xref: /linux/tools/perf/util/mmap.c (revision a9e900bc5c5914aca750afafa459363e575d3046)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011-2017, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4  *
5  * Parts came from evlist.c builtin-{top,stat,record}.c, see those files for further
6  * copyright notes.
7  */
8 
9 #include <sys/mman.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <asm/bug.h>
13 #include <linux/zalloc.h>
14 #include <stdlib.h>
15 #include <string.h>
16 #include <unistd.h> // sysconf()
17 #include <perf/mmap.h>
18 #ifdef HAVE_LIBNUMA_SUPPORT
19 #include <numaif.h>
20 #endif
21 #include "cpumap.h"
22 #include "debug.h"
23 #include "event.h"
24 #include "mmap.h"
25 #include "../perf.h"
26 #include <internal/lib.h> /* page_size */
27 #include <linux/bitmap.h>
28 
29 #define MASK_SIZE 1023
30 void mmap_cpu_mask__scnprintf(struct mmap_cpu_mask *mask, const char *tag)
31 {
32 	char buf[MASK_SIZE + 1];
33 	size_t len;
34 
35 	len = bitmap_scnprintf(mask->bits, mask->nbits, buf, MASK_SIZE);
36 	buf[len] = '\0';
37 	pr_debug("%p: %s mask[%zd]: %s\n", mask, tag, mask->nbits, buf);
38 }
39 
40 size_t mmap__mmap_len(struct mmap *map)
41 {
42 	return perf_mmap__mmap_len(&map->core);
43 }
44 
45 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
46 			       struct auxtrace_mmap_params *mp __maybe_unused,
47 			       void *userpg __maybe_unused,
48 			       int fd __maybe_unused)
49 {
50 	return 0;
51 }
52 
53 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
54 {
55 }
56 
57 void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
58 				       off_t auxtrace_offset __maybe_unused,
59 				       unsigned int auxtrace_pages __maybe_unused,
60 				       bool auxtrace_overwrite __maybe_unused)
61 {
62 }
63 
64 void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
65 					  struct evlist *evlist __maybe_unused,
66 					  struct evsel *evsel __maybe_unused,
67 					  int idx __maybe_unused)
68 {
69 }
70 
71 #ifdef HAVE_AIO_SUPPORT
72 static int perf_mmap__aio_enabled(struct mmap *map)
73 {
74 	return map->aio.nr_cblocks > 0;
75 }
76 
77 #ifdef HAVE_LIBNUMA_SUPPORT
78 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
79 {
80 	map->aio.data[idx] = mmap(NULL, mmap__mmap_len(map), PROT_READ|PROT_WRITE,
81 				  MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
82 	if (map->aio.data[idx] == MAP_FAILED) {
83 		map->aio.data[idx] = NULL;
84 		return -1;
85 	}
86 
87 	return 0;
88 }
89 
90 static void perf_mmap__aio_free(struct mmap *map, int idx)
91 {
92 	if (!map->aio.data || !map->aio.data[idx])
93 		return;
94 	munmap(map->aio.data[idx], mmap__mmap_len(map));
95 	map->aio.data[idx] = NULL;
96 }
97 
98 static int perf_mmap__aio_bind(struct mmap *map, int idx, struct perf_cpu cpu, int affinity)
99 {
100 	void *data;
101 	size_t mmap_len;
102 	unsigned long *node_mask;
103 	unsigned long node_index;
104 	int err = 0;
105 
106 	if (affinity != PERF_AFFINITY_SYS && cpu__max_node() > 1) {
107 		int node;
108 
109 		data = map->aio.data[idx];
110 		mmap_len = mmap__mmap_len(map);
111 		node = cpu__get_node(cpu);
112 		/* -1 sign-extends to ULONG_MAX, wrapping bitmap_zalloc(0) and OOB __set_bit */
113 		if (node < 0)
114 			return 0;
115 		node_index = node;
116 		node_mask = bitmap_zalloc(node_index + 1);
117 		if (!node_mask) {
118 			pr_err("Failed to allocate node mask for mbind: error %m\n");
119 			return -1;
120 		}
121 		__set_bit(node_index, node_mask);
122 		if (mbind(data, mmap_len, MPOL_BIND, node_mask, node_index + 1 + 1, 0)) {
123 			pr_err("Failed to bind [%p-%p] AIO buffer to node %lu: error %m\n",
124 				data, data + mmap_len, node_index);
125 			err = -1;
126 		}
127 		bitmap_free(node_mask);
128 	}
129 
130 	return err;
131 }
132 #else /* !HAVE_LIBNUMA_SUPPORT */
133 static int perf_mmap__aio_alloc(struct mmap *map, int idx)
134 {
135 	map->aio.data[idx] = malloc(mmap__mmap_len(map));
136 	if (map->aio.data[idx] == NULL)
137 		return -1;
138 
139 	return 0;
140 }
141 
142 static void perf_mmap__aio_free(struct mmap *map, int idx)
143 {
144 	if (!map->aio.data)
145 		return;
146 	zfree(&(map->aio.data[idx]));
147 }
148 
149 static int perf_mmap__aio_bind(struct mmap *map __maybe_unused, int idx __maybe_unused,
150 		struct perf_cpu cpu __maybe_unused, int affinity __maybe_unused)
151 {
152 	return 0;
153 }
154 #endif
155 
156 static int perf_mmap__aio_mmap(struct mmap *map, struct mmap_params *mp)
157 {
158 	int delta_max, i, prio, ret;
159 
160 	map->aio.nr_cblocks = mp->nr_cblocks;
161 	if (map->aio.nr_cblocks) {
162 		map->aio.aiocb = calloc(map->aio.nr_cblocks, sizeof(struct aiocb *));
163 		if (!map->aio.aiocb) {
164 			pr_debug2("failed to allocate aiocb for data buffer, error %m\n");
165 			return -1;
166 		}
167 		map->aio.cblocks = calloc(map->aio.nr_cblocks, sizeof(struct aiocb));
168 		if (!map->aio.cblocks) {
169 			pr_debug2("failed to allocate cblocks for data buffer, error %m\n");
170 			return -1;
171 		}
172 		map->aio.data = calloc(map->aio.nr_cblocks, sizeof(void *));
173 		if (!map->aio.data) {
174 			pr_debug2("failed to allocate data buffer, error %m\n");
175 			return -1;
176 		}
177 		delta_max = sysconf(_SC_AIO_PRIO_DELTA_MAX);
178 		for (i = 0; i < map->aio.nr_cblocks; ++i) {
179 			ret = perf_mmap__aio_alloc(map, i);
180 			if (ret == -1) {
181 				pr_debug2("failed to allocate data buffer area, error %m");
182 				return -1;
183 			}
184 			ret = perf_mmap__aio_bind(map, i, map->core.cpu, mp->affinity);
185 			if (ret == -1)
186 				return -1;
187 			/*
188 			 * Use cblock.aio_fildes value different from -1
189 			 * to denote started aio write operation on the
190 			 * cblock so it requires explicit record__aio_sync()
191 			 * call prior the cblock may be reused again.
192 			 */
193 			map->aio.cblocks[i].aio_fildes = -1;
194 			/*
195 			 * Allocate cblocks with priority delta to have
196 			 * faster aio write system calls because queued requests
197 			 * are kept in separate per-prio queues and adding
198 			 * a new request will iterate thru shorter per-prio
199 			 * list. Blocks with numbers higher than
200 			 *  _SC_AIO_PRIO_DELTA_MAX go with priority 0.
201 			 */
202 			prio = delta_max - i;
203 			map->aio.cblocks[i].aio_reqprio = prio >= 0 ? prio : 0;
204 		}
205 	}
206 
207 	return 0;
208 }
209 
210 static void perf_mmap__aio_munmap(struct mmap *map)
211 {
212 	int i;
213 
214 	for (i = 0; i < map->aio.nr_cblocks; ++i)
215 		perf_mmap__aio_free(map, i);
216 	if (map->aio.data)
217 		zfree(&map->aio.data);
218 	zfree(&map->aio.cblocks);
219 	zfree(&map->aio.aiocb);
220 }
221 #else /* !HAVE_AIO_SUPPORT */
222 static int perf_mmap__aio_enabled(struct mmap *map __maybe_unused)
223 {
224 	return 0;
225 }
226 
227 static int perf_mmap__aio_mmap(struct mmap *map __maybe_unused,
228 			       struct mmap_params *mp __maybe_unused)
229 {
230 	return 0;
231 }
232 
233 static void perf_mmap__aio_munmap(struct mmap *map __maybe_unused)
234 {
235 }
236 #endif
237 
238 void mmap__munmap(struct mmap *map)
239 {
240 	bitmap_free(map->affinity_mask.bits);
241 	map->affinity_mask.bits = NULL;
242 	map->affinity_mask.nbits = 0;
243 
244 	zstd_fini(&map->zstd_data);
245 
246 	perf_mmap__aio_munmap(map);
247 	if (map->data != NULL) {
248 		munmap(map->data, mmap__mmap_len(map));
249 		map->data = NULL;
250 	}
251 	auxtrace_mmap__munmap(&map->auxtrace_mmap);
252 }
253 
254 static void build_node_mask(int node, struct mmap_cpu_mask *mask)
255 {
256 	int idx, nr_cpus;
257 	struct perf_cpu cpu;
258 	struct perf_cpu_map *cpu_map = cpu_map__online();
259 
260 	if (!cpu_map)
261 		return;
262 
263 	nr_cpus = perf_cpu_map__nr(cpu_map);
264 	for (idx = 0; idx < nr_cpus; idx++) {
265 		cpu = perf_cpu_map__cpu(cpu_map, idx); /* map c index to online cpu index */
266 		if (cpu__get_node(cpu) == node)
267 			__set_bit(cpu.cpu, mask->bits);
268 	}
269 	perf_cpu_map__put(cpu_map);
270 }
271 
272 static int perf_mmap__setup_affinity_mask(struct mmap *map, struct mmap_params *mp)
273 {
274 	map->affinity_mask.nbits = cpu__max_cpu().cpu;
275 	map->affinity_mask.bits = bitmap_zalloc(map->affinity_mask.nbits);
276 	if (!map->affinity_mask.bits)
277 		return -1;
278 
279 	if (mp->affinity == PERF_AFFINITY_NODE && cpu__max_node() > 1)
280 		build_node_mask(cpu__get_node(map->core.cpu), &map->affinity_mask);
281 	else if (mp->affinity == PERF_AFFINITY_CPU)
282 		__set_bit(map->core.cpu.cpu, map->affinity_mask.bits);
283 
284 	return 0;
285 }
286 
287 int mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, struct perf_cpu cpu)
288 {
289 	if (perf_mmap__mmap(&map->core, &mp->core, fd, cpu)) {
290 		pr_debug2("failed to mmap perf event ring buffer, error %d\n",
291 			  errno);
292 		return -1;
293 	}
294 
295 	if (mp->affinity != PERF_AFFINITY_SYS &&
296 		perf_mmap__setup_affinity_mask(map, mp)) {
297 		pr_debug2("failed to alloc mmap affinity mask, error %d\n",
298 			  errno);
299 		return -1;
300 	}
301 
302 	if (verbose == 2)
303 		mmap_cpu_mask__scnprintf(&map->affinity_mask, "mmap");
304 
305 	map->core.flush = mp->flush;
306 
307 	if (zstd_init(&map->zstd_data, mp->comp_level)) {
308 		pr_debug2("failed to init mmap compressor, error %d\n", errno);
309 		return -1;
310 	}
311 
312 	if (mp->comp_level && !perf_mmap__aio_enabled(map)) {
313 		map->data = mmap(NULL, mmap__mmap_len(map), PROT_READ|PROT_WRITE,
314 				 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
315 		if (map->data == MAP_FAILED) {
316 			pr_debug2("failed to mmap data buffer, error %d\n",
317 					errno);
318 			map->data = NULL;
319 			return -1;
320 		}
321 	}
322 
323 	if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
324 				&mp->auxtrace_mp, map->core.base, fd))
325 		return -1;
326 
327 	return perf_mmap__aio_mmap(map, mp);
328 }
329 
330 int perf_mmap__push(struct mmap *md, void *to,
331 		    int push(struct mmap *map, void *to, void *buf, size_t size))
332 {
333 	u64 head = perf_mmap__read_head(&md->core);
334 	unsigned char *data = md->core.base + page_size;
335 	unsigned long size;
336 	void *buf;
337 	int rc = 0;
338 
339 	rc = perf_mmap__read_init(&md->core);
340 	if (rc < 0)
341 		return (rc == -EAGAIN) ? 1 : -1;
342 
343 	size = md->core.end - md->core.start;
344 
345 	if ((md->core.start & md->core.mask) + size != (md->core.end & md->core.mask)) {
346 		buf = &data[md->core.start & md->core.mask];
347 		size = md->core.mask + 1 - (md->core.start & md->core.mask);
348 		md->core.start += size;
349 
350 		if (push(md, to, buf, size) < 0) {
351 			rc = -1;
352 			goto out;
353 		}
354 	}
355 
356 	buf = &data[md->core.start & md->core.mask];
357 	size = md->core.end - md->core.start;
358 	md->core.start += size;
359 
360 	if (push(md, to, buf, size) < 0) {
361 		rc = -1;
362 		goto out;
363 	}
364 
365 	md->core.prev = head;
366 	perf_mmap__consume(&md->core);
367 out:
368 	return rc;
369 }
370