xref: /openbmc/linux/net/core/xdp.c (revision ee7da21a) 
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* net/core/xdp.c
3  *
4  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
5  */
6 #include <linux/bpf.h>
7 #include <linux/filter.h>
8 #include <linux/types.h>
9 #include <linux/mm.h>
10 #include <linux/netdevice.h>
11 #include <linux/slab.h>
12 #include <linux/idr.h>
13 #include <linux/rhashtable.h>
14 #include <linux/bug.h>
15 #include <net/page_pool.h>
16 
17 #include <net/xdp.h>
18 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
19 #include <trace/events/xdp.h>
20 #include <net/xdp_sock_drv.h>
21 
22 #define REG_STATE_NEW		0x0
23 #define REG_STATE_REGISTERED	0x1
24 #define REG_STATE_UNREGISTERED	0x2
25 #define REG_STATE_UNUSED	0x3
26 
27 static DEFINE_IDA(mem_id_pool);
28 static DEFINE_MUTEX(mem_id_lock);
29 #define MEM_ID_MAX 0xFFFE
30 #define MEM_ID_MIN 1
31 static int mem_id_next = MEM_ID_MIN;
32 
33 static bool mem_id_init; /* false */
34 static struct rhashtable *mem_id_ht;
35 
36 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
37 {
38 	const u32 *k = data;
39 	const u32 key = *k;
40 
41 	BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
42 		     != sizeof(u32));
43 
44 	/* Use cyclic increasing ID as direct hash key */
45 	return key;
46 }
47 
48 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
49 			  const void *ptr)
50 {
51 	const struct xdp_mem_allocator *xa = ptr;
52 	u32 mem_id = *(u32 *)arg->key;
53 
54 	return xa->mem.id != mem_id;
55 }
56 
57 static const struct rhashtable_params mem_id_rht_params = {
58 	.nelem_hint = 64,
59 	.head_offset = offsetof(struct xdp_mem_allocator, node),
60 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
61 	.key_len = sizeof_field(struct xdp_mem_allocator, mem.id),
62 	.max_size = MEM_ID_MAX,
63 	.min_size = 8,
64 	.automatic_shrinking = true,
65 	.hashfn    = xdp_mem_id_hashfn,
66 	.obj_cmpfn = xdp_mem_id_cmp,
67 };
68 
69 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
70 {
71 	struct xdp_mem_allocator *xa;
72 
73 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
74 
75 	/* Allow this ID to be reused */
76 	ida_simple_remove(&mem_id_pool, xa->mem.id);
77 
78 	kfree(xa);
79 }
80 
81 static void mem_xa_remove(struct xdp_mem_allocator *xa)
82 {
83 	trace_mem_disconnect(xa);
84 
85 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
86 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
87 }
88 
89 static void mem_allocator_disconnect(void *allocator)
90 {
91 	struct xdp_mem_allocator *xa;
92 	struct rhashtable_iter iter;
93 
94 	mutex_lock(&mem_id_lock);
95 
96 	rhashtable_walk_enter(mem_id_ht, &iter);
97 	do {
98 		rhashtable_walk_start(&iter);
99 
100 		while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
101 			if (xa->allocator == allocator)
102 				mem_xa_remove(xa);
103 		}
104 
105 		rhashtable_walk_stop(&iter);
106 
107 	} while (xa == ERR_PTR(-EAGAIN));
108 	rhashtable_walk_exit(&iter);
109 
110 	mutex_unlock(&mem_id_lock);
111 }
112 
113 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
114 {
115 	struct xdp_mem_allocator *xa;
116 	int type = xdp_rxq->mem.type;
117 	int id = xdp_rxq->mem.id;
118 
119 	/* Reset mem info to defaults */
120 	xdp_rxq->mem.id = 0;
121 	xdp_rxq->mem.type = 0;
122 
123 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
124 		WARN(1, "Missing register, driver bug");
125 		return;
126 	}
127 
128 	if (id == 0)
129 		return;
130 
131 	if (type == MEM_TYPE_PAGE_POOL) {
132 		rcu_read_lock();
133 		xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
134 		page_pool_destroy(xa->page_pool);
135 		rcu_read_unlock();
136 	}
137 }
138 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
139 
140 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
141 {
142 	/* Simplify driver cleanup code paths, allow unreg "unused" */
143 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
144 		return;
145 
146 	WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
147 
148 	xdp_rxq_info_unreg_mem_model(xdp_rxq);
149 
150 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
151 	xdp_rxq->dev = NULL;
152 }
153 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
154 
155 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
156 {
157 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
158 }
159 
160 /* Returns 0 on success, negative on failure */
161 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
162 		     struct net_device *dev, u32 queue_index, unsigned int napi_id)
163 {
164 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
165 		WARN(1, "Driver promised not to register this");
166 		return -EINVAL;
167 	}
168 
169 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
170 		WARN(1, "Missing unregister, handled but fix driver");
171 		xdp_rxq_info_unreg(xdp_rxq);
172 	}
173 
174 	if (!dev) {
175 		WARN(1, "Missing net_device from driver");
176 		return -ENODEV;
177 	}
178 
179 	/* State either UNREGISTERED or NEW */
180 	xdp_rxq_info_init(xdp_rxq);
181 	xdp_rxq->dev = dev;
182 	xdp_rxq->queue_index = queue_index;
183 	xdp_rxq->napi_id = napi_id;
184 
185 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
186 	return 0;
187 }
188 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
189 
190 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
191 {
192 	xdp_rxq->reg_state = REG_STATE_UNUSED;
193 }
194 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
195 
196 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
197 {
198 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
199 }
200 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
201 
202 static int __mem_id_init_hash_table(void)
203 {
204 	struct rhashtable *rht;
205 	int ret;
206 
207 	if (unlikely(mem_id_init))
208 		return 0;
209 
210 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
211 	if (!rht)
212 		return -ENOMEM;
213 
214 	ret = rhashtable_init(rht, &mem_id_rht_params);
215 	if (ret < 0) {
216 		kfree(rht);
217 		return ret;
218 	}
219 	mem_id_ht = rht;
220 	smp_mb(); /* mutex lock should provide enough pairing */
221 	mem_id_init = true;
222 
223 	return 0;
224 }
225 
226 /* Allocate a cyclic ID that maps to allocator pointer.
227  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
228  *
229  * Caller must lock mem_id_lock.
230  */
231 static int __mem_id_cyclic_get(gfp_t gfp)
232 {
233 	int retries = 1;
234 	int id;
235 
236 again:
237 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
238 	if (id < 0) {
239 		if (id == -ENOSPC) {
240 			/* Cyclic allocator, reset next id */
241 			if (retries--) {
242 				mem_id_next = MEM_ID_MIN;
243 				goto again;
244 			}
245 		}
246 		return id; /* errno */
247 	}
248 	mem_id_next = id + 1;
249 
250 	return id;
251 }
252 
253 static bool __is_supported_mem_type(enum xdp_mem_type type)
254 {
255 	if (type == MEM_TYPE_PAGE_POOL)
256 		return is_page_pool_compiled_in();
257 
258 	if (type >= MEM_TYPE_MAX)
259 		return false;
260 
261 	return true;
262 }
263 
264 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
265 			       enum xdp_mem_type type, void *allocator)
266 {
267 	struct xdp_mem_allocator *xdp_alloc;
268 	gfp_t gfp = GFP_KERNEL;
269 	int id, errno, ret;
270 	void *ptr;
271 
272 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
273 		WARN(1, "Missing register, driver bug");
274 		return -EFAULT;
275 	}
276 
277 	if (!__is_supported_mem_type(type))
278 		return -EOPNOTSUPP;
279 
280 	xdp_rxq->mem.type = type;
281 
282 	if (!allocator) {
283 		if (type == MEM_TYPE_PAGE_POOL)
284 			return -EINVAL; /* Setup time check page_pool req */
285 		return 0;
286 	}
287 
288 	/* Delay init of rhashtable to save memory if feature isn't used */
289 	if (!mem_id_init) {
290 		mutex_lock(&mem_id_lock);
291 		ret = __mem_id_init_hash_table();
292 		mutex_unlock(&mem_id_lock);
293 		if (ret < 0) {
294 			WARN_ON(1);
295 			return ret;
296 		}
297 	}
298 
299 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
300 	if (!xdp_alloc)
301 		return -ENOMEM;
302 
303 	mutex_lock(&mem_id_lock);
304 	id = __mem_id_cyclic_get(gfp);
305 	if (id < 0) {
306 		errno = id;
307 		goto err;
308 	}
309 	xdp_rxq->mem.id = id;
310 	xdp_alloc->mem  = xdp_rxq->mem;
311 	xdp_alloc->allocator = allocator;
312 
313 	/* Insert allocator into ID lookup table */
314 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
315 	if (IS_ERR(ptr)) {
316 		ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
317 		xdp_rxq->mem.id = 0;
318 		errno = PTR_ERR(ptr);
319 		goto err;
320 	}
321 
322 	if (type == MEM_TYPE_PAGE_POOL)
323 		page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
324 
325 	mutex_unlock(&mem_id_lock);
326 
327 	trace_mem_connect(xdp_alloc, xdp_rxq);
328 	return 0;
329 err:
330 	mutex_unlock(&mem_id_lock);
331 	kfree(xdp_alloc);
332 	return errno;
333 }
334 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
335 
336 /* XDP RX runs under NAPI protection, and in different delivery error
337  * scenarios (e.g. queue full), it is possible to return the xdp_frame
338  * while still leveraging this protection.  The @napi_direct boolean
339  * is used for those calls sites.  Thus, allowing for faster recycling
340  * of xdp_frames/pages in those cases.
341  */
342 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
343 			 struct xdp_buff *xdp)
344 {
345 	struct xdp_mem_allocator *xa;
346 	struct page *page;
347 
348 	switch (mem->type) {
349 	case MEM_TYPE_PAGE_POOL:
350 		rcu_read_lock();
351 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
352 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
353 		page = virt_to_head_page(data);
354 		if (napi_direct && xdp_return_frame_no_direct())
355 			napi_direct = false;
356 		page_pool_put_full_page(xa->page_pool, page, napi_direct);
357 		rcu_read_unlock();
358 		break;
359 	case MEM_TYPE_PAGE_SHARED:
360 		page_frag_free(data);
361 		break;
362 	case MEM_TYPE_PAGE_ORDER0:
363 		page = virt_to_page(data); /* Assumes order0 page*/
364 		put_page(page);
365 		break;
366 	case MEM_TYPE_XSK_BUFF_POOL:
367 		/* NB! Only valid from an xdp_buff! */
368 		xsk_buff_free(xdp);
369 		break;
370 	default:
371 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
372 		WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
373 		break;
374 	}
375 }
376 
377 void xdp_return_frame(struct xdp_frame *xdpf)
378 {
379 	__xdp_return(xdpf->data, &xdpf->mem, false, NULL);
380 }
381 EXPORT_SYMBOL_GPL(xdp_return_frame);
382 
383 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
384 {
385 	__xdp_return(xdpf->data, &xdpf->mem, true, NULL);
386 }
387 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
388 
389 /* XDP bulk APIs introduce a defer/flush mechanism to return
390  * pages belonging to the same xdp_mem_allocator object
391  * (identified via the mem.id field) in bulk to optimize
392  * I-cache and D-cache.
393  * The bulk queue size is set to 16 to be aligned to how
394  * XDP_REDIRECT bulking works. The bulk is flushed when
395  * it is full or when mem.id changes.
396  * xdp_frame_bulk is usually stored/allocated on the function
397  * call-stack to avoid locking penalties.
398  */
399 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
400 {
401 	struct xdp_mem_allocator *xa = bq->xa;
402 
403 	if (unlikely(!xa || !bq->count))
404 		return;
405 
406 	page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
407 	/* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
408 	bq->count = 0;
409 }
410 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
411 
412 /* Must be called with rcu_read_lock held */
413 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
414 			   struct xdp_frame_bulk *bq)
415 {
416 	struct xdp_mem_info *mem = &xdpf->mem;
417 	struct xdp_mem_allocator *xa;
418 
419 	if (mem->type != MEM_TYPE_PAGE_POOL) {
420 		__xdp_return(xdpf->data, &xdpf->mem, false, NULL);
421 		return;
422 	}
423 
424 	xa = bq->xa;
425 	if (unlikely(!xa)) {
426 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
427 		bq->count = 0;
428 		bq->xa = xa;
429 	}
430 
431 	if (bq->count == XDP_BULK_QUEUE_SIZE)
432 		xdp_flush_frame_bulk(bq);
433 
434 	if (unlikely(mem->id != xa->mem.id)) {
435 		xdp_flush_frame_bulk(bq);
436 		bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
437 	}
438 
439 	bq->q[bq->count++] = xdpf->data;
440 }
441 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
442 
443 void xdp_return_buff(struct xdp_buff *xdp)
444 {
445 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
446 }
447 
448 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
449 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
450 {
451 	struct xdp_mem_allocator *xa;
452 	struct page *page;
453 
454 	rcu_read_lock();
455 	xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
456 	page = virt_to_head_page(data);
457 	if (xa)
458 		page_pool_release_page(xa->page_pool, page);
459 	rcu_read_unlock();
460 }
461 EXPORT_SYMBOL_GPL(__xdp_release_frame);
462 
463 void xdp_attachment_setup(struct xdp_attachment_info *info,
464 			  struct netdev_bpf *bpf)
465 {
466 	if (info->prog)
467 		bpf_prog_put(info->prog);
468 	info->prog = bpf->prog;
469 	info->flags = bpf->flags;
470 }
471 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
472 
473 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
474 {
475 	unsigned int metasize, totsize;
476 	void *addr, *data_to_copy;
477 	struct xdp_frame *xdpf;
478 	struct page *page;
479 
480 	/* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
481 	metasize = xdp_data_meta_unsupported(xdp) ? 0 :
482 		   xdp->data - xdp->data_meta;
483 	totsize = xdp->data_end - xdp->data + metasize;
484 
485 	if (sizeof(*xdpf) + totsize > PAGE_SIZE)
486 		return NULL;
487 
488 	page = dev_alloc_page();
489 	if (!page)
490 		return NULL;
491 
492 	addr = page_to_virt(page);
493 	xdpf = addr;
494 	memset(xdpf, 0, sizeof(*xdpf));
495 
496 	addr += sizeof(*xdpf);
497 	data_to_copy = metasize ? xdp->data_meta : xdp->data;
498 	memcpy(addr, data_to_copy, totsize);
499 
500 	xdpf->data = addr + metasize;
501 	xdpf->len = totsize - metasize;
502 	xdpf->headroom = 0;
503 	xdpf->metasize = metasize;
504 	xdpf->frame_sz = PAGE_SIZE;
505 	xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
506 
507 	xsk_buff_free(xdp);
508 	return xdpf;
509 }
510 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
511 
512 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
513 void xdp_warn(const char *msg, const char *func, const int line)
514 {
515 	WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
516 };
517 EXPORT_SYMBOL_GPL(xdp_warn);
518 
519 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
520 {
521 	n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
522 				      n_skb, skbs);
523 	if (unlikely(!n_skb))
524 		return -ENOMEM;
525 
526 	return 0;
527 }
528 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
529 
530 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
531 					   struct sk_buff *skb,
532 					   struct net_device *dev)
533 {
534 	unsigned int headroom, frame_size;
535 	void *hard_start;
536 
537 	/* Part of headroom was reserved to xdpf */
538 	headroom = sizeof(*xdpf) + xdpf->headroom;
539 
540 	/* Memory size backing xdp_frame data already have reserved
541 	 * room for build_skb to place skb_shared_info in tailroom.
542 	 */
543 	frame_size = xdpf->frame_sz;
544 
545 	hard_start = xdpf->data - headroom;
546 	skb = build_skb_around(skb, hard_start, frame_size);
547 	if (unlikely(!skb))
548 		return NULL;
549 
550 	skb_reserve(skb, headroom);
551 	__skb_put(skb, xdpf->len);
552 	if (xdpf->metasize)
553 		skb_metadata_set(skb, xdpf->metasize);
554 
555 	/* Essential SKB info: protocol and skb->dev */
556 	skb->protocol = eth_type_trans(skb, dev);
557 
558 	/* Optional SKB info, currently missing:
559 	 * - HW checksum info		(skb->ip_summed)
560 	 * - HW RX hash			(skb_set_hash)
561 	 * - RX ring dev queue index	(skb_record_rx_queue)
562 	 */
563 
564 	/* Until page_pool get SKB return path, release DMA here */
565 	xdp_release_frame(xdpf);
566 
567 	/* Allow SKB to reuse area used by xdp_frame */
568 	xdp_scrub_frame(xdpf);
569 
570 	return skb;
571 }
572 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
573 
574 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
575 					 struct net_device *dev)
576 {
577 	struct sk_buff *skb;
578 
579 	skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
580 	if (unlikely(!skb))
581 		return NULL;
582 
583 	memset(skb, 0, offsetof(struct sk_buff, tail));
584 
585 	return __xdp_build_skb_from_frame(xdpf, skb, dev);
586 }
587 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
588 
589 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
590 {
591 	unsigned int headroom, totalsize;
592 	struct xdp_frame *nxdpf;
593 	struct page *page;
594 	void *addr;
595 
596 	headroom = xdpf->headroom + sizeof(*xdpf);
597 	totalsize = headroom + xdpf->len;
598 
599 	if (unlikely(totalsize > PAGE_SIZE))
600 		return NULL;
601 	page = dev_alloc_page();
602 	if (!page)
603 		return NULL;
604 	addr = page_to_virt(page);
605 
606 	memcpy(addr, xdpf, totalsize);
607 
608 	nxdpf = addr;
609 	nxdpf->data = addr + headroom;
610 	nxdpf->frame_sz = PAGE_SIZE;
611 	nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
612 	nxdpf->mem.id = 0;
613 
614 	return nxdpf;
615 }
616