Re: [PATCH] arm64: use sregs for syscall probe registers

["van der Linden, Frank" <fllinden@amazon.com>]

On 4/15/19, 2:32 PM, "William Cohen" <wcohen@redhat.com> wrote:

On 4/12/19 1:30 PM, Frank van der Linden wrote:
> Since syscall wrappers are now active on arm64 (4.19+), arguments need to
> be retrieved the right way, by checking if there is a saved set of system
> call registers, and using them if there are.
> ---
>  tapset/arm64/registers.stp | 50 ++++++++++++++++++++++++++++------------------
>  1 file changed, 31 insertions(+), 19 deletions(-)
> 
> diff --git a/tapset/arm64/registers.stp b/tapset/arm64/registers.stp
> index b2e56495d..b001b8efe 100644
> --- a/tapset/arm64/registers.stp
> +++ b/tapset/arm64/registers.stp
> @@ -107,6 +107,36 @@ function u_register:long (name:string) {
>  	return _stp_register(name, 0)
>  }
>  
> +function _stp_arg_register:long (argnum:long) %{ /* pure */
> +	long val;
> +	struct pt_regs *regs;
> +
> +	if (STAP_ARG_argnum < 1 || STAP_ARG_argnum > 8) {
> +		snprintf(CONTEXT->error_buffer, sizeof(CONTEXT->error_buffer),
> +				"Cannot access arg(%lld)",
> +				(long long)STAP_ARG_argnum);
> +		CONTEXT->last_error = CONTEXT->error_buffer;
> +		return;
> +	}
> +
> +	/* syscall-in-pt_regs mode, 4.19+ */
> +	if (CONTEXT->sregs) {
> +		regs = CONTEXT->sregs;
> +	} else {
> +		regs = (CONTEXT->user_mode_p ? CONTEXT->uregs : CONTEXT->kregs);
> +	}
> +
> +	if (!regs) {
> +		CONTEXT->last_error = "No registers available in this context";
> +		return;
> +	}
> +
> +	memcpy(&val, ((char *)regs) + ((STAP_ARG_argnum - 1) * sizeof (long)),
> +	    sizeof (long));
> +
> +	STAP_RETVALUE = val;

> Could the memcpy be eliminated and STAP_RETVALUE statements be simplified to something like:
>
> STAP_RETVALUE = regs->regs[STAP_ARG_argnum-1];
>
>-Will

Probably. I was following existing practice and used memcpy. See also _stp_get_register_by_offset() for various architectures: they all use memcpy. I am not sure why. Alignment paranoia?

Frank