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_unreg_mem_model(struct xdp_mem_info *mem)
115 struct xdp_mem_allocator *xa;
116 int type = mem->type;
119 /* Reset mem info to defaults */
126 if (type == MEM_TYPE_PAGE_POOL) {
128 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
129 page_pool_destroy(xa->page_pool);
133 EXPORT_SYMBOL_GPL(xdp_unreg_mem_model);
135 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
137 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
138 WARN(1, "Missing register, driver bug");
142 xdp_unreg_mem_model(&xdp_rxq->mem);
144 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
146 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
148 /* Simplify driver cleanup code paths, allow unreg "unused" */
149 if (xdp_rxq->reg_state == REG_STATE_UNUSED)
152 xdp_rxq_info_unreg_mem_model(xdp_rxq);
154 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
157 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
159 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
161 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
164 /* Returns 0 on success, negative on failure */
165 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
166 struct net_device *dev, u32 queue_index, unsigned int napi_id)
169 WARN(1, "Missing net_device from driver");
173 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
174 WARN(1, "Driver promised not to register this");
178 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
179 WARN(1, "Missing unregister, handled but fix driver");
180 xdp_rxq_info_unreg(xdp_rxq);
183 /* State either UNREGISTERED or NEW */
184 xdp_rxq_info_init(xdp_rxq);
186 xdp_rxq->queue_index = queue_index;
187 xdp_rxq->napi_id = napi_id;
189 xdp_rxq->reg_state = REG_STATE_REGISTERED;
192 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
194 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
196 xdp_rxq->reg_state = REG_STATE_UNUSED;
198 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
200 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
202 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
204 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
206 static int __mem_id_init_hash_table(void)
208 struct rhashtable *rht;
211 if (unlikely(mem_id_init))
214 rht = kzalloc(sizeof(*rht), GFP_KERNEL);
218 ret = rhashtable_init(rht, &mem_id_rht_params);
224 smp_mb(); /* mutex lock should provide enough pairing */
230 /* Allocate a cyclic ID that maps to allocator pointer.
231 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
233 * Caller must lock mem_id_lock.
235 static int __mem_id_cyclic_get(gfp_t gfp)
241 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
244 /* Cyclic allocator, reset next id */
246 mem_id_next = MEM_ID_MIN;
250 return id; /* errno */
252 mem_id_next = id + 1;
257 static bool __is_supported_mem_type(enum xdp_mem_type type)
259 if (type == MEM_TYPE_PAGE_POOL)
260 return is_page_pool_compiled_in();
262 if (type >= MEM_TYPE_MAX)
268 static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem,
269 enum xdp_mem_type type,
272 struct xdp_mem_allocator *xdp_alloc;
273 gfp_t gfp = GFP_KERNEL;
277 if (!__is_supported_mem_type(type))
278 return ERR_PTR(-EOPNOTSUPP);
283 if (type == MEM_TYPE_PAGE_POOL)
284 return ERR_PTR(-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);
301 return ERR_PTR(-ENOMEM);
303 mutex_lock(&mem_id_lock);
304 id = __mem_id_cyclic_get(gfp);
310 xdp_alloc->mem = *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, mem->id);
318 errno = PTR_ERR(ptr);
322 if (type == MEM_TYPE_PAGE_POOL)
323 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem);
325 mutex_unlock(&mem_id_lock);
329 mutex_unlock(&mem_id_lock);
331 return ERR_PTR(errno);
334 int xdp_reg_mem_model(struct xdp_mem_info *mem,
335 enum xdp_mem_type type, void *allocator)
337 struct xdp_mem_allocator *xdp_alloc;
339 xdp_alloc = __xdp_reg_mem_model(mem, type, allocator);
340 if (IS_ERR(xdp_alloc))
341 return PTR_ERR(xdp_alloc);
344 EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
346 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
347 enum xdp_mem_type type, void *allocator)
349 struct xdp_mem_allocator *xdp_alloc;
351 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
352 WARN(1, "Missing register, driver bug");
356 xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator);
357 if (IS_ERR(xdp_alloc))
358 return PTR_ERR(xdp_alloc);
360 trace_mem_connect(xdp_alloc, xdp_rxq);
364 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
366 /* XDP RX runs under NAPI protection, and in different delivery error
367 * scenarios (e.g. queue full), it is possible to return the xdp_frame
368 * while still leveraging this protection. The @napi_direct boolean
369 * is used for those calls sites. Thus, allowing for faster recycling
370 * of xdp_frames/pages in those cases.
372 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
373 struct xdp_buff *xdp)
375 struct xdp_mem_allocator *xa;
379 case MEM_TYPE_PAGE_POOL:
381 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
382 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
383 page = virt_to_head_page(data);
384 if (napi_direct && xdp_return_frame_no_direct())
386 page_pool_put_full_page(xa->page_pool, page, napi_direct);
389 case MEM_TYPE_PAGE_SHARED:
390 page_frag_free(data);
392 case MEM_TYPE_PAGE_ORDER0:
393 page = virt_to_page(data); /* Assumes order0 page*/
396 case MEM_TYPE_XSK_BUFF_POOL:
397 /* NB! Only valid from an xdp_buff! */
401 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
402 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
407 void xdp_return_frame(struct xdp_frame *xdpf)
409 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
411 EXPORT_SYMBOL_GPL(xdp_return_frame);
413 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
415 __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
417 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
419 /* XDP bulk APIs introduce a defer/flush mechanism to return
420 * pages belonging to the same xdp_mem_allocator object
421 * (identified via the mem.id field) in bulk to optimize
422 * I-cache and D-cache.
423 * The bulk queue size is set to 16 to be aligned to how
424 * XDP_REDIRECT bulking works. The bulk is flushed when
425 * it is full or when mem.id changes.
426 * xdp_frame_bulk is usually stored/allocated on the function
427 * call-stack to avoid locking penalties.
429 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
431 struct xdp_mem_allocator *xa = bq->xa;
433 if (unlikely(!xa || !bq->count))
436 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
437 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
440 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
442 /* Must be called with rcu_read_lock held */
443 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
444 struct xdp_frame_bulk *bq)
446 struct xdp_mem_info *mem = &xdpf->mem;
447 struct xdp_mem_allocator *xa;
449 if (mem->type != MEM_TYPE_PAGE_POOL) {
450 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
456 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
461 if (bq->count == XDP_BULK_QUEUE_SIZE)
462 xdp_flush_frame_bulk(bq);
464 if (unlikely(mem->id != xa->mem.id)) {
465 xdp_flush_frame_bulk(bq);
466 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
469 bq->q[bq->count++] = xdpf->data;
471 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
473 void xdp_return_buff(struct xdp_buff *xdp)
475 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
478 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
479 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
481 struct xdp_mem_allocator *xa;
485 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
486 page = virt_to_head_page(data);
488 page_pool_release_page(xa->page_pool, page);
491 EXPORT_SYMBOL_GPL(__xdp_release_frame);
493 void xdp_attachment_setup(struct xdp_attachment_info *info,
494 struct netdev_bpf *bpf)
497 bpf_prog_put(info->prog);
498 info->prog = bpf->prog;
499 info->flags = bpf->flags;
501 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
503 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
505 unsigned int metasize, totsize;
506 void *addr, *data_to_copy;
507 struct xdp_frame *xdpf;
510 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
511 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
512 xdp->data - xdp->data_meta;
513 totsize = xdp->data_end - xdp->data + metasize;
515 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
518 page = dev_alloc_page();
522 addr = page_to_virt(page);
524 memset(xdpf, 0, sizeof(*xdpf));
526 addr += sizeof(*xdpf);
527 data_to_copy = metasize ? xdp->data_meta : xdp->data;
528 memcpy(addr, data_to_copy, totsize);
530 xdpf->data = addr + metasize;
531 xdpf->len = totsize - metasize;
533 xdpf->metasize = metasize;
534 xdpf->frame_sz = PAGE_SIZE;
535 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
540 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
542 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
543 void xdp_warn(const char *msg, const char *func, const int line)
545 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
547 EXPORT_SYMBOL_GPL(xdp_warn);
549 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
551 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
553 if (unlikely(!n_skb))
558 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
560 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
562 struct net_device *dev)
564 unsigned int headroom, frame_size;
567 /* Part of headroom was reserved to xdpf */
568 headroom = sizeof(*xdpf) + xdpf->headroom;
570 /* Memory size backing xdp_frame data already have reserved
571 * room for build_skb to place skb_shared_info in tailroom.
573 frame_size = xdpf->frame_sz;
575 hard_start = xdpf->data - headroom;
576 skb = build_skb_around(skb, hard_start, frame_size);
580 skb_reserve(skb, headroom);
581 __skb_put(skb, xdpf->len);
583 skb_metadata_set(skb, xdpf->metasize);
585 /* Essential SKB info: protocol and skb->dev */
586 skb->protocol = eth_type_trans(skb, dev);
588 /* Optional SKB info, currently missing:
589 * - HW checksum info (skb->ip_summed)
590 * - HW RX hash (skb_set_hash)
591 * - RX ring dev queue index (skb_record_rx_queue)
594 /* Until page_pool get SKB return path, release DMA here */
595 xdp_release_frame(xdpf);
597 /* Allow SKB to reuse area used by xdp_frame */
598 xdp_scrub_frame(xdpf);
602 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
604 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
605 struct net_device *dev)
609 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
613 memset(skb, 0, offsetof(struct sk_buff, tail));
615 return __xdp_build_skb_from_frame(xdpf, skb, dev);
617 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
619 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
621 unsigned int headroom, totalsize;
622 struct xdp_frame *nxdpf;
626 headroom = xdpf->headroom + sizeof(*xdpf);
627 totalsize = headroom + xdpf->len;
629 if (unlikely(totalsize > PAGE_SIZE))
631 page = dev_alloc_page();
634 addr = page_to_virt(page);
636 memcpy(addr, xdpf, totalsize);
639 nxdpf->data = addr + headroom;
640 nxdpf->frame_sz = PAGE_SIZE;
641 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;