xref: /openbmc/linux/net/core/xdp.c (revision 586b4106)
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/btf.h>
8 #include <linux/btf_ids.h>
9 #include <linux/filter.h>
10 #include <linux/types.h>
11 #include <linux/mm.h>
12 #include <linux/netdevice.h>
13 #include <linux/slab.h>
14 #include <linux/idr.h>
15 #include <linux/rhashtable.h>
16 #include <linux/bug.h>
17 #include <net/page_pool/helpers.h>
18 
19 #include <net/xdp.h>
20 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
21 #include <trace/events/xdp.h>
22 #include <net/xdp_sock_drv.h>
23 
24 #define REG_STATE_NEW		0x0
25 #define REG_STATE_REGISTERED	0x1
26 #define REG_STATE_UNREGISTERED	0x2
27 #define REG_STATE_UNUSED	0x3
28 
29 static DEFINE_IDA(mem_id_pool);
30 static DEFINE_MUTEX(mem_id_lock);
31 #define MEM_ID_MAX 0xFFFE
32 #define MEM_ID_MIN 1
33 static int mem_id_next = MEM_ID_MIN;
34 
35 static bool mem_id_init; /* false */
36 static struct rhashtable *mem_id_ht;
37 
38 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
39 {
40 	const u32 *k = data;
41 	const u32 key = *k;
42 
43 	BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
44 		     != sizeof(u32));
45 
46 	/* Use cyclic increasing ID as direct hash key */
47 	return key;
48 }
49 
50 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
51 			  const void *ptr)
52 {
53 	const struct xdp_mem_allocator *xa = ptr;
54 	u32 mem_id = *(u32 *)arg->key;
55 
56 	return xa->mem.id != mem_id;
57 }
58 
59 static const struct rhashtable_params mem_id_rht_params = {
60 	.nelem_hint = 64,
61 	.head_offset = offsetof(struct xdp_mem_allocator, node),
62 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
63 	.key_len = sizeof_field(struct xdp_mem_allocator, mem.id),
64 	.max_size = MEM_ID_MAX,
65 	.min_size = 8,
66 	.automatic_shrinking = true,
67 	.hashfn    = xdp_mem_id_hashfn,
68 	.obj_cmpfn = xdp_mem_id_cmp,
69 };
70 
71 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
72 {
73 	struct xdp_mem_allocator *xa;
74 
75 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
76 
77 	/* Allow this ID to be reused */
78 	ida_simple_remove(&mem_id_pool, xa->mem.id);
79 
80 	kfree(xa);
81 }
82 
83 static void mem_xa_remove(struct xdp_mem_allocator *xa)
84 {
85 	trace_mem_disconnect(xa);
86 
87 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
88 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
89 }
90 
91 static void mem_allocator_disconnect(void *allocator)
92 {
93 	struct xdp_mem_allocator *xa;
94 	struct rhashtable_iter iter;
95 
96 	mutex_lock(&mem_id_lock);
97 
98 	rhashtable_walk_enter(mem_id_ht, &iter);
99 	do {
100 		rhashtable_walk_start(&iter);
101 
102 		while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
103 			if (xa->allocator == allocator)
104 				mem_xa_remove(xa);
105 		}
106 
107 		rhashtable_walk_stop(&iter);
108 
109 	} while (xa == ERR_PTR(-EAGAIN));
110 	rhashtable_walk_exit(&iter);
111 
112 	mutex_unlock(&mem_id_lock);
113 }
114 
115 void xdp_unreg_mem_model(struct xdp_mem_info *mem)
116 {
117 	struct xdp_mem_allocator *xa;
118 	int type = mem->type;
119 	int id = mem->id;
120 
121 	/* Reset mem info to defaults */
122 	mem->id = 0;
123 	mem->type = 0;
124 
125 	if (id == 0)
126 		return;
127 
128 	if (type == MEM_TYPE_PAGE_POOL) {
129 		rcu_read_lock();
130 		xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
131 		page_pool_destroy(xa->page_pool);
132 		rcu_read_unlock();
133 	}
134 }
135 EXPORT_SYMBOL_GPL(xdp_unreg_mem_model);
136 
137 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
138 {
139 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
140 		WARN(1, "Missing register, driver bug");
141 		return;
142 	}
143 
144 	xdp_unreg_mem_model(&xdp_rxq->mem);
145 }
146 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
147 
148 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
149 {
150 	/* Simplify driver cleanup code paths, allow unreg "unused" */
151 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
152 		return;
153 
154 	xdp_rxq_info_unreg_mem_model(xdp_rxq);
155 
156 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
157 	xdp_rxq->dev = NULL;
158 }
159 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
160 
161 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
162 {
163 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
164 }
165 
166 /* Returns 0 on success, negative on failure */
167 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
168 		       struct net_device *dev, u32 queue_index,
169 		       unsigned int napi_id, u32 frag_size)
170 {
171 	if (!dev) {
172 		WARN(1, "Missing net_device from driver");
173 		return -ENODEV;
174 	}
175 
176 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
177 		WARN(1, "Driver promised not to register this");
178 		return -EINVAL;
179 	}
180 
181 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
182 		WARN(1, "Missing unregister, handled but fix driver");
183 		xdp_rxq_info_unreg(xdp_rxq);
184 	}
185 
186 	/* State either UNREGISTERED or NEW */
187 	xdp_rxq_info_init(xdp_rxq);
188 	xdp_rxq->dev = dev;
189 	xdp_rxq->queue_index = queue_index;
190 	xdp_rxq->napi_id = napi_id;
191 	xdp_rxq->frag_size = frag_size;
192 
193 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
194 	return 0;
195 }
196 EXPORT_SYMBOL_GPL(__xdp_rxq_info_reg);
197 
198 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
199 {
200 	xdp_rxq->reg_state = REG_STATE_UNUSED;
201 }
202 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
203 
204 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
205 {
206 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
207 }
208 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
209 
210 static int __mem_id_init_hash_table(void)
211 {
212 	struct rhashtable *rht;
213 	int ret;
214 
215 	if (unlikely(mem_id_init))
216 		return 0;
217 
218 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
219 	if (!rht)
220 		return -ENOMEM;
221 
222 	ret = rhashtable_init(rht, &mem_id_rht_params);
223 	if (ret < 0) {
224 		kfree(rht);
225 		return ret;
226 	}
227 	mem_id_ht = rht;
228 	smp_mb(); /* mutex lock should provide enough pairing */
229 	mem_id_init = true;
230 
231 	return 0;
232 }
233 
234 /* Allocate a cyclic ID that maps to allocator pointer.
235  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
236  *
237  * Caller must lock mem_id_lock.
238  */
239 static int __mem_id_cyclic_get(gfp_t gfp)
240 {
241 	int retries = 1;
242 	int id;
243 
244 again:
245 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
246 	if (id < 0) {
247 		if (id == -ENOSPC) {
248 			/* Cyclic allocator, reset next id */
249 			if (retries--) {
250 				mem_id_next = MEM_ID_MIN;
251 				goto again;
252 			}
253 		}
254 		return id; /* errno */
255 	}
256 	mem_id_next = id + 1;
257 
258 	return id;
259 }
260 
261 static bool __is_supported_mem_type(enum xdp_mem_type type)
262 {
263 	if (type == MEM_TYPE_PAGE_POOL)
264 		return is_page_pool_compiled_in();
265 
266 	if (type >= MEM_TYPE_MAX)
267 		return false;
268 
269 	return true;
270 }
271 
272 static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem,
273 						     enum xdp_mem_type type,
274 						     void *allocator)
275 {
276 	struct xdp_mem_allocator *xdp_alloc;
277 	gfp_t gfp = GFP_KERNEL;
278 	int id, errno, ret;
279 	void *ptr;
280 
281 	if (!__is_supported_mem_type(type))
282 		return ERR_PTR(-EOPNOTSUPP);
283 
284 	mem->type = type;
285 
286 	if (!allocator) {
287 		if (type == MEM_TYPE_PAGE_POOL)
288 			return ERR_PTR(-EINVAL); /* Setup time check page_pool req */
289 		return NULL;
290 	}
291 
292 	/* Delay init of rhashtable to save memory if feature isn't used */
293 	if (!mem_id_init) {
294 		mutex_lock(&mem_id_lock);
295 		ret = __mem_id_init_hash_table();
296 		mutex_unlock(&mem_id_lock);
297 		if (ret < 0)
298 			return ERR_PTR(ret);
299 	}
300 
301 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
302 	if (!xdp_alloc)
303 		return ERR_PTR(-ENOMEM);
304 
305 	mutex_lock(&mem_id_lock);
306 	id = __mem_id_cyclic_get(gfp);
307 	if (id < 0) {
308 		errno = id;
309 		goto err;
310 	}
311 	mem->id = id;
312 	xdp_alloc->mem = *mem;
313 	xdp_alloc->allocator = allocator;
314 
315 	/* Insert allocator into ID lookup table */
316 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
317 	if (IS_ERR(ptr)) {
318 		ida_simple_remove(&mem_id_pool, mem->id);
319 		mem->id = 0;
320 		errno = PTR_ERR(ptr);
321 		goto err;
322 	}
323 
324 	if (type == MEM_TYPE_PAGE_POOL)
325 		page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem);
326 
327 	mutex_unlock(&mem_id_lock);
328 
329 	return xdp_alloc;
330 err:
331 	mutex_unlock(&mem_id_lock);
332 	kfree(xdp_alloc);
333 	return ERR_PTR(errno);
334 }
335 
336 int xdp_reg_mem_model(struct xdp_mem_info *mem,
337 		      enum xdp_mem_type type, void *allocator)
338 {
339 	struct xdp_mem_allocator *xdp_alloc;
340 
341 	xdp_alloc = __xdp_reg_mem_model(mem, type, allocator);
342 	if (IS_ERR(xdp_alloc))
343 		return PTR_ERR(xdp_alloc);
344 	return 0;
345 }
346 EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
347 
348 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
349 			       enum xdp_mem_type type, void *allocator)
350 {
351 	struct xdp_mem_allocator *xdp_alloc;
352 
353 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
354 		WARN(1, "Missing register, driver bug");
355 		return -EFAULT;
356 	}
357 
358 	xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator);
359 	if (IS_ERR(xdp_alloc))
360 		return PTR_ERR(xdp_alloc);
361 
362 	if (trace_mem_connect_enabled() && xdp_alloc)
363 		trace_mem_connect(xdp_alloc, xdp_rxq);
364 	return 0;
365 }
366 
367 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
368 
369 /* XDP RX runs under NAPI protection, and in different delivery error
370  * scenarios (e.g. queue full), it is possible to return the xdp_frame
371  * while still leveraging this protection.  The @napi_direct boolean
372  * is used for those calls sites.  Thus, allowing for faster recycling
373  * of xdp_frames/pages in those cases.
374  */
375 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
376 		  struct xdp_buff *xdp)
377 {
378 	struct page *page;
379 
380 	switch (mem->type) {
381 	case MEM_TYPE_PAGE_POOL:
382 		page = virt_to_head_page(data);
383 		if (napi_direct && xdp_return_frame_no_direct())
384 			napi_direct = false;
385 		/* No need to check ((page->pp_magic & ~0x3UL) == PP_SIGNATURE)
386 		 * as mem->type knows this a page_pool page
387 		 */
388 		page_pool_put_full_page(page->pp, page, napi_direct);
389 		break;
390 	case MEM_TYPE_PAGE_SHARED:
391 		page_frag_free(data);
392 		break;
393 	case MEM_TYPE_PAGE_ORDER0:
394 		page = virt_to_page(data); /* Assumes order0 page*/
395 		put_page(page);
396 		break;
397 	case MEM_TYPE_XSK_BUFF_POOL:
398 		/* NB! Only valid from an xdp_buff! */
399 		xsk_buff_free(xdp);
400 		break;
401 	default:
402 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
403 		WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
404 		break;
405 	}
406 }
407 
408 void xdp_return_frame(struct xdp_frame *xdpf)
409 {
410 	struct skb_shared_info *sinfo;
411 	int i;
412 
413 	if (likely(!xdp_frame_has_frags(xdpf)))
414 		goto out;
415 
416 	sinfo = xdp_get_shared_info_from_frame(xdpf);
417 	for (i = 0; i < sinfo->nr_frags; i++) {
418 		struct page *page = skb_frag_page(&sinfo->frags[i]);
419 
420 		__xdp_return(page_address(page), &xdpf->mem, false, NULL);
421 	}
422 out:
423 	__xdp_return(xdpf->data, &xdpf->mem, false, NULL);
424 }
425 EXPORT_SYMBOL_GPL(xdp_return_frame);
426 
427 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
428 {
429 	struct skb_shared_info *sinfo;
430 	int i;
431 
432 	if (likely(!xdp_frame_has_frags(xdpf)))
433 		goto out;
434 
435 	sinfo = xdp_get_shared_info_from_frame(xdpf);
436 	for (i = 0; i < sinfo->nr_frags; i++) {
437 		struct page *page = skb_frag_page(&sinfo->frags[i]);
438 
439 		__xdp_return(page_address(page), &xdpf->mem, true, NULL);
440 	}
441 out:
442 	__xdp_return(xdpf->data, &xdpf->mem, true, NULL);
443 }
444 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
445 
446 /* XDP bulk APIs introduce a defer/flush mechanism to return
447  * pages belonging to the same xdp_mem_allocator object
448  * (identified via the mem.id field) in bulk to optimize
449  * I-cache and D-cache.
450  * The bulk queue size is set to 16 to be aligned to how
451  * XDP_REDIRECT bulking works. The bulk is flushed when
452  * it is full or when mem.id changes.
453  * xdp_frame_bulk is usually stored/allocated on the function
454  * call-stack to avoid locking penalties.
455  */
456 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
457 {
458 	struct xdp_mem_allocator *xa = bq->xa;
459 
460 	if (unlikely(!xa || !bq->count))
461 		return;
462 
463 	page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
464 	/* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
465 	bq->count = 0;
466 }
467 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
468 
469 /* Must be called with rcu_read_lock held */
470 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
471 			   struct xdp_frame_bulk *bq)
472 {
473 	struct xdp_mem_info *mem = &xdpf->mem;
474 	struct xdp_mem_allocator *xa;
475 
476 	if (mem->type != MEM_TYPE_PAGE_POOL) {
477 		xdp_return_frame(xdpf);
478 		return;
479 	}
480 
481 	xa = bq->xa;
482 	if (unlikely(!xa)) {
483 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
484 		bq->count = 0;
485 		bq->xa = xa;
486 	}
487 
488 	if (bq->count == XDP_BULK_QUEUE_SIZE)
489 		xdp_flush_frame_bulk(bq);
490 
491 	if (unlikely(mem->id != xa->mem.id)) {
492 		xdp_flush_frame_bulk(bq);
493 		bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
494 	}
495 
496 	if (unlikely(xdp_frame_has_frags(xdpf))) {
497 		struct skb_shared_info *sinfo;
498 		int i;
499 
500 		sinfo = xdp_get_shared_info_from_frame(xdpf);
501 		for (i = 0; i < sinfo->nr_frags; i++) {
502 			skb_frag_t *frag = &sinfo->frags[i];
503 
504 			bq->q[bq->count++] = skb_frag_address(frag);
505 			if (bq->count == XDP_BULK_QUEUE_SIZE)
506 				xdp_flush_frame_bulk(bq);
507 		}
508 	}
509 	bq->q[bq->count++] = xdpf->data;
510 }
511 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
512 
513 void xdp_return_buff(struct xdp_buff *xdp)
514 {
515 	struct skb_shared_info *sinfo;
516 	int i;
517 
518 	if (likely(!xdp_buff_has_frags(xdp)))
519 		goto out;
520 
521 	sinfo = xdp_get_shared_info_from_buff(xdp);
522 	for (i = 0; i < sinfo->nr_frags; i++) {
523 		struct page *page = skb_frag_page(&sinfo->frags[i]);
524 
525 		__xdp_return(page_address(page), &xdp->rxq->mem, true, xdp);
526 	}
527 out:
528 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
529 }
530 EXPORT_SYMBOL_GPL(xdp_return_buff);
531 
532 void xdp_attachment_setup(struct xdp_attachment_info *info,
533 			  struct netdev_bpf *bpf)
534 {
535 	if (info->prog)
536 		bpf_prog_put(info->prog);
537 	info->prog = bpf->prog;
538 	info->flags = bpf->flags;
539 }
540 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
541 
542 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
543 {
544 	unsigned int metasize, totsize;
545 	void *addr, *data_to_copy;
546 	struct xdp_frame *xdpf;
547 	struct page *page;
548 
549 	/* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
550 	metasize = xdp_data_meta_unsupported(xdp) ? 0 :
551 		   xdp->data - xdp->data_meta;
552 	totsize = xdp->data_end - xdp->data + metasize;
553 
554 	if (sizeof(*xdpf) + totsize > PAGE_SIZE)
555 		return NULL;
556 
557 	page = dev_alloc_page();
558 	if (!page)
559 		return NULL;
560 
561 	addr = page_to_virt(page);
562 	xdpf = addr;
563 	memset(xdpf, 0, sizeof(*xdpf));
564 
565 	addr += sizeof(*xdpf);
566 	data_to_copy = metasize ? xdp->data_meta : xdp->data;
567 	memcpy(addr, data_to_copy, totsize);
568 
569 	xdpf->data = addr + metasize;
570 	xdpf->len = totsize - metasize;
571 	xdpf->headroom = 0;
572 	xdpf->metasize = metasize;
573 	xdpf->frame_sz = PAGE_SIZE;
574 	xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
575 
576 	xsk_buff_free(xdp);
577 	return xdpf;
578 }
579 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
580 
581 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
582 void xdp_warn(const char *msg, const char *func, const int line)
583 {
584 	WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
585 };
586 EXPORT_SYMBOL_GPL(xdp_warn);
587 
588 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
589 {
590 	n_skb = kmem_cache_alloc_bulk(skbuff_cache, gfp, n_skb, skbs);
591 	if (unlikely(!n_skb))
592 		return -ENOMEM;
593 
594 	return 0;
595 }
596 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
597 
598 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
599 					   struct sk_buff *skb,
600 					   struct net_device *dev)
601 {
602 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
603 	unsigned int headroom, frame_size;
604 	void *hard_start;
605 	u8 nr_frags;
606 
607 	/* xdp frags frame */
608 	if (unlikely(xdp_frame_has_frags(xdpf)))
609 		nr_frags = sinfo->nr_frags;
610 
611 	/* Part of headroom was reserved to xdpf */
612 	headroom = sizeof(*xdpf) + xdpf->headroom;
613 
614 	/* Memory size backing xdp_frame data already have reserved
615 	 * room for build_skb to place skb_shared_info in tailroom.
616 	 */
617 	frame_size = xdpf->frame_sz;
618 
619 	hard_start = xdpf->data - headroom;
620 	skb = build_skb_around(skb, hard_start, frame_size);
621 	if (unlikely(!skb))
622 		return NULL;
623 
624 	skb_reserve(skb, headroom);
625 	__skb_put(skb, xdpf->len);
626 	if (xdpf->metasize)
627 		skb_metadata_set(skb, xdpf->metasize);
628 
629 	if (unlikely(xdp_frame_has_frags(xdpf)))
630 		xdp_update_skb_shared_info(skb, nr_frags,
631 					   sinfo->xdp_frags_size,
632 					   nr_frags * xdpf->frame_sz,
633 					   xdp_frame_is_frag_pfmemalloc(xdpf));
634 
635 	/* Essential SKB info: protocol and skb->dev */
636 	skb->protocol = eth_type_trans(skb, dev);
637 
638 	/* Optional SKB info, currently missing:
639 	 * - HW checksum info		(skb->ip_summed)
640 	 * - HW RX hash			(skb_set_hash)
641 	 * - RX ring dev queue index	(skb_record_rx_queue)
642 	 */
643 
644 	if (xdpf->mem.type == MEM_TYPE_PAGE_POOL)
645 		skb_mark_for_recycle(skb);
646 
647 	/* Allow SKB to reuse area used by xdp_frame */
648 	xdp_scrub_frame(xdpf);
649 
650 	return skb;
651 }
652 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
653 
654 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
655 					 struct net_device *dev)
656 {
657 	struct sk_buff *skb;
658 
659 	skb = kmem_cache_alloc(skbuff_cache, GFP_ATOMIC);
660 	if (unlikely(!skb))
661 		return NULL;
662 
663 	memset(skb, 0, offsetof(struct sk_buff, tail));
664 
665 	return __xdp_build_skb_from_frame(xdpf, skb, dev);
666 }
667 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
668 
669 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
670 {
671 	unsigned int headroom, totalsize;
672 	struct xdp_frame *nxdpf;
673 	struct page *page;
674 	void *addr;
675 
676 	headroom = xdpf->headroom + sizeof(*xdpf);
677 	totalsize = headroom + xdpf->len;
678 
679 	if (unlikely(totalsize > PAGE_SIZE))
680 		return NULL;
681 	page = dev_alloc_page();
682 	if (!page)
683 		return NULL;
684 	addr = page_to_virt(page);
685 
686 	memcpy(addr, xdpf, totalsize);
687 
688 	nxdpf = addr;
689 	nxdpf->data = addr + headroom;
690 	nxdpf->frame_sz = PAGE_SIZE;
691 	nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
692 	nxdpf->mem.id = 0;
693 
694 	return nxdpf;
695 }
696 
697 __diag_push();
698 __diag_ignore_all("-Wmissing-prototypes",
699 		  "Global functions as their definitions will be in vmlinux BTF");
700 
701 /**
702  * bpf_xdp_metadata_rx_timestamp - Read XDP frame RX timestamp.
703  * @ctx: XDP context pointer.
704  * @timestamp: Return value pointer.
705  *
706  * Return:
707  * * Returns 0 on success or ``-errno`` on error.
708  * * ``-EOPNOTSUPP`` : means device driver does not implement kfunc
709  * * ``-ENODATA``    : means no RX-timestamp available for this frame
710  */
711 __bpf_kfunc int bpf_xdp_metadata_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
712 {
713 	return -EOPNOTSUPP;
714 }
715 
716 /**
717  * bpf_xdp_metadata_rx_hash - Read XDP frame RX hash.
718  * @ctx: XDP context pointer.
719  * @hash: Return value pointer.
720  * @rss_type: Return value pointer for RSS type.
721  *
722  * The RSS hash type (@rss_type) specifies what portion of packet headers NIC
723  * hardware used when calculating RSS hash value.  The RSS type can be decoded
724  * via &enum xdp_rss_hash_type either matching on individual L3/L4 bits
725  * ``XDP_RSS_L*`` or by combined traditional *RSS Hashing Types*
726  * ``XDP_RSS_TYPE_L*``.
727  *
728  * Return:
729  * * Returns 0 on success or ``-errno`` on error.
730  * * ``-EOPNOTSUPP`` : means device driver doesn't implement kfunc
731  * * ``-ENODATA``    : means no RX-hash available for this frame
732  */
733 __bpf_kfunc int bpf_xdp_metadata_rx_hash(const struct xdp_md *ctx, u32 *hash,
734 					 enum xdp_rss_hash_type *rss_type)
735 {
736 	return -EOPNOTSUPP;
737 }
738 
739 __diag_pop();
740 
741 BTF_SET8_START(xdp_metadata_kfunc_ids)
742 #define XDP_METADATA_KFUNC(_, name) BTF_ID_FLAGS(func, name, KF_TRUSTED_ARGS)
743 XDP_METADATA_KFUNC_xxx
744 #undef XDP_METADATA_KFUNC
745 BTF_SET8_END(xdp_metadata_kfunc_ids)
746 
747 static const struct btf_kfunc_id_set xdp_metadata_kfunc_set = {
748 	.owner = THIS_MODULE,
749 	.set   = &xdp_metadata_kfunc_ids,
750 };
751 
752 BTF_ID_LIST(xdp_metadata_kfunc_ids_unsorted)
753 #define XDP_METADATA_KFUNC(name, str) BTF_ID(func, str)
754 XDP_METADATA_KFUNC_xxx
755 #undef XDP_METADATA_KFUNC
756 
757 u32 bpf_xdp_metadata_kfunc_id(int id)
758 {
759 	/* xdp_metadata_kfunc_ids is sorted and can't be used */
760 	return xdp_metadata_kfunc_ids_unsorted[id];
761 }
762 
763 bool bpf_dev_bound_kfunc_id(u32 btf_id)
764 {
765 	return btf_id_set8_contains(&xdp_metadata_kfunc_ids, btf_id);
766 }
767 
768 static int __init xdp_metadata_init(void)
769 {
770 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_XDP, &xdp_metadata_kfunc_set);
771 }
772 late_initcall(xdp_metadata_init);
773 
774 void xdp_set_features_flag(struct net_device *dev, xdp_features_t val)
775 {
776 	val &= NETDEV_XDP_ACT_MASK;
777 	if (dev->xdp_features == val)
778 		return;
779 
780 	dev->xdp_features = val;
781 
782 	if (dev->reg_state == NETREG_REGISTERED)
783 		call_netdevice_notifiers(NETDEV_XDP_FEAT_CHANGE, dev);
784 }
785 EXPORT_SYMBOL_GPL(xdp_set_features_flag);
786 
787 void xdp_features_set_redirect_target(struct net_device *dev, bool support_sg)
788 {
789 	xdp_features_t val = (dev->xdp_features | NETDEV_XDP_ACT_NDO_XMIT);
790 
791 	if (support_sg)
792 		val |= NETDEV_XDP_ACT_NDO_XMIT_SG;
793 	xdp_set_features_flag(dev, val);
794 }
795 EXPORT_SYMBOL_GPL(xdp_features_set_redirect_target);
796 
797 void xdp_features_clear_redirect_target(struct net_device *dev)
798 {
799 	xdp_features_t val = dev->xdp_features;
800 
801 	val &= ~(NETDEV_XDP_ACT_NDO_XMIT | NETDEV_XDP_ACT_NDO_XMIT_SG);
802 	xdp_set_features_flag(dev, val);
803 }
804 EXPORT_SYMBOL_GPL(xdp_features_clear_redirect_target);
805