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