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);