1 // SPDX-License-Identifier: GPL-2.0
2 #include <stddef.h>
3 #include <stdlib.h>
4 #include <string.h>
5 #include <errno.h>
6 #include <sys/types.h>
7 #include <sys/stat.h>
8 #include <unistd.h>
9 #include <api/fs/fs.h>
10 #include <linux/kernel.h>
11 #include "map_symbol.h"
12 #include "mem-events.h"
13 #include "debug.h"
14 #include "symbol.h"
15 #include "pmu.h"
16 #include "pmu-hybrid.h"
17
18 unsigned int perf_mem_events__loads_ldlat = 30;
19
20 #define E(t, n, s) { .tag = t, .name = n, .sysfs_name = s }
21
22 static struct perf_mem_event perf_mem_events[PERF_MEM_EVENTS__MAX] = {
23 E("ldlat-loads", "cpu/mem-loads,ldlat=%u/P", "cpu/events/mem-loads"),
24 E("ldlat-stores", "cpu/mem-stores/P", "cpu/events/mem-stores"),
25 E(NULL, NULL, NULL),
26 };
27 #undef E
28
29 static char mem_loads_name[100];
30 static bool mem_loads_name__init;
31
perf_mem_events__ptr(int i)32 struct perf_mem_event * __weak perf_mem_events__ptr(int i)
33 {
34 if (i >= PERF_MEM_EVENTS__MAX)
35 return NULL;
36
37 return &perf_mem_events[i];
38 }
39
perf_mem_events__name(int i,char * pmu_name __maybe_unused)40 char * __weak perf_mem_events__name(int i, char *pmu_name __maybe_unused)
41 {
42 struct perf_mem_event *e = perf_mem_events__ptr(i);
43
44 if (!e)
45 return NULL;
46
47 if (i == PERF_MEM_EVENTS__LOAD) {
48 if (!mem_loads_name__init) {
49 mem_loads_name__init = true;
50 scnprintf(mem_loads_name, sizeof(mem_loads_name),
51 e->name, perf_mem_events__loads_ldlat);
52 }
53 return mem_loads_name;
54 }
55
56 return (char *)e->name;
57 }
58
is_mem_loads_aux_event(struct evsel * leader __maybe_unused)59 __weak bool is_mem_loads_aux_event(struct evsel *leader __maybe_unused)
60 {
61 return false;
62 }
63
perf_mem_events__parse(const char * str)64 int perf_mem_events__parse(const char *str)
65 {
66 char *tok, *saveptr = NULL;
67 bool found = false;
68 char *buf;
69 int j;
70
71 /* We need buffer that we know we can write to. */
72 buf = malloc(strlen(str) + 1);
73 if (!buf)
74 return -ENOMEM;
75
76 strcpy(buf, str);
77
78 tok = strtok_r((char *)buf, ",", &saveptr);
79
80 while (tok) {
81 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
82 struct perf_mem_event *e = perf_mem_events__ptr(j);
83
84 if (!e->tag)
85 continue;
86
87 if (strstr(e->tag, tok))
88 e->record = found = true;
89 }
90
91 tok = strtok_r(NULL, ",", &saveptr);
92 }
93
94 free(buf);
95
96 if (found)
97 return 0;
98
99 pr_err("failed: event '%s' not found, use '-e list' to get list of available events\n", str);
100 return -1;
101 }
102
perf_mem_event__supported(const char * mnt,char * sysfs_name)103 static bool perf_mem_event__supported(const char *mnt, char *sysfs_name)
104 {
105 char path[PATH_MAX];
106 struct stat st;
107
108 scnprintf(path, PATH_MAX, "%s/devices/%s", mnt, sysfs_name);
109 return !stat(path, &st);
110 }
111
perf_mem_events__init(void)112 int perf_mem_events__init(void)
113 {
114 const char *mnt = sysfs__mount();
115 bool found = false;
116 int j;
117
118 if (!mnt)
119 return -ENOENT;
120
121 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
122 struct perf_mem_event *e = perf_mem_events__ptr(j);
123 struct perf_pmu *pmu;
124 char sysfs_name[100];
125
126 /*
127 * If the event entry isn't valid, skip initialization
128 * and "e->supported" will keep false.
129 */
130 if (!e->tag)
131 continue;
132
133 if (!perf_pmu__has_hybrid()) {
134 scnprintf(sysfs_name, sizeof(sysfs_name),
135 e->sysfs_name, "cpu");
136 e->supported = perf_mem_event__supported(mnt, sysfs_name);
137 } else {
138 perf_pmu__for_each_hybrid_pmu(pmu) {
139 scnprintf(sysfs_name, sizeof(sysfs_name),
140 e->sysfs_name, pmu->name);
141 e->supported |= perf_mem_event__supported(mnt, sysfs_name);
142 }
143 }
144
145 if (e->supported)
146 found = true;
147 }
148
149 return found ? 0 : -ENOENT;
150 }
151
perf_mem_events__list(void)152 void perf_mem_events__list(void)
153 {
154 int j;
155
156 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
157 struct perf_mem_event *e = perf_mem_events__ptr(j);
158
159 fprintf(stderr, "%-*s%-*s%s",
160 e->tag ? 13 : 0,
161 e->tag ? : "",
162 e->tag && verbose > 0 ? 25 : 0,
163 e->tag && verbose > 0 ? perf_mem_events__name(j, NULL) : "",
164 e->supported ? ": available\n" : "");
165 }
166 }
167
perf_mem_events__print_unsupport_hybrid(struct perf_mem_event * e,int idx)168 static void perf_mem_events__print_unsupport_hybrid(struct perf_mem_event *e,
169 int idx)
170 {
171 const char *mnt = sysfs__mount();
172 char sysfs_name[100];
173 struct perf_pmu *pmu;
174
175 perf_pmu__for_each_hybrid_pmu(pmu) {
176 scnprintf(sysfs_name, sizeof(sysfs_name), e->sysfs_name,
177 pmu->name);
178 if (!perf_mem_event__supported(mnt, sysfs_name)) {
179 pr_err("failed: event '%s' not supported\n",
180 perf_mem_events__name(idx, pmu->name));
181 }
182 }
183 }
184
perf_mem_events__record_args(const char ** rec_argv,int * argv_nr,char ** rec_tmp,int * tmp_nr)185 int perf_mem_events__record_args(const char **rec_argv, int *argv_nr,
186 char **rec_tmp, int *tmp_nr)
187 {
188 int i = *argv_nr, k = 0;
189 struct perf_mem_event *e;
190 struct perf_pmu *pmu;
191 char *s;
192
193 for (int j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
194 e = perf_mem_events__ptr(j);
195 if (!e->record)
196 continue;
197
198 if (!perf_pmu__has_hybrid()) {
199 if (!e->supported) {
200 pr_err("failed: event '%s' not supported\n",
201 perf_mem_events__name(j, NULL));
202 return -1;
203 }
204
205 rec_argv[i++] = "-e";
206 rec_argv[i++] = perf_mem_events__name(j, NULL);
207 } else {
208 if (!e->supported) {
209 perf_mem_events__print_unsupport_hybrid(e, j);
210 return -1;
211 }
212
213 perf_pmu__for_each_hybrid_pmu(pmu) {
214 rec_argv[i++] = "-e";
215 s = perf_mem_events__name(j, pmu->name);
216 if (s) {
217 s = strdup(s);
218 if (!s)
219 return -1;
220
221 rec_argv[i++] = s;
222 rec_tmp[k++] = s;
223 }
224 }
225 }
226 }
227
228 *argv_nr = i;
229 *tmp_nr = k;
230 return 0;
231 }
232
233 static const char * const tlb_access[] = {
234 "N/A",
235 "HIT",
236 "MISS",
237 "L1",
238 "L2",
239 "Walker",
240 "Fault",
241 };
242
perf_mem__tlb_scnprintf(char * out,size_t sz,struct mem_info * mem_info)243 int perf_mem__tlb_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
244 {
245 size_t l = 0, i;
246 u64 m = PERF_MEM_TLB_NA;
247 u64 hit, miss;
248
249 sz -= 1; /* -1 for null termination */
250 out[0] = '\0';
251
252 if (mem_info)
253 m = mem_info->data_src.mem_dtlb;
254
255 hit = m & PERF_MEM_TLB_HIT;
256 miss = m & PERF_MEM_TLB_MISS;
257
258 /* already taken care of */
259 m &= ~(PERF_MEM_TLB_HIT|PERF_MEM_TLB_MISS);
260
261 for (i = 0; m && i < ARRAY_SIZE(tlb_access); i++, m >>= 1) {
262 if (!(m & 0x1))
263 continue;
264 if (l) {
265 strcat(out, " or ");
266 l += 4;
267 }
268 l += scnprintf(out + l, sz - l, tlb_access[i]);
269 }
270 if (*out == '\0')
271 l += scnprintf(out, sz - l, "N/A");
272 if (hit)
273 l += scnprintf(out + l, sz - l, " hit");
274 if (miss)
275 l += scnprintf(out + l, sz - l, " miss");
276
277 return l;
278 }
279
280 static const char * const mem_lvl[] = {
281 "N/A",
282 "HIT",
283 "MISS",
284 "L1",
285 "LFB/MAB",
286 "L2",
287 "L3",
288 "Local RAM",
289 "Remote RAM (1 hop)",
290 "Remote RAM (2 hops)",
291 "Remote Cache (1 hop)",
292 "Remote Cache (2 hops)",
293 "I/O",
294 "Uncached",
295 };
296
297 static const char * const mem_lvlnum[] = {
298 [PERF_MEM_LVLNUM_CXL] = "CXL",
299 [PERF_MEM_LVLNUM_IO] = "I/O",
300 [PERF_MEM_LVLNUM_ANY_CACHE] = "Any cache",
301 [PERF_MEM_LVLNUM_LFB] = "LFB/MAB",
302 [PERF_MEM_LVLNUM_RAM] = "RAM",
303 [PERF_MEM_LVLNUM_PMEM] = "PMEM",
304 [PERF_MEM_LVLNUM_NA] = "N/A",
305 };
306
307 static const char * const mem_hops[] = {
308 "N/A",
309 /*
310 * While printing, 'Remote' will be added to represent
311 * 'Remote core, same node' accesses as remote field need
312 * to be set with mem_hops field.
313 */
314 "core, same node",
315 "node, same socket",
316 "socket, same board",
317 "board",
318 };
319
perf_mem__op_scnprintf(char * out,size_t sz,struct mem_info * mem_info)320 static int perf_mem__op_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
321 {
322 u64 op = PERF_MEM_LOCK_NA;
323 int l;
324
325 if (mem_info)
326 op = mem_info->data_src.mem_op;
327
328 if (op & PERF_MEM_OP_NA)
329 l = scnprintf(out, sz, "N/A");
330 else if (op & PERF_MEM_OP_LOAD)
331 l = scnprintf(out, sz, "LOAD");
332 else if (op & PERF_MEM_OP_STORE)
333 l = scnprintf(out, sz, "STORE");
334 else if (op & PERF_MEM_OP_PFETCH)
335 l = scnprintf(out, sz, "PFETCH");
336 else if (op & PERF_MEM_OP_EXEC)
337 l = scnprintf(out, sz, "EXEC");
338 else
339 l = scnprintf(out, sz, "No");
340
341 return l;
342 }
343
perf_mem__lvl_scnprintf(char * out,size_t sz,struct mem_info * mem_info)344 int perf_mem__lvl_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
345 {
346 size_t i, l = 0;
347 u64 m = PERF_MEM_LVL_NA;
348 u64 hit, miss;
349 int printed = 0;
350
351 if (mem_info)
352 m = mem_info->data_src.mem_lvl;
353
354 sz -= 1; /* -1 for null termination */
355 out[0] = '\0';
356
357 hit = m & PERF_MEM_LVL_HIT;
358 miss = m & PERF_MEM_LVL_MISS;
359
360 /* already taken care of */
361 m &= ~(PERF_MEM_LVL_HIT|PERF_MEM_LVL_MISS);
362
363 if (mem_info && mem_info->data_src.mem_remote) {
364 strcat(out, "Remote ");
365 l += 7;
366 }
367
368 /*
369 * Incase mem_hops field is set, we can skip printing data source via
370 * PERF_MEM_LVL namespace.
371 */
372 if (mem_info && mem_info->data_src.mem_hops) {
373 l += scnprintf(out + l, sz - l, "%s ", mem_hops[mem_info->data_src.mem_hops]);
374 } else {
375 for (i = 0; m && i < ARRAY_SIZE(mem_lvl); i++, m >>= 1) {
376 if (!(m & 0x1))
377 continue;
378 if (printed++) {
379 strcat(out, " or ");
380 l += 4;
381 }
382 l += scnprintf(out + l, sz - l, mem_lvl[i]);
383 }
384 }
385
386 if (mem_info && mem_info->data_src.mem_lvl_num) {
387 int lvl = mem_info->data_src.mem_lvl_num;
388 if (printed++) {
389 strcat(out, " or ");
390 l += 4;
391 }
392 if (mem_lvlnum[lvl])
393 l += scnprintf(out + l, sz - l, mem_lvlnum[lvl]);
394 else
395 l += scnprintf(out + l, sz - l, "L%d", lvl);
396 }
397
398 if (l == 0)
399 l += scnprintf(out + l, sz - l, "N/A");
400 if (hit)
401 l += scnprintf(out + l, sz - l, " hit");
402 if (miss)
403 l += scnprintf(out + l, sz - l, " miss");
404
405 return l;
406 }
407
408 static const char * const snoop_access[] = {
409 "N/A",
410 "None",
411 "Hit",
412 "Miss",
413 "HitM",
414 };
415
416 static const char * const snoopx_access[] = {
417 "Fwd",
418 "Peer",
419 };
420
perf_mem__snp_scnprintf(char * out,size_t sz,struct mem_info * mem_info)421 int perf_mem__snp_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
422 {
423 size_t i, l = 0;
424 u64 m = PERF_MEM_SNOOP_NA;
425
426 sz -= 1; /* -1 for null termination */
427 out[0] = '\0';
428
429 if (mem_info)
430 m = mem_info->data_src.mem_snoop;
431
432 for (i = 0; m && i < ARRAY_SIZE(snoop_access); i++, m >>= 1) {
433 if (!(m & 0x1))
434 continue;
435 if (l) {
436 strcat(out, " or ");
437 l += 4;
438 }
439 l += scnprintf(out + l, sz - l, snoop_access[i]);
440 }
441
442 m = 0;
443 if (mem_info)
444 m = mem_info->data_src.mem_snoopx;
445
446 for (i = 0; m && i < ARRAY_SIZE(snoopx_access); i++, m >>= 1) {
447 if (!(m & 0x1))
448 continue;
449
450 if (l) {
451 strcat(out, " or ");
452 l += 4;
453 }
454 l += scnprintf(out + l, sz - l, snoopx_access[i]);
455 }
456
457 if (*out == '\0')
458 l += scnprintf(out, sz - l, "N/A");
459
460 return l;
461 }
462
perf_mem__lck_scnprintf(char * out,size_t sz,struct mem_info * mem_info)463 int perf_mem__lck_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
464 {
465 u64 mask = PERF_MEM_LOCK_NA;
466 int l;
467
468 if (mem_info)
469 mask = mem_info->data_src.mem_lock;
470
471 if (mask & PERF_MEM_LOCK_NA)
472 l = scnprintf(out, sz, "N/A");
473 else if (mask & PERF_MEM_LOCK_LOCKED)
474 l = scnprintf(out, sz, "Yes");
475 else
476 l = scnprintf(out, sz, "No");
477
478 return l;
479 }
480
perf_mem__blk_scnprintf(char * out,size_t sz,struct mem_info * mem_info)481 int perf_mem__blk_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
482 {
483 size_t l = 0;
484 u64 mask = PERF_MEM_BLK_NA;
485
486 sz -= 1; /* -1 for null termination */
487 out[0] = '\0';
488
489 if (mem_info)
490 mask = mem_info->data_src.mem_blk;
491
492 if (!mask || (mask & PERF_MEM_BLK_NA)) {
493 l += scnprintf(out + l, sz - l, " N/A");
494 return l;
495 }
496 if (mask & PERF_MEM_BLK_DATA)
497 l += scnprintf(out + l, sz - l, " Data");
498 if (mask & PERF_MEM_BLK_ADDR)
499 l += scnprintf(out + l, sz - l, " Addr");
500
501 return l;
502 }
503
perf_script__meminfo_scnprintf(char * out,size_t sz,struct mem_info * mem_info)504 int perf_script__meminfo_scnprintf(char *out, size_t sz, struct mem_info *mem_info)
505 {
506 int i = 0;
507
508 i += scnprintf(out, sz, "|OP ");
509 i += perf_mem__op_scnprintf(out + i, sz - i, mem_info);
510 i += scnprintf(out + i, sz - i, "|LVL ");
511 i += perf_mem__lvl_scnprintf(out + i, sz, mem_info);
512 i += scnprintf(out + i, sz - i, "|SNP ");
513 i += perf_mem__snp_scnprintf(out + i, sz - i, mem_info);
514 i += scnprintf(out + i, sz - i, "|TLB ");
515 i += perf_mem__tlb_scnprintf(out + i, sz - i, mem_info);
516 i += scnprintf(out + i, sz - i, "|LCK ");
517 i += perf_mem__lck_scnprintf(out + i, sz - i, mem_info);
518 i += scnprintf(out + i, sz - i, "|BLK ");
519 i += perf_mem__blk_scnprintf(out + i, sz - i, mem_info);
520
521 return i;
522 }
523
c2c_decode_stats(struct c2c_stats * stats,struct mem_info * mi)524 int c2c_decode_stats(struct c2c_stats *stats, struct mem_info *mi)
525 {
526 union perf_mem_data_src *data_src = &mi->data_src;
527 u64 daddr = mi->daddr.addr;
528 u64 op = data_src->mem_op;
529 u64 lvl = data_src->mem_lvl;
530 u64 snoop = data_src->mem_snoop;
531 u64 snoopx = data_src->mem_snoopx;
532 u64 lock = data_src->mem_lock;
533 u64 blk = data_src->mem_blk;
534 /*
535 * Skylake might report unknown remote level via this
536 * bit, consider it when evaluating remote HITMs.
537 *
538 * Incase of power, remote field can also be used to denote cache
539 * accesses from the another core of same node. Hence, setting
540 * mrem only when HOPS is zero along with set remote field.
541 */
542 bool mrem = (data_src->mem_remote && !data_src->mem_hops);
543 int err = 0;
544
545 #define HITM_INC(__f) \
546 do { \
547 stats->__f++; \
548 stats->tot_hitm++; \
549 } while (0)
550
551 #define PEER_INC(__f) \
552 do { \
553 stats->__f++; \
554 stats->tot_peer++; \
555 } while (0)
556
557 #define P(a, b) PERF_MEM_##a##_##b
558
559 stats->nr_entries++;
560
561 if (lock & P(LOCK, LOCKED)) stats->locks++;
562
563 if (blk & P(BLK, DATA)) stats->blk_data++;
564 if (blk & P(BLK, ADDR)) stats->blk_addr++;
565
566 if (op & P(OP, LOAD)) {
567 /* load */
568 stats->load++;
569
570 if (!daddr) {
571 stats->ld_noadrs++;
572 return -1;
573 }
574
575 if (lvl & P(LVL, HIT)) {
576 if (lvl & P(LVL, UNC)) stats->ld_uncache++;
577 if (lvl & P(LVL, IO)) stats->ld_io++;
578 if (lvl & P(LVL, LFB)) stats->ld_fbhit++;
579 if (lvl & P(LVL, L1 )) stats->ld_l1hit++;
580 if (lvl & P(LVL, L2)) {
581 stats->ld_l2hit++;
582
583 if (snoopx & P(SNOOPX, PEER))
584 PEER_INC(lcl_peer);
585 }
586 if (lvl & P(LVL, L3 )) {
587 if (snoop & P(SNOOP, HITM))
588 HITM_INC(lcl_hitm);
589 else
590 stats->ld_llchit++;
591
592 if (snoopx & P(SNOOPX, PEER))
593 PEER_INC(lcl_peer);
594 }
595
596 if (lvl & P(LVL, LOC_RAM)) {
597 stats->lcl_dram++;
598 if (snoop & P(SNOOP, HIT))
599 stats->ld_shared++;
600 else
601 stats->ld_excl++;
602 }
603
604 if ((lvl & P(LVL, REM_RAM1)) ||
605 (lvl & P(LVL, REM_RAM2)) ||
606 mrem) {
607 stats->rmt_dram++;
608 if (snoop & P(SNOOP, HIT))
609 stats->ld_shared++;
610 else
611 stats->ld_excl++;
612 }
613 }
614
615 if ((lvl & P(LVL, REM_CCE1)) ||
616 (lvl & P(LVL, REM_CCE2)) ||
617 mrem) {
618 if (snoop & P(SNOOP, HIT)) {
619 stats->rmt_hit++;
620 } else if (snoop & P(SNOOP, HITM)) {
621 HITM_INC(rmt_hitm);
622 } else if (snoopx & P(SNOOPX, PEER)) {
623 stats->rmt_hit++;
624 PEER_INC(rmt_peer);
625 }
626 }
627
628 if ((lvl & P(LVL, MISS)))
629 stats->ld_miss++;
630
631 } else if (op & P(OP, STORE)) {
632 /* store */
633 stats->store++;
634
635 if (!daddr) {
636 stats->st_noadrs++;
637 return -1;
638 }
639
640 if (lvl & P(LVL, HIT)) {
641 if (lvl & P(LVL, UNC)) stats->st_uncache++;
642 if (lvl & P(LVL, L1 )) stats->st_l1hit++;
643 }
644 if (lvl & P(LVL, MISS))
645 if (lvl & P(LVL, L1)) stats->st_l1miss++;
646 if (lvl & P(LVL, NA))
647 stats->st_na++;
648 } else {
649 /* unparsable data_src? */
650 stats->noparse++;
651 return -1;
652 }
653
654 if (!mi->daddr.ms.map || !mi->iaddr.ms.map) {
655 stats->nomap++;
656 return -1;
657 }
658
659 #undef P
660 #undef HITM_INC
661 return err;
662 }
663
c2c_add_stats(struct c2c_stats * stats,struct c2c_stats * add)664 void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add)
665 {
666 stats->nr_entries += add->nr_entries;
667
668 stats->locks += add->locks;
669 stats->store += add->store;
670 stats->st_uncache += add->st_uncache;
671 stats->st_noadrs += add->st_noadrs;
672 stats->st_l1hit += add->st_l1hit;
673 stats->st_l1miss += add->st_l1miss;
674 stats->st_na += add->st_na;
675 stats->load += add->load;
676 stats->ld_excl += add->ld_excl;
677 stats->ld_shared += add->ld_shared;
678 stats->ld_uncache += add->ld_uncache;
679 stats->ld_io += add->ld_io;
680 stats->ld_miss += add->ld_miss;
681 stats->ld_noadrs += add->ld_noadrs;
682 stats->ld_fbhit += add->ld_fbhit;
683 stats->ld_l1hit += add->ld_l1hit;
684 stats->ld_l2hit += add->ld_l2hit;
685 stats->ld_llchit += add->ld_llchit;
686 stats->lcl_hitm += add->lcl_hitm;
687 stats->rmt_hitm += add->rmt_hitm;
688 stats->tot_hitm += add->tot_hitm;
689 stats->lcl_peer += add->lcl_peer;
690 stats->rmt_peer += add->rmt_peer;
691 stats->tot_peer += add->tot_peer;
692 stats->rmt_hit += add->rmt_hit;
693 stats->lcl_dram += add->lcl_dram;
694 stats->rmt_dram += add->rmt_dram;
695 stats->blk_data += add->blk_data;
696 stats->blk_addr += add->blk_addr;
697 stats->nomap += add->nomap;
698 stats->noparse += add->noparse;
699 }
700