Btrfs: Use a mutex in the extent buffer for tree block locking

This replaces the use of the page cache lock bit for locking, which wasn't
suitable for block size < page size and couldn't be used recursively.

The mutexes alone don't fix either problem, but they are the first step.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index 32bb4ed..7380449 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -2690,6 +2690,7 @@
 	eb = kmem_cache_zalloc(extent_buffer_cache, mask);
 	eb->start = start;
 	eb->len = len;
+	mutex_init(&eb->mutex);
 	spin_lock_irqsave(&leak_lock, flags);
 	list_add(&eb->leak_list, &buffers);
 	spin_unlock_irqrestore(&leak_lock, flags);
@@ -2837,6 +2838,7 @@
 
 	for (i = 0; i < num_pages; i++) {
 		page = extent_buffer_page(eb, i);
+		lock_page(page);
 		if (i == 0)
 			set_page_extent_head(page, eb->len);
 		else
@@ -2854,6 +2856,7 @@
 			end  = start + PAGE_CACHE_SIZE - 1;
 			if (test_range_bit(tree, start, end,
 					   EXTENT_DIRTY, 0)) {
+				unlock_page(page);
 				continue;
 			}
 		}
@@ -2865,6 +2868,7 @@
 						PAGECACHE_TAG_DIRTY);
 		}
 		read_unlock_irq(&page->mapping->tree_lock);
+		unlock_page(page);
 	}
 	return 0;
 }
@@ -2893,12 +2897,17 @@
 		 * on us if the page isn't already dirty.
 		 */
 		if (i == 0) {
+			lock_page(page);
 			set_page_extent_head(page, eb->len);
 		} else if (PagePrivate(page) &&
 			   page->private != EXTENT_PAGE_PRIVATE) {
+			lock_page(page);
 			set_page_extent_mapped(page);
+			unlock_page(page);
 		}
 		__set_page_dirty_nobuffers(extent_buffer_page(eb, i));
+		if (i == 0)
+			unlock_page(page);
 	}
 	return set_extent_dirty(tree, eb->start,
 				eb->start + eb->len - 1, GFP_NOFS);