1 // SPDX-License-Identifier: GPL-2.0
7 #include "util/evlist.h" // for struct evsel_str_handler
8 #include "util/evsel.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/target.h"
13 #include "util/callchain.h"
14 #include "util/lock-contention.h"
15 #include "util/bpf_skel/lock_data.h"
17 #include <subcmd/pager.h>
18 #include <subcmd/parse-options.h>
19 #include "util/trace-event.h"
20 #include "util/tracepoint.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/tool.h"
25 #include "util/data.h"
26 #include "util/string2.h"
28 #include "util/util.h"
30 #include <sys/types.h>
31 #include <sys/prctl.h>
32 #include <semaphore.h>
37 #include <linux/list.h>
38 #include <linux/hash.h>
39 #include <linux/kernel.h>
40 #include <linux/zalloc.h>
41 #include <linux/err.h>
42 #include <linux/stringify.h>
44 static struct perf_session *session;
45 static struct target target;
47 /* based on kernel/lockdep.c */
48 #define LOCKHASH_BITS 12
49 #define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
51 static struct hlist_head *lockhash_table;
53 #define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
54 #define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
56 static struct rb_root thread_stats;
58 static bool combine_locks;
59 static bool show_thread_stats;
60 static bool show_lock_addrs;
61 static bool show_lock_owner;
63 static unsigned long bpf_map_entries = MAX_ENTRIES;
64 static int max_stack_depth = CONTENTION_STACK_DEPTH;
65 static int stack_skip = CONTENTION_STACK_SKIP;
66 static int print_nr_entries = INT_MAX / 2;
67 static LIST_HEAD(callstack_filters);
69 struct callstack_filter {
70 struct list_head list;
74 static struct lock_filter filters;
76 static enum lock_aggr_mode aggr_mode = LOCK_AGGR_ADDR;
78 static bool needs_callstack(void)
80 return !list_empty(&callstack_filters);
83 static struct thread_stat *thread_stat_find(u32 tid)
86 struct thread_stat *st;
88 node = thread_stats.rb_node;
90 st = container_of(node, struct thread_stat, rb);
93 else if (tid < st->tid)
96 node = node->rb_right;
102 static void thread_stat_insert(struct thread_stat *new)
104 struct rb_node **rb = &thread_stats.rb_node;
105 struct rb_node *parent = NULL;
106 struct thread_stat *p;
109 p = container_of(*rb, struct thread_stat, rb);
112 if (new->tid < p->tid)
113 rb = &(*rb)->rb_left;
114 else if (new->tid > p->tid)
115 rb = &(*rb)->rb_right;
117 BUG_ON("inserting invalid thread_stat\n");
120 rb_link_node(&new->rb, parent, rb);
121 rb_insert_color(&new->rb, &thread_stats);
124 static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
126 struct thread_stat *st;
128 st = thread_stat_find(tid);
132 st = zalloc(sizeof(struct thread_stat));
134 pr_err("memory allocation failed\n");
139 INIT_LIST_HEAD(&st->seq_list);
141 thread_stat_insert(st);
146 static struct thread_stat *thread_stat_findnew_first(u32 tid);
147 static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
148 thread_stat_findnew_first;
150 static struct thread_stat *thread_stat_findnew_first(u32 tid)
152 struct thread_stat *st;
154 st = zalloc(sizeof(struct thread_stat));
156 pr_err("memory allocation failed\n");
160 INIT_LIST_HEAD(&st->seq_list);
162 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
163 rb_insert_color(&st->rb, &thread_stats);
165 thread_stat_findnew = thread_stat_findnew_after_first;
169 /* build simple key function one is bigger than two */
170 #define SINGLE_KEY(member) \
171 static int lock_stat_key_ ## member(struct lock_stat *one, \
172 struct lock_stat *two) \
174 return one->member > two->member; \
177 SINGLE_KEY(nr_acquired)
178 SINGLE_KEY(nr_contended)
179 SINGLE_KEY(avg_wait_time)
180 SINGLE_KEY(wait_time_total)
181 SINGLE_KEY(wait_time_max)
183 static int lock_stat_key_wait_time_min(struct lock_stat *one,
184 struct lock_stat *two)
186 u64 s1 = one->wait_time_min;
187 u64 s2 = two->wait_time_min;
188 if (s1 == ULLONG_MAX)
190 if (s2 == ULLONG_MAX)
197 * name: the value for specify by user
198 * this should be simpler than raw name of member
199 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
202 /* header: the string printed on the header line */
204 /* len: the printing width of the field */
206 /* key: a pointer to function to compare two lock stats for sorting */
207 int (*key)(struct lock_stat*, struct lock_stat*);
208 /* print: a pointer to function to print a given lock stats */
209 void (*print)(struct lock_key*, struct lock_stat*);
210 /* list: list entry to link this */
211 struct list_head list;
214 static void lock_stat_key_print_time(unsigned long long nsec, int len)
216 static const struct {
220 { 1e9 * 3600, "h " },
228 for (int i = 0; table[i].unit; i++) {
229 if (nsec < table[i].base)
232 pr_info("%*.2f %s", len - 3, nsec / table[i].base, table[i].unit);
236 pr_info("%*llu %s", len - 3, nsec, "ns");
239 #define PRINT_KEY(member) \
240 static void lock_stat_key_print_ ## member(struct lock_key *key, \
241 struct lock_stat *ls) \
243 pr_info("%*llu", key->len, (unsigned long long)ls->member); \
246 #define PRINT_TIME(member) \
247 static void lock_stat_key_print_ ## member(struct lock_key *key, \
248 struct lock_stat *ls) \
250 lock_stat_key_print_time((unsigned long long)ls->member, key->len); \
253 PRINT_KEY(nr_acquired)
254 PRINT_KEY(nr_contended)
255 PRINT_TIME(avg_wait_time)
256 PRINT_TIME(wait_time_total)
257 PRINT_TIME(wait_time_max)
259 static void lock_stat_key_print_wait_time_min(struct lock_key *key,
260 struct lock_stat *ls)
262 u64 wait_time = ls->wait_time_min;
264 if (wait_time == ULLONG_MAX)
267 lock_stat_key_print_time(wait_time, key->len);
271 static const char *sort_key = "acquired";
273 static int (*compare)(struct lock_stat *, struct lock_stat *);
275 static struct rb_root sorted; /* place to store intermediate data */
276 static struct rb_root result; /* place to store sorted data */
278 static LIST_HEAD(lock_keys);
279 static const char *output_fields;
281 #define DEF_KEY_LOCK(name, header, fn_suffix, len) \
282 { #name, header, len, lock_stat_key_ ## fn_suffix, lock_stat_key_print_ ## fn_suffix, {} }
283 static struct lock_key report_keys[] = {
284 DEF_KEY_LOCK(acquired, "acquired", nr_acquired, 10),
285 DEF_KEY_LOCK(contended, "contended", nr_contended, 10),
286 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12),
287 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12),
288 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12),
289 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12),
291 /* extra comparisons much complicated should be here */
295 static struct lock_key contention_keys[] = {
296 DEF_KEY_LOCK(contended, "contended", nr_contended, 10),
297 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12),
298 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12),
299 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12),
300 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12),
302 /* extra comparisons much complicated should be here */
306 static int select_key(bool contention)
309 struct lock_key *keys = report_keys;
312 keys = contention_keys;
314 for (i = 0; keys[i].name; i++) {
315 if (!strcmp(keys[i].name, sort_key)) {
316 compare = keys[i].key;
318 /* selected key should be in the output fields */
319 if (list_empty(&keys[i].list))
320 list_add_tail(&keys[i].list, &lock_keys);
326 pr_err("Unknown compare key: %s\n", sort_key);
330 static int add_output_field(bool contention, char *name)
333 struct lock_key *keys = report_keys;
336 keys = contention_keys;
338 for (i = 0; keys[i].name; i++) {
339 if (strcmp(keys[i].name, name))
342 /* prevent double link */
343 if (list_empty(&keys[i].list))
344 list_add_tail(&keys[i].list, &lock_keys);
349 pr_err("Unknown output field: %s\n", name);
353 static int setup_output_field(bool contention, const char *str)
355 char *tok, *tmp, *orig;
357 struct lock_key *keys = report_keys;
360 keys = contention_keys;
362 /* no output field given: use all of them */
364 for (i = 0; keys[i].name; i++)
365 list_add_tail(&keys[i].list, &lock_keys);
369 for (i = 0; keys[i].name; i++)
370 INIT_LIST_HEAD(&keys[i].list);
372 orig = tmp = strdup(str);
376 while ((tok = strsep(&tmp, ",")) != NULL){
377 ret = add_output_field(contention, tok);
386 static void combine_lock_stats(struct lock_stat *st)
388 struct rb_node **rb = &sorted.rb_node;
389 struct rb_node *parent = NULL;
394 p = container_of(*rb, struct lock_stat, rb);
397 if (st->name && p->name)
398 ret = strcmp(st->name, p->name);
400 ret = !!st->name - !!p->name;
403 p->nr_acquired += st->nr_acquired;
404 p->nr_contended += st->nr_contended;
405 p->wait_time_total += st->wait_time_total;
408 p->avg_wait_time = p->wait_time_total / p->nr_contended;
410 if (p->wait_time_min > st->wait_time_min)
411 p->wait_time_min = st->wait_time_min;
412 if (p->wait_time_max < st->wait_time_max)
413 p->wait_time_max = st->wait_time_max;
415 p->broken |= st->broken;
421 rb = &(*rb)->rb_left;
423 rb = &(*rb)->rb_right;
426 rb_link_node(&st->rb, parent, rb);
427 rb_insert_color(&st->rb, &sorted);
430 static void insert_to_result(struct lock_stat *st,
431 int (*bigger)(struct lock_stat *, struct lock_stat *))
433 struct rb_node **rb = &result.rb_node;
434 struct rb_node *parent = NULL;
437 if (combine_locks && st->combined)
441 p = container_of(*rb, struct lock_stat, rb);
445 rb = &(*rb)->rb_left;
447 rb = &(*rb)->rb_right;
450 rb_link_node(&st->rb, parent, rb);
451 rb_insert_color(&st->rb, &result);
454 /* returns left most element of result, and erase it */
455 static struct lock_stat *pop_from_result(void)
457 struct rb_node *node = result.rb_node;
462 while (node->rb_left)
463 node = node->rb_left;
465 rb_erase(node, &result);
466 return container_of(node, struct lock_stat, rb);
469 struct lock_stat *lock_stat_find(u64 addr)
471 struct hlist_head *entry = lockhashentry(addr);
472 struct lock_stat *ret;
474 hlist_for_each_entry(ret, entry, hash_entry) {
475 if (ret->addr == addr)
481 struct lock_stat *lock_stat_findnew(u64 addr, const char *name, int flags)
483 struct hlist_head *entry = lockhashentry(addr);
484 struct lock_stat *ret, *new;
486 hlist_for_each_entry(ret, entry, hash_entry) {
487 if (ret->addr == addr)
491 new = zalloc(sizeof(struct lock_stat));
496 new->name = strdup(name);
503 new->wait_time_min = ULLONG_MAX;
505 hlist_add_head(&new->hash_entry, entry);
509 pr_err("memory allocation failed\n");
513 bool match_callstack_filter(struct machine *machine, u64 *callstack)
519 if (list_empty(&callstack_filters))
522 for (int i = 0; i < max_stack_depth; i++) {
523 struct callstack_filter *filter;
525 if (!callstack || !callstack[i])
529 sym = machine__find_kernel_symbol(machine, ip, &kmap);
533 list_for_each_entry(filter, &callstack_filters, list) {
534 if (strstr(sym->name, filter->name))
541 struct trace_lock_handler {
542 /* it's used on CONFIG_LOCKDEP */
543 int (*acquire_event)(struct evsel *evsel,
544 struct perf_sample *sample);
546 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */
547 int (*acquired_event)(struct evsel *evsel,
548 struct perf_sample *sample);
550 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */
551 int (*contended_event)(struct evsel *evsel,
552 struct perf_sample *sample);
554 /* it's used on CONFIG_LOCKDEP */
555 int (*release_event)(struct evsel *evsel,
556 struct perf_sample *sample);
558 /* it's used when CONFIG_LOCKDEP is off */
559 int (*contention_begin_event)(struct evsel *evsel,
560 struct perf_sample *sample);
562 /* it's used when CONFIG_LOCKDEP is off */
563 int (*contention_end_event)(struct evsel *evsel,
564 struct perf_sample *sample);
567 static struct lock_seq_stat *get_seq(struct thread_stat *ts, u64 addr)
569 struct lock_seq_stat *seq;
571 list_for_each_entry(seq, &ts->seq_list, list) {
572 if (seq->addr == addr)
576 seq = zalloc(sizeof(struct lock_seq_stat));
578 pr_err("memory allocation failed\n");
581 seq->state = SEQ_STATE_UNINITIALIZED;
584 list_add(&seq->list, &ts->seq_list);
596 static int bad_hist[BROKEN_MAX];
603 static int get_key_by_aggr_mode_simple(u64 *key, u64 addr, u32 tid)
612 case LOCK_AGGR_CALLER:
614 pr_err("Invalid aggregation mode: %d\n", aggr_mode);
620 static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample);
622 static int get_key_by_aggr_mode(u64 *key, u64 addr, struct evsel *evsel,
623 struct perf_sample *sample)
625 if (aggr_mode == LOCK_AGGR_CALLER) {
626 *key = callchain_id(evsel, sample);
629 return get_key_by_aggr_mode_simple(key, addr, sample->tid);
632 static int report_lock_acquire_event(struct evsel *evsel,
633 struct perf_sample *sample)
635 struct lock_stat *ls;
636 struct thread_stat *ts;
637 struct lock_seq_stat *seq;
638 const char *name = evsel__strval(evsel, sample, "name");
639 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
640 int flag = evsel__intval(evsel, sample, "flags");
644 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
648 ls = lock_stat_findnew(key, name, 0);
652 ts = thread_stat_findnew(sample->tid);
656 seq = get_seq(ts, addr);
660 switch (seq->state) {
661 case SEQ_STATE_UNINITIALIZED:
662 case SEQ_STATE_RELEASED:
664 seq->state = SEQ_STATE_ACQUIRING;
668 if (flag & READ_LOCK)
670 seq->state = SEQ_STATE_READ_ACQUIRED;
675 case SEQ_STATE_READ_ACQUIRED:
676 if (flag & READ_LOCK) {
684 case SEQ_STATE_ACQUIRED:
685 case SEQ_STATE_ACQUIRING:
686 case SEQ_STATE_CONTENDED:
688 /* broken lock sequence */
691 bad_hist[BROKEN_ACQUIRE]++;
693 list_del_init(&seq->list);
697 BUG_ON("Unknown state of lock sequence found!\n");
702 seq->prev_event_time = sample->time;
707 static int report_lock_acquired_event(struct evsel *evsel,
708 struct perf_sample *sample)
710 struct lock_stat *ls;
711 struct thread_stat *ts;
712 struct lock_seq_stat *seq;
714 const char *name = evsel__strval(evsel, sample, "name");
715 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
719 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
723 ls = lock_stat_findnew(key, name, 0);
727 ts = thread_stat_findnew(sample->tid);
731 seq = get_seq(ts, addr);
735 switch (seq->state) {
736 case SEQ_STATE_UNINITIALIZED:
737 /* orphan event, do nothing */
739 case SEQ_STATE_ACQUIRING:
741 case SEQ_STATE_CONTENDED:
742 contended_term = sample->time - seq->prev_event_time;
743 ls->wait_time_total += contended_term;
744 if (contended_term < ls->wait_time_min)
745 ls->wait_time_min = contended_term;
746 if (ls->wait_time_max < contended_term)
747 ls->wait_time_max = contended_term;
749 case SEQ_STATE_RELEASED:
750 case SEQ_STATE_ACQUIRED:
751 case SEQ_STATE_READ_ACQUIRED:
752 /* broken lock sequence */
755 bad_hist[BROKEN_ACQUIRED]++;
757 list_del_init(&seq->list);
761 BUG_ON("Unknown state of lock sequence found!\n");
765 seq->state = SEQ_STATE_ACQUIRED;
767 ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
768 seq->prev_event_time = sample->time;
773 static int report_lock_contended_event(struct evsel *evsel,
774 struct perf_sample *sample)
776 struct lock_stat *ls;
777 struct thread_stat *ts;
778 struct lock_seq_stat *seq;
779 const char *name = evsel__strval(evsel, sample, "name");
780 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
784 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
788 ls = lock_stat_findnew(key, name, 0);
792 ts = thread_stat_findnew(sample->tid);
796 seq = get_seq(ts, addr);
800 switch (seq->state) {
801 case SEQ_STATE_UNINITIALIZED:
802 /* orphan event, do nothing */
804 case SEQ_STATE_ACQUIRING:
806 case SEQ_STATE_RELEASED:
807 case SEQ_STATE_ACQUIRED:
808 case SEQ_STATE_READ_ACQUIRED:
809 case SEQ_STATE_CONTENDED:
810 /* broken lock sequence */
813 bad_hist[BROKEN_CONTENDED]++;
815 list_del_init(&seq->list);
819 BUG_ON("Unknown state of lock sequence found!\n");
823 seq->state = SEQ_STATE_CONTENDED;
825 ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
826 seq->prev_event_time = sample->time;
831 static int report_lock_release_event(struct evsel *evsel,
832 struct perf_sample *sample)
834 struct lock_stat *ls;
835 struct thread_stat *ts;
836 struct lock_seq_stat *seq;
837 const char *name = evsel__strval(evsel, sample, "name");
838 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
842 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
846 ls = lock_stat_findnew(key, name, 0);
850 ts = thread_stat_findnew(sample->tid);
854 seq = get_seq(ts, addr);
858 switch (seq->state) {
859 case SEQ_STATE_UNINITIALIZED:
861 case SEQ_STATE_ACQUIRED:
863 case SEQ_STATE_READ_ACQUIRED:
865 BUG_ON(seq->read_count < 0);
866 if (seq->read_count) {
871 case SEQ_STATE_ACQUIRING:
872 case SEQ_STATE_CONTENDED:
873 case SEQ_STATE_RELEASED:
874 /* broken lock sequence */
877 bad_hist[BROKEN_RELEASE]++;
881 BUG_ON("Unknown state of lock sequence found!\n");
887 list_del_init(&seq->list);
893 static int get_symbol_name_offset(struct map *map, struct symbol *sym, u64 ip,
898 if (map == NULL || sym == NULL) {
903 offset = map__map_ip(map, ip) - sym->start;
906 return scnprintf(buf, size, "%s+%#lx", sym->name, offset);
908 return strlcpy(buf, sym->name, size);
910 static int lock_contention_caller(struct evsel *evsel, struct perf_sample *sample,
913 struct thread *thread;
914 struct callchain_cursor *cursor;
915 struct machine *machine = &session->machines.host;
920 /* lock names will be replaced to task name later */
921 if (show_thread_stats)
924 thread = machine__findnew_thread(machine, -1, sample->pid);
928 cursor = get_tls_callchain_cursor();
930 /* use caller function name from the callchain */
931 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
932 NULL, NULL, max_stack_depth);
938 callchain_cursor_commit(cursor);
942 struct callchain_cursor_node *node;
944 node = callchain_cursor_current(cursor);
948 /* skip first few entries - for lock functions */
949 if (++skip <= stack_skip)
953 if (sym && !machine__is_lock_function(machine, node->ip)) {
954 get_symbol_name_offset(node->ms.map, sym, node->ip,
960 callchain_cursor_advance(cursor);
965 static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample)
967 struct callchain_cursor *cursor;
968 struct machine *machine = &session->machines.host;
969 struct thread *thread;
974 thread = machine__findnew_thread(machine, -1, sample->pid);
978 cursor = get_tls_callchain_cursor();
979 /* use caller function name from the callchain */
980 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
981 NULL, NULL, max_stack_depth);
987 callchain_cursor_commit(cursor);
990 struct callchain_cursor_node *node;
992 node = callchain_cursor_current(cursor);
996 /* skip first few entries - for lock functions */
997 if (++skip <= stack_skip)
1000 if (node->ms.sym && machine__is_lock_function(machine, node->ip))
1003 hash ^= hash_long((unsigned long)node->ip, 64);
1006 callchain_cursor_advance(cursor);
1011 static u64 *get_callstack(struct perf_sample *sample, int max_stack)
1017 callstack = calloc(max_stack, sizeof(*callstack));
1018 if (callstack == NULL)
1021 for (i = 0, c = 0; i < sample->callchain->nr && c < max_stack; i++) {
1022 u64 ip = sample->callchain->ips[i];
1024 if (ip >= PERF_CONTEXT_MAX)
1027 callstack[c++] = ip;
1032 static int report_lock_contention_begin_event(struct evsel *evsel,
1033 struct perf_sample *sample)
1035 struct lock_stat *ls;
1036 struct thread_stat *ts;
1037 struct lock_seq_stat *seq;
1038 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1039 unsigned int flags = evsel__intval(evsel, sample, "flags");
1042 static bool kmap_loaded;
1043 struct machine *machine = &session->machines.host;
1047 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1052 unsigned long *addrs;
1054 /* make sure it loads the kernel map to find lock symbols */
1055 map__load(machine__kernel_map(machine));
1058 /* convert (kernel) symbols to addresses */
1059 for (i = 0; i < filters.nr_syms; i++) {
1060 sym = machine__find_kernel_symbol_by_name(machine,
1064 pr_warning("ignore unknown symbol: %s\n",
1069 addrs = realloc(filters.addrs,
1070 (filters.nr_addrs + 1) * sizeof(*addrs));
1071 if (addrs == NULL) {
1072 pr_warning("memory allocation failure\n");
1076 addrs[filters.nr_addrs++] = map__unmap_ip(kmap, sym->start);
1077 filters.addrs = addrs;
1081 ls = lock_stat_find(key);
1084 const char *name = "";
1086 switch (aggr_mode) {
1087 case LOCK_AGGR_ADDR:
1088 sym = machine__find_kernel_symbol(machine, key, &kmap);
1092 case LOCK_AGGR_CALLER:
1094 if (lock_contention_caller(evsel, sample, buf, sizeof(buf)) < 0)
1097 case LOCK_AGGR_TASK:
1102 ls = lock_stat_findnew(key, name, flags);
1107 if (filters.nr_types) {
1110 for (i = 0; i < filters.nr_types; i++) {
1111 if (flags == filters.types[i]) {
1121 if (filters.nr_addrs) {
1124 for (i = 0; i < filters.nr_addrs; i++) {
1125 if (addr == filters.addrs[i]) {
1135 if (needs_callstack()) {
1136 u64 *callstack = get_callstack(sample, max_stack_depth);
1137 if (callstack == NULL)
1140 if (!match_callstack_filter(machine, callstack)) {
1145 if (ls->callstack == NULL)
1146 ls->callstack = callstack;
1151 ts = thread_stat_findnew(sample->tid);
1155 seq = get_seq(ts, addr);
1159 switch (seq->state) {
1160 case SEQ_STATE_UNINITIALIZED:
1161 case SEQ_STATE_ACQUIRED:
1163 case SEQ_STATE_CONTENDED:
1165 * It can have nested contention begin with mutex spinning,
1166 * then we would use the original contention begin event and
1167 * ignore the second one.
1170 case SEQ_STATE_ACQUIRING:
1171 case SEQ_STATE_READ_ACQUIRED:
1172 case SEQ_STATE_RELEASED:
1173 /* broken lock sequence */
1176 bad_hist[BROKEN_CONTENDED]++;
1178 list_del_init(&seq->list);
1182 BUG_ON("Unknown state of lock sequence found!\n");
1186 if (seq->state != SEQ_STATE_CONTENDED) {
1187 seq->state = SEQ_STATE_CONTENDED;
1188 seq->prev_event_time = sample->time;
1195 static int report_lock_contention_end_event(struct evsel *evsel,
1196 struct perf_sample *sample)
1198 struct lock_stat *ls;
1199 struct thread_stat *ts;
1200 struct lock_seq_stat *seq;
1202 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1206 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1210 ls = lock_stat_find(key);
1214 ts = thread_stat_find(sample->tid);
1218 seq = get_seq(ts, addr);
1222 switch (seq->state) {
1223 case SEQ_STATE_UNINITIALIZED:
1225 case SEQ_STATE_CONTENDED:
1226 contended_term = sample->time - seq->prev_event_time;
1227 ls->wait_time_total += contended_term;
1228 if (contended_term < ls->wait_time_min)
1229 ls->wait_time_min = contended_term;
1230 if (ls->wait_time_max < contended_term)
1231 ls->wait_time_max = contended_term;
1233 case SEQ_STATE_ACQUIRING:
1234 case SEQ_STATE_ACQUIRED:
1235 case SEQ_STATE_READ_ACQUIRED:
1236 case SEQ_STATE_RELEASED:
1237 /* broken lock sequence */
1240 bad_hist[BROKEN_ACQUIRED]++;
1242 list_del_init(&seq->list);
1246 BUG_ON("Unknown state of lock sequence found!\n");
1250 seq->state = SEQ_STATE_ACQUIRED;
1252 ls->avg_wait_time = ls->wait_time_total/ls->nr_acquired;
1257 /* lock oriented handlers */
1258 /* TODO: handlers for CPU oriented, thread oriented */
1259 static struct trace_lock_handler report_lock_ops = {
1260 .acquire_event = report_lock_acquire_event,
1261 .acquired_event = report_lock_acquired_event,
1262 .contended_event = report_lock_contended_event,
1263 .release_event = report_lock_release_event,
1264 .contention_begin_event = report_lock_contention_begin_event,
1265 .contention_end_event = report_lock_contention_end_event,
1268 static struct trace_lock_handler contention_lock_ops = {
1269 .contention_begin_event = report_lock_contention_begin_event,
1270 .contention_end_event = report_lock_contention_end_event,
1274 static struct trace_lock_handler *trace_handler;
1276 static int evsel__process_lock_acquire(struct evsel *evsel, struct perf_sample *sample)
1278 if (trace_handler->acquire_event)
1279 return trace_handler->acquire_event(evsel, sample);
1283 static int evsel__process_lock_acquired(struct evsel *evsel, struct perf_sample *sample)
1285 if (trace_handler->acquired_event)
1286 return trace_handler->acquired_event(evsel, sample);
1290 static int evsel__process_lock_contended(struct evsel *evsel, struct perf_sample *sample)
1292 if (trace_handler->contended_event)
1293 return trace_handler->contended_event(evsel, sample);
1297 static int evsel__process_lock_release(struct evsel *evsel, struct perf_sample *sample)
1299 if (trace_handler->release_event)
1300 return trace_handler->release_event(evsel, sample);
1304 static int evsel__process_contention_begin(struct evsel *evsel, struct perf_sample *sample)
1306 if (trace_handler->contention_begin_event)
1307 return trace_handler->contention_begin_event(evsel, sample);
1311 static int evsel__process_contention_end(struct evsel *evsel, struct perf_sample *sample)
1313 if (trace_handler->contention_end_event)
1314 return trace_handler->contention_end_event(evsel, sample);
1318 static void print_bad_events(int bad, int total)
1320 /* Output for debug, this have to be removed */
1323 const char *name[4] =
1324 { "acquire", "acquired", "contended", "release" };
1326 for (i = 0; i < BROKEN_MAX; i++)
1327 broken += bad_hist[i];
1329 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1332 pr_info("\n=== output for debug ===\n\n");
1333 pr_info("bad: %d, total: %d\n", bad, total);
1334 pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
1335 pr_info("histogram of events caused bad sequence\n");
1336 for (i = 0; i < BROKEN_MAX; i++)
1337 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
1340 /* TODO: various way to print, coloring, nano or milli sec */
1341 static void print_result(void)
1343 struct lock_stat *st;
1344 struct lock_key *key;
1346 int bad, total, printed;
1349 pr_info("%20s ", "Name");
1350 list_for_each_entry(key, &lock_keys, list)
1351 pr_info("%*s ", key->len, key->header);
1355 bad = total = printed = 0;
1356 while ((st = pop_from_result())) {
1360 if (!st->nr_acquired)
1363 bzero(cut_name, 20);
1365 if (strlen(st->name) < 20) {
1366 /* output raw name */
1367 const char *name = st->name;
1369 if (show_thread_stats) {
1372 /* st->addr contains tid of thread */
1373 t = perf_session__findnew(session, st->addr);
1374 name = thread__comm_str(t);
1377 pr_info("%20s ", name);
1379 strncpy(cut_name, st->name, 16);
1383 cut_name[19] = '\0';
1384 /* cut off name for saving output style */
1385 pr_info("%20s ", cut_name);
1388 list_for_each_entry(key, &lock_keys, list) {
1389 key->print(key, st);
1394 if (++printed >= print_nr_entries)
1398 print_bad_events(bad, total);
1401 static bool info_threads, info_map;
1403 static void dump_threads(void)
1405 struct thread_stat *st;
1406 struct rb_node *node;
1409 pr_info("%10s: comm\n", "Thread ID");
1411 node = rb_first(&thread_stats);
1413 st = container_of(node, struct thread_stat, rb);
1414 t = perf_session__findnew(session, st->tid);
1415 pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
1416 node = rb_next(node);
1421 static int compare_maps(struct lock_stat *a, struct lock_stat *b)
1425 if (a->name && b->name)
1426 ret = strcmp(a->name, b->name);
1428 ret = !!a->name - !!b->name;
1431 return a->addr < b->addr;
1436 static void dump_map(void)
1439 struct lock_stat *st;
1441 pr_info("Address of instance: name of class\n");
1442 for (i = 0; i < LOCKHASH_SIZE; i++) {
1443 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1444 insert_to_result(st, compare_maps);
1448 while ((st = pop_from_result()))
1449 pr_info(" %#llx: %s\n", (unsigned long long)st->addr, st->name);
1452 static int dump_info(void)
1462 pr_err("Unknown type of information\n");
1468 static const struct evsel_str_handler lock_tracepoints[] = {
1469 { "lock:lock_acquire", evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
1470 { "lock:lock_acquired", evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1471 { "lock:lock_contended", evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1472 { "lock:lock_release", evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
1475 static const struct evsel_str_handler contention_tracepoints[] = {
1476 { "lock:contention_begin", evsel__process_contention_begin, },
1477 { "lock:contention_end", evsel__process_contention_end, },
1480 static int process_event_update(struct perf_tool *tool,
1481 union perf_event *event,
1482 struct evlist **pevlist)
1486 ret = perf_event__process_event_update(tool, event, pevlist);
1490 /* this can return -EEXIST since we call it for each evsel */
1491 perf_session__set_tracepoints_handlers(session, lock_tracepoints);
1492 perf_session__set_tracepoints_handlers(session, contention_tracepoints);
1496 typedef int (*tracepoint_handler)(struct evsel *evsel,
1497 struct perf_sample *sample);
1499 static int process_sample_event(struct perf_tool *tool __maybe_unused,
1500 union perf_event *event,
1501 struct perf_sample *sample,
1502 struct evsel *evsel,
1503 struct machine *machine)
1506 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1509 if (thread == NULL) {
1510 pr_debug("problem processing %d event, skipping it.\n",
1511 event->header.type);
1515 if (evsel->handler != NULL) {
1516 tracepoint_handler f = evsel->handler;
1517 err = f(evsel, sample);
1520 thread__put(thread);
1525 static void combine_result(void)
1528 struct lock_stat *st;
1533 for (i = 0; i < LOCKHASH_SIZE; i++) {
1534 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1535 combine_lock_stats(st);
1540 static void sort_result(void)
1543 struct lock_stat *st;
1545 for (i = 0; i < LOCKHASH_SIZE; i++) {
1546 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1547 insert_to_result(st, compare);
1552 static const struct {
1556 } lock_type_table[] = {
1557 { 0, "semaphore", "semaphore" },
1558 { LCB_F_SPIN, "spinlock", "spinlock" },
1559 { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock" },
1560 { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock" },
1561 { LCB_F_READ, "rwsem:R", "rwsem" },
1562 { LCB_F_WRITE, "rwsem:W", "rwsem" },
1563 { LCB_F_RT, "rt-mutex", "rt-mutex" },
1564 { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt" },
1565 { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt" },
1566 { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem" },
1567 { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem" },
1568 { LCB_F_MUTEX, "mutex", "mutex" },
1569 { LCB_F_MUTEX | LCB_F_SPIN, "mutex", "mutex" },
1570 /* alias for get_type_flag() */
1571 { LCB_F_MUTEX | LCB_F_SPIN, "mutex-spin", "mutex" },
1574 static const char *get_type_str(unsigned int flags)
1576 flags &= LCB_F_MAX_FLAGS - 1;
1578 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1579 if (lock_type_table[i].flags == flags)
1580 return lock_type_table[i].str;
1585 static const char *get_type_name(unsigned int flags)
1587 flags &= LCB_F_MAX_FLAGS - 1;
1589 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1590 if (lock_type_table[i].flags == flags)
1591 return lock_type_table[i].name;
1596 static unsigned int get_type_flag(const char *str)
1598 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1599 if (!strcmp(lock_type_table[i].name, str))
1600 return lock_type_table[i].flags;
1602 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1603 if (!strcmp(lock_type_table[i].str, str))
1604 return lock_type_table[i].flags;
1609 static void lock_filter_finish(void)
1611 zfree(&filters.types);
1612 filters.nr_types = 0;
1614 zfree(&filters.addrs);
1615 filters.nr_addrs = 0;
1617 for (int i = 0; i < filters.nr_syms; i++)
1618 free(filters.syms[i]);
1620 zfree(&filters.syms);
1621 filters.nr_syms = 0;
1624 static void sort_contention_result(void)
1629 static void print_bpf_events(int total, struct lock_contention_fails *fails)
1631 /* Output for debug, this have to be removed */
1632 int broken = fails->task + fails->stack + fails->time + fails->data;
1634 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1638 pr_info("\n=== output for debug ===\n\n");
1639 pr_info("bad: %d, total: %d\n", broken, total);
1640 pr_info("bad rate: %.2f %%\n", (double)broken / (double)total * 100);
1642 pr_info("histogram of failure reasons\n");
1643 pr_info(" %10s: %d\n", "task", fails->task);
1644 pr_info(" %10s: %d\n", "stack", fails->stack);
1645 pr_info(" %10s: %d\n", "time", fails->time);
1646 pr_info(" %10s: %d\n", "data", fails->data);
1649 static void print_contention_result(struct lock_contention *con)
1651 struct lock_stat *st;
1652 struct lock_key *key;
1653 int bad, total, printed;
1656 list_for_each_entry(key, &lock_keys, list)
1657 pr_info("%*s ", key->len, key->header);
1659 switch (aggr_mode) {
1660 case LOCK_AGGR_TASK:
1661 pr_info(" %10s %s\n\n", "pid",
1662 show_lock_owner ? "owner" : "comm");
1664 case LOCK_AGGR_CALLER:
1665 pr_info(" %10s %s\n\n", "type", "caller");
1667 case LOCK_AGGR_ADDR:
1668 pr_info(" %16s %s\n\n", "address", "symbol");
1675 bad = total = printed = 0;
1677 while ((st = pop_from_result())) {
1681 total += use_bpf ? st->nr_contended : 1;
1685 if (!st->wait_time_total)
1688 list_for_each_entry(key, &lock_keys, list) {
1689 key->print(key, st);
1693 switch (aggr_mode) {
1694 case LOCK_AGGR_CALLER:
1695 pr_info(" %10s %s\n", get_type_str(st->flags), st->name);
1697 case LOCK_AGGR_TASK:
1699 t = perf_session__findnew(session, pid);
1700 pr_info(" %10d %s\n",
1701 pid, pid == -1 ? "Unknown" : thread__comm_str(t));
1703 case LOCK_AGGR_ADDR:
1704 pr_info(" %016llx %s (%s)\n", (unsigned long long)st->addr,
1705 st->name, get_type_name(st->flags));
1711 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) {
1717 for (int i = 0; i < max_stack_depth; i++) {
1718 if (!st->callstack || !st->callstack[i])
1721 ip = st->callstack[i];
1722 sym = machine__find_kernel_symbol(con->machine, ip, &kmap);
1723 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf));
1724 pr_info("\t\t\t%#lx %s\n", (unsigned long)ip, buf);
1728 if (++printed >= print_nr_entries)
1732 if (print_nr_entries) {
1733 /* update the total/bad stats */
1734 while ((st = pop_from_result())) {
1735 total += use_bpf ? st->nr_contended : 1;
1740 /* some entries are collected but hidden by the callstack filter */
1741 total += con->nr_filtered;
1744 print_bpf_events(total, &con->fails);
1746 print_bad_events(bad, total);
1751 static int __cmd_report(bool display_info)
1754 struct perf_tool eops = {
1755 .attr = perf_event__process_attr,
1756 .event_update = process_event_update,
1757 .sample = process_sample_event,
1758 .comm = perf_event__process_comm,
1759 .mmap = perf_event__process_mmap,
1760 .namespaces = perf_event__process_namespaces,
1761 .tracing_data = perf_event__process_tracing_data,
1762 .ordered_events = true,
1764 struct perf_data data = {
1766 .mode = PERF_DATA_MODE_READ,
1770 session = perf_session__new(&data, &eops);
1771 if (IS_ERR(session)) {
1772 pr_err("Initializing perf session failed\n");
1773 return PTR_ERR(session);
1776 /* for lock function check */
1777 symbol_conf.allow_aliases = true;
1778 symbol__init(&session->header.env);
1780 if (!data.is_pipe) {
1781 if (!perf_session__has_traces(session, "lock record"))
1784 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
1785 pr_err("Initializing perf session tracepoint handlers failed\n");
1789 if (perf_session__set_tracepoints_handlers(session, contention_tracepoints)) {
1790 pr_err("Initializing perf session tracepoint handlers failed\n");
1795 if (setup_output_field(false, output_fields))
1798 if (select_key(false))
1801 if (show_thread_stats)
1802 aggr_mode = LOCK_AGGR_TASK;
1804 err = perf_session__process_events(session);
1809 if (display_info) /* used for info subcommand */
1818 perf_session__delete(session);
1822 static void sighandler(int sig __maybe_unused)
1826 static int check_lock_contention_options(const struct option *options,
1827 const char * const *usage)
1830 if (show_thread_stats && show_lock_addrs) {
1831 pr_err("Cannot use thread and addr mode together\n");
1832 parse_options_usage(usage, options, "threads", 0);
1833 parse_options_usage(NULL, options, "lock-addr", 0);
1837 if (show_lock_owner && !use_bpf) {
1838 pr_err("Lock owners are available only with BPF\n");
1839 parse_options_usage(usage, options, "lock-owner", 0);
1840 parse_options_usage(NULL, options, "use-bpf", 0);
1844 if (show_lock_owner && show_lock_addrs) {
1845 pr_err("Cannot use owner and addr mode together\n");
1846 parse_options_usage(usage, options, "lock-owner", 0);
1847 parse_options_usage(NULL, options, "lock-addr", 0);
1851 if (show_lock_owner)
1852 show_thread_stats = true;
1857 static int __cmd_contention(int argc, const char **argv)
1860 struct perf_tool eops = {
1861 .attr = perf_event__process_attr,
1862 .event_update = process_event_update,
1863 .sample = process_sample_event,
1864 .comm = perf_event__process_comm,
1865 .mmap = perf_event__process_mmap,
1866 .tracing_data = perf_event__process_tracing_data,
1867 .ordered_events = true,
1869 struct perf_data data = {
1871 .mode = PERF_DATA_MODE_READ,
1874 struct lock_contention con = {
1876 .map_nr_entries = bpf_map_entries,
1877 .max_stack = max_stack_depth,
1878 .stack_skip = stack_skip,
1879 .filters = &filters,
1880 .save_callstack = needs_callstack(),
1881 .owner = show_lock_owner,
1884 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table));
1885 if (!lockhash_table)
1888 con.result = &lockhash_table[0];
1890 session = perf_session__new(use_bpf ? NULL : &data, &eops);
1891 if (IS_ERR(session)) {
1892 pr_err("Initializing perf session failed\n");
1893 err = PTR_ERR(session);
1897 con.machine = &session->machines.host;
1899 con.aggr_mode = aggr_mode = show_thread_stats ? LOCK_AGGR_TASK :
1900 show_lock_addrs ? LOCK_AGGR_ADDR : LOCK_AGGR_CALLER;
1902 if (con.aggr_mode == LOCK_AGGR_CALLER)
1903 con.save_callstack = true;
1905 /* for lock function check */
1906 symbol_conf.allow_aliases = true;
1907 symbol__init(&session->header.env);
1910 err = target__validate(&target);
1914 target__strerror(&target, err, errbuf, 512);
1915 pr_err("%s\n", errbuf);
1919 signal(SIGINT, sighandler);
1920 signal(SIGCHLD, sighandler);
1921 signal(SIGTERM, sighandler);
1923 con.evlist = evlist__new();
1924 if (con.evlist == NULL) {
1929 err = evlist__create_maps(con.evlist, &target);
1934 err = evlist__prepare_workload(con.evlist, &target,
1940 if (lock_contention_prepare(&con) < 0) {
1941 pr_err("lock contention BPF setup failed\n");
1944 } else if (!data.is_pipe) {
1945 if (!perf_session__has_traces(session, "lock record"))
1948 if (!evlist__find_evsel_by_str(session->evlist,
1949 "lock:contention_begin")) {
1950 pr_err("lock contention evsel not found\n");
1954 if (perf_session__set_tracepoints_handlers(session,
1955 contention_tracepoints)) {
1956 pr_err("Initializing perf session tracepoint handlers failed\n");
1961 if (setup_output_field(true, output_fields))
1964 if (select_key(true))
1968 lock_contention_start();
1970 evlist__start_workload(con.evlist);
1972 /* wait for signal */
1975 lock_contention_stop();
1976 lock_contention_read(&con);
1978 err = perf_session__process_events(session);
1985 sort_contention_result();
1986 print_contention_result(&con);
1989 lock_filter_finish();
1990 evlist__delete(con.evlist);
1991 lock_contention_finish();
1992 perf_session__delete(session);
1993 zfree(&lockhash_table);
1998 static int __cmd_record(int argc, const char **argv)
2000 const char *record_args[] = {
2001 "record", "-R", "-m", "1024", "-c", "1", "--synth", "task",
2003 const char *callgraph_args[] = {
2004 "--call-graph", "fp," __stringify(CONTENTION_STACK_DEPTH),
2006 unsigned int rec_argc, i, j, ret;
2007 unsigned int nr_tracepoints;
2008 unsigned int nr_callgraph_args = 0;
2009 const char **rec_argv;
2010 bool has_lock_stat = true;
2012 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
2013 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
2014 pr_debug("tracepoint %s is not enabled. "
2015 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
2016 lock_tracepoints[i].name);
2017 has_lock_stat = false;
2025 for (i = 0; i < ARRAY_SIZE(contention_tracepoints); i++) {
2026 if (!is_valid_tracepoint(contention_tracepoints[i].name)) {
2027 pr_err("tracepoint %s is not enabled.\n",
2028 contention_tracepoints[i].name);
2033 nr_callgraph_args = ARRAY_SIZE(callgraph_args);
2036 rec_argc = ARRAY_SIZE(record_args) + nr_callgraph_args + argc - 1;
2039 nr_tracepoints = ARRAY_SIZE(lock_tracepoints);
2041 nr_tracepoints = ARRAY_SIZE(contention_tracepoints);
2043 /* factor of 2 is for -e in front of each tracepoint */
2044 rec_argc += 2 * nr_tracepoints;
2046 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2050 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2051 rec_argv[i] = strdup(record_args[i]);
2053 for (j = 0; j < nr_tracepoints; j++) {
2054 const char *ev_name;
2057 ev_name = strdup(lock_tracepoints[j].name);
2059 ev_name = strdup(contention_tracepoints[j].name);
2064 rec_argv[i++] = "-e";
2065 rec_argv[i++] = ev_name;
2068 for (j = 0; j < nr_callgraph_args; j++, i++)
2069 rec_argv[i] = callgraph_args[j];
2071 for (j = 1; j < (unsigned int)argc; j++, i++)
2072 rec_argv[i] = argv[j];
2074 BUG_ON(i != rec_argc);
2076 ret = cmd_record(i, rec_argv);
2081 static int parse_map_entry(const struct option *opt, const char *str,
2082 int unset __maybe_unused)
2084 unsigned long *len = (unsigned long *)opt->value;
2089 val = strtoul(str, &endptr, 0);
2090 if (*endptr != '\0' || errno != 0) {
2091 pr_err("invalid BPF map length: %s\n", str);
2099 static int parse_max_stack(const struct option *opt, const char *str,
2100 int unset __maybe_unused)
2102 unsigned long *len = (unsigned long *)opt->value;
2107 val = strtol(str, &endptr, 0);
2108 if (*endptr != '\0' || errno != 0) {
2109 pr_err("invalid max stack depth: %s\n", str);
2113 if (val < 0 || val > sysctl__max_stack()) {
2114 pr_err("invalid max stack depth: %ld\n", val);
2122 static bool add_lock_type(unsigned int flags)
2126 tmp = realloc(filters.types, (filters.nr_types + 1) * sizeof(*filters.types));
2130 tmp[filters.nr_types++] = flags;
2131 filters.types = tmp;
2135 static int parse_lock_type(const struct option *opt __maybe_unused, const char *str,
2136 int unset __maybe_unused)
2138 char *s, *tmp, *tok;
2145 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2146 unsigned int flags = get_type_flag(tok);
2149 pr_err("Unknown lock flags: %s\n", tok);
2154 if (!add_lock_type(flags)) {
2164 static bool add_lock_addr(unsigned long addr)
2168 tmp = realloc(filters.addrs, (filters.nr_addrs + 1) * sizeof(*filters.addrs));
2170 pr_err("Memory allocation failure\n");
2174 tmp[filters.nr_addrs++] = addr;
2175 filters.addrs = tmp;
2179 static bool add_lock_sym(char *name)
2182 char *sym = strdup(name);
2185 pr_err("Memory allocation failure\n");
2189 tmp = realloc(filters.syms, (filters.nr_syms + 1) * sizeof(*filters.syms));
2191 pr_err("Memory allocation failure\n");
2196 tmp[filters.nr_syms++] = sym;
2201 static int parse_lock_addr(const struct option *opt __maybe_unused, const char *str,
2202 int unset __maybe_unused)
2204 char *s, *tmp, *tok;
2212 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2215 addr = strtoul(tok, &end, 16);
2217 if (!add_lock_addr(addr)) {
2225 * At this moment, we don't have kernel symbols. Save the symbols
2226 * in a separate list and resolve them to addresses later.
2228 if (!add_lock_sym(tok)) {
2238 static int parse_call_stack(const struct option *opt __maybe_unused, const char *str,
2239 int unset __maybe_unused)
2241 char *s, *tmp, *tok;
2248 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2249 struct callstack_filter *entry;
2251 entry = malloc(sizeof(*entry) + strlen(tok) + 1);
2252 if (entry == NULL) {
2253 pr_err("Memory allocation failure\n");
2257 strcpy(entry->name, tok);
2258 list_add_tail(&entry->list, &callstack_filters);
2265 int cmd_lock(int argc, const char **argv)
2267 const struct option lock_options[] = {
2268 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2269 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
2270 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
2271 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
2272 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2273 "file", "vmlinux pathname"),
2274 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
2275 "file", "kallsyms pathname"),
2276 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
2280 const struct option info_options[] = {
2281 OPT_BOOLEAN('t', "threads", &info_threads,
2282 "dump thread list in perf.data"),
2283 OPT_BOOLEAN('m', "map", &info_map,
2284 "map of lock instances (address:name table)"),
2285 OPT_PARENT(lock_options)
2288 const struct option report_options[] = {
2289 OPT_STRING('k', "key", &sort_key, "acquired",
2290 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2291 OPT_STRING('F', "field", &output_fields, NULL,
2292 "output fields (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2294 OPT_BOOLEAN('c', "combine-locks", &combine_locks,
2295 "combine locks in the same class"),
2296 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2297 "show per-thread lock stats"),
2298 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2299 OPT_PARENT(lock_options)
2302 struct option contention_options[] = {
2303 OPT_STRING('k', "key", &sort_key, "wait_total",
2304 "key for sorting (contended / wait_total / wait_max / wait_min / avg_wait)"),
2305 OPT_STRING('F', "field", &output_fields, "contended,wait_total,wait_max,avg_wait",
2306 "output fields (contended / wait_total / wait_max / wait_min / avg_wait)"),
2307 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2308 "show per-thread lock stats"),
2309 OPT_BOOLEAN('b', "use-bpf", &use_bpf, "use BPF program to collect lock contention stats"),
2310 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2311 "System-wide collection from all CPUs"),
2312 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2313 "List of cpus to monitor"),
2314 OPT_STRING('p', "pid", &target.pid, "pid",
2315 "Trace on existing process id"),
2316 OPT_STRING(0, "tid", &target.tid, "tid",
2317 "Trace on existing thread id (exclusive to --pid)"),
2318 OPT_CALLBACK('M', "map-nr-entries", &bpf_map_entries, "num",
2319 "Max number of BPF map entries", parse_map_entry),
2320 OPT_CALLBACK(0, "max-stack", &max_stack_depth, "num",
2321 "Set the maximum stack depth when collecting lopck contention, "
2322 "Default: " __stringify(CONTENTION_STACK_DEPTH), parse_max_stack),
2323 OPT_INTEGER(0, "stack-skip", &stack_skip,
2324 "Set the number of stack depth to skip when finding a lock caller, "
2325 "Default: " __stringify(CONTENTION_STACK_SKIP)),
2326 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2327 OPT_BOOLEAN('l', "lock-addr", &show_lock_addrs, "show lock stats by address"),
2328 OPT_CALLBACK('Y', "type-filter", NULL, "FLAGS",
2329 "Filter specific type of locks", parse_lock_type),
2330 OPT_CALLBACK('L', "lock-filter", NULL, "ADDRS/NAMES",
2331 "Filter specific address/symbol of locks", parse_lock_addr),
2332 OPT_CALLBACK('S', "callstack-filter", NULL, "NAMES",
2333 "Filter specific function in the callstack", parse_call_stack),
2334 OPT_BOOLEAN('o', "lock-owner", &show_lock_owner, "show lock owners instead of waiters"),
2335 OPT_PARENT(lock_options)
2338 const char * const info_usage[] = {
2339 "perf lock info [<options>]",
2342 const char *const lock_subcommands[] = { "record", "report", "script",
2343 "info", "contention", NULL };
2344 const char *lock_usage[] = {
2348 const char * const report_usage[] = {
2349 "perf lock report [<options>]",
2352 const char * const contention_usage[] = {
2353 "perf lock contention [<options>]",
2359 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table));
2360 if (!lockhash_table)
2363 for (i = 0; i < LOCKHASH_SIZE; i++)
2364 INIT_HLIST_HEAD(lockhash_table + i);
2366 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
2367 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2369 usage_with_options(lock_usage, lock_options);
2371 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2372 return __cmd_record(argc, argv);
2373 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) {
2374 trace_handler = &report_lock_ops;
2376 argc = parse_options(argc, argv,
2377 report_options, report_usage, 0);
2379 usage_with_options(report_usage, report_options);
2381 rc = __cmd_report(false);
2382 } else if (!strcmp(argv[0], "script")) {
2383 /* Aliased to 'perf script' */
2384 rc = cmd_script(argc, argv);
2385 } else if (!strcmp(argv[0], "info")) {
2387 argc = parse_options(argc, argv,
2388 info_options, info_usage, 0);
2390 usage_with_options(info_usage, info_options);
2392 /* recycling report_lock_ops */
2393 trace_handler = &report_lock_ops;
2394 rc = __cmd_report(true);
2395 } else if (strlen(argv[0]) > 2 && strstarts("contention", argv[0])) {
2396 trace_handler = &contention_lock_ops;
2397 sort_key = "wait_total";
2398 output_fields = "contended,wait_total,wait_max,avg_wait";
2400 #ifndef HAVE_BPF_SKEL
2401 set_option_nobuild(contention_options, 'b', "use-bpf",
2402 "no BUILD_BPF_SKEL=1", false);
2405 argc = parse_options(argc, argv, contention_options,
2406 contention_usage, 0);
2409 if (check_lock_contention_options(contention_options,
2410 contention_usage) < 0)
2413 rc = __cmd_contention(argc, argv);
2415 usage_with_options(lock_usage, lock_options);
2418 zfree(&lockhash_table);