Re: [PATCH] Docs/admin-guide/cgroup-v2: fix memory.stat doc details

[Doehyun Baek]

> ...but what do you mean by this?
> As I'm looking at the code in obj_cgroup_charge_zswap() and
> memcg_page_state_output_unit(), I'd say those are pages and the docs is
> thus alright.
>
> Thanks,
> Michal

Thanks for taking a look.

I agree that the counters are pages internally. I was talking about what
gets printed in memory.stat.

The internal updates are page-count based:

    mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
    if (size == PAGE_SIZE)
        mod_memcg_state(memcg, MEMCG_ZSWAP_INCOMP, 1);

However, both zswapped and zswap_incomp are memory_stats[] entries, so
memory.stat prints them through memcg_page_state_output(). Since
MEMCG_ZSWAP_INCOMP is not special-cased as a raw count, the stored page
count is multiplied by the default PAGE_SIZE unit and exported as bytes.

    unsigned long memcg_page_state_output(struct mem_cgroup *memcg, int item)
    {
        return memcg_page_state(memcg, item) *
        memcg_page_state_output_unit(item);
    }

Separately, this matches the existing documentation style for zswapped,
whose exported value is described as a memory amount:

    zswapped
        Amount of application memory swapped out to zswap.

Since zswap_incomp follows the same memory.stat output path, I think its
documentation should describe the exported value as a memory amount too.

I also boot-tested this in QEMU with the current tree and zswap enabled.
With incompressible pages pushed into zswap, memory.stat showed:

    zswap 87822336
    zswapped 87822336
    zswap_incomp 87822336

The zswap_incomp value there is byte-valued; it is not a plain page
count. It also matches zswapped in this all-incompressible case, which
is consistent with both being exported as memory amounts.

Best,
Doehyun Baek