[PATCH v5] arm64: Introduce IRQ stack
jungseoklee85 at gmail.com
Wed Oct 21 07:56:13 PDT 2015
On Oct 21, 2015, at 1:04 AM, James Morse wrote:
> On 20/10/15 16:05, Jungseok Lee wrote:
>> On Oct 20, 2015, at 7:05 PM, James Morse wrote:
>>> On 17/10/15 15:27, Jungseok Lee wrote:
>>>> diff --git a/arch/arm64/kernel/irq.c b/arch/arm64/kernel/irq.c
>>>> index 9f17ec0..13fe8f4 100644
>>>> --- a/arch/arm64/kernel/irq.c
>>>> +++ b/arch/arm64/kernel/irq.c
>>>> @@ -30,6 +30,8 @@
>>>> unsigned long irq_err_count;
>>>> +DEFINE_PER_CPU(void *, irq_stacks);
>>>> int arch_show_interrupts(struct seq_file *p, int prec)
>>>> show_ipi_list(p, prec);
>>>> @@ -47,9 +49,31 @@ void __init set_handle_irq(void (*handle_irq)(struct pt_regs *))
>>>> handle_arch_irq = handle_irq;
>>>> +static char boot_irq_stack[IRQ_STACK_SIZE] __aligned(IRQ_STACK_SIZE);
>>> Is kmalloc() not available this early? Regardless:
>>> As Catalin is happy with the Image size increase , this could become
>>> something like:
>>>> DEFINE_PER_CPU(union thread_union, irq_stack);
>>> Which will remove the need to __alloc_irq_stack()s.
>> We cannot rely on static allocation using percpu in case of 4KB page system.
>> Since ARM64 utilizes generic setup_per_cpu_areas(), tpidr_el1 is PAGE_SIZE
>> aligned. That is, IRQ stack is allocated with PAGE_SIZE alignment for secondary
>> cores. However, the top-bit method works well under the assumption that IRQ
>> stack is IRQ_STACK_SIZE aligned. It leads to IRQ re-entrance check failure.
> Yikes! That is nasty... well caught!
> Now I understand why you had the per-cpu version #ifdef'd in your example
> hunk earlier!
> Do we need the irq stack to be aligned like this? It was originally for the
> old implementation of current_thread_info(), which this patch changes.
Not necessarily, but the alignment restriction helps us to simplify IRQ
re-entrance check and linkage between a process stack and IRQ one.
> If its just the re-entrance check that needs the alignment, maybe the
> irq_count approach is better (but count late not early), and drop the
> alignment requirement on interrupt stacks. We know re-entrant irqs will
> keep sp_el0, so the new current_thread_info() still works.
Hmm.. I cannot image how simple this logic is without implementation detail.
We should consider the number of memory access such as pointer read under
count based approach. In this context, I guess a static allocation is better
than dynamic one.
> I think Catalin's comment is to count like x86 (64 bit version) does in
> arch/x86/entry/entry_64.S:do_softirq_own_stack, and treat this as a
> re-entrance flag in entry.S.
> task stacks still need to be aligned, as when user space is interrupted, we
> have a kernel stack, and no idea where its struct task_struct is, unless we
> know it was originally THREAD_SIZE aligned.
More information about the linux-arm-kernel