Re: [PATCH v2] mm/page_alloc: use existing highatomic reserves on the buddy fastpath

[Johannes Weiner <hannes@cmpxchg.org>]

On Wed, Jun 17, 2026 at 04:49:58PM -0700, JP Kobryn wrote:
> ALLOC_HIGHATOMIC currently provides both access to MIGRATE_HIGHATOMIC free
> pages and permission to create new highatomic pageblock reserves. This
> makes it unsuitable for the fastpath.
> 
> However, the fastpath can reach rmqueue_buddy() while MIGRATE_HIGHATOMIC
> reserves have free pages available. In this situation, the allocation can
> fall back to other migratetypes without trying those reserves first.
> 
> Allow high-priority non-blocking allocations above order-0 and up to the
> costly order to use existing MIGRATE_HIGHATOMIC reserves on the buddy
> fastpath. Change the semantics of ALLOC_HIGHATOMIC so that it only allows
> access to the reserves without permission to grow them. Add a new flag
> ALLOC_HIGHATOMIC_RESERVE that specifically allows growing the reserves.
> 
> A UDP receive workload was run with free MIGRATE_HIGHATOMIC pageblocks
> available in the target zone. Before this patch, the workload did not
> consume these blocks. With this patch, eligible order-1 allocations
> reaching the buddy path consumed existing MIGRATE_HIGHATOMIC pageblocks,
> with no highatomic misses observed. The workload did not grow highatomic
> reserves and NAPI page-frag allocations remained healthy with no failures
> or order-0 fallbacks.

Thanks for digging deeper into this! That's a great find.

> Signed-off-by: JP Kobryn <jp.kobryn@linux.dev>
> Reviewed-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
> ---
> v2:
>   - decouple use semantics from ALLOC_HIGHATOMIC_RESERVE
>   - update changelog to reflect above change and reword test paragraph
>   - adjust comment in PCP path
>   - rebase onto Linus' tree ~v7.2-rc1
> 
> v1: https://lore.kernel.org/linux-mm/20260616191420.52556-1-jp.kobryn@linux.dev/
> 
>  mm/internal.h   |  1 +
>  mm/page_alloc.c | 30 +++++++++++++++++++++++++-----
>  2 files changed, 26 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/internal.h b/mm/internal.h
> index 5a2ddcf68e0b..6700659615e8 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -1478,6 +1478,7 @@ unsigned int reclaim_clean_pages_from_list(struct zone *zone,
>  #define ALLOC_HIGHATOMIC	0x200 /* Allows access to MIGRATE_HIGHATOMIC */
>  #define ALLOC_TRYLOCK		0x400 /* Only use spin_trylock in allocation path */
>  #define ALLOC_KSWAPD		0x800 /* allow waking of kswapd, __GFP_KSWAPD_RECLAIM set */
> +#define ALLOC_HIGHATOMIC_RESERVE	0x1000 /* Allows growing MIGRATE_HIGHATOMIC reserves */
>  
>  /* Flags that allow allocations below the min watermark. */
>  #define ALLOC_RESERVES (ALLOC_NON_BLOCK|ALLOC_MIN_RESERVE|ALLOC_HIGHATOMIC|ALLOC_OOM)
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index d49c254174da..ed919e2ac99a 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3238,7 +3238,7 @@ struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
>  	 * If this is a high-order atomic allocation then check
>  	 * if the pageblock should be reserved for the future
>  	 */
> -	if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
> +	if (unlikely(alloc_flags & ALLOC_HIGHATOMIC_RESERVE))
>  		reserve_highatomic_pageblock(page, order, zone);

You could check ALLOC_WMARK_MIN to determine the slowpath. This way
you wouldn't need another alloc flag:

	/* Slowpath (precarious) high-atomic allocation. Maybe reserve block */
	if (unlikely((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_WMARK_MIN)) == (ALLOC_HIGHATOMIC|ALLOC_WMARK_MIN)))
		reserve_highatomic_pageblock(page, order, zone);

[ we really ought to generalize gfp_has_flags() ... ]

>  	__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
> @@ -3320,8 +3320,9 @@ struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
>  			 *
>  			 * Instead, direct it towards the reserves by
>  			 * returning NULL, which will make the caller fall
> -			 * back to rmqueue_buddy. This will try to use the
> -			 * reserves first and grow them if needed.
> +			 * back to rmqueue_buddy. There it will try to use
> +			 * the reserves first and grow them if needed and
> +			 * permitted by the ALLOC_HIGHATOMIC_RESERVE flag.
>  			 */
>  			if (alloc_flags & ALLOC_HIGHATOMIC)
>  				return NULL;
> @@ -3768,6 +3769,24 @@ alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
>  	return alloc_flags;
>  }
>  
> +/*
> + * Let high-priority non-blocking allocations above order-0 and up
> + * to the costly order try to use existing MIGRATE_HIGHATOMIC
> + * reserves on the fastpath.
> + */
> +static inline unsigned int
> +alloc_flags_highatomic_fastpath(gfp_t gfp_mask, unsigned int order)
> +{
> +	if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
> +		return 0;

There seems to be a mismatch between this and gfp_to_alloc_flags()
(slowpath), where slowpath is still allowed to tap highatomic reserves
for costly orders. Is that on purpose?

> +	if (!(gfp_mask & __GFP_HIGH))
> +		return 0;
> +	if (gfp_mask & (__GFP_DIRECT_RECLAIM | __GFP_NOMEMALLOC))
> +		return 0;

This duplicates gfp_to_alloc_flags() logic which seems fragile. How
about the below:

> +
> +	return ALLOC_HIGHATOMIC;
> +}
> +
>  /* Must be called after current_gfp_context() which can change gfp_mask */
>  static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
>  						  unsigned int alloc_flags)
> @@ -4495,7 +4514,7 @@ gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
>  			alloc_flags |= ALLOC_NON_BLOCK;
>  
>  			if (order > 0 && (alloc_flags & ALLOC_MIN_RESERVE))
> -				alloc_flags |= ALLOC_HIGHATOMIC;
> +				alloc_flags |= (ALLOC_HIGHATOMIC | ALLOC_HIGHATOMIC_RESERVE);
>  		}
>  
>  		/*
> @@ -5215,7 +5234,8 @@ struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
>  	 * Forbid the first pass from falling back to types that fragment
>  	 * memory until all local zones are considered.
>  	 */
> -	alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp);
> +	alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp) |
> +			alloc_flags_highatomic_fastpath(alloc_gfp, order);

	alloc_flags |= gfp_to_alloc_flags(gfp, order) & ALLOC_HIGHATOMIC;

Or factor gfp_to_alloc_flags_nonblocking() from gfp_to_alloc_flags()
and reuse that here, to save a few cycles in the fast path.

>  	/* First allocation attempt */
>  	page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
> -- 
> 2.54.0