1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin diff command: Analyze two perf.data input files, look up and read
6 * DSOs and symbol information, sort them and produce a diff.
11 #include "util/debug.h"
12 #include "util/event.h"
13 #include "util/hist.h"
14 #include "util/evsel.h"
15 #include "util/evlist.h"
16 #include "util/session.h"
17 #include "util/tool.h"
18 #include "util/sort.h"
19 #include "util/srcline.h"
20 #include "util/symbol.h"
21 #include "util/data.h"
22 #include "util/config.h"
23 #include "util/time-utils.h"
24 #include "util/annotate.h"
26 #include "util/spark.h"
27 #include "util/block-info.h"
28 #include <linux/err.h>
29 #include <linux/zalloc.h>
30 #include <subcmd/pager.h>
31 #include <subcmd/parse-options.h>
39 struct perf_tool tool;
41 struct perf_time_interval *ptime_range;
47 /* Diff command specific HPP columns. */
49 PERF_HPP_DIFF__BASELINE,
50 PERF_HPP_DIFF__PERIOD,
51 PERF_HPP_DIFF__PERIOD_BASELINE,
54 PERF_HPP_DIFF__WEIGHTED_DIFF,
55 PERF_HPP_DIFF__FORMULA,
56 PERF_HPP_DIFF__DELTA_ABS,
57 PERF_HPP_DIFF__CYCLES,
58 PERF_HPP_DIFF__CYCLES_HIST,
60 PERF_HPP_DIFF__MAX_INDEX
64 struct perf_hpp_fmt fmt;
71 struct perf_session *session;
72 struct perf_data data;
75 struct diff_hpp_fmt fmt[PERF_HPP_DIFF__MAX_INDEX];
78 static struct data__file *data__files;
79 static int data__files_cnt;
81 #define data__for_each_file_start(i, d, s) \
82 for (i = s, d = &data__files[s]; \
83 i < data__files_cnt; \
84 i++, d = &data__files[i])
86 #define data__for_each_file(i, d) data__for_each_file_start(i, d, 0)
87 #define data__for_each_file_new(i, d) data__for_each_file_start(i, d, 1)
90 static bool show_period;
91 static bool show_formula;
92 static bool show_baseline_only;
93 static bool cycles_hist;
94 static unsigned int sort_compute = 1;
96 static s64 compute_wdiff_w1;
97 static s64 compute_wdiff_w2;
99 static const char *cpu_list;
100 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
105 COMPUTE_WEIGHTED_DIFF,
111 const char *compute_names[COMPUTE_MAX] = {
112 [COMPUTE_DELTA] = "delta",
113 [COMPUTE_DELTA_ABS] = "delta-abs",
114 [COMPUTE_RATIO] = "ratio",
115 [COMPUTE_WEIGHTED_DIFF] = "wdiff",
116 [COMPUTE_CYCLES] = "cycles",
119 static int compute = COMPUTE_DELTA_ABS;
121 static int compute_2_hpp[COMPUTE_MAX] = {
122 [COMPUTE_DELTA] = PERF_HPP_DIFF__DELTA,
123 [COMPUTE_DELTA_ABS] = PERF_HPP_DIFF__DELTA_ABS,
124 [COMPUTE_RATIO] = PERF_HPP_DIFF__RATIO,
125 [COMPUTE_WEIGHTED_DIFF] = PERF_HPP_DIFF__WEIGHTED_DIFF,
126 [COMPUTE_CYCLES] = PERF_HPP_DIFF__CYCLES,
129 #define MAX_COL_WIDTH 70
131 static struct header_column {
134 } columns[PERF_HPP_DIFF__MAX_INDEX] = {
135 [PERF_HPP_DIFF__BASELINE] = {
138 [PERF_HPP_DIFF__PERIOD] = {
142 [PERF_HPP_DIFF__PERIOD_BASELINE] = {
143 .name = "Base period",
146 [PERF_HPP_DIFF__DELTA] = {
150 [PERF_HPP_DIFF__DELTA_ABS] = {
154 [PERF_HPP_DIFF__RATIO] = {
158 [PERF_HPP_DIFF__WEIGHTED_DIFF] = {
159 .name = "Weighted diff",
162 [PERF_HPP_DIFF__FORMULA] = {
164 .width = MAX_COL_WIDTH,
166 [PERF_HPP_DIFF__CYCLES] = {
167 .name = "[Program Block Range] Cycles Diff",
170 [PERF_HPP_DIFF__CYCLES_HIST] = {
171 .name = "stddev/Hist",
172 .width = NUM_SPARKS + 9,
176 static int setup_compute_opt_wdiff(char *opt)
186 w2_str = strchr(opt, ',');
194 compute_wdiff_w1 = strtol(w1_str, NULL, 10);
195 compute_wdiff_w2 = strtol(w2_str, NULL, 10);
197 if (!compute_wdiff_w1 || !compute_wdiff_w2)
200 pr_debug("compute wdiff w1(%" PRId64 ") w2(%" PRId64 ")\n",
201 compute_wdiff_w1, compute_wdiff_w2);
207 pr_err("Failed: wrong weight data, use 'wdiff:w1,w2'\n");
212 static int setup_compute_opt(char *opt)
214 if (compute == COMPUTE_WEIGHTED_DIFF)
215 return setup_compute_opt_wdiff(opt);
218 pr_err("Failed: extra option specified '%s'", opt);
225 static int setup_compute(const struct option *opt, const char *str,
226 int unset __maybe_unused)
228 int *cp = (int *) opt->value;
229 char *cstr = (char *) str;
239 option = strchr(str, ':');
241 unsigned len = option++ - str;
244 * The str data are not writeable, so we need
245 * to use another buffer.
248 /* No option value is longer. */
249 if (len >= sizeof(buf))
252 strncpy(buf, str, len);
257 for (i = 0; i < COMPUTE_MAX; i++)
258 if (!strcmp(cstr, compute_names[i])) {
260 return setup_compute_opt(option);
263 pr_err("Failed: '%s' is not computation method "
264 "(use 'delta','ratio' or 'wdiff')\n", str);
268 static double period_percent(struct hist_entry *he, u64 period)
270 u64 total = hists__total_period(he->hists);
272 return (period * 100.0) / total;
275 static double compute_delta(struct hist_entry *he, struct hist_entry *pair)
277 double old_percent = period_percent(he, he->stat.period);
278 double new_percent = period_percent(pair, pair->stat.period);
280 pair->diff.period_ratio_delta = new_percent - old_percent;
281 pair->diff.computed = true;
282 return pair->diff.period_ratio_delta;
285 static double compute_ratio(struct hist_entry *he, struct hist_entry *pair)
287 double old_period = he->stat.period ?: 1;
288 double new_period = pair->stat.period;
290 pair->diff.computed = true;
291 pair->diff.period_ratio = new_period / old_period;
292 return pair->diff.period_ratio;
295 static s64 compute_wdiff(struct hist_entry *he, struct hist_entry *pair)
297 u64 old_period = he->stat.period;
298 u64 new_period = pair->stat.period;
300 pair->diff.computed = true;
301 pair->diff.wdiff = new_period * compute_wdiff_w2 -
302 old_period * compute_wdiff_w1;
304 return pair->diff.wdiff;
307 static int formula_delta(struct hist_entry *he, struct hist_entry *pair,
308 char *buf, size_t size)
310 u64 he_total = he->hists->stats.total_period;
311 u64 pair_total = pair->hists->stats.total_period;
313 if (symbol_conf.filter_relative) {
314 he_total = he->hists->stats.total_non_filtered_period;
315 pair_total = pair->hists->stats.total_non_filtered_period;
317 return scnprintf(buf, size,
318 "(%" PRIu64 " * 100 / %" PRIu64 ") - "
319 "(%" PRIu64 " * 100 / %" PRIu64 ")",
320 pair->stat.period, pair_total,
321 he->stat.period, he_total);
324 static int formula_ratio(struct hist_entry *he, struct hist_entry *pair,
325 char *buf, size_t size)
327 double old_period = he->stat.period;
328 double new_period = pair->stat.period;
330 return scnprintf(buf, size, "%.0F / %.0F", new_period, old_period);
333 static int formula_wdiff(struct hist_entry *he, struct hist_entry *pair,
334 char *buf, size_t size)
336 u64 old_period = he->stat.period;
337 u64 new_period = pair->stat.period;
339 return scnprintf(buf, size,
340 "(%" PRIu64 " * " "%" PRId64 ") - (%" PRIu64 " * " "%" PRId64 ")",
341 new_period, compute_wdiff_w2, old_period, compute_wdiff_w1);
344 static int formula_fprintf(struct hist_entry *he, struct hist_entry *pair,
345 char *buf, size_t size)
349 case COMPUTE_DELTA_ABS:
350 return formula_delta(he, pair, buf, size);
352 return formula_ratio(he, pair, buf, size);
353 case COMPUTE_WEIGHTED_DIFF:
354 return formula_wdiff(he, pair, buf, size);
362 static void *block_hist_zalloc(size_t size)
364 struct block_hist *bh;
366 bh = zalloc(size + sizeof(*bh));
373 static void block_hist_free(void *he)
375 struct block_hist *bh;
377 bh = container_of(he, struct block_hist, he);
378 hists__delete_entries(&bh->block_hists);
382 struct hist_entry_ops block_hist_ops = {
383 .new = block_hist_zalloc,
384 .free = block_hist_free,
387 static int diff__process_sample_event(struct perf_tool *tool,
388 union perf_event *event,
389 struct perf_sample *sample,
391 struct machine *machine)
393 struct perf_diff *pdiff = container_of(tool, struct perf_diff, tool);
394 struct addr_location al;
395 struct hists *hists = evsel__hists(evsel);
398 if (perf_time__ranges_skip_sample(pdiff->ptime_range, pdiff->range_num,
403 if (machine__resolve(machine, &al, sample) < 0) {
404 pr_warning("problem processing %d event, skipping it.\n",
409 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap)) {
414 if (compute != COMPUTE_CYCLES) {
415 if (!hists__add_entry(hists, &al, NULL, NULL, NULL, sample,
417 pr_warning("problem incrementing symbol period, "
422 if (!hists__add_entry_ops(hists, &block_hist_ops, &al, NULL,
423 NULL, NULL, sample, true)) {
424 pr_warning("problem incrementing symbol period, "
429 hist__account_cycles(sample->branch_stack, &al, sample, false,
434 * The total_period is updated here before going to the output
435 * tree since normally only the baseline hists will call
436 * hists__output_resort() and precompute needs the total
437 * period in order to sort entries by percentage delta.
439 hists->stats.total_period += sample->period;
441 hists->stats.total_non_filtered_period += sample->period;
444 addr_location__put(&al);
448 static struct perf_diff pdiff = {
450 .sample = diff__process_sample_event,
451 .mmap = perf_event__process_mmap,
452 .mmap2 = perf_event__process_mmap2,
453 .comm = perf_event__process_comm,
454 .exit = perf_event__process_exit,
455 .fork = perf_event__process_fork,
456 .lost = perf_event__process_lost,
457 .namespaces = perf_event__process_namespaces,
458 .ordered_events = true,
459 .ordering_requires_timestamps = true,
463 static struct evsel *evsel_match(struct evsel *evsel,
464 struct evlist *evlist)
468 evlist__for_each_entry(evlist, e) {
469 if (perf_evsel__match2(evsel, e))
476 static void perf_evlist__collapse_resort(struct evlist *evlist)
480 evlist__for_each_entry(evlist, evsel) {
481 struct hists *hists = evsel__hists(evsel);
483 hists__collapse_resort(hists, NULL);
487 static struct data__file *fmt_to_data_file(struct perf_hpp_fmt *fmt)
489 struct diff_hpp_fmt *dfmt = container_of(fmt, struct diff_hpp_fmt, fmt);
490 void *ptr = dfmt - dfmt->idx;
491 struct data__file *d = container_of(ptr, struct data__file, fmt);
496 static struct hist_entry*
497 get_pair_data(struct hist_entry *he, struct data__file *d)
499 if (hist_entry__has_pairs(he)) {
500 struct hist_entry *pair;
502 list_for_each_entry(pair, &he->pairs.head, pairs.node)
503 if (pair->hists == d->hists)
510 static struct hist_entry*
511 get_pair_fmt(struct hist_entry *he, struct diff_hpp_fmt *dfmt)
513 struct data__file *d = fmt_to_data_file(&dfmt->fmt);
515 return get_pair_data(he, d);
518 static void hists__baseline_only(struct hists *hists)
520 struct rb_root_cached *root;
521 struct rb_node *next;
523 if (hists__has(hists, need_collapse))
524 root = &hists->entries_collapsed;
526 root = hists->entries_in;
528 next = rb_first_cached(root);
529 while (next != NULL) {
530 struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node_in);
532 next = rb_next(&he->rb_node_in);
533 if (!hist_entry__next_pair(he)) {
534 rb_erase_cached(&he->rb_node_in, root);
535 hist_entry__delete(he);
540 static int64_t block_cycles_diff_cmp(struct hist_entry *left,
541 struct hist_entry *right)
543 bool pairs_left = hist_entry__has_pairs(left);
544 bool pairs_right = hist_entry__has_pairs(right);
547 if (!pairs_left && !pairs_right)
550 l = labs(left->diff.cycles);
551 r = labs(right->diff.cycles);
555 static int64_t block_sort(struct perf_hpp_fmt *fmt __maybe_unused,
556 struct hist_entry *left, struct hist_entry *right)
558 return block_cycles_diff_cmp(right, left);
561 static void init_block_hist(struct block_hist *bh)
563 __hists__init(&bh->block_hists, &bh->block_list);
564 perf_hpp_list__init(&bh->block_list);
566 INIT_LIST_HEAD(&bh->block_fmt.list);
567 INIT_LIST_HEAD(&bh->block_fmt.sort_list);
568 bh->block_fmt.cmp = block_info__cmp;
569 bh->block_fmt.sort = block_sort;
570 perf_hpp_list__register_sort_field(&bh->block_list,
575 static int block_pair_cmp(struct hist_entry *a, struct hist_entry *b)
577 struct block_info *bi_a = a->block_info;
578 struct block_info *bi_b = b->block_info;
581 if (!bi_a->sym || !bi_b->sym)
584 cmp = strcmp(bi_a->sym->name, bi_b->sym->name);
586 if ((!cmp) && (bi_a->start == bi_b->start) && (bi_a->end == bi_b->end))
592 static struct hist_entry *get_block_pair(struct hist_entry *he,
593 struct hists *hists_pair)
595 struct rb_root_cached *root = hists_pair->entries_in;
596 struct rb_node *next = rb_first_cached(root);
599 while (next != NULL) {
600 struct hist_entry *he_pair = rb_entry(next, struct hist_entry,
603 next = rb_next(&he_pair->rb_node_in);
605 cmp = block_pair_cmp(he_pair, he);
613 static void init_spark_values(unsigned long *svals, int num)
615 for (int i = 0; i < num; i++)
619 static void update_spark_value(unsigned long *svals, int num,
620 struct stats *stats, u64 val)
628 static void compute_cycles_diff(struct hist_entry *he,
629 struct hist_entry *pair)
631 pair->diff.computed = true;
632 if (pair->block_info->num && he->block_info->num) {
634 pair->block_info->cycles_aggr / pair->block_info->num_aggr -
635 he->block_info->cycles_aggr / he->block_info->num_aggr;
640 init_stats(&pair->diff.stats);
641 init_spark_values(pair->diff.svals, NUM_SPARKS);
643 for (int i = 0; i < pair->block_info->num; i++) {
646 if (i >= he->block_info->num || i >= NUM_SPARKS)
649 val = labs(pair->block_info->cycles_spark[i] -
650 he->block_info->cycles_spark[i]);
652 update_spark_value(pair->diff.svals, NUM_SPARKS,
653 &pair->diff.stats, val);
654 update_stats(&pair->diff.stats, val);
659 static void block_hists_match(struct hists *hists_base,
660 struct hists *hists_pair)
662 struct rb_root_cached *root = hists_base->entries_in;
663 struct rb_node *next = rb_first_cached(root);
665 while (next != NULL) {
666 struct hist_entry *he = rb_entry(next, struct hist_entry,
668 struct hist_entry *pair = get_block_pair(he, hists_pair);
670 next = rb_next(&he->rb_node_in);
673 hist_entry__add_pair(pair, he);
674 compute_cycles_diff(he, pair);
679 static void hists__precompute(struct hists *hists)
681 struct rb_root_cached *root;
682 struct rb_node *next;
684 if (hists__has(hists, need_collapse))
685 root = &hists->entries_collapsed;
687 root = hists->entries_in;
689 next = rb_first_cached(root);
690 while (next != NULL) {
691 struct block_hist *bh, *pair_bh;
692 struct hist_entry *he, *pair;
693 struct data__file *d;
696 he = rb_entry(next, struct hist_entry, rb_node_in);
697 next = rb_next(&he->rb_node_in);
699 if (compute == COMPUTE_CYCLES) {
700 bh = container_of(he, struct block_hist, he);
702 block_info__process_sym(he, bh, NULL, 0);
705 data__for_each_file_new(i, d) {
706 pair = get_pair_data(he, d);
712 case COMPUTE_DELTA_ABS:
713 compute_delta(he, pair);
716 compute_ratio(he, pair);
718 case COMPUTE_WEIGHTED_DIFF:
719 compute_wdiff(he, pair);
722 pair_bh = container_of(pair, struct block_hist,
724 init_block_hist(pair_bh);
725 block_info__process_sym(pair, pair_bh, NULL, 0);
727 bh = container_of(he, struct block_hist, he);
729 if (bh->valid && pair_bh->valid) {
730 block_hists_match(&bh->block_hists,
731 &pair_bh->block_hists);
732 hists__output_resort(&pair_bh->block_hists,
743 static int64_t cmp_doubles(double l, double r)
754 __hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
760 double l = left->diff.period_ratio_delta;
761 double r = right->diff.period_ratio_delta;
763 return cmp_doubles(l, r);
765 case COMPUTE_DELTA_ABS:
767 double l = fabs(left->diff.period_ratio_delta);
768 double r = fabs(right->diff.period_ratio_delta);
770 return cmp_doubles(l, r);
774 double l = left->diff.period_ratio;
775 double r = right->diff.period_ratio;
777 return cmp_doubles(l, r);
779 case COMPUTE_WEIGHTED_DIFF:
781 s64 l = left->diff.wdiff;
782 s64 r = right->diff.wdiff;
794 hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
797 bool pairs_left = hist_entry__has_pairs(left);
798 bool pairs_right = hist_entry__has_pairs(right);
799 struct hist_entry *p_right, *p_left;
801 if (!pairs_left && !pairs_right)
804 if (!pairs_left || !pairs_right)
805 return pairs_left ? -1 : 1;
807 p_left = get_pair_data(left, &data__files[sort_idx]);
808 p_right = get_pair_data(right, &data__files[sort_idx]);
810 if (!p_left && !p_right)
813 if (!p_left || !p_right)
814 return p_left ? -1 : 1;
817 * We have 2 entries of same kind, let's
818 * make the data comparison.
820 return __hist_entry__cmp_compute(p_left, p_right, c);
824 hist_entry__cmp_compute_idx(struct hist_entry *left, struct hist_entry *right,
827 struct hist_entry *p_right, *p_left;
829 p_left = get_pair_data(left, &data__files[sort_idx]);
830 p_right = get_pair_data(right, &data__files[sort_idx]);
832 if (!p_left && !p_right)
835 if (!p_left || !p_right)
836 return p_left ? -1 : 1;
838 if (c != COMPUTE_DELTA && c != COMPUTE_DELTA_ABS) {
840 * The delta can be computed without the baseline, but
841 * others are not. Put those entries which have no
844 if (left->dummy && right->dummy)
847 if (left->dummy || right->dummy)
848 return left->dummy ? 1 : -1;
851 return __hist_entry__cmp_compute(p_left, p_right, c);
855 hist_entry__cmp_nop(struct perf_hpp_fmt *fmt __maybe_unused,
856 struct hist_entry *left __maybe_unused,
857 struct hist_entry *right __maybe_unused)
863 hist_entry__cmp_baseline(struct perf_hpp_fmt *fmt __maybe_unused,
864 struct hist_entry *left, struct hist_entry *right)
866 if (left->stat.period == right->stat.period)
868 return left->stat.period > right->stat.period ? 1 : -1;
872 hist_entry__cmp_delta(struct perf_hpp_fmt *fmt,
873 struct hist_entry *left, struct hist_entry *right)
875 struct data__file *d = fmt_to_data_file(fmt);
877 return hist_entry__cmp_compute(right, left, COMPUTE_DELTA, d->idx);
881 hist_entry__cmp_delta_abs(struct perf_hpp_fmt *fmt,
882 struct hist_entry *left, struct hist_entry *right)
884 struct data__file *d = fmt_to_data_file(fmt);
886 return hist_entry__cmp_compute(right, left, COMPUTE_DELTA_ABS, d->idx);
890 hist_entry__cmp_ratio(struct perf_hpp_fmt *fmt,
891 struct hist_entry *left, struct hist_entry *right)
893 struct data__file *d = fmt_to_data_file(fmt);
895 return hist_entry__cmp_compute(right, left, COMPUTE_RATIO, d->idx);
899 hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt,
900 struct hist_entry *left, struct hist_entry *right)
902 struct data__file *d = fmt_to_data_file(fmt);
904 return hist_entry__cmp_compute(right, left, COMPUTE_WEIGHTED_DIFF, d->idx);
908 hist_entry__cmp_delta_idx(struct perf_hpp_fmt *fmt __maybe_unused,
909 struct hist_entry *left, struct hist_entry *right)
911 return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA,
916 hist_entry__cmp_delta_abs_idx(struct perf_hpp_fmt *fmt __maybe_unused,
917 struct hist_entry *left, struct hist_entry *right)
919 return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA_ABS,
924 hist_entry__cmp_ratio_idx(struct perf_hpp_fmt *fmt __maybe_unused,
925 struct hist_entry *left, struct hist_entry *right)
927 return hist_entry__cmp_compute_idx(right, left, COMPUTE_RATIO,
932 hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused,
933 struct hist_entry *left, struct hist_entry *right)
935 return hist_entry__cmp_compute_idx(right, left, COMPUTE_WEIGHTED_DIFF,
939 static void hists__process(struct hists *hists)
941 if (show_baseline_only)
942 hists__baseline_only(hists);
944 hists__precompute(hists);
945 hists__output_resort(hists, NULL);
947 if (compute == COMPUTE_CYCLES)
948 symbol_conf.report_block = true;
950 hists__fprintf(hists, !quiet, 0, 0, 0, stdout,
951 !symbol_conf.use_callchain);
954 static void data__fprintf(void)
956 struct data__file *d;
959 fprintf(stdout, "# Data files:\n");
961 data__for_each_file(i, d)
962 fprintf(stdout, "# [%d] %s %s\n",
963 d->idx, d->data.path,
964 !d->idx ? "(Baseline)" : "");
966 fprintf(stdout, "#\n");
969 static void data_process(void)
971 struct evlist *evlist_base = data__files[0].session->evlist;
972 struct evsel *evsel_base;
975 evlist__for_each_entry(evlist_base, evsel_base) {
976 struct hists *hists_base = evsel__hists(evsel_base);
977 struct data__file *d;
980 data__for_each_file_new(i, d) {
981 struct evlist *evlist = d->session->evlist;
985 evsel = evsel_match(evsel_base, evlist);
989 hists = evsel__hists(evsel);
992 hists__match(hists_base, hists);
994 if (!show_baseline_only)
995 hists__link(hists_base, hists);
999 fprintf(stdout, "%s# Event '%s'\n#\n", first ? "" : "\n",
1000 perf_evsel__name(evsel_base));
1005 if (verbose > 0 || ((data__files_cnt > 2) && !quiet))
1008 /* Don't sort callchain for perf diff */
1009 perf_evsel__reset_sample_bit(evsel_base, CALLCHAIN);
1011 hists__process(hists_base);
1015 static void data__free(struct data__file *d)
1019 for (col = 0; col < PERF_HPP_DIFF__MAX_INDEX; col++) {
1020 struct diff_hpp_fmt *fmt = &d->fmt[col];
1022 zfree(&fmt->header);
1026 static int abstime_str_dup(char **pstr)
1030 if (pdiff.time_str && strchr(pdiff.time_str, ':')) {
1031 str = strdup(pdiff.time_str);
1040 static int parse_absolute_time(struct data__file *d, char **pstr)
1046 * Absolute timestamp for one file has the format: a.b,c.d
1047 * For multiple files, the format is: a.b,c.d:a.b,c.d
1049 p = strchr(*pstr, ':');
1052 pr_err("Invalid time string\n");
1059 pr_err("Invalid time string\n");
1064 ret = perf_time__parse_for_ranges(*pstr, d->session,
1079 static int parse_percent_time(struct data__file *d)
1083 ret = perf_time__parse_for_ranges(pdiff.time_str, d->session,
1090 static int parse_time_str(struct data__file *d, char *abstime_ostr,
1091 char **pabstime_tmp)
1096 ret = parse_absolute_time(d, pabstime_tmp);
1097 else if (pdiff.time_str)
1098 ret = parse_percent_time(d);
1103 static int check_file_brstack(void)
1105 struct data__file *d;
1109 data__for_each_file(i, d) {
1110 d->session = perf_session__new(&d->data, false, &pdiff.tool);
1111 if (IS_ERR(d->session)) {
1112 pr_err("Failed to open %s\n", d->data.path);
1113 return PTR_ERR(d->session);
1116 has_br_stack = perf_header__has_feat(&d->session->header,
1117 HEADER_BRANCH_STACK);
1118 perf_session__delete(d->session);
1123 /* Set only all files having branch stacks */
1124 pdiff.has_br_stack = true;
1128 static int __cmd_diff(void)
1130 struct data__file *d;
1132 char *abstime_ostr, *abstime_tmp;
1134 ret = abstime_str_dup(&abstime_ostr);
1138 abstime_tmp = abstime_ostr;
1141 data__for_each_file(i, d) {
1142 d->session = perf_session__new(&d->data, false, &pdiff.tool);
1143 if (IS_ERR(d->session)) {
1144 ret = PTR_ERR(d->session);
1145 pr_err("Failed to open %s\n", d->data.path);
1149 if (pdiff.time_str) {
1150 ret = parse_time_str(d, abstime_ostr, &abstime_tmp);
1156 ret = perf_session__cpu_bitmap(d->session, cpu_list,
1162 ret = perf_session__process_events(d->session);
1164 pr_err("Failed to process %s\n", d->data.path);
1168 perf_evlist__collapse_resort(d->session->evlist);
1170 if (pdiff.ptime_range)
1171 zfree(&pdiff.ptime_range);
1177 data__for_each_file(i, d) {
1178 perf_session__delete(d->session);
1184 if (pdiff.ptime_range)
1185 zfree(&pdiff.ptime_range);
1193 static const char * const diff_usage[] = {
1194 "perf diff [<options>] [old_file] [new_file]",
1198 static const struct option options[] = {
1199 OPT_INCR('v', "verbose", &verbose,
1200 "be more verbose (show symbol address, etc)"),
1201 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1202 OPT_BOOLEAN('b', "baseline-only", &show_baseline_only,
1203 "Show only items with match in baseline"),
1204 OPT_CALLBACK('c', "compute", &compute,
1205 "delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs),cycles",
1206 "Entries differential computation selection",
1208 OPT_BOOLEAN('p', "period", &show_period,
1209 "Show period values."),
1210 OPT_BOOLEAN('F', "formula", &show_formula,
1212 OPT_BOOLEAN(0, "cycles-hist", &cycles_hist,
1213 "Show cycles histogram and standard deviation "
1214 "- WARNING: use only with -c cycles."),
1215 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1216 "dump raw trace in ASCII"),
1217 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1218 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1219 "file", "kallsyms pathname"),
1220 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1221 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1222 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1223 "only consider symbols in these dsos"),
1224 OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1225 "only consider symbols in these comms"),
1226 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1227 "only consider these symbols"),
1228 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1229 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1230 " Please refer the man page for the complete list."),
1231 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1232 "separator for columns, no spaces will be added between "
1233 "columns '.' is reserved."),
1234 OPT_CALLBACK(0, "symfs", NULL, "directory",
1235 "Look for files with symbols relative to this directory",
1236 symbol__config_symfs),
1237 OPT_UINTEGER('o', "order", &sort_compute, "Specify compute sorting."),
1238 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1239 "How to display percentage of filtered entries", parse_filter_percentage),
1240 OPT_STRING(0, "time", &pdiff.time_str, "str",
1241 "Time span (time percent or absolute timestamp)"),
1242 OPT_STRING(0, "cpu", &cpu_list, "cpu", "list of cpus to profile"),
1243 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1244 "only consider symbols in these pids"),
1245 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1246 "only consider symbols in these tids"),
1250 static double baseline_percent(struct hist_entry *he)
1252 u64 total = hists__total_period(he->hists);
1254 return 100.0 * he->stat.period / total;
1257 static int hpp__color_baseline(struct perf_hpp_fmt *fmt,
1258 struct perf_hpp *hpp, struct hist_entry *he)
1260 struct diff_hpp_fmt *dfmt =
1261 container_of(fmt, struct diff_hpp_fmt, fmt);
1262 double percent = baseline_percent(he);
1263 char pfmt[20] = " ";
1266 scnprintf(pfmt, 20, "%%%d.2f%%%%", dfmt->header_width - 1);
1267 return percent_color_snprintf(hpp->buf, hpp->size,
1270 return scnprintf(hpp->buf, hpp->size, "%*s",
1271 dfmt->header_width, pfmt);
1274 static int hpp__entry_baseline(struct hist_entry *he, char *buf, size_t size)
1276 double percent = baseline_percent(he);
1277 const char *fmt = symbol_conf.field_sep ? "%.2f" : "%6.2f%%";
1281 ret = scnprintf(buf, size, fmt, percent);
1286 static int cycles_printf(struct hist_entry *he, struct hist_entry *pair,
1287 struct perf_hpp *hpp, int width)
1289 struct block_hist *bh = container_of(he, struct block_hist, he);
1290 struct block_hist *bh_pair = container_of(pair, struct block_hist, he);
1291 struct hist_entry *block_he;
1292 struct block_info *bi;
1294 char *start_line, *end_line;
1296 block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1303 * Avoid printing the warning "addr2line_init failed for ..."
1305 symbol_conf.disable_add2line_warn = true;
1307 bi = block_he->block_info;
1309 start_line = map__srcline(he->ms.map, bi->sym->start + bi->start,
1312 end_line = map__srcline(he->ms.map, bi->sym->start + bi->end,
1315 if ((start_line != SRCLINE_UNKNOWN) && (end_line != SRCLINE_UNKNOWN)) {
1316 scnprintf(buf, sizeof(buf), "[%s -> %s] %4ld",
1317 start_line, end_line, block_he->diff.cycles);
1319 scnprintf(buf, sizeof(buf), "[%7lx -> %7lx] %4ld",
1320 bi->start, bi->end, block_he->diff.cycles);
1323 free_srcline(start_line);
1324 free_srcline(end_line);
1326 return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1329 static int __hpp__color_compare(struct perf_hpp_fmt *fmt,
1330 struct perf_hpp *hpp, struct hist_entry *he,
1331 int comparison_method)
1333 struct diff_hpp_fmt *dfmt =
1334 container_of(fmt, struct diff_hpp_fmt, fmt);
1335 struct hist_entry *pair = get_pair_fmt(he, dfmt);
1338 char pfmt[20] = " ";
1341 if (comparison_method == COMPUTE_CYCLES) {
1342 struct block_hist *bh;
1344 bh = container_of(he, struct block_hist, he);
1352 switch (comparison_method) {
1354 if (pair->diff.computed)
1355 diff = pair->diff.period_ratio_delta;
1357 diff = compute_delta(he, pair);
1359 scnprintf(pfmt, 20, "%%%+d.2f%%%%", dfmt->header_width - 1);
1360 return percent_color_snprintf(hpp->buf, hpp->size,
1365 if (pair->diff.computed)
1366 diff = pair->diff.period_ratio;
1368 diff = compute_ratio(he, pair);
1370 scnprintf(pfmt, 20, "%%%d.6f", dfmt->header_width);
1371 return value_color_snprintf(hpp->buf, hpp->size,
1373 case COMPUTE_WEIGHTED_DIFF:
1376 if (pair->diff.computed)
1377 wdiff = pair->diff.wdiff;
1379 wdiff = compute_wdiff(he, pair);
1381 scnprintf(pfmt, 20, "%%14ld", dfmt->header_width);
1382 return color_snprintf(hpp->buf, hpp->size,
1383 get_percent_color(wdiff),
1385 case COMPUTE_CYCLES:
1386 return cycles_printf(he, pair, hpp, dfmt->header_width);
1391 return scnprintf(hpp->buf, hpp->size, "%*s",
1392 dfmt->header_width, "N/A");
1394 return scnprintf(hpp->buf, hpp->size, "%*s",
1395 dfmt->header_width, pfmt);
1398 static int hpp__color_delta(struct perf_hpp_fmt *fmt,
1399 struct perf_hpp *hpp, struct hist_entry *he)
1401 return __hpp__color_compare(fmt, hpp, he, COMPUTE_DELTA);
1404 static int hpp__color_ratio(struct perf_hpp_fmt *fmt,
1405 struct perf_hpp *hpp, struct hist_entry *he)
1407 return __hpp__color_compare(fmt, hpp, he, COMPUTE_RATIO);
1410 static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,
1411 struct perf_hpp *hpp, struct hist_entry *he)
1413 return __hpp__color_compare(fmt, hpp, he, COMPUTE_WEIGHTED_DIFF);
1416 static int hpp__color_cycles(struct perf_hpp_fmt *fmt,
1417 struct perf_hpp *hpp, struct hist_entry *he)
1419 return __hpp__color_compare(fmt, hpp, he, COMPUTE_CYCLES);
1422 static int all_zero(unsigned long *vals, int len)
1426 for (i = 0; i < len; i++)
1432 static int print_cycles_spark(char *bf, int size, unsigned long *svals, u64 n)
1441 if (all_zero(svals, n))
1444 printed = print_spark(bf, size, svals, n);
1445 printed += scnprintf(bf + printed, size - printed, " ");
1449 static int hpp__color_cycles_hist(struct perf_hpp_fmt *fmt,
1450 struct perf_hpp *hpp, struct hist_entry *he)
1452 struct diff_hpp_fmt *dfmt =
1453 container_of(fmt, struct diff_hpp_fmt, fmt);
1454 struct hist_entry *pair = get_pair_fmt(he, dfmt);
1455 struct block_hist *bh = container_of(he, struct block_hist, he);
1456 struct block_hist *bh_pair;
1457 struct hist_entry *block_he;
1458 char spark[32], buf[128];
1469 bh_pair = container_of(pair, struct block_hist, he);
1471 block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1477 ret = print_cycles_spark(spark, sizeof(spark), block_he->diff.svals,
1478 block_he->diff.stats.n);
1480 r = rel_stddev_stats(stddev_stats(&block_he->diff.stats),
1481 avg_stats(&block_he->diff.stats));
1485 * Padding spaces if number of sparks less than NUM_SPARKS
1486 * otherwise the output is not aligned.
1488 pad = NUM_SPARKS - ((ret - 1) / 3);
1489 scnprintf(buf, sizeof(buf), "%s%5.1f%% %s", "\u00B1", r, spark);
1490 ret = scnprintf(hpp->buf, hpp->size, "%*s",
1491 dfmt->header_width, buf);
1494 ret += scnprintf(hpp->buf + ret, hpp->size - ret,
1502 return scnprintf(hpp->buf, hpp->size, "%*s",
1503 dfmt->header_width, " ");
1507 hpp__entry_unpair(struct hist_entry *he, int idx, char *buf, size_t size)
1510 case PERF_HPP_DIFF__PERIOD_BASELINE:
1511 scnprintf(buf, size, "%" PRIu64, he->stat.period);
1520 hpp__entry_pair(struct hist_entry *he, struct hist_entry *pair,
1521 int idx, char *buf, size_t size)
1528 case PERF_HPP_DIFF__DELTA:
1529 case PERF_HPP_DIFF__DELTA_ABS:
1530 if (pair->diff.computed)
1531 diff = pair->diff.period_ratio_delta;
1533 diff = compute_delta(he, pair);
1535 scnprintf(buf, size, "%+4.2F%%", diff);
1538 case PERF_HPP_DIFF__RATIO:
1539 /* No point for ratio number if we are dummy.. */
1541 scnprintf(buf, size, "N/A");
1545 if (pair->diff.computed)
1546 ratio = pair->diff.period_ratio;
1548 ratio = compute_ratio(he, pair);
1551 scnprintf(buf, size, "%14.6F", ratio);
1554 case PERF_HPP_DIFF__WEIGHTED_DIFF:
1555 /* No point for wdiff number if we are dummy.. */
1557 scnprintf(buf, size, "N/A");
1561 if (pair->diff.computed)
1562 wdiff = pair->diff.wdiff;
1564 wdiff = compute_wdiff(he, pair);
1567 scnprintf(buf, size, "%14ld", wdiff);
1570 case PERF_HPP_DIFF__FORMULA:
1571 formula_fprintf(he, pair, buf, size);
1574 case PERF_HPP_DIFF__PERIOD:
1575 scnprintf(buf, size, "%" PRIu64, pair->stat.period);
1584 __hpp__entry_global(struct hist_entry *he, struct diff_hpp_fmt *dfmt,
1585 char *buf, size_t size)
1587 struct hist_entry *pair = get_pair_fmt(he, dfmt);
1588 int idx = dfmt->idx;
1590 /* baseline is special */
1591 if (idx == PERF_HPP_DIFF__BASELINE)
1592 hpp__entry_baseline(he, buf, size);
1595 hpp__entry_pair(he, pair, idx, buf, size);
1597 hpp__entry_unpair(he, idx, buf, size);
1601 static int hpp__entry_global(struct perf_hpp_fmt *_fmt, struct perf_hpp *hpp,
1602 struct hist_entry *he)
1604 struct diff_hpp_fmt *dfmt =
1605 container_of(_fmt, struct diff_hpp_fmt, fmt);
1606 char buf[MAX_COL_WIDTH] = " ";
1608 __hpp__entry_global(he, dfmt, buf, MAX_COL_WIDTH);
1610 if (symbol_conf.field_sep)
1611 return scnprintf(hpp->buf, hpp->size, "%s", buf);
1613 return scnprintf(hpp->buf, hpp->size, "%*s",
1614 dfmt->header_width, buf);
1617 static int hpp__header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1618 struct hists *hists __maybe_unused,
1619 int line __maybe_unused,
1620 int *span __maybe_unused)
1622 struct diff_hpp_fmt *dfmt =
1623 container_of(fmt, struct diff_hpp_fmt, fmt);
1625 BUG_ON(!dfmt->header);
1626 return scnprintf(hpp->buf, hpp->size, dfmt->header);
1629 static int hpp__width(struct perf_hpp_fmt *fmt,
1630 struct perf_hpp *hpp __maybe_unused,
1631 struct hists *hists __maybe_unused)
1633 struct diff_hpp_fmt *dfmt =
1634 container_of(fmt, struct diff_hpp_fmt, fmt);
1636 BUG_ON(dfmt->header_width <= 0);
1637 return dfmt->header_width;
1640 static void init_header(struct data__file *d, struct diff_hpp_fmt *dfmt)
1642 #define MAX_HEADER_NAME 100
1643 char buf_indent[MAX_HEADER_NAME];
1644 char buf[MAX_HEADER_NAME];
1645 const char *header = NULL;
1648 BUG_ON(dfmt->idx >= PERF_HPP_DIFF__MAX_INDEX);
1649 header = columns[dfmt->idx].name;
1650 width = columns[dfmt->idx].width;
1652 /* Only our defined HPP fmts should appear here. */
1655 if (data__files_cnt > 2)
1656 scnprintf(buf, MAX_HEADER_NAME, "%s/%d", header, d->idx);
1658 #define NAME (data__files_cnt > 2 ? buf : header)
1659 dfmt->header_width = width;
1660 width = (int) strlen(NAME);
1661 if (dfmt->header_width < width)
1662 dfmt->header_width = width;
1664 scnprintf(buf_indent, MAX_HEADER_NAME, "%*s",
1665 dfmt->header_width, NAME);
1667 dfmt->header = strdup(buf_indent);
1668 #undef MAX_HEADER_NAME
1672 static void data__hpp_register(struct data__file *d, int idx)
1674 struct diff_hpp_fmt *dfmt = &d->fmt[idx];
1675 struct perf_hpp_fmt *fmt = &dfmt->fmt;
1679 fmt->header = hpp__header;
1680 fmt->width = hpp__width;
1681 fmt->entry = hpp__entry_global;
1682 fmt->cmp = hist_entry__cmp_nop;
1683 fmt->collapse = hist_entry__cmp_nop;
1685 /* TODO more colors */
1687 case PERF_HPP_DIFF__BASELINE:
1688 fmt->color = hpp__color_baseline;
1689 fmt->sort = hist_entry__cmp_baseline;
1691 case PERF_HPP_DIFF__DELTA:
1692 fmt->color = hpp__color_delta;
1693 fmt->sort = hist_entry__cmp_delta;
1695 case PERF_HPP_DIFF__RATIO:
1696 fmt->color = hpp__color_ratio;
1697 fmt->sort = hist_entry__cmp_ratio;
1699 case PERF_HPP_DIFF__WEIGHTED_DIFF:
1700 fmt->color = hpp__color_wdiff;
1701 fmt->sort = hist_entry__cmp_wdiff;
1703 case PERF_HPP_DIFF__DELTA_ABS:
1704 fmt->color = hpp__color_delta;
1705 fmt->sort = hist_entry__cmp_delta_abs;
1707 case PERF_HPP_DIFF__CYCLES:
1708 fmt->color = hpp__color_cycles;
1709 fmt->sort = hist_entry__cmp_nop;
1711 case PERF_HPP_DIFF__CYCLES_HIST:
1712 fmt->color = hpp__color_cycles_hist;
1713 fmt->sort = hist_entry__cmp_nop;
1716 fmt->sort = hist_entry__cmp_nop;
1720 init_header(d, dfmt);
1721 perf_hpp__column_register(fmt);
1722 perf_hpp__register_sort_field(fmt);
1725 static int ui_init(void)
1727 struct data__file *d;
1728 struct perf_hpp_fmt *fmt;
1731 data__for_each_file(i, d) {
1734 * Baseline or compute realted columns:
1736 * PERF_HPP_DIFF__BASELINE
1737 * PERF_HPP_DIFF__DELTA
1738 * PERF_HPP_DIFF__RATIO
1739 * PERF_HPP_DIFF__WEIGHTED_DIFF
1740 * PERF_HPP_DIFF__CYCLES
1742 data__hpp_register(d, i ? compute_2_hpp[compute] :
1743 PERF_HPP_DIFF__BASELINE);
1745 if (cycles_hist && i)
1746 data__hpp_register(d, PERF_HPP_DIFF__CYCLES_HIST);
1751 * PERF_HPP_DIFF__FORMULA
1752 * PERF_HPP_DIFF__PERIOD
1753 * PERF_HPP_DIFF__PERIOD_BASELINE
1755 if (show_formula && i)
1756 data__hpp_register(d, PERF_HPP_DIFF__FORMULA);
1759 data__hpp_register(d, i ? PERF_HPP_DIFF__PERIOD :
1760 PERF_HPP_DIFF__PERIOD_BASELINE);
1767 * Prepend an fmt to sort on columns at 'sort_compute' first.
1768 * This fmt is added only to the sort list but not to the
1769 * output fields list.
1771 * Note that this column (data) can be compared twice - one
1772 * for this 'sort_compute' fmt and another for the normal
1773 * diff_hpp_fmt. But it shouldn't a problem as most entries
1774 * will be sorted out by first try or baseline and comparing
1775 * is not a costly operation.
1777 fmt = zalloc(sizeof(*fmt));
1779 pr_err("Memory allocation failed\n");
1783 fmt->cmp = hist_entry__cmp_nop;
1784 fmt->collapse = hist_entry__cmp_nop;
1788 fmt->sort = hist_entry__cmp_delta_idx;
1791 fmt->sort = hist_entry__cmp_ratio_idx;
1793 case COMPUTE_WEIGHTED_DIFF:
1794 fmt->sort = hist_entry__cmp_wdiff_idx;
1796 case COMPUTE_DELTA_ABS:
1797 fmt->sort = hist_entry__cmp_delta_abs_idx;
1799 case COMPUTE_CYCLES:
1801 * Should set since 'fmt->sort' is called without
1802 * checking valid during sorting
1804 fmt->sort = hist_entry__cmp_nop;
1810 perf_hpp__prepend_sort_field(fmt);
1814 static int data_init(int argc, const char **argv)
1816 struct data__file *d;
1817 static const char *defaults[] = {
1821 bool use_default = true;
1824 data__files_cnt = 2;
1828 defaults[1] = argv[0];
1830 data__files_cnt = argc;
1831 use_default = false;
1833 } else if (perf_guest) {
1834 defaults[0] = "perf.data.host";
1835 defaults[1] = "perf.data.guest";
1838 if (sort_compute >= (unsigned int) data__files_cnt) {
1839 pr_err("Order option out of limit.\n");
1843 data__files = zalloc(sizeof(*data__files) * data__files_cnt);
1847 data__for_each_file(i, d) {
1848 struct perf_data *data = &d->data;
1850 data->path = use_default ? defaults[i] : argv[i];
1851 data->mode = PERF_DATA_MODE_READ,
1852 data->force = force,
1860 static int diff__config(const char *var, const char *value,
1861 void *cb __maybe_unused)
1863 if (!strcmp(var, "diff.order")) {
1865 if (perf_config_int(&ret, var, value) < 0)
1870 if (!strcmp(var, "diff.compute")) {
1871 if (!strcmp(value, "delta")) {
1872 compute = COMPUTE_DELTA;
1873 } else if (!strcmp(value, "delta-abs")) {
1874 compute = COMPUTE_DELTA_ABS;
1875 } else if (!strcmp(value, "ratio")) {
1876 compute = COMPUTE_RATIO;
1877 } else if (!strcmp(value, "wdiff")) {
1878 compute = COMPUTE_WEIGHTED_DIFF;
1880 pr_err("Invalid compute method: %s\n", value);
1888 int cmd_diff(int argc, const char **argv)
1890 int ret = hists__init();
1895 perf_config(diff__config, NULL);
1897 argc = parse_options(argc, argv, options, diff_usage, 0);
1900 perf_quiet_option();
1902 if (cycles_hist && (compute != COMPUTE_CYCLES))
1903 usage_with_options(diff_usage, options);
1905 symbol__annotation_init();
1907 if (symbol__init(NULL) < 0)
1910 if (data_init(argc, argv) < 0)
1913 if (check_file_brstack() < 0)
1916 if (compute == COMPUTE_CYCLES && !pdiff.has_br_stack)
1922 sort__mode = SORT_MODE__DIFF;
1924 if (setup_sorting(NULL) < 0)
1925 usage_with_options(diff_usage, options);
1929 sort__setup_elide(NULL);
1931 return __cmd_diff();