xref: /openbmc/linux/net/core/xdp.c (revision a266ef69b890f099069cf51bb40572611c435a54)
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_ids.h>
8 #include <linux/filter.h>
9 #include <linux/types.h>
10 #include <linux/mm.h>
11 #include <linux/netdevice.h>
12 #include <linux/slab.h>
13 #include <linux/idr.h>
14 #include <linux/rhashtable.h>
15 #include <linux/bug.h>
16 #include <net/page_pool.h>
17 
18 #include <net/xdp.h>
19 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
20 #include <trace/events/xdp.h>
21 #include <net/xdp_sock_drv.h>
22 
23 #define REG_STATE_NEW		0x0
24 #define REG_STATE_REGISTERED	0x1
25 #define REG_STATE_UNREGISTERED	0x2
26 #define REG_STATE_UNUSED	0x3
27 
28 static DEFINE_IDA(mem_id_pool);
29 static DEFINE_MUTEX(mem_id_lock);
30 #define MEM_ID_MAX 0xFFFE
31 #define MEM_ID_MIN 1
32 static int mem_id_next = MEM_ID_MIN;
33 
34 static bool mem_id_init; /* false */
35 static struct rhashtable *mem_id_ht;
36 
37 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
38 {
39 	const u32 *k = data;
40 	const u32 key = *k;
41 
42 	BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
43 		     != sizeof(u32));
44 
45 	/* Use cyclic increasing ID as direct hash key */
46 	return key;
47 }
48 
49 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
50 			  const void *ptr)
51 {
52 	const struct xdp_mem_allocator *xa = ptr;
53 	u32 mem_id = *(u32 *)arg->key;
54 
55 	return xa->mem.id != mem_id;
56 }
57 
58 static const struct rhashtable_params mem_id_rht_params = {
59 	.nelem_hint = 64,
60 	.head_offset = offsetof(struct xdp_mem_allocator, node),
61 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
62 	.key_len = sizeof_field(struct xdp_mem_allocator, mem.id),
63 	.max_size = MEM_ID_MAX,
64 	.min_size = 8,
65 	.automatic_shrinking = true,
66 	.hashfn    = xdp_mem_id_hashfn,
67 	.obj_cmpfn = xdp_mem_id_cmp,
68 };
69 
70 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
71 {
72 	struct xdp_mem_allocator *xa;
73 
74 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
75 
76 	/* Allow this ID to be reused */
77 	ida_simple_remove(&mem_id_pool, xa->mem.id);
78 
79 	kfree(xa);
80 }
81 
82 static void mem_xa_remove(struct xdp_mem_allocator *xa)
83 {
84 	trace_mem_disconnect(xa);
85 
86 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
87 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
88 }
89 
90 static void mem_allocator_disconnect(void *allocator)
91 {
92 	struct xdp_mem_allocator *xa;
93 	struct rhashtable_iter iter;
94 
95 	mutex_lock(&mem_id_lock);
96 
97 	rhashtable_walk_enter(mem_id_ht, &iter);
98 	do {
99 		rhashtable_walk_start(&iter);
100 
101 		while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
102 			if (xa->allocator == allocator)
103 				mem_xa_remove(xa);
104 		}
105 
106 		rhashtable_walk_stop(&iter);
107 
108 	} while (xa == ERR_PTR(-EAGAIN));
109 	rhashtable_walk_exit(&iter);
110 
111 	mutex_unlock(&mem_id_lock);
112 }
113 
114 void xdp_unreg_mem_model(struct xdp_mem_info *mem)
115 {
116 	struct xdp_mem_allocator *xa;
117 	int type = mem->type;
118 	int id = mem->id;
119 
120 	/* Reset mem info to defaults */
121 	mem->id = 0;
122 	mem->type = 0;
123 
124 	if (id == 0)
125 		return;
126 
127 	if (type == MEM_TYPE_PAGE_POOL) {
128 		rcu_read_lock();
129 		xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
130 		page_pool_destroy(xa->page_pool);
131 		rcu_read_unlock();
132 	}
133 }
134 EXPORT_SYMBOL_GPL(xdp_unreg_mem_model);
135 
136 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
137 {
138 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
139 		WARN(1, "Missing register, driver bug");
140 		return;
141 	}
142 
143 	xdp_unreg_mem_model(&xdp_rxq->mem);
144 }
145 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
146 
147 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
148 {
149 	/* Simplify driver cleanup code paths, allow unreg "unused" */
150 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
151 		return;
152 
153 	xdp_rxq_info_unreg_mem_model(xdp_rxq);
154 
155 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
156 	xdp_rxq->dev = NULL;
157 }
158 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
159 
160 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
161 {
162 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
163 }
164 
165 /* Returns 0 on success, negative on failure */
166 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
167 		       struct net_device *dev, u32 queue_index,
168 		       unsigned int napi_id, u32 frag_size)
169 {
170 	if (!dev) {
171 		WARN(1, "Missing net_device from driver");
172 		return -ENODEV;
173 	}
174 
175 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
176 		WARN(1, "Driver promised not to register this");
177 		return -EINVAL;
178 	}
179 
180 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
181 		WARN(1, "Missing unregister, handled but fix driver");
182 		xdp_rxq_info_unreg(xdp_rxq);
183 	}
184 
185 	/* State either UNREGISTERED or NEW */
186 	xdp_rxq_info_init(xdp_rxq);
187 	xdp_rxq->dev = dev;
188 	xdp_rxq->queue_index = queue_index;
189 	xdp_rxq->napi_id = napi_id;
190 	xdp_rxq->frag_size = frag_size;
191 
192 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
193 	return 0;
194 }
195 EXPORT_SYMBOL_GPL(__xdp_rxq_info_reg);
196 
197 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
198 {
199 	xdp_rxq->reg_state = REG_STATE_UNUSED;
200 }
201 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
202 
203 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
204 {
205 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
206 }
207 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
208 
209 static int __mem_id_init_hash_table(void)
210 {
211 	struct rhashtable *rht;
212 	int ret;
213 
214 	if (unlikely(mem_id_init))
215 		return 0;
216 
217 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
218 	if (!rht)
219 		return -ENOMEM;
220 
221 	ret = rhashtable_init(rht, &mem_id_rht_params);
222 	if (ret < 0) {
223 		kfree(rht);
224 		return ret;
225 	}
226 	mem_id_ht = rht;
227 	smp_mb(); /* mutex lock should provide enough pairing */
228 	mem_id_init = true;
229 
230 	return 0;
231 }
232 
233 /* Allocate a cyclic ID that maps to allocator pointer.
234  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
235  *
236  * Caller must lock mem_id_lock.
237  */
238 static int __mem_id_cyclic_get(gfp_t gfp)
239 {
240 	int retries = 1;
241 	int id;
242 
243 again:
244 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
245 	if (id < 0) {
246 		if (id == -ENOSPC) {
247 			/* Cyclic allocator, reset next id */
248 			if (retries--) {
249 				mem_id_next = MEM_ID_MIN;
250 				goto again;
251 			}
252 		}
253 		return id; /* errno */
254 	}
255 	mem_id_next = id + 1;
256 
257 	return id;
258 }
259 
260 static bool __is_supported_mem_type(enum xdp_mem_type type)
261 {
262 	if (type == MEM_TYPE_PAGE_POOL)
263 		return is_page_pool_compiled_in();
264 
265 	if (type >= MEM_TYPE_MAX)
266 		return false;
267 
268 	return true;
269 }
270 
271 static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem,
272 						     enum xdp_mem_type type,
273 						     void *allocator)
274 {
275 	struct xdp_mem_allocator *xdp_alloc;
276 	gfp_t gfp = GFP_KERNEL;
277 	int id, errno, ret;
278 	void *ptr;
279 
280 	if (!__is_supported_mem_type(type))
281 		return ERR_PTR(-EOPNOTSUPP);
282 
283 	mem->type = type;
284 
285 	if (!allocator) {
286 		if (type == MEM_TYPE_PAGE_POOL)
287 			return ERR_PTR(-EINVAL); /* Setup time check page_pool req */
288 		return NULL;
289 	}
290 
291 	/* Delay init of rhashtable to save memory if feature isn't used */
292 	if (!mem_id_init) {
293 		mutex_lock(&mem_id_lock);
294 		ret = __mem_id_init_hash_table();
295 		mutex_unlock(&mem_id_lock);
296 		if (ret < 0) {
297 			WARN_ON(1);
298 			return ERR_PTR(ret);
299 		}
300 	}
301 
302 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
303 	if (!xdp_alloc)
304 		return ERR_PTR(-ENOMEM);
305 
306 	mutex_lock(&mem_id_lock);
307 	id = __mem_id_cyclic_get(gfp);
308 	if (id < 0) {
309 		errno = id;
310 		goto err;
311 	}
312 	mem->id = id;
313 	xdp_alloc->mem = *mem;
314 	xdp_alloc->allocator = allocator;
315 
316 	/* Insert allocator into ID lookup table */
317 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
318 	if (IS_ERR(ptr)) {
319 		ida_simple_remove(&mem_id_pool, mem->id);
320 		mem->id = 0;
321 		errno = PTR_ERR(ptr);
322 		goto err;
323 	}
324 
325 	if (type == MEM_TYPE_PAGE_POOL)
326 		page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem);
327 
328 	mutex_unlock(&mem_id_lock);
329 
330 	return xdp_alloc;
331 err:
332 	mutex_unlock(&mem_id_lock);
333 	kfree(xdp_alloc);
334 	return ERR_PTR(errno);
335 }
336 
337 int xdp_reg_mem_model(struct xdp_mem_info *mem,
338 		      enum xdp_mem_type type, void *allocator)
339 {
340 	struct xdp_mem_allocator *xdp_alloc;
341 
342 	xdp_alloc = __xdp_reg_mem_model(mem, type, allocator);
343 	if (IS_ERR(xdp_alloc))
344 		return PTR_ERR(xdp_alloc);
345 	return 0;
346 }
347 EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
348 
349 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
350 			       enum xdp_mem_type type, void *allocator)
351 {
352 	struct xdp_mem_allocator *xdp_alloc;
353 
354 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
355 		WARN(1, "Missing register, driver bug");
356 		return -EFAULT;
357 	}
358 
359 	xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator);
360 	if (IS_ERR(xdp_alloc))
361 		return PTR_ERR(xdp_alloc);
362 
363 	if (trace_mem_connect_enabled() && xdp_alloc)
364 		trace_mem_connect(xdp_alloc, xdp_rxq);
365 	return 0;
366 }
367 
368 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
369 
370 /* XDP RX runs under NAPI protection, and in different delivery error
371  * scenarios (e.g. queue full), it is possible to return the xdp_frame
372  * while still leveraging this protection.  The @napi_direct boolean
373  * is used for those calls sites.  Thus, allowing for faster recycling
374  * of xdp_frames/pages in those cases.
375  */
376 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
377 		  struct xdp_buff *xdp)
378 {
379 	struct page *page;
380 
381 	switch (mem->type) {
382 	case MEM_TYPE_PAGE_POOL:
383 		page = virt_to_head_page(data);
384 		if (napi_direct && xdp_return_frame_no_direct())
385 			napi_direct = false;
386 		/* No need to check ((page->pp_magic & ~0x3UL) == PP_SIGNATURE)
387 		 * as mem->type knows this a page_pool page
388 		 */
389 		page_pool_put_full_page(page->pp, page, napi_direct);
390 		break;
391 	case MEM_TYPE_PAGE_SHARED:
392 		page_frag_free(data);
393 		break;
394 	case MEM_TYPE_PAGE_ORDER0:
395 		page = virt_to_page(data); /* Assumes order0 page*/
396 		put_page(page);
397 		break;
398 	case MEM_TYPE_XSK_BUFF_POOL:
399 		/* NB! Only valid from an xdp_buff! */
400 		xsk_buff_free(xdp);
401 		break;
402 	default:
403 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
404 		WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
405 		break;
406 	}
407 }
408 
409 void xdp_return_frame(struct xdp_frame *xdpf)
410 {
411 	struct skb_shared_info *sinfo;
412 	int i;
413 
414 	if (likely(!xdp_frame_has_frags(xdpf)))
415 		goto out;
416 
417 	sinfo = xdp_get_shared_info_from_frame(xdpf);
418 	for (i = 0; i < sinfo->nr_frags; i++) {
419 		struct page *page = skb_frag_page(&sinfo->frags[i]);
420 
421 		__xdp_return(page_address(page), &xdpf->mem, false, NULL);
422 	}
423 out:
424 	__xdp_return(xdpf->data, &xdpf->mem, false, NULL);
425 }
426 EXPORT_SYMBOL_GPL(xdp_return_frame);
427 
428 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
429 {
430 	struct skb_shared_info *sinfo;
431 	int i;
432 
433 	if (likely(!xdp_frame_has_frags(xdpf)))
434 		goto out;
435 
436 	sinfo = xdp_get_shared_info_from_frame(xdpf);
437 	for (i = 0; i < sinfo->nr_frags; i++) {
438 		struct page *page = skb_frag_page(&sinfo->frags[i]);
439 
440 		__xdp_return(page_address(page), &xdpf->mem, true, NULL);
441 	}
442 out:
443 	__xdp_return(xdpf->data, &xdpf->mem, true, NULL);
444 }
445 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
446 
447 /* XDP bulk APIs introduce a defer/flush mechanism to return
448  * pages belonging to the same xdp_mem_allocator object
449  * (identified via the mem.id field) in bulk to optimize
450  * I-cache and D-cache.
451  * The bulk queue size is set to 16 to be aligned to how
452  * XDP_REDIRECT bulking works. The bulk is flushed when
453  * it is full or when mem.id changes.
454  * xdp_frame_bulk is usually stored/allocated on the function
455  * call-stack to avoid locking penalties.
456  */
457 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
458 {
459 	struct xdp_mem_allocator *xa = bq->xa;
460 
461 	if (unlikely(!xa || !bq->count))
462 		return;
463 
464 	page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
465 	/* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
466 	bq->count = 0;
467 }
468 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
469 
470 /* Must be called with rcu_read_lock held */
471 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
472 			   struct xdp_frame_bulk *bq)
473 {
474 	struct xdp_mem_info *mem = &xdpf->mem;
475 	struct xdp_mem_allocator *xa;
476 
477 	if (mem->type != MEM_TYPE_PAGE_POOL) {
478 		xdp_return_frame(xdpf);
479 		return;
480 	}
481 
482 	xa = bq->xa;
483 	if (unlikely(!xa)) {
484 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
485 		bq->count = 0;
486 		bq->xa = xa;
487 	}
488 
489 	if (bq->count == XDP_BULK_QUEUE_SIZE)
490 		xdp_flush_frame_bulk(bq);
491 
492 	if (unlikely(mem->id != xa->mem.id)) {
493 		xdp_flush_frame_bulk(bq);
494 		bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
495 	}
496 
497 	if (unlikely(xdp_frame_has_frags(xdpf))) {
498 		struct skb_shared_info *sinfo;
499 		int i;
500 
501 		sinfo = xdp_get_shared_info_from_frame(xdpf);
502 		for (i = 0; i < sinfo->nr_frags; i++) {
503 			skb_frag_t *frag = &sinfo->frags[i];
504 
505 			bq->q[bq->count++] = skb_frag_address(frag);
506 			if (bq->count == XDP_BULK_QUEUE_SIZE)
507 				xdp_flush_frame_bulk(bq);
508 		}
509 	}
510 	bq->q[bq->count++] = xdpf->data;
511 }
512 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
513 
514 void xdp_return_buff(struct xdp_buff *xdp)
515 {
516 	struct skb_shared_info *sinfo;
517 	int i;
518 
519 	if (likely(!xdp_buff_has_frags(xdp)))
520 		goto out;
521 
522 	sinfo = xdp_get_shared_info_from_buff(xdp);
523 	for (i = 0; i < sinfo->nr_frags; i++) {
524 		struct page *page = skb_frag_page(&sinfo->frags[i]);
525 
526 		__xdp_return(page_address(page), &xdp->rxq->mem, true, xdp);
527 	}
528 out:
529 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
530 }
531 EXPORT_SYMBOL_GPL(xdp_return_buff);
532 
533 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
534 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
535 {
536 	struct xdp_mem_allocator *xa;
537 	struct page *page;
538 
539 	rcu_read_lock();
540 	xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
541 	page = virt_to_head_page(data);
542 	if (xa)
543 		page_pool_release_page(xa->page_pool, page);
544 	rcu_read_unlock();
545 }
546 EXPORT_SYMBOL_GPL(__xdp_release_frame);
547 
548 void xdp_attachment_setup(struct xdp_attachment_info *info,
549 			  struct netdev_bpf *bpf)
550 {
551 	if (info->prog)
552 		bpf_prog_put(info->prog);
553 	info->prog = bpf->prog;
554 	info->flags = bpf->flags;
555 }
556 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
557 
558 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
559 {
560 	unsigned int metasize, totsize;
561 	void *addr, *data_to_copy;
562 	struct xdp_frame *xdpf;
563 	struct page *page;
564 
565 	/* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
566 	metasize = xdp_data_meta_unsupported(xdp) ? 0 :
567 		   xdp->data - xdp->data_meta;
568 	totsize = xdp->data_end - xdp->data + metasize;
569 
570 	if (sizeof(*xdpf) + totsize > PAGE_SIZE)
571 		return NULL;
572 
573 	page = dev_alloc_page();
574 	if (!page)
575 		return NULL;
576 
577 	addr = page_to_virt(page);
578 	xdpf = addr;
579 	memset(xdpf, 0, sizeof(*xdpf));
580 
581 	addr += sizeof(*xdpf);
582 	data_to_copy = metasize ? xdp->data_meta : xdp->data;
583 	memcpy(addr, data_to_copy, totsize);
584 
585 	xdpf->data = addr + metasize;
586 	xdpf->len = totsize - metasize;
587 	xdpf->headroom = 0;
588 	xdpf->metasize = metasize;
589 	xdpf->frame_sz = PAGE_SIZE;
590 	xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
591 
592 	xsk_buff_free(xdp);
593 	return xdpf;
594 }
595 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
596 
597 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
598 void xdp_warn(const char *msg, const char *func, const int line)
599 {
600 	WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
601 };
602 EXPORT_SYMBOL_GPL(xdp_warn);
603 
604 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
605 {
606 	n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
607 				      n_skb, skbs);
608 	if (unlikely(!n_skb))
609 		return -ENOMEM;
610 
611 	return 0;
612 }
613 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
614 
615 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
616 					   struct sk_buff *skb,
617 					   struct net_device *dev)
618 {
619 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
620 	unsigned int headroom, frame_size;
621 	void *hard_start;
622 	u8 nr_frags;
623 
624 	/* xdp frags frame */
625 	if (unlikely(xdp_frame_has_frags(xdpf)))
626 		nr_frags = sinfo->nr_frags;
627 
628 	/* Part of headroom was reserved to xdpf */
629 	headroom = sizeof(*xdpf) + xdpf->headroom;
630 
631 	/* Memory size backing xdp_frame data already have reserved
632 	 * room for build_skb to place skb_shared_info in tailroom.
633 	 */
634 	frame_size = xdpf->frame_sz;
635 
636 	hard_start = xdpf->data - headroom;
637 	skb = build_skb_around(skb, hard_start, frame_size);
638 	if (unlikely(!skb))
639 		return NULL;
640 
641 	skb_reserve(skb, headroom);
642 	__skb_put(skb, xdpf->len);
643 	if (xdpf->metasize)
644 		skb_metadata_set(skb, xdpf->metasize);
645 
646 	if (unlikely(xdp_frame_has_frags(xdpf)))
647 		xdp_update_skb_shared_info(skb, nr_frags,
648 					   sinfo->xdp_frags_size,
649 					   nr_frags * xdpf->frame_sz,
650 					   xdp_frame_is_frag_pfmemalloc(xdpf));
651 
652 	/* Essential SKB info: protocol and skb->dev */
653 	skb->protocol = eth_type_trans(skb, dev);
654 
655 	/* Optional SKB info, currently missing:
656 	 * - HW checksum info		(skb->ip_summed)
657 	 * - HW RX hash			(skb_set_hash)
658 	 * - RX ring dev queue index	(skb_record_rx_queue)
659 	 */
660 
661 	/* Until page_pool get SKB return path, release DMA here */
662 	xdp_release_frame(xdpf);
663 
664 	/* Allow SKB to reuse area used by xdp_frame */
665 	xdp_scrub_frame(xdpf);
666 
667 	return skb;
668 }
669 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
670 
671 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
672 					 struct net_device *dev)
673 {
674 	struct sk_buff *skb;
675 
676 	skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
677 	if (unlikely(!skb))
678 		return NULL;
679 
680 	memset(skb, 0, offsetof(struct sk_buff, tail));
681 
682 	return __xdp_build_skb_from_frame(xdpf, skb, dev);
683 }
684 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
685 
686 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
687 {
688 	unsigned int headroom, totalsize;
689 	struct xdp_frame *nxdpf;
690 	struct page *page;
691 	void *addr;
692 
693 	headroom = xdpf->headroom + sizeof(*xdpf);
694 	totalsize = headroom + xdpf->len;
695 
696 	if (unlikely(totalsize > PAGE_SIZE))
697 		return NULL;
698 	page = dev_alloc_page();
699 	if (!page)
700 		return NULL;
701 	addr = page_to_virt(page);
702 
703 	memcpy(addr, xdpf, totalsize);
704 
705 	nxdpf = addr;
706 	nxdpf->data = addr + headroom;
707 	nxdpf->frame_sz = PAGE_SIZE;
708 	nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
709 	nxdpf->mem.id = 0;
710 
711 	return nxdpf;
712 }
713 
714 __diag_push();
715 __diag_ignore_all("-Wmissing-prototypes",
716 		  "Global functions as their definitions will be in vmlinux BTF");
717 
718 /**
719  * bpf_xdp_metadata_rx_timestamp - Read XDP frame RX timestamp.
720  * @ctx: XDP context pointer.
721  * @timestamp: Return value pointer.
722  *
723  * Returns 0 on success or ``-errno`` on error.
724  */
725 int bpf_xdp_metadata_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
726 {
727 	return -EOPNOTSUPP;
728 }
729 
730 /**
731  * bpf_xdp_metadata_rx_hash - Read XDP frame RX hash.
732  * @ctx: XDP context pointer.
733  * @hash: Return value pointer.
734  *
735  * Returns 0 on success or ``-errno`` on error.
736  */
737 int bpf_xdp_metadata_rx_hash(const struct xdp_md *ctx, u32 *hash)
738 {
739 	return -EOPNOTSUPP;
740 }
741 
742 __diag_pop();
743 
744 BTF_SET8_START(xdp_metadata_kfunc_ids)
745 #define XDP_METADATA_KFUNC(_, name) BTF_ID_FLAGS(func, name, 0)
746 XDP_METADATA_KFUNC_xxx
747 #undef XDP_METADATA_KFUNC
748 BTF_SET8_END(xdp_metadata_kfunc_ids)
749 
750 static const struct btf_kfunc_id_set xdp_metadata_kfunc_set = {
751 	.owner = THIS_MODULE,
752 	.set   = &xdp_metadata_kfunc_ids,
753 };
754 
755 BTF_ID_LIST(xdp_metadata_kfunc_ids_unsorted)
756 #define XDP_METADATA_KFUNC(name, str) BTF_ID(func, str)
757 XDP_METADATA_KFUNC_xxx
758 #undef XDP_METADATA_KFUNC
759 
760 u32 bpf_xdp_metadata_kfunc_id(int id)
761 {
762 	/* xdp_metadata_kfunc_ids is sorted and can't be used */
763 	return xdp_metadata_kfunc_ids_unsorted[id];
764 }
765 
766 bool bpf_dev_bound_kfunc_id(u32 btf_id)
767 {
768 	return btf_id_set8_contains(&xdp_metadata_kfunc_ids, btf_id);
769 }
770 
771 static int __init xdp_metadata_init(void)
772 {
773 	return register_btf_kfunc_id_set(BPF_PROG_TYPE_XDP, &xdp_metadata_kfunc_set);
774 }
775 late_initcall(xdp_metadata_init);
776