On Wed, Jul 24, 2024 at 08:21:01PM GMT, Roman Gushchin wrote:
> Struct page_counter is used by memory and hugetlb cgroups to track
> the resource usage. Memory cgroups are using up to four different
> counters (on cgroup v1). Hugetlb cgroups are using 2 * HUGE_MAX_HSTATE
> counters.
>
> Every page_counter structure (except the one belonging to the root
> cgroup) contains a pointer to the corresponding parent structure.
> It's obviously wasteful because for each cgroup there are several
> similar chains of page_counter structures. It's not particularly
> cache-effective too.
>
> Also every page_counter structure contains fields required for
> tracking the usage of the memory protection, while it's used only
> for the main ("memory") counter and only by memory cgroups.
>
> To address these issues let's make page_counter structures to support
> N counters and use a single page_counter structure for a memory cgroup
> and 2 page_counter structures for a hugetlb cgroup (to track the
> actual usage and reservations).
>
> N is determined at the compile time depending on the configuration and
> HUGE_MAX_HSTATE value. HUGE_MAX_HSTATE is relatively small on all
> existing platforms (max is 5 on sparc), so it's not too far from 4
> memcg memory types.
>
> Memory min/low functionality is supported only for the first counter.
>
> Memory savings (on x86_64):
>
> |-----------------------------+----------------------------+----------------------------|
> | | memcg | hugetlb |
> |-----------------------------+----------------------------+----------------------------|
> | original, CONFIG_MEMCG_V1=n | size: 1600, cachelines: 25 | size: 1280, cachelines: 20 |
> | patched, CONFIG_MEMCG_V1=n | size: 1472, cachelines: 23 | size: 1024, cachelines: 16 |
> |-----------------------------+----------------------------+----------------------------|
> | original, CONFIG_MEMCG_V1=y | size: 2176, cachelines: 34 | size: 1280, cachelines: 20 |
> | patched, CONFIG_MEMCG_V1=y | size: 1664, cachelines: 26 | size: 1152, cachelines: 18 |
> |-----------------------------+----------------------------+----------------------------|
>
> So it's 10-35% reduction in the number of cache lines depending on
> the kernel configuration.
>
> Signed-off-by: Roman Gushchin <roman.gushchin@xxxxxxxxx>
Acked-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
