[ECOS] Thread suspend when scheduler is locked.

[ECOS] Thread suspend when scheduler is locked.

[Sergei Organov]

Hello,

In the following sequence:

  cyg_scheduler_lock();
  cyg_thread_suspend(thread);

the second call returns immediately that is very surprising. Is it a bug
or fature?

-- Sergei.



-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: Thread suspend when scheduler is locked.

[Sergei Organov]

A few clarifications below.

Sergei Organov <osv@javad.com> writes:
> Hello,
>
> In the following sequence:
>
>   cyg_scheduler_lock();
>   cyg_thread_suspend(thread);
>
> the second call returns immediately that is very surprising. Is it a bug
> or fature?

1. 'thread' is the current thread being executed so the suspend call asks
    thread for suspending itself.

2. cyg_thread_delay(delay) returns immediately no matter what is 'delay'
   value.

3. On the other hand, waiting on a primitive such as closed semaphore
   works fine.

Looks like a bug to me.

-- Sergei.


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Re: Thread suspend when scheduler is locked.

[Nick Garnett]

On 21/06/12 10:00, Sergei Organov wrote:
> A few clarifications below.
> 
> Sergei Organov <osv@javad.com> writes:
>> Hello,
>>
>> In the following sequence:
>>
>>   cyg_scheduler_lock();
>>   cyg_thread_suspend(thread);
>>
>> the second call returns immediately that is very surprising. Is it a bug
>> or fature?
> 
> 1. 'thread' is the current thread being executed so the suspend call asks
>     thread for suspending itself.
> 
> 2. cyg_thread_delay(delay) returns immediately no matter what is 'delay'
>    value.
> 
> 3. On the other hand, waiting on a primitive such as closed semaphore
>    works fine.
> 
> Looks like a bug to me.


The form of many of these simple functions is to lock the scheduler,
alter whatever state is appropriate, and then unlock the scheduler, with
the understanding that any scheduling may happen at that point.
Semaphore waits and other things are slightly different because they
have changes to state to make after the wait, and need to retain the
lock across it.

In the cases you cite, any scheduling consequences of the calls made
while the scheduler lock is claimed will occur when you finally unlock
the scheduler yourself.

One can argue that this is useful functionality, since it allows you to
make a number of system state changes atomically under the scheduler
lock. However, I don't actually know of any examples where this has been
done, except in DSRs.

Essentially locking the scheduler yourself is like running in a DSR,
there is only a limited number of calls it makes sense to use, and their
effects may be deferred until the scheduler lock returns to zero.

While it might be possible to rewrite some of these routines to work
even when the scheduler lock is already claimed, it would make them
considerably more complex, larger, and thus less maintainable.



-- 
Nick Garnett                                      eCos Kernel Architect
eCosCentric Limited    http://www.eCosCentric.com      The eCos experts
Barnwell House, Barnwell Drive, Cambridge, UK.     Tel: +44 1223 245571
Registered in England and Wales:                        Reg No: 4422071

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: Thread suspend when scheduler is locked.

[Sergei Organov]

Nick Garnett <nickg@calivar.com> writes:

> On 21/06/12 10:00, Sergei Organov wrote:
>> A few clarifications below.
>> 
>> Sergei Organov <osv@javad.com> writes:
>>> Hello,
>>>
>>> In the following sequence:
>>>
>>>   cyg_scheduler_lock();
>>>   cyg_thread_suspend(thread);
>>>
>>> the second call returns immediately that is very surprising. Is it a bug
>>> or fature?
>> 
>> 1. 'thread' is the current thread being executed so the suspend call asks
>>     thread for suspending itself.
>> 
>> 2. cyg_thread_delay(delay) returns immediately no matter what is 'delay'
>>    value.
>> 
>> 3. On the other hand, waiting on a primitive such as closed semaphore
>>    works fine.
>> 
>> Looks like a bug to me.
>
>

[...]

> Essentially locking the scheduler yourself is like running in a DSR,
> there is only a limited number of calls it makes sense to use, and their
> effects may be deferred until the scheduler lock returns to zero.

So, there is no way in eCos to just disable _preemption_? I.e.,
(temporarily) switch from preemptive to cooperative multi-threading?

-- Sergei.


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Re: Thread suspend when scheduler is locked.

[Nick Garnett]

On 21/06/12 12:51, Sergei Organov wrote:
> Nick Garnett <nickg@calivar.com> writes:

>> Essentially locking the scheduler yourself is like running in a DSR,
>> there is only a limited number of calls it makes sense to use, and their
>> effects may be deferred until the scheduler lock returns to zero.
> 
> So, there is no way in eCos to just disable _preemption_? I.e.,
> (temporarily) switch from preemptive to cooperative multi-threading?

Not really. You could try implementing a new scheduler that only did
context switches when a yield function is called. If you disable
timeslicing and run all your threads at the same priority they will
essentially run round-robin and only lose the CPU when they explicitly
suspend, although higher priority threads will still preempt. But the
usual way to make a set of threads work cooperatively is to use a mutex
to serialize their execution.

Also note that the scheduler lock will also delay or block execution of
DSRs. Which is probably not what you want.

Trying to make eCos work as a cooperatively scheduled system is likely
to be more trouble than it is worth. It wasn't really designed to work
that way.



-- 
Nick Garnett                                      eCos Kernel Architect
eCosCentric Limited    http://www.eCosCentric.com      The eCos experts
Barnwell House, Barnwell Drive, Cambridge, UK.     Tel: +44 1223 245571
Registered in England and Wales:                        Reg No: 4422071

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: Thread suspend when scheduler is locked.

[Sergei Organov]

Nick Garnett <nickg@calivar.com> writes:

> On 21/06/12 12:51, Sergei Organov wrote:
>> Nick Garnett <nickg@calivar.com> writes:
>
>>> Essentially locking the scheduler yourself is like running in a DSR,
>>> there is only a limited number of calls it makes sense to use, and their
>>> effects may be deferred until the scheduler lock returns to zero.
>> 
>> So, there is no way in eCos to just disable _preemption_? I.e.,
>> (temporarily) switch from preemptive to cooperative multi-threading?
>
> Not really. You could try implementing a new scheduler that only did
> context switches when a yield function is called. If you disable
> timeslicing and run all your threads at the same priority they will
> essentially run round-robin and only lose the CPU when they explicitly
> suspend, although higher priority threads will still preempt.

No, I need full-featured preemptive scheduling with multiple real-time
thread priorities. I thought "disable preemption" feature is pretty
standard among RTOSes and have it in my RTOS abstraction layer,
implemented as a call to cyg_scheduler_lock() in eCos as disabling
preemption is what description of the call claims.

In fact I use this feature for very special purpose and can re-design my
code to avoid it.

-- Sergei.


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Re: Thread suspend when scheduler is locked.

[Jonathan Larmour]

On 21/06/12 15:57, Sergei Organov wrote:
> Nick Garnett <nickg@calivar.com> writes:
> 
>> On 21/06/12 12:51, Sergei Organov wrote:
>>> Nick Garnett <nickg@calivar.com> writes:
>>
>>>> Essentially locking the scheduler yourself is like running in a DSR,
>>>> there is only a limited number of calls it makes sense to use, and their
>>>> effects may be deferred until the scheduler lock returns to zero.
>>>
>>> So, there is no way in eCos to just disable _preemption_? I.e.,
>>> (temporarily) switch from preemptive to cooperative multi-threading?
>>
>> Not really. You could try implementing a new scheduler that only did
>> context switches when a yield function is called. If you disable
>> timeslicing and run all your threads at the same priority they will
>> essentially run round-robin and only lose the CPU when they explicitly
>> suspend, although higher priority threads will still preempt.
> 
> No, I need full-featured preemptive scheduling with multiple real-time
> thread priorities. I thought "disable preemption" feature is pretty
> standard among RTOSes and have it in my RTOS abstraction layer,
> implemented as a call to cyg_scheduler_lock() in eCos as disabling
> preemption is what description of the call claims.

The problem with any sort of abstraction layer is matching what you mean with
what exactly is provided. In particular does disabling pre-emption mean that
DSRs should be disabled too? From what you're saying, I think you mean that
DSRs should still be allowed.

So if you just want to make sure no other threads can pre-empt this one, then
surely the answer is to temporarily change its thread priority to 1, and
change it back when you're done. And reserve priority 1 for threads in this
situation.

If multiple threads can disable pre-emption at the same time, but still be
able to suspend themselves, then you would have to work out what that really
is meant to mean. It's not clear what would be correct behaviour in that
situation in general, so you would have to decide what that means in your
specific case.

> In fact I use this feature for very special purpose and can re-design my
> code to avoid it.

Good, but I do wonder if that could have been avoided...

Jifl
-- 
eCosCentric Limited      http://www.eCosCentric.com/     The eCos experts
Barnwell House, Barnwell Drive, Cambridge, UK.       Tel: +44 1223 245571
Registered in England and Wales: Reg No 4422071.
------["Si fractum non sit, noli id reficere"]------       Opinions==mine

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: Thread suspend when scheduler is locked.

[Sergei Organov]

Jonathan Larmour <jifl@eCosCentric.com> writes:
> On 21/06/12 15:57, Sergei Organov wrote:
>> Nick Garnett <nickg@calivar.com> writes:
>> 
>>> On 21/06/12 12:51, Sergei Organov wrote:
>>>> Nick Garnett <nickg@calivar.com> writes:
>>>
>>>>> Essentially locking the scheduler yourself is like running in a DSR,
>>>>> there is only a limited number of calls it makes sense to use, and their
>>>>> effects may be deferred until the scheduler lock returns to zero.
>>>>
>>>> So, there is no way in eCos to just disable _preemption_? I.e.,
>>>> (temporarily) switch from preemptive to cooperative multi-threading?
>>>
>>> Not really. You could try implementing a new scheduler that only did
>>> context switches when a yield function is called. If you disable
>>> timeslicing and run all your threads at the same priority they will
>>> essentially run round-robin and only lose the CPU when they explicitly
>>> suspend, although higher priority threads will still preempt.
>> 
>> No, I need full-featured preemptive scheduling with multiple real-time
>> thread priorities. I thought "disable preemption" feature is pretty
>> standard among RTOSes and have it in my RTOS abstraction layer,
>> implemented as a call to cyg_scheduler_lock() in eCos as disabling
>> preemption is what description of the call claims.
>
> The problem with any sort of abstraction layer is matching what you mean with
> what exactly is provided. In particular does disabling pre-emption mean that
> DSRs should be disabled too? From what you're saying, I think you mean that
> DSRs should still be allowed.

No, in fact I didn't care. For 2-level interrupts servicing eCos
implements, complete set of features would be:

1. Disable/restore ISRs (also effectively does 2 and 3)
2. Disable/restore DSRs (also effectively does 3)
3. Disable/restore threads switching

all these states being thread-local. Current scheduler lock feature is
close to (2), except that it introduces surprising behavior I was
referring to.

The main idea behind disabling preemption is that as soon as system
function that could lead to DSRs invokation and/or task switch is called
explicitly, DSRs are run and taks will be switched. This is what does
not happen in some of current eCos system functions.

Note that even though DSRs are disabled when scheduler is locked, given
scheduler lock is thread-local, even timed waits should work out of the
box for the thread that locks scheduler as scheduler is still enabled
when this thread sleeps.

> So if you just want to make sure no other threads can pre-empt this one, then
> surely the answer is to temporarily change its thread priority to 1, and
> change it back when you're done. And reserve priority 1 for threads in this
> situation.

Yes, reserving highest priority for that purpose would work, provided
scheduler supports dynamic priority change.

> If multiple threads can disable pre-emption at the same time, but still be
> able to suspend themselves, then you would have to work out what that really
> is meant to mean. It's not clear what would be correct behaviour in that
> situation in general, so you would have to decide what that means in your
> specific case.
>
>> In fact I use this feature for very special purpose and can re-design my
>> code to avoid it.
>
> Good, but I do wonder if that could have been avoided...

I believe it could. Hopefully it's just a matter of adding a few
explicit calls to reschedule() where appropriate in the eCos kernel code
as waits on primitives such as semaphores do work as expected, so all
the required functionality is apparently already there.

-- Sergei.


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss