On Wed, Sep 01, 2021 at 04:05:40PM +0200, Michal Hocko wrote:
[SNIP]
>
> This looks better than the previous attempt. It would be still better to
> solve this at the page allocator layer. The slowpath is already doing
> this for the nodemask. E.g.
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index eeb3a9cb36bb..a3193134540d 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -4929,6 +4929,17 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> if (!ac->preferred_zoneref->zone)
> goto nopage;
>
> + /*
> + * Check for insane configurations where the cpuset doesn't contain any suitable
> + * zone to satisfy the request - e.g. kernel allocations from MOVABLE nodes only
> + */
> + if (cpusets_enabled() && (gfp_mask & __GFP_HARDWALL)) {
> + struct zoneref *z = first_zones_zonelist(ac->zonelist, ac->highest_zoneidx,
> + &cpuset_current_mems_allowed);
> + if (!z->zone)
> + goto nopage;
> + }
> +
> if (alloc_flags & ALLOC_KSWAPD)
> wake_all_kswapds(order, gfp_mask, ac);
Thanks for the suggestion! It dose bail out early skipping the kswapd,
direct reclaim and compaction.
I also looked at prepare_alloc_pages() which does some cpuset check
and zone initialization, but I'd better leave it alone as it's in a
real hot path, while here is in slowpath anyway.
Will run some page fault benchmark cases with this patch.
Thanks,
Feng
> if this is seen as an additional overhead for an insane configuration
> then we can add insane_cpusets_enabled() which would be a static branch
> enabled when somebody actually tries to configure movable only cpusets
> or potentially other dubious usage.
> --
> Michal Hocko
> SUSE Labs
