[PATCH] arm64: kernel: Use a separate stack for irq interrupts.

[James Morse]

On 09/09/15 14:22, Jungseok Lee wrote:
> On Sep 9, 2015, at 1:47 AM, James Morse wrote:
>> On 08/09/15 15:54, Jungseok Lee wrote:
>>> On Sep 7, 2015, at 11:36 PM, James Morse wrote:
>>>> diff --git a/arch/arm64/kernel/irq.c b/arch/arm64/kernel/irq.c
>>>> index 463fa2e7e34c..10b57a006da8 100644
>>>> --- a/arch/arm64/kernel/irq.c
>>>> +++ b/arch/arm64/kernel/irq.c
>>>> @@ -26,11 +26,14 @@
>>>> #include <linux/smp.h>
>>>> #include <linux/init.h>
>>>> #include <linux/irqchip.h>
>>>> +#include <linux/percpu.h>
>>>> #include <linux/seq_file.h>
>>>> #include <linux/ratelimit.h>
>>>>
>>>> unsigned long irq_err_count;
>>>>
>>>> +DEFINE_PER_CPU(unsigned long, irq_sp) = 0;
>>>> +
>>>> int arch_show_interrupts(struct seq_file *p, int prec)
>>>> {
>>>> #ifdef CONFIG_SMP
>>>> @@ -55,6 +58,10 @@ void __init init_IRQ(void)
>>>> 	irqchip_init();
>>>> 	if (!handle_arch_irq)
>>>> 		panic("No interrupt controller found.");
>>>> +
>>>> +	/* Allocate an irq stack for the boot cpu */
>>>> +	if (alloc_irq_stack(smp_processor_id()))
>>>> +		panic("Failed to allocate irq stack for boot cpu.");
>>>> }
>>>>
>>>> #ifdef CONFIG_HOTPLUG_CPU
>>>> @@ -117,3 +124,48 @@ void migrate_irqs(void)
>>>> 	local_irq_restore(flags);
>>>> }
>>>> #endif /* CONFIG_HOTPLUG_CPU */
>>>> +
>>>> +/* Allocate an irq_stack for a cpu that is about to be brought up. */
>>>> +int alloc_irq_stack(unsigned int cpu)
>>>> +{
>>>> +	struct page *irq_stack_page;
>>>> +	union thread_union *irq_stack;
>>>> +
>>>> +	/* reuse stack allocated previously */
>>>> +	if (per_cpu(irq_sp, cpu))
>>>> +		return 0;
>>>
>>> I'd like to avoid even this simple check since CPU hotplug could be heavily
>>> used for power management.
>>
>> I don't think its a problem:
>> __cpu_up() contains a call to wait_for_completion_timeout() (which could
>> eventually end up in the scheduler), so I don't think it could ever be on a
>> 'really hot' path.
>>
>> For really frequent hotplug-like power management, cpu_suspend() makes use
>> of firmware support to power-off cores - from what I can see it doesn't use
>> __cpu_up().
> 
> In case of some platforms, CPU hotplug is triggered via sysfs for power management
> based on user data. What is advantage of putting stack allocation into this path?

It will only happen for CPUs that are brought up.


> IRQ stack allocation is an critical one-shot operation. So, there would be no issue
> to give this work to a booting core. 

I agree, but:

>From include/linux/cpumask.h:
> *  If HOTPLUG is enabled, then cpu_possible_mask is forced to have
> *  all NR_CPUS bits set, otherwise it is just the set of CPUs that
> *  ACPI reports present at boot.

(This doesn't seem to happen with DT - but might with ACPI.)

NR_CPUs could be much bigger than the number of cpus the system ever has.
Allocating a stack for every possible cpu would waste memory. It is better
to do it just-in-time, when we know the memory will be used.

This already happens for the per-cpu idle task, (please check I traced
these through correctly!)
_cpu_up()
  idle_thread_get()
    init_idle()
      fork_idle()
        copy_process()
          dup_task_struct()
            alloc_task_struct_node()
            alloc_thread_info_node()
            arch_dup_task_struct()

So plenty of memory-allocation occurs during _cpu_up(), idle_init() checks
whether the idle task has already been created.


>>>> +
>>>> +	irq_stack_page = alloc_kmem_pages(THREADINFO_GFP, THREAD_SIZE_ORDER);
>>>> +	if (!irq_stack_page) {
>>>> +		pr_crit("CPU%u: failed to allocate irq stack for cpu %u\n",
>>>> +			smp_processor_id(), cpu);
>>>> +		return -ENOMEM;
>>>> +	}
>>>> +	irq_stack = page_address(irq_stack_page);
> 
> I forgot leaving a comment here. How about using __get_free_pages? It returns an
> address instead of page. 

Good idea, I was paranoid about the stack not being correctly aligned (as
current_thread_info() relies on this). I lifted that alloc_kmem_pages line
from kernel/fork.c.


>>>> +
>>>> +	per_cpu(irq_sp, cpu) = (unsigned long)irq_stack->stack
>>>> +			       + THREAD_START_SP;
>>>> +
>>>> +	return 0;
>>>> +}
>>>> +
>>>> +/*
>>>> + * Copy struct thread_info between two stacks, and update current->stack.
>>>> + * This is used when moving to the irq exception stack.
>>>> + * Changing current->stack is necessary so that non-arch thread_info writers
>>>> + * don't use the new thread_info->task->stack to find the old thread_info when
>>>> + * setting flags like TIF_NEED_RESCHED...
>>>> + */
>>>> +asmlinkage int irq_copy_thread_info(unsigned long dst_sp, unsigned long src_sp)
>>>> +{
>>>> +	struct thread_info *src = get_thread_info(src_sp);
>>>> +	struct thread_info *dst = get_thread_info(dst_sp);
>>>> +
>>>> +	if (dst == src)
>>>> +		return 0;
>>>> +
>>>> +	*dst = *src;
>>>> +	current->stack = dst;
>>>> +
>>>> +	return 1;
>>>> +}
>>>
>>> Although this is 32-byte memcpy, sizeof(struct thread_info), this function is called
>>> twice to handle a single interrupt. Isn's it too expensive?
>>>
>>> This is a major difference between my approach and this patch. I think we should get
>>> some feedbacks on struct thread_info information management for v2 preparation.
>>
>> Agreed.
>>
>> The other difference, as Akashi Takahiro pointed out, is the behaviour of
>> object_is_on_stack(). What should this return for an object on an irq stack?
> 
> 0 should be returned in that case to align with x86 behavior according to check_stack()
> context and the commit, 81520a1b0649d0.

Unless we can update ftrace on x86 too! :)


Thanks for your comments,


James