Re: [RFC PATCH 0/7] Reviving the slab destructor to tackle the percpu allocator scalability problem

[Harry Yoo <harry.yoo@oracle.com>]

On Thu, Apr 24, 2025 at 08:47:05AM -1000, Tejun Heo wrote:
> On Thu, Apr 24, 2025 at 05:07:48PM +0900, Harry Yoo wrote:
> ...
> > Eighteen years later, Mateusz Guzik proposed [1] re-introducing a slab
> > constructor/destructor pair to mitigate the global serialization point
> > (pcpu_alloc_mutex) that occurs when each slab object allocates and frees
> > percpu memory during its lifetime.
> > 
> > Consider mm_struct: it allocates two percpu regions (mm_cid and rss_stat),
> > so each allocate–free cycle requires two expensive acquire/release on
> > that mutex.
> 
> When percpu allocator was first introduced, the use cases were a lot more
> limited, so the single mutex and expensive alloc/free paths weren't a
> problem. We keep using percpu memory for more and more things, and I always
> thought we'd eventually need a more sophisticated allocator something with
> object caching.

Yeah, when you first write an allocator, it's an overkill to make it
too scalable. But over time, as with other allocators, more users show up
that require a more sophisticated allocator.

> I don't exactly know what that should look like but maybe a
> simplified version of sl*b serving power of two sizes should do or maybe it
> needs to be smaller and more adaptive. We'd need to collect some data to
> decide which way to go.

I'm not sure what kind of data we need — maybe allocation size distributions,
or more profiling data on workloads that contend on percpu allocator's locks?

> Improving percpu allocator in general is obviously a heavier lift but that
> may be a better long-term direction.

Yeah, if that's doable. But until then I think it still makes sense to cache
it within slab objects, ...or probably even after improving the percpu
allocator? It's a still churn that's incurred during each object's lifetime
regardless. (Need some data to see if justifiable)

And, as Mateusz explained, the percpu allocator isn’t the only motivation
for the ctor/dtor pair. Other expensive serializations like pgd_lock and
percpu_counters_lock are other motivations to do this.

-- 
Cheers,
Harry / Hyeonggon