Re: [PATCH 1/2] alpha/ptrace: Record and handle the absence of switch_stack

[Al Viro]

On Wed, Jun 16, 2021 at 01:31:52PM -0500, Eric W. Biederman wrote:

> +.macro	SAVE_SWITCH_STACK
> +	DO_SWITCH_STACK
> +1:	ldl_l	$1, TI_FLAGS($8)
> +	bis	$1, _TIF_ALLREGS_SAVED, $1
> +	stl_c	$1, TI_FLAGS($8)
> +	beq	$1, 2f
> +.subsection 2
> +2:	br	1b
> +.previous
> +.endm

What the hell?  *IF* you are going to go that way, at least put it into
->status, not ->flag - those are thread-synchronous and do not require that
kind of masturbation.

> +.macro	RESTORE_SWITCH_STACK
> +1:	ldl_l	$1, TI_FLAGS($8)
> +	bic	$1, _TIF_ALLREGS_SAVED, $1
> +	stl_c	$1, TI_FLAGS($8)
> +	beq	$1, 2f
> +.subsection 2
> +2:	br	1b
> +.previous
> +	UNDO_SWITCH_STACK
> +.endm

Ditto.  What do you need that flag for, anyway?

> @@ -117,7 +117,13 @@ get_reg_addr(struct task_struct * task, unsigned long regno)
>  		zero = 0;
>  		addr = &zero;
>  	} else {
> -		addr = task_stack_page(task) + regoff[regno];
> +		int off = regoff[regno];
> +		if (WARN_ON_ONCE((off < PT_REG(r0)) &&
> +				!test_ti_thread_flag(task_thread_info(task),
> +						     TIF_ALLREGS_SAVED)))
> +			addr = &zero;
> +		else
> +			addr = task_stack_page(task) + off;

A sanity check in slow path, buggering the hell out of a fast path?