On Fri, Jan 22, 2021 at 02:05:47PM +0100, Jann Horn wrote:

On Thu, Jan 21, 2021 at 7:19 PM Vlastimil Babka [off-list ref] wrote:

quoted

On 1/21/21 11:01 AM, Christoph Lameter wrote:

quoted

On Thu, 21 Jan 2021, Bharata B Rao wrote:

quoted

quoted

The problem is that calculate_order() is called a number of times
before secondaries CPUs are booted and it returns 1 instead of 224.
This makes the use of num_online_cpus() irrelevant for those cases

After adding in my command line "slub_min_objects=36" which equals to
4 * (fls(num_online_cpus()) + 1) with a correct num_online_cpus == 224
, the regression diseapears:

9 iterations of hackbench -l 16000 -g 16: 3.201sec (+/- 0.90%)

I'm surprised that hackbench is that sensitive to slab performance, anyway. It's
supposed to be a scheduler benchmark? What exactly is going on?

Uuuh, I think powerpc doesn't have cmpxchg_double?

"vgrep cmpxchg_double arch/" just spits out arm64, s390 and x86? And
<https://liblfds.org/mediawiki/index.php?title=Article:CAS_and_LL/SC_Implementation_Details_by_Processor_family>
says under "POWERPC": "no DW LL/SC"

So powerpc is probably hitting the page-bitlock-based implementation
all the time for stuff like __slub_free()? Do you have detailed
profiling results from "perf top" or something like that?

I can check that, but the current patch was aimed at reducing
the page order of the slub caches so that they don't end up
consuming more memory on 64K systems.

(I actually have some WIP patches and a design document for getting
rid of cmpxchg_double in struct page that I hacked together in the
last couple days; I'm currently in the process of sending them over to
some other folks in the company who hopefully have cycles to
review/polish/benchmark them so that they can be upstreamed, assuming
that those folks think they're important enough. I don't have the
cycles for it...)

Sounds interesting, will keep a watch to see its effect on powerpc.

Regards,
Bharata.