Re: [PATCH 0/3] TLB flush multiple pages per IPI v5

[Vlastimil Babka]

On 25.6.2015 20:36, Linus Torvalds wrote:
> 
> On Jun 25, 2015 04:48, "Ingo Molnar" <mingo@xxxxxxxxxx
> <mailto:mingo@xxxxxxxxxx>> wrote:
>>
>>  - 1x, 2x, 3x, 4x means up to 4 adjacent 4K vmalloc()-ed pages are accessed, the
>>    first byte in each
> 
> So that test is a bit unfair. From previous timing of Intel TLB fills, I can
> tell you that Intel is particularly good at doing adjacent entries.
> 
> That's independent of the fact that page tables have very good locality (if they
> are the radix tree type - the hashed page tables that ppc uses are shit). So
> when filling adjacent entries, you take the cache misses for the page tables
> only once, but even aside from that, Intel send to do particularly well at the
> "next page" TLB fill case

AFAIK that's because they also cache partial translations, so if the first 3
levels are the same (as they mostly are for the "next page" scenario) it will
only have to look at the last level of pages tables. AMD does that too.

> Now, I think that's a reasonably common case, and I'm not saying that it's
> unfair to compare for that reason, but it does highlight the good case for TLB
> walking.
> 
> So I would suggest you highlight the bad case too: use invlpg to invalidate
> *one* TLB entry, and then walk four non-adjacent entries. And compare *that* to
> the full TLB flush.
> 
> Now, I happen to still believe in the full flush, but let's not pick benchmarks
> that might not show the advantages of the finer granularity.
> 
>         Linus
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/