The table entries were updated field by field. There were two issues:
1. bitwise operations are read-modify-write sequences and not atomic,
   nor efficient.
2. the above non-atomic operations causes entries out of synchronization
   and inconsistency.
This patch combines the fields into a one-go 32bit entry update to avoid
inconsistency and as a bonus save CPU cycles.

Suggested-by: Gavin Hu <gavin...@arm.com>
Signed-off-by: Ruifeng Wang <ruifeng.w...@arm.com>
Reviewed-by: Gavin Hu <gavin...@arm.com>
---
 lib/librte_lpm/rte_lpm.c | 35 +++++++++++++++++++++++++----------
 1 file changed, 25 insertions(+), 10 deletions(-)

diff --git a/lib/librte_lpm/rte_lpm.c b/lib/librte_lpm/rte_lpm.c
index d86248713..95e2f75aa 100644
--- a/lib/librte_lpm/rte_lpm.c
+++ b/lib/librte_lpm/rte_lpm.c
@@ -906,9 +906,14 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t 
ip_masked, uint8_t depth,
 
                /* Set tbl8 entry. */
                for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
-                       lpm->tbl8[i].depth = depth;
-                       lpm->tbl8[i].next_hop = next_hop;
-                       lpm->tbl8[i].valid = VALID;
+                       struct rte_lpm_tbl_entry_v20 new_tbl8_entry = {
+                               .valid = VALID,
+                               .depth = depth,
+                               .valid_group = lpm->tbl8[i].valid_group,
+                       };
+                       new_tbl8_entry.next_hop = next_hop;
+                       __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
+                                       __ATOMIC_RELAXED);
                }
 
                /*
@@ -943,19 +948,29 @@ add_depth_big_v20(struct rte_lpm_v20 *lpm, uint32_t 
ip_masked, uint8_t depth,
 
                /* Populate new tbl8 with tbl24 value. */
                for (i = tbl8_group_start; i < tbl8_group_end; i++) {
-                       lpm->tbl8[i].valid = VALID;
-                       lpm->tbl8[i].depth = lpm->tbl24[tbl24_index].depth;
-                       lpm->tbl8[i].next_hop =
-                                       lpm->tbl24[tbl24_index].next_hop;
+                       struct rte_lpm_tbl_entry_v20 new_tbl8_entry = {
+                               .valid = VALID,
+                               .depth = lpm->tbl24[tbl24_index].depth,
+                               .valid_group = lpm->tbl8[i].valid_group,
+                       };
+                       new_tbl8_entry.next_hop =
+                               lpm->tbl24[tbl24_index].next_hop;
+                       __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
+                                       __ATOMIC_RELAXED);
                }
 
                tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 
                /* Insert new rule into the tbl8 entry. */
                for (i = tbl8_index; i < tbl8_index + tbl8_range; i++) {
-                       lpm->tbl8[i].valid = VALID;
-                       lpm->tbl8[i].depth = depth;
-                       lpm->tbl8[i].next_hop = next_hop;
+                       struct rte_lpm_tbl_entry_v20 new_tbl8_entry = {
+                               .valid = VALID,
+                               .depth = depth,
+                               .valid_group = lpm->tbl8[i].valid_group,
+                       };
+                       new_tbl8_entry.next_hop = next_hop;
+                       __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
+                                       __ATOMIC_RELAXED);
                }
 
                /*
-- 
2.17.1

Reply via email to