Re: [PATCH v2 3/3] x86/pti/64: Remove the SYSCALL64 entry trampoline

[Andy Lutomirski]

On Fri, Sep 7, 2018 at 5:04 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Sep 7, 2018 at 12:54 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>>
>> >  - We execute from an extra page and read from another extra page
>> > during the syscall.  (The latter is because we need to use a relative
>> > addressing mode to find sp1 -- it's the same *cacheline* we'd use
>> > anyway, but we're accessing it using an alias, so it's an extra TLB
>> > entry.)
>>
>> Ok, but is this really an issue with PTI?
>
> I'd expect it to be *more* of an issue with PTI, since you're already
> wasting TLB entries due to the whole "two different page tables".
>
> Sure, address space ID's save you from reloading them all the time,
> but don't help with capacity.
>
> But yeah, in the sense of "with PTI, all kernel entries are slow
> anyway, so none of this matters" is probably correct in a very real
> sense.
>
> That said, the real reason I like Andy's patch series is that I think
> it's simpler than the alternatives (including the current setup). No
> subtle mappings, no nothing. It removes a lot more lines than it adds,
> and half the lines that it *does* add are comments.
>
> Virtual mapping tricks may be cool, but in the end, not having to use
> them is better still, I think.
>

If (and this is a *big* if) all the percpu data is within 2GB of the
entry text, then we could avoid this extra TLB reference by accessing
it directly instead of using an alias.

I suppose the summary is that the retpoline-free trampoline variant is
even more complicated than the code I'm removing in this series, and
that it would be at best a teeny tiny win.  Once all the Spectre dust
settles and we get fixed CPUs, we could consider re-optimizing this.