xref: /openbmc/linux/kernel/bpf/ringbuf.c (revision 26cfd12b)
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 <uapi/linux/btf.h>
12 
13 #define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
14 
15 /* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
16 #define RINGBUF_PGOFF \
17 	(offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
18 /* consumer page and producer page */
19 #define RINGBUF_POS_PAGES 2
20 
21 #define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
22 
23 /* Maximum size of ring buffer area is limited by 32-bit page offset within
24  * record header, counted in pages. Reserve 8 bits for extensibility, and take
25  * into account few extra pages for consumer/producer pages and
26  * non-mmap()'able parts. This gives 64GB limit, which seems plenty for single
27  * ring buffer.
28  */
29 #define RINGBUF_MAX_DATA_SZ \
30 	(((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
31 
32 struct bpf_ringbuf {
33 	wait_queue_head_t waitq;
34 	struct irq_work work;
35 	u64 mask;
36 	struct page **pages;
37 	int nr_pages;
38 	spinlock_t spinlock ____cacheline_aligned_in_smp;
39 	/* Consumer and producer counters are put into separate pages to allow
40 	 * mapping consumer page as r/w, but restrict producer page to r/o.
41 	 * This protects producer position from being modified by user-space
42 	 * application and ruining in-kernel position tracking.
43 	 */
44 	unsigned long consumer_pos __aligned(PAGE_SIZE);
45 	unsigned long producer_pos __aligned(PAGE_SIZE);
46 	char data[] __aligned(PAGE_SIZE);
47 };
48 
49 struct bpf_ringbuf_map {
50 	struct bpf_map map;
51 	struct bpf_map_memory memory;
52 	struct bpf_ringbuf *rb;
53 };
54 
55 /* 8-byte ring buffer record header structure */
56 struct bpf_ringbuf_hdr {
57 	u32 len;
58 	u32 pg_off;
59 };
60 
61 static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
62 {
63 	const gfp_t flags = GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
64 			    __GFP_ZERO;
65 	int nr_meta_pages = RINGBUF_PGOFF + RINGBUF_POS_PAGES;
66 	int nr_data_pages = data_sz >> PAGE_SHIFT;
67 	int nr_pages = nr_meta_pages + nr_data_pages;
68 	struct page **pages, *page;
69 	struct bpf_ringbuf *rb;
70 	size_t array_size;
71 	int i;
72 
73 	/* Each data page is mapped twice to allow "virtual"
74 	 * continuous read of samples wrapping around the end of ring
75 	 * buffer area:
76 	 * ------------------------------------------------------
77 	 * | meta pages |  real data pages  |  same data pages  |
78 	 * ------------------------------------------------------
79 	 * |            | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
80 	 * ------------------------------------------------------
81 	 * |            | TA             DA | TA             DA |
82 	 * ------------------------------------------------------
83 	 *                               ^^^^^^^
84 	 *                                  |
85 	 * Here, no need to worry about special handling of wrapped-around
86 	 * data due to double-mapped data pages. This works both in kernel and
87 	 * when mmap()'ed in user-space, simplifying both kernel and
88 	 * user-space implementations significantly.
89 	 */
90 	array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
91 	if (array_size > PAGE_SIZE)
92 		pages = vmalloc_node(array_size, numa_node);
93 	else
94 		pages = kmalloc_node(array_size, flags, numa_node);
95 	if (!pages)
96 		return NULL;
97 
98 	for (i = 0; i < nr_pages; i++) {
99 		page = alloc_pages_node(numa_node, flags, 0);
100 		if (!page) {
101 			nr_pages = i;
102 			goto err_free_pages;
103 		}
104 		pages[i] = page;
105 		if (i >= nr_meta_pages)
106 			pages[nr_data_pages + i] = page;
107 	}
108 
109 	rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
110 		  VM_ALLOC | VM_USERMAP, PAGE_KERNEL);
111 	if (rb) {
112 		rb->pages = pages;
113 		rb->nr_pages = nr_pages;
114 		return rb;
115 	}
116 
117 err_free_pages:
118 	for (i = 0; i < nr_pages; i++)
119 		__free_page(pages[i]);
120 	kvfree(pages);
121 	return NULL;
122 }
123 
124 static void bpf_ringbuf_notify(struct irq_work *work)
125 {
126 	struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
127 
128 	wake_up_all(&rb->waitq);
129 }
130 
131 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
132 {
133 	struct bpf_ringbuf *rb;
134 
135 	if (!data_sz || !PAGE_ALIGNED(data_sz))
136 		return ERR_PTR(-EINVAL);
137 
138 #ifdef CONFIG_64BIT
139 	/* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
140 	if (data_sz > RINGBUF_MAX_DATA_SZ)
141 		return ERR_PTR(-E2BIG);
142 #endif
143 
144 	rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
145 	if (!rb)
146 		return ERR_PTR(-ENOMEM);
147 
148 	spin_lock_init(&rb->spinlock);
149 	init_waitqueue_head(&rb->waitq);
150 	init_irq_work(&rb->work, bpf_ringbuf_notify);
151 
152 	rb->mask = data_sz - 1;
153 	rb->consumer_pos = 0;
154 	rb->producer_pos = 0;
155 
156 	return rb;
157 }
158 
159 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
160 {
161 	struct bpf_ringbuf_map *rb_map;
162 	u64 cost;
163 	int err;
164 
165 	if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
166 		return ERR_PTR(-EINVAL);
167 
168 	if (attr->key_size || attr->value_size ||
169 	    attr->max_entries == 0 || !PAGE_ALIGNED(attr->max_entries))
170 		return ERR_PTR(-EINVAL);
171 
172 	rb_map = kzalloc(sizeof(*rb_map), GFP_USER);
173 	if (!rb_map)
174 		return ERR_PTR(-ENOMEM);
175 
176 	bpf_map_init_from_attr(&rb_map->map, attr);
177 
178 	cost = sizeof(struct bpf_ringbuf_map) +
179 	       sizeof(struct bpf_ringbuf) +
180 	       attr->max_entries;
181 	err = bpf_map_charge_init(&rb_map->map.memory, cost);
182 	if (err)
183 		goto err_free_map;
184 
185 	rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
186 	if (IS_ERR(rb_map->rb)) {
187 		err = PTR_ERR(rb_map->rb);
188 		goto err_uncharge;
189 	}
190 
191 	return &rb_map->map;
192 
193 err_uncharge:
194 	bpf_map_charge_finish(&rb_map->map.memory);
195 err_free_map:
196 	kfree(rb_map);
197 	return ERR_PTR(err);
198 }
199 
200 static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
201 {
202 	/* copy pages pointer and nr_pages to local variable, as we are going
203 	 * to unmap rb itself with vunmap() below
204 	 */
205 	struct page **pages = rb->pages;
206 	int i, nr_pages = rb->nr_pages;
207 
208 	vunmap(rb);
209 	for (i = 0; i < nr_pages; i++)
210 		__free_page(pages[i]);
211 	kvfree(pages);
212 }
213 
214 static void ringbuf_map_free(struct bpf_map *map)
215 {
216 	struct bpf_ringbuf_map *rb_map;
217 
218 	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
219 	 * so the programs (can be more than one that used this map) were
220 	 * disconnected from events. Wait for outstanding critical sections in
221 	 * these programs to complete
222 	 */
223 	synchronize_rcu();
224 
225 	rb_map = container_of(map, struct bpf_ringbuf_map, map);
226 	bpf_ringbuf_free(rb_map->rb);
227 	kfree(rb_map);
228 }
229 
230 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
231 {
232 	return ERR_PTR(-ENOTSUPP);
233 }
234 
235 static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
236 				   u64 flags)
237 {
238 	return -ENOTSUPP;
239 }
240 
241 static int ringbuf_map_delete_elem(struct bpf_map *map, void *key)
242 {
243 	return -ENOTSUPP;
244 }
245 
246 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
247 				    void *next_key)
248 {
249 	return -ENOTSUPP;
250 }
251 
252 static size_t bpf_ringbuf_mmap_page_cnt(const struct bpf_ringbuf *rb)
253 {
254 	size_t data_pages = (rb->mask + 1) >> PAGE_SHIFT;
255 
256 	/* consumer page + producer page + 2 x data pages */
257 	return RINGBUF_POS_PAGES + 2 * data_pages;
258 }
259 
260 static int ringbuf_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
261 {
262 	struct bpf_ringbuf_map *rb_map;
263 	size_t mmap_sz;
264 
265 	rb_map = container_of(map, struct bpf_ringbuf_map, map);
266 	mmap_sz = bpf_ringbuf_mmap_page_cnt(rb_map->rb) << PAGE_SHIFT;
267 
268 	if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) > mmap_sz)
269 		return -EINVAL;
270 
271 	return remap_vmalloc_range(vma, rb_map->rb,
272 				   vma->vm_pgoff + RINGBUF_PGOFF);
273 }
274 
275 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
276 {
277 	unsigned long cons_pos, prod_pos;
278 
279 	cons_pos = smp_load_acquire(&rb->consumer_pos);
280 	prod_pos = smp_load_acquire(&rb->producer_pos);
281 	return prod_pos - cons_pos;
282 }
283 
284 static __poll_t ringbuf_map_poll(struct bpf_map *map, struct file *filp,
285 				 struct poll_table_struct *pts)
286 {
287 	struct bpf_ringbuf_map *rb_map;
288 
289 	rb_map = container_of(map, struct bpf_ringbuf_map, map);
290 	poll_wait(filp, &rb_map->rb->waitq, pts);
291 
292 	if (ringbuf_avail_data_sz(rb_map->rb))
293 		return EPOLLIN | EPOLLRDNORM;
294 	return 0;
295 }
296 
297 const struct bpf_map_ops ringbuf_map_ops = {
298 	.map_alloc = ringbuf_map_alloc,
299 	.map_free = ringbuf_map_free,
300 	.map_mmap = ringbuf_map_mmap,
301 	.map_poll = ringbuf_map_poll,
302 	.map_lookup_elem = ringbuf_map_lookup_elem,
303 	.map_update_elem = ringbuf_map_update_elem,
304 	.map_delete_elem = ringbuf_map_delete_elem,
305 	.map_get_next_key = ringbuf_map_get_next_key,
306 };
307 
308 /* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
309  * calculate offset from record metadata to ring buffer in pages, rounded
310  * down. This page offset is stored as part of record metadata and allows to
311  * restore struct bpf_ringbuf * from record pointer. This page offset is
312  * stored at offset 4 of record metadata header.
313  */
314 static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
315 				     struct bpf_ringbuf_hdr *hdr)
316 {
317 	return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
318 }
319 
320 /* Given pointer to ring buffer record header, restore pointer to struct
321  * bpf_ringbuf itself by using page offset stored at offset 4
322  */
323 static struct bpf_ringbuf *
324 bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
325 {
326 	unsigned long addr = (unsigned long)(void *)hdr;
327 	unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
328 
329 	return (void*)((addr & PAGE_MASK) - off);
330 }
331 
332 static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
333 {
334 	unsigned long cons_pos, prod_pos, new_prod_pos, flags;
335 	u32 len, pg_off;
336 	struct bpf_ringbuf_hdr *hdr;
337 
338 	if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
339 		return NULL;
340 
341 	len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
342 	cons_pos = smp_load_acquire(&rb->consumer_pos);
343 
344 	if (in_nmi()) {
345 		if (!spin_trylock_irqsave(&rb->spinlock, flags))
346 			return NULL;
347 	} else {
348 		spin_lock_irqsave(&rb->spinlock, flags);
349 	}
350 
351 	prod_pos = rb->producer_pos;
352 	new_prod_pos = prod_pos + len;
353 
354 	/* check for out of ringbuf space by ensuring producer position
355 	 * doesn't advance more than (ringbuf_size - 1) ahead
356 	 */
357 	if (new_prod_pos - cons_pos > rb->mask) {
358 		spin_unlock_irqrestore(&rb->spinlock, flags);
359 		return NULL;
360 	}
361 
362 	hdr = (void *)rb->data + (prod_pos & rb->mask);
363 	pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
364 	hdr->len = size | BPF_RINGBUF_BUSY_BIT;
365 	hdr->pg_off = pg_off;
366 
367 	/* pairs with consumer's smp_load_acquire() */
368 	smp_store_release(&rb->producer_pos, new_prod_pos);
369 
370 	spin_unlock_irqrestore(&rb->spinlock, flags);
371 
372 	return (void *)hdr + BPF_RINGBUF_HDR_SZ;
373 }
374 
375 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
376 {
377 	struct bpf_ringbuf_map *rb_map;
378 
379 	if (unlikely(flags))
380 		return 0;
381 
382 	rb_map = container_of(map, struct bpf_ringbuf_map, map);
383 	return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
384 }
385 
386 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
387 	.func		= bpf_ringbuf_reserve,
388 	.ret_type	= RET_PTR_TO_ALLOC_MEM_OR_NULL,
389 	.arg1_type	= ARG_CONST_MAP_PTR,
390 	.arg2_type	= ARG_CONST_ALLOC_SIZE_OR_ZERO,
391 	.arg3_type	= ARG_ANYTHING,
392 };
393 
394 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
395 {
396 	unsigned long rec_pos, cons_pos;
397 	struct bpf_ringbuf_hdr *hdr;
398 	struct bpf_ringbuf *rb;
399 	u32 new_len;
400 
401 	hdr = sample - BPF_RINGBUF_HDR_SZ;
402 	rb = bpf_ringbuf_restore_from_rec(hdr);
403 	new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
404 	if (discard)
405 		new_len |= BPF_RINGBUF_DISCARD_BIT;
406 
407 	/* update record header with correct final size prefix */
408 	xchg(&hdr->len, new_len);
409 
410 	/* if consumer caught up and is waiting for our record, notify about
411 	 * new data availability
412 	 */
413 	rec_pos = (void *)hdr - (void *)rb->data;
414 	cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
415 
416 	if (flags & BPF_RB_FORCE_WAKEUP)
417 		irq_work_queue(&rb->work);
418 	else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
419 		irq_work_queue(&rb->work);
420 }
421 
422 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
423 {
424 	bpf_ringbuf_commit(sample, flags, false /* discard */);
425 	return 0;
426 }
427 
428 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
429 	.func		= bpf_ringbuf_submit,
430 	.ret_type	= RET_VOID,
431 	.arg1_type	= ARG_PTR_TO_ALLOC_MEM,
432 	.arg2_type	= ARG_ANYTHING,
433 };
434 
435 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
436 {
437 	bpf_ringbuf_commit(sample, flags, true /* discard */);
438 	return 0;
439 }
440 
441 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
442 	.func		= bpf_ringbuf_discard,
443 	.ret_type	= RET_VOID,
444 	.arg1_type	= ARG_PTR_TO_ALLOC_MEM,
445 	.arg2_type	= ARG_ANYTHING,
446 };
447 
448 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
449 	   u64, flags)
450 {
451 	struct bpf_ringbuf_map *rb_map;
452 	void *rec;
453 
454 	if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
455 		return -EINVAL;
456 
457 	rb_map = container_of(map, struct bpf_ringbuf_map, map);
458 	rec = __bpf_ringbuf_reserve(rb_map->rb, size);
459 	if (!rec)
460 		return -EAGAIN;
461 
462 	memcpy(rec, data, size);
463 	bpf_ringbuf_commit(rec, flags, false /* discard */);
464 	return 0;
465 }
466 
467 const struct bpf_func_proto bpf_ringbuf_output_proto = {
468 	.func		= bpf_ringbuf_output,
469 	.ret_type	= RET_INTEGER,
470 	.arg1_type	= ARG_CONST_MAP_PTR,
471 	.arg2_type	= ARG_PTR_TO_MEM,
472 	.arg3_type	= ARG_CONST_SIZE_OR_ZERO,
473 	.arg4_type	= ARG_ANYTHING,
474 };
475 
476 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
477 {
478 	struct bpf_ringbuf *rb;
479 
480 	rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
481 
482 	switch (flags) {
483 	case BPF_RB_AVAIL_DATA:
484 		return ringbuf_avail_data_sz(rb);
485 	case BPF_RB_RING_SIZE:
486 		return rb->mask + 1;
487 	case BPF_RB_CONS_POS:
488 		return smp_load_acquire(&rb->consumer_pos);
489 	case BPF_RB_PROD_POS:
490 		return smp_load_acquire(&rb->producer_pos);
491 	default:
492 		return 0;
493 	}
494 }
495 
496 const struct bpf_func_proto bpf_ringbuf_query_proto = {
497 	.func		= bpf_ringbuf_query,
498 	.ret_type	= RET_INTEGER,
499 	.arg1_type	= ARG_CONST_MAP_PTR,
500 	.arg2_type	= ARG_ANYTHING,
501 };
502