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
xdp_mem_id_hashfn(const void * data,u32 len,u32 seed)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
xdp_mem_id_cmp(struct rhashtable_compare_arg * arg,const void * ptr)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
__xdp_mem_allocator_rcu_free(struct rcu_head * rcu)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
mem_xa_remove(struct xdp_mem_allocator * xa)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
mem_allocator_disconnect(void * allocator)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
xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info * xdp_rxq)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
xdp_rxq_info_unreg(struct xdp_rxq_info * xdp_rxq)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 xdp_rxq_info_unreg_mem_model(xdp_rxq);
147
148 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
149 xdp_rxq->dev = NULL;
150 }
151 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
152
xdp_rxq_info_init(struct xdp_rxq_info * xdp_rxq)153 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
154 {
155 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
156 }
157
158 /* Returns 0 on success, negative on failure */
xdp_rxq_info_reg(struct xdp_rxq_info * xdp_rxq,struct net_device * dev,u32 queue_index,unsigned int napi_id)159 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
160 struct net_device *dev, u32 queue_index, unsigned int napi_id)
161 {
162 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
163 WARN(1, "Driver promised not to register this");
164 return -EINVAL;
165 }
166
167 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
168 WARN(1, "Missing unregister, handled but fix driver");
169 xdp_rxq_info_unreg(xdp_rxq);
170 }
171
172 if (!dev) {
173 WARN(1, "Missing net_device from driver");
174 return -ENODEV;
175 }
176
177 /* State either UNREGISTERED or NEW */
178 xdp_rxq_info_init(xdp_rxq);
179 xdp_rxq->dev = dev;
180 xdp_rxq->queue_index = queue_index;
181 xdp_rxq->napi_id = napi_id;
182
183 xdp_rxq->reg_state = REG_STATE_REGISTERED;
184 return 0;
185 }
186 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
187
xdp_rxq_info_unused(struct xdp_rxq_info * xdp_rxq)188 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
189 {
190 xdp_rxq->reg_state = REG_STATE_UNUSED;
191 }
192 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
193
xdp_rxq_info_is_reg(struct xdp_rxq_info * xdp_rxq)194 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
195 {
196 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
197 }
198 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
199
__mem_id_init_hash_table(void)200 static int __mem_id_init_hash_table(void)
201 {
202 struct rhashtable *rht;
203 int ret;
204
205 if (unlikely(mem_id_init))
206 return 0;
207
208 rht = kzalloc(sizeof(*rht), GFP_KERNEL);
209 if (!rht)
210 return -ENOMEM;
211
212 ret = rhashtable_init(rht, &mem_id_rht_params);
213 if (ret < 0) {
214 kfree(rht);
215 return ret;
216 }
217 mem_id_ht = rht;
218 smp_mb(); /* mutex lock should provide enough pairing */
219 mem_id_init = true;
220
221 return 0;
222 }
223
224 /* Allocate a cyclic ID that maps to allocator pointer.
225 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
226 *
227 * Caller must lock mem_id_lock.
228 */
__mem_id_cyclic_get(gfp_t gfp)229 static int __mem_id_cyclic_get(gfp_t gfp)
230 {
231 int retries = 1;
232 int id;
233
234 again:
235 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
236 if (id < 0) {
237 if (id == -ENOSPC) {
238 /* Cyclic allocator, reset next id */
239 if (retries--) {
240 mem_id_next = MEM_ID_MIN;
241 goto again;
242 }
243 }
244 return id; /* errno */
245 }
246 mem_id_next = id + 1;
247
248 return id;
249 }
250
__is_supported_mem_type(enum xdp_mem_type type)251 static bool __is_supported_mem_type(enum xdp_mem_type type)
252 {
253 if (type == MEM_TYPE_PAGE_POOL)
254 return is_page_pool_compiled_in();
255
256 if (type >= MEM_TYPE_MAX)
257 return false;
258
259 return true;
260 }
261
xdp_rxq_info_reg_mem_model(struct xdp_rxq_info * xdp_rxq,enum xdp_mem_type type,void * allocator)262 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
263 enum xdp_mem_type type, void *allocator)
264 {
265 struct xdp_mem_allocator *xdp_alloc;
266 gfp_t gfp = GFP_KERNEL;
267 int id, errno, ret;
268 void *ptr;
269
270 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
271 WARN(1, "Missing register, driver bug");
272 return -EFAULT;
273 }
274
275 if (!__is_supported_mem_type(type))
276 return -EOPNOTSUPP;
277
278 xdp_rxq->mem.type = type;
279
280 if (!allocator) {
281 if (type == MEM_TYPE_PAGE_POOL)
282 return -EINVAL; /* Setup time check page_pool req */
283 return 0;
284 }
285
286 /* Delay init of rhashtable to save memory if feature isn't used */
287 if (!mem_id_init) {
288 mutex_lock(&mem_id_lock);
289 ret = __mem_id_init_hash_table();
290 mutex_unlock(&mem_id_lock);
291 if (ret < 0) {
292 WARN_ON(1);
293 return ret;
294 }
295 }
296
297 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
298 if (!xdp_alloc)
299 return -ENOMEM;
300
301 mutex_lock(&mem_id_lock);
302 id = __mem_id_cyclic_get(gfp);
303 if (id < 0) {
304 errno = id;
305 goto err;
306 }
307 xdp_rxq->mem.id = id;
308 xdp_alloc->mem = xdp_rxq->mem;
309 xdp_alloc->allocator = allocator;
310
311 /* Insert allocator into ID lookup table */
312 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
313 if (IS_ERR(ptr)) {
314 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
315 xdp_rxq->mem.id = 0;
316 errno = PTR_ERR(ptr);
317 goto err;
318 }
319
320 if (type == MEM_TYPE_PAGE_POOL)
321 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
322
323 mutex_unlock(&mem_id_lock);
324
325 trace_mem_connect(xdp_alloc, xdp_rxq);
326 return 0;
327 err:
328 mutex_unlock(&mem_id_lock);
329 kfree(xdp_alloc);
330 return errno;
331 }
332 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
333
334 /* XDP RX runs under NAPI protection, and in different delivery error
335 * scenarios (e.g. queue full), it is possible to return the xdp_frame
336 * while still leveraging this protection. The @napi_direct boolean
337 * is used for those calls sites. Thus, allowing for faster recycling
338 * of xdp_frames/pages in those cases.
339 */
__xdp_return(void * data,struct xdp_mem_info * mem,bool napi_direct,struct xdp_buff * xdp)340 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
341 struct xdp_buff *xdp)
342 {
343 struct xdp_mem_allocator *xa;
344 struct page *page;
345
346 switch (mem->type) {
347 case MEM_TYPE_PAGE_POOL:
348 rcu_read_lock();
349 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
350 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
351 page = virt_to_head_page(data);
352 if (napi_direct && xdp_return_frame_no_direct())
353 napi_direct = false;
354 page_pool_put_full_page(xa->page_pool, page, napi_direct);
355 rcu_read_unlock();
356 break;
357 case MEM_TYPE_PAGE_SHARED:
358 page_frag_free(data);
359 break;
360 case MEM_TYPE_PAGE_ORDER0:
361 page = virt_to_page(data); /* Assumes order0 page*/
362 put_page(page);
363 break;
364 case MEM_TYPE_XSK_BUFF_POOL:
365 /* NB! Only valid from an xdp_buff! */
366 xsk_buff_free(xdp);
367 break;
368 default:
369 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
370 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
371 break;
372 }
373 }
374
xdp_return_frame(struct xdp_frame * xdpf)375 void xdp_return_frame(struct xdp_frame *xdpf)
376 {
377 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
378 }
379 EXPORT_SYMBOL_GPL(xdp_return_frame);
380
xdp_return_frame_rx_napi(struct xdp_frame * xdpf)381 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
382 {
383 __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
384 }
385 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
386
387 /* XDP bulk APIs introduce a defer/flush mechanism to return
388 * pages belonging to the same xdp_mem_allocator object
389 * (identified via the mem.id field) in bulk to optimize
390 * I-cache and D-cache.
391 * The bulk queue size is set to 16 to be aligned to how
392 * XDP_REDIRECT bulking works. The bulk is flushed when
393 * it is full or when mem.id changes.
394 * xdp_frame_bulk is usually stored/allocated on the function
395 * call-stack to avoid locking penalties.
396 */
xdp_flush_frame_bulk(struct xdp_frame_bulk * bq)397 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
398 {
399 struct xdp_mem_allocator *xa = bq->xa;
400
401 if (unlikely(!xa || !bq->count))
402 return;
403
404 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
405 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
406 bq->count = 0;
407 }
408 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
409
410 /* Must be called with rcu_read_lock held */
xdp_return_frame_bulk(struct xdp_frame * xdpf,struct xdp_frame_bulk * bq)411 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
412 struct xdp_frame_bulk *bq)
413 {
414 struct xdp_mem_info *mem = &xdpf->mem;
415 struct xdp_mem_allocator *xa;
416
417 if (mem->type != MEM_TYPE_PAGE_POOL) {
418 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
419 return;
420 }
421
422 xa = bq->xa;
423 if (unlikely(!xa)) {
424 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
425 bq->count = 0;
426 bq->xa = xa;
427 }
428
429 if (bq->count == XDP_BULK_QUEUE_SIZE)
430 xdp_flush_frame_bulk(bq);
431
432 if (unlikely(mem->id != xa->mem.id)) {
433 xdp_flush_frame_bulk(bq);
434 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
435 }
436
437 bq->q[bq->count++] = xdpf->data;
438 }
439 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
440
xdp_return_buff(struct xdp_buff * xdp)441 void xdp_return_buff(struct xdp_buff *xdp)
442 {
443 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
444 }
445
446 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
__xdp_release_frame(void * data,struct xdp_mem_info * mem)447 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
448 {
449 struct xdp_mem_allocator *xa;
450 struct page *page;
451
452 rcu_read_lock();
453 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
454 page = virt_to_head_page(data);
455 if (xa)
456 page_pool_release_page(xa->page_pool, page);
457 rcu_read_unlock();
458 }
459 EXPORT_SYMBOL_GPL(__xdp_release_frame);
460
xdp_attachment_setup(struct xdp_attachment_info * info,struct netdev_bpf * bpf)461 void xdp_attachment_setup(struct xdp_attachment_info *info,
462 struct netdev_bpf *bpf)
463 {
464 if (info->prog)
465 bpf_prog_put(info->prog);
466 info->prog = bpf->prog;
467 info->flags = bpf->flags;
468 }
469 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
470
xdp_convert_zc_to_xdp_frame(struct xdp_buff * xdp)471 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
472 {
473 unsigned int metasize, totsize;
474 void *addr, *data_to_copy;
475 struct xdp_frame *xdpf;
476 struct page *page;
477
478 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
479 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
480 xdp->data - xdp->data_meta;
481 totsize = xdp->data_end - xdp->data + metasize;
482
483 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
484 return NULL;
485
486 page = dev_alloc_page();
487 if (!page)
488 return NULL;
489
490 addr = page_to_virt(page);
491 xdpf = addr;
492 memset(xdpf, 0, sizeof(*xdpf));
493
494 addr += sizeof(*xdpf);
495 data_to_copy = metasize ? xdp->data_meta : xdp->data;
496 memcpy(addr, data_to_copy, totsize);
497
498 xdpf->data = addr + metasize;
499 xdpf->len = totsize - metasize;
500 xdpf->headroom = 0;
501 xdpf->metasize = metasize;
502 xdpf->frame_sz = PAGE_SIZE;
503 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
504
505 xsk_buff_free(xdp);
506 return xdpf;
507 }
508 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
509
510 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
xdp_warn(const char * msg,const char * func,const int line)511 void xdp_warn(const char *msg, const char *func, const int line)
512 {
513 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
514 };
515 EXPORT_SYMBOL_GPL(xdp_warn);
516
xdp_alloc_skb_bulk(void ** skbs,int n_skb,gfp_t gfp)517 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
518 {
519 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
520 n_skb, skbs);
521 if (unlikely(!n_skb))
522 return -ENOMEM;
523
524 return 0;
525 }
526 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
527
__xdp_build_skb_from_frame(struct xdp_frame * xdpf,struct sk_buff * skb,struct net_device * dev)528 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
529 struct sk_buff *skb,
530 struct net_device *dev)
531 {
532 unsigned int headroom, frame_size;
533 void *hard_start;
534
535 /* Part of headroom was reserved to xdpf */
536 headroom = sizeof(*xdpf) + xdpf->headroom;
537
538 /* Memory size backing xdp_frame data already have reserved
539 * room for build_skb to place skb_shared_info in tailroom.
540 */
541 frame_size = xdpf->frame_sz;
542
543 hard_start = xdpf->data - headroom;
544 skb = build_skb_around(skb, hard_start, frame_size);
545 if (unlikely(!skb))
546 return NULL;
547
548 skb_reserve(skb, headroom);
549 __skb_put(skb, xdpf->len);
550 if (xdpf->metasize)
551 skb_metadata_set(skb, xdpf->metasize);
552
553 /* Essential SKB info: protocol and skb->dev */
554 skb->protocol = eth_type_trans(skb, dev);
555
556 /* Optional SKB info, currently missing:
557 * - HW checksum info (skb->ip_summed)
558 * - HW RX hash (skb_set_hash)
559 * - RX ring dev queue index (skb_record_rx_queue)
560 */
561
562 /* Until page_pool get SKB return path, release DMA here */
563 xdp_release_frame(xdpf);
564
565 /* Allow SKB to reuse area used by xdp_frame */
566 xdp_scrub_frame(xdpf);
567
568 return skb;
569 }
570 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
571
xdp_build_skb_from_frame(struct xdp_frame * xdpf,struct net_device * dev)572 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
573 struct net_device *dev)
574 {
575 struct sk_buff *skb;
576
577 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
578 if (unlikely(!skb))
579 return NULL;
580
581 memset(skb, 0, offsetof(struct sk_buff, tail));
582
583 return __xdp_build_skb_from_frame(xdpf, skb, dev);
584 }
585 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
586
xdpf_clone(struct xdp_frame * xdpf)587 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
588 {
589 unsigned int headroom, totalsize;
590 struct xdp_frame *nxdpf;
591 struct page *page;
592 void *addr;
593
594 headroom = xdpf->headroom + sizeof(*xdpf);
595 totalsize = headroom + xdpf->len;
596
597 if (unlikely(totalsize > PAGE_SIZE))
598 return NULL;
599 page = dev_alloc_page();
600 if (!page)
601 return NULL;
602 addr = page_to_virt(page);
603
604 memcpy(addr, xdpf, totalsize);
605
606 nxdpf = addr;
607 nxdpf->data = addr + headroom;
608 nxdpf->frame_sz = PAGE_SIZE;
609 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
610 nxdpf->mem.id = 0;
611
612 return nxdpf;
613 }
614