Re: [RFC PATCH 0/4] mm/vmscan: honour node reclaim limits per type

On Fri 21-08-26 16:17:37, Ridong Chen wrote:
> From: Ridong Chen <[email protected]>
> 
> min_unmapped_pages and min_slab_pages are documented as per-type limits,
> but node reclaim treats them as one combined gate: once either is
> exceeded, shrink_node() reclaims slab, file and anon together and pushes
> the other type below its limit. Per-node proactive reclaim reuses the
> same gate and fares worse -- with page cache and slab both under their
> limits it reclaims nothing and returns -EAGAIN on an anon-heavy node [1].
> 
> This series gates each type separately via two scan_control flags
> (skip_slab_reclaim, skip_file_reclaim) set only on the node reclaim path,
> drops the combined gate, and extends node_reclaim()'s early bail to check
> anon. The flags default to zero, so kswapd, direct, memcg and drop_caches
> are unaffected.

You are explaining what but missing the most important part _Why_ do we
need to have this addressed? Is this just addressing Sashiko review
refernced below? Is there any real usecase where the current behavior
matters?

> Tested on QEMU (x86_64, 2 NUMA nodes), A/B kernels differing only in this
> series.
> 
> Proactive reclaim (echo to node/reclaim) on an anon-heavy node, file and
> slab under their limits:
> 
>   metric              before   after
>   -----------------   ------   -----------
>   pages reclaimed     0 MiB    254 MiB anon
>   return value        -EAGAIN  0
> 
> Node reclaim (zone_reclaim_mode) with one type under its limit -- the type
> under its limit must be left alone:
> 
>   type under limit    before          after
>   -----------------   -------------   --------
>   slab                645 scans       0 scans
>   page cache          83 MiB scanned  0 MiB
> 
> [1] https://sashiko.dev/#/patchset/[email protected]
> 
> Ridong Chen (4):
>   mm/vmscan: only reclaim slab in node reclaim when over min_slab_pages
>   mm/vmscan: only reclaim file pages in node reclaim when over
>     min_unmapped_pages
>   mm/vmscan: drop the combined limit gate in __node_reclaim()
>   mm/vmscan: do not skip node reclaim when only anon is reclaimable
> 
>  mm/vmscan.c | 89 ++++++++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 67 insertions(+), 22 deletions(-)
> 
> -- 
> 2.34.1

-- 
Michal Hocko
SUSE Labs