net/xdp/xsk_queue.h

   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 bool xskq_cons_read_addr_unchecked(struct xsk_queue *q, u64 *addr)
 115 {
 116         struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
 117
 118         if (q->cached_cons != q->cached_prod) {
 119                 u32 idx = q->cached_cons & q->ring_mask;
 120
 121                 *addr = ring->desc[idx];
 122                 return true;
 123         }
 124
 125         return false;
 126 }
 127
 128 static inline bool xp_aligned_validate_desc(struct xsk_buff_pool *pool,
 129                                             struct xdp_desc *desc)
 130 {
 131         u64 chunk, chunk_end;
 132
 133         chunk = xp_aligned_extract_addr(pool, desc->addr);
 134         if (likely(desc->len)) {
 135                 chunk_end = xp_aligned_extract_addr(pool, desc->addr + desc->len - 1);
 136                 if (chunk != chunk_end)
 137                         return false;
 138         }
 139
 140         if (chunk >= pool->addrs_cnt)
 141                 return false;
 142
 143         if (desc->options)
 144                 return false;
 145         return true;
 146 }
 147
 148 static inline bool xp_unaligned_validate_desc(struct xsk_buff_pool *pool,
 149                                               struct xdp_desc *desc)
 150 {
 151         u64 addr, base_addr;
 152
 153         base_addr = xp_unaligned_extract_addr(desc->addr);
 154         addr = xp_unaligned_add_offset_to_addr(desc->addr);
 155
 156         if (desc->len > pool->chunk_size)
 157                 return false;
 158
 159         if (base_addr >= pool->addrs_cnt || addr >= pool->addrs_cnt ||
 160             xp_desc_crosses_non_contig_pg(pool, addr, desc->len))
 161                 return false;
 162
 163         if (desc->options)
 164                 return false;
 165         return true;
 166 }
 167
 168 static inline bool xp_validate_desc(struct xsk_buff_pool *pool,
 169                                     struct xdp_desc *desc)
 170 {
 171         return pool->unaligned ? xp_unaligned_validate_desc(pool, desc) :
 172                 xp_aligned_validate_desc(pool, desc);
 173 }
 174
 175 static inline bool xskq_cons_is_valid_desc(struct xsk_queue *q,
 176                                            struct xdp_desc *d,
 177                                            struct xsk_buff_pool *pool)
 178 {
 179         if (!xp_validate_desc(pool, d)) {
 180                 q->invalid_descs++;
 181                 return false;
 182         }
 183         return true;
 184 }
 185
 186 static inline bool xskq_cons_read_desc(struct xsk_queue *q,
 187                                        struct xdp_desc *desc,
 188                                        struct xsk_buff_pool *pool)
 189 {
 190         while (q->cached_cons != q->cached_prod) {
 191                 struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
 192                 u32 idx = q->cached_cons & q->ring_mask;
 193
 194                 *desc = ring->desc[idx];
 195                 if (xskq_cons_is_valid_desc(q, desc, pool))
 196                         return true;
 197
 198                 q->cached_cons++;
 199         }
 200
 201         return false;
 202 }
 203
 204 static inline u32 xskq_cons_read_desc_batch(struct xsk_queue *q,
 205                                             struct xdp_desc *descs,
 206                                             struct xsk_buff_pool *pool, u32 max)
 207 {
 208         u32 cached_cons = q->cached_cons, nb_entries = 0;
 209
 210         while (cached_cons != q->cached_prod && nb_entries < max) {
 211                 struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
 212                 u32 idx = cached_cons & q->ring_mask;
 213
 214                 descs[nb_entries] = ring->desc[idx];
 215                 if (unlikely(!xskq_cons_is_valid_desc(q, &descs[nb_entries], pool))) {
 216                         /* Skip the entry */
 217                         cached_cons++;
 218                         continue;
 219                 }
 220
 221                 nb_entries++;
 222                 cached_cons++;
 223         }
 224
 225         return nb_entries;
 226 }
 227
 228 /* Functions for consumers */
 229
 230 static inline void __xskq_cons_release(struct xsk_queue *q)
 231 {
 232         smp_store_release(&q->ring->consumer, q->cached_cons); /* D, matchees A */
 233 }
 234
 235 static inline void __xskq_cons_peek(struct xsk_queue *q)
 236 {
 237         /* Refresh the local pointer */
 238         q->cached_prod = smp_load_acquire(&q->ring->producer);  /* C, matches B */
 239 }
 240
 241 static inline void xskq_cons_get_entries(struct xsk_queue *q)
 242 {
 243         __xskq_cons_release(q);
 244         __xskq_cons_peek(q);
 245 }
 246
 247 static inline u32 xskq_cons_nb_entries(struct xsk_queue *q, u32 max)
 248 {
 249         u32 entries = q->cached_prod - q->cached_cons;
 250
 251         if (entries >= max)
 252                 return max;
 253
 254         __xskq_cons_peek(q);
 255         entries = q->cached_prod - q->cached_cons;
 256
 257         return entries >= max ? max : entries;
 258 }
 259
 260 static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt)
 261 {
 262         return xskq_cons_nb_entries(q, cnt) >= cnt ? true : false;
 263 }
 264
 265 static inline bool xskq_cons_peek_addr_unchecked(struct xsk_queue *q, u64 *addr)
 266 {
 267         if (q->cached_prod == q->cached_cons)
 268                 xskq_cons_get_entries(q);
 269         return xskq_cons_read_addr_unchecked(q, addr);
 270 }
 271
 272 static inline bool xskq_cons_peek_desc(struct xsk_queue *q,
 273                                        struct xdp_desc *desc,
 274                                        struct xsk_buff_pool *pool)
 275 {
 276         if (q->cached_prod == q->cached_cons)
 277                 xskq_cons_get_entries(q);
 278         return xskq_cons_read_desc(q, desc, pool);
 279 }
 280
 281 static inline u32 xskq_cons_peek_desc_batch(struct xsk_queue *q, struct xdp_desc *descs,
 282                                             struct xsk_buff_pool *pool, u32 max)
 283 {
 284         u32 entries = xskq_cons_nb_entries(q, max);
 285
 286         return xskq_cons_read_desc_batch(q, descs, pool, entries);
 287 }
 288
 289 /* To improve performance in the xskq_cons_release functions, only update local state here.
 290  * Reflect this to global state when we get new entries from the ring in
 291  * xskq_cons_get_entries() and whenever Rx or Tx processing are completed in the NAPI loop.
 292  */
 293 static inline void xskq_cons_release(struct xsk_queue *q)
 294 {
 295         q->cached_cons++;
 296 }
 297
 298 static inline void xskq_cons_release_n(struct xsk_queue *q, u32 cnt)
 299 {
 300         q->cached_cons += cnt;
 301 }
 302
 303 static inline bool xskq_cons_is_full(struct xsk_queue *q)
 304 {
 305         /* No barriers needed since data is not accessed */
 306         return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer) ==
 307                 q->nentries;
 308 }
 309
 310 static inline u32 xskq_cons_present_entries(struct xsk_queue *q)
 311 {
 312         /* No barriers needed since data is not accessed */
 313         return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer);
 314 }
 315
 316 /* Functions for producers */
 317
 318 static inline u32 xskq_prod_nb_free(struct xsk_queue *q, u32 max)
 319 {
 320         u32 free_entries = q->nentries - (q->cached_prod - q->cached_cons);
 321
 322         if (free_entries >= max)
 323                 return max;
 324
 325         /* Refresh the local tail pointer */
 326         q->cached_cons = READ_ONCE(q->ring->consumer);
 327         free_entries = q->nentries - (q->cached_prod - q->cached_cons);
 328
 329         return free_entries >= max ? max : free_entries;
 330 }
 331
 332 static inline bool xskq_prod_is_full(struct xsk_queue *q)
 333 {
 334         return xskq_prod_nb_free(q, 1) ? false : true;
 335 }
 336
 337 static inline void xskq_prod_cancel(struct xsk_queue *q)
 338 {
 339         q->cached_prod--;
 340 }
 341
 342 static inline int xskq_prod_reserve(struct xsk_queue *q)
 343 {
 344         if (xskq_prod_is_full(q))
 345                 return -ENOSPC;
 346
 347         /* A, matches D */
 348         q->cached_prod++;
 349         return 0;
 350 }
 351
 352 static inline int xskq_prod_reserve_addr(struct xsk_queue *q, u64 addr)
 353 {
 354         struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
 355
 356         if (xskq_prod_is_full(q))
 357                 return -ENOSPC;
 358
 359         /* A, matches D */
 360         ring->desc[q->cached_prod++ & q->ring_mask] = addr;
 361         return 0;
 362 }
 363
 364 static inline u32 xskq_prod_reserve_addr_batch(struct xsk_queue *q, struct xdp_desc *descs,
 365                                                u32 max)
 366 {
 367         struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
 368         u32 nb_entries, i, cached_prod;
 369
 370         nb_entries = xskq_prod_nb_free(q, max);
 371
 372         /* A, matches D */
 373         cached_prod = q->cached_prod;
 374         for (i = 0; i < nb_entries; i++)
 375                 ring->desc[cached_prod++ & q->ring_mask] = descs[i].addr;
 376         q->cached_prod = cached_prod;
 377
 378         return nb_entries;
 379 }
 380
 381 static inline int xskq_prod_reserve_desc(struct xsk_queue *q,
 382                                          u64 addr, u32 len)
 383 {
 384         struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
 385         u32 idx;
 386
 387         if (xskq_prod_is_full(q))
 388                 return -ENOSPC;
 389
 390         /* A, matches D */
 391         idx = q->cached_prod++ & q->ring_mask;
 392         ring->desc[idx].addr = addr;
 393         ring->desc[idx].len = len;
 394
 395         return 0;
 396 }
 397
 398 static inline void __xskq_prod_submit(struct xsk_queue *q, u32 idx)
 399 {
 400         smp_store_release(&q->ring->producer, idx); /* B, matches C */
 401 }
 402
 403 static inline void xskq_prod_submit(struct xsk_queue *q)
 404 {
 405         __xskq_prod_submit(q, q->cached_prod);
 406 }
 407
 408 static inline void xskq_prod_submit_addr(struct xsk_queue *q, u64 addr)
 409 {
 410         struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
 411         u32 idx = q->ring->producer;
 412
 413         ring->desc[idx++ & q->ring_mask] = addr;
 414
 415         __xskq_prod_submit(q, idx);
 416 }
 417
 418 static inline void xskq_prod_submit_n(struct xsk_queue *q, u32 nb_entries)
 419 {
 420         __xskq_prod_submit(q, q->ring->producer + nb_entries);
 421 }
 422
 423 static inline bool xskq_prod_is_empty(struct xsk_queue *q)
 424 {
 425         /* No barriers needed since data is not accessed */
 426         return READ_ONCE(q->ring->consumer) == READ_ONCE(q->ring->producer);
 427 }
 428
 429 /* For both producers and consumers */
 430
 431 static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
 432 {
 433         return q ? q->invalid_descs : 0;
 434 }
 435
 436 static inline u64 xskq_nb_queue_empty_descs(struct xsk_queue *q)
 437 {
 438         return q ? q->queue_empty_descs : 0;
 439 }
 440
 441 struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
 442 void xskq_destroy(struct xsk_queue *q_ops);
 443
 444 #endif /* _LINUX_XSK_QUEUE_H */