On Oct 21, 2021, at 8:04 PM, Andrew Morton [off-list ref] wrote:

On Thu, 21 Oct 2021 05:21:07 -0700 Nadav Amit [off-list ref] wrote:

quoted

This patch-set is intended to remove unnecessary TLB flushes. It is
based on feedback from v1 and several bugs I found in v1 myself.

Basically, there are 3 optimizations in this patch-set:
1. Avoiding TLB flushes on change_huge_pmd() that are only needed to
  prevent the A/D bits from changing.
2. Use TLB batching infrastructure to batch flushes across VMAs and
  do better/fewer flushes.
3. Avoid TLB flushes on permission demotion.

Andrea asked for the aforementioned (2) to come after (3), but this
is not simple (specifically since change_prot_numa() needs the number
of pages affected).

[1/5] appears to be a significant fix which should probably be
backported into -stable kernels.  If you agree with this then I suggest
it be prepared as a standalone patch, separate from the other four
patches.  With a cc:stable.

There is no functionality bug in the kernel. The Knights Landing bug
was circumvented eventually by changing the swap entry structure so
the access/dirty bits would not overlap with the swap entry data.

And the remaining patches are a performance optimization.  Has any
attempt been made to quantify the benefits?

I included some data before [1]. In general the cost that is saved
is the cost of a TLB flush/shootdown.

I will modify my benchmark to test huge-pages (which were not
included in the previous patch-set) and send results later. I would
also try nodejs to see if there is a significant enough benefit.
Nodejs crashed before (hence the 3rd patch added here), as it
exec-protects/unprotects pages - I will see if the benefit shows in
the benchmarks.

[ The motivation behind the patches is to later introduce userfaultfd
writeprotectv interface, and for my use-case that is under 
development this proved to improve performance considerably. ]



[1] https://lore.kernel.org/linux-mm/DA49DBBB-FFEE-4ACC-BB6C-364D07533C5E@vmware.com/ (local)