lpm6: add incremental update on delete
authorAlex Kiselev <alex@therouter.net>
Tue, 18 Sep 2018 10:57:17 +0000 (13:57 +0300)
committerThomas Monjalon <thomas@monjalon.net>
Wed, 19 Sep 2018 15:11:37 +0000 (17:11 +0200)
Rework the delete function and add additional
internal data structures to support incremental
LPM tree update rather than full tree rebuild.

Signed-off-by: Alex Kiselev <alex@therouter.net>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
lib/librte_lpm/rte_lpm6.c

index fd86b05..6212003 100644 (file)
@@ -41,6 +41,7 @@
 #define BYTES2_SIZE                              16
 
 #define RULE_HASH_TABLE_EXTRA_SPACE              64
+#define TBL24_IND                        UINT32_MAX
 
 #define lpm6_tbl8_gindex next_hop
 
@@ -81,6 +82,17 @@ struct rte_lpm6_rule_key {
        uint8_t depth; /**< Rule depth. */
 };
 
+/* Header of tbl8 */
+struct rte_lpm_tbl8_hdr {
+       uint32_t owner_tbl_ind; /**< owner table: TBL24_IND if owner is tbl24,
+                                 *  otherwise index of tbl8
+                                 */
+       uint32_t owner_entry_ind; /**< index of the owner table entry where
+                                   *  pointer to the tbl8 is stored
+                                   */
+       uint32_t ref_cnt; /**< table reference counter */
+};
+
 /** LPM6 structure. */
 struct rte_lpm6 {
        /* LPM metadata. */
@@ -88,12 +100,17 @@ struct rte_lpm6 {
        uint32_t max_rules;              /**< Max number of rules. */
        uint32_t used_rules;             /**< Used rules so far. */
        uint32_t number_tbl8s;           /**< Number of tbl8s to allocate. */
-       uint32_t next_tbl8;              /**< Next tbl8 to be used. */
 
        /* LPM Tables. */
        struct rte_hash *rules_tbl; /**< LPM rules. */
        struct rte_lpm6_tbl_entry tbl24[RTE_LPM6_TBL24_NUM_ENTRIES]
                        __rte_cache_aligned; /**< LPM tbl24 table. */
+
+       uint32_t *tbl8_pool; /**< pool of indexes of free tbl8s */
+       uint32_t tbl8_pool_pos; /**< current position in the tbl8 pool */
+
+       struct rte_lpm_tbl8_hdr *tbl8_hdrs; /* array of tbl8 headers */
+
        struct rte_lpm6_tbl_entry tbl8[0]
                        __rte_cache_aligned; /**< LPM tbl8 table. */
 };
@@ -142,6 +159,59 @@ rule_hash(const void *data, __rte_unused uint32_t data_len,
        return rte_jhash(data, sizeof(struct rte_lpm6_rule_key), init_val);
 }
 
+/*
+ * Init pool of free tbl8 indexes
+ */
+static void
+tbl8_pool_init(struct rte_lpm6 *lpm)
+{
+       uint32_t i;
+
+       /* put entire range of indexes to the tbl8 pool */
+       for (i = 0; i < lpm->number_tbl8s; i++)
+               lpm->tbl8_pool[i] = i;
+
+       lpm->tbl8_pool_pos = 0;
+}
+
+/*
+ * Get an index of a free tbl8 from the pool
+ */
+static inline uint32_t
+tbl8_get(struct rte_lpm6 *lpm, uint32_t *tbl8_ind)
+{
+       if (lpm->tbl8_pool_pos == lpm->number_tbl8s)
+               /* no more free tbl8 */
+               return -ENOSPC;
+
+       /* next index */
+       *tbl8_ind = lpm->tbl8_pool[lpm->tbl8_pool_pos++];
+       return 0;
+}
+
+/*
+ * Put an index of a free tbl8 back to the pool
+ */
+static inline uint32_t
+tbl8_put(struct rte_lpm6 *lpm, uint32_t tbl8_ind)
+{
+       if (lpm->tbl8_pool_pos == 0)
+               /* pool is full */
+               return -ENOSPC;
+
+       lpm->tbl8_pool[--lpm->tbl8_pool_pos] = tbl8_ind;
+       return 0;
+}
+
+/*
+ * Returns number of tbl8s available in the pool
+ */
+static inline uint32_t
+tbl8_available(struct rte_lpm6 *lpm)
+{
+       return lpm->number_tbl8s - lpm->tbl8_pool_pos;
+}
+
 /*
  * Init a rule key.
  *       note that ip must be already masked
@@ -182,6 +252,8 @@ rte_lpm6_create(const char *name, int socket_id,
        uint64_t mem_size;
        struct rte_lpm6_list *lpm_list;
        struct rte_hash *rules_tbl = NULL;
+       uint32_t *tbl8_pool = NULL;
+       struct rte_lpm_tbl8_hdr *tbl8_hdrs = NULL;
 
        lpm_list = RTE_TAILQ_CAST(rte_lpm6_tailq.head, rte_lpm6_list);
 
@@ -216,6 +288,28 @@ rte_lpm6_create(const char *name, int socket_id,
                goto fail_wo_unlock;
        }
 
+       /* allocate tbl8 indexes pool */
+       tbl8_pool = rte_malloc(NULL,
+                       sizeof(uint32_t) * config->number_tbl8s,
+                       RTE_CACHE_LINE_SIZE);
+       if (tbl8_pool == NULL) {
+               RTE_LOG(ERR, LPM, "LPM tbl8 pool allocation failed: %s (%d)",
+                                 rte_strerror(rte_errno), rte_errno);
+               rte_errno = ENOMEM;
+               goto fail_wo_unlock;
+       }
+
+       /* allocate tbl8 headers */
+       tbl8_hdrs = rte_malloc(NULL,
+                       sizeof(struct rte_lpm_tbl8_hdr) * config->number_tbl8s,
+                       RTE_CACHE_LINE_SIZE);
+       if (tbl8_hdrs == NULL) {
+               RTE_LOG(ERR, LPM, "LPM tbl8 headers allocation failed: %s (%d)",
+                                 rte_strerror(rte_errno), rte_errno);
+               rte_errno = ENOMEM;
+               goto fail_wo_unlock;
+       }
+
        snprintf(mem_name, sizeof(mem_name), "LPM_%s", name);
 
        /* Determine the amount of memory to allocate. */
@@ -260,6 +354,11 @@ rte_lpm6_create(const char *name, int socket_id,
        lpm->number_tbl8s = config->number_tbl8s;
        snprintf(lpm->name, sizeof(lpm->name), "%s", name);
        lpm->rules_tbl = rules_tbl;
+       lpm->tbl8_pool = tbl8_pool;
+       lpm->tbl8_hdrs = tbl8_hdrs;
+
+       /* init the stack */
+       tbl8_pool_init(lpm);
 
        te->data = (void *) lpm;
 
@@ -271,6 +370,8 @@ fail:
        rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 
 fail_wo_unlock:
+       rte_free(tbl8_hdrs);
+       rte_free(tbl8_pool);
        rte_hash_free(rules_tbl);
 
        return NULL;
@@ -332,6 +433,8 @@ rte_lpm6_free(struct rte_lpm6 *lpm)
 
        rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 
+       rte_free(lpm->tbl8_hdrs);
+       rte_free(lpm->tbl8_pool);
        rte_hash_free(lpm->rules_tbl);
        rte_free(lpm);
        rte_free(te);
@@ -420,24 +523,24 @@ rule_add(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth, uint32_t next_hop)
  * in the IP address returns a match.
  */
 static void
-expand_rule(struct rte_lpm6 *lpm, uint32_t tbl8_gindex, uint8_t depth,
-               uint32_t next_hop)
+expand_rule(struct rte_lpm6 *lpm, uint32_t tbl8_gindex, uint8_t old_depth,
+               uint8_t new_depth, uint32_t next_hop, uint8_t valid)
 {
        uint32_t tbl8_group_end, tbl8_gindex_next, j;
 
        tbl8_group_end = tbl8_gindex + RTE_LPM6_TBL8_GROUP_NUM_ENTRIES;
 
        struct rte_lpm6_tbl_entry new_tbl8_entry = {
-               .valid = VALID,
-               .valid_group = VALID,
-               .depth = depth,
+               .valid = valid,
+               .valid_group = valid,
+               .depth = new_depth,
                .next_hop = next_hop,
                .ext_entry = 0,
        };
 
        for (j = tbl8_gindex; j < tbl8_group_end; j++) {
                if (!lpm->tbl8[j].valid || (lpm->tbl8[j].ext_entry == 0
-                               && lpm->tbl8[j].depth <= depth)) {
+                               && lpm->tbl8[j].depth <= old_depth)) {
 
                        lpm->tbl8[j] = new_tbl8_entry;
 
@@ -445,11 +548,101 @@ expand_rule(struct rte_lpm6 *lpm, uint32_t tbl8_gindex, uint8_t depth,
 
                        tbl8_gindex_next = lpm->tbl8[j].lpm6_tbl8_gindex
                                        * RTE_LPM6_TBL8_GROUP_NUM_ENTRIES;
-                       expand_rule(lpm, tbl8_gindex_next, depth, next_hop);
+                       expand_rule(lpm, tbl8_gindex_next, old_depth, new_depth,
+                                       next_hop, valid);
                }
        }
 }
 
+/*
+ * Init a tbl8 header
+ */
+static inline void
+init_tbl8_header(struct rte_lpm6 *lpm, uint32_t tbl_ind,
+               uint32_t owner_tbl_ind, uint32_t owner_entry_ind)
+{
+       struct rte_lpm_tbl8_hdr *tbl_hdr = &lpm->tbl8_hdrs[tbl_ind];
+       tbl_hdr->owner_tbl_ind = owner_tbl_ind;
+       tbl_hdr->owner_entry_ind = owner_entry_ind;
+       tbl_hdr->ref_cnt = 0;
+}
+
+/*
+ * Calculate index to the table based on the number and position
+ * of the bytes being inspected in this step.
+ */
+static uint32_t
+get_bitshift(const uint8_t *ip, uint8_t first_byte, uint8_t bytes)
+{
+       uint32_t entry_ind, i;
+       int8_t bitshift;
+
+       entry_ind = 0;
+       for (i = first_byte; i < (uint32_t)(first_byte + bytes); i++) {
+               bitshift = (int8_t)((bytes - i)*BYTE_SIZE);
+
+               if (bitshift < 0)
+                       bitshift = 0;
+               entry_ind = entry_ind | ip[i-1] << bitshift;
+       }
+
+       return entry_ind;
+}
+
+/*
+ * Simulate adding a new route to the LPM counting number
+ * of new tables that will be needed
+ *
+ * It returns 0 on success, or 1 if
+ * the process needs to be continued by calling the function again.
+ */
+static inline int
+simulate_add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
+               struct rte_lpm6_tbl_entry **next_tbl, const uint8_t *ip,
+               uint8_t bytes, uint8_t first_byte, uint8_t depth,
+               uint32_t *need_tbl_nb)
+{
+       uint32_t entry_ind;
+       uint8_t bits_covered;
+       uint32_t next_tbl_ind;
+
+       /*
+        * Calculate index to the table based on the number and position
+        * of the bytes being inspected in this step.
+        */
+       entry_ind = get_bitshift(ip, first_byte, bytes);
+
+       /* Number of bits covered in this step */
+       bits_covered = (uint8_t)((bytes+first_byte-1)*BYTE_SIZE);
+
+       if (depth <= bits_covered) {
+               *need_tbl_nb = 0;
+               return 0;
+       }
+
+       if (tbl[entry_ind].valid == 0 || tbl[entry_ind].ext_entry == 0) {
+               /* from this point on a new table is needed on each level
+                * that is not covered yet
+                */
+               depth -= bits_covered;
+               uint32_t cnt = depth >> 3; /* depth / BYTE_SIZE */
+               if (depth & 7) /* 0b00000111 */
+                       /* if depth % 8 > 0 then one more table is needed
+                        * for those last bits
+                        */
+                       cnt++;
+
+               *need_tbl_nb = cnt;
+               return 0;
+       }
+
+       next_tbl_ind = tbl[entry_ind].lpm6_tbl8_gindex;
+       *next_tbl = &(lpm->tbl8[next_tbl_ind *
+               RTE_LPM6_TBL8_GROUP_NUM_ENTRIES]);
+       *need_tbl_nb = 0;
+       return 1;
+}
+
 /*
  * Partially adds a new route to the data structure (tbl24+tbl8s).
  * It returns 0 on success, a negative number on failure, or 1 if
@@ -457,25 +650,21 @@ expand_rule(struct rte_lpm6 *lpm, uint32_t tbl8_gindex, uint8_t depth,
  */
 static inline int
 add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
-               struct rte_lpm6_tbl_entry **tbl_next, uint8_t *ip, uint8_t bytes,
-               uint8_t first_byte, uint8_t depth, uint32_t next_hop)
+               uint32_t tbl_ind, struct rte_lpm6_tbl_entry **next_tbl,
+               uint32_t *next_tbl_ind, uint8_t *ip, uint8_t bytes,
+               uint8_t first_byte, uint8_t depth, uint32_t next_hop,
+               uint8_t is_new_rule)
 {
-       uint32_t tbl_index, tbl_range, tbl8_group_start, tbl8_group_end, i;
-       int32_t tbl8_gindex;
-       int8_t bitshift;
+       uint32_t entry_ind, tbl_range, tbl8_group_start, tbl8_group_end, i;
+       uint32_t tbl8_gindex;
        uint8_t bits_covered;
+       int ret;
 
        /*
         * Calculate index to the table based on the number and position
         * of the bytes being inspected in this step.
         */
-       tbl_index = 0;
-       for (i = first_byte; i < (uint32_t)(first_byte + bytes); i++) {
-               bitshift = (int8_t)((bytes - i)*BYTE_SIZE);
-
-               if (bitshift < 0) bitshift = 0;
-               tbl_index = tbl_index | ip[i-1] << bitshift;
-       }
+       entry_ind = get_bitshift(ip, first_byte, bytes);
 
        /* Number of bits covered in this step */
        bits_covered = (uint8_t)((bytes+first_byte-1)*BYTE_SIZE);
@@ -487,7 +676,7 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
        if (depth <= bits_covered) {
                tbl_range = 1 << (bits_covered - depth);
 
-               for (i = tbl_index; i < (tbl_index + tbl_range); i++) {
+               for (i = entry_ind; i < (entry_ind + tbl_range); i++) {
                        if (!tbl[i].valid || (tbl[i].ext_entry == 0 &&
                                        tbl[i].depth <= depth)) {
 
@@ -509,10 +698,15 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
                                 */
                                tbl8_gindex = tbl[i].lpm6_tbl8_gindex *
                                                RTE_LPM6_TBL8_GROUP_NUM_ENTRIES;
-                               expand_rule(lpm, tbl8_gindex, depth, next_hop);
+                               expand_rule(lpm, tbl8_gindex, depth, depth,
+                                               next_hop, VALID);
                        }
                }
 
+               /* update tbl8 rule reference counter */
+               if (tbl_ind != TBL24_IND && is_new_rule)
+                       lpm->tbl8_hdrs[tbl_ind].ref_cnt++;
+
                return 0;
        }
        /*
@@ -521,12 +715,24 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
         */
        else {
                /* If it's invalid a new tbl8 is needed */
-               if (!tbl[tbl_index].valid) {
-                       if (lpm->next_tbl8 < lpm->number_tbl8s)
-                               tbl8_gindex = (lpm->next_tbl8)++;
-                       else
+               if (!tbl[entry_ind].valid) {
+                       /* get a new table */
+                       ret = tbl8_get(lpm, &tbl8_gindex);
+                       if (ret != 0)
                                return -ENOSPC;
 
+                       /* invalidate all new tbl8 entries */
+                       tbl8_group_start = tbl8_gindex *
+                                       RTE_LPM6_TBL8_GROUP_NUM_ENTRIES;
+                       memset(&lpm->tbl8[tbl8_group_start], 0,
+                                         RTE_LPM6_TBL8_GROUP_NUM_ENTRIES);
+
+                       /* init the new table's header:
+                        *   save the reference to the owner table
+                        */
+                       init_tbl8_header(lpm, tbl8_gindex, tbl_ind, entry_ind);
+
+                       /* reference to a new tbl8 */
                        struct rte_lpm6_tbl_entry new_tbl_entry = {
                                .lpm6_tbl8_gindex = tbl8_gindex,
                                .depth = 0,
@@ -535,17 +741,20 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
                                .ext_entry = 1,
                        };
 
-                       tbl[tbl_index] = new_tbl_entry;
+                       tbl[entry_ind] = new_tbl_entry;
+
+                       /* update the current table's reference counter */
+                       if (tbl_ind != TBL24_IND)
+                               lpm->tbl8_hdrs[tbl_ind].ref_cnt++;
                }
                /*
-                * If it's valid but not extended the rule that was stored *
+                * If it's valid but not extended the rule that was stored
                 * here needs to be moved to the next table.
                 */
-               else if (tbl[tbl_index].ext_entry == 0) {
-                       /* Search for free tbl8 group. */
-                       if (lpm->next_tbl8 < lpm->number_tbl8s)
-                               tbl8_gindex = (lpm->next_tbl8)++;
-                       else
+               else if (tbl[entry_ind].ext_entry == 0) {
+                       /* get a new tbl8 index */
+                       ret = tbl8_get(lpm, &tbl8_gindex);
+                       if (ret != 0)
                                return -ENOSPC;
 
                        tbl8_group_start = tbl8_gindex *
@@ -553,13 +762,22 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
                        tbl8_group_end = tbl8_group_start +
                                        RTE_LPM6_TBL8_GROUP_NUM_ENTRIES;
 
+                       struct rte_lpm6_tbl_entry tbl_entry = {
+                               .next_hop = tbl[entry_ind].next_hop,
+                               .depth = tbl[entry_ind].depth,
+                               .valid = VALID,
+                               .valid_group = VALID,
+                               .ext_entry = 0
+                       };
+
                        /* Populate new tbl8 with tbl value. */
-                       for (i = tbl8_group_start; i < tbl8_group_end; i++) {
-                               lpm->tbl8[i].valid = VALID;
-                               lpm->tbl8[i].depth = tbl[tbl_index].depth;
-                               lpm->tbl8[i].next_hop = tbl[tbl_index].next_hop;
-                               lpm->tbl8[i].ext_entry = 0;
-                       }
+                       for (i = tbl8_group_start; i < tbl8_group_end; i++)
+                               lpm->tbl8[i] = tbl_entry;
+
+                       /* init the new table's header:
+                        *   save the reference to the owner table
+                        */
+                       init_tbl8_header(lpm, tbl8_gindex, tbl_ind, entry_ind);
 
                        /*
                         * Update tbl entry to point to new tbl8 entry. Note: The
@@ -574,11 +792,16 @@ add_step(struct rte_lpm6 *lpm, struct rte_lpm6_tbl_entry *tbl,
                                .ext_entry = 1,
                        };
 
-                       tbl[tbl_index] = new_tbl_entry;
+                       tbl[entry_ind] = new_tbl_entry;
+
+                       /* update the current table's reference counter */
+                       if (tbl_ind != TBL24_IND)
+                               lpm->tbl8_hdrs[tbl_ind].ref_cnt++;
                }
 
-               *tbl_next = &(lpm->tbl8[tbl[tbl_index].lpm6_tbl8_gindex *
-                               RTE_LPM6_TBL8_GROUP_NUM_ENTRIES]);
+               *next_tbl_ind = tbl[entry_ind].lpm6_tbl8_gindex;
+               *next_tbl = &(lpm->tbl8[*next_tbl_ind *
+                                 RTE_LPM6_TBL8_GROUP_NUM_ENTRIES]);
        }
 
        return 1;
@@ -595,13 +818,56 @@ rte_lpm6_add_v20(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth,
 }
 VERSION_SYMBOL(rte_lpm6_add, _v20, 2.0);
 
+
+/*
+ * Simulate adding a route to LPM
+ *
+ *     Returns:
+ *    0 on success
+ *    -ENOSPC not enought tbl8 left
+ */
+static int
+simulate_add(struct rte_lpm6 *lpm, const uint8_t *masked_ip, uint8_t depth)
+{
+       struct rte_lpm6_tbl_entry *tbl;
+       struct rte_lpm6_tbl_entry *tbl_next = NULL;
+       int ret, i;
+
+       /* number of new tables needed for a step */
+       uint32_t need_tbl_nb;
+       /* total number of new tables needed */
+       uint32_t total_need_tbl_nb;
+
+       /* Inspect the first three bytes through tbl24 on the first step. */
+       ret = simulate_add_step(lpm, lpm->tbl24, &tbl_next, masked_ip,
+                       ADD_FIRST_BYTE, 1, depth, &need_tbl_nb);
+       total_need_tbl_nb = need_tbl_nb;
+       /*
+        * Inspect one by one the rest of the bytes until
+        * the process is completed.
+        */
+       for (i = ADD_FIRST_BYTE; i < RTE_LPM6_IPV6_ADDR_SIZE && ret == 1; i++) {
+               tbl = tbl_next;
+               ret = simulate_add_step(lpm, tbl, &tbl_next, masked_ip, 1,
+                               (uint8_t)(i+1), depth, &need_tbl_nb);
+               total_need_tbl_nb += need_tbl_nb;
+       }
+
+       if (tbl8_available(lpm) < total_need_tbl_nb)
+               /* not enought tbl8 to add a rule */
+               return -ENOSPC;
+
+       return 0;
+}
+
 int
 rte_lpm6_add_v1705(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth,
                uint32_t next_hop)
 {
        struct rte_lpm6_tbl_entry *tbl;
        struct rte_lpm6_tbl_entry *tbl_next = NULL;
-       int32_t ret;
+       /* init to avoid compiler warning */
+       uint32_t tbl_next_num = 123456;
        int status;
        uint8_t masked_ip[RTE_LPM6_IPV6_ADDR_SIZE];
        int i;
@@ -614,21 +880,23 @@ rte_lpm6_add_v1705(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth,
        ip6_copy_addr(masked_ip, ip);
        ip6_mask_addr(masked_ip, depth);
 
-       /* Add the rule to the rule table. */
-       ret = rule_add(lpm, masked_ip, depth, next_hop);
-       /* If there is no space available for new rule return error. */
+       /* Simulate adding a new route */
+       int ret = simulate_add(lpm, masked_ip, depth);
        if (ret < 0)
                return ret;
 
+       /* Add the rule to the rule table. */
+       int is_new_rule = rule_add(lpm, masked_ip, depth, next_hop);
+       /* If there is no space available for new rule return error. */
+       if (is_new_rule < 0)
+               return is_new_rule;
+
        /* Inspect the first three bytes through tbl24 on the first step. */
        tbl = lpm->tbl24;
-       status = add_step (lpm, tbl, &tbl_next, masked_ip, ADD_FIRST_BYTE, 1,
-                       depth, next_hop);
-       if (status < 0) {
-               rte_lpm6_delete(lpm, masked_ip, depth);
-
-               return status;
-       }
+       status = add_step(lpm, tbl, TBL24_IND, &tbl_next, &tbl_next_num,
+                       masked_ip, ADD_FIRST_BYTE, 1, depth, next_hop,
+                       is_new_rule);
+       assert(status >= 0);
 
        /*
         * Inspect one by one the rest of the bytes until
@@ -636,13 +904,10 @@ rte_lpm6_add_v1705(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth,
         */
        for (i = ADD_FIRST_BYTE; i < RTE_LPM6_IPV6_ADDR_SIZE && status == 1; i++) {
                tbl = tbl_next;
-               status = add_step (lpm, tbl, &tbl_next, masked_ip, 1, (uint8_t)(i+1),
-                               depth, next_hop);
-               if (status < 0) {
-                       rte_lpm6_delete(lpm, masked_ip, depth);
-
-                       return status;
-               }
+               status = add_step(lpm, tbl, tbl_next_num, &tbl_next,
+                               &tbl_next_num, masked_ip, 1, (uint8_t)(i+1),
+                               depth, next_hop, is_new_rule);
+               assert(status >= 0);
        }
 
        return status;
@@ -717,9 +982,8 @@ rte_lpm6_lookup_v1705(const struct rte_lpm6 *lpm, uint8_t *ip,
        uint32_t tbl24_index;
 
        /* DEBUG: Check user input arguments. */
-       if ((lpm == NULL) || (ip == NULL) || (next_hop == NULL)) {
+       if ((lpm == NULL) || (ip == NULL) || (next_hop == NULL))
                return -EINVAL;
-       }
 
        first_byte = LOOKUP_FIRST_BYTE;
        tbl24_index = (ip[0] << BYTES2_SIZE) | (ip[1] << BYTE_SIZE) | ip[2];
@@ -755,9 +1019,8 @@ rte_lpm6_lookup_bulk_func_v20(const struct rte_lpm6 *lpm,
        int status;
 
        /* DEBUG: Check user input arguments. */
-       if ((lpm == NULL) || (ips == NULL) || (next_hops == NULL)) {
+       if ((lpm == NULL) || (ips == NULL) || (next_hops == NULL))
                return -EINVAL;
-       }
 
        for (i = 0; i < n; i++) {
                first_byte = LOOKUP_FIRST_BYTE;
@@ -900,62 +1163,22 @@ rule_delete(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth)
        return ret;
 }
 
-/*
- * Deletes a rule
- */
-int
-rte_lpm6_delete(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth)
-{
-       uint8_t masked_ip[RTE_LPM6_IPV6_ADDR_SIZE];
-       int ret;
-
-       /*
-        * Check input arguments.
-        */
-       if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM6_MAX_DEPTH)) {
-               return -EINVAL;
-       }
-
-       /* Copy the IP and mask it to avoid modifying user's input data. */
-       ip6_copy_addr(masked_ip, ip);
-       ip6_mask_addr(masked_ip, depth);
-
-       /* Delete the rule from the rule table. */
-       ret = rule_delete(lpm, masked_ip, depth);
-       if (ret < 0)
-               return -ENOENT;
-
-       /*
-        * Set all the table entries to 0 (ie delete every rule
-        * from the data structure.
-        */
-       lpm->next_tbl8 = 0;
-       memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
-       memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0])
-                       * RTE_LPM6_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
-
-       /*
-        * Add every rule again (except for the one that was removed from
-        * the rules table).
-        */
-       rebuild_lpm(lpm);
-
-       return 0;
-}
-
 /*
  * Deletes a group of rules
+ *
+ * Note that the function rebuilds the lpm table,
+ * rather than doing incremental updates like
+ * the regular delete function
  */
 int
 rte_lpm6_delete_bulk_func(struct rte_lpm6 *lpm,
-               uint8_t ips[][RTE_LPM6_IPV6_ADDR_SIZE], uint8_t *depths, unsigned n)
+               uint8_t ips[][RTE_LPM6_IPV6_ADDR_SIZE], uint8_t *depths,
+               unsigned n)
 {
        uint8_t masked_ip[RTE_LPM6_IPV6_ADDR_SIZE];
        unsigned i;
 
-       /*
-        * Check input arguments.
-        */
+       /* Check input arguments. */
        if ((lpm == NULL) || (ips == NULL) || (depths == NULL))
                return -EINVAL;
 
@@ -969,10 +1192,10 @@ rte_lpm6_delete_bulk_func(struct rte_lpm6 *lpm,
         * Set all the table entries to 0 (ie delete every rule
         * from the data structure.
         */
-       lpm->next_tbl8 = 0;
        memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
        memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0])
                        * RTE_LPM6_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
+       tbl8_pool_init(lpm);
 
        /*
         * Add every rule again (except for the ones that were removed from
@@ -992,9 +1215,6 @@ rte_lpm6_delete_all(struct rte_lpm6 *lpm)
        /* Zero used rules counter. */
        lpm->used_rules = 0;
 
-       /* Zero next tbl8 index. */
-       lpm->next_tbl8 = 0;
-
        /* Zero tbl24. */
        memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
 
@@ -1002,6 +1222,268 @@ rte_lpm6_delete_all(struct rte_lpm6 *lpm)
        memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0]) *
                        RTE_LPM6_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
 
+       /* init pool of free tbl8 indexes */
+       tbl8_pool_init(lpm);
+
        /* Delete all rules form the rules table. */
        rte_hash_reset(lpm->rules_tbl);
 }
+
+/*
+ * Convert a depth to a one byte long mask
+ *   Example: 4 will be converted to 0xF0
+ */
+static uint8_t __attribute__((pure))
+depth_to_mask_1b(uint8_t depth)
+{
+       /* To calculate a mask start with a 1 on the left hand side and right
+        * shift while populating the left hand side with 1's
+        */
+       return (signed char)0x80 >> (depth - 1);
+}
+
+/*
+ * Find a less specific rule
+ */
+static int
+rule_find_less_specific(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth,
+       struct rte_lpm6_rule *rule)
+{
+       int ret;
+       uint32_t next_hop;
+       uint8_t mask;
+       struct rte_lpm6_rule_key rule_key;
+
+       if (depth == 1)
+               return 0;
+
+       rule_key_init(&rule_key, ip, depth);
+
+       while (depth > 1) {
+               depth--;
+
+               /* each iteration zero one more bit of the key */
+               mask = depth & 7; /* depth % BYTE_SIZE */
+               if (mask > 0)
+                       mask = depth_to_mask_1b(mask);
+
+               rule_key.depth = depth;
+               rule_key.ip[depth >> 3] &= mask;
+
+               ret = rule_find_with_key(lpm, &rule_key, &next_hop);
+               if (ret) {
+                       rule->depth = depth;
+                       ip6_copy_addr(rule->ip, rule_key.ip);
+                       rule->next_hop = next_hop;
+                       return 1;
+               }
+       }
+
+       return 0;
+}
+
+/*
+ * Find range of tbl8 cells occupied by a rule
+ */
+static void
+rule_find_range(struct rte_lpm6 *lpm, const uint8_t *ip, uint8_t depth,
+                 struct rte_lpm6_tbl_entry **from,
+                 struct rte_lpm6_tbl_entry **to,
+                 uint32_t *out_tbl_ind)
+{
+       uint32_t ind;
+       uint32_t first_3bytes = (uint32_t)ip[0] << 16 | ip[1] << 8 | ip[2];
+
+       if (depth <= 24) {
+               /* rule is within the top level */
+               ind = first_3bytes;
+               *from = &lpm->tbl24[ind];
+               ind += (1 << (24 - depth)) - 1;
+               *to = &lpm->tbl24[ind];
+               *out_tbl_ind = TBL24_IND;
+       } else {
+               /* top level entry */
+               struct rte_lpm6_tbl_entry *tbl = &lpm->tbl24[first_3bytes];
+               assert(tbl->ext_entry == 1);
+               /* first tbl8 */
+               uint32_t tbl_ind = tbl->lpm6_tbl8_gindex;
+               tbl = &lpm->tbl8[tbl_ind *
+                               RTE_LPM6_TBL8_GROUP_NUM_ENTRIES];
+               /* current ip byte, the top level is already behind */
+               uint8_t byte = 3;
+               /* minus top level */
+               depth -= 24;
+
+               /* interate through levels (tbl8s)
+                * until we reach the last one
+                */
+               while (depth > 8) {
+                       tbl += ip[byte];
+                       assert(tbl->ext_entry == 1);
+                       /* go to the next level/tbl8 */
+                       tbl_ind = tbl->lpm6_tbl8_gindex;
+                       tbl = &lpm->tbl8[tbl_ind *
+                                       RTE_LPM6_TBL8_GROUP_NUM_ENTRIES];
+                       byte += 1;
+                       depth -= 8;
+               }
+
+               /* last level/tbl8 */
+               ind = ip[byte] & depth_to_mask_1b(depth);
+               *from = &tbl[ind];
+               ind += (1 << (8 - depth)) - 1;
+               *to = &tbl[ind];
+               *out_tbl_ind = tbl_ind;
+       }
+}
+
+/*
+ * Remove a table from the LPM tree
+ */
+static void
+remove_tbl(struct rte_lpm6 *lpm, struct rte_lpm_tbl8_hdr *tbl_hdr,
+                 uint32_t tbl_ind, struct rte_lpm6_rule *lsp_rule)
+{
+       struct rte_lpm6_tbl_entry *owner_entry;
+
+       if (tbl_hdr->owner_tbl_ind == TBL24_IND)
+               owner_entry = &lpm->tbl24[tbl_hdr->owner_entry_ind];
+       else {
+               uint32_t owner_tbl_ind = tbl_hdr->owner_tbl_ind;
+               owner_entry = &lpm->tbl8[
+                       owner_tbl_ind * RTE_LPM6_TBL8_GROUP_NUM_ENTRIES +
+                       tbl_hdr->owner_entry_ind];
+
+               struct rte_lpm_tbl8_hdr *owner_tbl_hdr =
+                       &lpm->tbl8_hdrs[owner_tbl_ind];
+               if (--owner_tbl_hdr->ref_cnt == 0)
+                       remove_tbl(lpm, owner_tbl_hdr, owner_tbl_ind, lsp_rule);
+       }
+
+       assert(owner_entry->ext_entry == 1);
+
+       /* unlink the table */
+       if (lsp_rule != NULL) {
+               struct rte_lpm6_tbl_entry new_tbl_entry = {
+                       .next_hop = lsp_rule->next_hop,
+                       .depth = lsp_rule->depth,
+                       .valid = VALID,
+                       .valid_group = VALID,
+                       .ext_entry = 0
+               };
+
+               *owner_entry = new_tbl_entry;
+       } else {
+               struct rte_lpm6_tbl_entry new_tbl_entry = {
+                       .next_hop = 0,
+                       .depth = 0,
+                       .valid = INVALID,
+                       .valid_group = INVALID,
+                       .ext_entry = 0
+               };
+
+               *owner_entry = new_tbl_entry;
+       }
+
+       /* return the table to the pool */
+       tbl8_put(lpm, tbl_ind);
+}
+
+/*
+ * Deletes a rule
+ */
+int
+rte_lpm6_delete(struct rte_lpm6 *lpm, uint8_t *ip, uint8_t depth)
+{
+       uint8_t masked_ip[RTE_LPM6_IPV6_ADDR_SIZE];
+       struct rte_lpm6_rule lsp_rule_obj;
+       struct rte_lpm6_rule *lsp_rule;
+       int ret;
+       uint32_t tbl_ind;
+       struct rte_lpm6_tbl_entry *from, *to;
+
+       /* Check input arguments. */
+       if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM6_MAX_DEPTH))
+               return -EINVAL;
+
+       /* Copy the IP and mask it to avoid modifying user's input data. */
+       ip6_copy_addr(masked_ip, ip);
+       ip6_mask_addr(masked_ip, depth);
+
+       /* Delete the rule from the rule table. */
+       ret = rule_delete(lpm, masked_ip, depth);
+       if (ret < 0)
+               return -ENOENT;
+
+       /* find rule cells */
+       rule_find_range(lpm, masked_ip, depth, &from, &to, &tbl_ind);
+
+       /* find a less specific rule (a rule with smaller depth)
+        * note: masked_ip will be modified, don't use it anymore
+        */
+       ret = rule_find_less_specific(lpm, masked_ip, depth,
+                       &lsp_rule_obj);
+       lsp_rule = ret ? &lsp_rule_obj : NULL;
+
+       /* decrement the table rule counter,
+        * note that tbl24 doesn't have a header
+        */
+       if (tbl_ind != TBL24_IND) {
+               struct rte_lpm_tbl8_hdr *tbl_hdr = &lpm->tbl8_hdrs[tbl_ind];
+               if (--tbl_hdr->ref_cnt == 0) {
+                       /* remove the table */
+                       remove_tbl(lpm, tbl_hdr, tbl_ind, lsp_rule);
+                       return 0;
+               }
+       }
+
+       /* iterate rule cells */
+       for (; from <= to; from++)
+               if (from->ext_entry == 1) {
+                       /* reference to a more specific space
+                        * of the prefix/rule. Entries in a more
+                        * specific space that are not used by
+                        * a more specific prefix must be occupied
+                        * by the prefix
+                        */
+                       if (lsp_rule != NULL)
+                               expand_rule(lpm,
+                                       from->lpm6_tbl8_gindex *
+                                       RTE_LPM6_TBL8_GROUP_NUM_ENTRIES,
+                                       depth, lsp_rule->depth,
+                                       lsp_rule->next_hop, VALID);
+                       else
+                               /* since the prefix has no less specific prefix,
+                                * its more specific space must be invalidated
+                                */
+                               expand_rule(lpm,
+                                       from->lpm6_tbl8_gindex *
+                                       RTE_LPM6_TBL8_GROUP_NUM_ENTRIES,
+                                       depth, 0, 0, INVALID);
+               } else if (from->depth == depth) {
+                       /* entry is not a reference and belongs to the prefix */
+                       if (lsp_rule != NULL) {
+                               struct rte_lpm6_tbl_entry new_tbl_entry = {
+                                       .next_hop = lsp_rule->next_hop,
+                                       .depth = lsp_rule->depth,
+                                       .valid = VALID,
+                                       .valid_group = VALID,
+                                       .ext_entry = 0
+                               };
+
+                               *from = new_tbl_entry;
+                       } else {
+                               struct rte_lpm6_tbl_entry new_tbl_entry = {
+                                       .next_hop = 0,
+                                       .depth = 0,
+                                       .valid = INVALID,
+                                       .valid_group = INVALID,
+                                       .ext_entry = 0
+                               };
+
+                               *from = new_tbl_entry;
+                       }
+               }
+
+       return 0;
+}