net/mlx5: remove redundant flag in device config
[dpdk.git] / drivers / net / mlx5 / mlx5_utils.c
index c92914c..e4e66ae 100644 (file)
  */
 
 #include <rte_malloc.h>
-#include <rte_hash_crc.h>
 
-#include "mlx5_utils.h"
+#include <mlx5_malloc.h>
 
-struct mlx5_hlist *
-mlx5_hlist_create(const char *name, uint32_t size)
-{
-       struct mlx5_hlist *h;
-       uint32_t act_size;
-       uint32_t alloc_size;
+#include "mlx5_utils.h"
 
-       if (!size)
-               return NULL;
-       /* Align to the next power of 2, 32bits integer is enough now. */
-       if (!rte_is_power_of_2(size)) {
-               act_size = rte_align32pow2(size);
-               DRV_LOG(WARNING, "Size 0x%" PRIX32 " is not power of 2, will "
-                       "be aligned to 0x%" PRIX32 ".\n", size, act_size);
-       } else {
-               act_size = size;
-       }
-       alloc_size = sizeof(struct mlx5_hlist) +
-                    sizeof(struct mlx5_hlist_head) * act_size;
-       /* Using zmalloc, then no need to initialize the heads. */
-       h = rte_zmalloc(name, alloc_size, RTE_CACHE_LINE_SIZE);
-       if (!h) {
-               DRV_LOG(ERR, "No memory for hash list %s creation\n",
-                       name ? name : "None");
-               return NULL;
-       }
-       if (name)
-               snprintf(h->name, MLX5_HLIST_NAMESIZE, "%s", name);
-       h->table_sz = act_size;
-       h->mask = act_size - 1;
-       DRV_LOG(DEBUG, "Hash list with %s size 0x%" PRIX32 " is created.\n",
-               h->name, act_size);
-       return h;
-}
+/********************* Indexed pool **********************/
 
-struct mlx5_hlist_entry *
-mlx5_hlist_lookup(struct mlx5_hlist *h, uint64_t key)
+static inline void
+mlx5_ipool_lock(struct mlx5_indexed_pool *pool)
 {
-       uint32_t idx;
-       struct mlx5_hlist_head *first;
-       struct mlx5_hlist_entry *node;
-
-       MLX5_ASSERT(h);
-       idx = rte_hash_crc_8byte(key, 0) & h->mask;
-       first = &h->heads[idx];
-       LIST_FOREACH(node, first, next) {
-               if (node->key == key)
-                       return node;
-       }
-       return NULL;
+       if (pool->cfg.need_lock)
+               rte_spinlock_lock(&pool->rsz_lock);
 }
 
-int
-mlx5_hlist_insert(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry)
+static inline void
+mlx5_ipool_unlock(struct mlx5_indexed_pool *pool)
 {
-       uint32_t idx;
-       struct mlx5_hlist_head *first;
-       struct mlx5_hlist_entry *node;
-
-       MLX5_ASSERT(h && entry);
-       idx = rte_hash_crc_8byte(entry->key, 0) & h->mask;
-       first = &h->heads[idx];
-       /* No need to reuse the lookup function. */
-       LIST_FOREACH(node, first, next) {
-               if (node->key == entry->key)
-                       return -EEXIST;
-       }
-       LIST_INSERT_HEAD(first, entry, next);
-       return 0;
+       if (pool->cfg.need_lock)
+               rte_spinlock_unlock(&pool->rsz_lock);
 }
 
-void
-mlx5_hlist_remove(struct mlx5_hlist *h __rte_unused,
-                 struct mlx5_hlist_entry *entry)
+static inline uint32_t
+mlx5_trunk_idx_get(struct mlx5_indexed_pool *pool, uint32_t entry_idx)
 {
-       MLX5_ASSERT(entry && entry->next.le_prev);
-       LIST_REMOVE(entry, next);
-       /* Set to NULL to get rid of removing action for more than once. */
-       entry->next.le_prev = NULL;
-}
+       struct mlx5_indexed_pool_config *cfg = &pool->cfg;
+       uint32_t trunk_idx = 0;
+       uint32_t i;
 
-void
-mlx5_hlist_destroy(struct mlx5_hlist *h,
-                  mlx5_hlist_destroy_callback_fn cb, void *ctx)
-{
-       uint32_t idx;
-       struct mlx5_hlist_entry *entry;
-
-       MLX5_ASSERT(h);
-       for (idx = 0; idx < h->table_sz; ++idx) {
-               /* no LIST_FOREACH_SAFE, using while instead */
-               while (!LIST_EMPTY(&h->heads[idx])) {
-                       entry = LIST_FIRST(&h->heads[idx]);
-                       LIST_REMOVE(entry, next);
-                       /*
-                        * The owner of whole element which contains data entry
-                        * is the user, so it's the user's duty to do the clean
-                        * up and the free work because someone may not put the
-                        * hlist entry at the beginning(suggested to locate at
-                        * the beginning). Or else the default free function
-                        * will be used.
-                        */
-                       if (cb)
-                               cb(entry, ctx);
-                       else
-                               rte_free(entry);
+       if (!cfg->grow_trunk)
+               return entry_idx / cfg->trunk_size;
+       if (entry_idx >= pool->grow_tbl[cfg->grow_trunk - 1]) {
+               trunk_idx = (entry_idx - pool->grow_tbl[cfg->grow_trunk - 1]) /
+                           (cfg->trunk_size << (cfg->grow_shift *
+                           cfg->grow_trunk)) + cfg->grow_trunk;
+       } else {
+               for (i = 0; i < cfg->grow_trunk; i++) {
+                       if (entry_idx < pool->grow_tbl[i])
+                               break;
                }
+               trunk_idx = i;
        }
-       rte_free(h);
+       return trunk_idx;
 }
 
-static inline void
-mlx5_ipool_lock(struct mlx5_indexed_pool *pool)
+static inline uint32_t
+mlx5_trunk_size_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 {
-       if (pool->cfg.need_lock)
-               rte_spinlock_lock(&pool->lock);
+       struct mlx5_indexed_pool_config *cfg = &pool->cfg;
+
+       return cfg->trunk_size << (cfg->grow_shift *
+              (trunk_idx > cfg->grow_trunk ? cfg->grow_trunk : trunk_idx));
 }
 
-static inline void
-mlx5_ipool_unlock(struct mlx5_indexed_pool *pool)
+static inline uint32_t
+mlx5_trunk_idx_offset_get(struct mlx5_indexed_pool *pool, uint32_t trunk_idx)
 {
-       if (pool->cfg.need_lock)
-               rte_spinlock_unlock(&pool->lock);
+       struct mlx5_indexed_pool_config *cfg = &pool->cfg;
+       uint32_t offset = 0;
+
+       if (!trunk_idx)
+               return 0;
+       if (!cfg->grow_trunk)
+               return cfg->trunk_size * trunk_idx;
+       if (trunk_idx < cfg->grow_trunk)
+               offset = pool->grow_tbl[trunk_idx - 1];
+       else
+               offset = pool->grow_tbl[cfg->grow_trunk - 1] +
+                        (cfg->trunk_size << (cfg->grow_shift *
+                        cfg->grow_trunk)) * (trunk_idx - cfg->grow_trunk);
+       return offset;
 }
 
 struct mlx5_indexed_pool *
 mlx5_ipool_create(struct mlx5_indexed_pool_config *cfg)
 {
        struct mlx5_indexed_pool *pool;
+       uint32_t i;
 
-       if (!cfg || !cfg->size || (!cfg->malloc ^ !cfg->free) ||
+       if (!cfg || (!cfg->malloc ^ !cfg->free) ||
+           (cfg->per_core_cache && cfg->release_mem_en) ||
            (cfg->trunk_size && ((cfg->trunk_size & (cfg->trunk_size - 1)) ||
            ((__builtin_ffs(cfg->trunk_size) + TRUNK_IDX_BITS) > 32))))
                return NULL;
-       pool = rte_zmalloc("mlx5_ipool", sizeof(*pool), RTE_CACHE_LINE_SIZE);
+       pool = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pool) + cfg->grow_trunk *
+                          sizeof(pool->grow_tbl[0]), RTE_CACHE_LINE_SIZE,
+                          SOCKET_ID_ANY);
        if (!pool)
                return NULL;
        pool->cfg = *cfg;
        if (!pool->cfg.trunk_size)
                pool->cfg.trunk_size = MLX5_IPOOL_DEFAULT_TRUNK_SIZE;
        if (!cfg->malloc && !cfg->free) {
-               pool->cfg.malloc = rte_malloc_socket;
-               pool->cfg.free = rte_free;
+               pool->cfg.malloc = mlx5_malloc;
+               pool->cfg.free = mlx5_free;
        }
-       pool->free_list = TRUNK_INVALID;
        if (pool->cfg.need_lock)
-               rte_spinlock_init(&pool->lock);
+               rte_spinlock_init(&pool->rsz_lock);
+       /*
+        * Initialize the dynamic grow trunk size lookup table to have a quick
+        * lookup for the trunk entry index offset.
+        */
+       for (i = 0; i < cfg->grow_trunk; i++) {
+               pool->grow_tbl[i] = cfg->trunk_size << (cfg->grow_shift * i);
+               if (i > 0)
+                       pool->grow_tbl[i] += pool->grow_tbl[i - 1];
+       }
+       if (!pool->cfg.max_idx)
+               pool->cfg.max_idx =
+                       mlx5_trunk_idx_offset_get(pool, TRUNK_MAX_IDX + 1);
+       if (!cfg->per_core_cache)
+               pool->free_list = TRUNK_INVALID;
+       rte_spinlock_init(&pool->lcore_lock);
        return pool;
 }
 
@@ -164,20 +125,22 @@ mlx5_ipool_grow(struct mlx5_indexed_pool *pool)
        struct mlx5_indexed_trunk **trunk_tmp;
        struct mlx5_indexed_trunk **p;
        size_t trunk_size = 0;
+       size_t data_size;
        size_t bmp_size;
-       uint32_t idx;
+       uint32_t idx, cur_max_idx, i;
 
-       if (pool->n_trunk_valid == TRUNK_MAX_IDX)
+       cur_max_idx = mlx5_trunk_idx_offset_get(pool, pool->n_trunk_valid);
+       if (pool->n_trunk_valid == TRUNK_MAX_IDX ||
+           cur_max_idx >= pool->cfg.max_idx)
                return -ENOMEM;
        if (pool->n_trunk_valid == pool->n_trunk) {
                /* No free trunk flags, expand trunk list. */
                int n_grow = pool->n_trunk_valid ? pool->n_trunk :
                             RTE_CACHE_LINE_SIZE / sizeof(void *);
 
-               p = pool->cfg.malloc(pool->cfg.type,
-                                (pool->n_trunk_valid + n_grow) *
-                                sizeof(struct mlx5_indexed_trunk *),
-                                RTE_CACHE_LINE_SIZE, rte_socket_id());
+               p = pool->cfg.malloc(0, (pool->n_trunk_valid + n_grow) *
+                                    sizeof(struct mlx5_indexed_trunk *),
+                                    RTE_CACHE_LINE_SIZE, rte_socket_id());
                if (!p)
                        return -ENOMEM;
                if (pool->trunks)
@@ -191,25 +154,41 @@ mlx5_ipool_grow(struct mlx5_indexed_pool *pool)
                        pool->cfg.free(trunk_tmp);
                pool->n_trunk += n_grow;
        }
-       idx = pool->n_trunk_valid;
+       if (!pool->cfg.release_mem_en) {
+               idx = pool->n_trunk_valid;
+       } else {
+               /* Find the first available slot in trunk list */
+               for (idx = 0; idx < pool->n_trunk; idx++)
+                       if (pool->trunks[idx] == NULL)
+                               break;
+       }
        trunk_size += sizeof(*trunk);
-       bmp_size = rte_bitmap_get_memory_footprint(pool->cfg.trunk_size);
-       trunk_size += pool->cfg.trunk_size * pool->cfg.size + bmp_size;
-       trunk = pool->cfg.malloc(pool->cfg.type, trunk_size,
+       data_size = mlx5_trunk_size_get(pool, idx);
+       bmp_size = rte_bitmap_get_memory_footprint(data_size);
+       /* rte_bitmap requires memory cacheline aligned. */
+       trunk_size += RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size);
+       trunk_size += bmp_size;
+       trunk = pool->cfg.malloc(0, trunk_size,
                                 RTE_CACHE_LINE_SIZE, rte_socket_id());
        if (!trunk)
                return -ENOMEM;
        pool->trunks[idx] = trunk;
        trunk->idx = idx;
-       trunk->free = pool->cfg.trunk_size;
+       trunk->free = data_size;
        trunk->prev = TRUNK_INVALID;
        trunk->next = TRUNK_INVALID;
        MLX5_ASSERT(pool->free_list == TRUNK_INVALID);
        pool->free_list = idx;
        /* Mark all entries as available. */
-       trunk->bmp = rte_bitmap_init_with_all_set(pool->cfg.trunk_size,
-                    &trunk->data[pool->cfg.trunk_size  * pool->cfg.size],
+       trunk->bmp = rte_bitmap_init_with_all_set(data_size, &trunk->data
+                    [RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size)],
                     bmp_size);
+       /* Clear the overhead bits in the trunk if it happens. */
+       if (cur_max_idx + data_size > pool->cfg.max_idx) {
+               for (i = pool->cfg.max_idx - cur_max_idx; i < data_size; i++)
+                       rte_bitmap_clear(trunk->bmp, i);
+       }
+       MLX5_ASSERT(trunk->bmp);
        pool->n_trunk_valid++;
 #ifdef POOL_DEBUG
        pool->trunk_new++;
@@ -218,6 +197,316 @@ mlx5_ipool_grow(struct mlx5_indexed_pool *pool)
        return 0;
 }
 
+static inline struct mlx5_indexed_cache *
+mlx5_ipool_update_global_cache(struct mlx5_indexed_pool *pool, int cidx)
+{
+       struct mlx5_indexed_cache *gc, *lc, *olc = NULL;
+
+       lc = pool->cache[cidx]->lc;
+       gc = __atomic_load_n(&pool->gc, __ATOMIC_RELAXED);
+       if (gc && lc != gc) {
+               mlx5_ipool_lock(pool);
+               if (lc && !(--lc->ref_cnt))
+                       olc = lc;
+               lc = pool->gc;
+               lc->ref_cnt++;
+               pool->cache[cidx]->lc = lc;
+               mlx5_ipool_unlock(pool);
+               if (olc)
+                       pool->cfg.free(olc);
+       }
+       return lc;
+}
+
+static uint32_t
+mlx5_ipool_allocate_from_global(struct mlx5_indexed_pool *pool, int cidx)
+{
+       struct mlx5_indexed_trunk *trunk;
+       struct mlx5_indexed_cache *p, *lc, *olc = NULL;
+       size_t trunk_size = 0;
+       size_t data_size;
+       uint32_t cur_max_idx, trunk_idx, trunk_n;
+       uint32_t fetch_size, ts_idx, i;
+       int n_grow;
+
+check_again:
+       p = NULL;
+       fetch_size = 0;
+       /*
+        * Fetch new index from global if possible. First round local
+        * cache will be NULL.
+        */
+       lc = pool->cache[cidx]->lc;
+       mlx5_ipool_lock(pool);
+       /* Try to update local cache first. */
+       if (likely(pool->gc)) {
+               if (lc != pool->gc) {
+                       if (lc && !(--lc->ref_cnt))
+                               olc = lc;
+                       lc = pool->gc;
+                       lc->ref_cnt++;
+                       pool->cache[cidx]->lc = lc;
+               }
+               if (lc->len) {
+                       /* Use the updated local cache to fetch index. */
+                       fetch_size = pool->cfg.per_core_cache >> 2;
+                       if (lc->len < fetch_size)
+                               fetch_size = lc->len;
+                       lc->len -= fetch_size;
+                       memcpy(pool->cache[cidx]->idx, &lc->idx[lc->len],
+                              sizeof(uint32_t) * fetch_size);
+               }
+       }
+       mlx5_ipool_unlock(pool);
+       if (unlikely(olc)) {
+               pool->cfg.free(olc);
+               olc = NULL;
+       }
+       if (fetch_size) {
+               pool->cache[cidx]->len = fetch_size - 1;
+               return pool->cache[cidx]->idx[pool->cache[cidx]->len];
+       }
+       trunk_idx = lc ? __atomic_load_n(&lc->n_trunk_valid,
+                        __ATOMIC_ACQUIRE) : 0;
+       trunk_n = lc ? lc->n_trunk : 0;
+       cur_max_idx = mlx5_trunk_idx_offset_get(pool, trunk_idx);
+       /* Check if index reach maximum. */
+       if (trunk_idx == TRUNK_MAX_IDX ||
+           cur_max_idx >= pool->cfg.max_idx)
+               return 0;
+       /* No enough space in trunk array, resize the trunks array. */
+       if (trunk_idx == trunk_n) {
+               n_grow = trunk_idx ? trunk_idx :
+                            RTE_CACHE_LINE_SIZE / sizeof(void *);
+               cur_max_idx = mlx5_trunk_idx_offset_get(pool, trunk_n + n_grow);
+               /* Resize the trunk array. */
+               p = pool->cfg.malloc(0, ((trunk_idx + n_grow) *
+                       sizeof(struct mlx5_indexed_trunk *)) +
+                       (cur_max_idx * sizeof(uint32_t)) + sizeof(*p),
+                       RTE_CACHE_LINE_SIZE, rte_socket_id());
+               if (!p)
+                       return 0;
+               p->trunks = (struct mlx5_indexed_trunk **)&p->idx[cur_max_idx];
+               if (lc)
+                       memcpy(p->trunks, lc->trunks, trunk_idx *
+                      sizeof(struct mlx5_indexed_trunk *));
+#ifdef RTE_LIBRTE_MLX5_DEBUG
+               memset(RTE_PTR_ADD(p->trunks, trunk_idx * sizeof(void *)), 0,
+                       n_grow * sizeof(void *));
+#endif
+               p->n_trunk_valid = trunk_idx;
+               p->n_trunk = trunk_n + n_grow;
+               p->len = 0;
+       }
+       /* Prepare the new trunk. */
+       trunk_size = sizeof(*trunk);
+       data_size = mlx5_trunk_size_get(pool, trunk_idx);
+       trunk_size += RTE_CACHE_LINE_ROUNDUP(data_size * pool->cfg.size);
+       trunk = pool->cfg.malloc(0, trunk_size,
+                                RTE_CACHE_LINE_SIZE, rte_socket_id());
+       if (unlikely(!trunk)) {
+               pool->cfg.free(p);
+               return 0;
+       }
+       trunk->idx = trunk_idx;
+       trunk->free = data_size;
+       mlx5_ipool_lock(pool);
+       /*
+        * Double check if trunks has been updated or have available index.
+        * During the new trunk allocate, index may still be flushed to the
+        * global cache. So also need to check the pool->gc->len.
+        */
+       if (pool->gc && (lc != pool->gc ||
+           lc->n_trunk_valid != trunk_idx ||
+           pool->gc->len)) {
+               mlx5_ipool_unlock(pool);
+               if (p)
+                       pool->cfg.free(p);
+               pool->cfg.free(trunk);
+               goto check_again;
+       }
+       /* Resize the trunk array and update local cache first.  */
+       if (p) {
+               if (lc && !(--lc->ref_cnt))
+                       olc = lc;
+               lc = p;
+               lc->ref_cnt = 1;
+               pool->cache[cidx]->lc = lc;
+               __atomic_store_n(&pool->gc, p, __ATOMIC_RELAXED);
+       }
+       /* Add trunk to trunks array. */
+       lc->trunks[trunk_idx] = trunk;
+       __atomic_fetch_add(&lc->n_trunk_valid, 1, __ATOMIC_RELAXED);
+       /* Enqueue half of the index to global. */
+       ts_idx = mlx5_trunk_idx_offset_get(pool, trunk_idx) + 1;
+       fetch_size = trunk->free >> 1;
+       for (i = 0; i < fetch_size; i++)
+               lc->idx[i] = ts_idx + i;
+       lc->len = fetch_size;
+       mlx5_ipool_unlock(pool);
+       /* Copy left half - 1 to local cache index array. */
+       pool->cache[cidx]->len = trunk->free - fetch_size - 1;
+       ts_idx += fetch_size;
+       for (i = 0; i < pool->cache[cidx]->len; i++)
+               pool->cache[cidx]->idx[i] = ts_idx + i;
+       if (olc)
+               pool->cfg.free(olc);
+       return ts_idx + i;
+}
+
+static void *
+_mlx5_ipool_get_cache(struct mlx5_indexed_pool *pool, int cidx, uint32_t idx)
+{
+       struct mlx5_indexed_trunk *trunk;
+       struct mlx5_indexed_cache *lc;
+       uint32_t trunk_idx;
+       uint32_t entry_idx;
+
+       MLX5_ASSERT(idx);
+       if (unlikely(!pool->cache[cidx])) {
+               pool->cache[cidx] = pool->cfg.malloc(MLX5_MEM_ZERO,
+                       sizeof(struct mlx5_ipool_per_lcore) +
+                       (pool->cfg.per_core_cache * sizeof(uint32_t)),
+                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
+               if (!pool->cache[cidx]) {
+                       DRV_LOG(ERR, "Ipool cache%d allocate failed\n", cidx);
+                       return NULL;
+               }
+       }
+       lc = mlx5_ipool_update_global_cache(pool, cidx);
+       idx -= 1;
+       trunk_idx = mlx5_trunk_idx_get(pool, idx);
+       trunk = lc->trunks[trunk_idx];
+       MLX5_ASSERT(trunk);
+       entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk_idx);
+       return &trunk->data[entry_idx * pool->cfg.size];
+}
+
+static void *
+mlx5_ipool_get_cache(struct mlx5_indexed_pool *pool, uint32_t idx)
+{
+       void *entry;
+       int cidx;
+
+       cidx = rte_lcore_index(rte_lcore_id());
+       if (unlikely(cidx == -1)) {
+               cidx = RTE_MAX_LCORE;
+               rte_spinlock_lock(&pool->lcore_lock);
+       }
+       entry = _mlx5_ipool_get_cache(pool, cidx, idx);
+       if (unlikely(cidx == RTE_MAX_LCORE))
+               rte_spinlock_unlock(&pool->lcore_lock);
+       return entry;
+}
+
+
+static void *
+_mlx5_ipool_malloc_cache(struct mlx5_indexed_pool *pool, int cidx,
+                        uint32_t *idx)
+{
+       if (unlikely(!pool->cache[cidx])) {
+               pool->cache[cidx] = pool->cfg.malloc(MLX5_MEM_ZERO,
+                       sizeof(struct mlx5_ipool_per_lcore) +
+                       (pool->cfg.per_core_cache * sizeof(uint32_t)),
+                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
+               if (!pool->cache[cidx]) {
+                       DRV_LOG(ERR, "Ipool cache%d allocate failed\n", cidx);
+                       return NULL;
+               }
+       } else if (pool->cache[cidx]->len) {
+               pool->cache[cidx]->len--;
+               *idx = pool->cache[cidx]->idx[pool->cache[cidx]->len];
+               return _mlx5_ipool_get_cache(pool, cidx, *idx);
+       }
+       /* Not enough idx in global cache. Keep fetching from global. */
+       *idx = mlx5_ipool_allocate_from_global(pool, cidx);
+       if (unlikely(!(*idx)))
+               return NULL;
+       return _mlx5_ipool_get_cache(pool, cidx, *idx);
+}
+
+static void *
+mlx5_ipool_malloc_cache(struct mlx5_indexed_pool *pool, uint32_t *idx)
+{
+       void *entry;
+       int cidx;
+
+       cidx = rte_lcore_index(rte_lcore_id());
+       if (unlikely(cidx == -1)) {
+               cidx = RTE_MAX_LCORE;
+               rte_spinlock_lock(&pool->lcore_lock);
+       }
+       entry = _mlx5_ipool_malloc_cache(pool, cidx, idx);
+       if (unlikely(cidx == RTE_MAX_LCORE))
+               rte_spinlock_unlock(&pool->lcore_lock);
+       return entry;
+}
+
+static void
+_mlx5_ipool_free_cache(struct mlx5_indexed_pool *pool, int cidx, uint32_t idx)
+{
+       struct mlx5_ipool_per_lcore *ilc;
+       struct mlx5_indexed_cache *gc, *olc = NULL;
+       uint32_t reclaim_num = 0;
+
+       MLX5_ASSERT(idx);
+       /*
+        * When index was allocated on core A but freed on core B. In this
+        * case check if local cache on core B was allocated before.
+        */
+       if (unlikely(!pool->cache[cidx])) {
+               pool->cache[cidx] = pool->cfg.malloc(MLX5_MEM_ZERO,
+                       sizeof(struct mlx5_ipool_per_lcore) +
+                       (pool->cfg.per_core_cache * sizeof(uint32_t)),
+                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
+               if (!pool->cache[cidx]) {
+                       DRV_LOG(ERR, "Ipool cache%d allocate failed\n", cidx);
+                       return;
+               }
+       }
+       /* Try to enqueue to local index cache. */
+       if (pool->cache[cidx]->len < pool->cfg.per_core_cache) {
+               pool->cache[cidx]->idx[pool->cache[cidx]->len] = idx;
+               pool->cache[cidx]->len++;
+               return;
+       }
+       ilc = pool->cache[cidx];
+       reclaim_num = pool->cfg.per_core_cache >> 2;
+       ilc->len -= reclaim_num;
+       /* Local index cache full, try with global index cache. */
+       mlx5_ipool_lock(pool);
+       gc = pool->gc;
+       if (ilc->lc != gc) {
+               if (!(--ilc->lc->ref_cnt))
+                       olc = ilc->lc;
+               gc->ref_cnt++;
+               ilc->lc = gc;
+       }
+       memcpy(&gc->idx[gc->len], &ilc->idx[ilc->len],
+              reclaim_num * sizeof(uint32_t));
+       gc->len += reclaim_num;
+       mlx5_ipool_unlock(pool);
+       if (olc)
+               pool->cfg.free(olc);
+       pool->cache[cidx]->idx[pool->cache[cidx]->len] = idx;
+       pool->cache[cidx]->len++;
+}
+
+static void
+mlx5_ipool_free_cache(struct mlx5_indexed_pool *pool, uint32_t idx)
+{
+       int cidx;
+
+       cidx = rte_lcore_index(rte_lcore_id());
+       if (unlikely(cidx == -1)) {
+               cidx = RTE_MAX_LCORE;
+               rte_spinlock_lock(&pool->lcore_lock);
+       }
+       _mlx5_ipool_free_cache(pool, cidx, idx);
+       if (unlikely(cidx == RTE_MAX_LCORE))
+               rte_spinlock_unlock(&pool->lcore_lock);
+}
+
 void *
 mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 {
@@ -226,6 +515,8 @@ mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
        uint32_t iidx = 0;
        void *p;
 
+       if (pool->cfg.per_core_cache)
+               return mlx5_ipool_malloc_cache(pool, idx);
        mlx5_ipool_lock(pool);
        if (pool->free_list == TRUNK_INVALID) {
                /* If no available trunks, grow new. */
@@ -244,10 +535,15 @@ mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
        MLX5_ASSERT(slab);
        iidx += __builtin_ctzll(slab);
        MLX5_ASSERT(iidx != UINT32_MAX);
-       MLX5_ASSERT(iidx < pool->cfg.trunk_size);
+       MLX5_ASSERT(iidx < mlx5_trunk_size_get(pool, trunk->idx));
        rte_bitmap_clear(trunk->bmp, iidx);
        p = &trunk->data[iidx * pool->cfg.size];
-       iidx += trunk->idx * pool->cfg.trunk_size;
+       /*
+        * The ipool index should grow continually from small to big,
+        * some features as metering only accept limited bits of index.
+        * Random index with MSB set may be rejected.
+        */
+       iidx += mlx5_trunk_idx_offset_get(pool, trunk->idx);
        iidx += 1; /* non-zero index. */
        trunk->free--;
 #ifdef POOL_DEBUG
@@ -276,7 +572,7 @@ mlx5_ipool_zmalloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
 {
        void *entry = mlx5_ipool_malloc(pool, idx);
 
-       if (entry)
+       if (entry && pool->cfg.size)
                memset(entry, 0, pool->cfg.size);
        return entry;
 }
@@ -286,21 +582,50 @@ mlx5_ipool_free(struct mlx5_indexed_pool *pool, uint32_t idx)
 {
        struct mlx5_indexed_trunk *trunk;
        uint32_t trunk_idx;
+       uint32_t entry_idx;
 
        if (!idx)
                return;
+       if (pool->cfg.per_core_cache) {
+               mlx5_ipool_free_cache(pool, idx);
+               return;
+       }
        idx -= 1;
        mlx5_ipool_lock(pool);
-       trunk_idx = idx / pool->cfg.trunk_size;
-       if (trunk_idx >= pool->n_trunk_valid)
+       trunk_idx = mlx5_trunk_idx_get(pool, idx);
+       if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
+           (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
                goto out;
        trunk = pool->trunks[trunk_idx];
-       if (!trunk || trunk_idx != trunk->idx ||
-           rte_bitmap_get(trunk->bmp, idx % pool->cfg.trunk_size))
+       if (!trunk)
                goto out;
-       rte_bitmap_set(trunk->bmp, idx % pool->cfg.trunk_size);
+       entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
+       if (trunk_idx != trunk->idx ||
+           rte_bitmap_get(trunk->bmp, entry_idx))
+               goto out;
+       rte_bitmap_set(trunk->bmp, entry_idx);
        trunk->free++;
-       if (trunk->free == 1) {
+       if (pool->cfg.release_mem_en && trunk->free == mlx5_trunk_size_get
+          (pool, trunk->idx)) {
+               if (pool->free_list == trunk->idx)
+                       pool->free_list = trunk->next;
+               if (trunk->next != TRUNK_INVALID)
+                       pool->trunks[trunk->next]->prev = trunk->prev;
+               if (trunk->prev != TRUNK_INVALID)
+                       pool->trunks[trunk->prev]->next = trunk->next;
+               pool->cfg.free(trunk);
+               pool->trunks[trunk_idx] = NULL;
+               pool->n_trunk_valid--;
+#ifdef POOL_DEBUG
+               pool->trunk_avail--;
+               pool->trunk_free++;
+#endif
+               if (pool->n_trunk_valid == 0) {
+                       pool->cfg.free(pool->trunks);
+                       pool->trunks = NULL;
+                       pool->n_trunk = 0;
+               }
+       } else if (trunk->free == 1) {
                /* Put into free trunk list head. */
                MLX5_ASSERT(pool->free_list != trunk->idx);
                trunk->next = pool->free_list;
@@ -326,19 +651,26 @@ mlx5_ipool_get(struct mlx5_indexed_pool *pool, uint32_t idx)
        struct mlx5_indexed_trunk *trunk;
        void *p = NULL;
        uint32_t trunk_idx;
+       uint32_t entry_idx;
 
        if (!idx)
                return NULL;
+       if (pool->cfg.per_core_cache)
+               return mlx5_ipool_get_cache(pool, idx);
        idx -= 1;
        mlx5_ipool_lock(pool);
-       trunk_idx = idx / pool->cfg.trunk_size;
-       if (trunk_idx >= pool->n_trunk_valid)
+       trunk_idx = mlx5_trunk_idx_get(pool, idx);
+       if ((!pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk_valid) ||
+           (pool->cfg.release_mem_en && trunk_idx >= pool->n_trunk))
                goto out;
        trunk = pool->trunks[trunk_idx];
-       if (!trunk || trunk_idx != trunk->idx ||
-           rte_bitmap_get(trunk->bmp, idx % pool->cfg.trunk_size))
+       if (!trunk)
+               goto out;
+       entry_idx = idx - mlx5_trunk_idx_offset_get(pool, trunk->idx);
+       if (trunk_idx != trunk->idx ||
+           rte_bitmap_get(trunk->bmp, entry_idx))
                goto out;
-       p = &trunk->data[(idx % pool->cfg.trunk_size) * pool->cfg.size];
+       p = &trunk->data[entry_idx * pool->cfg.size];
 out:
        mlx5_ipool_unlock(pool);
        return p;
@@ -347,23 +679,134 @@ out:
 int
 mlx5_ipool_destroy(struct mlx5_indexed_pool *pool)
 {
-       struct mlx5_indexed_trunk **trunks;
-       uint32_t i;
+       struct mlx5_indexed_trunk **trunks = NULL;
+       struct mlx5_indexed_cache *gc = pool->gc;
+       uint32_t i, n_trunk_valid = 0;
 
        MLX5_ASSERT(pool);
        mlx5_ipool_lock(pool);
-       trunks = pool->trunks;
-       for (i = 0; i < pool->n_trunk; i++) {
+       if (pool->cfg.per_core_cache) {
+               for (i = 0; i <= RTE_MAX_LCORE; i++) {
+                       /*
+                        * Free only old global cache. Pool gc will be
+                        * freed at last.
+                        */
+                       if (pool->cache[i]) {
+                               if (pool->cache[i]->lc &&
+                                   pool->cache[i]->lc != pool->gc &&
+                                   (!(--pool->cache[i]->lc->ref_cnt)))
+                                       pool->cfg.free(pool->cache[i]->lc);
+                               pool->cfg.free(pool->cache[i]);
+                       }
+               }
+               if (gc) {
+                       trunks = gc->trunks;
+                       n_trunk_valid = gc->n_trunk_valid;
+               }
+       } else {
+               gc = NULL;
+               trunks = pool->trunks;
+               n_trunk_valid = pool->n_trunk_valid;
+       }
+       for (i = 0; i < n_trunk_valid; i++) {
                if (trunks[i])
                        pool->cfg.free(trunks[i]);
        }
-       if (!pool->trunks)
-               pool->cfg.free(pool->trunks);
+       if (!gc && trunks)
+               pool->cfg.free(trunks);
+       if (gc)
+               pool->cfg.free(gc);
        mlx5_ipool_unlock(pool);
-       rte_free(pool);
+       mlx5_free(pool);
        return 0;
 }
 
+void
+mlx5_ipool_flush_cache(struct mlx5_indexed_pool *pool)
+{
+       uint32_t i, j;
+       struct mlx5_indexed_cache *gc;
+       struct rte_bitmap *ibmp;
+       uint32_t bmp_num, mem_size;
+
+       if (!pool->cfg.per_core_cache)
+               return;
+       gc = pool->gc;
+       if (!gc)
+               return;
+       /* Reset bmp. */
+       bmp_num = mlx5_trunk_idx_offset_get(pool, gc->n_trunk_valid);
+       mem_size = rte_bitmap_get_memory_footprint(bmp_num);
+       pool->bmp_mem = pool->cfg.malloc(MLX5_MEM_ZERO, mem_size,
+                                        RTE_CACHE_LINE_SIZE, rte_socket_id());
+       if (!pool->bmp_mem) {
+               DRV_LOG(ERR, "Ipool bitmap mem allocate failed.\n");
+               return;
+       }
+       ibmp = rte_bitmap_init_with_all_set(bmp_num, pool->bmp_mem, mem_size);
+       if (!ibmp) {
+               pool->cfg.free(pool->bmp_mem);
+               pool->bmp_mem = NULL;
+               DRV_LOG(ERR, "Ipool bitmap create failed.\n");
+               return;
+       }
+       pool->ibmp = ibmp;
+       /* Clear global cache. */
+       for (i = 0; i < gc->len; i++)
+               rte_bitmap_clear(ibmp, gc->idx[i] - 1);
+       /* Clear core cache. */
+       for (i = 0; i < RTE_MAX_LCORE + 1; i++) {
+               struct mlx5_ipool_per_lcore *ilc = pool->cache[i];
+
+               if (!ilc)
+                       continue;
+               for (j = 0; j < ilc->len; j++)
+                       rte_bitmap_clear(ibmp, ilc->idx[j] - 1);
+       }
+}
+
+static void *
+mlx5_ipool_get_next_cache(struct mlx5_indexed_pool *pool, uint32_t *pos)
+{
+       struct rte_bitmap *ibmp;
+       uint64_t slab = 0;
+       uint32_t iidx = *pos;
+
+       ibmp = pool->ibmp;
+       if (!ibmp || !rte_bitmap_scan(ibmp, &iidx, &slab)) {
+               if (pool->bmp_mem) {
+                       pool->cfg.free(pool->bmp_mem);
+                       pool->bmp_mem = NULL;
+                       pool->ibmp = NULL;
+               }
+               return NULL;
+       }
+       iidx += __builtin_ctzll(slab);
+       rte_bitmap_clear(ibmp, iidx);
+       iidx++;
+       *pos = iidx;
+       return mlx5_ipool_get_cache(pool, iidx);
+}
+
+void *
+mlx5_ipool_get_next(struct mlx5_indexed_pool *pool, uint32_t *pos)
+{
+       uint32_t idx = *pos;
+       void *entry;
+
+       if (pool->cfg.per_core_cache)
+               return mlx5_ipool_get_next_cache(pool, pos);
+       while (idx <= mlx5_trunk_idx_offset_get(pool, pool->n_trunk)) {
+               entry = mlx5_ipool_get(pool, idx);
+               if (entry) {
+                       *pos = idx;
+                       return entry;
+               }
+               idx++;
+       }
+       return NULL;
+}
+
 void
 mlx5_ipool_dump(struct mlx5_indexed_pool *pool)
 {
@@ -378,3 +821,407 @@ mlx5_ipool_dump(struct mlx5_indexed_pool *pool)
               pool->trunk_empty, pool->trunk_avail, pool->trunk_free);
 #endif
 }
+
+struct mlx5_l3t_tbl *
+mlx5_l3t_create(enum mlx5_l3t_type type)
+{
+       struct mlx5_l3t_tbl *tbl;
+       struct mlx5_indexed_pool_config l3t_ip_cfg = {
+               .trunk_size = 16,
+               .grow_trunk = 6,
+               .grow_shift = 1,
+               .need_lock = 0,
+               .release_mem_en = 1,
+               .malloc = mlx5_malloc,
+               .free = mlx5_free,
+       };
+
+       if (type >= MLX5_L3T_TYPE_MAX) {
+               rte_errno = EINVAL;
+               return NULL;
+       }
+       tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(struct mlx5_l3t_tbl), 1,
+                         SOCKET_ID_ANY);
+       if (!tbl) {
+               rte_errno = ENOMEM;
+               return NULL;
+       }
+       tbl->type = type;
+       switch (type) {
+       case MLX5_L3T_TYPE_WORD:
+               l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_word);
+               l3t_ip_cfg.type = "mlx5_l3t_e_tbl_w";
+               break;
+       case MLX5_L3T_TYPE_DWORD:
+               l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_dword);
+               l3t_ip_cfg.type = "mlx5_l3t_e_tbl_dw";
+               break;
+       case MLX5_L3T_TYPE_QWORD:
+               l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_qword);
+               l3t_ip_cfg.type = "mlx5_l3t_e_tbl_qw";
+               break;
+       default:
+               l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_ptr);
+               l3t_ip_cfg.type = "mlx5_l3t_e_tbl_tpr";
+               break;
+       }
+       rte_spinlock_init(&tbl->sl);
+       tbl->eip = mlx5_ipool_create(&l3t_ip_cfg);
+       if (!tbl->eip) {
+               rte_errno = ENOMEM;
+               mlx5_free(tbl);
+               tbl = NULL;
+       }
+       return tbl;
+}
+
+void
+mlx5_l3t_destroy(struct mlx5_l3t_tbl *tbl)
+{
+       struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
+       uint32_t i, j;
+
+       if (!tbl)
+               return;
+       g_tbl = tbl->tbl;
+       if (g_tbl) {
+               for (i = 0; i < MLX5_L3T_GT_SIZE; i++) {
+                       m_tbl = g_tbl->tbl[i];
+                       if (!m_tbl)
+                               continue;
+                       for (j = 0; j < MLX5_L3T_MT_SIZE; j++) {
+                               if (!m_tbl->tbl[j])
+                                       continue;
+                               MLX5_ASSERT(!((struct mlx5_l3t_entry_word *)
+                                           m_tbl->tbl[j])->ref_cnt);
+                               mlx5_ipool_free(tbl->eip,
+                                               ((struct mlx5_l3t_entry_word *)
+                                               m_tbl->tbl[j])->idx);
+                               m_tbl->tbl[j] = 0;
+                               if (!(--m_tbl->ref_cnt))
+                                       break;
+                       }
+                       MLX5_ASSERT(!m_tbl->ref_cnt);
+                       mlx5_free(g_tbl->tbl[i]);
+                       g_tbl->tbl[i] = 0;
+                       if (!(--g_tbl->ref_cnt))
+                               break;
+               }
+               MLX5_ASSERT(!g_tbl->ref_cnt);
+               mlx5_free(tbl->tbl);
+               tbl->tbl = 0;
+       }
+       mlx5_ipool_destroy(tbl->eip);
+       mlx5_free(tbl);
+}
+
+static int32_t
+__l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
+               union mlx5_l3t_data *data)
+{
+       struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
+       struct mlx5_l3t_entry_word *w_e_tbl;
+       struct mlx5_l3t_entry_dword *dw_e_tbl;
+       struct mlx5_l3t_entry_qword *qw_e_tbl;
+       struct mlx5_l3t_entry_ptr *ptr_e_tbl;
+       void *e_tbl;
+       uint32_t entry_idx;
+
+       g_tbl = tbl->tbl;
+       if (!g_tbl)
+               return -1;
+       m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
+       if (!m_tbl)
+               return -1;
+       e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
+       if (!e_tbl)
+               return -1;
+       entry_idx = idx & MLX5_L3T_ET_MASK;
+       switch (tbl->type) {
+       case MLX5_L3T_TYPE_WORD:
+               w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
+               data->word = w_e_tbl->entry[entry_idx].data;
+               if (w_e_tbl->entry[entry_idx].data)
+                       w_e_tbl->entry[entry_idx].ref_cnt++;
+               break;
+       case MLX5_L3T_TYPE_DWORD:
+               dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
+               data->dword = dw_e_tbl->entry[entry_idx].data;
+               if (dw_e_tbl->entry[entry_idx].data)
+                       dw_e_tbl->entry[entry_idx].ref_cnt++;
+               break;
+       case MLX5_L3T_TYPE_QWORD:
+               qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
+               data->qword = qw_e_tbl->entry[entry_idx].data;
+               if (qw_e_tbl->entry[entry_idx].data)
+                       qw_e_tbl->entry[entry_idx].ref_cnt++;
+               break;
+       default:
+               ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
+               data->ptr = ptr_e_tbl->entry[entry_idx].data;
+               if (ptr_e_tbl->entry[entry_idx].data)
+                       ptr_e_tbl->entry[entry_idx].ref_cnt++;
+               break;
+       }
+       return 0;
+}
+
+int32_t
+mlx5_l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
+                  union mlx5_l3t_data *data)
+{
+       int ret;
+
+       rte_spinlock_lock(&tbl->sl);
+       ret = __l3t_get_entry(tbl, idx, data);
+       rte_spinlock_unlock(&tbl->sl);
+       return ret;
+}
+
+int32_t
+mlx5_l3t_clear_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx)
+{
+       struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
+       struct mlx5_l3t_entry_word *w_e_tbl;
+       struct mlx5_l3t_entry_dword *dw_e_tbl;
+       struct mlx5_l3t_entry_qword *qw_e_tbl;
+       struct mlx5_l3t_entry_ptr *ptr_e_tbl;
+       void *e_tbl;
+       uint32_t entry_idx;
+       uint64_t ref_cnt;
+       int32_t ret = -1;
+
+       rte_spinlock_lock(&tbl->sl);
+       g_tbl = tbl->tbl;
+       if (!g_tbl)
+               goto out;
+       m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
+       if (!m_tbl)
+               goto out;
+       e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
+       if (!e_tbl)
+               goto out;
+       entry_idx = idx & MLX5_L3T_ET_MASK;
+       switch (tbl->type) {
+       case MLX5_L3T_TYPE_WORD:
+               w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
+               MLX5_ASSERT(w_e_tbl->entry[entry_idx].ref_cnt);
+               ret = --w_e_tbl->entry[entry_idx].ref_cnt;
+               if (ret)
+                       goto out;
+               w_e_tbl->entry[entry_idx].data = 0;
+               ref_cnt = --w_e_tbl->ref_cnt;
+               break;
+       case MLX5_L3T_TYPE_DWORD:
+               dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
+               MLX5_ASSERT(dw_e_tbl->entry[entry_idx].ref_cnt);
+               ret = --dw_e_tbl->entry[entry_idx].ref_cnt;
+               if (ret)
+                       goto out;
+               dw_e_tbl->entry[entry_idx].data = 0;
+               ref_cnt = --dw_e_tbl->ref_cnt;
+               break;
+       case MLX5_L3T_TYPE_QWORD:
+               qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
+               MLX5_ASSERT(qw_e_tbl->entry[entry_idx].ref_cnt);
+               ret = --qw_e_tbl->entry[entry_idx].ref_cnt;
+               if (ret)
+                       goto out;
+               qw_e_tbl->entry[entry_idx].data = 0;
+               ref_cnt = --qw_e_tbl->ref_cnt;
+               break;
+       default:
+               ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
+               MLX5_ASSERT(ptr_e_tbl->entry[entry_idx].ref_cnt);
+               ret = --ptr_e_tbl->entry[entry_idx].ref_cnt;
+               if (ret)
+                       goto out;
+               ptr_e_tbl->entry[entry_idx].data = NULL;
+               ref_cnt = --ptr_e_tbl->ref_cnt;
+               break;
+       }
+       if (!ref_cnt) {
+               mlx5_ipool_free(tbl->eip,
+                               ((struct mlx5_l3t_entry_word *)e_tbl)->idx);
+               m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
+                                                                       NULL;
+               if (!(--m_tbl->ref_cnt)) {
+                       mlx5_free(m_tbl);
+                       g_tbl->tbl
+                       [(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] = NULL;
+                       if (!(--g_tbl->ref_cnt)) {
+                               mlx5_free(g_tbl);
+                               tbl->tbl = 0;
+                       }
+               }
+       }
+out:
+       rte_spinlock_unlock(&tbl->sl);
+       return ret;
+}
+
+static int32_t
+__l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
+               union mlx5_l3t_data *data)
+{
+       struct mlx5_l3t_level_tbl *g_tbl, *m_tbl;
+       struct mlx5_l3t_entry_word *w_e_tbl;
+       struct mlx5_l3t_entry_dword *dw_e_tbl;
+       struct mlx5_l3t_entry_qword *qw_e_tbl;
+       struct mlx5_l3t_entry_ptr *ptr_e_tbl;
+       void *e_tbl;
+       uint32_t entry_idx, tbl_idx = 0;
+
+       /* Check the global table, create it if empty. */
+       g_tbl = tbl->tbl;
+       if (!g_tbl) {
+               g_tbl = mlx5_malloc(MLX5_MEM_ZERO,
+                                   sizeof(struct mlx5_l3t_level_tbl) +
+                                   sizeof(void *) * MLX5_L3T_GT_SIZE, 1,
+                                   SOCKET_ID_ANY);
+               if (!g_tbl) {
+                       rte_errno = ENOMEM;
+                       return -1;
+               }
+               tbl->tbl = g_tbl;
+       }
+       /*
+        * Check the middle table, create it if empty. Ref_cnt will be
+        * increased if new sub table created.
+        */
+       m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
+       if (!m_tbl) {
+               m_tbl = mlx5_malloc(MLX5_MEM_ZERO,
+                                   sizeof(struct mlx5_l3t_level_tbl) +
+                                   sizeof(void *) * MLX5_L3T_MT_SIZE, 1,
+                                   SOCKET_ID_ANY);
+               if (!m_tbl) {
+                       rte_errno = ENOMEM;
+                       return -1;
+               }
+               g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK] =
+                                                                       m_tbl;
+               g_tbl->ref_cnt++;
+       }
+       /*
+        * Check the entry table, create it if empty. Ref_cnt will be
+        * increased if new sub entry table created.
+        */
+       e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
+       if (!e_tbl) {
+               e_tbl = mlx5_ipool_zmalloc(tbl->eip, &tbl_idx);
+               if (!e_tbl) {
+                       rte_errno = ENOMEM;
+                       return -1;
+               }
+               ((struct mlx5_l3t_entry_word *)e_tbl)->idx = tbl_idx;
+               m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK] =
+                                                                       e_tbl;
+               m_tbl->ref_cnt++;
+       }
+       entry_idx = idx & MLX5_L3T_ET_MASK;
+       switch (tbl->type) {
+       case MLX5_L3T_TYPE_WORD:
+               w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
+               if (w_e_tbl->entry[entry_idx].data) {
+                       data->word = w_e_tbl->entry[entry_idx].data;
+                       w_e_tbl->entry[entry_idx].ref_cnt++;
+                       rte_errno = EEXIST;
+                       return -1;
+               }
+               w_e_tbl->entry[entry_idx].data = data->word;
+               w_e_tbl->entry[entry_idx].ref_cnt = 1;
+               w_e_tbl->ref_cnt++;
+               break;
+       case MLX5_L3T_TYPE_DWORD:
+               dw_e_tbl = (struct mlx5_l3t_entry_dword *)e_tbl;
+               if (dw_e_tbl->entry[entry_idx].data) {
+                       data->dword = dw_e_tbl->entry[entry_idx].data;
+                       dw_e_tbl->entry[entry_idx].ref_cnt++;
+                       rte_errno = EEXIST;
+                       return -1;
+               }
+               dw_e_tbl->entry[entry_idx].data = data->dword;
+               dw_e_tbl->entry[entry_idx].ref_cnt = 1;
+               dw_e_tbl->ref_cnt++;
+               break;
+       case MLX5_L3T_TYPE_QWORD:
+               qw_e_tbl = (struct mlx5_l3t_entry_qword *)e_tbl;
+               if (qw_e_tbl->entry[entry_idx].data) {
+                       data->qword = qw_e_tbl->entry[entry_idx].data;
+                       qw_e_tbl->entry[entry_idx].ref_cnt++;
+                       rte_errno = EEXIST;
+                       return -1;
+               }
+               qw_e_tbl->entry[entry_idx].data = data->qword;
+               qw_e_tbl->entry[entry_idx].ref_cnt = 1;
+               qw_e_tbl->ref_cnt++;
+               break;
+       default:
+               ptr_e_tbl = (struct mlx5_l3t_entry_ptr *)e_tbl;
+               if (ptr_e_tbl->entry[entry_idx].data) {
+                       data->ptr = ptr_e_tbl->entry[entry_idx].data;
+                       ptr_e_tbl->entry[entry_idx].ref_cnt++;
+                       rte_errno = EEXIST;
+                       return -1;
+               }
+               ptr_e_tbl->entry[entry_idx].data = data->ptr;
+               ptr_e_tbl->entry[entry_idx].ref_cnt = 1;
+               ptr_e_tbl->ref_cnt++;
+               break;
+       }
+       return 0;
+}
+
+int32_t
+mlx5_l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
+                  union mlx5_l3t_data *data)
+{
+       int ret;
+
+       rte_spinlock_lock(&tbl->sl);
+       ret = __l3t_set_entry(tbl, idx, data);
+       rte_spinlock_unlock(&tbl->sl);
+       return ret;
+}
+
+int32_t
+mlx5_l3t_prepare_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
+                      union mlx5_l3t_data *data,
+                      mlx5_l3t_alloc_callback_fn cb, void *ctx)
+{
+       int32_t ret;
+
+       rte_spinlock_lock(&tbl->sl);
+       /* Check if entry data is ready. */
+       ret = __l3t_get_entry(tbl, idx, data);
+       if (!ret) {
+               switch (tbl->type) {
+               case MLX5_L3T_TYPE_WORD:
+                       if (data->word)
+                               goto out;
+                       break;
+               case MLX5_L3T_TYPE_DWORD:
+                       if (data->dword)
+                               goto out;
+                       break;
+               case MLX5_L3T_TYPE_QWORD:
+                       if (data->qword)
+                               goto out;
+                       break;
+               default:
+                       if (data->ptr)
+                               goto out;
+                       break;
+               }
+       }
+       /* Entry data is not ready, use user callback to create it. */
+       ret = cb(ctx, data);
+       if (ret)
+               goto out;
+       /* Save the new allocated data to entry. */
+       ret = __l3t_set_entry(tbl, idx, data);
+out:
+       rte_spinlock_unlock(&tbl->sl);
+       return ret;
+}