On 07/15/2018 04:04 AM, Kees Cook wrote:
> On Wed, Jul 4, 2018 at 10:47 AM, Vlastimil Babka <vbabka@xxxxxxx> wrote:
>> On 07/04/2018 06:52 PM, Kees Cook wrote:
>>> This produces less efficient code in the general case, and I'd like to
>>> keep the general case (hardening enabled) as fast as possible.
>>
>> How specifically is the code less efficient? It should be always a
>> static key check (no-op thanks to the code patching involved) and a
>> function call in the "hardening enabled" case, just in different order.
>> And in either case compiled out if it's a constant.
>
> My understanding from reading the jump label comments[1] is that on
> order produces:
>
> NOP
> do normal thing
> label1:
> do rest of function
> RET
> label2:
> do exceptional thing
> jump label1
>
> where "NOP" is changed to "JMP label2" when toggled, and the other is:
>
> JMP label1
> do exceptional thing
> JMP label2
> label1:
> do normal thing
> label2:
> do rest of function
> RET
>
> where "JMP label1" is changed to NOP when toggled. (i.e. does the
> default do NOP, thing, function, or does the default to JMP, thing,
> JMP, function)
My mistake, sorry. I didn't mean to change likely() to unlikely(). Also
I didn't negate the condition. The correct code is:
if (!__builtin_constant_p(n) &&
!static_branch_unlikely(&bypass_usercopy_checks))
__check_object_size(ptr, n, to_user);
I've test-compiled it, did objdump -d and checked few call sites and they
seem to be preceded just y NOP, so it's the first case you mentioned above,
as expected - calling __check_object_size() is treated as the "normal thing".
