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 <net/page_pool.h>
17 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
18 #include <trace/events/xdp.h>
20 #define REG_STATE_NEW 0x0
21 #define REG_STATE_REGISTERED 0x1
22 #define REG_STATE_UNREGISTERED 0x2
23 #define REG_STATE_UNUSED 0x3
25 static DEFINE_IDA(mem_id_pool);
26 static DEFINE_MUTEX(mem_id_lock);
27 #define MEM_ID_MAX 0xFFFE
29 static int mem_id_next = MEM_ID_MIN;
31 static bool mem_id_init; /* false */
32 static struct rhashtable *mem_id_ht;
34 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
39 BUILD_BUG_ON(FIELD_SIZEOF(struct xdp_mem_allocator, mem.id)
42 /* Use cyclic increasing ID as direct hash key */
46 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
49 const struct xdp_mem_allocator *xa = ptr;
50 u32 mem_id = *(u32 *)arg->key;
52 return xa->mem.id != mem_id;
55 static const struct rhashtable_params mem_id_rht_params = {
57 .head_offset = offsetof(struct xdp_mem_allocator, node),
58 .key_offset = offsetof(struct xdp_mem_allocator, mem.id),
59 .key_len = FIELD_SIZEOF(struct xdp_mem_allocator, mem.id),
60 .max_size = MEM_ID_MAX,
62 .automatic_shrinking = true,
63 .hashfn = xdp_mem_id_hashfn,
64 .obj_cmpfn = xdp_mem_id_cmp,
67 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
69 struct xdp_mem_allocator *xa;
71 xa = container_of(rcu, struct xdp_mem_allocator, rcu);
73 /* Allow this ID to be reused */
74 ida_simple_remove(&mem_id_pool, xa->mem.id);
78 xa->mem.type = 0xF0F0;
79 xa->allocator = (void *)0xDEAD9001;
84 static void mem_xa_remove(struct xdp_mem_allocator *xa)
86 trace_mem_disconnect(xa);
88 if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
89 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
92 static void mem_allocator_disconnect(void *allocator)
94 struct xdp_mem_allocator *xa;
95 struct rhashtable_iter iter;
97 mutex_lock(&mem_id_lock);
99 rhashtable_walk_enter(mem_id_ht, &iter);
101 rhashtable_walk_start(&iter);
103 while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
104 if (xa->allocator == allocator)
108 rhashtable_walk_stop(&iter);
110 } while (xa == ERR_PTR(-EAGAIN));
111 rhashtable_walk_exit(&iter);
113 mutex_unlock(&mem_id_lock);
116 static void mem_id_disconnect(int id)
118 struct xdp_mem_allocator *xa;
120 mutex_lock(&mem_id_lock);
122 xa = rhashtable_lookup_fast(mem_id_ht, &id, mem_id_rht_params);
124 mutex_unlock(&mem_id_lock);
125 WARN(1, "Request remove non-existing id(%d), driver bug?", id);
129 trace_mem_disconnect(xa);
131 if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
132 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
134 mutex_unlock(&mem_id_lock);
137 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
139 struct xdp_mem_allocator *xa;
140 int id = xdp_rxq->mem.id;
142 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
143 WARN(1, "Missing register, driver bug");
150 if (xdp_rxq->mem.type == MEM_TYPE_ZERO_COPY)
151 return mem_id_disconnect(id);
153 if (xdp_rxq->mem.type == MEM_TYPE_PAGE_POOL) {
155 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
156 page_pool_destroy(xa->page_pool);
160 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
162 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
164 /* Simplify driver cleanup code paths, allow unreg "unused" */
165 if (xdp_rxq->reg_state == REG_STATE_UNUSED)
168 WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
170 xdp_rxq_info_unreg_mem_model(xdp_rxq);
172 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
175 /* Reset mem info to defaults */
177 xdp_rxq->mem.type = 0;
179 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
181 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
183 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
186 /* Returns 0 on success, negative on failure */
187 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
188 struct net_device *dev, u32 queue_index)
190 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
191 WARN(1, "Driver promised not to register this");
195 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
196 WARN(1, "Missing unregister, handled but fix driver");
197 xdp_rxq_info_unreg(xdp_rxq);
201 WARN(1, "Missing net_device from driver");
205 /* State either UNREGISTERED or NEW */
206 xdp_rxq_info_init(xdp_rxq);
208 xdp_rxq->queue_index = queue_index;
210 xdp_rxq->reg_state = REG_STATE_REGISTERED;
213 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
215 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
217 xdp_rxq->reg_state = REG_STATE_UNUSED;
219 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
221 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
223 return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
225 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
227 static int __mem_id_init_hash_table(void)
229 struct rhashtable *rht;
232 if (unlikely(mem_id_init))
235 rht = kzalloc(sizeof(*rht), GFP_KERNEL);
239 ret = rhashtable_init(rht, &mem_id_rht_params);
245 smp_mb(); /* mutex lock should provide enough pairing */
251 /* Allocate a cyclic ID that maps to allocator pointer.
252 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
254 * Caller must lock mem_id_lock.
256 static int __mem_id_cyclic_get(gfp_t gfp)
262 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
265 /* Cyclic allocator, reset next id */
267 mem_id_next = MEM_ID_MIN;
271 return id; /* errno */
273 mem_id_next = id + 1;
278 static bool __is_supported_mem_type(enum xdp_mem_type type)
280 if (type == MEM_TYPE_PAGE_POOL)
281 return is_page_pool_compiled_in();
283 if (type >= MEM_TYPE_MAX)
289 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
290 enum xdp_mem_type type, void *allocator)
292 struct xdp_mem_allocator *xdp_alloc;
293 gfp_t gfp = GFP_KERNEL;
297 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
298 WARN(1, "Missing register, driver bug");
302 if (!__is_supported_mem_type(type))
305 xdp_rxq->mem.type = type;
308 if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
309 return -EINVAL; /* Setup time check page_pool req */
313 /* Delay init of rhashtable to save memory if feature isn't used */
315 mutex_lock(&mem_id_lock);
316 ret = __mem_id_init_hash_table();
317 mutex_unlock(&mem_id_lock);
324 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
328 mutex_lock(&mem_id_lock);
329 id = __mem_id_cyclic_get(gfp);
334 xdp_rxq->mem.id = id;
335 xdp_alloc->mem = xdp_rxq->mem;
336 xdp_alloc->allocator = allocator;
338 /* Insert allocator into ID lookup table */
339 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
341 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
343 errno = PTR_ERR(ptr);
347 if (type == MEM_TYPE_PAGE_POOL)
348 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
350 mutex_unlock(&mem_id_lock);
352 trace_mem_connect(xdp_alloc, xdp_rxq);
355 mutex_unlock(&mem_id_lock);
359 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
361 /* XDP RX runs under NAPI protection, and in different delivery error
362 * scenarios (e.g. queue full), it is possible to return the xdp_frame
363 * while still leveraging this protection. The @napi_direct boolian
364 * is used for those calls sites. Thus, allowing for faster recycling
365 * of xdp_frames/pages in those cases.
367 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
368 unsigned long handle)
370 struct xdp_mem_allocator *xa;
374 case MEM_TYPE_PAGE_POOL:
376 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
377 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
378 page = virt_to_head_page(data);
379 napi_direct &= !xdp_return_frame_no_direct();
380 page_pool_put_page(xa->page_pool, page, napi_direct);
383 case MEM_TYPE_PAGE_SHARED:
384 page_frag_free(data);
386 case MEM_TYPE_PAGE_ORDER0:
387 page = virt_to_page(data); /* Assumes order0 page*/
390 case MEM_TYPE_ZERO_COPY:
391 /* NB! Only valid from an xdp_buff! */
393 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
394 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
395 xa->zc_alloc->free(xa->zc_alloc, handle);
398 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
403 void xdp_return_frame(struct xdp_frame *xdpf)
405 __xdp_return(xdpf->data, &xdpf->mem, false, 0);
407 EXPORT_SYMBOL_GPL(xdp_return_frame);
409 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
411 __xdp_return(xdpf->data, &xdpf->mem, true, 0);
413 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
415 void xdp_return_buff(struct xdp_buff *xdp)
417 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
419 EXPORT_SYMBOL_GPL(xdp_return_buff);
421 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
422 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
424 struct xdp_mem_allocator *xa;
428 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
429 page = virt_to_head_page(data);
431 page_pool_release_page(xa->page_pool, page);
434 EXPORT_SYMBOL_GPL(__xdp_release_frame);
436 int xdp_attachment_query(struct xdp_attachment_info *info,
437 struct netdev_bpf *bpf)
439 bpf->prog_id = info->prog ? info->prog->aux->id : 0;
440 bpf->prog_flags = info->prog ? info->flags : 0;
443 EXPORT_SYMBOL_GPL(xdp_attachment_query);
445 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
446 struct netdev_bpf *bpf)
448 if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
449 NL_SET_ERR_MSG(bpf->extack,
450 "program loaded with different flags");
455 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
457 void xdp_attachment_setup(struct xdp_attachment_info *info,
458 struct netdev_bpf *bpf)
461 bpf_prog_put(info->prog);
462 info->prog = bpf->prog;
463 info->flags = bpf->flags;
465 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
467 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
469 unsigned int metasize, totsize;
470 void *addr, *data_to_copy;
471 struct xdp_frame *xdpf;
474 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
475 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
476 xdp->data - xdp->data_meta;
477 totsize = xdp->data_end - xdp->data + metasize;
479 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
482 page = dev_alloc_page();
486 addr = page_to_virt(page);
488 memset(xdpf, 0, sizeof(*xdpf));
490 addr += sizeof(*xdpf);
491 data_to_copy = metasize ? xdp->data_meta : xdp->data;
492 memcpy(addr, data_to_copy, totsize);
494 xdpf->data = addr + metasize;
495 xdpf->len = totsize - metasize;
497 xdpf->metasize = metasize;
498 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
500 xdp_return_buff(xdp);
503 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);