pipeline: support non-incremental table updates
authorCristian Dumitrescu <cristian.dumitrescu@intel.com>
Tue, 16 Feb 2021 20:46:44 +0000 (20:46 +0000)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 23 Mar 2021 17:45:44 +0000 (18:45 +0100)
Some table types (e.g. exact match/hash) allow for incremental table
updates, while others (e.g. wildcard match/ACL) do not. The former is
already supported, the latter is enabled by this patch.

Signed-off-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
lib/librte_pipeline/rte_swx_ctl.c

index 4a416bc..6bef9c3 100644 (file)
@@ -42,11 +42,38 @@ struct table {
        struct rte_swx_table_ops ops;
        struct rte_swx_table_params params;
 
+       /* Set of "stable" keys: these keys are currently part of the table;
+        * these keys will be preserved with no action data changes after the
+        * next commit.
+        */
        struct rte_swx_table_entry_list entries;
+
+       /* Set of new keys: these keys are currently NOT part of the table;
+        * these keys will be added to the table on the next commit, if
+        * the commit operation is successful.
+        */
        struct rte_swx_table_entry_list pending_add;
+
+       /* Set of keys to be modified: these keys are currently part of the
+        * table; these keys are still going to be part of the table after the
+        * next commit, but their action data will be modified if the commit
+        * operation is successful. The modify0 list contains the keys with the
+        * current action data, the modify1 list contains the keys with the
+        * modified action data.
+        */
        struct rte_swx_table_entry_list pending_modify0;
        struct rte_swx_table_entry_list pending_modify1;
+
+       /* Set of keys to be deleted: these keys are currently part of the
+        * table; these keys are to be deleted from the table on the next
+        * commit, if the commit operation is successful.
+        */
        struct rte_swx_table_entry_list pending_delete;
+
+       /* The pending default action: this is NOT the current default action;
+        * this will be the new default action after the next commit, if the
+        * next commit operation is successful.
+        */
        struct rte_swx_table_entry *pending_default;
 
        int is_stub;
@@ -609,6 +636,31 @@ table_pending_default_free(struct table *table)
        table->pending_default = NULL;
 }
 
+static int
+table_is_update_pending(struct table *table, int consider_pending_default)
+{
+       struct rte_swx_table_entry *e;
+       uint32_t n = 0;
+
+       /* Pending add. */
+       TAILQ_FOREACH(e, &table->pending_add, node)
+               n++;
+
+       /* Pending modify. */
+       TAILQ_FOREACH(e, &table->pending_modify1, node)
+               n++;
+
+       /* Pending delete. */
+       TAILQ_FOREACH(e, &table->pending_delete, node)
+               n++;
+
+       /* Pending default. */
+       if (consider_pending_default && table->pending_default)
+               n++;
+
+       return n;
+}
+
 static void
 table_free(struct rte_swx_ctl_pipeline *ctl)
 {
@@ -680,7 +732,7 @@ table_state_create(struct rte_swx_ctl_pipeline *ctl)
                struct rte_swx_table_state *ts_next = &ctl->ts_next[i];
 
                /* Table object. */
-               if (!table->is_stub) {
+               if (!table->is_stub && table->ops.add) {
                        ts_next->obj = table->ops.create(&table->params,
                                                         &table->entries,
                                                         table->info.args,
@@ -691,6 +743,9 @@ table_state_create(struct rte_swx_ctl_pipeline *ctl)
                        }
                }
 
+               if (!table->is_stub && !table->ops.add)
+                       ts_next->obj = ts->obj;
+
                /* Default action data: duplicate from current table state. */
                ts_next->default_action_data =
                        malloc(table->params.action_data_size);
@@ -1114,54 +1169,173 @@ rte_swx_ctl_pipeline_table_default_entry_add(struct rte_swx_ctl_pipeline *ctl,
        return 0;
 }
 
+
+static void
+table_entry_list_free(struct rte_swx_table_entry_list *list)
+{
+       for ( ; ; ) {
+               struct rte_swx_table_entry *entry;
+
+               entry = TAILQ_FIRST(list);
+               if (!entry)
+                       break;
+
+               TAILQ_REMOVE(list, entry, node);
+               table_entry_free(entry);
+       }
+}
+
+static int
+table_entry_list_duplicate(struct rte_swx_ctl_pipeline *ctl,
+                          uint32_t table_id,
+                          struct rte_swx_table_entry_list *dst,
+                          struct rte_swx_table_entry_list *src)
+{
+       struct rte_swx_table_entry *src_entry;
+
+       TAILQ_FOREACH(src_entry, src, node) {
+               struct rte_swx_table_entry *dst_entry;
+
+               dst_entry = table_entry_duplicate(ctl, table_id, src_entry, 1, 1);
+               if (!dst_entry)
+                       goto error;
+
+               TAILQ_INSERT_TAIL(dst, dst_entry, node);
+       }
+
+       return 0;
+
+error:
+       table_entry_list_free(dst);
+       return -ENOMEM;
+}
+
+/* This commit stage contains all the operations that can fail; in case ANY of
+ * them fails for ANY table, ALL of them are rolled back for ALL the tables.
+ */
 static int
-table_rollfwd0(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
+table_rollfwd0(struct rte_swx_ctl_pipeline *ctl,
+              uint32_t table_id,
+              uint32_t after_swap)
 {
        struct table *table = &ctl->tables[table_id];
+       struct rte_swx_table_state *ts = &ctl->ts[table_id];
        struct rte_swx_table_state *ts_next = &ctl->ts_next[table_id];
-       struct rte_swx_table_entry *entry;
 
-       /* Reset counters. */
-       table->n_add = 0;
-       table->n_modify = 0;
-       table->n_delete = 0;
+       if (table->is_stub || !table_is_update_pending(table, 0))
+               return 0;
 
-       /* Add pending rules. */
-       TAILQ_FOREACH(entry, &table->pending_add, node) {
-               int status;
+       /*
+        * Current table supports incremental update.
+        */
+       if (table->ops.add) {
+               /* Reset counters. */
+               table->n_add = 0;
+               table->n_modify = 0;
+               table->n_delete = 0;
 
-               status = table->ops.add(ts_next->obj, entry);
-               if (status)
-                       return status;
+               /* Add pending rules. */
+               struct rte_swx_table_entry *entry;
 
-               table->n_add++;
-       }
+               TAILQ_FOREACH(entry, &table->pending_add, node) {
+                       int status;
 
-       /* Modify pending rules. */
-       TAILQ_FOREACH(entry, &table->pending_modify1, node) {
-               int status;
+                       status = table->ops.add(ts_next->obj, entry);
+                       if (status)
+                               return status;
 
-               status = table->ops.add(ts_next->obj, entry);
-               if (status)
-                       return status;
+                       table->n_add++;
+               }
+
+               /* Modify pending rules. */
+               TAILQ_FOREACH(entry, &table->pending_modify1, node) {
+                       int status;
+
+                       status = table->ops.add(ts_next->obj, entry);
+                       if (status)
+                               return status;
+
+                       table->n_modify++;
+               }
+
+               /* Delete pending rules. */
+               TAILQ_FOREACH(entry, &table->pending_delete, node) {
+                       int status;
 
-               table->n_modify++;
+                       status = table->ops.del(ts_next->obj, entry);
+                       if (status)
+                               return status;
+
+                       table->n_delete++;
+               }
+
+               return 0;
        }
 
-       /* Delete pending rules. */
-       TAILQ_FOREACH(entry, &table->pending_delete, node) {
+       /*
+        * Current table does NOT support incremental update.
+        */
+       if (!after_swap) {
+               struct rte_swx_table_entry_list list;
                int status;
 
-               status = table->ops.del(ts_next->obj, entry);
+               /* Create updated list of entries included. */
+               TAILQ_INIT(&list);
+
+               status = table_entry_list_duplicate(ctl,
+                                                   table_id,
+                                                   &list,
+                                                   &table->entries);
+               if (status)
+                       goto error;
+
+               status = table_entry_list_duplicate(ctl,
+                                                   table_id,
+                                                   &list,
+                                                   &table->pending_add);
+               if (status)
+                       goto error;
+
+               status = table_entry_list_duplicate(ctl,
+                                                   table_id,
+                                                   &list,
+                                                   &table->pending_modify1);
                if (status)
-                       return status;
+                       goto error;
+
+               /* Create new table object with the updates included. */
+               ts_next->obj = table->ops.create(&table->params,
+                                                &list,
+                                                table->info.args,
+                                                ctl->numa_node);
+               if (!ts_next->obj) {
+                       status = -ENODEV;
+                       goto error;
+               }
+
+               table_entry_list_free(&list);
+
+               return 0;
 
-               table->n_delete++;
+error:
+               table_entry_list_free(&list);
+               return status;
        }
 
+       /* Free the old table object. */
+       if (ts_next->obj && table->ops.free)
+               table->ops.free(ts_next->obj);
+
+       /* Copy over the new table object. */
+       ts_next->obj = ts->obj;
+
        return 0;
 }
 
+/* This commit stage contains all the operations that cannot fail. They are
+ * executed only if the previous stage was successful for ALL the tables. Hence,
+ * none of these operations has to be rolled back for ANY table.
+ */
 static void
 table_rollfwd1(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
 {
@@ -1186,6 +1360,10 @@ table_rollfwd1(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
        ts_next->default_action_id = action_id;
 }
 
+/* This last commit stage is simply finalizing a successful commit operation.
+ * This stage is only executed if all the previous stages were successful. This
+ * stage cannot fail.
+ */
 static void
 table_rollfwd2(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
 {
@@ -1212,43 +1390,66 @@ table_rollfwd2(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
        table_pending_default_free(table);
 }
 
+/* The rollback stage is only executed when the commit failed, i.e. ANY of the
+ * commit operations that can fail did fail for ANY table. It reverts ALL the
+ * tables to their state before the commit started, as if the commit never
+ * happened.
+ */
 static void
 table_rollback(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
 {
        struct table *table = &ctl->tables[table_id];
        struct rte_swx_table_state *ts_next = &ctl->ts_next[table_id];
-       struct rte_swx_table_entry *entry;
 
-       /* Add back all the entries that were just deleted. */
-       TAILQ_FOREACH(entry, &table->pending_delete, node) {
-               if (!table->n_delete)
-                       break;
+       if (table->is_stub || !table_is_update_pending(table, 0))
+               return;
 
-               table->ops.add(ts_next->obj, entry);
-               table->n_delete--;
-       }
+       if (table->ops.add) {
+               struct rte_swx_table_entry *entry;
 
-       /* Add back the old copy for all the entries that were just
-        * modified.
-        */
-       TAILQ_FOREACH(entry, &table->pending_modify0, node) {
-               if (!table->n_modify)
-                       break;
+               /* Add back all the entries that were just deleted. */
+               TAILQ_FOREACH(entry, &table->pending_delete, node) {
+                       if (!table->n_delete)
+                               break;
 
-               table->ops.add(ts_next->obj, entry);
-               table->n_modify--;
-       }
+                       table->ops.add(ts_next->obj, entry);
+                       table->n_delete--;
+               }
 
-       /* Delete all the entries that were just added. */
-       TAILQ_FOREACH(entry, &table->pending_add, node) {
-               if (!table->n_add)
-                       break;
+               /* Add back the old copy for all the entries that were just
+                * modified.
+                */
+               TAILQ_FOREACH(entry, &table->pending_modify0, node) {
+                       if (!table->n_modify)
+                               break;
+
+                       table->ops.add(ts_next->obj, entry);
+                       table->n_modify--;
+               }
 
-               table->ops.del(ts_next->obj, entry);
-               table->n_add--;
+               /* Delete all the entries that were just added. */
+               TAILQ_FOREACH(entry, &table->pending_add, node) {
+                       if (!table->n_add)
+                               break;
+
+                       table->ops.del(ts_next->obj, entry);
+                       table->n_add--;
+               }
+       } else {
+               struct rte_swx_table_state *ts = &ctl->ts[table_id];
+
+               /* Free the new table object, as update was cancelled. */
+               if (ts_next->obj && table->ops.free)
+                       table->ops.free(ts_next->obj);
+
+               /* Reinstate the old table object. */
+               ts_next->obj = ts->obj;
        }
 }
 
+/* This stage is conditionally executed (as instructed by the user) after a
+ * failed commit operation to remove ALL the pending work for ALL the tables.
+ */
 static void
 table_abort(struct rte_swx_ctl_pipeline *ctl, uint32_t table_id)
 {
@@ -1290,7 +1491,7 @@ rte_swx_ctl_pipeline_commit(struct rte_swx_ctl_pipeline *ctl, int abort_on_fail)
         * ts.
         */
        for (i = 0; i < ctl->info.n_tables; i++) {
-               status = table_rollfwd0(ctl, i);
+               status = table_rollfwd0(ctl, i, 0);
                if (status)
                        goto rollback;
        }
@@ -1310,7 +1511,7 @@ rte_swx_ctl_pipeline_commit(struct rte_swx_ctl_pipeline *ctl, int abort_on_fail)
        /* Operate the changes on the current ts_next, which is the previous ts.
         */
        for (i = 0; i < ctl->info.n_tables; i++) {
-               table_rollfwd0(ctl, i);
+               table_rollfwd0(ctl, i, 1);
                table_rollfwd1(ctl, i);
                table_rollfwd2(ctl, i);
        }
@@ -1444,11 +1645,11 @@ rte_swx_ctl_pipeline_table_entry_read(struct rte_swx_ctl_pipeline *ctl,
                                mask = field_hton(mask, mf->n_bits);
                }
 
-                       /* Copy to entry. */
-                       if (entry->key_mask)
-                               memcpy(&entry->key_mask[offset],
-                                      (uint8_t *)&mask,
-                                      mf->n_bits / 8);
+               /* Copy to entry. */
+               if (entry->key_mask)
+                       memcpy(&entry->key_mask[offset],
+                              (uint8_t *)&mask,
+                              mf->n_bits / 8);
 
                /*
                 * Value.