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