Re: [patch v2 4/4] ptrace: Add support for generic PTRACE_GETREGSET/PTRACE_SETREGSET

["H.J. Lu" <hjl.tools@gmail.com>]

On Tue, Feb 9, 2010 at 5:52 PM, Roland McGrath <roland@redhat.com> wrote:
>> 'addr' parameter for the ptrace system call encode the REGSET type and
>> the buffer length. 'data' parameter points to the user buffer.
>>
>>       ptrace(PTRACE_GETREGSET/PTRACE_SETREGSET, pid,
>>              (NT_TYPE << 20) | buf_length, buf);
>
> IMHO this bit swizzling is a non-starter.  The NT_* codes can use a full 32
> bits.  NT_PRXFPREG uses 31 bits.  The comments about ignoring the high bits
> for this as a special case just seem insane to me.  Pass a full 32 bit word
> for NT_* and a full word for the buffer size.  What's so terrible about
> just having an obvious and comprehensible API?
>
> IMHO if you insist on an insane bit swizzling approach, you should mix the
> buffer size bits into the request code, leaving the "addr" argument free
> for the unmolested NT_* code.  There is just no earthly reason that we
> should suddenly impose a new and arcane constraint on what NT_* values can
> be, even if there is no particular reason for future values not to be small.
>
>> +int generic_ptrace_regset(struct task_struct *child, long request, long addr,
>> +                       long data);
>
> There is no need for a global function for this.  It should all be static
> in kernel/ptrace.c, called only by ptrace_request()/compat_ptrace_request().
>

Won't it be called by ptrace emulation in utrace?


-- 
H.J.