Skip to contentSkip to conversation

cpusets,isolcpus: resolve conflict between cpusets and isolcpus

PATCH In review [email protected] 3 participants 10 messages Feb 23, 2015 – Feb 25, 2015

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover cpusets,isolcpus: resolve conflict between cpusets and isolcpus

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

#1 lore raw

PATCH 1/2 cpusets,isolcpus: exclude isolcpus from load balancing in cpusets

From: Rik van Riel <[email protected]>

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

Cc Peter Zijlstra <[email protected]>
Cc Clark Williams <[email protected]>
Cc Li Zefan <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Luiz Capitulino <[email protected]>
---
 include/linux/sched.h |  2 ++
 kernel/cpuset.c       | 13 +++++++++++--
 kernel/sched/core.c   |  2 +-
 3 files changed, 14 insertions(+), 3 deletions(-)
include/linux/sched.h +2 -0
index cb5cdc777c8a..af1b32a5ddcc 100644
@@ -1038,6 +1038,8 @@ static inline struct cpumask *sched_domain_span(struct sched_domain *sd)
 extern void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 				    struct sched_domain_attr *dattr_new);
 
+extern cpumask_var_t cpu_isolated_map;
+
 /* Allocate an array of sched domains, for partition_sched_domains(). */
 cpumask_var_t *alloc_sched_domains(unsigned int ndoms);
 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms);
kernel/cpuset.c +11 -2
index 64b257f6bca2..1ad63fa37cb4 100644
@@ -625,6 +625,7 @@ static int generate_sched_domains(cpumask_var_t **domains,
 	int csn;		/* how many cpuset ptrs in csa so far */
 	int i, j, k;		/* indices for partition finding loops */
 	cpumask_var_t *doms;	/* resulting partition; i.e. sched domains */
+	cpumask_var_t non_isolated_cpus;  /* load balanced CPUs */
 	struct sched_domain_attr *dattr;  /* attributes for custom domains */
 	int ndoms = 0;		/* number of sched domains in result */
 	int nslot;		/* next empty doms[] struct cpumask slot */
@@ -634,6 +635,10 @@ static int generate_sched_domains(cpumask_var_t **domains,
 	dattr = NULL;
 	csa = NULL;
 
+	if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
+		goto done;
+	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
+
 	/* Special case for the 99% of systems with one, full, sched domain */
 	if (is_sched_load_balance(&top_cpuset)) {
 		ndoms = 1;
@@ -646,7 +651,8 @@ static int generate_sched_domains(cpumask_var_t **domains,
 			*dattr = SD_ATTR_INIT;
 			update_domain_attr_tree(dattr, &top_cpuset);
 		}
-		cpumask_copy(doms[0], top_cpuset.effective_cpus);
+		cpumask_and(doms[0], top_cpuset.effective_cpus,
+				     non_isolated_cpus);
 
 		goto done;
 	}
@@ -669,7 +675,8 @@ static int generate_sched_domains(cpumask_var_t **domains,
 		 * the corresponding sched domain.
 		 */
 		if (!cpumask_empty(cp->cpus_allowed) &&
-		    !is_sched_load_balance(cp))
+		    !(is_sched_load_balance(cp) &&
+		      cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
 			continue;
 
 		if (is_sched_load_balance(cp))
@@ -751,6 +758,7 @@ static int generate_sched_domains(cpumask_var_t **domains,
 
 			if (apn == b->pn) {
 				cpumask_or(dp, dp, b->effective_cpus);
+				cpumask_and(dp, dp, non_isolated_cpus);
 				if (dattr)
 					update_domain_attr_tree(dattr + nslot, b);
 
@@ -763,6 +771,7 @@ static int generate_sched_domains(cpumask_var_t **domains,
 	BUG_ON(nslot != ndoms);
 
 done:
+	free_cpumask_var(non_isolated_cpus);
 	kfree(csa);
 
 	/*
kernel/sched/core.c +1 -1
index 97fe79cf613e..6069f3703240 100644
@@ -5831,7 +5831,7 @@ cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 }
 
 /* cpus with isolated domains */
-static cpumask_var_t cpu_isolated_map;
+cpumask_var_t cpu_isolated_map;
 
 /* Setup the mask of cpus configured for isolated domains */
 static int __init isolated_cpu_setup(char *str)
Signature · 1 line
1.9.3
Tested-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] cpusets,isolcpus: exclude isolcpus from load balancing in cpusets

Quoted 18 lines · On Mon, 23 Feb 2015, [email protected] wrote:

From: Rik van Riel <[email protected]>

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

Quoted 7 lines
Cc Peter Zijlstra <[email protected]>
Cc Clark Williams <[email protected]>
Cc Li Zefan <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Luiz Capitulino <[email protected]>
#3 lore raw

PATCH 2/2 cpusets,isolcpus: add file to show isolated cpus in cpuset

In reply to [email protected]

From: Rik van Riel <[email protected]>

The previous patch makes it so the code skips over isolcpus when building scheduler load balancing domains. This makes it hard to see for a user which of the CPUs in a cpuset are participating in load balancing, and which ones are isolated cpus.

Add a cpuset.isolcpus file with info on which cpus in a cpuset are isolated CPUs.

This file is read-only for now. In the future we could extend things so isolcpus can be changed at run time, for the root (system wide) cpuset only.

Cc Peter Zijlstra <[email protected]>
Cc Clark Williams <[email protected]>
Cc Li Zefan <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Luiz Capitulino <[email protected]>
---
 kernel/cpuset.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
kernel/cpuset.c +27 -0
index 1ad63fa37cb4..19ad5d3377f8 100644
@@ -1563,6 +1563,7 @@ typedef enum {
 	FILE_MEMORY_PRESSURE,
 	FILE_SPREAD_PAGE,
 	FILE_SPREAD_SLAB,
+	FILE_ISOLCPUS,
 } cpuset_filetype_t;
 
 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
@@ -1704,6 +1705,23 @@ static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
 	return retval ?: nbytes;
 }
 
+static size_t cpuset_sprintf_isolcpus(char *s, ssize_t pos, struct cpuset *cs)
+{
+	cpumask_var_t my_isolated_cpus;
+	ssize_t count;
+	
+	if (!alloc_cpumask_var(&my_isolated_cpus, GFP_KERNEL))
+		return 0;
+
+	cpumask_and(my_isolated_cpus, cs->cpus_allowed, cpu_isolated_map);
+
+	count = cpulist_scnprintf(s, pos, my_isolated_cpus);
+
+	free_cpumask_var(my_isolated_cpus);
+
+	return count;
+}
+
 /*
  * These ascii lists should be read in a single call, by using a user
  * buffer large enough to hold the entire map.  If read in smaller
@@ -1738,6 +1756,9 @@ static int cpuset_common_seq_show(struct seq_file *sf, void *v)
 	case FILE_EFFECTIVE_MEMLIST:
 		s += nodelist_scnprintf(s, count, cs->effective_mems);
 		break;
+	case FILE_ISOLCPUS:
+		s += cpuset_sprintf_isolcpus(s, count, cs);
+		break;
 	default:
 		ret = -EINVAL;
 		goto out_unlock;
@@ -1906,6 +1927,12 @@ static struct cftype files[] = {
 		.private = FILE_MEMORY_PRESSURE_ENABLED,
 	},
 
+	{
+		.name = "isolcpus",
+		.seq_show = cpuset_common_seq_show,
+		.private = FILE_ISOLCPUS,
+	},
+
 	{ }	/* terminate */
 };
 
Signature · 1 line
1.9.3
#4 lore raw

Re: [PATCH 2/2] cpusets,isolcpus: add file to show isolated cpus in cpuset

Quoted 38 lines · On Mon, 23 Feb 2015, [email protected] wrote:

From: Rik van Riel <[email protected]>

The previous patch makes it so the code skips over isolcpus when building scheduler load balancing domains. This makes it hard to see for a user which of the CPUs in a cpuset are participating in load balancing, and which ones are isolated cpus.

Add a cpuset.isolcpus file with info on which cpus in a cpuset are isolated CPUs.

This file is read-only for now. In the future we could extend things so isolcpus can be changed at run time, for the root (system wide) cpuset only.

Cc Peter Zijlstra <[email protected]>
Cc Clark Williams <[email protected]>
Cc Li Zefan <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Luiz Capitulino <[email protected]>
---
 kernel/cpuset.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
kernel/cpuset.c +1 -0
index 1ad63fa37cb4..19ad5d3377f8 100644
@@ -1563,6 +1563,7 @@ typedef enum {
 	FILE_MEMORY_PRESSURE,
 	FILE_SPREAD_PAGE,
 	FILE_SPREAD_SLAB,
+	FILE_ISOLCPUS,
 } cpuset_filetype_t;
 
 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,

kernel/cpuset.c static ssize_t cpuset_write_resmask(struct kernfs_open_file *of, · line 1712

Quoted 8 lines
kernel/cpuset.c +5 -0
@@ -1704,6 +1705,23 @@ static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
 	return retval ?: nbytes;
 }
 
+static size_t cpuset_sprintf_isolcpus(char *s, ssize_t pos, struct cpuset *cs)
+{
+	cpumask_var_t my_isolated_cpus;
+	ssize_t count;
+	

Whitespace.

Quoted 16 lines
kernel/cpuset.c +12 -0
+	if (!alloc_cpumask_var(&my_isolated_cpus, GFP_KERNEL))
+		return 0;
+
+	cpumask_and(my_isolated_cpus, cs->cpus_allowed, cpu_isolated_map);
+
+	count = cpulist_scnprintf(s, pos, my_isolated_cpus);
+
+	free_cpumask_var(my_isolated_cpus);
+
+	return count;
+}
+
 /*
  * These ascii lists should be read in a single call, by using a user
  * buffer large enough to hold the entire map.  If read in smaller

kernel/cpuset.c static int cpuset_common_seq_show(struct seq_file *sf, void *v) · line 1761

Quoted 6 lines
kernel/cpuset.c +3 -0
@@ -1738,6 +1756,9 @@ static int cpuset_common_seq_show(struct seq_file *sf, void *v)
 	case FILE_EFFECTIVE_MEMLIST:
 		s += nodelist_scnprintf(s, count, cs->effective_mems);
 		break;
+	case FILE_ISOLCPUS:
+		s += cpuset_sprintf_isolcpus(s, count, cs);
+		break;

This patch looks fine, and I think cpuset.effective_cpus and cpuset.isolcpus can be used well together, but will need updating now that commit e8e6d97c9b ("cpuset: use %*pb[l] to print bitmaps including cpumasks and nodemasks") has been merged which reworks this function.

It's a little unfortunate, though, that the user sees Cpus_allowed, cpuset.cpus, and cpuset.effective_cpus that include isolcpus and then have to check another cpulist for the isolcpus to see their sched domain, though.

Quoted 16 lines
kernel/cpuset.c +6 -0
 	default:
 		ret = -EINVAL;
 		goto out_unlock;
@@ -1906,6 +1927,12 @@ static struct cftype files[] = {
 		.private = FILE_MEMORY_PRESSURE_ENABLED,
 	},
 
+	{
+		.name = "isolcpus",
+		.seq_show = cpuset_common_seq_show,
+		.private = FILE_ISOLCPUS,
+	},
+
 	{ }	/* terminate */
 };
 
#5 lore raw

Re: [PATCH 2/2] cpusets,isolcpus: add file to show isolated cpus in cpuset

Quoted 70 lines · David Rientjes
Quoted 38 lines · On Mon, 23 Feb 2015, [email protected] wrote:

From: Rik van Riel <[email protected]>

The previous patch makes it so the code skips over isolcpus when building scheduler load balancing domains. This makes it hard to see for a user which of the CPUs in a cpuset are participating in load balancing, and which ones are isolated cpus.

Add a cpuset.isolcpus file with info on which cpus in a cpuset are isolated CPUs.

This file is read-only for now. In the future we could extend things so isolcpus can be changed at run time, for the root (system wide) cpuset only.

Cc Peter Zijlstra <[email protected]>
Cc Clark Williams <[email protected]>
Cc Li Zefan <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Luiz Capitulino <[email protected]>
---
 kernel/cpuset.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
kernel/cpuset.c +1 -0
index 1ad63fa37cb4..19ad5d3377f8 100644
@@ -1563,6 +1563,7 @@ typedef enum {
 	FILE_MEMORY_PRESSURE,
 	FILE_SPREAD_PAGE,
 	FILE_SPREAD_SLAB,
+	FILE_ISOLCPUS,
 } cpuset_filetype_t;
 
 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,

kernel/cpuset.c static ssize_t cpuset_write_resmask(struct kernfs_open_file *of, · line 1712

Quoted 8 lines
kernel/cpuset.c +5 -0
@@ -1704,6 +1705,23 @@ static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
 	return retval ?: nbytes;
 }
 
+static size_t cpuset_sprintf_isolcpus(char *s, ssize_t pos, struct cpuset *cs)
+{
+	cpumask_var_t my_isolated_cpus;
+	ssize_t count;
+	

Whitespace.

Quoted 19 lines
kernel/cpuset.c +12 -0
+	if (!alloc_cpumask_var(&my_isolated_cpus, GFP_KERNEL))
+		return 0;
+
+	cpumask_and(my_isolated_cpus, cs->cpus_allowed, cpu_isolated_map);
+
+	count = cpulist_scnprintf(s, pos, my_isolated_cpus);
+
+	free_cpumask_var(my_isolated_cpus);
+
+	return count;
+}
+
 /*
  * These ascii lists should be read in a single call, by using a user
  * buffer large enough to hold the entire map.  If read in smaller
@@ -1738,6 +1756,9 @@ static int cpuset_common_seq_show(struct seq_file *sf, void *v)
 	case FILE_EFFECTIVE_MEMLIST:
 		s += nodelist_scnprintf(s, count, cs->effective_mems);
 		break;
Quoted 8 lines

kernel/cpuset.c static int cpuset_common_seq_show(struct seq_file *sf, void *v) · line 1761

Quoted 3 lines
kernel/cpuset.c +3 -0
+	case FILE_ISOLCPUS:
+		s += cpuset_sprintf_isolcpus(s, count, cs);
+		break;

This patch looks fine, and I think cpuset.effective_cpus and cpuset.isolcpus can be used well together, but will need updating now that commit e8e6d97c9b ("cpuset: use %*pb[l] to print bitmaps including cpumasks and nodemasks") has been merged which reworks this function.

I will take a look at that changeset. It was not in the tip tree I worked against.

Expect a v2 :)

Quoted 4 lines

It's a little unfortunate, though, that the user sees Cpus_allowed, cpuset.cpus, and cpuset.effective_cpus that include isolcpus and then have to check another cpulist for the isolcpus to see their sched domain, though.

Agreed, but all the alternatives I could think of would break the userspace API, leaving this as the best way to go.

Signature · 1 line
All rights reversed
#6 lore raw
In reply to [email protected]
Quoted 13 lines · On Mon, 2015-02-23 at 16:45 -0500, [email protected] wrote:

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

Quoted 2 lines

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

3/3: nohz_full cpus become part of that unified isolated map?

-Mike

#7 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 16 lines · Mike Galbraith
Quoted 15 lines · On Mon, 2015-02-23 at 16:45 -0500, [email protected] wrote:

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

Quoted 5 lines
Quoted 3 lines

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

3/3: nohz_full cpus become part of that unified isolated map?

There may be use cases where users want nohz_full, but still want the scheduler to automatically load balance the CPU.

I am not sure whether we want nohz_full and isolcpus to always overlap 100%.

On the other hand, any CPU that is isolated with isolcpus= probably wants nohz_full...

- --
All rights reversed
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1

iQEcBAEBAgAGBQJU7IbwAAoJEM553pKExN6DxpAIAIt3Wp1fYhyTiceCZPZj/75y aNdpa1tsdyZmC3UoqHlWPajhU9kz3LV88gkDuRVLkBSIbAdc+Krpj0QwU80SBpn8 MIRkzlDE5pHwqgpNEmY0dTI8OP/BWH6SzgkbAqACTeffR8glz49ELFL2IK9hSl4P 2j5gOc1sgBD24cpqComw0qpIJwhRfDTr270zHPzEcqwESYLD57Z6AZxLuz8UjDnD vgvmz5+zCeVKfPWfFSCUHGDZ56PuQAvQk0olAVp5pd6wwGoPyAMNJm12RgE2ru0M 8y/xyHAhbtzdV7XsdfBWWe6F4jmodvjhKKtqRdwGzQSGkJzxGjdlFBeSkOkeBsk= =4dML
-----END PGP SIGNATURE-----

#8 lore raw
Quoted 23 lines · Rik van Riel
Quoted 16 lines · Mike Galbraith
Quoted 15 lines · On Mon, 2015-02-23 at 16:45 -0500, [email protected] wrote:

Ensure that cpus specified with the isolcpus= boot commandline option stay outside of the load balancing in the kernel scheduler.

Operations like load balancing can introduce unwanted latencies, which is exactly what the isolcpus= commandline is there to prevent.

Previously, simply creating a new cpuset, without even touching the cpuset.cpus field inside the new cpuset, would undo the effects of isolcpus=, by creating a scheduler domain spanning the whole system, and setting up load balancing inside that domain. The cpuset root cpuset.cpus file is read-only, so there was not even a way to undo that effect.

Quoted 5 lines
Quoted 3 lines

This does not impact the majority of cpusets users, since isolcpus= is a fairly specialized feature used for realtime purposes.

3/3: nohz_full cpus become part of that unified isolated map?

Quoted 8 lines

There may be use cases where users want nohz_full, but still want the scheduler to automatically load balance the CPU.

I am not sure whether we want nohz_full and isolcpus to always overlap 100%.

On the other hand, any CPU that is isolated with isolcpus= probably wants nohz_full...

I can't imagine caring deeply about the tiny interference of the tick, yet not caring about the massive interference of load balancing.

-Mike

#9 lore raw
In reply to Rik van Riel
Quoted 2 lines · Rik van Riel

On the other hand, any CPU that is isolated with isolcpus= probably wants nohz_full...

Not here. I isolate (via cpusets) for a 60 core rt load, but it's not single task/core, and doesn't like the nohz_full overhead.

-Mike