xref: /linux/net/xdp/xsk_queue.h (revision 4d374ba0bf30a2a372167ee4b7cdd527e7b47b3b)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* XDP user-space ring structure
3  * Copyright(c) 2018 Intel Corporation.
4  */
5 
6 #ifndef _LINUX_XSK_QUEUE_H
7 #define _LINUX_XSK_QUEUE_H
8 
9 #include <linux/types.h>
10 #include <linux/if_xdp.h>
11 #include <net/xdp_sock.h>
12 
13 #define RX_BATCH_SIZE 16
14 #define LAZY_UPDATE_THRESHOLD 128
15 
16 struct xdp_ring {
17 	u32 producer ____cacheline_aligned_in_smp;
18 	u32 consumer ____cacheline_aligned_in_smp;
19 	u32 flags;
20 };
21 
22 /* Used for the RX and TX queues for packets */
23 struct xdp_rxtx_ring {
24 	struct xdp_ring ptrs;
25 	struct xdp_desc desc[0] ____cacheline_aligned_in_smp;
26 };
27 
28 /* Used for the fill and completion queues for buffers */
29 struct xdp_umem_ring {
30 	struct xdp_ring ptrs;
31 	u64 desc[0] ____cacheline_aligned_in_smp;
32 };
33 
34 struct xsk_queue {
35 	u64 chunk_mask;
36 	u64 size;
37 	u32 ring_mask;
38 	u32 nentries;
39 	u32 prod_head;
40 	u32 prod_tail;
41 	u32 cons_head;
42 	u32 cons_tail;
43 	struct xdp_ring *ring;
44 	u64 invalid_descs;
45 };
46 
47 /* The structure of the shared state of the rings are the same as the
48  * ring buffer in kernel/events/ring_buffer.c. For the Rx and completion
49  * ring, the kernel is the producer and user space is the consumer. For
50  * the Tx and fill rings, the kernel is the consumer and user space is
51  * the producer.
52  *
53  * producer                         consumer
54  *
55  * if (LOAD ->consumer) {           LOAD ->producer
56  *                    (A)           smp_rmb()       (C)
57  *    STORE $data                   LOAD $data
58  *    smp_wmb()       (B)           smp_mb()        (D)
59  *    STORE ->producer              STORE ->consumer
60  * }
61  *
62  * (A) pairs with (D), and (B) pairs with (C).
63  *
64  * Starting with (B), it protects the data from being written after
65  * the producer pointer. If this barrier was missing, the consumer
66  * could observe the producer pointer being set and thus load the data
67  * before the producer has written the new data. The consumer would in
68  * this case load the old data.
69  *
70  * (C) protects the consumer from speculatively loading the data before
71  * the producer pointer actually has been read. If we do not have this
72  * barrier, some architectures could load old data as speculative loads
73  * are not discarded as the CPU does not know there is a dependency
74  * between ->producer and data.
75  *
76  * (A) is a control dependency that separates the load of ->consumer
77  * from the stores of $data. In case ->consumer indicates there is no
78  * room in the buffer to store $data we do not. So no barrier is needed.
79  *
80  * (D) protects the load of the data to be observed to happen after the
81  * store of the consumer pointer. If we did not have this memory
82  * barrier, the producer could observe the consumer pointer being set
83  * and overwrite the data with a new value before the consumer got the
84  * chance to read the old value. The consumer would thus miss reading
85  * the old entry and very likely read the new entry twice, once right
86  * now and again after circling through the ring.
87  */
88 
89 /* Common functions operating for both RXTX and umem queues */
90 
91 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
92 {
93 	return q ? q->invalid_descs : 0;
94 }
95 
96 static inline u32 xskq_nb_avail(struct xsk_queue *q, u32 dcnt)
97 {
98 	u32 entries = q->prod_tail - q->cons_tail;
99 
100 	if (entries == 0) {
101 		/* Refresh the local pointer */
102 		q->prod_tail = READ_ONCE(q->ring->producer);
103 		entries = q->prod_tail - q->cons_tail;
104 	}
105 
106 	return (entries > dcnt) ? dcnt : entries;
107 }
108 
109 static inline u32 xskq_nb_free(struct xsk_queue *q, u32 producer, u32 dcnt)
110 {
111 	u32 free_entries = q->nentries - (producer - q->cons_tail);
112 
113 	if (free_entries >= dcnt)
114 		return free_entries;
115 
116 	/* Refresh the local tail pointer */
117 	q->cons_tail = READ_ONCE(q->ring->consumer);
118 	return q->nentries - (producer - q->cons_tail);
119 }
120 
121 static inline bool xskq_has_addrs(struct xsk_queue *q, u32 cnt)
122 {
123 	u32 entries = q->prod_tail - q->cons_tail;
124 
125 	if (entries >= cnt)
126 		return true;
127 
128 	/* Refresh the local pointer. */
129 	q->prod_tail = READ_ONCE(q->ring->producer);
130 	entries = q->prod_tail - q->cons_tail;
131 
132 	return entries >= cnt;
133 }
134 
135 /* UMEM queue */
136 
137 static inline bool xskq_is_valid_addr(struct xsk_queue *q, u64 addr)
138 {
139 	if (addr >= q->size) {
140 		q->invalid_descs++;
141 		return false;
142 	}
143 
144 	return true;
145 }
146 
147 static inline u64 *xskq_validate_addr(struct xsk_queue *q, u64 *addr)
148 {
149 	while (q->cons_tail != q->cons_head) {
150 		struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
151 		unsigned int idx = q->cons_tail & q->ring_mask;
152 
153 		*addr = READ_ONCE(ring->desc[idx]) & q->chunk_mask;
154 		if (xskq_is_valid_addr(q, *addr))
155 			return addr;
156 
157 		q->cons_tail++;
158 	}
159 
160 	return NULL;
161 }
162 
163 static inline u64 *xskq_peek_addr(struct xsk_queue *q, u64 *addr)
164 {
165 	if (q->cons_tail == q->cons_head) {
166 		smp_mb(); /* D, matches A */
167 		WRITE_ONCE(q->ring->consumer, q->cons_tail);
168 		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
169 
170 		/* Order consumer and data */
171 		smp_rmb();
172 	}
173 
174 	return xskq_validate_addr(q, addr);
175 }
176 
177 static inline void xskq_discard_addr(struct xsk_queue *q)
178 {
179 	q->cons_tail++;
180 }
181 
182 static inline int xskq_produce_addr(struct xsk_queue *q, u64 addr)
183 {
184 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
185 
186 	if (xskq_nb_free(q, q->prod_tail, 1) == 0)
187 		return -ENOSPC;
188 
189 	/* A, matches D */
190 	ring->desc[q->prod_tail++ & q->ring_mask] = addr;
191 
192 	/* Order producer and data */
193 	smp_wmb(); /* B, matches C */
194 
195 	WRITE_ONCE(q->ring->producer, q->prod_tail);
196 	return 0;
197 }
198 
199 static inline int xskq_produce_addr_lazy(struct xsk_queue *q, u64 addr)
200 {
201 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
202 
203 	if (xskq_nb_free(q, q->prod_head, LAZY_UPDATE_THRESHOLD) == 0)
204 		return -ENOSPC;
205 
206 	/* A, matches D */
207 	ring->desc[q->prod_head++ & q->ring_mask] = addr;
208 	return 0;
209 }
210 
211 static inline void xskq_produce_flush_addr_n(struct xsk_queue *q,
212 					     u32 nb_entries)
213 {
214 	/* Order producer and data */
215 	smp_wmb(); /* B, matches C */
216 
217 	q->prod_tail += nb_entries;
218 	WRITE_ONCE(q->ring->producer, q->prod_tail);
219 }
220 
221 static inline int xskq_reserve_addr(struct xsk_queue *q)
222 {
223 	if (xskq_nb_free(q, q->prod_head, 1) == 0)
224 		return -ENOSPC;
225 
226 	/* A, matches D */
227 	q->prod_head++;
228 	return 0;
229 }
230 
231 /* Rx/Tx queue */
232 
233 static inline bool xskq_is_valid_desc(struct xsk_queue *q, struct xdp_desc *d)
234 {
235 	if (!xskq_is_valid_addr(q, d->addr))
236 		return false;
237 
238 	if (((d->addr + d->len) & q->chunk_mask) != (d->addr & q->chunk_mask) ||
239 	    d->options) {
240 		q->invalid_descs++;
241 		return false;
242 	}
243 
244 	return true;
245 }
246 
247 static inline struct xdp_desc *xskq_validate_desc(struct xsk_queue *q,
248 						  struct xdp_desc *desc)
249 {
250 	while (q->cons_tail != q->cons_head) {
251 		struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
252 		unsigned int idx = q->cons_tail & q->ring_mask;
253 
254 		*desc = READ_ONCE(ring->desc[idx]);
255 		if (xskq_is_valid_desc(q, desc))
256 			return desc;
257 
258 		q->cons_tail++;
259 	}
260 
261 	return NULL;
262 }
263 
264 static inline struct xdp_desc *xskq_peek_desc(struct xsk_queue *q,
265 					      struct xdp_desc *desc)
266 {
267 	if (q->cons_tail == q->cons_head) {
268 		smp_mb(); /* D, matches A */
269 		WRITE_ONCE(q->ring->consumer, q->cons_tail);
270 		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
271 
272 		/* Order consumer and data */
273 		smp_rmb(); /* C, matches B */
274 	}
275 
276 	return xskq_validate_desc(q, desc);
277 }
278 
279 static inline void xskq_discard_desc(struct xsk_queue *q)
280 {
281 	q->cons_tail++;
282 }
283 
284 static inline int xskq_produce_batch_desc(struct xsk_queue *q,
285 					  u64 addr, u32 len)
286 {
287 	struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
288 	unsigned int idx;
289 
290 	if (xskq_nb_free(q, q->prod_head, 1) == 0)
291 		return -ENOSPC;
292 
293 	/* A, matches D */
294 	idx = (q->prod_head++) & q->ring_mask;
295 	ring->desc[idx].addr = addr;
296 	ring->desc[idx].len = len;
297 
298 	return 0;
299 }
300 
301 static inline void xskq_produce_flush_desc(struct xsk_queue *q)
302 {
303 	/* Order producer and data */
304 	smp_wmb(); /* B, matches C */
305 
306 	q->prod_tail = q->prod_head;
307 	WRITE_ONCE(q->ring->producer, q->prod_tail);
308 }
309 
310 static inline bool xskq_full_desc(struct xsk_queue *q)
311 {
312 	return xskq_nb_avail(q, q->nentries) == q->nentries;
313 }
314 
315 static inline bool xskq_empty_desc(struct xsk_queue *q)
316 {
317 	return xskq_nb_free(q, q->prod_tail, q->nentries) == q->nentries;
318 }
319 
320 void xskq_set_umem(struct xsk_queue *q, u64 size, u64 chunk_mask);
321 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
322 void xskq_destroy(struct xsk_queue *q_ops);
323 
324 /* Executed by the core when the entire UMEM gets freed */
325 void xsk_reuseq_destroy(struct xdp_umem *umem);
326 
327 #endif /* _LINUX_XSK_QUEUE_H */
328