Skip to contentSkip to conversation

sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

PATCH In review Libo Chen 4 participants 9 messages Apr 24, 2025

Subsystems (4)
lore
Share & tools
lore

Libo Chen

author
#0 lore raw

v5 cover sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via
cpuset.mems
sched/numa: Add tracepoint that tracks the skipping of numa balancing
due to cpuset memory pinning

 include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
 kernel/sched/fair.c          |  9 +++++++++
 2 files changed, 42 insertions(+)
Signature · 1 line
2.43.5
author
#1 lore raw

PATCH v5 1/2 sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

When the memory of the current task is pinned to one NUMA node by cgroup, there is no point in continuing the rest of VMA scanning and hinting page faults as they will just be overhead. With this change, there will be no more unnecessary PTE updates or page faults in this scenario.

We have seen up to a 6x improvement on a typical java workload running on VMs with memory and CPU pinned to one NUMA node via cpuset in a two-socket AARCH64 system. With the same pinning, on a 18-cores-per-socket Intel platform, we have seen 20% improvment in a microbench that creates a 30-vCPU selftest KVM guest with 4GB memory, where each vCPU reads 4KB pages in a fixed number of loops.

---
 kernel/sched/fair.c | 7 +++++++
 1 file changed, 7 insertions(+)
kernel/sched/fair.c +7 -0
index e43993a4e580..c9903b1b3948 100644
@@ -3329,6 +3329,13 @@ static void task_numa_work(struct callback_head *work)
 	if (p->flags & PF_EXITING)
 		return;
 
+	/*
+	 * Memory is pinned to only one NUMA node via cpuset.mems, naturally
+	 * no page can be migrated.
+	 */
+	if (cpusets_enabled() && nodes_weight(cpuset_current_mems_allowed) == 1)
+		return;
+
 	if (!mm->numa_next_scan) {
 		mm->numa_next_scan = now +
 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
Signature · 1 line
2.43.5
author
#2 lore raw

PATCH v5 2/2 sched/numa: Add tracepoint that tracks the skipping of numa balancing due to cpuset memory pinning

In reply to Libo Chen

Unlike sched_skip_vma_numa tracepoint which tracks skipped VMAs, this tracks the task subjected to cpuset.mems pinning and prints out its allowed memory node mask.

---
 include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
 kernel/sched/fair.c          |  4 +++-
 2 files changed, 36 insertions(+), 1 deletion(-)
include/trace/events/sched.h +33 -0
index 8994e97d86c1..ff3990318aec 100644
@@ -745,6 +745,39 @@ TRACE_EVENT(sched_skip_vma_numa,
 		  __entry->vm_end,
 		  __print_symbolic(__entry->reason, NUMAB_SKIP_REASON))
 );
+
+TRACE_EVENT(sched_skip_cpuset_numa,
+
+	TP_PROTO(struct task_struct *tsk, nodemask_t *mem_allowed_ptr),
+
+	TP_ARGS(tsk, mem_allowed_ptr),
+
+	TP_STRUCT__entry(
+		__array( char,		comm,		TASK_COMM_LEN		)
+		__field( pid_t,		pid					)
+		__field( pid_t,		tgid					)
+		__field( pid_t,		ngid					)
+		__array( unsigned long, mem_allowed, BITS_TO_LONGS(MAX_NUMNODES))
+	),
+
+	TP_fast_assign(
+		memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
+		__entry->pid		 = task_pid_nr(tsk);
+		__entry->tgid		 = task_tgid_nr(tsk);
+		__entry->ngid		 = task_numa_group_id(tsk);
+		BUILD_BUG_ON(sizeof(nodemask_t) != \
+			     BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long));
+		memcpy(__entry->mem_allowed, mem_allowed_ptr->bits,
+		       sizeof(__entry->mem_allowed));
+	),
+
+	TP_printk("comm=%s pid=%d tgid=%d ngid=%d mem_nodes_allowed=%*pbl",
+		  __entry->comm,
+		  __entry->pid,
+		  __entry->tgid,
+		  __entry->ngid,
+		  MAX_NUMNODES, __entry->mem_allowed)
+);
 #endif /* CONFIG_NUMA_BALANCING */
 
 /*
kernel/sched/fair.c +3 -1
index c9903b1b3948..cc892961ce15 100644
@@ -3333,8 +3333,10 @@ static void task_numa_work(struct callback_head *work)
 	 * Memory is pinned to only one NUMA node via cpuset.mems, naturally
 	 * no page can be migrated.
 	 */
-	if (cpusets_enabled() && nodes_weight(cpuset_current_mems_allowed) == 1)
+	if (cpusets_enabled() && nodes_weight(cpuset_current_mems_allowed) == 1) {
+		trace_sched_skip_cpuset_numa(current, &cpuset_current_mems_allowed);
 		return;
+	}
 
 	if (!mm->numa_next_scan) {
 		mm->numa_next_scan = now +
Signature · 1 line
2.43.5
reviewer Tested-by on the v5 series
#3 lore raw
In reply to Libo Chen

Hello Libo,

Quoted 8 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

I tested the series with hackbench running on a dual socket system with memory pinned to one node and I could see the skip_cpuset_numa traces being logged:

 sched-messaging-9430    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9430 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9640    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9640 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9645    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9645 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9637    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9637 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9629    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9629 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9639    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9639 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9630    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9630 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9487    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9487 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9635    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9635 tgid=9007 ngid=0 mem_nodes_allowed=0
 sched-messaging-9647    ...: sched_skip_cpuset_numa: comm=sched-messaging pid=9647 tgid=9007 ngid=0 mem_nodes_allowed=0

...

Feel free to add:

Signature · 2 lines
Thanks and Regards,
Prateek
Quoted 24 lines

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via
cpuset.mems
sched/numa: Add tracepoint that tracks the skipping of numa balancing
due to cpuset memory pinning

  include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
  kernel/sched/fair.c          |  9 +++++++++
  2 files changed, 42 insertions(+)
Tested-by Chen Yu Tested-by K Prateek Nayak on the v5 series
#4 lore raw
In reply to Libo Chen
Quoted 28 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via
cpuset.mems
sched/numa: Add tracepoint that tracks the skipping of numa balancing
due to cpuset memory pinning

Quoted 4 lines
  include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
  kernel/sched/fair.c          |  9 +++++++++
  2 files changed, 42 insertions(+)

Hello Libo,

For some reason I am not able to apply this patch. I am trying to test the boot warning[1].

I am trying to apply on top of next-20250423. Below is the error. Am I missing anything?

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

Error:

git am -i v5_20250423_libo_chen_sched_numa_skip_vma_scanning_on_memory_pinned_to_one_numa_node_via_cpuset_mems.mbx Commit Body is:
--------------------------
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

When the memory of the current task is pinned to one NUMA node by cgroup, there is no point in continuing the rest of VMA scanning and hinting page faults as they will just be overhead. With this change, there will be no more unnecessary PTE updates or page faults in this scenario.

We have seen up to a 6x improvement on a typical java workload running on VMs with memory and CPU pinned to one NUMA node via cpuset in a two-socket AARCH64 system. With the same pinning, on a 18-cores-per-socket Intel platform, we have seen 20% improvment in a microbench that creates a 30-vCPU selftest KVM guest with 4GB memory, where each vCPU reads 4KB pages in a fixed number of loops.

--------------------------
Apply? [y]es/[n]o/[e]dit/[v]iew patch/[a]ccept all: a
Applying: sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems
error: patch failed: kernel/sched/fair.c:3329
error: kernel/sched/fair.c: patch does not apply
Patch failed at 0001 sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

Regards,

Venkat.

author
#5 lore raw
Quoted 64 lines · Venkat Rao Bagalkote
Quoted 28 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
   sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems
   sched/numa: Add tracepoint that tracks the skipping of numa balancing due to cpuset memory pinning

Quoted 4 lines

  include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++ kernel/sched/fair.c          |  9 +++++++++

  2 files changed, 42 insertions(+)

Hello Libo,

For some reason I am not able to apply this patch. I am trying to test the boot warning[1].

I am trying to apply on top of next-20250423. Below is the error. Am I missing anything?

[1]: https://urldefense.com/v3/__https://lore.kernel.org/all/[email protected]/__;!!ACWV5N9M2RV99hQ!IQpY9WDL1O3ppDekb1PpaTYJ98ehOXL6dNIkx02MPN84bCieT18zCh7WSouHctEGpwG2rtpZB42l7b5mkMFb$
Error:

git am -i v5_20250423_libo_chen_sched_numa_skip_vma_scanning_on_memory_pinned_to_one_numa_node_via_cpuset_mems.mbx Commit Body is:
--------------------------
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

When the memory of the current task is pinned to one NUMA node by cgroup, there is no point in continuing the rest of VMA scanning and hinting page faults as they will just be overhead. With this change, there will be no more unnecessary PTE updates or page faults in this scenario.

We have seen up to a 6x improvement on a typical java workload running on VMs with memory and CPU pinned to one NUMA node via cpuset in a two-socket AARCH64 system. With the same pinning, on a 18-cores-per-socket Intel platform, we have seen 20% improvment in a microbench that creates a 30-vCPU selftest KVM guest with 4GB memory, where each vCPU reads 4KB pages in a fixed number of loops.

Quoted 8 lines

--------------------------
Apply? [y]es/[n]o/[e]dit/[v]iew patch/[a]ccept all: a
Applying: sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems
error: patch failed: kernel/sched/fair.c:3329
error: kernel/sched/fair.c: patch does not apply
Patch failed at 0001 sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

Hi Venkat,

I just did git am -i t.mbox on top of next-20250423, not sure why but the second patch was ahead of the first patch in apply order, have you made sure the second patch was not applied before the first one?

- Libo

Quoted 6 lines

Regards,

Venkat.

#6 lore raw
Quoted 72 lines · Libo Chen
Quoted 62 lines · Venkat Rao Bagalkote
Quoted 28 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
   sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems
   sched/numa: Add tracepoint that tracks the skipping of numa balancing due to cpuset memory pinning

Quoted 4 lines
   include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
   kernel/sched/fair.c          |  9 +++++++++
   2 files changed, 42 insertions(+)

Hello Libo,

For some reason I am not able to apply this patch. I am trying to test the boot warning[1].

I am trying to apply on top of next-20250423. Below is the error. Am I missing anything?

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

git am -i v5_20250423_libo_chen_sched_numa_skip_vma_scanning_on_memory_pinned_to_one_numa_node_via_cpuset_mems.mbx Commit Body is:
--------------------------
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

When the memory of the current task is pinned to one NUMA node by cgroup, there is no point in continuing the rest of VMA scanning and hinting page faults as they will just be overhead. With this change, there will be no more unnecessary PTE updates or page faults in this scenario.

We have seen up to a 6x improvement on a typical java workload running on VMs with memory and CPU pinned to one NUMA node via cpuset in a two-socket AARCH64 system. With the same pinning, on a 18-cores-per-socket Intel platform, we have seen 20% improvment in a microbench that creates a 30-vCPU selftest KVM guest with 4GB memory, where each vCPU reads 4KB pages in a fixed number of loops.

Quoted 8 lines

--------------------------
Apply? [y]es/[n]o/[e]dit/[v]iew patch/[a]ccept all: a
Applying: sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems
error: patch failed: kernel/sched/fair.c:3329
error: kernel/sched/fair.c: patch does not apply
Patch failed at 0001 sched/numa: Skip VMA scanning on memory pinned to one NUMA node via cpuset.mems

Quoted 6 lines

Hi Venkat,

I just did git am -i t.mbox on top of next-20250423, not sure why but the second patch was ahead of the first patch in apply order, have you made sure the second patch was not applied before the first one?

- Libo

Hi Libo,

Apolozies!!!

I freshly cloned and tried and it worked now. So, please ignore my earlier mail.

Regards,

Venkat.

Quoted 7 lines
Quoted 6 lines

Regards,

Venkat.

Tested-by on the v5 series
#7 lore raw
In reply to Libo Chen
Quoted 28 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via
cpuset.mems
sched/numa: Add tracepoint that tracks the skipping of numa balancing
due to cpuset memory pinning

Quoted 4 lines
  include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
  kernel/sched/fair.c          |  9 +++++++++
  2 files changed, 42 insertions(+)

Tested on top of next-20250424. The boot warning[1] is fixed with this version.

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

Tested-by on the v5 series
#8 lore raw
In reply to Libo Chen
Quoted 28 lines · Libo Chen

v1->v2:
1. add perf improvment numbers in commit log. Yet to find perf diff on will-it-scale, so not included here. Plan to run more workloads.
2. add tracepoint.
3. To peterz's comment, this will make it impossible to attract tasks to those memory just like other VMA skippings. This is the current implementation, I think we can improve that in the future, but at the moment it's probabaly better to keep it consistent.

v2->v3:
1. add enable_cpuset() based on Mel's suggestion but again I think it's redundant.
2. print out nodemask with %*p.. format in the tracepoint.

v3->v4:
1. fix an unsafe dereference of a pointer to content not on ring buffer, namely mem_allowed_ptr in the tracepoint.

v4->v5:
1. add BUILD_BUG_ON() in TP_fast_assign() to guard against future changes (particularly in size) in nodemask_t.

Libo Chen (2):
sched/numa: Skip VMA scanning on memory pinned to one NUMA node via
cpuset.mems
sched/numa: Add tracepoint that tracks the skipping of numa balancing
due to cpuset memory pinning

Quoted 4 lines
  include/trace/events/sched.h | 33 +++++++++++++++++++++++++++++++++
  kernel/sched/fair.c          |  9 +++++++++
  2 files changed, 42 insertions(+)

Tested the above patch on top of next-20250424 and it fixes the boot warning on IBM Power server. Hence,

Regards,

Venkat.