1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* XDP user-space ring structure
3 * Copyright(c) 2018 Intel Corporation.
6 #ifndef _LINUX_XSK_QUEUE_H
7 #define _LINUX_XSK_QUEUE_H
9 #include <linux/types.h>
10 #include <linux/if_xdp.h>
11 #include <net/xdp_sock.h>
12 #include <net/xsk_buff_pool.h>
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.
21 u32 pad1 ____cacheline_aligned_in_smp;
22 u32 consumer ____cacheline_aligned_in_smp;
23 u32 pad2 ____cacheline_aligned_in_smp;
25 u32 pad3 ____cacheline_aligned_in_smp;
28 /* Used for the RX and TX queues for packets */
29 struct xdp_rxtx_ring {
31 struct xdp_desc desc[] ____cacheline_aligned_in_smp;
34 /* Used for the fill and completion queues for buffers */
35 struct xdp_umem_ring {
37 u64 desc[] ____cacheline_aligned_in_smp;
45 struct xdp_ring *ring;
47 u64 queue_empty_descs;
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.
59 * if (LOAD ->consumer) { (A) LOAD.acq ->producer (C)
60 * STORE $data LOAD $data
61 * STORE.rel ->producer (B) STORE.rel ->consumer (D)
64 * (A) pairs with (D), and (B) pairs with (C).
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.
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.
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.
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.
92 /* The operations on the rings are the following:
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
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.
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.
109 * The function names below reflect these operations.
112 /* Functions that read and validate content from consumer rings. */
114 static inline void __xskq_cons_read_addr_unchecked(struct xsk_queue *q, u32 cached_cons, u64 *addr)
116 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
117 u32 idx = cached_cons & q->ring_mask;
119 *addr = ring->desc[idx];
122 static inline bool xskq_cons_read_addr_unchecked(struct xsk_queue *q, u64 *addr)
124 if (q->cached_cons != q->cached_prod) {
125 __xskq_cons_read_addr_unchecked(q, q->cached_cons, addr);
132 static inline bool xp_aligned_validate_desc(struct xsk_buff_pool *pool,
133 struct xdp_desc *desc)
135 u64 chunk, chunk_end;
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)
144 if (chunk >= pool->addrs_cnt)
152 static inline bool xp_unaligned_validate_desc(struct xsk_buff_pool *pool,
153 struct xdp_desc *desc)
157 base_addr = xp_unaligned_extract_addr(desc->addr);
158 addr = xp_unaligned_add_offset_to_addr(desc->addr);
160 if (desc->len > pool->chunk_size)
163 if (base_addr >= pool->addrs_cnt || addr >= pool->addrs_cnt ||
164 xp_desc_crosses_non_contig_pg(pool, addr, desc->len))
172 static inline bool xp_validate_desc(struct xsk_buff_pool *pool,
173 struct xdp_desc *desc)
175 return pool->unaligned ? xp_unaligned_validate_desc(pool, desc) :
176 xp_aligned_validate_desc(pool, desc);
179 static inline bool xskq_cons_is_valid_desc(struct xsk_queue *q,
181 struct xsk_buff_pool *pool)
183 if (!xp_validate_desc(pool, d)) {
190 static inline bool xskq_cons_read_desc(struct xsk_queue *q,
191 struct xdp_desc *desc,
192 struct xsk_buff_pool *pool)
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;
198 *desc = ring->desc[idx];
199 if (xskq_cons_is_valid_desc(q, desc, pool))
208 static inline void xskq_cons_release_n(struct xsk_queue *q, u32 cnt)
210 q->cached_cons += cnt;
213 static inline u32 xskq_cons_read_desc_batch(struct xsk_queue *q, struct xsk_buff_pool *pool,
216 u32 cached_cons = q->cached_cons, nb_entries = 0;
217 struct xdp_desc *descs = pool->tx_descs;
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;
223 descs[nb_entries] = ring->desc[idx];
224 if (unlikely(!xskq_cons_is_valid_desc(q, &descs[nb_entries], pool))) {
234 /* Release valid plus any invalid entries */
235 xskq_cons_release_n(q, cached_cons - q->cached_cons);
239 /* Functions for consumers */
241 static inline void __xskq_cons_release(struct xsk_queue *q)
243 smp_store_release(&q->ring->consumer, q->cached_cons); /* D, matchees A */
246 static inline void __xskq_cons_peek(struct xsk_queue *q)
248 /* Refresh the local pointer */
249 q->cached_prod = smp_load_acquire(&q->ring->producer); /* C, matches B */
252 static inline void xskq_cons_get_entries(struct xsk_queue *q)
254 __xskq_cons_release(q);
258 static inline u32 xskq_cons_nb_entries(struct xsk_queue *q, u32 max)
260 u32 entries = q->cached_prod - q->cached_cons;
266 entries = q->cached_prod - q->cached_cons;
268 return entries >= max ? max : entries;
271 static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt)
273 return xskq_cons_nb_entries(q, cnt) >= cnt;
276 static inline bool xskq_cons_peek_addr_unchecked(struct xsk_queue *q, u64 *addr)
278 if (q->cached_prod == q->cached_cons)
279 xskq_cons_get_entries(q);
280 return xskq_cons_read_addr_unchecked(q, addr);
283 static inline bool xskq_cons_peek_desc(struct xsk_queue *q,
284 struct xdp_desc *desc,
285 struct xsk_buff_pool *pool)
287 if (q->cached_prod == q->cached_cons)
288 xskq_cons_get_entries(q);
289 return xskq_cons_read_desc(q, desc, pool);
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.
296 static inline void xskq_cons_release(struct xsk_queue *q)
301 static inline u32 xskq_cons_present_entries(struct xsk_queue *q)
303 /* No barriers needed since data is not accessed */
304 return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer);
307 /* Functions for producers */
309 static inline u32 xskq_prod_nb_free(struct xsk_queue *q, u32 max)
311 u32 free_entries = q->nentries - (q->cached_prod - q->cached_cons);
313 if (free_entries >= max)
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);
320 return free_entries >= max ? max : free_entries;
323 static inline bool xskq_prod_is_full(struct xsk_queue *q)
325 return xskq_prod_nb_free(q, 1) ? false : true;
328 static inline void xskq_prod_cancel(struct xsk_queue *q)
333 static inline int xskq_prod_reserve(struct xsk_queue *q)
335 if (xskq_prod_is_full(q))
343 static inline int xskq_prod_reserve_addr(struct xsk_queue *q, u64 addr)
345 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
347 if (xskq_prod_is_full(q))
351 ring->desc[q->cached_prod++ & q->ring_mask] = addr;
355 static inline void xskq_prod_write_addr_batch(struct xsk_queue *q, struct xdp_desc *descs,
358 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
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;
368 static inline int xskq_prod_reserve_desc(struct xsk_queue *q,
371 struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
374 if (xskq_prod_is_full(q))
378 idx = q->cached_prod++ & q->ring_mask;
379 ring->desc[idx].addr = addr;
380 ring->desc[idx].len = len;
385 static inline void __xskq_prod_submit(struct xsk_queue *q, u32 idx)
387 smp_store_release(&q->ring->producer, idx); /* B, matches C */
390 static inline void xskq_prod_submit(struct xsk_queue *q)
392 __xskq_prod_submit(q, q->cached_prod);
395 static inline void xskq_prod_submit_addr(struct xsk_queue *q, u64 addr)
397 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
398 u32 idx = q->ring->producer;
400 ring->desc[idx++ & q->ring_mask] = addr;
402 __xskq_prod_submit(q, idx);
405 static inline void xskq_prod_submit_n(struct xsk_queue *q, u32 nb_entries)
407 __xskq_prod_submit(q, q->ring->producer + nb_entries);
410 static inline bool xskq_prod_is_empty(struct xsk_queue *q)
412 /* No barriers needed since data is not accessed */
413 return READ_ONCE(q->ring->consumer) == READ_ONCE(q->ring->producer);
416 /* For both producers and consumers */
418 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
420 return q ? q->invalid_descs : 0;
423 static inline u64 xskq_nb_queue_empty_descs(struct xsk_queue *q)
425 return q ? q->queue_empty_descs : 0;
428 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
429 void xskq_destroy(struct xsk_queue *q_ops);
431 #endif /* _LINUX_XSK_QUEUE_H */