Merge pull request #15206 from anitazha/systoomd-v0

systemd-oomd

src/basic/cgroup-util.c

@@ -1685,6 +1685,26 @@ int cg_get_attribute_as_uint64(const char *controller, const char *path, const c
         return 0;
 }
 
+int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret) {
+        _cleanup_free_ char *value = NULL;
+        int r;
+
+        assert(ret);
+
+        r = cg_get_attribute(controller, path, attribute, &value);
+        if (r == -ENOENT)
+                return -ENODATA;
+        if (r < 0)
+                return r;
+
+        r = parse_boolean(value);
+        if (r < 0)
+                return r;
+
+        *ret = r;
+        return 0;
+}
+
 int cg_get_keyed_attribute_full(
                 const char *controller,
                 const char *path,
@@ -2161,3 +2181,10 @@ CGroupMask get_cpu_accounting_mask(void) {
 bool cpu_accounting_is_cheap(void) {
         return get_cpu_accounting_mask() == 0;
 }
+
+static const char* const managed_oom_mode_table[_MANAGED_OOM_MODE_MAX] = {
+        [MANAGED_OOM_AUTO] = "auto",
+        [MANAGED_OOM_KILL] = "kill",
+};
+
+DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode, ManagedOOMMode);

src/basic/cgroup-util.h

@@ -208,6 +208,9 @@ static inline int cg_get_keyed_attribute_graceful(
 
 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret);
 
+/* Does a parse_boolean() on the attribute contents and sets ret accordingly */
+int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret);
+
 int cg_set_access(const char *controller, const char *path, uid_t uid, gid_t gid);
 
 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags);
@@ -275,3 +278,13 @@ CGroupController cgroup_controller_from_string(const char *s) _pure_;
 
 bool is_cgroup_fs(const struct statfs *s);
 bool fd_is_cgroup_fs(int fd);
+
+typedef enum ManagedOOMMode {
+        MANAGED_OOM_AUTO,
+        MANAGED_OOM_KILL,
+        _MANAGED_OOM_MODE_MAX,
+        _MANAGED_OOM_MODE_INVALID = -1,
+} ManagedOOMMode;
+
+const char* managed_oom_mode_to_string(ManagedOOMMode m) _const_;
+ManagedOOMMode managed_oom_mode_from_string(const char *s) _pure_;

src/basic/def.h

@@ -63,3 +63,5 @@
                 .un.sun_family = AF_UNIX,                       \
                 .un.sun_path = "\0/org/freedesktop/plymouthd",  \
         }
+
+#define VARLINK_ADDR_PATH_MANAGED_OOM "/run/systemd/io.system.ManagedOOM"

src/basic/linux/loadavg.h

@@ -0,0 +1,48 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _LINUX_SCHED_LOADAVG_H
+#define _LINUX_SCHED_LOADAVG_H
+
+/*
+ * These are the constant used to fake the fixed-point load-average
+ * counting. Some notes:
+ *  - 11 bit fractions expand to 22 bits by the multiplies: this gives
+ *    a load-average precision of 10 bits integer + 11 bits fractional
+ *  - if you want to count load-averages more often, you need more
+ *    precision, or rounding will get you. With 2-second counting freq,
+ *    the EXP_n values would be 1981, 2034 and 2043 if still using only
+ *    11 bit fractions.
+ */
+extern unsigned long avenrun[];                 /* Load averages */
+extern void get_avenrun(unsigned long *loads, unsigned long offset, int shift);
+
+#define FSHIFT          11                      /* nr of bits of precision */
+#define FIXED_1         (1<<FSHIFT)             /* 1.0 as fixed-point */
+#define LOAD_FREQ       (5*HZ+1)                /* 5 sec intervals */
+#define EXP_1           1884                    /* 1/exp(5sec/1min) as fixed-point */
+#define EXP_5           2014                    /* 1/exp(5sec/5min) */
+#define EXP_15          2037                    /* 1/exp(5sec/15min) */
+
+/*
+ * a1 = a0 * e + a * (1 - e)
+ */
+static inline unsigned long
+calc_load(unsigned long load, unsigned long exp, unsigned long active)
+{
+        unsigned long newload;
+
+        newload = load * exp + active * (FIXED_1 - exp);
+        if (active >= load)
+                newload += FIXED_1-1;
+
+        return newload / FIXED_1;
+}
+
+extern unsigned long calc_load_n(unsigned long load, unsigned long exp,
+                                 unsigned long active, unsigned int n);
+
+#define LOAD_INT(x) ((x) >> FSHIFT)
+#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
+
+extern void calc_global_load(unsigned long ticks);
+
+#endif /* _LINUX_SCHED_LOADAVG_H */

src/basic/meson.build

@@ -111,6 +111,7 @@ basic_sources = files('''
         linux/ipv6_route.h
         linux/l2tp.h
         linux/libc-compat.h
+        linux/loadavg.h
         linux/netdevice.h
         linux/netlink.h
         linux/rtnetlink.h

src/basic/parse-util.c

@@ -862,3 +862,45 @@ int parse_oom_score_adjust(const char *s, int *ret) {
         *ret = v;
         return 0;
 }
+
+int store_loadavg_fixed_point(unsigned long i, unsigned long f, loadavg_t *ret) {
+        assert(ret);
+
+        if (i >= (~0UL << FSHIFT))
+                return -ERANGE;
+
+        i = i << FSHIFT;
+        f = DIV_ROUND_UP((f << FSHIFT), 100);
+
+        if (f >= FIXED_1)
+                return -ERANGE;
+
+        *ret = i | f;
+        return 0;
+}
+
+int parse_loadavg_fixed_point(const char *s, loadavg_t *ret) {
+        const char *d, *f_str, *i_str;
+        unsigned long i, f;
+        int r;
+
+        assert(s);
+        assert(ret);
+
+        d = strchr(s, '.');
+        if (!d)
+                return -EINVAL;
+
+        i_str = strndupa(s, d - s);
+        f_str = d + 1;
+
+        r = safe_atolu_full(i_str, 10, &i);
+        if (r < 0)
+                return r;
+
+        r = safe_atolu_full(f_str, 10, &f);
+        if (r < 0)
+                return r;
+
+        return store_loadavg_fixed_point(i, f, ret);
+}

src/basic/parse-util.h

@@ -3,12 +3,15 @@
 
 #include <inttypes.h>
 #include <limits.h>
+#include <linux/loadavg.h>
 #include <stddef.h>
 #include <stdint.h>
 #include <sys/types.h>
 
 #include "macro.h"
 
+typedef unsigned long loadavg_t;
+
 int parse_boolean(const char *v) _pure_;
 int parse_dev(const char *s, dev_t *ret);
 int parse_pid(const char *s, pid_t* ret_pid);
@@ -88,18 +91,18 @@ static inline int safe_atoux64(const char *s, uint64_t *ret) {
 }
 
 #if LONG_MAX == INT_MAX
-static inline int safe_atolu(const char *s, unsigned long *ret_u) {
+static inline int safe_atolu_full(const char *s, unsigned base, long unsigned *ret_u) {
         assert_cc(sizeof(unsigned long) == sizeof(unsigned));
-        return safe_atou(s, (unsigned*) ret_u);
+        return safe_atou_full(s, base, (unsigned*) ret_u);
 }
 static inline int safe_atoli(const char *s, long int *ret_u) {
         assert_cc(sizeof(long int) == sizeof(int));
         return safe_atoi(s, (int*) ret_u);
 }
 #else
-static inline int safe_atolu(const char *s, unsigned long *ret_u) {
+static inline int safe_atolu_full(const char *s, unsigned base, unsigned long *ret_u) {
         assert_cc(sizeof(unsigned long) == sizeof(unsigned long long));
-        return safe_atollu(s, (unsigned long long*) ret_u);
+        return safe_atollu_full(s, base, (unsigned long long*) ret_u);
 }
 static inline int safe_atoli(const char *s, long int *ret_u) {
         assert_cc(sizeof(long int) == sizeof(long long int));
@@ -107,6 +110,10 @@ static inline int safe_atoli(const char *s, long int *ret_u) {
 }
 #endif
 
+static inline int safe_atolu(const char *s, unsigned long *ret_u) {
+        return safe_atolu_full(s, 0, ret_u);
+}
+
 #if SIZE_MAX == UINT_MAX
 static inline int safe_atozu(const char *s, size_t *ret_u) {
         assert_cc(sizeof(size_t) == sizeof(unsigned));
@@ -137,3 +144,8 @@ int parse_ip_port_range(const char *s, uint16_t *low, uint16_t *high);
 int parse_ip_prefix_length(const char *s, int *ret);
 
 int parse_oom_score_adjust(const char *s, int *ret);
+
+/* Given a Linux load average (e.g. decimal number 34.89 where 34 is passed as i and 89 is passed as f), convert it
+ * to a loadavg_t. */
+int store_loadavg_fixed_point(unsigned long i, unsigned long f, loadavg_t *ret);
+int parse_loadavg_fixed_point(const char *s, loadavg_t *ret);

src/core/cgroup.c

@@ -128,6 +128,9 @@ void cgroup_context_init(CGroupContext *c) {
                 .startup_blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID,
 
                 .tasks_max = TASKS_MAX_UNSET,
+
+                .moom_swap = MANAGED_OOM_AUTO,
+                .moom_mem_pressure = MANAGED_OOM_AUTO,
         };
 }
 
@@ -411,7 +414,10 @@ void cgroup_context_dump(Unit *u, FILE* f, const char *prefix) {
                 "%sTasksMax: %" PRIu64 "\n"
                 "%sDevicePolicy: %s\n"
                 "%sDisableControllers: %s\n"
-                "%sDelegate: %s\n",
+                "%sDelegate: %s\n"
+                "%sManagedOOMSwap: %s\n"
+                "%sManagedOOMMemoryPressure: %s\n"
+                "%sManagedOOMMemoryPressureLimitPercent: %d%%\n",
                 prefix, yes_no(c->cpu_accounting),
                 prefix, yes_no(c->io_accounting),
                 prefix, yes_no(c->blockio_accounting),
@@ -441,7 +447,10 @@ void cgroup_context_dump(Unit *u, FILE* f, const char *prefix) {
                 prefix, tasks_max_resolve(&c->tasks_max),
                 prefix, cgroup_device_policy_to_string(c->device_policy),
                 prefix, strempty(disable_controllers_str),
-                prefix, yes_no(c->delegate));
+                prefix, yes_no(c->delegate),
+                prefix, managed_oom_mode_to_string(c->moom_swap),
+                prefix, managed_oom_mode_to_string(c->moom_mem_pressure),
+                prefix, c->moom_mem_pressure_limit);
 
         if (c->delegate) {
                 _cleanup_free_ char *t = NULL;
@@ -2672,6 +2681,47 @@ static void unit_remove_from_cgroup_empty_queue(Unit *u) {
         u->in_cgroup_empty_queue = false;
 }
 
+int unit_check_oomd_kill(Unit *u) {
+        _cleanup_free_ char *value = NULL;
+        bool increased;
+        uint64_t n = 0;
+        int r;
+
+        if (!u->cgroup_path)
+                return 0;
+
+        r = cg_all_unified();
+        if (r < 0)
+                return log_unit_debug_errno(u, r, "Couldn't determine whether we are in all unified mode: %m");
+        else if (r == 0)
+                return 0;
+
+        r = cg_get_xattr_malloc(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "user.systemd_oomd_kill", &value);
+        if (r < 0 && r != -ENODATA)
+                return r;
+
+        if (!isempty(value)) {
+                 r = safe_atou64(value, &n);
+                 if (r < 0)
+                         return r;
+        }
+
+        increased = n > u->managed_oom_kill_last;
+        u->managed_oom_kill_last = n;
+
+        if (!increased)
+                return 0;
+
+        if (n > 0)
+                log_struct(LOG_NOTICE,
+                           "MESSAGE_ID=" SD_MESSAGE_UNIT_OOMD_KILL_STR,
+                           LOG_UNIT_ID(u),
+                           LOG_UNIT_INVOCATION_ID(u),
+                           LOG_UNIT_MESSAGE(u, "systemd-oomd killed %"PRIu64" process(es) in this unit.", n));
+
+        return 1;
+}
+
 int unit_check_oom(Unit *u) {
         _cleanup_free_ char *oom_kill = NULL;
         bool increased;

src/core/cgroup.h

@@ -159,6 +159,11 @@ struct CGroupContext {
 
         /* Common */
         TasksMax tasks_max;
+
+        /* Settings for systemd-oomd */
+        ManagedOOMMode moom_swap;
+        ManagedOOMMode moom_mem_pressure;
+        int moom_mem_pressure_limit;
 };
 
 /* Used when querying IP accounting data */
@@ -224,6 +229,7 @@ int unit_watch_cgroup(Unit *u);
 int unit_watch_cgroup_memory(Unit *u);
 
 void unit_add_to_cgroup_empty_queue(Unit *u);
+int unit_check_oomd_kill(Unit *u);
 int unit_check_oom(Unit *u);
 
 int unit_attach_pids_to_cgroup(Unit *u, Set *pids, const char *suffix_path);

src/core/core-varlink.c

@@ -2,6 +2,7 @@
 
 #include "core-varlink.h"
 #include "mkdir.h"
+#include "strv.h"
 #include "user-util.h"
 #include "varlink.h"
 
@@ -15,6 +16,11 @@ typedef struct LookupParameters {
         const char *service;
 } LookupParameters;
 
+static const char* const managed_oom_mode_properties[] = {
+        "ManagedOOMSwap",
+        "ManagedOOMMemoryPressure",
+};
+
 static int build_user_json(const char *user_name, uid_t uid, JsonVariant **ret) {
         assert(user_name);
         assert(uid_is_valid(uid));
@@ -45,6 +51,150 @@ static bool user_match_lookup_parameters(LookupParameters *p, const char *name,
         return true;
 }
 
+static int build_managed_oom_json_array_element(Unit *u, const char *property, JsonVariant **ret_v) {
+        _cleanup_(json_variant_unrefp) JsonVariant *v = NULL;
+        CGroupContext *c;
+        const char *mode;
+        int r;
+
+        assert(u);
+        assert(property);
+        assert(ret_v);
+
+        if (!UNIT_VTABLE(u)->can_set_managed_oom)
+                return -EOPNOTSUPP;
+
+        c = unit_get_cgroup_context(u);
+        if (!c)
+                return -EINVAL;
+
+        if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(u)))
+                /* systemd-oomd should always treat inactive units as though they didn't enable any action since they
+                 * should not have a valid cgroup */
+                mode = managed_oom_mode_to_string(MANAGED_OOM_AUTO);
+        else if (streq(property, "ManagedOOMSwap"))
+                mode = managed_oom_mode_to_string(c->moom_swap);
+        else if (streq(property, "ManagedOOMMemoryPressure"))
+                mode = managed_oom_mode_to_string(c->moom_mem_pressure);
+        else
+                return -EINVAL;
+
+        r = json_build(&v, JSON_BUILD_OBJECT(
+                                JSON_BUILD_PAIR("mode", JSON_BUILD_STRING(mode)),
+                                JSON_BUILD_PAIR("path", JSON_BUILD_STRING(u->cgroup_path)),
+                                JSON_BUILD_PAIR("property", JSON_BUILD_STRING(property)),
+                                JSON_BUILD_PAIR("limit", JSON_BUILD_UNSIGNED(c->moom_mem_pressure_limit))));
+
+        *ret_v = TAKE_PTR(v);
+        return r;
+}
+
+int manager_varlink_send_managed_oom_update(Unit *u) {
+        _cleanup_(json_variant_unrefp) JsonVariant *arr = NULL, *v = NULL;
+        CGroupContext *c;
+        int r;
+
+        assert(u);
+
+        if (!UNIT_VTABLE(u)->can_set_managed_oom || !u->manager || !u->manager->managed_oom_varlink_request || !u->cgroup_path)
+                return 0;
+
+        c = unit_get_cgroup_context(u);
+        if (!c)
+                return 0;
+
+        r = json_build(&arr, JSON_BUILD_EMPTY_ARRAY);
+        if (r < 0)
+                return r;
+
+        for (size_t i = 0; i < ELEMENTSOF(managed_oom_mode_properties); i++) {
+                _cleanup_(json_variant_unrefp) JsonVariant *e = NULL;
+
+                r = build_managed_oom_json_array_element(u, managed_oom_mode_properties[i], &e);
+                if (r < 0)
+                        return r;
+
+                r = json_variant_append_array(&arr, e);
+                if (r < 0)
+                        return r;
+        }
+
+        r = json_build(&v, JSON_BUILD_OBJECT(JSON_BUILD_PAIR("cgroups", JSON_BUILD_VARIANT(arr))));
+        if (r < 0)
+                return r;
+
+        return varlink_notify(u->manager->managed_oom_varlink_request, v);
+}
+
+static int vl_method_subscribe_managed_oom_cgroups(
+                Varlink *link,
+                JsonVariant *parameters,
+                VarlinkMethodFlags flags,
+                void *userdata) {
+        static const UnitType supported_unit_types[] = { UNIT_SLICE, UNIT_SERVICE, UNIT_SCOPE };
+        _cleanup_(json_variant_unrefp) JsonVariant *v = NULL, *arr = NULL;
+        Manager *m = userdata;
+        int r;
+
+        assert(link);
+        assert(m);
+
+        if (json_variant_elements(parameters) > 0)
+                return varlink_error_invalid_parameter(link, parameters);
+
+        /* We only take one subscriber for this method so return an error if there's already an existing one.
+         * This shouldn't happen since systemd-oomd is the only client of this method. */
+        if (FLAGS_SET(flags, VARLINK_METHOD_MORE) && m->managed_oom_varlink_request)
+                return varlink_error(m->managed_oom_varlink_request, VARLINK_ERROR_SUBSCRIPTION_TAKEN, NULL);
+
+        r = json_build(&arr, JSON_BUILD_EMPTY_ARRAY);
+        if (r < 0)
+                return r;
+
+        for (size_t i = 0; i < ELEMENTSOF(supported_unit_types); i++) {
+                Unit *u;
+
+                LIST_FOREACH(units_by_type, u, m->units_by_type[supported_unit_types[i]]) {
+                        CGroupContext *c;
+
+                        if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(u)))
+                                continue;
+
+                        c = unit_get_cgroup_context(u);
+                        if (!c)
+                                continue;
+
+                        for (size_t j = 0; j < ELEMENTSOF(managed_oom_mode_properties); j++) {
+                                _cleanup_(json_variant_unrefp) JsonVariant *e = NULL;
+
+                                /* For the initial varlink call we only care about units that enabled (i.e. mode is not
+                                 * set to "auto") oomd properties. */
+                                if (!(streq(managed_oom_mode_properties[j], "ManagedOOMSwap") && c->moom_swap == MANAGED_OOM_KILL) &&
+                                    !(streq(managed_oom_mode_properties[j], "ManagedOOMMemoryPressure") && c->moom_mem_pressure == MANAGED_OOM_KILL))
+                                        continue;
+
+                                r = build_managed_oom_json_array_element(u, managed_oom_mode_properties[j], &e);
+                                if (r < 0)
+                                        return r;
+
+                                r = json_variant_append_array(&arr, e);
+                                if (r < 0)
+                                        return r;
+                        }
+                }
+        }
+
+        r = json_build(&v, JSON_BUILD_OBJECT(JSON_BUILD_PAIR("cgroups", JSON_BUILD_VARIANT(arr))));
+        if (r < 0)
+                return r;
+
+        if (!FLAGS_SET(flags, VARLINK_METHOD_MORE))
+                return varlink_reply(link, v);
+
+        m->managed_oom_varlink_request = varlink_ref(link);
+        return varlink_notify(m->managed_oom_varlink_request, v);
+}
+
 static int vl_method_get_user_record(Varlink *link, JsonVariant *parameters, VarlinkMethodFlags flags, void *userdata) {
 
         static const JsonDispatch dispatch_table[] = {
@@ -262,6 +412,17 @@ static int vl_method_get_memberships(Varlink *link, JsonVariant *parameters, Var
         return varlink_error(link, "io.systemd.UserDatabase.NoRecordFound", NULL);
 }
 
+static void vl_disconnect(VarlinkServer *s, Varlink *link, void *userdata) {
+        Manager *m = userdata;
+
+        assert(m);
+        assert(s);
+        assert(link);
+
+        if (link == m->managed_oom_varlink_request)
+                m->managed_oom_varlink_request = varlink_unref(link);
+}
+
 int manager_varlink_init(Manager *m) {
         _cleanup_(varlink_server_unrefp) VarlinkServer *s = NULL;
         int r;
@@ -284,16 +445,25 @@ int manager_varlink_init(Manager *m) {
                         s,
                         "io.systemd.UserDatabase.GetUserRecord",  vl_method_get_user_record,
                         "io.systemd.UserDatabase.GetGroupRecord", vl_method_get_group_record,
-                        "io.systemd.UserDatabase.GetMemberships", vl_method_get_memberships);
+                        "io.systemd.UserDatabase.GetMemberships", vl_method_get_memberships,
+                        "io.systemd.ManagedOOM.SubscribeManagedOOMCGroups",  vl_method_subscribe_managed_oom_cgroups);
         if (r < 0)
                 return log_error_errno(r, "Failed to register varlink methods: %m");
 
+        r = varlink_server_bind_disconnect(s, vl_disconnect);
+        if (r < 0)
+                return log_error_errno(r, "Failed to register varlink disconnect handler: %m");
+
         if (!MANAGER_IS_TEST_RUN(m)) {
                 (void) mkdir_p_label("/run/systemd/userdb", 0755);
 
                 r = varlink_server_listen_address(s, "/run/systemd/userdb/io.systemd.DynamicUser", 0666);
                 if (r < 0)
                         return log_error_errno(r, "Failed to bind to varlink socket: %m");
+
+                r = varlink_server_listen_address(s, VARLINK_ADDR_PATH_MANAGED_OOM, 0666);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to bind to varlink socket: %m");
         }
 
         r = varlink_server_attach_event(s, m->event, SD_EVENT_PRIORITY_NORMAL);
@@ -307,5 +477,11 @@ int manager_varlink_init(Manager *m) {
 void manager_varlink_done(Manager *m) {
         assert(m);
 
+        /* Send the final message if we still have a subscribe request open. */
+        if (m->managed_oom_varlink_request) {
+                (void) varlink_error(m->managed_oom_varlink_request, VARLINK_ERROR_DISCONNECTED, NULL);
+                m->managed_oom_varlink_request = varlink_unref(m->managed_oom_varlink_request);
+        }
+
         m->varlink_server = varlink_server_unref(m->varlink_server);
 }

src/core/core-varlink.h

@@ -5,3 +5,8 @@
 
 int manager_varlink_init(Manager *m);
 void manager_varlink_done(Manager *m);
+
+/* The manager is expected to send an update to systemd-oomd if one of the following occurs:
+ * - The value of ManagedOOM*= properties change
+ * - A unit with ManagedOOM*= properties changes unit active state */
+int manager_varlink_send_managed_oom_update(Unit *u);

src/core/dbus-cgroup.c

@@ -8,6 +8,7 @@
 #include "bus-get-properties.h"
 #include "cgroup-util.h"
 #include "cgroup.h"
+#include "core-varlink.h"
 #include "dbus-cgroup.h"
 #include "dbus-util.h"
 #include "errno-util.h"
@@ -19,6 +20,7 @@
 BUS_DEFINE_PROPERTY_GET(bus_property_get_tasks_max, "t", TasksMax, tasks_max_resolve);
 
 static BUS_DEFINE_PROPERTY_GET_ENUM(property_get_cgroup_device_policy, cgroup_device_policy, CGroupDevicePolicy);
+static BUS_DEFINE_PROPERTY_GET_ENUM(property_get_managed_oom_mode, managed_oom_mode, ManagedOOMMode);
 
 static int property_get_cgroup_mask(
                 sd_bus *bus,
@@ -391,6 +393,9 @@ const sd_bus_vtable bus_cgroup_vtable[] = {
         SD_BUS_PROPERTY("IPIngressFilterPath", "as", NULL, offsetof(CGroupContext, ip_filters_ingress), 0),
         SD_BUS_PROPERTY("IPEgressFilterPath", "as", NULL, offsetof(CGroupContext, ip_filters_egress), 0),
         SD_BUS_PROPERTY("DisableControllers", "as", property_get_cgroup_mask, offsetof(CGroupContext, disable_controllers), 0),
+        SD_BUS_PROPERTY("ManagedOOMSwap", "s", property_get_managed_oom_mode, offsetof(CGroupContext, moom_swap), 0),
+        SD_BUS_PROPERTY("ManagedOOMMemoryPressure", "s", property_get_managed_oom_mode, offsetof(CGroupContext, moom_mem_pressure), 0),
+        SD_BUS_PROPERTY("ManagedOOMMemoryPressureLimitPercent", "s", bus_property_get_percent, offsetof(CGroupContext, moom_mem_pressure_limit), 0),
         SD_BUS_VTABLE_END
 };
 
@@ -1667,6 +1672,45 @@ int bus_cgroup_set_property(
                 return 1;
         }
 
+        if (STR_IN_SET(name, "ManagedOOMSwap", "ManagedOOMMemoryPressure")) {
+                ManagedOOMMode *cgroup_mode = streq(name, "ManagedOOMSwap") ? &c->moom_swap : &c->moom_mem_pressure;
+                ManagedOOMMode m;
+                const char *mode;
+
+                if (!UNIT_VTABLE(u)->can_set_managed_oom)
+                        return sd_bus_error_setf(error, SD_BUS_ERROR_INVALID_ARGS, "Cannot set %s for this unit type", name);
+
+                r = sd_bus_message_read(message, "s", &mode);
+                if (r < 0)
+                        return r;
+
+                m = managed_oom_mode_from_string(mode);
+                if (m < 0)
+                        return -EINVAL;
+
+                if (!UNIT_WRITE_FLAGS_NOOP(flags)) {
+                        *cgroup_mode = m;
+                        unit_write_settingf(u, flags, name, "%s=%s", name, mode);
+                }
+
+                (void) manager_varlink_send_managed_oom_update(u);
+                return 1;
+        }
+
+        if (streq(name, "ManagedOOMMemoryPressureLimitPercent")) {
+                if (!UNIT_VTABLE(u)->can_set_managed_oom)
+                        return sd_bus_error_setf(error, SD_BUS_ERROR_INVALID_ARGS, "Cannot set %s for this unit type", name);
+
+                r = bus_set_transient_percent(u, name, &c->moom_mem_pressure_limit, message, flags, error);
+                if (r < 0)
+                        return r;
+
+                if (c->moom_mem_pressure == MANAGED_OOM_KILL)
+                        (void) manager_varlink_send_managed_oom_update(u);
+
+                return 1;
+        }
+
         if (streq(name, "DisableControllers") || (u->transient && u->load_state == UNIT_STUB))
                 return bus_cgroup_set_transient_property(u, c, name, message, flags, error);
 

src/core/dbus-util.c

@@ -91,6 +91,35 @@ int bus_set_transient_bool(
         return 1;
 }
 
+int bus_set_transient_percent(
+                Unit *u,
+                const char *name,
+                int *p,
+                sd_bus_message *message,
+                UnitWriteFlags flags,
+                sd_bus_error *error) {
+
+        const char *v;
+        int r;
+
+        assert(p);
+
+        r = sd_bus_message_read(message, "s", &v);
+        if (r < 0)
+                return r;
+
+        r = parse_percent(v);
+        if (r < 0)
+                return r;
+
+        if (!UNIT_WRITE_FLAGS_NOOP(flags)) {
+                *p = r;
+                unit_write_settingf(u, flags, name, "%s=%d%%", name, r);
+        }
+
+        return 1;
+}
+
 int bus_set_transient_usec_internal(
                 Unit *u,
                 const char *name,

src/core/dbus-util.h

@@ -240,6 +240,7 @@ int bus_set_transient_user_relaxed(Unit *u, const char *name, char **p, sd_bus_m
 int bus_set_transient_path(Unit *u, const char *name, char **p, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error);
 int bus_set_transient_string(Unit *u, const char *name, char **p, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error);
 int bus_set_transient_bool(Unit *u, const char *name, bool *p, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error);
+int bus_set_transient_percent(Unit *u, const char *name, int *p, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error);
 int bus_set_transient_usec_internal(Unit *u, const char *name, usec_t *p, bool fix_0, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error);
 static inline int bus_set_transient_usec(Unit *u, const char *name, usec_t *p, sd_bus_message *message, UnitWriteFlags flags, sd_bus_error *error) {
         return bus_set_transient_usec_internal(u, name, p, false, message, flags, error);

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

@@ -224,6 +224,9 @@ $1.IPAddressAllow,               config_parse_ip_address_access,     0,
 $1.IPAddressDeny,                config_parse_ip_address_access,     0,                             offsetof($1, cgroup_context.ip_address_deny)
 $1.IPIngressFilterPath,          config_parse_ip_filter_bpf_progs,   0,                             offsetof($1, cgroup_context.ip_filters_ingress)
 $1.IPEgressFilterPath,           config_parse_ip_filter_bpf_progs,   0,                             offsetof($1, cgroup_context.ip_filters_egress)
+$1.ManagedOOMSwap,               config_parse_managed_oom_mode,      0,                             offsetof($1, cgroup_context.moom_swap)
+$1.ManagedOOMMemoryPressure,     config_parse_managed_oom_mode,      0,                             offsetof($1, cgroup_context.moom_mem_pressure)
+$1.ManagedOOMMemoryPressureLimitPercent,config_parse_managed_oom_mem_pressure_limit,0,              offsetof($1, cgroup_context)
 $1.NetClass,                     config_parse_warn_compat,           DISABLED_LEGACY,               0'
 )m4_dnl
 Unit.Description,                config_parse_unit_string_printf,    0,                             offsetof(Unit, description)

src/core/load-fragment.c

@@ -26,6 +26,7 @@
 #include "capability-util.h"
 #include "cgroup-setup.h"
 #include "conf-parser.h"
+#include "core-varlink.h"
 #include "cpu-set-util.h"
 #include "env-util.h"
 #include "errno-list.h"
@@ -3812,6 +3813,86 @@ int config_parse_delegate(
         return 0;
 }
 
+int config_parse_managed_oom_mode(
+                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) {
+        Unit *u = userdata;
+        ManagedOOMMode *mode = data, m;
+        UnitType t;
+
+        t = unit_name_to_type(unit);
+        assert(t != _UNIT_TYPE_INVALID);
+
+        if (!unit_vtable[t]->can_set_managed_oom)
+                return log_syntax(unit, LOG_WARNING, filename, line, 0, "%s= is not supported for this unit type, ignoring.", lvalue);
+
+        if (isempty(rvalue)) {
+                *mode = MANAGED_OOM_AUTO;
+                goto finish;
+        }
+
+        m = managed_oom_mode_from_string(rvalue);
+        if (m < 0) {
+                log_syntax(unit, LOG_WARNING, filename, line, 0, "Invalid syntax, ignoring: %s", rvalue);
+                return 0;
+        }
+        *mode = m;
+
+finish:
+        (void) manager_varlink_send_managed_oom_update(u);
+        return 0;
+}
+
+int config_parse_managed_oom_mem_pressure_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) {
+        Unit *u = userdata;
+        CGroupContext *c = data;
+        UnitType t;
+        int r;
+
+        t = unit_name_to_type(unit);
+        assert(t != _UNIT_TYPE_INVALID);
+
+        if (!unit_vtable[t]->can_set_managed_oom)
+                return log_syntax(unit, LOG_WARNING, filename, line, 0, "%s= is not supported for this unit type, ignoring.", lvalue);
+
+        if (isempty(rvalue)) {
+                c->moom_mem_pressure_limit = 0;
+                goto finish;
+        }
+
+        r = parse_percent(rvalue);
+        if (r < 0) {
+                log_syntax(unit, LOG_WARNING, filename, line, r, "Failed to parse limit percent value, ignoring: %s", rvalue);
+                return 0;
+        }
+
+        c->moom_mem_pressure_limit = r;
+
+finish:
+        /* Only update the limit if memory pressure detection is enabled because the information is irrelevant otherwise */
+        if (c->moom_mem_pressure == MANAGED_OOM_KILL)
+                (void) manager_varlink_send_managed_oom_update(u);
+        return 0;
+}
+
 int config_parse_device_allow(
                 const char *unit,
                 const char *filename,

src/core/load-fragment.h

@@ -76,6 +76,8 @@ CONFIG_PARSER_PROTOTYPE(config_parse_cpu_shares);
 CONFIG_PARSER_PROTOTYPE(config_parse_memory_limit);
 CONFIG_PARSER_PROTOTYPE(config_parse_tasks_max);
 CONFIG_PARSER_PROTOTYPE(config_parse_delegate);
+CONFIG_PARSER_PROTOTYPE(config_parse_managed_oom_mode);
+CONFIG_PARSER_PROTOTYPE(config_parse_managed_oom_mem_pressure_limit);
 CONFIG_PARSER_PROTOTYPE(config_parse_device_policy);
 CONFIG_PARSER_PROTOTYPE(config_parse_device_allow);
 CONFIG_PARSER_PROTOTYPE(config_parse_io_device_latency);

src/core/manager.c

@@ -2578,6 +2578,11 @@ static int manager_dispatch_sigchld(sd_event_source *source, void *userdata) {
                          * We only do this for the cgroup the PID belonged to. */
                         (void) unit_check_oom(u1);
 
+                        /* This only logs for now. In the future when the interface for kills/notifications
+                         * is more stable we can extend service results table similar to how kernel oom kills
+                         * are managed. */
+                        (void) unit_check_oomd_kill(u1);
+
                         manager_invoke_sigchld_event(m, u1, &si);
                 }
                 if (u2)

src/core/manager.h

@@ -434,6 +434,8 @@ struct Manager {
         bool honor_device_enumeration;
 
         VarlinkServer *varlink_server;
+        /* Only systemd-oomd should be using this to subscribe to changes in ManagedOOM settings */
+        Varlink *managed_oom_varlink_request;
 };
 
 static inline usec_t manager_default_timeout_abort_usec(Manager *m) {

src/core/scope.c

@@ -621,6 +621,7 @@ const UnitVTable scope_vtable = {
         .can_delegate = true,
         .can_fail = true,
         .once_only = true,
+        .can_set_managed_oom = true,
 
         .init = scope_init,
         .load = scope_load,

src/core/service.c

@@ -4533,6 +4533,7 @@ const UnitVTable service_vtable = {
         .can_transient = true,
         .can_delegate = true,
         .can_fail = true,
+        .can_set_managed_oom = true,
 
         .init = service_init,
         .done = service_done,

src/core/slice.c

@@ -435,6 +435,7 @@ const UnitVTable slice_vtable = {
         .private_section = "Slice",
 
         .can_transient = true,
+        .can_set_managed_oom = true,
 
         .init = slice_init,
         .load = slice_load,

src/core/unit.c

@@ -15,6 +15,7 @@
 #include "bus-util.h"
 #include "cgroup-setup.h"
 #include "cgroup-util.h"
+#include "core-varlink.h"
 #include "dbus-unit.h"
 #include "dbus.h"
 #include "dropin.h"
@@ -1573,6 +1574,31 @@ static int unit_add_mount_dependencies(Unit *u) {
         return 0;
 }
 
+static int unit_add_oomd_dependencies(Unit *u) {
+        CGroupContext *c;
+        bool wants_oomd;
+        int r;
+
+        assert(u);
+
+        if (!u->default_dependencies)
+                return 0;
+
+        c = unit_get_cgroup_context(u);
+        if (!c)
+                return 0;
+
+        wants_oomd = (c->moom_swap == MANAGED_OOM_KILL || c->moom_mem_pressure == MANAGED_OOM_KILL);
+        if (!wants_oomd)
+                return 0;
+
+        r = unit_add_two_dependencies_by_name(u, UNIT_AFTER, UNIT_WANTS, "systemd-oomd.service", true, UNIT_DEPENDENCY_FILE);
+        if (r < 0)
+                return r;
+
+        return 0;
+}
+
 static int unit_add_startup_units(Unit *u) {
         CGroupContext *c;
 
@@ -1633,6 +1659,10 @@ int unit_load(Unit *u) {
                 if (r < 0)
                         goto fail;
 
+                r = unit_add_oomd_dependencies(u);
+                if (r < 0)
+                        goto fail;
+
                 r = unit_add_startup_units(u);
                 if (r < 0)
                         goto fail;
@@ -2592,6 +2622,18 @@ void unit_notify(Unit *u, UnitActiveState os, UnitActiveState ns, UnitNotifyFlag
          * the bus queue, so that any job change signal queued will force out the unit change signal first. */
         unit_add_to_dbus_queue(u);
 
+        /* Update systemd-oomd on the property/state change */
+        if (os != ns) {
+                /* Always send an update if the unit is going into an inactive state so systemd-oomd knows to stop
+                 * monitoring.
+                 * Also send an update whenever the unit goes active; this is to handle a case where an override file
+                 * sets one of the ManagedOOM*= properties to "kill", then later removes it. systemd-oomd needs to
+                 * know to stop monitoring when the unit changes from "kill" -> "auto" on daemon-reload, but we don't
+                 * have the information on the property. Thus, indiscriminately send an update. */
+                if (UNIT_IS_INACTIVE_OR_FAILED(ns) || ns == UNIT_ACTIVE)
+                        (void) manager_varlink_send_managed_oom_update(u);
+        }
+
         /* Update timestamps for state changes */
         if (!MANAGER_IS_RELOADING(m)) {
                 dual_timestamp_get(&u->state_change_timestamp);
@@ -3558,6 +3600,9 @@ int unit_serialize(Unit *u, FILE *f, FDSet *fds, bool serialize_jobs) {
         if (u->cpu_usage_last != NSEC_INFINITY)
                 (void) serialize_item_format(f, "cpu-usage-last", "%" PRIu64, u->cpu_usage_last);
 
+        if (u->managed_oom_kill_last > 0)
+                (void) serialize_item_format(f, "managed-oom-kill-last", "%" PRIu64, u->managed_oom_kill_last);
+
         if (u->oom_kill_last > 0)
                 (void) serialize_item_format(f, "oom-kill-last", "%" PRIu64, u->oom_kill_last);
 
@@ -3803,6 +3848,14 @@ int unit_deserialize(Unit *u, FILE *f, FDSet *fds) {
 
                         continue;
 
+                } else if (streq(l, "managed-oom-kill-last")) {
+
+                        r = safe_atou64(v, &u->managed_oom_kill_last);
+                        if (r < 0)
+                                log_unit_debug(u, "Failed to read managed OOM kill last %s, ignoring.", v);
+
+                        continue;
+
                 } else if (streq(l, "oom-kill-last")) {
 
                         r = safe_atou64(v, &u->oom_kill_last);

src/core/unit.h

@@ -260,7 +260,10 @@ typedef struct Unit {
         nsec_t cpu_usage_base;
         nsec_t cpu_usage_last; /* the most recently read value */
 
-        /* The  current counter of the oom_kill field in the memory.events cgroup attribute */
+        /* The current counter of processes sent SIGKILL by systemd-oomd */
+        uint64_t managed_oom_kill_last;
+
+        /* The current counter of the oom_kill field in the memory.events cgroup attribute */
         uint64_t oom_kill_last;
 
         /* Where the io.stat data was at the time the unit was started */
@@ -625,6 +628,9 @@ typedef struct UnitVTable {
 
         /* True if queued jobs of this type should be GC'ed if no other job needs them anymore */
         bool gc_jobs:1;
+
+        /* True if systemd-oomd can monitor and act on this unit's recursive children's cgroup(s)  */
+        bool can_set_managed_oom:1;
 } UnitVTable;
 
 extern const UnitVTable * const unit_vtable[_UNIT_TYPE_MAX];

src/oom/meson.build

@@ -0,0 +1,34 @@
+# SPDX-License-Identifier: LGPL-2.1+
+
+systemd_oomd_sources = files('''
+        oomd-manager-bus.c
+        oomd-manager-bus.h
+        oomd-manager.c
+        oomd-manager.h
+        oomd-util.c
+        oomd-util.h
+        oomd.c
+'''.split())
+
+oomctl_sources = files('''
+        oomctl.c
+'''.split())
+
+if conf.get('ENABLE_OOMD') == 1
+        tests += [
+                [['src/oom/test-oomd-util.c',
+                  'src/oom/oomd-util.c',
+                  'src/oom/oomd-util.h'],
+                 [],
+                 []]
+        ]
+
+        install_data('org.freedesktop.oom1.conf',
+                     install_dir : dbuspolicydir)
+
+        install_data('org.freedesktop.oom1.service',
+                     install_dir : dbussystemservicedir)
+
+        install_data('oomd.conf',
+                     install_dir : pkgsysconfdir)
+endif

src/oom/oomctl.c

@@ -0,0 +1,138 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <getopt.h>
+#include <unistd.h>
+
+#include "bus-error.h"
+#include "copy.h"
+#include "main-func.h"
+#include "pretty-print.h"
+#include "terminal-util.h"
+#include "verbs.h"
+
+static PagerFlags arg_pager_flags = 0;
+
+static int help(int argc, char *argv[], void *userdata) {
+        _cleanup_free_ char *link = NULL;
+        int r;
+
+        (void) pager_open(arg_pager_flags);
+
+        r = terminal_urlify_man("oomctl", "1", &link);
+        if (r < 0)
+                return log_oom();
+
+        printf("%1$s [OPTIONS...] COMMAND ...\n\n"
+               "%2$sManage or inspect the userspace OOM killer.%3$s\n"
+               "\n%4$sCommands:%5$s\n"
+               "  dump                        Output the current state of systemd-oomd\n"
+               "\n%4$sOptions:%5$s\n"
+               "  -h --help                   Show this help\n"
+               "     --version                Show package version\n"
+               "     --no-pager               Do not pipe output into a pager\n"
+               "\nSee the %6$s for details.\n"
+               , program_invocation_short_name
+               , ansi_highlight(), ansi_normal()
+               , ansi_underline(), ansi_normal()
+               , link
+        );
+
+        return 0;
+}
+
+static int dump_state(int argc, char *argv[], void *userdata) {
+        _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
+        _cleanup_(sd_bus_message_unrefp) sd_bus_message *reply = NULL;
+        _cleanup_(sd_bus_flush_close_unrefp) sd_bus *bus = NULL;
+        int fd = -1;
+        int r;
+
+        r = sd_bus_open_system(&bus);
+        if (r < 0)
+                return log_error_errno(r, "Failed to connect system bus: %m");
+
+        (void) pager_open(arg_pager_flags);
+
+        r = sd_bus_call_method(
+                        bus,
+                        "org.freedesktop.oom1",
+                        "/org/freedesktop/oom1",
+                        "org.freedesktop.oom1.Manager",
+                        "DumpByFileDescriptor",
+                        &error,
+                        &reply,
+                        NULL);
+        if (r < 0)
+                return log_error_errno(r, "Failed to dump context: %s", bus_error_message(&error, r));
+
+        r = sd_bus_message_read(reply, "h", &fd);
+        if (r < 0)
+                return bus_log_parse_error(r);
+
+        fflush(stdout);
+        return copy_bytes(fd, STDOUT_FILENO, (uint64_t) -1, 0);
+}
+
+static int parse_argv(int argc, char *argv[]) {
+        enum {
+                ARG_VERSION = 0x100,
+                ARG_NO_PAGER,
+        };
+
+        static const struct option options[] = {
+                { "help",     no_argument, NULL, 'h'          },
+                { "version",  no_argument, NULL, ARG_VERSION  },
+                { "no-pager", no_argument, NULL, ARG_NO_PAGER },
+                {}
+        };
+
+        int c;
+
+        assert(argc >= 0);
+        assert(argv);
+
+        while ((c = getopt_long(argc, argv, "h", options, NULL)) >= 0)
+
+                switch (c) {
+
+                        case 'h':
+                                return help(0, NULL, NULL);
+
+                        case ARG_VERSION:
+                                return version();
+
+                        case ARG_NO_PAGER:
+                                arg_pager_flags |= PAGER_DISABLE;
+                                break;
+
+                        case '?':
+                                return -EINVAL;
+
+                        default:
+                                assert_not_reached("Invalid option passed.");
+                }
+
+        return 1;
+}
+
+static int run(int argc, char* argv[]) {
+        static const Verb verbs[] = {
+                { "help", VERB_ANY, VERB_ANY, 0,            help       },
+                { "dump", VERB_ANY, 1,        VERB_DEFAULT, dump_state },
+                {}
+        };
+
+        int r;
+
+        log_show_color(true);
+        log_parse_environment();
+        log_open();
+
+        r = parse_argv(argc, argv);
+        if (r <= 0)
+                return r;
+
+        return dispatch_verb(argc, argv, verbs, NULL);
+}
+
+DEFINE_MAIN_FUNCTION(run);

src/oom/oomd-manager-bus.c

@@ -0,0 +1,36 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <linux/capability.h>
+
+#include "bus-common-errors.h"
+#include "bus-polkit.h"
+#include "fd-util.h"
+#include "oomd-manager-bus.h"
+#include "oomd-manager.h"
+#include "user-util.h"
+
+static int bus_method_dump_by_fd(sd_bus_message *message, void *userdata, sd_bus_error *error) {
+        _cleanup_free_ char *dump = NULL;
+        _cleanup_close_ int fd = -1;
+        Manager *m = userdata;
+        int r;
+
+        assert(message);
+        assert(m);
+
+        r = manager_get_dump_string(m, &dump);
+        if (r < 0)
+                return r;
+
+        fd = acquire_data_fd(dump, strlen(dump), 0);
+        if (fd < 0)
+                return fd;
+
+        return sd_bus_reply_method_return(message, "h", fd);
+}
+
+const sd_bus_vtable manager_vtable[] = {
+        SD_BUS_VTABLE_START(0),
+        SD_BUS_METHOD("DumpByFileDescriptor", NULL, "h", bus_method_dump_by_fd, SD_BUS_VTABLE_UNPRIVILEGED),
+        SD_BUS_VTABLE_END
+};

src/oom/oomd-manager-bus.h

@@ -0,0 +1,8 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+#pragma once
+
+#include "sd-bus.h"
+
+typedef struct Manager Manager;
+
+extern const sd_bus_vtable manager_vtable[];

src/oom/oomd-manager.c

@@ -0,0 +1,549 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include "bus-log-control-api.h"
+#include "bus-util.h"
+#include "bus-polkit.h"
+#include "cgroup-util.h"
+#include "fd-util.h"
+#include "fileio.h"
+#include "oomd-manager-bus.h"
+#include "oomd-manager.h"
+#include "path-util.h"
+
+typedef struct ManagedOOMReply {
+        ManagedOOMMode mode;
+        char *path;
+        char *property;
+        unsigned limit;
+} ManagedOOMReply;
+
+static void managed_oom_reply_destroy(ManagedOOMReply *reply) {
+        assert(reply);
+        free(reply->path);
+        free(reply->property);
+}
+
+static int managed_oom_mode(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
+        ManagedOOMMode *mode = userdata, m;
+        const char *s;
+
+        assert(mode);
+        assert_se(s = json_variant_string(v));
+
+        m = managed_oom_mode_from_string(s);
+        if (m < 0)
+                return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL), "%s is not a valid ManagedOOMMode", s);
+
+        *mode = m;
+        return 0;
+}
+
+static int process_managed_oom_reply(
+                Varlink *link,
+                JsonVariant *parameters,
+                const char *error_id,
+                VarlinkReplyFlags flags,
+                void *userdata) {
+        JsonVariant *c, *cgroups;
+        Manager *m = userdata;
+        int r = 0;
+
+        assert(m);
+
+        static const JsonDispatch dispatch_table[] = {
+                { "mode",     JSON_VARIANT_STRING,   managed_oom_mode,       offsetof(ManagedOOMReply, mode),     JSON_MANDATORY },
+                { "path",     JSON_VARIANT_STRING,   json_dispatch_string,   offsetof(ManagedOOMReply, path),     JSON_MANDATORY },
+                { "property", JSON_VARIANT_STRING,   json_dispatch_string,   offsetof(ManagedOOMReply, property), JSON_MANDATORY },
+                { "limit",    JSON_VARIANT_UNSIGNED, json_dispatch_unsigned, offsetof(ManagedOOMReply, limit),    0 },
+                {},
+        };
+
+        if (error_id) {
+                r = -EIO;
+                log_debug("Error getting ManagedOOM cgroups: %s", error_id);
+                goto finish;
+        }
+
+        cgroups = json_variant_by_key(parameters, "cgroups");
+        if (!cgroups) {
+                r = -EINVAL;
+                goto finish;
+        }
+
+        /* Skip malformed elements and keep processing in case the others are good */
+        JSON_VARIANT_ARRAY_FOREACH(c, cgroups) {
+                _cleanup_(managed_oom_reply_destroy) ManagedOOMReply reply = {};
+                OomdCGroupContext *ctx;
+                Hashmap *monitor_hm;
+                loadavg_t limit;
+                int ret;
+
+                if (!json_variant_is_object(c))
+                        continue;
+
+                ret = json_dispatch(c, dispatch_table, NULL, 0, &reply);
+                if (ret == -ENOMEM) {
+                        r = ret;
+                        goto finish;
+                } else if (ret < 0)
+                        continue;
+
+                monitor_hm = streq(reply.property, "ManagedOOMSwap") ?
+                                m->monitored_swap_cgroup_contexts : m->monitored_mem_pressure_cgroup_contexts;
+
+                if (reply.mode == MANAGED_OOM_AUTO) {
+                        (void) oomd_cgroup_context_free(hashmap_remove(monitor_hm, reply.path));
+                        continue;
+                }
+
+                limit = m->default_mem_pressure_limit;
+
+                if (streq(reply.property, "ManagedOOMMemoryPressure")) {
+                        if (reply.limit > 100)
+                                continue;
+                        else if (reply.limit != 0) {
+                                ret = store_loadavg_fixed_point((unsigned long) reply.limit, 0, &limit);
+                                if (ret < 0)
+                                        continue;
+                        }
+                }
+
+                ret = oomd_insert_cgroup_context(NULL, monitor_hm, reply.path);
+                if (ret == -ENOMEM) {
+                        r = ret;
+                        goto finish;
+                }
+
+                /* Always update the limit in case it was changed. For non-memory pressure detection the value is
+                 * ignored so always updating it here is not a problem. */
+                ctx = hashmap_get(monitor_hm, reply.path);
+                if (ctx)
+                        ctx->mem_pressure_limit = limit;
+        }
+
+finish:
+        if (!FLAGS_SET(flags, VARLINK_REPLY_CONTINUES))
+                m->varlink = varlink_close_unref(link);
+
+        return r;
+}
+
+/* Fill `new_h` with `path`'s descendent OomdCGroupContexts. Only include descendent cgroups that are possible
+ * candidates for action. That is, only leaf cgroups or cgroups with memory.oom.group set to "1".
+ *
+ * This function ignores most errors in order to handle cgroups that may have been cleaned up while populating
+ * the hashmap.
+ *
+ * `new_h` is of the form { key: cgroup paths -> value: OomdCGroupContext } */
+static int recursively_get_cgroup_context(Hashmap *new_h, const char *path) {
+        _cleanup_free_ char *subpath = NULL;
+        _cleanup_closedir_ DIR *d = NULL;
+        int r;
+
+        assert(new_h);
+        assert(path);
+
+        r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
+        if (r < 0)
+                return r;
+
+        r = cg_read_subgroup(d, &subpath);
+        if (r < 0)
+                return r;
+        else if (r == 0) { /* No subgroups? We're a leaf node */
+                r = oomd_insert_cgroup_context(NULL, new_h, path);
+                return (r == -ENOMEM) ? r : 0;
+        }
+
+        do {
+                _cleanup_free_ char *cg_path = NULL;
+                bool oom_group;
+
+                cg_path = path_join(empty_to_root(path), subpath);
+                if (!cg_path)
+                        return -ENOMEM;
+
+                subpath = mfree(subpath);
+
+                r = cg_get_attribute_as_bool("memory", cg_path, "memory.oom.group", &oom_group);
+                /* The cgroup might be gone. Skip it as a candidate since we can't get information on it. */
+                if (r < 0)
+                        return (r == -ENOMEM) ? r : 0;
+
+                if (oom_group) {
+                        r = oomd_insert_cgroup_context(NULL, new_h, cg_path);
+                        if (r == -ENOMEM)
+                                return r;
+                } else {
+                        r = recursively_get_cgroup_context(new_h, cg_path);
+                        if (r == -ENOMEM)
+                                return r;
+                }
+        } while ((r = cg_read_subgroup(d, &subpath)) > 0);
+
+        return 0;
+}
+
+static int update_monitored_cgroup_contexts(Hashmap **monitored_cgroups) {
+        _cleanup_hashmap_free_ Hashmap *new_base = NULL;
+        OomdCGroupContext *ctx;
+        int r;
+
+        assert(monitored_cgroups);
+
+        new_base = hashmap_new(&oomd_cgroup_ctx_hash_ops);
+        if (!new_base)
+                return -ENOMEM;
+
+        HASHMAP_FOREACH(ctx, *monitored_cgroups) {
+                /* Skip most errors since the cgroup we're trying to update might not exist anymore. */
+                r = oomd_insert_cgroup_context(*monitored_cgroups, new_base, ctx->path);
+                if (r == -ENOMEM)
+                        return r;
+        }
+
+        hashmap_free(*monitored_cgroups);
+        *monitored_cgroups = TAKE_PTR(new_base);
+
+        return 0;
+}
+
+static int get_monitored_cgroup_contexts_candidates(Hashmap *monitored_cgroups, Hashmap **ret_candidates) {
+        _cleanup_hashmap_free_ Hashmap *candidates = NULL;
+        OomdCGroupContext *ctx;
+        int r;
+
+        assert(monitored_cgroups);
+        assert(ret_candidates);
+
+        candidates = hashmap_new(&oomd_cgroup_ctx_hash_ops);
+        if (!candidates)
+                return -ENOMEM;
+
+        HASHMAP_FOREACH(ctx, monitored_cgroups) {
+                r = recursively_get_cgroup_context(candidates, ctx->path);
+                if (r == -ENOMEM)
+                        return r;
+        }
+
+        *ret_candidates = TAKE_PTR(candidates);
+
+        return 0;
+}
+
+static int acquire_managed_oom_connect(Manager *m) {
+        _cleanup_(varlink_close_unrefp) Varlink *link = NULL;
+        int r;
+
+        assert(m);
+        assert(m->event);
+
+        r = varlink_connect_address(&link, VARLINK_ADDR_PATH_MANAGED_OOM);
+        if (r < 0)
+                return log_error_errno(r, "Failed to connect to %s: %m", VARLINK_ADDR_PATH_MANAGED_OOM);
+
+        (void) varlink_set_userdata(link, m);
+        (void) varlink_set_description(link, "oomd");
+        (void) varlink_set_relative_timeout(link, USEC_INFINITY);
+
+        r = varlink_attach_event(link, m->event, SD_EVENT_PRIORITY_NORMAL);
+        if (r < 0)
+                return log_error_errno(r, "Failed to attach varlink connection to event loop: %m");
+
+        r = varlink_bind_reply(link, process_managed_oom_reply);
+        if (r < 0)
+                return log_error_errno(r, "Failed to bind reply callback: %m");
+
+        r = varlink_observe(link, "io.systemd.ManagedOOM.SubscribeManagedOOMCGroups", NULL);
+        if (r < 0)
+                return log_error_errno(r, "Failed to observe varlink call: %m");
+
+        m->varlink = TAKE_PTR(link);
+        return 0;
+}
+
+static int monitor_cgroup_contexts_handler(sd_event_source *s, uint64_t usec, void *userdata) {
+        _cleanup_set_free_ Set *targets = NULL;
+        Manager *m = userdata;
+        usec_t usec_now;
+        int r;
+
+        assert(s);
+        assert(userdata);
+
+        /* Reset timer */
+        r = sd_event_now(sd_event_source_get_event(s), CLOCK_MONOTONIC, &usec_now);
+        if (r < 0)
+                return log_error_errno(r, "Failed to reset event timer");
+
+        r = sd_event_source_set_time_relative(s, INTERVAL_USEC);
+        if (r < 0)
+                return log_error_errno(r, "Failed to set relative time for timer");
+
+        /* Reconnect if our connection dropped */
+        if (!m->varlink) {
+                r = acquire_managed_oom_connect(m);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to acquire varlink connection");
+        }
+
+        /* Update the cgroups used for detection/action */
+        r = update_monitored_cgroup_contexts(&m->monitored_swap_cgroup_contexts);
+        if (r == -ENOMEM)
+                return log_error_errno(r, "Failed to update monitored swap cgroup contexts");
+
+        r = update_monitored_cgroup_contexts(&m->monitored_mem_pressure_cgroup_contexts);
+        if (r == -ENOMEM)
+                return log_error_errno(r, "Failed to update monitored memory pressure cgroup contexts");
+
+        r = oomd_system_context_acquire("/proc/swaps", &m->system_context);
+        /* If there aren't units depending on swap actions, the only error we exit on is ENOMEM */
+        if (r == -ENOMEM || (r < 0 && !hashmap_isempty(m->monitored_swap_cgroup_contexts)))
+                return log_error_errno(r, "Failed to acquire system context");
+
+        /* If we're still recovering from a kill, don't try to kill again yet */
+        if (m->post_action_delay_start > 0) {
+                if (m->post_action_delay_start + POST_ACTION_DELAY_USEC > usec_now)
+                        return 0;
+                else
+                        m->post_action_delay_start = 0;
+        }
+
+        r = oomd_pressure_above(m->monitored_mem_pressure_cgroup_contexts, PRESSURE_DURATION_USEC, &targets);
+        if (r == -ENOMEM)
+                return log_error_errno(r, "Failed to check if memory pressure exceeded limits");
+        else if (r == 1) {
+                /* Check if there was reclaim activity in the last interval. The concern is the following case:
+                 * Pressure climbed, a lot of high-frequency pages were reclaimed, and we killed the offending
+                 * cgroup. Even after this, well-behaved processes will fault in recently resident pages and
+                 * this will cause pressure to remain high. Thus if there isn't any reclaim pressure, no need
+                 * to kill something (it won't help anyways). */
+                if (oomd_memory_reclaim(m->monitored_mem_pressure_cgroup_contexts)) {
+                        _cleanup_hashmap_free_ Hashmap *candidates = NULL;
+                        OomdCGroupContext *t;
+
+                        r = get_monitored_cgroup_contexts_candidates(m->monitored_mem_pressure_cgroup_contexts, &candidates);
+                        if (r == -ENOMEM)
+                                return log_error_errno(r, "Failed to get monitored memory pressure cgroup candidates");
+
+                        SET_FOREACH(t, targets) {
+                                log_notice("Memory pressure for %s is greater than %lu for more than %"PRIu64" seconds and there was reclaim activity",
+                                        t->path, LOAD_INT(t->mem_pressure_limit), PRESSURE_DURATION_USEC / USEC_PER_SEC);
+
+                                r = oomd_kill_by_pgscan(candidates, t->path, m->dry_run);
+                                if (r == -ENOMEM)
+                                        return log_error_errno(r, "Failed to kill cgroup processes by pgscan");
+                                if (r < 0)
+                                        log_info("Failed to kill any cgroup(s) under %s based on pressure", t->path);
+                                else {
+                                        /* Don't act on all the high pressure cgroups at once; return as soon as we kill one */
+                                        m->post_action_delay_start = usec_now;
+                                        return 0;
+                                }
+                        }
+                }
+        }
+
+        if (oomd_swap_free_below(&m->system_context, (100 - m->swap_used_limit))) {
+                _cleanup_hashmap_free_ Hashmap *candidates = NULL;
+
+                log_notice("Swap used (%"PRIu64") / total (%"PRIu64") is more than %u%%",
+                        m->system_context.swap_used, m->system_context.swap_total, m->swap_used_limit);
+
+                r = get_monitored_cgroup_contexts_candidates(m->monitored_swap_cgroup_contexts, &candidates);
+                if (r == -ENOMEM)
+                        return log_error_errno(r, "Failed to get monitored swap cgroup candidates");
+
+                r = oomd_kill_by_swap_usage(candidates, m->dry_run);
+                if (r == -ENOMEM)
+                        return log_error_errno(r, "Failed to kill cgroup processes by swap usage");
+                if (r < 0)
+                        log_info("Failed to kill any cgroup(s) based on swap");
+                else {
+                        m->post_action_delay_start = usec_now;
+                        return 0;
+                }
+        }
+
+        return 0;
+}
+
+static int monitor_cgroup_contexts(Manager *m) {
+        _cleanup_(sd_event_source_unrefp) sd_event_source *s = NULL;
+        int r;
+
+        assert(m);
+        assert(m->event);
+
+        r = sd_event_add_time(m->event, &s, CLOCK_MONOTONIC, 0, 0, monitor_cgroup_contexts_handler, m);
+        if (r < 0)
+                return r;
+
+        r = sd_event_source_set_exit_on_failure(s, true);
+        if (r < 0)
+                return r;
+
+        r = sd_event_source_set_enabled(s, SD_EVENT_ON);
+        if (r < 0)
+                return r;
+
+        (void) sd_event_source_set_description(s, "oomd-timer");
+
+        m->cgroup_context_event_source = TAKE_PTR(s);
+        return 0;
+}
+
+void manager_free(Manager *m) {
+        assert(m);
+
+        varlink_close_unref(m->varlink);
+        sd_event_source_unref(m->cgroup_context_event_source);
+        sd_event_unref(m->event);
+
+        bus_verify_polkit_async_registry_free(m->polkit_registry);
+        sd_bus_flush_close_unref(m->bus);
+
+        hashmap_free(m->monitored_swap_cgroup_contexts);
+        hashmap_free(m->monitored_mem_pressure_cgroup_contexts);
+
+        free(m);
+}
+
+int manager_new(Manager **ret) {
+        _cleanup_(manager_freep) Manager *m = NULL;
+        int r;
+
+        assert(ret);
+
+        m = new0(Manager, 1);
+        if (!m)
+                return -ENOMEM;
+
+        r = sd_event_default(&m->event);
+        if (r < 0)
+                return r;
+
+        (void) sd_event_set_watchdog(m->event, true);
+
+        r = sd_event_add_signal(m->event, NULL, SIGINT, NULL, NULL);
+        if (r < 0)
+                return r;
+
+        r = sd_event_add_signal(m->event, NULL, SIGTERM, NULL, NULL);
+        if (r < 0)
+                return r;
+
+        m->monitored_swap_cgroup_contexts = hashmap_new(&oomd_cgroup_ctx_hash_ops);
+        if (!m->monitored_swap_cgroup_contexts)
+                return -ENOMEM;
+
+        m->monitored_mem_pressure_cgroup_contexts = hashmap_new(&oomd_cgroup_ctx_hash_ops);
+        if (!m->monitored_mem_pressure_cgroup_contexts)
+                return -ENOMEM;
+
+        *ret = TAKE_PTR(m);
+        return 0;
+}
+
+static int manager_connect_bus(Manager *m) {
+        int r;
+
+        assert(m);
+        assert(!m->bus);
+
+        r = bus_open_system_watch_bind_with_description(&m->bus, "bus-api-oom");
+        if (r < 0)
+                return log_error_errno(r, "Failed to connect to bus: %m");
+
+        r = sd_bus_add_object_vtable(m->bus, NULL, "/org/freedesktop/oom1", "org.freedesktop.oom1.Manager", manager_vtable, m);
+        if (r < 0)
+                return log_error_errno(r, "Failed to add manager object vtable: %m");
+
+        r = bus_log_control_api_register(m->bus);
+        if (r < 0)
+                return r;
+
+        r = sd_bus_request_name_async(m->bus, NULL, "org.freedesktop.oom1", 0, NULL, NULL);
+        if (r < 0)
+                return log_error_errno(r, "Failed to request name: %m");
+
+        r = sd_bus_attach_event(m->bus, m->event, 0);
+        if (r < 0)
+                return log_error_errno(r, "Failed to attach bus to event loop: %m");
+
+        return 0;
+}
+
+int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit) {
+        unsigned long l;
+        int r;
+
+        assert(m);
+
+        m->dry_run = dry_run;
+
+        m->swap_used_limit = swap_used_limit != -1 ? swap_used_limit : DEFAULT_SWAP_USED_LIMIT;
+        assert(m->swap_used_limit <= 100);
+
+        l = mem_pressure_limit != -1 ? mem_pressure_limit : DEFAULT_MEM_PRESSURE_LIMIT;
+        r = store_loadavg_fixed_point(l, 0, &m->default_mem_pressure_limit);
+        if (r < 0)
+                return r;
+
+        r = manager_connect_bus(m);
+        if (r < 0)
+                return r;
+
+        r = acquire_managed_oom_connect(m);
+        if (r < 0)
+                return r;
+
+        r = monitor_cgroup_contexts(m);
+        if (r < 0)
+                return r;
+
+        return 0;
+}
+
+int manager_get_dump_string(Manager *m, char **ret) {
+        _cleanup_free_ char *dump = NULL;
+        _cleanup_fclose_ FILE *f = NULL;
+        OomdCGroupContext *c;
+        size_t size;
+        char *key;
+        int r;
+
+        assert(m);
+        assert(ret);
+
+        f = open_memstream_unlocked(&dump, &size);
+        if (!f)
+                return -errno;
+
+        fprintf(f,
+                "Dry Run: %s\n"
+                "Swap Used Limit: %u%%\n"
+                "Default Memory Pressure Limit: %lu%%\n"
+                "System Context:\n",
+                yes_no(m->dry_run),
+                m->swap_used_limit,
+                LOAD_INT(m->default_mem_pressure_limit));
+        oomd_dump_system_context(&m->system_context, f, "\t");
+
+        fprintf(f, "Swap Monitored CGroups:\n");
+        HASHMAP_FOREACH_KEY(c, key, m->monitored_swap_cgroup_contexts)
+                oomd_dump_swap_cgroup_context(c, f, "\t");
+
+        fprintf(f, "Memory Pressure Monitored CGroups:\n");
+        HASHMAP_FOREACH_KEY(c, key, m->monitored_mem_pressure_cgroup_contexts)
+                oomd_dump_memory_pressure_cgroup_context(c, f, "\t");
+
+        r = fflush_and_check(f);
+        if (r < 0)
+                return r;
+
+        f = safe_fclose(f);
+
+        *ret = TAKE_PTR(dump);
+        return 0;
+}

src/oom/oomd-manager.h

@@ -0,0 +1,60 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+#pragma once
+
+#include "conf-parser.h"
+#include "oomd-util.h"
+#include "sd-event.h"
+#include "varlink.h"
+
+/* Polling interval for monitoring stats */
+#define INTERVAL_USEC (1 * USEC_PER_SEC)
+
+/* Used to weight the averages */
+#define AVERAGE_SIZE_DECAY 4
+
+/* Take action if 10s of memory pressure > 60 for more than 30s. We use the "full" value from PSI so this is the
+ * percentage of time all tasks were delayed (i.e. unproductive).
+ * Generally 60 or higher might be acceptable for something like system.slice with no memory.high set; processes in
+ * system.slice are assumed to be less latency sensitive. */
+#define PRESSURE_DURATION_USEC (30 * USEC_PER_SEC)
+#define DEFAULT_MEM_PRESSURE_LIMIT 60
+#define DEFAULT_SWAP_USED_LIMIT 90
+
+#define POST_ACTION_DELAY_USEC (15 * USEC_PER_SEC)
+
+typedef struct Manager Manager;
+
+struct Manager {
+        sd_bus *bus;
+        sd_event *event;
+
+        Hashmap *polkit_registry;
+
+        bool dry_run;
+        unsigned swap_used_limit;
+        loadavg_t default_mem_pressure_limit;
+
+        /* k: cgroup paths -> v: OomdCGroupContext
+         * Used to detect when to take action. */
+        Hashmap *monitored_swap_cgroup_contexts;
+        Hashmap *monitored_mem_pressure_cgroup_contexts;
+
+        OomdSystemContext system_context;
+
+        usec_t post_action_delay_start;
+
+        sd_event_source *cgroup_context_event_source;
+
+        Varlink *varlink;
+};
+
+void manager_free(Manager *m);
+DEFINE_TRIVIAL_CLEANUP_FUNC(Manager*, manager_free);
+
+int manager_new(Manager **ret);
+
+int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit);
+
+int manager_get_dump_string(Manager *m, char **ret);
+
+CONFIG_PARSER_PROTOTYPE(config_parse_oomd_default);

src/oom/oomd-util.c

@@ -0,0 +1,451 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <sys/xattr.h>
+#include <unistd.h>
+
+#include "cgroup-util.h"
+#include "fd-util.h"
+#include "format-util.h"
+#include "oomd-util.h"
+#include "parse-util.h"
+#include "path-util.h"
+#include "procfs-util.h"
+#include "signal-util.h"
+#include "sort-util.h"
+#include "stat-util.h"
+#include "stdio-util.h"
+
+DEFINE_HASH_OPS_WITH_VALUE_DESTRUCTOR(
+                oomd_cgroup_ctx_hash_ops,
+                char,
+                string_hash_func,
+                string_compare_func,
+                OomdCGroupContext,
+                oomd_cgroup_context_free);
+
+static int log_kill(pid_t pid, int sig, void *userdata) {
+        log_debug("oomd attempting to kill " PID_FMT " with %s", pid, signal_to_string(sig));
+        return 0;
+}
+
+static int increment_oomd_xattr(const char *path, const char *xattr, uint64_t num_procs_killed) {
+        _cleanup_free_ char *value = NULL;
+        char buf[DECIMAL_STR_MAX(uint64_t) + 1];
+        uint64_t curr_count = 0;
+        int r;
+
+        assert(path);
+        assert(xattr);
+
+        r = cg_get_xattr_malloc(SYSTEMD_CGROUP_CONTROLLER, path, xattr, &value);
+        if (r < 0 && r != -ENODATA)
+                return r;
+
+        if (!isempty(value)) {
+                 r = safe_atou64(value, &curr_count);
+                 if (r < 0)
+                         return r;
+        }
+
+        if (curr_count > UINT64_MAX - num_procs_killed)
+                return -EOVERFLOW;
+
+        xsprintf(buf, "%"PRIu64, curr_count + num_procs_killed);
+        r = cg_set_xattr(SYSTEMD_CGROUP_CONTROLLER, path, xattr, buf, strlen(buf), 0);
+        if (r < 0)
+                return r;
+
+        return 0;
+}
+
+OomdCGroupContext *oomd_cgroup_context_free(OomdCGroupContext *ctx) {
+        if (!ctx)
+                return NULL;
+
+        free(ctx->path);
+        return mfree(ctx);
+}
+
+int oomd_pressure_above(Hashmap *h, usec_t duration, Set **ret) {
+        _cleanup_set_free_ Set *targets = NULL;
+        OomdCGroupContext *ctx;
+        char *key;
+        int r;
+
+        assert(h);
+        assert(ret);
+
+        targets = set_new(NULL);
+        if (!targets)
+                return -ENOMEM;
+
+        HASHMAP_FOREACH_KEY(ctx, key, h) {
+                if (ctx->memory_pressure.avg10 > ctx->mem_pressure_limit) {
+                        usec_t diff;
+
+                        if (ctx->last_hit_mem_pressure_limit == 0)
+                                ctx->last_hit_mem_pressure_limit = now(CLOCK_MONOTONIC);
+
+                        diff = now(CLOCK_MONOTONIC) - ctx->last_hit_mem_pressure_limit;
+                        if (diff >= duration) {
+                                r = set_put(targets, ctx);
+                                if (r < 0)
+                                        return -ENOMEM;
+                        }
+                } else
+                        ctx->last_hit_mem_pressure_limit = 0;
+        }
+
+        if (!set_isempty(targets)) {
+                *ret = TAKE_PTR(targets);
+                return 1;
+        }
+
+        *ret = NULL;
+        return 0;
+}
+
+bool oomd_memory_reclaim(Hashmap *h) {
+        uint64_t pgscan = 0, pgscan_of = 0, last_pgscan = 0, last_pgscan_of = 0;
+        OomdCGroupContext *ctx;
+
+        assert(h);
+
+        /* If sum of all the current pgscan values are greater than the sum of all the last_pgscan values,
+         * there was reclaim activity. Used along with pressure checks to decide whether to take action. */
+
+        HASHMAP_FOREACH(ctx, h) {
+                uint64_t sum;
+
+                sum = pgscan + ctx->pgscan;
+                if (sum < pgscan || sum < ctx->pgscan)
+                        pgscan_of++; /* count overflows */
+                pgscan = sum;
+
+                sum = last_pgscan + ctx->last_pgscan;
+                if (sum < last_pgscan || sum < ctx->last_pgscan)
+                        last_pgscan_of++; /* count overflows */
+                last_pgscan = sum;
+        }
+
+        /* overflow counts are the same, return sums comparison */
+        if (last_pgscan_of == pgscan_of)
+                return pgscan > last_pgscan;
+
+        return pgscan_of > last_pgscan_of;
+}
+
+bool oomd_swap_free_below(const OomdSystemContext *ctx, uint64_t threshold_percent) {
+        uint64_t swap_threshold;
+
+        assert(ctx);
+        assert(threshold_percent <= 100);
+
+        swap_threshold = ctx->swap_total * threshold_percent / ((uint64_t) 100);
+        return (ctx->swap_total - ctx->swap_used) < swap_threshold;
+}
+
+int oomd_sort_cgroup_contexts(Hashmap *h, oomd_compare_t compare_func, const char *prefix, OomdCGroupContext ***ret) {
+        _cleanup_free_ OomdCGroupContext **sorted = NULL;
+        OomdCGroupContext *item;
+        size_t k = 0;
+
+        assert(h);
+        assert(compare_func);
+        assert(ret);
+
+        sorted = new0(OomdCGroupContext*, hashmap_size(h));
+        if (!sorted)
+                return -ENOMEM;
+
+        HASHMAP_FOREACH(item, h) {
+                if (item->path && prefix && !path_startswith(item->path, prefix))
+                        continue;
+
+                sorted[k++] = item;
+        }
+
+        typesafe_qsort(sorted, k, compare_func);
+
+        *ret = TAKE_PTR(sorted);
+
+        assert(k <= INT_MAX);
+        return (int) k;
+}
+
+int oomd_cgroup_kill(const char *path, bool recurse, bool dry_run) {
+        _cleanup_set_free_ Set *pids_killed = NULL;
+        int r;
+
+        assert(path);
+
+        if (dry_run) {
+                _cleanup_free_ char *cg_path = NULL;
+
+                r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &cg_path);
+                if (r < 0)
+                        return r;
+
+                log_debug("oomd dry-run: Would have tried to kill %s with recurse=%s", cg_path, true_false(recurse));
+                return 0;
+        }
+
+        pids_killed = set_new(NULL);
+        if (!pids_killed)
+                return -ENOMEM;
+
+        if (recurse)
+                r = cg_kill_recursive(SYSTEMD_CGROUP_CONTROLLER, path, SIGKILL, CGROUP_IGNORE_SELF, pids_killed, log_kill, NULL);
+        else
+                r = cg_kill(SYSTEMD_CGROUP_CONTROLLER, path, SIGKILL, CGROUP_IGNORE_SELF, pids_killed, log_kill, NULL);
+        if (r < 0)
+                return r;
+
+        r = increment_oomd_xattr(path, "user.systemd_oomd_kill", set_size(pids_killed));
+        if (r < 0)
+                log_debug_errno(r, "Failed to set user.systemd_oomd_kill on kill: %m");
+
+        return set_size(pids_killed) != 0;
+}
+
+int oomd_kill_by_pgscan(Hashmap *h, const char *prefix, bool dry_run) {
+        _cleanup_free_ OomdCGroupContext **sorted = NULL;
+        int r;
+
+        assert(h);
+
+        r = oomd_sort_cgroup_contexts(h, compare_pgscan, prefix, &sorted);
+        if (r < 0)
+                return r;
+
+        for (int i = 0; i < r; i++) {
+                if (sorted[i]->pgscan == 0)
+                        break;
+
+                r = oomd_cgroup_kill(sorted[i]->path, true, dry_run);
+                if (r > 0 || r == -ENOMEM)
+                        break;
+        }
+
+        return r;
+}
+
+int oomd_kill_by_swap_usage(Hashmap *h, bool dry_run) {
+        _cleanup_free_ OomdCGroupContext **sorted = NULL;
+        int r;
+
+        assert(h);
+
+        r = oomd_sort_cgroup_contexts(h, compare_swap_usage, NULL, &sorted);
+        if (r < 0)
+                return r;
+
+        /* Try to kill cgroups with non-zero swap usage until we either succeed in
+         * killing or we get to a cgroup with no swap usage. */
+        for (int i = 0; i < r; i++) {
+                if (sorted[i]->swap_usage == 0)
+                        break;
+
+                r = oomd_cgroup_kill(sorted[i]->path, true, dry_run);
+                if (r > 0 || r == -ENOMEM)
+                        break;
+        }
+
+        return r;
+}
+
+int oomd_cgroup_context_acquire(const char *path, OomdCGroupContext **ret) {
+        _cleanup_(oomd_cgroup_context_freep) OomdCGroupContext *ctx = NULL;
+        _cleanup_free_ char *p = NULL, *val = NULL;
+        bool is_root;
+        int r;
+
+        assert(path);
+        assert(ret);
+
+        ctx = new0(OomdCGroupContext, 1);
+        if (!ctx)
+                return -ENOMEM;
+
+        is_root = empty_or_root(path);
+
+        r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, "memory.pressure", &p);
+        if (r < 0)
+                return log_debug_errno(r, "Error getting cgroup memory pressure path from %s: %m", path);
+
+        r = read_resource_pressure(p, PRESSURE_TYPE_FULL, &ctx->memory_pressure);
+        if (r < 0)
+                return log_debug_errno(r, "Error parsing memory pressure from %s: %m", p);
+
+        if (is_root) {
+                r = procfs_memory_get_used(&ctx->current_memory_usage);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting memory used from procfs: %m");
+        } else {
+                r = cg_get_attribute_as_uint64(SYSTEMD_CGROUP_CONTROLLER, path, "memory.current", &ctx->current_memory_usage);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting memory.current from %s: %m", path);
+
+                r = cg_get_attribute_as_uint64(SYSTEMD_CGROUP_CONTROLLER, path, "memory.min", &ctx->memory_min);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting memory.min from %s: %m", path);
+
+                r = cg_get_attribute_as_uint64(SYSTEMD_CGROUP_CONTROLLER, path, "memory.low", &ctx->memory_low);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting memory.low from %s: %m", path);
+
+                r = cg_get_attribute_as_uint64(SYSTEMD_CGROUP_CONTROLLER, path, "memory.swap.current", &ctx->swap_usage);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting memory.swap.current from %s: %m", path);
+
+                r = cg_get_keyed_attribute(SYSTEMD_CGROUP_CONTROLLER, path, "memory.stat", STRV_MAKE("pgscan"), &val);
+                if (r < 0)
+                        return log_debug_errno(r, "Error getting pgscan from memory.stat under %s: %m", path);
+
+                r = safe_atou64(val, &ctx->pgscan);
+                if (r < 0)
+                        return log_debug_errno(r, "Error converting pgscan value to uint64_t: %m");
+        }
+
+        ctx->path = strdup(empty_to_root(path));
+        if (!ctx->path)
+                return -ENOMEM;
+
+        *ret = TAKE_PTR(ctx);
+        return 0;
+}
+
+int oomd_system_context_acquire(const char *proc_swaps_path, OomdSystemContext *ret) {
+        _cleanup_fclose_ FILE *f = NULL;
+        OomdSystemContext ctx = {};
+        int r;
+
+        assert(proc_swaps_path);
+        assert(ret);
+
+        f = fopen(proc_swaps_path, "re");
+        if (!f)
+                return -errno;
+
+        (void) fscanf(f, "%*s %*s %*s %*s %*s\n");
+
+        for (;;) {
+                uint64_t total, used;
+
+                r = fscanf(f,
+                           "%*s "          /* device/file */
+                           "%*s "          /* type of swap */
+                           "%" PRIu64 " "  /* swap size */
+                           "%" PRIu64 " "  /* used */
+                           "%*s\n",        /* priority */
+                           &total, &used);
+
+                if (r == EOF && feof(f))
+                         break;
+
+                if (r != 2) {
+                        if (ferror(f))
+                                return log_debug_errno(errno, "Error reading from %s: %m", proc_swaps_path);
+
+                        return log_debug_errno(SYNTHETIC_ERRNO(EIO),
+                                               "Failed to parse values from %s: %m", proc_swaps_path);
+                }
+
+                ctx.swap_total += total * 1024U;
+                ctx.swap_used += used * 1024U;
+        }
+
+        *ret = ctx;
+        return 0;
+}
+
+int oomd_insert_cgroup_context(Hashmap *old_h, Hashmap *new_h, const char *path) {
+        _cleanup_(oomd_cgroup_context_freep) OomdCGroupContext *curr_ctx = NULL;
+        OomdCGroupContext *old_ctx, *ctx;
+        int r;
+
+        assert(new_h);
+        assert(path);
+
+        r = oomd_cgroup_context_acquire(path, &curr_ctx);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to get OomdCGroupContext for %s: %m", path);
+
+        old_ctx = hashmap_get(old_h, path);
+        if (old_ctx) {
+                curr_ctx->last_pgscan = old_ctx->pgscan;
+                curr_ctx->mem_pressure_limit = old_ctx->mem_pressure_limit;
+                curr_ctx->last_hit_mem_pressure_limit = old_ctx->last_hit_mem_pressure_limit;
+        }
+
+        ctx = TAKE_PTR(curr_ctx);
+        r = hashmap_put(new_h, ctx->path, ctx);
+        if (r < 0)
+                return r;
+
+        return 0;
+}
+
+void oomd_dump_swap_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix) {
+        char swap[FORMAT_BYTES_MAX];
+
+        assert(ctx);
+        assert(f);
+
+        if (!empty_or_root(ctx->path))
+                fprintf(f,
+                        "%sPath: %s\n"
+                        "%s\tSwap Usage: %s\n",
+                        strempty(prefix), ctx->path,
+                        strempty(prefix), format_bytes(swap, sizeof(swap), ctx->swap_usage));
+        else
+                fprintf(f,
+                        "%sPath: %s\n"
+                        "%s\tSwap Usage: (see System Context)\n",
+                        strempty(prefix), ctx->path,
+                        strempty(prefix));
+}
+
+void oomd_dump_memory_pressure_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix) {
+        char tbuf[FORMAT_TIMESPAN_MAX], mem_use[FORMAT_BYTES_MAX];
+        char mem_min[FORMAT_BYTES_MAX], mem_low[FORMAT_BYTES_MAX];
+
+        assert(ctx);
+        assert(f);
+
+        fprintf(f,
+                "%sPath: %s\n"
+                "%s\tMemory Pressure Limit: %lu%%\n"
+                "%s\tPressure: Avg10: %lu.%02lu Avg60: %lu.%02lu Avg300: %lu.%02lu Total: %s\n"
+                "%s\tCurrent Memory Usage: %s\n",
+                strempty(prefix), ctx->path,
+                strempty(prefix), LOAD_INT(ctx->mem_pressure_limit),
+                strempty(prefix),
+                LOAD_INT(ctx->memory_pressure.avg10), LOAD_FRAC(ctx->memory_pressure.avg10),
+                LOAD_INT(ctx->memory_pressure.avg60), LOAD_FRAC(ctx->memory_pressure.avg60),
+                LOAD_INT(ctx->memory_pressure.avg300), LOAD_FRAC(ctx->memory_pressure.avg300),
+                format_timespan(tbuf, sizeof(tbuf), ctx->memory_pressure.total, USEC_PER_SEC),
+                strempty(prefix), format_bytes(mem_use, sizeof(mem_use), ctx->current_memory_usage));
+
+        if (!empty_or_root(ctx->path))
+                fprintf(f,
+                        "%s\tMemory Min: %s\n"
+                        "%s\tMemory Low: %s\n"
+                        "%s\tPgscan: %" PRIu64 "\n",
+                        strempty(prefix), format_bytes_cgroup_protection(mem_min, sizeof(mem_min), ctx->memory_min),
+                        strempty(prefix), format_bytes_cgroup_protection(mem_low, sizeof(mem_low), ctx->memory_low),
+                        strempty(prefix), ctx->pgscan);
+}
+
+void oomd_dump_system_context(const OomdSystemContext *ctx, FILE *f, const char *prefix) {
+        char used[FORMAT_BYTES_MAX], total[FORMAT_BYTES_MAX];
+
+        assert(ctx);
+        assert(f);
+
+        fprintf(f,
+                "%sSwap: Used: %s Total: %s\n",
+                strempty(prefix),
+                format_bytes(used, sizeof(used), ctx->swap_used),
+                format_bytes(total, sizeof(total), ctx->swap_total));
+}

src/oom/oomd-util.h

@@ -0,0 +1,112 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+#pragma once
+
+#include <stdbool.h>
+
+#include "hashmap.h"
+#include "psi-util.h"
+
+#define GROWING_SIZE_PERCENTILE 80
+
+extern const struct hash_ops oomd_cgroup_ctx_hash_ops;
+
+typedef struct OomdCGroupContext OomdCGroupContext;
+typedef struct OomdSystemContext OomdSystemContext;
+
+typedef int (oomd_compare_t)(OomdCGroupContext * const *, OomdCGroupContext * const *);
+
+struct OomdCGroupContext {
+        char *path;
+
+        ResourcePressure memory_pressure;
+
+        uint64_t current_memory_usage;
+
+        uint64_t memory_min;
+        uint64_t memory_low;
+        uint64_t swap_usage;
+
+        uint64_t last_pgscan;
+        uint64_t pgscan;
+
+        /* These are only used by oomd_pressure_above for acting on high memory pressure. */
+        loadavg_t mem_pressure_limit;
+        usec_t last_hit_mem_pressure_limit;
+};
+
+struct OomdSystemContext {
+        uint64_t swap_total;
+        uint64_t swap_used;
+};
+
+OomdCGroupContext *oomd_cgroup_context_free(OomdCGroupContext *ctx);
+DEFINE_TRIVIAL_CLEANUP_FUNC(OomdCGroupContext*, oomd_cgroup_context_free);
+
+/* All hashmaps used with these functions are expected to be of the form
+ * key: cgroup paths -> value: OomdCGroupContext. */
+
+/* Scans all the OomdCGroupContexts in `h` and returns 1 and a set of pointers to those OomdCGroupContexts in `ret`
+ * if any of them have exceeded their supplied memory pressure limits for the `duration` length of time.
+ * `last_hit_mem_pressure_limit` is updated accordingly for each entry when the limit is exceeded, and when it returns
+ * below the limit.
+ * Returns 0 and sets `ret` to an empty set if no entries exceeded limits for `duration`.
+ * Returns -ENOMEM for allocation errors. */
+int oomd_pressure_above(Hashmap *h, usec_t duration, Set **ret);
+
+/* Sum up current OomdCGroupContexts' pgscan values and last interval's pgscan values in `h`. Returns true if the
+ * current sum is higher than the last interval's sum (there was some reclaim activity). */
+bool oomd_memory_reclaim(Hashmap *h);
+
+/* Returns true if the amount of swap free is below the percentage of swap specified by `threshold_percent`. */
+bool oomd_swap_free_below(const OomdSystemContext *ctx, uint64_t threshold_percent);
+
+static inline int compare_pgscan(OomdCGroupContext * const *c1, OomdCGroupContext * const *c2) {
+        assert(c1);
+        assert(c2);
+
+        if ((*c1)->pgscan > (*c2)->pgscan)
+                return -1;
+        else if ((*c1)->pgscan < (*c2)->pgscan)
+                return 1;
+        else
+                return 0;
+}
+
+static inline int compare_swap_usage(OomdCGroupContext * const *c1, OomdCGroupContext * const *c2) {
+        assert(c1);
+        assert(c2);
+
+        if ((*c1)->swap_usage > (*c2)->swap_usage)
+                return -1;
+        else if ((*c1)->swap_usage < (*c2)->swap_usage)
+                return 1;
+        else
+                return 0;
+}
+
+/* Get an array of OomdCGroupContexts from `h`, qsorted from largest to smallest values according to `compare_func`.
+ * If `prefix` is not NULL, only include OomdCGroupContexts whose paths start with prefix. Otherwise all paths are sorted.
+ * Returns the number of sorted items; negative on error. */
+int oomd_sort_cgroup_contexts(Hashmap *h, oomd_compare_t compare_func, const char *prefix, OomdCGroupContext ***ret);
+
+/* Returns a negative value on error, 0 if no processes were killed, or 1 if processes were killed. */
+int oomd_cgroup_kill(const char *path, bool recurse, bool dry_run);
+
+/* The following oomd_kill_by_* functions return 1 if processes were killed, or negative otherwise. */
+/* If `prefix` is supplied, only cgroups whose paths start with `prefix` are eligible candidates. Otherwise,
+ * everything in `h` is a candidate. */
+int oomd_kill_by_pgscan(Hashmap *h, const char *prefix, bool dry_run);
+int oomd_kill_by_swap_usage(Hashmap *h, bool dry_run);
+
+int oomd_cgroup_context_acquire(const char *path, OomdCGroupContext **ret);
+int oomd_system_context_acquire(const char *proc_swaps_path, OomdSystemContext *ret);
+
+/* Get the OomdCGroupContext of `path` and insert it into `new_h`. The key for the inserted context will be `path`.
+ *
+ * `old_h` is used to get data used to calculate prior interval information. `old_h` can be NULL in which case there
+ * was no prior data to reference. */
+int oomd_insert_cgroup_context(Hashmap *old_h, Hashmap *new_h, const char *path);
+
+void oomd_dump_swap_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix);
+void oomd_dump_memory_pressure_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix);
+void oomd_dump_system_context(const OomdSystemContext *ctx, FILE *f, const char *prefix);

src/oom/oomd.c

@@ -0,0 +1,144 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <getopt.h>
+
+#include "cgroup-util.h"
+#include "conf-parser.h"
+#include "daemon-util.h"
+#include "log.h"
+#include "main-func.h"
+#include "oomd-manager.h"
+#include "parse-util.h"
+#include "pretty-print.c"
+#include "psi-util.h"
+#include "signal-util.h"
+
+static bool arg_dry_run = false;
+static int arg_swap_used_limit = -1;
+static int arg_mem_pressure_limit = -1;
+
+static int parse_config(void) {
+        static const ConfigTableItem items[] = {
+                { "OOM", "SwapUsedLimitPercent",              config_parse_percent, 0, &arg_swap_used_limit    },
+                { "OOM", "DefaultMemoryPressureLimitPercent", config_parse_percent, 0, &arg_mem_pressure_limit },
+                {}
+        };
+
+        return config_parse_many_nulstr(PKGSYSCONFDIR "/oomd.conf",
+                                        CONF_PATHS_NULSTR("systemd/oomd.conf.d"),
+                                        "OOM\0",
+                                        config_item_table_lookup,
+                                        items,
+                                        CONFIG_PARSE_WARN,
+                                        NULL,
+                                        NULL);
+}
+
+static int help(void) {
+        _cleanup_free_ char *link = NULL;
+        int r;
+
+        r = terminal_urlify_man("systemd-oomd", "1", &link);
+        if (r < 0)
+                return log_oom();
+
+        printf("%s [OPTIONS...]\n\n"
+                        "Run the userspace out-of-memory (OOM) killer.\n\n"
+                        "  -h --help     Show this help\n"
+                        "     --dry-run  Log write/destructive actions instead of doing them\n"
+                        "\nSee the %s for details.\n"
+                        , program_invocation_short_name
+                        , link
+              );
+
+        return 0;
+}
+
+static int parse_argv(int argc, char *argv[]) {
+        enum {
+                ARG_DRY_RUN,
+        };
+
+        static const struct option options[] = {
+                { "help",    no_argument, NULL, 'h'         },
+                { "dry-run", no_argument, NULL, ARG_DRY_RUN },
+                {}
+        };
+
+        int c;
+
+        assert(argc >= 0);
+        assert(argv);
+
+        while ((c = getopt_long(argc, argv, "h", options, NULL)) >= 0)
+
+                switch (c) {
+
+                        case 'h':
+                                return help();
+
+                        case ARG_DRY_RUN:
+                                arg_dry_run = true;
+                                break;
+
+                        case '?':
+                                return -EINVAL;
+
+                        default:
+                                assert_not_reached("Invalid option passed.");
+                }
+
+        return 1;
+}
+
+static int run(int argc, char *argv[]) {
+        _cleanup_(notify_on_cleanup) const char *notify_msg = NULL;
+        _cleanup_(manager_freep) Manager *m = NULL;
+        int r;
+
+        log_setup_service();
+
+        r = parse_argv(argc, argv);
+        if (r <= 0)
+                return r;
+
+        r = parse_config();
+        if (r < 0)
+                return r;
+
+        /* Do some basic requirement checks for running systemd-oomd. It's not exhaustive as some of the other
+         * requirements do not have a reliable means to check for in code. */
+        if (access("/proc/swaps", F_OK) < 0)
+                return log_error_errno(errno, "Swap not enabled: %m");
+
+        if (!is_pressure_supported())
+                return log_error_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "Pressure Stall Information (PSI) is not supported");
+
+        r = cg_all_unified();
+        if (r < 0)
+                return log_error_errno(r, "Failed to determine whether the unified cgroups hierarchy is used: %m");
+        if (r == 0)
+                return log_error_errno(SYNTHETIC_ERRNO(EOPNOTSUPP), "Requires the unified cgroups hierarchy");
+
+        assert_se(sigprocmask_many(SIG_BLOCK, NULL, SIGTERM, SIGINT, -1) >= 0);
+
+        r = manager_new(&m);
+        if (r < 0)
+                return log_error_errno(r, "Failed to create manager: %m");
+
+        r = manager_start(m, arg_dry_run, arg_swap_used_limit, arg_mem_pressure_limit);
+        if (r < 0)
+                return log_error_errno(r, "Failed to start up daemon: %m");
+
+        notify_msg = notify_start(NOTIFY_READY, NOTIFY_STOPPING);
+
+        log_info("systemd-oomd starting%s!", arg_dry_run ? " in dry run mode" : "");
+
+        r = sd_event_loop(m->event);
+        if (r < 0)
+                return log_error_errno(r, "Event loop failed: %m");
+
+        return 0;
+}
+
+DEFINE_MAIN_FUNCTION(run);

src/oom/oomd.conf

@@ -0,0 +1,16 @@
+#  This file is part of systemd.
+#
+#  systemd is free software; you can redistribute it and/or modify it
+#  under the terms of the GNU Lesser General Public License as published by
+#  the Free Software Foundation; either version 2.1 of the License, or
+#  (at your option) any later version.
+#
+# Entries in this file show the compile time defaults.
+# You can change settings by editing this file.
+# Defaults can be restored by simply deleting this file.
+#
+# See oomd.conf(5) for details
+
+[OOM]
+#SwapUsedLimitPercent=90%
+#DefaultMemoryPressureLimitPercent=60%

src/oom/org.freedesktop.oom1.conf

@@ -0,0 +1,47 @@
+<?xml version="1.0"?> <!--*-nxml-*-->
+<!DOCTYPE busconfig PUBLIC "-//freedesktop//DTD D-BUS Bus Configuration 1.0//EN"
+        "http://www.freedesktop.org/standards/dbus/1.0/busconfig.dtd">
+
+<!-- SPDX-License-Identifier: LGPL-2.1+ -->
+
+<busconfig>
+
+        <policy user="systemd-oom">
+                <allow own="org.freedesktop.oom1"/>
+                <allow send_destination="org.freedesktop.oom1"/>
+                <allow receive_sender="org.freedesktop.oom1"/>
+        </policy>
+
+        <policy user="root">
+                <allow send_destination="org.freedesktop.oom1"/>
+        </policy>
+
+        <policy context="default">
+                <deny send_destination="org.freedesktop.oom1"/>
+
+                <!-- Generic interfaces -->
+
+                <allow send_destination="org.freedesktop.oom1"
+                       send_interface="org.freedesktop.DBus.Introspectable"/>
+
+                <allow send_destination="org.freedesktop.oom1"
+                       send_interface="org.freedesktop.DBus.Peer"/>
+
+                <allow send_destination="org.freedesktop.oom1"
+                       send_interface="org.freedesktop.DBus.Properties"
+                       send_member="Get"/>
+
+                <allow send_destination="org.freedesktop.oom1"
+                       send_interface="org.freedesktop.DBus.Properties"
+                       send_member="GetAll"/>
+
+                <!-- Manager interface -->
+
+                <allow send_destination="org.freedesktop.oom1"
+                       send_interface="org.freedesktop.oom1.Manager"
+                       send_member="DumpByFileDescriptor"/>
+
+                <allow receive_sender="org.freedesktop.oom1"/>
+        </policy>
+
+</busconfig>

src/oom/org.freedesktop.oom1.service

@@ -0,0 +1,14 @@
+#  SPDX-License-Identifier: LGPL-2.1+
+#
+#  This file is part of systemd.
+#
+#  systemd is free software; you can redistribute it and/or modify it
+#  under the terms of the GNU Lesser General Public License as published by
+#  the Free Software Foundation; either version 2.1 of the License, or
+#  (at your option) any later version.
+
+[D-BUS Service]
+Name=org.freedesktop.oom1
+Exec=/bin/false
+User=root
+SystemdService=dbus-org.freedesktop.oom1.service

src/oom/test-oomd-util.c

@@ -0,0 +1,348 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <unistd.h>
+
+#include "alloc-util.h"
+#include "cgroup-setup.h"
+#include "cgroup-util.h"
+#include "fileio.h"
+#include "fs-util.h"
+#include "oomd-util.h"
+#include "parse-util.h"
+#include "path-util.h"
+#include "string-util.h"
+#include "strv.h"
+#include "tests.h"
+
+static int fork_and_sleep(unsigned sleep_min) {
+        usec_t n, timeout, ts;
+
+        pid_t pid = fork();
+        assert_se(pid >= 0);
+
+        if (pid == 0) {
+                timeout = sleep_min * USEC_PER_MINUTE;
+                ts = now(CLOCK_MONOTONIC);
+                while (true) {
+                        n = now(CLOCK_MONOTONIC);
+                        if (ts + timeout < n) {
+                                log_error("Child timed out waiting to be killed");
+                                abort();
+                        }
+                        sleep(1);
+                }
+        }
+
+        return pid;
+}
+
+static void test_oomd_cgroup_kill(void) {
+        _cleanup_free_ char *cgroup_root = NULL, *cgroup = NULL;
+        int pid[2];
+
+        if (geteuid() != 0)
+                return (void) log_tests_skipped("not root");
+
+        if (cg_all_unified() <= 0)
+                return (void) log_tests_skipped("cgroups are not running in unified mode");
+
+        assert_se(cg_pid_get_path(NULL, 0, &cgroup_root) >= 0);
+
+        /* Create another cgroup below this one for the pids we forked off. We need this to be managed
+         * by the test so that pid1 doesn't delete it before we can read the xattrs. */
+        cgroup = path_join(cgroup_root, "oomdkilltest");
+        assert(cgroup);
+
+        /* If we don't have permissions to set xattrs we're likely in a userns or missing capabilities */
+        if (cg_set_xattr(SYSTEMD_CGROUP_CONTROLLER, cgroup, "user.oomd_test", "test", 4, 0) == -EPERM)
+                return (void) log_tests_skipped("no permissions to set user xattrs");
+
+        /* Do this twice to also check the increment behavior on the xattrs */
+        for (int i = 0; i < 2; i++) {
+                _cleanup_free_ char *v = NULL;
+                int r;
+
+                for (int j = 0; j < 2; j++) {
+                        pid[j] = fork_and_sleep(5);
+                        assert_se(cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, cgroup, pid[j]) >= 0);
+                }
+
+                r = oomd_cgroup_kill(cgroup, false /* recurse */, false /* dry run */);
+                if (r <= 0) {
+                        log_debug_errno(r, "Failed to kill processes under %s: %m", cgroup);
+                        abort();
+                }
+
+                /* Wait a bit since processes may take some time to be cleaned up. */
+                sleep(2);
+                assert_se(cg_is_empty(SYSTEMD_CGROUP_CONTROLLER, cgroup) == true);
+
+                assert_se(cg_get_xattr_malloc(SYSTEMD_CGROUP_CONTROLLER, cgroup, "user.systemd_oomd_kill", &v) >= 0);
+                assert_se(memcmp(v, i == 0 ? "2" : "4", 2) == 0);
+        }
+}
+
+static void test_oomd_cgroup_context_acquire_and_insert(void) {
+        _cleanup_hashmap_free_ Hashmap *h1 = NULL, *h2 = NULL;
+        _cleanup_(oomd_cgroup_context_freep) OomdCGroupContext *ctx = NULL;
+        _cleanup_free_ char *cgroup = NULL;
+        OomdCGroupContext *c1, *c2;
+
+        if (geteuid() != 0)
+                return (void) log_tests_skipped("not root");
+
+        if (!is_pressure_supported())
+                return (void) log_tests_skipped("system does not support pressure");
+
+        if (cg_all_unified() <= 0)
+                return (void) log_tests_skipped("cgroups are not running in unified mode");
+
+        assert_se(cg_pid_get_path(NULL, 0, &cgroup) >= 0);
+
+        assert_se(oomd_cgroup_context_acquire(cgroup, &ctx) == 0);
+
+        assert_se(streq(ctx->path, cgroup));
+        assert_se(ctx->memory_pressure.avg10 == 0);
+        assert_se(ctx->memory_pressure.avg60 == 0);
+        assert_se(ctx->memory_pressure.avg300 == 0);
+        assert_se(ctx->memory_pressure.total == 0);
+        assert_se(ctx->current_memory_usage > 0);
+        assert_se(ctx->memory_min == 0);
+        assert_se(ctx->memory_low == 0);
+        assert_se(ctx->swap_usage == 0);
+        assert_se(ctx->last_pgscan == 0);
+        assert_se(ctx->pgscan == 0);
+        ctx = oomd_cgroup_context_free(ctx);
+
+        /* Test the root cgroup */
+        assert_se(oomd_cgroup_context_acquire("", &ctx) == 0);
+        assert_se(streq(ctx->path, "/"));
+        assert_se(ctx->current_memory_usage > 0);
+
+        /* Test hashmap inserts */
+        assert_se(h1 = hashmap_new(&oomd_cgroup_ctx_hash_ops));
+        assert_se(oomd_insert_cgroup_context(NULL, h1, cgroup) == 0);
+        c1 = hashmap_get(h1, cgroup);
+        assert_se(c1);
+
+         /* make sure certain values from h1 get updated in h2 */
+        c1->pgscan = 5555;
+        c1->mem_pressure_limit = 6789;
+        c1->last_hit_mem_pressure_limit = 42;
+        assert_se(h2 = hashmap_new(&oomd_cgroup_ctx_hash_ops));
+        assert_se(oomd_insert_cgroup_context(h1, h2, cgroup) == 0);
+        c1 = hashmap_get(h1, cgroup);
+        c2 = hashmap_get(h2, cgroup);
+        assert_se(c1);
+        assert_se(c2);
+        assert_se(c1 != c2);
+        assert_se(c2->last_pgscan == 5555);
+        assert_se(c2->mem_pressure_limit == 6789);
+        assert_se(c2->last_hit_mem_pressure_limit == 42);
+}
+
+static void test_oomd_system_context_acquire(void) {
+        _cleanup_(unlink_tempfilep) char path[] = "/oomdgetsysctxtestXXXXXX";
+        OomdSystemContext ctx;
+
+        if (geteuid() != 0)
+                return (void) log_tests_skipped("not root");
+
+        assert_se(mkstemp(path));
+
+        assert_se(oomd_system_context_acquire("/verylikelynonexistentpath", &ctx) == -ENOENT);
+
+        assert_se(oomd_system_context_acquire(path, &ctx) == 0);
+        assert_se(ctx.swap_total == 0);
+        assert_se(ctx.swap_used == 0);
+
+        assert_se(write_string_file(path, "some\nwords\nacross\nmultiple\nlines", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(oomd_system_context_acquire(path, &ctx) == 0);
+        assert_se(ctx.swap_total == 0);
+        assert_se(ctx.swap_used == 0);
+
+        assert_se(write_string_file(path, "Filename                                Type            Size    Used    Priority\n"
+                                          "/swapvol/swapfile                       file            18971644        0       -3\n"
+                                          "/dev/vda2                               partition       1999868 993780  -2", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(oomd_system_context_acquire(path, &ctx) == 0);
+        assert_se(ctx.swap_total == 21474828288);
+        assert_se(ctx.swap_used == 1017630720);
+}
+
+static void test_oomd_pressure_above(void) {
+        _cleanup_hashmap_free_ Hashmap *h1 = NULL, *h2 = NULL;
+        _cleanup_set_free_ Set *t1 = NULL, *t2 = NULL, *t3 = NULL;
+        OomdCGroupContext ctx[2], *c;
+        loadavg_t threshold;
+
+        assert_se(store_loadavg_fixed_point(80, 0, &threshold) == 0);
+
+        /* /herp.slice */
+        assert_se(store_loadavg_fixed_point(99, 99, &(ctx[0].memory_pressure.avg10)) == 0);
+        assert_se(store_loadavg_fixed_point(99, 99, &(ctx[0].memory_pressure.avg60)) == 0);
+        assert_se(store_loadavg_fixed_point(99, 99, &(ctx[0].memory_pressure.avg300)) == 0);
+        ctx[0].mem_pressure_limit = threshold;
+
+        /* /derp.slice */
+        assert_se(store_loadavg_fixed_point(1, 11, &(ctx[1].memory_pressure.avg10)) == 0);
+        assert_se(store_loadavg_fixed_point(1, 11, &(ctx[1].memory_pressure.avg60)) == 0);
+        assert_se(store_loadavg_fixed_point(1, 11, &(ctx[1].memory_pressure.avg300)) == 0);
+        ctx[1].mem_pressure_limit = threshold;
+
+
+        /* High memory pressure */
+        assert_se(h1 = hashmap_new(&string_hash_ops));
+        assert_se(hashmap_put(h1, "/herp.slice", &ctx[0]) >= 0);
+        assert_se(oomd_pressure_above(h1, 0 /* duration */, &t1) == 1);
+        assert_se(set_contains(t1, &ctx[0]) == true);
+        assert_se(c = hashmap_get(h1, "/herp.slice"));
+        assert_se(c->last_hit_mem_pressure_limit > 0);
+
+        /* Low memory pressure */
+        assert_se(h2 = hashmap_new(&string_hash_ops));
+        assert_se(hashmap_put(h2, "/derp.slice", &ctx[1]) >= 0);
+        assert_se(oomd_pressure_above(h2, 0 /* duration */, &t2) == 0);
+        assert_se(t2 == NULL);
+        assert_se(c = hashmap_get(h2, "/derp.slice"));
+        assert_se(c->last_hit_mem_pressure_limit == 0);
+
+        /* High memory pressure w/ multiple cgroups */
+        assert_se(hashmap_put(h1, "/derp.slice", &ctx[1]) >= 0);
+        assert_se(oomd_pressure_above(h1, 0 /* duration */, &t3) == 1);
+        assert_se(set_contains(t3, &ctx[0]) == true);
+        assert_se(set_size(t3) == 1);
+        assert_se(c = hashmap_get(h1, "/herp.slice"));
+        assert_se(c->last_hit_mem_pressure_limit > 0);
+        assert_se(c = hashmap_get(h1, "/derp.slice"));
+        assert_se(c->last_hit_mem_pressure_limit == 0);
+}
+
+static void test_oomd_memory_reclaim(void) {
+        _cleanup_hashmap_free_ Hashmap *h1 = NULL;
+        char **paths = STRV_MAKE("/0.slice",
+                                 "/1.slice",
+                                 "/2.slice",
+                                 "/3.slice",
+                                 "/4.slice");
+
+        OomdCGroupContext ctx[5] = {
+                { .path = paths[0],
+                  .last_pgscan = 100,
+                  .pgscan = 100 },
+                { .path = paths[1],
+                  .last_pgscan = 100,
+                  .pgscan = 100 },
+                { .path = paths[2],
+                  .last_pgscan = 77,
+                  .pgscan = 33 },
+                { .path = paths[3],
+                  .last_pgscan = UINT64_MAX,
+                  .pgscan = 100 },
+                { .path = paths[4],
+                  .last_pgscan = 100,
+                  .pgscan = UINT64_MAX },
+        };
+
+        assert_se(h1 = hashmap_new(&string_hash_ops));
+        assert_se(hashmap_put(h1, paths[0], &ctx[0]) >= 0);
+        assert_se(hashmap_put(h1, paths[1], &ctx[1]) >= 0);
+        assert_se(oomd_memory_reclaim(h1) == false);
+
+        assert_se(hashmap_put(h1, paths[2], &ctx[2]) >= 0);
+        assert_se(oomd_memory_reclaim(h1) == false);
+
+        assert_se(hashmap_put(h1, paths[4], &ctx[4]) >= 0);
+        assert_se(oomd_memory_reclaim(h1) == true);
+
+        assert_se(hashmap_put(h1, paths[3], &ctx[3]) >= 0);
+        assert_se(oomd_memory_reclaim(h1) == false);
+}
+
+static void test_oomd_swap_free_below(void) {
+        OomdSystemContext ctx = (OomdSystemContext) {
+                .swap_total = 20971512 * 1024U,
+                .swap_used = 20971440 * 1024U,
+        };
+        assert_se(oomd_swap_free_below(&ctx, 20) == true);
+
+        ctx = (OomdSystemContext) {
+                .swap_total = 20971512 * 1024U,
+                .swap_used = 3310136 * 1024U,
+        };
+        assert_se(oomd_swap_free_below(&ctx, 20) == false);
+}
+
+static void test_oomd_sort_cgroups(void) {
+        _cleanup_hashmap_free_ Hashmap *h = NULL;
+        _cleanup_free_ OomdCGroupContext **sorted_cgroups;
+        char **paths = STRV_MAKE("/herp.slice",
+                                 "/herp.slice/derp.scope",
+                                 "/herp.slice/derp.scope/sheep.service",
+                                 "/zupa.slice");
+
+        OomdCGroupContext ctx[4] = {
+                { .path = paths[0],
+                  .swap_usage = 20,
+                  .pgscan = 60 },
+                { .path = paths[1],
+                  .swap_usage = 60,
+                  .pgscan = 40 },
+                { .path = paths[2],
+                  .swap_usage = 40,
+                  .pgscan = 20 },
+                { .path = paths[3],
+                  .swap_usage = 10,
+                  .pgscan = 80 },
+        };
+
+        assert_se(h = hashmap_new(&string_hash_ops));
+
+        assert_se(hashmap_put(h, "/herp.slice", &ctx[0]) >= 0);
+        assert_se(hashmap_put(h, "/herp.slice/derp.scope", &ctx[1]) >= 0);
+        assert_se(hashmap_put(h, "/herp.slice/derp.scope/sheep.service", &ctx[2]) >= 0);
+        assert_se(hashmap_put(h, "/zupa.slice", &ctx[3]) >= 0);
+
+        assert_se(oomd_sort_cgroup_contexts(h, compare_swap_usage, NULL, &sorted_cgroups) == 4);
+        assert_se(sorted_cgroups[0] == &ctx[1]);
+        assert_se(sorted_cgroups[1] == &ctx[2]);
+        assert_se(sorted_cgroups[2] == &ctx[0]);
+        assert_se(sorted_cgroups[3] == &ctx[3]);
+        sorted_cgroups = mfree(sorted_cgroups);
+
+        assert_se(oomd_sort_cgroup_contexts(h, compare_pgscan, NULL, &sorted_cgroups) == 4);
+        assert_se(sorted_cgroups[0] == &ctx[3]);
+        assert_se(sorted_cgroups[1] == &ctx[0]);
+        assert_se(sorted_cgroups[2] == &ctx[1]);
+        assert_se(sorted_cgroups[3] == &ctx[2]);
+        sorted_cgroups = mfree(sorted_cgroups);
+
+        assert_se(oomd_sort_cgroup_contexts(h, compare_pgscan, "/herp.slice/derp.scope", &sorted_cgroups) == 2);
+        assert_se(sorted_cgroups[0] == &ctx[1]);
+        assert_se(sorted_cgroups[1] == &ctx[2]);
+        assert_se(sorted_cgroups[2] == 0);
+        assert_se(sorted_cgroups[3] == 0);
+        sorted_cgroups = mfree(sorted_cgroups);
+}
+
+int main(void) {
+        int r;
+
+        test_setup_logging(LOG_DEBUG);
+
+        test_oomd_system_context_acquire();
+        test_oomd_pressure_above();
+        test_oomd_memory_reclaim();
+        test_oomd_swap_free_below();
+        test_oomd_sort_cgroups();
+
+        /* The following tests operate on live cgroups */
+
+        r = enter_cgroup_root(NULL);
+        if (r < 0)
+                return log_tests_skipped_errno(r, "failed to enter a test cgroup scope");
+
+        test_oomd_cgroup_kill();
+        test_oomd_cgroup_context_acquire_and_insert();
+
+        return 0;
+}

src/shared/bus-get-properties.c

@@ -2,6 +2,7 @@
 
 #include "bus-get-properties.h"
 #include "rlimit-util.h"
+#include "stdio-util.h"
 #include "string-util.h"
 
 int bus_property_get_bool(
@@ -54,6 +55,23 @@ int bus_property_get_id128(
                 return sd_bus_message_append_array(reply, 'y', id->bytes, 16);
 }
 
+int bus_property_get_percent(
+                sd_bus *bus,
+                const char *path,
+                const char *interface,
+                const char *property,
+                sd_bus_message *reply,
+                void *userdata,
+                sd_bus_error *error) {
+
+        char pstr[DECIMAL_STR_MAX(int) + 2];
+        int p = *(int*) userdata;
+
+        xsprintf(pstr, "%d%%", p);
+
+        return sd_bus_message_append_basic(reply, 's', pstr);
+}
+
 #if __SIZEOF_SIZE_T__ != 8
 int bus_property_get_size(
                 sd_bus *bus,

src/shared/bus-get-properties.h

@@ -8,6 +8,7 @@
 int bus_property_get_bool(sd_bus *bus, const char *path, const char *interface, const char *property, sd_bus_message *reply, void *userdata, sd_bus_error *error);
 int bus_property_set_bool(sd_bus *bus, const char *path, const char *interface, const char *property, sd_bus_message *value, void *userdata, sd_bus_error *error);
 int bus_property_get_id128(sd_bus *bus, const char *path, const char *interface, const char *property, sd_bus_message *reply, void *userdata, sd_bus_error *error);
+int bus_property_get_percent(sd_bus *bus, const char *path, const char *interface, const char *property, sd_bus_message *reply, void *userdata, sd_bus_error *error);
 
 #define bus_property_get_usec ((sd_bus_property_get_t) NULL)
 #define bus_property_set_usec ((sd_bus_property_set_t) NULL)

src/shared/bus-unit-util.c

@@ -432,7 +432,11 @@ static int bus_append_ip_address_access(sd_bus_message *m, int family, const uni
 static int bus_append_cgroup_property(sd_bus_message *m, const char *field, const char *eq) {
         int r;
 
-        if (STR_IN_SET(field, "DevicePolicy", "Slice"))
+        if (STR_IN_SET(field, "DevicePolicy",
+                              "Slice",
+                              "ManagedOOMSwap",
+                              "ManagedOOMMemoryPressure",
+                              "ManagedOOMMemoryPressureLimitPercent"))
                 return bus_append_string(m, field, eq);
 
         if (STR_IN_SET(field, "CPUAccounting",

src/shared/conf-parser.c

@@ -1243,3 +1243,5 @@ int config_parse_vlanprotocol(const char* unit,
 
         return 0;
 }
+
+DEFINE_CONFIG_PARSE(config_parse_percent, parse_percent, "Failed to parse percent value");

src/shared/conf-parser.h

@@ -147,6 +147,7 @@ CONFIG_PARSER_PROTOTYPE(config_parse_ip_port);
 CONFIG_PARSER_PROTOTYPE(config_parse_mtu);
 CONFIG_PARSER_PROTOTYPE(config_parse_rlimit);
 CONFIG_PARSER_PROTOTYPE(config_parse_vlanprotocol);
+CONFIG_PARSER_PROTOTYPE(config_parse_percent);
 
 typedef enum Disabled {
         DISABLED_CONFIGURATION,

src/shared/meson.build

@@ -189,6 +189,8 @@ shared_sources = files('''
         pkcs11-util.h
         pretty-print.c
         pretty-print.h
+        psi-util.c
+        psi-util.h
         ptyfwd.c
         ptyfwd.h
         pwquality-util.c

src/shared/psi-util.c

@@ -0,0 +1,118 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <stdio.h>
+#include <unistd.h>
+
+#include "alloc-util.h"
+#include "extract-word.h"
+#include "fd-util.h"
+#include "fileio.h"
+#include "parse-util.h"
+#include "psi-util.h"
+#include "string-util.h"
+#include "stat-util.h"
+#include "strv.h"
+
+int read_resource_pressure(const char *path, PressureType type, ResourcePressure *ret) {
+        _cleanup_free_ char *line = NULL;
+        _cleanup_fclose_ FILE *f = NULL;
+        unsigned field_filled = 0;
+        ResourcePressure rp = {};
+        const char *t, *cline;
+        char *word;
+        int r;
+
+        assert(path);
+        assert(IN_SET(type, PRESSURE_TYPE_SOME, PRESSURE_TYPE_FULL));
+        assert(ret);
+
+        if (type == PRESSURE_TYPE_SOME)
+                t = "some";
+        else if (type == PRESSURE_TYPE_FULL)
+                t = "full";
+        else
+                return -EINVAL;
+
+        r = fopen_unlocked(path, "re", &f);
+         if (r < 0)
+                 return r;
+
+        for (;;) {
+                _cleanup_free_ char *l = NULL;
+                char *w;
+
+                r = read_line(f, LONG_LINE_MAX, &l);
+                if (r < 0)
+                        return r;
+                if (r == 0)
+                        break;
+
+                w = first_word(l, t);
+                if (w) {
+                        line = TAKE_PTR(l);
+                        cline = w;
+                        break;
+                }
+        }
+
+        if (!line)
+                return -ENODATA;
+
+        /* extracts either avgX=Y.Z or total=X */
+        while ((r = extract_first_word(&cline, &word, NULL, 0)) > 0) {
+                _cleanup_free_ char *w = word;
+                const char *v;
+
+                if ((v = startswith(w, "avg10="))) {
+                        if (field_filled & (1U << 0))
+                                return -EINVAL;
+
+                        field_filled |= 1U << 0;
+                        r = parse_loadavg_fixed_point(v, &rp.avg10);
+                } else if ((v = startswith(w, "avg60="))) {
+                        if (field_filled & (1U << 1))
+                                return -EINVAL;
+
+                        field_filled |= 1U << 1;
+                        r = parse_loadavg_fixed_point(v, &rp.avg60);
+                } else if ((v = startswith(w, "avg300="))) {
+                        if (field_filled & (1U << 2))
+                                return -EINVAL;
+
+                        field_filled |= 1U << 2;
+                        r = parse_loadavg_fixed_point(v, &rp.avg300);
+                } else if ((v = startswith(w, "total="))) {
+                        if (field_filled & (1U << 3))
+                                return -EINVAL;
+
+                        field_filled |= 1U << 3;
+                        r = safe_atou64(v, &rp.total);
+                } else
+                        continue;
+
+                if (r < 0)
+                        return r;
+        }
+
+        if (r < 0)
+                return r;
+
+        if (field_filled != 15U)
+                return -EINVAL;
+
+        *ret = rp;
+        return 0;
+}
+
+int is_pressure_supported(void) {
+        const char *p;
+
+        FOREACH_STRING(p, "/proc/pressure/cpu", "/proc/pressure/io", "/proc/pressure/memory")
+                if (access(p, F_OK) < 0) {
+                        if (errno == ENOENT)
+                                return 0;
+                        return -errno;
+                }
+
+        return 1;
+}

src/shared/psi-util.h

@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+#pragma once
+
+#include <stdbool.h>
+
+#include "parse-util.h"
+#include "time-util.h"
+
+typedef enum PressureType {
+        PRESSURE_TYPE_SOME,
+        PRESSURE_TYPE_FULL,
+} PressureType;
+
+/* Averages are stored in fixed-point with 11 bit fractions */
+typedef struct ResourcePressure {
+        loadavg_t avg10;
+        loadavg_t avg60;
+        loadavg_t avg300;
+        usec_t total;
+} ResourcePressure;
+
+/** Upstream 4.20+ format
+ *
+ *  some avg10=0.22 avg60=0.17 avg300=1.11 total=58761459
+ *  full avg10=0.23 avg60=0.16 avg300=1.08 total=58464525
+ */
+int read_resource_pressure(const char *path, PressureType type, ResourcePressure *ret);
+
+/* Was the kernel compiled with CONFIG_PSI=y? 1 if yes, 0 if not, negative on error. */
+int is_pressure_supported(void);

src/shared/tests.c

@@ -254,7 +254,7 @@ static int allocate_scope(void) {
         return 0;
 }
 
-int enter_cgroup_subroot(char **ret_cgroup) {
+static int enter_cgroup(char **ret_cgroup, bool enter_subroot) {
         _cleanup_free_ char *cgroup_root = NULL, *cgroup_subroot = NULL;
         CGroupMask supported;
         int r;
@@ -268,7 +268,13 @@ int enter_cgroup_subroot(char **ret_cgroup) {
                 return log_warning_errno(r, "cg_pid_get_path(NULL, 0, ...) failed: %m");
         assert(r >= 0);
 
-        assert_se(asprintf(&cgroup_subroot, "%s/%" PRIx64, cgroup_root, random_u64()) >= 0);
+        if (enter_subroot)
+                assert_se(asprintf(&cgroup_subroot, "%s/%" PRIx64, cgroup_root, random_u64()) >= 0);
+        else {
+                cgroup_subroot = strdup(cgroup_root);
+                assert_se(cgroup_subroot != NULL);
+        }
+
         assert_se(cg_mask_supported(&supported) >= 0);
 
         /* If this fails, then we don't mind as the later cgroup operations will fail too, and it's fine if
@@ -287,3 +293,11 @@ int enter_cgroup_subroot(char **ret_cgroup) {
 
         return 0;
 }
+
+int enter_cgroup_subroot(char **ret_cgroup) {
+        return enter_cgroup(ret_cgroup, true);
+}
+
+int enter_cgroup_root(char **ret_cgroup) {
+        return enter_cgroup(ret_cgroup, false);
+}

src/shared/tests.h

@@ -20,6 +20,7 @@ static inline bool manager_errno_skip_test(int r) {
 
 char* setup_fake_runtime_dir(void);
 int enter_cgroup_subroot(char **ret_cgroup);
+int enter_cgroup_root(char **ret_cgroup);
 int get_testdata_dir(const char *suffix, char **ret);
 const char* get_catalog_dir(void);
 bool slow_tests_enabled(void);

src/shared/varlink.c

@@ -418,6 +418,11 @@ static int varlink_test_disconnect(Varlink *v) {
         if (IN_SET(v->state, VARLINK_IDLE_CLIENT) && (v->write_disconnected || v->got_pollhup))
                 goto disconnect;
 
+        /* The server is still expecting to write more, but its write end is disconnected and it got a POLLHUP
+         * (i.e. from a disconnected client), so disconnect. */
+        if (IN_SET(v->state, VARLINK_PENDING_METHOD, VARLINK_PENDING_METHOD_MORE) && v->write_disconnected && v->got_pollhup)
+                goto disconnect;
+
         return 0;
 
 disconnect:

src/shared/varlink.h

@@ -171,3 +171,4 @@ DEFINE_TRIVIAL_CLEANUP_FUNC(VarlinkServer *, varlink_server_unref);
 #define VARLINK_ERROR_METHOD_NOT_FOUND "org.varlink.service.MethodNotFound"
 #define VARLINK_ERROR_METHOD_NOT_IMPLEMENTED "org.varlink.service.MethodNotImplemented"
 #define VARLINK_ERROR_INVALID_PARAMETER "org.varlink.service.InvalidParameter"
+#define VARLINK_ERROR_SUBSCRIPTION_TAKEN "org.varlink.service.SubscriptionTaken"

src/systemd/sd-messages.h

@@ -127,6 +127,9 @@ _SD_BEGIN_DECLARATIONS;
 #define SD_MESSAGE_OVERMOUNTING           SD_ID128_MAKE(1d,ee,03,69,c7,fc,47,36,b7,09,9b,38,ec,b4,6e,e7)
 #define SD_MESSAGE_OVERMOUNTING_STR       SD_ID128_MAKE_STR(1d,ee,03,69,c7,fc,47,36,b7,09,9b,38,ec,b4,6e,e7)
 
+#define SD_MESSAGE_UNIT_OOMD_KILL         SD_ID128_MAKE(d9,89,61,1b,15,e4,4c,9d,bf,31,e3,c8,12,56,e4,ed)
+#define SD_MESSAGE_UNIT_OOMD_KILL_STR     SD_ID128_MAKE_STR(d9,89,61,1b,15,e4,4c,9d,bf,31,e3,c8,12,56,e4,ed)
+
 #define SD_MESSAGE_UNIT_OUT_OF_MEMORY     SD_ID128_MAKE(fe,6f,aa,94,e7,77,46,63,a0,da,52,71,78,91,d8,ef)
 #define SD_MESSAGE_UNIT_OUT_OF_MEMORY_STR SD_ID128_MAKE_STR(fe,6f,aa,94,e7,77,46,63,a0,da,52,71,78,91,d8,ef)
 

src/test/meson.build

@@ -804,6 +804,10 @@ tests += [
         [['src/test/test-local-addresses.c'],
          [],
          []],
+
+        [['src/test/test-psi-util.c'],
+         [],
+         []],
 ]
 
 ############################################################

src/test/test-parse-util.c

@@ -1,6 +1,7 @@
 /* SPDX-License-Identifier: LGPL-2.1+ */
 
 #include <errno.h>
+#include <linux/loadavg.h>
 #include <locale.h>
 #include <math.h>
 #include <sys/socket.h>
@@ -929,6 +930,42 @@ static void test_parse_mtu(void) {
         assert_se(parse_mtu(AF_UNSPEC, "", &mtu) == -EINVAL);
 }
 
+static void test_parse_loadavg_fixed_point(void) {
+        loadavg_t fp;
+
+        assert_se(parse_loadavg_fixed_point("1.23", &fp) == 0);
+        assert_se(LOAD_INT(fp) == 1);
+        assert_se(LOAD_FRAC(fp) == 23);
+
+        assert_se(parse_loadavg_fixed_point("1.80", &fp) == 0);
+        assert_se(LOAD_INT(fp) == 1);
+        assert_se(LOAD_FRAC(fp) == 80);
+
+        assert_se(parse_loadavg_fixed_point("0.07", &fp) == 0);
+        assert_se(LOAD_INT(fp) == 0);
+        assert_se(LOAD_FRAC(fp) == 7);
+
+        assert_se(parse_loadavg_fixed_point("0.00", &fp) == 0);
+        assert_se(LOAD_INT(fp) == 0);
+        assert_se(LOAD_FRAC(fp) == 0);
+
+        assert_se(parse_loadavg_fixed_point("4096.57", &fp) == 0);
+        assert_se(LOAD_INT(fp) == 4096);
+        assert_se(LOAD_FRAC(fp) == 57);
+
+        /* Caps out at 2 digit fracs */
+        assert_se(parse_loadavg_fixed_point("1.100", &fp) == -ERANGE);
+
+        assert_se(parse_loadavg_fixed_point("4096.4096", &fp) == -ERANGE);
+        assert_se(parse_loadavg_fixed_point("-4000.5", &fp) == -ERANGE);
+        assert_se(parse_loadavg_fixed_point("18446744073709551615.5", &fp) == -ERANGE);
+        assert_se(parse_loadavg_fixed_point("foobar", &fp) == -EINVAL);
+        assert_se(parse_loadavg_fixed_point("3333", &fp) == -EINVAL);
+        assert_se(parse_loadavg_fixed_point("1.2.3", &fp) == -EINVAL);
+        assert_se(parse_loadavg_fixed_point(".", &fp) == -EINVAL);
+        assert_se(parse_loadavg_fixed_point("", &fp) == -EINVAL);
+}
+
 int main(int argc, char *argv[]) {
         log_parse_environment();
         log_open();
@@ -955,6 +992,7 @@ int main(int argc, char *argv[]) {
         test_parse_errno();
         test_parse_syscall_and_errno();
         test_parse_mtu();
+        test_parse_loadavg_fixed_point();
 
         return 0;
 }

src/test/test-psi-util.c

@@ -0,0 +1,80 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <linux/loadavg.h>
+
+#include "alloc-util.h"
+#include "fileio.h"
+#include "fs-util.h"
+#include "psi-util.h"
+#include "tests.h"
+
+static void test_read_mem_pressure(void) {
+        _cleanup_(unlink_tempfilep) char path[] = "/tmp/pressurereadtestXXXXXX";
+        ResourcePressure rp;
+
+        if (geteuid() != 0)
+                return (void) log_tests_skipped("not root");
+
+        assert_se(mkstemp(path));
+
+        assert_se(read_resource_pressure("/verylikelynonexistentpath", PRESSURE_TYPE_SOME, &rp) < 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) < 0);
+
+        assert_se(write_string_file(path, "herpdederp\n", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) < 0);
+
+        /* Pressure file with some invalid values*/
+        assert_se(write_string_file(path, "some avg10=0.22=55 avg60=0.17=8 avg300=1.11=00 total=58761459\n"
+                                          "full avg10=0.23=55 avg60=0.16=8 avg300=1.08=00 total=58464525", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) < 0);
+
+        /* Same pressure valid values as below but with duplicate avg60 field */
+        assert_se(write_string_file(path, "some avg10=0.22 avg60=0.17 avg60=0.18 avg300=1.11 total=58761459\n"
+                                          "full avg10=0.23 avg60=0.16 avg300=1.08 total=58464525", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) < 0);
+
+        assert_se(write_string_file(path, "some avg10=0.22 avg60=0.17 avg300=1.11 total=58761459\n"
+                                          "full avg10=0.23 avg60=0.16 avg300=1.08 total=58464525", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) == 0);
+        assert_se(LOAD_INT(rp.avg10) == 0);
+        assert_se(LOAD_FRAC(rp.avg10) == 22);
+        assert_se(LOAD_INT(rp.avg60) == 0);
+        assert_se(LOAD_FRAC(rp.avg60) == 17);
+        assert_se(LOAD_INT(rp.avg300) == 1);
+        assert_se(LOAD_FRAC(rp.avg300) == 11);
+        assert_se(rp.total == 58761459);
+        assert(read_resource_pressure(path, PRESSURE_TYPE_FULL, &rp) == 0);
+        assert_se(LOAD_INT(rp.avg10) == 0);
+        assert_se(LOAD_FRAC(rp.avg10) == 23);
+        assert_se(LOAD_INT(rp.avg60) == 0);
+        assert_se(LOAD_FRAC(rp.avg60) == 16);
+        assert_se(LOAD_INT(rp.avg300) == 1);
+        assert_se(LOAD_FRAC(rp.avg300) == 8);
+        assert_se(rp.total == 58464525);
+
+        /* Pressure file with extra unsupported fields */
+        assert_se(write_string_file(path, "some avg5=0.55 avg10=0.22 avg60=0.17 avg300=1.11 total=58761459\n"
+                                          "full avg10=0.23 avg60=0.16 avg300=1.08 avg600=2.00 total=58464525", WRITE_STRING_FILE_CREATE) == 0);
+        assert_se(read_resource_pressure(path, PRESSURE_TYPE_SOME, &rp) == 0);
+        assert_se(LOAD_INT(rp.avg10) == 0);
+        assert_se(LOAD_FRAC(rp.avg10) == 22);
+        assert_se(LOAD_INT(rp.avg60) == 0);
+        assert_se(LOAD_FRAC(rp.avg60) == 17);
+        assert_se(LOAD_INT(rp.avg300) == 1);
+        assert_se(LOAD_FRAC(rp.avg300) == 11);
+        assert_se(rp.total == 58761459);
+        assert(read_resource_pressure(path, PRESSURE_TYPE_FULL, &rp) == 0);
+        assert_se(LOAD_INT(rp.avg10) == 0);
+        assert_se(LOAD_FRAC(rp.avg10) == 23);
+        assert_se(LOAD_INT(rp.avg60) == 0);
+        assert_se(LOAD_FRAC(rp.avg60) == 16);
+        assert_se(LOAD_INT(rp.avg300) == 1);
+        assert_se(LOAD_FRAC(rp.avg300) == 8);
+        assert_se(rp.total == 58464525);
+}
+
+int main(void) {
+        test_setup_logging(LOG_DEBUG);
+        test_read_mem_pressure();
+        return 0;
+}

src/test/test-tables.c

@@ -3,6 +3,7 @@
 #include "architecture.h"
 #include "automount.h"
 #include "cgroup.h"
+#include "cgroup-util.h"
 #include "compress.h"
 #include "condition.h"
 #include "device-private.h"
@@ -71,6 +72,7 @@ int main(int argc, char **argv) {
         test_table(locale_variable, VARIABLE_LC);
         test_table(log_target, LOG_TARGET);
         test_table(mac_address_policy, MAC_ADDRESS_POLICY);
+        test_table(managed_oom_mode, MANAGED_OOM_MODE);
         test_table(manager_state, MANAGER_STATE);
         test_table(manager_timestamp, MANAGER_TIMESTAMP);
         test_table(mount_exec_command, MOUNT_EXEC_COMMAND);