Re: [PATCH 13/22] mm/hugetlb: restrict pool growth to common memory
From: Richard Cheng
Date: Wed Nov 18 2026 - 02:32:31 EST
On Sat, Nov 14, 2026 at 02:09:50AM +0800, Gregory Price (Meta) wrote:
> Fresh HugeTLB pool folios come from ordinary allocations, which only
> N_MEMORY_COMMON nodes serve.
>
> Restrict HugeTLB allocations and pool-growth to N_MEMORY_COMMON nodes.
>
> Keep the original allowed mask for pool shrinking and demotion so
> existing pool folios can still be found on any online-memory node.
>
> The result of this change is that hugepage allocations will fail on
> nodes without N_MEMORY_COMMON. This restriction can be loosened later
> when/if it grows a user.
>
> All memory is common at this point, so behavior is unchanged.
>
> Signed-off-by: Gregory Price (Meta) <gourry@xxxxxxxxxx>
> ---
> mm/hugetlb.c | 44 ++++++++++++++++++++------------------------
> 1 file changed, 20 insertions(+), 24 deletions(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index cb2de24affdc4..cc1468f68901b 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -3452,13 +3452,13 @@ static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
> gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
>
> folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
> - &node_states[N_MEMORY], NULL);
> + &node_states[N_MEMORY_COMMON], NULL);
> if (!folio && !list_empty(&folio_list) &&
> hugetlb_vmemmap_optimizable(h)) {
> prep_and_add_allocated_folios(h, &folio_list);
> INIT_LIST_HEAD(&folio_list);
> folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
> - &node_states[N_MEMORY], NULL);
> + &node_states[N_MEMORY_COMMON], NULL);
> }
> if (!folio)
> break;
> @@ -3528,7 +3528,7 @@ static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned l
> prep_and_add_allocated_folios(h, &folio_list);
> INIT_LIST_HEAD(&folio_list);
> }
> - folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
> + folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY_COMMON],
> &node_alloc_noretry, &next_node);
> if (!folio)
> break;
> @@ -3672,13 +3672,12 @@ static void __init hugetlb_init_hstates(void)
>
> for_each_hstate(h) {
> /*
> - * Always reset to first_memory_node here, even if
> - * next_nid_to_alloc was set before - we can't
> - * reference hugetlb_bootmem_nodes after init, and
> - * first_memory_node is right for all further allocations.
> + * Always reset the allocation cursor here, even if it was set
> + * before: hugetlb_bootmem_nodes cannot be referenced after init,
> + * and further pool allocations use common memory.
> */
> - h->next_nid_to_alloc = first_memory_node;
> - h->next_nid_to_free = first_memory_node;
> + h->next_nid_to_alloc = first_node_state(N_MEMORY_COMMON);
> + h->next_nid_to_free = first_node_state(N_MEMORY_COMMON);
>
> /* oversize hugepages were init'ed in early boot */
> if (!hstate_is_gigantic(h))
> @@ -3802,23 +3801,21 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> unsigned long allocated;
> struct folio *folio;
> LIST_HEAD(page_list);
> - NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
> + NODEMASK_ALLOC(struct nodemask_scratch, scratch, GFP_KERNEL);
>
> - /*
> - * Bit mask controlling how hard we retry per-node allocations.
> - * If we can not allocate the bit mask, do not attempt to allocate
> - * the requested huge pages.
> - */
> - if (node_alloc_noretry)
> - nodes_clear(*node_alloc_noretry);
> - else
> + /* Scratch masks for allocation eligibility and per-node retries. */
> + if (!scratch)
> return -ENOMEM;
> + nodes_clear(scratch->mask2);
>
> /*
> * resize_lock mutex prevents concurrent adjustments to number of
> * pages in hstate via the proc/sysfs interfaces.
> */
> mutex_lock(&h->resize_lock);
> +
> + /* Fresh pool pages come from the ordinary allocation domain. */
> + nodes_and(scratch->mask1, *nodes_allowed, node_states[N_MEMORY_COMMON]);
For a global request the caller passes &node_states[N_MEMORY], so before this patch
the grow loop read the live mask on every alloc_pool_huge_folio().
Now the mask is sampled once, before the loop.
The loop drops hugetlb_lock and can run for a long time. If a node gets its first
memoroy in that window, the old code would start using it. The new code won't, until
the next write to nr_hugepages.
So the pool can end up uneven, or smaller than requested if the other nodes are full.
I think the nodes_and() should be placed inside the loop to keep the old behavior ?
Best regards,
Richard Cheng.
> flush_free_hpage_work(h);
> spin_lock_irq(&hugetlb_lock);
>
> @@ -3855,7 +3852,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> if (count > persistent_huge_pages(h)) {
> spin_unlock_irq(&hugetlb_lock);
> mutex_unlock(&h->resize_lock);
> - NODEMASK_FREE(node_alloc_noretry);
> + NODEMASK_FREE(scratch);
> return -EINVAL;
> }
> /* Fall through to decrease pool */
> @@ -3873,7 +3870,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> * within all the constraints specified by the sysctls.
> */
> while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
> - if (!adjust_pool_surplus(h, nodes_allowed, -1))
> + if (!adjust_pool_surplus(h, &scratch->mask1, -1))
> break;
> }
>
> @@ -3889,9 +3886,8 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> /* yield cpu to avoid soft lockup */
> cond_resched();
>
> - folio = alloc_pool_huge_folio(h, nodes_allowed,
> - node_alloc_noretry,
> - &h->next_nid_to_alloc);
> + folio = alloc_pool_huge_folio(h, &scratch->mask1, &scratch->mask2,
> + &h->next_nid_to_alloc);
> if (!folio) {
> prep_and_add_allocated_folios(h, &page_list);
> spin_lock_irq(&hugetlb_lock);
> @@ -3978,7 +3974,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> spin_unlock_irq(&hugetlb_lock);
> mutex_unlock(&h->resize_lock);
>
> - NODEMASK_FREE(node_alloc_noretry);
> + NODEMASK_FREE(scratch);
>
> return 0;
> }
> --
> 2.53.1-Meta
>
>