From a735699a8287c19e043b7d2fe9a387a3938e1e2f Mon Sep 17 00:00:00 2001

From: =?UTF-8?q?Michal=20Sekleta=CC=81r?= <msekleta@redhat.com>

Date: Mon, 18 Nov 2019 12:50:11 +0100

Subject: [PATCH] core: introduce NUMAPolicy and NUMAMask options

Make possible to set NUMA allocation policy for manager. Manager's

policy is by default inherited to all forked off processes. However, it

is possible to override the policy on per-service basis. Currently we

support, these policies: default, prefer, bind, interleave, local.

See man 2 set_mempolicy for details on each policy.

Overall NUMA policy actually consists of two parts. Policy itself and

bitmask representing NUMA nodes where is policy effective. Node mask can

be specified using related option, NUMAMask. Default mask can be

overwritten on per-service level.

(cherry-picked from commit fe9c54b2188e6cd23262a319f96b13215f2c5e9c)

Resolves: #1734787

---

 man/systemd-system.conf.xml           | 19 ++++++

 man/systemd.exec.xml                  | 28 +++++++++

 meson.build                           |  4 ++

 src/basic/cpu-set-util.c              | 91 +++++++++++++++++++++++++++

 src/basic/cpu-set-util.h              | 28 +++++++++

 src/basic/exit-status.c               |  3 +

 src/basic/exit-status.h               |  1 +

 src/basic/missing_syscall.h           | 43 +++++++++++++

 src/core/dbus-execute.c               | 65 ++++++++++++++++++-

 src/core/execute.c                    | 20 ++++++

 src/core/execute.h                    |  1 +

 src/core/load-fragment-gperf.gperf.m4 |  2 +

 src/core/load-fragment.c              | 28 +++++++++

 src/core/load-fragment.h              |  2 +

 src/core/main.c                       | 27 ++++++++

 src/core/system.conf.in               |  2 +

 src/shared/bus-unit-util.c            | 28 +++++++++

 src/systemctl/systemctl.c             | 18 +++++-

 18 files changed, 405 insertions(+), 5 deletions(-)

diff --git a/man/systemd-system.conf.xml b/man/systemd-system.conf.xml

index ab23779ec0..988c4e7665 100644

--- a/man/systemd-system.conf.xml

+++ b/man/systemd-system.conf.xml

@@ -132,6 +132,25 @@

         anymore.</para></listitem>

       </varlistentry>

+      <varlistentry>

+        <term><varname>NUMAPolicy=</varname></term>

+

+        <listitem><para>Configures the NUMA memory policy for the service manager and the default NUMA memory policy

+        for all forked off processes. Individual services may override the default policy with the

+        <varname>NUMAPolicy=</varname> setting in unit files, see

+        <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>

+      </varlistentry>

+

+      <varlistentry>

+        <term><varname>NUMAMask=</varname></term>

+

+        <listitem><para>Configures the NUMA node mask that will be associated with the selected NUMA policy. Note that

+        <option>default</option> and <option>local</option> NUMA policies don't require explicit NUMA node mask and

+        value of the option can be empty. Similarly to <varname>NUMAPolicy=</varname>, value can be overriden

+        by individual services in unit files, see

+        <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>

+      </varlistentry>

+

       <varlistentry>

         <term><varname>RuntimeWatchdogSec=</varname></term>

         <term><varname>ShutdownWatchdogSec=</varname></term>

diff --git a/man/systemd.exec.xml b/man/systemd.exec.xml

index 342b8385bc..87fb8b34f4 100644

--- a/man/systemd.exec.xml

+++ b/man/systemd.exec.xml

@@ -710,6 +710,28 @@ CapabilityBoundingSet=~CAP_B CAP_C</programlisting>

         details.</para></listitem>

       </varlistentry>

+      <varlistentry>

+        <term><varname>NUMAPolicy=</varname></term>

+

+        <listitem><para>Controls the NUMA memory policy of the executed processes. Takes a policy type, one of:

+        <option>default</option>, <option>preferred</option>, <option>bind</option>, <option>interleave</option> and

+        <option>local</option>. A list of NUMA nodes that should be associated with the policy must be specified

+        in <varname>NUMAMask=</varname>. For more details on each policy please see,

+        <citerefentry><refentrytitle>set_mempolicy</refentrytitle><manvolnum>2</manvolnum></citerefentry>. For overall

+        overview of NUMA support in Linux see,

+        <citerefentry><refentrytitle>numa</refentrytitle><manvolnum>7</manvolnum></citerefentry>

+        </para></listitem>

+      </varlistentry>

+

+      <varlistentry>

+        <term><varname>NUMAMask=</varname></term>

+

+        <listitem><para>Controls the NUMA node list which will be applied alongside with selected NUMA policy.

+        Takes a list of NUMA nodes and has the same syntax as a list of CPUs for <varname>CPUAffinity=</varname>

+        option. Note that the list of NUMA nodes is not required for <option>default</option> and <option>local</option>

+        policies and for <option>preferred</option> policy we expect a single NUMA node.</para></listitem>

+      </varlistentry>

+

       <varlistentry>

         <term><varname>IOSchedulingClass=</varname></term>

@@ -2709,6 +2731,12 @@ StandardInputData=SWNrIHNpdHplIGRhIHVuJyBlc3NlIEtsb3BzLAp1ZmYgZWVtYWwga2xvcHAncy

             <entry><constant>EXIT_CONFIGURATION_DIRECTORY</constant></entry>

             <entry>Failed to set up unit's configuration directory. See <varname>ConfigurationDirectory=</varname> above.</entry>

           </row>

+          <row>

+            <entry>242</entry>

+            <entry><constant>EXIT_NUMA_POLICY</constant></entry>

+            <entry>Failed to set up unit's NUMA memory policy. See <varname>NUMAPolicy=</varname> and <varname>NUMAMask=</varname>above.</entry>

+          </row>

+

       </tgroup>

diff --git a/meson.build b/meson.build

index 613a5133b6..fe82ca4ac2 100644

--- a/meson.build

+++ b/meson.build

@@ -501,6 +501,10 @@ foreach ident : [

                                  #include <unistd.h>'''],

         ['explicit_bzero' ,   '''#include <string.h>'''],

         ['reallocarray',      '''#include <malloc.h>'''],

+        ['set_mempolicy',     '''#include <stdlib.h>

+                                 #include <unistd.h>'''],

+        ['get_mempolicy',     '''#include <stdlib.h>

+                                 #include <unistd.h>'''],

 ]

         have = cc.has_function(ident[0], prefix : ident[1], args : '-D_GNU_SOURCE')

diff --git a/src/basic/cpu-set-util.c b/src/basic/cpu-set-util.c

index 103b9703b3..36cb017ae7 100644

--- a/src/basic/cpu-set-util.c

+++ b/src/basic/cpu-set-util.c

@@ -10,11 +10,17 @@

 #include "alloc-util.h"

 #include "cpu-set-util.h"

+#include "dirent-util.h"

 #include "extract-word.h"

+#include "fd-util.h"

 #include "log.h"

 #include "macro.h"

+#include "missing.h"

 #include "parse-util.h"

+#include "stat-util.h"

 #include "string-util.h"

+#include "string-table.h"

+#include "strv.h"

 #include "util.h"

 char* cpu_set_to_string(const CPUSet *a) {

@@ -290,3 +296,88 @@ int cpu_set_from_dbus(const uint8_t *bits, size_t size, CPUSet *set) {

         s = (CPUSet) {};

         return 0;

 }

+

+bool numa_policy_is_valid(const NUMAPolicy *policy) {

+        assert(policy);

+

+        if (!mpol_is_valid(numa_policy_get_type(policy)))

+                return false;

+

+        if (!policy->nodes.set &&

+            !IN_SET(numa_policy_get_type(policy), MPOL_DEFAULT, MPOL_LOCAL, MPOL_PREFERRED))

+                return false;

+

+        if (policy->nodes.set &&

+            numa_policy_get_type(policy) == MPOL_PREFERRED &&

+            CPU_COUNT_S(policy->nodes.allocated, policy->nodes.set) != 1)

+                return false;

+

+        return true;

+}

+

+static int numa_policy_to_mempolicy(const NUMAPolicy *policy, unsigned long *ret_maxnode, unsigned long **ret_nodes) {

+        unsigned node, bits = 0, ulong_bits;

+        _cleanup_free_ unsigned long *out = NULL;

+

+        assert(policy);

+        assert(ret_maxnode);

+        assert(ret_nodes);

+

+        if (IN_SET(numa_policy_get_type(policy), MPOL_DEFAULT, MPOL_LOCAL) ||

+            (numa_policy_get_type(policy) == MPOL_PREFERRED && !policy->nodes.set)) {

+                *ret_nodes = NULL;

+                *ret_maxnode = 0;

+                return 0;

+        }

+

+        bits = policy->nodes.allocated * 8;

+        ulong_bits = sizeof(unsigned long) * 8;

+

+        out = new0(unsigned long, DIV_ROUND_UP(policy->nodes.allocated, sizeof(unsigned long)));

+        if (!out)

+                return -ENOMEM;

+

+        /* We don't make any assumptions about internal type libc is using to store NUMA node mask.

+           Hence we need to convert the node mask to the representation expected by set_mempolicy() */

+        for (node = 0; node < bits; node++)

+                if (CPU_ISSET_S(node, policy->nodes.allocated, policy->nodes.set))

+                        out[node / ulong_bits] |= 1ul << (node % ulong_bits);

+

+        *ret_nodes = TAKE_PTR(out);

+        *ret_maxnode = bits + 1;

+        return 0;

+}

+

+int apply_numa_policy(const NUMAPolicy *policy) {

+        int r;

+        _cleanup_free_ unsigned long *nodes = NULL;

+        unsigned long maxnode;

+

+        assert(policy);

+

+        if (get_mempolicy(NULL, NULL, 0, 0, 0) < 0 && errno == ENOSYS)

+                return -EOPNOTSUPP;

+

+        if (!numa_policy_is_valid(policy))

+                return -EINVAL;

+

+        r = numa_policy_to_mempolicy(policy, &maxnode, &nodes);

+        if (r < 0)

+                return r;

+

+        r = set_mempolicy(numa_policy_get_type(policy), nodes, maxnode);

+        if (r < 0)

+                return -errno;

+

+        return 0;

+}

+

+static const char* const mpol_table[] = {

+        [MPOL_DEFAULT]    = "default",

+        [MPOL_PREFERRED]  = "preferred",

+        [MPOL_BIND]       = "bind",

+        [MPOL_INTERLEAVE] = "interleave",

+        [MPOL_LOCAL]      = "local",

+};

+

+DEFINE_STRING_TABLE_LOOKUP(mpol, int);

diff --git a/src/basic/cpu-set-util.h b/src/basic/cpu-set-util.h

index ec640b2ec9..295028cb54 100644

--- a/src/basic/cpu-set-util.h

+++ b/src/basic/cpu-set-util.h

@@ -8,6 +8,7 @@

 #include <sched.h>

 #include "macro.h"

+#include "missing.h"

 /* This wraps the libc interface with a variable to keep the allocated size. */

 typedef struct CPUSet {

@@ -52,3 +53,30 @@ int cpu_set_to_dbus(const CPUSet *set, uint8_t **ret, size_t *allocated);

 int cpu_set_from_dbus(const uint8_t *bits, size_t size, CPUSet *set);

 int cpus_in_affinity_mask(void);

+

+static inline bool mpol_is_valid(int t) {

+        return t >= MPOL_DEFAULT && t <= MPOL_LOCAL;

+}

+

+typedef struct NUMAPolicy {

+        /* Always use numa_policy_get_type() to read the value */

+        int type;

+        CPUSet nodes;

+} NUMAPolicy;

+

+bool numa_policy_is_valid(const NUMAPolicy *p);

+

+static inline int numa_policy_get_type(const NUMAPolicy *p) {

+        return p->type < 0 ? (p->nodes.set ? MPOL_PREFERRED : -1) : p->type;

+}

+

+static inline void numa_policy_reset(NUMAPolicy *p) {

+        assert(p);

+        cpu_set_reset(&p->nodes);

+        p->type = -1;

+}

+

+int apply_numa_policy(const NUMAPolicy *policy);

+

+const char* mpol_to_string(int i) _const_;

+int mpol_from_string(const char *s) _pure_;

diff --git a/src/basic/exit-status.c b/src/basic/exit-status.c

index 21af8c4c71..0a7a53b73d 100644

--- a/src/basic/exit-status.c

+++ b/src/basic/exit-status.c

@@ -155,6 +155,9 @@ const char* exit_status_to_string(int status, ExitStatusLevel level) {

                 case EXIT_CONFIGURATION_DIRECTORY:

                         return "CONFIGURATION_DIRECTORY";

+

+                case EXIT_NUMA_POLICY:

+                        return "NUMA_POLICY";

                 }

         }

diff --git a/src/basic/exit-status.h b/src/basic/exit-status.h

index c41e8b82c3..dc284aacb1 100644

--- a/src/basic/exit-status.h

+++ b/src/basic/exit-status.h

@@ -69,6 +69,7 @@ enum {

         EXIT_CACHE_DIRECTORY,

         EXIT_LOGS_DIRECTORY, /* 240 */

         EXIT_CONFIGURATION_DIRECTORY,

+        EXIT_NUMA_POLICY,

 };

 typedef enum ExitStatusLevel {

diff --git a/src/basic/missing_syscall.h b/src/basic/missing_syscall.h

index 93c60458bf..014dd2b326 100644

--- a/src/basic/missing_syscall.h

+++ b/src/basic/missing_syscall.h

@@ -428,3 +428,46 @@ static inline ssize_t missing_statx(int dfd, const char *filename, unsigned flag

 #  define statx missing_statx

 #endif

+

+#if !HAVE_SET_MEMPOLICY

+

+enum {

+        MPOL_DEFAULT,

+        MPOL_PREFERRED,

+        MPOL_BIND,

+        MPOL_INTERLEAVE,

+        MPOL_LOCAL,

+};

+

+static inline long missing_set_mempolicy(int mode, const unsigned long *nodemask,

+                           unsigned long maxnode) {

+        long i;

+#  ifdef __NR_set_mempolicy

+        i = syscall(__NR_set_mempolicy, mode, nodemask, maxnode);

+#  else

+        errno = ENOSYS;

+        i = -1;

+#  endif

+        return i;

+}

+

+#  define set_mempolicy missing_set_mempolicy

+#endif

+

+

+#if !HAVE_GET_MEMPOLICY

+static inline long missing_get_mempolicy(int *mode, unsigned long *nodemask,

+                           unsigned long maxnode, void *addr,

+                           unsigned long flags) {

+        long i;

+#  ifdef __NR_get_mempolicy

+        i = syscall(__NR_get_mempolicy, mode, nodemask, maxnode, addr, flags);

+#  else

+        errno = ENOSYS;

+        i = -1;

+#  endif

+        return i;

+}

+

+#define get_mempolicy missing_get_mempolicy

+#endif

diff --git a/src/core/dbus-execute.c b/src/core/dbus-execute.c

index 50ea71a281..198f149210 100644

--- a/src/core/dbus-execute.c

+++ b/src/core/dbus-execute.c

@@ -223,6 +223,48 @@ static int property_get_cpu_affinity(

         return sd_bus_message_append_array(reply, 'y', c->cpu_set.set, c->cpu_set.allocated);

 }

+static int property_get_numa_mask(

+                sd_bus *bus,

+                const char *path,

+                const char *interface,

+                const char *property,

+                sd_bus_message *reply,

+                void *userdata,

+                sd_bus_error *error) {

+

+        ExecContext *c = userdata;

+        _cleanup_free_ uint8_t *array = NULL;

+        size_t allocated;

+

+        assert(bus);

+        assert(reply);

+        assert(c);

+

+        (void) cpu_set_to_dbus(&c->numa_policy.nodes, &array, &allocated);

+

+        return sd_bus_message_append_array(reply, 'y', array, allocated);

+}

+

+static int property_get_numa_policy(

+                sd_bus *bus,

+                const char *path,

+                const char *interface,

+                const char *property,

+                sd_bus_message *reply,

+                void *userdata,

+                sd_bus_error *error) {

+        ExecContext *c = userdata;

+        int32_t policy;

+

+        assert(bus);

+        assert(reply);

+        assert(c);

+

+        policy = numa_policy_get_type(&c->numa_policy);

+

+        return sd_bus_message_append_basic(reply, 'i', &policy);

+}

+

 static int property_get_timer_slack_nsec(

                 sd_bus *bus,

                 const char *path,

@@ -698,6 +740,8 @@ const sd_bus_vtable bus_exec_vtable[] = {

         SD_BUS_PROPERTY("CPUSchedulingPolicy", "i", property_get_cpu_sched_policy, 0, SD_BUS_VTABLE_PROPERTY_CONST),

         SD_BUS_PROPERTY("CPUSchedulingPriority", "i", property_get_cpu_sched_priority, 0, SD_BUS_VTABLE_PROPERTY_CONST),

         SD_BUS_PROPERTY("CPUAffinity", "ay", property_get_cpu_affinity, 0, SD_BUS_VTABLE_PROPERTY_CONST),

+        SD_BUS_PROPERTY("NUMAPolicy", "i", property_get_numa_policy, 0, SD_BUS_VTABLE_PROPERTY_CONST),

+        SD_BUS_PROPERTY("NUMAMask", "ay", property_get_numa_mask, 0, SD_BUS_VTABLE_PROPERTY_CONST),

         SD_BUS_PROPERTY("TimerSlackNSec", "t", property_get_timer_slack_nsec, 0, SD_BUS_VTABLE_PROPERTY_CONST),

         SD_BUS_PROPERTY("CPUSchedulingResetOnFork", "b", bus_property_get_bool, offsetof(ExecContext, cpu_sched_reset_on_fork), SD_BUS_VTABLE_PROPERTY_CONST),

         SD_BUS_PROPERTY("NonBlocking", "b", bus_property_get_bool, offsetof(ExecContext, non_blocking), SD_BUS_VTABLE_PROPERTY_CONST),

@@ -1550,9 +1594,10 @@ int bus_exec_context_set_transient_property(

                 return 1;

         }

 #endif

-        if (streq(name, "CPUAffinity")) {

+        if (STR_IN_SET(name, "CPUAffinity", "NUMAMask")) {

                 const void *a;

                 size_t n;

+                bool affinity = streq(name, "CPUAffinity");

                 _cleanup_(cpu_set_reset) CPUSet set = {};

                 r = sd_bus_message_read_array(message, 'y', &a, &n);

@@ -1565,7 +1610,7 @@ int bus_exec_context_set_transient_property(

                 if (!UNIT_WRITE_FLAGS_NOOP(flags)) {

                         if (n == 0) {

-                                cpu_set_reset(&c->cpu_set);

+                                cpu_set_reset(affinity ? &c->cpu_set : &c->numa_policy.nodes);

                                 unit_write_settingf(u, flags, name, "%s=", name);

                         } else {

                                 _cleanup_free_ char *str = NULL;

@@ -1577,7 +1622,7 @@ int bus_exec_context_set_transient_property(

                                 /* We forego any optimizations here, and always create the structure using

                                  * cpu_set_add_all(), because we don't want to care if the existing size we

                                  * got over dbus is appropriate. */

-                                r = cpu_set_add_all(&c->cpu_set, &set);

+                                r = cpu_set_add_all(affinity ? &c->cpu_set : &c->numa_policy.nodes, &set);

                                 if (r < 0)

                                         return r;

@@ -1587,6 +1632,20 @@ int bus_exec_context_set_transient_property(

                 return 1;

+        } else if (streq(name, "NUMAPolicy")) {

+                int32_t type;

+

+                r = sd_bus_message_read(message, "i", &type);

+                if (r < 0)

+                        return r;

+

+                if (!mpol_is_valid(type))

+                        return sd_bus_error_setf(error, SD_BUS_ERROR_INVALID_ARGS, "Invalid NUMAPolicy value: %i", type);

+

+                if (!UNIT_WRITE_FLAGS_NOOP(flags))

+                        c->numa_policy.type = type;

+

+                return 1;

         } else if (streq(name, "IOSchedulingClass")) {

                 int32_t q;

diff --git a/src/core/execute.c b/src/core/execute.c

index bc26aa66e7..56aa89e1ec 100644

--- a/src/core/execute.c

+++ b/src/core/execute.c

@@ -2997,6 +2997,16 @@ static int exec_child(

                         return log_unit_error_errno(unit, errno, "Failed to set up CPU affinity: %m");

                 }

+        if (mpol_is_valid(numa_policy_get_type(&context->numa_policy))) {

+                r = apply_numa_policy(&context->numa_policy);

+                if (r == -EOPNOTSUPP)

+                        log_unit_debug_errno(unit, r, "NUMA support not available, ignoring.");

+                else if (r < 0) {

+                        *exit_status = EXIT_NUMA_POLICY;

+                        return log_unit_error_errno(unit, r, "Failed to set NUMA memory policy: %m");

+                }

+        }

+

         if (context->ioprio_set)

                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, context->ioprio) < 0) {

                         *exit_status = EXIT_IOPRIO;

@@ -3651,6 +3661,7 @@ void exec_context_init(ExecContext *c) {

         assert_cc(NAMESPACE_FLAGS_INITIAL != NAMESPACE_FLAGS_ALL);

         c->restrict_namespaces = NAMESPACE_FLAGS_INITIAL;

         c->log_level_max = -1;

+        numa_policy_reset(&c->numa_policy);

 }

 void exec_context_done(ExecContext *c) {

@@ -3695,6 +3706,7 @@ void exec_context_done(ExecContext *c) {

         c->n_temporary_filesystems = 0;

         cpu_set_reset(&c->cpu_set);

+        numa_policy_reset(&c->numa_policy);

         c->utmp_id = mfree(c->utmp_id);

         c->selinux_context = mfree(c->selinux_context);

@@ -4104,6 +4116,14 @@ void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix) {

                 fprintf(f, "%sCPUAffinity: %s\n", prefix, affinity);

         }

+        if (mpol_is_valid(numa_policy_get_type(&c->numa_policy))) {

+                _cleanup_free_ char *nodes = NULL;

+

+                nodes = cpu_set_to_range_string(&c->numa_policy.nodes);

+                fprintf(f, "%sNUMAPolicy: %s\n", prefix, mpol_to_string(numa_policy_get_type(&c->numa_policy)));

+                fprintf(f, "%sNUMAMask: %s\n", prefix, strnull(nodes));

+        }

+

         if (c->timer_slack_nsec != NSEC_INFINITY)

                 fprintf(f, "%sTimerSlackNSec: "NSEC_FMT "\n", prefix, c->timer_slack_nsec);

diff --git a/src/core/execute.h b/src/core/execute.h

index e1e7a494cd..b2eb55f8f5 100644

--- a/src/core/execute.h

+++ b/src/core/execute.h

@@ -150,6 +150,7 @@ struct ExecContext {

         int cpu_sched_priority;

         CPUSet cpu_set;

+        NUMAPolicy numa_policy;

         ExecInput std_input;

         ExecOutput std_output;

diff --git a/src/core/load-fragment-gperf.gperf.m4 b/src/core/load-fragment-gperf.gperf.m4

index 1066bcfb8f..cdf4d14c4e 100644

--- a/src/core/load-fragment-gperf.gperf.m4

+++ b/src/core/load-fragment-gperf.gperf.m4

@@ -36,6 +36,8 @@ $1.CPUSchedulingPolicy,          config_parse_exec_cpu_sched_policy, 0,

 $1.CPUSchedulingPriority,        config_parse_exec_cpu_sched_prio,   0,                             offsetof($1, exec_context)

 $1.CPUSchedulingResetOnFork,     config_parse_bool,                  0,                             offsetof($1, exec_context.cpu_sched_reset_on_fork)

 $1.CPUAffinity,                  config_parse_exec_cpu_affinity,     0,                             offsetof($1, exec_context)

+$1.NUMAPolicy,                   config_parse_numa_policy,           0,                             offsetof($1, exec_context.numa_policy.type)

+$1.NUMAMask,                     config_parse_numa_mask,             0,                             offsetof($1, exec_context.numa_policy)

 $1.UMask,                        config_parse_mode,                  0,                             offsetof($1, exec_context.umask)

 $1.Environment,                  config_parse_environ,               0,                             offsetof($1, exec_context.environment)

 $1.EnvironmentFile,              config_parse_unit_env_file,         0,                             offsetof($1, exec_context.environment_files)

diff --git a/src/core/load-fragment.c b/src/core/load-fragment.c

index 34ae834188..35dd595098 100644

--- a/src/core/load-fragment.c

+++ b/src/core/load-fragment.c

@@ -93,6 +93,7 @@ DEFINE_CONFIG_PARSE_PTR(config_parse_blockio_weight, cg_blkio_weight_parse, uint

 DEFINE_CONFIG_PARSE_PTR(config_parse_cg_weight, cg_weight_parse, uint64_t, "Invalid weight");

 DEFINE_CONFIG_PARSE_PTR(config_parse_cpu_shares, cg_cpu_shares_parse, uint64_t, "Invalid CPU shares");

 DEFINE_CONFIG_PARSE_PTR(config_parse_exec_mount_flags, mount_propagation_flags_from_string, unsigned long, "Failed to parse mount flag");

+DEFINE_CONFIG_PARSE_ENUM_WITH_DEFAULT(config_parse_numa_policy, mpol, int, -1, "Invalid NUMA policy type");

 int config_parse_unit_deps(

                 const char *unit,

@@ -1159,6 +1160,33 @@ int config_parse_exec_cpu_sched_policy(const char *unit,

         return 0;

 }

+int config_parse_numa_mask(const char *unit,

+                           const char *filename,

+                           unsigned line,

+                           const char *section,

+                           unsigned section_line,

+                           const char *lvalue,

+                           int ltype,

+                           const char *rvalue,

+                           void *data,

+                           void *userdata) {

+        int r;

+        NUMAPolicy *p = data;

+

+        assert(filename);

+        assert(lvalue);

+        assert(rvalue);

+        assert(data);

+

+        r = parse_cpu_set_extend(rvalue, &p->nodes, true, unit, filename, line, lvalue);

+        if (r < 0) {

+                log_syntax(unit, LOG_ERR, filename, line, r, "Failed to parse NUMA node mask, ignoring: %s", rvalue);

+                return 0;

+        }

+

+        return r;

+}

+

 int config_parse_exec_cpu_sched_prio(const char *unit,

                                      const char *filename,

                                      unsigned line,

diff --git a/src/core/load-fragment.h b/src/core/load-fragment.h

index dad281ef72..f2ca1b8ee7 100644

--- a/src/core/load-fragment.h

+++ b/src/core/load-fragment.h

@@ -102,6 +102,8 @@ CONFIG_PARSER_PROTOTYPE(config_parse_job_timeout_sec);

 CONFIG_PARSER_PROTOTYPE(config_parse_job_running_timeout_sec);

 CONFIG_PARSER_PROTOTYPE(config_parse_log_extra_fields);

 CONFIG_PARSER_PROTOTYPE(config_parse_collect_mode);

+CONFIG_PARSER_PROTOTYPE(config_parse_numa_policy);

+CONFIG_PARSER_PROTOTYPE(config_parse_numa_mask);

 /* gperf prototypes */

 const struct ConfigPerfItem* load_fragment_gperf_lookup(const char *key, GPERF_LEN_TYPE length);

diff --git a/src/core/main.c b/src/core/main.c

index c74dc641c1..83f9dd5878 100644

--- a/src/core/main.c

+++ b/src/core/main.c

@@ -134,6 +134,7 @@ static uint64_t arg_default_tasks_max;

 static sd_id128_t arg_machine_id;

 static EmergencyAction arg_cad_burst_action;

 static CPUSet arg_cpu_affinity;

+static NUMAPolicy arg_numa_policy;

 static int parse_configuration(void);

@@ -660,6 +661,8 @@ static int parse_config_file(void) {

                 { "Manager", "ShowStatus",                config_parse_show_status,      0, &arg_show_status                       },

                 { "Manager", "CPUAffinity",               config_parse_cpu_affinity2,    0, &arg_cpu_affinity                      },

                 { "Manager", "JoinControllers",           config_parse_join_controllers, 0, &arg_join_controllers                  },

+                { "Manager", "NUMAPolicy",                config_parse_numa_policy,      0, &arg_numa_policy.type                  },

+                { "Manager", "NUMAMask",                  config_parse_numa_mask,        0, &arg_numa_policy                       },

                 { "Manager", "RuntimeWatchdogSec",        config_parse_sec,              0, &arg_runtime_watchdog                  },

                 { "Manager", "ShutdownWatchdogSec",       config_parse_sec,              0, &arg_shutdown_watchdog                 },

                 { "Manager", "WatchdogDevice",            config_parse_path,             0, &arg_watchdog_device                   },

@@ -1501,6 +1504,27 @@ static void update_cpu_affinity(bool skip_setup) {

                 log_warning_errno(errno, "Failed to set CPU affinity: %m");

 }

+static void update_numa_policy(bool skip_setup) {

+        int r;

+        _cleanup_free_ char *nodes = NULL;

+        const char * policy = NULL;

+

+        if (skip_setup || !mpol_is_valid(numa_policy_get_type(&arg_numa_policy)))

+                return;

+

+        if (DEBUG_LOGGING) {

+                policy = mpol_to_string(numa_policy_get_type(&arg_numa_policy));

+                nodes = cpu_set_to_range_string(&arg_numa_policy.nodes);

+                log_debug("Setting NUMA policy to %s, with nodes %s.", strnull(policy), strnull(nodes));

+        }

+

+        r = apply_numa_policy(&arg_numa_policy);

+        if (r == -EOPNOTSUPP)

+                log_debug_errno(r, "NUMA support not available, ignoring.");

+        else if (r < 0)

+                log_warning_errno(r, "Failed to set NUMA memory policy: %m");

+}

+

 static void do_reexecute(

                 int argc,

                 char *argv[],

@@ -1672,6 +1696,7 @@ static int invoke_main_loop(

                         set_manager_defaults(m);

                         update_cpu_affinity(false);

+                        update_numa_policy(false);

                         if (saved_log_level >= 0)

                                 manager_override_log_level(m, saved_log_level);

@@ -1832,6 +1857,7 @@ static int initialize_runtime(

                 return 0;

         update_cpu_affinity(skip_setup);

+        update_numa_policy(skip_setup);

         if (arg_system) {

                 /* Make sure we leave a core dump without panicing the kernel. */

@@ -2011,6 +2037,7 @@ static void reset_arguments(void) {

         arg_cad_burst_action = EMERGENCY_ACTION_REBOOT_FORCE;

         cpu_set_reset(&arg_cpu_affinity);

+        numa_policy_reset(&arg_numa_policy);

 }

 static int parse_configuration(void) {

diff --git a/src/core/system.conf.in b/src/core/system.conf.in

index 653ec6b8c9..0d93fbf147 100644

--- a/src/core/system.conf.in

+++ b/src/core/system.conf.in

@@ -24,6 +24,8 @@

 #CtrlAltDelBurstAction=reboot-force

 #CPUAffinity=1 2

 #JoinControllers=cpu,cpuacct net_cls,net_prio

+#NUMAPolicy=default

+#NUMAMask=

 #RuntimeWatchdogSec=0

 #ShutdownWatchdogSec=10min

 #CapabilityBoundingSet=

diff --git a/src/shared/bus-unit-util.c b/src/shared/bus-unit-util.c

index ec8732c226..055edd6e22 100644

--- a/src/shared/bus-unit-util.c

+++ b/src/shared/bus-unit-util.c

@@ -947,6 +947,34 @@ static int bus_append_execute_property(sd_bus_message *m, const char *field, con

                 return bus_append_byte_array(m, field, array, allocated);

         }

+        if (streq(field, "NUMAPolicy")) {

+                r = mpol_from_string(eq);

+                if (r < 0)

+                        return log_error_errno(r, "Failed to parse %s value: %s", field, eq);

+

+                r = sd_bus_message_append(m, "(sv)", field, "i", (int32_t) r);

+                if (r < 0)

+                        return bus_log_create_error(r);

+

+                return 1;

+        }

+

+        if (streq(field, "NUMAMask")) {

+                _cleanup_(cpu_set_reset) CPUSet nodes = {};

+                _cleanup_free_ uint8_t *array = NULL;