Michael Neuhauser wrote:
On Mon, 2005-03-21 at 22:05, Jan Kiszka wrote:
Michael Neuhauser schrieb:
On Mon, 2005-03-21 at 17:54, Jan Kiszka wrote:
rtai_oldnames.h in magma-CVS is broken. This patch fixes it for x86,
other architectures may require a check as well.
With this patch applied, RTnet at least loads on top of magma now.
Jan
[...]
-- rtai_oldnames.h 5 Jan 2005 16:48:31 -0000 1.4
+++ rtai_oldnames.h 21 Mar 2005 16:36:29 -0000
@@ -37,9 +37,9 @@
#define IFLAG RTAI_IFLAG
#define hard_cli() rtai_cli()
#define hard_sti() rtai_sti()
-#define hard_save_flags_and_cli(x) rtai_local_irq_save(x)
-#define hard_restore_flags(x) rtai_local_irq_restore(x)
-#define hard_save_flags(x) rtai_local_irq_flags(x)
+#define hard_save_flags_and_cli(x) rtai_save_flags_and_cli(x)
+#define hard_restore_flags(x) rtai_restore_flags(x)
+#define hard_save_flags(x) rtai_save_flags(x)
#define hard_cpu_id adeos_processor_id
#endif /* __KERNEL__ */
Hm, there is definitely something wrong here (rtai_local_irq_*() are not
defined for i386). But I'm not sure if your patch is correct either.
It lets RTnet compile, isn't this enough? ;)
It's a start :-)
Seriously:
What is expected from the hard_*() macros? (I'm not so familiar with
"historical" RTAI but would like to get it right for ARM.) If they
should disable the CPU's interrupt enabled flag (as their name suggests)
then it should be
#define hard_cli() rtai_hw_cli()
#define hard_sti() rtai_hw_sti()
#define hard_save_flags_and_cli() rtai_hw_save_flags_and_cli()
...
I think hard_* were originally (rthal) intended to touch the hardware
directly. With adeos, it's sufficient (I think it's actually the only
correct way) to stall the pipeline in front of RTAI.
If you know what you are doing it is perfectly OK to hard disable IRQs
(at least I see it that way).
In fact, that's ok for the use case RTnet.
Your patch is only correct if the hard_*() macros are expected to
prevent interrupts to be delivered to RTAI, but nothing more, i.e. you
can't use them to protect time-critical hardware operations (if only the
RTAI pipeline is stalled, interrupts are still handled by Adeos, they
are just not dispatched to RTAI).
So, let's walk the #define chain of x86-vesuvio (assuming it is correct):
Wait a minute, are we talking about vesuvio or magma/vulcano now?
hard_cli() -> rtai_cli() -> adeos_stall_pipeline_from(&rtai_domain)
That's what I did to fiddle out the hopefully correct meaning for my patch.
The define chain is not the problem. The problem seems to me, that it
seems unclear what the hard_*() macros should provide:
* protection against IRQ-delivery to RTAI
or
* protection against IRQs happening at all
If RTAI is the highest domain in the interrupt pipeline, Adeos will use
hw masking directly, so the effect is basically the same. But bypassing
Adeos is in general a bad idea, because this would prevent you to stack
other domains before RTAI for instrumentation/debugging purpose, since
the virtualization would be lost.
The only situation where immediate cli/sti is a requirement concerns
sections of code that could conflict with any code run on behalf of
__adeos_handle_irq, basically the acknowledge handlers, and by extension
short sections of code directly fiddling with the hw that might behave
incorrectly if interrupted at all. To find those, just build a SMP
kernel on a UP box to activate the spinlocks, and let the deadlocks
happen. In all other cases, it's usually better to let Adeos handle the
interrupt mask, this is actually what it is there for.
I don't expect you to answer this, but maybe someone with a big picture
view (Paolo, are you listening?) might shed some light on it.
Mike
--
Philippe.
