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 id = xdp_rxq->mem.id;
117
118 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
119 WARN(1, "Missing register, driver bug");
120 return;
121 }
122
123 if (id == 0)
124 return;
125
126 if (xdp_rxq->mem.type == MEM_TYPE_PAGE_POOL) {
127 rcu_read_lock();
128 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
129 page_pool_destroy(xa->page_pool);
130 rcu_read_unlock();
131 }
132 }
133 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
134
xdp_rxq_info_unreg(struct xdp_rxq_info * xdp_rxq)135 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
136 {
137 /* Simplify driver cleanup code paths, allow unreg "unused" */
138 if (xdp_rxq->reg_state == REG_STATE_UNUSED)
139 return;
140
141 WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
142
143 xdp_rxq_info_unreg_mem_model(xdp_rxq);
144
145 xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
146 xdp_rxq->dev = NULL;
147
148 /* Reset mem info to defaults */
149 xdp_rxq->mem.id = 0;
150 xdp_rxq->mem.type = 0;
151 }
152 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
153
xdp_rxq_info_init(struct xdp_rxq_info * xdp_rxq)154 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
155 {
156 memset(xdp_rxq, 0, sizeof(*xdp_rxq));
157 }
158
159 /* Returns 0 on success, negative on failure */
xdp_rxq_info_reg(struct xdp_rxq_info * xdp_rxq,struct net_device * dev,u32 queue_index)160 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
161 struct net_device *dev, u32 queue_index)
162 {
163 if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
164 WARN(1, "Driver promised not to register this");
165 return -EINVAL;
166 }
167
168 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
169 WARN(1, "Missing unregister, handled but fix driver");
170 xdp_rxq_info_unreg(xdp_rxq);
171 }
172
173 if (!dev) {
174 WARN(1, "Missing net_device from driver");
175 return -ENODEV;
176 }
177
178 /* State either UNREGISTERED or NEW */
179 xdp_rxq_info_init(xdp_rxq);
180 xdp_rxq->dev = dev;
181 xdp_rxq->queue_index = queue_index;
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
xdp_return_buff(struct xdp_buff * xdp)387 void xdp_return_buff(struct xdp_buff *xdp)
388 {
389 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
390 }
391
392 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
__xdp_release_frame(void * data,struct xdp_mem_info * mem)393 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
394 {
395 struct xdp_mem_allocator *xa;
396 struct page *page;
397
398 rcu_read_lock();
399 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
400 page = virt_to_head_page(data);
401 if (xa)
402 page_pool_release_page(xa->page_pool, page);
403 rcu_read_unlock();
404 }
405 EXPORT_SYMBOL_GPL(__xdp_release_frame);
406
xdp_attachment_setup(struct xdp_attachment_info * info,struct netdev_bpf * bpf)407 void xdp_attachment_setup(struct xdp_attachment_info *info,
408 struct netdev_bpf *bpf)
409 {
410 if (info->prog)
411 bpf_prog_put(info->prog);
412 info->prog = bpf->prog;
413 info->flags = bpf->flags;
414 }
415 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
416
xdp_convert_zc_to_xdp_frame(struct xdp_buff * xdp)417 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
418 {
419 unsigned int metasize, totsize;
420 void *addr, *data_to_copy;
421 struct xdp_frame *xdpf;
422 struct page *page;
423
424 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
425 metasize = xdp_data_meta_unsupported(xdp) ? 0 :
426 xdp->data - xdp->data_meta;
427 totsize = xdp->data_end - xdp->data + metasize;
428
429 if (sizeof(*xdpf) + totsize > PAGE_SIZE)
430 return NULL;
431
432 page = dev_alloc_page();
433 if (!page)
434 return NULL;
435
436 addr = page_to_virt(page);
437 xdpf = addr;
438 memset(xdpf, 0, sizeof(*xdpf));
439
440 addr += sizeof(*xdpf);
441 data_to_copy = metasize ? xdp->data_meta : xdp->data;
442 memcpy(addr, data_to_copy, totsize);
443
444 xdpf->data = addr + metasize;
445 xdpf->len = totsize - metasize;
446 xdpf->headroom = 0;
447 xdpf->metasize = metasize;
448 xdpf->frame_sz = PAGE_SIZE;
449 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
450
451 xsk_buff_free(xdp);
452 return xdpf;
453 }
454 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
455
456 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
xdp_warn(const char * msg,const char * func,const int line)457 void xdp_warn(const char *msg, const char *func, const int line)
458 {
459 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
460 };
461 EXPORT_SYMBOL_GPL(xdp_warn);
462