]> git.karo-electronics.de Git - karo-tx-linux.git/blobdiff - fs/dcache.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[karo-tx-linux.git] / fs / dcache.c
index 40707d88a9452aea7c995198c2d94bfb1e9775c5..06f65857a855725247c1190d243c0e19cccd8570 100644 (file)
@@ -150,7 +150,7 @@ static long get_nr_dentry_unused(void)
        return sum < 0 ? 0 : sum;
 }
 
-int proc_nr_dentry(ctl_table *table, int write, void __user *buffer,
+int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
                   size_t *lenp, loff_t *ppos)
 {
        dentry_stat.nr_dentry = get_nr_dentry();
@@ -246,16 +246,8 @@ static void __d_free(struct rcu_head *head)
        kmem_cache_free(dentry_cache, dentry); 
 }
 
-/*
- * no locks, please.
- */
-static void d_free(struct dentry *dentry)
+static void dentry_free(struct dentry *dentry)
 {
-       BUG_ON((int)dentry->d_lockref.count > 0);
-       this_cpu_dec(nr_dentry);
-       if (dentry->d_op && dentry->d_op->d_release)
-               dentry->d_op->d_release(dentry);
-
        /* if dentry was never visible to RCU, immediate free is OK */
        if (!(dentry->d_flags & DCACHE_RCUACCESS))
                __d_free(&dentry->d_u.d_rcu);
@@ -403,56 +395,6 @@ static void dentry_lru_add(struct dentry *dentry)
                d_lru_add(dentry);
 }
 
-/*
- * Remove a dentry with references from the LRU.
- *
- * If we are on the shrink list, then we can get to try_prune_one_dentry() and
- * lose our last reference through the parent walk. In this case, we need to
- * remove ourselves from the shrink list, not the LRU.
- */
-static void dentry_lru_del(struct dentry *dentry)
-{
-       if (dentry->d_flags & DCACHE_LRU_LIST) {
-               if (dentry->d_flags & DCACHE_SHRINK_LIST)
-                       return d_shrink_del(dentry);
-               d_lru_del(dentry);
-       }
-}
-
-/**
- * d_kill - kill dentry and return parent
- * @dentry: dentry to kill
- * @parent: parent dentry
- *
- * The dentry must already be unhashed and removed from the LRU.
- *
- * If this is the root of the dentry tree, return NULL.
- *
- * dentry->d_lock and parent->d_lock must be held by caller, and are dropped by
- * d_kill.
- */
-static struct dentry *d_kill(struct dentry *dentry, struct dentry *parent)
-       __releases(dentry->d_lock)
-       __releases(parent->d_lock)
-       __releases(dentry->d_inode->i_lock)
-{
-       list_del(&dentry->d_u.d_child);
-       /*
-        * Inform d_walk() that we are no longer attached to the
-        * dentry tree
-        */
-       dentry->d_flags |= DCACHE_DENTRY_KILLED;
-       if (parent)
-               spin_unlock(&parent->d_lock);
-       dentry_iput(dentry);
-       /*
-        * dentry_iput drops the locks, at which point nobody (except
-        * transient RCU lookups) can reach this dentry.
-        */
-       d_free(dentry);
-       return parent;
-}
-
 /**
  * d_drop - drop a dentry
  * @dentry: dentry to drop
@@ -499,37 +441,12 @@ void d_drop(struct dentry *dentry)
 }
 EXPORT_SYMBOL(d_drop);
 
-/*
- * Finish off a dentry we've decided to kill.
- * dentry->d_lock must be held, returns with it unlocked.
- * If ref is non-zero, then decrement the refcount too.
- * Returns dentry requiring refcount drop, or NULL if we're done.
- */
-static struct dentry *
-dentry_kill(struct dentry *dentry, int unlock_on_failure)
-       __releases(dentry->d_lock)
+static void __dentry_kill(struct dentry *dentry)
 {
-       struct inode *inode;
-       struct dentry *parent;
-
-       inode = dentry->d_inode;
-       if (inode && !spin_trylock(&inode->i_lock)) {
-relock:
-               if (unlock_on_failure) {
-                       spin_unlock(&dentry->d_lock);
-                       cpu_relax();
-               }
-               return dentry; /* try again with same dentry */
-       }
-       if (IS_ROOT(dentry))
-               parent = NULL;
-       else
+       struct dentry *parent = NULL;
+       bool can_free = true;
+       if (!IS_ROOT(dentry))
                parent = dentry->d_parent;
-       if (parent && !spin_trylock(&parent->d_lock)) {
-               if (inode)
-                       spin_unlock(&inode->i_lock);
-               goto relock;
-       }
 
        /*
         * The dentry is now unrecoverably dead to the world.
@@ -543,10 +460,105 @@ relock:
        if ((dentry->d_flags & DCACHE_OP_PRUNE) && !d_unhashed(dentry))
                dentry->d_op->d_prune(dentry);
 
-       dentry_lru_del(dentry);
+       if (dentry->d_flags & DCACHE_LRU_LIST) {
+               if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
+                       d_lru_del(dentry);
+       }
        /* if it was on the hash then remove it */
        __d_drop(dentry);
-       return d_kill(dentry, parent);
+       list_del(&dentry->d_u.d_child);
+       /*
+        * Inform d_walk() that we are no longer attached to the
+        * dentry tree
+        */
+       dentry->d_flags |= DCACHE_DENTRY_KILLED;
+       if (parent)
+               spin_unlock(&parent->d_lock);
+       dentry_iput(dentry);
+       /*
+        * dentry_iput drops the locks, at which point nobody (except
+        * transient RCU lookups) can reach this dentry.
+        */
+       BUG_ON((int)dentry->d_lockref.count > 0);
+       this_cpu_dec(nr_dentry);
+       if (dentry->d_op && dentry->d_op->d_release)
+               dentry->d_op->d_release(dentry);
+
+       spin_lock(&dentry->d_lock);
+       if (dentry->d_flags & DCACHE_SHRINK_LIST) {
+               dentry->d_flags |= DCACHE_MAY_FREE;
+               can_free = false;
+       }
+       spin_unlock(&dentry->d_lock);
+       if (likely(can_free))
+               dentry_free(dentry);
+}
+
+/*
+ * Finish off a dentry we've decided to kill.
+ * dentry->d_lock must be held, returns with it unlocked.
+ * If ref is non-zero, then decrement the refcount too.
+ * Returns dentry requiring refcount drop, or NULL if we're done.
+ */
+static struct dentry *dentry_kill(struct dentry *dentry)
+       __releases(dentry->d_lock)
+{
+       struct inode *inode = dentry->d_inode;
+       struct dentry *parent = NULL;
+
+       if (inode && unlikely(!spin_trylock(&inode->i_lock)))
+               goto failed;
+
+       if (!IS_ROOT(dentry)) {
+               parent = dentry->d_parent;
+               if (unlikely(!spin_trylock(&parent->d_lock))) {
+                       if (inode)
+                               spin_unlock(&inode->i_lock);
+                       goto failed;
+               }
+       }
+
+       __dentry_kill(dentry);
+       return parent;
+
+failed:
+       spin_unlock(&dentry->d_lock);
+       cpu_relax();
+       return dentry; /* try again with same dentry */
+}
+
+static inline struct dentry *lock_parent(struct dentry *dentry)
+{
+       struct dentry *parent = dentry->d_parent;
+       if (IS_ROOT(dentry))
+               return NULL;
+       if (unlikely((int)dentry->d_lockref.count < 0))
+               return NULL;
+       if (likely(spin_trylock(&parent->d_lock)))
+               return parent;
+       rcu_read_lock();
+       spin_unlock(&dentry->d_lock);
+again:
+       parent = ACCESS_ONCE(dentry->d_parent);
+       spin_lock(&parent->d_lock);
+       /*
+        * We can't blindly lock dentry until we are sure
+        * that we won't violate the locking order.
+        * Any changes of dentry->d_parent must have
+        * been done with parent->d_lock held, so
+        * spin_lock() above is enough of a barrier
+        * for checking if it's still our child.
+        */
+       if (unlikely(parent != dentry->d_parent)) {
+               spin_unlock(&parent->d_lock);
+               goto again;
+       }
+       rcu_read_unlock();
+       if (parent != dentry)
+               spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
+       else
+               parent = NULL;
+       return parent;
 }
 
 /* 
@@ -602,7 +614,7 @@ repeat:
        return;
 
 kill_it:
-       dentry = dentry_kill(dentry, 1);
+       dentry = dentry_kill(dentry);
        if (dentry)
                goto repeat;
 }
@@ -815,64 +827,15 @@ restart:
 }
 EXPORT_SYMBOL(d_prune_aliases);
 
-/*
- * Try to throw away a dentry - free the inode, dput the parent.
- * Requires dentry->d_lock is held, and dentry->d_count == 0.
- * Releases dentry->d_lock.
- *
- * This may fail if locks cannot be acquired no problem, just try again.
- */
-static struct dentry * try_prune_one_dentry(struct dentry *dentry)
-       __releases(dentry->d_lock)
-{
-       struct dentry *parent;
-
-       parent = dentry_kill(dentry, 0);
-       /*
-        * If dentry_kill returns NULL, we have nothing more to do.
-        * if it returns the same dentry, trylocks failed. In either
-        * case, just loop again.
-        *
-        * Otherwise, we need to prune ancestors too. This is necessary
-        * to prevent quadratic behavior of shrink_dcache_parent(), but
-        * is also expected to be beneficial in reducing dentry cache
-        * fragmentation.
-        */
-       if (!parent)
-               return NULL;
-       if (parent == dentry)
-               return dentry;
-
-       /* Prune ancestors. */
-       dentry = parent;
-       while (dentry) {
-               if (lockref_put_or_lock(&dentry->d_lockref))
-                       return NULL;
-               dentry = dentry_kill(dentry, 1);
-       }
-       return NULL;
-}
-
 static void shrink_dentry_list(struct list_head *list)
 {
-       struct dentry *dentry;
-
-       rcu_read_lock();
-       for (;;) {
-               dentry = list_entry_rcu(list->prev, struct dentry, d_lru);
-               if (&dentry->d_lru == list)
-                       break; /* empty */
+       struct dentry *dentry, *parent;
 
-               /*
-                * Get the dentry lock, and re-verify that the dentry is
-                * this on the shrinking list. If it is, we know that
-                * DCACHE_SHRINK_LIST and DCACHE_LRU_LIST are set.
-                */
+       while (!list_empty(list)) {
+               struct inode *inode;
+               dentry = list_entry(list->prev, struct dentry, d_lru);
                spin_lock(&dentry->d_lock);
-               if (dentry != list_entry(list->prev, struct dentry, d_lru)) {
-                       spin_unlock(&dentry->d_lock);
-                       continue;
-               }
+               parent = lock_parent(dentry);
 
                /*
                 * The dispose list is isolated and dentries are not accounted
@@ -885,30 +848,63 @@ static void shrink_dentry_list(struct list_head *list)
                 * We found an inuse dentry which was not removed from
                 * the LRU because of laziness during lookup. Do not free it.
                 */
-               if (dentry->d_lockref.count) {
+               if ((int)dentry->d_lockref.count > 0) {
                        spin_unlock(&dentry->d_lock);
+                       if (parent)
+                               spin_unlock(&parent->d_lock);
                        continue;
                }
-               rcu_read_unlock();
 
-               /*
-                * If 'try_to_prune()' returns a dentry, it will
-                * be the same one we passed in, and d_lock will
-                * have been held the whole time, so it will not
-                * have been added to any other lists. We failed
-                * to get the inode lock.
-                *
-                * We just add it back to the shrink list.
-                */
-               dentry = try_prune_one_dentry(dentry);
 
-               rcu_read_lock();
-               if (dentry) {
+               if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) {
+                       bool can_free = dentry->d_flags & DCACHE_MAY_FREE;
+                       spin_unlock(&dentry->d_lock);
+                       if (parent)
+                               spin_unlock(&parent->d_lock);
+                       if (can_free)
+                               dentry_free(dentry);
+                       continue;
+               }
+
+               inode = dentry->d_inode;
+               if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
                        d_shrink_add(dentry, list);
                        spin_unlock(&dentry->d_lock);
+                       if (parent)
+                               spin_unlock(&parent->d_lock);
+                       continue;
+               }
+
+               __dentry_kill(dentry);
+
+               /*
+                * We need to prune ancestors too. This is necessary to prevent
+                * quadratic behavior of shrink_dcache_parent(), but is also
+                * expected to be beneficial in reducing dentry cache
+                * fragmentation.
+                */
+               dentry = parent;
+               while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) {
+                       parent = lock_parent(dentry);
+                       if (dentry->d_lockref.count != 1) {
+                               dentry->d_lockref.count--;
+                               spin_unlock(&dentry->d_lock);
+                               if (parent)
+                                       spin_unlock(&parent->d_lock);
+                               break;
+                       }
+                       inode = dentry->d_inode;        /* can't be NULL */
+                       if (unlikely(!spin_trylock(&inode->i_lock))) {
+                               spin_unlock(&dentry->d_lock);
+                               if (parent)
+                                       spin_unlock(&parent->d_lock);
+                               cpu_relax();
+                               continue;
+                       }
+                       __dentry_kill(dentry);
+                       dentry = parent;
                }
        }
-       rcu_read_unlock();
 }
 
 static enum lru_status
@@ -1261,34 +1257,23 @@ static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
        if (data->start == dentry)
                goto out;
 
-       /*
-        * move only zero ref count dentries to the dispose list.
-        *
-        * Those which are presently on the shrink list, being processed
-        * by shrink_dentry_list(), shouldn't be moved.  Otherwise the
-        * loop in shrink_dcache_parent() might not make any progress
-        * and loop forever.
-        */
-       if (dentry->d_lockref.count) {
-               dentry_lru_del(dentry);
-       } else if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
-               /*
-                * We can't use d_lru_shrink_move() because we
-                * need to get the global LRU lock and do the
-                * LRU accounting.
-                */
-               d_lru_del(dentry);
-               d_shrink_add(dentry, &data->dispose);
+       if (dentry->d_flags & DCACHE_SHRINK_LIST) {
                data->found++;
-               ret = D_WALK_NORETRY;
+       } else {
+               if (dentry->d_flags & DCACHE_LRU_LIST)
+                       d_lru_del(dentry);
+               if (!dentry->d_lockref.count) {
+                       d_shrink_add(dentry, &data->dispose);
+                       data->found++;
+               }
        }
        /*
         * We can return to the caller if we have found some (this
         * ensures forward progress). We'll be coming back to find
         * the rest.
         */
-       if (data->found && need_resched())
-               ret = D_WALK_QUIT;
+       if (!list_empty(&data->dispose))
+               ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
 out:
        return ret;
 }
@@ -1318,45 +1303,35 @@ void shrink_dcache_parent(struct dentry *parent)
 }
 EXPORT_SYMBOL(shrink_dcache_parent);
 
-static enum d_walk_ret umount_collect(void *_data, struct dentry *dentry)
+static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
 {
-       struct select_data *data = _data;
-       enum d_walk_ret ret = D_WALK_CONTINUE;
+       /* it has busy descendents; complain about those instead */
+       if (!list_empty(&dentry->d_subdirs))
+               return D_WALK_CONTINUE;
 
-       if (dentry->d_lockref.count) {
-               dentry_lru_del(dentry);
-               if (likely(!list_empty(&dentry->d_subdirs)))
-                       goto out;
-               if (dentry == data->start && dentry->d_lockref.count == 1)
-                       goto out;
-               printk(KERN_ERR
-                      "BUG: Dentry %p{i=%lx,n=%s}"
-                      " still in use (%d)"
-                      " [unmount of %s %s]\n",
+       /* root with refcount 1 is fine */
+       if (dentry == _data && dentry->d_lockref.count == 1)
+               return D_WALK_CONTINUE;
+
+       printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
+                       " still in use (%d) [unmount of %s %s]\n",
                       dentry,
                       dentry->d_inode ?
                       dentry->d_inode->i_ino : 0UL,
-                      dentry->d_name.name,
+                      dentry,
                       dentry->d_lockref.count,
                       dentry->d_sb->s_type->name,
                       dentry->d_sb->s_id);
-               BUG();
-       } else if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
-               /*
-                * We can't use d_lru_shrink_move() because we
-                * need to get the global LRU lock and do the
-                * LRU accounting.
-                */
-               if (dentry->d_flags & DCACHE_LRU_LIST)
-                       d_lru_del(dentry);
-               d_shrink_add(dentry, &data->dispose);
-               data->found++;
-               ret = D_WALK_NORETRY;
-       }
-out:
-       if (data->found && need_resched())
-               ret = D_WALK_QUIT;
-       return ret;
+       WARN_ON(1);
+       return D_WALK_CONTINUE;
+}
+
+static void do_one_tree(struct dentry *dentry)
+{
+       shrink_dcache_parent(dentry);
+       d_walk(dentry, dentry, umount_check, NULL);
+       d_drop(dentry);
+       dput(dentry);
 }
 
 /*
@@ -1366,40 +1341,15 @@ void shrink_dcache_for_umount(struct super_block *sb)
 {
        struct dentry *dentry;
 
-       if (down_read_trylock(&sb->s_umount))
-               BUG();
+       WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
 
        dentry = sb->s_root;
        sb->s_root = NULL;
-       for (;;) {
-               struct select_data data;
-
-               INIT_LIST_HEAD(&data.dispose);
-               data.start = dentry;
-               data.found = 0;
-
-               d_walk(dentry, &data, umount_collect, NULL);
-               if (!data.found)
-                       break;
-
-               shrink_dentry_list(&data.dispose);
-               cond_resched();
-       }
-       d_drop(dentry);
-       dput(dentry);
+       do_one_tree(dentry);
 
        while (!hlist_bl_empty(&sb->s_anon)) {
-               struct select_data data;
-               dentry = hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash);
-
-               INIT_LIST_HEAD(&data.dispose);
-               data.start = NULL;
-               data.found = 0;
-
-               d_walk(dentry, &data, umount_collect, NULL);
-               if (data.found)
-                       shrink_dentry_list(&data.dispose);
-               cond_resched();
+               dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash));
+               do_one_tree(dentry);
        }
 }
 
@@ -1647,8 +1597,7 @@ static void __d_instantiate(struct dentry *dentry, struct inode *inode)
        unsigned add_flags = d_flags_for_inode(inode);
 
        spin_lock(&dentry->d_lock);
-       dentry->d_flags &= ~DCACHE_ENTRY_TYPE;
-       dentry->d_flags |= add_flags;
+       __d_set_type(dentry, add_flags);
        if (inode)
                hlist_add_head(&dentry->d_alias, &inode->i_dentry);
        dentry->d_inode = inode;