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
 20static struct lock_contention_bpf *skel;
 21
 22int 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
182int lock_contention_start(void)
183{
184	skel->bss->enabled = 1;
185	return 0;
186}
187
188int lock_contention_stop(void)
189{
190	skel->bss->enabled = 0;
191	return 0;
192}
193
194static const char *lock_contention_get_name(struct lock_contention *con,
195					    struct contention_key *key,
196					    u64 *stack_trace, u32 flags)
197{
198	int idx = 0;
199	u64 addr;
200	const char *name = "";
201	static char name_buf[KSYM_NAME_LEN];
202	struct symbol *sym;
203	struct map *kmap;
204	struct machine *machine = con->machine;
205
206	if (con->aggr_mode == LOCK_AGGR_TASK) {
207		struct contention_task_data task;
208		int pid = key->pid;
209		int task_fd = bpf_map__fd(skel->maps.task_data);
210
211		/* do not update idle comm which contains CPU number */
212		if (pid) {
213			struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid);
214
215			if (t == NULL)
216				return name;
217			if (!bpf_map_lookup_elem(task_fd, &pid, &task) &&
218			    thread__set_comm(t, task.comm, /*timestamp=*/0))
219				name = task.comm;
220		}
221		return name;
222	}
223
224	if (con->aggr_mode == LOCK_AGGR_ADDR) {
225		int lock_fd = bpf_map__fd(skel->maps.lock_syms);
226
227		/* per-process locks set upper bits of the flags */
228		if (flags & LCD_F_MMAP_LOCK)
229			return "mmap_lock";
230		if (flags & LCD_F_SIGHAND_LOCK)
231			return "siglock";
232
233		/* global locks with symbols */
234		sym = machine__find_kernel_symbol(machine, key->lock_addr_or_cgroup, &kmap);
235		if (sym)
236			return sym->name;
237
238		/* try semi-global locks collected separately */
239		if (!bpf_map_lookup_elem(lock_fd, &key->lock_addr_or_cgroup, &flags)) {
240			if (flags == LOCK_CLASS_RQLOCK)
241				return "rq_lock";
242		}
243
244		return "";
245	}
246
247	if (con->aggr_mode == LOCK_AGGR_CGROUP) {
248		u64 cgrp_id = key->lock_addr_or_cgroup;
249		struct cgroup *cgrp = __cgroup__find(&con->cgroups, cgrp_id);
250
251		if (cgrp)
252			return cgrp->name;
253
254		snprintf(name_buf, sizeof(name_buf), "cgroup:%" PRIu64 "", cgrp_id);
255		return name_buf;
256	}
257
258	/* LOCK_AGGR_CALLER: skip lock internal functions */
259	while (machine__is_lock_function(machine, stack_trace[idx]) &&
260	       idx < con->max_stack - 1)
261		idx++;
262
263	addr = stack_trace[idx];
264	sym = machine__find_kernel_symbol(machine, addr, &kmap);
265
266	if (sym) {
267		unsigned long offset;
268
269		offset = map__map_ip(kmap, addr) - sym->start;
270
271		if (offset == 0)
272			return sym->name;
273
274		snprintf(name_buf, sizeof(name_buf), "%s+%#lx", sym->name, offset);
275	} else {
276		snprintf(name_buf, sizeof(name_buf), "%#lx", (unsigned long)addr);
277	}
278
279	return name_buf;
280}
281
282int lock_contention_read(struct lock_contention *con)
283{
284	int fd, stack, err = 0;
285	struct contention_key *prev_key, key = {};
286	struct contention_data data = {};
287	struct lock_stat *st = NULL;
288	struct machine *machine = con->machine;
289	u64 *stack_trace;
290	size_t stack_size = con->max_stack * sizeof(*stack_trace);
291
292	fd = bpf_map__fd(skel->maps.lock_stat);
293	stack = bpf_map__fd(skel->maps.stacks);
294
295	con->fails.task = skel->bss->task_fail;
296	con->fails.stack = skel->bss->stack_fail;
297	con->fails.time = skel->bss->time_fail;
298	con->fails.data = skel->bss->data_fail;
299
300	stack_trace = zalloc(stack_size);
301	if (stack_trace == NULL)
302		return -1;
303
304	if (con->aggr_mode == LOCK_AGGR_TASK) {
305		struct thread *idle = __machine__findnew_thread(machine,
306								/*pid=*/0,
307								/*tid=*/0);
308		thread__set_comm(idle, "swapper", /*timestamp=*/0);
309	}
310
311	if (con->aggr_mode == LOCK_AGGR_ADDR) {
312		DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts,
313			.flags = BPF_F_TEST_RUN_ON_CPU,
314		);
315		int prog_fd = bpf_program__fd(skel->progs.collect_lock_syms);
316
317		bpf_prog_test_run_opts(prog_fd, &opts);
318	}
319
320	/* make sure it loads the kernel map */
321	maps__load_first(machine->kmaps);
322
323	prev_key = NULL;
324	while (!bpf_map_get_next_key(fd, prev_key, &key)) {
325		s64 ls_key;
326		const char *name;
327
328		/* to handle errors in the loop body */
329		err = -1;
330
331		bpf_map_lookup_elem(fd, &key, &data);
332		if (con->save_callstack) {
333			bpf_map_lookup_elem(stack, &key.stack_id, stack_trace);
334
335			if (!match_callstack_filter(machine, stack_trace)) {
336				con->nr_filtered += data.count;
337				goto next;
338			}
339		}
340
341		switch (con->aggr_mode) {
342		case LOCK_AGGR_CALLER:
343			ls_key = key.stack_id;
344			break;
345		case LOCK_AGGR_TASK:
346			ls_key = key.pid;
347			break;
348		case LOCK_AGGR_ADDR:
349		case LOCK_AGGR_CGROUP:
350			ls_key = key.lock_addr_or_cgroup;
351			break;
352		default:
353			goto next;
354		}
355
356		st = lock_stat_find(ls_key);
357		if (st != NULL) {
358			st->wait_time_total += data.total_time;
359			if (st->wait_time_max < data.max_time)
360				st->wait_time_max = data.max_time;
361			if (st->wait_time_min > data.min_time)
362				st->wait_time_min = data.min_time;
363
364			st->nr_contended += data.count;
365			if (st->nr_contended)
366				st->avg_wait_time = st->wait_time_total / st->nr_contended;
367			goto next;
368		}
369
370		name = lock_contention_get_name(con, &key, stack_trace, data.flags);
371		st = lock_stat_findnew(ls_key, name, data.flags);
372		if (st == NULL)
373			break;
374
375		st->nr_contended = data.count;
376		st->wait_time_total = data.total_time;
377		st->wait_time_max = data.max_time;
378		st->wait_time_min = data.min_time;
379
380		if (data.count)
381			st->avg_wait_time = data.total_time / data.count;
382
383		if (con->aggr_mode == LOCK_AGGR_CALLER && verbose > 0) {
384			st->callstack = memdup(stack_trace, stack_size);
385			if (st->callstack == NULL)
386				break;
387		}
388
389next:
390		prev_key = &key;
391
392		/* we're fine now, reset the error */
393		err = 0;
394	}
395
396	free(stack_trace);
397
398	return err;
399}
400
401int lock_contention_finish(struct lock_contention *con)
402{
403	if (skel) {
404		skel->bss->enabled = 0;
405		lock_contention_bpf__destroy(skel);
406	}
407
408	while (!RB_EMPTY_ROOT(&con->cgroups)) {
409		struct rb_node *node = rb_first(&con->cgroups);
410		struct cgroup *cgrp = rb_entry(node, struct cgroup, node);
411
412		rb_erase(node, &con->cgroups);
413		cgroup__put(cgrp);
414	}
415
416	return 0;
417}