xref: /openbmc/linux/net/core/xdp.c (revision 3e30a927) 
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 <net/page_pool.h>
15 
16 #include <net/xdp.h>
17 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
18 #include <trace/events/xdp.h>
19 
20 #define REG_STATE_NEW		0x0
21 #define REG_STATE_REGISTERED	0x1
22 #define REG_STATE_UNREGISTERED	0x2
23 #define REG_STATE_UNUSED	0x3
24 
25 static DEFINE_IDA(mem_id_pool);
26 static DEFINE_MUTEX(mem_id_lock);
27 #define MEM_ID_MAX 0xFFFE
28 #define MEM_ID_MIN 1
29 static int mem_id_next = MEM_ID_MIN;
30 
31 static bool mem_id_init; /* false */
32 static struct rhashtable *mem_id_ht;
33 
34 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
35 {
36 	const u32 *k = data;
37 	const u32 key = *k;
38 
39 	BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id)
40 		     != sizeof(u32));
41 
42 	/* Use cyclic increasing ID as direct hash key */
43 	return key;
44 }
45 
46 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
47 			  const void *ptr)
48 {
49 	const struct xdp_mem_allocator *xa = ptr;
50 	u32 mem_id = *(u32 *)arg->key;
51 
52 	return xa->mem.id != mem_id;
53 }
54 
55 static const struct rhashtable_params mem_id_rht_params = {
56 	.nelem_hint = 64,
57 	.head_offset = offsetof(struct xdp_mem_allocator, node),
58 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
59 	.key_len = sizeof_field(struct xdp_mem_allocator, mem.id),
60 	.max_size = MEM_ID_MAX,
61 	.min_size = 8,
62 	.automatic_shrinking = true,
63 	.hashfn    = xdp_mem_id_hashfn,
64 	.obj_cmpfn = xdp_mem_id_cmp,
65 };
66 
67 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
68 {
69 	struct xdp_mem_allocator *xa;
70 
71 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
72 
73 	/* Allow this ID to be reused */
74 	ida_simple_remove(&mem_id_pool, xa->mem.id);
75 
76 	kfree(xa);
77 }
78 
79 static void mem_xa_remove(struct xdp_mem_allocator *xa)
80 {
81 	trace_mem_disconnect(xa);
82 
83 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
84 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
85 }
86 
87 static void mem_allocator_disconnect(void *allocator)
88 {
89 	struct xdp_mem_allocator *xa;
90 	struct rhashtable_iter iter;
91 
92 	mutex_lock(&mem_id_lock);
93 
94 	rhashtable_walk_enter(mem_id_ht, &iter);
95 	do {
96 		rhashtable_walk_start(&iter);
97 
98 		while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
99 			if (xa->allocator == allocator)
100 				mem_xa_remove(xa);
101 		}
102 
103 		rhashtable_walk_stop(&iter);
104 
105 	} while (xa == ERR_PTR(-EAGAIN));
106 	rhashtable_walk_exit(&iter);
107 
108 	mutex_unlock(&mem_id_lock);
109 }
110 
111 static void mem_id_disconnect(int id)
112 {
113 	struct xdp_mem_allocator *xa;
114 
115 	mutex_lock(&mem_id_lock);
116 
117 	xa = rhashtable_lookup_fast(mem_id_ht, &id, mem_id_rht_params);
118 	if (!xa) {
119 		mutex_unlock(&mem_id_lock);
120 		WARN(1, "Request remove non-existing id(%d), driver bug?", id);
121 		return;
122 	}
123 
124 	trace_mem_disconnect(xa);
125 
126 	if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
127 		call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
128 
129 	mutex_unlock(&mem_id_lock);
130 }
131 
132 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
133 {
134 	struct xdp_mem_allocator *xa;
135 	int id = xdp_rxq->mem.id;
136 
137 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
138 		WARN(1, "Missing register, driver bug");
139 		return;
140 	}
141 
142 	if (id == 0)
143 		return;
144 
145 	if (xdp_rxq->mem.type == MEM_TYPE_ZERO_COPY)
146 		return mem_id_disconnect(id);
147 
148 	if (xdp_rxq->mem.type == MEM_TYPE_PAGE_POOL) {
149 		rcu_read_lock();
150 		xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
151 		page_pool_destroy(xa->page_pool);
152 		rcu_read_unlock();
153 	}
154 }
155 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
156 
157 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
158 {
159 	/* Simplify driver cleanup code paths, allow unreg "unused" */
160 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
161 		return;
162 
163 	WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
164 
165 	xdp_rxq_info_unreg_mem_model(xdp_rxq);
166 
167 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
168 	xdp_rxq->dev = NULL;
169 
170 	/* Reset mem info to defaults */
171 	xdp_rxq->mem.id = 0;
172 	xdp_rxq->mem.type = 0;
173 }
174 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
175 
176 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
177 {
178 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
179 }
180 
181 /* Returns 0 on success, negative on failure */
182 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
183 		     struct net_device *dev, u32 queue_index)
184 {
185 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
186 		WARN(1, "Driver promised not to register this");
187 		return -EINVAL;
188 	}
189 
190 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
191 		WARN(1, "Missing unregister, handled but fix driver");
192 		xdp_rxq_info_unreg(xdp_rxq);
193 	}
194 
195 	if (!dev) {
196 		WARN(1, "Missing net_device from driver");
197 		return -ENODEV;
198 	}
199 
200 	/* State either UNREGISTERED or NEW */
201 	xdp_rxq_info_init(xdp_rxq);
202 	xdp_rxq->dev = dev;
203 	xdp_rxq->queue_index = queue_index;
204 
205 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
206 	return 0;
207 }
208 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
209 
210 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
211 {
212 	xdp_rxq->reg_state = REG_STATE_UNUSED;
213 }
214 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
215 
216 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
217 {
218 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
219 }
220 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
221 
222 static int __mem_id_init_hash_table(void)
223 {
224 	struct rhashtable *rht;
225 	int ret;
226 
227 	if (unlikely(mem_id_init))
228 		return 0;
229 
230 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
231 	if (!rht)
232 		return -ENOMEM;
233 
234 	ret = rhashtable_init(rht, &mem_id_rht_params);
235 	if (ret < 0) {
236 		kfree(rht);
237 		return ret;
238 	}
239 	mem_id_ht = rht;
240 	smp_mb(); /* mutex lock should provide enough pairing */
241 	mem_id_init = true;
242 
243 	return 0;
244 }
245 
246 /* Allocate a cyclic ID that maps to allocator pointer.
247  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
248  *
249  * Caller must lock mem_id_lock.
250  */
251 static int __mem_id_cyclic_get(gfp_t gfp)
252 {
253 	int retries = 1;
254 	int id;
255 
256 again:
257 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
258 	if (id < 0) {
259 		if (id == -ENOSPC) {
260 			/* Cyclic allocator, reset next id */
261 			if (retries--) {
262 				mem_id_next = MEM_ID_MIN;
263 				goto again;
264 			}
265 		}
266 		return id; /* errno */
267 	}
268 	mem_id_next = id + 1;
269 
270 	return id;
271 }
272 
273 static bool __is_supported_mem_type(enum xdp_mem_type type)
274 {
275 	if (type == MEM_TYPE_PAGE_POOL)
276 		return is_page_pool_compiled_in();
277 
278 	if (type >= MEM_TYPE_MAX)
279 		return false;
280 
281 	return true;
282 }
283 
284 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
285 			       enum xdp_mem_type type, void *allocator)
286 {
287 	struct xdp_mem_allocator *xdp_alloc;
288 	gfp_t gfp = GFP_KERNEL;
289 	int id, errno, ret;
290 	void *ptr;
291 
292 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
293 		WARN(1, "Missing register, driver bug");
294 		return -EFAULT;
295 	}
296 
297 	if (!__is_supported_mem_type(type))
298 		return -EOPNOTSUPP;
299 
300 	xdp_rxq->mem.type = type;
301 
302 	if (!allocator) {
303 		if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
304 			return -EINVAL; /* Setup time check page_pool req */
305 		return 0;
306 	}
307 
308 	/* Delay init of rhashtable to save memory if feature isn't used */
309 	if (!mem_id_init) {
310 		mutex_lock(&mem_id_lock);
311 		ret = __mem_id_init_hash_table();
312 		mutex_unlock(&mem_id_lock);
313 		if (ret < 0) {
314 			WARN_ON(1);
315 			return ret;
316 		}
317 	}
318 
319 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
320 	if (!xdp_alloc)
321 		return -ENOMEM;
322 
323 	mutex_lock(&mem_id_lock);
324 	id = __mem_id_cyclic_get(gfp);
325 	if (id < 0) {
326 		errno = id;
327 		goto err;
328 	}
329 	xdp_rxq->mem.id = id;
330 	xdp_alloc->mem  = xdp_rxq->mem;
331 	xdp_alloc->allocator = allocator;
332 
333 	/* Insert allocator into ID lookup table */
334 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
335 	if (IS_ERR(ptr)) {
336 		ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
337 		xdp_rxq->mem.id = 0;
338 		errno = PTR_ERR(ptr);
339 		goto err;
340 	}
341 
342 	if (type == MEM_TYPE_PAGE_POOL)
343 		page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
344 
345 	mutex_unlock(&mem_id_lock);
346 
347 	trace_mem_connect(xdp_alloc, xdp_rxq);
348 	return 0;
349 err:
350 	mutex_unlock(&mem_id_lock);
351 	kfree(xdp_alloc);
352 	return errno;
353 }
354 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
355 
356 /* XDP RX runs under NAPI protection, and in different delivery error
357  * scenarios (e.g. queue full), it is possible to return the xdp_frame
358  * while still leveraging this protection.  The @napi_direct boolean
359  * is used for those calls sites.  Thus, allowing for faster recycling
360  * of xdp_frames/pages in those cases.
361  */
362 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
363 			 unsigned long handle)
364 {
365 	struct xdp_mem_allocator *xa;
366 	struct page *page;
367 
368 	switch (mem->type) {
369 	case MEM_TYPE_PAGE_POOL:
370 		rcu_read_lock();
371 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
372 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
373 		page = virt_to_head_page(data);
374 		napi_direct &= !xdp_return_frame_no_direct();
375 		page_pool_put_page(xa->page_pool, page, napi_direct);
376 		rcu_read_unlock();
377 		break;
378 	case MEM_TYPE_PAGE_SHARED:
379 		page_frag_free(data);
380 		break;
381 	case MEM_TYPE_PAGE_ORDER0:
382 		page = virt_to_page(data); /* Assumes order0 page*/
383 		put_page(page);
384 		break;
385 	case MEM_TYPE_ZERO_COPY:
386 		/* NB! Only valid from an xdp_buff! */
387 		rcu_read_lock();
388 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
389 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
390 		xa->zc_alloc->free(xa->zc_alloc, handle);
391 		rcu_read_unlock();
392 	default:
393 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
394 		break;
395 	}
396 }
397 
398 void xdp_return_frame(struct xdp_frame *xdpf)
399 {
400 	__xdp_return(xdpf->data, &xdpf->mem, false, 0);
401 }
402 EXPORT_SYMBOL_GPL(xdp_return_frame);
403 
404 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
405 {
406 	__xdp_return(xdpf->data, &xdpf->mem, true, 0);
407 }
408 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
409 
410 void xdp_return_buff(struct xdp_buff *xdp)
411 {
412 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
413 }
414 EXPORT_SYMBOL_GPL(xdp_return_buff);
415 
416 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
417 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
418 {
419 	struct xdp_mem_allocator *xa;
420 	struct page *page;
421 
422 	rcu_read_lock();
423 	xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
424 	page = virt_to_head_page(data);
425 	if (xa)
426 		page_pool_release_page(xa->page_pool, page);
427 	rcu_read_unlock();
428 }
429 EXPORT_SYMBOL_GPL(__xdp_release_frame);
430 
431 int xdp_attachment_query(struct xdp_attachment_info *info,
432 			 struct netdev_bpf *bpf)
433 {
434 	bpf->prog_id = info->prog ? info->prog->aux->id : 0;
435 	bpf->prog_flags = info->prog ? info->flags : 0;
436 	return 0;
437 }
438 EXPORT_SYMBOL_GPL(xdp_attachment_query);
439 
440 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
441 			     struct netdev_bpf *bpf)
442 {
443 	if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
444 		NL_SET_ERR_MSG(bpf->extack,
445 			       "program loaded with different flags");
446 		return false;
447 	}
448 	return true;
449 }
450 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
451 
452 void xdp_attachment_setup(struct xdp_attachment_info *info,
453 			  struct netdev_bpf *bpf)
454 {
455 	if (info->prog)
456 		bpf_prog_put(info->prog);
457 	info->prog = bpf->prog;
458 	info->flags = bpf->flags;
459 }
460 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
461 
462 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
463 {
464 	unsigned int metasize, totsize;
465 	void *addr, *data_to_copy;
466 	struct xdp_frame *xdpf;
467 	struct page *page;
468 
469 	/* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
470 	metasize = xdp_data_meta_unsupported(xdp) ? 0 :
471 		   xdp->data - xdp->data_meta;
472 	totsize = xdp->data_end - xdp->data + metasize;
473 
474 	if (sizeof(*xdpf) + totsize > PAGE_SIZE)
475 		return NULL;
476 
477 	page = dev_alloc_page();
478 	if (!page)
479 		return NULL;
480 
481 	addr = page_to_virt(page);
482 	xdpf = addr;
483 	memset(xdpf, 0, sizeof(*xdpf));
484 
485 	addr += sizeof(*xdpf);
486 	data_to_copy = metasize ? xdp->data_meta : xdp->data;
487 	memcpy(addr, data_to_copy, totsize);
488 
489 	xdpf->data = addr + metasize;
490 	xdpf->len = totsize - metasize;
491 	xdpf->headroom = 0;
492 	xdpf->metasize = metasize;
493 	xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
494 
495 	xdp_return_buff(xdp);
496 	return xdpf;
497 }
498 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
499