daandemeyer / rpms / systemd

Forked from rpms/systemd 2 years ago
Clone
4fbe94
From a735699a8287c19e043b7d2fe9a387a3938e1e2f Mon Sep 17 00:00:00 2001
4fbe94
From: =?UTF-8?q?Michal=20Sekleta=CC=81r?= <msekleta@redhat.com>
4fbe94
Date: Mon, 18 Nov 2019 12:50:11 +0100
4fbe94
Subject: [PATCH] core: introduce NUMAPolicy and NUMAMask options
4fbe94
4fbe94
Make possible to set NUMA allocation policy for manager. Manager's
4fbe94
policy is by default inherited to all forked off processes. However, it
4fbe94
is possible to override the policy on per-service basis. Currently we
4fbe94
support, these policies: default, prefer, bind, interleave, local.
4fbe94
See man 2 set_mempolicy for details on each policy.
4fbe94
4fbe94
Overall NUMA policy actually consists of two parts. Policy itself and
4fbe94
bitmask representing NUMA nodes where is policy effective. Node mask can
4fbe94
be specified using related option, NUMAMask. Default mask can be
4fbe94
overwritten on per-service level.
4fbe94
4fbe94
(cherry-picked from commit fe9c54b2188e6cd23262a319f96b13215f2c5e9c)
4fbe94
4fbe94
Resolves: #1734787
4fbe94
---
4fbe94
 man/systemd-system.conf.xml           | 19 ++++++
4fbe94
 man/systemd.exec.xml                  | 28 +++++++++
4fbe94
 meson.build                           |  4 ++
4fbe94
 src/basic/cpu-set-util.c              | 91 +++++++++++++++++++++++++++
4fbe94
 src/basic/cpu-set-util.h              | 28 +++++++++
4fbe94
 src/basic/exit-status.c               |  3 +
4fbe94
 src/basic/exit-status.h               |  1 +
4fbe94
 src/basic/missing_syscall.h           | 43 +++++++++++++
4fbe94
 src/core/dbus-execute.c               | 65 ++++++++++++++++++-
4fbe94
 src/core/execute.c                    | 20 ++++++
4fbe94
 src/core/execute.h                    |  1 +
4fbe94
 src/core/load-fragment-gperf.gperf.m4 |  2 +
4fbe94
 src/core/load-fragment.c              | 28 +++++++++
4fbe94
 src/core/load-fragment.h              |  2 +
4fbe94
 src/core/main.c                       | 27 ++++++++
4fbe94
 src/core/system.conf.in               |  2 +
4fbe94
 src/shared/bus-unit-util.c            | 28 +++++++++
4fbe94
 src/systemctl/systemctl.c             | 18 +++++-
4fbe94
 18 files changed, 405 insertions(+), 5 deletions(-)
4fbe94
4fbe94
diff --git a/man/systemd-system.conf.xml b/man/systemd-system.conf.xml
4fbe94
index ab23779ec0..988c4e7665 100644
4fbe94
--- a/man/systemd-system.conf.xml
4fbe94
+++ b/man/systemd-system.conf.xml
4fbe94
@@ -132,6 +132,25 @@
4fbe94
         anymore.</para></listitem>
4fbe94
       </varlistentry>
4fbe94
 
4fbe94
+      <varlistentry>
4fbe94
+        <term><varname>NUMAPolicy=</varname></term>
4fbe94
+
4fbe94
+        <listitem><para>Configures the NUMA memory policy for the service manager and the default NUMA memory policy
4fbe94
+        for all forked off processes. Individual services may override the default policy with the
4fbe94
+        <varname>NUMAPolicy=</varname> setting in unit files, see
4fbe94
+        <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>
4fbe94
+      </varlistentry>
4fbe94
+
4fbe94
+      <varlistentry>
4fbe94
+        <term><varname>NUMAMask=</varname></term>
4fbe94
+
4fbe94
+        <listitem><para>Configures the NUMA node mask that will be associated with the selected NUMA policy. Note that
4fbe94
+        <option>default</option> and <option>local</option> NUMA policies don't require explicit NUMA node mask and
4fbe94
+        value of the option can be empty. Similarly to <varname>NUMAPolicy=</varname>, value can be overriden
4fbe94
+        by individual services in unit files, see
4fbe94
+        <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>
4fbe94
+      </varlistentry>
4fbe94
+
4fbe94
       <varlistentry>
4fbe94
         <term><varname>RuntimeWatchdogSec=</varname></term>
4fbe94
         <term><varname>ShutdownWatchdogSec=</varname></term>
4fbe94
diff --git a/man/systemd.exec.xml b/man/systemd.exec.xml
4fbe94
index 342b8385bc..87fb8b34f4 100644
4fbe94
--- a/man/systemd.exec.xml
4fbe94
+++ b/man/systemd.exec.xml
4fbe94
@@ -710,6 +710,28 @@ CapabilityBoundingSet=~CAP_B CAP_C</programlisting>
4fbe94
         details.</para></listitem>
4fbe94
       </varlistentry>
4fbe94
 
4fbe94
+      <varlistentry>
4fbe94
+        <term><varname>NUMAPolicy=</varname></term>
4fbe94
+
4fbe94
+        <listitem><para>Controls the NUMA memory policy of the executed processes. Takes a policy type, one of:
4fbe94
+        <option>default</option>, <option>preferred</option>, <option>bind</option>, <option>interleave</option> and
4fbe94
+        <option>local</option>. A list of NUMA nodes that should be associated with the policy must be specified
4fbe94
+        in <varname>NUMAMask=</varname>. For more details on each policy please see,
4fbe94
+        <citerefentry><refentrytitle>set_mempolicy</refentrytitle><manvolnum>2</manvolnum></citerefentry>. For overall
4fbe94
+        overview of NUMA support in Linux see,
4fbe94
+        <citerefentry><refentrytitle>numa</refentrytitle><manvolnum>7</manvolnum></citerefentry>
4fbe94
+        </para></listitem>
4fbe94
+      </varlistentry>
4fbe94
+
4fbe94
+      <varlistentry>
4fbe94
+        <term><varname>NUMAMask=</varname></term>
4fbe94
+
4fbe94
+        <listitem><para>Controls the NUMA node list which will be applied alongside with selected NUMA policy.
4fbe94
+        Takes a list of NUMA nodes and has the same syntax as a list of CPUs for <varname>CPUAffinity=</varname>
4fbe94
+        option. Note that the list of NUMA nodes is not required for <option>default</option> and <option>local</option>
4fbe94
+        policies and for <option>preferred</option> policy we expect a single NUMA node.</para></listitem>
4fbe94
+      </varlistentry>
4fbe94
+
4fbe94
       <varlistentry>
4fbe94
         <term><varname>IOSchedulingClass=</varname></term>
4fbe94
 
4fbe94
@@ -2709,6 +2731,12 @@ StandardInputData=SWNrIHNpdHplIGRhIHVuJyBlc3NlIEtsb3BzLAp1ZmYgZWVtYWwga2xvcHAncy
4fbe94
             <entry><constant>EXIT_CONFIGURATION_DIRECTORY</constant></entry>
4fbe94
             <entry>Failed to set up unit's configuration directory. See <varname>ConfigurationDirectory=</varname> above.</entry>
4fbe94
           </row>
4fbe94
+          <row>
4fbe94
+            <entry>242</entry>
4fbe94
+            <entry><constant>EXIT_NUMA_POLICY</constant></entry>
4fbe94
+            <entry>Failed to set up unit's NUMA memory policy. See <varname>NUMAPolicy=</varname> and <varname>NUMAMask=</varname>above.</entry>
4fbe94
+          </row>
4fbe94
+
4fbe94
         
4fbe94
       </tgroup>
4fbe94
     
4fbe94
diff --git a/meson.build b/meson.build
4fbe94
index 613a5133b6..fe82ca4ac2 100644
4fbe94
--- a/meson.build
4fbe94
+++ b/meson.build
4fbe94
@@ -501,6 +501,10 @@ foreach ident : [
4fbe94
                                  #include <unistd.h>'''],
4fbe94
         ['explicit_bzero' ,   '''#include <string.h>'''],
4fbe94
         ['reallocarray',      '''#include <malloc.h>'''],
4fbe94
+        ['set_mempolicy',     '''#include <stdlib.h>
4fbe94
+                                 #include <unistd.h>'''],
4fbe94
+        ['get_mempolicy',     '''#include <stdlib.h>
4fbe94
+                                 #include <unistd.h>'''],
4fbe94
 ]
4fbe94
 
4fbe94
         have = cc.has_function(ident[0], prefix : ident[1], args : '-D_GNU_SOURCE')
4fbe94
diff --git a/src/basic/cpu-set-util.c b/src/basic/cpu-set-util.c
4fbe94
index 103b9703b3..36cb017ae7 100644
4fbe94
--- a/src/basic/cpu-set-util.c
4fbe94
+++ b/src/basic/cpu-set-util.c
4fbe94
@@ -10,11 +10,17 @@
4fbe94
 
4fbe94
 #include "alloc-util.h"
4fbe94
 #include "cpu-set-util.h"
4fbe94
+#include "dirent-util.h"
4fbe94
 #include "extract-word.h"
4fbe94
+#include "fd-util.h"
4fbe94
 #include "log.h"
4fbe94
 #include "macro.h"
4fbe94
+#include "missing.h"
4fbe94
 #include "parse-util.h"
4fbe94
+#include "stat-util.h"
4fbe94
 #include "string-util.h"
4fbe94
+#include "string-table.h"
4fbe94
+#include "strv.h"
4fbe94
 #include "util.h"
4fbe94
 
4fbe94
 char* cpu_set_to_string(const CPUSet *a) {
4fbe94
@@ -290,3 +296,88 @@ int cpu_set_from_dbus(const uint8_t *bits, size_t size, CPUSet *set) {
4fbe94
         s = (CPUSet) {};
4fbe94
         return 0;
4fbe94
 }
4fbe94
+
4fbe94
+bool numa_policy_is_valid(const NUMAPolicy *policy) {
4fbe94
+        assert(policy);
4fbe94
+
4fbe94
+        if (!mpol_is_valid(numa_policy_get_type(policy)))
4fbe94
+                return false;
4fbe94
+
4fbe94
+        if (!policy->nodes.set &&
4fbe94
+            !IN_SET(numa_policy_get_type(policy), MPOL_DEFAULT, MPOL_LOCAL, MPOL_PREFERRED))
4fbe94
+                return false;
4fbe94
+
4fbe94
+        if (policy->nodes.set &&
4fbe94
+            numa_policy_get_type(policy) == MPOL_PREFERRED &&
4fbe94
+            CPU_COUNT_S(policy->nodes.allocated, policy->nodes.set) != 1)
4fbe94
+                return false;
4fbe94
+
4fbe94
+        return true;
4fbe94
+}
4fbe94
+
4fbe94
+static int numa_policy_to_mempolicy(const NUMAPolicy *policy, unsigned long *ret_maxnode, unsigned long **ret_nodes) {
4fbe94
+        unsigned node, bits = 0, ulong_bits;
4fbe94
+        _cleanup_free_ unsigned long *out = NULL;
4fbe94
+
4fbe94
+        assert(policy);
4fbe94
+        assert(ret_maxnode);
4fbe94
+        assert(ret_nodes);
4fbe94
+
4fbe94
+        if (IN_SET(numa_policy_get_type(policy), MPOL_DEFAULT, MPOL_LOCAL) ||
4fbe94
+            (numa_policy_get_type(policy) == MPOL_PREFERRED && !policy->nodes.set)) {
4fbe94
+                *ret_nodes = NULL;
4fbe94
+                *ret_maxnode = 0;
4fbe94
+                return 0;
4fbe94
+        }
4fbe94
+
4fbe94
+        bits = policy->nodes.allocated * 8;
4fbe94
+        ulong_bits = sizeof(unsigned long) * 8;
4fbe94
+
4fbe94
+        out = new0(unsigned long, DIV_ROUND_UP(policy->nodes.allocated, sizeof(unsigned long)));
4fbe94
+        if (!out)
4fbe94
+                return -ENOMEM;
4fbe94
+
4fbe94
+        /* We don't make any assumptions about internal type libc is using to store NUMA node mask.
4fbe94
+           Hence we need to convert the node mask to the representation expected by set_mempolicy() */
4fbe94
+        for (node = 0; node < bits; node++)
4fbe94
+                if (CPU_ISSET_S(node, policy->nodes.allocated, policy->nodes.set))
4fbe94
+                        out[node / ulong_bits] |= 1ul << (node % ulong_bits);
4fbe94
+
4fbe94
+        *ret_nodes = TAKE_PTR(out);
4fbe94
+        *ret_maxnode = bits + 1;
4fbe94
+        return 0;
4fbe94
+}
4fbe94
+
4fbe94
+int apply_numa_policy(const NUMAPolicy *policy) {
4fbe94
+        int r;
4fbe94
+        _cleanup_free_ unsigned long *nodes = NULL;
4fbe94
+        unsigned long maxnode;
4fbe94
+
4fbe94
+        assert(policy);
4fbe94
+
4fbe94
+        if (get_mempolicy(NULL, NULL, 0, 0, 0) < 0 && errno == ENOSYS)
4fbe94
+                return -EOPNOTSUPP;
4fbe94
+
4fbe94
+        if (!numa_policy_is_valid(policy))
4fbe94
+                return -EINVAL;
4fbe94
+
4fbe94
+        r = numa_policy_to_mempolicy(policy, &maxnode, &nodes);
4fbe94
+        if (r < 0)
4fbe94
+                return r;
4fbe94
+
4fbe94
+        r = set_mempolicy(numa_policy_get_type(policy), nodes, maxnode);
4fbe94
+        if (r < 0)
4fbe94
+                return -errno;
4fbe94
+
4fbe94
+        return 0;
4fbe94
+}
4fbe94
+
4fbe94
+static const char* const mpol_table[] = {
4fbe94
+        [MPOL_DEFAULT]    = "default",
4fbe94
+        [MPOL_PREFERRED]  = "preferred",
4fbe94
+        [MPOL_BIND]       = "bind",
4fbe94
+        [MPOL_INTERLEAVE] = "interleave",
4fbe94
+        [MPOL_LOCAL]      = "local",
4fbe94
+};
4fbe94
+
4fbe94
+DEFINE_STRING_TABLE_LOOKUP(mpol, int);
4fbe94
diff --git a/src/basic/cpu-set-util.h b/src/basic/cpu-set-util.h
4fbe94
index ec640b2ec9..295028cb54 100644
4fbe94
--- a/src/basic/cpu-set-util.h
4fbe94
+++ b/src/basic/cpu-set-util.h
4fbe94
@@ -8,6 +8,7 @@
4fbe94
 #include <sched.h>
4fbe94
 
4fbe94
 #include "macro.h"
4fbe94
+#include "missing.h"
4fbe94
 
4fbe94
 /* This wraps the libc interface with a variable to keep the allocated size. */
4fbe94
 typedef struct CPUSet {
4fbe94
@@ -52,3 +53,30 @@ int cpu_set_to_dbus(const CPUSet *set, uint8_t **ret, size_t *allocated);
4fbe94
 int cpu_set_from_dbus(const uint8_t *bits, size_t size, CPUSet *set);
4fbe94
 
4fbe94
 int cpus_in_affinity_mask(void);
4fbe94
+
4fbe94
+static inline bool mpol_is_valid(int t) {
4fbe94
+        return t >= MPOL_DEFAULT && t <= MPOL_LOCAL;
4fbe94
+}
4fbe94
+
4fbe94
+typedef struct NUMAPolicy {
4fbe94
+        /* Always use numa_policy_get_type() to read the value */
4fbe94
+        int type;
4fbe94
+        CPUSet nodes;
4fbe94
+} NUMAPolicy;
4fbe94
+
4fbe94
+bool numa_policy_is_valid(const NUMAPolicy *p);
4fbe94
+
4fbe94
+static inline int numa_policy_get_type(const NUMAPolicy *p) {
4fbe94
+        return p->type < 0 ? (p->nodes.set ? MPOL_PREFERRED : -1) : p->type;
4fbe94
+}
4fbe94
+
4fbe94
+static inline void numa_policy_reset(NUMAPolicy *p) {
4fbe94
+        assert(p);
4fbe94
+        cpu_set_reset(&p->nodes);
4fbe94
+        p->type = -1;
4fbe94
+}
4fbe94
+
4fbe94
+int apply_numa_policy(const NUMAPolicy *policy);
4fbe94
+
4fbe94
+const char* mpol_to_string(int i) _const_;
4fbe94
+int mpol_from_string(const char *s) _pure_;
4fbe94
diff --git a/src/basic/exit-status.c b/src/basic/exit-status.c
4fbe94
index 21af8c4c71..0a7a53b73d 100644
4fbe94
--- a/src/basic/exit-status.c
4fbe94
+++ b/src/basic/exit-status.c
4fbe94
@@ -155,6 +155,9 @@ const char* exit_status_to_string(int status, ExitStatusLevel level) {
4fbe94
 
4fbe94
                 case EXIT_CONFIGURATION_DIRECTORY:
4fbe94
                         return "CONFIGURATION_DIRECTORY";
4fbe94
+
4fbe94
+                case EXIT_NUMA_POLICY:
4fbe94
+                        return "NUMA_POLICY";
4fbe94
                 }
4fbe94
         }
4fbe94
 
4fbe94
diff --git a/src/basic/exit-status.h b/src/basic/exit-status.h
4fbe94
index c41e8b82c3..dc284aacb1 100644
4fbe94
--- a/src/basic/exit-status.h
4fbe94
+++ b/src/basic/exit-status.h
4fbe94
@@ -69,6 +69,7 @@ enum {
4fbe94
         EXIT_CACHE_DIRECTORY,
4fbe94
         EXIT_LOGS_DIRECTORY, /* 240 */
4fbe94
         EXIT_CONFIGURATION_DIRECTORY,
4fbe94
+        EXIT_NUMA_POLICY,
4fbe94
 };
4fbe94
 
4fbe94
 typedef enum ExitStatusLevel {
4fbe94
diff --git a/src/basic/missing_syscall.h b/src/basic/missing_syscall.h
4fbe94
index 93c60458bf..014dd2b326 100644
4fbe94
--- a/src/basic/missing_syscall.h
4fbe94
+++ b/src/basic/missing_syscall.h
4fbe94
@@ -428,3 +428,46 @@ static inline ssize_t missing_statx(int dfd, const char *filename, unsigned flag
4fbe94
 
4fbe94
 #  define statx missing_statx
4fbe94
 #endif
4fbe94
+
4fbe94
+#if !HAVE_SET_MEMPOLICY
4fbe94
+
4fbe94
+enum {
4fbe94
+        MPOL_DEFAULT,
4fbe94
+        MPOL_PREFERRED,
4fbe94
+        MPOL_BIND,
4fbe94
+        MPOL_INTERLEAVE,
4fbe94
+        MPOL_LOCAL,
4fbe94
+};
4fbe94
+
4fbe94
+static inline long missing_set_mempolicy(int mode, const unsigned long *nodemask,
4fbe94
+                           unsigned long maxnode) {
4fbe94
+        long i;
4fbe94
+#  ifdef __NR_set_mempolicy
4fbe94
+        i = syscall(__NR_set_mempolicy, mode, nodemask, maxnode);
4fbe94
+#  else
4fbe94
+        errno = ENOSYS;
4fbe94
+        i = -1;
4fbe94
+#  endif
4fbe94
+        return i;
4fbe94
+}
4fbe94
+
4fbe94
+#  define set_mempolicy missing_set_mempolicy
4fbe94
+#endif
4fbe94
+
4fbe94
+
4fbe94
+#if !HAVE_GET_MEMPOLICY
4fbe94
+static inline long missing_get_mempolicy(int *mode, unsigned long *nodemask,
4fbe94
+                           unsigned long maxnode, void *addr,
4fbe94
+                           unsigned long flags) {
4fbe94
+        long i;
4fbe94
+#  ifdef __NR_get_mempolicy
4fbe94
+        i = syscall(__NR_get_mempolicy, mode, nodemask, maxnode, addr, flags);
4fbe94
+#  else
4fbe94
+        errno = ENOSYS;
4fbe94
+        i = -1;
4fbe94
+#  endif
4fbe94
+        return i;
4fbe94
+}
4fbe94
+
4fbe94
+#define get_mempolicy missing_get_mempolicy
4fbe94
+#endif
4fbe94
diff --git a/src/core/dbus-execute.c b/src/core/dbus-execute.c
4fbe94
index 50ea71a281..198f149210 100644
4fbe94
--- a/src/core/dbus-execute.c
4fbe94
+++ b/src/core/dbus-execute.c
4fbe94
@@ -223,6 +223,48 @@ static int property_get_cpu_affinity(
4fbe94
         return sd_bus_message_append_array(reply, 'y', c->cpu_set.set, c->cpu_set.allocated);
4fbe94
 }
4fbe94
 
4fbe94
+static int property_get_numa_mask(
4fbe94
+                sd_bus *bus,
4fbe94
+                const char *path,
4fbe94
+                const char *interface,
4fbe94
+                const char *property,
4fbe94
+                sd_bus_message *reply,
4fbe94
+                void *userdata,
4fbe94
+                sd_bus_error *error) {
4fbe94
+
4fbe94
+        ExecContext *c = userdata;
4fbe94
+        _cleanup_free_ uint8_t *array = NULL;
4fbe94
+        size_t allocated;
4fbe94
+
4fbe94
+        assert(bus);
4fbe94
+        assert(reply);
4fbe94
+        assert(c);
4fbe94
+
4fbe94
+        (void) cpu_set_to_dbus(&c->numa_policy.nodes, &array, &allocated);
4fbe94
+
4fbe94
+        return sd_bus_message_append_array(reply, 'y', array, allocated);
4fbe94
+}
4fbe94
+
4fbe94
+static int property_get_numa_policy(
4fbe94
+                sd_bus *bus,
4fbe94
+                const char *path,
4fbe94
+                const char *interface,
4fbe94
+                const char *property,
4fbe94
+                sd_bus_message *reply,
4fbe94
+                void *userdata,
4fbe94
+                sd_bus_error *error) {
4fbe94
+        ExecContext *c = userdata;
4fbe94
+        int32_t policy;
4fbe94
+
4fbe94
+        assert(bus);
4fbe94
+        assert(reply);
4fbe94
+        assert(c);
4fbe94
+
4fbe94
+        policy = numa_policy_get_type(&c->numa_policy);
4fbe94
+
4fbe94
+        return sd_bus_message_append_basic(reply, 'i', &policy);
4fbe94
+}
4fbe94
+
4fbe94
 static int property_get_timer_slack_nsec(
4fbe94
                 sd_bus *bus,
4fbe94
                 const char *path,
4fbe94
@@ -698,6 +740,8 @@ const sd_bus_vtable bus_exec_vtable[] = {
4fbe94
         SD_BUS_PROPERTY("CPUSchedulingPolicy", "i", property_get_cpu_sched_policy, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
         SD_BUS_PROPERTY("CPUSchedulingPriority", "i", property_get_cpu_sched_priority, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
         SD_BUS_PROPERTY("CPUAffinity", "ay", property_get_cpu_affinity, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
+        SD_BUS_PROPERTY("NUMAPolicy", "i", property_get_numa_policy, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
+        SD_BUS_PROPERTY("NUMAMask", "ay", property_get_numa_mask, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
         SD_BUS_PROPERTY("TimerSlackNSec", "t", property_get_timer_slack_nsec, 0, SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
         SD_BUS_PROPERTY("CPUSchedulingResetOnFork", "b", bus_property_get_bool, offsetof(ExecContext, cpu_sched_reset_on_fork), SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
         SD_BUS_PROPERTY("NonBlocking", "b", bus_property_get_bool, offsetof(ExecContext, non_blocking), SD_BUS_VTABLE_PROPERTY_CONST),
4fbe94
@@ -1550,9 +1594,10 @@ int bus_exec_context_set_transient_property(
4fbe94
                 return 1;
4fbe94
         }
4fbe94
 #endif
4fbe94
-        if (streq(name, "CPUAffinity")) {
4fbe94
+        if (STR_IN_SET(name, "CPUAffinity", "NUMAMask")) {
4fbe94
                 const void *a;
4fbe94
                 size_t n;
4fbe94
+                bool affinity = streq(name, "CPUAffinity");
4fbe94
                 _cleanup_(cpu_set_reset) CPUSet set = {};
4fbe94
 
4fbe94
                 r = sd_bus_message_read_array(message, 'y', &a, &n);
4fbe94
@@ -1565,7 +1610,7 @@ int bus_exec_context_set_transient_property(
4fbe94
 
4fbe94
                 if (!UNIT_WRITE_FLAGS_NOOP(flags)) {
4fbe94
                         if (n == 0) {
4fbe94
-                                cpu_set_reset(&c->cpu_set);
4fbe94
+                                cpu_set_reset(affinity ? &c->cpu_set : &c->numa_policy.nodes);
4fbe94
                                 unit_write_settingf(u, flags, name, "%s=", name);
4fbe94
                         } else {
4fbe94
                                 _cleanup_free_ char *str = NULL;
4fbe94
@@ -1577,7 +1622,7 @@ int bus_exec_context_set_transient_property(
4fbe94
                                 /* We forego any optimizations here, and always create the structure using
4fbe94
                                  * cpu_set_add_all(), because we don't want to care if the existing size we
4fbe94
                                  * got over dbus is appropriate. */
4fbe94
-                                r = cpu_set_add_all(&c->cpu_set, &set);
4fbe94
+                                r = cpu_set_add_all(affinity ? &c->cpu_set : &c->numa_policy.nodes, &set);
4fbe94
                                 if (r < 0)
4fbe94
                                         return r;
4fbe94
 
4fbe94
@@ -1587,6 +1632,20 @@ int bus_exec_context_set_transient_property(
4fbe94
 
4fbe94
                 return 1;
4fbe94
 
4fbe94
+        } else if (streq(name, "NUMAPolicy")) {
4fbe94
+                int32_t type;
4fbe94
+
4fbe94
+                r = sd_bus_message_read(message, "i", &type);
4fbe94
+                if (r < 0)
4fbe94
+                        return r;
4fbe94
+
4fbe94
+                if (!mpol_is_valid(type))
4fbe94
+                        return sd_bus_error_setf(error, SD_BUS_ERROR_INVALID_ARGS, "Invalid NUMAPolicy value: %i", type);
4fbe94
+
4fbe94
+                if (!UNIT_WRITE_FLAGS_NOOP(flags))
4fbe94
+                        c->numa_policy.type = type;
4fbe94
+
4fbe94
+                return 1;
4fbe94
         } else if (streq(name, "IOSchedulingClass")) {
4fbe94
                 int32_t q;
4fbe94
 
4fbe94
diff --git a/src/core/execute.c b/src/core/execute.c
4fbe94
index bc26aa66e7..56aa89e1ec 100644
4fbe94
--- a/src/core/execute.c
4fbe94
+++ b/src/core/execute.c
4fbe94
@@ -2997,6 +2997,16 @@ static int exec_child(
4fbe94
                         return log_unit_error_errno(unit, errno, "Failed to set up CPU affinity: %m");
4fbe94
                 }
4fbe94
 
4fbe94
+        if (mpol_is_valid(numa_policy_get_type(&context->numa_policy))) {
4fbe94
+                r = apply_numa_policy(&context->numa_policy);
4fbe94
+                if (r == -EOPNOTSUPP)
4fbe94
+                        log_unit_debug_errno(unit, r, "NUMA support not available, ignoring.");
4fbe94
+                else if (r < 0) {
4fbe94
+                        *exit_status = EXIT_NUMA_POLICY;
4fbe94
+                        return log_unit_error_errno(unit, r, "Failed to set NUMA memory policy: %m");
4fbe94
+                }
4fbe94
+        }
4fbe94
+
4fbe94
         if (context->ioprio_set)
4fbe94
                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, context->ioprio) < 0) {
4fbe94
                         *exit_status = EXIT_IOPRIO;
4fbe94
@@ -3651,6 +3661,7 @@ void exec_context_init(ExecContext *c) {
4fbe94
         assert_cc(NAMESPACE_FLAGS_INITIAL != NAMESPACE_FLAGS_ALL);
4fbe94
         c->restrict_namespaces = NAMESPACE_FLAGS_INITIAL;
4fbe94
         c->log_level_max = -1;
4fbe94
+        numa_policy_reset(&c->numa_policy);
4fbe94
 }
4fbe94
 
4fbe94
 void exec_context_done(ExecContext *c) {
4fbe94
@@ -3695,6 +3706,7 @@ void exec_context_done(ExecContext *c) {
4fbe94
         c->n_temporary_filesystems = 0;
4fbe94
 
4fbe94
         cpu_set_reset(&c->cpu_set);
4fbe94
+        numa_policy_reset(&c->numa_policy);
4fbe94
 
4fbe94
         c->utmp_id = mfree(c->utmp_id);
4fbe94
         c->selinux_context = mfree(c->selinux_context);
4fbe94
@@ -4104,6 +4116,14 @@ void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix) {
4fbe94
                 fprintf(f, "%sCPUAffinity: %s\n", prefix, affinity);
4fbe94
         }
4fbe94
 
4fbe94
+        if (mpol_is_valid(numa_policy_get_type(&c->numa_policy))) {
4fbe94
+                _cleanup_free_ char *nodes = NULL;
4fbe94
+
4fbe94
+                nodes = cpu_set_to_range_string(&c->numa_policy.nodes);
4fbe94
+                fprintf(f, "%sNUMAPolicy: %s\n", prefix, mpol_to_string(numa_policy_get_type(&c->numa_policy)));
4fbe94
+                fprintf(f, "%sNUMAMask: %s\n", prefix, strnull(nodes));
4fbe94
+        }
4fbe94
+
4fbe94
         if (c->timer_slack_nsec != NSEC_INFINITY)
4fbe94
                 fprintf(f, "%sTimerSlackNSec: "NSEC_FMT "\n", prefix, c->timer_slack_nsec);
4fbe94
 
4fbe94
diff --git a/src/core/execute.h b/src/core/execute.h
4fbe94
index e1e7a494cd..b2eb55f8f5 100644
4fbe94
--- a/src/core/execute.h
4fbe94
+++ b/src/core/execute.h
4fbe94
@@ -150,6 +150,7 @@ struct ExecContext {
4fbe94
         int cpu_sched_priority;
4fbe94
 
4fbe94
         CPUSet cpu_set;
4fbe94
+        NUMAPolicy numa_policy;
4fbe94
 
4fbe94
         ExecInput std_input;
4fbe94
         ExecOutput std_output;
4fbe94
diff --git a/src/core/load-fragment-gperf.gperf.m4 b/src/core/load-fragment-gperf.gperf.m4
4fbe94
index 1066bcfb8f..cdf4d14c4e 100644
4fbe94
--- a/src/core/load-fragment-gperf.gperf.m4
4fbe94
+++ b/src/core/load-fragment-gperf.gperf.m4
4fbe94
@@ -36,6 +36,8 @@ $1.CPUSchedulingPolicy,          config_parse_exec_cpu_sched_policy, 0,
4fbe94
 $1.CPUSchedulingPriority,        config_parse_exec_cpu_sched_prio,   0,                             offsetof($1, exec_context)
4fbe94
 $1.CPUSchedulingResetOnFork,     config_parse_bool,                  0,                             offsetof($1, exec_context.cpu_sched_reset_on_fork)
4fbe94
 $1.CPUAffinity,                  config_parse_exec_cpu_affinity,     0,                             offsetof($1, exec_context)
4fbe94
+$1.NUMAPolicy,                   config_parse_numa_policy,           0,                             offsetof($1, exec_context.numa_policy.type)
4fbe94
+$1.NUMAMask,                     config_parse_numa_mask,             0,                             offsetof($1, exec_context.numa_policy)
4fbe94
 $1.UMask,                        config_parse_mode,                  0,                             offsetof($1, exec_context.umask)
4fbe94
 $1.Environment,                  config_parse_environ,               0,                             offsetof($1, exec_context.environment)
4fbe94
 $1.EnvironmentFile,              config_parse_unit_env_file,         0,                             offsetof($1, exec_context.environment_files)
4fbe94
diff --git a/src/core/load-fragment.c b/src/core/load-fragment.c
4fbe94
index 34ae834188..35dd595098 100644
4fbe94
--- a/src/core/load-fragment.c
4fbe94
+++ b/src/core/load-fragment.c
4fbe94
@@ -93,6 +93,7 @@ DEFINE_CONFIG_PARSE_PTR(config_parse_blockio_weight, cg_blkio_weight_parse, uint
4fbe94
 DEFINE_CONFIG_PARSE_PTR(config_parse_cg_weight, cg_weight_parse, uint64_t, "Invalid weight");
4fbe94
 DEFINE_CONFIG_PARSE_PTR(config_parse_cpu_shares, cg_cpu_shares_parse, uint64_t, "Invalid CPU shares");
4fbe94
 DEFINE_CONFIG_PARSE_PTR(config_parse_exec_mount_flags, mount_propagation_flags_from_string, unsigned long, "Failed to parse mount flag");
4fbe94
+DEFINE_CONFIG_PARSE_ENUM_WITH_DEFAULT(config_parse_numa_policy, mpol, int, -1, "Invalid NUMA policy type");
4fbe94
 
4fbe94
 int config_parse_unit_deps(
4fbe94
                 const char *unit,
4fbe94
@@ -1159,6 +1160,33 @@ int config_parse_exec_cpu_sched_policy(const char *unit,
4fbe94
         return 0;
4fbe94
 }
4fbe94
 
4fbe94
+int config_parse_numa_mask(const char *unit,
4fbe94
+                           const char *filename,
4fbe94
+                           unsigned line,
4fbe94
+                           const char *section,
4fbe94
+                           unsigned section_line,
4fbe94
+                           const char *lvalue,
4fbe94
+                           int ltype,
4fbe94
+                           const char *rvalue,
4fbe94
+                           void *data,
4fbe94
+                           void *userdata) {
4fbe94
+        int r;
4fbe94
+        NUMAPolicy *p = data;
4fbe94
+
4fbe94
+        assert(filename);
4fbe94
+        assert(lvalue);
4fbe94
+        assert(rvalue);
4fbe94
+        assert(data);
4fbe94
+
4fbe94
+        r = parse_cpu_set_extend(rvalue, &p->nodes, true, unit, filename, line, lvalue);
4fbe94
+        if (r < 0) {
4fbe94
+                log_syntax(unit, LOG_ERR, filename, line, r, "Failed to parse NUMA node mask, ignoring: %s", rvalue);
4fbe94
+                return 0;
4fbe94
+        }
4fbe94
+
4fbe94
+        return r;
4fbe94
+}
4fbe94
+
4fbe94
 int config_parse_exec_cpu_sched_prio(const char *unit,
4fbe94
                                      const char *filename,
4fbe94
                                      unsigned line,
4fbe94
diff --git a/src/core/load-fragment.h b/src/core/load-fragment.h
4fbe94
index dad281ef72..f2ca1b8ee7 100644
4fbe94
--- a/src/core/load-fragment.h
4fbe94
+++ b/src/core/load-fragment.h
4fbe94
@@ -102,6 +102,8 @@ CONFIG_PARSER_PROTOTYPE(config_parse_job_timeout_sec);
4fbe94
 CONFIG_PARSER_PROTOTYPE(config_parse_job_running_timeout_sec);
4fbe94
 CONFIG_PARSER_PROTOTYPE(config_parse_log_extra_fields);
4fbe94
 CONFIG_PARSER_PROTOTYPE(config_parse_collect_mode);
4fbe94
+CONFIG_PARSER_PROTOTYPE(config_parse_numa_policy);
4fbe94
+CONFIG_PARSER_PROTOTYPE(config_parse_numa_mask);
4fbe94
 
4fbe94
 /* gperf prototypes */
4fbe94
 const struct ConfigPerfItem* load_fragment_gperf_lookup(const char *key, GPERF_LEN_TYPE length);
4fbe94
diff --git a/src/core/main.c b/src/core/main.c
4fbe94
index c74dc641c1..83f9dd5878 100644
4fbe94
--- a/src/core/main.c
4fbe94
+++ b/src/core/main.c
4fbe94
@@ -134,6 +134,7 @@ static uint64_t arg_default_tasks_max;
4fbe94
 static sd_id128_t arg_machine_id;
4fbe94
 static EmergencyAction arg_cad_burst_action;
4fbe94
 static CPUSet arg_cpu_affinity;
4fbe94
+static NUMAPolicy arg_numa_policy;
4fbe94
 
4fbe94
 static int parse_configuration(void);
4fbe94
 
4fbe94
@@ -660,6 +661,8 @@ static int parse_config_file(void) {
4fbe94
                 { "Manager", "ShowStatus",                config_parse_show_status,      0, &arg_show_status                       },
4fbe94
                 { "Manager", "CPUAffinity",               config_parse_cpu_affinity2,    0, &arg_cpu_affinity                      },
4fbe94
                 { "Manager", "JoinControllers",           config_parse_join_controllers, 0, &arg_join_controllers                  },
4fbe94
+                { "Manager", "NUMAPolicy",                config_parse_numa_policy,      0, &arg_numa_policy.type                  },
4fbe94
+                { "Manager", "NUMAMask",                  config_parse_numa_mask,        0, &arg_numa_policy                       },
4fbe94
                 { "Manager", "RuntimeWatchdogSec",        config_parse_sec,              0, &arg_runtime_watchdog                  },
4fbe94
                 { "Manager", "ShutdownWatchdogSec",       config_parse_sec,              0, &arg_shutdown_watchdog                 },
4fbe94
                 { "Manager", "WatchdogDevice",            config_parse_path,             0, &arg_watchdog_device                   },
4fbe94
@@ -1501,6 +1504,27 @@ static void update_cpu_affinity(bool skip_setup) {
4fbe94
                 log_warning_errno(errno, "Failed to set CPU affinity: %m");
4fbe94
 }
4fbe94
 
4fbe94
+static void update_numa_policy(bool skip_setup) {
4fbe94
+        int r;
4fbe94
+        _cleanup_free_ char *nodes = NULL;
4fbe94
+        const char * policy = NULL;
4fbe94
+
4fbe94
+        if (skip_setup || !mpol_is_valid(numa_policy_get_type(&arg_numa_policy)))
4fbe94
+                return;
4fbe94
+
4fbe94
+        if (DEBUG_LOGGING) {
4fbe94
+                policy = mpol_to_string(numa_policy_get_type(&arg_numa_policy));
4fbe94
+                nodes = cpu_set_to_range_string(&arg_numa_policy.nodes);
4fbe94
+                log_debug("Setting NUMA policy to %s, with nodes %s.", strnull(policy), strnull(nodes));
4fbe94
+        }
4fbe94
+
4fbe94
+        r = apply_numa_policy(&arg_numa_policy);
4fbe94
+        if (r == -EOPNOTSUPP)
4fbe94
+                log_debug_errno(r, "NUMA support not available, ignoring.");
4fbe94
+        else if (r < 0)
4fbe94
+                log_warning_errno(r, "Failed to set NUMA memory policy: %m");
4fbe94
+}
4fbe94
+
4fbe94
 static void do_reexecute(
4fbe94
                 int argc,
4fbe94
                 char *argv[],
4fbe94
@@ -1672,6 +1696,7 @@ static int invoke_main_loop(
4fbe94
                         set_manager_defaults(m);
4fbe94
 
4fbe94
                         update_cpu_affinity(false);
4fbe94
+                        update_numa_policy(false);
4fbe94
 
4fbe94
                         if (saved_log_level >= 0)
4fbe94
                                 manager_override_log_level(m, saved_log_level);
4fbe94
@@ -1832,6 +1857,7 @@ static int initialize_runtime(
4fbe94
                 return 0;
4fbe94
 
4fbe94
         update_cpu_affinity(skip_setup);
4fbe94
+        update_numa_policy(skip_setup);
4fbe94
 
4fbe94
         if (arg_system) {
4fbe94
                 /* Make sure we leave a core dump without panicing the kernel. */
4fbe94
@@ -2011,6 +2037,7 @@ static void reset_arguments(void) {
4fbe94
         arg_cad_burst_action = EMERGENCY_ACTION_REBOOT_FORCE;
4fbe94
 
4fbe94
         cpu_set_reset(&arg_cpu_affinity);
4fbe94
+        numa_policy_reset(&arg_numa_policy);
4fbe94
 }
4fbe94
 
4fbe94
 static int parse_configuration(void) {
4fbe94
diff --git a/src/core/system.conf.in b/src/core/system.conf.in
4fbe94
index 653ec6b8c9..0d93fbf147 100644
4fbe94
--- a/src/core/system.conf.in
4fbe94
+++ b/src/core/system.conf.in
4fbe94
@@ -24,6 +24,8 @@
4fbe94
 #CtrlAltDelBurstAction=reboot-force
4fbe94
 #CPUAffinity=1 2
4fbe94
 #JoinControllers=cpu,cpuacct net_cls,net_prio
4fbe94
+#NUMAPolicy=default
4fbe94
+#NUMAMask=
4fbe94
 #RuntimeWatchdogSec=0
4fbe94
 #ShutdownWatchdogSec=10min
4fbe94
 #CapabilityBoundingSet=
4fbe94
diff --git a/src/shared/bus-unit-util.c b/src/shared/bus-unit-util.c
4fbe94
index ec8732c226..055edd6e22 100644
4fbe94
--- a/src/shared/bus-unit-util.c
4fbe94
+++ b/src/shared/bus-unit-util.c
4fbe94
@@ -947,6 +947,34 @@ static int bus_append_execute_property(sd_bus_message *m, const char *field, con
4fbe94
                 return bus_append_byte_array(m, field, array, allocated);
4fbe94
         }
4fbe94
 
4fbe94
+        if (streq(field, "NUMAPolicy")) {
4fbe94
+                r = mpol_from_string(eq);
4fbe94
+                if (r < 0)
4fbe94
+                        return log_error_errno(r, "Failed to parse %s value: %s", field, eq);
4fbe94
+
4fbe94
+                r = sd_bus_message_append(m, "(sv)", field, "i", (int32_t) r);
4fbe94
+                if (r < 0)
4fbe94
+                        return bus_log_create_error(r);
4fbe94
+
4fbe94
+                return 1;
4fbe94
+        }
4fbe94
+
4fbe94
+        if (streq(field, "NUMAMask")) {
4fbe94
+                _cleanup_(cpu_set_reset) CPUSet nodes = {};
4fbe94
+                _cleanup_free_ uint8_t *array = NULL;
4fbe94
+                size_t allocated;
4fbe94
+
4fbe94
+                r = parse_cpu_set(eq, &nodes);
4fbe94
+                if (r < 0)
4fbe94
+                        return log_error_errno(r, "Failed to parse %s value: %s", field, eq);
4fbe94
+
4fbe94
+                r = cpu_set_to_dbus(&nodes, &array, &allocated);
4fbe94
+                if (r < 0)
4fbe94
+                        return log_error_errno(r, "Failed to serialize NUMAMask: %m");
4fbe94
+
4fbe94
+                return bus_append_byte_array(m, field, array, allocated);
4fbe94
+        }
4fbe94
+
4fbe94
         if (STR_IN_SET(field, "RestrictAddressFamilies", "SystemCallFilter")) {
4fbe94
                 int whitelist = 1;
4fbe94
                 const char *p = eq;
4fbe94
diff --git a/src/systemctl/systemctl.c b/src/systemctl/systemctl.c
4fbe94
index 0154b300a3..7274921e6d 100644
4fbe94
--- a/src/systemctl/systemctl.c
4fbe94
+++ b/src/systemctl/systemctl.c
4fbe94
@@ -4573,6 +4573,20 @@ static int print_property(const char *name, sd_bus_message *m, bool value, bool
4fbe94
 
4fbe94
         switch (bus_type) {
4fbe94
 
4fbe94
+        case SD_BUS_TYPE_INT32:
4fbe94
+                if (streq(name, "NUMAPolicy")) {
4fbe94
+                        int32_t i;
4fbe94
+
4fbe94
+                        r = sd_bus_message_read_basic(m, bus_type, &i);
4fbe94
+                        if (r < 0)
4fbe94
+                                return r;
4fbe94
+
4fbe94
+                        print_prop(name, "%s", strna(mpol_to_string(i)));
4fbe94
+
4fbe94
+                        return 1;
4fbe94
+                }
4fbe94
+                break;
4fbe94
+
4fbe94
         case SD_BUS_TYPE_STRUCT:
4fbe94
 
4fbe94
                 if (contents[0] == SD_BUS_TYPE_UINT32 && streq(name, "Job")) {
4fbe94
@@ -4878,7 +4892,7 @@ static int print_property(const char *name, sd_bus_message *m, bool value, bool
4fbe94
                         print_prop(name, "%s", h);
4fbe94
 
4fbe94
                         return 1;
4fbe94
-                } else if (contents[0] == SD_BUS_TYPE_BYTE && streq(name, "CPUAffinity")) {
4fbe94
+                } else if (contents[0] == SD_BUS_TYPE_BYTE && STR_IN_SET(name, "CPUAffinity", "NUMAMask")) {
4fbe94
                         _cleanup_free_ char *affinity = NULL;
4fbe94
                         _cleanup_(cpu_set_reset) CPUSet set = {};
4fbe94
                         const void *a;
4fbe94
@@ -4890,7 +4904,7 @@ static int print_property(const char *name, sd_bus_message *m, bool value, bool
4fbe94
 
4fbe94
                         r = cpu_set_from_dbus(a, n, &set);
4fbe94
                         if (r < 0)
4fbe94
-                                return log_error_errno(r, "Failed to deserialize CPUAffinity: %m");
4fbe94
+                                return log_error_errno(r, "Failed to deserialize %s: %m", name);
4fbe94
 
4fbe94
                         affinity = cpu_set_to_range_string(&set);
4fbe94
                         if (!affinity)