svn commit: r339616 - in head: share/man/man4 sys/arm64/arm64 sys/kern sys/vm sys/x86/acpica
Mark Johnston
markj at FreeBSD.org
Mon Oct 22 20:13:53 UTC 2018
Author: markj
Date: Mon Oct 22 20:13:51 2018
New Revision: 339616
URL: https://svnweb.freebsd.org/changeset/base/339616
Log:
Make it possible to disable NUMA support with a tunable.
This provides a chicken switch for anyone negatively impacted by
enabling NUMA in the amd64 GENERIC kernel configuration. With
NUMA disabled at boot-time, information about the NUMA topology
is not exposed to the rest of the kernel, and all of physical
memory is viewed as coming from a single domain.
This method still has some performance overhead relative to disabling
NUMA support at compile time.
PR: 231460
Reviewed by: alc, gallatin, kib
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17439
Modified:
head/share/man/man4/numa.4
head/sys/arm64/arm64/mp_machdep.c
head/sys/kern/kern_cpuset.c
head/sys/vm/vm_phys.c
head/sys/x86/acpica/srat.c
Modified: head/share/man/man4/numa.4
==============================================================================
--- head/share/man/man4/numa.4 Mon Oct 22 20:00:43 2018 (r339615)
+++ head/share/man/man4/numa.4 Mon Oct 22 20:13:51 2018 (r339616)
@@ -24,18 +24,16 @@
.\"
.\" $FreeBSD$
.\"
-.Dd July 10, 2018
+.Dd October 22, 2018
.Dt NUMA 4
.Os
.Sh NAME
.Nm NUMA
.Nd Non-Uniform Memory Access
.Sh SYNOPSIS
-.Cd options SMP
-.Cd options MAXMEMDOM=16
+.Cd options MAXMEMDOM
+.Cd options NUMA
.Pp
-.In sys/cpuset.h
-.In sys/bus.h
.Sh DESCRIPTION
Non-Uniform Memory Access is a computer architecture design which
involves unequal costs between processors, memory and IO devices
@@ -47,14 +45,26 @@ architecture, the latency to access specific memory or
depends upon which processor the memory or device is attached to.
Accessing memory local to a processor is faster than accessing memory
that is connected to one of the other processors.
+.Fx
+implements NUMA-aware memory allocation policies.
+By default it attempts to ensure that allocations are balanced across
+each domain.
+Users may override the default domain selection policy using
+.Xr cpuset 1 .
.Pp
.Nm
-is enabled when the
+support is enabled when the
.Cd NUMA
-option is used in a kernel configuration
-file and the
+option is specified in the kernel configuration file.
+Each platform defines the
.Cd MAXMEMDOM
-option is set to a value greater than 1.
+constant, which specifies the maximum number of supported NUMA domains.
+This constant may be specified in the kernel configuration file.
+.Nm
+support can be disabled at boot time by setting the
+.Va vm.numa.disabled
+tunable to 1.
+Other values for this tunable are currently ignored.
.Pp
Thread and process
.Nm
@@ -128,7 +138,7 @@ tool first appeared in
.Fx 11.0
and were removed in
.Fx 12.0 .
-Current implementation appeared in
+The current implementation appeared in
.Fx 12.0 .
.Pp
.Sh AUTHORS
Modified: head/sys/arm64/arm64/mp_machdep.c
==============================================================================
--- head/sys/arm64/arm64/mp_machdep.c Mon Oct 22 20:00:43 2018 (r339615)
+++ head/sys/arm64/arm64/mp_machdep.c Mon Oct 22 20:13:51 2018 (r339616)
@@ -576,11 +576,12 @@ cpu_init_fdt(u_int id, phandle_t node, u_int addr_size
return (FALSE);
/* Try to read the numa node of this cpu */
- if (OF_getencprop(node, "numa-node-id", &domain, sizeof(domain)) > 0) {
- __pcpu[id].pc_domain = domain;
- if (domain < MAXMEMDOM)
- CPU_SET(id, &cpuset_domain[domain]);
- }
+ if (vm_ndomains == 1 ||
+ OF_getencprop(node, "numa-node-id", &domain, sizeof(domain)) <= 0)
+ domain = 0;
+ __pcpu[id].pc_domain = domain;
+ if (domain < MAXMEMDOM)
+ CPU_SET(id, &cpuset_domain[domain]);
return (TRUE);
}
Modified: head/sys/kern/kern_cpuset.c
==============================================================================
--- head/sys/kern/kern_cpuset.c Mon Oct 22 20:00:43 2018 (r339615)
+++ head/sys/kern/kern_cpuset.c Mon Oct 22 20:13:51 2018 (r339616)
@@ -458,6 +458,12 @@ _domainset_create(struct domainset *domain, struct dom
struct domainset *ndomain;
int i, j, max;
+ KASSERT(domain->ds_cnt <= vm_ndomains,
+ ("invalid domain count in domainset %p", domain));
+ KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER ||
+ domain->ds_prefer < vm_ndomains,
+ ("invalid preferred domain in domains %p", domain));
+
mtx_lock_spin(&cpuset_lock);
LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
if (domainset_equal(ndomain, domain))
Modified: head/sys/vm/vm_phys.c
==============================================================================
--- head/sys/vm/vm_phys.c Mon Oct 22 20:00:43 2018 (r339615)
+++ head/sys/vm/vm_phys.c Mon Oct 22 20:13:51 2018 (r339616)
@@ -597,11 +597,22 @@ vm_phys_register_domains(int ndomains, struct mem_affi
int *locality)
{
#ifdef NUMA
- int i;
+ int d, i;
- vm_ndomains = ndomains;
- mem_affinity = affinity;
- mem_locality = locality;
+ /*
+ * For now the only override value that we support is 1, which
+ * effectively disables NUMA-awareness in the allocators.
+ */
+ d = 0;
+ TUNABLE_INT_FETCH("vm.numa.disabled", &d);
+ if (d)
+ ndomains = 1;
+
+ if (ndomains > 1) {
+ vm_ndomains = ndomains;
+ mem_affinity = affinity;
+ mem_locality = locality;
+ }
for (i = 0; i < vm_ndomains; i++)
DOMAINSET_SET(i, &all_domains);
Modified: head/sys/x86/acpica/srat.c
==============================================================================
--- head/sys/x86/acpica/srat.c Mon Oct 22 20:00:43 2018 (r339615)
+++ head/sys/x86/acpica/srat.c Mon Oct 22 20:13:51 2018 (r339616)
@@ -535,11 +535,7 @@ srat_set_cpus(void *dummy)
if (!cpu->enabled)
panic("SRAT: CPU with APIC ID %u is not known",
pc->pc_apic_id);
-#ifdef NUMA
- pc->pc_domain = cpu->domain;
-#else
- pc->pc_domain = 0;
-#endif
+ pc->pc_domain = vm_ndomains > 1 ? cpu->domain : 0;
CPU_SET(i, &cpuset_domain[pc->pc_domain]);
if (bootverbose)
printf("SRAT: CPU %u has memory domain %d\n", i,
@@ -564,7 +560,7 @@ acpi_map_pxm_to_vm_domainid(int pxm)
for (i = 0; i < ndomain; i++) {
if (domain_pxm[i] == pxm)
- return (i);
+ return (vm_ndomains > 1 ? i : 0);
}
return (-1);
More information about the svn-src-all
mailing list