arch: call arch_smp_init() directly, do not use SYS_INIT

[nashif]

Move this to a call in the init process. arch_* calls are no services
and should be called consistently during initialization.

Signed-off-by: Anas Nashif anas.nashif@intel.com

[evgeniy-paltsev]

Is there a reason for calling it in other place (earlier)? We call INIT_LEVEL_SMP here:

zephyr/kernel/init.c

Line 451 in 593dfe1

z_sys_init_run_level(INIT_LEVEL_SMP);

This breaks SMP on ARC as we need to call ARC arch_smp_init after we initialize secondary CPU cores (after z_smp_init is called)

[andyross]

Then maybe we want to move this into soc/arch layers and out of arch_interface.h entirely? Really there's zero kernel involvement to the function, it's purely there so the platform layers can do their setup, but when/how isn't really something the OS can (or wants to) decide on.

[nashif]

now on arc, arch_smp_init is called in PRE_KERNEL_1., so I am not following.

INIT_LEVEL_SMP has been something we wanted to get rid of, if I am not mistaken, we just have 1-2 users of that level.

[tbursztyka]

SMP level and actually initializing smp is indeed completely different.

[andyross]

Agreed, this is sort of a cargo-cult thing; not sure where the pattern came from originally.

[nashif]

@evgeniy-paltsev can you please take another look?

[evgeniy-paltsev]

Ok, initially I was thinking about

zephyr/drivers/interrupt_controller/intc_arcv2_irq_unit.c

Line 80 in a35d5e7

SYS_INIT(arc_shared_intc_update_post_smp, SMP, 0);

as it's only SYS_INIT with SMP level for ARC, but we indeed changing the different part, so my initial comment is not relevant.

However, I've tested this on ARC SMP platforms - and I've got lot's of test failures (~120 failed test per platform). I need to investigate this.

[nashif]

However, I've tested this on ARC SMP platforms - and I've got lot's of test failures (~120 failed test per platform). I need to investigate this.

can you try moving:

#if defined(CONFIG_SMP)
        arch_smp_init();
#endif

to be between PRE_KERNEL_1 and PRE_KERNEL_2? I think it is related to the timer initialization.

[nashif]

No need for a DNM, it is already blocked by your review.

[evgeniy-paltsev]

Now it works.

In the previous implementation there was a conflict of initialization between arch_smp_init and a timer driver for ARC.