mm: __set_page_dirty uses spin_lock_irqsave instead of spin_lock_irq
KOSAKI Motohiro [Thu, 6 Feb 2014 20:04:28 +0000 (12:04 -0800)]
commit 227d53b397a32a7614667b3ecaf1d89902fb6c12 upstream.

To use spin_{un}lock_irq is dangerous if caller disabled interrupt.
During aio buffer migration, we have a possibility to see the following
call stack.

aio_migratepage  [disable interrupt]
  migrate_page_copy
    clear_page_dirty_for_io
      set_page_dirty
        __set_page_dirty_buffers
          __set_page_dirty
            spin_lock_irq

This mean, current aio migration is a deadlockable.  spin_lock_irqsave
is a safer alternative and we should use it.

Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reported-by: David Rientjes rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/buffer.c

index d2a4d1b..75964d7 100644 (file)
@@ -620,14 +620,16 @@ EXPORT_SYMBOL(mark_buffer_dirty_inode);
 static void __set_page_dirty(struct page *page,
                struct address_space *mapping, int warn)
 {
-       spin_lock_irq(&mapping->tree_lock);
+       unsigned long flags;
+
+       spin_lock_irqsave(&mapping->tree_lock, flags);
        if (page->mapping) {    /* Race with truncate? */
                WARN_ON_ONCE(warn && !PageUptodate(page));
                account_page_dirtied(page, mapping);
                radix_tree_tag_set(&mapping->page_tree,
                                page_index(page), PAGECACHE_TAG_DIRTY);
        }
-       spin_unlock_irq(&mapping->tree_lock);
+       spin_unlock_irqrestore(&mapping->tree_lock, flags);
        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 }