1
2
3
4
5
6
7
8
9#include <sys/mman.h>
10#include <inttypes.h>
11#include <asm/bug.h>
12#ifdef HAVE_LIBNUMA_SUPPORT
13#include <numaif.h>
14#endif
15#include "debug.h"
16#include "event.h"
17#include "mmap.h"
18#include "util.h"
19
20size_t perf_mmap__mmap_len(struct perf_mmap *map)
21{
22 return map->mask + 1 + page_size;
23}
24
25
26static union perf_event *perf_mmap__read(struct perf_mmap *map,
27 u64 *startp, u64 end)
28{
29 unsigned char *data = map->base + page_size;
30 union perf_event *event = NULL;
31 int diff = end - *startp;
32
33 if (diff >= (int)sizeof(event->header)) {
34 size_t size;
35
36 event = (union perf_event *)&data[*startp & map->mask];
37 size = event->header.size;
38
39 if (size < sizeof(event->header) || diff < (int)size)
40 return NULL;
41
42
43
44
45
46 if ((*startp & map->mask) + size != ((*startp + size) & map->mask)) {
47 unsigned int offset = *startp;
48 unsigned int len = min(sizeof(*event), size), cpy;
49 void *dst = map->event_copy;
50
51 do {
52 cpy = min(map->mask + 1 - (offset & map->mask), len);
53 memcpy(dst, &data[offset & map->mask], cpy);
54 offset += cpy;
55 dst += cpy;
56 len -= cpy;
57 } while (len);
58
59 event = (union perf_event *)map->event_copy;
60 }
61
62 *startp += size;
63 }
64
65 return event;
66}
67
68
69
70
71
72
73
74
75
76
77
78
79
80union perf_event *perf_mmap__read_event(struct perf_mmap *map)
81{
82 union perf_event *event;
83
84
85
86
87 if (!refcount_read(&map->refcnt))
88 return NULL;
89
90
91 if (!map->overwrite)
92 map->end = perf_mmap__read_head(map);
93
94 event = perf_mmap__read(map, &map->start, map->end);
95
96 if (!map->overwrite)
97 map->prev = map->start;
98
99 return event;
100}
101
102static bool perf_mmap__empty(struct perf_mmap *map)
103{
104 return perf_mmap__read_head(map) == map->prev && !map->auxtrace_mmap.base;
105}
106
107void perf_mmap__get(struct perf_mmap *map)
108{
109 refcount_inc(&map->refcnt);
110}
111
112void perf_mmap__put(struct perf_mmap *map)
113{
114 BUG_ON(map->base && refcount_read(&map->refcnt) == 0);
115
116 if (refcount_dec_and_test(&map->refcnt))
117 perf_mmap__munmap(map);
118}
119
120void perf_mmap__consume(struct perf_mmap *map)
121{
122 if (!map->overwrite) {
123 u64 old = map->prev;
124
125 perf_mmap__write_tail(map, old);
126 }
127
128 if (refcount_read(&map->refcnt) == 1 && perf_mmap__empty(map))
129 perf_mmap__put(map);
130}
131
132int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
133 struct auxtrace_mmap_params *mp __maybe_unused,
134 void *userpg __maybe_unused,
135 int fd __maybe_unused)
136{
137 return 0;
138}
139
140void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
141{
142}
143
144void __weak auxtrace_mmap_params__init(struct auxtrace_mmap_params *mp __maybe_unused,
145 off_t auxtrace_offset __maybe_unused,
146 unsigned int auxtrace_pages __maybe_unused,
147 bool auxtrace_overwrite __maybe_unused)
148{
149}
150
151void __weak auxtrace_mmap_params__set_idx(struct auxtrace_mmap_params *mp __maybe_unused,
152 struct perf_evlist *evlist __maybe_unused,
153 int idx __maybe_unused,
154 bool per_cpu __maybe_unused)
155{
156}
157
158#ifdef HAVE_AIO_SUPPORT
159static int perf_mmap__aio_enabled(struct perf_mmap *map)
160{
161 return map->aio.nr_cblocks > 0;
162}
163
164#ifdef HAVE_LIBNUMA_SUPPORT
165static int perf_mmap__aio_alloc(struct perf_mmap *map, int idx)
166{
167 map->aio.data[idx] = mmap(NULL, perf_mmap__mmap_len(map), PROT_READ|PROT_WRITE,
168 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
169 if (map->aio.data[idx] == MAP_FAILED) {
170 map->aio.data[idx] = NULL;
171 return -1;
172 }
173
174 return 0;
175}
176
177static void perf_mmap__aio_free(struct perf_mmap *map, int idx)
178{
179 if (map->aio.data[idx]) {
180 munmap(map->aio.data[idx], perf_mmap__mmap_len(map));
181 map->aio.data[idx] = NULL;
182 }
183}
184
185static int perf_mmap__aio_bind(struct perf_mmap *map, int idx, int cpu, int affinity)
186{
187 void *data;
188 size_t mmap_len;
189 unsigned long node_mask;
190
191 if (affinity != PERF_AFFINITY_SYS && cpu__max_node() > 1) {
192 data = map->aio.data[idx];
193 mmap_len = perf_mmap__mmap_len(map);
194 node_mask = 1UL << cpu__get_node(cpu);
195 if (mbind(data, mmap_len, MPOL_BIND, &node_mask, 1, 0)) {
196 pr_err("Failed to bind [%p-%p] AIO buffer to node %d: error %m\n",
197 data, data + mmap_len, cpu__get_node(cpu));
198 return -1;
199 }
200 }
201
202 return 0;
203}
204#else
205static int perf_mmap__aio_alloc(struct perf_mmap *map, int idx)
206{
207 map->aio.data[idx] = malloc(perf_mmap__mmap_len(map));
208 if (map->aio.data[idx] == NULL)
209 return -1;
210
211 return 0;
212}
213
214static void perf_mmap__aio_free(struct perf_mmap *map, int idx)
215{
216 zfree(&(map->aio.data[idx]));
217}
218
219static int perf_mmap__aio_bind(struct perf_mmap *map __maybe_unused, int idx __maybe_unused,
220 int cpu __maybe_unused, int affinity __maybe_unused)
221{
222 return 0;
223}
224#endif
225
226static int perf_mmap__aio_mmap(struct perf_mmap *map, struct mmap_params *mp)
227{
228 int delta_max, i, prio, ret;
229
230 map->aio.nr_cblocks = mp->nr_cblocks;
231 if (map->aio.nr_cblocks) {
232 map->aio.aiocb = calloc(map->aio.nr_cblocks, sizeof(struct aiocb *));
233 if (!map->aio.aiocb) {
234 pr_debug2("failed to allocate aiocb for data buffer, error %m\n");
235 return -1;
236 }
237 map->aio.cblocks = calloc(map->aio.nr_cblocks, sizeof(struct aiocb));
238 if (!map->aio.cblocks) {
239 pr_debug2("failed to allocate cblocks for data buffer, error %m\n");
240 return -1;
241 }
242 map->aio.data = calloc(map->aio.nr_cblocks, sizeof(void *));
243 if (!map->aio.data) {
244 pr_debug2("failed to allocate data buffer, error %m\n");
245 return -1;
246 }
247 delta_max = sysconf(_SC_AIO_PRIO_DELTA_MAX);
248 for (i = 0; i < map->aio.nr_cblocks; ++i) {
249 ret = perf_mmap__aio_alloc(map, i);
250 if (ret == -1) {
251 pr_debug2("failed to allocate data buffer area, error %m");
252 return -1;
253 }
254 ret = perf_mmap__aio_bind(map, i, map->cpu, mp->affinity);
255 if (ret == -1)
256 return -1;
257
258
259
260
261
262
263 map->aio.cblocks[i].aio_fildes = -1;
264
265
266
267
268
269
270
271
272 prio = delta_max - i;
273 map->aio.cblocks[i].aio_reqprio = prio >= 0 ? prio : 0;
274 }
275 }
276
277 return 0;
278}
279
280static void perf_mmap__aio_munmap(struct perf_mmap *map)
281{
282 int i;
283
284 for (i = 0; i < map->aio.nr_cblocks; ++i)
285 perf_mmap__aio_free(map, i);
286 if (map->aio.data)
287 zfree(&map->aio.data);
288 zfree(&map->aio.cblocks);
289 zfree(&map->aio.aiocb);
290}
291#else
292static int perf_mmap__aio_enabled(struct perf_mmap *map __maybe_unused)
293{
294 return 0;
295}
296
297static int perf_mmap__aio_mmap(struct perf_mmap *map __maybe_unused,
298 struct mmap_params *mp __maybe_unused)
299{
300 return 0;
301}
302
303static void perf_mmap__aio_munmap(struct perf_mmap *map __maybe_unused)
304{
305}
306#endif
307
308void perf_mmap__munmap(struct perf_mmap *map)
309{
310 perf_mmap__aio_munmap(map);
311 if (map->data != NULL) {
312 munmap(map->data, perf_mmap__mmap_len(map));
313 map->data = NULL;
314 }
315 if (map->base != NULL) {
316 munmap(map->base, perf_mmap__mmap_len(map));
317 map->base = NULL;
318 map->fd = -1;
319 refcount_set(&map->refcnt, 0);
320 }
321 auxtrace_mmap__munmap(&map->auxtrace_mmap);
322}
323
324static void build_node_mask(int node, cpu_set_t *mask)
325{
326 int c, cpu, nr_cpus;
327 const struct cpu_map *cpu_map = NULL;
328
329 cpu_map = cpu_map__online();
330 if (!cpu_map)
331 return;
332
333 nr_cpus = cpu_map__nr(cpu_map);
334 for (c = 0; c < nr_cpus; c++) {
335 cpu = cpu_map->map[c];
336 if (cpu__get_node(cpu) == node)
337 CPU_SET(cpu, mask);
338 }
339}
340
341static void perf_mmap__setup_affinity_mask(struct perf_mmap *map, struct mmap_params *mp)
342{
343 CPU_ZERO(&map->affinity_mask);
344 if (mp->affinity == PERF_AFFINITY_NODE && cpu__max_node() > 1)
345 build_node_mask(cpu__get_node(map->cpu), &map->affinity_mask);
346 else if (mp->affinity == PERF_AFFINITY_CPU)
347 CPU_SET(map->cpu, &map->affinity_mask);
348}
349
350int perf_mmap__mmap(struct perf_mmap *map, struct mmap_params *mp, int fd, int cpu)
351{
352
353
354
355
356
357
358
359
360
361
362
363
364
365 refcount_set(&map->refcnt, 2);
366 map->prev = 0;
367 map->mask = mp->mask;
368 map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
369 MAP_SHARED, fd, 0);
370 if (map->base == MAP_FAILED) {
371 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
372 errno);
373 map->base = NULL;
374 return -1;
375 }
376 map->fd = fd;
377 map->cpu = cpu;
378
379 perf_mmap__setup_affinity_mask(map, mp);
380
381 map->flush = mp->flush;
382
383 map->comp_level = mp->comp_level;
384
385 if (map->comp_level && !perf_mmap__aio_enabled(map)) {
386 map->data = mmap(NULL, perf_mmap__mmap_len(map), PROT_READ|PROT_WRITE,
387 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
388 if (map->data == MAP_FAILED) {
389 pr_debug2("failed to mmap data buffer, error %d\n",
390 errno);
391 map->data = NULL;
392 return -1;
393 }
394 }
395
396 if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
397 &mp->auxtrace_mp, map->base, fd))
398 return -1;
399
400 return perf_mmap__aio_mmap(map, mp);
401}
402
403static int overwrite_rb_find_range(void *buf, int mask, u64 *start, u64 *end)
404{
405 struct perf_event_header *pheader;
406 u64 evt_head = *start;
407 int size = mask + 1;
408
409 pr_debug2("%s: buf=%p, start=%"PRIx64"\n", __func__, buf, *start);
410 pheader = (struct perf_event_header *)(buf + (*start & mask));
411 while (true) {
412 if (evt_head - *start >= (unsigned int)size) {
413 pr_debug("Finished reading overwrite ring buffer: rewind\n");
414 if (evt_head - *start > (unsigned int)size)
415 evt_head -= pheader->size;
416 *end = evt_head;
417 return 0;
418 }
419
420 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
421
422 if (pheader->size == 0) {
423 pr_debug("Finished reading overwrite ring buffer: get start\n");
424 *end = evt_head;
425 return 0;
426 }
427
428 evt_head += pheader->size;
429 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
430 }
431 WARN_ONCE(1, "Shouldn't get here\n");
432 return -1;
433}
434
435
436
437
438static int __perf_mmap__read_init(struct perf_mmap *md)
439{
440 u64 head = perf_mmap__read_head(md);
441 u64 old = md->prev;
442 unsigned char *data = md->base + page_size;
443 unsigned long size;
444
445 md->start = md->overwrite ? head : old;
446 md->end = md->overwrite ? old : head;
447
448 if ((md->end - md->start) < md->flush)
449 return -EAGAIN;
450
451 size = md->end - md->start;
452 if (size > (unsigned long)(md->mask) + 1) {
453 if (!md->overwrite) {
454 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
455
456 md->prev = head;
457 perf_mmap__consume(md);
458 return -EAGAIN;
459 }
460
461
462
463
464
465 if (overwrite_rb_find_range(data, md->mask, &md->start, &md->end))
466 return -EINVAL;
467 }
468
469 return 0;
470}
471
472int perf_mmap__read_init(struct perf_mmap *map)
473{
474
475
476
477 if (!refcount_read(&map->refcnt))
478 return -ENOENT;
479
480 return __perf_mmap__read_init(map);
481}
482
483int perf_mmap__push(struct perf_mmap *md, void *to,
484 int push(struct perf_mmap *map, void *to, void *buf, size_t size))
485{
486 u64 head = perf_mmap__read_head(md);
487 unsigned char *data = md->base + page_size;
488 unsigned long size;
489 void *buf;
490 int rc = 0;
491
492 rc = perf_mmap__read_init(md);
493 if (rc < 0)
494 return (rc == -EAGAIN) ? 1 : -1;
495
496 size = md->end - md->start;
497
498 if ((md->start & md->mask) + size != (md->end & md->mask)) {
499 buf = &data[md->start & md->mask];
500 size = md->mask + 1 - (md->start & md->mask);
501 md->start += size;
502
503 if (push(md, to, buf, size) < 0) {
504 rc = -1;
505 goto out;
506 }
507 }
508
509 buf = &data[md->start & md->mask];
510 size = md->end - md->start;
511 md->start += size;
512
513 if (push(md, to, buf, size) < 0) {
514 rc = -1;
515 goto out;
516 }
517
518 md->prev = head;
519 perf_mmap__consume(md);
520out:
521 return rc;
522}
523
524
525
526
527
528
529
530void perf_mmap__read_done(struct perf_mmap *map)
531{
532
533
534
535 if (!refcount_read(&map->refcnt))
536 return;
537
538 map->prev = perf_mmap__read_head(map);
539}
540