chiark / gitweb /
core: add cgroup CPU controller support on the unified hierarchy
[elogind.git] / src / core / cgroup.c
index 35b862d..b53b35d 100644 (file)
@@ -1,5 +1,3 @@
-/*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
-
 /***
   This file is part of systemd.
 
 #include <fcntl.h>
 #include <fnmatch.h>
 
-#include "path-util.h"
-#include "special.h"
+#include "alloc-util.h"
 #include "cgroup-util.h"
 #include "cgroup.h"
+#include "fd-util.h"
+#include "fileio.h"
+#include "fs-util.h"
+#include "parse-util.h"
+#include "path-util.h"
+#include "process-util.h"
+//#include "special.h"
+#include "string-table.h"
+#include "string-util.h"
+#include "stdio-util.h"
 
 #define CGROUP_CPU_QUOTA_PERIOD_USEC ((usec_t) 100 * USEC_PER_MSEC)
 
+#if 0 /// UNNEEDED by elogind
+static void cgroup_compat_warn(void) {
+        static bool cgroup_compat_warned = false;
+
+        if (cgroup_compat_warned)
+                return;
+
+        log_warning("cgroup compatibility translation between legacy and unified hierarchy settings activated. See cgroup-compat debug messages for details.");
+        cgroup_compat_warned = true;
+}
+
+#define log_cgroup_compat(unit, fmt, ...) do {                                  \
+                cgroup_compat_warn();                                           \
+                log_unit_debug(unit, "cgroup-compat: " fmt, ##__VA_ARGS__);     \
+        } while (false)
+
 void cgroup_context_init(CGroupContext *c) {
         assert(c);
 
         /* Initialize everything to the kernel defaults, assuming the
          * structure is preinitialized to 0 */
 
-        c->cpu_shares = (unsigned long) -1;
-        c->startup_cpu_shares = (unsigned long) -1;
-        c->memory_limit = (uint64_t) -1;
-        c->blockio_weight = (unsigned long) -1;
-        c->startup_blockio_weight = (unsigned long) -1;
-
+        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->memory_high = CGROUP_LIMIT_MAX;
+        c->memory_max = CGROUP_LIMIT_MAX;
+
+        c->memory_limit = CGROUP_LIMIT_MAX;
+
+        c->io_weight = CGROUP_WEIGHT_INVALID;
+        c->startup_io_weight = CGROUP_WEIGHT_INVALID;
+
+        c->blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
+        c->startup_blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
+
+        c->tasks_max = (uint64_t) -1;
 }
 
 void cgroup_context_free_device_allow(CGroupContext *c, CGroupDeviceAllow *a) {
@@ -53,6 +88,24 @@ void cgroup_context_free_device_allow(CGroupContext *c, CGroupDeviceAllow *a) {
         free(a);
 }
 
+void cgroup_context_free_io_device_weight(CGroupContext *c, CGroupIODeviceWeight *w) {
+        assert(c);
+        assert(w);
+
+        LIST_REMOVE(device_weights, c->io_device_weights, w);
+        free(w->path);
+        free(w);
+}
+
+void cgroup_context_free_io_device_limit(CGroupContext *c, CGroupIODeviceLimit *l) {
+        assert(c);
+        assert(l);
+
+        LIST_REMOVE(device_limits, c->io_device_limits, l);
+        free(l->path);
+        free(l);
+}
+
 void cgroup_context_free_blockio_device_weight(CGroupContext *c, CGroupBlockIODeviceWeight *w) {
         assert(c);
         assert(w);
@@ -74,6 +127,12 @@ void cgroup_context_free_blockio_device_bandwidth(CGroupContext *c, CGroupBlockI
 void cgroup_context_done(CGroupContext *c) {
         assert(c);
 
+        while (c->io_device_weights)
+                cgroup_context_free_io_device_weight(c, c->io_device_weights);
+
+        while (c->io_device_limits)
+                cgroup_context_free_io_device_limit(c, c->io_device_limits);
+
         while (c->blockio_device_weights)
                 cgroup_context_free_blockio_device_weight(c, c->blockio_device_weights);
 
@@ -85,6 +144,8 @@ void cgroup_context_done(CGroupContext *c) {
 }
 
 void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
+        CGroupIODeviceLimit *il;
+        CGroupIODeviceWeight *iw;
         CGroupBlockIODeviceBandwidth *b;
         CGroupBlockIODeviceWeight *w;
         CGroupDeviceAllow *a;
@@ -97,25 +158,45 @@ void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
 
         fprintf(f,
                 "%sCPUAccounting=%s\n"
+                "%sIOAccounting=%s\n"
                 "%sBlockIOAccounting=%s\n"
                 "%sMemoryAccounting=%s\n"
-                "%sCPUShares=%lu\n"
-                "%sStartupCPUShares=%lu\n"
+                "%sTasksAccounting=%s\n"
+                "%sCPUWeight=%" PRIu64 "\n"
+                "%sStartupCPUWeight=%" PRIu64 "\n"
+                "%sCPUShares=%" PRIu64 "\n"
+                "%sStartupCPUShares=%" PRIu64 "\n"
                 "%sCPUQuotaPerSecSec=%s\n"
-                "%sBlockIOWeight=%lu\n"
-                "%sStartupBlockIOWeight=%lu\n"
+                "%sIOWeight=%" PRIu64 "\n"
+                "%sStartupIOWeight=%" PRIu64 "\n"
+                "%sBlockIOWeight=%" PRIu64 "\n"
+                "%sStartupBlockIOWeight=%" PRIu64 "\n"
+                "%sMemoryLow=%" PRIu64 "\n"
+                "%sMemoryHigh=%" PRIu64 "\n"
+                "%sMemoryMax=%" PRIu64 "\n"
                 "%sMemoryLimit=%" PRIu64 "\n"
+                "%sTasksMax=%" PRIu64 "\n"
                 "%sDevicePolicy=%s\n"
                 "%sDelegate=%s\n",
                 prefix, yes_no(c->cpu_accounting),
+                prefix, yes_no(c->io_accounting),
                 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),
+                prefix, c->io_weight,
+                prefix, c->startup_io_weight,
                 prefix, c->blockio_weight,
                 prefix, c->startup_blockio_weight,
+                prefix, c->memory_low,
+                prefix, c->memory_high,
+                prefix, c->memory_max,
                 prefix, c->memory_limit,
+                prefix, c->tasks_max,
                 prefix, cgroup_device_policy_to_string(c->device_policy),
                 prefix, yes_no(c->delegate));
 
@@ -126,9 +207,30 @@ void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
                         a->path,
                         a->r ? "r" : "", a->w ? "w" : "", a->m ? "m" : "");
 
+        LIST_FOREACH(device_weights, iw, c->io_device_weights)
+                fprintf(f,
+                        "%sIODeviceWeight=%s %" PRIu64,
+                        prefix,
+                        iw->path,
+                        iw->weight);
+
+        LIST_FOREACH(device_limits, il, c->io_device_limits) {
+                char buf[FORMAT_BYTES_MAX];
+                CGroupIOLimitType type;
+
+                for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
+                        if (il->limits[type] != cgroup_io_limit_defaults[type])
+                                fprintf(f,
+                                        "%s%s=%s %s\n",
+                                        prefix,
+                                        cgroup_io_limit_type_to_string(type),
+                                        il->path,
+                                        format_bytes(buf, sizeof(buf), il->limits[type]));
+        }
+
         LIST_FOREACH(device_weights, w, c->blockio_device_weights)
                 fprintf(f,
-                        "%sBlockIODeviceWeight=%s %lu",
+                        "%sBlockIODeviceWeight=%s %" PRIu64,
                         prefix,
                         w->path,
                         w->weight);
@@ -136,16 +238,22 @@ void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
         LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
                 char buf[FORMAT_BYTES_MAX];
 
-                fprintf(f,
-                        "%s%s=%s %s\n",
-                        prefix,
-                        b->read ? "BlockIOReadBandwidth" : "BlockIOWriteBandwidth",
-                        b->path,
-                        format_bytes(buf, sizeof(buf), b->bandwidth));
+                if (b->rbps != CGROUP_LIMIT_MAX)
+                        fprintf(f,
+                                "%sBlockIOReadBandwidth=%s %s\n",
+                                prefix,
+                                b->path,
+                                format_bytes(buf, sizeof(buf), b->rbps));
+                if (b->wbps != CGROUP_LIMIT_MAX)
+                        fprintf(f,
+                                "%sBlockIOWriteBandwidth=%s %s\n",
+                                prefix,
+                                b->path,
+                                format_bytes(buf, sizeof(buf), b->wbps));
         }
 }
 
-static int lookup_blkio_device(const char *p, dev_t *dev) {
+static int lookup_block_device(const char *p, dev_t *dev) {
         struct stat st;
         int r;
 
@@ -200,7 +308,8 @@ static int whitelist_device(const char *path, const char *node, const char *acc)
 
         r = cg_set_attribute("devices", path, "devices.allow", buf);
         if (r < 0)
-                log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set devices.allow on %s: %s", path, strerror(-r));
+                log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                               "Failed to set devices.allow on %s: %m", path);
 
         return r;
 }
@@ -270,7 +379,8 @@ static int whitelist_major(const char *path, const char *name, char type, const
 
                 r = cg_set_attribute("devices", path, "devices.allow", buf);
                 if (r < 0)
-                        log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set devices.allow on %s: %s", path, strerror(-r));
+                        log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                       "Failed to set devices.allow on %s: %m", path);
         }
 
         return 0;
@@ -280,113 +390,515 @@ fail:
         return -errno;
 }
 
-void cgroup_context_apply(CGroupContext *c, CGroupControllerMask mask, const char *path, ManagerState state) {
+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 ||
+                c->startup_io_weight != CGROUP_WEIGHT_INVALID ||
+                c->io_device_weights ||
+                c->io_device_limits;
+}
+
+static bool cgroup_context_has_blockio_config(CGroupContext *c) {
+        return c->blockio_accounting ||
+                c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
+                c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
+                c->blockio_device_weights ||
+                c->blockio_device_bandwidths;
+}
+
+static uint64_t cgroup_context_io_weight(CGroupContext *c, ManagerState state) {
+        if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
+            c->startup_io_weight != CGROUP_WEIGHT_INVALID)
+                return c->startup_io_weight;
+        else if (c->io_weight != CGROUP_WEIGHT_INVALID)
+                return c->io_weight;
+        else
+                return CGROUP_WEIGHT_DEFAULT;
+}
+
+static uint64_t cgroup_context_blkio_weight(CGroupContext *c, ManagerState state) {
+        if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
+            c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
+                return c->startup_blockio_weight;
+        else if (c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
+                return c->blockio_weight;
+        else
+                return CGROUP_BLKIO_WEIGHT_DEFAULT;
+}
+
+static uint64_t cgroup_weight_blkio_to_io(uint64_t blkio_weight) {
+        return CLAMP(blkio_weight * CGROUP_WEIGHT_DEFAULT / CGROUP_BLKIO_WEIGHT_DEFAULT,
+                     CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
+}
+
+static uint64_t cgroup_weight_io_to_blkio(uint64_t io_weight) {
+        return CLAMP(io_weight * CGROUP_BLKIO_WEIGHT_DEFAULT / CGROUP_WEIGHT_DEFAULT,
+                     CGROUP_BLKIO_WEIGHT_MIN, CGROUP_BLKIO_WEIGHT_MAX);
+}
+
+static void cgroup_apply_io_device_weight(Unit *u, const char *dev_path, uint64_t io_weight) {
+        char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
+        dev_t dev;
+        int r;
+
+        r = lookup_block_device(dev_path, &dev);
+        if (r < 0)
+                return;
+
+        xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), io_weight);
+        r = cg_set_attribute("io", u->cgroup_path, "io.weight", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set io.weight: %m");
+}
+
+static void cgroup_apply_blkio_device_weight(Unit *u, const char *dev_path, uint64_t blkio_weight) {
+        char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
+        dev_t dev;
+        int r;
+
+        r = lookup_block_device(dev_path, &dev);
+        if (r < 0)
+                return;
+
+        xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), blkio_weight);
+        r = cg_set_attribute("blkio", u->cgroup_path, "blkio.weight_device", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set blkio.weight_device: %m");
+}
+
+static unsigned cgroup_apply_io_device_limit(Unit *u, const char *dev_path, uint64_t *limits) {
+        char limit_bufs[_CGROUP_IO_LIMIT_TYPE_MAX][DECIMAL_STR_MAX(uint64_t)];
+        char buf[DECIMAL_STR_MAX(dev_t)*2+2+(6+DECIMAL_STR_MAX(uint64_t)+1)*4];
+        CGroupIOLimitType type;
+        dev_t dev;
+        unsigned n = 0;
+        int r;
+
+        r = lookup_block_device(dev_path, &dev);
+        if (r < 0)
+                return 0;
+
+        for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++) {
+                if (limits[type] != cgroup_io_limit_defaults[type]) {
+                        xsprintf(limit_bufs[type], "%" PRIu64, limits[type]);
+                        n++;
+                } else {
+                        xsprintf(limit_bufs[type], "%s", limits[type] == CGROUP_LIMIT_MAX ? "max" : "0");
+                }
+        }
+
+        xsprintf(buf, "%u:%u rbps=%s wbps=%s riops=%s wiops=%s\n", major(dev), minor(dev),
+                 limit_bufs[CGROUP_IO_RBPS_MAX], limit_bufs[CGROUP_IO_WBPS_MAX],
+                 limit_bufs[CGROUP_IO_RIOPS_MAX], limit_bufs[CGROUP_IO_WIOPS_MAX]);
+        r = cg_set_attribute("io", u->cgroup_path, "io.max", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set io.max: %m");
+        return n;
+}
+
+static unsigned cgroup_apply_blkio_device_limit(Unit *u, const char *dev_path, uint64_t rbps, uint64_t wbps) {
+        char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
+        dev_t dev;
+        unsigned n = 0;
+        int r;
+
+        r = lookup_block_device(dev_path, &dev);
+        if (r < 0)
+                return 0;
+
+        if (rbps != CGROUP_LIMIT_MAX)
+                n++;
+        sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), rbps);
+        r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.read_bps_device", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set blkio.throttle.read_bps_device: %m");
+
+        if (wbps != CGROUP_LIMIT_MAX)
+                n++;
+        sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), wbps);
+        r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.write_bps_device", buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set blkio.throttle.write_bps_device: %m");
+
+        return n;
+}
+
+static bool cgroup_context_has_unified_memory_config(CGroupContext *c) {
+        return c->memory_low > 0 || c->memory_high != CGROUP_LIMIT_MAX || c->memory_max != CGROUP_LIMIT_MAX;
+}
+
+static void cgroup_apply_unified_memory_limit(Unit *u, const char *file, uint64_t v) {
+        char buf[DECIMAL_STR_MAX(uint64_t) + 1] = "max";
+        int r;
+
+        if (v != CGROUP_LIMIT_MAX)
+                xsprintf(buf, "%" PRIu64 "\n", v);
+
+        r = cg_set_attribute("memory", u->cgroup_path, file, buf);
+        if (r < 0)
+                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                              "Failed to set %s: %m", file);
+}
+
+static void cgroup_context_apply(Unit *u, CGroupMask mask, ManagerState state) {
+        const char *path;
+        CGroupContext *c;
         bool is_root;
         int r;
 
+        assert(u);
+
+        c = unit_get_cgroup_context(u);
+        path = u->cgroup_path;
+
         assert(c);
         assert(path);
 
         if (mask == 0)
                 return;
 
-        /* Some cgroup attributes are not support on the root cgroup,
+        /* Some cgroup attributes are not supported on the root cgroup,
          * hence silently ignore */
         is_root = isempty(path) || path_equal(path, "/");
+        if (is_root)
+                /* Make sure we don't try to display messages with an empty path. */
+                path = "/";
 
-        if ((mask & CGROUP_CPU) && !is_root) {
-                char buf[MAX(DECIMAL_STR_MAX(unsigned long), DECIMAL_STR_MAX(usec_t)) + 1];
+        /* We generally ignore errors caused by read-only mounted
+         * cgroup trees (assuming we are running in a container then),
+         * and missing cgroups, i.e. EROFS and ENOENT. */
 
-                sprintf(buf, "%lu\n",
-                        IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) && c->startup_cpu_shares != (unsigned long) -1 ? c->startup_cpu_shares :
-                        c->cpu_shares != (unsigned long) -1 ? c->cpu_shares : 1024);
-                r = cg_set_attribute("cpu", path, "cpu.shares", buf);
-                if (r < 0)
-                        log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set cpu.shares on %s: %s", path, strerror(-r));
+        if ((mask & CGROUP_MASK_CPU) && !is_root) {
+                bool has_weight = cgroup_context_has_cpu_weight(c);
+                bool has_shares = cgroup_context_has_cpu_shares(c);
 
-                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_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set cpu.cfs_period_us on %s: %s", path, strerror(-r));
+                if (cg_unified() > 0) {
+                        uint64_t weight;
 
-                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_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set cpu.cfs_quota_us on %s: %s", path, strerror(-r));
+                        if (has_weight)
+                                weight = cgroup_context_cpu_weight(c, state);
+                        else if (has_shares) {
+                                uint64_t shares = cgroup_context_cpu_shares(c, state);
+
+                                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_BLKIO) {
-                char buf[MAX3(DECIMAL_STR_MAX(unsigned long)+1,
-                              DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(unsigned long)*1,
-                              DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1)];
-                CGroupBlockIODeviceWeight *w;
-                CGroupBlockIODeviceBandwidth *b;
+        if (mask & CGROUP_MASK_IO) {
+                bool has_io = cgroup_context_has_io_config(c);
+                bool has_blockio = cgroup_context_has_blockio_config(c);
 
                 if (!is_root) {
-                        sprintf(buf, "%lu\n", IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) && c->startup_blockio_weight != (unsigned long) -1 ? c->startup_blockio_weight :
-                                c->blockio_weight != (unsigned long) -1 ? c->blockio_weight : 1000);
-                        r = cg_set_attribute("blkio", path, "blkio.weight", buf);
+                        char buf[8+DECIMAL_STR_MAX(uint64_t)+1];
+                        uint64_t weight;
+
+                        if (has_io)
+                                weight = cgroup_context_io_weight(c, state);
+                        else if (has_blockio) {
+                                uint64_t blkio_weight = cgroup_context_blkio_weight(c, state);
+
+                                weight = cgroup_weight_blkio_to_io(blkio_weight);
+
+                                log_cgroup_compat(u, "Applying [Startup]BlockIOWeight %" PRIu64 " as [Startup]IOWeight %" PRIu64,
+                                                  blkio_weight, weight);
+                        } else
+                                weight = CGROUP_WEIGHT_DEFAULT;
+
+                        xsprintf(buf, "default %" PRIu64 "\n", weight);
+                        r = cg_set_attribute("io", path, "io.weight", buf);
                         if (r < 0)
-                                log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set blkio.weight on %s: %s", path, strerror(-r));
+                                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                              "Failed to set io.weight: %m");
 
-                        /* FIXME: no way to reset this list */
-                        LIST_FOREACH(device_weights, w, c->blockio_device_weights) {
-                                dev_t dev;
+                        if (has_io) {
+                                CGroupIODeviceWeight *w;
 
-                                r = lookup_blkio_device(w->path, &dev);
-                                if (r < 0)
-                                        continue;
+                                /* FIXME: no way to reset this list */
+                                LIST_FOREACH(device_weights, w, c->io_device_weights)
+                                        cgroup_apply_io_device_weight(u, w->path, w->weight);
+                        } else if (has_blockio) {
+                                CGroupBlockIODeviceWeight *w;
 
-                                sprintf(buf, "%u:%u %lu", major(dev), minor(dev), w->weight);
-                                r = cg_set_attribute("blkio", path, "blkio.weight_device", buf);
-                                if (r < 0)
-                                        log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set blkio.weight_device on %s: %s", path, strerror(-r));
+                                /* FIXME: no way to reset this list */
+                                LIST_FOREACH(device_weights, w, c->blockio_device_weights) {
+                                        weight = cgroup_weight_blkio_to_io(w->weight);
+
+                                        log_cgroup_compat(u, "Applying BlockIODeviceWeight %" PRIu64 " as IODeviceWeight %" PRIu64 " for %s",
+                                                          w->weight, weight, w->path);
+
+                                        cgroup_apply_io_device_weight(u, w->path, weight);
+                                }
                         }
                 }
 
-                /* FIXME: no way to reset this list */
-                LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
-                        const char *a;
-                        dev_t dev;
+                /* Apply limits and free ones without config. */
+                if (has_io) {
+                        CGroupIODeviceLimit *l, *next;
 
-                        r = lookup_blkio_device(b->path, &dev);
-                        if (r < 0)
-                                continue;
+                        LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
+                                if (!cgroup_apply_io_device_limit(u, l->path, l->limits))
+                                        cgroup_context_free_io_device_limit(c, l);
+                        }
+                } else if (has_blockio) {
+                        CGroupBlockIODeviceBandwidth *b, *next;
+
+                        LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths) {
+                                uint64_t limits[_CGROUP_IO_LIMIT_TYPE_MAX];
+                                CGroupIOLimitType type;
+
+                                for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
+                                        limits[type] = cgroup_io_limit_defaults[type];
+
+                                limits[CGROUP_IO_RBPS_MAX] = b->rbps;
+                                limits[CGROUP_IO_WBPS_MAX] = b->wbps;
 
-                        a = b->read ? "blkio.throttle.read_bps_device" : "blkio.throttle.write_bps_device";
+                                log_cgroup_compat(u, "Applying BlockIO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as IO{Read|Write}BandwidthMax for %s",
+                                                  b->rbps, b->wbps, b->path);
 
-                        sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), b->bandwidth);
-                        r = cg_set_attribute("blkio", path, a, buf);
+                                if (!cgroup_apply_io_device_limit(u, b->path, limits))
+                                        cgroup_context_free_blockio_device_bandwidth(c, b);
+                        }
+                }
+        }
+
+        if (mask & CGROUP_MASK_BLKIO) {
+                bool has_io = cgroup_context_has_io_config(c);
+                bool has_blockio = cgroup_context_has_blockio_config(c);
+
+                if (!is_root) {
+                        char buf[DECIMAL_STR_MAX(uint64_t)+1];
+                        uint64_t weight;
+
+                        if (has_blockio)
+                                weight = cgroup_context_blkio_weight(c, state);
+                        else if (has_io) {
+                                uint64_t io_weight = cgroup_context_io_weight(c, state);
+
+                                weight = cgroup_weight_io_to_blkio(cgroup_context_io_weight(c, state));
+
+                                log_cgroup_compat(u, "Applying [Startup]IOWeight %" PRIu64 " as [Startup]BlockIOWeight %" PRIu64,
+                                                  io_weight, weight);
+                        } else
+                                weight = CGROUP_BLKIO_WEIGHT_DEFAULT;
+
+                        xsprintf(buf, "%" PRIu64 "\n", weight);
+                        r = cg_set_attribute("blkio", path, "blkio.weight", buf);
                         if (r < 0)
-                                log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set %s on %s: %s", a, path, strerror(-r));
+                                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                              "Failed to set blkio.weight: %m");
+
+                        if (has_blockio) {
+                                CGroupBlockIODeviceWeight *w;
+
+                                /* FIXME: no way to reset this list */
+                                LIST_FOREACH(device_weights, w, c->blockio_device_weights)
+                                        cgroup_apply_blkio_device_weight(u, w->path, w->weight);
+                        } else if (has_io) {
+                                CGroupIODeviceWeight *w;
+
+                                /* FIXME: no way to reset this list */
+                                LIST_FOREACH(device_weights, w, c->io_device_weights) {
+                                        weight = cgroup_weight_io_to_blkio(w->weight);
+
+                                        log_cgroup_compat(u, "Applying IODeviceWeight %" PRIu64 " as BlockIODeviceWeight %" PRIu64 " for %s",
+                                                          w->weight, weight, w->path);
+
+                                        cgroup_apply_blkio_device_weight(u, w->path, weight);
+                                }
+                        }
+                }
+
+                /* Apply limits and free ones without config. */
+                if (has_blockio) {
+                        CGroupBlockIODeviceBandwidth *b, *next;
+
+                        LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths) {
+                                if (!cgroup_apply_blkio_device_limit(u, b->path, b->rbps, b->wbps))
+                                        cgroup_context_free_blockio_device_bandwidth(c, b);
+                        }
+                } else if (has_io) {
+                        CGroupIODeviceLimit *l, *next;
+
+                        LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
+                                log_cgroup_compat(u, "Applying IO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as BlockIO{Read|Write}BandwidthMax for %s",
+                                                  l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX], l->path);
+
+                                if (!cgroup_apply_blkio_device_limit(u, l->path, l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX]))
+                                        cgroup_context_free_io_device_limit(c, l);
+                        }
                 }
         }
 
-        if (mask & CGROUP_MEMORY) {
-                if (c->memory_limit != (uint64_t) -1) {
+        if ((mask & CGROUP_MASK_MEMORY) && !is_root) {
+                if (cg_unified() > 0) {
+                        uint64_t max = c->memory_max;
+
+                        if (cgroup_context_has_unified_memory_config(c))
+                                max = c->memory_max;
+                        else {
+                                max = c->memory_limit;
+
+                                if (max != CGROUP_LIMIT_MAX)
+                                        log_cgroup_compat(u, "Applying MemoryLimit %" PRIu64 " as MemoryMax", max);
+                        }
+
+                        cgroup_apply_unified_memory_limit(u, "memory.low", c->memory_low);
+                        cgroup_apply_unified_memory_limit(u, "memory.high", c->memory_high);
+                        cgroup_apply_unified_memory_limit(u, "memory.max", max);
+                } else {
                         char buf[DECIMAL_STR_MAX(uint64_t) + 1];
+                        uint64_t val = c->memory_limit;
 
-                        sprintf(buf, "%" PRIu64 "\n", c->memory_limit);
-                        r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
-                } else
-                        r = cg_set_attribute("memory", path, "memory.limit_in_bytes", "-1");
+                        if (val == CGROUP_LIMIT_MAX) {
+                                val = c->memory_max;
 
-                if (r < 0)
-                        log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to set memory.limit_in_bytes on %s: %s", path, strerror(-r));
+                                if (val != CGROUP_LIMIT_MAX)
+                                        log_cgroup_compat(u, "Applying MemoryMax %" PRIi64 " as MemoryLimit", c->memory_max);
+                        }
+
+                        if (val == CGROUP_LIMIT_MAX)
+                                strncpy(buf, "-1\n", sizeof(buf));
+                        else
+                                xsprintf(buf, "%" PRIu64 "\n", val);
+
+                        r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
+                        if (r < 0)
+                                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                              "Failed to set memory.limit_in_bytes: %m");
+                }
         }
 
-        if ((mask & CGROUP_DEVICE) && !is_root) {
+        if ((mask & CGROUP_MASK_DEVICES) && !is_root) {
                 CGroupDeviceAllow *a;
 
+                /* Changing the devices list of a populated cgroup
+                 * might result in EINVAL, hence ignore EINVAL
+                 * here. */
+
                 if (c->device_allow || c->device_policy != CGROUP_AUTO)
                         r = cg_set_attribute("devices", path, "devices.deny", "a");
                 else
                         r = cg_set_attribute("devices", path, "devices.allow", "a");
                 if (r < 0)
-                        log_full(r == -ENOENT ? LOG_DEBUG : LOG_WARNING, "Failed to reset devices.list on %s: %s", path, strerror(-r));
+                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                      "Failed to reset devices.list: %m");
 
                 if (c->device_policy == CGROUP_CLOSED ||
                     (c->device_policy == CGROUP_AUTO && c->device_allow)) {
@@ -397,7 +909,11 @@ void cgroup_context_apply(CGroupContext *c, CGroupControllerMask mask, const cha
                                 "/dev/random\0" "rwm\0"
                                 "/dev/urandom\0" "rwm\0"
                                 "/dev/tty\0" "rwm\0"
-                                "/dev/pts/ptmx\0" "rw\0"; /* /dev/pts/ptmx may not be duplicated, but accessed */
+                                "/dev/pts/ptmx\0" "rw\0" /* /dev/pts/ptmx may not be duplicated, but accessed */
+                                /* Allow /run/elogind/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
+                                /* Allow /run/systemd/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
+                                "/run/systemd/inaccessible/chr\0" "rwm\0"
+                                "/run/systemd/inaccessible/blk\0" "rwm\0";
 
                         const char *x, *y;
 
@@ -432,66 +948,92 @@ void cgroup_context_apply(CGroupContext *c, CGroupControllerMask mask, const cha
                         else if (startswith(a->path, "char-"))
                                 whitelist_major(path, a->path + 5, 'c', acc);
                         else
-                                log_debug("Ignoring device %s while writing cgroup attribute.", a->path);
+                                log_unit_debug(u, "Ignoring device %s while writing cgroup attribute.", a->path);
                 }
         }
+
+        if ((mask & CGROUP_MASK_PIDS) && !is_root) {
+
+                if (c->tasks_max != (uint64_t) -1) {
+                        char buf[DECIMAL_STR_MAX(uint64_t) + 2];
+
+                        sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
+                        r = cg_set_attribute("pids", path, "pids.max", buf);
+                } else
+                        r = cg_set_attribute("pids", path, "pids.max", "max");
+
+                if (r < 0)
+                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                      "Failed to set pids.max: %m");
+        }
 }
 
-CGroupControllerMask cgroup_context_get_mask(CGroupContext *c) {
-        CGroupControllerMask mask = 0;
+CGroupMask cgroup_context_get_mask(CGroupContext *c) {
+        CGroupMask mask = 0;
 
         /* Figure out which controllers we need */
 
         if (c->cpu_accounting ||
-            c->cpu_shares != (unsigned long) -1 ||
-            c->startup_cpu_shares != (unsigned long) -1 ||
+            cgroup_context_has_cpu_weight(c) ||
+            cgroup_context_has_cpu_shares(c) ||
             c->cpu_quota_per_sec_usec != USEC_INFINITY)
-                mask |= CGROUP_CPUACCT | CGROUP_CPU;
+                mask |= CGROUP_MASK_CPUACCT | CGROUP_MASK_CPU;
 
-        if (c->blockio_accounting ||
-            c->blockio_weight != (unsigned long) -1 ||
-            c->startup_blockio_weight != (unsigned long) -1 ||
-            c->blockio_device_weights ||
-            c->blockio_device_bandwidths)
-                mask |= CGROUP_BLKIO;
+        if (cgroup_context_has_io_config(c) || cgroup_context_has_blockio_config(c))
+                mask |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
 
         if (c->memory_accounting ||
-            c->memory_limit != (uint64_t) -1)
-                mask |= CGROUP_MEMORY;
+            c->memory_limit != CGROUP_LIMIT_MAX ||
+            cgroup_context_has_unified_memory_config(c))
+                mask |= CGROUP_MASK_MEMORY;
 
         if (c->device_allow ||
             c->device_policy != CGROUP_AUTO)
-                mask |= CGROUP_DEVICE;
+                mask |= CGROUP_MASK_DEVICES;
+
+        if (c->tasks_accounting ||
+            c->tasks_max != (uint64_t) -1)
+                mask |= CGROUP_MASK_PIDS;
 
         return mask;
 }
 
-CGroupControllerMask unit_get_cgroup_mask(Unit *u) {
+CGroupMask unit_get_own_mask(Unit *u) {
         CGroupContext *c;
 
+        /* Returns the mask of controllers the unit needs for itself */
+
         c = unit_get_cgroup_context(u);
         if (!c)
                 return 0;
 
         /* If delegation is turned on, then turn on all cgroups,
-         * unless the process we fork into it is known to drop
-         * privileges anyway, and shouldn't get access to the
-         * controllers anyway. */
+         * unless we are on the legacy hierarchy and the process we
+         * fork into it is known to drop privileges, and hence
+         * shouldn't get access to the controllers.
+         *
+         * Note that on the unified hierarchy it is safe to delegate
+         * controllers to unprivileged services. */
 
         if (c->delegate) {
                 ExecContext *e;
 
                 e = unit_get_exec_context(u);
-                if (!e || exec_context_maintains_privileges(e))
-                        return _CGROUP_CONTROLLER_MASK_ALL;
+                if (!e ||
+                    exec_context_maintains_privileges(e) ||
+                    cg_unified() > 0)
+                        return _CGROUP_MASK_ALL;
         }
 
         return cgroup_context_get_mask(c);
 }
 
-CGroupControllerMask unit_get_members_mask(Unit *u) {
+CGroupMask unit_get_members_mask(Unit *u) {
         assert(u);
 
+        /* Returns the mask of controllers all of the unit's children
+         * require, merged */
+
         if (u->cgroup_members_mask_valid)
                 return u->cgroup_members_mask;
 
@@ -510,7 +1052,7 @@ CGroupControllerMask unit_get_members_mask(Unit *u) {
                                 continue;
 
                         u->cgroup_members_mask |=
-                                unit_get_cgroup_mask(member) |
+                                unit_get_own_mask(member) |
                                 unit_get_members_mask(member);
                 }
         }
@@ -519,19 +1061,52 @@ CGroupControllerMask unit_get_members_mask(Unit *u) {
         return u->cgroup_members_mask;
 }
 
-CGroupControllerMask unit_get_siblings_mask(Unit *u) {
+CGroupMask unit_get_siblings_mask(Unit *u) {
         assert(u);
 
+        /* Returns the mask of controllers all of the unit's siblings
+         * require, i.e. the members mask of the unit's parent slice
+         * if there is one. */
+
         if (UNIT_ISSET(u->slice))
                 return unit_get_members_mask(UNIT_DEREF(u->slice));
 
-        return unit_get_cgroup_mask(u) | unit_get_members_mask(u);
+        return unit_get_own_mask(u) | unit_get_members_mask(u);
+}
+
+CGroupMask unit_get_subtree_mask(Unit *u) {
+
+        /* Returns the mask of this subtree, meaning of the group
+         * itself and its children. */
+
+        return unit_get_own_mask(u) | unit_get_members_mask(u);
+}
+
+CGroupMask unit_get_target_mask(Unit *u) {
+        CGroupMask mask;
+
+        /* This returns the cgroup mask of all controllers to enable
+         * for a specific cgroup, i.e. everything it needs itself,
+         * plus all that its children need, plus all that its siblings
+         * need. This is primarily useful on the legacy cgroup
+         * hierarchy, where we need to duplicate each cgroup in each
+         * hierarchy that shall be enabled for it. */
+
+        mask = unit_get_own_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
+        mask &= u->manager->cgroup_supported;
+
+        return mask;
 }
 
-CGroupControllerMask unit_get_target_mask(Unit *u) {
-        CGroupControllerMask mask;
+CGroupMask unit_get_enable_mask(Unit *u) {
+        CGroupMask mask;
+
+        /* This returns the cgroup mask of all controllers to enable
+         * for the children of a specific cgroup. This is primarily
+         * useful for the unified cgroup hierarchy, where each cgroup
+         * controls which controllers are enabled for its children. */
 
-        mask = unit_get_cgroup_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
+        mask = unit_get_members_mask(u);
         mask &= u->manager->cgroup_supported;
 
         return mask;
@@ -540,13 +1115,13 @@ CGroupControllerMask unit_get_target_mask(Unit *u) {
 /* Recurse from a unit up through its containing slices, propagating
  * mask bits upward. A unit is also member of itself. */
 void unit_update_cgroup_members_masks(Unit *u) {
-        CGroupControllerMask m;
+        CGroupMask m;
         bool more;
 
         assert(u);
 
         /* Calculate subtree mask */
-        m = unit_get_cgroup_mask(u) | unit_get_members_mask(u);
+        m = unit_get_subtree_mask(u);
 
         /* See if anything changed from the previous invocation. If
          * not, we're done. */
@@ -586,7 +1161,7 @@ void unit_update_cgroup_members_masks(Unit *u) {
         }
 }
 
-static const char *migrate_callback(CGroupControllerMask mask, void *userdata) {
+static const char *migrate_callback(CGroupMask mask, void *userdata) {
         Unit *u = userdata;
 
         assert(mask != 0);
@@ -604,55 +1179,168 @@ static const char *migrate_callback(CGroupControllerMask mask, void *userdata) {
         return NULL;
 }
 
-static int unit_create_cgroups(Unit *u, CGroupControllerMask mask) {
-        CGroupContext *c;
+char *unit_default_cgroup_path(Unit *u) {
+        _cleanup_free_ char *escaped = NULL, *slice = NULL;
         int r;
 
         assert(u);
 
-        c = unit_get_cgroup_context(u);
-        if (!c)
-                return 0;
+        if (unit_has_name(u, SPECIAL_ROOT_SLICE))
+                return strdup(u->manager->cgroup_root);
 
-        if (!u->cgroup_path) {
-                _cleanup_free_ char *path = NULL;
+        if (UNIT_ISSET(u->slice) && !unit_has_name(UNIT_DEREF(u->slice), SPECIAL_ROOT_SLICE)) {
+                r = cg_slice_to_path(UNIT_DEREF(u->slice)->id, &slice);
+                if (r < 0)
+                        return NULL;
+        }
 
-                path = unit_default_cgroup_path(u);
-                if (!path)
-                        return log_oom();
+        escaped = cg_escape(u->id);
+        if (!escaped)
+                return NULL;
 
-                r = hashmap_put(u->manager->cgroup_unit, path, u);
-                if (r < 0) {
-                        log_error(r == -EEXIST ? "cgroup %s exists already: %s" : "hashmap_put failed for %s: %s", path, strerror(-r));
-                        return r;
-                }
-                if (r > 0) {
-                        u->cgroup_path = path;
-                        path = NULL;
-                }
-        }
+        if (slice)
+                return strjoin(u->manager->cgroup_root, "/", slice, "/", escaped, NULL);
+        else
+                return strjoin(u->manager->cgroup_root, "/", escaped, NULL);
+}
 
-        /* First, create our own group */
-        r = cg_create_everywhere(u->manager->cgroup_supported, mask, u->cgroup_path);
-        if (r < 0)
-                return log_error_errno(r, "Failed to create cgroup %s: %m", u->cgroup_path);
+int unit_set_cgroup_path(Unit *u, const char *path) {
+        _cleanup_free_ char *p = NULL;
+        int r;
 
-        /* Keep track that this is now realized */
-        u->cgroup_realized = true;
-        u->cgroup_realized_mask = mask;
+        assert(u);
 
-        if (u->type != UNIT_SLICE && !c->delegate) {
+        if (path) {
+                p = strdup(path);
+                if (!p)
+                        return -ENOMEM;
+        } else
+                p = NULL;
 
-                /* Then, possibly move things over, but not if
-                 * subgroups may contain processes, which is the case
-                 * for slice and delegation units. */
-                r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
+        if (streq_ptr(u->cgroup_path, p))
+                return 0;
+
+        if (p) {
+                r = hashmap_put(u->manager->cgroup_unit, p, u);
                 if (r < 0)
-                        log_warning_errno(r, "Failed to migrate cgroup from to %s: %m", u->cgroup_path);
+                        return r;
         }
 
-        return 0;
-}
+        unit_release_cgroup(u);
+
+        u->cgroup_path = p;
+        p = NULL;
+
+        return 1;
+}
+
+int unit_watch_cgroup(Unit *u) {
+        _cleanup_free_ char *events = NULL;
+        int r;
+
+        assert(u);
+
+        if (!u->cgroup_path)
+                return 0;
+
+        if (u->cgroup_inotify_wd >= 0)
+                return 0;
+
+        /* Only applies to the unified hierarchy */
+        r = cg_unified();
+        if (r < 0)
+                return log_unit_error_errno(u, r, "Failed detect whether the unified hierarchy is used: %m");
+        if (r == 0)
+                return 0;
+
+        /* Don't watch the root slice, it's pointless. */
+        if (unit_has_name(u, SPECIAL_ROOT_SLICE))
+                return 0;
+
+        r = hashmap_ensure_allocated(&u->manager->cgroup_inotify_wd_unit, &trivial_hash_ops);
+        if (r < 0)
+                return log_oom();
+
+        r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "cgroup.events", &events);
+        if (r < 0)
+                return log_oom();
+
+        u->cgroup_inotify_wd = inotify_add_watch(u->manager->cgroup_inotify_fd, events, IN_MODIFY);
+        if (u->cgroup_inotify_wd < 0) {
+
+                if (errno == ENOENT) /* If the directory is already
+                                      * gone we don't need to track
+                                      * it, so this is not an error */
+                        return 0;
+
+                return log_unit_error_errno(u, errno, "Failed to add inotify watch descriptor for control group %s: %m", u->cgroup_path);
+        }
+
+        r = hashmap_put(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd), u);
+        if (r < 0)
+                return log_unit_error_errno(u, r, "Failed to add inotify watch descriptor to hash map: %m");
+
+        return 0;
+}
+
+static int unit_create_cgroup(
+                Unit *u,
+                CGroupMask target_mask,
+                CGroupMask enable_mask) {
+
+        CGroupContext *c;
+        int r;
+
+        assert(u);
+
+        c = unit_get_cgroup_context(u);
+        if (!c)
+                return 0;
+
+        if (!u->cgroup_path) {
+                _cleanup_free_ char *path = NULL;
+
+                path = unit_default_cgroup_path(u);
+                if (!path)
+                        return log_oom();
+
+                r = unit_set_cgroup_path(u, path);
+                if (r == -EEXIST)
+                        return log_unit_error_errno(u, r, "Control group %s exists already.", path);
+                if (r < 0)
+                        return log_unit_error_errno(u, r, "Failed to set unit's control group path to %s: %m", path);
+        }
+
+        /* First, create our own group */
+        r = cg_create_everywhere(u->manager->cgroup_supported, target_mask, u->cgroup_path);
+        if (r < 0)
+                return log_unit_error_errno(u, r, "Failed to create cgroup %s: %m", u->cgroup_path);
+
+        /* Start watching it */
+        (void) unit_watch_cgroup(u);
+
+        /* Enable all controllers we need */
+        r = cg_enable_everywhere(u->manager->cgroup_supported, enable_mask, u->cgroup_path);
+        if (r < 0)
+                log_unit_warning_errno(u, r, "Failed to enable controllers on cgroup %s, ignoring: %m", u->cgroup_path);
+
+        /* Keep track that this is now realized */
+        u->cgroup_realized = true;
+        u->cgroup_realized_mask = target_mask;
+        u->cgroup_enabled_mask = enable_mask;
+
+        if (u->type != UNIT_SLICE && !c->delegate) {
+
+                /* Then, possibly move things over, but not if
+                 * subgroups may contain processes, which is the case
+                 * for slice and delegation units. */
+                r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
+                if (r < 0)
+                        log_unit_warning_errno(u, r, "Failed to migrate cgroup from to %s, ignoring: %m", u->cgroup_path);
+        }
+
+        return 0;
+}
 
 int unit_attach_pids_to_cgroup(Unit *u) {
         int r;
@@ -669,10 +1357,10 @@ int unit_attach_pids_to_cgroup(Unit *u) {
         return 0;
 }
 
-static bool unit_has_mask_realized(Unit *u, CGroupControllerMask mask) {
+static bool unit_has_mask_realized(Unit *u, CGroupMask target_mask, CGroupMask enable_mask) {
         assert(u);
 
-        return u->cgroup_realized && u->cgroup_realized_mask == mask;
+        return u->cgroup_realized && u->cgroup_realized_mask == target_mask && u->cgroup_enabled_mask == enable_mask;
 }
 
 /* Check if necessary controllers and attributes for a unit are in place.
@@ -682,7 +1370,7 @@ static bool unit_has_mask_realized(Unit *u, CGroupControllerMask mask) {
  *
  * Returns 0 on success and < 0 on failure. */
 static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
-        CGroupControllerMask mask;
+        CGroupMask target_mask, enable_mask;
         int r;
 
         assert(u);
@@ -692,9 +1380,10 @@ static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
                 u->in_cgroup_queue = false;
         }
 
-        mask = unit_get_target_mask(u);
+        target_mask = unit_get_target_mask(u);
+        enable_mask = unit_get_enable_mask(u);
 
-        if (unit_has_mask_realized(u, mask))
+        if (unit_has_mask_realized(u, target_mask, enable_mask))
                 return 0;
 
         /* First, realize parents */
@@ -705,12 +1394,12 @@ static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
         }
 
         /* And then do the real work */
-        r = unit_create_cgroups(u, mask);
+        r = unit_create_cgroup(u, target_mask, enable_mask);
         if (r < 0)
                 return r;
 
         /* Finally, apply the necessary attributes. */
-        cgroup_context_apply(unit_get_cgroup_context(u), mask, u->cgroup_path, state);
+        cgroup_context_apply(u, target_mask, state);
 
         return 0;
 }
@@ -737,7 +1426,7 @@ unsigned manager_dispatch_cgroup_queue(Manager *m) {
 
                 r = unit_realize_cgroup_now(i, state);
                 if (r < 0)
-                        log_warning_errno(r, "Failed to realize cgroups for queued unit %s: %m", i->id);
+                        log_warning_errno(r, "Failed to realize cgroups for queued unit %s, ignoring: %m", i->id);
 
                 n++;
         }
@@ -773,7 +1462,7 @@ static void unit_queue_siblings(Unit *u) {
                         /* If the unit doesn't need any new controllers
                          * and has current ones realized, it doesn't need
                          * any changes. */
-                        if (unit_has_mask_realized(m, unit_get_target_mask(m)))
+                        if (unit_has_mask_realized(m, unit_get_target_mask(m), unit_get_enable_mask(m)))
                                 continue;
 
                         unit_add_to_cgroup_queue(m);
@@ -784,12 +1473,9 @@ static void unit_queue_siblings(Unit *u) {
 }
 
 int unit_realize_cgroup(Unit *u) {
-        CGroupContext *c;
-
         assert(u);
 
-        c = unit_get_cgroup_context(u);
-        if (!c)
+        if (!UNIT_HAS_CGROUP_CONTEXT(u))
                 return 0;
 
         /* So, here's the deal: when realizing the cgroups for this
@@ -810,39 +1496,68 @@ int unit_realize_cgroup(Unit *u) {
         return unit_realize_cgroup_now(u, manager_state(u->manager));
 }
 
-void unit_destroy_cgroup_if_empty(Unit *u) {
+void unit_release_cgroup(Unit *u) {
+        assert(u);
+
+        /* Forgets all cgroup details for this cgroup */
+
+        if (u->cgroup_path) {
+                (void) hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
+                u->cgroup_path = mfree(u->cgroup_path);
+        }
+
+        if (u->cgroup_inotify_wd >= 0) {
+                if (inotify_rm_watch(u->manager->cgroup_inotify_fd, u->cgroup_inotify_wd) < 0)
+                        log_unit_debug_errno(u, errno, "Failed to remove cgroup inotify watch %i for %s, ignoring", u->cgroup_inotify_wd, u->id);
+
+                (void) hashmap_remove(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd));
+                u->cgroup_inotify_wd = -1;
+        }
+}
+
+void unit_prune_cgroup(Unit *u) {
         int r;
+        bool is_root_slice;
 
         assert(u);
 
+        /* Removes the cgroup, if empty and possible, and stops watching it. */
+
         if (!u->cgroup_path)
                 return;
 
-        r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !unit_has_name(u, SPECIAL_ROOT_SLICE));
+        is_root_slice = unit_has_name(u, SPECIAL_ROOT_SLICE);
+
+        r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !is_root_slice);
         if (r < 0) {
-                log_debug_errno(r, "Failed to destroy cgroup %s: %m", u->cgroup_path);
+                log_unit_debug_errno(u, r, "Failed to destroy cgroup %s, ignoring: %m", u->cgroup_path);
                 return;
         }
 
-        hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
+        if (is_root_slice)
+                return;
+
+        unit_release_cgroup(u);
 
-        free(u->cgroup_path);
-        u->cgroup_path = NULL;
         u->cgroup_realized = false;
         u->cgroup_realized_mask = 0;
+        u->cgroup_enabled_mask = 0;
 }
 
-pid_t unit_search_main_pid(Unit *u) {
+int unit_search_main_pid(Unit *u, pid_t *ret) {
         _cleanup_fclose_ FILE *f = NULL;
         pid_t pid = 0, npid, mypid;
+        int r;
 
         assert(u);
+        assert(ret);
 
         if (!u->cgroup_path)
-                return 0;
+                return -ENXIO;
 
-        if (cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f) < 0)
-                return 0;
+        r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f);
+        if (r < 0)
+                return r;
 
         mypid = getpid();
         while (cg_read_pid(f, &npid) > 0)  {
@@ -852,93 +1567,302 @@ pid_t unit_search_main_pid(Unit *u) {
                         continue;
 
                 /* Ignore processes that aren't our kids */
-                if (get_parent_of_pid(npid, &ppid) >= 0 && ppid != mypid)
+                if (get_process_ppid(npid, &ppid) >= 0 && ppid != mypid)
                         continue;
 
-                if (pid != 0) {
+                if (pid != 0)
                         /* Dang, there's more than one daemonized PID
                         in this group, so we don't know what process
                         is the main process. */
-                        pid = 0;
-                        break;
-                }
+
+                        return -ENODATA;
 
                 pid = npid;
         }
 
-        return pid;
+        *ret = pid;
+        return 0;
+}
+
+static int unit_watch_pids_in_path(Unit *u, const char *path) {
+        _cleanup_closedir_ DIR *d = NULL;
+        _cleanup_fclose_ FILE *f = NULL;
+        int ret = 0, r;
+
+        assert(u);
+        assert(path);
+
+        r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, path, &f);
+        if (r < 0)
+                ret = r;
+        else {
+                pid_t pid;
+
+                while ((r = cg_read_pid(f, &pid)) > 0) {
+                        r = unit_watch_pid(u, pid);
+                        if (r < 0 && ret >= 0)
+                                ret = r;
+                }
+
+                if (r < 0 && ret >= 0)
+                        ret = r;
+        }
+
+        r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
+        if (r < 0) {
+                if (ret >= 0)
+                        ret = r;
+        } else {
+                char *fn;
+
+                while ((r = cg_read_subgroup(d, &fn)) > 0) {
+                        _cleanup_free_ char *p = NULL;
+
+                        p = strjoin(path, "/", fn, NULL);
+                        free(fn);
+
+                        if (!p)
+                                return -ENOMEM;
+
+                        r = unit_watch_pids_in_path(u, p);
+                        if (r < 0 && ret >= 0)
+                                ret = r;
+                }
+
+                if (r < 0 && ret >= 0)
+                        ret = r;
+        }
+
+        return ret;
+}
+
+int unit_watch_all_pids(Unit *u) {
+        assert(u);
+
+        /* Adds all PIDs from our cgroup to the set of PIDs we
+         * watch. This is a fallback logic for cases where we do not
+         * get reliable cgroup empty notifications: we try to use
+         * SIGCHLD as replacement. */
+
+        if (!u->cgroup_path)
+                return -ENOENT;
+
+        if (cg_unified() > 0) /* On unified we can use proper notifications */
+                return 0;
+
+        return unit_watch_pids_in_path(u, u->cgroup_path);
 }
 
+int unit_notify_cgroup_empty(Unit *u) {
+        int r;
+
+        assert(u);
+
+        if (!u->cgroup_path)
+                return 0;
+
+        r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path);
+        if (r <= 0)
+                return r;
+
+        unit_add_to_gc_queue(u);
+
+        if (UNIT_VTABLE(u)->notify_cgroup_empty)
+                UNIT_VTABLE(u)->notify_cgroup_empty(u);
+
+        return 0;
+}
+
+static int on_cgroup_inotify_event(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
+        Manager *m = userdata;
+
+        assert(s);
+        assert(fd >= 0);
+        assert(m);
+
+        for (;;) {
+                union inotify_event_buffer buffer;
+                struct inotify_event *e;
+                ssize_t l;
+
+                l = read(fd, &buffer, sizeof(buffer));
+                if (l < 0) {
+                        if (errno == EINTR || errno == EAGAIN)
+                                return 0;
+
+                        return log_error_errno(errno, "Failed to read control group inotify events: %m");
+                }
+
+                FOREACH_INOTIFY_EVENT(e, buffer, l) {
+                        Unit *u;
+
+                        if (e->wd < 0)
+                                /* Queue overflow has no watch descriptor */
+                                continue;
+
+                        if (e->mask & IN_IGNORED)
+                                /* The watch was just removed */
+                                continue;
+
+                        u = hashmap_get(m->cgroup_inotify_wd_unit, INT_TO_PTR(e->wd));
+                        if (!u) /* Not that inotify might deliver
+                                 * events for a watch even after it
+                                 * was removed, because it was queued
+                                 * before the removal. Let's ignore
+                                 * this here safely. */
+                                continue;
+
+                        (void) unit_notify_cgroup_empty(u);
+                }
+        }
+}
+#endif // 0
+
 int manager_setup_cgroup(Manager *m) {
         _cleanup_free_ char *path = NULL;
-        int r;
+        CGroupController c;
+        int r, unified;
+        char *e;
 
         assert(m);
 
         /* 1. Determine hierarchy */
-        free(m->cgroup_root);
-        m->cgroup_root = NULL;
-
+        m->cgroup_root = mfree(m->cgroup_root);
         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &m->cgroup_root);
         if (r < 0)
                 return log_error_errno(r, "Cannot determine cgroup we are running in: %m");
 
-        /* LEGACY: Already in /system.slice? If so, let's cut this
-         * off. This is to support live upgrades from older systemd
-         * versions where PID 1 was moved there. */
-        if (m->running_as == SYSTEMD_SYSTEM) {
-                char *e;
+#if 0 /// elogind does not support systemd scopes and slices
+        /* Chop off the init scope, if we are already located in it */
+        e = endswith(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
 
+        /* LEGACY: Also chop off the system slice if we are in
+         * it. This is to support live upgrades from older systemd
+         * versions where PID 1 was moved there. Also see
+         * cg_get_root_path(). */
+        if (!e && MANAGER_IS_SYSTEM(m)) {
                 e = endswith(m->cgroup_root, "/" SPECIAL_SYSTEM_SLICE);
                 if (!e)
-                        e = endswith(m->cgroup_root, "/system");
-                if (e)
-                        *e = 0;
+                        e = endswith(m->cgroup_root, "/system"); /* even more legacy */
         }
+        if (e)
+                *e = 0;
+#endif // 0
 
         /* And make sure to store away the root value without trailing
          * slash, even for the root dir, so that we can easily prepend
          * it everywhere. */
-        if (streq(m->cgroup_root, "/"))
-                m->cgroup_root[0] = 0;
+        while ((e = endswith(m->cgroup_root, "/")))
+                *e = 0;
+        log_debug_elogind("Cgroup Controller \"%s\" -> root \"%s\"",
+                          SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root);
 
         /* 2. Show data */
         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, NULL, &path);
         if (r < 0)
                 return log_error_errno(r, "Cannot find cgroup mount point: %m");
 
-        log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER ". File system hierarchy is at %s.", path);
+        unified = cg_unified();
+        if (unified < 0)
+                return log_error_errno(r, "Couldn't determine if we are running in the unified hierarchy: %m");
+        if (unified > 0)
+                log_debug("Unified cgroup hierarchy is located at %s.", path);
+        else
+                log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER ". File system hierarchy is at %s.", path);
+
         if (!m->test_run) {
+                const char *scope_path;
 
                 /* 3. Install agent */
-                if (m->running_as == SYSTEMD_SYSTEM) {
+                if (unified) {
+
+                        /* In the unified hierarchy we can get
+                         * cgroup empty notifications via inotify. */
+
+#if 0 /// elogind does not support the unified hierarchy, yet.
+                        m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
+                        safe_close(m->cgroup_inotify_fd);
+
+                        m->cgroup_inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
+                        if (m->cgroup_inotify_fd < 0)
+                                return log_error_errno(errno, "Failed to create control group inotify object: %m");
+
+                        r = sd_event_add_io(m->event, &m->cgroup_inotify_event_source, m->cgroup_inotify_fd, EPOLLIN, on_cgroup_inotify_event, m);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to watch control group inotify object: %m");
+
+                        /* Process cgroup empty notifications early, but after service notifications and SIGCHLD. Also
+                         * see handling of cgroup agent notifications, for the classic cgroup hierarchy support. */
+                        r = sd_event_source_set_priority(m->cgroup_inotify_event_source, SD_EVENT_PRIORITY_NORMAL-5);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to set priority of inotify event source: %m");
+
+                        (void) sd_event_source_set_description(m->cgroup_inotify_event_source, "cgroup-inotify");
+
+#else
+                        return log_error_errno(EOPNOTSUPP, "Unified cgroup hierarchy not supported: %m");
+#endif // 0
+                } else if (MANAGER_IS_SYSTEM(m)) {
+
+                        /* On the legacy hierarchy we only get
+                         * notifications via cgroup agents. (Which
+                         * isn't really reliable, since it does not
+                         * generate events when control groups with
+                         * children run empty. */
+
                         r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH);
                         if (r < 0)
                                 log_warning_errno(r, "Failed to install release agent, ignoring: %m");
                         else if (r > 0)
                                 log_debug("Installed release agent.");
-                        else
+                        else if (r == 0)
                                 log_debug("Release agent already installed.");
                 }
 
-                /* 4. Make sure we are in the root cgroup */
-                r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, 0);
+#if 0 /// elogind is not meant to run in systemd init scope
+                /* 4. Make sure we are in the special "init.scope" unit in the root slice. */
+                scope_path = strjoina(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
+                r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
+#else
+                if (streq(SYSTEMD_CGROUP_CONTROLLER, "name=elogind"))
+                        // we are our own cgroup controller
+                        scope_path = strjoina("");
+                else if (streq(m->cgroup_root, "/elogind"))
+                        // root already is our cgroup
+                        scope_path = strjoina(m->cgroup_root);
+                else
+                        // we have to create our own group
+                        scope_path = strjoina(m->cgroup_root, "/elogind");
+                r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
+#endif // 0
+                if (r < 0)
+                        return log_error_errno(r, "Failed to create %s control group: %m", scope_path);
+                log_debug_elogind("Created control group \"%s\"", scope_path);
+
+                /* also, move all other userspace processes remaining
+                 * in the root cgroup into that scope. */
+                r = cg_migrate(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
                 if (r < 0)
-                        return log_error_errno(r, "Failed to create root cgroup hierarchy: %m");
+                        log_warning_errno(r, "Couldn't move remaining userspace processes, ignoring: %m");
 
                 /* 5. And pin it, so that it cannot be unmounted */
                 safe_close(m->pin_cgroupfs_fd);
-
                 m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK);
                 if (m->pin_cgroupfs_fd < 0)
                         return log_error_errno(errno, "Failed to open pin file: %m");
 
-                /* 6.  Always enable hierarchial support if it exists... */
-                cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
+                /* 6.  Always enable hierarchical support if it exists... */
+                if (!unified)
+                        (void) cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
         }
 
         /* 7. Figure out which controllers are supported */
-        m->cgroup_supported = cg_mask_supported();
+        r = cg_mask_supported(&m->cgroup_supported);
+        if (r < 0)
+                return log_error_errno(r, "Failed to determine supported controllers: %m");
+
+        for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++)
+                log_debug("Controller '%s' supported: %s", cgroup_controller_to_string(c), yes_no(m->cgroup_supported & CGROUP_CONTROLLER_TO_MASK(c)));
 
         return 0;
 }
@@ -949,14 +1873,21 @@ void manager_shutdown_cgroup(Manager *m, bool delete) {
         /* We can't really delete the group, since we are in it. But
          * let's trim it. */
         if (delete && m->cgroup_root)
-                cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
+                (void) cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
+
+#if 0 /// elogind does not support the unified hierarchy, yet.
+        m->cgroup_inotify_wd_unit = hashmap_free(m->cgroup_inotify_wd_unit);
+
+        m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
+        m->cgroup_inotify_fd = safe_close(m->cgroup_inotify_fd);
+#endif // 0
 
         m->pin_cgroupfs_fd = safe_close(m->pin_cgroupfs_fd);
 
-        free(m->cgroup_root);
-        m->cgroup_root = NULL;
+        m->cgroup_root = mfree(m->cgroup_root);
 }
 
+#if 0 /// UNNEEDED by elogind
 Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
         char *p;
         Unit *u;
@@ -973,8 +1904,8 @@ Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
                 char *e;
 
                 e = strrchr(p, '/');
-                if (e == p || !e)
-                        return NULL;
+                if (!e || e == p)
+                        return hashmap_get(m->cgroup_unit, SPECIAL_ROOT_SLICE);
 
                 *e = 0;
 
@@ -984,13 +1915,13 @@ Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
         }
 }
 
-Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
+Unit *manager_get_unit_by_pid_cgroup(Manager *m, pid_t pid) {
         _cleanup_free_ char *cgroup = NULL;
         int r;
 
         assert(m);
 
-        if (pid <= 1)
+        if (pid <= 0)
                 return NULL;
 
         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
@@ -1000,27 +1931,235 @@ Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
         return manager_get_unit_by_cgroup(m, cgroup);
 }
 
+Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
+        Unit *u;
+
+        assert(m);
+
+        if (pid <= 0)
+                return NULL;
+
+        if (pid == 1)
+                return hashmap_get(m->units, SPECIAL_INIT_SCOPE);
+
+        u = hashmap_get(m->watch_pids1, PID_TO_PTR(pid));
+        if (u)
+                return u;
+
+        u = hashmap_get(m->watch_pids2, PID_TO_PTR(pid));
+        if (u)
+                return u;
+
+        return manager_get_unit_by_pid_cgroup(m, pid);
+}
+#endif // 0
+
+#if 0 /// elogind must substitute this with its own variant
 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
         Unit *u;
-        int r;
 
         assert(m);
         assert(cgroup);
 
+        log_debug("Got cgroup empty notification for: %s", cgroup);
+
         u = manager_get_unit_by_cgroup(m, cgroup);
-        if (u) {
-                r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, true);
-                if (r > 0) {
-                        if (UNIT_VTABLE(u)->notify_cgroup_empty)
-                                UNIT_VTABLE(u)->notify_cgroup_empty(u);
+        if (!u)
+                return 0;
 
-                        unit_add_to_gc_queue(u);
-                }
+        return unit_notify_cgroup_empty(u);
+}
+#else
+int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
+        Session *s;
+
+        assert(m);
+        assert(cgroup);
+
+        log_debug("Got cgroup empty notification for: %s", cgroup);
+
+        s = hashmap_get(m->sessions, cgroup);
+
+        if (s) {
+                session_finalize(s);
+                session_free(s);
+        } else
+                log_warning("Session not found: %s", cgroup);
+
+        return 0;
+}
+#endif // 0
+
+#if 0 /// UNNEEDED by elogind
+int unit_get_memory_current(Unit *u, uint64_t *ret) {
+        _cleanup_free_ char *v = NULL;
+        int r;
+
+        assert(u);
+        assert(ret);
+
+        if (!u->cgroup_path)
+                return -ENODATA;
+
+        if ((u->cgroup_realized_mask & CGROUP_MASK_MEMORY) == 0)
+                return -ENODATA;
+
+        if (cg_unified() <= 0)
+                r = cg_get_attribute("memory", u->cgroup_path, "memory.usage_in_bytes", &v);
+        else
+                r = cg_get_attribute("memory", u->cgroup_path, "memory.current", &v);
+        if (r == -ENOENT)
+                return -ENODATA;
+        if (r < 0)
+                return r;
+
+        return safe_atou64(v, ret);
+}
+
+int unit_get_tasks_current(Unit *u, uint64_t *ret) {
+        _cleanup_free_ char *v = NULL;
+        int r;
+
+        assert(u);
+        assert(ret);
+
+        if (!u->cgroup_path)
+                return -ENODATA;
+
+        if ((u->cgroup_realized_mask & CGROUP_MASK_PIDS) == 0)
+                return -ENODATA;
+
+        r = cg_get_attribute("pids", u->cgroup_path, "pids.current", &v);
+        if (r == -ENOENT)
+                return -ENODATA;
+        if (r < 0)
+                return r;
+
+        return safe_atou64(v, ret);
+}
+
+static int unit_get_cpu_usage_raw(Unit *u, nsec_t *ret) {
+        _cleanup_free_ char *v = NULL;
+        uint64_t ns;
+        int r;
+
+        assert(u);
+        assert(ret);
+
+        if (!u->cgroup_path)
+                return -ENODATA;
+
+        if (cg_unified() > 0) {
+                const char *keys[] = { "usage_usec", NULL };
+                _cleanup_free_ char *val = NULL;
+                uint64_t us;
+
+                if ((u->cgroup_realized_mask & CGROUP_MASK_CPU) == 0)
+                        return -ENODATA;
+
+                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;
 }
 
+int unit_get_cpu_usage(Unit *u, nsec_t *ret) {
+        nsec_t ns;
+        int r;
+
+        r = unit_get_cpu_usage_raw(u, &ns);
+        if (r < 0)
+                return r;
+
+        if (ns > u->cpu_usage_base)
+                ns -= u->cpu_usage_base;
+        else
+                ns = 0;
+
+        *ret = ns;
+        return 0;
+}
+
+int unit_reset_cpu_usage(Unit *u) {
+        nsec_t ns;
+        int r;
+
+        assert(u);
+
+        r = unit_get_cpu_usage_raw(u, &ns);
+        if (r < 0) {
+                u->cpu_usage_base = 0;
+                return r;
+        }
+
+        u->cpu_usage_base = ns;
+        return 0;
+}
+
+bool unit_cgroup_delegate(Unit *u) {
+        CGroupContext *c;
+
+        assert(u);
+
+        c = unit_get_cgroup_context(u);
+        if (!c)
+                return false;
+
+        return c->delegate;
+}
+
+void unit_invalidate_cgroup(Unit *u, CGroupMask m) {
+        assert(u);
+
+        if (!UNIT_HAS_CGROUP_CONTEXT(u))
+                return;
+
+        if (m == 0)
+                return;
+
+        /* always invalidate compat pairs together */
+        if (m & (CGROUP_MASK_IO | CGROUP_MASK_BLKIO))
+                m |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
+
+        if ((u->cgroup_realized_mask & m) == 0)
+                return;
+
+        u->cgroup_realized_mask &= ~m;
+        unit_add_to_cgroup_queue(u);
+}
+
+void manager_invalidate_startup_units(Manager *m) {
+        Iterator i;
+        Unit *u;
+
+        assert(m);
+
+        SET_FOREACH(u, m->startup_units, i)
+                unit_invalidate_cgroup(u, CGROUP_MASK_CPU|CGROUP_MASK_IO|CGROUP_MASK_BLKIO);
+}
+
 static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] = {
         [CGROUP_AUTO] = "auto",
         [CGROUP_CLOSED] = "closed",
@@ -1028,3 +2167,4 @@ static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] =
 };
 
 DEFINE_STRING_TABLE_LOOKUP(cgroup_device_policy, CGroupDevicePolicy);
+#endif // 0