Re: [PATCH 1/1] mm: slub: Introduce one knob to control the track of slub object

[Zhenhua Huang <quic_zhenhuah@quicinc.com>]

On 2025/7/23 19:38, Kent Overstreet wrote:
> On Wed, Jul 23, 2025 at 06:19:45PM +0900, Harry Yoo wrote:
>> The subject is a bit misleading. I think it should be something like
>> "alloc_tag: add an option to disable slab object accounting".
>>
>> On Wed, Jul 23, 2025 at 04:03:28PM +0800, Zhenhua Huang wrote:
>>> Mem profiling feature tracks both "alloc_slab_page"(page level) and slub
>>> object level allocations. To track object level allocations,
>>> slabobj_ext consumes 16 bytes per object for profiling slub object if
>>> CONFIG_MEMCG is set.
>>> Based on the data I've collected, this overhead accounts for approximately
>>> 5.7% of slub memory usage — a considerable cost.
>>> w/ noslub  slub_debug=-
>>> Slab:              87520 kB
>>> w/o noslub slub_debug=-
>>> Slab:              92812 kB
>>
>> Yes, the cost is not small and I hate that we have to pay 16 bytes of
>> memory overhead for each slab object when both memcg and memory profiling
>> are enabled.
> 
> I believe we did something about this for page_obj_ext; the exact
> pointer compression scheme we went with escapes me at the moment.
> 

Hi Kent,

I recall that it used page flags for compression — not actual pointers, 
but rather *idx*. Since every page has a corresponding struct page, that 
makes this approach feasible. However, it seems this assumption doesn't 
hold for slab objects.

> We did it for page and not slab because page_obj_ext is a large fixed
> size overhead and the page allocator is slower anyways, but it's
> conceivable we could do the same for slub if the memory overhead vs. cpu
> overhead tradeoff is worth it.
> 
> And - pointer compression is a valuable technique in general; coming up
> with some fast general purpose code (perhaps involving virtual mappings,
> we're not so limited on virtual address space as we used to be) might be
> worth someone's time exploring.
>