netdata/collectors/ebpf.plugin/ebpf_oomkill.c

// SPDX-License-Identifier: GPL-3.0-or-later

#include "ebpf.h"
#include "ebpf_oomkill.h"

struct config oomkill_config = { .first_section = NULL,
    .last_section = NULL,
    .mutex = NETDATA_MUTEX_INITIALIZER,
    .index = { .avl_tree = { .root = NULL, .compar = appconfig_section_compare },
        .rwlock = AVL_LOCK_INITIALIZER } };

#define OOMKILL_MAP_KILLCNT 0
static ebpf_local_maps_t oomkill_maps[] = {
    {
        .name = "tbl_oomkill",
        .internal_input = NETDATA_OOMKILL_MAX_ENTRIES,
        .user_input = 0,
        .type = NETDATA_EBPF_MAP_STATIC,
        .map_fd = ND_EBPF_MAP_FD_NOT_INITIALIZED,
#ifdef LIBBPF_MAJOR_VERSION
        .map_type = BPF_MAP_TYPE_PERCPU_HASH
#endif
    },
    /* end */
    {
        .name = NULL,
        .internal_input = 0,
        .user_input = 0,
        .type = NETDATA_EBPF_MAP_CONTROLLER,
        .map_fd = ND_EBPF_MAP_FD_NOT_INITIALIZED,
#ifdef LIBBPF_MAJOR_VERSION
        .map_type = BPF_MAP_TYPE_PERCPU_HASH
#endif
    }
};

static ebpf_tracepoint_t oomkill_tracepoints[] = {
    {.enabled = false, .class = "oom", .event = "mark_victim"},
    /* end */
    {.enabled = false, .class = NULL, .event = NULL}
};

static netdata_publish_syscall_t oomkill_publish_aggregated = {.name = "oomkill", .dimension = "oomkill",
                                                               .algorithm = "absolute",
                                                               .next = NULL};

/**
 * Clean up the main thread.
 *
 * @param ptr thread data.
 */
static void oomkill_cleanup(void *ptr)
{
    ebpf_module_t *em = (ebpf_module_t *)ptr;

    if (em->objects)
        ebpf_unload_legacy_code(em->objects, em->probe_links);

    pthread_mutex_lock(&ebpf_exit_cleanup);
    em->enabled = NETDATA_THREAD_EBPF_STOPPED;
    pthread_mutex_unlock(&ebpf_exit_cleanup);
}

static void oomkill_write_data(int32_t *keys, uint32_t total)
{
    // for each app, see if it was OOM killed. record as 1 if so otherwise 0.
    struct ebpf_target *w;
    for (w = apps_groups_root_target; w != NULL; w = w->next) {
        if (likely(w->exposed && w->processes)) {
            bool was_oomkilled = false;
            struct ebpf_pid_on_target *pids = w->root_pid;
            while (pids) {
                uint32_t j;
                for (j = 0; j < total; j++) {
                    if (pids->pid == keys[j]) {
                        was_oomkilled = true;
                        // set to 0 so we consider it "done".
                        keys[j] = 0;
                        goto write_dim;
                    }
                }
                pids = pids->next;
            }

        write_dim:;
            write_chart_dimension(w->name, was_oomkilled);
        }
    }

    // for any remaining keys for which we couldn't find a group, this could be
    // for various reasons, but the primary one is that the PID has not yet
    // been picked up by the process thread when parsing the proc filesystem.
    // since it's been OOM killed, it will never be parsed in the future, so
    // we have no choice but to dump it into `other`.
    uint32_t j;
    uint32_t rem_count = 0;
    for (j = 0; j < total; j++) {
        int32_t key = keys[j];
        if (key != 0) {
            rem_count += 1;
        }
    }
    if (rem_count > 0) {
        write_chart_dimension("other", rem_count);
    }
}

/**
 * Create specific OOMkill charts
 *
 * Create charts for cgroup/application.
 *
 * @param type the chart type.
 * @param update_every value to overwrite the update frequency set by the server.
 */
static void ebpf_create_specific_oomkill_charts(char *type, int update_every)
{
    ebpf_create_chart(type, NETDATA_OOMKILL_CHART, "OOM kills. This chart is provided by eBPF plugin.",
                      EBPF_COMMON_DIMENSION_KILLS, NETDATA_EBPF_MEMORY_GROUP,
                      NETDATA_CGROUP_OOMKILLS_CONTEXT, NETDATA_EBPF_CHART_TYPE_LINE,
                      NETDATA_CHART_PRIO_CGROUPS_CONTAINERS + 5600,
                      ebpf_create_global_dimension,
                      &oomkill_publish_aggregated, 1, update_every, NETDATA_EBPF_MODULE_NAME_OOMKILL);
}

/**
 *  Create Systemd OOMkill Charts
 *
 *  Create charts when systemd is enabled
 *
 *  @param update_every value to overwrite the update frequency set by the server.
 **/
static void ebpf_create_systemd_oomkill_charts(int update_every)
{
    ebpf_create_charts_on_systemd(NETDATA_OOMKILL_CHART, "OOM kills. This chart is provided by eBPF plugin.",
                                  EBPF_COMMON_DIMENSION_KILLS, NETDATA_EBPF_MEMORY_GROUP,
                                  NETDATA_EBPF_CHART_TYPE_LINE, 20191,
                                  ebpf_algorithms[NETDATA_EBPF_INCREMENTAL_IDX], NULL,
                                  NETDATA_EBPF_MODULE_NAME_OOMKILL, update_every);
}

/**
 * Send Systemd charts
 *
 * Send collected data to Netdata.
 */
static void ebpf_send_systemd_oomkill_charts()
{
    ebpf_cgroup_target_t *ect;
    write_begin_chart(NETDATA_SERVICE_FAMILY, NETDATA_OOMKILL_CHART);
    for (ect = ebpf_cgroup_pids; ect ; ect = ect->next) {
        if (unlikely(ect->systemd) && unlikely(ect->updated)) {
            write_chart_dimension(ect->name, (long long) ect->oomkill);
            ect->oomkill = 0;
        }
    }
    write_end_chart();
}

/*
 * Send Specific OOMkill data
 *
 * Send data for specific cgroup/apps.
 *
 * @param type   chart type
 * @param value  value for oomkill
 */
static void ebpf_send_specific_oomkill_data(char *type, int value)
{
    write_begin_chart(type, NETDATA_OOMKILL_CHART);
    write_chart_dimension(oomkill_publish_aggregated.name, (long long)value);
    write_end_chart();
}

/**
 * Create specific OOMkill charts
 *
 * Create charts for cgroup/application.
 *
 * @param type the chart type.
 * @param update_every value to overwrite the update frequency set by the server.
 */
static void ebpf_obsolete_specific_oomkill_charts(char *type, int update_every)
{
    ebpf_write_chart_obsolete(type, NETDATA_OOMKILL_CHART, "OOM kills. This chart is provided by eBPF plugin.",
                              EBPF_COMMON_DIMENSION_KILLS, NETDATA_EBPF_MEMORY_GROUP,
                              NETDATA_EBPF_CHART_TYPE_LINE, NETDATA_CGROUP_OOMKILLS_CONTEXT,
                              NETDATA_CHART_PRIO_CGROUPS_CONTAINERS + 5600, update_every);
}

/**
 * Send data to Netdata calling auxiliary functions.
 *
 * @param update_every value to overwrite the update frequency set by the server.
*/
void ebpf_oomkill_send_cgroup_data(int update_every)
{
    if (!ebpf_cgroup_pids)
        return;

    pthread_mutex_lock(&mutex_cgroup_shm);
    ebpf_cgroup_target_t *ect;

    int has_systemd = shm_ebpf_cgroup.header->systemd_enabled;
    if (has_systemd) {
        if (send_cgroup_chart) {
            ebpf_create_systemd_oomkill_charts(update_every);
        }
        ebpf_send_systemd_oomkill_charts();
    }

    for (ect = ebpf_cgroup_pids; ect ; ect = ect->next) {
        if (ect->systemd)
            continue;

        if (!(ect->flags & NETDATA_EBPF_CGROUP_HAS_OOMKILL_CHART) && ect->updated) {
            ebpf_create_specific_oomkill_charts(ect->name, update_every);
            ect->flags |= NETDATA_EBPF_CGROUP_HAS_OOMKILL_CHART;
        }

        if (ect->flags & NETDATA_EBPF_CGROUP_HAS_OOMKILL_CHART && ect->updated) {
            ebpf_send_specific_oomkill_data(ect->name, ect->oomkill);
        } else {
            ebpf_obsolete_specific_oomkill_charts(ect->name, update_every);
            ect->flags &= ~NETDATA_EBPF_CGROUP_HAS_OOMKILL_CHART;
        }
    }

    pthread_mutex_unlock(&mutex_cgroup_shm);
}

/**
 * Read data
 *
 * Read OOMKILL events from table.
 *
 * @param keys vector where data will be stored
 *
 * @return It returns the number of read elements
 */
static uint32_t oomkill_read_data(int32_t *keys)
{
    // the first `i` entries of `keys` will contain the currently active PIDs
    // in the eBPF map.
    uint32_t i = 0;

    uint32_t curr_key = 0;
    uint32_t key = 0;
    int mapfd = oomkill_maps[OOMKILL_MAP_KILLCNT].map_fd;
    while (bpf_map_get_next_key(mapfd, &curr_key, &key) == 0) {
        curr_key = key;

        keys[i] = (int32_t)key;
        i += 1;

        // delete this key now that we've recorded its existence. there's no
        // race here, as the same PID will only get OOM killed once.
        int test = bpf_map_delete_elem(mapfd, &key);
        if (unlikely(test < 0)) {
            // since there's only 1 thread doing these deletions, it should be
            // impossible to get this condition.
            error("key unexpectedly not available for deletion.");
        }
    }

    return i;
}

/**
 * Update cgroup
 *
 * Update cgroup data based in
 *
 * @param keys  vector with pids that had oomkill event
 * @param total number of elements in keys vector.
 */
static void ebpf_update_oomkill_cgroup(int32_t *keys, uint32_t total)
{
    ebpf_cgroup_target_t *ect;
    for (ect = ebpf_cgroup_pids; ect; ect = ect->next) {
        ect->oomkill = 0;
        struct pid_on_target2 *pids;
        for (pids = ect->pids; pids; pids = pids->next) {
            uint32_t j;
            int32_t pid = pids->pid;
            for (j = 0; j < total; j++) {
                if (pid == keys[j]) {
                    ect->oomkill = 1;
                    break;
                }
            }
        }
    }
}

/**
* Main loop for this collector.
 *
 * @param em the thread main structure.
*/
static void oomkill_collector(ebpf_module_t *em)
{
    int cgroups = em->cgroup_charts;
    int update_every = em->update_every;
    int32_t keys[NETDATA_OOMKILL_MAX_ENTRIES];
    memset(keys, 0, sizeof(keys));

    // loop and read until ebpf plugin is closed.
    heartbeat_t hb;
    heartbeat_init(&hb);
    int counter = update_every - 1;
    while (!ebpf_exit_plugin) {
        (void)heartbeat_next(&hb, USEC_PER_SEC);
        if (ebpf_exit_plugin || ++counter != update_every)
            continue;

        counter = 0;

        uint32_t count = oomkill_read_data(keys);
        if (!count)
            continue;

        pthread_mutex_lock(&collect_data_mutex);
        pthread_mutex_lock(&lock);
        if (cgroups) {
            ebpf_update_oomkill_cgroup(keys, count);
            // write everything from the ebpf map.
            ebpf_oomkill_send_cgroup_data(update_every);
        }

        if (em->apps_charts & NETDATA_EBPF_APPS_FLAG_CHART_CREATED) {
            write_begin_chart(NETDATA_APPS_FAMILY, NETDATA_OOMKILL_CHART);
            oomkill_write_data(keys, count);
            write_end_chart();
        }
        pthread_mutex_unlock(&lock);
        pthread_mutex_unlock(&collect_data_mutex);
    }
}

/**
 * Create apps charts
 *
 * Call ebpf_create_chart to create the charts on apps submenu.
 *
 * @param em a pointer to the structure with the default values.
 */
void ebpf_oomkill_create_apps_charts(struct ebpf_module *em, void *ptr)
{
    struct ebpf_target *root = ptr;
    ebpf_create_charts_on_apps(NETDATA_OOMKILL_CHART,
                               "OOM kills",
                               EBPF_COMMON_DIMENSION_KILLS,
                               "mem",
                               NETDATA_EBPF_CHART_TYPE_STACKED,
                               20020,
                               ebpf_algorithms[NETDATA_EBPF_ABSOLUTE_IDX],
                               root, em->update_every, NETDATA_EBPF_MODULE_NAME_OOMKILL);

    em->apps_charts |= NETDATA_EBPF_APPS_FLAG_CHART_CREATED;
}

/**
 * OOM kill tracking thread.
 *
 * @param ptr a `ebpf_module_t *`.
 * @return always NULL.
 */
void *ebpf_oomkill_thread(void *ptr)
{
    netdata_thread_cleanup_push(oomkill_cleanup, ptr);

    ebpf_module_t *em = (ebpf_module_t *)ptr;
    em->maps = oomkill_maps;

#define NETDATA_DEFAULT_OOM_DISABLED_MSG "Disabling OOMKILL thread, because"
    if (unlikely(!ebpf_all_pids || !em->apps_charts)) {
        // When we are not running integration with apps, we won't fill necessary variables for this thread to run, so
        // we need to disable it.
        pthread_mutex_lock(&ebpf_exit_cleanup);
        if (em->enabled)
            info("%s apps integration is completely disabled.", NETDATA_DEFAULT_OOM_DISABLED_MSG);
        pthread_mutex_unlock(&ebpf_exit_cleanup);

        goto endoomkill;
    } else if (running_on_kernel < NETDATA_EBPF_KERNEL_4_14) {
        pthread_mutex_lock(&ebpf_exit_cleanup);
        if (em->enabled)
            info("%s kernel does not have necessary tracepoints.", NETDATA_DEFAULT_OOM_DISABLED_MSG);
        pthread_mutex_unlock(&ebpf_exit_cleanup);

        goto endoomkill;
    }

    if (ebpf_enable_tracepoints(oomkill_tracepoints) == 0) {
        goto endoomkill;
    }

#ifdef LIBBPF_MAJOR_VERSION
    ebpf_define_map_type(em->maps, em->maps_per_core, running_on_kernel);
#endif
    em->probe_links = ebpf_load_program(ebpf_plugin_dir, em, running_on_kernel, isrh, &em->objects);
    if (!em->probe_links) {
        goto endoomkill;
    }

    pthread_mutex_lock(&lock);
    ebpf_update_stats(&plugin_statistics, em);
    ebpf_update_kernel_memory_with_vector(&plugin_statistics, em->maps);
    pthread_mutex_unlock(&lock);

    oomkill_collector(em);

endoomkill:
    ebpf_update_disabled_plugin_stats(em);

    netdata_thread_cleanup_pop(1);

    return NULL;
}