On 2021-07-29 16:17:26 [+0200], Vlastimil Babka wrote:

I forgot to point that out in the cover letter, but with v3 this change to
raw_spinlock_t is AFAICS no longer possible (at least with
CONFIG_SLUB_CPU_PARTIAL) because in put_cpu_partial() we now take the local_lock
and it can be called from get_partial_node() which takes the list_lock.

I saw increased latency numbers with CONFIG_SLUB_CPU_PARTIAL before it
got disabled for other reasons so I'm not too sad if it remains
disabled.

Hm but SLUB should never call the page allocator from under list_lock in my series?

oh yes. I run into CPU_PARTIAL instead. Sorry for the confusion. So
without PARTIAL it says "24,643 CPUs utilized" and no warnings :)

Sebastian