On 07/13/2016 08:21 AM, Michal Hocko wrote:
>> > This adds a tiny amount of overhead to all pte_none() checks.
>> > I doubt we'll be able to measure it anywhere.
> It would be better to introduce the overhead only for the affected
> cpu models but I guess this is also acceptable. Would it be too
> complicated to use alternatives for that?
The patch as it stands ends up doing a one-instruction change in
pte_none(). It goes from
64c8: 48 85 ff test %rdi,%rdi
to
64a8: 48 f7 c7 9f ff ff ff test $0xffffffffffffff9f,%rdi
So it essentially eats 4 bytes of icache more than it did before. But,
it's the same number of instructions, and I can't imagine that the CPU
will have any more trouble with a test against an immediate than a test
against 0.
We could theoretically do alternatives for this, but we would at *best*
end up with 4 bytes of noops. So, unless the processor likes decoding 4
noops better than 4 bytes of immediate as part of an instruction, we'll
not win anything. *Plus* the ugliness of the assembly that we'll need
to have the compiler guarantee that the PTE ends up in %rdi.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>