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 u32 xskq_cons_read_desc_batch(struct xsk_queue *q, struct xsk_buff_pool *pool, 209 u32 max) 210 { 211 u32 cached_cons = q->cached_cons, nb_entries = 0; 212 struct xdp_desc *descs = pool->tx_descs; 213 214 while (cached_cons != q->cached_prod && nb_entries < max) { 215 struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring; 216 u32 idx = cached_cons & q->ring_mask; 217 218 descs[nb_entries] = ring->desc[idx]; 219 if (unlikely(!xskq_cons_is_valid_desc(q, &descs[nb_entries], pool))) { 220 /* Skip the entry */ 221 cached_cons++; 222 continue; 223 } 224 225 nb_entries++; 226 cached_cons++; 227 } 228 229 return nb_entries; 230 } 231 232 /* Functions for consumers */ 233 234 static inline void __xskq_cons_release(struct xsk_queue *q) 235 { 236 smp_store_release(&q->ring->consumer, q->cached_cons); /* D, matchees A */ 237 } 238 239 static inline void __xskq_cons_peek(struct xsk_queue *q) 240 { 241 /* Refresh the local pointer */ 242 q->cached_prod = smp_load_acquire(&q->ring->producer); /* C, matches B */ 243 } 244 245 static inline void xskq_cons_get_entries(struct xsk_queue *q) 246 { 247 __xskq_cons_release(q); 248 __xskq_cons_peek(q); 249 } 250 251 static inline u32 xskq_cons_nb_entries(struct xsk_queue *q, u32 max) 252 { 253 u32 entries = q->cached_prod - q->cached_cons; 254 255 if (entries >= max) 256 return max; 257 258 __xskq_cons_peek(q); 259 entries = q->cached_prod - q->cached_cons; 260 261 return entries >= max ? max : entries; 262 } 263 264 static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt) 265 { 266 return xskq_cons_nb_entries(q, cnt) >= cnt; 267 } 268 269 static inline bool xskq_cons_peek_addr_unchecked(struct xsk_queue *q, u64 *addr) 270 { 271 if (q->cached_prod == q->cached_cons) 272 xskq_cons_get_entries(q); 273 return xskq_cons_read_addr_unchecked(q, addr); 274 } 275 276 static inline bool xskq_cons_peek_desc(struct xsk_queue *q, 277 struct xdp_desc *desc, 278 struct xsk_buff_pool *pool) 279 { 280 if (q->cached_prod == q->cached_cons) 281 xskq_cons_get_entries(q); 282 return xskq_cons_read_desc(q, desc, pool); 283 } 284 285 /* To improve performance in the xskq_cons_release functions, only update local state here. 286 * Reflect this to global state when we get new entries from the ring in 287 * xskq_cons_get_entries() and whenever Rx or Tx processing are completed in the NAPI loop. 288 */ 289 static inline void xskq_cons_release(struct xsk_queue *q) 290 { 291 q->cached_cons++; 292 } 293 294 static inline void xskq_cons_release_n(struct xsk_queue *q, u32 cnt) 295 { 296 q->cached_cons += cnt; 297 } 298 299 static inline u32 xskq_cons_present_entries(struct xsk_queue *q) 300 { 301 /* No barriers needed since data is not accessed */ 302 return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer); 303 } 304 305 /* Functions for producers */ 306 307 static inline u32 xskq_prod_nb_free(struct xsk_queue *q, u32 max) 308 { 309 u32 free_entries = q->nentries - (q->cached_prod - q->cached_cons); 310 311 if (free_entries >= max) 312 return max; 313 314 /* Refresh the local tail pointer */ 315 q->cached_cons = READ_ONCE(q->ring->consumer); 316 free_entries = q->nentries - (q->cached_prod - q->cached_cons); 317 318 return free_entries >= max ? max : free_entries; 319 } 320 321 static inline bool xskq_prod_is_full(struct xsk_queue *q) 322 { 323 return xskq_prod_nb_free(q, 1) ? false : true; 324 } 325 326 static inline void xskq_prod_cancel(struct xsk_queue *q) 327 { 328 q->cached_prod--; 329 } 330 331 static inline int xskq_prod_reserve(struct xsk_queue *q) 332 { 333 if (xskq_prod_is_full(q)) 334 return -ENOSPC; 335 336 /* A, matches D */ 337 q->cached_prod++; 338 return 0; 339 } 340 341 static inline int xskq_prod_reserve_addr(struct xsk_queue *q, u64 addr) 342 { 343 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring; 344 345 if (xskq_prod_is_full(q)) 346 return -ENOSPC; 347 348 /* A, matches D */ 349 ring->desc[q->cached_prod++ & q->ring_mask] = addr; 350 return 0; 351 } 352 353 static inline u32 xskq_prod_reserve_addr_batch(struct xsk_queue *q, struct xdp_desc *descs, 354 u32 max) 355 { 356 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring; 357 u32 nb_entries, i, cached_prod; 358 359 nb_entries = xskq_prod_nb_free(q, max); 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 return nb_entries; 368 } 369 370 static inline int xskq_prod_reserve_desc(struct xsk_queue *q, 371 u64 addr, u32 len) 372 { 373 struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring; 374 u32 idx; 375 376 if (xskq_prod_is_full(q)) 377 return -ENOBUFS; 378 379 /* A, matches D */ 380 idx = q->cached_prod++ & q->ring_mask; 381 ring->desc[idx].addr = addr; 382 ring->desc[idx].len = len; 383 384 return 0; 385 } 386 387 static inline void __xskq_prod_submit(struct xsk_queue *q, u32 idx) 388 { 389 smp_store_release(&q->ring->producer, idx); /* B, matches C */ 390 } 391 392 static inline void xskq_prod_submit(struct xsk_queue *q) 393 { 394 __xskq_prod_submit(q, q->cached_prod); 395 } 396 397 static inline void xskq_prod_submit_addr(struct xsk_queue *q, u64 addr) 398 { 399 struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring; 400 u32 idx = q->ring->producer; 401 402 ring->desc[idx++ & q->ring_mask] = addr; 403 404 __xskq_prod_submit(q, idx); 405 } 406 407 static inline void xskq_prod_submit_n(struct xsk_queue *q, u32 nb_entries) 408 { 409 __xskq_prod_submit(q, q->ring->producer + nb_entries); 410 } 411 412 static inline bool xskq_prod_is_empty(struct xsk_queue *q) 413 { 414 /* No barriers needed since data is not accessed */ 415 return READ_ONCE(q->ring->consumer) == READ_ONCE(q->ring->producer); 416 } 417 418 /* For both producers and consumers */ 419 420 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q) 421 { 422 return q ? q->invalid_descs : 0; 423 } 424 425 static inline u64 xskq_nb_queue_empty_descs(struct xsk_queue *q) 426 { 427 return q ? q->queue_empty_descs : 0; 428 } 429 430 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue); 431 void xskq_destroy(struct xsk_queue *q_ops); 432 433 #endif /* _LINUX_XSK_QUEUE_H */ 434