btrfs: fix dead lock while running replace and defrag concurrently
authorGui Hecheng <guihc.fnst@cn.fujitsu.com>
Mon, 10 Nov 2014 07:36:08 +0000 (15:36 +0800)
committerChris Mason <clm@fb.com>
Fri, 21 Nov 2014 01:20:08 +0000 (17:20 -0800)
This can be reproduced by fstests: btrfs/070

The scenario is like the following:

replace worker thread defrag thread
--------------------- -------------
copy_nocow_pages_worker btrfs_defrag_file
  copy_nocow_pages_for_inode     ...
  btrfs_writepages
  |A| lock_extent_bits     extent_write_cache_pages
|B|   lock_page
__extent_writepage
...   writepage_delalloc
    find_lock_delalloc_range
|B|        lock_extent_bits
  find_or_create_page
    pagecache_get_page
  |A| lock_page

This leads to an ABBA pattern deadlock. To fix it,
o we just change it to an AABB pattern which means to @unlock_extent_bits()
  before we @lock_page(), and in this way the @extent_read_full_page_nolock()
  is no longer in an locked context, so change it back to @extent_read_full_page()
  to regain protection.

o Since we @unlock_extent_bits() earlier, then before @write_page_nocow(),
  the extent may not really point at the physical block we want, so we
  have to check it before write.

Signed-off-by: Gui Hecheng <guihc.fnst@cn.fujitsu.com>
Tested-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/scrub.c

index efa08311382725c6770f3899033778ba8e3129ac..4325bb0111d9d9e6003ee2c0050fc71d35ecca92 100644 (file)
@@ -3310,6 +3310,50 @@ out:
        scrub_pending_trans_workers_dec(sctx);
 }
 
+static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
+                                u64 logical)
+{
+       struct extent_state *cached_state = NULL;
+       struct btrfs_ordered_extent *ordered;
+       struct extent_io_tree *io_tree;
+       struct extent_map *em;
+       u64 lockstart = start, lockend = start + len - 1;
+       int ret = 0;
+
+       io_tree = &BTRFS_I(inode)->io_tree;
+
+       lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
+       ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
+       if (ordered) {
+               btrfs_put_ordered_extent(ordered);
+               ret = 1;
+               goto out_unlock;
+       }
+
+       em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
+       if (IS_ERR(em)) {
+               ret = PTR_ERR(em);
+               goto out_unlock;
+       }
+
+       /*
+        * This extent does not actually cover the logical extent anymore,
+        * move on to the next inode.
+        */
+       if (em->block_start > logical ||
+           em->block_start + em->block_len < logical + len) {
+               free_extent_map(em);
+               ret = 1;
+               goto out_unlock;
+       }
+       free_extent_map(em);
+
+out_unlock:
+       unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
+                            GFP_NOFS);
+       return ret;
+}
+
 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
                                      struct scrub_copy_nocow_ctx *nocow_ctx)
 {
@@ -3318,13 +3362,10 @@ static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
        struct inode *inode;
        struct page *page;
        struct btrfs_root *local_root;
-       struct btrfs_ordered_extent *ordered;
-       struct extent_map *em;
-       struct extent_state *cached_state = NULL;
        struct extent_io_tree *io_tree;
        u64 physical_for_dev_replace;
+       u64 nocow_ctx_logical;
        u64 len = nocow_ctx->len;
-       u64 lockstart = offset, lockend = offset + len - 1;
        unsigned long index;
        int srcu_index;
        int ret = 0;
@@ -3356,30 +3397,13 @@ static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 
        physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
        io_tree = &BTRFS_I(inode)->io_tree;
+       nocow_ctx_logical = nocow_ctx->logical;
 
-       lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
-       ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
-       if (ordered) {
-               btrfs_put_ordered_extent(ordered);
-               goto out_unlock;
-       }
-
-       em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
-       if (IS_ERR(em)) {
-               ret = PTR_ERR(em);
-               goto out_unlock;
-       }
-
-       /*
-        * This extent does not actually cover the logical extent anymore,
-        * move on to the next inode.
-        */
-       if (em->block_start > nocow_ctx->logical ||
-           em->block_start + em->block_len < nocow_ctx->logical + len) {
-               free_extent_map(em);
-               goto out_unlock;
+       ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
+       if (ret) {
+               ret = ret > 0 ? 0 : ret;
+               goto out;
        }
-       free_extent_map(em);
 
        while (len >= PAGE_CACHE_SIZE) {
                index = offset >> PAGE_CACHE_SHIFT;
@@ -3396,7 +3420,7 @@ again:
                                goto next_page;
                } else {
                        ClearPageError(page);
-                       err = extent_read_full_page_nolock(io_tree, page,
+                       err = extent_read_full_page(io_tree, page,
                                                           btrfs_get_extent,
                                                           nocow_ctx->mirror_num);
                        if (err) {
@@ -3421,6 +3445,14 @@ again:
                                goto next_page;
                        }
                }
+
+               ret = check_extent_to_block(inode, offset, len,
+                                           nocow_ctx_logical);
+               if (ret) {
+                       ret = ret > 0 ? 0 : ret;
+                       goto next_page;
+               }
+
                err = write_page_nocow(nocow_ctx->sctx,
                                       physical_for_dev_replace, page);
                if (err)
@@ -3434,12 +3466,10 @@ next_page:
 
                offset += PAGE_CACHE_SIZE;
                physical_for_dev_replace += PAGE_CACHE_SIZE;
+               nocow_ctx_logical += PAGE_CACHE_SIZE;
                len -= PAGE_CACHE_SIZE;
        }
        ret = COPY_COMPLETE;
-out_unlock:
-       unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
-                            GFP_NOFS);
 out:
        mutex_unlock(&inode->i_mutex);
        iput(inode);