xref: /openbmc/linux/net/xdp/xsk_queue.h (revision 3bf90eca)
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 #include <net/xsk_buff_pool.h>
13 
14 #include "xsk.h"
15 
16 struct xdp_ring {
17 	u32 producer ____cacheline_aligned_in_smp;
18 	/* Hinder the adjacent cache prefetcher to prefetch the consumer
19 	 * pointer if the producer pointer is touched and vice versa.
20 	 */
21 	u32 pad1 ____cacheline_aligned_in_smp;
22 	u32 consumer ____cacheline_aligned_in_smp;
23 	u32 pad2 ____cacheline_aligned_in_smp;
24 	u32 flags;
25 	u32 pad3 ____cacheline_aligned_in_smp;
26 };
27 
28 /* Used for the RX and TX queues for packets */
29 struct xdp_rxtx_ring {
30 	struct xdp_ring ptrs;
31 	struct xdp_desc desc[] ____cacheline_aligned_in_smp;
32 };
33 
34 /* Used for the fill and completion queues for buffers */
35 struct xdp_umem_ring {
36 	struct xdp_ring ptrs;
37 	u64 desc[] ____cacheline_aligned_in_smp;
38 };
39 
40 struct xsk_queue {
41 	u32 ring_mask;
42 	u32 nentries;
43 	u32 cached_prod;
44 	u32 cached_cons;
45 	struct xdp_ring *ring;
46 	u64 invalid_descs;
47 	u64 queue_empty_descs;
48 };
49 
50 /* The structure of the shared state of the rings are a simple
51  * circular buffer, as outlined in
52  * Documentation/core-api/circular-buffers.rst. For the Rx and
53  * completion ring, the kernel is the producer and user space is the
54  * consumer. For the Tx and fill rings, the kernel is the consumer and
55  * user space is the producer.
56  *
57  * producer                         consumer
58  *
59  * if (LOAD ->consumer) {  (A)      LOAD.acq ->producer  (C)
60  *    STORE $data                   LOAD $data
61  *    STORE.rel ->producer (B)      STORE.rel ->consumer (D)
62  * }
63  *
64  * (A) pairs with (D), and (B) pairs with (C).
65  *
66  * Starting with (B), it protects the data from being written after
67  * the producer pointer. If this barrier was missing, the consumer
68  * could observe the producer pointer being set and thus load the data
69  * before the producer has written the new data. The consumer would in
70  * this case load the old data.
71  *
72  * (C) protects the consumer from speculatively loading the data before
73  * the producer pointer actually has been read. If we do not have this
74  * barrier, some architectures could load old data as speculative loads
75  * are not discarded as the CPU does not know there is a dependency
76  * between ->producer and data.
77  *
78  * (A) is a control dependency that separates the load of ->consumer
79  * from the stores of $data. In case ->consumer indicates there is no
80  * room in the buffer to store $data we do not. The dependency will
81  * order both of the stores after the loads. So no barrier is needed.
82  *
83  * (D) protects the load of the data to be observed to happen after the
84  * store of the consumer pointer. If we did not have this memory
85  * barrier, the producer could observe the consumer pointer being set
86  * and overwrite the data with a new value before the consumer got the
87  * chance to read the old value. The consumer would thus miss reading
88  * the old entry and very likely read the new entry twice, once right
89  * now and again after circling through the ring.
90  */
91 
92 /* The operations on the rings are the following:
93  *
94  * producer                           consumer
95  *
96  * RESERVE entries                    PEEK in the ring for entries
97  * WRITE data into the ring           READ data from the ring
98  * SUBMIT entries                     RELEASE entries
99  *
100  * The producer reserves one or more entries in the ring. It can then
101  * fill in these entries and finally submit them so that they can be
102  * seen and read by the consumer.
103  *
104  * The consumer peeks into the ring to see if the producer has written
105  * any new entries. If so, the consumer can then read these entries
106  * and when it is done reading them release them back to the producer
107  * so that the producer can use these slots to fill in new entries.
108  *
109  * The function names below reflect these operations.
110  */
111 
112 /* Functions that read and validate content from consumer rings. */
113 
114 static inline void __xskq_cons_read_addr_unchecked(struct xsk_queue *q, u32 cached_cons, u64 *addr)
115 {
116 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
117 	u32 idx = cached_cons & q->ring_mask;
118 
119 	*addr = ring->desc[idx];
120 }
121 
122 static inline bool xskq_cons_read_addr_unchecked(struct xsk_queue *q, u64 *addr)
123 {
124 	if (q->cached_cons != q->cached_prod) {
125 		__xskq_cons_read_addr_unchecked(q, q->cached_cons, addr);
126 		return true;
127 	}
128 
129 	return false;
130 }
131 
132 static inline bool xp_aligned_validate_desc(struct xsk_buff_pool *pool,
133 					    struct xdp_desc *desc)
134 {
135 	u64 chunk, chunk_end;
136 
137 	chunk = xp_aligned_extract_addr(pool, desc->addr);
138 	if (likely(desc->len)) {
139 		chunk_end = xp_aligned_extract_addr(pool, desc->addr + desc->len - 1);
140 		if (chunk != chunk_end)
141 			return false;
142 	}
143 
144 	if (chunk >= pool->addrs_cnt)
145 		return false;
146 
147 	if (desc->options)
148 		return false;
149 	return true;
150 }
151 
152 static inline bool xp_unaligned_validate_desc(struct xsk_buff_pool *pool,
153 					      struct xdp_desc *desc)
154 {
155 	u64 addr, base_addr;
156 
157 	base_addr = xp_unaligned_extract_addr(desc->addr);
158 	addr = xp_unaligned_add_offset_to_addr(desc->addr);
159 
160 	if (desc->len > pool->chunk_size)
161 		return false;
162 
163 	if (base_addr >= pool->addrs_cnt || addr >= pool->addrs_cnt ||
164 	    xp_desc_crosses_non_contig_pg(pool, addr, desc->len))
165 		return false;
166 
167 	if (desc->options)
168 		return false;
169 	return true;
170 }
171 
172 static inline bool xp_validate_desc(struct xsk_buff_pool *pool,
173 				    struct xdp_desc *desc)
174 {
175 	return pool->unaligned ? xp_unaligned_validate_desc(pool, desc) :
176 		xp_aligned_validate_desc(pool, desc);
177 }
178 
179 static inline bool xskq_cons_is_valid_desc(struct xsk_queue *q,
180 					   struct xdp_desc *d,
181 					   struct xsk_buff_pool *pool)
182 {
183 	if (!xp_validate_desc(pool, d)) {
184 		q->invalid_descs++;
185 		return false;
186 	}
187 	return true;
188 }
189 
190 static inline bool xskq_cons_read_desc(struct xsk_queue *q,
191 				       struct xdp_desc *desc,
192 				       struct xsk_buff_pool *pool)
193 {
194 	while (q->cached_cons != q->cached_prod) {
195 		struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
196 		u32 idx = q->cached_cons & q->ring_mask;
197 
198 		*desc = ring->desc[idx];
199 		if (xskq_cons_is_valid_desc(q, desc, pool))
200 			return true;
201 
202 		q->cached_cons++;
203 	}
204 
205 	return false;
206 }
207 
208 static inline void xskq_cons_release_n(struct xsk_queue *q, u32 cnt)
209 {
210 	q->cached_cons += cnt;
211 }
212 
213 static inline u32 xskq_cons_read_desc_batch(struct xsk_queue *q, struct xsk_buff_pool *pool,
214 					    u32 max)
215 {
216 	u32 cached_cons = q->cached_cons, nb_entries = 0;
217 	struct xdp_desc *descs = pool->tx_descs;
218 
219 	while (cached_cons != q->cached_prod && nb_entries < max) {
220 		struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
221 		u32 idx = cached_cons & q->ring_mask;
222 
223 		descs[nb_entries] = ring->desc[idx];
224 		if (unlikely(!xskq_cons_is_valid_desc(q, &descs[nb_entries], pool))) {
225 			/* Skip the entry */
226 			cached_cons++;
227 			continue;
228 		}
229 
230 		nb_entries++;
231 		cached_cons++;
232 	}
233 
234 	/* Release valid plus any invalid entries */
235 	xskq_cons_release_n(q, cached_cons - q->cached_cons);
236 	return nb_entries;
237 }
238 
239 /* Functions for consumers */
240 
241 static inline void __xskq_cons_release(struct xsk_queue *q)
242 {
243 	smp_store_release(&q->ring->consumer, q->cached_cons); /* D, matchees A */
244 }
245 
246 static inline void __xskq_cons_peek(struct xsk_queue *q)
247 {
248 	/* Refresh the local pointer */
249 	q->cached_prod = smp_load_acquire(&q->ring->producer);  /* C, matches B */
250 }
251 
252 static inline void xskq_cons_get_entries(struct xsk_queue *q)
253 {
254 	__xskq_cons_release(q);
255 	__xskq_cons_peek(q);
256 }
257 
258 static inline u32 xskq_cons_nb_entries(struct xsk_queue *q, u32 max)
259 {
260 	u32 entries = q->cached_prod - q->cached_cons;
261 
262 	if (entries >= max)
263 		return max;
264 
265 	__xskq_cons_peek(q);
266 	entries = q->cached_prod - q->cached_cons;
267 
268 	return entries >= max ? max : entries;
269 }
270 
271 static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt)
272 {
273 	return xskq_cons_nb_entries(q, cnt) >= cnt;
274 }
275 
276 static inline bool xskq_cons_peek_addr_unchecked(struct xsk_queue *q, u64 *addr)
277 {
278 	if (q->cached_prod == q->cached_cons)
279 		xskq_cons_get_entries(q);
280 	return xskq_cons_read_addr_unchecked(q, addr);
281 }
282 
283 static inline bool xskq_cons_peek_desc(struct xsk_queue *q,
284 				       struct xdp_desc *desc,
285 				       struct xsk_buff_pool *pool)
286 {
287 	if (q->cached_prod == q->cached_cons)
288 		xskq_cons_get_entries(q);
289 	return xskq_cons_read_desc(q, desc, pool);
290 }
291 
292 /* To improve performance in the xskq_cons_release functions, only update local state here.
293  * Reflect this to global state when we get new entries from the ring in
294  * xskq_cons_get_entries() and whenever Rx or Tx processing are completed in the NAPI loop.
295  */
296 static inline void xskq_cons_release(struct xsk_queue *q)
297 {
298 	q->cached_cons++;
299 }
300 
301 static inline u32 xskq_cons_present_entries(struct xsk_queue *q)
302 {
303 	/* No barriers needed since data is not accessed */
304 	return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer);
305 }
306 
307 /* Functions for producers */
308 
309 static inline u32 xskq_prod_nb_free(struct xsk_queue *q, u32 max)
310 {
311 	u32 free_entries = q->nentries - (q->cached_prod - q->cached_cons);
312 
313 	if (free_entries >= max)
314 		return max;
315 
316 	/* Refresh the local tail pointer */
317 	q->cached_cons = READ_ONCE(q->ring->consumer);
318 	free_entries = q->nentries - (q->cached_prod - q->cached_cons);
319 
320 	return free_entries >= max ? max : free_entries;
321 }
322 
323 static inline bool xskq_prod_is_full(struct xsk_queue *q)
324 {
325 	return xskq_prod_nb_free(q, 1) ? false : true;
326 }
327 
328 static inline void xskq_prod_cancel(struct xsk_queue *q)
329 {
330 	q->cached_prod--;
331 }
332 
333 static inline int xskq_prod_reserve(struct xsk_queue *q)
334 {
335 	if (xskq_prod_is_full(q))
336 		return -ENOSPC;
337 
338 	/* A, matches D */
339 	q->cached_prod++;
340 	return 0;
341 }
342 
343 static inline int xskq_prod_reserve_addr(struct xsk_queue *q, u64 addr)
344 {
345 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
346 
347 	if (xskq_prod_is_full(q))
348 		return -ENOSPC;
349 
350 	/* A, matches D */
351 	ring->desc[q->cached_prod++ & q->ring_mask] = addr;
352 	return 0;
353 }
354 
355 static inline void xskq_prod_write_addr_batch(struct xsk_queue *q, struct xdp_desc *descs,
356 					      u32 nb_entries)
357 {
358 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
359 	u32 i, cached_prod;
360 
361 	/* A, matches D */
362 	cached_prod = q->cached_prod;
363 	for (i = 0; i < nb_entries; i++)
364 		ring->desc[cached_prod++ & q->ring_mask] = descs[i].addr;
365 	q->cached_prod = cached_prod;
366 }
367 
368 static inline int xskq_prod_reserve_desc(struct xsk_queue *q,
369 					 u64 addr, u32 len)
370 {
371 	struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
372 	u32 idx;
373 
374 	if (xskq_prod_is_full(q))
375 		return -ENOBUFS;
376 
377 	/* A, matches D */
378 	idx = q->cached_prod++ & q->ring_mask;
379 	ring->desc[idx].addr = addr;
380 	ring->desc[idx].len = len;
381 
382 	return 0;
383 }
384 
385 static inline void __xskq_prod_submit(struct xsk_queue *q, u32 idx)
386 {
387 	smp_store_release(&q->ring->producer, idx); /* B, matches C */
388 }
389 
390 static inline void xskq_prod_submit(struct xsk_queue *q)
391 {
392 	__xskq_prod_submit(q, q->cached_prod);
393 }
394 
395 static inline void xskq_prod_submit_addr(struct xsk_queue *q, u64 addr)
396 {
397 	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
398 	u32 idx = q->ring->producer;
399 
400 	ring->desc[idx++ & q->ring_mask] = addr;
401 
402 	__xskq_prod_submit(q, idx);
403 }
404 
405 static inline void xskq_prod_submit_n(struct xsk_queue *q, u32 nb_entries)
406 {
407 	__xskq_prod_submit(q, q->ring->producer + nb_entries);
408 }
409 
410 static inline bool xskq_prod_is_empty(struct xsk_queue *q)
411 {
412 	/* No barriers needed since data is not accessed */
413 	return READ_ONCE(q->ring->consumer) == READ_ONCE(q->ring->producer);
414 }
415 
416 /* For both producers and consumers */
417 
418 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
419 {
420 	return q ? q->invalid_descs : 0;
421 }
422 
423 static inline u64 xskq_nb_queue_empty_descs(struct xsk_queue *q)
424 {
425 	return q ? q->queue_empty_descs : 0;
426 }
427 
428 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
429 void xskq_destroy(struct xsk_queue *q_ops);
430 
431 #endif /* _LINUX_XSK_QUEUE_H */
432