xref: /linux/tools/lib/bpf/ringbuf.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2 /*
3  * Ring buffer operations.
4  *
5  * Copyright (C) 2020 Facebook, Inc.
6  */
7 #ifndef _GNU_SOURCE
8 #define _GNU_SOURCE
9 #endif
10 #include <stdlib.h>
11 #include <stdio.h>
12 #include <errno.h>
13 #include <unistd.h>
14 #include <linux/err.h>
15 #include <linux/bpf.h>
16 #include <asm/barrier.h>
17 #include <sys/mman.h>
18 #include <sys/epoll.h>
19 #include <time.h>
20 
21 #include "libbpf.h"
22 #include "libbpf_internal.h"
23 #include "bpf.h"
24 #include "str_error.h"
25 
26 struct ring {
27 	ring_buffer_sample_fn sample_cb;
28 	void *ctx;
29 	void *data;
30 	unsigned long *consumer_pos;
31 	unsigned long *producer_pos;
32 	unsigned long mask;
33 	int map_fd;
34 };
35 
36 struct ring_buffer {
37 	struct epoll_event *events;
38 	struct ring **rings;
39 	size_t page_size;
40 	int epoll_fd;
41 	int ring_cnt;
42 };
43 
44 struct user_ring_buffer {
45 	struct epoll_event event;
46 	unsigned long *consumer_pos;
47 	unsigned long *producer_pos;
48 	void *data;
49 	unsigned long mask;
50 	size_t page_size;
51 	int map_fd;
52 	int epoll_fd;
53 };
54 
55 /* 8-byte ring buffer header structure */
56 struct ringbuf_hdr {
57 	__u32 len;
58 	__u32 pad;
59 };
60 
61 static void ringbuf_free_ring(struct ring_buffer *rb, struct ring *r)
62 {
63 	if (r->consumer_pos) {
64 		munmap(r->consumer_pos, rb->page_size);
65 		r->consumer_pos = NULL;
66 	}
67 	if (r->producer_pos) {
68 		munmap(r->producer_pos, rb->page_size + 2 * (r->mask + 1));
69 		r->producer_pos = NULL;
70 	}
71 
72 	free(r);
73 }
74 
75 /* Add extra RINGBUF maps to this ring buffer manager */
76 int ring_buffer__add(struct ring_buffer *rb, int map_fd,
77 		     ring_buffer_sample_fn sample_cb, void *ctx)
78 {
79 	struct bpf_map_info info;
80 	__u32 len = sizeof(info);
81 	struct epoll_event *e;
82 	struct ring *r;
83 	__u64 mmap_sz;
84 	void *tmp;
85 	int err;
86 
87 	memset(&info, 0, sizeof(info));
88 
89 	err = bpf_map_get_info_by_fd(map_fd, &info, &len);
90 	if (err) {
91 		err = -errno;
92 		pr_warn("ringbuf: failed to get map info for fd=%d: %s\n",
93 			map_fd, errstr(err));
94 		return libbpf_err(err);
95 	}
96 
97 	if (info.type != BPF_MAP_TYPE_RINGBUF) {
98 		pr_warn("ringbuf: map fd=%d is not BPF_MAP_TYPE_RINGBUF\n",
99 			map_fd);
100 		return libbpf_err(-EINVAL);
101 	}
102 
103 	tmp = libbpf_reallocarray(rb->rings, rb->ring_cnt + 1, sizeof(*rb->rings));
104 	if (!tmp)
105 		return libbpf_err(-ENOMEM);
106 	rb->rings = tmp;
107 
108 	tmp = libbpf_reallocarray(rb->events, rb->ring_cnt + 1, sizeof(*rb->events));
109 	if (!tmp)
110 		return libbpf_err(-ENOMEM);
111 	rb->events = tmp;
112 
113 	r = calloc(1, sizeof(*r));
114 	if (!r)
115 		return libbpf_err(-ENOMEM);
116 	rb->rings[rb->ring_cnt] = r;
117 
118 	r->map_fd = map_fd;
119 	r->sample_cb = sample_cb;
120 	r->ctx = ctx;
121 	r->mask = info.max_entries - 1;
122 
123 	/* Map writable consumer page */
124 	tmp = mmap(NULL, rb->page_size, PROT_READ | PROT_WRITE, MAP_SHARED, map_fd, 0);
125 	if (tmp == MAP_FAILED) {
126 		err = -errno;
127 		pr_warn("ringbuf: failed to mmap consumer page for map fd=%d: %s\n",
128 			map_fd, errstr(err));
129 		goto err_out;
130 	}
131 	r->consumer_pos = tmp;
132 
133 	/* Map read-only producer page and data pages. We map twice as big
134 	 * data size to allow simple reading of samples that wrap around the
135 	 * end of a ring buffer. See kernel implementation for details.
136 	 */
137 	mmap_sz = rb->page_size + 2 * (__u64)info.max_entries;
138 	if (mmap_sz != (__u64)(size_t)mmap_sz) {
139 		err = -E2BIG;
140 		pr_warn("ringbuf: ring buffer size (%u) is too big\n", info.max_entries);
141 		goto err_out;
142 	}
143 	tmp = mmap(NULL, (size_t)mmap_sz, PROT_READ, MAP_SHARED, map_fd, rb->page_size);
144 	if (tmp == MAP_FAILED) {
145 		err = -errno;
146 		pr_warn("ringbuf: failed to mmap data pages for map fd=%d: %s\n",
147 			map_fd, errstr(err));
148 		goto err_out;
149 	}
150 	r->producer_pos = tmp;
151 	r->data = tmp + rb->page_size;
152 
153 	e = &rb->events[rb->ring_cnt];
154 	memset(e, 0, sizeof(*e));
155 
156 	e->events = EPOLLIN;
157 	e->data.fd = rb->ring_cnt;
158 	if (epoll_ctl(rb->epoll_fd, EPOLL_CTL_ADD, map_fd, e) < 0) {
159 		err = -errno;
160 		pr_warn("ringbuf: failed to epoll add map fd=%d: %s\n",
161 			map_fd, errstr(err));
162 		goto err_out;
163 	}
164 
165 	rb->ring_cnt++;
166 	return 0;
167 
168 err_out:
169 	ringbuf_free_ring(rb, r);
170 	return libbpf_err(err);
171 }
172 
173 void ring_buffer__free(struct ring_buffer *rb)
174 {
175 	int i;
176 
177 	if (!rb)
178 		return;
179 
180 	for (i = 0; i < rb->ring_cnt; ++i)
181 		ringbuf_free_ring(rb, rb->rings[i]);
182 	if (rb->epoll_fd >= 0)
183 		close(rb->epoll_fd);
184 
185 	free(rb->events);
186 	free(rb->rings);
187 	free(rb);
188 }
189 
190 struct ring_buffer *
191 ring_buffer__new(int map_fd, ring_buffer_sample_fn sample_cb, void *ctx,
192 		 const struct ring_buffer_opts *opts)
193 {
194 	struct ring_buffer *rb;
195 	int err;
196 
197 	if (!OPTS_VALID(opts, ring_buffer_opts))
198 		return errno = EINVAL, NULL;
199 
200 	rb = calloc(1, sizeof(*rb));
201 	if (!rb)
202 		return errno = ENOMEM, NULL;
203 
204 	rb->page_size = getpagesize();
205 
206 	rb->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
207 	if (rb->epoll_fd < 0) {
208 		err = -errno;
209 		pr_warn("ringbuf: failed to create epoll instance: %s\n", errstr(err));
210 		goto err_out;
211 	}
212 
213 	err = ring_buffer__add(rb, map_fd, sample_cb, ctx);
214 	if (err)
215 		goto err_out;
216 
217 	return rb;
218 
219 err_out:
220 	ring_buffer__free(rb);
221 	return errno = -err, NULL;
222 }
223 
224 static inline int roundup_len(__u32 len)
225 {
226 	/* clear out top 2 bits (discard and busy, if set) */
227 	len <<= 2;
228 	len >>= 2;
229 	/* add length prefix */
230 	len += BPF_RINGBUF_HDR_SZ;
231 	/* round up to 8 byte alignment */
232 	return (len + 7) / 8 * 8;
233 }
234 
235 static int64_t ringbuf_process_ring(struct ring *r, size_t n)
236 {
237 	int *len_ptr, len, err;
238 	/* 64-bit to avoid overflow in case of extreme application behavior */
239 	int64_t cnt = 0;
240 	unsigned long cons_pos, prod_pos;
241 	bool got_new_data;
242 	void *sample;
243 
244 	cons_pos = smp_load_acquire(r->consumer_pos);
245 	do {
246 		got_new_data = false;
247 		prod_pos = smp_load_acquire(r->producer_pos);
248 		while (cons_pos < prod_pos) {
249 			len_ptr = r->data + (cons_pos & r->mask);
250 			len = smp_load_acquire(len_ptr);
251 
252 			/* sample not committed yet, bail out for now */
253 			if (len & BPF_RINGBUF_BUSY_BIT)
254 				goto done;
255 
256 			got_new_data = true;
257 			cons_pos += roundup_len(len);
258 
259 			if ((len & BPF_RINGBUF_DISCARD_BIT) == 0) {
260 				sample = (void *)len_ptr + BPF_RINGBUF_HDR_SZ;
261 				err = r->sample_cb(r->ctx, sample, len);
262 				if (err < 0) {
263 					/* update consumer pos and bail out */
264 					smp_store_release(r->consumer_pos,
265 							  cons_pos);
266 					return err;
267 				}
268 				cnt++;
269 			}
270 
271 			smp_store_release(r->consumer_pos, cons_pos);
272 
273 			if (cnt >= n)
274 				goto done;
275 		}
276 	} while (got_new_data);
277 done:
278 	return cnt;
279 }
280 
281 /* Consume available ring buffer(s) data without event polling, up to n
282  * records.
283  *
284  * Returns number of records consumed across all registered ring buffers (or
285  * n, whichever is less), or negative number if any of the callbacks return
286  * error.
287  */
288 int ring_buffer__consume_n(struct ring_buffer *rb, size_t n)
289 {
290 	int64_t err, res = 0;
291 	int i;
292 
293 	for (i = 0; i < rb->ring_cnt; i++) {
294 		struct ring *ring = rb->rings[i];
295 
296 		err = ringbuf_process_ring(ring, n);
297 		if (err < 0)
298 			return libbpf_err(err);
299 		res += err;
300 		n -= err;
301 
302 		if (n == 0)
303 			break;
304 	}
305 	return res > INT_MAX ? INT_MAX : res;
306 }
307 
308 /* Consume available ring buffer(s) data without event polling.
309  * Returns number of records consumed across all registered ring buffers (or
310  * INT_MAX, whichever is less), or negative number if any of the callbacks
311  * return error.
312  */
313 int ring_buffer__consume(struct ring_buffer *rb)
314 {
315 	int64_t err, res = 0;
316 	int i;
317 
318 	for (i = 0; i < rb->ring_cnt; i++) {
319 		struct ring *ring = rb->rings[i];
320 
321 		err = ringbuf_process_ring(ring, INT_MAX);
322 		if (err < 0)
323 			return libbpf_err(err);
324 		res += err;
325 		if (res > INT_MAX) {
326 			res = INT_MAX;
327 			break;
328 		}
329 	}
330 	return res;
331 }
332 
333 /* Poll for available data and consume records, if any are available.
334  * Returns number of records consumed (or INT_MAX, whichever is less), or
335  * negative number, if any of the registered callbacks returned error.
336  */
337 int ring_buffer__poll(struct ring_buffer *rb, int timeout_ms)
338 {
339 	int i, cnt;
340 	int64_t err, res = 0;
341 
342 	cnt = epoll_wait(rb->epoll_fd, rb->events, rb->ring_cnt, timeout_ms);
343 	if (cnt < 0)
344 		return libbpf_err(-errno);
345 
346 	for (i = 0; i < cnt; i++) {
347 		__u32 ring_id = rb->events[i].data.fd;
348 		struct ring *ring = rb->rings[ring_id];
349 
350 		err = ringbuf_process_ring(ring, INT_MAX);
351 		if (err < 0)
352 			return libbpf_err(err);
353 		res += err;
354 	}
355 	if (res > INT_MAX)
356 		res = INT_MAX;
357 	return res;
358 }
359 
360 /* Get an fd that can be used to sleep until data is available in the ring(s) */
361 int ring_buffer__epoll_fd(const struct ring_buffer *rb)
362 {
363 	return rb->epoll_fd;
364 }
365 
366 struct ring *ring_buffer__ring(struct ring_buffer *rb, unsigned int idx)
367 {
368 	if (idx >= rb->ring_cnt)
369 		return errno = ERANGE, NULL;
370 
371 	return rb->rings[idx];
372 }
373 
374 unsigned long ring__consumer_pos(const struct ring *r)
375 {
376 	/* Synchronizes with smp_store_release() in ringbuf_process_ring(). */
377 	return smp_load_acquire(r->consumer_pos);
378 }
379 
380 unsigned long ring__producer_pos(const struct ring *r)
381 {
382 	/* Synchronizes with smp_store_release() in __bpf_ringbuf_reserve() in
383 	 * the kernel.
384 	 */
385 	return smp_load_acquire(r->producer_pos);
386 }
387 
388 size_t ring__avail_data_size(const struct ring *r)
389 {
390 	unsigned long cons_pos, prod_pos;
391 
392 	cons_pos = ring__consumer_pos(r);
393 	prod_pos = ring__producer_pos(r);
394 	return prod_pos - cons_pos;
395 }
396 
397 size_t ring__size(const struct ring *r)
398 {
399 	return r->mask + 1;
400 }
401 
402 int ring__map_fd(const struct ring *r)
403 {
404 	return r->map_fd;
405 }
406 
407 int ring__consume_n(struct ring *r, size_t n)
408 {
409 	int64_t res;
410 
411 	res = ringbuf_process_ring(r, n);
412 	if (res < 0)
413 		return libbpf_err(res);
414 
415 	return res > INT_MAX ? INT_MAX : res;
416 }
417 
418 int ring__consume(struct ring *r)
419 {
420 	return ring__consume_n(r, INT_MAX);
421 }
422 
423 static void user_ringbuf_unmap_ring(struct user_ring_buffer *rb)
424 {
425 	if (rb->consumer_pos) {
426 		munmap(rb->consumer_pos, rb->page_size);
427 		rb->consumer_pos = NULL;
428 	}
429 	if (rb->producer_pos) {
430 		munmap(rb->producer_pos, rb->page_size + 2 * (rb->mask + 1));
431 		rb->producer_pos = NULL;
432 	}
433 }
434 
435 void user_ring_buffer__free(struct user_ring_buffer *rb)
436 {
437 	if (!rb)
438 		return;
439 
440 	user_ringbuf_unmap_ring(rb);
441 
442 	if (rb->epoll_fd >= 0)
443 		close(rb->epoll_fd);
444 
445 	free(rb);
446 }
447 
448 static int user_ringbuf_map(struct user_ring_buffer *rb, int map_fd)
449 {
450 	struct bpf_map_info info;
451 	__u32 len = sizeof(info);
452 	__u64 mmap_sz;
453 	void *tmp;
454 	struct epoll_event *rb_epoll;
455 	int err;
456 
457 	memset(&info, 0, sizeof(info));
458 
459 	err = bpf_map_get_info_by_fd(map_fd, &info, &len);
460 	if (err) {
461 		err = -errno;
462 		pr_warn("user ringbuf: failed to get map info for fd=%d: %s\n",
463 			map_fd, errstr(err));
464 		return err;
465 	}
466 
467 	if (info.type != BPF_MAP_TYPE_USER_RINGBUF) {
468 		pr_warn("user ringbuf: map fd=%d is not BPF_MAP_TYPE_USER_RINGBUF\n", map_fd);
469 		return -EINVAL;
470 	}
471 
472 	rb->map_fd = map_fd;
473 	rb->mask = info.max_entries - 1;
474 
475 	/* Map read-only consumer page */
476 	tmp = mmap(NULL, rb->page_size, PROT_READ, MAP_SHARED, map_fd, 0);
477 	if (tmp == MAP_FAILED) {
478 		err = -errno;
479 		pr_warn("user ringbuf: failed to mmap consumer page for map fd=%d: %s\n",
480 			map_fd, errstr(err));
481 		return err;
482 	}
483 	rb->consumer_pos = tmp;
484 
485 	/* Map read-write the producer page and data pages. We map the data
486 	 * region as twice the total size of the ring buffer to allow the
487 	 * simple reading and writing of samples that wrap around the end of
488 	 * the buffer.  See the kernel implementation for details.
489 	 */
490 	mmap_sz = rb->page_size + 2 * (__u64)info.max_entries;
491 	if (mmap_sz != (__u64)(size_t)mmap_sz) {
492 		pr_warn("user ringbuf: ring buf size (%u) is too big\n", info.max_entries);
493 		return -E2BIG;
494 	}
495 	tmp = mmap(NULL, (size_t)mmap_sz, PROT_READ | PROT_WRITE, MAP_SHARED,
496 		   map_fd, rb->page_size);
497 	if (tmp == MAP_FAILED) {
498 		err = -errno;
499 		pr_warn("user ringbuf: failed to mmap data pages for map fd=%d: %s\n",
500 			map_fd, errstr(err));
501 		return err;
502 	}
503 
504 	rb->producer_pos = tmp;
505 	rb->data = tmp + rb->page_size;
506 
507 	rb_epoll = &rb->event;
508 	rb_epoll->events = EPOLLOUT;
509 	if (epoll_ctl(rb->epoll_fd, EPOLL_CTL_ADD, map_fd, rb_epoll) < 0) {
510 		err = -errno;
511 		pr_warn("user ringbuf: failed to epoll add map fd=%d: %s\n", map_fd, errstr(err));
512 		return err;
513 	}
514 
515 	return 0;
516 }
517 
518 struct user_ring_buffer *
519 user_ring_buffer__new(int map_fd, const struct user_ring_buffer_opts *opts)
520 {
521 	struct user_ring_buffer *rb;
522 	int err;
523 
524 	if (!OPTS_VALID(opts, user_ring_buffer_opts))
525 		return errno = EINVAL, NULL;
526 
527 	rb = calloc(1, sizeof(*rb));
528 	if (!rb)
529 		return errno = ENOMEM, NULL;
530 
531 	rb->page_size = getpagesize();
532 
533 	rb->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
534 	if (rb->epoll_fd < 0) {
535 		err = -errno;
536 		pr_warn("user ringbuf: failed to create epoll instance: %s\n", errstr(err));
537 		goto err_out;
538 	}
539 
540 	err = user_ringbuf_map(rb, map_fd);
541 	if (err)
542 		goto err_out;
543 
544 	return rb;
545 
546 err_out:
547 	user_ring_buffer__free(rb);
548 	return errno = -err, NULL;
549 }
550 
551 static void user_ringbuf_commit(struct user_ring_buffer *rb, void *sample, bool discard)
552 {
553 	__u32 new_len;
554 	struct ringbuf_hdr *hdr;
555 	uintptr_t hdr_offset;
556 
557 	hdr_offset = rb->mask + 1 + (sample - rb->data) - BPF_RINGBUF_HDR_SZ;
558 	hdr = rb->data + (hdr_offset & rb->mask);
559 
560 	new_len = hdr->len & ~BPF_RINGBUF_BUSY_BIT;
561 	if (discard)
562 		new_len |= BPF_RINGBUF_DISCARD_BIT;
563 
564 	/* Synchronizes with smp_load_acquire() in __bpf_user_ringbuf_peek() in
565 	 * the kernel.
566 	 */
567 	__atomic_exchange_n(&hdr->len, new_len, __ATOMIC_ACQ_REL);
568 }
569 
570 void user_ring_buffer__discard(struct user_ring_buffer *rb, void *sample)
571 {
572 	user_ringbuf_commit(rb, sample, true);
573 }
574 
575 void user_ring_buffer__submit(struct user_ring_buffer *rb, void *sample)
576 {
577 	user_ringbuf_commit(rb, sample, false);
578 }
579 
580 void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
581 {
582 	__u32 avail_size, total_size, max_size;
583 	/* 64-bit to avoid overflow in case of extreme application behavior */
584 	__u64 cons_pos, prod_pos;
585 	struct ringbuf_hdr *hdr;
586 
587 	/* The top two bits are used as special flags */
588 	if (size & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT))
589 		return errno = E2BIG, NULL;
590 
591 	/* Synchronizes with smp_store_release() in __bpf_user_ringbuf_peek() in
592 	 * the kernel.
593 	 */
594 	cons_pos = smp_load_acquire(rb->consumer_pos);
595 	/* Synchronizes with smp_store_release() in user_ringbuf_commit() */
596 	prod_pos = smp_load_acquire(rb->producer_pos);
597 
598 	max_size = rb->mask + 1;
599 	avail_size = max_size - (prod_pos - cons_pos);
600 	/* Round up total size to a multiple of 8. */
601 	total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
602 
603 	if (total_size > max_size)
604 		return errno = E2BIG, NULL;
605 
606 	if (avail_size < total_size)
607 		return errno = ENOSPC, NULL;
608 
609 	hdr = rb->data + (prod_pos & rb->mask);
610 	hdr->len = size | BPF_RINGBUF_BUSY_BIT;
611 	hdr->pad = 0;
612 
613 	/* Synchronizes with smp_load_acquire() in __bpf_user_ringbuf_peek() in
614 	 * the kernel.
615 	 */
616 	smp_store_release(rb->producer_pos, prod_pos + total_size);
617 
618 	return (void *)rb->data + ((prod_pos + BPF_RINGBUF_HDR_SZ) & rb->mask);
619 }
620 
621 static __u64 ns_elapsed_timespec(const struct timespec *start, const struct timespec *end)
622 {
623 	__u64 start_ns, end_ns, ns_per_s = 1000000000;
624 
625 	start_ns = (__u64)start->tv_sec * ns_per_s + start->tv_nsec;
626 	end_ns = (__u64)end->tv_sec * ns_per_s + end->tv_nsec;
627 
628 	return end_ns - start_ns;
629 }
630 
631 void *user_ring_buffer__reserve_blocking(struct user_ring_buffer *rb, __u32 size, int timeout_ms)
632 {
633 	void *sample;
634 	int err, ms_remaining = timeout_ms;
635 	struct timespec start;
636 
637 	if (timeout_ms < 0 && timeout_ms != -1)
638 		return errno = EINVAL, NULL;
639 
640 	if (timeout_ms != -1) {
641 		err = clock_gettime(CLOCK_MONOTONIC, &start);
642 		if (err)
643 			return NULL;
644 	}
645 
646 	do {
647 		int cnt, ms_elapsed;
648 		struct timespec curr;
649 		__u64 ns_per_ms = 1000000;
650 
651 		sample = user_ring_buffer__reserve(rb, size);
652 		if (sample)
653 			return sample;
654 		else if (errno != ENOSPC)
655 			return NULL;
656 
657 		/* The kernel guarantees at least one event notification
658 		 * delivery whenever at least one sample is drained from the
659 		 * ring buffer in an invocation to bpf_ringbuf_drain(). Other
660 		 * additional events may be delivered at any time, but only one
661 		 * event is guaranteed per bpf_ringbuf_drain() invocation,
662 		 * provided that a sample is drained, and the BPF program did
663 		 * not pass BPF_RB_NO_WAKEUP to bpf_ringbuf_drain(). If
664 		 * BPF_RB_FORCE_WAKEUP is passed to bpf_ringbuf_drain(), a
665 		 * wakeup event will be delivered even if no samples are
666 		 * drained.
667 		 */
668 		cnt = epoll_wait(rb->epoll_fd, &rb->event, 1, ms_remaining);
669 		if (cnt < 0)
670 			return NULL;
671 
672 		if (timeout_ms == -1)
673 			continue;
674 
675 		err = clock_gettime(CLOCK_MONOTONIC, &curr);
676 		if (err)
677 			return NULL;
678 
679 		ms_elapsed = ns_elapsed_timespec(&start, &curr) / ns_per_ms;
680 		ms_remaining = timeout_ms - ms_elapsed;
681 	} while (ms_remaining > 0);
682 
683 	/* Try one more time to reserve a sample after the specified timeout has elapsed. */
684 	return user_ring_buffer__reserve(rb, size);
685 }
686