2 * This is rewrite of original c2c tool introduced in here:
3 * http://lwn.net/Articles/588866/
5 * The original tool was changed to fit in current perf state.
8 * Don Zickus <dzickus@redhat.com>
9 * Dick Fowles <fowles@inreach.com>
10 * Joe Mario <jmario@redhat.com>
12 #include <linux/compiler.h>
13 #include <linux/kernel.h>
14 #include <linux/stringify.h>
19 #include <subcmd/parse-options.h>
20 #include "mem-events.h"
30 #include "ui/browsers/hists.h"
35 struct perf_hpp_list list
;
36 struct c2c_stats stats
;
39 struct compute_stats
{
40 struct stats lcl_hitm
;
41 struct stats rmt_hitm
;
45 struct c2c_hist_entry
{
46 struct c2c_hists
*hists
;
47 struct c2c_stats stats
;
48 unsigned long *cpuset
;
49 struct c2c_stats
*node_stats
;
50 unsigned int cacheline_idx
;
52 struct compute_stats cstats
;
56 * because of its callchain dynamic entry
61 static char const *coalesce_default
= "pid,iaddr";
64 struct perf_tool tool
;
65 struct c2c_hists hists
;
67 unsigned long **nodes
;
79 /* HITM shared clines stats */
80 struct c2c_stats hitm_stats
;
98 static const char *display_str
[DISPLAY_MAX
] = {
99 [DISPLAY_LCL
] = "Local",
100 [DISPLAY_RMT
] = "Remote",
101 [DISPLAY_TOT
] = "Total",
104 static const struct option c2c_options
[] = {
105 OPT_INCR('v', "verbose", &verbose
, "be more verbose (show counter open errors, etc)"),
109 static struct perf_c2c c2c
;
111 static void *c2c_he_zalloc(size_t size
)
113 struct c2c_hist_entry
*c2c_he
;
115 c2c_he
= zalloc(size
+ sizeof(*c2c_he
));
119 c2c_he
->cpuset
= bitmap_alloc(c2c
.cpus_cnt
);
123 c2c_he
->node_stats
= zalloc(c2c
.nodes_cnt
* sizeof(*c2c_he
->node_stats
));
124 if (!c2c_he
->node_stats
)
127 init_stats(&c2c_he
->cstats
.lcl_hitm
);
128 init_stats(&c2c_he
->cstats
.rmt_hitm
);
129 init_stats(&c2c_he
->cstats
.load
);
134 static void c2c_he_free(void *he
)
136 struct c2c_hist_entry
*c2c_he
;
138 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
140 hists__delete_entries(&c2c_he
->hists
->hists
);
144 free(c2c_he
->cpuset
);
145 free(c2c_he
->node_stats
);
149 static struct hist_entry_ops c2c_entry_ops
= {
150 .new = c2c_he_zalloc
,
154 static int c2c_hists__init(struct c2c_hists
*hists
,
156 int nr_header_lines
);
158 static struct c2c_hists
*
159 he__get_c2c_hists(struct hist_entry
*he
,
163 struct c2c_hist_entry
*c2c_he
;
164 struct c2c_hists
*hists
;
167 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
169 return c2c_he
->hists
;
171 hists
= c2c_he
->hists
= zalloc(sizeof(*hists
));
175 ret
= c2c_hists__init(hists
, sort
, nr_header_lines
);
184 static void c2c_he__set_cpu(struct c2c_hist_entry
*c2c_he
,
185 struct perf_sample
*sample
)
187 if (WARN_ONCE(sample
->cpu
== (unsigned int) -1,
188 "WARNING: no sample cpu value"))
191 set_bit(sample
->cpu
, c2c_he
->cpuset
);
194 static void compute_stats(struct c2c_hist_entry
*c2c_he
,
195 struct c2c_stats
*stats
,
198 struct compute_stats
*cstats
= &c2c_he
->cstats
;
201 update_stats(&cstats
->rmt_hitm
, weight
);
202 else if (stats
->lcl_hitm
)
203 update_stats(&cstats
->lcl_hitm
, weight
);
204 else if (stats
->load
)
205 update_stats(&cstats
->load
, weight
);
208 static int process_sample_event(struct perf_tool
*tool __maybe_unused
,
209 union perf_event
*event
,
210 struct perf_sample
*sample
,
211 struct perf_evsel
*evsel
,
212 struct machine
*machine
)
214 struct c2c_hists
*c2c_hists
= &c2c
.hists
;
215 struct c2c_hist_entry
*c2c_he
;
216 struct c2c_stats stats
= { .nr_entries
= 0, };
217 struct hist_entry
*he
;
218 struct addr_location al
;
219 struct mem_info
*mi
, *mi_dup
;
222 if (machine__resolve(machine
, &al
, sample
) < 0) {
223 pr_debug("problem processing %d event, skipping it.\n",
228 ret
= sample__resolve_callchain(sample
, &callchain_cursor
, NULL
,
229 evsel
, &al
, sysctl_perf_event_max_stack
);
233 mi
= sample__resolve_mem(sample
, &al
);
237 mi_dup
= memdup(mi
, sizeof(*mi
));
241 c2c_decode_stats(&stats
, mi
);
243 he
= hists__add_entry_ops(&c2c_hists
->hists
, &c2c_entry_ops
,
249 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
250 c2c_add_stats(&c2c_he
->stats
, &stats
);
251 c2c_add_stats(&c2c_hists
->stats
, &stats
);
253 c2c_he__set_cpu(c2c_he
, sample
);
255 hists__inc_nr_samples(&c2c_hists
->hists
, he
->filtered
);
256 ret
= hist_entry__append_callchain(he
, sample
);
260 * There's already been warning about missing
261 * sample's cpu value. Let's account all to
262 * node 0 in this case, without any further
265 * Doing node stats only for single callchain data.
267 int cpu
= sample
->cpu
== (unsigned int) -1 ? 0 : sample
->cpu
;
268 int node
= c2c
.cpu2node
[cpu
];
272 mi_dup
= memdup(mi
, sizeof(*mi
));
276 c2c_hists
= he__get_c2c_hists(he
, c2c
.cl_sort
, 2);
280 he
= hists__add_entry_ops(&c2c_hists
->hists
, &c2c_entry_ops
,
286 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
287 c2c_add_stats(&c2c_he
->stats
, &stats
);
288 c2c_add_stats(&c2c_hists
->stats
, &stats
);
289 c2c_add_stats(&c2c_he
->node_stats
[node
], &stats
);
291 compute_stats(c2c_he
, &stats
, sample
->weight
);
293 c2c_he__set_cpu(c2c_he
, sample
);
295 hists__inc_nr_samples(&c2c_hists
->hists
, he
->filtered
);
296 ret
= hist_entry__append_callchain(he
, sample
);
300 addr_location__put(&al
);
311 static struct perf_c2c c2c
= {
313 .sample
= process_sample_event
,
314 .mmap
= perf_event__process_mmap
,
315 .mmap2
= perf_event__process_mmap2
,
316 .comm
= perf_event__process_comm
,
317 .exit
= perf_event__process_exit
,
318 .fork
= perf_event__process_fork
,
319 .lost
= perf_event__process_lost
,
320 .ordered_events
= true,
321 .ordering_requires_timestamps
= true,
325 static const char * const c2c_usage
[] = {
326 "perf c2c {record|report}",
330 static const char * const __usage_report
[] = {
335 static const char * const *report_c2c_usage
= __usage_report
;
337 #define C2C_HEADER_MAX 2
343 } line
[C2C_HEADER_MAX
];
346 struct c2c_dimension
{
347 struct c2c_header header
;
350 struct sort_entry
*se
;
352 int64_t (*cmp
)(struct perf_hpp_fmt
*fmt
,
353 struct hist_entry
*, struct hist_entry
*);
354 int (*entry
)(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
355 struct hist_entry
*he
);
356 int (*color
)(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
357 struct hist_entry
*he
);
361 struct perf_hpp_fmt fmt
;
362 struct c2c_dimension
*dim
;
365 #define SYMBOL_WIDTH 30
367 static struct c2c_dimension dim_symbol
;
368 static struct c2c_dimension dim_srcline
;
370 static int symbol_width(struct hists
*hists
, struct sort_entry
*se
)
372 int width
= hists__col_len(hists
, se
->se_width_idx
);
374 if (!c2c
.symbol_full
)
375 width
= MIN(width
, SYMBOL_WIDTH
);
380 static int c2c_width(struct perf_hpp_fmt
*fmt
,
381 struct perf_hpp
*hpp __maybe_unused
,
384 struct c2c_fmt
*c2c_fmt
;
385 struct c2c_dimension
*dim
;
387 c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
390 if (dim
== &dim_symbol
|| dim
== &dim_srcline
)
391 return symbol_width(hists
, dim
->se
);
393 return dim
->se
? hists__col_len(hists
, dim
->se
->se_width_idx
) :
397 static int c2c_header(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
398 struct hists
*hists
, int line
, int *span
)
400 struct perf_hpp_list
*hpp_list
= hists
->hpp_list
;
401 struct c2c_fmt
*c2c_fmt
;
402 struct c2c_dimension
*dim
;
403 const char *text
= NULL
;
404 int width
= c2c_width(fmt
, hpp
, hists
);
406 c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
410 text
= dim
->header
.line
[line
].text
;
411 /* Use the last line from sort_entry if not defined. */
412 if (!text
&& (line
== hpp_list
->nr_header_lines
- 1))
413 text
= dim
->se
->se_header
;
415 text
= dim
->header
.line
[line
].text
;
421 *span
= dim
->header
.line
[line
].span
;
428 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, text
);
431 #define HEX_STR(__s, __v) \
433 scnprintf(__s, sizeof(__s), "0x%" PRIx64, __v); \
438 dcacheline_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
439 struct hist_entry
*left
, struct hist_entry
*right
)
441 return sort__dcacheline_cmp(left
, right
);
444 static int dcacheline_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
445 struct hist_entry
*he
)
448 int width
= c2c_width(fmt
, hpp
, he
->hists
);
452 addr
= cl_address(he
->mem_info
->daddr
.addr
);
454 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, HEX_STR(buf
, addr
));
457 static int offset_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
458 struct hist_entry
*he
)
461 int width
= c2c_width(fmt
, hpp
, he
->hists
);
465 addr
= cl_offset(he
->mem_info
->daddr
.al_addr
);
467 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, HEX_STR(buf
, addr
));
471 offset_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
472 struct hist_entry
*left
, struct hist_entry
*right
)
474 uint64_t l
= 0, r
= 0;
477 l
= cl_offset(left
->mem_info
->daddr
.addr
);
479 r
= cl_offset(right
->mem_info
->daddr
.addr
);
481 return (int64_t)(r
- l
);
485 iaddr_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
486 struct hist_entry
*he
)
489 int width
= c2c_width(fmt
, hpp
, he
->hists
);
493 addr
= he
->mem_info
->iaddr
.addr
;
495 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, HEX_STR(buf
, addr
));
499 iaddr_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
500 struct hist_entry
*left
, struct hist_entry
*right
)
502 return sort__iaddr_cmp(left
, right
);
506 tot_hitm_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
507 struct hist_entry
*he
)
509 struct c2c_hist_entry
*c2c_he
;
510 int width
= c2c_width(fmt
, hpp
, he
->hists
);
511 unsigned int tot_hitm
;
513 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
514 tot_hitm
= c2c_he
->stats
.lcl_hitm
+ c2c_he
->stats
.rmt_hitm
;
516 return scnprintf(hpp
->buf
, hpp
->size
, "%*u", width
, tot_hitm
);
520 tot_hitm_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
521 struct hist_entry
*left
, struct hist_entry
*right
)
523 struct c2c_hist_entry
*c2c_left
;
524 struct c2c_hist_entry
*c2c_right
;
525 unsigned int tot_hitm_left
;
526 unsigned int tot_hitm_right
;
528 c2c_left
= container_of(left
, struct c2c_hist_entry
, he
);
529 c2c_right
= container_of(right
, struct c2c_hist_entry
, he
);
531 tot_hitm_left
= c2c_left
->stats
.lcl_hitm
+ c2c_left
->stats
.rmt_hitm
;
532 tot_hitm_right
= c2c_right
->stats
.lcl_hitm
+ c2c_right
->stats
.rmt_hitm
;
534 return tot_hitm_left
- tot_hitm_right
;
537 #define STAT_FN_ENTRY(__f) \
539 __f ## _entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, \
540 struct hist_entry *he) \
542 struct c2c_hist_entry *c2c_he; \
543 int width = c2c_width(fmt, hpp, he->hists); \
545 c2c_he = container_of(he, struct c2c_hist_entry, he); \
546 return scnprintf(hpp->buf, hpp->size, "%*u", width, \
547 c2c_he->stats.__f); \
550 #define STAT_FN_CMP(__f) \
552 __f ## _cmp(struct perf_hpp_fmt *fmt __maybe_unused, \
553 struct hist_entry *left, struct hist_entry *right) \
555 struct c2c_hist_entry *c2c_left, *c2c_right; \
557 c2c_left = container_of(left, struct c2c_hist_entry, he); \
558 c2c_right = container_of(right, struct c2c_hist_entry, he); \
559 return c2c_left->stats.__f - c2c_right->stats.__f; \
562 #define STAT_FN(__f) \
577 static uint64_t llc_miss(struct c2c_stats
*stats
)
581 llcmiss
= stats
->lcl_dram
+
590 ld_llcmiss_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
591 struct hist_entry
*he
)
593 struct c2c_hist_entry
*c2c_he
;
594 int width
= c2c_width(fmt
, hpp
, he
->hists
);
596 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
598 return scnprintf(hpp
->buf
, hpp
->size
, "%*lu", width
,
599 llc_miss(&c2c_he
->stats
));
603 ld_llcmiss_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
604 struct hist_entry
*left
, struct hist_entry
*right
)
606 struct c2c_hist_entry
*c2c_left
;
607 struct c2c_hist_entry
*c2c_right
;
609 c2c_left
= container_of(left
, struct c2c_hist_entry
, he
);
610 c2c_right
= container_of(right
, struct c2c_hist_entry
, he
);
612 return llc_miss(&c2c_left
->stats
) - llc_miss(&c2c_right
->stats
);
615 static uint64_t total_records(struct c2c_stats
*stats
)
617 uint64_t lclmiss
, ldcnt
, total
;
619 lclmiss
= stats
->lcl_dram
+
639 tot_recs_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
640 struct hist_entry
*he
)
642 struct c2c_hist_entry
*c2c_he
;
643 int width
= c2c_width(fmt
, hpp
, he
->hists
);
646 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
647 tot_recs
= total_records(&c2c_he
->stats
);
649 return scnprintf(hpp
->buf
, hpp
->size
, "%*" PRIu64
, width
, tot_recs
);
653 tot_recs_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
654 struct hist_entry
*left
, struct hist_entry
*right
)
656 struct c2c_hist_entry
*c2c_left
;
657 struct c2c_hist_entry
*c2c_right
;
658 uint64_t tot_recs_left
;
659 uint64_t tot_recs_right
;
661 c2c_left
= container_of(left
, struct c2c_hist_entry
, he
);
662 c2c_right
= container_of(right
, struct c2c_hist_entry
, he
);
664 tot_recs_left
= total_records(&c2c_left
->stats
);
665 tot_recs_right
= total_records(&c2c_right
->stats
);
667 return tot_recs_left
- tot_recs_right
;
670 static uint64_t total_loads(struct c2c_stats
*stats
)
672 uint64_t lclmiss
, ldcnt
;
674 lclmiss
= stats
->lcl_dram
+
690 tot_loads_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
691 struct hist_entry
*he
)
693 struct c2c_hist_entry
*c2c_he
;
694 int width
= c2c_width(fmt
, hpp
, he
->hists
);
697 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
698 tot_recs
= total_loads(&c2c_he
->stats
);
700 return scnprintf(hpp
->buf
, hpp
->size
, "%*" PRIu64
, width
, tot_recs
);
704 tot_loads_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
705 struct hist_entry
*left
, struct hist_entry
*right
)
707 struct c2c_hist_entry
*c2c_left
;
708 struct c2c_hist_entry
*c2c_right
;
709 uint64_t tot_recs_left
;
710 uint64_t tot_recs_right
;
712 c2c_left
= container_of(left
, struct c2c_hist_entry
, he
);
713 c2c_right
= container_of(right
, struct c2c_hist_entry
, he
);
715 tot_recs_left
= total_loads(&c2c_left
->stats
);
716 tot_recs_right
= total_loads(&c2c_right
->stats
);
718 return tot_recs_left
- tot_recs_right
;
721 typedef double (get_percent_cb
)(struct c2c_hist_entry
*);
724 percent_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
725 struct hist_entry
*he
, get_percent_cb get_percent
)
727 struct c2c_hist_entry
*c2c_he
;
728 int width
= c2c_width(fmt
, hpp
, he
->hists
);
731 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
732 per
= get_percent(c2c_he
);
734 #ifdef HAVE_SLANG_SUPPORT
736 return __hpp__slsmg_color_printf(hpp
, "%*.2f%%", width
- 1, per
);
738 return hpp_color_scnprintf(hpp
, "%*.2f%%", width
- 1, per
);
741 static double percent_hitm(struct c2c_hist_entry
*c2c_he
)
743 struct c2c_hists
*hists
;
744 struct c2c_stats
*stats
;
745 struct c2c_stats
*total
;
749 hists
= container_of(c2c_he
->he
.hists
, struct c2c_hists
, hists
);
750 stats
= &c2c_he
->stats
;
751 total
= &hists
->stats
;
753 switch (c2c
.display
) {
755 st
= stats
->rmt_hitm
;
756 tot
= total
->rmt_hitm
;
759 st
= stats
->lcl_hitm
;
760 tot
= total
->lcl_hitm
;
763 st
= stats
->tot_hitm
;
764 tot
= total
->tot_hitm
;
769 p
= tot
? (double) st
/ tot
: 0;
774 #define PERC_STR(__s, __v) \
776 scnprintf(__s, sizeof(__s), "%.2F%%", __v); \
781 percent_hitm_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
782 struct hist_entry
*he
)
784 struct c2c_hist_entry
*c2c_he
;
785 int width
= c2c_width(fmt
, hpp
, he
->hists
);
789 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
790 per
= percent_hitm(c2c_he
);
791 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, PERC_STR(buf
, per
));
795 percent_hitm_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
796 struct hist_entry
*he
)
798 return percent_color(fmt
, hpp
, he
, percent_hitm
);
802 percent_hitm_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
803 struct hist_entry
*left
, struct hist_entry
*right
)
805 struct c2c_hist_entry
*c2c_left
;
806 struct c2c_hist_entry
*c2c_right
;
810 c2c_left
= container_of(left
, struct c2c_hist_entry
, he
);
811 c2c_right
= container_of(right
, struct c2c_hist_entry
, he
);
813 per_left
= percent_hitm(c2c_left
);
814 per_right
= percent_hitm(c2c_right
);
816 return per_left
- per_right
;
819 static struct c2c_stats
*he_stats(struct hist_entry
*he
)
821 struct c2c_hist_entry
*c2c_he
;
823 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
824 return &c2c_he
->stats
;
827 static struct c2c_stats
*total_stats(struct hist_entry
*he
)
829 struct c2c_hists
*hists
;
831 hists
= container_of(he
->hists
, struct c2c_hists
, hists
);
832 return &hists
->stats
;
835 static double percent(int st
, int tot
)
837 return tot
? 100. * (double) st
/ (double) tot
: 0;
840 #define PERCENT(__h, __f) percent(he_stats(__h)->__f, total_stats(__h)->__f)
842 #define PERCENT_FN(__f) \
843 static double percent_ ## __f(struct c2c_hist_entry *c2c_he) \
845 struct c2c_hists *hists; \
847 hists = container_of(c2c_he->he.hists, struct c2c_hists, hists); \
848 return percent(c2c_he->stats.__f, hists->stats.__f); \
854 PERCENT_FN(st_l1miss
)
857 percent_rmt_hitm_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
858 struct hist_entry
*he
)
860 int width
= c2c_width(fmt
, hpp
, he
->hists
);
861 double per
= PERCENT(he
, rmt_hitm
);
864 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, PERC_STR(buf
, per
));
868 percent_rmt_hitm_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
869 struct hist_entry
*he
)
871 return percent_color(fmt
, hpp
, he
, percent_rmt_hitm
);
875 percent_rmt_hitm_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
876 struct hist_entry
*left
, struct hist_entry
*right
)
881 per_left
= PERCENT(left
, lcl_hitm
);
882 per_right
= PERCENT(right
, lcl_hitm
);
884 return per_left
- per_right
;
888 percent_lcl_hitm_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
889 struct hist_entry
*he
)
891 int width
= c2c_width(fmt
, hpp
, he
->hists
);
892 double per
= PERCENT(he
, lcl_hitm
);
895 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, PERC_STR(buf
, per
));
899 percent_lcl_hitm_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
900 struct hist_entry
*he
)
902 return percent_color(fmt
, hpp
, he
, percent_lcl_hitm
);
906 percent_lcl_hitm_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
907 struct hist_entry
*left
, struct hist_entry
*right
)
912 per_left
= PERCENT(left
, lcl_hitm
);
913 per_right
= PERCENT(right
, lcl_hitm
);
915 return per_left
- per_right
;
919 percent_stores_l1hit_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
920 struct hist_entry
*he
)
922 int width
= c2c_width(fmt
, hpp
, he
->hists
);
923 double per
= PERCENT(he
, st_l1hit
);
926 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, PERC_STR(buf
, per
));
930 percent_stores_l1hit_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
931 struct hist_entry
*he
)
933 return percent_color(fmt
, hpp
, he
, percent_st_l1hit
);
937 percent_stores_l1hit_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
938 struct hist_entry
*left
, struct hist_entry
*right
)
943 per_left
= PERCENT(left
, st_l1hit
);
944 per_right
= PERCENT(right
, st_l1hit
);
946 return per_left
- per_right
;
950 percent_stores_l1miss_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
951 struct hist_entry
*he
)
953 int width
= c2c_width(fmt
, hpp
, he
->hists
);
954 double per
= PERCENT(he
, st_l1miss
);
957 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, PERC_STR(buf
, per
));
961 percent_stores_l1miss_color(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
962 struct hist_entry
*he
)
964 return percent_color(fmt
, hpp
, he
, percent_st_l1miss
);
968 percent_stores_l1miss_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
969 struct hist_entry
*left
, struct hist_entry
*right
)
974 per_left
= PERCENT(left
, st_l1miss
);
975 per_right
= PERCENT(right
, st_l1miss
);
977 return per_left
- per_right
;
984 pid_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
985 struct hist_entry
*he
)
987 int width
= c2c_width(fmt
, hpp
, he
->hists
);
989 return scnprintf(hpp
->buf
, hpp
->size
, "%*d", width
, he
->thread
->pid_
);
993 pid_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
994 struct hist_entry
*left
, struct hist_entry
*right
)
996 return left
->thread
->pid_
- right
->thread
->pid_
;
1000 empty_cmp(struct perf_hpp_fmt
*fmt __maybe_unused
,
1001 struct hist_entry
*left __maybe_unused
,
1002 struct hist_entry
*right __maybe_unused
)
1008 node_entry(struct perf_hpp_fmt
*fmt __maybe_unused
, struct perf_hpp
*hpp
,
1009 struct hist_entry
*he
)
1011 struct c2c_hist_entry
*c2c_he
;
1016 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1018 for (node
= 0; node
< c2c
.nodes_cnt
; node
++) {
1019 DECLARE_BITMAP(set
, c2c
.cpus_cnt
);
1021 bitmap_zero(set
, c2c
.cpus_cnt
);
1022 bitmap_and(set
, c2c_he
->cpuset
, c2c
.nodes
[node
], c2c
.cpus_cnt
);
1024 if (!bitmap_weight(set
, c2c
.cpus_cnt
)) {
1025 if (c2c
.node_info
== 1) {
1026 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%21s", " ");
1027 advance_hpp(hpp
, ret
);
1033 ret
= scnprintf(hpp
->buf
, hpp
->size
, " ");
1034 advance_hpp(hpp
, ret
);
1037 switch (c2c
.node_info
) {
1039 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%2d", node
);
1040 advance_hpp(hpp
, ret
);
1044 int num
= bitmap_weight(c2c_he
->cpuset
, c2c
.cpus_cnt
);
1045 struct c2c_stats
*stats
= &c2c_he
->node_stats
[node
];
1047 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%2d{%2d ", node
, num
);
1048 advance_hpp(hpp
, ret
);
1050 #define DISPLAY_HITM(__h) \
1051 if (c2c_he->stats.__h> 0) { \
1052 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%% ", \
1053 percent(stats->__h, c2c_he->stats.__h));\
1055 ret = scnprintf(hpp->buf, hpp->size, "%6s ", "n/a"); \
1058 switch (c2c
.display
) {
1060 DISPLAY_HITM(rmt_hitm
);
1063 DISPLAY_HITM(lcl_hitm
);
1066 DISPLAY_HITM(tot_hitm
);
1073 advance_hpp(hpp
, ret
);
1075 if (c2c_he
->stats
.store
> 0) {
1076 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%5.1f%%}",
1077 percent(stats
->store
, c2c_he
->stats
.store
));
1079 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%6s}", "n/a");
1082 advance_hpp(hpp
, ret
);
1086 ret
= scnprintf(hpp
->buf
, hpp
->size
, "%2d{", node
);
1087 advance_hpp(hpp
, ret
);
1089 ret
= bitmap_scnprintf(set
, c2c
.cpus_cnt
, hpp
->buf
, hpp
->size
);
1090 advance_hpp(hpp
, ret
);
1092 ret
= scnprintf(hpp
->buf
, hpp
->size
, "}");
1093 advance_hpp(hpp
, ret
);
1106 mean_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
1107 struct hist_entry
*he
, double mean
)
1109 int width
= c2c_width(fmt
, hpp
, he
->hists
);
1112 scnprintf(buf
, 10, "%6.0f", mean
);
1113 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, buf
);
1116 #define MEAN_ENTRY(__func, __val) \
1118 __func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he) \
1120 struct c2c_hist_entry *c2c_he; \
1121 c2c_he = container_of(he, struct c2c_hist_entry, he); \
1122 return mean_entry(fmt, hpp, he, avg_stats(&c2c_he->cstats.__val)); \
1125 MEAN_ENTRY(mean_rmt_entry
, rmt_hitm
);
1126 MEAN_ENTRY(mean_lcl_entry
, lcl_hitm
);
1127 MEAN_ENTRY(mean_load_entry
, load
);
1130 cpucnt_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
1131 struct hist_entry
*he
)
1133 struct c2c_hist_entry
*c2c_he
;
1134 int width
= c2c_width(fmt
, hpp
, he
->hists
);
1137 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1139 scnprintf(buf
, 10, "%d", bitmap_weight(c2c_he
->cpuset
, c2c
.cpus_cnt
));
1140 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, buf
);
1144 cl_idx_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
1145 struct hist_entry
*he
)
1147 struct c2c_hist_entry
*c2c_he
;
1148 int width
= c2c_width(fmt
, hpp
, he
->hists
);
1151 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1153 scnprintf(buf
, 10, "%u", c2c_he
->cacheline_idx
);
1154 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, buf
);
1158 cl_idx_empty_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
1159 struct hist_entry
*he
)
1161 int width
= c2c_width(fmt
, hpp
, he
->hists
);
1163 return scnprintf(hpp
->buf
, hpp
->size
, "%*s", width
, "");
1166 #define HEADER_LOW(__h) \
1173 #define HEADER_BOTH(__h0, __h1) \
1183 #define HEADER_SPAN(__h0, __h1, __s) \
1194 #define HEADER_SPAN_LOW(__h) \
1201 static struct c2c_dimension dim_dcacheline
= {
1202 .header
= HEADER_LOW("Cacheline"),
1203 .name
= "dcacheline",
1204 .cmp
= dcacheline_cmp
,
1205 .entry
= dcacheline_entry
,
1209 static struct c2c_header header_offset_tui
= HEADER_LOW("Off");
1211 static struct c2c_dimension dim_offset
= {
1212 .header
= HEADER_BOTH("Data address", "Offset"),
1215 .entry
= offset_entry
,
1219 static struct c2c_dimension dim_iaddr
= {
1220 .header
= HEADER_LOW("Code address"),
1223 .entry
= iaddr_entry
,
1227 static struct c2c_dimension dim_tot_hitm
= {
1228 .header
= HEADER_SPAN("----- LLC Load Hitm -----", "Total", 2),
1230 .cmp
= tot_hitm_cmp
,
1231 .entry
= tot_hitm_entry
,
1235 static struct c2c_dimension dim_lcl_hitm
= {
1236 .header
= HEADER_SPAN_LOW("Lcl"),
1238 .cmp
= lcl_hitm_cmp
,
1239 .entry
= lcl_hitm_entry
,
1243 static struct c2c_dimension dim_rmt_hitm
= {
1244 .header
= HEADER_SPAN_LOW("Rmt"),
1246 .cmp
= rmt_hitm_cmp
,
1247 .entry
= rmt_hitm_entry
,
1251 static struct c2c_dimension dim_cl_rmt_hitm
= {
1252 .header
= HEADER_SPAN("----- HITM -----", "Rmt", 1),
1253 .name
= "cl_rmt_hitm",
1254 .cmp
= rmt_hitm_cmp
,
1255 .entry
= rmt_hitm_entry
,
1259 static struct c2c_dimension dim_cl_lcl_hitm
= {
1260 .header
= HEADER_SPAN_LOW("Lcl"),
1261 .name
= "cl_lcl_hitm",
1262 .cmp
= lcl_hitm_cmp
,
1263 .entry
= lcl_hitm_entry
,
1267 static struct c2c_dimension dim_stores
= {
1268 .header
= HEADER_SPAN("---- Store Reference ----", "Total", 2),
1271 .entry
= store_entry
,
1275 static struct c2c_dimension dim_stores_l1hit
= {
1276 .header
= HEADER_SPAN_LOW("L1Hit"),
1277 .name
= "stores_l1hit",
1278 .cmp
= st_l1hit_cmp
,
1279 .entry
= st_l1hit_entry
,
1283 static struct c2c_dimension dim_stores_l1miss
= {
1284 .header
= HEADER_SPAN_LOW("L1Miss"),
1285 .name
= "stores_l1miss",
1286 .cmp
= st_l1miss_cmp
,
1287 .entry
= st_l1miss_entry
,
1291 static struct c2c_dimension dim_cl_stores_l1hit
= {
1292 .header
= HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1293 .name
= "cl_stores_l1hit",
1294 .cmp
= st_l1hit_cmp
,
1295 .entry
= st_l1hit_entry
,
1299 static struct c2c_dimension dim_cl_stores_l1miss
= {
1300 .header
= HEADER_SPAN_LOW("L1 Miss"),
1301 .name
= "cl_stores_l1miss",
1302 .cmp
= st_l1miss_cmp
,
1303 .entry
= st_l1miss_entry
,
1307 static struct c2c_dimension dim_ld_fbhit
= {
1308 .header
= HEADER_SPAN("----- Core Load Hit -----", "FB", 2),
1310 .cmp
= ld_fbhit_cmp
,
1311 .entry
= ld_fbhit_entry
,
1315 static struct c2c_dimension dim_ld_l1hit
= {
1316 .header
= HEADER_SPAN_LOW("L1"),
1318 .cmp
= ld_l1hit_cmp
,
1319 .entry
= ld_l1hit_entry
,
1323 static struct c2c_dimension dim_ld_l2hit
= {
1324 .header
= HEADER_SPAN_LOW("L2"),
1326 .cmp
= ld_l2hit_cmp
,
1327 .entry
= ld_l2hit_entry
,
1331 static struct c2c_dimension dim_ld_llchit
= {
1332 .header
= HEADER_SPAN("-- LLC Load Hit --", "Llc", 1),
1333 .name
= "ld_lclhit",
1334 .cmp
= ld_llchit_cmp
,
1335 .entry
= ld_llchit_entry
,
1339 static struct c2c_dimension dim_ld_rmthit
= {
1340 .header
= HEADER_SPAN_LOW("Rmt"),
1341 .name
= "ld_rmthit",
1343 .entry
= rmt_hit_entry
,
1347 static struct c2c_dimension dim_ld_llcmiss
= {
1348 .header
= HEADER_BOTH("LLC", "Ld Miss"),
1349 .name
= "ld_llcmiss",
1350 .cmp
= ld_llcmiss_cmp
,
1351 .entry
= ld_llcmiss_entry
,
1355 static struct c2c_dimension dim_tot_recs
= {
1356 .header
= HEADER_BOTH("Total", "records"),
1358 .cmp
= tot_recs_cmp
,
1359 .entry
= tot_recs_entry
,
1363 static struct c2c_dimension dim_tot_loads
= {
1364 .header
= HEADER_BOTH("Total", "Loads"),
1365 .name
= "tot_loads",
1366 .cmp
= tot_loads_cmp
,
1367 .entry
= tot_loads_entry
,
1371 static struct c2c_header percent_hitm_header
[] = {
1372 [DISPLAY_LCL
] = HEADER_BOTH("Lcl", "Hitm"),
1373 [DISPLAY_RMT
] = HEADER_BOTH("Rmt", "Hitm"),
1374 [DISPLAY_TOT
] = HEADER_BOTH("Tot", "Hitm"),
1377 static struct c2c_dimension dim_percent_hitm
= {
1378 .name
= "percent_hitm",
1379 .cmp
= percent_hitm_cmp
,
1380 .entry
= percent_hitm_entry
,
1381 .color
= percent_hitm_color
,
1385 static struct c2c_dimension dim_percent_rmt_hitm
= {
1386 .header
= HEADER_SPAN("----- HITM -----", "Rmt", 1),
1387 .name
= "percent_rmt_hitm",
1388 .cmp
= percent_rmt_hitm_cmp
,
1389 .entry
= percent_rmt_hitm_entry
,
1390 .color
= percent_rmt_hitm_color
,
1394 static struct c2c_dimension dim_percent_lcl_hitm
= {
1395 .header
= HEADER_SPAN_LOW("Lcl"),
1396 .name
= "percent_lcl_hitm",
1397 .cmp
= percent_lcl_hitm_cmp
,
1398 .entry
= percent_lcl_hitm_entry
,
1399 .color
= percent_lcl_hitm_color
,
1403 static struct c2c_dimension dim_percent_stores_l1hit
= {
1404 .header
= HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1405 .name
= "percent_stores_l1hit",
1406 .cmp
= percent_stores_l1hit_cmp
,
1407 .entry
= percent_stores_l1hit_entry
,
1408 .color
= percent_stores_l1hit_color
,
1412 static struct c2c_dimension dim_percent_stores_l1miss
= {
1413 .header
= HEADER_SPAN_LOW("L1 Miss"),
1414 .name
= "percent_stores_l1miss",
1415 .cmp
= percent_stores_l1miss_cmp
,
1416 .entry
= percent_stores_l1miss_entry
,
1417 .color
= percent_stores_l1miss_color
,
1421 static struct c2c_dimension dim_dram_lcl
= {
1422 .header
= HEADER_SPAN("--- Load Dram ----", "Lcl", 1),
1424 .cmp
= lcl_dram_cmp
,
1425 .entry
= lcl_dram_entry
,
1429 static struct c2c_dimension dim_dram_rmt
= {
1430 .header
= HEADER_SPAN_LOW("Rmt"),
1432 .cmp
= rmt_dram_cmp
,
1433 .entry
= rmt_dram_entry
,
1437 static struct c2c_dimension dim_pid
= {
1438 .header
= HEADER_LOW("Pid"),
1445 static struct c2c_dimension dim_tid
= {
1446 .header
= HEADER_LOW("Tid"),
1451 static struct c2c_dimension dim_symbol
= {
1456 static struct c2c_dimension dim_dso
= {
1457 .header
= HEADER_BOTH("Shared", "Object"),
1462 static struct c2c_header header_node
[3] = {
1464 HEADER_LOW("Node{cpus %hitms %stores}"),
1465 HEADER_LOW("Node{cpu list}"),
1468 static struct c2c_dimension dim_node
= {
1471 .entry
= node_entry
,
1475 static struct c2c_dimension dim_mean_rmt
= {
1476 .header
= HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2),
1479 .entry
= mean_rmt_entry
,
1483 static struct c2c_dimension dim_mean_lcl
= {
1484 .header
= HEADER_SPAN_LOW("lcl hitm"),
1487 .entry
= mean_lcl_entry
,
1491 static struct c2c_dimension dim_mean_load
= {
1492 .header
= HEADER_SPAN_LOW("load"),
1493 .name
= "mean_load",
1495 .entry
= mean_load_entry
,
1499 static struct c2c_dimension dim_cpucnt
= {
1500 .header
= HEADER_BOTH("cpu", "cnt"),
1503 .entry
= cpucnt_entry
,
1507 static struct c2c_dimension dim_srcline
= {
1508 .name
= "cl_srcline",
1509 .se
= &sort_srcline
,
1512 static struct c2c_dimension dim_dcacheline_idx
= {
1513 .header
= HEADER_LOW("Index"),
1516 .entry
= cl_idx_entry
,
1520 static struct c2c_dimension dim_dcacheline_num
= {
1521 .header
= HEADER_LOW("Num"),
1524 .entry
= cl_idx_entry
,
1528 static struct c2c_dimension dim_dcacheline_num_empty
= {
1529 .header
= HEADER_LOW("Num"),
1530 .name
= "cl_num_empty",
1532 .entry
= cl_idx_empty_entry
,
1536 static struct c2c_dimension
*dimensions
[] = {
1548 &dim_cl_stores_l1hit
,
1549 &dim_cl_stores_l1miss
,
1559 &dim_percent_rmt_hitm
,
1560 &dim_percent_lcl_hitm
,
1561 &dim_percent_stores_l1hit
,
1562 &dim_percent_stores_l1miss
,
1575 &dim_dcacheline_idx
,
1576 &dim_dcacheline_num
,
1577 &dim_dcacheline_num_empty
,
1581 static void fmt_free(struct perf_hpp_fmt
*fmt
)
1583 struct c2c_fmt
*c2c_fmt
;
1585 c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
1589 static bool fmt_equal(struct perf_hpp_fmt
*a
, struct perf_hpp_fmt
*b
)
1591 struct c2c_fmt
*c2c_a
= container_of(a
, struct c2c_fmt
, fmt
);
1592 struct c2c_fmt
*c2c_b
= container_of(b
, struct c2c_fmt
, fmt
);
1594 return c2c_a
->dim
== c2c_b
->dim
;
1597 static struct c2c_dimension
*get_dimension(const char *name
)
1601 for (i
= 0; dimensions
[i
]; i
++) {
1602 struct c2c_dimension
*dim
= dimensions
[i
];
1604 if (!strcmp(dim
->name
, name
))
1611 static int c2c_se_entry(struct perf_hpp_fmt
*fmt
, struct perf_hpp
*hpp
,
1612 struct hist_entry
*he
)
1614 struct c2c_fmt
*c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
1615 struct c2c_dimension
*dim
= c2c_fmt
->dim
;
1616 size_t len
= fmt
->user_len
;
1619 len
= hists__col_len(he
->hists
, dim
->se
->se_width_idx
);
1621 if (dim
== &dim_symbol
|| dim
== &dim_srcline
)
1622 len
= symbol_width(he
->hists
, dim
->se
);
1625 return dim
->se
->se_snprintf(he
, hpp
->buf
, hpp
->size
, len
);
1628 static int64_t c2c_se_cmp(struct perf_hpp_fmt
*fmt
,
1629 struct hist_entry
*a
, struct hist_entry
*b
)
1631 struct c2c_fmt
*c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
1632 struct c2c_dimension
*dim
= c2c_fmt
->dim
;
1634 return dim
->se
->se_cmp(a
, b
);
1637 static int64_t c2c_se_collapse(struct perf_hpp_fmt
*fmt
,
1638 struct hist_entry
*a
, struct hist_entry
*b
)
1640 struct c2c_fmt
*c2c_fmt
= container_of(fmt
, struct c2c_fmt
, fmt
);
1641 struct c2c_dimension
*dim
= c2c_fmt
->dim
;
1642 int64_t (*collapse_fn
)(struct hist_entry
*, struct hist_entry
*);
1644 collapse_fn
= dim
->se
->se_collapse
?: dim
->se
->se_cmp
;
1645 return collapse_fn(a
, b
);
1648 static struct c2c_fmt
*get_format(const char *name
)
1650 struct c2c_dimension
*dim
= get_dimension(name
);
1651 struct c2c_fmt
*c2c_fmt
;
1652 struct perf_hpp_fmt
*fmt
;
1657 c2c_fmt
= zalloc(sizeof(*c2c_fmt
));
1663 fmt
= &c2c_fmt
->fmt
;
1664 INIT_LIST_HEAD(&fmt
->list
);
1665 INIT_LIST_HEAD(&fmt
->sort_list
);
1667 fmt
->cmp
= dim
->se
? c2c_se_cmp
: dim
->cmp
;
1668 fmt
->sort
= dim
->se
? c2c_se_cmp
: dim
->cmp
;
1669 fmt
->color
= dim
->se
? NULL
: dim
->color
;
1670 fmt
->entry
= dim
->se
? c2c_se_entry
: dim
->entry
;
1671 fmt
->header
= c2c_header
;
1672 fmt
->width
= c2c_width
;
1673 fmt
->collapse
= dim
->se
? c2c_se_collapse
: dim
->cmp
;
1674 fmt
->equal
= fmt_equal
;
1675 fmt
->free
= fmt_free
;
1680 static int c2c_hists__init_output(struct perf_hpp_list
*hpp_list
, char *name
)
1682 struct c2c_fmt
*c2c_fmt
= get_format(name
);
1686 return output_field_add(hpp_list
, name
);
1689 perf_hpp_list__column_register(hpp_list
, &c2c_fmt
->fmt
);
1693 static int c2c_hists__init_sort(struct perf_hpp_list
*hpp_list
, char *name
)
1695 struct c2c_fmt
*c2c_fmt
= get_format(name
);
1696 struct c2c_dimension
*dim
;
1700 return sort_dimension__add(hpp_list
, name
, NULL
, 0);
1704 if (dim
== &dim_dso
)
1707 perf_hpp_list__register_sort_field(hpp_list
, &c2c_fmt
->fmt
);
1711 #define PARSE_LIST(_list, _fn) \
1719 for (tok = strtok_r((char *)_list, ", ", &tmp); \
1720 tok; tok = strtok_r(NULL, ", ", &tmp)) { \
1721 ret = _fn(hpp_list, tok); \
1722 if (ret == -EINVAL) { \
1723 error("Invalid --fields key: `%s'", tok); \
1725 } else if (ret == -ESRCH) { \
1726 error("Unknown --fields key: `%s'", tok); \
1732 static int hpp_list__parse(struct perf_hpp_list
*hpp_list
,
1733 const char *output_
,
1736 char *output
= output_
? strdup(output_
) : NULL
;
1737 char *sort
= sort_
? strdup(sort_
) : NULL
;
1740 PARSE_LIST(output
, c2c_hists__init_output
);
1741 PARSE_LIST(sort
, c2c_hists__init_sort
);
1743 /* copy sort keys to output fields */
1744 perf_hpp__setup_output_field(hpp_list
);
1747 * We dont need other sorting keys other than those
1748 * we already specified. It also really slows down
1749 * the processing a lot with big number of output
1750 * fields, so switching this off for c2c.
1754 /* and then copy output fields to sort keys */
1755 perf_hpp__append_sort_keys(&hists
->list
);
1763 static int c2c_hists__init(struct c2c_hists
*hists
,
1765 int nr_header_lines
)
1767 __hists__init(&hists
->hists
, &hists
->list
);
1770 * Initialize only with sort fields, we need to resort
1771 * later anyway, and that's where we add output fields
1774 perf_hpp_list__init(&hists
->list
);
1776 /* Overload number of header lines.*/
1777 hists
->list
.nr_header_lines
= nr_header_lines
;
1779 return hpp_list__parse(&hists
->list
, NULL
, sort
);
1782 static int c2c_hists__reinit(struct c2c_hists
*c2c_hists
,
1786 perf_hpp__reset_output_field(&c2c_hists
->list
);
1787 return hpp_list__parse(&c2c_hists
->list
, output
, sort
);
1790 #define DISPLAY_LINE_LIMIT 0.0005
1792 static bool he__display(struct hist_entry
*he
, struct c2c_stats
*stats
)
1794 struct c2c_hist_entry
*c2c_he
;
1800 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1802 #define FILTER_HITM(__h) \
1804 ld_dist = ((double)c2c_he->stats.__h / stats->__h); \
1805 if (ld_dist < DISPLAY_LINE_LIMIT) \
1806 he->filtered = HIST_FILTER__C2C; \
1808 he->filtered = HIST_FILTER__C2C; \
1811 switch (c2c
.display
) {
1813 FILTER_HITM(lcl_hitm
);
1816 FILTER_HITM(rmt_hitm
);
1819 FILTER_HITM(tot_hitm
);
1826 return he
->filtered
== 0;
1829 static inline int valid_hitm_or_store(struct hist_entry
*he
)
1831 struct c2c_hist_entry
*c2c_he
;
1834 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1835 has_hitm
= c2c
.display
== DISPLAY_TOT
? c2c_he
->stats
.tot_hitm
:
1836 c2c
.display
== DISPLAY_LCL
? c2c_he
->stats
.lcl_hitm
:
1837 c2c_he
->stats
.rmt_hitm
;
1838 return has_hitm
|| c2c_he
->stats
.store
;
1841 static void calc_width(struct hist_entry
*he
)
1843 struct c2c_hists
*c2c_hists
;
1845 c2c_hists
= container_of(he
->hists
, struct c2c_hists
, hists
);
1846 hists__calc_col_len(&c2c_hists
->hists
, he
);
1849 static int filter_cb(struct hist_entry
*he
)
1851 if (c2c
.show_src
&& !he
->srcline
)
1852 he
->srcline
= hist_entry__get_srcline(he
);
1856 if (!valid_hitm_or_store(he
))
1857 he
->filtered
= HIST_FILTER__C2C
;
1862 static int resort_cl_cb(struct hist_entry
*he
)
1864 struct c2c_hist_entry
*c2c_he
;
1865 struct c2c_hists
*c2c_hists
;
1866 bool display
= he__display(he
, &c2c
.hitm_stats
);
1868 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1869 c2c_hists
= c2c_he
->hists
;
1873 if (display
&& c2c_hists
) {
1874 static unsigned int idx
;
1876 c2c_he
->cacheline_idx
= idx
++;
1878 c2c_hists__reinit(c2c_hists
, c2c
.cl_output
, c2c
.cl_resort
);
1880 hists__collapse_resort(&c2c_hists
->hists
, NULL
);
1881 hists__output_resort_cb(&c2c_hists
->hists
, NULL
, filter_cb
);
1887 static void setup_nodes_header(void)
1889 dim_node
.header
= header_node
[c2c
.node_info
];
1892 static int setup_nodes(struct perf_session
*session
)
1894 struct numa_node
*n
;
1895 unsigned long **nodes
;
1899 if (c2c
.node_info
> 2)
1902 c2c
.nodes_cnt
= session
->header
.env
.nr_numa_nodes
;
1903 c2c
.cpus_cnt
= session
->header
.env
.nr_cpus_online
;
1905 n
= session
->header
.env
.numa_nodes
;
1909 nodes
= zalloc(sizeof(unsigned long *) * c2c
.nodes_cnt
);
1915 cpu2node
= zalloc(sizeof(int) * c2c
.cpus_cnt
);
1919 for (cpu
= 0; cpu
< c2c
.cpus_cnt
; cpu
++)
1922 c2c
.cpu2node
= cpu2node
;
1924 for (node
= 0; node
< c2c
.nodes_cnt
; node
++) {
1925 struct cpu_map
*map
= n
[node
].map
;
1928 set
= bitmap_alloc(c2c
.cpus_cnt
);
1932 for (cpu
= 0; cpu
< map
->nr
; cpu
++) {
1933 set_bit(map
->map
[cpu
], set
);
1935 if (WARN_ONCE(cpu2node
[map
->map
[cpu
]] != -1, "node/cpu topology bug"))
1938 cpu2node
[map
->map
[cpu
]] = node
;
1944 setup_nodes_header();
1948 #define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm)
1950 static int resort_hitm_cb(struct hist_entry
*he
)
1952 struct c2c_hist_entry
*c2c_he
;
1953 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
1955 if (HAS_HITMS(c2c_he
)) {
1956 c2c
.shared_clines
++;
1957 c2c_add_stats(&c2c
.hitm_stats
, &c2c_he
->stats
);
1963 static int hists__iterate_cb(struct hists
*hists
, hists__resort_cb_t cb
)
1965 struct rb_node
*next
= rb_first(&hists
->entries
);
1969 struct hist_entry
*he
;
1971 he
= rb_entry(next
, struct hist_entry
, rb_node
);
1975 next
= rb_next(&he
->rb_node
);
1981 static void print_c2c__display_stats(FILE *out
)
1984 struct c2c_stats
*stats
= &c2c
.hists
.stats
;
1986 llc_misses
= stats
->lcl_dram
+
1991 fprintf(out
, "=================================================\n");
1992 fprintf(out
, " Trace Event Information \n");
1993 fprintf(out
, "=================================================\n");
1994 fprintf(out
, " Total records : %10d\n", stats
->nr_entries
);
1995 fprintf(out
, " Locked Load/Store Operations : %10d\n", stats
->locks
);
1996 fprintf(out
, " Load Operations : %10d\n", stats
->load
);
1997 fprintf(out
, " Loads - uncacheable : %10d\n", stats
->ld_uncache
);
1998 fprintf(out
, " Loads - IO : %10d\n", stats
->ld_io
);
1999 fprintf(out
, " Loads - Miss : %10d\n", stats
->ld_miss
);
2000 fprintf(out
, " Loads - no mapping : %10d\n", stats
->ld_noadrs
);
2001 fprintf(out
, " Load Fill Buffer Hit : %10d\n", stats
->ld_fbhit
);
2002 fprintf(out
, " Load L1D hit : %10d\n", stats
->ld_l1hit
);
2003 fprintf(out
, " Load L2D hit : %10d\n", stats
->ld_l2hit
);
2004 fprintf(out
, " Load LLC hit : %10d\n", stats
->ld_llchit
+ stats
->lcl_hitm
);
2005 fprintf(out
, " Load Local HITM : %10d\n", stats
->lcl_hitm
);
2006 fprintf(out
, " Load Remote HITM : %10d\n", stats
->rmt_hitm
);
2007 fprintf(out
, " Load Remote HIT : %10d\n", stats
->rmt_hit
);
2008 fprintf(out
, " Load Local DRAM : %10d\n", stats
->lcl_dram
);
2009 fprintf(out
, " Load Remote DRAM : %10d\n", stats
->rmt_dram
);
2010 fprintf(out
, " Load MESI State Exclusive : %10d\n", stats
->ld_excl
);
2011 fprintf(out
, " Load MESI State Shared : %10d\n", stats
->ld_shared
);
2012 fprintf(out
, " Load LLC Misses : %10d\n", llc_misses
);
2013 fprintf(out
, " LLC Misses to Local DRAM : %10.1f%%\n", ((double)stats
->lcl_dram
/(double)llc_misses
) * 100.);
2014 fprintf(out
, " LLC Misses to Remote DRAM : %10.1f%%\n", ((double)stats
->rmt_dram
/(double)llc_misses
) * 100.);
2015 fprintf(out
, " LLC Misses to Remote cache (HIT) : %10.1f%%\n", ((double)stats
->rmt_hit
/(double)llc_misses
) * 100.);
2016 fprintf(out
, " LLC Misses to Remote cache (HITM) : %10.1f%%\n", ((double)stats
->rmt_hitm
/(double)llc_misses
) * 100.);
2017 fprintf(out
, " Store Operations : %10d\n", stats
->store
);
2018 fprintf(out
, " Store - uncacheable : %10d\n", stats
->st_uncache
);
2019 fprintf(out
, " Store - no mapping : %10d\n", stats
->st_noadrs
);
2020 fprintf(out
, " Store L1D Hit : %10d\n", stats
->st_l1hit
);
2021 fprintf(out
, " Store L1D Miss : %10d\n", stats
->st_l1miss
);
2022 fprintf(out
, " No Page Map Rejects : %10d\n", stats
->nomap
);
2023 fprintf(out
, " Unable to parse data source : %10d\n", stats
->noparse
);
2026 static void print_shared_cacheline_info(FILE *out
)
2028 struct c2c_stats
*stats
= &c2c
.hitm_stats
;
2029 int hitm_cnt
= stats
->lcl_hitm
+ stats
->rmt_hitm
;
2031 fprintf(out
, "=================================================\n");
2032 fprintf(out
, " Global Shared Cache Line Event Information \n");
2033 fprintf(out
, "=================================================\n");
2034 fprintf(out
, " Total Shared Cache Lines : %10d\n", c2c
.shared_clines
);
2035 fprintf(out
, " Load HITs on shared lines : %10d\n", stats
->load
);
2036 fprintf(out
, " Fill Buffer Hits on shared lines : %10d\n", stats
->ld_fbhit
);
2037 fprintf(out
, " L1D hits on shared lines : %10d\n", stats
->ld_l1hit
);
2038 fprintf(out
, " L2D hits on shared lines : %10d\n", stats
->ld_l2hit
);
2039 fprintf(out
, " LLC hits on shared lines : %10d\n", stats
->ld_llchit
+ stats
->lcl_hitm
);
2040 fprintf(out
, " Locked Access on shared lines : %10d\n", stats
->locks
);
2041 fprintf(out
, " Store HITs on shared lines : %10d\n", stats
->store
);
2042 fprintf(out
, " Store L1D hits on shared lines : %10d\n", stats
->st_l1hit
);
2043 fprintf(out
, " Total Merged records : %10d\n", hitm_cnt
+ stats
->store
);
2046 static void print_cacheline(struct c2c_hists
*c2c_hists
,
2047 struct hist_entry
*he_cl
,
2048 struct perf_hpp_list
*hpp_list
,
2052 struct perf_hpp hpp
= {
2059 hists__fprintf_headers(&c2c_hists
->hists
, out
);
2065 fprintf(out
, " -------------------------------------------------------------\n");
2066 __hist_entry__snprintf(he_cl
, &hpp
, hpp_list
);
2067 fprintf(out
, "%s\n", bf
);
2068 fprintf(out
, " -------------------------------------------------------------\n");
2070 hists__fprintf(&c2c_hists
->hists
, false, 0, 0, 0, out
, true);
2073 static void print_pareto(FILE *out
)
2075 struct perf_hpp_list hpp_list
;
2079 perf_hpp_list__init(&hpp_list
);
2080 ret
= hpp_list__parse(&hpp_list
,
2089 if (WARN_ONCE(ret
, "failed to setup sort entries\n"))
2092 nd
= rb_first(&c2c
.hists
.hists
.entries
);
2094 for (; nd
; nd
= rb_next(nd
)) {
2095 struct hist_entry
*he
= rb_entry(nd
, struct hist_entry
, rb_node
);
2096 struct c2c_hist_entry
*c2c_he
;
2101 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
2102 print_cacheline(c2c_he
->hists
, he
, &hpp_list
, out
);
2106 static void print_c2c_info(FILE *out
, struct perf_session
*session
)
2108 struct perf_evlist
*evlist
= session
->evlist
;
2109 struct perf_evsel
*evsel
;
2112 fprintf(out
, "=================================================\n");
2113 fprintf(out
, " c2c details \n");
2114 fprintf(out
, "=================================================\n");
2116 evlist__for_each_entry(evlist
, evsel
) {
2117 fprintf(out
, "%-36s: %s\n", first
? " Events" : "",
2118 perf_evsel__name(evsel
));
2121 fprintf(out
, " Cachelines sort on : %s HITMs\n",
2122 display_str
[c2c
.display
]);
2123 fprintf(out
, " Cacheline data grouping : %s\n", c2c
.cl_sort
);
2126 static void perf_c2c__hists_fprintf(FILE *out
, struct perf_session
*session
)
2130 print_c2c__display_stats(out
);
2132 print_shared_cacheline_info(out
);
2134 print_c2c_info(out
, session
);
2140 fprintf(out
, "=================================================\n");
2141 fprintf(out
, " Shared Data Cache Line Table \n");
2142 fprintf(out
, "=================================================\n");
2143 fprintf(out
, "#\n");
2145 hists__fprintf(&c2c
.hists
.hists
, true, 0, 0, 0, stdout
, false);
2148 fprintf(out
, "=================================================\n");
2149 fprintf(out
, " Shared Cache Line Distribution Pareto \n");
2150 fprintf(out
, "=================================================\n");
2151 fprintf(out
, "#\n");
2156 #ifdef HAVE_SLANG_SUPPORT
2157 static void c2c_browser__update_nr_entries(struct hist_browser
*hb
)
2160 struct rb_node
*nd
= rb_first(&hb
->hists
->entries
);
2163 struct hist_entry
*he
= rb_entry(nd
, struct hist_entry
, rb_node
);
2171 hb
->nr_non_filtered_entries
= nr_entries
;
2174 struct c2c_cacheline_browser
{
2175 struct hist_browser hb
;
2176 struct hist_entry
*he
;
2180 perf_c2c_cacheline_browser__title(struct hist_browser
*browser
,
2181 char *bf
, size_t size
)
2183 struct c2c_cacheline_browser
*cl_browser
;
2184 struct hist_entry
*he
;
2187 cl_browser
= container_of(browser
, struct c2c_cacheline_browser
, hb
);
2188 he
= cl_browser
->he
;
2191 addr
= cl_address(he
->mem_info
->daddr
.addr
);
2193 scnprintf(bf
, size
, "Cacheline 0x%lx", addr
);
2197 static struct c2c_cacheline_browser
*
2198 c2c_cacheline_browser__new(struct hists
*hists
, struct hist_entry
*he
)
2200 struct c2c_cacheline_browser
*browser
;
2202 browser
= zalloc(sizeof(*browser
));
2204 hist_browser__init(&browser
->hb
, hists
);
2205 browser
->hb
.c2c_filter
= true;
2206 browser
->hb
.title
= perf_c2c_cacheline_browser__title
;
2213 static int perf_c2c__browse_cacheline(struct hist_entry
*he
)
2215 struct c2c_hist_entry
*c2c_he
;
2216 struct c2c_hists
*c2c_hists
;
2217 struct c2c_cacheline_browser
*cl_browser
;
2218 struct hist_browser
*browser
;
2221 " ENTER Togle callchains (if present) \n"
2222 " n Togle Node details info \n"
2223 " s Togle full lenght of symbol and source line columns \n"
2224 " q Return back to cacheline list \n";
2226 /* Display compact version first. */
2227 c2c
.symbol_full
= false;
2229 c2c_he
= container_of(he
, struct c2c_hist_entry
, he
);
2230 c2c_hists
= c2c_he
->hists
;
2232 cl_browser
= c2c_cacheline_browser__new(&c2c_hists
->hists
, he
);
2233 if (cl_browser
== NULL
)
2236 browser
= &cl_browser
->hb
;
2238 /* reset abort key so that it can get Ctrl-C as a key */
2240 SLang_init_tty(0, 0, 0);
2242 c2c_browser__update_nr_entries(browser
);
2245 key
= hist_browser__run(browser
, "? - help");
2249 c2c
.symbol_full
= !c2c
.symbol_full
;
2252 c2c
.node_info
= (c2c
.node_info
+ 1) % 3;
2253 setup_nodes_header();
2258 ui_browser__help_window(&browser
->b
, help
);
2270 static int perf_c2c_browser__title(struct hist_browser
*browser
,
2271 char *bf
, size_t size
)
2274 "Shared Data Cache Line Table "
2275 "(%lu entries, sorted on %s HITMs)",
2276 browser
->nr_non_filtered_entries
,
2277 display_str
[c2c
.display
]);
2281 static struct hist_browser
*
2282 perf_c2c_browser__new(struct hists
*hists
)
2284 struct hist_browser
*browser
= hist_browser__new(hists
);
2287 browser
->title
= perf_c2c_browser__title
;
2288 browser
->c2c_filter
= true;
2294 static int perf_c2c__hists_browse(struct hists
*hists
)
2296 struct hist_browser
*browser
;
2299 " d Display cacheline details \n"
2300 " ENTER Togle callchains (if present) \n"
2303 browser
= perf_c2c_browser__new(hists
);
2304 if (browser
== NULL
)
2307 /* reset abort key so that it can get Ctrl-C as a key */
2309 SLang_init_tty(0, 0, 0);
2311 c2c_browser__update_nr_entries(browser
);
2314 key
= hist_browser__run(browser
, "? - help");
2320 perf_c2c__browse_cacheline(browser
->he_selection
);
2323 ui_browser__help_window(&browser
->b
, help
);
2331 hist_browser__delete(browser
);
2335 static void perf_c2c_display(struct perf_session
*session
)
2338 perf_c2c__hists_fprintf(stdout
, session
);
2340 perf_c2c__hists_browse(&c2c
.hists
.hists
);
2343 static void perf_c2c_display(struct perf_session
*session
)
2346 perf_c2c__hists_fprintf(stdout
, session
);
2348 #endif /* HAVE_SLANG_SUPPORT */
2350 static void ui_quirks(void)
2352 if (!c2c
.use_stdio
) {
2353 dim_offset
.width
= 5;
2354 dim_offset
.header
= header_offset_tui
;
2357 dim_percent_hitm
.header
= percent_hitm_header
[c2c
.display
];
2360 #define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
2362 const char callchain_help
[] = "Display call graph (stack chain/backtrace):\n\n"
2363 CALLCHAIN_REPORT_HELP
2364 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT
;
2367 parse_callchain_opt(const struct option
*opt
, const char *arg
, int unset
)
2369 struct callchain_param
*callchain
= opt
->value
;
2371 callchain
->enabled
= !unset
;
2376 symbol_conf
.use_callchain
= false;
2377 callchain
->mode
= CHAIN_NONE
;
2381 return parse_callchain_report_opt(arg
);
2384 static int setup_callchain(struct perf_evlist
*evlist
)
2386 u64 sample_type
= perf_evlist__combined_sample_type(evlist
);
2387 enum perf_call_graph_mode mode
= CALLCHAIN_NONE
;
2389 if ((sample_type
& PERF_SAMPLE_REGS_USER
) &&
2390 (sample_type
& PERF_SAMPLE_STACK_USER
))
2391 mode
= CALLCHAIN_DWARF
;
2392 else if (sample_type
& PERF_SAMPLE_BRANCH_STACK
)
2393 mode
= CALLCHAIN_LBR
;
2394 else if (sample_type
& PERF_SAMPLE_CALLCHAIN
)
2395 mode
= CALLCHAIN_FP
;
2397 if (!callchain_param
.enabled
&&
2398 callchain_param
.mode
!= CHAIN_NONE
&&
2399 mode
!= CALLCHAIN_NONE
) {
2400 symbol_conf
.use_callchain
= true;
2401 if (callchain_register_param(&callchain_param
) < 0) {
2402 ui__error("Can't register callchain params.\n");
2407 callchain_param
.record_mode
= mode
;
2408 callchain_param
.min_percent
= 0;
2412 static int setup_display(const char *str
)
2414 const char *display
= str
?: "tot";
2416 if (!strcmp(display
, "tot"))
2417 c2c
.display
= DISPLAY_TOT
;
2418 else if (!strcmp(display
, "rmt"))
2419 c2c
.display
= DISPLAY_RMT
;
2420 else if (!strcmp(display
, "lcl"))
2421 c2c
.display
= DISPLAY_LCL
;
2423 pr_err("failed: unknown display type: %s\n", str
);
2430 #define for_each_token(__tok, __buf, __sep, __tmp) \
2431 for (__tok = strtok_r(__buf, __sep, &__tmp); __tok; \
2432 __tok = strtok_r(NULL, __sep, &__tmp))
2434 static int build_cl_output(char *cl_sort
, bool no_source
)
2436 char *tok
, *tmp
, *buf
= strdup(cl_sort
);
2437 bool add_pid
= false;
2438 bool add_tid
= false;
2439 bool add_iaddr
= false;
2440 bool add_sym
= false;
2441 bool add_dso
= false;
2442 bool add_src
= false;
2447 for_each_token(tok
, buf
, ",", tmp
) {
2448 if (!strcmp(tok
, "tid")) {
2450 } else if (!strcmp(tok
, "pid")) {
2452 } else if (!strcmp(tok
, "iaddr")) {
2456 add_src
= no_source
? false : true;
2457 } else if (!strcmp(tok
, "dso")) {
2459 } else if (strcmp(tok
, "offset")) {
2460 pr_err("unrecognized sort token: %s\n", tok
);
2465 if (asprintf(&c2c
.cl_output
,
2466 "%s%s%s%s%s%s%s%s%s%s",
2467 c2c
.use_stdio
? "cl_num_empty," : "",
2470 "percent_stores_l1hit,"
2471 "percent_stores_l1miss,"
2473 add_pid
? "pid," : "",
2474 add_tid
? "tid," : "",
2475 add_iaddr
? "iaddr," : "",
2481 add_sym
? "symbol," : "",
2482 add_dso
? "dso," : "",
2483 add_src
? "cl_srcline," : "",
2487 c2c
.show_src
= add_src
;
2493 static int setup_coalesce(const char *coalesce
, bool no_source
)
2495 const char *c
= coalesce
?: coalesce_default
;
2497 if (asprintf(&c2c
.cl_sort
, "offset,%s", c
) < 0)
2500 if (build_cl_output(c2c
.cl_sort
, no_source
))
2503 if (asprintf(&c2c
.cl_resort
, "offset,%s",
2504 c2c
.display
== DISPLAY_TOT
?
2506 c2c
.display
== DISPLAY_RMT
?
2507 "rmt_hitm,lcl_hitm" :
2508 "lcl_hitm,rmt_hitm") < 0)
2511 pr_debug("coalesce sort fields: %s\n", c2c
.cl_sort
);
2512 pr_debug("coalesce resort fields: %s\n", c2c
.cl_resort
);
2513 pr_debug("coalesce output fields: %s\n", c2c
.cl_output
);
2517 static int perf_c2c__report(int argc
, const char **argv
)
2519 struct perf_session
*session
;
2520 struct ui_progress prog
;
2521 struct perf_data_file file
= {
2522 .mode
= PERF_DATA_MODE_READ
,
2524 char callchain_default_opt
[] = CALLCHAIN_DEFAULT_OPT
;
2525 const char *display
= NULL
;
2526 const char *coalesce
= NULL
;
2527 bool no_source
= false;
2528 const struct option options
[] = {
2529 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
2530 "file", "vmlinux pathname"),
2531 OPT_STRING('i', "input", &input_name
, "file",
2532 "the input file to process"),
2533 OPT_INCR('N', "node-info", &c2c
.node_info
,
2534 "show extra node info in report (repeat for more info)"),
2535 #ifdef HAVE_SLANG_SUPPORT
2536 OPT_BOOLEAN(0, "stdio", &c2c
.use_stdio
, "Use the stdio interface"),
2538 OPT_BOOLEAN(0, "stats", &c2c
.stats_only
,
2539 "Use the stdio interface"),
2540 OPT_BOOLEAN(0, "full-symbols", &c2c
.symbol_full
,
2541 "Display full length of symbols"),
2542 OPT_BOOLEAN(0, "no-source", &no_source
,
2543 "Do not display Source Line column"),
2544 OPT_BOOLEAN(0, "show-all", &c2c
.show_all
,
2545 "Show all captured HITM lines."),
2546 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param
,
2547 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
2548 callchain_help
, &parse_callchain_opt
,
2549 callchain_default_opt
),
2550 OPT_STRING('d', "display", &display
, "Switch HITM output type", "lcl,rmt"),
2551 OPT_STRING('c', "coalesce", &coalesce
, "coalesce fields",
2552 "coalesce fields: pid,tid,iaddr,dso"),
2553 OPT_BOOLEAN('f', "force", &symbol_conf
.force
, "don't complain, do it"),
2554 OPT_PARENT(c2c_options
),
2559 argc
= parse_options(argc
, argv
, options
, report_c2c_usage
,
2560 PARSE_OPT_STOP_AT_NON_OPTION
);
2562 usage_with_options(report_c2c_usage
, options
);
2565 c2c
.use_stdio
= true;
2567 if (!input_name
|| !strlen(input_name
))
2568 input_name
= "perf.data";
2570 file
.path
= input_name
;
2571 file
.force
= symbol_conf
.force
;
2573 err
= setup_display(display
);
2577 err
= setup_coalesce(coalesce
, no_source
);
2579 pr_debug("Failed to initialize hists\n");
2583 err
= c2c_hists__init(&c2c
.hists
, "dcacheline", 2);
2585 pr_debug("Failed to initialize hists\n");
2589 session
= perf_session__new(&file
, 0, &c2c
.tool
);
2590 if (session
== NULL
) {
2591 pr_debug("No memory for session\n");
2595 err
= setup_nodes(session
);
2597 pr_err("Failed setup nodes\n");
2601 err
= setup_callchain(session
->evlist
);
2605 if (symbol__init(&session
->header
.env
) < 0)
2608 /* No pipe support at the moment. */
2609 if (perf_data_file__is_pipe(session
->file
)) {
2610 pr_debug("No pipe support at the moment.\n");
2619 setup_browser(false);
2621 err
= perf_session__process_events(session
);
2623 pr_err("failed to process sample\n");
2627 c2c_hists__reinit(&c2c
.hists
,
2632 "tot_hitm,lcl_hitm,rmt_hitm,"
2633 "stores,stores_l1hit,stores_l1miss,"
2634 "dram_lcl,dram_rmt,"
2637 "ld_fbhit,ld_l1hit,ld_l2hit,"
2638 "ld_lclhit,ld_rmthit",
2639 c2c
.display
== DISPLAY_TOT
? "tot_hitm" :
2640 c2c
.display
== DISPLAY_LCL
? "lcl_hitm" : "rmt_hitm"
2643 ui_progress__init(&prog
, c2c
.hists
.hists
.nr_entries
, "Sorting...");
2645 hists__collapse_resort(&c2c
.hists
.hists
, NULL
);
2646 hists__output_resort_cb(&c2c
.hists
.hists
, &prog
, resort_hitm_cb
);
2647 hists__iterate_cb(&c2c
.hists
.hists
, resort_cl_cb
);
2649 ui_progress__finish();
2653 perf_c2c_display(session
);
2656 perf_session__delete(session
);
2661 static int parse_record_events(const struct option
*opt
,
2662 const char *str
, int unset __maybe_unused
)
2664 bool *event_set
= (bool *) opt
->value
;
2667 return perf_mem_events__parse(str
);
2671 static const char * const __usage_record
[] = {
2672 "perf c2c record [<options>] [<command>]",
2673 "perf c2c record [<options>] -- <command> [<options>]",
2677 static const char * const *record_mem_usage
= __usage_record
;
2679 static int perf_c2c__record(int argc
, const char **argv
)
2681 int rec_argc
, i
= 0, j
;
2682 const char **rec_argv
;
2684 bool all_user
= false, all_kernel
= false;
2685 bool event_set
= false;
2686 struct option options
[] = {
2687 OPT_CALLBACK('e', "event", &event_set
, "event",
2688 "event selector. Use 'perf mem record -e list' to list available events",
2689 parse_record_events
),
2690 OPT_BOOLEAN('u', "all-user", &all_user
, "collect only user level data"),
2691 OPT_BOOLEAN('k', "all-kernel", &all_kernel
, "collect only kernel level data"),
2692 OPT_UINTEGER('l', "ldlat", &perf_mem_events__loads_ldlat
, "setup mem-loads latency"),
2693 OPT_PARENT(c2c_options
),
2697 if (perf_mem_events__init()) {
2698 pr_err("failed: memory events not supported\n");
2702 argc
= parse_options(argc
, argv
, options
, record_mem_usage
,
2703 PARSE_OPT_KEEP_UNKNOWN
);
2705 rec_argc
= argc
+ 10; /* max number of arguments */
2706 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
2710 rec_argv
[i
++] = "record";
2713 perf_mem_events
[PERF_MEM_EVENTS__LOAD
].record
= true;
2714 perf_mem_events
[PERF_MEM_EVENTS__STORE
].record
= true;
2717 if (perf_mem_events
[PERF_MEM_EVENTS__LOAD
].record
)
2718 rec_argv
[i
++] = "-W";
2720 rec_argv
[i
++] = "-d";
2721 rec_argv
[i
++] = "--sample-cpu";
2723 for (j
= 0; j
< PERF_MEM_EVENTS__MAX
; j
++) {
2724 if (!perf_mem_events
[j
].record
)
2727 if (!perf_mem_events
[j
].supported
) {
2728 pr_err("failed: event '%s' not supported\n",
2729 perf_mem_events
[j
].name
);
2733 rec_argv
[i
++] = "-e";
2734 rec_argv
[i
++] = perf_mem_events__name(j
);
2738 rec_argv
[i
++] = "--all-user";
2741 rec_argv
[i
++] = "--all-kernel";
2743 for (j
= 0; j
< argc
; j
++, i
++)
2744 rec_argv
[i
] = argv
[j
];
2747 pr_debug("calling: ");
2751 while (rec_argv
[j
]) {
2752 pr_debug("%s ", rec_argv
[j
]);
2758 ret
= cmd_record(i
, rec_argv
, NULL
);
2763 int cmd_c2c(int argc
, const char **argv
, const char *prefix __maybe_unused
)
2765 argc
= parse_options(argc
, argv
, c2c_options
, c2c_usage
,
2766 PARSE_OPT_STOP_AT_NON_OPTION
);
2769 usage_with_options(c2c_usage
, c2c_options
);
2771 if (!strncmp(argv
[0], "rec", 3)) {
2772 return perf_c2c__record(argc
, argv
);
2773 } else if (!strncmp(argv
[0], "rep", 3)) {
2774 return perf_c2c__report(argc
, argv
);
2776 usage_with_options(c2c_usage
, c2c_options
);