Btrfs: fix ->iterate_shared() by upgrading i_rwsem for delayed nodes
authorOmar Sandoval <osandov@fb.com>
Fri, 20 May 2016 20:50:33 +0000 (13:50 -0700)
committerChris Mason <clm@fb.com>
Sat, 25 Jun 2016 13:20:10 +0000 (06:20 -0700)
Commit fe742fd4f90f ("Revert "btrfs: switch to ->iterate_shared()"")
backed out the conversion to ->iterate_shared() for Btrfs because the
delayed inode handling in btrfs_real_readdir() is racy. However, we can
still do readdir in parallel if there are no delayed nodes.

This is a temporary fix which upgrades the shared inode lock to an
exclusive lock only when we have delayed items until we come up with a
more complete solution. While we're here, rename the
btrfs_{get,put}_delayed_items functions to make it very clear that
they're just for readdir.

Tested with xfstests and by doing a parallel kernel build:

while make tinyconfig && make -j4 && git clean dqfx; do
:
done

along with a bunch of parallel finds in another shell:

while true; do
for ((i=0; i<4; i++)); do
find . >/dev/null &
done
wait
done

Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/delayed-inode.c
fs/btrfs/delayed-inode.h
fs/btrfs/inode.c

index 61561c2a3f96da0917a0819fb9db48972385496d..d3aaabbfada0adcbb09936a5d47063a4adec65dd 100644 (file)
@@ -1606,15 +1606,23 @@ int btrfs_inode_delayed_dir_index_count(struct inode *inode)
        return 0;
 }
 
-void btrfs_get_delayed_items(struct inode *inode, struct list_head *ins_list,
-                            struct list_head *del_list)
+bool btrfs_readdir_get_delayed_items(struct inode *inode,
+                                    struct list_head *ins_list,
+                                    struct list_head *del_list)
 {
        struct btrfs_delayed_node *delayed_node;
        struct btrfs_delayed_item *item;
 
        delayed_node = btrfs_get_delayed_node(inode);
        if (!delayed_node)
-               return;
+               return false;
+
+       /*
+        * We can only do one readdir with delayed items at a time because of
+        * item->readdir_list.
+        */
+       inode_unlock_shared(inode);
+       inode_lock(inode);
 
        mutex_lock(&delayed_node->mutex);
        item = __btrfs_first_delayed_insertion_item(delayed_node);
@@ -1641,10 +1649,13 @@ void btrfs_get_delayed_items(struct inode *inode, struct list_head *ins_list,
         * requeue or dequeue this delayed node.
         */
        atomic_dec(&delayed_node->refs);
+
+       return true;
 }
 
-void btrfs_put_delayed_items(struct list_head *ins_list,
-                            struct list_head *del_list)
+void btrfs_readdir_put_delayed_items(struct inode *inode,
+                                    struct list_head *ins_list,
+                                    struct list_head *del_list)
 {
        struct btrfs_delayed_item *curr, *next;
 
@@ -1659,6 +1670,12 @@ void btrfs_put_delayed_items(struct list_head *ins_list,
                if (atomic_dec_and_test(&curr->refs))
                        kfree(curr);
        }
+
+       /*
+        * The VFS is going to do up_read(), so we need to downgrade back to a
+        * read lock.
+        */
+       downgrade_write(&inode->i_rwsem);
 }
 
 int btrfs_should_delete_dir_index(struct list_head *del_list,
index 0167853c84aea2d93a8fdd0314342e0a1ef42d1b..2495b3d4075f81538fde3ad0ca020648c164db91 100644 (file)
@@ -137,10 +137,12 @@ void btrfs_kill_all_delayed_nodes(struct btrfs_root *root);
 void btrfs_destroy_delayed_inodes(struct btrfs_root *root);
 
 /* Used for readdir() */
-void btrfs_get_delayed_items(struct inode *inode, struct list_head *ins_list,
-                            struct list_head *del_list);
-void btrfs_put_delayed_items(struct list_head *ins_list,
-                            struct list_head *del_list);
+bool btrfs_readdir_get_delayed_items(struct inode *inode,
+                                    struct list_head *ins_list,
+                                    struct list_head *del_list);
+void btrfs_readdir_put_delayed_items(struct inode *inode,
+                                    struct list_head *ins_list,
+                                    struct list_head *del_list);
 int btrfs_should_delete_dir_index(struct list_head *del_list,
                                  u64 index);
 int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
index d2be95cfb6d156f4323cd33733e51c20533fcb11..969a25c5abcbc59ae91f19ac9d16e2ce5bb5030c 100644 (file)
@@ -5757,6 +5757,7 @@ static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
        int name_len;
        int is_curr = 0;        /* ctx->pos points to the current index? */
        bool emitted;
+       bool put = false;
 
        /* FIXME, use a real flag for deciding about the key type */
        if (root->fs_info->tree_root == root)
@@ -5774,7 +5775,8 @@ static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
        if (key_type == BTRFS_DIR_INDEX_KEY) {
                INIT_LIST_HEAD(&ins_list);
                INIT_LIST_HEAD(&del_list);
-               btrfs_get_delayed_items(inode, &ins_list, &del_list);
+               put = btrfs_readdir_get_delayed_items(inode, &ins_list,
+                                                     &del_list);
        }
 
        key.type = key_type;
@@ -5921,8 +5923,8 @@ next:
 nopos:
        ret = 0;
 err:
-       if (key_type == BTRFS_DIR_INDEX_KEY)
-               btrfs_put_delayed_items(&ins_list, &del_list);
+       if (put)
+               btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
        btrfs_free_path(path);
        return ret;
 }
@@ -10534,7 +10536,7 @@ static const struct inode_operations btrfs_dir_ro_inode_operations = {
 static const struct file_operations btrfs_dir_file_operations = {
        .llseek         = generic_file_llseek,
        .read           = generic_read_dir,
-       .iterate        = btrfs_real_readdir,
+       .iterate_shared = btrfs_real_readdir,
        .unlocked_ioctl = btrfs_ioctl,
 #ifdef CONFIG_COMPAT
        .compat_ioctl   = btrfs_compat_ioctl,