xref: /openbmc/linux/net/xdp/xsk_queue.h (revision 4e95bc26)
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 };
20 
21 /* Used for the RX and TX queues for packets */
22 struct xdp_rxtx_ring {
23 	struct xdp_ring ptrs;
24 	struct xdp_desc desc[0] ____cacheline_aligned_in_smp;
25 };
26 
27 /* Used for the fill and completion queues for buffers */
28 struct xdp_umem_ring {
29 	struct xdp_ring ptrs;
30 	u64 desc[0] ____cacheline_aligned_in_smp;
31 };
32 
33 struct xsk_queue {
34 	u64 chunk_mask;
35 	u64 size;
36 	u32 ring_mask;
37 	u32 nentries;
38 	u32 prod_head;
39 	u32 prod_tail;
40 	u32 cons_head;
41 	u32 cons_tail;
42 	struct xdp_ring *ring;
43 	u64 invalid_descs;
44 };
45 
46 /* The structure of the shared state of the rings are the same as the
47  * ring buffer in kernel/events/ring_buffer.c. For the Rx and completion
48  * ring, the kernel is the producer and user space is the consumer. For
49  * the Tx and fill rings, the kernel is the consumer and user space is
50  * the producer.
51  *
52  * producer                         consumer
53  *
54  * if (LOAD ->consumer) {           LOAD ->producer
55  *                    (A)           smp_rmb()       (C)
56  *    STORE $data                   LOAD $data
57  *    smp_wmb()       (B)           smp_mb()        (D)
58  *    STORE ->producer              STORE ->consumer
59  * }
60  *
61  * (A) pairs with (D), and (B) pairs with (C).
62  *
63  * Starting with (B), it protects the data from being written after
64  * the producer pointer. If this barrier was missing, the consumer
65  * could observe the producer pointer being set and thus load the data
66  * before the producer has written the new data. The consumer would in
67  * this case load the old data.
68  *
69  * (C) protects the consumer from speculatively loading the data before
70  * the producer pointer actually has been read. If we do not have this
71  * barrier, some architectures could load old data as speculative loads
72  * are not discarded as the CPU does not know there is a dependency
73  * between ->producer and data.
74  *
75  * (A) is a control dependency that separates the load of ->consumer
76  * from the stores of $data. In case ->consumer indicates there is no
77  * room in the buffer to store $data we do not. So no barrier is needed.
78  *
79  * (D) protects the load of the data to be observed to happen after the
80  * store of the consumer pointer. If we did not have this memory
81  * barrier, the producer could observe the consumer pointer being set
82  * and overwrite the data with a new value before the consumer got the
83  * chance to read the old value. The consumer would thus miss reading
84  * the old entry and very likely read the new entry twice, once right
85  * now and again after circling through the ring.
86  */
87 
88 /* Common functions operating for both RXTX and umem queues */
89 
90 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
91 {
92 	return q ? q->invalid_descs : 0;
93 }
94 
95 static inline u32 xskq_nb_avail(struct xsk_queue *q, u32 dcnt)
96 {
97 	u32 entries = q->prod_tail - q->cons_tail;
98 
99 	if (entries == 0) {
100 		/* Refresh the local pointer */
101 		q->prod_tail = READ_ONCE(q->ring->producer);
102 		entries = q->prod_tail - q->cons_tail;
103 	}
104 
105 	return (entries > dcnt) ? dcnt : entries;
106 }
107 
108 static inline u32 xskq_nb_free(struct xsk_queue *q, u32 producer, u32 dcnt)
109 {
110 	u32 free_entries = q->nentries - (producer - q->cons_tail);
111 
112 	if (free_entries >= dcnt)
113 		return free_entries;
114 
115 	/* Refresh the local tail pointer */
116 	q->cons_tail = READ_ONCE(q->ring->consumer);
117 	return q->nentries - (producer - q->cons_tail);
118 }
119 
120 static inline bool xskq_has_addrs(struct xsk_queue *q, u32 cnt)
121 {
122 	u32 entries = q->prod_tail - q->cons_tail;
123 
124 	if (entries >= cnt)
125 		return true;
126 
127 	/* Refresh the local pointer. */
128 	q->prod_tail = READ_ONCE(q->ring->producer);
129 	entries = q->prod_tail - q->cons_tail;
130 
131 	return entries >= cnt;
132 }
133 
134 /* UMEM queue */
135 
136 static inline bool xskq_is_valid_addr(struct xsk_queue *q, u64 addr)
137 {
138 	if (addr >= q->size) {
139 		q->invalid_descs++;
140 		return false;
141 	}
142 
143 	return true;
144 }
145 
146 static inline u64 *xskq_validate_addr(struct xsk_queue *q, u64 *addr)
147 {
148 	while (q->cons_tail != q->cons_head) {
149 		struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
150 		unsigned int idx = q->cons_tail & q->ring_mask;
151 
152 		*addr = READ_ONCE(ring->desc[idx]) & q->chunk_mask;
153 		if (xskq_is_valid_addr(q, *addr))
154 			return addr;
155 
156 		q->cons_tail++;
157 	}
158 
159 	return NULL;
160 }
161 
162 static inline u64 *xskq_peek_addr(struct xsk_queue *q, u64 *addr)
163 {
164 	if (q->cons_tail == q->cons_head) {
165 		smp_mb(); /* D, matches A */
166 		WRITE_ONCE(q->ring->consumer, q->cons_tail);
167 		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
168 
169 		/* Order consumer and data */
170 		smp_rmb();
171 	}
172 
173 	return xskq_validate_addr(q, addr);
174 }
175 
176 static inline void xskq_discard_addr(struct xsk_queue *q)
177 {
178 	q->cons_tail++;
179 }
180 
181 static inline int xskq_produce_addr(struct xsk_queue *q, u64 addr)
182 {
183 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
184 
185 	if (xskq_nb_free(q, q->prod_tail, 1) == 0)
186 		return -ENOSPC;
187 
188 	/* A, matches D */
189 	ring->desc[q->prod_tail++ & q->ring_mask] = addr;
190 
191 	/* Order producer and data */
192 	smp_wmb(); /* B, matches C */
193 
194 	WRITE_ONCE(q->ring->producer, q->prod_tail);
195 	return 0;
196 }
197 
198 static inline int xskq_produce_addr_lazy(struct xsk_queue *q, u64 addr)
199 {
200 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
201 
202 	if (xskq_nb_free(q, q->prod_head, LAZY_UPDATE_THRESHOLD) == 0)
203 		return -ENOSPC;
204 
205 	/* A, matches D */
206 	ring->desc[q->prod_head++ & q->ring_mask] = addr;
207 	return 0;
208 }
209 
210 static inline void xskq_produce_flush_addr_n(struct xsk_queue *q,
211 					     u32 nb_entries)
212 {
213 	/* Order producer and data */
214 	smp_wmb(); /* B, matches C */
215 
216 	q->prod_tail += nb_entries;
217 	WRITE_ONCE(q->ring->producer, q->prod_tail);
218 }
219 
220 static inline int xskq_reserve_addr(struct xsk_queue *q)
221 {
222 	if (xskq_nb_free(q, q->prod_head, 1) == 0)
223 		return -ENOSPC;
224 
225 	/* A, matches D */
226 	q->prod_head++;
227 	return 0;
228 }
229 
230 /* Rx/Tx queue */
231 
232 static inline bool xskq_is_valid_desc(struct xsk_queue *q, struct xdp_desc *d)
233 {
234 	if (!xskq_is_valid_addr(q, d->addr))
235 		return false;
236 
237 	if (((d->addr + d->len) & q->chunk_mask) != (d->addr & q->chunk_mask) ||
238 	    d->options) {
239 		q->invalid_descs++;
240 		return false;
241 	}
242 
243 	return true;
244 }
245 
246 static inline struct xdp_desc *xskq_validate_desc(struct xsk_queue *q,
247 						  struct xdp_desc *desc)
248 {
249 	while (q->cons_tail != q->cons_head) {
250 		struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
251 		unsigned int idx = q->cons_tail & q->ring_mask;
252 
253 		*desc = READ_ONCE(ring->desc[idx]);
254 		if (xskq_is_valid_desc(q, desc))
255 			return desc;
256 
257 		q->cons_tail++;
258 	}
259 
260 	return NULL;
261 }
262 
263 static inline struct xdp_desc *xskq_peek_desc(struct xsk_queue *q,
264 					      struct xdp_desc *desc)
265 {
266 	if (q->cons_tail == q->cons_head) {
267 		smp_mb(); /* D, matches A */
268 		WRITE_ONCE(q->ring->consumer, q->cons_tail);
269 		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
270 
271 		/* Order consumer and data */
272 		smp_rmb(); /* C, matches B */
273 	}
274 
275 	return xskq_validate_desc(q, desc);
276 }
277 
278 static inline void xskq_discard_desc(struct xsk_queue *q)
279 {
280 	q->cons_tail++;
281 }
282 
283 static inline int xskq_produce_batch_desc(struct xsk_queue *q,
284 					  u64 addr, u32 len)
285 {
286 	struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
287 	unsigned int idx;
288 
289 	if (xskq_nb_free(q, q->prod_head, 1) == 0)
290 		return -ENOSPC;
291 
292 	/* A, matches D */
293 	idx = (q->prod_head++) & q->ring_mask;
294 	ring->desc[idx].addr = addr;
295 	ring->desc[idx].len = len;
296 
297 	return 0;
298 }
299 
300 static inline void xskq_produce_flush_desc(struct xsk_queue *q)
301 {
302 	/* Order producer and data */
303 	smp_wmb(); /* B, matches C */
304 
305 	q->prod_tail = q->prod_head;
306 	WRITE_ONCE(q->ring->producer, q->prod_tail);
307 }
308 
309 static inline bool xskq_full_desc(struct xsk_queue *q)
310 {
311 	return xskq_nb_avail(q, q->nentries) == q->nentries;
312 }
313 
314 static inline bool xskq_empty_desc(struct xsk_queue *q)
315 {
316 	return xskq_nb_free(q, q->prod_tail, q->nentries) == q->nentries;
317 }
318 
319 void xskq_set_umem(struct xsk_queue *q, u64 size, u64 chunk_mask);
320 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
321 void xskq_destroy(struct xsk_queue *q_ops);
322 
323 /* Executed by the core when the entire UMEM gets freed */
324 void xsk_reuseq_destroy(struct xdp_umem *umem);
325 
326 #endif /* _LINUX_XSK_QUEUE_H */
327