Loading...
1#include <linux/bpf.h>
2#include <linux/btf.h>
3#include <linux/err.h>
4#include <linux/irq_work.h>
5#include <linux/slab.h>
6#include <linux/filter.h>
7#include <linux/mm.h>
8#include <linux/vmalloc.h>
9#include <linux/wait.h>
10#include <linux/poll.h>
11#include <linux/kmemleak.h>
12#include <uapi/linux/btf.h>
13#include <linux/btf_ids.h>
14
15#define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
16
17/* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
18#define RINGBUF_PGOFF \
19 (offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
20/* consumer page and producer page */
21#define RINGBUF_POS_PAGES 2
22
23#define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
24
25/* Maximum size of ring buffer area is limited by 32-bit page offset within
26 * record header, counted in pages. Reserve 8 bits for extensibility, and take
27 * into account few extra pages for consumer/producer pages and
28 * non-mmap()'able parts. This gives 64GB limit, which seems plenty for single
29 * ring buffer.
30 */
31#define RINGBUF_MAX_DATA_SZ \
32 (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
33
34struct bpf_ringbuf {
35 wait_queue_head_t waitq;
36 struct irq_work work;
37 u64 mask;
38 struct page **pages;
39 int nr_pages;
40 spinlock_t spinlock ____cacheline_aligned_in_smp;
41 /* For user-space producer ring buffers, an atomic_t busy bit is used
42 * to synchronize access to the ring buffers in the kernel, rather than
43 * the spinlock that is used for kernel-producer ring buffers. This is
44 * done because the ring buffer must hold a lock across a BPF program's
45 * callback:
46 *
47 * __bpf_user_ringbuf_peek() // lock acquired
48 * -> program callback_fn()
49 * -> __bpf_user_ringbuf_sample_release() // lock released
50 *
51 * It is unsafe and incorrect to hold an IRQ spinlock across what could
52 * be a long execution window, so we instead simply disallow concurrent
53 * access to the ring buffer by kernel consumers, and return -EBUSY from
54 * __bpf_user_ringbuf_peek() if the busy bit is held by another task.
55 */
56 atomic_t busy ____cacheline_aligned_in_smp;
57 /* Consumer and producer counters are put into separate pages to
58 * allow each position to be mapped with different permissions.
59 * This prevents a user-space application from modifying the
60 * position and ruining in-kernel tracking. The permissions of the
61 * pages depend on who is producing samples: user-space or the
62 * kernel.
63 *
64 * Kernel-producer
65 * ---------------
66 * The producer position and data pages are mapped as r/o in
67 * userspace. For this approach, bits in the header of samples are
68 * used to signal to user-space, and to other producers, whether a
69 * sample is currently being written.
70 *
71 * User-space producer
72 * -------------------
73 * Only the page containing the consumer position is mapped r/o in
74 * user-space. User-space producers also use bits of the header to
75 * communicate to the kernel, but the kernel must carefully check and
76 * validate each sample to ensure that they're correctly formatted, and
77 * fully contained within the ring buffer.
78 */
79 unsigned long consumer_pos __aligned(PAGE_SIZE);
80 unsigned long producer_pos __aligned(PAGE_SIZE);
81 char data[] __aligned(PAGE_SIZE);
82};
83
84struct bpf_ringbuf_map {
85 struct bpf_map map;
86 struct bpf_ringbuf *rb;
87};
88
89/* 8-byte ring buffer record header structure */
90struct bpf_ringbuf_hdr {
91 u32 len;
92 u32 pg_off;
93};
94
95static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
96{
97 const gfp_t flags = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL |
98 __GFP_NOWARN | __GFP_ZERO;
99 int nr_meta_pages = RINGBUF_PGOFF + RINGBUF_POS_PAGES;
100 int nr_data_pages = data_sz >> PAGE_SHIFT;
101 int nr_pages = nr_meta_pages + nr_data_pages;
102 struct page **pages, *page;
103 struct bpf_ringbuf *rb;
104 size_t array_size;
105 int i;
106
107 /* Each data page is mapped twice to allow "virtual"
108 * continuous read of samples wrapping around the end of ring
109 * buffer area:
110 * ------------------------------------------------------
111 * | meta pages | real data pages | same data pages |
112 * ------------------------------------------------------
113 * | | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
114 * ------------------------------------------------------
115 * | | TA DA | TA DA |
116 * ------------------------------------------------------
117 * ^^^^^^^
118 * |
119 * Here, no need to worry about special handling of wrapped-around
120 * data due to double-mapped data pages. This works both in kernel and
121 * when mmap()'ed in user-space, simplifying both kernel and
122 * user-space implementations significantly.
123 */
124 array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
125 pages = bpf_map_area_alloc(array_size, numa_node);
126 if (!pages)
127 return NULL;
128
129 for (i = 0; i < nr_pages; i++) {
130 page = alloc_pages_node(numa_node, flags, 0);
131 if (!page) {
132 nr_pages = i;
133 goto err_free_pages;
134 }
135 pages[i] = page;
136 if (i >= nr_meta_pages)
137 pages[nr_data_pages + i] = page;
138 }
139
140 rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
141 VM_MAP | VM_USERMAP, PAGE_KERNEL);
142 if (rb) {
143 kmemleak_not_leak(pages);
144 rb->pages = pages;
145 rb->nr_pages = nr_pages;
146 return rb;
147 }
148
149err_free_pages:
150 for (i = 0; i < nr_pages; i++)
151 __free_page(pages[i]);
152 bpf_map_area_free(pages);
153 return NULL;
154}
155
156static void bpf_ringbuf_notify(struct irq_work *work)
157{
158 struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
159
160 wake_up_all(&rb->waitq);
161}
162
163static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
164{
165 struct bpf_ringbuf *rb;
166
167 rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
168 if (!rb)
169 return NULL;
170
171 spin_lock_init(&rb->spinlock);
172 atomic_set(&rb->busy, 0);
173 init_waitqueue_head(&rb->waitq);
174 init_irq_work(&rb->work, bpf_ringbuf_notify);
175
176 rb->mask = data_sz - 1;
177 rb->consumer_pos = 0;
178 rb->producer_pos = 0;
179
180 return rb;
181}
182
183static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
184{
185 struct bpf_ringbuf_map *rb_map;
186
187 if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
188 return ERR_PTR(-EINVAL);
189
190 if (attr->key_size || attr->value_size ||
191 !is_power_of_2(attr->max_entries) ||
192 !PAGE_ALIGNED(attr->max_entries))
193 return ERR_PTR(-EINVAL);
194
195#ifdef CONFIG_64BIT
196 /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
197 if (attr->max_entries > RINGBUF_MAX_DATA_SZ)
198 return ERR_PTR(-E2BIG);
199#endif
200
201 rb_map = bpf_map_area_alloc(sizeof(*rb_map), NUMA_NO_NODE);
202 if (!rb_map)
203 return ERR_PTR(-ENOMEM);
204
205 bpf_map_init_from_attr(&rb_map->map, attr);
206
207 rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
208 if (!rb_map->rb) {
209 bpf_map_area_free(rb_map);
210 return ERR_PTR(-ENOMEM);
211 }
212
213 return &rb_map->map;
214}
215
216static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
217{
218 /* copy pages pointer and nr_pages to local variable, as we are going
219 * to unmap rb itself with vunmap() below
220 */
221 struct page **pages = rb->pages;
222 int i, nr_pages = rb->nr_pages;
223
224 vunmap(rb);
225 for (i = 0; i < nr_pages; i++)
226 __free_page(pages[i]);
227 bpf_map_area_free(pages);
228}
229
230static void ringbuf_map_free(struct bpf_map *map)
231{
232 struct bpf_ringbuf_map *rb_map;
233
234 rb_map = container_of(map, struct bpf_ringbuf_map, map);
235 bpf_ringbuf_free(rb_map->rb);
236 bpf_map_area_free(rb_map);
237}
238
239static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
240{
241 return ERR_PTR(-ENOTSUPP);
242}
243
244static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
245 u64 flags)
246{
247 return -ENOTSUPP;
248}
249
250static int ringbuf_map_delete_elem(struct bpf_map *map, void *key)
251{
252 return -ENOTSUPP;
253}
254
255static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
256 void *next_key)
257{
258 return -ENOTSUPP;
259}
260
261static int ringbuf_map_mmap_kern(struct bpf_map *map, struct vm_area_struct *vma)
262{
263 struct bpf_ringbuf_map *rb_map;
264
265 rb_map = container_of(map, struct bpf_ringbuf_map, map);
266
267 if (vma->vm_flags & VM_WRITE) {
268 /* allow writable mapping for the consumer_pos only */
269 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
270 return -EPERM;
271 } else {
272 vma->vm_flags &= ~VM_MAYWRITE;
273 }
274 /* remap_vmalloc_range() checks size and offset constraints */
275 return remap_vmalloc_range(vma, rb_map->rb,
276 vma->vm_pgoff + RINGBUF_PGOFF);
277}
278
279static int ringbuf_map_mmap_user(struct bpf_map *map, struct vm_area_struct *vma)
280{
281 struct bpf_ringbuf_map *rb_map;
282
283 rb_map = container_of(map, struct bpf_ringbuf_map, map);
284
285 if (vma->vm_flags & VM_WRITE) {
286 if (vma->vm_pgoff == 0)
287 /* Disallow writable mappings to the consumer pointer,
288 * and allow writable mappings to both the producer
289 * position, and the ring buffer data itself.
290 */
291 return -EPERM;
292 } else {
293 vma->vm_flags &= ~VM_MAYWRITE;
294 }
295 /* remap_vmalloc_range() checks size and offset constraints */
296 return remap_vmalloc_range(vma, rb_map->rb, vma->vm_pgoff + RINGBUF_PGOFF);
297}
298
299static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
300{
301 unsigned long cons_pos, prod_pos;
302
303 cons_pos = smp_load_acquire(&rb->consumer_pos);
304 prod_pos = smp_load_acquire(&rb->producer_pos);
305 return prod_pos - cons_pos;
306}
307
308static u32 ringbuf_total_data_sz(const struct bpf_ringbuf *rb)
309{
310 return rb->mask + 1;
311}
312
313static __poll_t ringbuf_map_poll_kern(struct bpf_map *map, struct file *filp,
314 struct poll_table_struct *pts)
315{
316 struct bpf_ringbuf_map *rb_map;
317
318 rb_map = container_of(map, struct bpf_ringbuf_map, map);
319 poll_wait(filp, &rb_map->rb->waitq, pts);
320
321 if (ringbuf_avail_data_sz(rb_map->rb))
322 return EPOLLIN | EPOLLRDNORM;
323 return 0;
324}
325
326static __poll_t ringbuf_map_poll_user(struct bpf_map *map, struct file *filp,
327 struct poll_table_struct *pts)
328{
329 struct bpf_ringbuf_map *rb_map;
330
331 rb_map = container_of(map, struct bpf_ringbuf_map, map);
332 poll_wait(filp, &rb_map->rb->waitq, pts);
333
334 if (ringbuf_avail_data_sz(rb_map->rb) < ringbuf_total_data_sz(rb_map->rb))
335 return EPOLLOUT | EPOLLWRNORM;
336 return 0;
337}
338
339BTF_ID_LIST_SINGLE(ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
340const struct bpf_map_ops ringbuf_map_ops = {
341 .map_meta_equal = bpf_map_meta_equal,
342 .map_alloc = ringbuf_map_alloc,
343 .map_free = ringbuf_map_free,
344 .map_mmap = ringbuf_map_mmap_kern,
345 .map_poll = ringbuf_map_poll_kern,
346 .map_lookup_elem = ringbuf_map_lookup_elem,
347 .map_update_elem = ringbuf_map_update_elem,
348 .map_delete_elem = ringbuf_map_delete_elem,
349 .map_get_next_key = ringbuf_map_get_next_key,
350 .map_btf_id = &ringbuf_map_btf_ids[0],
351};
352
353BTF_ID_LIST_SINGLE(user_ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
354const struct bpf_map_ops user_ringbuf_map_ops = {
355 .map_meta_equal = bpf_map_meta_equal,
356 .map_alloc = ringbuf_map_alloc,
357 .map_free = ringbuf_map_free,
358 .map_mmap = ringbuf_map_mmap_user,
359 .map_poll = ringbuf_map_poll_user,
360 .map_lookup_elem = ringbuf_map_lookup_elem,
361 .map_update_elem = ringbuf_map_update_elem,
362 .map_delete_elem = ringbuf_map_delete_elem,
363 .map_get_next_key = ringbuf_map_get_next_key,
364 .map_btf_id = &user_ringbuf_map_btf_ids[0],
365};
366
367/* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
368 * calculate offset from record metadata to ring buffer in pages, rounded
369 * down. This page offset is stored as part of record metadata and allows to
370 * restore struct bpf_ringbuf * from record pointer. This page offset is
371 * stored at offset 4 of record metadata header.
372 */
373static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
374 struct bpf_ringbuf_hdr *hdr)
375{
376 return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
377}
378
379/* Given pointer to ring buffer record header, restore pointer to struct
380 * bpf_ringbuf itself by using page offset stored at offset 4
381 */
382static struct bpf_ringbuf *
383bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
384{
385 unsigned long addr = (unsigned long)(void *)hdr;
386 unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
387
388 return (void*)((addr & PAGE_MASK) - off);
389}
390
391static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
392{
393 unsigned long cons_pos, prod_pos, new_prod_pos, flags;
394 u32 len, pg_off;
395 struct bpf_ringbuf_hdr *hdr;
396
397 if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
398 return NULL;
399
400 len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
401 if (len > ringbuf_total_data_sz(rb))
402 return NULL;
403
404 cons_pos = smp_load_acquire(&rb->consumer_pos);
405
406 if (in_nmi()) {
407 if (!spin_trylock_irqsave(&rb->spinlock, flags))
408 return NULL;
409 } else {
410 spin_lock_irqsave(&rb->spinlock, flags);
411 }
412
413 prod_pos = rb->producer_pos;
414 new_prod_pos = prod_pos + len;
415
416 /* check for out of ringbuf space by ensuring producer position
417 * doesn't advance more than (ringbuf_size - 1) ahead
418 */
419 if (new_prod_pos - cons_pos > rb->mask) {
420 spin_unlock_irqrestore(&rb->spinlock, flags);
421 return NULL;
422 }
423
424 hdr = (void *)rb->data + (prod_pos & rb->mask);
425 pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
426 hdr->len = size | BPF_RINGBUF_BUSY_BIT;
427 hdr->pg_off = pg_off;
428
429 /* pairs with consumer's smp_load_acquire() */
430 smp_store_release(&rb->producer_pos, new_prod_pos);
431
432 spin_unlock_irqrestore(&rb->spinlock, flags);
433
434 return (void *)hdr + BPF_RINGBUF_HDR_SZ;
435}
436
437BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
438{
439 struct bpf_ringbuf_map *rb_map;
440
441 if (unlikely(flags))
442 return 0;
443
444 rb_map = container_of(map, struct bpf_ringbuf_map, map);
445 return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
446}
447
448const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
449 .func = bpf_ringbuf_reserve,
450 .ret_type = RET_PTR_TO_RINGBUF_MEM_OR_NULL,
451 .arg1_type = ARG_CONST_MAP_PTR,
452 .arg2_type = ARG_CONST_ALLOC_SIZE_OR_ZERO,
453 .arg3_type = ARG_ANYTHING,
454};
455
456static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
457{
458 unsigned long rec_pos, cons_pos;
459 struct bpf_ringbuf_hdr *hdr;
460 struct bpf_ringbuf *rb;
461 u32 new_len;
462
463 hdr = sample - BPF_RINGBUF_HDR_SZ;
464 rb = bpf_ringbuf_restore_from_rec(hdr);
465 new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
466 if (discard)
467 new_len |= BPF_RINGBUF_DISCARD_BIT;
468
469 /* update record header with correct final size prefix */
470 xchg(&hdr->len, new_len);
471
472 /* if consumer caught up and is waiting for our record, notify about
473 * new data availability
474 */
475 rec_pos = (void *)hdr - (void *)rb->data;
476 cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
477
478 if (flags & BPF_RB_FORCE_WAKEUP)
479 irq_work_queue(&rb->work);
480 else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
481 irq_work_queue(&rb->work);
482}
483
484BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
485{
486 bpf_ringbuf_commit(sample, flags, false /* discard */);
487 return 0;
488}
489
490const struct bpf_func_proto bpf_ringbuf_submit_proto = {
491 .func = bpf_ringbuf_submit,
492 .ret_type = RET_VOID,
493 .arg1_type = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
494 .arg2_type = ARG_ANYTHING,
495};
496
497BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
498{
499 bpf_ringbuf_commit(sample, flags, true /* discard */);
500 return 0;
501}
502
503const struct bpf_func_proto bpf_ringbuf_discard_proto = {
504 .func = bpf_ringbuf_discard,
505 .ret_type = RET_VOID,
506 .arg1_type = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
507 .arg2_type = ARG_ANYTHING,
508};
509
510BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
511 u64, flags)
512{
513 struct bpf_ringbuf_map *rb_map;
514 void *rec;
515
516 if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
517 return -EINVAL;
518
519 rb_map = container_of(map, struct bpf_ringbuf_map, map);
520 rec = __bpf_ringbuf_reserve(rb_map->rb, size);
521 if (!rec)
522 return -EAGAIN;
523
524 memcpy(rec, data, size);
525 bpf_ringbuf_commit(rec, flags, false /* discard */);
526 return 0;
527}
528
529const struct bpf_func_proto bpf_ringbuf_output_proto = {
530 .func = bpf_ringbuf_output,
531 .ret_type = RET_INTEGER,
532 .arg1_type = ARG_CONST_MAP_PTR,
533 .arg2_type = ARG_PTR_TO_MEM | MEM_RDONLY,
534 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
535 .arg4_type = ARG_ANYTHING,
536};
537
538BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
539{
540 struct bpf_ringbuf *rb;
541
542 rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
543
544 switch (flags) {
545 case BPF_RB_AVAIL_DATA:
546 return ringbuf_avail_data_sz(rb);
547 case BPF_RB_RING_SIZE:
548 return ringbuf_total_data_sz(rb);
549 case BPF_RB_CONS_POS:
550 return smp_load_acquire(&rb->consumer_pos);
551 case BPF_RB_PROD_POS:
552 return smp_load_acquire(&rb->producer_pos);
553 default:
554 return 0;
555 }
556}
557
558const struct bpf_func_proto bpf_ringbuf_query_proto = {
559 .func = bpf_ringbuf_query,
560 .ret_type = RET_INTEGER,
561 .arg1_type = ARG_CONST_MAP_PTR,
562 .arg2_type = ARG_ANYTHING,
563};
564
565BPF_CALL_4(bpf_ringbuf_reserve_dynptr, struct bpf_map *, map, u32, size, u64, flags,
566 struct bpf_dynptr_kern *, ptr)
567{
568 struct bpf_ringbuf_map *rb_map;
569 void *sample;
570 int err;
571
572 if (unlikely(flags)) {
573 bpf_dynptr_set_null(ptr);
574 return -EINVAL;
575 }
576
577 err = bpf_dynptr_check_size(size);
578 if (err) {
579 bpf_dynptr_set_null(ptr);
580 return err;
581 }
582
583 rb_map = container_of(map, struct bpf_ringbuf_map, map);
584
585 sample = __bpf_ringbuf_reserve(rb_map->rb, size);
586 if (!sample) {
587 bpf_dynptr_set_null(ptr);
588 return -EINVAL;
589 }
590
591 bpf_dynptr_init(ptr, sample, BPF_DYNPTR_TYPE_RINGBUF, 0, size);
592
593 return 0;
594}
595
596const struct bpf_func_proto bpf_ringbuf_reserve_dynptr_proto = {
597 .func = bpf_ringbuf_reserve_dynptr,
598 .ret_type = RET_INTEGER,
599 .arg1_type = ARG_CONST_MAP_PTR,
600 .arg2_type = ARG_ANYTHING,
601 .arg3_type = ARG_ANYTHING,
602 .arg4_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | MEM_UNINIT,
603};
604
605BPF_CALL_2(bpf_ringbuf_submit_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
606{
607 if (!ptr->data)
608 return 0;
609
610 bpf_ringbuf_commit(ptr->data, flags, false /* discard */);
611
612 bpf_dynptr_set_null(ptr);
613
614 return 0;
615}
616
617const struct bpf_func_proto bpf_ringbuf_submit_dynptr_proto = {
618 .func = bpf_ringbuf_submit_dynptr,
619 .ret_type = RET_VOID,
620 .arg1_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
621 .arg2_type = ARG_ANYTHING,
622};
623
624BPF_CALL_2(bpf_ringbuf_discard_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
625{
626 if (!ptr->data)
627 return 0;
628
629 bpf_ringbuf_commit(ptr->data, flags, true /* discard */);
630
631 bpf_dynptr_set_null(ptr);
632
633 return 0;
634}
635
636const struct bpf_func_proto bpf_ringbuf_discard_dynptr_proto = {
637 .func = bpf_ringbuf_discard_dynptr,
638 .ret_type = RET_VOID,
639 .arg1_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
640 .arg2_type = ARG_ANYTHING,
641};
642
643static int __bpf_user_ringbuf_peek(struct bpf_ringbuf *rb, void **sample, u32 *size)
644{
645 int err;
646 u32 hdr_len, sample_len, total_len, flags, *hdr;
647 u64 cons_pos, prod_pos;
648
649 /* Synchronizes with smp_store_release() in user-space producer. */
650 prod_pos = smp_load_acquire(&rb->producer_pos);
651 if (prod_pos % 8)
652 return -EINVAL;
653
654 /* Synchronizes with smp_store_release() in __bpf_user_ringbuf_sample_release() */
655 cons_pos = smp_load_acquire(&rb->consumer_pos);
656 if (cons_pos >= prod_pos)
657 return -ENODATA;
658
659 hdr = (u32 *)((uintptr_t)rb->data + (uintptr_t)(cons_pos & rb->mask));
660 /* Synchronizes with smp_store_release() in user-space producer. */
661 hdr_len = smp_load_acquire(hdr);
662 flags = hdr_len & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT);
663 sample_len = hdr_len & ~flags;
664 total_len = round_up(sample_len + BPF_RINGBUF_HDR_SZ, 8);
665
666 /* The sample must fit within the region advertised by the producer position. */
667 if (total_len > prod_pos - cons_pos)
668 return -EINVAL;
669
670 /* The sample must fit within the data region of the ring buffer. */
671 if (total_len > ringbuf_total_data_sz(rb))
672 return -E2BIG;
673
674 /* The sample must fit into a struct bpf_dynptr. */
675 err = bpf_dynptr_check_size(sample_len);
676 if (err)
677 return -E2BIG;
678
679 if (flags & BPF_RINGBUF_DISCARD_BIT) {
680 /* If the discard bit is set, the sample should be skipped.
681 *
682 * Update the consumer pos, and return -EAGAIN so the caller
683 * knows to skip this sample and try to read the next one.
684 */
685 smp_store_release(&rb->consumer_pos, cons_pos + total_len);
686 return -EAGAIN;
687 }
688
689 if (flags & BPF_RINGBUF_BUSY_BIT)
690 return -ENODATA;
691
692 *sample = (void *)((uintptr_t)rb->data +
693 (uintptr_t)((cons_pos + BPF_RINGBUF_HDR_SZ) & rb->mask));
694 *size = sample_len;
695 return 0;
696}
697
698static void __bpf_user_ringbuf_sample_release(struct bpf_ringbuf *rb, size_t size, u64 flags)
699{
700 u64 consumer_pos;
701 u32 rounded_size = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
702
703 /* Using smp_load_acquire() is unnecessary here, as the busy-bit
704 * prevents another task from writing to consumer_pos after it was read
705 * by this task with smp_load_acquire() in __bpf_user_ringbuf_peek().
706 */
707 consumer_pos = rb->consumer_pos;
708 /* Synchronizes with smp_load_acquire() in user-space producer. */
709 smp_store_release(&rb->consumer_pos, consumer_pos + rounded_size);
710}
711
712BPF_CALL_4(bpf_user_ringbuf_drain, struct bpf_map *, map,
713 void *, callback_fn, void *, callback_ctx, u64, flags)
714{
715 struct bpf_ringbuf *rb;
716 long samples, discarded_samples = 0, ret = 0;
717 bpf_callback_t callback = (bpf_callback_t)callback_fn;
718 u64 wakeup_flags = BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP;
719 int busy = 0;
720
721 if (unlikely(flags & ~wakeup_flags))
722 return -EINVAL;
723
724 rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
725
726 /* If another consumer is already consuming a sample, wait for them to finish. */
727 if (!atomic_try_cmpxchg(&rb->busy, &busy, 1))
728 return -EBUSY;
729
730 for (samples = 0; samples < BPF_MAX_USER_RINGBUF_SAMPLES && ret == 0; samples++) {
731 int err;
732 u32 size;
733 void *sample;
734 struct bpf_dynptr_kern dynptr;
735
736 err = __bpf_user_ringbuf_peek(rb, &sample, &size);
737 if (err) {
738 if (err == -ENODATA) {
739 break;
740 } else if (err == -EAGAIN) {
741 discarded_samples++;
742 continue;
743 } else {
744 ret = err;
745 goto schedule_work_return;
746 }
747 }
748
749 bpf_dynptr_init(&dynptr, sample, BPF_DYNPTR_TYPE_LOCAL, 0, size);
750 ret = callback((uintptr_t)&dynptr, (uintptr_t)callback_ctx, 0, 0, 0);
751 __bpf_user_ringbuf_sample_release(rb, size, flags);
752 }
753 ret = samples - discarded_samples;
754
755schedule_work_return:
756 /* Prevent the clearing of the busy-bit from being reordered before the
757 * storing of any rb consumer or producer positions.
758 */
759 smp_mb__before_atomic();
760 atomic_set(&rb->busy, 0);
761
762 if (flags & BPF_RB_FORCE_WAKEUP)
763 irq_work_queue(&rb->work);
764 else if (!(flags & BPF_RB_NO_WAKEUP) && samples > 0)
765 irq_work_queue(&rb->work);
766 return ret;
767}
768
769const struct bpf_func_proto bpf_user_ringbuf_drain_proto = {
770 .func = bpf_user_ringbuf_drain,
771 .ret_type = RET_INTEGER,
772 .arg1_type = ARG_CONST_MAP_PTR,
773 .arg2_type = ARG_PTR_TO_FUNC,
774 .arg3_type = ARG_PTR_TO_STACK_OR_NULL,
775 .arg4_type = ARG_ANYTHING,
776};
1#include <linux/bpf.h>
2#include <linux/btf.h>
3#include <linux/err.h>
4#include <linux/irq_work.h>
5#include <linux/slab.h>
6#include <linux/filter.h>
7#include <linux/mm.h>
8#include <linux/vmalloc.h>
9#include <linux/wait.h>
10#include <linux/poll.h>
11#include <linux/kmemleak.h>
12#include <uapi/linux/btf.h>
13#include <linux/btf_ids.h>
14
15#define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
16
17/* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
18#define RINGBUF_PGOFF \
19 (offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
20/* consumer page and producer page */
21#define RINGBUF_POS_PAGES 2
22#define RINGBUF_NR_META_PAGES (RINGBUF_PGOFF + RINGBUF_POS_PAGES)
23
24#define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
25
26struct bpf_ringbuf {
27 wait_queue_head_t waitq;
28 struct irq_work work;
29 u64 mask;
30 struct page **pages;
31 int nr_pages;
32 raw_spinlock_t spinlock ____cacheline_aligned_in_smp;
33 /* For user-space producer ring buffers, an atomic_t busy bit is used
34 * to synchronize access to the ring buffers in the kernel, rather than
35 * the spinlock that is used for kernel-producer ring buffers. This is
36 * done because the ring buffer must hold a lock across a BPF program's
37 * callback:
38 *
39 * __bpf_user_ringbuf_peek() // lock acquired
40 * -> program callback_fn()
41 * -> __bpf_user_ringbuf_sample_release() // lock released
42 *
43 * It is unsafe and incorrect to hold an IRQ spinlock across what could
44 * be a long execution window, so we instead simply disallow concurrent
45 * access to the ring buffer by kernel consumers, and return -EBUSY from
46 * __bpf_user_ringbuf_peek() if the busy bit is held by another task.
47 */
48 atomic_t busy ____cacheline_aligned_in_smp;
49 /* Consumer and producer counters are put into separate pages to
50 * allow each position to be mapped with different permissions.
51 * This prevents a user-space application from modifying the
52 * position and ruining in-kernel tracking. The permissions of the
53 * pages depend on who is producing samples: user-space or the
54 * kernel. Note that the pending counter is placed in the same
55 * page as the producer, so that it shares the same cache line.
56 *
57 * Kernel-producer
58 * ---------------
59 * The producer position and data pages are mapped as r/o in
60 * userspace. For this approach, bits in the header of samples are
61 * used to signal to user-space, and to other producers, whether a
62 * sample is currently being written.
63 *
64 * User-space producer
65 * -------------------
66 * Only the page containing the consumer position is mapped r/o in
67 * user-space. User-space producers also use bits of the header to
68 * communicate to the kernel, but the kernel must carefully check and
69 * validate each sample to ensure that they're correctly formatted, and
70 * fully contained within the ring buffer.
71 */
72 unsigned long consumer_pos __aligned(PAGE_SIZE);
73 unsigned long producer_pos __aligned(PAGE_SIZE);
74 unsigned long pending_pos;
75 char data[] __aligned(PAGE_SIZE);
76};
77
78struct bpf_ringbuf_map {
79 struct bpf_map map;
80 struct bpf_ringbuf *rb;
81};
82
83/* 8-byte ring buffer record header structure */
84struct bpf_ringbuf_hdr {
85 u32 len;
86 u32 pg_off;
87};
88
89static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
90{
91 const gfp_t flags = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL |
92 __GFP_NOWARN | __GFP_ZERO;
93 int nr_meta_pages = RINGBUF_NR_META_PAGES;
94 int nr_data_pages = data_sz >> PAGE_SHIFT;
95 int nr_pages = nr_meta_pages + nr_data_pages;
96 struct page **pages, *page;
97 struct bpf_ringbuf *rb;
98 size_t array_size;
99 int i;
100
101 /* Each data page is mapped twice to allow "virtual"
102 * continuous read of samples wrapping around the end of ring
103 * buffer area:
104 * ------------------------------------------------------
105 * | meta pages | real data pages | same data pages |
106 * ------------------------------------------------------
107 * | | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
108 * ------------------------------------------------------
109 * | | TA DA | TA DA |
110 * ------------------------------------------------------
111 * ^^^^^^^
112 * |
113 * Here, no need to worry about special handling of wrapped-around
114 * data due to double-mapped data pages. This works both in kernel and
115 * when mmap()'ed in user-space, simplifying both kernel and
116 * user-space implementations significantly.
117 */
118 array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
119 pages = bpf_map_area_alloc(array_size, numa_node);
120 if (!pages)
121 return NULL;
122
123 for (i = 0; i < nr_pages; i++) {
124 page = alloc_pages_node(numa_node, flags, 0);
125 if (!page) {
126 nr_pages = i;
127 goto err_free_pages;
128 }
129 pages[i] = page;
130 if (i >= nr_meta_pages)
131 pages[nr_data_pages + i] = page;
132 }
133
134 rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
135 VM_MAP | VM_USERMAP, PAGE_KERNEL);
136 if (rb) {
137 kmemleak_not_leak(pages);
138 rb->pages = pages;
139 rb->nr_pages = nr_pages;
140 return rb;
141 }
142
143err_free_pages:
144 for (i = 0; i < nr_pages; i++)
145 __free_page(pages[i]);
146 bpf_map_area_free(pages);
147 return NULL;
148}
149
150static void bpf_ringbuf_notify(struct irq_work *work)
151{
152 struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
153
154 wake_up_all(&rb->waitq);
155}
156
157/* Maximum size of ring buffer area is limited by 32-bit page offset within
158 * record header, counted in pages. Reserve 8 bits for extensibility, and
159 * take into account few extra pages for consumer/producer pages and
160 * non-mmap()'able parts, the current maximum size would be:
161 *
162 * (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
163 *
164 * This gives 64GB limit, which seems plenty for single ring buffer. Now
165 * considering that the maximum value of data_sz is (4GB - 1), there
166 * will be no overflow, so just note the size limit in the comments.
167 */
168static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
169{
170 struct bpf_ringbuf *rb;
171
172 rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
173 if (!rb)
174 return NULL;
175
176 raw_spin_lock_init(&rb->spinlock);
177 atomic_set(&rb->busy, 0);
178 init_waitqueue_head(&rb->waitq);
179 init_irq_work(&rb->work, bpf_ringbuf_notify);
180
181 rb->mask = data_sz - 1;
182 rb->consumer_pos = 0;
183 rb->producer_pos = 0;
184 rb->pending_pos = 0;
185
186 return rb;
187}
188
189static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
190{
191 struct bpf_ringbuf_map *rb_map;
192
193 if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
194 return ERR_PTR(-EINVAL);
195
196 if (attr->key_size || attr->value_size ||
197 !is_power_of_2(attr->max_entries) ||
198 !PAGE_ALIGNED(attr->max_entries))
199 return ERR_PTR(-EINVAL);
200
201 rb_map = bpf_map_area_alloc(sizeof(*rb_map), NUMA_NO_NODE);
202 if (!rb_map)
203 return ERR_PTR(-ENOMEM);
204
205 bpf_map_init_from_attr(&rb_map->map, attr);
206
207 rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
208 if (!rb_map->rb) {
209 bpf_map_area_free(rb_map);
210 return ERR_PTR(-ENOMEM);
211 }
212
213 return &rb_map->map;
214}
215
216static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
217{
218 /* copy pages pointer and nr_pages to local variable, as we are going
219 * to unmap rb itself with vunmap() below
220 */
221 struct page **pages = rb->pages;
222 int i, nr_pages = rb->nr_pages;
223
224 vunmap(rb);
225 for (i = 0; i < nr_pages; i++)
226 __free_page(pages[i]);
227 bpf_map_area_free(pages);
228}
229
230static void ringbuf_map_free(struct bpf_map *map)
231{
232 struct bpf_ringbuf_map *rb_map;
233
234 rb_map = container_of(map, struct bpf_ringbuf_map, map);
235 bpf_ringbuf_free(rb_map->rb);
236 bpf_map_area_free(rb_map);
237}
238
239static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
240{
241 return ERR_PTR(-ENOTSUPP);
242}
243
244static long ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
245 u64 flags)
246{
247 return -ENOTSUPP;
248}
249
250static long ringbuf_map_delete_elem(struct bpf_map *map, void *key)
251{
252 return -ENOTSUPP;
253}
254
255static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
256 void *next_key)
257{
258 return -ENOTSUPP;
259}
260
261static int ringbuf_map_mmap_kern(struct bpf_map *map, struct vm_area_struct *vma)
262{
263 struct bpf_ringbuf_map *rb_map;
264
265 rb_map = container_of(map, struct bpf_ringbuf_map, map);
266
267 if (vma->vm_flags & VM_WRITE) {
268 /* allow writable mapping for the consumer_pos only */
269 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
270 return -EPERM;
271 }
272 /* remap_vmalloc_range() checks size and offset constraints */
273 return remap_vmalloc_range(vma, rb_map->rb,
274 vma->vm_pgoff + RINGBUF_PGOFF);
275}
276
277static int ringbuf_map_mmap_user(struct bpf_map *map, struct vm_area_struct *vma)
278{
279 struct bpf_ringbuf_map *rb_map;
280
281 rb_map = container_of(map, struct bpf_ringbuf_map, map);
282
283 if (vma->vm_flags & VM_WRITE) {
284 if (vma->vm_pgoff == 0)
285 /* Disallow writable mappings to the consumer pointer,
286 * and allow writable mappings to both the producer
287 * position, and the ring buffer data itself.
288 */
289 return -EPERM;
290 }
291 /* remap_vmalloc_range() checks size and offset constraints */
292 return remap_vmalloc_range(vma, rb_map->rb, vma->vm_pgoff + RINGBUF_PGOFF);
293}
294
295static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
296{
297 unsigned long cons_pos, prod_pos;
298
299 cons_pos = smp_load_acquire(&rb->consumer_pos);
300 prod_pos = smp_load_acquire(&rb->producer_pos);
301 return prod_pos - cons_pos;
302}
303
304static u32 ringbuf_total_data_sz(const struct bpf_ringbuf *rb)
305{
306 return rb->mask + 1;
307}
308
309static __poll_t ringbuf_map_poll_kern(struct bpf_map *map, struct file *filp,
310 struct poll_table_struct *pts)
311{
312 struct bpf_ringbuf_map *rb_map;
313
314 rb_map = container_of(map, struct bpf_ringbuf_map, map);
315 poll_wait(filp, &rb_map->rb->waitq, pts);
316
317 if (ringbuf_avail_data_sz(rb_map->rb))
318 return EPOLLIN | EPOLLRDNORM;
319 return 0;
320}
321
322static __poll_t ringbuf_map_poll_user(struct bpf_map *map, struct file *filp,
323 struct poll_table_struct *pts)
324{
325 struct bpf_ringbuf_map *rb_map;
326
327 rb_map = container_of(map, struct bpf_ringbuf_map, map);
328 poll_wait(filp, &rb_map->rb->waitq, pts);
329
330 if (ringbuf_avail_data_sz(rb_map->rb) < ringbuf_total_data_sz(rb_map->rb))
331 return EPOLLOUT | EPOLLWRNORM;
332 return 0;
333}
334
335static u64 ringbuf_map_mem_usage(const struct bpf_map *map)
336{
337 struct bpf_ringbuf *rb;
338 int nr_data_pages;
339 int nr_meta_pages;
340 u64 usage = sizeof(struct bpf_ringbuf_map);
341
342 rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
343 usage += (u64)rb->nr_pages << PAGE_SHIFT;
344 nr_meta_pages = RINGBUF_NR_META_PAGES;
345 nr_data_pages = map->max_entries >> PAGE_SHIFT;
346 usage += (nr_meta_pages + 2 * nr_data_pages) * sizeof(struct page *);
347 return usage;
348}
349
350BTF_ID_LIST_SINGLE(ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
351const struct bpf_map_ops ringbuf_map_ops = {
352 .map_meta_equal = bpf_map_meta_equal,
353 .map_alloc = ringbuf_map_alloc,
354 .map_free = ringbuf_map_free,
355 .map_mmap = ringbuf_map_mmap_kern,
356 .map_poll = ringbuf_map_poll_kern,
357 .map_lookup_elem = ringbuf_map_lookup_elem,
358 .map_update_elem = ringbuf_map_update_elem,
359 .map_delete_elem = ringbuf_map_delete_elem,
360 .map_get_next_key = ringbuf_map_get_next_key,
361 .map_mem_usage = ringbuf_map_mem_usage,
362 .map_btf_id = &ringbuf_map_btf_ids[0],
363};
364
365BTF_ID_LIST_SINGLE(user_ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
366const struct bpf_map_ops user_ringbuf_map_ops = {
367 .map_meta_equal = bpf_map_meta_equal,
368 .map_alloc = ringbuf_map_alloc,
369 .map_free = ringbuf_map_free,
370 .map_mmap = ringbuf_map_mmap_user,
371 .map_poll = ringbuf_map_poll_user,
372 .map_lookup_elem = ringbuf_map_lookup_elem,
373 .map_update_elem = ringbuf_map_update_elem,
374 .map_delete_elem = ringbuf_map_delete_elem,
375 .map_get_next_key = ringbuf_map_get_next_key,
376 .map_mem_usage = ringbuf_map_mem_usage,
377 .map_btf_id = &user_ringbuf_map_btf_ids[0],
378};
379
380/* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
381 * calculate offset from record metadata to ring buffer in pages, rounded
382 * down. This page offset is stored as part of record metadata and allows to
383 * restore struct bpf_ringbuf * from record pointer. This page offset is
384 * stored at offset 4 of record metadata header.
385 */
386static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
387 struct bpf_ringbuf_hdr *hdr)
388{
389 return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
390}
391
392/* Given pointer to ring buffer record header, restore pointer to struct
393 * bpf_ringbuf itself by using page offset stored at offset 4
394 */
395static struct bpf_ringbuf *
396bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
397{
398 unsigned long addr = (unsigned long)(void *)hdr;
399 unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
400
401 return (void*)((addr & PAGE_MASK) - off);
402}
403
404static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
405{
406 unsigned long cons_pos, prod_pos, new_prod_pos, pend_pos, flags;
407 struct bpf_ringbuf_hdr *hdr;
408 u32 len, pg_off, tmp_size, hdr_len;
409
410 if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
411 return NULL;
412
413 len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
414 if (len > ringbuf_total_data_sz(rb))
415 return NULL;
416
417 cons_pos = smp_load_acquire(&rb->consumer_pos);
418
419 if (in_nmi()) {
420 if (!raw_spin_trylock_irqsave(&rb->spinlock, flags))
421 return NULL;
422 } else {
423 raw_spin_lock_irqsave(&rb->spinlock, flags);
424 }
425
426 pend_pos = rb->pending_pos;
427 prod_pos = rb->producer_pos;
428 new_prod_pos = prod_pos + len;
429
430 while (pend_pos < prod_pos) {
431 hdr = (void *)rb->data + (pend_pos & rb->mask);
432 hdr_len = READ_ONCE(hdr->len);
433 if (hdr_len & BPF_RINGBUF_BUSY_BIT)
434 break;
435 tmp_size = hdr_len & ~BPF_RINGBUF_DISCARD_BIT;
436 tmp_size = round_up(tmp_size + BPF_RINGBUF_HDR_SZ, 8);
437 pend_pos += tmp_size;
438 }
439 rb->pending_pos = pend_pos;
440
441 /* check for out of ringbuf space:
442 * - by ensuring producer position doesn't advance more than
443 * (ringbuf_size - 1) ahead
444 * - by ensuring oldest not yet committed record until newest
445 * record does not span more than (ringbuf_size - 1)
446 */
447 if (new_prod_pos - cons_pos > rb->mask ||
448 new_prod_pos - pend_pos > rb->mask) {
449 raw_spin_unlock_irqrestore(&rb->spinlock, flags);
450 return NULL;
451 }
452
453 hdr = (void *)rb->data + (prod_pos & rb->mask);
454 pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
455 hdr->len = size | BPF_RINGBUF_BUSY_BIT;
456 hdr->pg_off = pg_off;
457
458 /* pairs with consumer's smp_load_acquire() */
459 smp_store_release(&rb->producer_pos, new_prod_pos);
460
461 raw_spin_unlock_irqrestore(&rb->spinlock, flags);
462
463 return (void *)hdr + BPF_RINGBUF_HDR_SZ;
464}
465
466BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
467{
468 struct bpf_ringbuf_map *rb_map;
469
470 if (unlikely(flags))
471 return 0;
472
473 rb_map = container_of(map, struct bpf_ringbuf_map, map);
474 return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
475}
476
477const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
478 .func = bpf_ringbuf_reserve,
479 .ret_type = RET_PTR_TO_RINGBUF_MEM_OR_NULL,
480 .arg1_type = ARG_CONST_MAP_PTR,
481 .arg2_type = ARG_CONST_ALLOC_SIZE_OR_ZERO,
482 .arg3_type = ARG_ANYTHING,
483};
484
485static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
486{
487 unsigned long rec_pos, cons_pos;
488 struct bpf_ringbuf_hdr *hdr;
489 struct bpf_ringbuf *rb;
490 u32 new_len;
491
492 hdr = sample - BPF_RINGBUF_HDR_SZ;
493 rb = bpf_ringbuf_restore_from_rec(hdr);
494 new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
495 if (discard)
496 new_len |= BPF_RINGBUF_DISCARD_BIT;
497
498 /* update record header with correct final size prefix */
499 xchg(&hdr->len, new_len);
500
501 /* if consumer caught up and is waiting for our record, notify about
502 * new data availability
503 */
504 rec_pos = (void *)hdr - (void *)rb->data;
505 cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
506
507 if (flags & BPF_RB_FORCE_WAKEUP)
508 irq_work_queue(&rb->work);
509 else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
510 irq_work_queue(&rb->work);
511}
512
513BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
514{
515 bpf_ringbuf_commit(sample, flags, false /* discard */);
516 return 0;
517}
518
519const struct bpf_func_proto bpf_ringbuf_submit_proto = {
520 .func = bpf_ringbuf_submit,
521 .ret_type = RET_VOID,
522 .arg1_type = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
523 .arg2_type = ARG_ANYTHING,
524};
525
526BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
527{
528 bpf_ringbuf_commit(sample, flags, true /* discard */);
529 return 0;
530}
531
532const struct bpf_func_proto bpf_ringbuf_discard_proto = {
533 .func = bpf_ringbuf_discard,
534 .ret_type = RET_VOID,
535 .arg1_type = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
536 .arg2_type = ARG_ANYTHING,
537};
538
539BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
540 u64, flags)
541{
542 struct bpf_ringbuf_map *rb_map;
543 void *rec;
544
545 if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
546 return -EINVAL;
547
548 rb_map = container_of(map, struct bpf_ringbuf_map, map);
549 rec = __bpf_ringbuf_reserve(rb_map->rb, size);
550 if (!rec)
551 return -EAGAIN;
552
553 memcpy(rec, data, size);
554 bpf_ringbuf_commit(rec, flags, false /* discard */);
555 return 0;
556}
557
558const struct bpf_func_proto bpf_ringbuf_output_proto = {
559 .func = bpf_ringbuf_output,
560 .ret_type = RET_INTEGER,
561 .arg1_type = ARG_CONST_MAP_PTR,
562 .arg2_type = ARG_PTR_TO_MEM | MEM_RDONLY,
563 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
564 .arg4_type = ARG_ANYTHING,
565};
566
567BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
568{
569 struct bpf_ringbuf *rb;
570
571 rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
572
573 switch (flags) {
574 case BPF_RB_AVAIL_DATA:
575 return ringbuf_avail_data_sz(rb);
576 case BPF_RB_RING_SIZE:
577 return ringbuf_total_data_sz(rb);
578 case BPF_RB_CONS_POS:
579 return smp_load_acquire(&rb->consumer_pos);
580 case BPF_RB_PROD_POS:
581 return smp_load_acquire(&rb->producer_pos);
582 default:
583 return 0;
584 }
585}
586
587const struct bpf_func_proto bpf_ringbuf_query_proto = {
588 .func = bpf_ringbuf_query,
589 .ret_type = RET_INTEGER,
590 .arg1_type = ARG_CONST_MAP_PTR,
591 .arg2_type = ARG_ANYTHING,
592};
593
594BPF_CALL_4(bpf_ringbuf_reserve_dynptr, struct bpf_map *, map, u32, size, u64, flags,
595 struct bpf_dynptr_kern *, ptr)
596{
597 struct bpf_ringbuf_map *rb_map;
598 void *sample;
599 int err;
600
601 if (unlikely(flags)) {
602 bpf_dynptr_set_null(ptr);
603 return -EINVAL;
604 }
605
606 err = bpf_dynptr_check_size(size);
607 if (err) {
608 bpf_dynptr_set_null(ptr);
609 return err;
610 }
611
612 rb_map = container_of(map, struct bpf_ringbuf_map, map);
613
614 sample = __bpf_ringbuf_reserve(rb_map->rb, size);
615 if (!sample) {
616 bpf_dynptr_set_null(ptr);
617 return -EINVAL;
618 }
619
620 bpf_dynptr_init(ptr, sample, BPF_DYNPTR_TYPE_RINGBUF, 0, size);
621
622 return 0;
623}
624
625const struct bpf_func_proto bpf_ringbuf_reserve_dynptr_proto = {
626 .func = bpf_ringbuf_reserve_dynptr,
627 .ret_type = RET_INTEGER,
628 .arg1_type = ARG_CONST_MAP_PTR,
629 .arg2_type = ARG_ANYTHING,
630 .arg3_type = ARG_ANYTHING,
631 .arg4_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | MEM_UNINIT | MEM_WRITE,
632};
633
634BPF_CALL_2(bpf_ringbuf_submit_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
635{
636 if (!ptr->data)
637 return 0;
638
639 bpf_ringbuf_commit(ptr->data, flags, false /* discard */);
640
641 bpf_dynptr_set_null(ptr);
642
643 return 0;
644}
645
646const struct bpf_func_proto bpf_ringbuf_submit_dynptr_proto = {
647 .func = bpf_ringbuf_submit_dynptr,
648 .ret_type = RET_VOID,
649 .arg1_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
650 .arg2_type = ARG_ANYTHING,
651};
652
653BPF_CALL_2(bpf_ringbuf_discard_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
654{
655 if (!ptr->data)
656 return 0;
657
658 bpf_ringbuf_commit(ptr->data, flags, true /* discard */);
659
660 bpf_dynptr_set_null(ptr);
661
662 return 0;
663}
664
665const struct bpf_func_proto bpf_ringbuf_discard_dynptr_proto = {
666 .func = bpf_ringbuf_discard_dynptr,
667 .ret_type = RET_VOID,
668 .arg1_type = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
669 .arg2_type = ARG_ANYTHING,
670};
671
672static int __bpf_user_ringbuf_peek(struct bpf_ringbuf *rb, void **sample, u32 *size)
673{
674 int err;
675 u32 hdr_len, sample_len, total_len, flags, *hdr;
676 u64 cons_pos, prod_pos;
677
678 /* Synchronizes with smp_store_release() in user-space producer. */
679 prod_pos = smp_load_acquire(&rb->producer_pos);
680 if (prod_pos % 8)
681 return -EINVAL;
682
683 /* Synchronizes with smp_store_release() in __bpf_user_ringbuf_sample_release() */
684 cons_pos = smp_load_acquire(&rb->consumer_pos);
685 if (cons_pos >= prod_pos)
686 return -ENODATA;
687
688 hdr = (u32 *)((uintptr_t)rb->data + (uintptr_t)(cons_pos & rb->mask));
689 /* Synchronizes with smp_store_release() in user-space producer. */
690 hdr_len = smp_load_acquire(hdr);
691 flags = hdr_len & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT);
692 sample_len = hdr_len & ~flags;
693 total_len = round_up(sample_len + BPF_RINGBUF_HDR_SZ, 8);
694
695 /* The sample must fit within the region advertised by the producer position. */
696 if (total_len > prod_pos - cons_pos)
697 return -EINVAL;
698
699 /* The sample must fit within the data region of the ring buffer. */
700 if (total_len > ringbuf_total_data_sz(rb))
701 return -E2BIG;
702
703 /* The sample must fit into a struct bpf_dynptr. */
704 err = bpf_dynptr_check_size(sample_len);
705 if (err)
706 return -E2BIG;
707
708 if (flags & BPF_RINGBUF_DISCARD_BIT) {
709 /* If the discard bit is set, the sample should be skipped.
710 *
711 * Update the consumer pos, and return -EAGAIN so the caller
712 * knows to skip this sample and try to read the next one.
713 */
714 smp_store_release(&rb->consumer_pos, cons_pos + total_len);
715 return -EAGAIN;
716 }
717
718 if (flags & BPF_RINGBUF_BUSY_BIT)
719 return -ENODATA;
720
721 *sample = (void *)((uintptr_t)rb->data +
722 (uintptr_t)((cons_pos + BPF_RINGBUF_HDR_SZ) & rb->mask));
723 *size = sample_len;
724 return 0;
725}
726
727static void __bpf_user_ringbuf_sample_release(struct bpf_ringbuf *rb, size_t size, u64 flags)
728{
729 u64 consumer_pos;
730 u32 rounded_size = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
731
732 /* Using smp_load_acquire() is unnecessary here, as the busy-bit
733 * prevents another task from writing to consumer_pos after it was read
734 * by this task with smp_load_acquire() in __bpf_user_ringbuf_peek().
735 */
736 consumer_pos = rb->consumer_pos;
737 /* Synchronizes with smp_load_acquire() in user-space producer. */
738 smp_store_release(&rb->consumer_pos, consumer_pos + rounded_size);
739}
740
741BPF_CALL_4(bpf_user_ringbuf_drain, struct bpf_map *, map,
742 void *, callback_fn, void *, callback_ctx, u64, flags)
743{
744 struct bpf_ringbuf *rb;
745 long samples, discarded_samples = 0, ret = 0;
746 bpf_callback_t callback = (bpf_callback_t)callback_fn;
747 u64 wakeup_flags = BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP;
748 int busy = 0;
749
750 if (unlikely(flags & ~wakeup_flags))
751 return -EINVAL;
752
753 rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
754
755 /* If another consumer is already consuming a sample, wait for them to finish. */
756 if (!atomic_try_cmpxchg(&rb->busy, &busy, 1))
757 return -EBUSY;
758
759 for (samples = 0; samples < BPF_MAX_USER_RINGBUF_SAMPLES && ret == 0; samples++) {
760 int err;
761 u32 size;
762 void *sample;
763 struct bpf_dynptr_kern dynptr;
764
765 err = __bpf_user_ringbuf_peek(rb, &sample, &size);
766 if (err) {
767 if (err == -ENODATA) {
768 break;
769 } else if (err == -EAGAIN) {
770 discarded_samples++;
771 continue;
772 } else {
773 ret = err;
774 goto schedule_work_return;
775 }
776 }
777
778 bpf_dynptr_init(&dynptr, sample, BPF_DYNPTR_TYPE_LOCAL, 0, size);
779 ret = callback((uintptr_t)&dynptr, (uintptr_t)callback_ctx, 0, 0, 0);
780 __bpf_user_ringbuf_sample_release(rb, size, flags);
781 }
782 ret = samples - discarded_samples;
783
784schedule_work_return:
785 /* Prevent the clearing of the busy-bit from being reordered before the
786 * storing of any rb consumer or producer positions.
787 */
788 atomic_set_release(&rb->busy, 0);
789
790 if (flags & BPF_RB_FORCE_WAKEUP)
791 irq_work_queue(&rb->work);
792 else if (!(flags & BPF_RB_NO_WAKEUP) && samples > 0)
793 irq_work_queue(&rb->work);
794 return ret;
795}
796
797const struct bpf_func_proto bpf_user_ringbuf_drain_proto = {
798 .func = bpf_user_ringbuf_drain,
799 .ret_type = RET_INTEGER,
800 .arg1_type = ARG_CONST_MAP_PTR,
801 .arg2_type = ARG_PTR_TO_FUNC,
802 .arg3_type = ARG_PTR_TO_STACK_OR_NULL,
803 .arg4_type = ARG_ANYTHING,
804};