xref: /linux/net/core/xdp.c (revision 05ee19c18c2bb3dea69e29219017367c4a77e65a) 
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 
21 #define REG_STATE_NEW		0x0
22 #define REG_STATE_REGISTERED	0x1
23 #define REG_STATE_UNREGISTERED	0x2
24 #define REG_STATE_UNUSED	0x3
25 
26 static DEFINE_IDA(mem_id_pool);
27 static DEFINE_MUTEX(mem_id_lock);
28 #define MEM_ID_MAX 0xFFFE
29 #define MEM_ID_MIN 1
30 static int mem_id_next = MEM_ID_MIN;
31 
32 static bool mem_id_init; /* false */
33 static struct rhashtable *mem_id_ht;
34 
35 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
36 {
37 	const u32 *k = data;
38 	const u32 key = *k;
39 
40 	BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
41 		     != sizeof(u32));
42 
43 	/* Use cyclic increasing ID as direct hash key */
44 	return key;
45 }
46 
47 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
48 			  const void *ptr)
49 {
50 	const struct xdp_mem_allocator *xa = ptr;
51 	u32 mem_id = *(u32 *)arg->key;
52 
53 	return xa->mem.id != mem_id;
54 }
55 
56 static const struct rhashtable_params mem_id_rht_params = {
57 	.nelem_hint = 64,
58 	.head_offset = offsetof(struct xdp_mem_allocator, node),
59 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
60 	.key_len = sizeof_field(struct xdp_mem_allocator, mem.id),
61 	.max_size = MEM_ID_MAX,
62 	.min_size = 8,
63 	.automatic_shrinking = true,
64 	.hashfn    = xdp_mem_id_hashfn,
65 	.obj_cmpfn = xdp_mem_id_cmp,
66 };
67 
68 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
69 {
70 	struct xdp_mem_allocator *xa;
71 
72 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
73 
74 	/* Allow this ID to be reused */
75 	ida_simple_remove(&mem_id_pool, xa->mem.id);
76 
77 	kfree(xa);
78 }
79 
80 static void mem_xa_remove(struct xdp_mem_allocator *xa)
81 {
82 	trace_mem_disconnect(xa);
83 
84 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
85 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
86 }
87 
88 static void mem_allocator_disconnect(void *allocator)
89 {
90 	struct xdp_mem_allocator *xa;
91 	struct rhashtable_iter iter;
92 
93 	mutex_lock(&mem_id_lock);
94 
95 	rhashtable_walk_enter(mem_id_ht, &iter);
96 	do {
97 		rhashtable_walk_start(&iter);
98 
99 		while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
100 			if (xa->allocator == allocator)
101 				mem_xa_remove(xa);
102 		}
103 
104 		rhashtable_walk_stop(&iter);
105 
106 	} while (xa == ERR_PTR(-EAGAIN));
107 	rhashtable_walk_exit(&iter);
108 
109 	mutex_unlock(&mem_id_lock);
110 }
111 
112 static void mem_id_disconnect(int id)
113 {
114 	struct xdp_mem_allocator *xa;
115 
116 	mutex_lock(&mem_id_lock);
117 
118 	xa = rhashtable_lookup_fast(mem_id_ht, &id, mem_id_rht_params);
119 	if (!xa) {
120 		mutex_unlock(&mem_id_lock);
121 		WARN(1, "Request remove non-existing id(%d), driver bug?", id);
122 		return;
123 	}
124 
125 	trace_mem_disconnect(xa);
126 
127 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
128 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
129 
130 	mutex_unlock(&mem_id_lock);
131 }
132 
133 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
134 {
135 	struct xdp_mem_allocator *xa;
136 	int id = xdp_rxq->mem.id;
137 
138 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
139 		WARN(1, "Missing register, driver bug");
140 		return;
141 	}
142 
143 	if (id == 0)
144 		return;
145 
146 	if (xdp_rxq->mem.type == MEM_TYPE_ZERO_COPY)
147 		return mem_id_disconnect(id);
148 
149 	if (xdp_rxq->mem.type == MEM_TYPE_PAGE_POOL) {
150 		rcu_read_lock();
151 		xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
152 		page_pool_destroy(xa->page_pool);
153 		rcu_read_unlock();
154 	}
155 }
156 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
157 
158 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
159 {
160 	/* Simplify driver cleanup code paths, allow unreg "unused" */
161 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
162 		return;
163 
164 	WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
165 
166 	xdp_rxq_info_unreg_mem_model(xdp_rxq);
167 
168 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
169 	xdp_rxq->dev = NULL;
170 
171 	/* Reset mem info to defaults */
172 	xdp_rxq->mem.id = 0;
173 	xdp_rxq->mem.type = 0;
174 }
175 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
176 
177 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
178 {
179 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
180 }
181 
182 /* Returns 0 on success, negative on failure */
183 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
184 		     struct net_device *dev, u32 queue_index)
185 {
186 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
187 		WARN(1, "Driver promised not to register this");
188 		return -EINVAL;
189 	}
190 
191 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
192 		WARN(1, "Missing unregister, handled but fix driver");
193 		xdp_rxq_info_unreg(xdp_rxq);
194 	}
195 
196 	if (!dev) {
197 		WARN(1, "Missing net_device from driver");
198 		return -ENODEV;
199 	}
200 
201 	/* State either UNREGISTERED or NEW */
202 	xdp_rxq_info_init(xdp_rxq);
203 	xdp_rxq->dev = dev;
204 	xdp_rxq->queue_index = queue_index;
205 
206 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
207 	return 0;
208 }
209 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
210 
211 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
212 {
213 	xdp_rxq->reg_state = REG_STATE_UNUSED;
214 }
215 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
216 
217 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
218 {
219 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
220 }
221 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
222 
223 static int __mem_id_init_hash_table(void)
224 {
225 	struct rhashtable *rht;
226 	int ret;
227 
228 	if (unlikely(mem_id_init))
229 		return 0;
230 
231 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
232 	if (!rht)
233 		return -ENOMEM;
234 
235 	ret = rhashtable_init(rht, &mem_id_rht_params);
236 	if (ret < 0) {
237 		kfree(rht);
238 		return ret;
239 	}
240 	mem_id_ht = rht;
241 	smp_mb(); /* mutex lock should provide enough pairing */
242 	mem_id_init = true;
243 
244 	return 0;
245 }
246 
247 /* Allocate a cyclic ID that maps to allocator pointer.
248  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
249  *
250  * Caller must lock mem_id_lock.
251  */
252 static int __mem_id_cyclic_get(gfp_t gfp)
253 {
254 	int retries = 1;
255 	int id;
256 
257 again:
258 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
259 	if (id < 0) {
260 		if (id == -ENOSPC) {
261 			/* Cyclic allocator, reset next id */
262 			if (retries--) {
263 				mem_id_next = MEM_ID_MIN;
264 				goto again;
265 			}
266 		}
267 		return id; /* errno */
268 	}
269 	mem_id_next = id + 1;
270 
271 	return id;
272 }
273 
274 static bool __is_supported_mem_type(enum xdp_mem_type type)
275 {
276 	if (type == MEM_TYPE_PAGE_POOL)
277 		return is_page_pool_compiled_in();
278 
279 	if (type >= MEM_TYPE_MAX)
280 		return false;
281 
282 	return true;
283 }
284 
285 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
286 			       enum xdp_mem_type type, void *allocator)
287 {
288 	struct xdp_mem_allocator *xdp_alloc;
289 	gfp_t gfp = GFP_KERNEL;
290 	int id, errno, ret;
291 	void *ptr;
292 
293 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
294 		WARN(1, "Missing register, driver bug");
295 		return -EFAULT;
296 	}
297 
298 	if (!__is_supported_mem_type(type))
299 		return -EOPNOTSUPP;
300 
301 	xdp_rxq->mem.type = type;
302 
303 	if (!allocator) {
304 		if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
305 			return -EINVAL; /* Setup time check page_pool req */
306 		return 0;
307 	}
308 
309 	/* Delay init of rhashtable to save memory if feature isn't used */
310 	if (!mem_id_init) {
311 		mutex_lock(&mem_id_lock);
312 		ret = __mem_id_init_hash_table();
313 		mutex_unlock(&mem_id_lock);
314 		if (ret < 0) {
315 			WARN_ON(1);
316 			return ret;
317 		}
318 	}
319 
320 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
321 	if (!xdp_alloc)
322 		return -ENOMEM;
323 
324 	mutex_lock(&mem_id_lock);
325 	id = __mem_id_cyclic_get(gfp);
326 	if (id < 0) {
327 		errno = id;
328 		goto err;
329 	}
330 	xdp_rxq->mem.id = id;
331 	xdp_alloc->mem  = xdp_rxq->mem;
332 	xdp_alloc->allocator = allocator;
333 
334 	/* Insert allocator into ID lookup table */
335 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
336 	if (IS_ERR(ptr)) {
337 		ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
338 		xdp_rxq->mem.id = 0;
339 		errno = PTR_ERR(ptr);
340 		goto err;
341 	}
342 
343 	if (type == MEM_TYPE_PAGE_POOL)
344 		page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
345 
346 	mutex_unlock(&mem_id_lock);
347 
348 	trace_mem_connect(xdp_alloc, xdp_rxq);
349 	return 0;
350 err:
351 	mutex_unlock(&mem_id_lock);
352 	kfree(xdp_alloc);
353 	return errno;
354 }
355 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
356 
357 /* XDP RX runs under NAPI protection, and in different delivery error
358  * scenarios (e.g. queue full), it is possible to return the xdp_frame
359  * while still leveraging this protection.  The @napi_direct boolean
360  * is used for those calls sites.  Thus, allowing for faster recycling
361  * of xdp_frames/pages in those cases.
362  */
363 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
364 			 unsigned long handle)
365 {
366 	struct xdp_mem_allocator *xa;
367 	struct page *page;
368 
369 	switch (mem->type) {
370 	case MEM_TYPE_PAGE_POOL:
371 		rcu_read_lock();
372 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
373 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
374 		page = virt_to_head_page(data);
375 		napi_direct &= !xdp_return_frame_no_direct();
376 		page_pool_put_full_page(xa->page_pool, page, napi_direct);
377 		rcu_read_unlock();
378 		break;
379 	case MEM_TYPE_PAGE_SHARED:
380 		page_frag_free(data);
381 		break;
382 	case MEM_TYPE_PAGE_ORDER0:
383 		page = virt_to_page(data); /* Assumes order0 page*/
384 		put_page(page);
385 		break;
386 	case MEM_TYPE_ZERO_COPY:
387 		/* NB! Only valid from an xdp_buff! */
388 		rcu_read_lock();
389 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
390 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
391 		xa->zc_alloc->free(xa->zc_alloc, handle);
392 		rcu_read_unlock();
393 	default:
394 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
395 		break;
396 	}
397 }
398 
399 void xdp_return_frame(struct xdp_frame *xdpf)
400 {
401 	__xdp_return(xdpf->data, &xdpf->mem, false, 0);
402 }
403 EXPORT_SYMBOL_GPL(xdp_return_frame);
404 
405 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
406 {
407 	__xdp_return(xdpf->data, &xdpf->mem, true, 0);
408 }
409 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
410 
411 void xdp_return_buff(struct xdp_buff *xdp)
412 {
413 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
414 }
415 EXPORT_SYMBOL_GPL(xdp_return_buff);
416 
417 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
418 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
419 {
420 	struct xdp_mem_allocator *xa;
421 	struct page *page;
422 
423 	rcu_read_lock();
424 	xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
425 	page = virt_to_head_page(data);
426 	if (xa)
427 		page_pool_release_page(xa->page_pool, page);
428 	rcu_read_unlock();
429 }
430 EXPORT_SYMBOL_GPL(__xdp_release_frame);
431 
432 int xdp_attachment_query(struct xdp_attachment_info *info,
433 			 struct netdev_bpf *bpf)
434 {
435 	bpf->prog_id = info->prog ? info->prog->aux->id : 0;
436 	bpf->prog_flags = info->prog ? info->flags : 0;
437 	return 0;
438 }
439 EXPORT_SYMBOL_GPL(xdp_attachment_query);
440 
441 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
442 			     struct netdev_bpf *bpf)
443 {
444 	if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
445 		NL_SET_ERR_MSG(bpf->extack,
446 			       "program loaded with different flags");
447 		return false;
448 	}
449 	return true;
450 }
451 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
452 
453 void xdp_attachment_setup(struct xdp_attachment_info *info,
454 			  struct netdev_bpf *bpf)
455 {
456 	if (info->prog)
457 		bpf_prog_put(info->prog);
458 	info->prog = bpf->prog;
459 	info->flags = bpf->flags;
460 }
461 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
462 
463 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
464 {
465 	unsigned int metasize, totsize;
466 	void *addr, *data_to_copy;
467 	struct xdp_frame *xdpf;
468 	struct page *page;
469 
470 	/* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
471 	metasize = xdp_data_meta_unsupported(xdp) ? 0 :
472 		   xdp->data - xdp->data_meta;
473 	totsize = xdp->data_end - xdp->data + metasize;
474 
475 	if (sizeof(*xdpf) + totsize > PAGE_SIZE)
476 		return NULL;
477 
478 	page = dev_alloc_page();
479 	if (!page)
480 		return NULL;
481 
482 	addr = page_to_virt(page);
483 	xdpf = addr;
484 	memset(xdpf, 0, sizeof(*xdpf));
485 
486 	addr += sizeof(*xdpf);
487 	data_to_copy = metasize ? xdp->data_meta : xdp->data;
488 	memcpy(addr, data_to_copy, totsize);
489 
490 	xdpf->data = addr + metasize;
491 	xdpf->len = totsize - metasize;
492 	xdpf->headroom = 0;
493 	xdpf->metasize = metasize;
494 	xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
495 
496 	xdp_return_buff(xdp);
497 	return xdpf;
498 }
499 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
500 
501 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
502 void xdp_warn(const char *msg, const char *func, const int line)
503 {
504 	WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
505 };
506 EXPORT_SYMBOL_GPL(xdp_warn);
507