Btrfs: be more selective in the defrag ioctl
authorChris Mason <chris.mason@oracle.com>
Wed, 10 Mar 2010 15:52:59 +0000 (10:52 -0500)
committerChris Mason <chris.mason@oracle.com>
Mon, 15 Mar 2010 15:00:10 +0000 (11:00 -0400)
The btrfs defrag ioctl had some bugs around delalloc accounting, and it
wasn't properly skipping pages that were not in the mapping.

It wasn't properly clearing the page checked flag, which could make the
writeback code ignore the page forever while pinning it as dirty.

This commit fixes those problems and makes defrag a little smarter.  It
skips holes and it doesn't waste time defragging large extents.  If a
tiny extent comes before a very large extent, it will defrag both of
them to make sure the tiny extent ends up next to something big.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/ioctl.c

index 7875a75..3a89cd7 100644 (file)
@@ -475,6 +475,73 @@ out_unlock:
        return error;
 }
 
+static int should_defrag_range(struct inode *inode, u64 start, u64 len,
+                              u64 *last_len, u64 *skip, u64 *defrag_end)
+{
+       struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
+       struct extent_map *em = NULL;
+       struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
+       int ret = 1;
+
+       /*
+        * make sure that once we start defragging and extent, we keep on
+        * defragging it
+        */
+       if (start < *defrag_end)
+               return 1;
+
+       *skip = 0;
+
+       /*
+        * hopefully we have this extent in the tree already, try without
+        * the full extent lock
+        */
+       read_lock(&em_tree->lock);
+       em = lookup_extent_mapping(em_tree, start, len);
+       read_unlock(&em_tree->lock);
+
+       if (!em) {
+               /* get the big lock and read metadata off disk */
+               lock_extent(io_tree, start, start + len - 1, GFP_NOFS);
+               em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
+               unlock_extent(io_tree, start, start + len - 1, GFP_NOFS);
+
+               if (!em)
+                       return 0;
+       }
+
+       /* this will cover holes, and inline extents */
+       if (em->block_start >= EXTENT_MAP_LAST_BYTE)
+               ret = 0;
+
+       /*
+        * we hit a real extent, if it is big don't bother defragging it again
+        */
+       if ((*last_len == 0 || *last_len >= 256 * 1024) &&
+           em->len >= 256 * 1024)
+               ret = 0;
+
+       /*
+        * last_len ends up being a counter of how many bytes we've defragged.
+        * every time we choose not to defrag an extent, we reset *last_len
+        * so that the next tiny extent will force a defrag.
+        *
+        * The end result of this is that tiny extents before a single big
+        * extent will force at least part of that big extent to be defragged.
+        */
+       if (ret) {
+               *last_len += len;
+               *defrag_end = extent_map_end(em);
+       } else {
+               *last_len = 0;
+               *skip = extent_map_end(em);
+               *defrag_end = 0;
+       }
+
+       free_extent_map(em);
+       return ret;
+}
+
 static int btrfs_defrag_file(struct file *file)
 {
        struct inode *inode = fdentry(file)->d_inode;
@@ -487,37 +554,86 @@ static int btrfs_defrag_file(struct file *file)
        unsigned long total_read = 0;
        u64 page_start;
        u64 page_end;
+       u64 last_len = 0;
+       u64 skip = 0;
+       u64 defrag_end = 0;
        unsigned long i;
        int ret;
 
-       ret = btrfs_check_data_free_space(root, inode, inode->i_size);
-       if (ret)
-               return -ENOSPC;
+       if (inode->i_size == 0)
+               return 0;
+
+       last_index = (inode->i_size - 1) >> PAGE_CACHE_SHIFT;
+       i = 0;
+       while (i <= last_index) {
+               if (!should_defrag_range(inode, (u64)i << PAGE_CACHE_SHIFT,
+                                       PAGE_CACHE_SIZE, &last_len, &skip,
+                                       &defrag_end)) {
+                       unsigned long next;
+                       /*
+                        * the should_defrag function tells us how much to skip
+                        * bump our counter by the suggested amount
+                        */
+                       next = (skip + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+                       i = max(i + 1, next);
+                       continue;
+               }
 
-       mutex_lock(&inode->i_mutex);
-       last_index = inode->i_size >> PAGE_CACHE_SHIFT;
-       for (i = 0; i <= last_index; i++) {
                if (total_read % ra_pages == 0) {
                        btrfs_force_ra(inode->i_mapping, &file->f_ra, file, i,
                                       min(last_index, i + ra_pages - 1));
                }
                total_read++;
+               mutex_lock(&inode->i_mutex);
+
+               ret = btrfs_check_data_free_space(root, inode, PAGE_CACHE_SIZE);
+               if (ret) {
+                       ret = -ENOSPC;
+                       break;
+               }
+
+               ret = btrfs_reserve_metadata_for_delalloc(root, inode, 1);
+               if (ret) {
+                       btrfs_free_reserved_data_space(root, inode,
+                                                      PAGE_CACHE_SIZE);
+                       ret = -ENOSPC;
+                       break;
+               }
 again:
+               if (inode->i_size == 0 ||
+                   i > ((inode->i_size - 1) >> PAGE_CACHE_SHIFT)) {
+                       ret = 0;
+                       goto err_reservations;
+               }
+
                page = grab_cache_page(inode->i_mapping, i);
                if (!page)
-                       goto out_unlock;
+                       goto err_reservations;
+
                if (!PageUptodate(page)) {
                        btrfs_readpage(NULL, page);
                        lock_page(page);
                        if (!PageUptodate(page)) {
                                unlock_page(page);
                                page_cache_release(page);
-                               goto out_unlock;
+                               goto err_reservations;
                        }
                }
 
+               if (page->mapping != inode->i_mapping) {
+                       unlock_page(page);
+                       page_cache_release(page);
+                       goto again;
+               }
+
                wait_on_page_writeback(page);
 
+               if (PageDirty(page)) {
+                       btrfs_free_reserved_data_space(root, inode,
+                                                      PAGE_CACHE_SIZE);
+                       goto loop_unlock;
+               }
+
                page_start = (u64)page->index << PAGE_CACHE_SHIFT;
                page_end = page_start + PAGE_CACHE_SIZE - 1;
                lock_extent(io_tree, page_start, page_end, GFP_NOFS);
@@ -538,18 +654,32 @@ again:
                 * page if it is dirtied again later
                 */
                clear_page_dirty_for_io(page);
+               clear_extent_bits(&BTRFS_I(inode)->io_tree, page_start,
+                                 page_end, EXTENT_DIRTY | EXTENT_DELALLOC |
+                                 EXTENT_DO_ACCOUNTING, GFP_NOFS);
 
                btrfs_set_extent_delalloc(inode, page_start, page_end);
+               ClearPageChecked(page);
                set_page_dirty(page);
                unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
+
+loop_unlock:
                unlock_page(page);
                page_cache_release(page);
+               mutex_unlock(&inode->i_mutex);
+
+               btrfs_unreserve_metadata_for_delalloc(root, inode, 1);
                balance_dirty_pages_ratelimited_nr(inode->i_mapping, 1);
+               i++;
        }
 
-out_unlock:
-       mutex_unlock(&inode->i_mutex);
        return 0;
+
+err_reservations:
+       mutex_unlock(&inode->i_mutex);
+       btrfs_free_reserved_data_space(root, inode, PAGE_CACHE_SIZE);
+       btrfs_unreserve_metadata_for_delalloc(root, inode, 1);
+       return ret;
 }
 
 static noinline int btrfs_ioctl_resize(struct btrfs_root *root,