xref: /openbmc/linux/net/core/xdp.c (revision b9b77222)
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/types.h>
7 #include <linux/mm.h>
8 #include <linux/slab.h>
9 #include <linux/idr.h>
10 #include <linux/rhashtable.h>
11 #include <net/page_pool.h>
12 
13 #include <net/xdp.h>
14 
15 #define REG_STATE_NEW		0x0
16 #define REG_STATE_REGISTERED	0x1
17 #define REG_STATE_UNREGISTERED	0x2
18 #define REG_STATE_UNUSED	0x3
19 
20 static DEFINE_IDA(mem_id_pool);
21 static DEFINE_MUTEX(mem_id_lock);
22 #define MEM_ID_MAX 0xFFFE
23 #define MEM_ID_MIN 1
24 static int mem_id_next = MEM_ID_MIN;
25 
26 static bool mem_id_init; /* false */
27 static struct rhashtable *mem_id_ht;
28 
29 struct xdp_mem_allocator {
30 	struct xdp_mem_info mem;
31 	union {
32 		void *allocator;
33 		struct page_pool *page_pool;
34 		struct zero_copy_allocator *zc_alloc;
35 	};
36 	struct rhash_head node;
37 	struct rcu_head rcu;
38 };
39 
40 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
41 {
42 	const u32 *k = data;
43 	const u32 key = *k;
44 
45 	BUILD_BUG_ON(FIELD_SIZEOF(struct xdp_mem_allocator, mem.id)
46 		     != sizeof(u32));
47 
48 	/* Use cyclic increasing ID as direct hash key, see rht_bucket_index */
49 	return key << RHT_HASH_RESERVED_SPACE;
50 }
51 
52 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
53 			  const void *ptr)
54 {
55 	const struct xdp_mem_allocator *xa = ptr;
56 	u32 mem_id = *(u32 *)arg->key;
57 
58 	return xa->mem.id != mem_id;
59 }
60 
61 static const struct rhashtable_params mem_id_rht_params = {
62 	.nelem_hint = 64,
63 	.head_offset = offsetof(struct xdp_mem_allocator, node),
64 	.key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
65 	.key_len = FIELD_SIZEOF(struct xdp_mem_allocator, mem.id),
66 	.max_size = MEM_ID_MAX,
67 	.min_size = 8,
68 	.automatic_shrinking = true,
69 	.hashfn    = xdp_mem_id_hashfn,
70 	.obj_cmpfn = xdp_mem_id_cmp,
71 };
72 
73 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
74 {
75 	struct xdp_mem_allocator *xa;
76 
77 	xa = container_of(rcu, struct xdp_mem_allocator, rcu);
78 
79 	/* Allow this ID to be reused */
80 	ida_simple_remove(&mem_id_pool, xa->mem.id);
81 
82 	/* Notice, driver is expected to free the *allocator,
83 	 * e.g. page_pool, and MUST also use RCU free.
84 	 */
85 
86 	/* Poison memory */
87 	xa->mem.id = 0xFFFF;
88 	xa->mem.type = 0xF0F0;
89 	xa->allocator = (void *)0xDEAD9001;
90 
91 	kfree(xa);
92 }
93 
94 static void __xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
95 {
96 	struct xdp_mem_allocator *xa;
97 	int id = xdp_rxq->mem.id;
98 	int err;
99 
100 	if (id == 0)
101 		return;
102 
103 	mutex_lock(&mem_id_lock);
104 
105 	xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
106 	if (!xa) {
107 		mutex_unlock(&mem_id_lock);
108 		return;
109 	}
110 
111 	err = rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params);
112 	WARN_ON(err);
113 
114 	call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
115 
116 	mutex_unlock(&mem_id_lock);
117 }
118 
119 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
120 {
121 	/* Simplify driver cleanup code paths, allow unreg "unused" */
122 	if (xdp_rxq->reg_state == REG_STATE_UNUSED)
123 		return;
124 
125 	WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
126 
127 	__xdp_rxq_info_unreg_mem_model(xdp_rxq);
128 
129 	xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
130 	xdp_rxq->dev = NULL;
131 
132 	/* Reset mem info to defaults */
133 	xdp_rxq->mem.id = 0;
134 	xdp_rxq->mem.type = 0;
135 }
136 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
137 
138 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
139 {
140 	memset(xdp_rxq, 0, sizeof(*xdp_rxq));
141 }
142 
143 /* Returns 0 on success, negative on failure */
144 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
145 		     struct net_device *dev, u32 queue_index)
146 {
147 	if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
148 		WARN(1, "Driver promised not to register this");
149 		return -EINVAL;
150 	}
151 
152 	if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
153 		WARN(1, "Missing unregister, handled but fix driver");
154 		xdp_rxq_info_unreg(xdp_rxq);
155 	}
156 
157 	if (!dev) {
158 		WARN(1, "Missing net_device from driver");
159 		return -ENODEV;
160 	}
161 
162 	/* State either UNREGISTERED or NEW */
163 	xdp_rxq_info_init(xdp_rxq);
164 	xdp_rxq->dev = dev;
165 	xdp_rxq->queue_index = queue_index;
166 
167 	xdp_rxq->reg_state = REG_STATE_REGISTERED;
168 	return 0;
169 }
170 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
171 
172 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
173 {
174 	xdp_rxq->reg_state = REG_STATE_UNUSED;
175 }
176 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
177 
178 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
179 {
180 	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
181 }
182 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
183 
184 static int __mem_id_init_hash_table(void)
185 {
186 	struct rhashtable *rht;
187 	int ret;
188 
189 	if (unlikely(mem_id_init))
190 		return 0;
191 
192 	rht = kzalloc(sizeof(*rht), GFP_KERNEL);
193 	if (!rht)
194 		return -ENOMEM;
195 
196 	ret = rhashtable_init(rht, &mem_id_rht_params);
197 	if (ret < 0) {
198 		kfree(rht);
199 		return ret;
200 	}
201 	mem_id_ht = rht;
202 	smp_mb(); /* mutex lock should provide enough pairing */
203 	mem_id_init = true;
204 
205 	return 0;
206 }
207 
208 /* Allocate a cyclic ID that maps to allocator pointer.
209  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
210  *
211  * Caller must lock mem_id_lock.
212  */
213 static int __mem_id_cyclic_get(gfp_t gfp)
214 {
215 	int retries = 1;
216 	int id;
217 
218 again:
219 	id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
220 	if (id < 0) {
221 		if (id == -ENOSPC) {
222 			/* Cyclic allocator, reset next id */
223 			if (retries--) {
224 				mem_id_next = MEM_ID_MIN;
225 				goto again;
226 			}
227 		}
228 		return id; /* errno */
229 	}
230 	mem_id_next = id + 1;
231 
232 	return id;
233 }
234 
235 static bool __is_supported_mem_type(enum xdp_mem_type type)
236 {
237 	if (type == MEM_TYPE_PAGE_POOL)
238 		return is_page_pool_compiled_in();
239 
240 	if (type >= MEM_TYPE_MAX)
241 		return false;
242 
243 	return true;
244 }
245 
246 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
247 			       enum xdp_mem_type type, void *allocator)
248 {
249 	struct xdp_mem_allocator *xdp_alloc;
250 	gfp_t gfp = GFP_KERNEL;
251 	int id, errno, ret;
252 	void *ptr;
253 
254 	if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
255 		WARN(1, "Missing register, driver bug");
256 		return -EFAULT;
257 	}
258 
259 	if (!__is_supported_mem_type(type))
260 		return -EOPNOTSUPP;
261 
262 	xdp_rxq->mem.type = type;
263 
264 	if (!allocator) {
265 		if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
266 			return -EINVAL; /* Setup time check page_pool req */
267 		return 0;
268 	}
269 
270 	/* Delay init of rhashtable to save memory if feature isn't used */
271 	if (!mem_id_init) {
272 		mutex_lock(&mem_id_lock);
273 		ret = __mem_id_init_hash_table();
274 		mutex_unlock(&mem_id_lock);
275 		if (ret < 0) {
276 			WARN_ON(1);
277 			return ret;
278 		}
279 	}
280 
281 	xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
282 	if (!xdp_alloc)
283 		return -ENOMEM;
284 
285 	mutex_lock(&mem_id_lock);
286 	id = __mem_id_cyclic_get(gfp);
287 	if (id < 0) {
288 		errno = id;
289 		goto err;
290 	}
291 	xdp_rxq->mem.id = id;
292 	xdp_alloc->mem  = xdp_rxq->mem;
293 	xdp_alloc->allocator = allocator;
294 
295 	/* Insert allocator into ID lookup table */
296 	ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
297 	if (IS_ERR(ptr)) {
298 		errno = PTR_ERR(ptr);
299 		goto err;
300 	}
301 
302 	mutex_unlock(&mem_id_lock);
303 
304 	return 0;
305 err:
306 	mutex_unlock(&mem_id_lock);
307 	kfree(xdp_alloc);
308 	return errno;
309 }
310 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
311 
312 /* XDP RX runs under NAPI protection, and in different delivery error
313  * scenarios (e.g. queue full), it is possible to return the xdp_frame
314  * while still leveraging this protection.  The @napi_direct boolian
315  * is used for those calls sites.  Thus, allowing for faster recycling
316  * of xdp_frames/pages in those cases.
317  */
318 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
319 			 unsigned long handle)
320 {
321 	struct xdp_mem_allocator *xa;
322 	struct page *page;
323 
324 	switch (mem->type) {
325 	case MEM_TYPE_PAGE_POOL:
326 		rcu_read_lock();
327 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
328 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
329 		page = virt_to_head_page(data);
330 		if (xa)
331 			page_pool_put_page(xa->page_pool, page, napi_direct);
332 		else
333 			put_page(page);
334 		rcu_read_unlock();
335 		break;
336 	case MEM_TYPE_PAGE_SHARED:
337 		page_frag_free(data);
338 		break;
339 	case MEM_TYPE_PAGE_ORDER0:
340 		page = virt_to_page(data); /* Assumes order0 page*/
341 		put_page(page);
342 		break;
343 	case MEM_TYPE_ZERO_COPY:
344 		/* NB! Only valid from an xdp_buff! */
345 		rcu_read_lock();
346 		/* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
347 		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
348 		xa->zc_alloc->free(xa->zc_alloc, handle);
349 		rcu_read_unlock();
350 	default:
351 		/* Not possible, checked in xdp_rxq_info_reg_mem_model() */
352 		break;
353 	}
354 }
355 
356 void xdp_return_frame(struct xdp_frame *xdpf)
357 {
358 	__xdp_return(xdpf->data, &xdpf->mem, false, 0);
359 }
360 EXPORT_SYMBOL_GPL(xdp_return_frame);
361 
362 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
363 {
364 	__xdp_return(xdpf->data, &xdpf->mem, true, 0);
365 }
366 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
367 
368 void xdp_return_buff(struct xdp_buff *xdp)
369 {
370 	__xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
371 }
372 EXPORT_SYMBOL_GPL(xdp_return_buff);
373