drivers/net/mlx5/mlx5_utils.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 Mellanox Technologies, Ltd
   3  */
   4
   5 #include <rte_malloc.h>
   6 #include <rte_hash_crc.h>
   7
   8 #include <mlx5_malloc.h>
   9
  10 #include "mlx5_utils.h"
  11
  12 struct mlx5_hlist *
  13 mlx5_hlist_create(const char *name, uint32_t size)
  14 {
  15         struct mlx5_hlist *h;
  16         uint32_t act_size;
  17         uint32_t alloc_size;
  18
  19         if (!size)
  20                 return NULL;
  21         /* Align to the next power of 2, 32bits integer is enough now. */
  22         if (!rte_is_power_of_2(size)) {
  23                 act_size = rte_align32pow2(size);
  24                 DRV_LOG(WARNING, "Size 0x%" PRIX32 " is not power of 2, will "
  25                         "be aligned to 0x%" PRIX32 ".", size, act_size);
  26         } else {
  27                 act_size = size;
  28         }
  29         alloc_size = sizeof(struct mlx5_hlist) +
  30                      sizeof(struct mlx5_hlist_head) * act_size;
  31         /* Using zmalloc, then no need to initialize the heads. */
  32         h = mlx5_malloc(MLX5_MEM_ZERO, alloc_size, RTE_CACHE_LINE_SIZE,
  33                         SOCKET_ID_ANY);
  34         if (!h) {
  35                 DRV_LOG(ERR, "No memory for hash list %s creation",
  36                         name ? name : "None");
  37                 return NULL;
  38         }
  39         if (name)
  40                 snprintf(h->name, MLX5_HLIST_NAMESIZE, "%s", name);
  41         h->table_sz = act_size;
  42         h->mask = act_size - 1;
  43         DRV_LOG(DEBUG, "Hash list with %s size 0x%" PRIX32 " is created.",
  44                 h->name, act_size);
  45         return h;
  46 }
  47
  48 struct mlx5_hlist_entry *
  49 mlx5_hlist_lookup(struct mlx5_hlist *h, uint64_t key)
  50 {
  51         uint32_t idx;
  52         struct mlx5_hlist_head *first;
  53         struct mlx5_hlist_entry *node;
  54
  55         MLX5_ASSERT(h);
  56         idx = rte_hash_crc_8byte(key, 0) & h->mask;
  57         first = &h->heads[idx];
  58         LIST_FOREACH(node, first, next) {
  59                 if (node->key == key)
  60                         return node;
  61         }
  62         return NULL;
  63 }
  64
  65 int
  66 mlx5_hlist_insert(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry)
  67 {
  68         uint32_t idx;
  69         struct mlx5_hlist_head *first;
  70         struct mlx5_hlist_entry *node;
  71
  72         MLX5_ASSERT(h && entry);
  73         idx = rte_hash_crc_8byte(entry->key, 0) & h->mask;
  74         first = &h->heads[idx];
  75         /* No need to reuse the lookup function. */
  76         LIST_FOREACH(node, first, next) {
  77                 if (node->key == entry->key)
  78                         return -EEXIST;
  79         }
  80         LIST_INSERT_HEAD(first, entry, next);
  81         return 0;
  82 }
  83
  84 struct mlx5_hlist_entry *
  85 mlx5_hlist_lookup_ex(struct mlx5_hlist *h, uint64_t key,
  86                      mlx5_hlist_match_callback_fn cb, void *ctx)
  87 {
  88         uint32_t idx;
  89         struct mlx5_hlist_head *first;
  90         struct mlx5_hlist_entry *node;
  91
  92         MLX5_ASSERT(h && cb && ctx);
  93         idx = rte_hash_crc_8byte(key, 0) & h->mask;
  94         first = &h->heads[idx];
  95         LIST_FOREACH(node, first, next) {
  96                 if (!cb(node, ctx))
  97                         return node;
  98         }
  99         return NULL;
 100 }
 101
 102 int
 103 mlx5_hlist_insert_ex(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry,
 104                      mlx5_hlist_match_callback_fn cb, void *ctx)
 105 {
 106         uint32_t idx;
 107         struct mlx5_hlist_head *first;
 108         struct mlx5_hlist_entry *node;
 109
 110         MLX5_ASSERT(h && entry && cb && ctx);
 111         idx = rte_hash_crc_8byte(entry->key, 0) & h->mask;
 112         first = &h->heads[idx];
 113         /* No need to reuse the lookup function. */
 114         LIST_FOREACH(node, first, next) {
 115                 if (!cb(node, ctx))
 116                         return -EEXIST;
 117         }
 118         LIST_INSERT_HEAD(first, entry, next);
 119         return 0;
 120 }
 121
 122 void
 123 mlx5_hlist_remove(struct mlx5_hlist *h __rte_unused,
 124                   struct mlx5_hlist_entry *entry)
 125 {
 126         MLX5_ASSERT(entry && entry->next.le_prev);
 127         LIST_REMOVE(entry, next);
 128         /* Set to NULL to get rid of removing action for more than once. */
 129         entry->next.le_prev = NULL;
 130 }
 131
 132 void
 133 mlx5_hlist_destroy(struct mlx5_hlist *h,
 134                    mlx5_hlist_destroy_callback_fn cb, void *ctx)
 135 {
 136         uint32_t idx;
 137         struct mlx5_hlist_entry *entry;
 138
 139         MLX5_ASSERT(h);
 140         for (idx = 0; idx < h->table_sz; ++idx) {
 141                 /* no LIST_FOREACH_SAFE, using while instead */
 142                 while (!LIST_EMPTY(&h->heads[idx])) {
 143                         entry = LIST_FIRST(&h->heads[idx]);
 144                         LIST_REMOVE(entry, next);
 145                         /*
 146                          * The owner of whole element which contains data entry
 147                          * is the user, so it's the user's duty to do the clean
 148                          * up and the free work because someone may not put the
 149                          * hlist entry at the beginning(suggested to locate at
 150                          * the beginning). Or else the default free function
 151                          * will be used.
 152                          */
 153                         if (cb)
 154                                 cb(entry, ctx);
 155                         else
 156                                 mlx5_free(entry);
 157                 }
 158         }
 159         mlx5_free(h);
 160 }
 161
 162 static inline void
 163 mlx5_ipool_lock(struct mlx5_indexed_pool *pool)
 164 {
 165         if (pool->cfg.need_lock)
 166                 rte_spinlock_lock(&pool->lock);
 167 }
 168
 169 static inline void
 170 mlx5_ipool_unlock(struct mlx5_indexed_pool *pool)
 171 {
 172         if (pool->cfg.need_lock)
 173                 rte_spinlock_unlock(&pool->lock);
 174 }
 175
 176 static inline uint32_t
 177 mlx5_trunk_idx_get(struct mlx5_indexed_pool *pool, uint32_t entry_idx)
 178 {
 179         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 180         uint32_t trunk_idx = 0;
 181         uint32_t i;
 182
 183         if (!cfg->grow_trunk)
 184                 return entry_idx / cfg->trunk_size;
 185         if (entry_idx >= pool->grow_tbl[cfg->grow_trunk - 1]) {
 186                 trunk_idx = (entry_idx - pool->grow_tbl[cfg->grow_trunk - 1]) /
 187                             (cfg->trunk_size << (cfg->grow_shift *
 188                             cfg->grow_trunk)) + cfg->grow_trunk;
 189         } else {
 190                 for (i = 0; i < cfg->grow_trunk; i++) {
 191                         if (entry_idx < pool->grow_tbl[i])
 192                                 break;
 193                 }
 194                 trunk_idx = i;
 195         }
 196         return trunk_idx;
 197 }
 198
 199 static inline uint32_t
 200 mlx5_trunk_size_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 201 {
 202         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 203
 204         return cfg->trunk_size << (cfg->grow_shift *
 205                (trunk_idx > cfg->grow_trunk ? cfg->grow_trunk : trunk_idx));
 206 }
 207
 208 static inline uint32_t
 209 mlx5_trunk_idx_offset_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 210 {
 211         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 212         uint32_t offset = 0;
 213
 214         if (!trunk_idx)
 215                 return 0;
 216         if (!cfg->grow_trunk)
 217                 return cfg->trunk_size * trunk_idx;
 218         if (trunk_idx < cfg->grow_trunk)
 219                 offset = pool->grow_tbl[trunk_idx - 1];
 220         else
 221                 offset = pool->grow_tbl[cfg->grow_trunk - 1] +
 222                          (cfg->trunk_size << (cfg->grow_shift *
 223                          cfg->grow_trunk)) * (trunk_idx - cfg->grow_trunk);
 224         return offset;
 225 }
 226
 227 struct mlx5_indexed_pool *
 228 mlx5_ipool_create(struct mlx5_indexed_pool_config *cfg)
 229 {
 230         struct mlx5_indexed_pool *pool;
 231         uint32_t i;
 232
 233         if (!cfg || !cfg->size || (!cfg->malloc ^ !cfg->free) ||
 234             (cfg->trunk_size && ((cfg->trunk_size & (cfg->trunk_size - 1)) ||
 235             ((__builtin_ffs(cfg->trunk_size) + TRUNK_IDX_BITS) > 32))))
 236                 return NULL;
 237         pool = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pool) + cfg->grow_trunk *
 238                            sizeof(pool->grow_tbl[0]), RTE_CACHE_LINE_SIZE,
 239                            SOCKET_ID_ANY);
 240         if (!pool)
 241                 return NULL;
 242         pool->cfg = *cfg;
 243         if (!pool->cfg.trunk_size)
 244                 pool->cfg.trunk_size = MLX5_IPOOL_DEFAULT_TRUNK_SIZE;
 245         if (!cfg->malloc && !cfg->free) {
 246                 pool->cfg.malloc = mlx5_malloc;
 247                 pool->cfg.free = mlx5_free;
 248         }
 249         pool->free_list = TRUNK_INVALID;
 250         if (pool->cfg.need_lock)
 251                 rte_spinlock_init(&pool->lock);
 252         /*
 253          * Initialize the dynamic grow trunk size lookup table to have a quick
 254          * lookup for the trunk entry index offset.
 255          */
 256         for (i = 0; i < cfg->grow_trunk; i++) {
 257                 pool->grow_tbl[i] = cfg->trunk_size << (cfg->grow_shift * i);
 258                 if (i > 0)
 259                         pool->grow_tbl[i] += pool->grow_tbl[i - 1];
 260         }
 261         return pool;
 262 }
 263
 264 static int
 265 mlx5_ipool_grow(struct mlx5_indexed_pool *pool)
 266 {
 267         struct mlx5_indexed_trunk *trunk;
 268         struct mlx5_indexed_trunk **trunk_tmp;
 269         struct mlx5_indexed_trunk **p;
 270         size_t trunk_size = 0;
 271         size_t data_size;
 272         size_t bmp_size;
 273         uint32_t idx;
 274
 275         if (pool->n_trunk_valid == TRUNK_MAX_IDX)
 276                 return -ENOMEM;
 277         if (pool->n_trunk_valid == pool->n_trunk) {
 278                 /* No free trunk flags, expand trunk list. */
 279                 int n_grow = pool->n_trunk_valid ? pool->n_trunk :
 280                              RTE_CACHE_LINE_SIZE / sizeof(void *);
 281
 282                 p = pool->cfg.malloc(0, (pool->n_trunk_valid + n_grow) *
 283                                      sizeof(struct mlx5_indexed_trunk *),
 284                                      RTE_CACHE_LINE_SIZE, rte_socket_id());
 285                 if (!p)
 286                         return -ENOMEM;
 287                 if (pool->trunks)
 288                         memcpy(p, pool->trunks, pool->n_trunk_valid *
 289                                sizeof(struct mlx5_indexed_trunk *));
 290                 memset(RTE_PTR_ADD(p, pool->n_trunk_valid * sizeof(void *)), 0,
 291                        n_grow * sizeof(void *));
 292                 trunk_tmp = pool->trunks;
 293                 pool->trunks = p;
 294                 if (trunk_tmp)
 295                         pool->cfg.free(trunk_tmp);
 296                 pool->n_trunk += n_grow;
 297         }
 298         if (!pool->cfg.release_mem_en) {
 299                 idx = pool->n_trunk_valid;
 300         } else {
 301                 /* Find the first available slot in trunk list */
 302                 for (idx = 0; idx < pool->n_trunk; idx++)
 303                         if (pool->trunks[idx] == NULL)
 304                                 break;
 305         }
 306         trunk_size += sizeof(*trunk);
 307         data_size = mlx5_trunk_size_get(pool, idx);
 308         bmp_size = rte_bitmap_get_memory_footprint(data_size);
 309         /* rte_bitmap requires memory cacheline aligned. */
 310         trunk_size += RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size);
 311         trunk_size += bmp_size;
 312         trunk = pool->cfg.malloc(0, trunk_size,
 313                                  RTE_CACHE_LINE_SIZE, rte_socket_id());
 314         if (!trunk)
 315                 return -ENOMEM;
 316         pool->trunks[idx] = trunk;
 317         trunk->idx = idx;
 318         trunk->free = data_size;
 319         trunk->prev = TRUNK_INVALID;
 320         trunk->next = TRUNK_INVALID;
 321         MLX5_ASSERT(pool->free_list == TRUNK_INVALID);
 322         pool->free_list = idx;
 323         /* Mark all entries as available. */
 324         trunk->bmp = rte_bitmap_init_with_all_set(data_size, &trunk->data
 325                      [RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size)],
 326                      bmp_size);
 327         MLX5_ASSERT(trunk->bmp);
 328         pool->n_trunk_valid++;
 329 #ifdef POOL_DEBUG
 330         pool->trunk_new++;
 331         pool->trunk_avail++;
 332 #endif
 333         return 0;
 334 }
 335
 336 void *
 337 mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 338 {
 339         struct mlx5_indexed_trunk *trunk;
 340         uint64_t slab = 0;
 341         uint32_t iidx = 0;
 342         void *p;
 343
 344         mlx5_ipool_lock(pool);
 345         if (pool->free_list == TRUNK_INVALID) {
 346                 /* If no available trunks, grow new. */
 347                 if (mlx5_ipool_grow(pool)) {
 348                         mlx5_ipool_unlock(pool);
 349                         return NULL;
 350                 }
 351         }
 352         MLX5_ASSERT(pool->free_list != TRUNK_INVALID);
 353         trunk = pool->trunks[pool->free_list];
 354         MLX5_ASSERT(trunk->free);
 355         if (!rte_bitmap_scan(trunk->bmp, &iidx, &slab)) {
 356                 mlx5_ipool_unlock(pool);
 357                 return NULL;
 358         }
 359         MLX5_ASSERT(slab);
 360         iidx += __builtin_ctzll(slab);
 361         MLX5_ASSERT(iidx != UINT32_MAX);
 362         MLX5_ASSERT(iidx < mlx5_trunk_size_get(pool, trunk->idx));
 363         rte_bitmap_clear(trunk->bmp, iidx);
 364         p = &trunk->data[iidx * pool->cfg.size];
 365         iidx += mlx5_trunk_idx_offset_get(pool, trunk->idx);
 366         iidx += 1; /* non-zero index. */
 367         trunk->free--;
 368 #ifdef POOL_DEBUG
 369         pool->n_entry++;
 370 #endif
 371         if (!trunk->free) {
 372                 /* Full trunk will be removed from free list in imalloc. */
 373                 MLX5_ASSERT(pool->free_list == trunk->idx);
 374                 pool->free_list = trunk->next;
 375                 if (trunk->next != TRUNK_INVALID)
 376                         pool->trunks[trunk->next]->prev = TRUNK_INVALID;
 377                 trunk->prev = TRUNK_INVALID;
 378                 trunk->next = TRUNK_INVALID;
 379 #ifdef POOL_DEBUG
 380                 pool->trunk_empty++;
 381                 pool->trunk_avail--;
 382 #endif
 383         }
 384         *idx = iidx;
 385         mlx5_ipool_unlock(pool);
 386         return p;
 387 }
 388
 389 void *
 390 mlx5_ipool_zmalloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 391 {
 392         void *entry = mlx5_ipool_malloc(pool, idx);
 393
 394         if (entry)
 395                 memset(entry, 0, pool->cfg.size);
 396         return entry;
 397 }
 398
 399 void
 400 mlx5_ipool_free(struct mlx5_indexed_pool *pool, uint32_t idx)
 401 {
 402         struct mlx5_indexed_trunk *trunk;
 403         uint32_t trunk_idx;
 404         uint32_t entry_idx;
 405
 406         if (!idx)
 407                 return;
 408         idx -= 1;
 409         mlx5_ipool_lock(pool);
 410         trunk_idx = mlx5_trunk_idx_get(pool, idx);
 411         if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
 412             (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
 413                 goto out;
 414         trunk = pool->trunks[trunk_idx];
 415         if (!trunk)
 416                 goto out;
 417         entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
 418         if (trunk_idx != trunk->idx ||
 419             rte_bitmap_get(trunk->bmp, entry_idx))
 420                 goto out;
 421         rte_bitmap_set(trunk->bmp, entry_idx);
 422         trunk->free++;
 423         if (pool->cfg.release_mem_en && trunk->free == mlx5_trunk_size_get
 424            (pool, trunk->idx)) {
 425                 if (pool->free_list == trunk->idx)
 426                         pool->free_list = trunk->next;
 427                 if (trunk->next != TRUNK_INVALID)
 428                         pool->trunks[trunk->next]->prev = trunk->prev;
 429                 if (trunk->prev != TRUNK_INVALID)
 430                         pool->trunks[trunk->prev]->next = trunk->next;
 431                 pool->cfg.free(trunk);
 432                 pool->trunks[trunk_idx] = NULL;
 433                 pool->n_trunk_valid--;
 434 #ifdef POOL_DEBUG
 435                 pool->trunk_avail--;
 436                 pool->trunk_free++;
 437 #endif
 438                 if (pool->n_trunk_valid == 0) {
 439                         pool->cfg.free(pool->trunks);
 440                         pool->trunks = NULL;
 441                         pool->n_trunk = 0;
 442                 }
 443         } else if (trunk->free == 1) {
 444                 /* Put into free trunk list head. */
 445                 MLX5_ASSERT(pool->free_list != trunk->idx);
 446                 trunk->next = pool->free_list;
 447                 trunk->prev = TRUNK_INVALID;
 448                 if (pool->free_list != TRUNK_INVALID)
 449                         pool->trunks[pool->free_list]->prev = trunk->idx;
 450                 pool->free_list = trunk->idx;
 451 #ifdef POOL_DEBUG
 452                 pool->trunk_empty--;
 453                 pool->trunk_avail++;
 454 #endif
 455         }
 456 #ifdef POOL_DEBUG
 457         pool->n_entry--;
 458 #endif
 459 out:
 460         mlx5_ipool_unlock(pool);
 461 }
 462
 463 void *
 464 mlx5_ipool_get(struct mlx5_indexed_pool *pool, uint32_t idx)
 465 {
 466         struct mlx5_indexed_trunk *trunk;
 467         void *p = NULL;
 468         uint32_t trunk_idx;
 469         uint32_t entry_idx;
 470
 471         if (!idx)
 472                 return NULL;
 473         idx -= 1;
 474         mlx5_ipool_lock(pool);
 475         trunk_idx = mlx5_trunk_idx_get(pool, idx);
 476         if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
 477             (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
 478                 goto out;
 479         trunk = pool->trunks[trunk_idx];
 480         if (!trunk)
 481                 goto out;
 482         entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
 483         if (trunk_idx != trunk->idx ||
 484             rte_bitmap_get(trunk->bmp, entry_idx))
 485                 goto out;
 486         p = &trunk->data[entry_idx * pool->cfg.size];
 487 out:
 488         mlx5_ipool_unlock(pool);
 489         return p;
 490 }
 491
 492 int
 493 mlx5_ipool_destroy(struct mlx5_indexed_pool *pool)
 494 {
 495         struct mlx5_indexed_trunk **trunks;
 496         uint32_t i;
 497
 498         MLX5_ASSERT(pool);
 499         mlx5_ipool_lock(pool);
 500         trunks = pool->trunks;
 501         for (i = 0; i < pool->n_trunk; i++) {
 502                 if (trunks[i])
 503                         pool->cfg.free(trunks[i]);
 504         }
 505         if (!pool->trunks)
 506                 pool->cfg.free(pool->trunks);
 507         mlx5_ipool_unlock(pool);
 508         mlx5_free(pool);
 509         return 0;
 510 }
 511
 512 void
 513 mlx5_ipool_dump(struct mlx5_indexed_pool *pool)
 514 {
 515         printf("Pool %s entry size %u, trunks %u, %d entry per trunk, "
 516                "total: %d\n",
 517                pool->cfg.type, pool->cfg.size, pool->n_trunk_valid,
 518                pool->cfg.trunk_size, pool->n_trunk_valid);
 519 #ifdef POOL_DEBUG
 520         printf("Pool %s entry %u, trunk alloc %u, empty: %u, "
 521                "available %u free %u\n",
 522                pool->cfg.type, pool->n_entry, pool->trunk_new,
 523                pool->trunk_empty, pool->trunk_avail, pool->trunk_free);
 524 #endif
 525 }
 526
 527 struct mlx5_l3t_tbl *
 528 mlx5_l3t_create(enum mlx5_l3t_type type)
 529 {
 530         struct mlx5_l3t_tbl *tbl;
 531         struct mlx5_indexed_pool_config l3t_ip_cfg = {
 532                 .trunk_size = 16,
 533                 .grow_trunk = 6,
 534                 .grow_shift = 1,
 535                 .need_lock = 0,
 536                 .release_mem_en = 1,
 537                 .malloc = mlx5_malloc,
 538                 .free = mlx5_free,
 539         };
 540
 541         if (type >= MLX5_L3T_TYPE_MAX) {
 542                 rte_errno = EINVAL;
 543                 return NULL;
 544         }
 545         tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(struct mlx5_l3t_tbl), 1,
 546                           SOCKET_ID_ANY);
 547         if (!tbl) {
 548                 rte_errno = ENOMEM;
 549                 return NULL;
 550         }
 551         tbl->type = type;
 552         switch (type) {
 553         case MLX5_L3T_TYPE_WORD:
 554                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_word) +
 555                                   sizeof(uint16_t) * MLX5_L3T_ET_SIZE;
 556                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_w";
 557                 break;
 558         case MLX5_L3T_TYPE_DWORD:
 559                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_dword) +
 560                                   sizeof(uint32_t) * MLX5_L3T_ET_SIZE;
 561                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_dw";
 562                 break;
 563         case MLX5_L3T_TYPE_QWORD:
 564                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_qword) +
 565                                   sizeof(uint64_t) * MLX5_L3T_ET_SIZE;
 566                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_qw";
 567                 break;
 568         default:
 569                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_ptr) +
 570                                   sizeof(void *) * MLX5_L3T_ET_SIZE;
 571                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_tpr";
 572                 break;
 573         }
 574         tbl->eip = mlx5_ipool_create(&l3t_ip_cfg);
 575         if (!tbl->eip) {
 576                 rte_errno = ENOMEM;
 577                 mlx5_free(tbl);
 578                 tbl = NULL;
 579         }
 580         return tbl;
 581 }
 582
 583 void
 584 mlx5_l3t_destroy(struct mlx5_l3t_tbl *tbl)
 585 {
 586         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 587         uint32_t i, j;
 588
 589         if (!tbl)
 590                 return;
 591         g_tbl = tbl->tbl;
 592         if (g_tbl) {
 593                 for (i = 0; i < MLX5_L3T_GT_SIZE; i++) {
 594                         m_tbl = g_tbl->tbl[i];
 595                         if (!m_tbl)
 596                                 continue;
 597                         for (j = 0; j < MLX5_L3T_MT_SIZE; j++) {
 598                                 if (!m_tbl->tbl[j])
 599                                         continue;
 600                                 MLX5_ASSERT(!((struct mlx5_l3t_entry_word *)
 601                                             m_tbl->tbl[j])->ref_cnt);
 602                                 mlx5_ipool_free(tbl->eip,
 603                                                 ((struct mlx5_l3t_entry_word *)
 604                                                 m_tbl->tbl[j])->idx);
 605                                 m_tbl->tbl[j] = 0;
 606                                 if (!(--m_tbl->ref_cnt))
 607                                         break;
 608                         }
 609                         MLX5_ASSERT(!m_tbl->ref_cnt);
 610                         mlx5_free(g_tbl->tbl[i]);
 611                         g_tbl->tbl[i] = 0;
 612                         if (!(--g_tbl->ref_cnt))
 613                                 break;
 614                 }
 615                 MLX5_ASSERT(!g_tbl->ref_cnt);
 616                 mlx5_free(tbl->tbl);
 617                 tbl->tbl = 0;
 618         }
 619         mlx5_ipool_destroy(tbl->eip);
 620         mlx5_free(tbl);
 621 }
 622
 623 uint32_t
 624 mlx5_l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
 625                    union mlx5_l3t_data *data)
 626 {
 627         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 628         void *e_tbl;
 629         uint32_t entry_idx;
 630
 631         g_tbl = tbl->tbl;
 632         if (!g_tbl)
 633                 return -1;
 634         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
 635         if (!m_tbl)
 636                 return -1;
 637         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
 638         if (!e_tbl)
 639                 return -1;
 640         entry_idx = idx & MLX5_L3T_ET_MASK;
 641         switch (tbl->type) {
 642         case MLX5_L3T_TYPE_WORD:
 643                 data->word = ((struct mlx5_l3t_entry_word *)e_tbl)->entry
 644                              [entry_idx];
 645                 break;
 646         case MLX5_L3T_TYPE_DWORD:
 647                 data->dword = ((struct mlx5_l3t_entry_dword *)e_tbl)->entry
 648                              [entry_idx];
 649                 break;
 650         case MLX5_L3T_TYPE_QWORD:
 651                 data->qword = ((struct mlx5_l3t_entry_qword *)e_tbl)->entry
 652                               [entry_idx];
 653                 break;
 654         default:
 655                 data->ptr = ((struct mlx5_l3t_entry_ptr *)e_tbl)->entry
 656                             [entry_idx];
 657                 break;
 658         }
 659         return 0;
 660 }
 661
 662 void
 663 mlx5_l3t_clear_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx)
 664 {
 665         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 666         struct mlx5_l3t_entry_word *w_e_tbl;
 667         struct mlx5_l3t_entry_dword *dw_e_tbl;
 668         struct mlx5_l3t_entry_qword *qw_e_tbl;
 669         struct mlx5_l3t_entry_ptr *ptr_e_tbl;
 670         void *e_tbl;
 671         uint32_t entry_idx;
 672         uint64_t ref_cnt;
 673
 674         g_tbl = tbl->tbl;
 675         if (!g_tbl)
 676                 return;
 677         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
 678         if (!m_tbl)
 679                 return;
 680         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
 681         if (!e_tbl)
 682                 return;
 683         entry_idx = idx & MLX5_L3T_ET_MASK;
 684         switch (tbl->type) {
 685         case MLX5_L3T_TYPE_WORD:
 686                 w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
 687                 w_e_tbl->entry[entry_idx] = 0;
 688                 ref_cnt = --w_e_tbl->ref_cnt;
 689                 break;
 690         case MLX5_L3T_TYPE_DWORD:
 691                 dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
 692                 dw_e_tbl->entry[entry_idx] = 0;
 693                 ref_cnt = --dw_e_tbl->ref_cnt;
 694                 break;
 695         case MLX5_L3T_TYPE_QWORD:
 696                 qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
 697                 qw_e_tbl->entry[entry_idx] = 0;
 698                 ref_cnt = --qw_e_tbl->ref_cnt;
 699                 break;
 700         default:
 701                 ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
 702                 ptr_e_tbl->entry[entry_idx] = NULL;
 703                 ref_cnt = --ptr_e_tbl->ref_cnt;
 704                 break;
 705         }
 706         if (!ref_cnt) {
 707                 mlx5_ipool_free(tbl->eip,
 708                                 ((struct mlx5_l3t_entry_word *)e_tbl)->idx);
 709                 m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
 710                                                                         NULL;
 711                 if (!(--m_tbl->ref_cnt)) {
 712                         mlx5_free(m_tbl);
 713                         g_tbl->tbl
 714                         [(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] = NULL;
 715                         if (!(--g_tbl->ref_cnt)) {
 716                                 mlx5_free(g_tbl);
 717                                 tbl->tbl = 0;
 718                         }
 719                 }
 720         }
 721 }
 722
 723 uint32_t
 724 mlx5_l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
 725                    union mlx5_l3t_data *data)
 726 {
 727         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 728         struct mlx5_l3t_entry_word *w_e_tbl;
 729         struct mlx5_l3t_entry_dword *dw_e_tbl;
 730         struct mlx5_l3t_entry_qword *qw_e_tbl;
 731         struct mlx5_l3t_entry_ptr *ptr_e_tbl;
 732         void *e_tbl;
 733         uint32_t entry_idx, tbl_idx = 0;
 734
 735         /* Check the global table, create it if empty. */
 736         g_tbl = tbl->tbl;
 737         if (!g_tbl) {
 738                 g_tbl = mlx5_malloc(MLX5_MEM_ZERO,
 739                                     sizeof(struct mlx5_l3t_level_tbl) +
 740                                     sizeof(void *) * MLX5_L3T_GT_SIZE, 1,
 741                                     SOCKET_ID_ANY);
 742                 if (!g_tbl) {
 743                         rte_errno = ENOMEM;
 744                         return -1;
 745                 }
 746                 tbl->tbl = g_tbl;
 747         }
 748         /*
 749          * Check the middle table, create it if empty. Ref_cnt will be
 750          * increased if new sub table created.
 751          */
 752         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
 753         if (!m_tbl) {
 754                 m_tbl = mlx5_malloc(MLX5_MEM_ZERO,
 755                                     sizeof(struct mlx5_l3t_level_tbl) +
 756                                     sizeof(void *) * MLX5_L3T_MT_SIZE, 1,
 757                                     SOCKET_ID_ANY);
 758                 if (!m_tbl) {
 759                         rte_errno = ENOMEM;
 760                         return -1;
 761                 }
 762                 g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] =
 763                                                                         m_tbl;
 764                 g_tbl->ref_cnt++;
 765         }
 766         /*
 767          * Check the entry table, create it if empty. Ref_cnt will be
 768          * increased if new sub entry table created.
 769          */
 770         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
 771         if (!e_tbl) {
 772                 e_tbl = mlx5_ipool_zmalloc(tbl->eip, &tbl_idx);
 773                 if (!e_tbl) {
 774                         rte_errno = ENOMEM;
 775                         return -1;
 776                 }
 777                 ((struct mlx5_l3t_entry_word *)e_tbl)->idx = tbl_idx;
 778                 m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
 779                                                                         e_tbl;
 780                 m_tbl->ref_cnt++;
 781         }
 782         entry_idx = idx & MLX5_L3T_ET_MASK;
 783         switch (tbl->type) {
 784         case MLX5_L3T_TYPE_WORD:
 785                 w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
 786                 w_e_tbl->entry[entry_idx] = data->word;
 787                 w_e_tbl->ref_cnt++;
 788                 break;
 789         case MLX5_L3T_TYPE_DWORD:
 790                 dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
 791                 dw_e_tbl->entry[entry_idx] = data->dword;
 792                 dw_e_tbl->ref_cnt++;
 793                 break;
 794         case MLX5_L3T_TYPE_QWORD:
 795                 qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
 796                 qw_e_tbl->entry[entry_idx] = data->qword;
 797                 qw_e_tbl->ref_cnt++;
 798                 break;
 799         default:
 800                 ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
 801                 ptr_e_tbl->entry[entry_idx] = data->ptr;
 802                 ptr_e_tbl->ref_cnt++;
 803                 break;
 804         }
 805         return 0;
 806 }