]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
list_lru: remove special case function list_lru_dispose_all.
authorGlauber Costa <glommer@gmail.com>
Thu, 1 Aug 2013 06:36:12 +0000 (16:36 +1000)
committerStephen Rothwell <sfr@canb.auug.org.au>
Fri, 2 Aug 2013 07:37:14 +0000 (17:37 +1000)
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code).  At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway.  However, it's
implementation is buggy in the following way:

When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST.  However, this is done without the
nlru->lock taken.  The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time.  When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is.  Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.

The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock).  To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.

Such mechanism already exists in the form of the list_lru_walk_cb
callback.  So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.

With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.

Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
fs/dcache.c
include/linux/list_lru.h
mm/list_lru.c

index 2275bd416736a32cf4ba64e4d3865b99186f62aa..21f36b4d95ccc7cb2ee66d07b605cc0689c49fa8 100644 (file)
@@ -910,27 +910,29 @@ long prune_dcache_sb(struct super_block *sb, unsigned long nr_to_scan)
        return freed;
 }
 
-/*
- * Mark all the dentries as on being the dispose list so we don't think they are
- * still on the LRU if we try to kill them from ascending the parent chain in
- * try_prune_one_dentry() rather than directly from the dispose list.
- */
-static void
-shrink_dcache_list(
-       struct list_head *dispose)
+static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
+                                               spinlock_t *lru_lock, void *arg)
 {
-       struct dentry *dentry;
+       struct list_head *freeable = arg;
+       struct dentry   *dentry = container_of(item, struct dentry, d_lru);
 
-       rcu_read_lock();
-       list_for_each_entry_rcu(dentry, dispose, d_lru) {
-               spin_lock(&dentry->d_lock);
-               dentry->d_flags |= DCACHE_SHRINK_LIST;
-               spin_unlock(&dentry->d_lock);
-       }
-       rcu_read_unlock();
-       shrink_dentry_list(dispose);
+       /*
+        * we are inverting the lru lock/dentry->d_lock here,
+        * so use a trylock. If we fail to get the lock, just skip
+        * it
+        */
+       if (!spin_trylock(&dentry->d_lock))
+               return LRU_SKIP;
+
+       dentry->d_flags |= DCACHE_SHRINK_LIST;
+       list_move_tail(&dentry->d_lru, freeable);
+       this_cpu_dec(nr_dentry_unused);
+       spin_unlock(&dentry->d_lock);
+
+       return LRU_REMOVED;
 }
 
+
 /**
  * shrink_dcache_sb - shrink dcache for a superblock
  * @sb: superblock
@@ -940,10 +942,17 @@ shrink_dcache_list(
  */
 void shrink_dcache_sb(struct super_block *sb)
 {
-       long disposed;
+       long freed;
 
-       disposed = list_lru_dispose_all(&sb->s_dentry_lru, shrink_dcache_list);
-       this_cpu_sub(nr_dentry_unused, disposed);
+       do {
+               LIST_HEAD(dispose);
+
+               freed = list_lru_walk(&sb->s_dentry_lru,
+                       dentry_lru_isolate_shrink, &dispose, UINT_MAX);
+
+               this_cpu_sub(nr_dentry_unused, freed);
+               shrink_dentry_list(&dispose);
+       } while (freed > 0);
 }
 EXPORT_SYMBOL(shrink_dcache_sb);
 
index 2fe13e1a809aab48c4948031a6f4e5680fe85df7..4d02ad3badab51b8621b2dbffa61db76a02d6784 100644 (file)
@@ -137,21 +137,4 @@ list_lru_walk(struct list_lru *lru, list_lru_walk_cb isolate,
        }
        return isolated;
 }
-
-typedef void (*list_lru_dispose_cb)(struct list_head *dispose_list);
-/**
- * list_lru_dispose_all: forceably flush all elements in an @lru
- * @lru: the lru pointer
- * @dispose: callback function to be called for each lru list.
- *
- * This function will forceably isolate all elements into the dispose list, and
- * call the @dispose callback to flush the list. Please note that the callback
- * should expect items in any state, clean or dirty, and be able to flush all of
- * them.
- *
- * Return value: how many objects were freed. It should be equal to all objects
- * in the list_lru.
- */
-unsigned long
-list_lru_dispose_all(struct list_lru *lru, list_lru_dispose_cb dispose);
 #endif /* _LRU_LIST_H */
index 86cb55464f71a892c7d765907ea4de5182102001..f91c24188573af8b01dccf9e2b3fab7586b18f8e 100644 (file)
@@ -112,48 +112,6 @@ restart:
 }
 EXPORT_SYMBOL_GPL(list_lru_walk_node);
 
-static unsigned long list_lru_dispose_all_node(struct list_lru *lru, int nid,
-                                              list_lru_dispose_cb dispose)
-{
-       struct list_lru_node    *nlru = &lru->node[nid];
-       LIST_HEAD(dispose_list);
-       unsigned long disposed = 0;
-
-       spin_lock(&nlru->lock);
-       while (!list_empty(&nlru->list)) {
-               list_splice_init(&nlru->list, &dispose_list);
-               disposed += nlru->nr_items;
-               nlru->nr_items = 0;
-               node_clear(nid, lru->active_nodes);
-               spin_unlock(&nlru->lock);
-
-               dispose(&dispose_list);
-
-               spin_lock(&nlru->lock);
-       }
-       spin_unlock(&nlru->lock);
-       return disposed;
-}
-
-unsigned long list_lru_dispose_all(struct list_lru *lru,
-                                  list_lru_dispose_cb dispose)
-{
-       unsigned long disposed;
-       unsigned long total = 0;
-       int nid;
-
-       do {
-               disposed = 0;
-               for_each_node_mask(nid, lru->active_nodes) {
-                       disposed += list_lru_dispose_all_node(lru, nid,
-                                                             dispose);
-               }
-               total += disposed;
-       } while (disposed != 0);
-
-       return total;
-}
-
 int list_lru_init(struct list_lru *lru)
 {
        int i;