lkml.org 
[lkml]   [2017]   [Jun]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
On Thu, Jun 22, 2017 at 5:21 AM, Thomas Gleixner <tglx@linutronix.de> wrote:
> On Wed, 21 Jun 2017, Andy Lutomirski wrote:
>> On Wed, Jun 21, 2017 at 6:38 AM, Thomas Gleixner <tglx@linutronix.de> wrote:
>> > That requires a conditional branch
>> >
>> > if (asid >= NR_DYNAMIC_ASIDS) {
>> > asid = 0;
>> > ....
>> > }
>> >
>> > The question is whether 4 IDs would be sufficient which trades the branch
>> > for a mask operation. Or you go for 8 and spend another cache line.
>>
>> Interesting. I'm inclined to either leave it at 6 or reduce it to 4
>> for now and to optimize later.
>
> :)
>
>> > Hmm. So this loop needs to be taken unconditionally even if the task stays
>> > on the same CPU. And of course the number of dynamic IDs has to be short in
>> > order to makes this loop suck performance wise.
>> >
>> > Something like the completely disfunctional below might be worthwhile to
>> > explore. At least arch/x86/mm/ compiles :)
>> >
>> > It gets rid of the loop search and lifts the limit of dynamic ids by
>> > trading it with a percpu variable in mm_context_t.
>>
>> That would work, but it would take a lot more memory on large systems
>> with lots of processes, and I'd also be concerned that we might run
>> out of dynamic percpu space.
>
> Yeah, did not think about the dynamic percpu space.
>
>> How about a different idea: make the percpu data structure look like a
>> 4-way set associative cache. The ctxs array could be, say, 1024
>> entries long without using crazy amounts of memory. We'd divide it
>> into 256 buckets, so you'd index it like ctxs[4*bucket + slot]. For
>> each mm, we choose a random bucket (from 0 through 256), and then we'd
>> just loop over the four slots in the bucket in choose_asid(). This
>> would require very slightly more arithmetic (I'd guess only one or two
>> cycles, though) but, critically, wouldn't touch any more cachelines.
>>
>> The downside of both of these approaches over the one in this patch is
>> that the change that the percpu cacheline we need is not in the cache
>> is quite a bit higher since it's potentially a different cacheline for
>> each mm. It would probably still be a win because avoiding the flush
>> is really quite valuable.
>>
>> What do you think? The added code would be tiny.
>
> That might be worth a try.
>
> Now one other optimization which should be trivial to add is to keep the 4
> asid context entries in cpu_tlbstate and cache the last asid in thread
> info. If that's still valid then use it otherwise unconditionally get a new
> one. That avoids the whole loop machinery and thread info is cache hot in
> the context switch anyway. Delta patch on top of your version below.

I'm not sure I understand. If an mm has ASID 0 on CPU 0 and ASID 1 on
CPU 1 and a thread in that mm bounces back and forth between those
CPUs, won't your patch cause it to flush every time?

--Andy

\
 
 \ /
  Last update: 2017-06-22 20:13    [W:0.363 / U:0.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site