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[LOCKHASH_SIZE];
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 = 10240;
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 verbose > 0 || !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 = &callchain_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 /* use caller function name from the callchain */
929 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
930 NULL, NULL, max_stack_depth);
936 callchain_cursor_commit(cursor);
940 struct callchain_cursor_node *node;
942 node = callchain_cursor_current(cursor);
946 /* skip first few entries - for lock functions */
947 if (++skip <= stack_skip)
951 if (sym && !machine__is_lock_function(machine, node->ip)) {
952 get_symbol_name_offset(node->ms.map, sym, node->ip,
958 callchain_cursor_advance(cursor);
963 static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample)
965 struct callchain_cursor *cursor = &callchain_cursor;
966 struct machine *machine = &session->machines.host;
967 struct thread *thread;
972 thread = machine__findnew_thread(machine, -1, sample->pid);
976 /* use caller function name from the callchain */
977 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
978 NULL, NULL, max_stack_depth);
984 callchain_cursor_commit(cursor);
987 struct callchain_cursor_node *node;
989 node = callchain_cursor_current(cursor);
993 /* skip first few entries - for lock functions */
994 if (++skip <= stack_skip)
997 if (node->ms.sym && machine__is_lock_function(machine, node->ip))
1000 hash ^= hash_long((unsigned long)node->ip, 64);
1003 callchain_cursor_advance(cursor);
1008 static u64 *get_callstack(struct perf_sample *sample, int max_stack)
1014 callstack = calloc(max_stack, sizeof(*callstack));
1015 if (callstack == NULL)
1018 for (i = 0, c = 0; i < sample->callchain->nr && c < max_stack; i++) {
1019 u64 ip = sample->callchain->ips[i];
1021 if (ip >= PERF_CONTEXT_MAX)
1024 callstack[c++] = ip;
1029 static int report_lock_contention_begin_event(struct evsel *evsel,
1030 struct perf_sample *sample)
1032 struct lock_stat *ls;
1033 struct thread_stat *ts;
1034 struct lock_seq_stat *seq;
1035 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1036 unsigned int flags = evsel__intval(evsel, sample, "flags");
1039 static bool kmap_loaded;
1040 struct machine *machine = &session->machines.host;
1044 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1049 unsigned long *addrs;
1051 /* make sure it loads the kernel map to find lock symbols */
1052 map__load(machine__kernel_map(machine));
1055 /* convert (kernel) symbols to addresses */
1056 for (i = 0; i < filters.nr_syms; i++) {
1057 sym = machine__find_kernel_symbol_by_name(machine,
1061 pr_warning("ignore unknown symbol: %s\n",
1066 addrs = realloc(filters.addrs,
1067 (filters.nr_addrs + 1) * sizeof(*addrs));
1068 if (addrs == NULL) {
1069 pr_warning("memory allocation failure\n");
1073 addrs[filters.nr_addrs++] = kmap->unmap_ip(kmap, sym->start);
1074 filters.addrs = addrs;
1078 ls = lock_stat_find(key);
1081 const char *name = "";
1083 switch (aggr_mode) {
1084 case LOCK_AGGR_ADDR:
1085 sym = machine__find_kernel_symbol(machine, key, &kmap);
1089 case LOCK_AGGR_CALLER:
1091 if (lock_contention_caller(evsel, sample, buf, sizeof(buf)) < 0)
1094 case LOCK_AGGR_TASK:
1099 ls = lock_stat_findnew(key, name, flags);
1104 if (filters.nr_types) {
1107 for (i = 0; i < filters.nr_types; i++) {
1108 if (flags == filters.types[i]) {
1118 if (filters.nr_addrs) {
1121 for (i = 0; i < filters.nr_addrs; i++) {
1122 if (addr == filters.addrs[i]) {
1132 if (needs_callstack()) {
1133 u64 *callstack = get_callstack(sample, max_stack_depth);
1134 if (callstack == NULL)
1137 if (!match_callstack_filter(machine, callstack)) {
1142 if (ls->callstack == NULL)
1143 ls->callstack = callstack;
1148 ts = thread_stat_findnew(sample->tid);
1152 seq = get_seq(ts, addr);
1156 switch (seq->state) {
1157 case SEQ_STATE_UNINITIALIZED:
1158 case SEQ_STATE_ACQUIRED:
1160 case SEQ_STATE_CONTENDED:
1162 * It can have nested contention begin with mutex spinning,
1163 * then we would use the original contention begin event and
1164 * ignore the second one.
1167 case SEQ_STATE_ACQUIRING:
1168 case SEQ_STATE_READ_ACQUIRED:
1169 case SEQ_STATE_RELEASED:
1170 /* broken lock sequence */
1173 bad_hist[BROKEN_CONTENDED]++;
1175 list_del_init(&seq->list);
1179 BUG_ON("Unknown state of lock sequence found!\n");
1183 if (seq->state != SEQ_STATE_CONTENDED) {
1184 seq->state = SEQ_STATE_CONTENDED;
1185 seq->prev_event_time = sample->time;
1192 static int report_lock_contention_end_event(struct evsel *evsel,
1193 struct perf_sample *sample)
1195 struct lock_stat *ls;
1196 struct thread_stat *ts;
1197 struct lock_seq_stat *seq;
1199 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1203 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1207 ls = lock_stat_find(key);
1211 ts = thread_stat_find(sample->tid);
1215 seq = get_seq(ts, addr);
1219 switch (seq->state) {
1220 case SEQ_STATE_UNINITIALIZED:
1222 case SEQ_STATE_CONTENDED:
1223 contended_term = sample->time - seq->prev_event_time;
1224 ls->wait_time_total += contended_term;
1225 if (contended_term < ls->wait_time_min)
1226 ls->wait_time_min = contended_term;
1227 if (ls->wait_time_max < contended_term)
1228 ls->wait_time_max = contended_term;
1230 case SEQ_STATE_ACQUIRING:
1231 case SEQ_STATE_ACQUIRED:
1232 case SEQ_STATE_READ_ACQUIRED:
1233 case SEQ_STATE_RELEASED:
1234 /* broken lock sequence */
1237 bad_hist[BROKEN_ACQUIRED]++;
1239 list_del_init(&seq->list);
1243 BUG_ON("Unknown state of lock sequence found!\n");
1247 seq->state = SEQ_STATE_ACQUIRED;
1249 ls->avg_wait_time = ls->wait_time_total/ls->nr_acquired;
1254 /* lock oriented handlers */
1255 /* TODO: handlers for CPU oriented, thread oriented */
1256 static struct trace_lock_handler report_lock_ops = {
1257 .acquire_event = report_lock_acquire_event,
1258 .acquired_event = report_lock_acquired_event,
1259 .contended_event = report_lock_contended_event,
1260 .release_event = report_lock_release_event,
1261 .contention_begin_event = report_lock_contention_begin_event,
1262 .contention_end_event = report_lock_contention_end_event,
1265 static struct trace_lock_handler contention_lock_ops = {
1266 .contention_begin_event = report_lock_contention_begin_event,
1267 .contention_end_event = report_lock_contention_end_event,
1271 static struct trace_lock_handler *trace_handler;
1273 static int evsel__process_lock_acquire(struct evsel *evsel, struct perf_sample *sample)
1275 if (trace_handler->acquire_event)
1276 return trace_handler->acquire_event(evsel, sample);
1280 static int evsel__process_lock_acquired(struct evsel *evsel, struct perf_sample *sample)
1282 if (trace_handler->acquired_event)
1283 return trace_handler->acquired_event(evsel, sample);
1287 static int evsel__process_lock_contended(struct evsel *evsel, struct perf_sample *sample)
1289 if (trace_handler->contended_event)
1290 return trace_handler->contended_event(evsel, sample);
1294 static int evsel__process_lock_release(struct evsel *evsel, struct perf_sample *sample)
1296 if (trace_handler->release_event)
1297 return trace_handler->release_event(evsel, sample);
1301 static int evsel__process_contention_begin(struct evsel *evsel, struct perf_sample *sample)
1303 if (trace_handler->contention_begin_event)
1304 return trace_handler->contention_begin_event(evsel, sample);
1308 static int evsel__process_contention_end(struct evsel *evsel, struct perf_sample *sample)
1310 if (trace_handler->contention_end_event)
1311 return trace_handler->contention_end_event(evsel, sample);
1315 static void print_bad_events(int bad, int total)
1317 /* Output for debug, this have to be removed */
1320 const char *name[4] =
1321 { "acquire", "acquired", "contended", "release" };
1323 for (i = 0; i < BROKEN_MAX; i++)
1324 broken += bad_hist[i];
1326 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1329 pr_info("\n=== output for debug ===\n\n");
1330 pr_info("bad: %d, total: %d\n", bad, total);
1331 pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
1332 pr_info("histogram of events caused bad sequence\n");
1333 for (i = 0; i < BROKEN_MAX; i++)
1334 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
1337 /* TODO: various way to print, coloring, nano or milli sec */
1338 static void print_result(void)
1340 struct lock_stat *st;
1341 struct lock_key *key;
1343 int bad, total, printed;
1346 pr_info("%20s ", "Name");
1347 list_for_each_entry(key, &lock_keys, list)
1348 pr_info("%*s ", key->len, key->header);
1352 bad = total = printed = 0;
1353 while ((st = pop_from_result())) {
1357 if (!st->nr_acquired)
1360 bzero(cut_name, 20);
1362 if (strlen(st->name) < 20) {
1363 /* output raw name */
1364 const char *name = st->name;
1366 if (show_thread_stats) {
1369 /* st->addr contains tid of thread */
1370 t = perf_session__findnew(session, st->addr);
1371 name = thread__comm_str(t);
1374 pr_info("%20s ", name);
1376 strncpy(cut_name, st->name, 16);
1380 cut_name[19] = '\0';
1381 /* cut off name for saving output style */
1382 pr_info("%20s ", cut_name);
1385 list_for_each_entry(key, &lock_keys, list) {
1386 key->print(key, st);
1391 if (++printed >= print_nr_entries)
1395 print_bad_events(bad, total);
1398 static bool info_threads, info_map;
1400 static void dump_threads(void)
1402 struct thread_stat *st;
1403 struct rb_node *node;
1406 pr_info("%10s: comm\n", "Thread ID");
1408 node = rb_first(&thread_stats);
1410 st = container_of(node, struct thread_stat, rb);
1411 t = perf_session__findnew(session, st->tid);
1412 pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
1413 node = rb_next(node);
1418 static int compare_maps(struct lock_stat *a, struct lock_stat *b)
1422 if (a->name && b->name)
1423 ret = strcmp(a->name, b->name);
1425 ret = !!a->name - !!b->name;
1428 return a->addr < b->addr;
1433 static void dump_map(void)
1436 struct lock_stat *st;
1438 pr_info("Address of instance: name of class\n");
1439 for (i = 0; i < LOCKHASH_SIZE; i++) {
1440 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1441 insert_to_result(st, compare_maps);
1445 while ((st = pop_from_result()))
1446 pr_info(" %#llx: %s\n", (unsigned long long)st->addr, st->name);
1449 static int dump_info(void)
1459 pr_err("Unknown type of information\n");
1465 static const struct evsel_str_handler lock_tracepoints[] = {
1466 { "lock:lock_acquire", evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
1467 { "lock:lock_acquired", evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1468 { "lock:lock_contended", evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1469 { "lock:lock_release", evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
1472 static const struct evsel_str_handler contention_tracepoints[] = {
1473 { "lock:contention_begin", evsel__process_contention_begin, },
1474 { "lock:contention_end", evsel__process_contention_end, },
1477 static int process_event_update(struct perf_tool *tool,
1478 union perf_event *event,
1479 struct evlist **pevlist)
1483 ret = perf_event__process_event_update(tool, event, pevlist);
1487 /* this can return -EEXIST since we call it for each evsel */
1488 perf_session__set_tracepoints_handlers(session, lock_tracepoints);
1489 perf_session__set_tracepoints_handlers(session, contention_tracepoints);
1493 typedef int (*tracepoint_handler)(struct evsel *evsel,
1494 struct perf_sample *sample);
1496 static int process_sample_event(struct perf_tool *tool __maybe_unused,
1497 union perf_event *event,
1498 struct perf_sample *sample,
1499 struct evsel *evsel,
1500 struct machine *machine)
1503 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1506 if (thread == NULL) {
1507 pr_debug("problem processing %d event, skipping it.\n",
1508 event->header.type);
1512 if (evsel->handler != NULL) {
1513 tracepoint_handler f = evsel->handler;
1514 err = f(evsel, sample);
1517 thread__put(thread);
1522 static void combine_result(void)
1525 struct lock_stat *st;
1530 for (i = 0; i < LOCKHASH_SIZE; i++) {
1531 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1532 combine_lock_stats(st);
1537 static void sort_result(void)
1540 struct lock_stat *st;
1542 for (i = 0; i < LOCKHASH_SIZE; i++) {
1543 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1544 insert_to_result(st, compare);
1549 static const struct {
1553 } lock_type_table[] = {
1554 { 0, "semaphore", "semaphore" },
1555 { LCB_F_SPIN, "spinlock", "spinlock" },
1556 { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock" },
1557 { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock" },
1558 { LCB_F_READ, "rwsem:R", "rwsem" },
1559 { LCB_F_WRITE, "rwsem:W", "rwsem" },
1560 { LCB_F_RT, "rt-mutex", "rt-mutex" },
1561 { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt" },
1562 { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt" },
1563 { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem" },
1564 { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem" },
1565 { LCB_F_MUTEX, "mutex", "mutex" },
1566 { LCB_F_MUTEX | LCB_F_SPIN, "mutex", "mutex" },
1567 /* alias for get_type_flag() */
1568 { LCB_F_MUTEX | LCB_F_SPIN, "mutex-spin", "mutex" },
1571 static const char *get_type_str(unsigned int flags)
1573 flags &= LCB_F_MAX_FLAGS - 1;
1575 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1576 if (lock_type_table[i].flags == flags)
1577 return lock_type_table[i].str;
1582 static const char *get_type_name(unsigned int flags)
1584 flags &= LCB_F_MAX_FLAGS - 1;
1586 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1587 if (lock_type_table[i].flags == flags)
1588 return lock_type_table[i].name;
1593 static unsigned int get_type_flag(const char *str)
1595 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1596 if (!strcmp(lock_type_table[i].name, str))
1597 return lock_type_table[i].flags;
1599 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1600 if (!strcmp(lock_type_table[i].str, str))
1601 return lock_type_table[i].flags;
1606 static void lock_filter_finish(void)
1608 zfree(&filters.types);
1609 filters.nr_types = 0;
1611 zfree(&filters.addrs);
1612 filters.nr_addrs = 0;
1614 for (int i = 0; i < filters.nr_syms; i++)
1615 free(filters.syms[i]);
1617 zfree(&filters.syms);
1618 filters.nr_syms = 0;
1621 static void sort_contention_result(void)
1626 static void print_bpf_events(int total, struct lock_contention_fails *fails)
1628 /* Output for debug, this have to be removed */
1629 int broken = fails->task + fails->stack + fails->time;
1631 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1635 pr_info("\n=== output for debug ===\n\n");
1636 pr_info("bad: %d, total: %d\n", broken, total);
1637 pr_info("bad rate: %.2f %%\n", (double)broken / (double)total * 100);
1639 pr_info("histogram of failure reasons\n");
1640 pr_info(" %10s: %d\n", "task", fails->task);
1641 pr_info(" %10s: %d\n", "stack", fails->stack);
1642 pr_info(" %10s: %d\n", "time", fails->time);
1644 static void print_contention_result(struct lock_contention *con)
1646 struct lock_stat *st;
1647 struct lock_key *key;
1648 int bad, total, printed;
1651 list_for_each_entry(key, &lock_keys, list)
1652 pr_info("%*s ", key->len, key->header);
1654 switch (aggr_mode) {
1655 case LOCK_AGGR_TASK:
1656 pr_info(" %10s %s\n\n", "pid",
1657 show_lock_owner ? "owner" : "comm");
1659 case LOCK_AGGR_CALLER:
1660 pr_info(" %10s %s\n\n", "type", "caller");
1662 case LOCK_AGGR_ADDR:
1663 pr_info(" %16s %s\n\n", "address", "symbol");
1670 bad = total = printed = 0;
1672 while ((st = pop_from_result())) {
1676 total += use_bpf ? st->nr_contended : 1;
1680 if (!st->wait_time_total)
1683 list_for_each_entry(key, &lock_keys, list) {
1684 key->print(key, st);
1688 switch (aggr_mode) {
1689 case LOCK_AGGR_CALLER:
1690 pr_info(" %10s %s\n", get_type_str(st->flags), st->name);
1692 case LOCK_AGGR_TASK:
1694 t = perf_session__findnew(session, pid);
1695 pr_info(" %10d %s\n",
1696 pid, pid == -1 ? "Unknown" : thread__comm_str(t));
1698 case LOCK_AGGR_ADDR:
1699 pr_info(" %016llx %s (%s)\n", (unsigned long long)st->addr,
1700 st->name, get_type_name(st->flags));
1706 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) {
1712 for (int i = 0; i < max_stack_depth; i++) {
1713 if (!st->callstack || !st->callstack[i])
1716 ip = st->callstack[i];
1717 sym = machine__find_kernel_symbol(con->machine, ip, &kmap);
1718 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf));
1719 pr_info("\t\t\t%#lx %s\n", (unsigned long)ip, buf);
1723 if (++printed >= print_nr_entries)
1728 print_bpf_events(total, &con->fails);
1730 print_bad_events(bad, total);
1735 static int __cmd_report(bool display_info)
1738 struct perf_tool eops = {
1739 .attr = perf_event__process_attr,
1740 .event_update = process_event_update,
1741 .sample = process_sample_event,
1742 .comm = perf_event__process_comm,
1743 .mmap = perf_event__process_mmap,
1744 .namespaces = perf_event__process_namespaces,
1745 .tracing_data = perf_event__process_tracing_data,
1746 .ordered_events = true,
1748 struct perf_data data = {
1750 .mode = PERF_DATA_MODE_READ,
1754 session = perf_session__new(&data, &eops);
1755 if (IS_ERR(session)) {
1756 pr_err("Initializing perf session failed\n");
1757 return PTR_ERR(session);
1760 /* for lock function check */
1761 symbol_conf.sort_by_name = true;
1762 symbol_conf.allow_aliases = true;
1763 symbol__init(&session->header.env);
1765 if (!data.is_pipe) {
1766 if (!perf_session__has_traces(session, "lock record"))
1769 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
1770 pr_err("Initializing perf session tracepoint handlers failed\n");
1774 if (perf_session__set_tracepoints_handlers(session, contention_tracepoints)) {
1775 pr_err("Initializing perf session tracepoint handlers failed\n");
1780 if (setup_output_field(false, output_fields))
1783 if (select_key(false))
1786 if (show_thread_stats)
1787 aggr_mode = LOCK_AGGR_TASK;
1789 err = perf_session__process_events(session);
1794 if (display_info) /* used for info subcommand */
1803 perf_session__delete(session);
1807 static void sighandler(int sig __maybe_unused)
1811 static int check_lock_contention_options(const struct option *options,
1812 const char * const *usage)
1815 if (show_thread_stats && show_lock_addrs) {
1816 pr_err("Cannot use thread and addr mode together\n");
1817 parse_options_usage(usage, options, "threads", 0);
1818 parse_options_usage(NULL, options, "lock-addr", 0);
1822 if (show_lock_owner && !use_bpf) {
1823 pr_err("Lock owners are available only with BPF\n");
1824 parse_options_usage(usage, options, "lock-owner", 0);
1825 parse_options_usage(NULL, options, "use-bpf", 0);
1829 if (show_lock_owner && show_lock_addrs) {
1830 pr_err("Cannot use owner and addr mode together\n");
1831 parse_options_usage(usage, options, "lock-owner", 0);
1832 parse_options_usage(NULL, options, "lock-addr", 0);
1836 if (show_lock_owner)
1837 show_thread_stats = true;
1842 static int __cmd_contention(int argc, const char **argv)
1845 struct perf_tool eops = {
1846 .attr = perf_event__process_attr,
1847 .event_update = process_event_update,
1848 .sample = process_sample_event,
1849 .comm = perf_event__process_comm,
1850 .mmap = perf_event__process_mmap,
1851 .tracing_data = perf_event__process_tracing_data,
1852 .ordered_events = true,
1854 struct perf_data data = {
1856 .mode = PERF_DATA_MODE_READ,
1859 struct lock_contention con = {
1861 .result = &lockhash_table[0],
1862 .map_nr_entries = bpf_map_entries,
1863 .max_stack = max_stack_depth,
1864 .stack_skip = stack_skip,
1865 .filters = &filters,
1866 .save_callstack = needs_callstack(),
1867 .owner = show_lock_owner,
1870 session = perf_session__new(use_bpf ? NULL : &data, &eops);
1871 if (IS_ERR(session)) {
1872 pr_err("Initializing perf session failed\n");
1873 return PTR_ERR(session);
1876 con.machine = &session->machines.host;
1878 con.aggr_mode = aggr_mode = show_thread_stats ? LOCK_AGGR_TASK :
1879 show_lock_addrs ? LOCK_AGGR_ADDR : LOCK_AGGR_CALLER;
1881 if (con.aggr_mode == LOCK_AGGR_CALLER)
1882 con.save_callstack = true;
1884 /* for lock function check */
1885 symbol_conf.sort_by_name = true;
1886 symbol_conf.allow_aliases = true;
1887 symbol__init(&session->header.env);
1890 err = target__validate(&target);
1894 target__strerror(&target, err, errbuf, 512);
1895 pr_err("%s\n", errbuf);
1899 signal(SIGINT, sighandler);
1900 signal(SIGCHLD, sighandler);
1901 signal(SIGTERM, sighandler);
1903 con.evlist = evlist__new();
1904 if (con.evlist == NULL) {
1909 err = evlist__create_maps(con.evlist, &target);
1914 err = evlist__prepare_workload(con.evlist, &target,
1920 if (lock_contention_prepare(&con) < 0) {
1921 pr_err("lock contention BPF setup failed\n");
1924 } else if (!data.is_pipe) {
1925 if (!perf_session__has_traces(session, "lock record"))
1928 if (!evlist__find_evsel_by_str(session->evlist,
1929 "lock:contention_begin")) {
1930 pr_err("lock contention evsel not found\n");
1934 if (perf_session__set_tracepoints_handlers(session,
1935 contention_tracepoints)) {
1936 pr_err("Initializing perf session tracepoint handlers failed\n");
1941 if (setup_output_field(true, output_fields))
1944 if (select_key(true))
1948 lock_contention_start();
1950 evlist__start_workload(con.evlist);
1952 /* wait for signal */
1955 lock_contention_stop();
1956 lock_contention_read(&con);
1958 err = perf_session__process_events(session);
1965 sort_contention_result();
1966 print_contention_result(&con);
1969 lock_filter_finish();
1970 evlist__delete(con.evlist);
1971 lock_contention_finish();
1972 perf_session__delete(session);
1977 static int __cmd_record(int argc, const char **argv)
1979 const char *record_args[] = {
1980 "record", "-R", "-m", "1024", "-c", "1", "--synth", "task",
1982 const char *callgraph_args[] = {
1983 "--call-graph", "fp," __stringify(CONTENTION_STACK_DEPTH),
1985 unsigned int rec_argc, i, j, ret;
1986 unsigned int nr_tracepoints;
1987 unsigned int nr_callgraph_args = 0;
1988 const char **rec_argv;
1989 bool has_lock_stat = true;
1991 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
1992 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
1993 pr_debug("tracepoint %s is not enabled. "
1994 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
1995 lock_tracepoints[i].name);
1996 has_lock_stat = false;
2004 for (i = 0; i < ARRAY_SIZE(contention_tracepoints); i++) {
2005 if (!is_valid_tracepoint(contention_tracepoints[i].name)) {
2006 pr_err("tracepoint %s is not enabled.\n",
2007 contention_tracepoints[i].name);
2012 nr_callgraph_args = ARRAY_SIZE(callgraph_args);
2015 rec_argc = ARRAY_SIZE(record_args) + nr_callgraph_args + argc - 1;
2018 nr_tracepoints = ARRAY_SIZE(lock_tracepoints);
2020 nr_tracepoints = ARRAY_SIZE(contention_tracepoints);
2022 /* factor of 2 is for -e in front of each tracepoint */
2023 rec_argc += 2 * nr_tracepoints;
2025 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2029 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2030 rec_argv[i] = strdup(record_args[i]);
2032 for (j = 0; j < nr_tracepoints; j++) {
2033 const char *ev_name;
2036 ev_name = strdup(lock_tracepoints[j].name);
2038 ev_name = strdup(contention_tracepoints[j].name);
2043 rec_argv[i++] = "-e";
2044 rec_argv[i++] = ev_name;
2047 for (j = 0; j < nr_callgraph_args; j++, i++)
2048 rec_argv[i] = callgraph_args[j];
2050 for (j = 1; j < (unsigned int)argc; j++, i++)
2051 rec_argv[i] = argv[j];
2053 BUG_ON(i != rec_argc);
2055 ret = cmd_record(i, rec_argv);
2060 static int parse_map_entry(const struct option *opt, const char *str,
2061 int unset __maybe_unused)
2063 unsigned long *len = (unsigned long *)opt->value;
2068 val = strtoul(str, &endptr, 0);
2069 if (*endptr != '\0' || errno != 0) {
2070 pr_err("invalid BPF map length: %s\n", str);
2078 static int parse_max_stack(const struct option *opt, const char *str,
2079 int unset __maybe_unused)
2081 unsigned long *len = (unsigned long *)opt->value;
2086 val = strtol(str, &endptr, 0);
2087 if (*endptr != '\0' || errno != 0) {
2088 pr_err("invalid max stack depth: %s\n", str);
2092 if (val < 0 || val > sysctl__max_stack()) {
2093 pr_err("invalid max stack depth: %ld\n", val);
2101 static bool add_lock_type(unsigned int flags)
2105 tmp = realloc(filters.types, (filters.nr_types + 1) * sizeof(*filters.types));
2109 tmp[filters.nr_types++] = flags;
2110 filters.types = tmp;
2114 static int parse_lock_type(const struct option *opt __maybe_unused, const char *str,
2115 int unset __maybe_unused)
2117 char *s, *tmp, *tok;
2124 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2125 unsigned int flags = get_type_flag(tok);
2128 pr_err("Unknown lock flags: %s\n", tok);
2133 if (!add_lock_type(flags)) {
2143 static bool add_lock_addr(unsigned long addr)
2147 tmp = realloc(filters.addrs, (filters.nr_addrs + 1) * sizeof(*filters.addrs));
2149 pr_err("Memory allocation failure\n");
2153 tmp[filters.nr_addrs++] = addr;
2154 filters.addrs = tmp;
2158 static bool add_lock_sym(char *name)
2161 char *sym = strdup(name);
2164 pr_err("Memory allocation failure\n");
2168 tmp = realloc(filters.syms, (filters.nr_syms + 1) * sizeof(*filters.syms));
2170 pr_err("Memory allocation failure\n");
2175 tmp[filters.nr_syms++] = sym;
2180 static int parse_lock_addr(const struct option *opt __maybe_unused, const char *str,
2181 int unset __maybe_unused)
2183 char *s, *tmp, *tok;
2191 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2194 addr = strtoul(tok, &end, 16);
2196 if (!add_lock_addr(addr)) {
2204 * At this moment, we don't have kernel symbols. Save the symbols
2205 * in a separate list and resolve them to addresses later.
2207 if (!add_lock_sym(tok)) {
2217 static int parse_call_stack(const struct option *opt __maybe_unused, const char *str,
2218 int unset __maybe_unused)
2220 char *s, *tmp, *tok;
2227 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2228 struct callstack_filter *entry;
2230 entry = malloc(sizeof(*entry) + strlen(tok) + 1);
2231 if (entry == NULL) {
2232 pr_err("Memory allocation failure\n");
2236 strcpy(entry->name, tok);
2237 list_add_tail(&entry->list, &callstack_filters);
2244 int cmd_lock(int argc, const char **argv)
2246 const struct option lock_options[] = {
2247 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2248 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
2249 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
2250 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
2251 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2252 "file", "vmlinux pathname"),
2253 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
2254 "file", "kallsyms pathname"),
2255 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
2259 const struct option info_options[] = {
2260 OPT_BOOLEAN('t', "threads", &info_threads,
2261 "dump thread list in perf.data"),
2262 OPT_BOOLEAN('m', "map", &info_map,
2263 "map of lock instances (address:name table)"),
2264 OPT_PARENT(lock_options)
2267 const struct option report_options[] = {
2268 OPT_STRING('k', "key", &sort_key, "acquired",
2269 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2270 OPT_STRING('F', "field", &output_fields, NULL,
2271 "output fields (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2273 OPT_BOOLEAN('c', "combine-locks", &combine_locks,
2274 "combine locks in the same class"),
2275 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2276 "show per-thread lock stats"),
2277 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2278 OPT_PARENT(lock_options)
2281 struct option contention_options[] = {
2282 OPT_STRING('k', "key", &sort_key, "wait_total",
2283 "key for sorting (contended / wait_total / wait_max / wait_min / avg_wait)"),
2284 OPT_STRING('F', "field", &output_fields, "contended,wait_total,wait_max,avg_wait",
2285 "output fields (contended / wait_total / wait_max / wait_min / avg_wait)"),
2286 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2287 "show per-thread lock stats"),
2288 OPT_BOOLEAN('b', "use-bpf", &use_bpf, "use BPF program to collect lock contention stats"),
2289 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2290 "System-wide collection from all CPUs"),
2291 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2292 "List of cpus to monitor"),
2293 OPT_STRING('p', "pid", &target.pid, "pid",
2294 "Trace on existing process id"),
2295 OPT_STRING(0, "tid", &target.tid, "tid",
2296 "Trace on existing thread id (exclusive to --pid)"),
2297 OPT_CALLBACK('M', "map-nr-entries", &bpf_map_entries, "num",
2298 "Max number of BPF map entries", parse_map_entry),
2299 OPT_CALLBACK(0, "max-stack", &max_stack_depth, "num",
2300 "Set the maximum stack depth when collecting lopck contention, "
2301 "Default: " __stringify(CONTENTION_STACK_DEPTH), parse_max_stack),
2302 OPT_INTEGER(0, "stack-skip", &stack_skip,
2303 "Set the number of stack depth to skip when finding a lock caller, "
2304 "Default: " __stringify(CONTENTION_STACK_SKIP)),
2305 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2306 OPT_BOOLEAN('l', "lock-addr", &show_lock_addrs, "show lock stats by address"),
2307 OPT_CALLBACK('Y', "type-filter", NULL, "FLAGS",
2308 "Filter specific type of locks", parse_lock_type),
2309 OPT_CALLBACK('L', "lock-filter", NULL, "ADDRS/NAMES",
2310 "Filter specific address/symbol of locks", parse_lock_addr),
2311 OPT_CALLBACK('S', "callstack-filter", NULL, "NAMES",
2312 "Filter specific function in the callstack", parse_call_stack),
2313 OPT_BOOLEAN('o', "lock-owner", &show_lock_owner, "show lock owners instead of waiters"),
2314 OPT_PARENT(lock_options)
2317 const char * const info_usage[] = {
2318 "perf lock info [<options>]",
2321 const char *const lock_subcommands[] = { "record", "report", "script",
2322 "info", "contention", NULL };
2323 const char *lock_usage[] = {
2327 const char * const report_usage[] = {
2328 "perf lock report [<options>]",
2331 const char * const contention_usage[] = {
2332 "perf lock contention [<options>]",
2338 for (i = 0; i < LOCKHASH_SIZE; i++)
2339 INIT_HLIST_HEAD(lockhash_table + i);
2341 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
2342 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2344 usage_with_options(lock_usage, lock_options);
2346 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2347 return __cmd_record(argc, argv);
2348 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) {
2349 trace_handler = &report_lock_ops;
2351 argc = parse_options(argc, argv,
2352 report_options, report_usage, 0);
2354 usage_with_options(report_usage, report_options);
2356 rc = __cmd_report(false);
2357 } else if (!strcmp(argv[0], "script")) {
2358 /* Aliased to 'perf script' */
2359 return cmd_script(argc, argv);
2360 } else if (!strcmp(argv[0], "info")) {
2362 argc = parse_options(argc, argv,
2363 info_options, info_usage, 0);
2365 usage_with_options(info_usage, info_options);
2367 /* recycling report_lock_ops */
2368 trace_handler = &report_lock_ops;
2369 rc = __cmd_report(true);
2370 } else if (strlen(argv[0]) > 2 && strstarts("contention", argv[0])) {
2371 trace_handler = &contention_lock_ops;
2372 sort_key = "wait_total";
2373 output_fields = "contended,wait_total,wait_max,avg_wait";
2375 #ifndef HAVE_BPF_SKEL
2376 set_option_nobuild(contention_options, 'b', "use-bpf",
2377 "NO_BPF_SKEL=1", false);
2380 argc = parse_options(argc, argv, contention_options,
2381 contention_usage, 0);
2384 if (check_lock_contention_options(contention_options,
2385 contention_usage) < 0)
2388 rc = __cmd_contention(argc, argv);
2390 usage_with_options(lock_usage, lock_options);