By distributing both the allocation and the initialization tasks across
multiple threads, the initialization of 2M hugetlb will be faster,
thereby improving the boot speed.

Here are some test results:
      test case        no patch(ms)   patched(ms)   saved
 ------------------- -------------- ------------- --------
  256c2T(4 node) 2M           3336          1051   68.52%
  128c1T(2 node) 2M           1943           716   63.15%

Signed-off-by: Gang Li <ligang.bdlg@bytedance.com>
Tested-by: David Rientjes <rientjes@google.com>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 mm/hugetlb.c | 73 ++++++++++++++++++++++++++++++++++++++++------------
 1 file changed, 56 insertions(+), 17 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index d1ce1a52ad504..3ce957b3e350b 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -35,6 +35,7 @@
 #include <linux/delayacct.h>
 #include <linux/memory.h>
 #include <linux/mm_inline.h>
+#include <linux/padata.h>
 
 #include <asm/page.h>
 #include <asm/pgalloc.h>
@@ -3510,6 +3511,30 @@ static void __init hugetlb_hstate_alloc_pages_errcheck(unsigned long allocated,
 	}
 }
 
+static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned long end, void *arg)
+{
+	struct hstate *h = (struct hstate *)arg;
+	int i, num = end - start;
+	nodemask_t node_alloc_noretry;
+	LIST_HEAD(folio_list);
+	int next_node = first_online_node;
+
+	/* Bit mask controlling how hard we retry per-node allocations.*/
+	nodes_clear(node_alloc_noretry);
+
+	for (i = 0; i < num; ++i) {
+		struct folio *folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
+						&node_alloc_noretry, &next_node);
+		if (!folio)
+			break;
+
+		list_move(&folio->lru, &folio_list);
+		cond_resched();
+	}
+
+	prep_and_add_allocated_folios(h, &folio_list);
+}
+
 static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h)
 {
 	unsigned long i;
@@ -3525,26 +3550,40 @@ static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h)
 
 static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h)
 {
-	unsigned long i;
-	struct folio *folio;
-	LIST_HEAD(folio_list);
-	nodemask_t node_alloc_noretry;
-
-	/* Bit mask controlling how hard we retry per-node allocations.*/
-	nodes_clear(node_alloc_noretry);
+	struct padata_mt_job job = {
+		.fn_arg		= h,
+		.align		= 1,
+		.numa_aware	= true
+	};
 
-	for (i = 0; i < h->max_huge_pages; ++i) {
-		folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
-						&node_alloc_noretry);
-		if (!folio)
-			break;
-		list_add(&folio->lru, &folio_list);
-		cond_resched();
-	}
+	job.thread_fn	= hugetlb_pages_alloc_boot_node;
+	job.start	= 0;
+	job.size	= h->max_huge_pages;
 
-	prep_and_add_allocated_folios(h, &folio_list);
+	/*
+	 * job.max_threads is twice the num_node_state(N_MEMORY),
+	 *
+	 * Tests below indicate that a multiplier of 2 significantly improves
+	 * performance, and although larger values also provide improvements,
+	 * the gains are marginal.
+	 *
+	 * Therefore, choosing 2 as the multiplier strikes a good balance between
+	 * enhancing parallel processing capabilities and maintaining efficient
+	 * resource management.
+	 *
+	 * +------------+-------+-------+-------+-------+-------+
+	 * | multiplier |   1   |   2   |   3   |   4   |   5   |
+	 * +------------+-------+-------+-------+-------+-------+
+	 * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms |
+	 * | 2T   4node | 979ms | 679ms | 543ms | 489ms | 481ms |
+	 * | 50G  2node | 71ms  | 44ms  | 37ms  | 30ms  | 31ms  |
+	 * +------------+-------+-------+-------+-------+-------+
+	 */
+	job.max_threads	= num_node_state(N_MEMORY) * 2;
+	job.min_chunk	= h->max_huge_pages / num_node_state(N_MEMORY) / 2;
+	padata_do_multithreaded(&job);
 
-	return i;
+	return h->nr_huge_pages;
 }
 
 /*
-- 
2.20.1

Hi,

On Thu, Feb 22, 2024 at 10:04:20PM +0800, Gang Li wrote:
> By distributing both the allocation and the initialization tasks across
> multiple threads, the initialization of 2M hugetlb will be faster,
> thereby improving the boot speed.
> 
> Here are some test results:
>       test case        no patch(ms)   patched(ms)   saved
>  ------------------- -------------- ------------- --------
>   256c2T(4 node) 2M           3336          1051   68.52%
>   128c1T(2 node) 2M           1943           716   63.15%

Great improvement, and glad to see the multithreading is useful here.

>  static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h)
>  {
> -	unsigned long i;
> -	struct folio *folio;
> -	LIST_HEAD(folio_list);
> -	nodemask_t node_alloc_noretry;
> -
> -	/* Bit mask controlling how hard we retry per-node allocations.*/
> -	nodes_clear(node_alloc_noretry);
> +	struct padata_mt_job job = {
> +		.fn_arg		= h,
> +		.align		= 1,
> +		.numa_aware	= true
> +	};
>  
> -	for (i = 0; i < h->max_huge_pages; ++i) {
> -		folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
> -						&node_alloc_noretry);
> -		if (!folio)
> -			break;
> -		list_add(&folio->lru, &folio_list);
> -		cond_resched();
> -	}
> +	job.thread_fn	= hugetlb_pages_alloc_boot_node;
> +	job.start	= 0;
> +	job.size	= h->max_huge_pages;
>  
> -	prep_and_add_allocated_folios(h, &folio_list);
> +	/*
> +	 * job.max_threads is twice the num_node_state(N_MEMORY),
> +	 *
> +	 * Tests below indicate that a multiplier of 2 significantly improves
> +	 * performance, and although larger values also provide improvements,
> +	 * the gains are marginal.
> +	 *
> +	 * Therefore, choosing 2 as the multiplier strikes a good balance between
> +	 * enhancing parallel processing capabilities and maintaining efficient
> +	 * resource management.
> +	 *
> +	 * +------------+-------+-------+-------+-------+-------+
> +	 * | multiplier |   1   |   2   |   3   |   4   |   5   |
> +	 * +------------+-------+-------+-------+-------+-------+
> +	 * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms |
> +	 * | 2T   4node | 979ms | 679ms | 543ms | 489ms | 481ms |
> +	 * | 50G  2node | 71ms  | 44ms  | 37ms  | 30ms  | 31ms  |
> +	 * +------------+-------+-------+-------+-------+-------+
> +	 */
> +	job.max_threads	= num_node_state(N_MEMORY) * 2;
> +	job.min_chunk	= h->max_huge_pages / num_node_state(N_MEMORY) / 2;

For a single huge page, we get min_chunk of 0.  padata doesn't
explicitly handle that, but 'align' happens to save us from div by 0
later on.  It's an odd case, something to fix if there were another
version.


Not sure what efficient resource management means here.  Avoiding lock
contention?  The system is waiting on this initialization to start pid
1.  On big systems, most CPUs will be idle, so why not use available
resources to optimize it more?  max_threads could scale with CPU count
rather than a magic multiplier.

With that said, the major gain is already there, so either way,

Acked-by: Daniel Jordan <daniel.m.jordan@oracle.com> # padata