xref: /linux/tools/perf/util/cpumap.c (revision 90e0d94d369d342e735a75174439482119b6c393)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <api/fs/fs.h>
3 #include "cpumap.h"
4 #include "debug.h"
5 #include "event.h"
6 #include <assert.h>
7 #include <dirent.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <linux/bitmap.h>
11 #include "asm/bug.h"
12 
13 #include <linux/ctype.h>
14 #include <linux/zalloc.h>
15 
16 static struct perf_cpu max_cpu_num;
17 static struct perf_cpu max_present_cpu_num;
18 static int max_node_num;
19 /**
20  * The numa node X as read from /sys/devices/system/node/nodeX indexed by the
21  * CPU number.
22  */
23 static int *cpunode_map;
24 
25 bool perf_record_cpu_map_data__test_bit(int i,
26 					const struct perf_record_cpu_map_data *data)
27 {
28 	int bit_word32 = i / 32;
29 	__u32 bit_mask32 = 1U << (i & 31);
30 	int bit_word64 = i / 64;
31 	__u64 bit_mask64 = ((__u64)1) << (i & 63);
32 
33 	return (data->mask32_data.long_size == 4)
34 		? (bit_word32 < data->mask32_data.nr) &&
35 		(data->mask32_data.mask[bit_word32] & bit_mask32) != 0
36 		: (bit_word64 < data->mask64_data.nr) &&
37 		(data->mask64_data.mask[bit_word64] & bit_mask64) != 0;
38 }
39 
40 /* Read ith mask value from data into the given 64-bit sized bitmap */
41 static void perf_record_cpu_map_data__read_one_mask(const struct perf_record_cpu_map_data *data,
42 						    int i, unsigned long *bitmap)
43 {
44 #if __SIZEOF_LONG__ == 8
45 	if (data->mask32_data.long_size == 4)
46 		bitmap[0] = data->mask32_data.mask[i];
47 	else
48 		bitmap[0] = data->mask64_data.mask[i];
49 #else
50 	if (data->mask32_data.long_size == 4) {
51 		bitmap[0] = data->mask32_data.mask[i];
52 		bitmap[1] = 0;
53 	} else {
54 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
55 		bitmap[0] = (unsigned long)(data->mask64_data.mask[i] >> 32);
56 		bitmap[1] = (unsigned long)data->mask64_data.mask[i];
57 #else
58 		bitmap[0] = (unsigned long)data->mask64_data.mask[i];
59 		bitmap[1] = (unsigned long)(data->mask64_data.mask[i] >> 32);
60 #endif
61 	}
62 #endif
63 }
64 static struct perf_cpu_map *cpu_map__from_entries(const struct perf_record_cpu_map_data *data)
65 {
66 	struct perf_cpu_map *map;
67 
68 	map = perf_cpu_map__empty_new(data->cpus_data.nr);
69 	if (map) {
70 		unsigned i;
71 
72 		for (i = 0; i < data->cpus_data.nr; i++) {
73 			/*
74 			 * Special treatment for -1, which is not real cpu number,
75 			 * and we need to use (int) -1 to initialize map[i],
76 			 * otherwise it would become 65535.
77 			 */
78 			if (data->cpus_data.cpu[i] == (u16) -1)
79 				map->map[i].cpu = -1;
80 			else
81 				map->map[i].cpu = (int) data->cpus_data.cpu[i];
82 		}
83 	}
84 
85 	return map;
86 }
87 
88 static struct perf_cpu_map *cpu_map__from_mask(const struct perf_record_cpu_map_data *data)
89 {
90 	DECLARE_BITMAP(local_copy, 64);
91 	int weight = 0, mask_nr = data->mask32_data.nr;
92 	struct perf_cpu_map *map;
93 
94 	for (int i = 0; i < mask_nr; i++) {
95 		perf_record_cpu_map_data__read_one_mask(data, i, local_copy);
96 		weight += bitmap_weight(local_copy, 64);
97 	}
98 
99 	map = perf_cpu_map__empty_new(weight);
100 	if (!map)
101 		return NULL;
102 
103 	for (int i = 0, j = 0; i < mask_nr; i++) {
104 		int cpus_per_i = (i * data->mask32_data.long_size  * BITS_PER_BYTE);
105 		int cpu;
106 
107 		perf_record_cpu_map_data__read_one_mask(data, i, local_copy);
108 		for_each_set_bit(cpu, local_copy, 64)
109 			map->map[j++].cpu = cpu + cpus_per_i;
110 	}
111 	return map;
112 
113 }
114 
115 struct perf_cpu_map *cpu_map__new_data(const struct perf_record_cpu_map_data *data)
116 {
117 	if (data->type == PERF_CPU_MAP__CPUS)
118 		return cpu_map__from_entries(data);
119 	else
120 		return cpu_map__from_mask(data);
121 }
122 
123 size_t cpu_map__fprintf(struct perf_cpu_map *map, FILE *fp)
124 {
125 #define BUFSIZE 1024
126 	char buf[BUFSIZE];
127 
128 	cpu_map__snprint(map, buf, sizeof(buf));
129 	return fprintf(fp, "%s\n", buf);
130 #undef BUFSIZE
131 }
132 
133 struct perf_cpu_map *perf_cpu_map__empty_new(int nr)
134 {
135 	struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + sizeof(int) * nr);
136 
137 	if (cpus != NULL) {
138 		int i;
139 
140 		cpus->nr = nr;
141 		for (i = 0; i < nr; i++)
142 			cpus->map[i].cpu = -1;
143 
144 		refcount_set(&cpus->refcnt, 1);
145 	}
146 
147 	return cpus;
148 }
149 
150 struct cpu_aggr_map *cpu_aggr_map__empty_new(int nr)
151 {
152 	struct cpu_aggr_map *cpus = malloc(sizeof(*cpus) + sizeof(struct aggr_cpu_id) * nr);
153 
154 	if (cpus != NULL) {
155 		int i;
156 
157 		cpus->nr = nr;
158 		for (i = 0; i < nr; i++)
159 			cpus->map[i] = aggr_cpu_id__empty();
160 
161 		refcount_set(&cpus->refcnt, 1);
162 	}
163 
164 	return cpus;
165 }
166 
167 static int cpu__get_topology_int(int cpu, const char *name, int *value)
168 {
169 	char path[PATH_MAX];
170 
171 	snprintf(path, PATH_MAX,
172 		"devices/system/cpu/cpu%d/topology/%s", cpu, name);
173 
174 	return sysfs__read_int(path, value);
175 }
176 
177 int cpu__get_socket_id(struct perf_cpu cpu)
178 {
179 	int value, ret = cpu__get_topology_int(cpu.cpu, "physical_package_id", &value);
180 	return ret ?: value;
181 }
182 
183 struct aggr_cpu_id aggr_cpu_id__socket(struct perf_cpu cpu, void *data __maybe_unused)
184 {
185 	struct aggr_cpu_id id = aggr_cpu_id__empty();
186 
187 	id.socket = cpu__get_socket_id(cpu);
188 	return id;
189 }
190 
191 static int aggr_cpu_id__cmp(const void *a_pointer, const void *b_pointer)
192 {
193 	struct aggr_cpu_id *a = (struct aggr_cpu_id *)a_pointer;
194 	struct aggr_cpu_id *b = (struct aggr_cpu_id *)b_pointer;
195 
196 	if (a->node != b->node)
197 		return a->node - b->node;
198 	else if (a->socket != b->socket)
199 		return a->socket - b->socket;
200 	else if (a->die != b->die)
201 		return a->die - b->die;
202 	else if (a->core != b->core)
203 		return a->core - b->core;
204 	else
205 		return a->thread - b->thread;
206 }
207 
208 struct cpu_aggr_map *cpu_aggr_map__new(const struct perf_cpu_map *cpus,
209 				       aggr_cpu_id_get_t get_id,
210 				       void *data)
211 {
212 	int idx;
213 	struct perf_cpu cpu;
214 	struct cpu_aggr_map *c = cpu_aggr_map__empty_new(cpus->nr);
215 
216 	if (!c)
217 		return NULL;
218 
219 	/* Reset size as it may only be partially filled */
220 	c->nr = 0;
221 
222 	perf_cpu_map__for_each_cpu(cpu, idx, cpus) {
223 		bool duplicate = false;
224 		struct aggr_cpu_id cpu_id = get_id(cpu, data);
225 
226 		for (int j = 0; j < c->nr; j++) {
227 			if (aggr_cpu_id__equal(&cpu_id, &c->map[j])) {
228 				duplicate = true;
229 				break;
230 			}
231 		}
232 		if (!duplicate) {
233 			c->map[c->nr] = cpu_id;
234 			c->nr++;
235 		}
236 	}
237 	/* Trim. */
238 	if (c->nr != cpus->nr) {
239 		struct cpu_aggr_map *trimmed_c =
240 			realloc(c,
241 				sizeof(struct cpu_aggr_map) + sizeof(struct aggr_cpu_id) * c->nr);
242 
243 		if (trimmed_c)
244 			c = trimmed_c;
245 	}
246 	/* ensure we process id in increasing order */
247 	qsort(c->map, c->nr, sizeof(struct aggr_cpu_id), aggr_cpu_id__cmp);
248 
249 	return c;
250 
251 }
252 
253 int cpu__get_die_id(struct perf_cpu cpu)
254 {
255 	int value, ret = cpu__get_topology_int(cpu.cpu, "die_id", &value);
256 
257 	return ret ?: value;
258 }
259 
260 struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data)
261 {
262 	struct aggr_cpu_id id;
263 	int die;
264 
265 	die = cpu__get_die_id(cpu);
266 	/* There is no die_id on legacy system. */
267 	if (die == -1)
268 		die = 0;
269 
270 	/*
271 	 * die_id is relative to socket, so start
272 	 * with the socket ID and then add die to
273 	 * make a unique ID.
274 	 */
275 	id = aggr_cpu_id__socket(cpu, data);
276 	if (aggr_cpu_id__is_empty(&id))
277 		return id;
278 
279 	id.die = die;
280 	return id;
281 }
282 
283 int cpu__get_core_id(struct perf_cpu cpu)
284 {
285 	int value, ret = cpu__get_topology_int(cpu.cpu, "core_id", &value);
286 	return ret ?: value;
287 }
288 
289 struct aggr_cpu_id aggr_cpu_id__core(struct perf_cpu cpu, void *data)
290 {
291 	struct aggr_cpu_id id;
292 	int core = cpu__get_core_id(cpu);
293 
294 	/* aggr_cpu_id__die returns a struct with socket and die set. */
295 	id = aggr_cpu_id__die(cpu, data);
296 	if (aggr_cpu_id__is_empty(&id))
297 		return id;
298 
299 	/*
300 	 * core_id is relative to socket and die, we need a global id.
301 	 * So we combine the result from cpu_map__get_die with the core id
302 	 */
303 	id.core = core;
304 	return id;
305 
306 }
307 
308 struct aggr_cpu_id aggr_cpu_id__cpu(struct perf_cpu cpu, void *data)
309 {
310 	struct aggr_cpu_id id;
311 
312 	/* aggr_cpu_id__core returns a struct with socket, die and core set. */
313 	id = aggr_cpu_id__core(cpu, data);
314 	if (aggr_cpu_id__is_empty(&id))
315 		return id;
316 
317 	id.cpu = cpu;
318 	return id;
319 
320 }
321 
322 struct aggr_cpu_id aggr_cpu_id__node(struct perf_cpu cpu, void *data __maybe_unused)
323 {
324 	struct aggr_cpu_id id = aggr_cpu_id__empty();
325 
326 	id.node = cpu__get_node(cpu);
327 	return id;
328 }
329 
330 /* setup simple routines to easily access node numbers given a cpu number */
331 static int get_max_num(char *path, int *max)
332 {
333 	size_t num;
334 	char *buf;
335 	int err = 0;
336 
337 	if (filename__read_str(path, &buf, &num))
338 		return -1;
339 
340 	buf[num] = '\0';
341 
342 	/* start on the right, to find highest node num */
343 	while (--num) {
344 		if ((buf[num] == ',') || (buf[num] == '-')) {
345 			num++;
346 			break;
347 		}
348 	}
349 	if (sscanf(&buf[num], "%d", max) < 1) {
350 		err = -1;
351 		goto out;
352 	}
353 
354 	/* convert from 0-based to 1-based */
355 	(*max)++;
356 
357 out:
358 	free(buf);
359 	return err;
360 }
361 
362 /* Determine highest possible cpu in the system for sparse allocation */
363 static void set_max_cpu_num(void)
364 {
365 	const char *mnt;
366 	char path[PATH_MAX];
367 	int ret = -1;
368 
369 	/* set up default */
370 	max_cpu_num.cpu = 4096;
371 	max_present_cpu_num.cpu = 4096;
372 
373 	mnt = sysfs__mountpoint();
374 	if (!mnt)
375 		goto out;
376 
377 	/* get the highest possible cpu number for a sparse allocation */
378 	ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/possible", mnt);
379 	if (ret >= PATH_MAX) {
380 		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);
381 		goto out;
382 	}
383 
384 	ret = get_max_num(path, &max_cpu_num.cpu);
385 	if (ret)
386 		goto out;
387 
388 	/* get the highest present cpu number for a sparse allocation */
389 	ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/present", mnt);
390 	if (ret >= PATH_MAX) {
391 		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);
392 		goto out;
393 	}
394 
395 	ret = get_max_num(path, &max_present_cpu_num.cpu);
396 
397 out:
398 	if (ret)
399 		pr_err("Failed to read max cpus, using default of %d\n", max_cpu_num.cpu);
400 }
401 
402 /* Determine highest possible node in the system for sparse allocation */
403 static void set_max_node_num(void)
404 {
405 	const char *mnt;
406 	char path[PATH_MAX];
407 	int ret = -1;
408 
409 	/* set up default */
410 	max_node_num = 8;
411 
412 	mnt = sysfs__mountpoint();
413 	if (!mnt)
414 		goto out;
415 
416 	/* get the highest possible cpu number for a sparse allocation */
417 	ret = snprintf(path, PATH_MAX, "%s/devices/system/node/possible", mnt);
418 	if (ret >= PATH_MAX) {
419 		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);
420 		goto out;
421 	}
422 
423 	ret = get_max_num(path, &max_node_num);
424 
425 out:
426 	if (ret)
427 		pr_err("Failed to read max nodes, using default of %d\n", max_node_num);
428 }
429 
430 int cpu__max_node(void)
431 {
432 	if (unlikely(!max_node_num))
433 		set_max_node_num();
434 
435 	return max_node_num;
436 }
437 
438 struct perf_cpu cpu__max_cpu(void)
439 {
440 	if (unlikely(!max_cpu_num.cpu))
441 		set_max_cpu_num();
442 
443 	return max_cpu_num;
444 }
445 
446 struct perf_cpu cpu__max_present_cpu(void)
447 {
448 	if (unlikely(!max_present_cpu_num.cpu))
449 		set_max_cpu_num();
450 
451 	return max_present_cpu_num;
452 }
453 
454 
455 int cpu__get_node(struct perf_cpu cpu)
456 {
457 	if (unlikely(cpunode_map == NULL)) {
458 		pr_debug("cpu_map not initialized\n");
459 		return -1;
460 	}
461 
462 	return cpunode_map[cpu.cpu];
463 }
464 
465 static int init_cpunode_map(void)
466 {
467 	int i;
468 
469 	set_max_cpu_num();
470 	set_max_node_num();
471 
472 	cpunode_map = calloc(max_cpu_num.cpu, sizeof(int));
473 	if (!cpunode_map) {
474 		pr_err("%s: calloc failed\n", __func__);
475 		return -1;
476 	}
477 
478 	for (i = 0; i < max_cpu_num.cpu; i++)
479 		cpunode_map[i] = -1;
480 
481 	return 0;
482 }
483 
484 int cpu__setup_cpunode_map(void)
485 {
486 	struct dirent *dent1, *dent2;
487 	DIR *dir1, *dir2;
488 	unsigned int cpu, mem;
489 	char buf[PATH_MAX];
490 	char path[PATH_MAX];
491 	const char *mnt;
492 	int n;
493 
494 	/* initialize globals */
495 	if (init_cpunode_map())
496 		return -1;
497 
498 	mnt = sysfs__mountpoint();
499 	if (!mnt)
500 		return 0;
501 
502 	n = snprintf(path, PATH_MAX, "%s/devices/system/node", mnt);
503 	if (n >= PATH_MAX) {
504 		pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);
505 		return -1;
506 	}
507 
508 	dir1 = opendir(path);
509 	if (!dir1)
510 		return 0;
511 
512 	/* walk tree and setup map */
513 	while ((dent1 = readdir(dir1)) != NULL) {
514 		if (dent1->d_type != DT_DIR || sscanf(dent1->d_name, "node%u", &mem) < 1)
515 			continue;
516 
517 		n = snprintf(buf, PATH_MAX, "%s/%s", path, dent1->d_name);
518 		if (n >= PATH_MAX) {
519 			pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX);
520 			continue;
521 		}
522 
523 		dir2 = opendir(buf);
524 		if (!dir2)
525 			continue;
526 		while ((dent2 = readdir(dir2)) != NULL) {
527 			if (dent2->d_type != DT_LNK || sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
528 				continue;
529 			cpunode_map[cpu] = mem;
530 		}
531 		closedir(dir2);
532 	}
533 	closedir(dir1);
534 	return 0;
535 }
536 
537 size_t cpu_map__snprint(struct perf_cpu_map *map, char *buf, size_t size)
538 {
539 	int i, start = -1;
540 	bool first = true;
541 	size_t ret = 0;
542 
543 #define COMMA first ? "" : ","
544 
545 	for (i = 0; i < map->nr + 1; i++) {
546 		struct perf_cpu cpu = { .cpu = INT_MAX };
547 		bool last = i == map->nr;
548 
549 		if (!last)
550 			cpu = map->map[i];
551 
552 		if (start == -1) {
553 			start = i;
554 			if (last) {
555 				ret += snprintf(buf + ret, size - ret,
556 						"%s%d", COMMA,
557 						map->map[i].cpu);
558 			}
559 		} else if (((i - start) != (cpu.cpu - map->map[start].cpu)) || last) {
560 			int end = i - 1;
561 
562 			if (start == end) {
563 				ret += snprintf(buf + ret, size - ret,
564 						"%s%d", COMMA,
565 						map->map[start].cpu);
566 			} else {
567 				ret += snprintf(buf + ret, size - ret,
568 						"%s%d-%d", COMMA,
569 						map->map[start].cpu, map->map[end].cpu);
570 			}
571 			first = false;
572 			start = i;
573 		}
574 	}
575 
576 #undef COMMA
577 
578 	pr_debug2("cpumask list: %s\n", buf);
579 	return ret;
580 }
581 
582 static char hex_char(unsigned char val)
583 {
584 	if (val < 10)
585 		return val + '0';
586 	if (val < 16)
587 		return val - 10 + 'a';
588 	return '?';
589 }
590 
591 size_t cpu_map__snprint_mask(struct perf_cpu_map *map, char *buf, size_t size)
592 {
593 	int i, cpu;
594 	char *ptr = buf;
595 	unsigned char *bitmap;
596 	struct perf_cpu last_cpu = perf_cpu_map__cpu(map, map->nr - 1);
597 
598 	if (buf == NULL)
599 		return 0;
600 
601 	bitmap = zalloc(last_cpu.cpu / 8 + 1);
602 	if (bitmap == NULL) {
603 		buf[0] = '\0';
604 		return 0;
605 	}
606 
607 	for (i = 0; i < map->nr; i++) {
608 		cpu = perf_cpu_map__cpu(map, i).cpu;
609 		bitmap[cpu / 8] |= 1 << (cpu % 8);
610 	}
611 
612 	for (cpu = last_cpu.cpu / 4 * 4; cpu >= 0; cpu -= 4) {
613 		unsigned char bits = bitmap[cpu / 8];
614 
615 		if (cpu % 8)
616 			bits >>= 4;
617 		else
618 			bits &= 0xf;
619 
620 		*ptr++ = hex_char(bits);
621 		if ((cpu % 32) == 0 && cpu > 0)
622 			*ptr++ = ',';
623 	}
624 	*ptr = '\0';
625 	free(bitmap);
626 
627 	buf[size - 1] = '\0';
628 	return ptr - buf;
629 }
630 
631 const struct perf_cpu_map *cpu_map__online(void) /* thread unsafe */
632 {
633 	static const struct perf_cpu_map *online = NULL;
634 
635 	if (!online)
636 		online = perf_cpu_map__new(NULL); /* from /sys/devices/system/cpu/online */
637 
638 	return online;
639 }
640 
641 bool aggr_cpu_id__equal(const struct aggr_cpu_id *a, const struct aggr_cpu_id *b)
642 {
643 	return a->thread == b->thread &&
644 		a->node == b->node &&
645 		a->socket == b->socket &&
646 		a->die == b->die &&
647 		a->core == b->core &&
648 		a->cpu.cpu == b->cpu.cpu;
649 }
650 
651 bool aggr_cpu_id__is_empty(const struct aggr_cpu_id *a)
652 {
653 	return a->thread == -1 &&
654 		a->node == -1 &&
655 		a->socket == -1 &&
656 		a->die == -1 &&
657 		a->core == -1 &&
658 		a->cpu.cpu == -1;
659 }
660 
661 struct aggr_cpu_id aggr_cpu_id__empty(void)
662 {
663 	struct aggr_cpu_id ret = {
664 		.thread = -1,
665 		.node = -1,
666 		.socket = -1,
667 		.die = -1,
668 		.core = -1,
669 		.cpu = (struct perf_cpu){ .cpu = -1 },
670 	};
671 	return ret;
672 }
673