--- /dev/null
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include "cgroup-util.h"
+#include "fd-util.h"
+#include "fileio.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);
+
+ 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;
+}
+
+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 = acquire_managed_oom_connect(m);
+ if (r < 0)
+ return r;
+
+ r = monitor_cgroup_contexts(m);
+ if (r < 0)
+ return r;
+
+ return 0;
+}
--- /dev/null
+/* 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);
projects / thirdparty / systemd.git / commitdiff
