librelist archives

« back to archive

ARM port

ARM port

From:
Kees Stoop
Date:
2010-04-04 @ 22:55
Hi All,

Has someone already been working on this? Well, I did....
To get this working, I would like to ask you some questions.
- I guess the ARM interrupt handler must be able to do 'nested' 
interrupts right?
- Does the timer interrupt (the one in atomport.c) ever return? I mean, 
this interrupt calls the scheduler and that switches to one of the 
thread routines isn't it?

Thanks,
Kees.

Re: [atomthreads] ARM port

From:
Kelvin Lawson
Date:
2010-04-04 @ 23:17
Hi Kees,

Good to hear from you.

A user has reported some success on ARM9 but I haven't yet integrated
that port into the official tree and completed all automated tests on
it. I am planning to work on an ARM7 or ARM9 port myself within the
next couple of weeks but meanwhile I'll help out wherever I can. Which
ARM device are you using?

> - I guess the ARM interrupt handler must be able to do 'nested'
> interrupts right?

Both of the current ports (ATmega and STM8) currently run in
non-nested mode. Changes may be necessary for nested interrupts and I
will make any kernel changes necessary for that when I add support for
nesting. Meanwhile there is no reason why you would need to use nested
interrupts for the benefit of the kernel, only if your application
requires it.

> - Does the timer interrupt (the one in atomport.c) ever return? I mean,
> this interrupt calls the scheduler and that switches to one of the
> thread routines isn't it?

This will depend on the port, but the way the ATmega and STM8 ports
work is that any interrupt handlers (including the timer interrupt)
which cause a thread switch, will immediately switch execution to the
new thread. The final instructions of the interrupt handler are not
executed until the original thread is scheduled back in. This is OK on
both of these architectures because it is possible to "complete" an
interrupt by regular return calls rather than "return-from-interrupt"
calls. There is some detailed discussion of this in the new STM8 port
(ports/stm8/atomport-asm.s). There may be more appropriate schemes for
other architectures.

Please let me know if I can help with any further porting issues.

Thanks,
Kelvin.

Re: [atomthreads] ARM port

From:
Kees Stoop
Date:
2010-04-05 @ 00:07
Hi Kelvin,

Wow, that is quick!

Well, I am working on a port for the ARM1026EJ an embedded core that my 
company is using in an ASIC. There shouldn't be too much difference 
between this and an ARM9 though. The ARM port is a little more 
complicated because of the various processor modes. Either I have to run 
the whole scheduler in IRQ mode and use the IRQ stack, or switch to 
system or user mode in the ARM's interrupt handler (In case I go for 
nested interrupts this is a must anyway.)

Anyway, thanks for the explanation. Well, I thought that the timer 
interrupt would never return and just switch to another thread, so that 
nested interrupts were a must in order to keep getting timer interrupts. 
Perhaps I should further check your code to see when and where the timer 
interrupt returns coz I still don't understand the mechanism on how to 
switch. When the timer interrupt returns when the original thread is 
scheduled back in, then something, some interrupt, should cause that. 
Since interrupts are disabled during the timer interrupt (unless I used 
nested interrupts) I don't see how is working.

Thanks,
Kees.


> Hi Kees,
>
> Good to hear from you.
>
> A user has reported some success on ARM9 but I haven't yet integrated
> that port into the official tree and completed all automated tests on
> it. I am planning to work on an ARM7 or ARM9 port myself within the
> next couple of weeks but meanwhile I'll help out wherever I can. Which
> ARM device are you using?
>
>    
>> - I guess the ARM interrupt handler must be able to do 'nested'
>> interrupts right?
>>      
> Both of the current ports (ATmega and STM8) currently run in
> non-nested mode. Changes may be necessary for nested interrupts and I
> will make any kernel changes necessary for that when I add support for
> nesting. Meanwhile there is no reason why you would need to use nested
> interrupts for the benefit of the kernel, only if your application
> requires it.
>
>    
>> - Does the timer interrupt (the one in atomport.c) ever return? I mean,
>> this interrupt calls the scheduler and that switches to one of the
>> thread routines isn't it?
>>      
> This will depend on the port, but the way the ATmega and STM8 ports
> work is that any interrupt handlers (including the timer interrupt)
> which cause a thread switch, will immediately switch execution to the
> new thread. The final instructions of the interrupt handler are not
> executed until the original thread is scheduled back in. This is OK on
> both of these architectures because it is possible to "complete" an
> interrupt by regular return calls rather than "return-from-interrupt"
> calls. There is some detailed discussion of this in the new STM8 port
> (ports/stm8/atomport-asm.s). There may be more appropriate schemes for
> other architectures.
>
> Please let me know if I can help with any further porting issues.
>
> Thanks,
> Kelvin.
>    

Re: [atomthreads] ARM port

From:
Kelvin Lawson
Date:
2010-04-07 @ 21:15
Hi Kees,

> Perhaps I should further check your code to see when and where the timer
> interrupt returns coz I still don't understand the mechanism on how to
> switch. When the timer interrupt returns when the original thread is
> scheduled back in, then something, some interrupt, should cause that.
> Since interrupts are disabled during the timer interrupt (unless I used
> nested interrupts) I don't see how is working.

I've made a note to add some more detail to the documentation
regarding context switches. But I'll try to give you a flavour now
(this applies to the current range of ports, new ports may context
switch slightly differently):

First of all, cooperative context switches occur with no interrupt at
all. If a thread schedules itself out (for example by calling
atomTimerDelay() to go to sleep) then the scheduler decides which
thread should be up next, and calls archContextSwitch() in order to
switch the register context to the new thread.

Now consider a thread (Thread B) is asleep on a timer, and is due to
be scheduled in when the next timer tick interrupt occurs:
* Thread A is currently running
* The timer tick interrupt occurs and the scheduler decides during the
timer interrupt that Thread B should be woken up. It calls
archContextSwitch() to do this.
* archContextSwitch saves the context of the previously running thread
(Thread A) and restores the context of the new thread (Thread B). On
exit from the timer interrupt, Thread B is now running. However the
latter instructions of the original interrupt handler are not
executed, because context has switched to Thread B during
archContextSwitch(). That thread went to sleep using atomTimerDelay()
so on exiting archContextSwitch() it goes back up the call stack
through atomTimerDelay(). The actual remaining instructions of the
original interrupt handler are not executed yet, because the call
stack of that interrupt was saved on the interrupted thread's stack
(Thread A). When Thread A is eventually scheduled back in, it returns
through the interrupt handler call stack.

This scheme is used for both the ATmega and STM8 ports - it may not be
possible to do it on all architectures. It is possible on both of
these architectures because interrupts can be exited much like
returning from regular functions. Taking ATmega as an example, the
difference between RETI and RET is whether interrupts are re-enabled
or not. Therefore you can exit the timer interrupt using just a RET,
because the new thread context restore re-enables interrupts as part
of its context restore anyway.

This is convenient because it reduces the context-switch routine and
allows it to be used for both cooperative switches and preemptive (via
interrupt) switches. On a cooperative switch it only needs to save
those registers which are not saved by the compiler on a function
call, because the call to archContextSwitch() is a regular C call. For
preemptive switches, the interrupt handler may already have saved some
registers, and when it calls out to the C routine archContextSwitch()
it will do the same as for cooperative switches (save registers which
cannot be clobbered by the called routine). This means that the same
small context-switch routine can be used in both cases, reducing
development and debug time, as well as avoiding (in the cooperative
case) unnecessary register saves where registers are already saved by
the compiler on calling out to archContextSwitch().

This is slightly complicated on architectures with separate stacks for
interrupts, and I intend to provide an example of this shortly.
Different context-switch schemes could be supported in architecture
ports while maintaining the same core kernel.

I have just started a Cortex M3 port, and intend to keep adding ports
which will cover some of the more unusual architectures to provide
wider examples for those porting to such architectures.

Thanks,
Kelvin.