Fix rcuja: handle concurrent updates
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Sat, 18 May 2013 15:18:52 +0000 (17:18 +0200)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Sat, 18 May 2013 15:18:52 +0000 (17:18 +0200)
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
rcuja/rcuja.c

index db5072a2133f469672c2010df73e2485e09f2a57..97381437368d295380de5230bc9c72ed9217a9ca 100644 (file)
@@ -1000,6 +1000,16 @@ int ja_attach_node(struct cds_ja *ja,
                }
        }
 
+       if (*node_flag_ptr != NULL) {
+               /*
+                * Attach point is non-NULL: it has been updated between
+                * RCU lookup and lock acquisition. We need to re-try
+                * lookup and attach.
+                */
+               ret = -EAGAIN;
+               goto unlock_parent;
+       }
+
        /* Create new branch, starting from bottom */
        CDS_INIT_HLIST_HEAD(&head);
        cds_hlist_add_head_rcu(&child_node->list, &head);
@@ -1062,6 +1072,7 @@ check_error:
                        assert(!tmpret);
                }
        }
+unlock_parent:
        if (parent_shadow_node)
                rcuja_shadow_unlock(parent_shadow_node);
 unlock_shadow:
@@ -1086,8 +1097,10 @@ int ja_chain_node(struct cds_ja *ja,
        struct cds_ja_shadow_node *shadow_node;
 
        shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag);
-       if (!shadow_node)
+       if (!shadow_node) {
+               dbg_printf("AGAIN3\n");
                return -EAGAIN;
+       }
        cds_hlist_add_head_rcu(&node->list, head);
        rcuja_shadow_unlock(shadow_node);
        return 0;
@@ -1103,8 +1116,9 @@ int cds_ja_add(struct cds_ja *ja, uint64_t key,
                *parent2_node_flag;
        int ret;
 
-       if (caa_unlikely(key > ja->key_max))
+       if (caa_unlikely(key > ja->key_max)) {
                return -EINVAL;
+       }
        tree_depth = ja->tree_depth;
 
 retry:
@@ -1156,7 +1170,7 @@ retry:
                        (struct cds_hlist_head *) node_flag_ptr,
                        new_node);
        }
-       if (ret == -EAGAIN)
+       if (ret == -EAGAIN || ret == -EEXIST)
                goto retry;
 end:
        return ret;
@@ -1169,6 +1183,9 @@ end:
  * ensure that when a match value -> pointer is found in a node, it is
  * _NEVER_ changed for that node without recompaction, and recompaction
  * reallocates the node.
+ * However, when a child is removed from "linear" nodes, its pointer
+ * is set to NULL. We therefore check, while holding the locks, if this
+ * pointer is NULL, and return -ENOENT to the caller if it is the case.
  */
 static
 int ja_detach_node(struct cds_ja *ja,
@@ -1240,6 +1257,15 @@ int ja_detach_node(struct cds_ja *ja,
                }
        }
 
+       /*
+        * Check if node has been removed between RCU lookup and lock
+        * acquisition.
+        */
+       if (!*node_flag_ptr) {
+               ret = -ENOENT;
+               goto end;
+       }
+
        /*
         * At this point, we want to delete all nodes that are about to
         * be removed from shadow_nodes (except the last one, which is
@@ -1264,6 +1290,8 @@ int ja_detach_node(struct cds_ja *ja,
                &iter_node_flag,        /* Old new parent ptr in its parent */
                shadow_nodes[nr_branch - 1],    /* of parent */
                n);
+       if (ret)
+               goto end;
 
        dbg_printf("ja_detach_node: publish %p instead of %p\n",
                iter_node_flag, *parent_node_flag_ptr);
@@ -1404,7 +1432,12 @@ retry:
                                &hlist_head, match);
                }
        }
-       if (ret == -EAGAIN)
+       /*
+        * Explanation of -ENOENT handling: caused by concurrent delete
+        * between RCU lookup and actual removal. Need to re-do the
+        * lookup and removal attempt.
+        */
+       if (ret == -EAGAIN || ret == -ENOENT)
                goto retry;
        return ret;
 }
This page took 0.027374 seconds and 4 git commands to generate.