xref: /openbmc/linux/tools/lib/perf/cpumap.c (revision c6acb1e7)
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 
13 struct 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 
26 static 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 
35 struct 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 
42 void 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 
48 static 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 
71 struct perf_cpu_map *perf_cpu_map__default_new(void)
72 {
73 	return cpu_map__default_new();
74 }
75 
76 static int cmp_int(const void *a, const void *b)
77 {
78 	return *(const int *)a - *(const int*)b;
79 }
80 
81 static struct perf_cpu_map *cpu_map__trim_new(int nr_cpus, int *tmp_cpus)
82 {
83 	size_t payload_size = nr_cpus * sizeof(int);
84 	struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + payload_size);
85 	int i, j;
86 
87 	if (cpus != NULL) {
88 		memcpy(cpus->map, tmp_cpus, payload_size);
89 		qsort(cpus->map, nr_cpus, sizeof(int), cmp_int);
90 		/* Remove dups */
91 		j = 0;
92 		for (i = 0; i < nr_cpus; i++) {
93 			if (i == 0 || cpus->map[i] != cpus->map[i - 1])
94 				cpus->map[j++] = cpus->map[i];
95 		}
96 		cpus->nr = j;
97 		assert(j <= nr_cpus);
98 		refcount_set(&cpus->refcnt, 1);
99 	}
100 
101 	return cpus;
102 }
103 
104 struct perf_cpu_map *perf_cpu_map__read(FILE *file)
105 {
106 	struct perf_cpu_map *cpus = NULL;
107 	int nr_cpus = 0;
108 	int *tmp_cpus = NULL, *tmp;
109 	int max_entries = 0;
110 	int n, cpu, prev;
111 	char sep;
112 
113 	sep = 0;
114 	prev = -1;
115 	for (;;) {
116 		n = fscanf(file, "%u%c", &cpu, &sep);
117 		if (n <= 0)
118 			break;
119 		if (prev >= 0) {
120 			int new_max = nr_cpus + cpu - prev - 1;
121 
122 			WARN_ONCE(new_max >= MAX_NR_CPUS, "Perf can support %d CPUs. "
123 							  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
124 
125 			if (new_max >= max_entries) {
126 				max_entries = new_max + MAX_NR_CPUS / 2;
127 				tmp = realloc(tmp_cpus, max_entries * sizeof(int));
128 				if (tmp == NULL)
129 					goto out_free_tmp;
130 				tmp_cpus = tmp;
131 			}
132 
133 			while (++prev < cpu)
134 				tmp_cpus[nr_cpus++] = prev;
135 		}
136 		if (nr_cpus == max_entries) {
137 			max_entries += MAX_NR_CPUS;
138 			tmp = realloc(tmp_cpus, max_entries * sizeof(int));
139 			if (tmp == NULL)
140 				goto out_free_tmp;
141 			tmp_cpus = tmp;
142 		}
143 
144 		tmp_cpus[nr_cpus++] = cpu;
145 		if (n == 2 && sep == '-')
146 			prev = cpu;
147 		else
148 			prev = -1;
149 		if (n == 1 || sep == '\n')
150 			break;
151 	}
152 
153 	if (nr_cpus > 0)
154 		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
155 	else
156 		cpus = cpu_map__default_new();
157 out_free_tmp:
158 	free(tmp_cpus);
159 	return cpus;
160 }
161 
162 static struct perf_cpu_map *cpu_map__read_all_cpu_map(void)
163 {
164 	struct perf_cpu_map *cpus = NULL;
165 	FILE *onlnf;
166 
167 	onlnf = fopen("/sys/devices/system/cpu/online", "r");
168 	if (!onlnf)
169 		return cpu_map__default_new();
170 
171 	cpus = perf_cpu_map__read(onlnf);
172 	fclose(onlnf);
173 	return cpus;
174 }
175 
176 struct perf_cpu_map *perf_cpu_map__new(const char *cpu_list)
177 {
178 	struct perf_cpu_map *cpus = NULL;
179 	unsigned long start_cpu, end_cpu = 0;
180 	char *p = NULL;
181 	int i, nr_cpus = 0;
182 	int *tmp_cpus = NULL, *tmp;
183 	int max_entries = 0;
184 
185 	if (!cpu_list)
186 		return cpu_map__read_all_cpu_map();
187 
188 	/*
189 	 * must handle the case of empty cpumap to cover
190 	 * TOPOLOGY header for NUMA nodes with no CPU
191 	 * ( e.g., because of CPU hotplug)
192 	 */
193 	if (!isdigit(*cpu_list) && *cpu_list != '\0')
194 		goto out;
195 
196 	while (isdigit(*cpu_list)) {
197 		p = NULL;
198 		start_cpu = strtoul(cpu_list, &p, 0);
199 		if (start_cpu >= INT_MAX
200 		    || (*p != '\0' && *p != ',' && *p != '-'))
201 			goto invalid;
202 
203 		if (*p == '-') {
204 			cpu_list = ++p;
205 			p = NULL;
206 			end_cpu = strtoul(cpu_list, &p, 0);
207 
208 			if (end_cpu >= INT_MAX || (*p != '\0' && *p != ','))
209 				goto invalid;
210 
211 			if (end_cpu < start_cpu)
212 				goto invalid;
213 		} else {
214 			end_cpu = start_cpu;
215 		}
216 
217 		WARN_ONCE(end_cpu >= MAX_NR_CPUS, "Perf can support %d CPUs. "
218 						  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
219 
220 		for (; start_cpu <= end_cpu; start_cpu++) {
221 			/* check for duplicates */
222 			for (i = 0; i < nr_cpus; i++)
223 				if (tmp_cpus[i] == (int)start_cpu)
224 					goto invalid;
225 
226 			if (nr_cpus == max_entries) {
227 				max_entries += MAX_NR_CPUS;
228 				tmp = realloc(tmp_cpus, max_entries * sizeof(int));
229 				if (tmp == NULL)
230 					goto invalid;
231 				tmp_cpus = tmp;
232 			}
233 			tmp_cpus[nr_cpus++] = (int)start_cpu;
234 		}
235 		if (*p)
236 			++p;
237 
238 		cpu_list = p;
239 	}
240 
241 	if (nr_cpus > 0)
242 		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
243 	else if (*cpu_list != '\0')
244 		cpus = cpu_map__default_new();
245 	else
246 		cpus = perf_cpu_map__dummy_new();
247 invalid:
248 	free(tmp_cpus);
249 out:
250 	return cpus;
251 }
252 
253 int perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx)
254 {
255 	if (cpus && idx < cpus->nr)
256 		return cpus->map[idx];
257 
258 	return -1;
259 }
260 
261 int perf_cpu_map__nr(const struct perf_cpu_map *cpus)
262 {
263 	return cpus ? cpus->nr : 1;
264 }
265 
266 bool perf_cpu_map__empty(const struct perf_cpu_map *map)
267 {
268 	return map ? map->map[0] == -1 : true;
269 }
270 
271 int perf_cpu_map__idx(struct perf_cpu_map *cpus, int cpu)
272 {
273 	int low = 0, high = cpus->nr;
274 
275 	while (low < high) {
276 		int idx = (low + high) / 2,
277 		    cpu_at_idx = cpus->map[idx];
278 
279 		if (cpu_at_idx == cpu)
280 			return idx;
281 
282 		if (cpu_at_idx > cpu)
283 			high = idx;
284 		else
285 			low = idx + 1;
286 	}
287 
288 	return -1;
289 }
290 
291 int perf_cpu_map__max(struct perf_cpu_map *map)
292 {
293 	// cpu_map__trim_new() qsort()s it, cpu_map__default_new() sorts it as well.
294 	return map->nr > 0 ? map->map[map->nr - 1] : -1;
295 }
296 
297 /*
298  * Merge two cpumaps
299  *
300  * orig either gets freed and replaced with a new map, or reused
301  * with no reference count change (similar to "realloc")
302  * other has its reference count increased.
303  */
304 
305 struct perf_cpu_map *perf_cpu_map__merge(struct perf_cpu_map *orig,
306 					 struct perf_cpu_map *other)
307 {
308 	int *tmp_cpus;
309 	int tmp_len;
310 	int i, j, k;
311 	struct perf_cpu_map *merged;
312 
313 	if (!orig && !other)
314 		return NULL;
315 	if (!orig) {
316 		perf_cpu_map__get(other);
317 		return other;
318 	}
319 	if (!other)
320 		return orig;
321 	if (orig->nr == other->nr &&
322 	    !memcmp(orig->map, other->map, orig->nr * sizeof(int)))
323 		return orig;
324 
325 	tmp_len = orig->nr + other->nr;
326 	tmp_cpus = malloc(tmp_len * sizeof(int));
327 	if (!tmp_cpus)
328 		return NULL;
329 
330 	/* Standard merge algorithm from wikipedia */
331 	i = j = k = 0;
332 	while (i < orig->nr && j < other->nr) {
333 		if (orig->map[i] <= other->map[j]) {
334 			if (orig->map[i] == other->map[j])
335 				j++;
336 			tmp_cpus[k++] = orig->map[i++];
337 		} else
338 			tmp_cpus[k++] = other->map[j++];
339 	}
340 
341 	while (i < orig->nr)
342 		tmp_cpus[k++] = orig->map[i++];
343 
344 	while (j < other->nr)
345 		tmp_cpus[k++] = other->map[j++];
346 	assert(k <= tmp_len);
347 
348 	merged = cpu_map__trim_new(k, tmp_cpus);
349 	free(tmp_cpus);
350 	perf_cpu_map__put(orig);
351 	return merged;
352 }
353