[RFC PATCH 2/2] arm64: support HAVE_IRQ_EXIT_ON_IRQ_STACK

[Qi Zheng]

On 2022/7/7 20:49, Arnd Bergmann wrote:
> On Thu, Jul 7, 2022 at 1:05 PM Qi Zheng <zhengqi.arch at bytedance.com> wrote:
>>
>> Since softirqs are handled on the per-CPU IRQ stack,
>> let's support HAVE_IRQ_EXIT_ON_IRQ_STACK which causes
>> the core code to invoke __do_softirq() directly without
>> going through do_softirq_own_stack().
>>
>> Signed-off-by: Qi Zheng <zhengqi.arch at bytedance.com>
> 
> I think the idea is right, but the extra function pointer adds more complexity
> than necessary:
> 
>>   static __always_inline void __el1_irq(struct pt_regs *regs,
>>                                        void (*handler)(struct pt_regs *))
>>   {
>>          enter_from_kernel_mode(regs);
>>
>> -       irq_enter_rcu();
>> -       do_interrupt_handler(regs, handler);
>> -       irq_exit_rcu();
>> +       do_interrupt_handler(regs, handler, irq_handler);
>>
>>          arm64_preempt_schedule_irq();
>>
>> @@ -699,9 +711,7 @@ static void noinstr el0_interrupt(struct pt_regs *regs,
>>          if (regs->pc & BIT(55))
>>                  arm64_apply_bp_hardening();
>>
>> -       irq_enter_rcu();
>> -       do_interrupt_handler(regs, handler);
>> -       irq_exit_rcu();
>> +       do_interrupt_handler(regs, handler, irq_handler);
>>
>>          exit_to_user_mode(regs);
>>   }
> 
> Would it be possible to instead pull out the call_on_irq_stack() so these
> two functions are instead called on the IRQ stack already?

Hi,

Do you mean to modify call_on_irq_stack()?

I have tried doing a conditional jump inside call_on_irq_stack() like
this:

--- a/arch/arm64/kernel/entry.S
+++ b/arch/arm64/kernel/entry.S
@@ -888,13 +888,22 @@ SYM_FUNC_START(call_on_irq_stack)

         /* Move to the new stack and call the function there */
         mov     sp, x16
-       blr     x1
+
+       cmp     x2, #1
+       b.eq    99f
+
+       blr     x1
+       b       999f
+
+99:    bl      irq_enter_rcu
+       blr     x1
+       bl      irq_exit_rcu

         /*
          * Restore the SP from the FP, and restore the FP and LR from 
the frame
          * record.
          */
-       mov     sp, x29
+999:   mov     sp, x29
         ldp     x29, x30, [sp], #16
  #ifdef CONFIG_SHADOW_CALL_STACK
         ldp     scs_sp, xzr, [sp], #16

But this also requires a new parameter in do_interrupt_handler.

I also considered implementing call_on_irq_stack() for nmi and irq
separately, but later think it's unnecessary.

> 
>          Arnd

Thanks,
Qi