xref: /openbmc/linux/tools/lib/bpf/ringbuf.c (revision 6de298ff)
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 
25 struct ring {
26 	ring_buffer_sample_fn sample_cb;
27 	void *ctx;
28 	void *data;
29 	unsigned long *consumer_pos;
30 	unsigned long *producer_pos;
31 	unsigned long mask;
32 	int map_fd;
33 };
34 
35 struct ring_buffer {
36 	struct epoll_event *events;
37 	struct ring *rings;
38 	size_t page_size;
39 	int epoll_fd;
40 	int ring_cnt;
41 };
42 
43 struct user_ring_buffer {
44 	struct epoll_event event;
45 	unsigned long *consumer_pos;
46 	unsigned long *producer_pos;
47 	void *data;
48 	unsigned long mask;
49 	size_t page_size;
50 	int map_fd;
51 	int epoll_fd;
52 };
53 
54 /* 8-byte ring buffer header structure */
55 struct ringbuf_hdr {
56 	__u32 len;
57 	__u32 pad;
58 };
59 
60 static void ringbuf_unmap_ring(struct ring_buffer *rb, struct ring *r)
61 {
62 	if (r->consumer_pos) {
63 		munmap(r->consumer_pos, rb->page_size);
64 		r->consumer_pos = NULL;
65 	}
66 	if (r->producer_pos) {
67 		munmap(r->producer_pos, rb->page_size + 2 * (r->mask + 1));
68 		r->producer_pos = NULL;
69 	}
70 }
71 
72 /* Add extra RINGBUF maps to this ring buffer manager */
73 int ring_buffer__add(struct ring_buffer *rb, int map_fd,
74 		     ring_buffer_sample_fn sample_cb, void *ctx)
75 {
76 	struct bpf_map_info info;
77 	__u32 len = sizeof(info);
78 	struct epoll_event *e;
79 	struct ring *r;
80 	__u64 mmap_sz;
81 	void *tmp;
82 	int err;
83 
84 	memset(&info, 0, sizeof(info));
85 
86 	err = bpf_map_get_info_by_fd(map_fd, &info, &len);
87 	if (err) {
88 		err = -errno;
89 		pr_warn("ringbuf: failed to get map info for fd=%d: %d\n",
90 			map_fd, err);
91 		return libbpf_err(err);
92 	}
93 
94 	if (info.type != BPF_MAP_TYPE_RINGBUF) {
95 		pr_warn("ringbuf: map fd=%d is not BPF_MAP_TYPE_RINGBUF\n",
96 			map_fd);
97 		return libbpf_err(-EINVAL);
98 	}
99 
100 	tmp = libbpf_reallocarray(rb->rings, rb->ring_cnt + 1, sizeof(*rb->rings));
101 	if (!tmp)
102 		return libbpf_err(-ENOMEM);
103 	rb->rings = tmp;
104 
105 	tmp = libbpf_reallocarray(rb->events, rb->ring_cnt + 1, sizeof(*rb->events));
106 	if (!tmp)
107 		return libbpf_err(-ENOMEM);
108 	rb->events = tmp;
109 
110 	r = &rb->rings[rb->ring_cnt];
111 	memset(r, 0, sizeof(*r));
112 
113 	r->map_fd = map_fd;
114 	r->sample_cb = sample_cb;
115 	r->ctx = ctx;
116 	r->mask = info.max_entries - 1;
117 
118 	/* Map writable consumer page */
119 	tmp = mmap(NULL, rb->page_size, PROT_READ | PROT_WRITE, MAP_SHARED, map_fd, 0);
120 	if (tmp == MAP_FAILED) {
121 		err = -errno;
122 		pr_warn("ringbuf: failed to mmap consumer page for map fd=%d: %d\n",
123 			map_fd, err);
124 		return libbpf_err(err);
125 	}
126 	r->consumer_pos = tmp;
127 
128 	/* Map read-only producer page and data pages. We map twice as big
129 	 * data size to allow simple reading of samples that wrap around the
130 	 * end of a ring buffer. See kernel implementation for details.
131 	 */
132 	mmap_sz = rb->page_size + 2 * (__u64)info.max_entries;
133 	if (mmap_sz != (__u64)(size_t)mmap_sz) {
134 		pr_warn("ringbuf: ring buffer size (%u) is too big\n", info.max_entries);
135 		return libbpf_err(-E2BIG);
136 	}
137 	tmp = mmap(NULL, (size_t)mmap_sz, PROT_READ, MAP_SHARED, map_fd, rb->page_size);
138 	if (tmp == MAP_FAILED) {
139 		err = -errno;
140 		ringbuf_unmap_ring(rb, r);
141 		pr_warn("ringbuf: failed to mmap data pages for map fd=%d: %d\n",
142 			map_fd, err);
143 		return libbpf_err(err);
144 	}
145 	r->producer_pos = tmp;
146 	r->data = tmp + rb->page_size;
147 
148 	e = &rb->events[rb->ring_cnt];
149 	memset(e, 0, sizeof(*e));
150 
151 	e->events = EPOLLIN;
152 	e->data.fd = rb->ring_cnt;
153 	if (epoll_ctl(rb->epoll_fd, EPOLL_CTL_ADD, map_fd, e) < 0) {
154 		err = -errno;
155 		ringbuf_unmap_ring(rb, r);
156 		pr_warn("ringbuf: failed to epoll add map fd=%d: %d\n",
157 			map_fd, err);
158 		return libbpf_err(err);
159 	}
160 
161 	rb->ring_cnt++;
162 	return 0;
163 }
164 
165 void ring_buffer__free(struct ring_buffer *rb)
166 {
167 	int i;
168 
169 	if (!rb)
170 		return;
171 
172 	for (i = 0; i < rb->ring_cnt; ++i)
173 		ringbuf_unmap_ring(rb, &rb->rings[i]);
174 	if (rb->epoll_fd >= 0)
175 		close(rb->epoll_fd);
176 
177 	free(rb->events);
178 	free(rb->rings);
179 	free(rb);
180 }
181 
182 struct ring_buffer *
183 ring_buffer__new(int map_fd, ring_buffer_sample_fn sample_cb, void *ctx,
184 		 const struct ring_buffer_opts *opts)
185 {
186 	struct ring_buffer *rb;
187 	int err;
188 
189 	if (!OPTS_VALID(opts, ring_buffer_opts))
190 		return errno = EINVAL, NULL;
191 
192 	rb = calloc(1, sizeof(*rb));
193 	if (!rb)
194 		return errno = ENOMEM, NULL;
195 
196 	rb->page_size = getpagesize();
197 
198 	rb->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
199 	if (rb->epoll_fd < 0) {
200 		err = -errno;
201 		pr_warn("ringbuf: failed to create epoll instance: %d\n", err);
202 		goto err_out;
203 	}
204 
205 	err = ring_buffer__add(rb, map_fd, sample_cb, ctx);
206 	if (err)
207 		goto err_out;
208 
209 	return rb;
210 
211 err_out:
212 	ring_buffer__free(rb);
213 	return errno = -err, NULL;
214 }
215 
216 static inline int roundup_len(__u32 len)
217 {
218 	/* clear out top 2 bits (discard and busy, if set) */
219 	len <<= 2;
220 	len >>= 2;
221 	/* add length prefix */
222 	len += BPF_RINGBUF_HDR_SZ;
223 	/* round up to 8 byte alignment */
224 	return (len + 7) / 8 * 8;
225 }
226 
227 static int64_t ringbuf_process_ring(struct ring *r)
228 {
229 	int *len_ptr, len, err;
230 	/* 64-bit to avoid overflow in case of extreme application behavior */
231 	int64_t cnt = 0;
232 	unsigned long cons_pos, prod_pos;
233 	bool got_new_data;
234 	void *sample;
235 
236 	cons_pos = smp_load_acquire(r->consumer_pos);
237 	do {
238 		got_new_data = false;
239 		prod_pos = smp_load_acquire(r->producer_pos);
240 		while (cons_pos < prod_pos) {
241 			len_ptr = r->data + (cons_pos & r->mask);
242 			len = smp_load_acquire(len_ptr);
243 
244 			/* sample not committed yet, bail out for now */
245 			if (len & BPF_RINGBUF_BUSY_BIT)
246 				goto done;
247 
248 			got_new_data = true;
249 			cons_pos += roundup_len(len);
250 
251 			if ((len & BPF_RINGBUF_DISCARD_BIT) == 0) {
252 				sample = (void *)len_ptr + BPF_RINGBUF_HDR_SZ;
253 				err = r->sample_cb(r->ctx, sample, len);
254 				if (err < 0) {
255 					/* update consumer pos and bail out */
256 					smp_store_release(r->consumer_pos,
257 							  cons_pos);
258 					return err;
259 				}
260 				cnt++;
261 			}
262 
263 			smp_store_release(r->consumer_pos, cons_pos);
264 		}
265 	} while (got_new_data);
266 done:
267 	return cnt;
268 }
269 
270 /* Consume available ring buffer(s) data without event polling.
271  * Returns number of records consumed across all registered ring buffers (or
272  * INT_MAX, whichever is less), or negative number if any of the callbacks
273  * return error.
274  */
275 int ring_buffer__consume(struct ring_buffer *rb)
276 {
277 	int64_t err, res = 0;
278 	int i;
279 
280 	for (i = 0; i < rb->ring_cnt; i++) {
281 		struct ring *ring = &rb->rings[i];
282 
283 		err = ringbuf_process_ring(ring);
284 		if (err < 0)
285 			return libbpf_err(err);
286 		res += err;
287 	}
288 	if (res > INT_MAX)
289 		return INT_MAX;
290 	return res;
291 }
292 
293 /* Poll for available data and consume records, if any are available.
294  * Returns number of records consumed (or INT_MAX, whichever is less), or
295  * negative number, if any of the registered callbacks returned error.
296  */
297 int ring_buffer__poll(struct ring_buffer *rb, int timeout_ms)
298 {
299 	int i, cnt;
300 	int64_t err, res = 0;
301 
302 	cnt = epoll_wait(rb->epoll_fd, rb->events, rb->ring_cnt, timeout_ms);
303 	if (cnt < 0)
304 		return libbpf_err(-errno);
305 
306 	for (i = 0; i < cnt; i++) {
307 		__u32 ring_id = rb->events[i].data.fd;
308 		struct ring *ring = &rb->rings[ring_id];
309 
310 		err = ringbuf_process_ring(ring);
311 		if (err < 0)
312 			return libbpf_err(err);
313 		res += err;
314 	}
315 	if (res > INT_MAX)
316 		return INT_MAX;
317 	return res;
318 }
319 
320 /* Get an fd that can be used to sleep until data is available in the ring(s) */
321 int ring_buffer__epoll_fd(const struct ring_buffer *rb)
322 {
323 	return rb->epoll_fd;
324 }
325 
326 static void user_ringbuf_unmap_ring(struct user_ring_buffer *rb)
327 {
328 	if (rb->consumer_pos) {
329 		munmap(rb->consumer_pos, rb->page_size);
330 		rb->consumer_pos = NULL;
331 	}
332 	if (rb->producer_pos) {
333 		munmap(rb->producer_pos, rb->page_size + 2 * (rb->mask + 1));
334 		rb->producer_pos = NULL;
335 	}
336 }
337 
338 void user_ring_buffer__free(struct user_ring_buffer *rb)
339 {
340 	if (!rb)
341 		return;
342 
343 	user_ringbuf_unmap_ring(rb);
344 
345 	if (rb->epoll_fd >= 0)
346 		close(rb->epoll_fd);
347 
348 	free(rb);
349 }
350 
351 static int user_ringbuf_map(struct user_ring_buffer *rb, int map_fd)
352 {
353 	struct bpf_map_info info;
354 	__u32 len = sizeof(info);
355 	__u64 mmap_sz;
356 	void *tmp;
357 	struct epoll_event *rb_epoll;
358 	int err;
359 
360 	memset(&info, 0, sizeof(info));
361 
362 	err = bpf_map_get_info_by_fd(map_fd, &info, &len);
363 	if (err) {
364 		err = -errno;
365 		pr_warn("user ringbuf: failed to get map info for fd=%d: %d\n", map_fd, err);
366 		return err;
367 	}
368 
369 	if (info.type != BPF_MAP_TYPE_USER_RINGBUF) {
370 		pr_warn("user ringbuf: map fd=%d is not BPF_MAP_TYPE_USER_RINGBUF\n", map_fd);
371 		return -EINVAL;
372 	}
373 
374 	rb->map_fd = map_fd;
375 	rb->mask = info.max_entries - 1;
376 
377 	/* Map read-only consumer page */
378 	tmp = mmap(NULL, rb->page_size, PROT_READ, MAP_SHARED, map_fd, 0);
379 	if (tmp == MAP_FAILED) {
380 		err = -errno;
381 		pr_warn("user ringbuf: failed to mmap consumer page for map fd=%d: %d\n",
382 			map_fd, err);
383 		return err;
384 	}
385 	rb->consumer_pos = tmp;
386 
387 	/* Map read-write the producer page and data pages. We map the data
388 	 * region as twice the total size of the ring buffer to allow the
389 	 * simple reading and writing of samples that wrap around the end of
390 	 * the buffer.  See the kernel implementation for details.
391 	 */
392 	mmap_sz = rb->page_size + 2 * (__u64)info.max_entries;
393 	if (mmap_sz != (__u64)(size_t)mmap_sz) {
394 		pr_warn("user ringbuf: ring buf size (%u) is too big\n", info.max_entries);
395 		return -E2BIG;
396 	}
397 	tmp = mmap(NULL, (size_t)mmap_sz, PROT_READ | PROT_WRITE, MAP_SHARED,
398 		   map_fd, rb->page_size);
399 	if (tmp == MAP_FAILED) {
400 		err = -errno;
401 		pr_warn("user ringbuf: failed to mmap data pages for map fd=%d: %d\n",
402 			map_fd, err);
403 		return err;
404 	}
405 
406 	rb->producer_pos = tmp;
407 	rb->data = tmp + rb->page_size;
408 
409 	rb_epoll = &rb->event;
410 	rb_epoll->events = EPOLLOUT;
411 	if (epoll_ctl(rb->epoll_fd, EPOLL_CTL_ADD, map_fd, rb_epoll) < 0) {
412 		err = -errno;
413 		pr_warn("user ringbuf: failed to epoll add map fd=%d: %d\n", map_fd, err);
414 		return err;
415 	}
416 
417 	return 0;
418 }
419 
420 struct user_ring_buffer *
421 user_ring_buffer__new(int map_fd, const struct user_ring_buffer_opts *opts)
422 {
423 	struct user_ring_buffer *rb;
424 	int err;
425 
426 	if (!OPTS_VALID(opts, user_ring_buffer_opts))
427 		return errno = EINVAL, NULL;
428 
429 	rb = calloc(1, sizeof(*rb));
430 	if (!rb)
431 		return errno = ENOMEM, NULL;
432 
433 	rb->page_size = getpagesize();
434 
435 	rb->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
436 	if (rb->epoll_fd < 0) {
437 		err = -errno;
438 		pr_warn("user ringbuf: failed to create epoll instance: %d\n", err);
439 		goto err_out;
440 	}
441 
442 	err = user_ringbuf_map(rb, map_fd);
443 	if (err)
444 		goto err_out;
445 
446 	return rb;
447 
448 err_out:
449 	user_ring_buffer__free(rb);
450 	return errno = -err, NULL;
451 }
452 
453 static void user_ringbuf_commit(struct user_ring_buffer *rb, void *sample, bool discard)
454 {
455 	__u32 new_len;
456 	struct ringbuf_hdr *hdr;
457 	uintptr_t hdr_offset;
458 
459 	hdr_offset = rb->mask + 1 + (sample - rb->data) - BPF_RINGBUF_HDR_SZ;
460 	hdr = rb->data + (hdr_offset & rb->mask);
461 
462 	new_len = hdr->len & ~BPF_RINGBUF_BUSY_BIT;
463 	if (discard)
464 		new_len |= BPF_RINGBUF_DISCARD_BIT;
465 
466 	/* Synchronizes with smp_load_acquire() in __bpf_user_ringbuf_peek() in
467 	 * the kernel.
468 	 */
469 	__atomic_exchange_n(&hdr->len, new_len, __ATOMIC_ACQ_REL);
470 }
471 
472 void user_ring_buffer__discard(struct user_ring_buffer *rb, void *sample)
473 {
474 	user_ringbuf_commit(rb, sample, true);
475 }
476 
477 void user_ring_buffer__submit(struct user_ring_buffer *rb, void *sample)
478 {
479 	user_ringbuf_commit(rb, sample, false);
480 }
481 
482 void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
483 {
484 	__u32 avail_size, total_size, max_size;
485 	/* 64-bit to avoid overflow in case of extreme application behavior */
486 	__u64 cons_pos, prod_pos;
487 	struct ringbuf_hdr *hdr;
488 
489 	/* The top two bits are used as special flags */
490 	if (size & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT))
491 		return errno = E2BIG, NULL;
492 
493 	/* Synchronizes with smp_store_release() in __bpf_user_ringbuf_peek() in
494 	 * the kernel.
495 	 */
496 	cons_pos = smp_load_acquire(rb->consumer_pos);
497 	/* Synchronizes with smp_store_release() in user_ringbuf_commit() */
498 	prod_pos = smp_load_acquire(rb->producer_pos);
499 
500 	max_size = rb->mask + 1;
501 	avail_size = max_size - (prod_pos - cons_pos);
502 	/* Round up total size to a multiple of 8. */
503 	total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
504 
505 	if (total_size > max_size)
506 		return errno = E2BIG, NULL;
507 
508 	if (avail_size < total_size)
509 		return errno = ENOSPC, NULL;
510 
511 	hdr = rb->data + (prod_pos & rb->mask);
512 	hdr->len = size | BPF_RINGBUF_BUSY_BIT;
513 	hdr->pad = 0;
514 
515 	/* Synchronizes with smp_load_acquire() in __bpf_user_ringbuf_peek() in
516 	 * the kernel.
517 	 */
518 	smp_store_release(rb->producer_pos, prod_pos + total_size);
519 
520 	return (void *)rb->data + ((prod_pos + BPF_RINGBUF_HDR_SZ) & rb->mask);
521 }
522 
523 static __u64 ns_elapsed_timespec(const struct timespec *start, const struct timespec *end)
524 {
525 	__u64 start_ns, end_ns, ns_per_s = 1000000000;
526 
527 	start_ns = (__u64)start->tv_sec * ns_per_s + start->tv_nsec;
528 	end_ns = (__u64)end->tv_sec * ns_per_s + end->tv_nsec;
529 
530 	return end_ns - start_ns;
531 }
532 
533 void *user_ring_buffer__reserve_blocking(struct user_ring_buffer *rb, __u32 size, int timeout_ms)
534 {
535 	void *sample;
536 	int err, ms_remaining = timeout_ms;
537 	struct timespec start;
538 
539 	if (timeout_ms < 0 && timeout_ms != -1)
540 		return errno = EINVAL, NULL;
541 
542 	if (timeout_ms != -1) {
543 		err = clock_gettime(CLOCK_MONOTONIC, &start);
544 		if (err)
545 			return NULL;
546 	}
547 
548 	do {
549 		int cnt, ms_elapsed;
550 		struct timespec curr;
551 		__u64 ns_per_ms = 1000000;
552 
553 		sample = user_ring_buffer__reserve(rb, size);
554 		if (sample)
555 			return sample;
556 		else if (errno != ENOSPC)
557 			return NULL;
558 
559 		/* The kernel guarantees at least one event notification
560 		 * delivery whenever at least one sample is drained from the
561 		 * ring buffer in an invocation to bpf_ringbuf_drain(). Other
562 		 * additional events may be delivered at any time, but only one
563 		 * event is guaranteed per bpf_ringbuf_drain() invocation,
564 		 * provided that a sample is drained, and the BPF program did
565 		 * not pass BPF_RB_NO_WAKEUP to bpf_ringbuf_drain(). If
566 		 * BPF_RB_FORCE_WAKEUP is passed to bpf_ringbuf_drain(), a
567 		 * wakeup event will be delivered even if no samples are
568 		 * drained.
569 		 */
570 		cnt = epoll_wait(rb->epoll_fd, &rb->event, 1, ms_remaining);
571 		if (cnt < 0)
572 			return NULL;
573 
574 		if (timeout_ms == -1)
575 			continue;
576 
577 		err = clock_gettime(CLOCK_MONOTONIC, &curr);
578 		if (err)
579 			return NULL;
580 
581 		ms_elapsed = ns_elapsed_timespec(&start, &curr) / ns_per_ms;
582 		ms_remaining = timeout_ms - ms_elapsed;
583 	} while (ms_remaining > 0);
584 
585 	/* Try one more time to reserve a sample after the specified timeout has elapsed. */
586 	return user_ring_buffer__reserve(rb, size);
587 }
588