[PATCH v3 0/5] ARM64: Add kernel probes(Kprobes) support

[David Long]

On 12/10/14 11:38, Steve Capper wrote:
> On Tue, Dec 09, 2014 at 09:27:18AM -0500, David Long wrote:
>> On 12/09/14 08:33, Steve Capper wrote:
>>> On Thu, Dec 04, 2014 at 08:53:03PM +0900, Masami Hiramatsu wrote:
>
> [...]
>
>>>
>>> Not sure if this is helpful, but the following also caused a crash for
>>> me:
>>>
>>> echo "p:trace_event_buffer_lock_reserve trace_event_buffer_lock_reserve" > /sys/kernel/debug/tracing/kprobe_events
>>> echo "p:memcpy memcpy" >> /sys/kernel/debug/tracing/kprobe_events
>>> echo 1 > /sys/kernel/debug/tracing/events/kprobes/enable
>>>
>>> [immediate crash]
>>>
>>> The crash point for me is in the arm64 ASID allocator, it again looks
>>> like the interrupts are in an unexpected state.
>>> (check_and_switch_context goes down the irqs disabled code path, I
>>> think incorrectly).
>>>
>>> This occurred for me both with and without the proposed irq saving fix.
>>>
>>> I will do some more digging.
>>>
>>
>> Thanks, more information is good.
>>
>
> Hi,
>
> Some good news, I think I've fixed the problem I've been experiencing.
>
> Basically, I've torn out all the interrupt save/restore and have
> narrowed the scope to just sandwich the instruction single-step. This
> simplifies a lot of logic, and I've now been able to perf record a
> kprobe on memcpy (and the trace_event_buffer_lock_reserve + memcpy
> test) without any issues on a Juno platform.
>
> I may have been somewhat over-zealous with the chainsaw, so please do
> put this fix through its paces.
>
> Cheers,
> --
> Steve
>
>
>  From d3f4d80ce19bec71bd03209beb2fbfd8084d6543 Mon Sep 17 00:00:00 2001
> From: Steve Capper <steve.capper at linaro.org>
> Date: Mon, 1 Dec 2014 11:30:10 +0000
> Subject: [PATCH] Fix the interrupt handling for kprobes
>
> ---
>   arch/arm64/kernel/kprobes.c | 16 ++--------------
>   1 file changed, 2 insertions(+), 14 deletions(-)
>
> diff --git a/arch/arm64/kernel/kprobes.c b/arch/arm64/kernel/kprobes.c
> index be7c330..d39d826 100644
> --- a/arch/arm64/kernel/kprobes.c
> +++ b/arch/arm64/kernel/kprobes.c
> @@ -229,10 +229,6 @@ skip_singlestep_missed(struct kprobe_ctlblk *kcb, struct pt_regs *regs)
>   {
>   	/* set return addr to next pc to continue */
>   	instruction_pointer(regs) += sizeof(kprobe_opcode_t);
> -
> -	if (kcb->kprobe_status != KPROBE_REENTER)
> -		kprobes_restore_local_irqflag(regs);
> -
>   }
>
>   static void __kprobes setup_singlestep(struct kprobe *p,
> @@ -259,7 +255,7 @@ static void __kprobes setup_singlestep(struct kprobe *p,
>   			spsr_set_debug_flag(regs, 0);
>
>   		/* IRQs and single stepping do not mix well. */
> -		local_irq_disable();
> +		kprobes_save_local_irqflag(regs);
>   		kernel_enable_single_step(regs);
>   		instruction_pointer(regs) = slot;
>   	} else	{
> @@ -326,7 +322,6 @@ post_kprobe_handler(struct kprobe_ctlblk *kcb, struct pt_regs *regs)
>   	}
>
>   	reset_current_kprobe();
> -	kprobes_restore_local_irqflag(regs);
>   }
>
>   int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)
> @@ -380,8 +375,6 @@ int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)
>   			return 1;
>
>   		break;
> -	default:
> -		break;
>   	}
>   	return 0;
>   }
> @@ -446,7 +439,6 @@ void __kprobes kprobe_handler(struct pt_regs *regs)
>   		 * handling of this interrupt is appropriate.
>   		 * Return back to original instruction, and continue.
>   		 */
> -		kprobes_restore_local_irqflag(regs);
>   		return;
>   	} else if (cur) {
>   		/* We probably hit a jprobe.  Call its break handler. */
> @@ -459,7 +451,6 @@ void __kprobes kprobe_handler(struct pt_regs *regs)
>   		/* breakpoint is removed, now in a race
>   		 * Return back to original instruction & continue.
>   		 */
> -		kprobes_restore_local_irqflag(regs);
>   	}
>   }
>
> @@ -485,6 +476,7 @@ kprobe_single_step_handler(struct pt_regs *regs, unsigned int esr)
>   	retval = kprobe_ss_hit(kcb, instruction_pointer(regs));
>
>   	if (retval == DBG_HOOK_HANDLED) {
> +		kprobes_restore_local_irqflag(regs);
>   		kernel_disable_single_step();
>
>   		if (kcb->kprobe_status == KPROBE_REENTER)
> @@ -499,7 +491,6 @@ kprobe_single_step_handler(struct pt_regs *regs, unsigned int esr)
>   static int __kprobes
>   kprobe_breakpoint_handler(struct pt_regs *regs, unsigned int esr)
>   {
> -	kprobes_save_local_irqflag(regs);
>   	kprobe_handler(regs);
>   	return DBG_HOOK_HANDLED;
>   }
> @@ -563,7 +554,6 @@ int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
>   		memcpy((void *)stack_addr, kcb->jprobes_stack,
>   		       MIN_STACK_SIZE(stack_addr));
>   		preempt_enable_no_resched();
> -		kprobes_restore_local_irqflag(regs);
>   		return 1;
>   	}
>   	return 0;
> @@ -655,8 +645,6 @@ trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
>   		kfree(ri);
>   	}
>
> -	kprobes_restore_local_irqflag(regs);
> -
>   	/* return 1 so that post handlers not called */
>   	return 1;
>   }
>

Sorry for the delay in responding.  The assumption with the existing 
code is that you can't enable single-stepping in MDSCR until you disable 
interrupts.  But since we get to that point (in this case) with debug 
exceptions masked in daif I suppose this is an unnecessary constraint, 
as you have demonstrated.  I do actually wonder if we need to set the 
bit in MDSCR_EL1 at all since we're also setting it in the spsr in the 
regs structure that will be restored when the eret is done, but it looks 
like kernel_enable_single_step() sets both.

-dl