Merge pull request #3905 from htejun/cgroup-v2-cpu

core: add cgroup CPU controller support on the unified hierarchy

(zj: merging not squashing to make it clear against which upstream this patch was developed.)

13 files changed, 413 insertions, 90 deletions

diff --git a/man/systemd.resource-control.xml b/man/systemd.resource-control.xml
index 0e98ca78b8..84dbfa2ff3 100644
--- a/man/systemd.resource-control.xml
+++ b/man/systemd.resource-control.xml
@@ -106,13 +106,21 @@
 
     <para>
       <variablelist>
+
         <varlistentry>
-          <term><option>IO</option></term>
+          <term><option>CPU</option></term>
           <listitem>
-            <para><varname>IO</varname> prefixed settings are superset of and replace <varname>BlockIO</varname>
-            prefixed ones. On unified hierarchy, IO resource control also applies to buffered writes.</para>
+            <para>Due to the lack of consensus in the kernel community, the CPU controller support on the unified
+            cgroup hierarchy requires out-of-tree kernel patches. See <ulink
+            url="https://git.kernel.org/cgit/linux/kernel/git/tj/cgroup.git/tree/Documentation/cgroup-v2-cpu.txt?h=cgroup-v2-cpu">cgroup-v2-cpu.txt</ulink>.</para>
+
+            <para><varname>CPUWeight=</varname> and <varname>StartupCPUWeight=</varname> replace
+            <varname>CPUShares=</varname> and <varname>StartupCPUShares=</varname>, respectively.</para>
+
+            <para>The <literal>cpuacct</literal> controller does not exist separately on the unified hierarchy.</para>
           </listitem>
         </varlistentry>
+
         <varlistentry>
           <term><option>Memory</option></term>
           <listitem>
@@ -120,6 +128,15 @@
             and <varname>MemoryHigh=</varname> are effective only on unified hierarchy.</para>
           </listitem>
         </varlistentry>
+
+        <varlistentry>
+          <term><option>IO</option></term>
+          <listitem>
+            <para><varname>IO</varname> prefixed settings are superset of and replace <varname>BlockIO</varname>
+            prefixed ones. On unified hierarchy, IO resource control also applies to buffered writes.</para>
+          </listitem>
+        </varlistentry>
+
       </variablelist>
     </para>
 
@@ -160,30 +177,49 @@
       </varlistentry>
 
       <varlistentry>
+        <term><varname>CPUWeight=<replaceable>weight</replaceable></varname></term>
+        <term><varname>StartupCPUWeight=<replaceable>weight</replaceable></varname></term>
+
+        <listitem>
+          <para>Assign the specified CPU time weight to the processes executed, if the unified control group hierarchy
+          is used on the system. These options take an integer value and control the <literal>cpu.weight</literal>
+          control group attribute. The allowed range is 1 to 10000. Defaults to 100. For details about this control
+          group attribute, see <ulink
+          url="https://www.kernel.org/doc/Documentation/cgroup-v2.txt">cgroup-v2.txt</ulink> and <ulink
+          url="https://www.kernel.org/doc/Documentation/scheduler/sched-design-CFS.txt">sched-design-CFS.txt</ulink>.
+          The available CPU time is split up among all units within one slice relative to their CPU time weight.</para>
+
+          <para>While <varname>StartupCPUWeight=</varname> only applies to the startup phase of the system,
+          <varname>CPUWeight=</varname> applies to normal runtime of the system, and if the former is not set also to
+          the startup phase. Using <varname>StartupCPUWeight=</varname> allows prioritizing specific services at
+          boot-up differently than during normal runtime.</para>
+
+          <para>Implies <literal>CPUAccounting=true</literal>.</para>
+
+          <para>These settings are supported only if the unified control group hierarchy is used.</para>
+        </listitem>
+      </varlistentry>
+
+      <varlistentry>
         <term><varname>CPUShares=<replaceable>weight</replaceable></varname></term>
         <term><varname>StartupCPUShares=<replaceable>weight</replaceable></varname></term>
 
         <listitem>
-          <para>Assign the specified CPU time share weight to the
-          processes executed. These options take an integer value and
-          control the <literal>cpu.shares</literal> control group
-          attribute. The allowed range is 2 to 262144. Defaults to
-          1024. For details about this control group attribute, see
-          <ulink
+          <para>Assign the specified CPU time share weight to the processes executed. These options take an integer
+          value and control the <literal>cpu.shares</literal> control group attribute. The allowed range is 2 to
+          262144. Defaults to 1024. For details about this control group attribute, see <ulink
           url="https://www.kernel.org/doc/Documentation/scheduler/sched-design-CFS.txt">sched-design-CFS.txt</ulink>.
-          The available CPU time is split up among all units within
-          one slice relative to their CPU time share weight.</para>
+          The available CPU time is split up among all units within one slice relative to their CPU time share
+          weight.</para>
 
-          <para>While <varname>StartupCPUShares=</varname> only
-          applies to the startup phase of the system,
-          <varname>CPUShares=</varname> applies to normal runtime of
-          the system, and if the former is not set also to the startup
-          phase. Using <varname>StartupCPUShares=</varname> allows
-          prioritizing specific services at boot-up differently than
-          during normal runtime.</para>
+          <para>While <varname>StartupCPUShares=</varname> only applies to the startup phase of the system,
+          <varname>CPUShares=</varname> applies to normal runtime of the system, and if the former is not set also to
+          the startup phase. Using <varname>StartupCPUShares=</varname> allows prioritizing specific services at
+          boot-up differently than during normal runtime.</para>
 
-          <para>These options imply
-          <literal>CPUAccounting=true</literal>.</para>
+          <para>Implies <literal>CPUAccounting=true</literal>.</para>
+
+          <para>These settings are supported only if the legacy control group hierarchy is used.</para>
         </listitem>
       </varlistentry>
 
@@ -191,22 +227,20 @@
         <term><varname>CPUQuota=</varname></term>
 
         <listitem>
-          <para>Assign the specified CPU time quota to the processes
-          executed. Takes a percentage value, suffixed with "%". The
-          percentage specifies how much CPU time the unit shall get at
-          maximum, relative to the total CPU time available on one
-          CPU. Use values &gt; 100% for allotting CPU time on more than
-          one CPU. This controls the
-          <literal>cpu.cfs_quota_us</literal> control group
-          attribute. For details about this control group attribute,
-          see <ulink
+          <para>Assign the specified CPU time quota to the processes executed. Takes a percentage value, suffixed with
+          "%". The percentage specifies how much CPU time the unit shall get at maximum, relative to the total CPU time
+          available on one CPU. Use values &gt; 100% for allotting CPU time on more than one CPU. This controls the
+          <literal>cpu.max</literal> attribute on the unified control group hierarchy and
+          <literal>cpu.cfs_quota_us</literal> on legacy. For details about these control group attributes, see <ulink
+          url="https://www.kernel.org/doc/Documentation/cgroup-v2.txt">cgroup-v2.txt</ulink> and <ulink
           url="https://www.kernel.org/doc/Documentation/scheduler/sched-design-CFS.txt">sched-design-CFS.txt</ulink>.</para>
 
-          <para>Example: <varname>CPUQuota=20%</varname> ensures that
-          the executed processes will never get more than 20% CPU time
-          on one CPU.</para>
+          <para>Example: <varname>CPUQuota=20%</varname> ensures that the executed processes will never get more than
+          20% CPU time on one CPU.</para>
 
           <para>Implies <literal>CPUAccounting=true</literal>.</para>
+
+          <para>This setting is supported on both unified and legacy control group hierarchies.</para>
         </listitem>
       </varlistentry>
 
diff --git a/src/basic/cgroup-util.c b/src/basic/cgroup-util.c
index 302b958d0d..25ef8a5c76 100644
--- a/src/basic/cgroup-util.c
+++ b/src/basic/cgroup-util.c
@@ -1905,6 +1905,49 @@ int cg_get_attribute(const char *controller, const char *path, const char *attri
         return read_one_line_file(p, ret);
 }
 
+int cg_get_keyed_attribute(const char *controller, const char *path, const char *attribute, const char **keys, char **values) {
+        _cleanup_free_ char *filename = NULL, *content = NULL;
+        char *line, *p;
+        int i, r;
+
+        for (i = 0; keys[i]; i++)
+                values[i] = NULL;
+
+        r = cg_get_path(controller, path, attribute, &filename);
+        if (r < 0)
+                return r;
+
+        r = read_full_file(filename, &content, NULL);
+        if (r < 0)
+                return r;
+
+        p = content;
+        while ((line = strsep(&p, "\n"))) {
+                char *key;
+
+                key = strsep(&line, " ");
+
+                for (i = 0; keys[i]; i++) {
+                        if (streq(key, keys[i])) {
+                                values[i] = strdup(line);
+                                break;
+                        }
+                }
+        }
+
+        for (i = 0; keys[i]; i++) {
+                if (!values[i]) {
+                        for (i = 0; keys[i]; i++) {
+                                free(values[i]);
+                                values[i] = NULL;
+                        }
+                        return -ENOENT;
+                }
+        }
+
+        return 0;
+}
+
 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
         CGroupController c;
         int r, unified;
@@ -2101,10 +2144,10 @@ int cg_mask_supported(CGroupMask *ret) {
                         mask |= CGROUP_CONTROLLER_TO_MASK(v);
                 }
 
-                /* Currently, we only support the memory, io and pids
+                /* Currently, we support the cpu, memory, io and pids
                  * controller in the unified hierarchy, mask
                  * everything else off. */
-                mask &= CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
+                mask &= CGROUP_MASK_CPU | CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
 
         } else {
                 CGroupController c;
diff --git a/src/basic/cgroup-util.h b/src/basic/cgroup-util.h
index ec5c715987..a509a1775c 100644
--- a/src/basic/cgroup-util.h
+++ b/src/basic/cgroup-util.h
@@ -169,6 +169,7 @@ int cg_create_and_attach(const char *controller, const char *path, pid_t pid);
 
 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value);
 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret);
+int cg_get_keyed_attribute(const char *controller, const char *path, const char *attribute, const char **keys, char **values);
 
 int cg_set_group_access(const char *controller, const char *path, mode_t mode, uid_t uid, gid_t gid);
 int cg_set_task_access(const char *controller, const char *path, mode_t mode, uid_t uid, gid_t gid);
diff --git a/src/cgtop/cgtop.c b/src/cgtop/cgtop.c
index c67b328b38..6045ae0437 100644
--- a/src/cgtop/cgtop.c
+++ b/src/cgtop/cgtop.c
@@ -208,24 +208,47 @@ static int process(
                 if (g->n_tasks > 0)
                         g->n_tasks_valid = true;
 
-        } else if (streq(controller, "cpuacct") && cg_unified() <= 0) {
+        } else if (streq(controller, "cpu") || streq(controller, "cpuacct")) {
                 _cleanup_free_ char *p = NULL, *v = NULL;
                 uint64_t new_usage;
                 nsec_t timestamp;
 
-                r = cg_get_path(controller, path, "cpuacct.usage", &p);
-                if (r < 0)
-                        return r;
+                if (cg_unified() > 0) {
+                        const char *keys[] = { "usage_usec", NULL };
+                        _cleanup_free_ char *val = NULL;
 
-                r = read_one_line_file(p, &v);
-                if (r == -ENOENT)
-                        return 0;
-                if (r < 0)
-                        return r;
+                        if (!streq(controller, "cpu"))
+                                return 0;
 
-                r = safe_atou64(v, &new_usage);
-                if (r < 0)
-                        return r;
+                        r = cg_get_keyed_attribute("cpu", path, "cpu.stat", keys, &val);
+                        if (r == -ENOENT)
+                                return 0;
+                        if (r < 0)
+                                return r;
+
+                        r = safe_atou64(val, &new_usage);
+                        if (r < 0)
+                                return r;
+
+                        new_usage *= NSEC_PER_USEC;
+                } else {
+                        if (!streq(controller, "cpuacct"))
+                                return 0;
+
+                        r = cg_get_path(controller, path, "cpuacct.usage", &p);
+                        if (r < 0)
+                                return r;
+
+                        r = read_one_line_file(p, &v);
+                        if (r == -ENOENT)
+                                return 0;
+                        if (r < 0)
+                                return r;
+
+                        r = safe_atou64(v, &new_usage);
+                        if (r < 0)
+                                return r;
+                }
 
                 timestamp = now_nsec(CLOCK_MONOTONIC);
 
@@ -449,6 +472,9 @@ static int refresh(const char *root, Hashmap *a, Hashmap *b, unsigned iteration)
         r = refresh_one(SYSTEMD_CGROUP_CONTROLLER, root, a, b, iteration, 0, NULL);
         if (r < 0)
                 return r;
+        r = refresh_one("cpu", root, a, b, iteration, 0, NULL);
+        if (r < 0)
+                return r;
         r = refresh_one("cpuacct", root, a, b, iteration, 0, NULL);
         if (r < 0)
                 return r;
diff --git a/src/core/cgroup.c b/src/core/cgroup.c
index c19e43f571..910a64b4da 100644
--- a/src/core/cgroup.c
+++ b/src/core/cgroup.c
@@ -57,9 +57,12 @@ void cgroup_context_init(CGroupContext *c) {
         /* Initialize everything to the kernel defaults, assuming the
          * structure is preinitialized to 0 */
 
+        c->cpu_weight = CGROUP_WEIGHT_INVALID;
+        c->startup_cpu_weight = CGROUP_WEIGHT_INVALID;
+        c->cpu_quota_per_sec_usec = USEC_INFINITY;
+
         c->cpu_shares = CGROUP_CPU_SHARES_INVALID;
         c->startup_cpu_shares = CGROUP_CPU_SHARES_INVALID;
-        c->cpu_quota_per_sec_usec = USEC_INFINITY;
 
         c->memory_high = CGROUP_LIMIT_MAX;
         c->memory_max = CGROUP_LIMIT_MAX;
@@ -158,6 +161,8 @@ void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
                 "%sBlockIOAccounting=%s\n"
                 "%sMemoryAccounting=%s\n"
                 "%sTasksAccounting=%s\n"
+                "%sCPUWeight=%" PRIu64 "\n"
+                "%sStartupCPUWeight=%" PRIu64 "\n"
                 "%sCPUShares=%" PRIu64 "\n"
                 "%sStartupCPUShares=%" PRIu64 "\n"
                 "%sCPUQuotaPerSecSec=%s\n"
@@ -177,6 +182,8 @@ void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
                 prefix, yes_no(c->blockio_accounting),
                 prefix, yes_no(c->memory_accounting),
                 prefix, yes_no(c->tasks_accounting),
+                prefix, c->cpu_weight,
+                prefix, c->startup_cpu_weight,
                 prefix, c->cpu_shares,
                 prefix, c->startup_cpu_shares,
                 prefix, format_timespan(u, sizeof(u), c->cpu_quota_per_sec_usec, 1),
@@ -382,6 +389,95 @@ fail:
         return -errno;
 }
 
+static bool cgroup_context_has_cpu_weight(CGroupContext *c) {
+        return c->cpu_weight != CGROUP_WEIGHT_INVALID ||
+                c->startup_cpu_weight != CGROUP_WEIGHT_INVALID;
+}
+
+static bool cgroup_context_has_cpu_shares(CGroupContext *c) {
+        return c->cpu_shares != CGROUP_CPU_SHARES_INVALID ||
+                c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID;
+}
+
+static uint64_t cgroup_context_cpu_weight(CGroupContext *c, ManagerState state) {
+        if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
+            c->startup_cpu_weight != CGROUP_WEIGHT_INVALID)
+                return c->startup_cpu_weight;
+        else if (c->cpu_weight != CGROUP_WEIGHT_INVALID)
+                return c->cpu_weight;
+        else
+                return CGROUP_WEIGHT_DEFAULT;
+}
+
+static uint64_t cgroup_context_cpu_shares(CGroupContext *c, ManagerState state) {
+        if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
+            c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID)
+                return c->startup_cpu_shares;
+        else if (c->cpu_shares != CGROUP_CPU_SHARES_INVALID)
+                return c->cpu_shares;
+        else
+                return CGROUP_CPU_SHARES_DEFAULT;
+}
+
+static void cgroup_apply_unified_cpu_config(Unit *u, uint64_t weight, uint64_t quota) {
+        char buf[MAX(DECIMAL_STR_MAX(uint64_t) + 1, (DECIMAL_STR_MAX(usec_t) + 1) * 2)];
+        int r;
+
+        xsprintf(buf, "%" PRIu64 "\n", weight);
+        r = cg_set_attribute("cpu", u->cgroup_path, "cpu.weight", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set cpu.weight: %m");
+
+        if (quota != USEC_INFINITY)
+                xsprintf(buf, USEC_FMT " " USEC_FMT "\n",
+                         quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC, CGROUP_CPU_QUOTA_PERIOD_USEC);
+        else
+                xsprintf(buf, "max " USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
+
+        r = cg_set_attribute("cpu", u->cgroup_path, "cpu.max", buf);
+
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set cpu.max: %m");
+}
+
+static void cgroup_apply_legacy_cpu_config(Unit *u, uint64_t shares, uint64_t quota) {
+        char buf[MAX(DECIMAL_STR_MAX(uint64_t), DECIMAL_STR_MAX(usec_t)) + 1];
+        int r;
+
+        xsprintf(buf, "%" PRIu64 "\n", shares);
+        r = cg_set_attribute("cpu", u->cgroup_path, "cpu.shares", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set cpu.shares: %m");
+
+        xsprintf(buf, USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
+        r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_period_us", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set cpu.cfs_period_us: %m");
+
+        if (quota != USEC_INFINITY) {
+                xsprintf(buf, USEC_FMT "\n", quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC);
+                r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", buf);
+        } else
+                r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", "-1");
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set cpu.cfs_quota_us: %m");
+}
+
+static uint64_t cgroup_cpu_shares_to_weight(uint64_t shares) {
+        return CLAMP(shares * CGROUP_WEIGHT_DEFAULT / CGROUP_CPU_SHARES_DEFAULT,
+                     CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
+}
+
+static uint64_t cgroup_cpu_weight_to_shares(uint64_t weight) {
+        return CLAMP(weight * CGROUP_CPU_SHARES_DEFAULT / CGROUP_WEIGHT_DEFAULT,
+                     CGROUP_CPU_SHARES_MIN, CGROUP_CPU_SHARES_MAX);
+}
+
 static bool cgroup_context_has_io_config(CGroupContext *c) {
         return c->io_accounting ||
                 c->io_weight != CGROUP_WEIGHT_INVALID ||
@@ -566,30 +662,42 @@ static void cgroup_context_apply(Unit *u, CGroupMask mask, ManagerState state) {
          * and missing cgroups, i.e. EROFS and ENOENT. */
 
         if ((mask & CGROUP_MASK_CPU) && !is_root) {
-                char buf[MAX(DECIMAL_STR_MAX(uint64_t), DECIMAL_STR_MAX(usec_t)) + 1];
+                bool has_weight = cgroup_context_has_cpu_weight(c);
+                bool has_shares = cgroup_context_has_cpu_shares(c);
 
-                sprintf(buf, "%" PRIu64 "\n",
-                        IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) && c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID ? c->startup_cpu_shares :
-                        c->cpu_shares != CGROUP_CPU_SHARES_INVALID ? c->cpu_shares : CGROUP_CPU_SHARES_DEFAULT);
-                r = cg_set_attribute("cpu", path, "cpu.shares", buf);
-                if (r < 0)
-                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
-                                      "Failed to set cpu.shares: %m");
+                if (cg_unified() > 0) {
+                        uint64_t weight;
 
-                sprintf(buf, USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
-                r = cg_set_attribute("cpu", path, "cpu.cfs_period_us", buf);
-                if (r < 0)
-                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
-                                      "Failed to set cpu.cfs_period_us: %m");
+                        if (has_weight)
+                                weight = cgroup_context_cpu_weight(c, state);
+                        else if (has_shares) {
+                                uint64_t shares = cgroup_context_cpu_shares(c, state);
 
-                if (c->cpu_quota_per_sec_usec != USEC_INFINITY) {
-                        sprintf(buf, USEC_FMT "\n", c->cpu_quota_per_sec_usec * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC);
-                        r = cg_set_attribute("cpu", path, "cpu.cfs_quota_us", buf);
-                } else
-                        r = cg_set_attribute("cpu", path, "cpu.cfs_quota_us", "-1");
-                if (r < 0)
-                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
-                                      "Failed to set cpu.cfs_quota_us: %m");
+                                weight = cgroup_cpu_shares_to_weight(shares);
+
+                                log_cgroup_compat(u, "Applying [Startup]CpuShares %" PRIu64 " as [Startup]CpuWeight %" PRIu64 " on %s",
+                                                  shares, weight, path);
+                        } else
+                                weight = CGROUP_WEIGHT_DEFAULT;
+
+                        cgroup_apply_unified_cpu_config(u, weight, c->cpu_quota_per_sec_usec);
+                } else {
+                        uint64_t shares;
+
+                        if (has_shares)
+                                shares = cgroup_context_cpu_shares(c, state);
+                        else if (has_weight) {
+                                uint64_t weight = cgroup_context_cpu_weight(c, state);
+
+                                shares = cgroup_cpu_weight_to_shares(weight);
+
+                                log_cgroup_compat(u, "Applying [Startup]CpuWeight %" PRIu64 " as [Startup]CpuShares %" PRIu64 " on %s",
+                                                  weight, shares, path);
+                        } else
+                                shares = CGROUP_CPU_SHARES_DEFAULT;
+
+                        cgroup_apply_legacy_cpu_config(u, shares, c->cpu_quota_per_sec_usec);
+                }
         }
 
         if (mask & CGROUP_MASK_IO) {
@@ -864,8 +972,8 @@ CGroupMask cgroup_context_get_mask(CGroupContext *c) {
         /* Figure out which controllers we need */
 
         if (c->cpu_accounting ||
-            c->cpu_shares != CGROUP_CPU_SHARES_INVALID ||
-            c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID ||
+            cgroup_context_has_cpu_weight(c) ||
+            cgroup_context_has_cpu_shares(c) ||
             c->cpu_quota_per_sec_usec != USEC_INFINITY)
                 mask |= CGROUP_MASK_CPUACCT | CGROUP_MASK_CPU;
 
@@ -1890,18 +1998,37 @@ static int unit_get_cpu_usage_raw(Unit *u, nsec_t *ret) {
         if (!u->cgroup_path)
                 return -ENODATA;
 
-        if ((u->cgroup_realized_mask & CGROUP_MASK_CPUACCT) == 0)
-                return -ENODATA;
+        if (cg_unified() > 0) {
+                const char *keys[] = { "usage_usec", NULL };
+                _cleanup_free_ char *val = NULL;
+                uint64_t us;
 
-        r = cg_get_attribute("cpuacct", u->cgroup_path, "cpuacct.usage", &v);
-        if (r == -ENOENT)
-                return -ENODATA;
-        if (r < 0)
-                return r;
+                if ((u->cgroup_realized_mask & CGROUP_MASK_CPU) == 0)
+                        return -ENODATA;
 
-        r = safe_atou64(v, &ns);
-        if (r < 0)
-                return r;
+                r = cg_get_keyed_attribute("cpu", u->cgroup_path, "cpu.stat", keys, &val);
+                if (r < 0)
+                        return r;
+
+                r = safe_atou64(val, &us);
+                if (r < 0)
+                        return r;
+
+                ns = us * NSEC_PER_USEC;
+        } else {
+                if ((u->cgroup_realized_mask & CGROUP_MASK_CPUACCT) == 0)
+                        return -ENODATA;
+
+                r = cg_get_attribute("cpuacct", u->cgroup_path, "cpuacct.usage", &v);
+                if (r == -ENOENT)
+                        return -ENODATA;
+                if (r < 0)
+                        return r;
+
+                r = safe_atou64(v, &ns);
+                if (r < 0)
+                        return r;
+        }
 
         *ret = ns;
         return 0;
@@ -1915,8 +2042,8 @@ int unit_get_cpu_usage(Unit *u, nsec_t *ret) {
         if (r < 0)
                 return r;
 
-        if (ns > u->cpuacct_usage_base)
-                ns -= u->cpuacct_usage_base;
+        if (ns > u->cpu_usage_base)
+                ns -= u->cpu_usage_base;
         else
                 ns = 0;
 
@@ -1932,11 +2059,11 @@ int unit_reset_cpu_usage(Unit *u) {
 
         r = unit_get_cpu_usage_raw(u, &ns);
         if (r < 0) {
-                u->cpuacct_usage_base = 0;
+                u->cpu_usage_base = 0;
                 return r;
         }
 
-        u->cpuacct_usage_base = ns;
+        u->cpu_usage_base = ns;
         return 0;
 }
 
diff --git a/src/core/cgroup.h b/src/core/cgroup.h
index a57403e79f..2fe9cc4039 100644
--- a/src/core/cgroup.h
+++ b/src/core/cgroup.h
@@ -89,6 +89,10 @@ struct CGroupContext {
         bool tasks_accounting;
 
         /* For unified hierarchy */
+        uint64_t cpu_weight;
+        uint64_t startup_cpu_weight;
+        usec_t cpu_quota_per_sec_usec;
+
         uint64_t io_weight;
         uint64_t startup_io_weight;
         LIST_HEAD(CGroupIODeviceWeight, io_device_weights);
@@ -101,7 +105,6 @@ struct CGroupContext {
         /* For legacy hierarchies */
         uint64_t cpu_shares;
         uint64_t startup_cpu_shares;
-        usec_t cpu_quota_per_sec_usec;
 
         uint64_t blockio_weight;
         uint64_t startup_blockio_weight;
diff --git a/src/core/dbus-cgroup.c b/src/core/dbus-cgroup.c
index 85b0c86a2f..2ca80d2996 100644
--- a/src/core/dbus-cgroup.c
+++ b/src/core/dbus-cgroup.c
@@ -210,6 +210,8 @@ const sd_bus_vtable bus_cgroup_vtable[] = {
         SD_BUS_VTABLE_START(0),
         SD_BUS_PROPERTY("Delegate", "b", bus_property_get_bool, offsetof(CGroupContext, delegate), 0),
         SD_BUS_PROPERTY("CPUAccounting", "b", bus_property_get_bool, offsetof(CGroupContext, cpu_accounting), 0),
+        SD_BUS_PROPERTY("CPUWeight", "t", NULL, offsetof(CGroupContext, cpu_weight), 0),
+        SD_BUS_PROPERTY("StartupCPUWeight", "t", NULL, offsetof(CGroupContext, startup_cpu_weight), 0),
         SD_BUS_PROPERTY("CPUShares", "t", NULL, offsetof(CGroupContext, cpu_shares), 0),
         SD_BUS_PROPERTY("StartupCPUShares", "t", NULL, offsetof(CGroupContext, startup_cpu_shares), 0),
         SD_BUS_PROPERTY("CPUQuotaPerSecUSec", "t", bus_property_get_usec, offsetof(CGroupContext, cpu_quota_per_sec_usec), 0),
@@ -303,6 +305,50 @@ int bus_cgroup_set_property(
 
                 return 1;
 
+        } else if (streq(name, "CPUWeight")) {
+                uint64_t weight;
+
+                r = sd_bus_message_read(message, "t", &weight);
+                if (r < 0)
+                        return r;
+
+                if (!CGROUP_WEIGHT_IS_OK(weight))
+                        return sd_bus_error_set_errnof(error, EINVAL, "CPUWeight value out of range");
+
+                if (mode != UNIT_CHECK) {
+                        c->cpu_weight = weight;
+                        unit_invalidate_cgroup(u, CGROUP_MASK_CPU);
+
+                        if (weight == CGROUP_WEIGHT_INVALID)
+                                unit_write_drop_in_private(u, mode, name, "CPUWeight=");
+                        else
+                                unit_write_drop_in_private_format(u, mode, name, "CPUWeight=%" PRIu64, weight);
+                }
+
+                return 1;
+
+        } else if (streq(name, "StartupCPUWeight")) {
+                uint64_t weight;
+
+                r = sd_bus_message_read(message, "t", &weight);
+                if (r < 0)
+                        return r;
+
+                if (!CGROUP_WEIGHT_IS_OK(weight))
+                        return sd_bus_error_set_errnof(error, EINVAL, "StartupCPUWeight value out of range");
+
+                if (mode != UNIT_CHECK) {
+                        c->startup_cpu_weight = weight;
+                        unit_invalidate_cgroup(u, CGROUP_MASK_CPU);
+
+                        if (weight == CGROUP_CPU_SHARES_INVALID)
+                                unit_write_drop_in_private(u, mode, name, "StartupCPUWeight=");
+                        else
+                                unit_write_drop_in_private_format(u, mode, name, "StartupCPUWeight=%" PRIu64, weight);
+                }
+
+                return 1;
+
         } else if (streq(name, "CPUShares")) {
                 uint64_t shares;
 
diff --git a/src/core/load-fragment-gperf.gperf.m4 b/src/core/load-fragment-gperf.gperf.m4
index 251155b428..05fe0df7e3 100644
--- a/src/core/load-fragment-gperf.gperf.m4
+++ b/src/core/load-fragment-gperf.gperf.m4
@@ -122,6 +122,8 @@ $1.KillSignal,                   config_parse_signal,                0,
 m4_define(`CGROUP_CONTEXT_CONFIG_ITEMS',
 `$1.Slice,                       config_parse_unit_slice,            0,                             0
 $1.CPUAccounting,                config_parse_bool,                  0,                             offsetof($1, cgroup_context.cpu_accounting)
+$1.CPUWeight,                    config_parse_cpu_weight,            0,                             offsetof($1, cgroup_context.cpu_weight)
+$1.StartupCPUWeight,             config_parse_cpu_weight,            0,                             offsetof($1, cgroup_context.startup_cpu_weight)
 $1.CPUShares,                    config_parse_cpu_shares,            0,                             offsetof($1, cgroup_context.cpu_shares)
 $1.StartupCPUShares,             config_parse_cpu_shares,            0,                             offsetof($1, cgroup_context.startup_cpu_shares)
 $1.CPUQuota,                     config_parse_cpu_quota,             0,                             offsetof($1, cgroup_context)
diff --git a/src/core/load-fragment.c b/src/core/load-fragment.c
index 420f368689..4ad6c4d79b 100644
--- a/src/core/load-fragment.c
+++ b/src/core/load-fragment.c
@@ -2851,6 +2851,34 @@ int config_parse_unit_slice(
 
 DEFINE_CONFIG_PARSE_ENUM(config_parse_device_policy, cgroup_device_policy, CGroupDevicePolicy, "Failed to parse device policy");
 
+int config_parse_cpu_weight(
+                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) {
+
+        uint64_t *weight = data;
+        int r;
+
+        assert(filename);
+        assert(lvalue);
+        assert(rvalue);
+
+        r = cg_weight_parse(rvalue, weight);
+        if (r < 0) {
+                log_syntax(unit, LOG_ERR, filename, line, r, "CPU weight '%s' invalid. Ignoring.", rvalue);
+                return 0;
+        }
+
+        return 0;
+}
+
 int config_parse_cpu_shares(
                 const char *unit,
                 const char *filename,
@@ -4191,6 +4219,7 @@ void unit_dump_config_items(FILE *f) {
                 { config_parse_address_families,      "FAMILIES" },
 #endif
                 { config_parse_cpu_shares,            "SHARES" },
+                { config_parse_cpu_weight,            "WEIGHT" },
                 { config_parse_memory_limit,          "LIMIT" },
                 { config_parse_device_allow,          "DEVICE" },
                 { config_parse_device_policy,         "POLICY" },
diff --git a/src/core/load-fragment.h b/src/core/load-fragment.h
index 213bce55a7..8b688740cf 100644
--- a/src/core/load-fragment.h
+++ b/src/core/load-fragment.h
@@ -81,6 +81,7 @@ int config_parse_syscall_errno(const char *unit, const char *filename, unsigned
 int config_parse_environ(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 config_parse_pass_environ(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 config_parse_unit_slice(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 config_parse_cpu_weight(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 config_parse_cpu_shares(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 config_parse_memory_limit(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 config_parse_tasks_max(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);
diff --git a/src/core/unit.c b/src/core/unit.c
index ff7c562fba..952604e0db 100644
--- a/src/core/unit.c
+++ b/src/core/unit.c
@@ -2608,7 +2608,7 @@ int unit_serialize(Unit *u, FILE *f, FDSet *fds, bool serialize_jobs) {
                 unit_serialize_item(u, f, "assert-result", yes_no(u->assert_result));
 
         unit_serialize_item(u, f, "transient", yes_no(u->transient));
-        unit_serialize_item_format(u, f, "cpuacct-usage-base", "%" PRIu64, u->cpuacct_usage_base);
+        unit_serialize_item_format(u, f, "cpu-usage-base", "%" PRIu64, u->cpu_usage_base);
 
         if (u->cgroup_path)
                 unit_serialize_item(u, f, "cgroup", u->cgroup_path);
@@ -2824,9 +2824,9 @@ int unit_deserialize(Unit *u, FILE *f, FDSet *fds) {
 
                         continue;
 
-                } else if (streq(l, "cpuacct-usage-base")) {
+                } else if (streq(l, "cpu-usage-base") || streq(l, "cpuacct-usage-base")) {
 
-                        r = safe_atou64(v, &u->cpuacct_usage_base);
+                        r = safe_atou64(v, &u->cpu_usage_base);
                         if (r < 0)
                                 log_unit_debug(u, "Failed to parse CPU usage %s, ignoring.", v);
 
diff --git a/src/core/unit.h b/src/core/unit.h
index 47eb8d50a6..a6c69938dd 100644
--- a/src/core/unit.h
+++ b/src/core/unit.h
@@ -184,8 +184,8 @@ struct Unit {
         UnitFileState unit_file_state;
         int unit_file_preset;
 
-        /* Where the cpuacct.usage cgroup counter was at the time the unit was started */
-        nsec_t cpuacct_usage_base;
+        /* Where the cpu.stat or cpuacct.usage was at the time the unit was started */
+        nsec_t cpu_usage_base;
 
         /* Counterparts in the cgroup filesystem */
         char *cgroup_path;
diff --git a/src/shared/bus-unit-util.c b/src/shared/bus-unit-util.c
index 7774d607c7..f9e12e0578 100644
--- a/src/shared/bus-unit-util.c
+++ b/src/shared/bus-unit-util.c
@@ -212,6 +212,17 @@ int bus_append_unit_property_assignment(sd_bus_message *m, const char *assignmen
 
                 r = sd_bus_message_append(m, "v", "b", r);
 
+        } else if (STR_IN_SET(field, "CPUWeight", "StartupCPUWeight")) {
+                uint64_t u;
+
+                r = cg_weight_parse(eq, &u);
+                if (r < 0) {
+                        log_error("Failed to parse %s value %s.", field, eq);
+                        return -EINVAL;
+                }
+
+                r = sd_bus_message_append(m, "v", "t", u);
+
         } else if (STR_IN_SET(field, "CPUShares", "StartupCPUShares")) {
                 uint64_t u;