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
   7 #include <mlx5_malloc.h>
   8
   9 #include "mlx5_utils.h"
  10
  11
  12 /********************* Cache list ************************/
  13
  14 static struct mlx5_cache_entry *
  15 mlx5_clist_default_create_cb(struct mlx5_cache_list *list,
  16                              struct mlx5_cache_entry *entry __rte_unused,
  17                              void *ctx __rte_unused)
  18 {
  19         return mlx5_malloc(MLX5_MEM_ZERO, list->entry_sz, 0, SOCKET_ID_ANY);
  20 }
  21
  22 static void
  23 mlx5_clist_default_remove_cb(struct mlx5_cache_list *list __rte_unused,
  24                              struct mlx5_cache_entry *entry)
  25 {
  26         mlx5_free(entry);
  27 }
  28
  29 int
  30 mlx5_cache_list_init(struct mlx5_cache_list *list, const char *name,
  31                      uint32_t entry_size, void *ctx,
  32                      mlx5_cache_create_cb cb_create,
  33                      mlx5_cache_match_cb cb_match,
  34                      mlx5_cache_remove_cb cb_remove)
  35 {
  36         MLX5_ASSERT(list);
  37         if (!cb_match || (!cb_create ^ !cb_remove))
  38                 return -1;
  39         if (name)
  40                 snprintf(list->name, sizeof(list->name), "%s", name);
  41         list->entry_sz = entry_size;
  42         list->ctx = ctx;
  43         list->cb_create = cb_create ? cb_create : mlx5_clist_default_create_cb;
  44         list->cb_match = cb_match;
  45         list->cb_remove = cb_remove ? cb_remove : mlx5_clist_default_remove_cb;
  46         rte_rwlock_init(&list->lock);
  47         DRV_LOG(DEBUG, "Cache list %s initialized.", list->name);
  48         LIST_INIT(&list->head);
  49         return 0;
  50 }
  51
  52 static struct mlx5_cache_entry *
  53 __cache_lookup(struct mlx5_cache_list *list, void *ctx, bool reuse)
  54 {
  55         struct mlx5_cache_entry *entry;
  56
  57         LIST_FOREACH(entry, &list->head, next) {
  58                 if (list->cb_match(list, entry, ctx))
  59                         continue;
  60                 if (reuse) {
  61                         __atomic_add_fetch(&entry->ref_cnt, 1,
  62                                            __ATOMIC_RELAXED);
  63                         DRV_LOG(DEBUG, "Cache list %s entry %p ref++: %u.",
  64                                 list->name, (void *)entry, entry->ref_cnt);
  65                 }
  66                 break;
  67         }
  68         return entry;
  69 }
  70
  71 static struct mlx5_cache_entry *
  72 cache_lookup(struct mlx5_cache_list *list, void *ctx, bool reuse)
  73 {
  74         struct mlx5_cache_entry *entry;
  75
  76         rte_rwlock_read_lock(&list->lock);
  77         entry = __cache_lookup(list, ctx, reuse);
  78         rte_rwlock_read_unlock(&list->lock);
  79         return entry;
  80 }
  81
  82 struct mlx5_cache_entry *
  83 mlx5_cache_lookup(struct mlx5_cache_list *list, void *ctx)
  84 {
  85         return cache_lookup(list, ctx, false);
  86 }
  87
  88 struct mlx5_cache_entry *
  89 mlx5_cache_register(struct mlx5_cache_list *list, void *ctx)
  90 {
  91         struct mlx5_cache_entry *entry;
  92         uint32_t prev_gen_cnt = 0;
  93
  94         MLX5_ASSERT(list);
  95         prev_gen_cnt = __atomic_load_n(&list->gen_cnt, __ATOMIC_ACQUIRE);
  96         /* Lookup with read lock, reuse if found. */
  97         entry = cache_lookup(list, ctx, true);
  98         if (entry)
  99                 return entry;
 100         /* Not found, append with write lock - block read from other threads. */
 101         rte_rwlock_write_lock(&list->lock);
 102         /* If list changed by other threads before lock, search again. */
 103         if (prev_gen_cnt != __atomic_load_n(&list->gen_cnt, __ATOMIC_ACQUIRE)) {
 104                 /* Lookup and reuse w/o read lock. */
 105                 entry = __cache_lookup(list, ctx, true);
 106                 if (entry)
 107                         goto done;
 108         }
 109         entry = list->cb_create(list, entry, ctx);
 110         if (!entry) {
 111                 DRV_LOG(ERR, "Failed to init cache list %s entry %p.",
 112                         list->name, (void *)entry);
 113                 goto done;
 114         }
 115         entry->ref_cnt = 1;
 116         LIST_INSERT_HEAD(&list->head, entry, next);
 117         __atomic_add_fetch(&list->gen_cnt, 1, __ATOMIC_RELEASE);
 118         __atomic_add_fetch(&list->count, 1, __ATOMIC_ACQUIRE);
 119         DRV_LOG(DEBUG, "Cache list %s entry %p new: %u.",
 120                 list->name, (void *)entry, entry->ref_cnt);
 121 done:
 122         rte_rwlock_write_unlock(&list->lock);
 123         return entry;
 124 }
 125
 126 int
 127 mlx5_cache_unregister(struct mlx5_cache_list *list,
 128                       struct mlx5_cache_entry *entry)
 129 {
 130         rte_rwlock_write_lock(&list->lock);
 131         MLX5_ASSERT(entry && entry->next.le_prev);
 132         DRV_LOG(DEBUG, "Cache list %s entry %p ref--: %u.",
 133                 list->name, (void *)entry, entry->ref_cnt);
 134         if (--entry->ref_cnt) {
 135                 rte_rwlock_write_unlock(&list->lock);
 136                 return 1;
 137         }
 138         __atomic_add_fetch(&list->gen_cnt, 1, __ATOMIC_ACQUIRE);
 139         __atomic_sub_fetch(&list->count, 1, __ATOMIC_ACQUIRE);
 140         LIST_REMOVE(entry, next);
 141         list->cb_remove(list, entry);
 142         rte_rwlock_write_unlock(&list->lock);
 143         DRV_LOG(DEBUG, "Cache list %s entry %p removed.",
 144                 list->name, (void *)entry);
 145         return 0;
 146 }
 147
 148 void
 149 mlx5_cache_list_destroy(struct mlx5_cache_list *list)
 150 {
 151         struct mlx5_cache_entry *entry;
 152
 153         MLX5_ASSERT(list);
 154         /* no LIST_FOREACH_SAFE, using while instead */
 155         while (!LIST_EMPTY(&list->head)) {
 156                 entry = LIST_FIRST(&list->head);
 157                 LIST_REMOVE(entry, next);
 158                 list->cb_remove(list, entry);
 159                 DRV_LOG(DEBUG, "Cache list %s entry %p destroyed.",
 160                         list->name, (void *)entry);
 161         }
 162         memset(list, 0, sizeof(*list));
 163 }
 164
 165 uint32_t
 166 mlx5_cache_list_get_entry_num(struct mlx5_cache_list *list)
 167 {
 168         MLX5_ASSERT(list);
 169         return __atomic_load_n(&list->count, __ATOMIC_RELAXED);
 170 }
 171
 172 /********************* Indexed pool **********************/
 173
 174 static inline void
 175 mlx5_ipool_lock(struct mlx5_indexed_pool *pool)
 176 {
 177         if (pool->cfg.need_lock)
 178                 rte_spinlock_lock(&pool->rsz_lock);
 179 }
 180
 181 static inline void
 182 mlx5_ipool_unlock(struct mlx5_indexed_pool *pool)
 183 {
 184         if (pool->cfg.need_lock)
 185                 rte_spinlock_unlock(&pool->rsz_lock);
 186 }
 187
 188 static inline uint32_t
 189 mlx5_trunk_idx_get(struct mlx5_indexed_pool *pool, uint32_t entry_idx)
 190 {
 191         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 192         uint32_t trunk_idx = 0;
 193         uint32_t i;
 194
 195         if (!cfg->grow_trunk)
 196                 return entry_idx / cfg->trunk_size;
 197         if (entry_idx >= pool->grow_tbl[cfg->grow_trunk - 1]) {
 198                 trunk_idx = (entry_idx - pool->grow_tbl[cfg->grow_trunk - 1]) /
 199                             (cfg->trunk_size << (cfg->grow_shift *
 200                             cfg->grow_trunk)) + cfg->grow_trunk;
 201         } else {
 202                 for (i = 0; i < cfg->grow_trunk; i++) {
 203                         if (entry_idx < pool->grow_tbl[i])
 204                                 break;
 205                 }
 206                 trunk_idx = i;
 207         }
 208         return trunk_idx;
 209 }
 210
 211 static inline uint32_t
 212 mlx5_trunk_size_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 213 {
 214         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 215
 216         return cfg->trunk_size << (cfg->grow_shift *
 217                (trunk_idx > cfg->grow_trunk ? cfg->grow_trunk : trunk_idx));
 218 }
 219
 220 static inline uint32_t
 221 mlx5_trunk_idx_offset_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 222 {
 223         struct mlx5_indexed_pool_config *cfg = &pool->cfg;
 224         uint32_t offset = 0;
 225
 226         if (!trunk_idx)
 227                 return 0;
 228         if (!cfg->grow_trunk)
 229                 return cfg->trunk_size * trunk_idx;
 230         if (trunk_idx < cfg->grow_trunk)
 231                 offset = pool->grow_tbl[trunk_idx - 1];
 232         else
 233                 offset = pool->grow_tbl[cfg->grow_trunk - 1] +
 234                          (cfg->trunk_size << (cfg->grow_shift *
 235                          cfg->grow_trunk)) * (trunk_idx - cfg->grow_trunk);
 236         return offset;
 237 }
 238
 239 struct mlx5_indexed_pool *
 240 mlx5_ipool_create(struct mlx5_indexed_pool_config *cfg)
 241 {
 242         struct mlx5_indexed_pool *pool;
 243         uint32_t i;
 244
 245         if (!cfg || (!cfg->malloc ^ !cfg->free) ||
 246             (cfg->per_core_cache && cfg->release_mem_en) ||
 247             (cfg->trunk_size && ((cfg->trunk_size & (cfg->trunk_size - 1)) ||
 248             ((__builtin_ffs(cfg->trunk_size) + TRUNK_IDX_BITS) > 32))))
 249                 return NULL;
 250         pool = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pool) + cfg->grow_trunk *
 251                            sizeof(pool->grow_tbl[0]), RTE_CACHE_LINE_SIZE,
 252                            SOCKET_ID_ANY);
 253         if (!pool)
 254                 return NULL;
 255         pool->cfg = *cfg;
 256         if (!pool->cfg.trunk_size)
 257                 pool->cfg.trunk_size = MLX5_IPOOL_DEFAULT_TRUNK_SIZE;
 258         if (!cfg->malloc && !cfg->free) {
 259                 pool->cfg.malloc = mlx5_malloc;
 260                 pool->cfg.free = mlx5_free;
 261         }
 262         if (pool->cfg.need_lock)
 263                 rte_spinlock_init(&pool->rsz_lock);
 264         /*
 265          * Initialize the dynamic grow trunk size lookup table to have a quick
 266          * lookup for the trunk entry index offset.
 267          */
 268         for (i = 0; i < cfg->grow_trunk; i++) {
 269                 pool->grow_tbl[i] = cfg->trunk_size << (cfg->grow_shift * i);
 270                 if (i > 0)
 271                         pool->grow_tbl[i] += pool->grow_tbl[i - 1];
 272         }
 273         if (!pool->cfg.max_idx)
 274                 pool->cfg.max_idx =
 275                         mlx5_trunk_idx_offset_get(pool, TRUNK_MAX_IDX + 1);
 276         if (!cfg->per_core_cache)
 277                 pool->free_list = TRUNK_INVALID;
 278         return pool;
 279 }
 280
 281 static int
 282 mlx5_ipool_grow(struct mlx5_indexed_pool *pool)
 283 {
 284         struct mlx5_indexed_trunk *trunk;
 285         struct mlx5_indexed_trunk **trunk_tmp;
 286         struct mlx5_indexed_trunk **p;
 287         size_t trunk_size = 0;
 288         size_t data_size;
 289         size_t bmp_size;
 290         uint32_t idx, cur_max_idx, i;
 291
 292         cur_max_idx = mlx5_trunk_idx_offset_get(pool, pool->n_trunk_valid);
 293         if (pool->n_trunk_valid == TRUNK_MAX_IDX ||
 294             cur_max_idx >= pool->cfg.max_idx)
 295                 return -ENOMEM;
 296         if (pool->n_trunk_valid == pool->n_trunk) {
 297                 /* No free trunk flags, expand trunk list. */
 298                 int n_grow = pool->n_trunk_valid ? pool->n_trunk :
 299                              RTE_CACHE_LINE_SIZE / sizeof(void *);
 300
 301                 p = pool->cfg.malloc(0, (pool->n_trunk_valid + n_grow) *
 302                                      sizeof(struct mlx5_indexed_trunk *),
 303                                      RTE_CACHE_LINE_SIZE, rte_socket_id());
 304                 if (!p)
 305                         return -ENOMEM;
 306                 if (pool->trunks)
 307                         memcpy(p, pool->trunks, pool->n_trunk_valid *
 308                                sizeof(struct mlx5_indexed_trunk *));
 309                 memset(RTE_PTR_ADD(p, pool->n_trunk_valid * sizeof(void *)), 0,
 310                        n_grow * sizeof(void *));
 311                 trunk_tmp = pool->trunks;
 312                 pool->trunks = p;
 313                 if (trunk_tmp)
 314                         pool->cfg.free(trunk_tmp);
 315                 pool->n_trunk += n_grow;
 316         }
 317         if (!pool->cfg.release_mem_en) {
 318                 idx = pool->n_trunk_valid;
 319         } else {
 320                 /* Find the first available slot in trunk list */
 321                 for (idx = 0; idx < pool->n_trunk; idx++)
 322                         if (pool->trunks[idx] == NULL)
 323                                 break;
 324         }
 325         trunk_size += sizeof(*trunk);
 326         data_size = mlx5_trunk_size_get(pool, idx);
 327         bmp_size = rte_bitmap_get_memory_footprint(data_size);
 328         /* rte_bitmap requires memory cacheline aligned. */
 329         trunk_size += RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size);
 330         trunk_size += bmp_size;
 331         trunk = pool->cfg.malloc(0, trunk_size,
 332                                  RTE_CACHE_LINE_SIZE, rte_socket_id());
 333         if (!trunk)
 334                 return -ENOMEM;
 335         pool->trunks[idx] = trunk;
 336         trunk->idx = idx;
 337         trunk->free = data_size;
 338         trunk->prev = TRUNK_INVALID;
 339         trunk->next = TRUNK_INVALID;
 340         MLX5_ASSERT(pool->free_list == TRUNK_INVALID);
 341         pool->free_list = idx;
 342         /* Mark all entries as available. */
 343         trunk->bmp = rte_bitmap_init_with_all_set(data_size, &trunk->data
 344                      [RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size)],
 345                      bmp_size);
 346         /* Clear the overhead bits in the trunk if it happens. */
 347         if (cur_max_idx + data_size > pool->cfg.max_idx) {
 348                 for (i = pool->cfg.max_idx - cur_max_idx; i < data_size; i++)
 349                         rte_bitmap_clear(trunk->bmp, i);
 350         }
 351         MLX5_ASSERT(trunk->bmp);
 352         pool->n_trunk_valid++;
 353 #ifdef POOL_DEBUG
 354         pool->trunk_new++;
 355         pool->trunk_avail++;
 356 #endif
 357         return 0;
 358 }
 359
 360 static inline struct mlx5_indexed_cache *
 361 mlx5_ipool_update_global_cache(struct mlx5_indexed_pool *pool, int cidx)
 362 {
 363         struct mlx5_indexed_cache *gc, *lc, *olc = NULL;
 364
 365         lc = pool->cache[cidx]->lc;
 366         gc = __atomic_load_n(&pool->gc, __ATOMIC_RELAXED);
 367         if (gc && lc != gc) {
 368                 mlx5_ipool_lock(pool);
 369                 if (lc && !(--lc->ref_cnt))
 370                         olc = lc;
 371                 lc = pool->gc;
 372                 lc->ref_cnt++;
 373                 pool->cache[cidx]->lc = lc;
 374                 mlx5_ipool_unlock(pool);
 375                 if (olc)
 376                         pool->cfg.free(olc);
 377         }
 378         return lc;
 379 }
 380
 381 static uint32_t
 382 mlx5_ipool_allocate_from_global(struct mlx5_indexed_pool *pool, int cidx)
 383 {
 384         struct mlx5_indexed_trunk *trunk;
 385         struct mlx5_indexed_cache *p, *lc, *olc = NULL;
 386         size_t trunk_size = 0;
 387         size_t data_size;
 388         uint32_t cur_max_idx, trunk_idx, trunk_n;
 389         uint32_t fetch_size, ts_idx, i;
 390         int n_grow;
 391
 392 check_again:
 393         p = NULL;
 394         fetch_size = 0;
 395         /*
 396          * Fetch new index from global if possible. First round local
 397          * cache will be NULL.
 398          */
 399         lc = pool->cache[cidx]->lc;
 400         mlx5_ipool_lock(pool);
 401         /* Try to update local cache first. */
 402         if (likely(pool->gc)) {
 403                 if (lc != pool->gc) {
 404                         if (lc && !(--lc->ref_cnt))
 405                                 olc = lc;
 406                         lc = pool->gc;
 407                         lc->ref_cnt++;
 408                         pool->cache[cidx]->lc = lc;
 409                 }
 410                 if (lc->len) {
 411                         /* Use the updated local cache to fetch index. */
 412                         fetch_size = pool->cfg.per_core_cache >> 2;
 413                         if (lc->len < fetch_size)
 414                                 fetch_size = lc->len;
 415                         lc->len -= fetch_size;
 416                         memcpy(pool->cache[cidx]->idx, &lc->idx[lc->len],
 417                                sizeof(uint32_t) * fetch_size);
 418                 }
 419         }
 420         mlx5_ipool_unlock(pool);
 421         if (unlikely(olc)) {
 422                 pool->cfg.free(olc);
 423                 olc = NULL;
 424         }
 425         if (fetch_size) {
 426                 pool->cache[cidx]->len = fetch_size - 1;
 427                 return pool->cache[cidx]->idx[pool->cache[cidx]->len];
 428         }
 429         trunk_idx = lc ? __atomic_load_n(&lc->n_trunk_valid,
 430                          __ATOMIC_ACQUIRE) : 0;
 431         trunk_n = lc ? lc->n_trunk : 0;
 432         cur_max_idx = mlx5_trunk_idx_offset_get(pool, trunk_idx);
 433         /* Check if index reach maximum. */
 434         if (trunk_idx == TRUNK_MAX_IDX ||
 435             cur_max_idx >= pool->cfg.max_idx)
 436                 return 0;
 437         /* No enough space in trunk array, resize the trunks array. */
 438         if (trunk_idx == trunk_n) {
 439                 n_grow = trunk_idx ? trunk_idx :
 440                              RTE_CACHE_LINE_SIZE / sizeof(void *);
 441                 cur_max_idx = mlx5_trunk_idx_offset_get(pool, trunk_n + n_grow);
 442                 /* Resize the trunk array. */
 443                 p = pool->cfg.malloc(0, ((trunk_idx + n_grow) *
 444                         sizeof(struct mlx5_indexed_trunk *)) +
 445                         (cur_max_idx * sizeof(uint32_t)) + sizeof(*p),
 446                         RTE_CACHE_LINE_SIZE, rte_socket_id());
 447                 if (!p)
 448                         return 0;
 449                 p->trunks = (struct mlx5_indexed_trunk **)&p->idx[cur_max_idx];
 450                 if (lc)
 451                         memcpy(p->trunks, lc->trunks, trunk_idx *
 452                        sizeof(struct mlx5_indexed_trunk *));
 453 #ifdef RTE_LIBRTE_MLX5_DEBUG
 454                 memset(RTE_PTR_ADD(p->trunks, trunk_idx * sizeof(void *)), 0,
 455                         n_grow * sizeof(void *));
 456 #endif
 457                 p->n_trunk_valid = trunk_idx;
 458                 p->n_trunk = trunk_n + n_grow;
 459                 p->len = 0;
 460         }
 461         /* Prepare the new trunk. */
 462         trunk_size = sizeof(*trunk);
 463         data_size = mlx5_trunk_size_get(pool, trunk_idx);
 464         trunk_size += RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size);
 465         trunk = pool->cfg.malloc(0, trunk_size,
 466                                  RTE_CACHE_LINE_SIZE, rte_socket_id());
 467         if (unlikely(!trunk)) {
 468                 pool->cfg.free(p);
 469                 return 0;
 470         }
 471         trunk->idx = trunk_idx;
 472         trunk->free = data_size;
 473         mlx5_ipool_lock(pool);
 474         /*
 475          * Double check if trunks has been updated or have available index.
 476          * During the new trunk allocate, index may still be flushed to the
 477          * global cache. So also need to check the pool->gc->len.
 478          */
 479         if (pool->gc && (lc != pool->gc ||
 480             lc->n_trunk_valid != trunk_idx ||
 481             pool->gc->len)) {
 482                 mlx5_ipool_unlock(pool);
 483                 if (p)
 484                         pool->cfg.free(p);
 485                 pool->cfg.free(trunk);
 486                 goto check_again;
 487         }
 488         /* Resize the trunk array and update local cache first.  */
 489         if (p) {
 490                 if (lc && !(--lc->ref_cnt))
 491                         olc = lc;
 492                 lc = p;
 493                 lc->ref_cnt = 1;
 494                 pool->cache[cidx]->lc = lc;
 495                 __atomic_store_n(&pool->gc, p, __ATOMIC_RELAXED);
 496         }
 497         /* Add trunk to trunks array. */
 498         lc->trunks[trunk_idx] = trunk;
 499         __atomic_fetch_add(&lc->n_trunk_valid, 1, __ATOMIC_RELAXED);
 500         /* Enqueue half of the index to global. */
 501         ts_idx = mlx5_trunk_idx_offset_get(pool, trunk_idx) + 1;
 502         fetch_size = trunk->free >> 1;
 503         for (i = 0; i < fetch_size; i++)
 504                 lc->idx[i] = ts_idx + i;
 505         lc->len = fetch_size;
 506         mlx5_ipool_unlock(pool);
 507         /* Copy left half - 1 to local cache index array. */
 508         pool->cache[cidx]->len = trunk->free - fetch_size - 1;
 509         ts_idx += fetch_size;
 510         for (i = 0; i < pool->cache[cidx]->len; i++)
 511                 pool->cache[cidx]->idx[i] = ts_idx + i;
 512         if (olc)
 513                 pool->cfg.free(olc);
 514         return ts_idx + i;
 515 }
 516
 517 static void *
 518 mlx5_ipool_get_cache(struct mlx5_indexed_pool *pool, uint32_t idx)
 519 {
 520         struct mlx5_indexed_trunk *trunk;
 521         struct mlx5_indexed_cache *lc;
 522         uint32_t trunk_idx;
 523         uint32_t entry_idx;
 524         int cidx;
 525
 526         MLX5_ASSERT(idx);
 527         cidx = rte_lcore_index(rte_lcore_id());
 528         if (unlikely(cidx == -1)) {
 529                 rte_errno = ENOTSUP;
 530                 return NULL;
 531         }
 532         lc = mlx5_ipool_update_global_cache(pool, cidx);
 533         idx -= 1;
 534         trunk_idx = mlx5_trunk_idx_get(pool, idx);
 535         trunk = lc->trunks[trunk_idx];
 536         MLX5_ASSERT(trunk);
 537         entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk_idx);
 538         return &trunk->data[entry_idx * pool->cfg.size];
 539 }
 540
 541 static void *
 542 mlx5_ipool_malloc_cache(struct mlx5_indexed_pool *pool, uint32_t *idx)
 543 {
 544         int cidx;
 545
 546         cidx = rte_lcore_index(rte_lcore_id());
 547         if (unlikely(cidx == -1)) {
 548                 rte_errno = ENOTSUP;
 549                 return NULL;
 550         }
 551         if (unlikely(!pool->cache[cidx])) {
 552                 pool->cache[cidx] = pool->cfg.malloc(MLX5_MEM_ZERO,
 553                         sizeof(struct mlx5_ipool_per_lcore) +
 554                         (pool->cfg.per_core_cache * sizeof(uint32_t)),
 555                         RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 556                 if (!pool->cache[cidx]) {
 557                         DRV_LOG(ERR, "Ipool cache%d allocate failed\n", cidx);
 558                         return NULL;
 559                 }
 560         } else if (pool->cache[cidx]->len) {
 561                 pool->cache[cidx]->len--;
 562                 *idx = pool->cache[cidx]->idx[pool->cache[cidx]->len];
 563                 return mlx5_ipool_get_cache(pool, *idx);
 564         }
 565         /* Not enough idx in global cache. Keep fetching from global. */
 566         *idx = mlx5_ipool_allocate_from_global(pool, cidx);
 567         if (unlikely(!(*idx)))
 568                 return NULL;
 569         return mlx5_ipool_get_cache(pool, *idx);
 570 }
 571
 572 static void
 573 mlx5_ipool_free_cache(struct mlx5_indexed_pool *pool, uint32_t idx)
 574 {
 575         int cidx;
 576         struct mlx5_ipool_per_lcore *ilc;
 577         struct mlx5_indexed_cache *gc, *olc = NULL;
 578         uint32_t reclaim_num = 0;
 579
 580         MLX5_ASSERT(idx);
 581         cidx = rte_lcore_index(rte_lcore_id());
 582         if (unlikely(cidx == -1)) {
 583                 rte_errno = ENOTSUP;
 584                 return;
 585         }
 586         /*
 587          * When index was allocated on core A but freed on core B. In this
 588          * case check if local cache on core B was allocated before.
 589          */
 590         if (unlikely(!pool->cache[cidx])) {
 591                 pool->cache[cidx] = pool->cfg.malloc(MLX5_MEM_ZERO,
 592                         sizeof(struct mlx5_ipool_per_lcore) +
 593                         (pool->cfg.per_core_cache * sizeof(uint32_t)),
 594                         RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 595                 if (!pool->cache[cidx]) {
 596                         DRV_LOG(ERR, "Ipool cache%d allocate failed\n", cidx);
 597                         return;
 598                 }
 599         }
 600         /* Try to enqueue to local index cache. */
 601         if (pool->cache[cidx]->len < pool->cfg.per_core_cache) {
 602                 pool->cache[cidx]->idx[pool->cache[cidx]->len] = idx;
 603                 pool->cache[cidx]->len++;
 604                 return;
 605         }
 606         ilc = pool->cache[cidx];
 607         reclaim_num = pool->cfg.per_core_cache >> 2;
 608         ilc->len -= reclaim_num;
 609         /* Local index cache full, try with global index cache. */
 610         mlx5_ipool_lock(pool);
 611         gc = pool->gc;
 612         if (ilc->lc != gc) {
 613                 if (!(--ilc->lc->ref_cnt))
 614                         olc = ilc->lc;
 615                 gc->ref_cnt++;
 616                 ilc->lc = gc;
 617         }
 618         memcpy(&gc->idx[gc->len], &ilc->idx[ilc->len],
 619                reclaim_num * sizeof(uint32_t));
 620         gc->len += reclaim_num;
 621         mlx5_ipool_unlock(pool);
 622         if (olc)
 623                 pool->cfg.free(olc);
 624         pool->cache[cidx]->idx[pool->cache[cidx]->len] = idx;
 625         pool->cache[cidx]->len++;
 626 }
 627
 628 void *
 629 mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 630 {
 631         struct mlx5_indexed_trunk *trunk;
 632         uint64_t slab = 0;
 633         uint32_t iidx = 0;
 634         void *p;
 635
 636         if (pool->cfg.per_core_cache)
 637                 return mlx5_ipool_malloc_cache(pool, idx);
 638         mlx5_ipool_lock(pool);
 639         if (pool->free_list == TRUNK_INVALID) {
 640                 /* If no available trunks, grow new. */
 641                 if (mlx5_ipool_grow(pool)) {
 642                         mlx5_ipool_unlock(pool);
 643                         return NULL;
 644                 }
 645         }
 646         MLX5_ASSERT(pool->free_list != TRUNK_INVALID);
 647         trunk = pool->trunks[pool->free_list];
 648         MLX5_ASSERT(trunk->free);
 649         if (!rte_bitmap_scan(trunk->bmp, &iidx, &slab)) {
 650                 mlx5_ipool_unlock(pool);
 651                 return NULL;
 652         }
 653         MLX5_ASSERT(slab);
 654         iidx += __builtin_ctzll(slab);
 655         MLX5_ASSERT(iidx != UINT32_MAX);
 656         MLX5_ASSERT(iidx < mlx5_trunk_size_get(pool, trunk->idx));
 657         rte_bitmap_clear(trunk->bmp, iidx);
 658         p = &trunk->data[iidx * pool->cfg.size];
 659         /*
 660          * The ipool index should grow continually from small to big,
 661          * some features as metering only accept limited bits of index.
 662          * Random index with MSB set may be rejected.
 663          */
 664         iidx += mlx5_trunk_idx_offset_get(pool, trunk->idx);
 665         iidx += 1; /* non-zero index. */
 666         trunk->free--;
 667 #ifdef POOL_DEBUG
 668         pool->n_entry++;
 669 #endif
 670         if (!trunk->free) {
 671                 /* Full trunk will be removed from free list in imalloc. */
 672                 MLX5_ASSERT(pool->free_list == trunk->idx);
 673                 pool->free_list = trunk->next;
 674                 if (trunk->next != TRUNK_INVALID)
 675                         pool->trunks[trunk->next]->prev = TRUNK_INVALID;
 676                 trunk->prev = TRUNK_INVALID;
 677                 trunk->next = TRUNK_INVALID;
 678 #ifdef POOL_DEBUG
 679                 pool->trunk_empty++;
 680                 pool->trunk_avail--;
 681 #endif
 682         }
 683         *idx = iidx;
 684         mlx5_ipool_unlock(pool);
 685         return p;
 686 }
 687
 688 void *
 689 mlx5_ipool_zmalloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 690 {
 691         void *entry = mlx5_ipool_malloc(pool, idx);
 692
 693         if (entry && pool->cfg.size)
 694                 memset(entry, 0, pool->cfg.size);
 695         return entry;
 696 }
 697
 698 void
 699 mlx5_ipool_free(struct mlx5_indexed_pool *pool, uint32_t idx)
 700 {
 701         struct mlx5_indexed_trunk *trunk;
 702         uint32_t trunk_idx;
 703         uint32_t entry_idx;
 704
 705         if (!idx)
 706                 return;
 707         if (pool->cfg.per_core_cache) {
 708                 mlx5_ipool_free_cache(pool, idx);
 709                 return;
 710         }
 711         idx -= 1;
 712         mlx5_ipool_lock(pool);
 713         trunk_idx = mlx5_trunk_idx_get(pool, idx);
 714         if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
 715             (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
 716                 goto out;
 717         trunk = pool->trunks[trunk_idx];
 718         if (!trunk)
 719                 goto out;
 720         entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
 721         if (trunk_idx != trunk->idx ||
 722             rte_bitmap_get(trunk->bmp, entry_idx))
 723                 goto out;
 724         rte_bitmap_set(trunk->bmp, entry_idx);
 725         trunk->free++;
 726         if (pool->cfg.release_mem_en && trunk->free == mlx5_trunk_size_get
 727            (pool, trunk->idx)) {
 728                 if (pool->free_list == trunk->idx)
 729                         pool->free_list = trunk->next;
 730                 if (trunk->next != TRUNK_INVALID)
 731                         pool->trunks[trunk->next]->prev = trunk->prev;
 732                 if (trunk->prev != TRUNK_INVALID)
 733                         pool->trunks[trunk->prev]->next = trunk->next;
 734                 pool->cfg.free(trunk);
 735                 pool->trunks[trunk_idx] = NULL;
 736                 pool->n_trunk_valid--;
 737 #ifdef POOL_DEBUG
 738                 pool->trunk_avail--;
 739                 pool->trunk_free++;
 740 #endif
 741                 if (pool->n_trunk_valid == 0) {
 742                         pool->cfg.free(pool->trunks);
 743                         pool->trunks = NULL;
 744                         pool->n_trunk = 0;
 745                 }
 746         } else if (trunk->free == 1) {
 747                 /* Put into free trunk list head. */
 748                 MLX5_ASSERT(pool->free_list != trunk->idx);
 749                 trunk->next = pool->free_list;
 750                 trunk->prev = TRUNK_INVALID;
 751                 if (pool->free_list != TRUNK_INVALID)
 752                         pool->trunks[pool->free_list]->prev = trunk->idx;
 753                 pool->free_list = trunk->idx;
 754 #ifdef POOL_DEBUG
 755                 pool->trunk_empty--;
 756                 pool->trunk_avail++;
 757 #endif
 758         }
 759 #ifdef POOL_DEBUG
 760         pool->n_entry--;
 761 #endif
 762 out:
 763         mlx5_ipool_unlock(pool);
 764 }
 765
 766 void *
 767 mlx5_ipool_get(struct mlx5_indexed_pool *pool, uint32_t idx)
 768 {
 769         struct mlx5_indexed_trunk *trunk;
 770         void *p = NULL;
 771         uint32_t trunk_idx;
 772         uint32_t entry_idx;
 773
 774         if (!idx)
 775                 return NULL;
 776         if (pool->cfg.per_core_cache)
 777                 return mlx5_ipool_get_cache(pool, idx);
 778         idx -= 1;
 779         mlx5_ipool_lock(pool);
 780         trunk_idx = mlx5_trunk_idx_get(pool, idx);
 781         if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
 782             (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
 783                 goto out;
 784         trunk = pool->trunks[trunk_idx];
 785         if (!trunk)
 786                 goto out;
 787         entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
 788         if (trunk_idx != trunk->idx ||
 789             rte_bitmap_get(trunk->bmp, entry_idx))
 790                 goto out;
 791         p = &trunk->data[entry_idx * pool->cfg.size];
 792 out:
 793         mlx5_ipool_unlock(pool);
 794         return p;
 795 }
 796
 797 int
 798 mlx5_ipool_destroy(struct mlx5_indexed_pool *pool)
 799 {
 800         struct mlx5_indexed_trunk **trunks = NULL;
 801         struct mlx5_indexed_cache *gc = pool->gc;
 802         uint32_t i, n_trunk_valid = 0;
 803
 804         MLX5_ASSERT(pool);
 805         mlx5_ipool_lock(pool);
 806         if (pool->cfg.per_core_cache) {
 807                 for (i = 0; i < RTE_MAX_LCORE; i++) {
 808                         /*
 809                          * Free only old global cache. Pool gc will be
 810                          * freed at last.
 811                          */
 812                         if (pool->cache[i]) {
 813                                 if (pool->cache[i]->lc &&
 814                                     pool->cache[i]->lc != pool->gc &&
 815                                     (!(--pool->cache[i]->lc->ref_cnt)))
 816                                         pool->cfg.free(pool->cache[i]->lc);
 817                                 pool->cfg.free(pool->cache[i]);
 818                         }
 819                 }
 820                 if (gc) {
 821                         trunks = gc->trunks;
 822                         n_trunk_valid = gc->n_trunk_valid;
 823                 }
 824         } else {
 825                 gc = NULL;
 826                 trunks = pool->trunks;
 827                 n_trunk_valid = pool->n_trunk_valid;
 828         }
 829         for (i = 0; i < n_trunk_valid; i++) {
 830                 if (trunks[i])
 831                         pool->cfg.free(trunks[i]);
 832         }
 833         if (!gc && trunks)
 834                 pool->cfg.free(trunks);
 835         if (gc)
 836                 pool->cfg.free(gc);
 837         mlx5_ipool_unlock(pool);
 838         mlx5_free(pool);
 839         return 0;
 840 }
 841
 842 void
 843 mlx5_ipool_dump(struct mlx5_indexed_pool *pool)
 844 {
 845         printf("Pool %s entry size %u, trunks %u, %d entry per trunk, "
 846                "total: %d\n",
 847                pool->cfg.type, pool->cfg.size, pool->n_trunk_valid,
 848                pool->cfg.trunk_size, pool->n_trunk_valid);
 849 #ifdef POOL_DEBUG
 850         printf("Pool %s entry %u, trunk alloc %u, empty: %u, "
 851                "available %u free %u\n",
 852                pool->cfg.type, pool->n_entry, pool->trunk_new,
 853                pool->trunk_empty, pool->trunk_avail, pool->trunk_free);
 854 #endif
 855 }
 856
 857 struct mlx5_l3t_tbl *
 858 mlx5_l3t_create(enum mlx5_l3t_type type)
 859 {
 860         struct mlx5_l3t_tbl *tbl;
 861         struct mlx5_indexed_pool_config l3t_ip_cfg = {
 862                 .trunk_size = 16,
 863                 .grow_trunk = 6,
 864                 .grow_shift = 1,
 865                 .need_lock = 0,
 866                 .release_mem_en = 1,
 867                 .malloc = mlx5_malloc,
 868                 .free = mlx5_free,
 869         };
 870
 871         if (type >= MLX5_L3T_TYPE_MAX) {
 872                 rte_errno = EINVAL;
 873                 return NULL;
 874         }
 875         tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(struct mlx5_l3t_tbl), 1,
 876                           SOCKET_ID_ANY);
 877         if (!tbl) {
 878                 rte_errno = ENOMEM;
 879                 return NULL;
 880         }
 881         tbl->type = type;
 882         switch (type) {
 883         case MLX5_L3T_TYPE_WORD:
 884                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_word);
 885                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_w";
 886                 break;
 887         case MLX5_L3T_TYPE_DWORD:
 888                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_dword);
 889                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_dw";
 890                 break;
 891         case MLX5_L3T_TYPE_QWORD:
 892                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_qword);
 893                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_qw";
 894                 break;
 895         default:
 896                 l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_ptr);
 897                 l3t_ip_cfg.type = "mlx5_l3t_e_tbl_tpr";
 898                 break;
 899         }
 900         rte_spinlock_init(&tbl->sl);
 901         tbl->eip = mlx5_ipool_create(&l3t_ip_cfg);
 902         if (!tbl->eip) {
 903                 rte_errno = ENOMEM;
 904                 mlx5_free(tbl);
 905                 tbl = NULL;
 906         }
 907         return tbl;
 908 }
 909
 910 void
 911 mlx5_l3t_destroy(struct mlx5_l3t_tbl *tbl)
 912 {
 913         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 914         uint32_t i, j;
 915
 916         if (!tbl)
 917                 return;
 918         g_tbl = tbl->tbl;
 919         if (g_tbl) {
 920                 for (i = 0; i < MLX5_L3T_GT_SIZE; i++) {
 921                         m_tbl = g_tbl->tbl[i];
 922                         if (!m_tbl)
 923                                 continue;
 924                         for (j = 0; j < MLX5_L3T_MT_SIZE; j++) {
 925                                 if (!m_tbl->tbl[j])
 926                                         continue;
 927                                 MLX5_ASSERT(!((struct mlx5_l3t_entry_word *)
 928                                             m_tbl->tbl[j])->ref_cnt);
 929                                 mlx5_ipool_free(tbl->eip,
 930                                                 ((struct mlx5_l3t_entry_word *)
 931                                                 m_tbl->tbl[j])->idx);
 932                                 m_tbl->tbl[j] = 0;
 933                                 if (!(--m_tbl->ref_cnt))
 934                                         break;
 935                         }
 936                         MLX5_ASSERT(!m_tbl->ref_cnt);
 937                         mlx5_free(g_tbl->tbl[i]);
 938                         g_tbl->tbl[i] = 0;
 939                         if (!(--g_tbl->ref_cnt))
 940                                 break;
 941                 }
 942                 MLX5_ASSERT(!g_tbl->ref_cnt);
 943                 mlx5_free(tbl->tbl);
 944                 tbl->tbl = 0;
 945         }
 946         mlx5_ipool_destroy(tbl->eip);
 947         mlx5_free(tbl);
 948 }
 949
 950 static int32_t
 951 __l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
 952                 union mlx5_l3t_data *data)
 953 {
 954         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
 955         struct mlx5_l3t_entry_word *w_e_tbl;
 956         struct mlx5_l3t_entry_dword *dw_e_tbl;
 957         struct mlx5_l3t_entry_qword *qw_e_tbl;
 958         struct mlx5_l3t_entry_ptr *ptr_e_tbl;
 959         void *e_tbl;
 960         uint32_t entry_idx;
 961
 962         g_tbl = tbl->tbl;
 963         if (!g_tbl)
 964                 return -1;
 965         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
 966         if (!m_tbl)
 967                 return -1;
 968         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
 969         if (!e_tbl)
 970                 return -1;
 971         entry_idx = idx & MLX5_L3T_ET_MASK;
 972         switch (tbl->type) {
 973         case MLX5_L3T_TYPE_WORD:
 974                 w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
 975                 data->word = w_e_tbl->entry[entry_idx].data;
 976                 if (w_e_tbl->entry[entry_idx].data)
 977                         w_e_tbl->entry[entry_idx].ref_cnt++;
 978                 break;
 979         case MLX5_L3T_TYPE_DWORD:
 980                 dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
 981                 data->dword = dw_e_tbl->entry[entry_idx].data;
 982                 if (dw_e_tbl->entry[entry_idx].data)
 983                         dw_e_tbl->entry[entry_idx].ref_cnt++;
 984                 break;
 985         case MLX5_L3T_TYPE_QWORD:
 986                 qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
 987                 data->qword = qw_e_tbl->entry[entry_idx].data;
 988                 if (qw_e_tbl->entry[entry_idx].data)
 989                         qw_e_tbl->entry[entry_idx].ref_cnt++;
 990                 break;
 991         default:
 992                 ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
 993                 data->ptr = ptr_e_tbl->entry[entry_idx].data;
 994                 if (ptr_e_tbl->entry[entry_idx].data)
 995                         ptr_e_tbl->entry[entry_idx].ref_cnt++;
 996                 break;
 997         }
 998         return 0;
 999 }
1000
1001 int32_t
1002 mlx5_l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
1003                    union mlx5_l3t_data *data)
1004 {
1005         int ret;
1006
1007         rte_spinlock_lock(&tbl->sl);
1008         ret = __l3t_get_entry(tbl, idx, data);
1009         rte_spinlock_unlock(&tbl->sl);
1010         return ret;
1011 }
1012
1013 int32_t
1014 mlx5_l3t_clear_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx)
1015 {
1016         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
1017         struct mlx5_l3t_entry_word *w_e_tbl;
1018         struct mlx5_l3t_entry_dword *dw_e_tbl;
1019         struct mlx5_l3t_entry_qword *qw_e_tbl;
1020         struct mlx5_l3t_entry_ptr *ptr_e_tbl;
1021         void *e_tbl;
1022         uint32_t entry_idx;
1023         uint64_t ref_cnt;
1024         int32_t ret = -1;
1025
1026         rte_spinlock_lock(&tbl->sl);
1027         g_tbl = tbl->tbl;
1028         if (!g_tbl)
1029                 goto out;
1030         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
1031         if (!m_tbl)
1032                 goto out;
1033         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
1034         if (!e_tbl)
1035                 goto out;
1036         entry_idx = idx & MLX5_L3T_ET_MASK;
1037         switch (tbl->type) {
1038         case MLX5_L3T_TYPE_WORD:
1039                 w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
1040                 MLX5_ASSERT(w_e_tbl->entry[entry_idx].ref_cnt);
1041                 ret = --w_e_tbl->entry[entry_idx].ref_cnt;
1042                 if (ret)
1043                         goto out;
1044                 w_e_tbl->entry[entry_idx].data = 0;
1045                 ref_cnt = --w_e_tbl->ref_cnt;
1046                 break;
1047         case MLX5_L3T_TYPE_DWORD:
1048                 dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
1049                 MLX5_ASSERT(dw_e_tbl->entry[entry_idx].ref_cnt);
1050                 ret = --dw_e_tbl->entry[entry_idx].ref_cnt;
1051                 if (ret)
1052                         goto out;
1053                 dw_e_tbl->entry[entry_idx].data = 0;
1054                 ref_cnt = --dw_e_tbl->ref_cnt;
1055                 break;
1056         case MLX5_L3T_TYPE_QWORD:
1057                 qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
1058                 MLX5_ASSERT(qw_e_tbl->entry[entry_idx].ref_cnt);
1059                 ret = --qw_e_tbl->entry[entry_idx].ref_cnt;
1060                 if (ret)
1061                         goto out;
1062                 qw_e_tbl->entry[entry_idx].data = 0;
1063                 ref_cnt = --qw_e_tbl->ref_cnt;
1064                 break;
1065         default:
1066                 ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
1067                 MLX5_ASSERT(ptr_e_tbl->entry[entry_idx].ref_cnt);
1068                 ret = --ptr_e_tbl->entry[entry_idx].ref_cnt;
1069                 if (ret)
1070                         goto out;
1071                 ptr_e_tbl->entry[entry_idx].data = NULL;
1072                 ref_cnt = --ptr_e_tbl->ref_cnt;
1073                 break;
1074         }
1075         if (!ref_cnt) {
1076                 mlx5_ipool_free(tbl->eip,
1077                                 ((struct mlx5_l3t_entry_word *)e_tbl)->idx);
1078                 m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
1079                                                                         NULL;
1080                 if (!(--m_tbl->ref_cnt)) {
1081                         mlx5_free(m_tbl);
1082                         g_tbl->tbl
1083                         [(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] = NULL;
1084                         if (!(--g_tbl->ref_cnt)) {
1085                                 mlx5_free(g_tbl);
1086                                 tbl->tbl = 0;
1087                         }
1088                 }
1089         }
1090 out:
1091         rte_spinlock_unlock(&tbl->sl);
1092         return ret;
1093 }
1094
1095 static int32_t
1096 __l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
1097                 union mlx5_l3t_data *data)
1098 {
1099         struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
1100         struct mlx5_l3t_entry_word *w_e_tbl;
1101         struct mlx5_l3t_entry_dword *dw_e_tbl;
1102         struct mlx5_l3t_entry_qword *qw_e_tbl;
1103         struct mlx5_l3t_entry_ptr *ptr_e_tbl;
1104         void *e_tbl;
1105         uint32_t entry_idx, tbl_idx = 0;
1106
1107         /* Check the global table, create it if empty. */
1108         g_tbl = tbl->tbl;
1109         if (!g_tbl) {
1110                 g_tbl = mlx5_malloc(MLX5_MEM_ZERO,
1111                                     sizeof(struct mlx5_l3t_level_tbl) +
1112                                     sizeof(void *) * MLX5_L3T_GT_SIZE, 1,
1113                                     SOCKET_ID_ANY);
1114                 if (!g_tbl) {
1115                         rte_errno = ENOMEM;
1116                         return -1;
1117                 }
1118                 tbl->tbl = g_tbl;
1119         }
1120         /*
1121          * Check the middle table, create it if empty. Ref_cnt will be
1122          * increased if new sub table created.
1123          */
1124         m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
1125         if (!m_tbl) {
1126                 m_tbl = mlx5_malloc(MLX5_MEM_ZERO,
1127                                     sizeof(struct mlx5_l3t_level_tbl) +
1128                                     sizeof(void *) * MLX5_L3T_MT_SIZE, 1,
1129                                     SOCKET_ID_ANY);
1130                 if (!m_tbl) {
1131                         rte_errno = ENOMEM;
1132                         return -1;
1133                 }
1134                 g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] =
1135                                                                         m_tbl;
1136                 g_tbl->ref_cnt++;
1137         }
1138         /*
1139          * Check the entry table, create it if empty. Ref_cnt will be
1140          * increased if new sub entry table created.
1141          */
1142         e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
1143         if (!e_tbl) {
1144                 e_tbl = mlx5_ipool_zmalloc(tbl->eip, &tbl_idx);
1145                 if (!e_tbl) {
1146                         rte_errno = ENOMEM;
1147                         return -1;
1148                 }
1149                 ((struct mlx5_l3t_entry_word *)e_tbl)->idx = tbl_idx;
1150                 m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
1151                                                                         e_tbl;
1152                 m_tbl->ref_cnt++;
1153         }
1154         entry_idx = idx & MLX5_L3T_ET_MASK;
1155         switch (tbl->type) {
1156         case MLX5_L3T_TYPE_WORD:
1157                 w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
1158                 if (w_e_tbl->entry[entry_idx].data) {
1159                         data->word = w_e_tbl->entry[entry_idx].data;
1160                         w_e_tbl->entry[entry_idx].ref_cnt++;
1161                         rte_errno = EEXIST;
1162                         return -1;
1163                 }
1164                 w_e_tbl->entry[entry_idx].data = data->word;
1165                 w_e_tbl->entry[entry_idx].ref_cnt = 1;
1166                 w_e_tbl->ref_cnt++;
1167                 break;
1168         case MLX5_L3T_TYPE_DWORD:
1169                 dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
1170                 if (dw_e_tbl->entry[entry_idx].data) {
1171                         data->dword = dw_e_tbl->entry[entry_idx].data;
1172                         dw_e_tbl->entry[entry_idx].ref_cnt++;
1173                         rte_errno = EEXIST;
1174                         return -1;
1175                 }
1176                 dw_e_tbl->entry[entry_idx].data = data->dword;
1177                 dw_e_tbl->entry[entry_idx].ref_cnt = 1;
1178                 dw_e_tbl->ref_cnt++;
1179                 break;
1180         case MLX5_L3T_TYPE_QWORD:
1181                 qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
1182                 if (qw_e_tbl->entry[entry_idx].data) {
1183                         data->qword = qw_e_tbl->entry[entry_idx].data;
1184                         qw_e_tbl->entry[entry_idx].ref_cnt++;
1185                         rte_errno = EEXIST;
1186                         return -1;
1187                 }
1188                 qw_e_tbl->entry[entry_idx].data = data->qword;
1189                 qw_e_tbl->entry[entry_idx].ref_cnt = 1;
1190                 qw_e_tbl->ref_cnt++;
1191                 break;
1192         default:
1193                 ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
1194                 if (ptr_e_tbl->entry[entry_idx].data) {
1195                         data->ptr = ptr_e_tbl->entry[entry_idx].data;
1196                         ptr_e_tbl->entry[entry_idx].ref_cnt++;
1197                         rte_errno = EEXIST;
1198                         return -1;
1199                 }
1200                 ptr_e_tbl->entry[entry_idx].data = data->ptr;
1201                 ptr_e_tbl->entry[entry_idx].ref_cnt = 1;
1202                 ptr_e_tbl->ref_cnt++;
1203                 break;
1204         }
1205         return 0;
1206 }
1207
1208 int32_t
1209 mlx5_l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
1210                    union mlx5_l3t_data *data)
1211 {
1212         int ret;
1213
1214         rte_spinlock_lock(&tbl->sl);
1215         ret = __l3t_set_entry(tbl, idx, data);
1216         rte_spinlock_unlock(&tbl->sl);
1217         return ret;
1218 }
1219
1220 int32_t
1221 mlx5_l3t_prepare_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
1222                        union mlx5_l3t_data *data,
1223                        mlx5_l3t_alloc_callback_fn cb, void *ctx)
1224 {
1225         int32_t ret;
1226
1227         rte_spinlock_lock(&tbl->sl);
1228         /* Check if entry data is ready. */
1229         ret = __l3t_get_entry(tbl, idx, data);
1230         if (!ret) {
1231                 switch (tbl->type) {
1232                 case MLX5_L3T_TYPE_WORD:
1233                         if (data->word)
1234                                 goto out;
1235                         break;
1236                 case MLX5_L3T_TYPE_DWORD:
1237                         if (data->dword)
1238                                 goto out;
1239                         break;
1240                 case MLX5_L3T_TYPE_QWORD:
1241                         if (data->qword)
1242                                 goto out;
1243                         break;
1244                 default:
1245                         if (data->ptr)
1246                                 goto out;
1247                         break;
1248                 }
1249         }
1250         /* Entry data is not ready, use user callback to create it. */
1251         ret = cb(ctx, data);
1252         if (ret)
1253                 goto out;
1254         /* Save the new allocated data to entry. */
1255         ret = __l3t_set_entry(tbl, idx, data);
1256 out:
1257         rte_spinlock_unlock(&tbl->sl);
1258         return ret;
1259 }