1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
7 #include <linux/filter.h>
8 #include <linux/types.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>
18 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
19 #include <trace/events/xdp.h>
20 #include <net/xdp_sock_drv.h>
22 #define REG_STATE_NEW 0x0
23 #define REG_STATE_REGISTERED 0x1
24 #define REG_STATE_UNREGISTERED 0x2
25 #define REG_STATE_UNUSED 0x3
27 static DEFINE_IDA(mem_id_pool);
28 static DEFINE_MUTEX(mem_id_lock);
29 #define MEM_ID_MAX 0xFFFE
31 static int mem_id_next = MEM_ID_MIN;
33 static bool mem_id_init; /* false */
34 static struct rhashtable *mem_id_ht;
36 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
41 BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
44 /* Use cyclic increasing ID as direct hash key */
48 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
51 const struct xdp_mem_allocator *xa = ptr;
52 u32 mem_id = *(u32 *)arg->key;
54 return xa->mem.id != mem_id;
57 static const struct rhashtable_params mem_id_rht_params = {
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,
64 .automatic_shrinking = true,
65 .hashfn = xdp_mem_id_hashfn,
66 .obj_cmpfn = xdp_mem_id_cmp,
69 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
71 struct xdp_mem_allocator *xa;
73 xa = container_of(rcu, struct xdp_mem_allocator, rcu);
75 /* Allow this ID to be reused */
76 ida_simple_remove(&mem_id_pool, xa->mem.id);
81 static void mem_xa_remove(struct xdp_mem_allocator *xa)
83 trace_mem_disconnect(xa);
85 if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
86 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
89 static void mem_allocator_disconnect(void *allocator)
91 struct xdp_mem_allocator *xa;
92 struct rhashtable_iter iter;
94 mutex_lock(&mem_id_lock);
96 rhashtable_walk_enter(mem_id_ht, &iter);
98 rhashtable_walk_start(&iter);
100 while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
101 if (xa->allocator == allocator)
105 rhashtable_walk_stop(&iter);
107 } while (xa == ERR_PTR(-EAGAIN));
108 rhashtable_walk_exit(&iter);
110 mutex_unlock(&mem_id_lock);
113 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
115 struct xdp_mem_allocator *xa;
116 int type = xdp_rxq->mem.type;
117 int id = xdp_rxq->mem.id;
119 /* Reset mem info to defaults */
121 xdp_rxq->mem.type = 0;
123 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
124 WARN(1, "Missing register, driver bug");
131 if (type == MEM_TYPE_PAGE_POOL) {
133 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
134 page_pool_destroy(xa->page_pool);
138 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
140 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
142 /* Simplify driver cleanup code paths, allow unreg "unused" */
143 if (xdp_rxq->reg_state == REG_STATE_UNUSED)
146 WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
148 xdp_rxq_info_unreg_mem_model(xdp_rxq);
150 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
153 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
155 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
157 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
160 /* Returns 0 on success, negative on failure */
161 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
162 struct net_device *dev, u32 queue_index, unsigned int napi_id)
164 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
165 WARN(1, "Driver promised not to register this");
169 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
170 WARN(1, "Missing unregister, handled but fix driver");
171 xdp_rxq_info_unreg(xdp_rxq);
175 WARN(1, "Missing net_device from driver");
179 /* State either UNREGISTERED or NEW */
180 xdp_rxq_info_init(xdp_rxq);
182 xdp_rxq->queue_index = queue_index;
183 xdp_rxq->napi_id = napi_id;
185 xdp_rxq->reg_state = REG_STATE_REGISTERED;
188 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
190 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
192 xdp_rxq->reg_state = REG_STATE_UNUSED;
194 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
196 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
198 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
200 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
202 static int __mem_id_init_hash_table(void)
204 struct rhashtable *rht;
207 if (unlikely(mem_id_init))
210 rht = kzalloc(sizeof(*rht), GFP_KERNEL);
214 ret = rhashtable_init(rht, &mem_id_rht_params);
220 smp_mb(); /* mutex lock should provide enough pairing */
226 /* Allocate a cyclic ID that maps to allocator pointer.
227 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
229 * Caller must lock mem_id_lock.
231 static int __mem_id_cyclic_get(gfp_t gfp)
237 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
240 /* Cyclic allocator, reset next id */
242 mem_id_next = MEM_ID_MIN;
246 return id; /* errno */
248 mem_id_next = id + 1;
253 static bool __is_supported_mem_type(enum xdp_mem_type type)
255 if (type == MEM_TYPE_PAGE_POOL)
256 return is_page_pool_compiled_in();
258 if (type >= MEM_TYPE_MAX)
264 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
265 enum xdp_mem_type type, void *allocator)
267 struct xdp_mem_allocator *xdp_alloc;
268 gfp_t gfp = GFP_KERNEL;
272 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
273 WARN(1, "Missing register, driver bug");
277 if (!__is_supported_mem_type(type))
280 xdp_rxq->mem.type = type;
283 if (type == MEM_TYPE_PAGE_POOL)
284 return -EINVAL; /* Setup time check page_pool req */
288 /* Delay init of rhashtable to save memory if feature isn't used */
290 mutex_lock(&mem_id_lock);
291 ret = __mem_id_init_hash_table();
292 mutex_unlock(&mem_id_lock);
299 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
303 mutex_lock(&mem_id_lock);
304 id = __mem_id_cyclic_get(gfp);
309 xdp_rxq->mem.id = id;
310 xdp_alloc->mem = xdp_rxq->mem;
311 xdp_alloc->allocator = allocator;
313 /* Insert allocator into ID lookup table */
314 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
316 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
318 errno = PTR_ERR(ptr);
322 if (type == MEM_TYPE_PAGE_POOL)
323 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
325 mutex_unlock(&mem_id_lock);
327 trace_mem_connect(xdp_alloc, xdp_rxq);
330 mutex_unlock(&mem_id_lock);
334 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
336 /* XDP RX runs under NAPI protection, and in different delivery error
337 * scenarios (e.g. queue full), it is possible to return the xdp_frame
338 * while still leveraging this protection. The @napi_direct boolean
339 * is used for those calls sites. Thus, allowing for faster recycling
340 * of xdp_frames/pages in those cases.
342 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
343 struct xdp_buff *xdp)
345 struct xdp_mem_allocator *xa;
349 case MEM_TYPE_PAGE_POOL:
351 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
352 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
353 page = virt_to_head_page(data);
354 if (napi_direct && xdp_return_frame_no_direct())
356 page_pool_put_full_page(xa->page_pool, page, napi_direct);
359 case MEM_TYPE_PAGE_SHARED:
360 page_frag_free(data);
362 case MEM_TYPE_PAGE_ORDER0:
363 page = virt_to_page(data); /* Assumes order0 page*/
366 case MEM_TYPE_XSK_BUFF_POOL:
367 /* NB! Only valid from an xdp_buff! */
371 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
372 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
377 void xdp_return_frame(struct xdp_frame *xdpf)
379 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
381 EXPORT_SYMBOL_GPL(xdp_return_frame);
383 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
385 __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
387 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
389 /* XDP bulk APIs introduce a defer/flush mechanism to return
390 * pages belonging to the same xdp_mem_allocator object
391 * (identified via the mem.id field) in bulk to optimize
392 * I-cache and D-cache.
393 * The bulk queue size is set to 16 to be aligned to how
394 * XDP_REDIRECT bulking works. The bulk is flushed when
395 * it is full or when mem.id changes.
396 * xdp_frame_bulk is usually stored/allocated on the function
397 * call-stack to avoid locking penalties.
399 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
401 struct xdp_mem_allocator *xa = bq->xa;
403 if (unlikely(!xa || !bq->count))
406 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
407 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
410 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
412 /* Must be called with rcu_read_lock held */
413 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
414 struct xdp_frame_bulk *bq)
416 struct xdp_mem_info *mem = &xdpf->mem;
417 struct xdp_mem_allocator *xa;
419 if (mem->type != MEM_TYPE_PAGE_POOL) {
420 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
426 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
431 if (bq->count == XDP_BULK_QUEUE_SIZE)
432 xdp_flush_frame_bulk(bq);
434 if (unlikely(mem->id != xa->mem.id)) {
435 xdp_flush_frame_bulk(bq);
436 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
439 bq->q[bq->count++] = xdpf->data;
441 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
443 void xdp_return_buff(struct xdp_buff *xdp)
445 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
448 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
449 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
451 struct xdp_mem_allocator *xa;
455 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
456 page = virt_to_head_page(data);
458 page_pool_release_page(xa->page_pool, page);
461 EXPORT_SYMBOL_GPL(__xdp_release_frame);
463 void xdp_attachment_setup(struct xdp_attachment_info *info,
464 struct netdev_bpf *bpf)
467 bpf_prog_put(info->prog);
468 info->prog = bpf->prog;
469 info->flags = bpf->flags;
471 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
473 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
475 unsigned int metasize, totsize;
476 void *addr, *data_to_copy;
477 struct xdp_frame *xdpf;
480 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
481 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
482 xdp->data - xdp->data_meta;
483 totsize = xdp->data_end - xdp->data + metasize;
485 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
488 page = dev_alloc_page();
492 addr = page_to_virt(page);
494 memset(xdpf, 0, sizeof(*xdpf));
496 addr += sizeof(*xdpf);
497 data_to_copy = metasize ? xdp->data_meta : xdp->data;
498 memcpy(addr, data_to_copy, totsize);
500 xdpf->data = addr + metasize;
501 xdpf->len = totsize - metasize;
503 xdpf->metasize = metasize;
504 xdpf->frame_sz = PAGE_SIZE;
505 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
510 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
512 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
513 void xdp_warn(const char *msg, const char *func, const int line)
515 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
517 EXPORT_SYMBOL_GPL(xdp_warn);
519 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
521 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
523 if (unlikely(!n_skb))
528 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
530 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
532 struct net_device *dev)
534 unsigned int headroom, frame_size;
537 /* Part of headroom was reserved to xdpf */
538 headroom = sizeof(*xdpf) + xdpf->headroom;
540 /* Memory size backing xdp_frame data already have reserved
541 * room for build_skb to place skb_shared_info in tailroom.
543 frame_size = xdpf->frame_sz;
545 hard_start = xdpf->data - headroom;
546 skb = build_skb_around(skb, hard_start, frame_size);
550 skb_reserve(skb, headroom);
551 __skb_put(skb, xdpf->len);
553 skb_metadata_set(skb, xdpf->metasize);
555 /* Essential SKB info: protocol and skb->dev */
556 skb->protocol = eth_type_trans(skb, dev);
558 /* Optional SKB info, currently missing:
559 * - HW checksum info (skb->ip_summed)
560 * - HW RX hash (skb_set_hash)
561 * - RX ring dev queue index (skb_record_rx_queue)
564 /* Until page_pool get SKB return path, release DMA here */
565 xdp_release_frame(xdpf);
567 /* Allow SKB to reuse area used by xdp_frame */
568 xdp_scrub_frame(xdpf);
572 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
574 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
575 struct net_device *dev)
579 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
583 memset(skb, 0, offsetof(struct sk_buff, tail));
585 return __xdp_build_skb_from_frame(xdpf, skb, dev);
587 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
589 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
591 unsigned int headroom, totalsize;
592 struct xdp_frame *nxdpf;
596 headroom = xdpf->headroom + sizeof(*xdpf);
597 totalsize = headroom + xdpf->len;
599 if (unlikely(totalsize > PAGE_SIZE))
601 page = dev_alloc_page();
604 addr = page_to_virt(page);
606 memcpy(addr, xdpf, totalsize);
609 nxdpf->data = addr + headroom;
610 nxdpf->frame_sz = PAGE_SIZE;
611 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;