1 | // SPDX-License-Identifier: GPL-2.0 |
2 | #include "util/cgroup.h" |
3 | #include "util/debug.h" |
4 | #include "util/evlist.h" |
5 | #include "util/machine.h" |
6 | #include "util/map.h" |
7 | #include "util/symbol.h" |
8 | #include "util/target.h" |
9 | #include "util/thread.h" |
10 | #include "util/thread_map.h" |
11 | #include "util/lock-contention.h" |
12 | #include <linux/zalloc.h> |
13 | #include <linux/string.h> |
14 | #include <bpf/bpf.h> |
15 | #include <inttypes.h> |
16 | |
17 | #include "bpf_skel/lock_contention.skel.h" |
18 | #include "bpf_skel/lock_data.h" |
19 | |
20 | static struct lock_contention_bpf *skel; |
21 | |
22 | int lock_contention_prepare(struct lock_contention *con) |
23 | { |
24 | int i, fd; |
25 | int ncpus = 1, ntasks = 1, ntypes = 1, naddrs = 1, ncgrps = 1; |
26 | struct evlist *evlist = con->evlist; |
27 | struct target *target = con->target; |
28 | |
29 | skel = lock_contention_bpf__open(); |
30 | if (!skel) { |
31 | pr_err("Failed to open lock-contention BPF skeleton\n" ); |
32 | return -1; |
33 | } |
34 | |
35 | bpf_map__set_value_size(skel->maps.stacks, con->max_stack * sizeof(u64)); |
36 | bpf_map__set_max_entries(skel->maps.lock_stat, con->map_nr_entries); |
37 | bpf_map__set_max_entries(skel->maps.tstamp, con->map_nr_entries); |
38 | |
39 | if (con->aggr_mode == LOCK_AGGR_TASK) |
40 | bpf_map__set_max_entries(skel->maps.task_data, con->map_nr_entries); |
41 | else |
42 | bpf_map__set_max_entries(skel->maps.task_data, 1); |
43 | |
44 | if (con->save_callstack) |
45 | bpf_map__set_max_entries(skel->maps.stacks, con->map_nr_entries); |
46 | else |
47 | bpf_map__set_max_entries(skel->maps.stacks, 1); |
48 | |
49 | if (target__has_cpu(target)) |
50 | ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus); |
51 | if (target__has_task(target)) |
52 | ntasks = perf_thread_map__nr(evlist->core.threads); |
53 | if (con->filters->nr_types) |
54 | ntypes = con->filters->nr_types; |
55 | if (con->filters->nr_cgrps) |
56 | ncgrps = con->filters->nr_cgrps; |
57 | |
58 | /* resolve lock name filters to addr */ |
59 | if (con->filters->nr_syms) { |
60 | struct symbol *sym; |
61 | struct map *kmap; |
62 | unsigned long *addrs; |
63 | |
64 | for (i = 0; i < con->filters->nr_syms; i++) { |
65 | sym = machine__find_kernel_symbol_by_name(con->machine, |
66 | con->filters->syms[i], |
67 | &kmap); |
68 | if (sym == NULL) { |
69 | pr_warning("ignore unknown symbol: %s\n" , |
70 | con->filters->syms[i]); |
71 | continue; |
72 | } |
73 | |
74 | addrs = realloc(con->filters->addrs, |
75 | (con->filters->nr_addrs + 1) * sizeof(*addrs)); |
76 | if (addrs == NULL) { |
77 | pr_warning("memory allocation failure\n" ); |
78 | continue; |
79 | } |
80 | |
81 | addrs[con->filters->nr_addrs++] = map__unmap_ip(kmap, sym->start); |
82 | con->filters->addrs = addrs; |
83 | } |
84 | naddrs = con->filters->nr_addrs; |
85 | } |
86 | |
87 | bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus); |
88 | bpf_map__set_max_entries(skel->maps.task_filter, ntasks); |
89 | bpf_map__set_max_entries(skel->maps.type_filter, ntypes); |
90 | bpf_map__set_max_entries(skel->maps.addr_filter, naddrs); |
91 | bpf_map__set_max_entries(skel->maps.cgroup_filter, ncgrps); |
92 | |
93 | if (lock_contention_bpf__load(skel) < 0) { |
94 | pr_err("Failed to load lock-contention BPF skeleton\n" ); |
95 | return -1; |
96 | } |
97 | |
98 | if (target__has_cpu(target)) { |
99 | u32 cpu; |
100 | u8 val = 1; |
101 | |
102 | skel->bss->has_cpu = 1; |
103 | fd = bpf_map__fd(skel->maps.cpu_filter); |
104 | |
105 | for (i = 0; i < ncpus; i++) { |
106 | cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu; |
107 | bpf_map_update_elem(fd, &cpu, &val, BPF_ANY); |
108 | } |
109 | } |
110 | |
111 | if (target__has_task(target)) { |
112 | u32 pid; |
113 | u8 val = 1; |
114 | |
115 | skel->bss->has_task = 1; |
116 | fd = bpf_map__fd(skel->maps.task_filter); |
117 | |
118 | for (i = 0; i < ntasks; i++) { |
119 | pid = perf_thread_map__pid(evlist->core.threads, i); |
120 | bpf_map_update_elem(fd, &pid, &val, BPF_ANY); |
121 | } |
122 | } |
123 | |
124 | if (target__none(target) && evlist->workload.pid > 0) { |
125 | u32 pid = evlist->workload.pid; |
126 | u8 val = 1; |
127 | |
128 | skel->bss->has_task = 1; |
129 | fd = bpf_map__fd(skel->maps.task_filter); |
130 | bpf_map_update_elem(fd, &pid, &val, BPF_ANY); |
131 | } |
132 | |
133 | if (con->filters->nr_types) { |
134 | u8 val = 1; |
135 | |
136 | skel->bss->has_type = 1; |
137 | fd = bpf_map__fd(skel->maps.type_filter); |
138 | |
139 | for (i = 0; i < con->filters->nr_types; i++) |
140 | bpf_map_update_elem(fd, &con->filters->types[i], &val, BPF_ANY); |
141 | } |
142 | |
143 | if (con->filters->nr_addrs) { |
144 | u8 val = 1; |
145 | |
146 | skel->bss->has_addr = 1; |
147 | fd = bpf_map__fd(skel->maps.addr_filter); |
148 | |
149 | for (i = 0; i < con->filters->nr_addrs; i++) |
150 | bpf_map_update_elem(fd, &con->filters->addrs[i], &val, BPF_ANY); |
151 | } |
152 | |
153 | if (con->filters->nr_cgrps) { |
154 | u8 val = 1; |
155 | |
156 | skel->bss->has_cgroup = 1; |
157 | fd = bpf_map__fd(skel->maps.cgroup_filter); |
158 | |
159 | for (i = 0; i < con->filters->nr_cgrps; i++) |
160 | bpf_map_update_elem(fd, &con->filters->cgrps[i], &val, BPF_ANY); |
161 | } |
162 | |
163 | /* these don't work well if in the rodata section */ |
164 | skel->bss->stack_skip = con->stack_skip; |
165 | skel->bss->aggr_mode = con->aggr_mode; |
166 | skel->bss->needs_callstack = con->save_callstack; |
167 | skel->bss->lock_owner = con->owner; |
168 | |
169 | if (con->aggr_mode == LOCK_AGGR_CGROUP) { |
170 | if (cgroup_is_v2("perf_event" )) |
171 | skel->bss->use_cgroup_v2 = 1; |
172 | |
173 | read_all_cgroups(&con->cgroups); |
174 | } |
175 | |
176 | bpf_program__set_autoload(skel->progs.collect_lock_syms, false); |
177 | |
178 | lock_contention_bpf__attach(skel); |
179 | return 0; |
180 | } |
181 | |
182 | /* |
183 | * Run the BPF program directly using BPF_PROG_TEST_RUN to update the end |
184 | * timestamp in ktime so that it can calculate delta easily. |
185 | */ |
186 | static void mark_end_timestamp(void) |
187 | { |
188 | DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, |
189 | .flags = BPF_F_TEST_RUN_ON_CPU, |
190 | ); |
191 | int prog_fd = bpf_program__fd(skel->progs.end_timestamp); |
192 | |
193 | bpf_prog_test_run_opts(prog_fd, &opts); |
194 | } |
195 | |
196 | static void update_lock_stat(int map_fd, int pid, u64 end_ts, |
197 | enum lock_aggr_mode aggr_mode, |
198 | struct tstamp_data *ts_data) |
199 | { |
200 | u64 delta; |
201 | struct contention_key stat_key = {}; |
202 | struct contention_data stat_data; |
203 | |
204 | if (ts_data->timestamp >= end_ts) |
205 | return; |
206 | |
207 | delta = end_ts - ts_data->timestamp; |
208 | |
209 | switch (aggr_mode) { |
210 | case LOCK_AGGR_CALLER: |
211 | stat_key.stack_id = ts_data->stack_id; |
212 | break; |
213 | case LOCK_AGGR_TASK: |
214 | stat_key.pid = pid; |
215 | break; |
216 | case LOCK_AGGR_ADDR: |
217 | stat_key.lock_addr_or_cgroup = ts_data->lock; |
218 | break; |
219 | case LOCK_AGGR_CGROUP: |
220 | /* TODO */ |
221 | return; |
222 | default: |
223 | return; |
224 | } |
225 | |
226 | if (bpf_map_lookup_elem(map_fd, &stat_key, &stat_data) < 0) |
227 | return; |
228 | |
229 | stat_data.total_time += delta; |
230 | stat_data.count++; |
231 | |
232 | if (delta > stat_data.max_time) |
233 | stat_data.max_time = delta; |
234 | if (delta < stat_data.min_time) |
235 | stat_data.min_time = delta; |
236 | |
237 | bpf_map_update_elem(map_fd, &stat_key, &stat_data, BPF_EXIST); |
238 | } |
239 | |
240 | /* |
241 | * Account entries in the tstamp map (which didn't see the corresponding |
242 | * lock:contention_end tracepoint) using end_ts. |
243 | */ |
244 | static void account_end_timestamp(struct lock_contention *con) |
245 | { |
246 | int ts_fd, stat_fd; |
247 | int *prev_key, key; |
248 | u64 end_ts = skel->bss->end_ts; |
249 | int total_cpus; |
250 | enum lock_aggr_mode aggr_mode = con->aggr_mode; |
251 | struct tstamp_data ts_data, *cpu_data; |
252 | |
253 | /* Iterate per-task tstamp map (key = TID) */ |
254 | ts_fd = bpf_map__fd(skel->maps.tstamp); |
255 | stat_fd = bpf_map__fd(skel->maps.lock_stat); |
256 | |
257 | prev_key = NULL; |
258 | while (!bpf_map_get_next_key(ts_fd, prev_key, &key)) { |
259 | if (bpf_map_lookup_elem(ts_fd, &key, &ts_data) == 0) { |
260 | int pid = key; |
261 | |
262 | if (aggr_mode == LOCK_AGGR_TASK && con->owner) |
263 | pid = ts_data.flags; |
264 | |
265 | update_lock_stat(map_fd: stat_fd, pid, end_ts, aggr_mode, |
266 | ts_data: &ts_data); |
267 | } |
268 | |
269 | prev_key = &key; |
270 | } |
271 | |
272 | /* Now it'll check per-cpu tstamp map which doesn't have TID. */ |
273 | if (aggr_mode == LOCK_AGGR_TASK || aggr_mode == LOCK_AGGR_CGROUP) |
274 | return; |
275 | |
276 | total_cpus = cpu__max_cpu().cpu; |
277 | ts_fd = bpf_map__fd(skel->maps.tstamp_cpu); |
278 | |
279 | cpu_data = calloc(total_cpus, sizeof(*cpu_data)); |
280 | if (cpu_data == NULL) |
281 | return; |
282 | |
283 | prev_key = NULL; |
284 | while (!bpf_map_get_next_key(ts_fd, prev_key, &key)) { |
285 | if (bpf_map_lookup_elem(ts_fd, &key, cpu_data) < 0) |
286 | goto next; |
287 | |
288 | for (int i = 0; i < total_cpus; i++) { |
289 | update_lock_stat(map_fd: stat_fd, pid: -1, end_ts, aggr_mode, |
290 | ts_data: &cpu_data[i]); |
291 | } |
292 | |
293 | next: |
294 | prev_key = &key; |
295 | } |
296 | free(cpu_data); |
297 | } |
298 | |
299 | int lock_contention_start(void) |
300 | { |
301 | skel->bss->enabled = 1; |
302 | return 0; |
303 | } |
304 | |
305 | int lock_contention_stop(void) |
306 | { |
307 | skel->bss->enabled = 0; |
308 | mark_end_timestamp(); |
309 | return 0; |
310 | } |
311 | |
312 | static const char *lock_contention_get_name(struct lock_contention *con, |
313 | struct contention_key *key, |
314 | u64 *stack_trace, u32 flags) |
315 | { |
316 | int idx = 0; |
317 | u64 addr; |
318 | const char *name = "" ; |
319 | static char name_buf[KSYM_NAME_LEN]; |
320 | struct symbol *sym; |
321 | struct map *kmap; |
322 | struct machine *machine = con->machine; |
323 | |
324 | if (con->aggr_mode == LOCK_AGGR_TASK) { |
325 | struct contention_task_data task; |
326 | int pid = key->pid; |
327 | int task_fd = bpf_map__fd(skel->maps.task_data); |
328 | |
329 | /* do not update idle comm which contains CPU number */ |
330 | if (pid) { |
331 | struct thread *t = machine__findnew_thread(machine, /*pid=*/-1, pid); |
332 | |
333 | if (t == NULL) |
334 | return name; |
335 | if (!bpf_map_lookup_elem(task_fd, &pid, &task) && |
336 | thread__set_comm(t, task.comm, /*timestamp=*/0)) |
337 | name = task.comm; |
338 | } |
339 | return name; |
340 | } |
341 | |
342 | if (con->aggr_mode == LOCK_AGGR_ADDR) { |
343 | int lock_fd = bpf_map__fd(skel->maps.lock_syms); |
344 | |
345 | /* per-process locks set upper bits of the flags */ |
346 | if (flags & LCD_F_MMAP_LOCK) |
347 | return "mmap_lock" ; |
348 | if (flags & LCD_F_SIGHAND_LOCK) |
349 | return "siglock" ; |
350 | |
351 | /* global locks with symbols */ |
352 | sym = machine__find_kernel_symbol(machine, key->lock_addr_or_cgroup, &kmap); |
353 | if (sym) |
354 | return sym->name; |
355 | |
356 | /* try semi-global locks collected separately */ |
357 | if (!bpf_map_lookup_elem(lock_fd, &key->lock_addr_or_cgroup, &flags)) { |
358 | if (flags == LOCK_CLASS_RQLOCK) |
359 | return "rq_lock" ; |
360 | } |
361 | |
362 | return "" ; |
363 | } |
364 | |
365 | if (con->aggr_mode == LOCK_AGGR_CGROUP) { |
366 | u64 cgrp_id = key->lock_addr_or_cgroup; |
367 | struct cgroup *cgrp = __cgroup__find(&con->cgroups, cgrp_id); |
368 | |
369 | if (cgrp) |
370 | return cgrp->name; |
371 | |
372 | snprintf(name_buf, sizeof(name_buf), "cgroup:%" PRIu64 "" , cgrp_id); |
373 | return name_buf; |
374 | } |
375 | |
376 | /* LOCK_AGGR_CALLER: skip lock internal functions */ |
377 | while (machine__is_lock_function(machine, stack_trace[idx]) && |
378 | idx < con->max_stack - 1) |
379 | idx++; |
380 | |
381 | addr = stack_trace[idx]; |
382 | sym = machine__find_kernel_symbol(machine, addr, &kmap); |
383 | |
384 | if (sym) { |
385 | unsigned long offset; |
386 | |
387 | offset = map__map_ip(kmap, addr) - sym->start; |
388 | |
389 | if (offset == 0) |
390 | return sym->name; |
391 | |
392 | snprintf(name_buf, sizeof(name_buf), "%s+%#lx" , sym->name, offset); |
393 | } else { |
394 | snprintf(name_buf, sizeof(name_buf), "%#lx" , (unsigned long)addr); |
395 | } |
396 | |
397 | return name_buf; |
398 | } |
399 | |
400 | int lock_contention_read(struct lock_contention *con) |
401 | { |
402 | int fd, stack, err = 0; |
403 | struct contention_key *prev_key, key = {}; |
404 | struct contention_data data = {}; |
405 | struct lock_stat *st = NULL; |
406 | struct machine *machine = con->machine; |
407 | u64 *stack_trace; |
408 | size_t stack_size = con->max_stack * sizeof(*stack_trace); |
409 | |
410 | fd = bpf_map__fd(skel->maps.lock_stat); |
411 | stack = bpf_map__fd(skel->maps.stacks); |
412 | |
413 | con->fails.task = skel->bss->task_fail; |
414 | con->fails.stack = skel->bss->stack_fail; |
415 | con->fails.time = skel->bss->time_fail; |
416 | con->fails.data = skel->bss->data_fail; |
417 | |
418 | stack_trace = zalloc(stack_size); |
419 | if (stack_trace == NULL) |
420 | return -1; |
421 | |
422 | account_end_timestamp(con); |
423 | |
424 | if (con->aggr_mode == LOCK_AGGR_TASK) { |
425 | struct thread *idle = machine__findnew_thread(machine, |
426 | /*pid=*/0, |
427 | /*tid=*/0); |
428 | thread__set_comm(idle, "swapper" , /*timestamp=*/0); |
429 | } |
430 | |
431 | if (con->aggr_mode == LOCK_AGGR_ADDR) { |
432 | DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, |
433 | .flags = BPF_F_TEST_RUN_ON_CPU, |
434 | ); |
435 | int prog_fd = bpf_program__fd(skel->progs.collect_lock_syms); |
436 | |
437 | bpf_prog_test_run_opts(prog_fd, &opts); |
438 | } |
439 | |
440 | /* make sure it loads the kernel map */ |
441 | maps__load_first(machine->kmaps); |
442 | |
443 | prev_key = NULL; |
444 | while (!bpf_map_get_next_key(fd, prev_key, &key)) { |
445 | s64 ls_key; |
446 | const char *name; |
447 | |
448 | /* to handle errors in the loop body */ |
449 | err = -1; |
450 | |
451 | bpf_map_lookup_elem(fd, &key, &data); |
452 | if (con->save_callstack) { |
453 | bpf_map_lookup_elem(stack, &key.stack_id, stack_trace); |
454 | |
455 | if (!match_callstack_filter(machine, stack_trace)) { |
456 | con->nr_filtered += data.count; |
457 | goto next; |
458 | } |
459 | } |
460 | |
461 | switch (con->aggr_mode) { |
462 | case LOCK_AGGR_CALLER: |
463 | ls_key = key.stack_id; |
464 | break; |
465 | case LOCK_AGGR_TASK: |
466 | ls_key = key.pid; |
467 | break; |
468 | case LOCK_AGGR_ADDR: |
469 | case LOCK_AGGR_CGROUP: |
470 | ls_key = key.lock_addr_or_cgroup; |
471 | break; |
472 | default: |
473 | goto next; |
474 | } |
475 | |
476 | st = lock_stat_find(ls_key); |
477 | if (st != NULL) { |
478 | st->wait_time_total += data.total_time; |
479 | if (st->wait_time_max < data.max_time) |
480 | st->wait_time_max = data.max_time; |
481 | if (st->wait_time_min > data.min_time) |
482 | st->wait_time_min = data.min_time; |
483 | |
484 | st->nr_contended += data.count; |
485 | if (st->nr_contended) |
486 | st->avg_wait_time = st->wait_time_total / st->nr_contended; |
487 | goto next; |
488 | } |
489 | |
490 | name = lock_contention_get_name(con, key: &key, stack_trace, flags: data.flags); |
491 | st = lock_stat_findnew(ls_key, name, data.flags); |
492 | if (st == NULL) |
493 | break; |
494 | |
495 | st->nr_contended = data.count; |
496 | st->wait_time_total = data.total_time; |
497 | st->wait_time_max = data.max_time; |
498 | st->wait_time_min = data.min_time; |
499 | |
500 | if (data.count) |
501 | st->avg_wait_time = data.total_time / data.count; |
502 | |
503 | if (con->aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { |
504 | st->callstack = memdup(stack_trace, stack_size); |
505 | if (st->callstack == NULL) |
506 | break; |
507 | } |
508 | |
509 | next: |
510 | prev_key = &key; |
511 | |
512 | /* we're fine now, reset the error */ |
513 | err = 0; |
514 | } |
515 | |
516 | free(stack_trace); |
517 | |
518 | return err; |
519 | } |
520 | |
521 | int lock_contention_finish(struct lock_contention *con) |
522 | { |
523 | if (skel) { |
524 | skel->bss->enabled = 0; |
525 | lock_contention_bpf__destroy(skel); |
526 | } |
527 | |
528 | while (!RB_EMPTY_ROOT(&con->cgroups)) { |
529 | struct rb_node *node = rb_first(&con->cgroups); |
530 | struct cgroup *cgrp = rb_entry(node, struct cgroup, node); |
531 | |
532 | rb_erase(node, &con->cgroups); |
533 | cgroup__put(cgrp); |
534 | } |
535 | |
536 | return 0; |
537 | } |
538 | |