On 06/04/2023 10:58 pm, Peter Zijlstra wrote:
> On Mon, Apr 03, 2023 at 10:06:01PM +0800, Lai Jiangshan wrote:
>> +/* Replicate the interrupted atomic-IST-entry's CLEAR_REGS macro. */
>> +static __always_inline void replicate_clear_regs(struct pt_regs *target)
>> +{
>> + target->di = 0;
>> + target->si = 0;
>> + target->dx = 0;
>> + target->cx = 0;
>> + target->ax = 0;
>> + target->r8 = 0;
>> + target->r9 = 0;
>> + target->r10 = 0;
>> + target->r11 = 0;
>> + target->bx = 0;
>> + target->bp = 0;
>> + target->r12 = 0;
>> + target->r13 = 0;
>> + target->r14 = 0;
>> + target->r15 = 0;
>> +}
> I think there's compilers smart enough to see through your attempts at
> avoiding mem{set,cpy}() there
Indeed. It took a little bit of convincing (needed 4 extra registers to
zero), but https://godbolt.org/z/7rvb8db66
Including Peter's other observation about speculation safety, you
basically can't have any C at all before passing IBRS/UNRET/whatever
else comes along in the future.
Otherwise, the compiler will make you wish you'd written it in asm the
first time around.
~Andrew
