From: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
Date: Sat, 22 Feb 2020 00:41:57 +0000
> On Fri, Feb 21, 2020 at 11:22:44AM -0800, David Miller wrote:
>> From: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
>> Date: Fri, 21 Feb 2020 18:59:03 +0000
>>
>> > Again, a couple of copy_regset_to_user(), but there's an additional
>> > twist - GETREGSET of 32bit task on sparc64 will use access_process_vm()
>> > when trying to fetch L0..L7/I0..I7 of other task, using copy_from_user()
>> > only when the target is equal to current. For sparc32 this is not
>> > true - it's always copy_from_user() there, so the values it reports
>> > for those registers have nothing to do with the target process. That
>> > part smells like a bug; by the time GETREGSET had been introduced
>> > sparc32 was not getting much attention, GETREGS worked just fine
>> > (not reporting L*/I* anyway) and for coredump it was accessing the
>> > caller's memory. Not sure if anyone cares at that point...
>>
>> That's definitely a bug and sparc64 is doing it correctly.
>
> OK... What does the comment in
> case PTRACE_GETREGS64:
> ret = copy_regset_to_user(child, view, REGSET_GENERAL,
> 1 * sizeof(u64),
> 15 * sizeof(u64),
> &pregs->u_regs[0]);
> if (!ret) {
> /* XXX doesn't handle 'y' register correctly XXX */
> ret = copy_regset_to_user(child, view, REGSET_GENERAL,
> 32 * sizeof(u64),
> 4 * sizeof(u64),
> &pregs->tstate);
> }
> break;
> refer to? The fact that you end up with 0 in pregs->y and Y in pregs->magic?
> In that case it's probably too late to do anything about that...
Yes, that's exactly what it's talking about since we have:
unsigned int y;
unsigned int magic;
and we're doing a 64-bit value copy.
