xref: /linux/tools/lib/perf/cpumap.c (revision 4359a011e259a4608afc7fb3635370c9d4ba5943) 
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 static struct perf_cpu_map *perf_cpu_map__alloc(int nr_cpus)
14 {
15 	struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + sizeof(struct perf_cpu) * nr_cpus);
16 
17 	if (cpus != NULL) {
18 		cpus->nr = nr_cpus;
19 		refcount_set(&cpus->refcnt, 1);
20 
21 	}
22 	return cpus;
23 }
24 
25 struct perf_cpu_map *perf_cpu_map__dummy_new(void)
26 {
27 	struct perf_cpu_map *cpus = perf_cpu_map__alloc(1);
28 
29 	if (cpus)
30 		cpus->map[0].cpu = -1;
31 
32 	return cpus;
33 }
34 
35 static void cpu_map__delete(struct perf_cpu_map *map)
36 {
37 	if (map) {
38 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
39 			  "cpu_map refcnt unbalanced\n");
40 		free(map);
41 	}
42 }
43 
44 struct perf_cpu_map *perf_cpu_map__get(struct perf_cpu_map *map)
45 {
46 	if (map)
47 		refcount_inc(&map->refcnt);
48 	return map;
49 }
50 
51 void perf_cpu_map__put(struct perf_cpu_map *map)
52 {
53 	if (map && refcount_dec_and_test(&map->refcnt))
54 		cpu_map__delete(map);
55 }
56 
57 static struct perf_cpu_map *cpu_map__default_new(void)
58 {
59 	struct perf_cpu_map *cpus;
60 	int nr_cpus;
61 
62 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
63 	if (nr_cpus < 0)
64 		return NULL;
65 
66 	cpus = perf_cpu_map__alloc(nr_cpus);
67 	if (cpus != NULL) {
68 		int i;
69 
70 		for (i = 0; i < nr_cpus; ++i)
71 			cpus->map[i].cpu = i;
72 	}
73 
74 	return cpus;
75 }
76 
77 struct perf_cpu_map *perf_cpu_map__default_new(void)
78 {
79 	return cpu_map__default_new();
80 }
81 
82 
83 static int cmp_cpu(const void *a, const void *b)
84 {
85 	const struct perf_cpu *cpu_a = a, *cpu_b = b;
86 
87 	return cpu_a->cpu - cpu_b->cpu;
88 }
89 
90 static struct perf_cpu_map *cpu_map__trim_new(int nr_cpus, const struct perf_cpu *tmp_cpus)
91 {
92 	size_t payload_size = nr_cpus * sizeof(struct perf_cpu);
93 	struct perf_cpu_map *cpus = perf_cpu_map__alloc(nr_cpus);
94 	int i, j;
95 
96 	if (cpus != NULL) {
97 		memcpy(cpus->map, tmp_cpus, payload_size);
98 		qsort(cpus->map, nr_cpus, sizeof(struct perf_cpu), cmp_cpu);
99 		/* Remove dups */
100 		j = 0;
101 		for (i = 0; i < nr_cpus; i++) {
102 			if (i == 0 || cpus->map[i].cpu != cpus->map[i - 1].cpu)
103 				cpus->map[j++].cpu = cpus->map[i].cpu;
104 		}
105 		cpus->nr = j;
106 		assert(j <= nr_cpus);
107 	}
108 	return cpus;
109 }
110 
111 struct perf_cpu_map *perf_cpu_map__read(FILE *file)
112 {
113 	struct perf_cpu_map *cpus = NULL;
114 	int nr_cpus = 0;
115 	struct perf_cpu *tmp_cpus = NULL, *tmp;
116 	int max_entries = 0;
117 	int n, cpu, prev;
118 	char sep;
119 
120 	sep = 0;
121 	prev = -1;
122 	for (;;) {
123 		n = fscanf(file, "%u%c", &cpu, &sep);
124 		if (n <= 0)
125 			break;
126 		if (prev >= 0) {
127 			int new_max = nr_cpus + cpu - prev - 1;
128 
129 			WARN_ONCE(new_max >= MAX_NR_CPUS, "Perf can support %d CPUs. "
130 							  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
131 
132 			if (new_max >= max_entries) {
133 				max_entries = new_max + MAX_NR_CPUS / 2;
134 				tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
135 				if (tmp == NULL)
136 					goto out_free_tmp;
137 				tmp_cpus = tmp;
138 			}
139 
140 			while (++prev < cpu)
141 				tmp_cpus[nr_cpus++].cpu = prev;
142 		}
143 		if (nr_cpus == max_entries) {
144 			max_entries += MAX_NR_CPUS;
145 			tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
146 			if (tmp == NULL)
147 				goto out_free_tmp;
148 			tmp_cpus = tmp;
149 		}
150 
151 		tmp_cpus[nr_cpus++].cpu = cpu;
152 		if (n == 2 && sep == '-')
153 			prev = cpu;
154 		else
155 			prev = -1;
156 		if (n == 1 || sep == '\n')
157 			break;
158 	}
159 
160 	if (nr_cpus > 0)
161 		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
162 	else
163 		cpus = cpu_map__default_new();
164 out_free_tmp:
165 	free(tmp_cpus);
166 	return cpus;
167 }
168 
169 static struct perf_cpu_map *cpu_map__read_all_cpu_map(void)
170 {
171 	struct perf_cpu_map *cpus = NULL;
172 	FILE *onlnf;
173 
174 	onlnf = fopen("/sys/devices/system/cpu/online", "r");
175 	if (!onlnf)
176 		return cpu_map__default_new();
177 
178 	cpus = perf_cpu_map__read(onlnf);
179 	fclose(onlnf);
180 	return cpus;
181 }
182 
183 struct perf_cpu_map *perf_cpu_map__new(const char *cpu_list)
184 {
185 	struct perf_cpu_map *cpus = NULL;
186 	unsigned long start_cpu, end_cpu = 0;
187 	char *p = NULL;
188 	int i, nr_cpus = 0;
189 	struct perf_cpu *tmp_cpus = NULL, *tmp;
190 	int max_entries = 0;
191 
192 	if (!cpu_list)
193 		return cpu_map__read_all_cpu_map();
194 
195 	/*
196 	 * must handle the case of empty cpumap to cover
197 	 * TOPOLOGY header for NUMA nodes with no CPU
198 	 * ( e.g., because of CPU hotplug)
199 	 */
200 	if (!isdigit(*cpu_list) && *cpu_list != '\0')
201 		goto out;
202 
203 	while (isdigit(*cpu_list)) {
204 		p = NULL;
205 		start_cpu = strtoul(cpu_list, &p, 0);
206 		if (start_cpu >= INT_MAX
207 		    || (*p != '\0' && *p != ',' && *p != '-'))
208 			goto invalid;
209 
210 		if (*p == '-') {
211 			cpu_list = ++p;
212 			p = NULL;
213 			end_cpu = strtoul(cpu_list, &p, 0);
214 
215 			if (end_cpu >= INT_MAX || (*p != '\0' && *p != ','))
216 				goto invalid;
217 
218 			if (end_cpu < start_cpu)
219 				goto invalid;
220 		} else {
221 			end_cpu = start_cpu;
222 		}
223 
224 		WARN_ONCE(end_cpu >= MAX_NR_CPUS, "Perf can support %d CPUs. "
225 						  "Consider raising MAX_NR_CPUS\n", MAX_NR_CPUS);
226 
227 		for (; start_cpu <= end_cpu; start_cpu++) {
228 			/* check for duplicates */
229 			for (i = 0; i < nr_cpus; i++)
230 				if (tmp_cpus[i].cpu == (int)start_cpu)
231 					goto invalid;
232 
233 			if (nr_cpus == max_entries) {
234 				max_entries += MAX_NR_CPUS;
235 				tmp = realloc(tmp_cpus, max_entries * sizeof(struct perf_cpu));
236 				if (tmp == NULL)
237 					goto invalid;
238 				tmp_cpus = tmp;
239 			}
240 			tmp_cpus[nr_cpus++].cpu = (int)start_cpu;
241 		}
242 		if (*p)
243 			++p;
244 
245 		cpu_list = p;
246 	}
247 
248 	if (nr_cpus > 0)
249 		cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
250 	else if (*cpu_list != '\0')
251 		cpus = cpu_map__default_new();
252 	else
253 		cpus = perf_cpu_map__dummy_new();
254 invalid:
255 	free(tmp_cpus);
256 out:
257 	return cpus;
258 }
259 
260 struct perf_cpu perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx)
261 {
262 	struct perf_cpu result = {
263 		.cpu = -1
264 	};
265 
266 	if (cpus && idx < cpus->nr)
267 		return cpus->map[idx];
268 
269 	return result;
270 }
271 
272 int perf_cpu_map__nr(const struct perf_cpu_map *cpus)
273 {
274 	return cpus ? cpus->nr : 1;
275 }
276 
277 bool perf_cpu_map__empty(const struct perf_cpu_map *map)
278 {
279 	return map ? map->map[0].cpu == -1 : true;
280 }
281 
282 int perf_cpu_map__idx(const struct perf_cpu_map *cpus, struct perf_cpu cpu)
283 {
284 	int low, high;
285 
286 	if (!cpus)
287 		return -1;
288 
289 	low = 0;
290 	high = cpus->nr;
291 	while (low < high) {
292 		int idx = (low + high) / 2;
293 		struct perf_cpu cpu_at_idx = cpus->map[idx];
294 
295 		if (cpu_at_idx.cpu == cpu.cpu)
296 			return idx;
297 
298 		if (cpu_at_idx.cpu > cpu.cpu)
299 			high = idx;
300 		else
301 			low = idx + 1;
302 	}
303 
304 	return -1;
305 }
306 
307 bool perf_cpu_map__has(const struct perf_cpu_map *cpus, struct perf_cpu cpu)
308 {
309 	return perf_cpu_map__idx(cpus, cpu) != -1;
310 }
311 
312 struct perf_cpu perf_cpu_map__max(const struct perf_cpu_map *map)
313 {
314 	struct perf_cpu result = {
315 		.cpu = -1
316 	};
317 
318 	// cpu_map__trim_new() qsort()s it, cpu_map__default_new() sorts it as well.
319 	return map->nr > 0 ? map->map[map->nr - 1] : result;
320 }
321 
322 /** Is 'b' a subset of 'a'. */
323 bool perf_cpu_map__is_subset(const struct perf_cpu_map *a, const struct perf_cpu_map *b)
324 {
325 	if (a == b || !b)
326 		return true;
327 	if (!a || b->nr > a->nr)
328 		return false;
329 
330 	for (int i = 0, j = 0; i < a->nr; i++) {
331 		if (a->map[i].cpu > b->map[j].cpu)
332 			return false;
333 		if (a->map[i].cpu == b->map[j].cpu) {
334 			j++;
335 			if (j == b->nr)
336 				return true;
337 		}
338 	}
339 	return false;
340 }
341 
342 /*
343  * Merge two cpumaps
344  *
345  * orig either gets freed and replaced with a new map, or reused
346  * with no reference count change (similar to "realloc")
347  * other has its reference count increased.
348  */
349 
350 struct perf_cpu_map *perf_cpu_map__merge(struct perf_cpu_map *orig,
351 					 struct perf_cpu_map *other)
352 {
353 	struct perf_cpu *tmp_cpus;
354 	int tmp_len;
355 	int i, j, k;
356 	struct perf_cpu_map *merged;
357 
358 	if (perf_cpu_map__is_subset(orig, other))
359 		return orig;
360 	if (perf_cpu_map__is_subset(other, orig)) {
361 		perf_cpu_map__put(orig);
362 		return perf_cpu_map__get(other);
363 	}
364 
365 	tmp_len = orig->nr + other->nr;
366 	tmp_cpus = malloc(tmp_len * sizeof(struct perf_cpu));
367 	if (!tmp_cpus)
368 		return NULL;
369 
370 	/* Standard merge algorithm from wikipedia */
371 	i = j = k = 0;
372 	while (i < orig->nr && j < other->nr) {
373 		if (orig->map[i].cpu <= other->map[j].cpu) {
374 			if (orig->map[i].cpu == other->map[j].cpu)
375 				j++;
376 			tmp_cpus[k++] = orig->map[i++];
377 		} else
378 			tmp_cpus[k++] = other->map[j++];
379 	}
380 
381 	while (i < orig->nr)
382 		tmp_cpus[k++] = orig->map[i++];
383 
384 	while (j < other->nr)
385 		tmp_cpus[k++] = other->map[j++];
386 	assert(k <= tmp_len);
387 
388 	merged = cpu_map__trim_new(k, tmp_cpus);
389 	free(tmp_cpus);
390 	perf_cpu_map__put(orig);
391 	return merged;
392 }
393