0001
0002
0003
0004
0005
0006
0007
0008 #include <time.h>
0009 #include <fcntl.h>
0010 #include <stdio.h>
0011 #include <unistd.h>
0012
0013 #include <linux/time64.h>
0014
0015 #include "util/debug.h"
0016 #include "util/kwork.h"
0017
0018 #include <bpf/bpf.h>
0019
0020 #include "util/bpf_skel/kwork_trace.skel.h"
0021
0022
0023
0024
0025 #define MAX_KWORKNAME 128
0026
0027 struct work_key {
0028 u32 type;
0029 u32 cpu;
0030 u64 id;
0031 };
0032
0033 struct report_data {
0034 u64 nr;
0035 u64 total_time;
0036 u64 max_time;
0037 u64 max_time_start;
0038 u64 max_time_end;
0039 };
0040
0041 struct kwork_class_bpf {
0042 struct kwork_class *class;
0043
0044 void (*load_prepare)(struct perf_kwork *kwork);
0045 int (*get_work_name)(struct work_key *key, char **ret_name);
0046 };
0047
0048 static struct kwork_trace_bpf *skel;
0049
0050 static struct timespec ts_start;
0051 static struct timespec ts_end;
0052
0053 void perf_kwork__trace_start(void)
0054 {
0055 clock_gettime(CLOCK_MONOTONIC, &ts_start);
0056 skel->bss->enabled = 1;
0057 }
0058
0059 void perf_kwork__trace_finish(void)
0060 {
0061 clock_gettime(CLOCK_MONOTONIC, &ts_end);
0062 skel->bss->enabled = 0;
0063 }
0064
0065 static int get_work_name_from_map(struct work_key *key, char **ret_name)
0066 {
0067 char name[MAX_KWORKNAME] = { 0 };
0068 int fd = bpf_map__fd(skel->maps.perf_kwork_names);
0069
0070 *ret_name = NULL;
0071
0072 if (fd < 0) {
0073 pr_debug("Invalid names map fd\n");
0074 return 0;
0075 }
0076
0077 if ((bpf_map_lookup_elem(fd, key, name) == 0) && (strlen(name) != 0)) {
0078 *ret_name = strdup(name);
0079 if (*ret_name == NULL) {
0080 pr_err("Failed to copy work name\n");
0081 return -1;
0082 }
0083 }
0084
0085 return 0;
0086 }
0087
0088 static void irq_load_prepare(struct perf_kwork *kwork)
0089 {
0090 if (kwork->report == KWORK_REPORT_RUNTIME) {
0091 bpf_program__set_autoload(skel->progs.report_irq_handler_entry, true);
0092 bpf_program__set_autoload(skel->progs.report_irq_handler_exit, true);
0093 }
0094 }
0095
0096 static struct kwork_class_bpf kwork_irq_bpf = {
0097 .load_prepare = irq_load_prepare,
0098 .get_work_name = get_work_name_from_map,
0099 };
0100
0101 static void softirq_load_prepare(struct perf_kwork *kwork)
0102 {
0103 if (kwork->report == KWORK_REPORT_RUNTIME) {
0104 bpf_program__set_autoload(skel->progs.report_softirq_entry, true);
0105 bpf_program__set_autoload(skel->progs.report_softirq_exit, true);
0106 } else if (kwork->report == KWORK_REPORT_LATENCY) {
0107 bpf_program__set_autoload(skel->progs.latency_softirq_raise, true);
0108 bpf_program__set_autoload(skel->progs.latency_softirq_entry, true);
0109 }
0110 }
0111
0112 static struct kwork_class_bpf kwork_softirq_bpf = {
0113 .load_prepare = softirq_load_prepare,
0114 .get_work_name = get_work_name_from_map,
0115 };
0116
0117 static void workqueue_load_prepare(struct perf_kwork *kwork)
0118 {
0119 if (kwork->report == KWORK_REPORT_RUNTIME) {
0120 bpf_program__set_autoload(skel->progs.report_workqueue_execute_start, true);
0121 bpf_program__set_autoload(skel->progs.report_workqueue_execute_end, true);
0122 } else if (kwork->report == KWORK_REPORT_LATENCY) {
0123 bpf_program__set_autoload(skel->progs.latency_workqueue_activate_work, true);
0124 bpf_program__set_autoload(skel->progs.latency_workqueue_execute_start, true);
0125 }
0126 }
0127
0128 static struct kwork_class_bpf kwork_workqueue_bpf = {
0129 .load_prepare = workqueue_load_prepare,
0130 .get_work_name = get_work_name_from_map,
0131 };
0132
0133 static struct kwork_class_bpf *
0134 kwork_class_bpf_supported_list[KWORK_CLASS_MAX] = {
0135 [KWORK_CLASS_IRQ] = &kwork_irq_bpf,
0136 [KWORK_CLASS_SOFTIRQ] = &kwork_softirq_bpf,
0137 [KWORK_CLASS_WORKQUEUE] = &kwork_workqueue_bpf,
0138 };
0139
0140 static bool valid_kwork_class_type(enum kwork_class_type type)
0141 {
0142 return type >= 0 && type < KWORK_CLASS_MAX ? true : false;
0143 }
0144
0145 static int setup_filters(struct perf_kwork *kwork)
0146 {
0147 u8 val = 1;
0148 int i, nr_cpus, key, fd;
0149 struct perf_cpu_map *map;
0150
0151 if (kwork->cpu_list != NULL) {
0152 fd = bpf_map__fd(skel->maps.perf_kwork_cpu_filter);
0153 if (fd < 0) {
0154 pr_debug("Invalid cpu filter fd\n");
0155 return -1;
0156 }
0157
0158 map = perf_cpu_map__new(kwork->cpu_list);
0159 if (map == NULL) {
0160 pr_debug("Invalid cpu_list\n");
0161 return -1;
0162 }
0163
0164 nr_cpus = libbpf_num_possible_cpus();
0165 for (i = 0; i < perf_cpu_map__nr(map); i++) {
0166 struct perf_cpu cpu = perf_cpu_map__cpu(map, i);
0167
0168 if (cpu.cpu >= nr_cpus) {
0169 perf_cpu_map__put(map);
0170 pr_err("Requested cpu %d too large\n", cpu.cpu);
0171 return -1;
0172 }
0173 bpf_map_update_elem(fd, &cpu.cpu, &val, BPF_ANY);
0174 }
0175 perf_cpu_map__put(map);
0176
0177 skel->bss->has_cpu_filter = 1;
0178 }
0179
0180 if (kwork->profile_name != NULL) {
0181 if (strlen(kwork->profile_name) >= MAX_KWORKNAME) {
0182 pr_err("Requested name filter %s too large, limit to %d\n",
0183 kwork->profile_name, MAX_KWORKNAME - 1);
0184 return -1;
0185 }
0186
0187 fd = bpf_map__fd(skel->maps.perf_kwork_name_filter);
0188 if (fd < 0) {
0189 pr_debug("Invalid name filter fd\n");
0190 return -1;
0191 }
0192
0193 key = 0;
0194 bpf_map_update_elem(fd, &key, kwork->profile_name, BPF_ANY);
0195
0196 skel->bss->has_name_filter = 1;
0197 }
0198
0199 return 0;
0200 }
0201
0202 int perf_kwork__trace_prepare_bpf(struct perf_kwork *kwork)
0203 {
0204 struct bpf_program *prog;
0205 struct kwork_class *class;
0206 struct kwork_class_bpf *class_bpf;
0207 enum kwork_class_type type;
0208
0209 skel = kwork_trace_bpf__open();
0210 if (!skel) {
0211 pr_debug("Failed to open kwork trace skeleton\n");
0212 return -1;
0213 }
0214
0215
0216
0217
0218
0219 bpf_object__for_each_program(prog, skel->obj)
0220 bpf_program__set_autoload(prog, false);
0221
0222 list_for_each_entry(class, &kwork->class_list, list) {
0223 type = class->type;
0224 if (!valid_kwork_class_type(type) ||
0225 (kwork_class_bpf_supported_list[type] == NULL)) {
0226 pr_err("Unsupported bpf trace class %s\n", class->name);
0227 goto out;
0228 }
0229
0230 class_bpf = kwork_class_bpf_supported_list[type];
0231 class_bpf->class = class;
0232
0233 if (class_bpf->load_prepare != NULL)
0234 class_bpf->load_prepare(kwork);
0235 }
0236
0237 if (kwork_trace_bpf__load(skel)) {
0238 pr_debug("Failed to load kwork trace skeleton\n");
0239 goto out;
0240 }
0241
0242 if (setup_filters(kwork))
0243 goto out;
0244
0245 if (kwork_trace_bpf__attach(skel)) {
0246 pr_debug("Failed to attach kwork trace skeleton\n");
0247 goto out;
0248 }
0249
0250 return 0;
0251
0252 out:
0253 kwork_trace_bpf__destroy(skel);
0254 return -1;
0255 }
0256
0257 static int add_work(struct perf_kwork *kwork,
0258 struct work_key *key,
0259 struct report_data *data)
0260 {
0261 struct kwork_work *work;
0262 struct kwork_class_bpf *bpf_trace;
0263 struct kwork_work tmp = {
0264 .id = key->id,
0265 .name = NULL,
0266 .cpu = key->cpu,
0267 };
0268 enum kwork_class_type type = key->type;
0269
0270 if (!valid_kwork_class_type(type)) {
0271 pr_debug("Invalid class type %d to add work\n", type);
0272 return -1;
0273 }
0274
0275 bpf_trace = kwork_class_bpf_supported_list[type];
0276 tmp.class = bpf_trace->class;
0277
0278 if ((bpf_trace->get_work_name != NULL) &&
0279 (bpf_trace->get_work_name(key, &tmp.name)))
0280 return -1;
0281
0282 work = perf_kwork_add_work(kwork, tmp.class, &tmp);
0283 if (work == NULL)
0284 return -1;
0285
0286 if (kwork->report == KWORK_REPORT_RUNTIME) {
0287 work->nr_atoms = data->nr;
0288 work->total_runtime = data->total_time;
0289 work->max_runtime = data->max_time;
0290 work->max_runtime_start = data->max_time_start;
0291 work->max_runtime_end = data->max_time_end;
0292 } else if (kwork->report == KWORK_REPORT_LATENCY) {
0293 work->nr_atoms = data->nr;
0294 work->total_latency = data->total_time;
0295 work->max_latency = data->max_time;
0296 work->max_latency_start = data->max_time_start;
0297 work->max_latency_end = data->max_time_end;
0298 } else {
0299 pr_debug("Invalid bpf report type %d\n", kwork->report);
0300 return -1;
0301 }
0302
0303 kwork->timestart = (u64)ts_start.tv_sec * NSEC_PER_SEC + ts_start.tv_nsec;
0304 kwork->timeend = (u64)ts_end.tv_sec * NSEC_PER_SEC + ts_end.tv_nsec;
0305
0306 return 0;
0307 }
0308
0309 int perf_kwork__report_read_bpf(struct perf_kwork *kwork)
0310 {
0311 struct report_data data;
0312 struct work_key key = {
0313 .type = 0,
0314 .cpu = 0,
0315 .id = 0,
0316 };
0317 struct work_key prev = {
0318 .type = 0,
0319 .cpu = 0,
0320 .id = 0,
0321 };
0322 int fd = bpf_map__fd(skel->maps.perf_kwork_report);
0323
0324 if (fd < 0) {
0325 pr_debug("Invalid report fd\n");
0326 return -1;
0327 }
0328
0329 while (!bpf_map_get_next_key(fd, &prev, &key)) {
0330 if ((bpf_map_lookup_elem(fd, &key, &data)) != 0) {
0331 pr_debug("Failed to lookup report elem\n");
0332 return -1;
0333 }
0334
0335 if ((data.nr != 0) && (add_work(kwork, &key, &data) != 0))
0336 return -1;
0337
0338 prev = key;
0339 }
0340 return 0;
0341 }
0342
0343 void perf_kwork__report_cleanup_bpf(void)
0344 {
0345 kwork_trace_bpf__destroy(skel);
0346 }