Jan Kiszka wrote:
> Philippe Gerum wrote:
>> Jan Kiszka wrote:
>>> Philippe Gerum wrote:
>>>> Jan Kiszka wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Philippe Gerum wrote:
>>>>>>> Jan Kiszka wrote:
>>>>>>>> Gilles Chanteperdrix wrote:
>>>>>>>>> Jan Kiszka wrote:
>>>>>>>>>> This series fixes the issues around rt_task_inquire I posted 
>>>>>>>>>> yesterday.
>>>>>>>>>> Additionally, it introduces an analogous services pthread_inquire_np 
>>>>>>>>>> for
>>>>>>>>>> the POSIX skin. That allows, among other things, to implement test 
>>>>>>>>>> cases
>>>>>>>>>> for the upcoming fast xnsynch/mutex patches.
>>>>>>>>> Ok. Since this applies only for debugging purpose, and displaying
>>>>>>>>> whether a task is in primary mode may be use badly by users, maybe we
>>>>>>>>> should make this service a shadow syscall, and not export any 
>>>>>>>>> interface
>>>>>>>>> to use it. This would further avoid duplication between the native and
>>>>>>>>> posix skins.
>>>>>>>> Debugging is not the holy, exclusive business of Xenomai hackers.
>>>>>>>> The inquire services are useful in libraries as well, when you want to
>>>>>>>> check if the caller complies to the call convention ("don't use in
>>>>>>>> primary mode", "caller's priority must not exceed X" or whatever).
>>>>>>>> That said, I'm open for unifying the code, maybe introducing some
>>>>>>>> xnthread_inquire.
>>>>>>> That would be much better than publishing an open interface to fiddle 
>>>>>>> even more
>>>>>>> with thread modes via rt_task_set_mode(). I would definitely merge that.
>>>>>> Code refactoring is no problem, will work that out. I just want to keep
>>>>>> the user interface.
>>>>> Looking into this, I come to the conclusion that xnthread_inquire is
>>>>> only then a gain if both rt_task_inquire and pthread_inquire_np use the
>>>>> same data structure layout. And that means that both need to use the
>>>>> same time encodings, not struct timespec vs. RTIME like it is now. That
>>>>> would not be beautiful, but feasible (e.g. picking __u64 as type,
>>>>> passing nanoseconds). Still, it does not yet convince me.
>>>> The idea behind xnthread_inquire() is not about replacing rt_task_inquire()
>>>> under the hood, but rather to get back only fundamental values, such as the
>>>> current thread status, in order to determine whether XNRELAX is set or not 
>>>> for
>>>> instance.
>>>> XENOMAI_SYSCALL1(__xn_sys_inquire) => 
>>>> xnthread_inquire(xnpod_current_thread())
>>> OK, slowly getting to your core. Assume we had __xn_sys_inquire, what
>>> should it return, and what should we mask from rt_task_inquire, and do
>>> we still want pthread_inquire_np.
>> __xn_sys_inquire would return both the informational and status bitmasks, for
>> debug and sanity check purposes.
>>> But I really dislike this approach of artificially
>>> complicating/crippling interfaces just to hide XNRELAX (I can't imagine
>>> you have problems with the other information rt_task_inquire returns).
>> It is not about crippling that interface at all, it is about not adding the
>> XNRELAX bit - since it is NOT there ATM - to prevent further misuse of the
> Look at the code, it _is_ practically there, it _is_ usable, but it is
> inverted (T_PRIMARY==XNRELAX). That's why I insist on the fact that the
> current rt_task_inquire implementation is broken when taking its
> documentation as a reference.
>> thread mode, which is a very real problem in lots of application code now.
>> Look, even Hannes who very well knows what co-kernel stuff is all about
>> misinterpreted the Xenomai API in that area:
>> http://www.captain.at/review-rtai-versus-xenomai.php
> Not directly his fault, he just copied our old, incorrect serial test
> code (/me failed to review it before it saw the light of the internet).
> But I told him ages ago to update his page.

The point still stands, whoever made the mistake, that mistake was due to a
misunderstanding of the API. This is the gist of the matter: let's remove the
source of such issue.

>> The reason is likely because people tend to think that if T_PRIMARY is
>> explicitly provided, that means that the real-time core does not handle the
>> issue implicitly, and then conclude that they ought to use 
>> rt_task_set_mode() to
>> do the job themselves. That's wrong, that's badly wrong.
> Again, don't shoot the messenger, let's discuss how to overcome the
> explicit mode switch interface! You are looking at the wrong spot, IMHO.

I'm not shooting the messenger, I'm just telling the messenger that his message
is currently wrong.

Please look at the code again:

- __rt_task_inquire() does not remap XNRELAX to some still undefined T_PRIMARY
bit because I never wanted to do so.

- The comments attached to the status bits in question do make pretty clear that
I did not want those statuses to exist actually, there were just meant to be
placeholders for rt_task_set_mode() only. T_JOINABLE is arguably a good
candidate for conversion to an actual status , that is not disputed.

#define T_PRIMARY  0x00000200   /* Recycle internal bits status which */
#define T_JOINABLE 0x00000400   /* won't be passed to the nucleus.  */

- __rt_task_set_mode() is currently trying to set T_PRIMARY within the old mode
mask to be returned to the user to be consistent with the input parameters that
may set it, but that fails, and beyond that, it's plain silly (I wrote this
code, so I can tell you it is, right?). It is silly because __rt_task_set_mode()
has to be tagged as __xn_exec_primary; so, what is the point?

What we need is:

- to prevent the thread mode from being explicitly manipulated from the public
interface. So we need to get rid of T_PRIMARY in rt_task_set_mode(). At the same
time, we still need to provide a way for low-level code to control that mode,
That service already exists, i.e. XENOMAI_SYSCALL(__xn_sys_migrate).

- a way to get the current thread status for debugging/sanity checks/logging
etc; but since we want to move the primary mode bit away from the public
interface of rt_task_set_mode(), and that is XENOMAI_SYSCALL(__xn_sys_inquire).
There would be no point in returning T_PRIMARY back explicitly via
rt_task_inquire(). It would even be counter-productive. I don't care if we don't
mask out XNRELAX and other unwanted bits from RT_TASK_INFO::status; the point is
that we should not mention those internal bits in the rt_task_inquire()

> BTW, RTnet no longer requires explicit mode switch. And for Xenomai 2.5,
> I will remove rtdm_device.open_rt/socket_rt as well as
> rtdm_operations.close_rt, ie. a whole bunch of handlers that motivated
> most of the explicit switches from my POV. So I would be fine with
> deprecating that switch service for 2.5 and maybe remove it with 2.6 (or
> whatever comes next).
>> Aside of this, using rt_task_inquire() in instrumentation/debug/sanity 
>> checking
>> code would be overkill most of the time. If the purpose is to implement
>> constructs like:
>>      if (!task_in_primary_mode()) {
>>              blah;
>>      }
>> Then, you likely don't want the overhead of copying useless things like the 
>> task
>> name, fetching the outstanding timeout values, or any other values we may 
>> want
>> to add in the future to RT_TASK_INFO (e.g. we did not use to send back any 
>> kind
>> of statistical information via that structure in 2.3.x, but we started to do 
>> so
>> with 2.4.x).
> That overhead is negligible compared to the anyway required syscall. If
> we needed to optimize debugging stuff (I doubt so), we would have to
> push that data (also the thread prio etc.) continuously to user space.

It is not about debugging stuff only, you said you wanted to add sanity checks
as well, like I just sketched above and will redo below:

        if (!task_in_primary_mode())
                return -CANTDOTHAT;

In that case, you certainly don't want to pay for such overhead. Therefore, what
you need is XENOMAI_SYSCALL(__xn_sys_inquire).

> Jan


Xenomai-core mailing list

Reply via email to