xref: /kernel/linux/linux-5.10/tools/lib/perf/cpumap.c (revision 8c2ecf20) 
1// SPDX-License-Identifier: GPL-2.0-only
2#include <perf/cpumap.h>
3#include <stdlib.h>
4#include <linux/refcount.h>
5#include <internal/cpumap.h>
6#include <asm/bug.h>
7#include <stdio.h>
8#include <string.h>
9#include <unistd.h>
10#include <ctype.h>
11#include <limits.h>
12
13struct perf_cpu_map *perf_cpu_map__dummy_new(void)
14{
15	struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + sizeof(int));
16
17	if (cpus != NULL) {
18		cpus->nr = 1;
19		cpus->map[0] = -1;
20		refcount_set(&cpus->refcnt, 1);
21	}
22
23	return cpus;
24}
25
26static void cpu_map__delete(struct perf_cpu_map *map)
27{
28	if (map) {
29		WARN_ONCE(refcount_read(&map->refcnt) != 0,
30			  "cpu_map refcnt unbalanced\n");
31		free(map);
32	}
33}
34
35struct perf_cpu_map *perf_cpu_map__get(struct perf_cpu_map *map)
36{
37	if (map)
38		refcount_inc(&map->refcnt);
39	return map;
40}
41
42void perf_cpu_map__put(struct perf_cpu_map *map)
43{
44	if (map && refcount_dec_and_test(&map->refcnt))
45		cpu_map__delete(map);
46}
47
48static struct perf_cpu_map *cpu_map__default_new(void)
49{
50	struct perf_cpu_map *cpus;
51	int nr_cpus;
52
53	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
54	if (nr_cpus < 0)
55		return NULL;
56
57	cpus = malloc(sizeof(*cpus) + nr_cpus * sizeof(int));
58	if (cpus != NULL) {
59		int i;
60
61		for (i = 0; i < nr_cpus; ++i)
62			cpus->map[i] = i;
63
64		cpus->nr = nr_cpus;
65		refcount_set(&cpus->refcnt, 1);
66	}
67
68	return cpus;
69}
70
71static int cmp_int(const void *a, const void *b)
72{
73	return *(const int *)a - *(const int*)b;
74}
75
76static struct perf_cpu_map *cpu_map__trim_new(int nr_cpus, int *tmp_cpus)
77{
78	size_t payload_size = nr_cpus * sizeof(int);
79	struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + payload_size);
80	int i, j;
81
82	if (cpus != NULL) {
83		memcpy(cpus->map, tmp_cpus, payload_size);
84		qsort(cpus->map, nr_cpus, sizeof(int), cmp_int);
85		/* Remove dups */
86		j = 0;
87		for (i = 0; i < nr_cpus; i++) {
88			if (i == 0 || cpus->map[i] != cpus->map[i - 1])
89				cpus->map[j++] = cpus->map[i];
90		}
91		cpus->nr = j;
92		assert(j <= nr_cpus);
93		refcount_set(&cpus->refcnt, 1);
94	}
95
96	return cpus;
97}
98
99struct perf_cpu_map *perf_cpu_map__read(FILE *file)
100{
101	struct perf_cpu_map *cpus = NULL;
102	int nr_cpus = 0;
103	int *tmp_cpus = NULL, *tmp;
104	int max_entries = 0;
105	int n, cpu, prev;
106	char sep;
107
108	sep = 0;
109	prev = -1;
110	for (;;) {
111		n = fscanf(file, "%u%c", &cpu, &sep);
112		if (n <= 0)
113			break;
114		if (prev >= 0) {
115			int new_max = nr_cpus + cpu - prev - 1;
116
117			WARN_ONCE(new_max >= MAX_NR_CPUS, "Perf can support %d CPUs. "
118							  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
119
120			if (new_max >= max_entries) {
121				max_entries = new_max + MAX_NR_CPUS / 2;
122				tmp = realloc(tmp_cpus, max_entries * sizeof(int));
123				if (tmp == NULL)
124					goto out_free_tmp;
125				tmp_cpus = tmp;
126			}
127
128			while (++prev < cpu)
129				tmp_cpus[nr_cpus++] = prev;
130		}
131		if (nr_cpus == max_entries) {
132			max_entries += MAX_NR_CPUS;
133			tmp = realloc(tmp_cpus, max_entries * sizeof(int));
134			if (tmp == NULL)
135				goto out_free_tmp;
136			tmp_cpus = tmp;
137		}
138
139		tmp_cpus[nr_cpus++] = cpu;
140		if (n == 2 && sep == '-')
141			prev = cpu;
142		else
143			prev = -1;
144		if (n == 1 || sep == '\n')
145			break;
146	}
147
148	if (nr_cpus > 0)
149		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
150	else
151		cpus = cpu_map__default_new();
152out_free_tmp:
153	free(tmp_cpus);
154	return cpus;
155}
156
157static struct perf_cpu_map *cpu_map__read_all_cpu_map(void)
158{
159	struct perf_cpu_map *cpus = NULL;
160	FILE *onlnf;
161
162	onlnf = fopen("/sys/devices/system/cpu/online", "r");
163	if (!onlnf)
164		return cpu_map__default_new();
165
166	cpus = perf_cpu_map__read(onlnf);
167	fclose(onlnf);
168	return cpus;
169}
170
171struct perf_cpu_map *perf_cpu_map__new(const char *cpu_list)
172{
173	struct perf_cpu_map *cpus = NULL;
174	unsigned long start_cpu, end_cpu = 0;
175	char *p = NULL;
176	int i, nr_cpus = 0;
177	int *tmp_cpus = NULL, *tmp;
178	int max_entries = 0;
179
180	if (!cpu_list)
181		return cpu_map__read_all_cpu_map();
182
183	/*
184	 * must handle the case of empty cpumap to cover
185	 * TOPOLOGY header for NUMA nodes with no CPU
186	 * ( e.g., because of CPU hotplug)
187	 */
188	if (!isdigit(*cpu_list) && *cpu_list != '\0')
189		goto out;
190
191	while (isdigit(*cpu_list)) {
192		p = NULL;
193		start_cpu = strtoul(cpu_list, &p, 0);
194		if (start_cpu >= INT_MAX
195		    || (*p != '\0' && *p != ',' && *p != '-'))
196			goto invalid;
197
198		if (*p == '-') {
199			cpu_list = ++p;
200			p = NULL;
201			end_cpu = strtoul(cpu_list, &p, 0);
202
203			if (end_cpu >= INT_MAX || (*p != '\0' && *p != ','))
204				goto invalid;
205
206			if (end_cpu < start_cpu)
207				goto invalid;
208		} else {
209			end_cpu = start_cpu;
210		}
211
212		WARN_ONCE(end_cpu >= MAX_NR_CPUS, "Perf can support %d CPUs. "
213						  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
214
215		for (; start_cpu <= end_cpu; start_cpu++) {
216			/* check for duplicates */
217			for (i = 0; i < nr_cpus; i++)
218				if (tmp_cpus[i] == (int)start_cpu)
219					goto invalid;
220
221			if (nr_cpus == max_entries) {
222				max_entries += MAX_NR_CPUS;
223				tmp = realloc(tmp_cpus, max_entries * sizeof(int));
224				if (tmp == NULL)
225					goto invalid;
226				tmp_cpus = tmp;
227			}
228			tmp_cpus[nr_cpus++] = (int)start_cpu;
229		}
230		if (*p)
231			++p;
232
233		cpu_list = p;
234	}
235
236	if (nr_cpus > 0)
237		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
238	else if (*cpu_list != '\0')
239		cpus = cpu_map__default_new();
240	else
241		cpus = perf_cpu_map__dummy_new();
242invalid:
243	free(tmp_cpus);
244out:
245	return cpus;
246}
247
248int perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx)
249{
250	if (cpus && idx < cpus->nr)
251		return cpus->map[idx];
252
253	return -1;
254}
255
256int perf_cpu_map__nr(const struct perf_cpu_map *cpus)
257{
258	return cpus ? cpus->nr : 1;
259}
260
261bool perf_cpu_map__empty(const struct perf_cpu_map *map)
262{
263	return map ? map->map[0] == -1 : true;
264}
265
266int perf_cpu_map__idx(struct perf_cpu_map *cpus, int cpu)
267{
268	int i;
269
270	for (i = 0; i < cpus->nr; ++i) {
271		if (cpus->map[i] == cpu)
272			return i;
273	}
274
275	return -1;
276}
277
278int perf_cpu_map__max(struct perf_cpu_map *map)
279{
280	int i, max = -1;
281
282	for (i = 0; i < map->nr; i++) {
283		if (map->map[i] > max)
284			max = map->map[i];
285	}
286
287	return max;
288}
289
290/*
291 * Merge two cpumaps
292 *
293 * orig either gets freed and replaced with a new map, or reused
294 * with no reference count change (similar to "realloc")
295 * other has its reference count increased.
296 */
297
298struct perf_cpu_map *perf_cpu_map__merge(struct perf_cpu_map *orig,
299					 struct perf_cpu_map *other)
300{
301	int *tmp_cpus;
302	int tmp_len;
303	int i, j, k;
304	struct perf_cpu_map *merged;
305
306	if (!orig && !other)
307		return NULL;
308	if (!orig) {
309		perf_cpu_map__get(other);
310		return other;
311	}
312	if (!other)
313		return orig;
314	if (orig->nr == other->nr &&
315	    !memcmp(orig->map, other->map, orig->nr * sizeof(int)))
316		return orig;
317
318	tmp_len = orig->nr + other->nr;
319	tmp_cpus = malloc(tmp_len * sizeof(int));
320	if (!tmp_cpus)
321		return NULL;
322
323	/* Standard merge algorithm from wikipedia */
324	i = j = k = 0;
325	while (i < orig->nr && j < other->nr) {
326		if (orig->map[i] <= other->map[j]) {
327			if (orig->map[i] == other->map[j])
328				j++;
329			tmp_cpus[k++] = orig->map[i++];
330		} else
331			tmp_cpus[k++] = other->map[j++];
332	}
333
334	while (i < orig->nr)
335		tmp_cpus[k++] = orig->map[i++];
336
337	while (j < other->nr)
338		tmp_cpus[k++] = other->map[j++];
339	assert(k <= tmp_len);
340
341	merged = cpu_map__trim_new(k, tmp_cpus);
342	free(tmp_cpus);
343	perf_cpu_map__put(orig);
344	return merged;
345}
346