[PATCH v3 09/13] arm64: debug: split single stepping exception entry

[Mark Rutland]

This looks good; one comment below.

On Mon, Jun 09, 2025 at 06:34:09PM +0100, Ada Couprie Diaz wrote:
> +void do_softstep(unsigned long esr, struct pt_regs *regs)
>  {
> -	/*
> -	 * If we are stepping a pending breakpoint, call the hw_breakpoint
> -	 * handler first.
> -	 */
> -	if (try_step_suspended_breakpoints(regs))
> -		return 0;
> -
>  	if (call_step_hook(regs, esr) == DBG_HOOK_HANDLED)
> -		return 0;
> +		return;
>  
>  	if (user_mode(regs)) {
>  		send_user_sigtrap(TRAP_TRACE);
> @@ -219,10 +212,8 @@ static int single_step_handler(unsigned long unused, unsigned long esr,
>  		 */
>  		set_regs_spsr_ss(regs);
>  	}
> -
> -	return 0;
>  }
> -NOKPROBE_SYMBOL(single_step_handler);
> +NOKPROBE_SYMBOL(do_softstep);

With the EL0/EL1 entry paths split up, it would be nice to split this
into separate do_el{0,1}_softstep() handlers, like we do for
do_el{0,1}_undef(). With the relevant portions of call_step_hook()
folded in, that'd leave us with:

| void do_el0_softstep(struct pt_regs *regs, unsigned long esr)
| {
| 	if (uprobe_singlestep_handler(regs, esr) == DBG_HOOK_HANDLED)
| 		return;
| 	
| 	send_user_sigtrap(TRAP_TRACE);
| 	/*
| 	 * ptrace will disable single step unless explicitly asked to
| 	 * re-enable it. For other clients, it makes sense to leave it
| 	 * enabled (i.e. rewind the controls to the active-not-pending
| 	 * state).
| 	 */
| 	 user_rewind_single_step(current);
| }
| 
| void do_el1_softstep(struct pt_regs *regs, unsigned long esr)
| {
| 	if (kgdb_singlestep_handler(regs, esr) == DBG_HOOK_HANDLED)
| 		return;
| 	
| 	pr_warn("Unexpected kernel single-step exception at EL1\n");
| 	/*
| 	 * Re-enable stepping since we know that we will be
| 	 * returning to regs.
| 	 */
| 	set_regs_spsr_ss(regs);
| }
| NOKPROBE_SYMBOL(do_el1_softstep):

Doing that would remove some indirection, and make this a bit easier to
follow. That would also permit the EL0 handler to be kprobed (which is
safe since there's no risk of recursion in the EL0 handler).

Mark.