1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <perf/cpumap.h>
4 #include <linux/refcount.h>
5 #include <internal/cpumap.h>
14 void perf_cpu_map__set_nr(struct perf_cpu_map *map, int nr_cpus)
16 RC_CHK_ACCESS(map)->nr = nr_cpus;
19 struct perf_cpu_map *perf_cpu_map__alloc(int nr_cpus)
21 RC_STRUCT(perf_cpu_map) *cpus;
22 struct perf_cpu_map *result;
27 cpus = malloc(sizeof(*cpus) + sizeof(struct perf_cpu) * nr_cpus);
28 if (ADD_RC_CHK(result, cpus)) {
30 refcount_set(&cpus->refcnt, 1);
35 struct perf_cpu_map *perf_cpu_map__new_any_cpu(void)
37 struct perf_cpu_map *cpus = perf_cpu_map__alloc(1);
40 RC_CHK_ACCESS(cpus)->map[0].cpu = -1;
45 static void cpu_map__delete(struct perf_cpu_map *map)
48 WARN_ONCE(refcount_read(perf_cpu_map__refcnt(map)) != 0,
49 "cpu_map refcnt unbalanced\n");
54 struct perf_cpu_map *perf_cpu_map__get(struct perf_cpu_map *map)
56 struct perf_cpu_map *result;
58 if (RC_CHK_GET(result, map))
59 refcount_inc(perf_cpu_map__refcnt(map));
64 void perf_cpu_map__put(struct perf_cpu_map *map)
67 if (refcount_dec_and_test(perf_cpu_map__refcnt(map)))
74 static struct perf_cpu_map *cpu_map__new_sysconf(void)
76 struct perf_cpu_map *cpus;
77 int nr_cpus, nr_cpus_conf;
79 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
83 nr_cpus_conf = sysconf(_SC_NPROCESSORS_CONF);
84 if (nr_cpus != nr_cpus_conf) {
85 pr_warning("Number of online CPUs (%d) differs from the number configured (%d) the CPU map will only cover the first %d CPUs.",
86 nr_cpus, nr_cpus_conf, nr_cpus);
89 cpus = perf_cpu_map__alloc(nr_cpus);
93 for (i = 0; i < nr_cpus; ++i)
94 RC_CHK_ACCESS(cpus)->map[i].cpu = i;
100 static struct perf_cpu_map *cpu_map__new_sysfs_online(void)
102 struct perf_cpu_map *cpus = NULL;
105 onlnf = fopen("/sys/devices/system/cpu/online", "r");
107 cpus = perf_cpu_map__read(onlnf);
113 struct perf_cpu_map *perf_cpu_map__new_online_cpus(void)
115 struct perf_cpu_map *cpus = cpu_map__new_sysfs_online();
120 return cpu_map__new_sysconf();
124 static int cmp_cpu(const void *a, const void *b)
126 const struct perf_cpu *cpu_a = a, *cpu_b = b;
128 return cpu_a->cpu - cpu_b->cpu;
131 static struct perf_cpu __perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx)
133 return RC_CHK_ACCESS(cpus)->map[idx];
136 static struct perf_cpu_map *cpu_map__trim_new(int nr_cpus, const struct perf_cpu *tmp_cpus)
138 size_t payload_size = nr_cpus * sizeof(struct perf_cpu);
139 struct perf_cpu_map *cpus = perf_cpu_map__alloc(nr_cpus);
143 memcpy(RC_CHK_ACCESS(cpus)->map, tmp_cpus, payload_size);
144 qsort(RC_CHK_ACCESS(cpus)->map, nr_cpus, sizeof(struct perf_cpu), cmp_cpu);
147 for (i = 0; i < nr_cpus; i++) {
149 __perf_cpu_map__cpu(cpus, i).cpu !=
150 __perf_cpu_map__cpu(cpus, i - 1).cpu) {
151 RC_CHK_ACCESS(cpus)->map[j++].cpu =
152 __perf_cpu_map__cpu(cpus, i).cpu;
155 perf_cpu_map__set_nr(cpus, j);
156 assert(j <= nr_cpus);
161 struct perf_cpu_map *perf_cpu_map__read(FILE *file)
163 struct perf_cpu_map *cpus = NULL;
165 struct perf_cpu *tmp_cpus = NULL, *tmp;
173 n = fscanf(file, "%u%c", &cpu, &sep);
177 int new_max = nr_cpus + cpu - prev - 1;
179 WARN_ONCE(new_max >= MAX_NR_CPUS, "Perf can support %d CPUs. "
180 "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
182 if (new_max >= max_entries) {
183 max_entries = new_max + MAX_NR_CPUS / 2;
184 tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
191 tmp_cpus[nr_cpus++].cpu = prev;
193 if (nr_cpus == max_entries) {
194 max_entries += MAX_NR_CPUS;
195 tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
201 tmp_cpus[nr_cpus++].cpu = cpu;
202 if (n == 2 && sep == '-')
206 if (n == 1 || sep == '\n')
211 cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
217 struct perf_cpu_map *perf_cpu_map__new(const char *cpu_list)
219 struct perf_cpu_map *cpus = NULL;
220 unsigned long start_cpu, end_cpu = 0;
223 struct perf_cpu *tmp_cpus = NULL, *tmp;
227 return perf_cpu_map__new_online_cpus();
230 * must handle the case of empty cpumap to cover
231 * TOPOLOGY header for NUMA nodes with no CPU
232 * ( e.g., because of CPU hotplug)
234 if (!isdigit(*cpu_list) && *cpu_list != '\0')
237 while (isdigit(*cpu_list)) {
239 start_cpu = strtoul(cpu_list, &p, 0);
240 if (start_cpu >= INT_MAX
241 || (*p != '\0' && *p != ',' && *p != '-'))
247 end_cpu = strtoul(cpu_list, &p, 0);
249 if (end_cpu >= INT_MAX || (*p != '\0' && *p != ','))
252 if (end_cpu < start_cpu)
258 WARN_ONCE(end_cpu >= MAX_NR_CPUS, "Perf can support %d CPUs. "
259 "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
261 for (; start_cpu <= end_cpu; start_cpu++) {
262 /* check for duplicates */
263 for (i = 0; i < nr_cpus; i++)
264 if (tmp_cpus[i].cpu == (int)start_cpu)
267 if (nr_cpus == max_entries) {
268 max_entries += MAX_NR_CPUS;
269 tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
274 tmp_cpus[nr_cpus++].cpu = (int)start_cpu;
283 cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
284 else if (*cpu_list != '\0') {
285 pr_warning("Unexpected characters at end of cpu list ('%s'), using online CPUs.",
287 cpus = perf_cpu_map__new_online_cpus();
289 cpus = perf_cpu_map__new_any_cpu();
296 static int __perf_cpu_map__nr(const struct perf_cpu_map *cpus)
298 return RC_CHK_ACCESS(cpus)->nr;
301 struct perf_cpu perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx)
303 struct perf_cpu result = {
307 if (cpus && idx < __perf_cpu_map__nr(cpus))
308 return __perf_cpu_map__cpu(cpus, idx);
313 int perf_cpu_map__nr(const struct perf_cpu_map *cpus)
315 return cpus ? __perf_cpu_map__nr(cpus) : 1;
318 bool perf_cpu_map__has_any_cpu_or_is_empty(const struct perf_cpu_map *map)
320 return map ? __perf_cpu_map__cpu(map, 0).cpu == -1 : true;
323 bool perf_cpu_map__is_any_cpu_or_is_empty(const struct perf_cpu_map *map)
328 return __perf_cpu_map__nr(map) == 1 && __perf_cpu_map__cpu(map, 0).cpu == -1;
331 bool perf_cpu_map__is_empty(const struct perf_cpu_map *map)
336 int perf_cpu_map__idx(const struct perf_cpu_map *cpus, struct perf_cpu cpu)
344 high = __perf_cpu_map__nr(cpus);
346 int idx = (low + high) / 2;
347 struct perf_cpu cpu_at_idx = __perf_cpu_map__cpu(cpus, idx);
349 if (cpu_at_idx.cpu == cpu.cpu)
352 if (cpu_at_idx.cpu > cpu.cpu)
361 bool perf_cpu_map__has(const struct perf_cpu_map *cpus, struct perf_cpu cpu)
363 return perf_cpu_map__idx(cpus, cpu) != -1;
366 bool perf_cpu_map__equal(const struct perf_cpu_map *lhs, const struct perf_cpu_map *rhs)
376 nr = __perf_cpu_map__nr(lhs);
377 if (nr != __perf_cpu_map__nr(rhs))
380 for (int idx = 0; idx < nr; idx++) {
381 if (__perf_cpu_map__cpu(lhs, idx).cpu != __perf_cpu_map__cpu(rhs, idx).cpu)
387 bool perf_cpu_map__has_any_cpu(const struct perf_cpu_map *map)
389 return map && __perf_cpu_map__cpu(map, 0).cpu == -1;
392 struct perf_cpu perf_cpu_map__min(const struct perf_cpu_map *map)
394 struct perf_cpu cpu, result = {
399 perf_cpu_map__for_each_cpu_skip_any(cpu, idx, map) {
406 struct perf_cpu perf_cpu_map__max(const struct perf_cpu_map *map)
408 struct perf_cpu result = {
412 // cpu_map__trim_new() qsort()s it, cpu_map__default_new() sorts it as well.
413 return __perf_cpu_map__nr(map) > 0
414 ? __perf_cpu_map__cpu(map, __perf_cpu_map__nr(map) - 1)
418 /** Is 'b' a subset of 'a'. */
419 bool perf_cpu_map__is_subset(const struct perf_cpu_map *a, const struct perf_cpu_map *b)
423 if (!a || __perf_cpu_map__nr(b) > __perf_cpu_map__nr(a))
426 for (int i = 0, j = 0; i < __perf_cpu_map__nr(a); i++) {
427 if (__perf_cpu_map__cpu(a, i).cpu > __perf_cpu_map__cpu(b, j).cpu)
429 if (__perf_cpu_map__cpu(a, i).cpu == __perf_cpu_map__cpu(b, j).cpu) {
431 if (j == __perf_cpu_map__nr(b))
441 * orig either gets freed and replaced with a new map, or reused
442 * with no reference count change (similar to "realloc")
443 * other has its reference count increased.
446 struct perf_cpu_map *perf_cpu_map__merge(struct perf_cpu_map *orig,
447 struct perf_cpu_map *other)
449 struct perf_cpu *tmp_cpus;
452 struct perf_cpu_map *merged;
454 if (perf_cpu_map__is_subset(orig, other))
456 if (perf_cpu_map__is_subset(other, orig)) {
457 perf_cpu_map__put(orig);
458 return perf_cpu_map__get(other);
461 tmp_len = __perf_cpu_map__nr(orig) + __perf_cpu_map__nr(other);
462 tmp_cpus = malloc(tmp_len * sizeof(struct perf_cpu));
466 /* Standard merge algorithm from wikipedia */
468 while (i < __perf_cpu_map__nr(orig) && j < __perf_cpu_map__nr(other)) {
469 if (__perf_cpu_map__cpu(orig, i).cpu <= __perf_cpu_map__cpu(other, j).cpu) {
470 if (__perf_cpu_map__cpu(orig, i).cpu == __perf_cpu_map__cpu(other, j).cpu)
472 tmp_cpus[k++] = __perf_cpu_map__cpu(orig, i++);
474 tmp_cpus[k++] = __perf_cpu_map__cpu(other, j++);
477 while (i < __perf_cpu_map__nr(orig))
478 tmp_cpus[k++] = __perf_cpu_map__cpu(orig, i++);
480 while (j < __perf_cpu_map__nr(other))
481 tmp_cpus[k++] = __perf_cpu_map__cpu(other, j++);
482 assert(k <= tmp_len);
484 merged = cpu_map__trim_new(k, tmp_cpus);
486 perf_cpu_map__put(orig);
490 struct perf_cpu_map *perf_cpu_map__intersect(struct perf_cpu_map *orig,
491 struct perf_cpu_map *other)
493 struct perf_cpu *tmp_cpus;
496 struct perf_cpu_map *merged = NULL;
498 if (perf_cpu_map__is_subset(other, orig))
499 return perf_cpu_map__get(orig);
500 if (perf_cpu_map__is_subset(orig, other))
501 return perf_cpu_map__get(other);
503 tmp_len = max(__perf_cpu_map__nr(orig), __perf_cpu_map__nr(other));
504 tmp_cpus = malloc(tmp_len * sizeof(struct perf_cpu));
509 while (i < __perf_cpu_map__nr(orig) && j < __perf_cpu_map__nr(other)) {
510 if (__perf_cpu_map__cpu(orig, i).cpu < __perf_cpu_map__cpu(other, j).cpu)
512 else if (__perf_cpu_map__cpu(orig, i).cpu > __perf_cpu_map__cpu(other, j).cpu)
516 tmp_cpus[k++] = __perf_cpu_map__cpu(orig, i++);
520 merged = cpu_map__trim_new(k, tmp_cpus);