Re: [PATCH v4 10/10] x86/mm: Try to preserve old TLB entries using PCID

[Andy Lutomirski <luto@kernel.org>]

On Wed, Jul 5, 2017 at 5:25 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> On Thu, Jun 29, 2017 at 08:53:22AM -0700, Andy Lutomirski wrote:
>> +static void choose_new_asid(struct mm_struct *next, u64 next_tlb_gen,
>> +                         u16 *new_asid, bool *need_flush)
>> +{
>> +     u16 asid;
>> +
>> +     if (!static_cpu_has(X86_FEATURE_PCID)) {
>> +             *new_asid = 0;
>> +             *need_flush = true;
>> +             return;
>> +     }
>> +
>> +     for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
>> +             if (this_cpu_read(cpu_tlbstate.ctxs[asid].ctx_id) !=
>> +                 next->context.ctx_id)
>> +                     continue;
>> +
>> +             *new_asid = asid;
>> +             *need_flush = (this_cpu_read(cpu_tlbstate.ctxs[asid].tlb_gen) <
>> +                            next_tlb_gen);
>> +             return;
>> +     }
>> +
>> +     /*
>> +      * We don't currently own an ASID slot on this CPU.
>> +      * Allocate a slot.
>> +      */
>> +     *new_asid = this_cpu_add_return(cpu_tlbstate.next_asid, 1) - 1;
>
> So this basically RR the ASID slots. Have you tried slightly more
> complex replacement policies like CLOCK ?

No, mainly because I'm lazy and because CLOCK requires scavenging a
bit.  (Which we can certainly do, but it will further complicate the
code.)  It could be worth playing with better replacement algorithms
as a followup, though.

I've also considered a slight elaboration of RR in which we make sure
not to reuse the most recent ASID slot, which would guarantee that, if
we switch from task A to B and back to A, we don't flush on the way
back to A.  (Currently, if B is not in the cache, there's a 1/6 chance
we'll flush on the way back.)