Re: [linus:master] [mm] 088b8aa537: vm-scalability.throughput -6.5% regression

[David Hildenbrand <david@redhat.com>]

On 21.11.22 04:03, kernel test robot wrote:
> Greeting,
> 
> FYI, we noticed a -6.5% regression of vm-scalability.throughput due to commit:
> 
> commit: 088b8aa537c2c767765f1c19b555f21ffe555786 ("mm: fix PageAnonExclusive clearing racing with concurrent RCU GUP-fast")
> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
> 
> in testcase: vm-scalability
> on test machine: 88 threads 2 sockets Intel(R) Xeon(R) Gold 6238M CPU @ 2.10GHz (Cascade Lake) with 128G memory
> with following parameters:
> 
> 	thp_enabled: never
> 	thp_defrag: never
> 	nr_task: 1
> 	nr_pmem: 2
> 	priority: 1
> 	test: swap-w-seq
> 	cpufreq_governor: performance
> 
> test-description: The motivation behind this suite is to exercise functions and regions of the mm/ of the Linux kernel which are of interest to us.
> test-url: https://git.kernel.org/cgit/linux/kernel/git/wfg/vm-scalability.git/
> 

Yes, page_try_share_anon_rmap() might now be a bit more expensive now, 
turning try_to_unmap_one() a bit more expensive. However, that patch 
also changes the unconditional TLB flush into a conditional TLB flush, 
so results might vary heavily between machines/architectures.

smp_mb__after_atomic() is a NOP on x86, so the smp_mb() before the 
page_maybe_dma_pinned() check would have to be responsible.

While there might certainly be ways for optimizing that further (e.g., 
if the ptep_get_and_clear() already implies a smp_mb()), the facts that:

(1) It's a swap micro-benchmark
(2) We have 3% stddev

Don't make me get active now ;)

-- 
Thanks,

David / dhildenb