struct kwork_work *key,
struct list_head *sort_list)
{
- struct kwork_work *work = NULL;
+ struct kwork_work *work = work_search(root, key, sort_list);
- work = work_search(root, key, sort_list);
if (work != NULL)
return work;
work = work_new(key);
- if (work == NULL)
- return NULL;
+ if (work)
+ work_insert(root, work, sort_list);
- work_insert(root, work, sort_list);
return work;
}
}
}
+struct kwork_work *perf_kwork_add_work(struct perf_kwork *kwork,
+ struct kwork_class *class,
+ struct kwork_work *key)
+{
+ struct kwork_work *work = NULL;
+
+ work = work_new(key);
+ if (work == NULL)
+ return NULL;
+
+ work_insert(&class->work_root, work, &kwork->cmp_id);
+ return work;
+}
+
+static void sig_handler(int sig)
+{
+ /*
+ * Simply capture termination signal so that
+ * the program can continue after pause returns
+ */
+ pr_debug("Captuer signal %d\n", sig);
+}
+
+static int perf_kwork__report_bpf(struct perf_kwork *kwork)
+{
+ int ret;
+
+ signal(SIGINT, sig_handler);
+ signal(SIGTERM, sig_handler);
+
+ ret = perf_kwork__trace_prepare_bpf(kwork);
+ if (ret)
+ return -1;
+
+ printf("Starting trace, Hit <Ctrl+C> to stop and report\n");
+
+ perf_kwork__trace_start();
+
+ /*
+ * a simple pause, wait here for stop signal
+ */
+ pause();
+
+ perf_kwork__trace_finish();
+
+ perf_kwork__report_read_bpf(kwork);
+
+ perf_kwork__report_cleanup_bpf();
+
+ return 0;
+}
+
static int perf_kwork__report(struct perf_kwork *kwork)
{
int ret;
struct rb_node *next;
struct kwork_work *work;
- ret = perf_kwork__read_events(kwork);
+ if (kwork->use_bpf)
+ ret = perf_kwork__report_bpf(kwork);
+ else
+ ret = perf_kwork__read_events(kwork);
+
if (ret != 0)
return -1;
"input file name"),
OPT_BOOLEAN('S', "with-summary", &kwork.summary,
"Show summary with statistics"),
+#ifdef HAVE_BPF_SKEL
+ OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf,
+ "Use BPF to measure kwork runtime"),
+#endif
OPT_PARENT(kwork_options)
};
const struct option latency_options[] = {
"Time span for analysis (start,stop)"),
OPT_STRING('i', "input", &input_name, "file",
"input file name"),
+#ifdef HAVE_BPF_SKEL
+ OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf,
+ "Use BPF to measure kwork latency"),
+#endif
OPT_PARENT(kwork_options)
};
const struct option timehist_options[] = {
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * bpf_kwork.c
+ *
+ * Copyright (c) 2022 Huawei Inc, Yang Jihong <yangjihong1@huawei.com>
+ */
+
+#include <time.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <unistd.h>
+
+#include <linux/time64.h>
+
+#include "util/debug.h"
+#include "util/kwork.h"
+
+#include <bpf/bpf.h>
+
+#include "util/bpf_skel/kwork_trace.skel.h"
+
+/*
+ * This should be in sync with "util/kwork_trace.bpf.c"
+ */
+#define MAX_KWORKNAME 128
+
+struct work_key {
+ u32 type;
+ u32 cpu;
+ u64 id;
+};
+
+struct report_data {
+ u64 nr;
+ u64 total_time;
+ u64 max_time;
+ u64 max_time_start;
+ u64 max_time_end;
+};
+
+struct kwork_class_bpf {
+ struct kwork_class *class;
+
+ void (*load_prepare)(struct perf_kwork *kwork);
+ int (*get_work_name)(struct work_key *key, char **ret_name);
+};
+
+static struct kwork_trace_bpf *skel;
+
+static struct timespec ts_start;
+static struct timespec ts_end;
+
+void perf_kwork__trace_start(void)
+{
+ clock_gettime(CLOCK_MONOTONIC, &ts_start);
+ skel->bss->enabled = 1;
+}
+
+void perf_kwork__trace_finish(void)
+{
+ clock_gettime(CLOCK_MONOTONIC, &ts_end);
+ skel->bss->enabled = 0;
+}
+
+static struct kwork_class_bpf *
+kwork_class_bpf_supported_list[KWORK_CLASS_MAX] = {
+ [KWORK_CLASS_IRQ] = NULL,
+ [KWORK_CLASS_SOFTIRQ] = NULL,
+ [KWORK_CLASS_WORKQUEUE] = NULL,
+};
+
+static bool valid_kwork_class_type(enum kwork_class_type type)
+{
+ return type >= 0 && type < KWORK_CLASS_MAX ? true : false;
+}
+
+static int setup_filters(struct perf_kwork *kwork)
+{
+ u8 val = 1;
+ int i, nr_cpus, key, fd;
+ struct perf_cpu_map *map;
+
+ if (kwork->cpu_list != NULL) {
+ fd = bpf_map__fd(skel->maps.perf_kwork_cpu_filter);
+ if (fd < 0) {
+ pr_debug("Invalid cpu filter fd\n");
+ return -1;
+ }
+
+ map = perf_cpu_map__new(kwork->cpu_list);
+ if (map == NULL) {
+ pr_debug("Invalid cpu_list\n");
+ return -1;
+ }
+
+ nr_cpus = libbpf_num_possible_cpus();
+ for (i = 0; i < perf_cpu_map__nr(map); i++) {
+ struct perf_cpu cpu = perf_cpu_map__cpu(map, i);
+
+ if (cpu.cpu >= nr_cpus) {
+ perf_cpu_map__put(map);
+ pr_err("Requested cpu %d too large\n", cpu.cpu);
+ return -1;
+ }
+ bpf_map_update_elem(fd, &cpu.cpu, &val, BPF_ANY);
+ }
+ perf_cpu_map__put(map);
+
+ skel->bss->has_cpu_filter = 1;
+ }
+
+ if (kwork->profile_name != NULL) {
+ if (strlen(kwork->profile_name) >= MAX_KWORKNAME) {
+ pr_err("Requested name filter %s too large, limit to %d\n",
+ kwork->profile_name, MAX_KWORKNAME - 1);
+ return -1;
+ }
+
+ fd = bpf_map__fd(skel->maps.perf_kwork_name_filter);
+ if (fd < 0) {
+ pr_debug("Invalid name filter fd\n");
+ return -1;
+ }
+
+ key = 0;
+ bpf_map_update_elem(fd, &key, kwork->profile_name, BPF_ANY);
+
+ skel->bss->has_name_filter = 1;
+ }
+
+ return 0;
+}
+
+int perf_kwork__trace_prepare_bpf(struct perf_kwork *kwork)
+{
+ struct bpf_program *prog;
+ struct kwork_class *class;
+ struct kwork_class_bpf *class_bpf;
+ enum kwork_class_type type;
+
+ skel = kwork_trace_bpf__open();
+ if (!skel) {
+ pr_debug("Failed to open kwork trace skeleton\n");
+ return -1;
+ }
+
+ /*
+ * set all progs to non-autoload,
+ * then set corresponding progs according to config
+ */
+ bpf_object__for_each_program(prog, skel->obj)
+ bpf_program__set_autoload(prog, false);
+
+ list_for_each_entry(class, &kwork->class_list, list) {
+ type = class->type;
+ if (!valid_kwork_class_type(type) ||
+ (kwork_class_bpf_supported_list[type] == NULL)) {
+ pr_err("Unsupported bpf trace class %s\n", class->name);
+ goto out;
+ }
+
+ class_bpf = kwork_class_bpf_supported_list[type];
+ class_bpf->class = class;
+
+ if (class_bpf->load_prepare != NULL)
+ class_bpf->load_prepare(kwork);
+ }
+
+ if (kwork_trace_bpf__load(skel)) {
+ pr_debug("Failed to load kwork trace skeleton\n");
+ goto out;
+ }
+
+ if (setup_filters(kwork))
+ goto out;
+
+ if (kwork_trace_bpf__attach(skel)) {
+ pr_debug("Failed to attach kwork trace skeleton\n");
+ goto out;
+ }
+
+ return 0;
+
+out:
+ kwork_trace_bpf__destroy(skel);
+ return -1;
+}
+
+static int add_work(struct perf_kwork *kwork,
+ struct work_key *key,
+ struct report_data *data)
+{
+ struct kwork_work *work;
+ struct kwork_class_bpf *bpf_trace;
+ struct kwork_work tmp = {
+ .id = key->id,
+ .name = NULL,
+ .cpu = key->cpu,
+ };
+ enum kwork_class_type type = key->type;
+
+ if (!valid_kwork_class_type(type)) {
+ pr_debug("Invalid class type %d to add work\n", type);
+ return -1;
+ }
+
+ bpf_trace = kwork_class_bpf_supported_list[type];
+ tmp.class = bpf_trace->class;
+
+ if ((bpf_trace->get_work_name != NULL) &&
+ (bpf_trace->get_work_name(key, &tmp.name)))
+ return -1;
+
+ work = perf_kwork_add_work(kwork, tmp.class, &tmp);
+ if (work == NULL)
+ return -1;
+
+ if (kwork->report == KWORK_REPORT_RUNTIME) {
+ work->nr_atoms = data->nr;
+ work->total_runtime = data->total_time;
+ work->max_runtime = data->max_time;
+ work->max_runtime_start = data->max_time_start;
+ work->max_runtime_end = data->max_time_end;
+ } else if (kwork->report == KWORK_REPORT_LATENCY) {
+ work->nr_atoms = data->nr;
+ work->total_latency = data->total_time;
+ work->max_latency = data->max_time;
+ work->max_latency_start = data->max_time_start;
+ work->max_latency_end = data->max_time_end;
+ } else {
+ pr_debug("Invalid bpf report type %d\n", kwork->report);
+ return -1;
+ }
+
+ kwork->timestart = (u64)ts_start.tv_sec * NSEC_PER_SEC + ts_start.tv_nsec;
+ kwork->timeend = (u64)ts_end.tv_sec * NSEC_PER_SEC + ts_end.tv_nsec;
+
+ return 0;
+}
+
+int perf_kwork__report_read_bpf(struct perf_kwork *kwork)
+{
+ struct report_data data;
+ struct work_key key = {
+ .type = 0,
+ .cpu = 0,
+ .id = 0,
+ };
+ struct work_key prev = {
+ .type = 0,
+ .cpu = 0,
+ .id = 0,
+ };
+ int fd = bpf_map__fd(skel->maps.perf_kwork_report);
+
+ if (fd < 0) {
+ pr_debug("Invalid report fd\n");
+ return -1;
+ }
+
+ while (!bpf_map_get_next_key(fd, &prev, &key)) {
+ if ((bpf_map_lookup_elem(fd, &key, &data)) != 0) {
+ pr_debug("Failed to lookup report elem\n");
+ return -1;
+ }
+
+ if ((data.nr != 0) && (add_work(kwork, &key, &data) != 0))
+ return -1;
+
+ prev = key;
+ }
+ return 0;
+}
+
+void perf_kwork__report_cleanup_bpf(void)
+{
+ kwork_trace_bpf__destroy(skel);
+}
--- /dev/null
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+// Copyright (c) 2022, Huawei
+
+#include "vmlinux.h"
+#include <bpf/bpf_helpers.h>
+#include <bpf/bpf_tracing.h>
+
+#define KWORK_COUNT 100
+#define MAX_KWORKNAME 128
+
+/*
+ * This should be in sync with "util/kwork.h"
+ */
+enum kwork_class_type {
+ KWORK_CLASS_IRQ,
+ KWORK_CLASS_SOFTIRQ,
+ KWORK_CLASS_WORKQUEUE,
+ KWORK_CLASS_MAX,
+};
+
+struct work_key {
+ __u32 type;
+ __u32 cpu;
+ __u64 id;
+};
+
+struct report_data {
+ __u64 nr;
+ __u64 total_time;
+ __u64 max_time;
+ __u64 max_time_start;
+ __u64 max_time_end;
+};
+
+struct {
+ __uint(type, BPF_MAP_TYPE_HASH);
+ __uint(key_size, sizeof(struct work_key));
+ __uint(value_size, MAX_KWORKNAME);
+ __uint(max_entries, KWORK_COUNT);
+} perf_kwork_names SEC(".maps");
+
+struct {
+ __uint(type, BPF_MAP_TYPE_HASH);
+ __uint(key_size, sizeof(struct work_key));
+ __uint(value_size, sizeof(__u64));
+ __uint(max_entries, KWORK_COUNT);
+} perf_kwork_time SEC(".maps");
+
+struct {
+ __uint(type, BPF_MAP_TYPE_HASH);
+ __uint(key_size, sizeof(struct work_key));
+ __uint(value_size, sizeof(struct report_data));
+ __uint(max_entries, KWORK_COUNT);
+} perf_kwork_report SEC(".maps");
+
+struct {
+ __uint(type, BPF_MAP_TYPE_HASH);
+ __uint(key_size, sizeof(__u32));
+ __uint(value_size, sizeof(__u8));
+ __uint(max_entries, 1);
+} perf_kwork_cpu_filter SEC(".maps");
+
+struct {
+ __uint(type, BPF_MAP_TYPE_ARRAY);
+ __uint(key_size, sizeof(__u32));
+ __uint(value_size, MAX_KWORKNAME);
+ __uint(max_entries, 1);
+} perf_kwork_name_filter SEC(".maps");
+
+int enabled = 0;
+int has_cpu_filter = 0;
+int has_name_filter = 0;
+
+char LICENSE[] SEC("license") = "Dual BSD/GPL";