lpm: avoid race conditions for v20
[dpdk.git] / lib / librte_lpm / rte_lpm.c
index b91f742..5bd8ab9 100644 (file)
@@ -21,6 +21,7 @@
 #include <rte_errno.h>
 #include <rte_rwlock.h>
 #include <rte_spinlock.h>
+#include <rte_tailq.h>
 
 #include "rte_lpm.h"
 
@@ -70,7 +71,7 @@ depth_to_mask(uint8_t depth)
 /*
  * Converts given depth value to its corresponding range value.
  */
-static inline uint32_t __attribute__((pure))
+static uint32_t __attribute__((pure))
 depth_to_range(uint8_t depth)
 {
        VERIFY_DEPTH(depth);
@@ -399,7 +400,7 @@ MAP_STATIC_SYMBOL(void rte_lpm_free(struct rte_lpm *lpm),
  * are stored in the rule table from 0 - 31.
  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
  */
-static inline int32_t
+static int32_t
 rule_add_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
        uint8_t next_hop)
 {
@@ -471,7 +472,7 @@ rule_add_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
        return rule_index;
 }
 
-static inline int32_t
+static int32_t
 rule_add_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
        uint32_t next_hop)
 {
@@ -547,7 +548,7 @@ rule_add_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
  * Delete a rule from the rule table.
  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
  */
-static inline void
+static void
 rule_delete_v20(struct rte_lpm_v20 *lpm, int32_t rule_index, uint8_t depth)
 {
        int i;
@@ -570,7 +571,7 @@ rule_delete_v20(struct rte_lpm_v20 *lpm, int32_t rule_index, uint8_t depth)
        lpm->rule_info[depth - 1].used_rules--;
 }
 
-static inline void
+static void
 rule_delete_v1604(struct rte_lpm *lpm, int32_t rule_index, uint8_t depth)
 {
        int i;
@@ -597,7 +598,7 @@ rule_delete_v1604(struct rte_lpm *lpm, int32_t rule_index, uint8_t depth)
  * Finds a rule in rule table.
  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
  */
-static inline int32_t
+static int32_t
 rule_find_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth)
 {
        uint32_t rule_gindex, last_rule, rule_index;
@@ -618,7 +619,7 @@ rule_find_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth)
        return -EINVAL;
 }
 
-static inline int32_t
+static int32_t
 rule_find_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)
 {
        uint32_t rule_gindex, last_rule, rule_index;
@@ -642,7 +643,7 @@ rule_find_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)
 /*
  * Find, clean and allocate a tbl8.
  */
-static inline int32_t
+static int32_t
 tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8)
 {
        uint32_t group_idx; /* tbl8 group index. */
@@ -669,7 +670,7 @@ tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8)
        return -ENOSPC;
 }
 
-static inline int32_t
+static int32_t
 tbl8_alloc_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)
 {
        uint32_t group_idx; /* tbl8 group index. */
@@ -695,21 +696,21 @@ tbl8_alloc_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)
        return -ENOSPC;
 }
 
-static inline void
+static void
 tbl8_free_v20(struct rte_lpm_tbl_entry_v20 *tbl8, uint32_t tbl8_group_start)
 {
        /* Set tbl8 group invalid*/
        tbl8[tbl8_group_start].valid_group = INVALID;
 }
 
-static inline void
+static void
 tbl8_free_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t tbl8_group_start)
 {
        /* Set tbl8 group invalid*/
        tbl8[tbl8_group_start].valid_group = INVALID;
 }
 
-static inline int32_t
+static __rte_noinline int32_t
 add_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip, uint8_t depth,
                uint8_t next_hop)
 {
@@ -737,7 +738,8 @@ add_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip, uint8_t depth,
                        /* Setting tbl24 entry in one go to avoid race
                         * conditions
                         */
-                       lpm->tbl24[i] = new_tbl24_entry;
+                       __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
+                                       __ATOMIC_RELEASE);
 
                        continue;
                }
@@ -777,7 +779,7 @@ add_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip, uint8_t depth,
        return 0;
 }
 
-static inline int32_t
+static __rte_noinline int32_t
 add_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
                uint32_t next_hop)
 {
@@ -806,7 +808,8 @@ add_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
                        /* Setting tbl24 entry in one go to avoid race
                         * conditions
                         */
-                       lpm->tbl24[i] = new_tbl24_entry;
+                       __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
+                                       __ATOMIC_RELEASE);
 
                        continue;
                }
@@ -846,7 +849,7 @@ add_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
        return 0;
 }
 
-static inline int32_t
+static __rte_noinline int32_t
 add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
                uint8_t next_hop)
 {
@@ -891,7 +894,8 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
                        .depth = 0,
                };
 
-               lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
+                               __ATOMIC_RELEASE);
 
        } /* If valid entry but not extended calculate the index into Table8. */
        else if (lpm->tbl24[tbl24_index].valid_group == 0) {
@@ -937,7 +941,8 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
                                .depth = 0,
                };
 
-               lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
+                               __ATOMIC_RELEASE);
 
        } else { /*
                * If it is valid, extended entry calculate the index into tbl8.
@@ -971,7 +976,7 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked, uint8_t depth,
        return 0;
 }
 
-static inline int32_t
+static __rte_noinline int32_t
 add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
                uint32_t next_hop)
 {
@@ -1017,7 +1022,11 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
                        .depth = 0,
                };
 
-               lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               /* The tbl24 entry must be written only after the
+                * tbl8 entries are written.
+                */
+               __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
+                               __ATOMIC_RELEASE);
 
        } /* If valid entry but not extended calculate the index into Table8. */
        else if (lpm->tbl24[tbl24_index].valid_group == 0) {
@@ -1063,7 +1072,11 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
                                .depth = 0,
                };
 
-               lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               /* The tbl24 entry must be written only after the
+                * tbl8 entries are written.
+                */
+               __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
+                               __ATOMIC_RELEASE);
 
        } else { /*
                * If it is valid, extended entry calculate the index into tbl8.
@@ -1244,7 +1257,7 @@ BIND_DEFAULT_SYMBOL(rte_lpm_is_rule_present, _v1604, 16.04);
 MAP_STATIC_SYMBOL(int rte_lpm_is_rule_present(struct rte_lpm *lpm, uint32_t ip,
                uint8_t depth, uint32_t *next_hop), rte_lpm_is_rule_present_v1604);
 
-static inline int32_t
+static int32_t
 find_previous_rule_v20(struct rte_lpm_v20 *lpm, uint32_t ip, uint8_t depth,
                uint8_t *sub_rule_depth)
 {
@@ -1266,7 +1279,7 @@ find_previous_rule_v20(struct rte_lpm_v20 *lpm, uint32_t ip, uint8_t depth,
        return -1;
 }
 
-static inline int32_t
+static int32_t
 find_previous_rule_v1604(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
                uint8_t *sub_rule_depth)
 {
@@ -1288,7 +1301,7 @@ find_previous_rule_v1604(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
        return -1;
 }
 
-static inline int32_t
+static int32_t
 delete_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
        uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 {
@@ -1311,7 +1324,15 @@ delete_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
 
                        if (lpm->tbl24[i].valid_group == 0 &&
                                        lpm->tbl24[i].depth <= depth) {
-                               lpm->tbl24[i].valid = INVALID;
+                               struct rte_lpm_tbl_entry_v20
+                                       zero_tbl24_entry = {
+                                               .valid = INVALID,
+                                               .depth = 0,
+                                               .valid_group = 0,
+                                       };
+                                       zero_tbl24_entry.next_hop = 0;
+                               __atomic_store(&lpm->tbl24[i],
+                                       &zero_tbl24_entry, __ATOMIC_RELEASE);
                        } else if (lpm->tbl24[i].valid_group == 1) {
                                /*
                                 * If TBL24 entry is extended, then there has
@@ -1356,7 +1377,8 @@ delete_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
 
                        if (lpm->tbl24[i].valid_group == 0 &&
                                        lpm->tbl24[i].depth <= depth) {
-                               lpm->tbl24[i] = new_tbl24_entry;
+                               __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
+                                               __ATOMIC_RELEASE);
                        } else  if (lpm->tbl24[i].valid_group == 1) {
                                /*
                                 * If TBL24 entry is extended, then there has
@@ -1381,7 +1403,7 @@ delete_depth_small_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
        return 0;
 }
 
-static inline int32_t
+static int32_t
 delete_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
        uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 {
@@ -1391,6 +1413,7 @@ delete_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
        /* Calculate the range and index into Table24. */
        tbl24_range = depth_to_range(depth);
        tbl24_index = (ip_masked >> 8);
+       struct rte_lpm_tbl_entry zero_tbl24_entry = {0};
 
        /*
         * Firstly check the sub_rule_index. A -1 indicates no replacement rule
@@ -1405,7 +1428,8 @@ delete_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
 
                        if (lpm->tbl24[i].valid_group == 0 &&
                                        lpm->tbl24[i].depth <= depth) {
-                               lpm->tbl24[i].valid = INVALID;
+                               __atomic_store(&lpm->tbl24[i],
+                                       &zero_tbl24_entry, __ATOMIC_RELEASE);
                        } else if (lpm->tbl24[i].valid_group == 1) {
                                /*
                                 * If TBL24 entry is extended, then there has
@@ -1450,7 +1474,8 @@ delete_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
 
                        if (lpm->tbl24[i].valid_group == 0 &&
                                        lpm->tbl24[i].depth <= depth) {
-                               lpm->tbl24[i] = new_tbl24_entry;
+                               __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
+                                               __ATOMIC_RELEASE);
                        } else  if (lpm->tbl24[i].valid_group == 1) {
                                /*
                                 * If TBL24 entry is extended, then there has
@@ -1483,7 +1508,7 @@ delete_depth_small_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
  * Return of value > -1 means tbl8 is in use but has all the same values and
  * thus can be recycled
  */
-static inline int32_t
+static int32_t
 tbl8_recycle_check_v20(struct rte_lpm_tbl_entry_v20 *tbl8,
                uint32_t tbl8_group_start)
 {
@@ -1530,7 +1555,7 @@ tbl8_recycle_check_v20(struct rte_lpm_tbl_entry_v20 *tbl8,
        return -EINVAL;
 }
 
-static inline int32_t
+static int32_t
 tbl8_recycle_check_v1604(struct rte_lpm_tbl_entry *tbl8,
                uint32_t tbl8_group_start)
 {
@@ -1577,7 +1602,7 @@ tbl8_recycle_check_v1604(struct rte_lpm_tbl_entry *tbl8,
        return -EINVAL;
 }
 
-static inline int32_t
+static int32_t
 delete_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
        uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 {
@@ -1635,8 +1660,11 @@ delete_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
        tbl8_recycle_index = tbl8_recycle_check_v20(lpm->tbl8, tbl8_group_start);
 
        if (tbl8_recycle_index == -EINVAL) {
-               /* Set tbl24 before freeing tbl8 to avoid race condition. */
+               /* Set tbl24 before freeing tbl8 to avoid race condition.
+                * Prevent the free of the tbl8 group from hoisting.
+                */
                lpm->tbl24[tbl24_index].valid = 0;
+               __atomic_thread_fence(__ATOMIC_RELEASE);
                tbl8_free_v20(lpm->tbl8, tbl8_group_start);
        } else if (tbl8_recycle_index > -1) {
                /* Update tbl24 entry. */
@@ -1647,15 +1675,18 @@ delete_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t ip_masked,
                        .depth = lpm->tbl8[tbl8_recycle_index].depth,
                };
 
-               /* Set tbl24 before freeing tbl8 to avoid race condition. */
+               /* Set tbl24 before freeing tbl8 to avoid race condition.
+                * Prevent the free of the tbl8 group from hoisting.
+                */
                lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               __atomic_thread_fence(__ATOMIC_RELEASE);
                tbl8_free_v20(lpm->tbl8, tbl8_group_start);
        }
 
        return 0;
 }
 
-static inline int32_t
+static int32_t
 delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
        uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 {
@@ -1713,8 +1744,11 @@ delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
        tbl8_recycle_index = tbl8_recycle_check_v1604(lpm->tbl8, tbl8_group_start);
 
        if (tbl8_recycle_index == -EINVAL) {
-               /* Set tbl24 before freeing tbl8 to avoid race condition. */
+               /* Set tbl24 before freeing tbl8 to avoid race condition.
+                * Prevent the free of the tbl8 group from hoisting.
+                */
                lpm->tbl24[tbl24_index].valid = 0;
+               __atomic_thread_fence(__ATOMIC_RELEASE);
                tbl8_free_v1604(lpm->tbl8, tbl8_group_start);
        } else if (tbl8_recycle_index > -1) {
                /* Update tbl24 entry. */
@@ -1725,8 +1759,11 @@ delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
                        .depth = lpm->tbl8[tbl8_recycle_index].depth,
                };
 
-               /* Set tbl24 before freeing tbl8 to avoid race condition. */
+               /* Set tbl24 before freeing tbl8 to avoid race condition.
+                * Prevent the free of the tbl8 group from hoisting.
+                */
                lpm->tbl24[tbl24_index] = new_tbl24_entry;
+               __atomic_thread_fence(__ATOMIC_RELEASE);
                tbl8_free_v1604(lpm->tbl8, tbl8_group_start);
        }
 #undef group_idx