Loading...
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
29static DEFINE_IDA(mem_id_pool);
30static DEFINE_MUTEX(mem_id_lock);
31#define MEM_ID_MAX 0xFFFE
32#define MEM_ID_MIN 1
33static int mem_id_next = MEM_ID_MIN;
34
35static bool mem_id_init; /* false */
36static struct rhashtable *mem_id_ht;
37
38static 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
50static 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
59static 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
71static 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
83static 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
91static 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
115void 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}
135EXPORT_SYMBOL_GPL(xdp_unreg_mem_model);
136
137void 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}
146EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
147
148void 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}
159EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
160
161static 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 */
167int __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}
196EXPORT_SYMBOL_GPL(__xdp_rxq_info_reg);
197
198void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
199{
200 xdp_rxq->reg_state = REG_STATE_UNUSED;
201}
202EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
203
204bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
205{
206 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
207}
208EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
209
210static 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 */
239static int __mem_id_cyclic_get(gfp_t gfp)
240{
241 int retries = 1;
242 int id;
243
244again:
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
261static 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
272static 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 WARN_ON(1);
299 return ERR_PTR(ret);
300 }
301 }
302
303 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
304 if (!xdp_alloc)
305 return ERR_PTR(-ENOMEM);
306
307 mutex_lock(&mem_id_lock);
308 id = __mem_id_cyclic_get(gfp);
309 if (id < 0) {
310 errno = id;
311 goto err;
312 }
313 mem->id = id;
314 xdp_alloc->mem = *mem;
315 xdp_alloc->allocator = allocator;
316
317 /* Insert allocator into ID lookup table */
318 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
319 if (IS_ERR(ptr)) {
320 ida_simple_remove(&mem_id_pool, mem->id);
321 mem->id = 0;
322 errno = PTR_ERR(ptr);
323 goto err;
324 }
325
326 if (type == MEM_TYPE_PAGE_POOL)
327 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem);
328
329 mutex_unlock(&mem_id_lock);
330
331 return xdp_alloc;
332err:
333 mutex_unlock(&mem_id_lock);
334 kfree(xdp_alloc);
335 return ERR_PTR(errno);
336}
337
338int xdp_reg_mem_model(struct xdp_mem_info *mem,
339 enum xdp_mem_type type, void *allocator)
340{
341 struct xdp_mem_allocator *xdp_alloc;
342
343 xdp_alloc = __xdp_reg_mem_model(mem, type, allocator);
344 if (IS_ERR(xdp_alloc))
345 return PTR_ERR(xdp_alloc);
346 return 0;
347}
348EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
349
350int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
351 enum xdp_mem_type type, void *allocator)
352{
353 struct xdp_mem_allocator *xdp_alloc;
354
355 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
356 WARN(1, "Missing register, driver bug");
357 return -EFAULT;
358 }
359
360 xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator);
361 if (IS_ERR(xdp_alloc))
362 return PTR_ERR(xdp_alloc);
363
364 if (trace_mem_connect_enabled() && xdp_alloc)
365 trace_mem_connect(xdp_alloc, xdp_rxq);
366 return 0;
367}
368
369EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
370
371/* XDP RX runs under NAPI protection, and in different delivery error
372 * scenarios (e.g. queue full), it is possible to return the xdp_frame
373 * while still leveraging this protection. The @napi_direct boolean
374 * is used for those calls sites. Thus, allowing for faster recycling
375 * of xdp_frames/pages in those cases.
376 */
377void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
378 struct xdp_buff *xdp)
379{
380 struct page *page;
381
382 switch (mem->type) {
383 case MEM_TYPE_PAGE_POOL:
384 page = virt_to_head_page(data);
385 if (napi_direct && xdp_return_frame_no_direct())
386 napi_direct = false;
387 /* No need to check ((page->pp_magic & ~0x3UL) == PP_SIGNATURE)
388 * as mem->type knows this a page_pool page
389 */
390 page_pool_put_full_page(page->pp, page, napi_direct);
391 break;
392 case MEM_TYPE_PAGE_SHARED:
393 page_frag_free(data);
394 break;
395 case MEM_TYPE_PAGE_ORDER0:
396 page = virt_to_page(data); /* Assumes order0 page*/
397 put_page(page);
398 break;
399 case MEM_TYPE_XSK_BUFF_POOL:
400 /* NB! Only valid from an xdp_buff! */
401 xsk_buff_free(xdp);
402 break;
403 default:
404 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
405 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
406 break;
407 }
408}
409
410void xdp_return_frame(struct xdp_frame *xdpf)
411{
412 struct skb_shared_info *sinfo;
413 int i;
414
415 if (likely(!xdp_frame_has_frags(xdpf)))
416 goto out;
417
418 sinfo = xdp_get_shared_info_from_frame(xdpf);
419 for (i = 0; i < sinfo->nr_frags; i++) {
420 struct page *page = skb_frag_page(&sinfo->frags[i]);
421
422 __xdp_return(page_address(page), &xdpf->mem, false, NULL);
423 }
424out:
425 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
426}
427EXPORT_SYMBOL_GPL(xdp_return_frame);
428
429void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
430{
431 struct skb_shared_info *sinfo;
432 int i;
433
434 if (likely(!xdp_frame_has_frags(xdpf)))
435 goto out;
436
437 sinfo = xdp_get_shared_info_from_frame(xdpf);
438 for (i = 0; i < sinfo->nr_frags; i++) {
439 struct page *page = skb_frag_page(&sinfo->frags[i]);
440
441 __xdp_return(page_address(page), &xdpf->mem, true, NULL);
442 }
443out:
444 __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
445}
446EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
447
448/* XDP bulk APIs introduce a defer/flush mechanism to return
449 * pages belonging to the same xdp_mem_allocator object
450 * (identified via the mem.id field) in bulk to optimize
451 * I-cache and D-cache.
452 * The bulk queue size is set to 16 to be aligned to how
453 * XDP_REDIRECT bulking works. The bulk is flushed when
454 * it is full or when mem.id changes.
455 * xdp_frame_bulk is usually stored/allocated on the function
456 * call-stack to avoid locking penalties.
457 */
458void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
459{
460 struct xdp_mem_allocator *xa = bq->xa;
461
462 if (unlikely(!xa || !bq->count))
463 return;
464
465 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
466 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
467 bq->count = 0;
468}
469EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
470
471/* Must be called with rcu_read_lock held */
472void xdp_return_frame_bulk(struct xdp_frame *xdpf,
473 struct xdp_frame_bulk *bq)
474{
475 struct xdp_mem_info *mem = &xdpf->mem;
476 struct xdp_mem_allocator *xa;
477
478 if (mem->type != MEM_TYPE_PAGE_POOL) {
479 xdp_return_frame(xdpf);
480 return;
481 }
482
483 xa = bq->xa;
484 if (unlikely(!xa)) {
485 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
486 bq->count = 0;
487 bq->xa = xa;
488 }
489
490 if (bq->count == XDP_BULK_QUEUE_SIZE)
491 xdp_flush_frame_bulk(bq);
492
493 if (unlikely(mem->id != xa->mem.id)) {
494 xdp_flush_frame_bulk(bq);
495 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
496 }
497
498 if (unlikely(xdp_frame_has_frags(xdpf))) {
499 struct skb_shared_info *sinfo;
500 int i;
501
502 sinfo = xdp_get_shared_info_from_frame(xdpf);
503 for (i = 0; i < sinfo->nr_frags; i++) {
504 skb_frag_t *frag = &sinfo->frags[i];
505
506 bq->q[bq->count++] = skb_frag_address(frag);
507 if (bq->count == XDP_BULK_QUEUE_SIZE)
508 xdp_flush_frame_bulk(bq);
509 }
510 }
511 bq->q[bq->count++] = xdpf->data;
512}
513EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
514
515void xdp_return_buff(struct xdp_buff *xdp)
516{
517 struct skb_shared_info *sinfo;
518 int i;
519
520 if (likely(!xdp_buff_has_frags(xdp)))
521 goto out;
522
523 sinfo = xdp_get_shared_info_from_buff(xdp);
524 for (i = 0; i < sinfo->nr_frags; i++) {
525 struct page *page = skb_frag_page(&sinfo->frags[i]);
526
527 __xdp_return(page_address(page), &xdp->rxq->mem, true, xdp);
528 }
529out:
530 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
531}
532EXPORT_SYMBOL_GPL(xdp_return_buff);
533
534void xdp_attachment_setup(struct xdp_attachment_info *info,
535 struct netdev_bpf *bpf)
536{
537 if (info->prog)
538 bpf_prog_put(info->prog);
539 info->prog = bpf->prog;
540 info->flags = bpf->flags;
541}
542EXPORT_SYMBOL_GPL(xdp_attachment_setup);
543
544struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
545{
546 unsigned int metasize, totsize;
547 void *addr, *data_to_copy;
548 struct xdp_frame *xdpf;
549 struct page *page;
550
551 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
552 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
553 xdp->data - xdp->data_meta;
554 totsize = xdp->data_end - xdp->data + metasize;
555
556 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
557 return NULL;
558
559 page = dev_alloc_page();
560 if (!page)
561 return NULL;
562
563 addr = page_to_virt(page);
564 xdpf = addr;
565 memset(xdpf, 0, sizeof(*xdpf));
566
567 addr += sizeof(*xdpf);
568 data_to_copy = metasize ? xdp->data_meta : xdp->data;
569 memcpy(addr, data_to_copy, totsize);
570
571 xdpf->data = addr + metasize;
572 xdpf->len = totsize - metasize;
573 xdpf->headroom = 0;
574 xdpf->metasize = metasize;
575 xdpf->frame_sz = PAGE_SIZE;
576 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
577
578 xsk_buff_free(xdp);
579 return xdpf;
580}
581EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
582
583/* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
584void xdp_warn(const char *msg, const char *func, const int line)
585{
586 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
587};
588EXPORT_SYMBOL_GPL(xdp_warn);
589
590int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
591{
592 n_skb = kmem_cache_alloc_bulk(skbuff_cache, gfp, n_skb, skbs);
593 if (unlikely(!n_skb))
594 return -ENOMEM;
595
596 return 0;
597}
598EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
599
600struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
601 struct sk_buff *skb,
602 struct net_device *dev)
603{
604 struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
605 unsigned int headroom, frame_size;
606 void *hard_start;
607 u8 nr_frags;
608
609 /* xdp frags frame */
610 if (unlikely(xdp_frame_has_frags(xdpf)))
611 nr_frags = sinfo->nr_frags;
612
613 /* Part of headroom was reserved to xdpf */
614 headroom = sizeof(*xdpf) + xdpf->headroom;
615
616 /* Memory size backing xdp_frame data already have reserved
617 * room for build_skb to place skb_shared_info in tailroom.
618 */
619 frame_size = xdpf->frame_sz;
620
621 hard_start = xdpf->data - headroom;
622 skb = build_skb_around(skb, hard_start, frame_size);
623 if (unlikely(!skb))
624 return NULL;
625
626 skb_reserve(skb, headroom);
627 __skb_put(skb, xdpf->len);
628 if (xdpf->metasize)
629 skb_metadata_set(skb, xdpf->metasize);
630
631 if (unlikely(xdp_frame_has_frags(xdpf)))
632 xdp_update_skb_shared_info(skb, nr_frags,
633 sinfo->xdp_frags_size,
634 nr_frags * xdpf->frame_sz,
635 xdp_frame_is_frag_pfmemalloc(xdpf));
636
637 /* Essential SKB info: protocol and skb->dev */
638 skb->protocol = eth_type_trans(skb, dev);
639
640 /* Optional SKB info, currently missing:
641 * - HW checksum info (skb->ip_summed)
642 * - HW RX hash (skb_set_hash)
643 * - RX ring dev queue index (skb_record_rx_queue)
644 */
645
646 if (xdpf->mem.type == MEM_TYPE_PAGE_POOL)
647 skb_mark_for_recycle(skb);
648
649 /* Allow SKB to reuse area used by xdp_frame */
650 xdp_scrub_frame(xdpf);
651
652 return skb;
653}
654EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
655
656struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
657 struct net_device *dev)
658{
659 struct sk_buff *skb;
660
661 skb = kmem_cache_alloc(skbuff_cache, GFP_ATOMIC);
662 if (unlikely(!skb))
663 return NULL;
664
665 memset(skb, 0, offsetof(struct sk_buff, tail));
666
667 return __xdp_build_skb_from_frame(xdpf, skb, dev);
668}
669EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
670
671struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
672{
673 unsigned int headroom, totalsize;
674 struct xdp_frame *nxdpf;
675 struct page *page;
676 void *addr;
677
678 headroom = xdpf->headroom + sizeof(*xdpf);
679 totalsize = headroom + xdpf->len;
680
681 if (unlikely(totalsize > PAGE_SIZE))
682 return NULL;
683 page = dev_alloc_page();
684 if (!page)
685 return NULL;
686 addr = page_to_virt(page);
687
688 memcpy(addr, xdpf, totalsize);
689
690 nxdpf = addr;
691 nxdpf->data = addr + headroom;
692 nxdpf->frame_sz = PAGE_SIZE;
693 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
694 nxdpf->mem.id = 0;
695
696 return nxdpf;
697}
698
699__bpf_kfunc_start_defs();
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/**
740 * bpf_xdp_metadata_rx_vlan_tag - Get XDP packet outermost VLAN tag
741 * @ctx: XDP context pointer.
742 * @vlan_proto: Destination pointer for VLAN Tag protocol identifier (TPID).
743 * @vlan_tci: Destination pointer for VLAN TCI (VID + DEI + PCP)
744 *
745 * In case of success, ``vlan_proto`` contains *Tag protocol identifier (TPID)*,
746 * usually ``ETH_P_8021Q`` or ``ETH_P_8021AD``, but some networks can use
747 * custom TPIDs. ``vlan_proto`` is stored in **network byte order (BE)**
748 * and should be used as follows:
749 * ``if (vlan_proto == bpf_htons(ETH_P_8021Q)) do_something();``
750 *
751 * ``vlan_tci`` contains the remaining 16 bits of a VLAN tag.
752 * Driver is expected to provide those in **host byte order (usually LE)**,
753 * so the bpf program should not perform byte conversion.
754 * According to 802.1Q standard, *VLAN TCI (Tag control information)*
755 * is a bit field that contains:
756 * *VLAN identifier (VID)* that can be read with ``vlan_tci & 0xfff``,
757 * *Drop eligible indicator (DEI)* - 1 bit,
758 * *Priority code point (PCP)* - 3 bits.
759 * For detailed meaning of DEI and PCP, please refer to other sources.
760 *
761 * Return:
762 * * Returns 0 on success or ``-errno`` on error.
763 * * ``-EOPNOTSUPP`` : device driver doesn't implement kfunc
764 * * ``-ENODATA`` : VLAN tag was not stripped or is not available
765 */
766__bpf_kfunc int bpf_xdp_metadata_rx_vlan_tag(const struct xdp_md *ctx,
767 __be16 *vlan_proto, u16 *vlan_tci)
768{
769 return -EOPNOTSUPP;
770}
771
772__bpf_kfunc_end_defs();
773
774BTF_SET8_START(xdp_metadata_kfunc_ids)
775#define XDP_METADATA_KFUNC(_, __, name, ___) BTF_ID_FLAGS(func, name, KF_TRUSTED_ARGS)
776XDP_METADATA_KFUNC_xxx
777#undef XDP_METADATA_KFUNC
778BTF_SET8_END(xdp_metadata_kfunc_ids)
779
780static const struct btf_kfunc_id_set xdp_metadata_kfunc_set = {
781 .owner = THIS_MODULE,
782 .set = &xdp_metadata_kfunc_ids,
783};
784
785BTF_ID_LIST(xdp_metadata_kfunc_ids_unsorted)
786#define XDP_METADATA_KFUNC(name, _, str, __) BTF_ID(func, str)
787XDP_METADATA_KFUNC_xxx
788#undef XDP_METADATA_KFUNC
789
790u32 bpf_xdp_metadata_kfunc_id(int id)
791{
792 /* xdp_metadata_kfunc_ids is sorted and can't be used */
793 return xdp_metadata_kfunc_ids_unsorted[id];
794}
795
796bool bpf_dev_bound_kfunc_id(u32 btf_id)
797{
798 return btf_id_set8_contains(&xdp_metadata_kfunc_ids, btf_id);
799}
800
801static int __init xdp_metadata_init(void)
802{
803 return register_btf_kfunc_id_set(BPF_PROG_TYPE_XDP, &xdp_metadata_kfunc_set);
804}
805late_initcall(xdp_metadata_init);
806
807void xdp_set_features_flag(struct net_device *dev, xdp_features_t val)
808{
809 val &= NETDEV_XDP_ACT_MASK;
810 if (dev->xdp_features == val)
811 return;
812
813 dev->xdp_features = val;
814
815 if (dev->reg_state == NETREG_REGISTERED)
816 call_netdevice_notifiers(NETDEV_XDP_FEAT_CHANGE, dev);
817}
818EXPORT_SYMBOL_GPL(xdp_set_features_flag);
819
820void xdp_features_set_redirect_target(struct net_device *dev, bool support_sg)
821{
822 xdp_features_t val = (dev->xdp_features | NETDEV_XDP_ACT_NDO_XMIT);
823
824 if (support_sg)
825 val |= NETDEV_XDP_ACT_NDO_XMIT_SG;
826 xdp_set_features_flag(dev, val);
827}
828EXPORT_SYMBOL_GPL(xdp_features_set_redirect_target);
829
830void xdp_features_clear_redirect_target(struct net_device *dev)
831{
832 xdp_features_t val = dev->xdp_features;
833
834 val &= ~(NETDEV_XDP_ACT_NDO_XMIT | NETDEV_XDP_ACT_NDO_XMIT_SG);
835 xdp_set_features_flag(dev, val);
836}
837EXPORT_SYMBOL_GPL(xdp_features_clear_redirect_target);
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
27static DEFINE_IDA(mem_id_pool);
28static DEFINE_MUTEX(mem_id_lock);
29#define MEM_ID_MAX 0xFFFE
30#define MEM_ID_MIN 1
31static int mem_id_next = MEM_ID_MIN;
32
33static bool mem_id_init; /* false */
34static struct rhashtable *mem_id_ht;
35
36static 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
48static 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
57static 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
69static 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
81static 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
89static 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
113void xdp_unreg_mem_model(struct xdp_mem_info *mem)
114{
115 struct xdp_mem_allocator *xa;
116 int type = mem->type;
117 int id = mem->id;
118
119 /* Reset mem info to defaults */
120 mem->id = 0;
121 mem->type = 0;
122
123 if (id == 0)
124 return;
125
126 if (type == MEM_TYPE_PAGE_POOL) {
127 rcu_read_lock();
128 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
129 page_pool_destroy(xa->page_pool);
130 rcu_read_unlock();
131 }
132}
133EXPORT_SYMBOL_GPL(xdp_unreg_mem_model);
134
135void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
136{
137 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
138 WARN(1, "Missing register, driver bug");
139 return;
140 }
141
142 xdp_unreg_mem_model(&xdp_rxq->mem);
143}
144EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
145
146void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
147{
148 /* Simplify driver cleanup code paths, allow unreg "unused" */
149 if (xdp_rxq->reg_state == REG_STATE_UNUSED)
150 return;
151
152 xdp_rxq_info_unreg_mem_model(xdp_rxq);
153
154 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
155 xdp_rxq->dev = NULL;
156}
157EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
158
159static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
160{
161 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
162}
163
164/* Returns 0 on success, negative on failure */
165int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
166 struct net_device *dev, u32 queue_index,
167 unsigned int napi_id, u32 frag_size)
168{
169 if (!dev) {
170 WARN(1, "Missing net_device from driver");
171 return -ENODEV;
172 }
173
174 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
175 WARN(1, "Driver promised not to register this");
176 return -EINVAL;
177 }
178
179 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
180 WARN(1, "Missing unregister, handled but fix driver");
181 xdp_rxq_info_unreg(xdp_rxq);
182 }
183
184 /* State either UNREGISTERED or NEW */
185 xdp_rxq_info_init(xdp_rxq);
186 xdp_rxq->dev = dev;
187 xdp_rxq->queue_index = queue_index;
188 xdp_rxq->napi_id = napi_id;
189 xdp_rxq->frag_size = frag_size;
190
191 xdp_rxq->reg_state = REG_STATE_REGISTERED;
192 return 0;
193}
194EXPORT_SYMBOL_GPL(__xdp_rxq_info_reg);
195
196void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
197{
198 xdp_rxq->reg_state = REG_STATE_UNUSED;
199}
200EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
201
202bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
203{
204 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
205}
206EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
207
208static int __mem_id_init_hash_table(void)
209{
210 struct rhashtable *rht;
211 int ret;
212
213 if (unlikely(mem_id_init))
214 return 0;
215
216 rht = kzalloc(sizeof(*rht), GFP_KERNEL);
217 if (!rht)
218 return -ENOMEM;
219
220 ret = rhashtable_init(rht, &mem_id_rht_params);
221 if (ret < 0) {
222 kfree(rht);
223 return ret;
224 }
225 mem_id_ht = rht;
226 smp_mb(); /* mutex lock should provide enough pairing */
227 mem_id_init = true;
228
229 return 0;
230}
231
232/* Allocate a cyclic ID that maps to allocator pointer.
233 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
234 *
235 * Caller must lock mem_id_lock.
236 */
237static int __mem_id_cyclic_get(gfp_t gfp)
238{
239 int retries = 1;
240 int id;
241
242again:
243 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
244 if (id < 0) {
245 if (id == -ENOSPC) {
246 /* Cyclic allocator, reset next id */
247 if (retries--) {
248 mem_id_next = MEM_ID_MIN;
249 goto again;
250 }
251 }
252 return id; /* errno */
253 }
254 mem_id_next = id + 1;
255
256 return id;
257}
258
259static bool __is_supported_mem_type(enum xdp_mem_type type)
260{
261 if (type == MEM_TYPE_PAGE_POOL)
262 return is_page_pool_compiled_in();
263
264 if (type >= MEM_TYPE_MAX)
265 return false;
266
267 return true;
268}
269
270static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem,
271 enum xdp_mem_type type,
272 void *allocator)
273{
274 struct xdp_mem_allocator *xdp_alloc;
275 gfp_t gfp = GFP_KERNEL;
276 int id, errno, ret;
277 void *ptr;
278
279 if (!__is_supported_mem_type(type))
280 return ERR_PTR(-EOPNOTSUPP);
281
282 mem->type = type;
283
284 if (!allocator) {
285 if (type == MEM_TYPE_PAGE_POOL)
286 return ERR_PTR(-EINVAL); /* Setup time check page_pool req */
287 return NULL;
288 }
289
290 /* Delay init of rhashtable to save memory if feature isn't used */
291 if (!mem_id_init) {
292 mutex_lock(&mem_id_lock);
293 ret = __mem_id_init_hash_table();
294 mutex_unlock(&mem_id_lock);
295 if (ret < 0) {
296 WARN_ON(1);
297 return ERR_PTR(ret);
298 }
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;
330err:
331 mutex_unlock(&mem_id_lock);
332 kfree(xdp_alloc);
333 return ERR_PTR(errno);
334}
335
336int 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}
346EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
347
348int 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
367EXPORT_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 */
375void __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
408void 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 }
422out:
423 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
424}
425EXPORT_SYMBOL_GPL(xdp_return_frame);
426
427void 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 }
441out:
442 __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
443}
444EXPORT_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 */
456void 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}
467EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
468
469/* Must be called with rcu_read_lock held */
470void 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}
511EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
512
513void 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 }
527out:
528 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
529}
530EXPORT_SYMBOL_GPL(xdp_return_buff);
531
532/* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
533void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
534{
535 struct xdp_mem_allocator *xa;
536 struct page *page;
537
538 rcu_read_lock();
539 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
540 page = virt_to_head_page(data);
541 if (xa)
542 page_pool_release_page(xa->page_pool, page);
543 rcu_read_unlock();
544}
545EXPORT_SYMBOL_GPL(__xdp_release_frame);
546
547void xdp_attachment_setup(struct xdp_attachment_info *info,
548 struct netdev_bpf *bpf)
549{
550 if (info->prog)
551 bpf_prog_put(info->prog);
552 info->prog = bpf->prog;
553 info->flags = bpf->flags;
554}
555EXPORT_SYMBOL_GPL(xdp_attachment_setup);
556
557struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
558{
559 unsigned int metasize, totsize;
560 void *addr, *data_to_copy;
561 struct xdp_frame *xdpf;
562 struct page *page;
563
564 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
565 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
566 xdp->data - xdp->data_meta;
567 totsize = xdp->data_end - xdp->data + metasize;
568
569 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
570 return NULL;
571
572 page = dev_alloc_page();
573 if (!page)
574 return NULL;
575
576 addr = page_to_virt(page);
577 xdpf = addr;
578 memset(xdpf, 0, sizeof(*xdpf));
579
580 addr += sizeof(*xdpf);
581 data_to_copy = metasize ? xdp->data_meta : xdp->data;
582 memcpy(addr, data_to_copy, totsize);
583
584 xdpf->data = addr + metasize;
585 xdpf->len = totsize - metasize;
586 xdpf->headroom = 0;
587 xdpf->metasize = metasize;
588 xdpf->frame_sz = PAGE_SIZE;
589 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
590
591 xsk_buff_free(xdp);
592 return xdpf;
593}
594EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
595
596/* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
597void xdp_warn(const char *msg, const char *func, const int line)
598{
599 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
600};
601EXPORT_SYMBOL_GPL(xdp_warn);
602
603int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
604{
605 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
606 n_skb, skbs);
607 if (unlikely(!n_skb))
608 return -ENOMEM;
609
610 return 0;
611}
612EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
613
614struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
615 struct sk_buff *skb,
616 struct net_device *dev)
617{
618 struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
619 unsigned int headroom, frame_size;
620 void *hard_start;
621 u8 nr_frags;
622
623 /* xdp frags frame */
624 if (unlikely(xdp_frame_has_frags(xdpf)))
625 nr_frags = sinfo->nr_frags;
626
627 /* Part of headroom was reserved to xdpf */
628 headroom = sizeof(*xdpf) + xdpf->headroom;
629
630 /* Memory size backing xdp_frame data already have reserved
631 * room for build_skb to place skb_shared_info in tailroom.
632 */
633 frame_size = xdpf->frame_sz;
634
635 hard_start = xdpf->data - headroom;
636 skb = build_skb_around(skb, hard_start, frame_size);
637 if (unlikely(!skb))
638 return NULL;
639
640 skb_reserve(skb, headroom);
641 __skb_put(skb, xdpf->len);
642 if (xdpf->metasize)
643 skb_metadata_set(skb, xdpf->metasize);
644
645 if (unlikely(xdp_frame_has_frags(xdpf)))
646 xdp_update_skb_shared_info(skb, nr_frags,
647 sinfo->xdp_frags_size,
648 nr_frags * xdpf->frame_sz,
649 xdp_frame_is_frag_pfmemalloc(xdpf));
650
651 /* Essential SKB info: protocol and skb->dev */
652 skb->protocol = eth_type_trans(skb, dev);
653
654 /* Optional SKB info, currently missing:
655 * - HW checksum info (skb->ip_summed)
656 * - HW RX hash (skb_set_hash)
657 * - RX ring dev queue index (skb_record_rx_queue)
658 */
659
660 /* Until page_pool get SKB return path, release DMA here */
661 xdp_release_frame(xdpf);
662
663 /* Allow SKB to reuse area used by xdp_frame */
664 xdp_scrub_frame(xdpf);
665
666 return skb;
667}
668EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
669
670struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
671 struct net_device *dev)
672{
673 struct sk_buff *skb;
674
675 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
676 if (unlikely(!skb))
677 return NULL;
678
679 memset(skb, 0, offsetof(struct sk_buff, tail));
680
681 return __xdp_build_skb_from_frame(xdpf, skb, dev);
682}
683EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
684
685struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
686{
687 unsigned int headroom, totalsize;
688 struct xdp_frame *nxdpf;
689 struct page *page;
690 void *addr;
691
692 headroom = xdpf->headroom + sizeof(*xdpf);
693 totalsize = headroom + xdpf->len;
694
695 if (unlikely(totalsize > PAGE_SIZE))
696 return NULL;
697 page = dev_alloc_page();
698 if (!page)
699 return NULL;
700 addr = page_to_virt(page);
701
702 memcpy(addr, xdpf, totalsize);
703
704 nxdpf = addr;
705 nxdpf->data = addr + headroom;
706 nxdpf->frame_sz = PAGE_SIZE;
707 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
708 nxdpf->mem.id = 0;
709
710 return nxdpf;
711}