xref: /openbmc/linux/net/core/xdp.c (revision a0ae2562c6c4b2721d9fddba63b7286c13517d9f)
1 /* net/core/xdp.c
2  *
3  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
4  * Released under terms in GPL version 2.  See COPYING.
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 
18 #define REG_STATE_NEW		0x0
19 #define REG_STATE_REGISTERED	0x1
20 #define REG_STATE_UNREGISTERED	0x2
21 #define REG_STATE_UNUSED	0x3
22 
23 static DEFINE_IDA(mem_id_pool);
24 static DEFINE_MUTEX(mem_id_lock);
25 #define MEM_ID_MAX 0xFFFE
26 #define MEM_ID_MIN 1
27 static int mem_id_next = MEM_ID_MIN;
28 
29 static bool mem_id_init; /* false */
30 static struct rhashtable *mem_id_ht;
31 
32 struct xdp_mem_allocator {
33 	struct xdp_mem_info mem;
34 	union {
35 		void *allocator;
36 		struct page_pool *page_pool;
37 		struct zero_copy_allocator *zc_alloc;
38 	};
39 	struct rhash_head node;
40 	struct rcu_head rcu;
41 };
42 
43 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
44 {
45 	const u32 *k = data;
46 	const u32 key = *k;
47 
48 	BUILD_BUG_ON(FIELD_SIZEOF(struct xdp_mem_allocator, mem.id)
49 		     != sizeof(u32));
50 
51 	/* Use cyclic increasing ID as direct hash key */
52 	return key;
53 }
54 
55 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
56 			  const void *ptr)
57 {
58 	const struct xdp_mem_allocator *xa = ptr;
59 	u32 mem_id = *(u32 *)arg->key;
60 
61 	return xa->mem.id != mem_id;
62 }
63 
64 static const struct rhashtable_params mem_id_rht_params = {
65 	.nelem_hint = 64,
66 	.head_offset = offsetof(struct xdp_mem_allocator, node),
67 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
68 	.key_len = FIELD_SIZEOF(struct xdp_mem_allocator, mem.id),
69 	.max_size = MEM_ID_MAX,
70 	.min_size = 8,
71 	.automatic_shrinking = true,
72 	.hashfn    = xdp_mem_id_hashfn,
73 	.obj_cmpfn = xdp_mem_id_cmp,
74 };
75 
76 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
77 {
78 	struct xdp_mem_allocator *xa;
79 
80 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
81 
82 	/* Allow this ID to be reused */
83 	ida_simple_remove(&mem_id_pool, xa->mem.id);
84 
85 	/* Notice, driver is expected to free the *allocator,
86 	 * e.g. page_pool, and MUST also use RCU free.
87 	 */
88 
89 	/* Poison memory */
90 	xa->mem.id = 0xFFFF;
91 	xa->mem.type = 0xF0F0;
92 	xa->allocator = (void *)0xDEAD9001;
93 
94 	kfree(xa);
95 }
96 
97 static void __xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
98 {
99 	struct xdp_mem_allocator *xa;
100 	int id = xdp_rxq->mem.id;
101 	int err;
102 
103 	if (id == 0)
104 		return;
105 
106 	mutex_lock(&mem_id_lock);
107 
108 	xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
109 	if (!xa) {
110 		mutex_unlock(&mem_id_lock);
111 		return;
112 	}
113 
114 	err = rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params);
115 	WARN_ON(err);
116 
117 	call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
118 
119 	mutex_unlock(&mem_id_lock);
120 }
121 
122 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
123 {
124 	/* Simplify driver cleanup code paths, allow unreg "unused" */
125 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
126 		return;
127 
128 	WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
129 
130 	__xdp_rxq_info_unreg_mem_model(xdp_rxq);
131 
132 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
133 	xdp_rxq->dev = NULL;
134 
135 	/* Reset mem info to defaults */
136 	xdp_rxq->mem.id = 0;
137 	xdp_rxq->mem.type = 0;
138 }
139 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
140 
141 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
142 {
143 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
144 }
145 
146 /* Returns 0 on success, negative on failure */
147 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
148 		     struct net_device *dev, u32 queue_index)
149 {
150 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
151 		WARN(1, "Driver promised not to register this");
152 		return -EINVAL;
153 	}
154 
155 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
156 		WARN(1, "Missing unregister, handled but fix driver");
157 		xdp_rxq_info_unreg(xdp_rxq);
158 	}
159 
160 	if (!dev) {
161 		WARN(1, "Missing net_device from driver");
162 		return -ENODEV;
163 	}
164 
165 	/* State either UNREGISTERED or NEW */
166 	xdp_rxq_info_init(xdp_rxq);
167 	xdp_rxq->dev = dev;
168 	xdp_rxq->queue_index = queue_index;
169 
170 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
171 	return 0;
172 }
173 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
174 
175 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
176 {
177 	xdp_rxq->reg_state = REG_STATE_UNUSED;
178 }
179 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
180 
181 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
182 {
183 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
184 }
185 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
186 
187 static int __mem_id_init_hash_table(void)
188 {
189 	struct rhashtable *rht;
190 	int ret;
191 
192 	if (unlikely(mem_id_init))
193 		return 0;
194 
195 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
196 	if (!rht)
197 		return -ENOMEM;
198 
199 	ret = rhashtable_init(rht, &mem_id_rht_params);
200 	if (ret < 0) {
201 		kfree(rht);
202 		return ret;
203 	}
204 	mem_id_ht = rht;
205 	smp_mb(); /* mutex lock should provide enough pairing */
206 	mem_id_init = true;
207 
208 	return 0;
209 }
210 
211 /* Allocate a cyclic ID that maps to allocator pointer.
212  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
213  *
214  * Caller must lock mem_id_lock.
215  */
216 static int __mem_id_cyclic_get(gfp_t gfp)
217 {
218 	int retries = 1;
219 	int id;
220 
221 again:
222 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
223 	if (id < 0) {
224 		if (id == -ENOSPC) {
225 			/* Cyclic allocator, reset next id */
226 			if (retries--) {
227 				mem_id_next = MEM_ID_MIN;
228 				goto again;
229 			}
230 		}
231 		return id; /* errno */
232 	}
233 	mem_id_next = id + 1;
234 
235 	return id;
236 }
237 
238 static bool __is_supported_mem_type(enum xdp_mem_type type)
239 {
240 	if (type == MEM_TYPE_PAGE_POOL)
241 		return is_page_pool_compiled_in();
242 
243 	if (type >= MEM_TYPE_MAX)
244 		return false;
245 
246 	return true;
247 }
248 
249 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
250 			       enum xdp_mem_type type, void *allocator)
251 {
252 	struct xdp_mem_allocator *xdp_alloc;
253 	gfp_t gfp = GFP_KERNEL;
254 	int id, errno, ret;
255 	void *ptr;
256 
257 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
258 		WARN(1, "Missing register, driver bug");
259 		return -EFAULT;
260 	}
261 
262 	if (!__is_supported_mem_type(type))
263 		return -EOPNOTSUPP;
264 
265 	xdp_rxq->mem.type = type;
266 
267 	if (!allocator) {
268 		if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
269 			return -EINVAL; /* Setup time check page_pool req */
270 		return 0;
271 	}
272 
273 	/* Delay init of rhashtable to save memory if feature isn't used */
274 	if (!mem_id_init) {
275 		mutex_lock(&mem_id_lock);
276 		ret = __mem_id_init_hash_table();
277 		mutex_unlock(&mem_id_lock);
278 		if (ret < 0) {
279 			WARN_ON(1);
280 			return ret;
281 		}
282 	}
283 
284 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
285 	if (!xdp_alloc)
286 		return -ENOMEM;
287 
288 	mutex_lock(&mem_id_lock);
289 	id = __mem_id_cyclic_get(gfp);
290 	if (id < 0) {
291 		errno = id;
292 		goto err;
293 	}
294 	xdp_rxq->mem.id = id;
295 	xdp_alloc->mem  = xdp_rxq->mem;
296 	xdp_alloc->allocator = allocator;
297 
298 	/* Insert allocator into ID lookup table */
299 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
300 	if (IS_ERR(ptr)) {
301 		errno = PTR_ERR(ptr);
302 		goto err;
303 	}
304 
305 	mutex_unlock(&mem_id_lock);
306 
307 	return 0;
308 err:
309 	mutex_unlock(&mem_id_lock);
310 	kfree(xdp_alloc);
311 	return errno;
312 }
313 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
314 
315 /* XDP RX runs under NAPI protection, and in different delivery error
316  * scenarios (e.g. queue full), it is possible to return the xdp_frame
317  * while still leveraging this protection.  The @napi_direct boolian
318  * is used for those calls sites.  Thus, allowing for faster recycling
319  * of xdp_frames/pages in those cases.
320  */
321 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
322 			 unsigned long handle)
323 {
324 	struct xdp_mem_allocator *xa;
325 	struct page *page;
326 
327 	switch (mem->type) {
328 	case MEM_TYPE_PAGE_POOL:
329 		rcu_read_lock();
330 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
331 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
332 		page = virt_to_head_page(data);
333 		if (xa)
334 			page_pool_put_page(xa->page_pool, page, napi_direct);
335 		else
336 			put_page(page);
337 		rcu_read_unlock();
338 		break;
339 	case MEM_TYPE_PAGE_SHARED:
340 		page_frag_free(data);
341 		break;
342 	case MEM_TYPE_PAGE_ORDER0:
343 		page = virt_to_page(data); /* Assumes order0 page*/
344 		put_page(page);
345 		break;
346 	case MEM_TYPE_ZERO_COPY:
347 		/* NB! Only valid from an xdp_buff! */
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 		xa->zc_alloc->free(xa->zc_alloc, handle);
352 		rcu_read_unlock();
353 	default:
354 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
355 		break;
356 	}
357 }
358 
359 void xdp_return_frame(struct xdp_frame *xdpf)
360 {
361 	__xdp_return(xdpf->data, &xdpf->mem, false, 0);
362 }
363 EXPORT_SYMBOL_GPL(xdp_return_frame);
364 
365 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
366 {
367 	__xdp_return(xdpf->data, &xdpf->mem, true, 0);
368 }
369 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
370 
371 void xdp_return_buff(struct xdp_buff *xdp)
372 {
373 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
374 }
375 EXPORT_SYMBOL_GPL(xdp_return_buff);
376 
377 int xdp_attachment_query(struct xdp_attachment_info *info,
378 			 struct netdev_bpf *bpf)
379 {
380 	bpf->prog_id = info->prog ? info->prog->aux->id : 0;
381 	bpf->prog_flags = info->prog ? info->flags : 0;
382 	return 0;
383 }
384 EXPORT_SYMBOL_GPL(xdp_attachment_query);
385 
386 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
387 			     struct netdev_bpf *bpf)
388 {
389 	if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
390 		NL_SET_ERR_MSG(bpf->extack,
391 			       "program loaded with different flags");
392 		return false;
393 	}
394 	return true;
395 }
396 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
397 
398 void xdp_attachment_setup(struct xdp_attachment_info *info,
399 			  struct netdev_bpf *bpf)
400 {
401 	if (info->prog)
402 		bpf_prog_put(info->prog);
403 	info->prog = bpf->prog;
404 	info->flags = bpf->flags;
405 }
406 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
407