context/FPU switches, I would not change XENO_OPT_SYS_STACKPOOL_SZ,
maybe only add a clear message when creation of kernel-space tasks fails.

-- 
                                                 Gilles.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Gilles Chanteperdrix wrote:
>>> Jan Kiszka wrote:
>>>> Hi Gilles,
>>>>
>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>> size accordingly.
>>> It depends on the arguments you pass to switchtest.
>> None, ie. the default settings.
> 
> Then 6 kernel-space tasks are created. Since switchtest is not the

6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
with >= 40k.

> average use of Xenomai, and is supposed to be used only for validating
> the context/FPU switches, I would not change XENO_OPT_SYS_STACKPOOL_SZ,
> maybe only add a clear message when creation of kernel-space tasks fails.

Well, isn't the switchtest part of xeno-test? Would be unfortunate to
require a kernel rebuild just for running the standard tests. I think we
don't loose much by slightly increasing the stack to, say, 64K on 32 bit
and 128K on 64 bit archs.

BTW, XNARCH_THREAD_STACKSZ for x86-64 equals the x86-32 bit value (4K)
while I don't think the have they same demands (nor the same resource
restrictions).

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Gilles Chanteperdrix]

Jan Kiszka wrote:
> Gilles Chanteperdrix wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Hi Gilles,
>>>>>
>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>> size accordingly.
>>>> It depends on the arguments you pass to switchtest.
>>> None, ie. the default settings.
>> Then 6 kernel-space tasks are created. Since switchtest is not the
> 
> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
> with >= 40k.
> 
>> average use of Xenomai, and is supposed to be used only for validating
>> the context/FPU switches, I would not change XENO_OPT_SYS_STACKPOOL_SZ,
>> maybe only add a clear message when creation of kernel-space tasks fails.
> 
> Well, isn't the switchtest part of xeno-test? Would be unfortunate to
> require a kernel rebuild just for running the standard tests. I think we
> don't loose much by slightly increasing the stack to, say, 64K on 32 bit
> and 128K on 64 bit archs.

The kernel rebuild is already required since the switchtest driver is
not compiled by default.

-- 
                                                 Gilles.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Gilles Chanteperdrix wrote:
>>> Jan Kiszka wrote:
>>>> Gilles Chanteperdrix wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Hi Gilles,
>>>>>>
>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>> size accordingly.
>>>>> It depends on the arguments you pass to switchtest.
>>>> None, ie. the default settings.
>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>> with >= 40k.
>>
>>> average use of Xenomai, and is supposed to be used only for validating
>>> the context/FPU switches, I would not change XENO_OPT_SYS_STACKPOOL_SZ,
>>> maybe only add a clear message when creation of kernel-space tasks fails.
>> Well, isn't the switchtest part of xeno-test? Would be unfortunate to
>> require a kernel rebuild just for running the standard tests. I think we
>> don't loose much by slightly increasing the stack to, say, 64K on 32 bit
>> and 128K on 64 bit archs.
> 
> The kernel rebuild is already required since the switchtest driver is
> not compiled by default.

Even if it were default off (which it isn't), requiring a second rebuild
due to this implicit dependency would remain, well, unfortunate.

I will propose a Kconfig adjustment later.

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Philippe Gerum]

Jan Kiszka wrote:
> Gilles Chanteperdrix wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Hi Gilles,
>>>>>
>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>> size accordingly.
>>>> It depends on the arguments you pass to switchtest.
>>> None, ie. the default settings.
>> Then 6 kernel-space tasks are created. Since switchtest is not the
> 
> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
> with >= 40k.
>

Actually, it is not really an overhead, but rather the fact that it wants at
least two initially free pages per heap.

>> average use of Xenomai, and is supposed to be used only for validating
>> the context/FPU switches, I would not change XENO_OPT_SYS_STACKPOOL_SZ,
>> maybe only add a clear message when creation of kernel-space tasks fails.
> 
> Well, isn't the switchtest part of xeno-test? Would be unfortunate to
> require a kernel rebuild just for running the standard tests. I think we
> don't loose much by slightly increasing the stack to, say, 64K on 32 bit
> and 128K on 64 bit archs.
> 
> BTW, XNARCH_THREAD_STACKSZ for x86-64 equals the x86-32 bit value (4K)
> while I don't think the have they same demands (nor the same resource
> restrictions).
> 
> Jan
> 


-- 
Philippe.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Philippe Gerum wrote:
> Jan Kiszka wrote:
>> Gilles Chanteperdrix wrote:
>>> Jan Kiszka wrote:
>>>> Gilles Chanteperdrix wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Hi Gilles,
>>>>>>
>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>> size accordingly.
>>>>> It depends on the arguments you pass to switchtest.
>>>> None, ie. the default settings.
>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>> with >= 40k.
>>
> 
> Actually, it is not really an overhead, but rather the fact that it wants at
> least two initially free pages per heap.

That would make 22K. The problem is that the management overhead is
rounded up to another full page, requiring a 8K allocation per 4K
request. Reminds me of TLSF - if I only had the time... :)

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Gilles Chanteperdrix]

Jan Kiszka wrote:
> Philippe Gerum wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Gilles Chanteperdrix wrote:
>>>>>> Jan Kiszka wrote:
>>>>>>> Hi Gilles,
>>>>>>>
>>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>>> size accordingly.
>>>>>> It depends on the arguments you pass to switchtest.
>>>>> None, ie. the default settings.
>>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>>> with >= 40k.
>>>
>> Actually, it is not really an overhead, but rather the fact that it wants at
>> least two initially free pages per heap.
> 
> That would make 22K. The problem is that the management overhead is
> rounded up to another full page, requiring a 8K allocation per 4K
> request. Reminds me of TLSF - if I only had the time... :)

When allocating 4K, only 4K are allocated.

I think the problem could rather be that you are running switchtest on a
multi-cpu system, in which same the number of cpu tasks created is 6 by cpu.

-- 
                                                 Gilles.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Philippe Gerum wrote:
>>> Jan Kiszka wrote:
>>>> Gilles Chanteperdrix wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Gilles Chanteperdrix wrote:
>>>>>>> Jan Kiszka wrote:
>>>>>>>> Hi Gilles,
>>>>>>>>
>>>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>>>> size accordingly.
>>>>>>> It depends on the arguments you pass to switchtest.
>>>>>> None, ie. the default settings.
>>>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>>>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>>>> with >= 40k.
>>>>
>>> Actually, it is not really an overhead, but rather the fact that it wants at
>>> least two initially free pages per heap.
>> That would make 22K. The problem is that the management overhead is
>> rounded up to another full page, requiring a 8K allocation per 4K
>> request. Reminds me of TLSF - if I only had the time... :)
> 
> When allocating 4K, only 4K are allocated.
> 
> I think the problem could rather be that you are running switchtest on a
> multi-cpu system, in which same the number of cpu tasks created is 6 by cpu.

That was in fact the problem here. OK, so we should just consider
adjusting the pool default for 8K-stack archs (PowerPC64, Blackfin, and
maybe also x86-64).

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Philippe Gerum]

Jan Kiszka wrote:
> Philippe Gerum wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Gilles Chanteperdrix wrote:
>>>>>> Jan Kiszka wrote:
>>>>>>> Hi Gilles,
>>>>>>>
>>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>>> size accordingly.
>>>>>> It depends on the arguments you pass to switchtest.
>>>>> None, ie. the default settings.
>>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>>> with >= 40k.
>>>
>> Actually, it is not really an overhead, but rather the fact that it wants at
>> least two initially free pages per heap.
> 
> That would make 22K. The problem is that the management overhead is
> rounded up to another full page, requiring a 8K allocation per 4K
> request.

Nope. An individual 4k request is going to pull 8 x 512 bytes pages from the
stack pool, not more.

 Reminds me of TLSF - if I only had the time... :)
> 

It looks like working properly for -solo.

> Jan
> 


-- 
Philippe.

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Jan Kiszka]

Philippe Gerum wrote:
> Jan Kiszka wrote:
>> Philippe Gerum wrote:
>>> Jan Kiszka wrote:
>>>> Gilles Chanteperdrix wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Gilles Chanteperdrix wrote:
>>>>>>> Jan Kiszka wrote:
>>>>>>>> Hi Gilles,
>>>>>>>>
>>>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>>>> size accordingly.
>>>>>>> It depends on the arguments you pass to switchtest.
>>>>>> None, ie. the default settings.
>>>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>>>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>>>> with >= 40k.
>>>>
>>> Actually, it is not really an overhead, but rather the fact that it wants at
>>> least two initially free pages per heap.
>> That would make 22K. The problem is that the management overhead is
>> rounded up to another full page, requiring a 8K allocation per 4K
>> request.
> 
> Nope. An individual 4k request is going to pull 8 x 512 bytes pages from the
> stack pool, not more.

Yeah, I see. The only "overhead" here was already paid via the pagemap.

> 
>  Reminds me of TLSF - if I only had the time... :)
> 
> It looks like working properly for -solo.

I think to remember your concerns were more about missing fragmentation,
size overhead and performance comparisons. A working version for
standard Xenomai was already available at that time (maybe not for all
archs, but that is surely quickly fixed).

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux

Re: [Xenomai-core] XENO_OPT_SYS_STACKPOOLSZ vs. switchtest

[Philippe Gerum]

Jan Kiszka wrote:
> Philippe Gerum wrote:
>> Jan Kiszka wrote:
>>> Philippe Gerum wrote:
>>>> Jan Kiszka wrote:
>>>>> Gilles Chanteperdrix wrote:
>>>>>> Jan Kiszka wrote:
>>>>>>> Gilles Chanteperdrix wrote:
>>>>>>>> Jan Kiszka wrote:
>>>>>>>>> Hi Gilles,
>>>>>>>>>
>>>>>>>>> how much XENO_OPT_SYS_STACKPOOLSZ do I need to run switchtest for
>>>>>>>>> default settings? At least on x86-64, the default 32K is not enough.
>>>>>>>>> Unless we talk about GB ;), maybe it makes sense to adjust the default
>>>>>>>>> size accordingly.
>>>>>>>> It depends on the arguments you pass to switchtest.
>>>>>>> None, ie. the default settings.
>>>>>> Then 6 kernel-space tasks are created. Since switchtest is not the
>>>>> 6*4 is 20k... Ah, the well-known allocator overhead, I guess. Will try
>>>>> with >= 40k.
>>>>>
>>>> Actually, it is not really an overhead, but rather the fact that it wants at
>>>> least two initially free pages per heap.
>>> That would make 22K. The problem is that the management overhead is
>>> rounded up to another full page, requiring a 8K allocation per 4K
>>> request.
>> Nope. An individual 4k request is going to pull 8 x 512 bytes pages from the
>> stack pool, not more.
> 
> Yeah, I see. The only "overhead" here was already paid via the pagemap.
> 
>>  Reminds me of TLSF - if I only had the time... :)
>>
>> It looks like working properly for -solo.
> 
> I think to remember your concerns were more about missing fragmentation,
> size overhead and performance comparisons. A working version for
> standard Xenomai was already available at that time (maybe not for all
> archs, but that is surely quickly fixed).
>

Locking and init fixes went to TLSF 2.4.4 in the recent months, so we would have
needed those anyway. It's quite late to merge TLSF in Xenomai 2.5, but I would
have no objection to make it the base allocator of the 3.x series. We would have
to extend it with:

- support for a few callouts, such as the validity checking function used in
xnhead_test_and_free(),
- shared memory export (this is quite allocator agnostic in fact)
- detailed error codes in xnheap_test_and_free/xnheap_free

> Jan
> 


-- 
Philippe.