list_lru: fix broken LRU_RETRY behaviour
[cascardo/linux.git] / mm / list_lru.c
index 1efe4ec..e77c29f 100644 (file)
@@ -73,19 +73,19 @@ list_lru_walk_node(struct list_lru *lru, int nid, list_lru_walk_cb isolate,
        struct list_lru_node    *nlru = &lru->node[nid];
        struct list_head *item, *n;
        unsigned long isolated = 0;
-       /*
-        * If we don't keep state of at which pass we are, we can loop at
-        * LRU_RETRY, since we have no guarantees that the caller will be able
-        * to do something other than retry on the next pass. We handle this by
-        * allowing at most one retry per object. This should not be altered
-        * by any condition other than LRU_RETRY.
-        */
-       bool first_pass = true;
 
        spin_lock(&nlru->lock);
 restart:
        list_for_each_safe(item, n, &nlru->list) {
                enum lru_status ret;
+
+               /*
+                * decrement nr_to_walk first so that we don't livelock if we
+                * get stuck on large numbesr of LRU_RETRY items
+                */
+               if (--(*nr_to_walk) == 0)
+                       break;
+
                ret = isolate(item, &nlru->lock, cb_arg);
                switch (ret) {
                case LRU_REMOVED:
@@ -100,19 +100,14 @@ restart:
                case LRU_SKIP:
                        break;
                case LRU_RETRY:
-                       if (!first_pass) {
-                               first_pass = true;
-                               break;
-                       }
-                       first_pass = false;
+                       /*
+                        * The lru lock has been dropped, our list traversal is
+                        * now invalid and so we have to restart from scratch.
+                        */
                        goto restart;
                default:
                        BUG();
                }
-
-               if ((*nr_to_walk)-- == 0)
-                       break;
-
        }
 
        spin_unlock(&nlru->lock);