net/mlx5: fix setting VF default MAC through representor
[dpdk.git] / drivers / net / mlx5 / mlx5_utils.c
index fefe833..a39b5ed 100644 (file)
@@ -9,25 +9,44 @@
 
 #include "mlx5_utils.h"
 
+/********************* Hash List **********************/
+
+static struct mlx5_hlist_entry *
+mlx5_hlist_default_create_cb(struct mlx5_hlist *h, uint64_t key __rte_unused,
+                            void *ctx __rte_unused)
+{
+       return mlx5_malloc(MLX5_MEM_ZERO, h->entry_sz, 0, SOCKET_ID_ANY);
+}
+
+static void
+mlx5_hlist_default_remove_cb(struct mlx5_hlist *h __rte_unused,
+                            struct mlx5_hlist_entry *entry)
+{
+       mlx5_free(entry);
+}
+
 struct mlx5_hlist *
-mlx5_hlist_create(const char *name, uint32_t size)
+mlx5_hlist_create(const char *name, uint32_t size, uint32_t entry_size,
+                 uint32_t flags, mlx5_hlist_create_cb cb_create,
+                 mlx5_hlist_match_cb cb_match, mlx5_hlist_remove_cb cb_remove)
 {
        struct mlx5_hlist *h;
        uint32_t act_size;
        uint32_t alloc_size;
+       uint32_t i;
 
-       if (!size)
+       if (!size || !cb_match || (!cb_create ^ !cb_remove))
                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 ".", size, act_size);
+               DRV_LOG(DEBUG, "Size 0x%" PRIX32 " is not power of 2, "
+                       "will be aligned to 0x%" PRIX32 ".", size, act_size);
        } else {
                act_size = size;
        }
        alloc_size = sizeof(struct mlx5_hlist) +
-                    sizeof(struct mlx5_hlist_head) * act_size;
+                    sizeof(struct mlx5_hlist_bucket) * act_size;
        /* Using zmalloc, then no need to initialize the heads. */
        h = mlx5_malloc(MLX5_MEM_ZERO, alloc_size, RTE_CACHE_LINE_SIZE,
                        SOCKET_ID_ANY);
@@ -40,107 +59,150 @@ mlx5_hlist_create(const char *name, uint32_t size)
                snprintf(h->name, MLX5_HLIST_NAMESIZE, "%s", name);
        h->table_sz = act_size;
        h->mask = act_size - 1;
+       h->entry_sz = entry_size;
+       h->direct_key = !!(flags & MLX5_HLIST_DIRECT_KEY);
+       h->write_most = !!(flags & MLX5_HLIST_WRITE_MOST);
+       h->cb_create = cb_create ? cb_create : mlx5_hlist_default_create_cb;
+       h->cb_match = cb_match;
+       h->cb_remove = cb_remove ? cb_remove : mlx5_hlist_default_remove_cb;
+       for (i = 0; i < act_size; i++)
+               rte_rwlock_init(&h->buckets[i].lock);
        DRV_LOG(DEBUG, "Hash list with %s size 0x%" PRIX32 " is created.",
                h->name, act_size);
        return h;
 }
 
-struct mlx5_hlist_entry *
-mlx5_hlist_lookup(struct mlx5_hlist *h, uint64_t key)
+static struct mlx5_hlist_entry *
+__hlist_lookup(struct mlx5_hlist *h, uint64_t key, uint32_t idx,
+              void *ctx, bool reuse)
 {
-       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];
+       first = &h->buckets[idx].head;
        LIST_FOREACH(node, first, next) {
-               if (node->key == key)
-                       return node;
+               if (!h->cb_match(h, node, key, ctx)) {
+                       if (reuse) {
+                               __atomic_add_fetch(&node->ref_cnt, 1,
+                                                  __ATOMIC_RELAXED);
+                               DRV_LOG(DEBUG, "Hash list %s entry %p "
+                                       "reuse: %u.",
+                                       h->name, (void *)node, node->ref_cnt);
+                       }
+                       break;
+               }
        }
-       return NULL;
+       return node;
 }
 
-int
-mlx5_hlist_insert(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry)
+static struct mlx5_hlist_entry *
+hlist_lookup(struct mlx5_hlist *h, uint64_t key, uint32_t idx,
+            void *ctx, bool reuse)
 {
-       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;
+       MLX5_ASSERT(h);
+       rte_rwlock_read_lock(&h->buckets[idx].lock);
+       node = __hlist_lookup(h, key, idx, ctx, reuse);
+       rte_rwlock_read_unlock(&h->buckets[idx].lock);
+       return node;
 }
 
 struct mlx5_hlist_entry *
-mlx5_hlist_lookup_ex(struct mlx5_hlist *h, uint64_t key,
-                    mlx5_hlist_match_callback_fn cb, void *ctx)
+mlx5_hlist_lookup(struct mlx5_hlist *h, uint64_t key, void *ctx)
 {
        uint32_t idx;
-       struct mlx5_hlist_head *first;
-       struct mlx5_hlist_entry *node;
 
-       MLX5_ASSERT(h && cb && ctx);
-       idx = rte_hash_crc_8byte(key, 0) & h->mask;
-       first = &h->heads[idx];
-       LIST_FOREACH(node, first, next) {
-               if (!cb(node, ctx))
-                       return node;
-       }
-       return NULL;
+       if (h->direct_key)
+               idx = (uint32_t)(key & h->mask);
+       else
+               idx = rte_hash_crc_8byte(key, 0) & h->mask;
+       return hlist_lookup(h, key, idx, ctx, false);
 }
 
-int
-mlx5_hlist_insert_ex(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry,
-                    mlx5_hlist_match_callback_fn cb, void *ctx)
+struct mlx5_hlist_entry*
+mlx5_hlist_register(struct mlx5_hlist *h, uint64_t key, void *ctx)
 {
        uint32_t idx;
        struct mlx5_hlist_head *first;
-       struct mlx5_hlist_entry *node;
+       struct mlx5_hlist_bucket *b;
+       struct mlx5_hlist_entry *entry;
+       uint32_t prev_gen_cnt = 0;
 
-       MLX5_ASSERT(h && entry && cb && ctx);
-       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 (!cb(node, ctx))
-                       return -EEXIST;
+       if (h->direct_key)
+               idx = (uint32_t)(key & h->mask);
+       else
+               idx = rte_hash_crc_8byte(key, 0) & h->mask;
+       MLX5_ASSERT(h);
+       b = &h->buckets[idx];
+       /* Use write lock directly for write-most list. */
+       if (!h->write_most) {
+               prev_gen_cnt = __atomic_load_n(&b->gen_cnt, __ATOMIC_ACQUIRE);
+               entry = hlist_lookup(h, key, idx, ctx, true);
+               if (entry)
+                       return entry;
+       }
+       rte_rwlock_write_lock(&b->lock);
+       /* Check if the list changed by other threads. */
+       if (h->write_most ||
+           prev_gen_cnt != __atomic_load_n(&b->gen_cnt, __ATOMIC_ACQUIRE)) {
+               entry = __hlist_lookup(h, key, idx, ctx, true);
+               if (entry)
+                       goto done;
+       }
+       first = &b->head;
+       entry = h->cb_create(h, key, ctx);
+       if (!entry) {
+               rte_errno = ENOMEM;
+               DRV_LOG(DEBUG, "Can't allocate hash list %s entry.", h->name);
+               goto done;
        }
+       entry->idx = idx;
+       entry->ref_cnt = 1;
        LIST_INSERT_HEAD(first, entry, next);
-       return 0;
+       __atomic_add_fetch(&b->gen_cnt, 1, __ATOMIC_ACQ_REL);
+       DRV_LOG(DEBUG, "Hash list %s entry %p new: %u.",
+               h->name, (void *)entry, entry->ref_cnt);
+done:
+       rte_rwlock_write_unlock(&b->lock);
+       return entry;
 }
 
-void
-mlx5_hlist_remove(struct mlx5_hlist *h __rte_unused,
-                 struct mlx5_hlist_entry *entry)
+int
+mlx5_hlist_unregister(struct mlx5_hlist *h, struct mlx5_hlist_entry *entry)
 {
-       MLX5_ASSERT(entry && entry->next.le_prev);
+       uint32_t idx = entry->idx;
+
+       rte_rwlock_write_lock(&h->buckets[idx].lock);
+       MLX5_ASSERT(entry && entry->ref_cnt && entry->next.le_prev);
+       DRV_LOG(DEBUG, "Hash list %s entry %p deref: %u.",
+               h->name, (void *)entry, entry->ref_cnt);
+       if (--entry->ref_cnt) {
+               rte_rwlock_write_unlock(&h->buckets[idx].lock);
+               return 1;
+       }
        LIST_REMOVE(entry, next);
        /* Set to NULL to get rid of removing action for more than once. */
        entry->next.le_prev = NULL;
+       h->cb_remove(h, entry);
+       rte_rwlock_write_unlock(&h->buckets[idx].lock);
+       DRV_LOG(DEBUG, "Hash list %s entry %p removed.",
+               h->name, (void *)entry);
+       return 0;
 }
 
 void
-mlx5_hlist_destroy(struct mlx5_hlist *h,
-                  mlx5_hlist_destroy_callback_fn cb, void *ctx)
+mlx5_hlist_destroy(struct mlx5_hlist *h)
 {
        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]);
+               /* No LIST_FOREACH_SAFE, using while instead. */
+               while (!LIST_EMPTY(&h->buckets[idx].head)) {
+                       entry = LIST_FIRST(&h->buckets[idx].head);
                        LIST_REMOVE(entry, next);
                        /*
                         * The owner of whole element which contains data entry
@@ -150,15 +212,174 @@ mlx5_hlist_destroy(struct mlx5_hlist *h,
                         * the beginning). Or else the default free function
                         * will be used.
                         */
-                       if (cb)
-                               cb(entry, ctx);
-                       else
-                               mlx5_free(entry);
+                       h->cb_remove(h, entry);
                }
        }
        mlx5_free(h);
 }
 
+/********************* Cache list ************************/
+
+static struct mlx5_cache_entry *
+mlx5_clist_default_create_cb(struct mlx5_cache_list *list,
+                            struct mlx5_cache_entry *entry __rte_unused,
+                            void *ctx __rte_unused)
+{
+       return mlx5_malloc(MLX5_MEM_ZERO, list->entry_sz, 0, SOCKET_ID_ANY);
+}
+
+static void
+mlx5_clist_default_remove_cb(struct mlx5_cache_list *list __rte_unused,
+                            struct mlx5_cache_entry *entry)
+{
+       mlx5_free(entry);
+}
+
+int
+mlx5_cache_list_init(struct mlx5_cache_list *list, const char *name,
+                    uint32_t entry_size, void *ctx,
+                    mlx5_cache_create_cb cb_create,
+                    mlx5_cache_match_cb cb_match,
+                    mlx5_cache_remove_cb cb_remove)
+{
+       MLX5_ASSERT(list);
+       if (!cb_match || (!cb_create ^ !cb_remove))
+               return -1;
+       if (name)
+               snprintf(list->name, sizeof(list->name), "%s", name);
+       list->entry_sz = entry_size;
+       list->ctx = ctx;
+       list->cb_create = cb_create ? cb_create : mlx5_clist_default_create_cb;
+       list->cb_match = cb_match;
+       list->cb_remove = cb_remove ? cb_remove : mlx5_clist_default_remove_cb;
+       rte_rwlock_init(&list->lock);
+       DRV_LOG(DEBUG, "Cache list %s initialized.", list->name);
+       LIST_INIT(&list->head);
+       return 0;
+}
+
+static struct mlx5_cache_entry *
+__cache_lookup(struct mlx5_cache_list *list, void *ctx, bool reuse)
+{
+       struct mlx5_cache_entry *entry;
+
+       LIST_FOREACH(entry, &list->head, next) {
+               if (list->cb_match(list, entry, ctx))
+                       continue;
+               if (reuse) {
+                       __atomic_add_fetch(&entry->ref_cnt, 1,
+                                          __ATOMIC_RELAXED);
+                       DRV_LOG(DEBUG, "Cache list %s entry %p ref++: %u.",
+                               list->name, (void *)entry, entry->ref_cnt);
+               }
+               break;
+       }
+       return entry;
+}
+
+static struct mlx5_cache_entry *
+cache_lookup(struct mlx5_cache_list *list, void *ctx, bool reuse)
+{
+       struct mlx5_cache_entry *entry;
+
+       rte_rwlock_read_lock(&list->lock);
+       entry = __cache_lookup(list, ctx, reuse);
+       rte_rwlock_read_unlock(&list->lock);
+       return entry;
+}
+
+struct mlx5_cache_entry *
+mlx5_cache_lookup(struct mlx5_cache_list *list, void *ctx)
+{
+       return cache_lookup(list, ctx, false);
+}
+
+struct mlx5_cache_entry *
+mlx5_cache_register(struct mlx5_cache_list *list, void *ctx)
+{
+       struct mlx5_cache_entry *entry;
+       uint32_t prev_gen_cnt = 0;
+
+       MLX5_ASSERT(list);
+       prev_gen_cnt = __atomic_load_n(&list->gen_cnt, __ATOMIC_ACQUIRE);
+       /* Lookup with read lock, reuse if found. */
+       entry = cache_lookup(list, ctx, true);
+       if (entry)
+               return entry;
+       /* Not found, append with write lock - block read from other threads. */
+       rte_rwlock_write_lock(&list->lock);
+       /* If list changed by other threads before lock, search again. */
+       if (prev_gen_cnt != __atomic_load_n(&list->gen_cnt, __ATOMIC_ACQUIRE)) {
+               /* Lookup and reuse w/o read lock. */
+               entry = __cache_lookup(list, ctx, true);
+               if (entry)
+                       goto done;
+       }
+       entry = list->cb_create(list, entry, ctx);
+       if (!entry) {
+               DRV_LOG(ERR, "Failed to init cache list %s entry %p.",
+                       list->name, (void *)entry);
+               goto done;
+       }
+       entry->ref_cnt = 1;
+       LIST_INSERT_HEAD(&list->head, entry, next);
+       __atomic_add_fetch(&list->gen_cnt, 1, __ATOMIC_RELEASE);
+       __atomic_add_fetch(&list->count, 1, __ATOMIC_ACQUIRE);
+       DRV_LOG(DEBUG, "Cache list %s entry %p new: %u.",
+               list->name, (void *)entry, entry->ref_cnt);
+done:
+       rte_rwlock_write_unlock(&list->lock);
+       return entry;
+}
+
+int
+mlx5_cache_unregister(struct mlx5_cache_list *list,
+                     struct mlx5_cache_entry *entry)
+{
+       rte_rwlock_write_lock(&list->lock);
+       MLX5_ASSERT(entry && entry->next.le_prev);
+       DRV_LOG(DEBUG, "Cache list %s entry %p ref--: %u.",
+               list->name, (void *)entry, entry->ref_cnt);
+       if (--entry->ref_cnt) {
+               rte_rwlock_write_unlock(&list->lock);
+               return 1;
+       }
+       __atomic_add_fetch(&list->gen_cnt, 1, __ATOMIC_ACQUIRE);
+       __atomic_sub_fetch(&list->count, 1, __ATOMIC_ACQUIRE);
+       LIST_REMOVE(entry, next);
+       list->cb_remove(list, entry);
+       rte_rwlock_write_unlock(&list->lock);
+       DRV_LOG(DEBUG, "Cache list %s entry %p removed.",
+               list->name, (void *)entry);
+       return 0;
+}
+
+void
+mlx5_cache_list_destroy(struct mlx5_cache_list *list)
+{
+       struct mlx5_cache_entry *entry;
+
+       MLX5_ASSERT(list);
+       /* no LIST_FOREACH_SAFE, using while instead */
+       while (!LIST_EMPTY(&list->head)) {
+               entry = LIST_FIRST(&list->head);
+               LIST_REMOVE(entry, next);
+               list->cb_remove(list, entry);
+               DRV_LOG(DEBUG, "Cache list %s entry %p destroyed.",
+                       list->name, (void *)entry);
+       }
+       memset(list, 0, sizeof(*list));
+}
+
+uint32_t
+mlx5_cache_list_get_entry_num(struct mlx5_cache_list *list)
+{
+       MLX5_ASSERT(list);
+       return __atomic_load_n(&list->count, __ATOMIC_RELAXED);
+}
+
+/********************* Indexed pool **********************/
+
 static inline void
 mlx5_ipool_lock(struct mlx5_indexed_pool *pool)
 {
@@ -230,7 +451,7 @@ 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->trunk_size && ((cfg->trunk_size & (cfg->trunk_size - 1)) ||
            ((__builtin_ffs(cfg->trunk_size) + TRUNK_IDX_BITS) > 32))))
                return NULL;
@@ -362,6 +583,11 @@ mlx5_ipool_malloc(struct mlx5_indexed_pool *pool, uint32_t *idx)
        MLX5_ASSERT(iidx < mlx5_trunk_size_get(pool, trunk->idx));
        rte_bitmap_clear(trunk->bmp, iidx);
        p = &trunk->data[iidx * pool->cfg.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--;
@@ -391,7 +617,7 @@ 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;
 }
@@ -551,26 +777,23 @@ mlx5_l3t_create(enum mlx5_l3t_type type)
        tbl->type = type;
        switch (type) {
        case MLX5_L3T_TYPE_WORD:
-               l3t_ip_cfg.size = sizeof(struct mlx5_l3t_entry_word) +
-                                 sizeof(uint16_t) * MLX5_L3T_ET_SIZE;
+               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) +
-                                 sizeof(uint32_t) * MLX5_L3T_ET_SIZE;
+               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) +
-                                 sizeof(uint64_t) * MLX5_L3T_ET_SIZE;
+               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) +
-                                 sizeof(void *) * MLX5_L3T_ET_SIZE;
+               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;
@@ -620,11 +843,15 @@ mlx5_l3t_destroy(struct mlx5_l3t_tbl *tbl)
        mlx5_free(tbl);
 }
 
-uint32_t
-mlx5_l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
-                  union mlx5_l3t_data *data)
+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;
 
@@ -640,26 +867,46 @@ mlx5_l3t_get_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
        entry_idx = idx & MLX5_L3T_ET_MASK;
        switch (tbl->type) {
        case MLX5_L3T_TYPE_WORD:
-               data->word = ((struct mlx5_l3t_entry_word *)e_tbl)->entry
-                            [entry_idx];
+               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:
-               data->dword = ((struct mlx5_l3t_entry_dword *)e_tbl)->entry
-                            [entry_idx];
+               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:
-               data->qword = ((struct mlx5_l3t_entry_qword *)e_tbl)->entry
-                             [entry_idx];
+               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:
-               data->ptr = ((struct mlx5_l3t_entry_ptr *)e_tbl)->entry
-                           [entry_idx];
+               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;
 }
 
-void
+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;
@@ -670,36 +917,54 @@ mlx5_l3t_clear_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx)
        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)
-               return;
+               goto out;
        m_tbl = g_tbl->tbl[(idx >> MLX5_L3T_GT_OFFSET) & MLX5_L3T_GT_MASK];
        if (!m_tbl)
-               return;
+               goto out;
        e_tbl = m_tbl->tbl[(idx >> MLX5_L3T_MT_OFFSET) & MLX5_L3T_MT_MASK];
        if (!e_tbl)
-               return;
+               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;
-               w_e_tbl->entry[entry_idx] = 0;
+               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;
-               dw_e_tbl->entry[entry_idx] = 0;
+               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;
-               qw_e_tbl->entry[entry_idx] = 0;
+               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;
-               ptr_e_tbl->entry[entry_idx] = NULL;
+               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;
        }
@@ -718,11 +983,14 @@ mlx5_l3t_clear_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx)
                        }
                }
        }
+out:
+       rte_spinlock_unlock(&tbl->sl);
+       return ret;
 }
 
-uint32_t
-mlx5_l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
-                  union mlx5_l3t_data *data)
+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;
@@ -783,24 +1051,105 @@ mlx5_l3t_set_entry(struct mlx5_l3t_tbl *tbl, uint32_t idx,
        switch (tbl->type) {
        case MLX5_L3T_TYPE_WORD:
                w_e_tbl = (struct mlx5_l3t_entry_word *)e_tbl;
-               w_e_tbl->entry[entry_idx] = data->word;
+               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;
-               dw_e_tbl->entry[entry_idx] = data->dword;
+               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;
-               qw_e_tbl->entry[entry_idx] = data->qword;
+               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;
-               ptr_e_tbl->entry[entry_idx] = data->ptr;
+               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;
+}