Skip to contentSkip to conversation

hugetlb: parallelize hugetlb page allocation on boot

RFC PATCH Superseded Gang Li 4 participants 14 messages Nov 23, 2023 – Nov 29, 2023

Subsystems (2)

v1 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Gang Li

author
#0 lore raw

cover hugetlb: parallelize hugetlb page allocation on boot

From: Gang Li <[email protected]>

Inspired by these patches [1][2], this series aims to speed up the initialization of hugetlb during the boot process through parallelization.

It is particularly effective in large systems. On a machine equipped with 1TB of memory and two NUMA nodes, the time for hugetlb initialization was reduced from 2 seconds to 1 second.

In the future, as memory continues to grow, more and more time can be saved.

This series currently focuses on optimizing 2MB hugetlb. Since gigantic pages are few in number, their optimization effects are not as pronounced. We may explore optimizations for gigantic pages in the future.

Thanks,
Gang Li

Gang Li (4):
hugetlb: code clean for hugetlb_hstate_alloc_pages
hugetlb: split hugetlb_hstate_alloc_pages
hugetlb: add timing to hugetlb allocations on boot
hugetlb: parallelize hugetlb page allocation

 mm/hugetlb.c | 191 ++++++++++++++++++++++++++++++++++++---------------
 1 file changed, 134 insertions(+), 57 deletions(-)
Signature · 1 line
2.20.1
author
#1 lore raw

RFC PATCH 1/4 hugetlb: code clean for hugetlb_hstate_alloc_pages

From: Gang Li <[email protected]>

This patch focus on cleaning up the code related to per node allocation and error reporting in the hugetlb alloc:

- hugetlb_hstate_alloc_pages_node_specific() to handle iterates through each online node and performs allocation if necessary.
- hugetlb_hstate_alloc_pages_report() report error during allocation. And the value of h->max_huge_pages is updated accordingly.

This patch has no functional changes.

---
 mm/hugetlb.c | 46 +++++++++++++++++++++++++++++-----------------
 1 file changed, 29 insertions(+), 17 deletions(-)
mm/hugetlb.c +29 -17
index c466551e2fd9..7af2ee08ad1b 100644
@@ -3482,6 +3482,33 @@ static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
 	h->max_huge_pages_node[nid] = i;
 }
 
+static bool __init hugetlb_hstate_alloc_pages_node_specific(struct hstate *h)
+{
+	int i;
+	bool node_specific_alloc = false;
+
+	for_each_online_node(i) {
+		if (h->max_huge_pages_node[i] > 0) {
+			hugetlb_hstate_alloc_pages_onenode(h, i);
+			node_specific_alloc = true;
+		}
+	}
+
+	return node_specific_alloc;
+}
+
+static void __init hugetlb_hstate_alloc_pages_report(unsigned long allocated, struct hstate *h)
+{
+	if (allocated < h->max_huge_pages) {
+		char buf[32];
+
+		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
+		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
+			h->max_huge_pages, buf, allocated);
+		h->max_huge_pages = allocated;
+	}
+}
+
 /*
  * NOTE: this routine is called in different contexts for gigantic and
  * non-gigantic pages.
@@ -3499,7 +3526,6 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 	struct folio *folio;
 	LIST_HEAD(folio_list);
 	nodemask_t *node_alloc_noretry;
-	bool node_specific_alloc = false;
 
 	/* skip gigantic hugepages allocation if hugetlb_cma enabled */
 	if (hstate_is_gigantic(h) && hugetlb_cma_size) {
@@ -3508,14 +3534,7 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 	}
 
 	/* do node specific alloc */
-	for_each_online_node(i) {
-		if (h->max_huge_pages_node[i] > 0) {
-			hugetlb_hstate_alloc_pages_onenode(h, i);
-			node_specific_alloc = true;
-		}
-	}
-
-	if (node_specific_alloc)
+	if (hugetlb_hstate_alloc_pages_node_specific(h))
 		return;
 
 	/* below will do all node balanced alloc */
@@ -3558,14 +3577,7 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 	/* list will be empty if hstate_is_gigantic */
 	prep_and_add_allocated_folios(h, &folio_list);
 
-	if (i < h->max_huge_pages) {
-		char buf[32];
-
-		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
-		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
-			h->max_huge_pages, buf, i);
-		h->max_huge_pages = i;
-	}
+	hugetlb_hstate_alloc_pages_report(i, h);
 	kfree(node_alloc_noretry);
 }
 
Signature · 1 line
2.20.1
author
#2 lore raw

RFC PATCH 2/4 hugetlb: split hugetlb_hstate_alloc_pages

In reply to Gang Li

From: Gang Li <[email protected]>

Split hugetlb_hstate_alloc_pages into gigantic and non-gigantic.

These patch has no functional changes.

---
 mm/hugetlb.c | 86 +++++++++++++++++++++++++++-------------------------
 1 file changed, 45 insertions(+), 41 deletions(-)
mm/hugetlb.c +45 -41
index 7af2ee08ad1b..7f9ff0855dd0 100644
@@ -3509,6 +3509,47 @@ static void __init hugetlb_hstate_alloc_pages_report(unsigned long allocated, st
 	}
 }
 
+static unsigned long __init hugetlb_hstate_alloc_pages_gigantic(struct hstate *h)
+{
+	unsigned long i;
+
+	for (i = 0; i < h->max_huge_pages; ++i) {
+		/*
+		 * gigantic pages not added to list as they are not
+		 * added to pools now.
+		 */
+		if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
+			break;
+		cond_resched();
+	}
+
+	return i;
+}
+
+static unsigned long __init hugetlb_hstate_alloc_pages_non_gigantic(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);
+
+	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();
+	}
+
+	prep_and_add_allocated_folios(h, &folio_list);
+
+	return i;
+}
+
 /*
  * NOTE: this routine is called in different contexts for gigantic and
  * non-gigantic pages.
@@ -3522,10 +3563,7 @@ static void __init hugetlb_hstate_alloc_pages_report(unsigned long allocated, st
  */
 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 {
-	unsigned long i;
-	struct folio *folio;
-	LIST_HEAD(folio_list);
-	nodemask_t *node_alloc_noretry;
+	unsigned long allocated;
 
 	/* skip gigantic hugepages allocation if hugetlb_cma enabled */
 	if (hstate_is_gigantic(h) && hugetlb_cma_size) {
@@ -3539,46 +3577,12 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 
 	/* below will do all node balanced alloc */
 	if (!hstate_is_gigantic(h)) {
-		/*
-		 * Bit mask controlling how hard we retry per-node allocations.
-		 * Ignore errors as lower level routines can deal with
-		 * node_alloc_noretry == NULL.  If this kmalloc fails at boot
-		 * time, we are likely in bigger trouble.
-		 */
-		node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
-						GFP_KERNEL);
+		allocated = hugetlb_hstate_alloc_pages_non_gigantic(h);
 	} else {
-		/* allocations done at boot time */
-		node_alloc_noretry = NULL;
-	}
-
-	/* bit mask controlling how hard we retry per-node allocations */
-	if (node_alloc_noretry)
-		nodes_clear(*node_alloc_noretry);
-
-	for (i = 0; i < h->max_huge_pages; ++i) {
-		if (hstate_is_gigantic(h)) {
-			/*
-			 * gigantic pages not added to list as they are not
-			 * added to pools now.
-			 */
-			if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
-				break;
-		} else {
-			folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
-							node_alloc_noretry);
-			if (!folio)
-				break;
-			list_add(&folio->lru, &folio_list);
-		}
-		cond_resched();
+		allocated = hugetlb_hstate_alloc_pages_gigantic(h);
 	}
 
-	/* list will be empty if hstate_is_gigantic */
-	prep_and_add_allocated_folios(h, &folio_list);
-
-	hugetlb_hstate_alloc_pages_report(i, h);
-	kfree(node_alloc_noretry);
+	hugetlb_hstate_alloc_pages_report(allocated, h);
 }
 
 static void __init hugetlb_init_hstates(void)
Signature · 1 line
2.20.1
author
#3 lore raw

RFC PATCH 3/4 hugetlb: add timing to hugetlb allocations on boot

In reply to Gang Li

From: Gang Li <[email protected]>

Add timing to hugetlb allocations for further optimization.

---
 mm/hugetlb.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
mm/hugetlb.c +3 -1
index 7f9ff0855dd0..ac8558724cc2 100644
@@ -3563,7 +3563,7 @@ static unsigned long __init hugetlb_hstate_alloc_pages_non_gigantic(struct hstat
  */
 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 {
-	unsigned long allocated;
+	unsigned long allocated, start;
 
 	/* skip gigantic hugepages allocation if hugetlb_cma enabled */
 	if (hstate_is_gigantic(h) && hugetlb_cma_size) {
@@ -3576,11 +3576,13 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 		return;
 
 	/* below will do all node balanced alloc */
+	start = jiffies;
 	if (!hstate_is_gigantic(h)) {
 		allocated = hugetlb_hstate_alloc_pages_non_gigantic(h);
 	} else {
 		allocated = hugetlb_hstate_alloc_pages_gigantic(h);
 	}
+	pr_info("HugeTLB: Allocation takes %u ms\n", jiffies_to_msecs(jiffies - start));
 
 	hugetlb_hstate_alloc_pages_report(allocated, h);
 }
Signature · 1 line
2.20.1
author
#4 lore raw

RFC PATCH 4/4 hugetlb: parallelize hugetlb page allocation

In reply to Gang Li

From: Gang Li <[email protected]>

By distributing the allocation across threads, large hugetlb configurations can allocate pages faster, improving boot speed.

---
 mm/hugetlb.c | 89 +++++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 74 insertions(+), 15 deletions(-)
mm/hugetlb.c +74 -15
index ac8558724cc2..df3fbe95989e 100644
@@ -3509,6 +3509,55 @@ static void __init hugetlb_hstate_alloc_pages_report(unsigned long allocated, st
 	}
 }
 
+struct hugetlb_work {
+	struct work_struct work;
+	struct hstate *h;
+	int num;
+	int nid;
+};
+
+static atomic_t hugetlb_hstate_alloc_n_undone __initdata;
+static __initdata DECLARE_COMPLETION(hugetlb_hstate_alloc_comp);
+
+static void __init hugetlb_alloc_node(struct work_struct *w)
+{
+	struct hugetlb_work *hw = container_of(w, struct hugetlb_work, work);
+	struct hstate *h = hw->h;
+	int i, num = hw->num;
+	nodemask_t node_alloc_noretry;
+	unsigned long flags;
+
+	/* 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);
+		if (!folio)
+			break;
+		spin_lock_irqsave(&hugetlb_lock, flags);
+		__prep_account_new_huge_page(h, folio_nid(folio));
+		enqueue_hugetlb_folio(h, folio);
+		spin_unlock_irqrestore(&hugetlb_lock, flags);
+		cond_resched();
+	}
+
+	if (atomic_dec_and_test(&hugetlb_hstate_alloc_n_undone))
+		complete(&hugetlb_hstate_alloc_comp);
+}
+
+static void __init hugetlb_vmemmap_optimize_node(struct work_struct *w)
+{
+	struct hugetlb_work *hw = container_of(w, struct hugetlb_work, work);
+	struct hstate *h = hw->h;
+	int nid = hw->nid;
+
+	hugetlb_vmemmap_optimize_folios(h, &h->hugepage_freelists[nid]);
+
+	if (atomic_dec_and_test(&hugetlb_hstate_alloc_n_undone))
+		complete(&hugetlb_hstate_alloc_comp);
+}
+
 static unsigned long __init hugetlb_hstate_alloc_pages_gigantic(struct hstate *h)
 {
 	unsigned long i;
@@ -3528,26 +3577,36 @@ static unsigned long __init hugetlb_hstate_alloc_pages_gigantic(struct hstate *h
 
 static unsigned long __init hugetlb_hstate_alloc_pages_non_gigantic(struct hstate *h)
 {
-	unsigned long i;
-	struct folio *folio;
-	LIST_HEAD(folio_list);
-	nodemask_t node_alloc_noretry;
+	int nid;
+	struct hugetlb_work *works;
 
-	/* Bit mask controlling how hard we retry per-node allocations.*/
-	nodes_clear(node_alloc_noretry);
+	works = kcalloc(num_node_state(N_MEMORY), sizeof(*works), GFP_KERNEL);
+	if (works == NULL) {
+		pr_warn("HugeTLB: allocating struct hugetlb_work failed.\n");
+		return 0;
+	}
 
-	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();
+	atomic_set(&hugetlb_hstate_alloc_n_undone, num_node_state(N_MEMORY));
+	for_each_node_state(nid, N_MEMORY) {
+		works[nid].h = h;
+		works[nid].num = h->max_huge_pages/num_node_state(N_MEMORY);
+		if (nid == 0)
+			works[nid].num += h->max_huge_pages % num_node_state(N_MEMORY);
+		INIT_WORK(&works[nid].work, hugetlb_alloc_node);
+		queue_work_node(nid, system_unbound_wq, &works[nid].work);
 	}
+	wait_for_completion(&hugetlb_hstate_alloc_comp);
 
-	prep_and_add_allocated_folios(h, &folio_list);
+	atomic_set(&hugetlb_hstate_alloc_n_undone, num_node_state(N_MEMORY));
+	for_each_node_state(nid, N_MEMORY) {
+		works[nid].nid = nid;
+		INIT_WORK(&works[nid].work, hugetlb_vmemmap_optimize_node);
+		queue_work_node(nid, system_unbound_wq, &works[nid].work);
+	}
+	wait_for_completion(&hugetlb_hstate_alloc_comp);
 
-	return i;
+	kfree(works);
+	return h->nr_huge_pages;
 }
 
 /*
Signature · 1 line
2.20.1
#6 lore raw
In reply to Gang Li
Quoted 9 lines · Gang Li

From: Gang Li <[email protected]>

Inspired by these patches [1][2], this series aims to speed up the initialization of hugetlb during the boot process through parallelization.

It is particularly effective in large systems. On a machine equipped with 1TB of memory and two NUMA nodes, the time for hugetlb initialization was reduced from 2 seconds to 1 second.

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

2 vs. 1 second on a 1 TiB system is usually really just noise.

Signature · 3 lines
Cheers,

David / dhildenb
#7 lore raw
Quoted 11 lines · David Hildenbrand
Quoted 9 lines · Gang Li

From: Gang Li <[email protected]>

Inspired by these patches [1][2], this series aims to speed up the initialization of hugetlb during the boot process through parallelization.

It is particularly effective in large systems. On a machine equipped with 1TB of memory and two NUMA nodes, the time for hugetlb initialization was reduced from 2 seconds to 1 second.

Quoted 6 lines

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

2 vs. 1 second on a 1 TiB system is usually really just noise.

The cost will continue to grow over time, so I presume that Gang is trying to get out in front of the issue even though it may not be a large savings today.

Running single boot tests, with the latest upstream kernel, allocating 1,440 1GB hugetlb pages on a 1.5TB AMD host appears to take 1.47s.

But allocating 11,776 1GB hugetlb pages on a 12TB Intel host takes 65.2s today with the current implementation.

So it's likely something worth optimizing.

Gang, I'm curious about this in the cover letter:

"""
This series currently focuses on optimizing 2MB hugetlb. Since gigantic pages are few in number, their optimization effects are not as pronounced. We may explore optimizations for gigantic pages in the future.
"""

For >1TB hosts, why the emphasis on 2MB hugetlb? :) I would have expected 1GB pages. Are you really allocating ~500k 2MB hugetlb pages?

So if the patchset optimizes for the more likely scenario on these large hosts, which would be 1GB pages, that would be great.

#8 lore raw
Quoted 24 lines · David Rientjes
Quoted 11 lines · David Hildenbrand
Quoted 9 lines · Gang Li

From: Gang Li <[email protected]>

Inspired by these patches [1][2], this series aims to speed up the initialization of hugetlb during the boot process through parallelization.

It is particularly effective in large systems. On a machine equipped with 1TB of memory and two NUMA nodes, the time for hugetlb initialization was reduced from 2 seconds to 1 second.

Quoted 6 lines

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

2 vs. 1 second on a 1 TiB system is usually really just noise.

The cost will continue to grow over time, so I presume that Gang is trying to get out in front of the issue even though it may not be a large savings today.

Quoted 5 lines

Running single boot tests, with the latest upstream kernel, allocating 1,440 1GB hugetlb pages on a 1.5TB AMD host appears to take 1.47s.

But allocating 11,776 1GB hugetlb pages on a 12TB Intel host takes 65.2s today with the current implementation.

And there, the 65.2s won't be noise because that 12TB system is up by a snap of a finger? :)

Signature · 3 lines
Cheers,

David / dhildenb
#9 lore raw
Quoted 25 lines · David Hildenbrand
Quoted 24 lines · David Rientjes
Quoted 11 lines · David Hildenbrand
Quoted 9 lines · Gang Li

From: Gang Li <[email protected]>

Inspired by these patches [1][2], this series aims to speed up the initialization of hugetlb during the boot process through parallelization.

It is particularly effective in large systems. On a machine equipped with 1TB of memory and two NUMA nodes, the time for hugetlb initialization was reduced from 2 seconds to 1 second.

Quoted 6 lines

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

2 vs. 1 second on a 1 TiB system is usually really just noise.

The cost will continue to grow over time, so I presume that Gang is trying to get out in front of the issue even though it may not be a large savings today.

Quoted 9 lines
Quoted 5 lines

Running single boot tests, with the latest upstream kernel, allocating 1,440 1GB hugetlb pages on a 1.5TB AMD host appears to take 1.47s.

But allocating 11,776 1GB hugetlb pages on a 12TB Intel host takes 65.2s today with the current implementation.

And there, the 65.2s won't be noise because that 12TB system is up by a snap of a finger? :)

In this single boot test, total boot time was 373.78s, so 1GB hugetlb allocation is 17.4% of that.

Would love to see what the numbers would look like if 1GB pages were supported.

#10 lore raw
Quoted 8 lines · David Rientjes
Quoted 3 lines · David Hildenbrand

And there, the 65.2s won't be noise because that 12TB system is up by a snap of a finger? :)

In this single boot test, total boot time was 373.78s, so 1GB hugetlb allocation is 17.4% of that.

Thank you for sharing these data. Currently, I don't have access to a machine of such large capacity, so the benefits in my tests are not as pronounced.

I believe testing on a system of this scale would yield significant benefits.

Quoted 4 lines

Would love to see what the numbers would look like if 1GB pages were supported.

Support for 1GB hugetlb is not yet perfect, so it wasn't included in v1. But I'm happy to refine and introduce 1GB hugetlb support in future versions.

#11 lore raw

Hi David Hildenbrand :),

Quoted 3 lines · David Hildenbrand

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

Regarding your concern about complexity, this is indeed something to consider. There is a precedent of parallelization in pgdata[1] which might be reused (or other methods) to reduce the complexity of this series.

[1] https://lore.kernel.org/all/[email protected]/

#12 lore raw
Quoted 11 lines · Gang Li

Hi David Hildenbrand :),

Quoted 3 lines · David Hildenbrand

Sorry to say, but why is that a scenario worth adding complexity for / optimizing for? You don't cover that, so there is a clear lack in the motivation.

Regarding your concern about complexity, this is indeed something to consider. There is a precedent of parallelization in pgdata[1] which might be reused (or other methods) to reduce the complexity of this series.

Yes, please!

Signature · 3 lines
Cheers,

David / dhildenb
#13 lore raw
In reply to Gang Li
Quoted 15 lines · Gang Li
Quoted 8 lines
Quoted 5 lines

And there, the 65.2s won't be noise because that 12TB system is up by a snap
of a finger? :)

In this single boot test, total boot time was 373.78s, so 1GB hugetlb allocation is 17.4% of that.

Thank you for sharing these data. Currently, I don't have access to a machine of such large capacity, so the benefits in my tests are not as pronounced.

I believe testing on a system of this scale would yield significant benefits.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

Would love to see what the numbers would look like if 1GB pages were supported.

Support for 1GB hugetlb is not yet perfect, so it wasn't included in v1. But I'm happy to refine and introduce 1GB hugetlb support in future versions.

That would be very appreciated, thank you! I'm happy to test and collect data for any proposed patch series on 12TB systems booted with a lot of 1GB hugetlb pages on the kernel command line.