Btrfs: fix preallocation and nodatacow checks in O_DIRECT
authorChris Mason <chris.mason@oracle.com>
Wed, 26 May 2010 15:04:10 +0000 (11:04 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 27 May 2010 01:34:45 +0000 (21:34 -0400)
The O_DIRECT code wasn't checking for multiple references
on preallocated or nodatacow extents.  This means it
wasn't honoring snapshots properly.

The fix here is to add an explicit check for multiple references
This also fixes the math for selecting the correct disk block,
making sure not to go past the end of the extent.

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

index 5d62f21..5ab120d 100644 (file)
@@ -5255,6 +5255,106 @@ out:
        return em;
 }
 
+/*
+ * returns 1 when the nocow is safe, < 1 on error, 0 if the
+ * block must be cow'd
+ */
+static noinline int can_nocow_odirect(struct btrfs_trans_handle *trans,
+                                     struct inode *inode, u64 offset, u64 len)
+{
+       struct btrfs_path *path;
+       int ret;
+       struct extent_buffer *leaf;
+       struct btrfs_root *root = BTRFS_I(inode)->root;
+       struct btrfs_file_extent_item *fi;
+       struct btrfs_key key;
+       u64 disk_bytenr;
+       u64 backref_offset;
+       u64 extent_end;
+       u64 num_bytes;
+       int slot;
+       int found_type;
+
+       path = btrfs_alloc_path();
+       if (!path)
+               return -ENOMEM;
+
+       ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
+                                      offset, 0);
+       if (ret < 0)
+               goto out;
+
+       slot = path->slots[0];
+       if (ret == 1) {
+               if (slot == 0) {
+                       /* can't find the item, must cow */
+                       ret = 0;
+                       goto out;
+               }
+               slot--;
+       }
+       ret = 0;
+       leaf = path->nodes[0];
+       btrfs_item_key_to_cpu(leaf, &key, slot);
+       if (key.objectid != inode->i_ino ||
+           key.type != BTRFS_EXTENT_DATA_KEY) {
+               /* not our file or wrong item type, must cow */
+               goto out;
+       }
+
+       if (key.offset > offset) {
+               /* Wrong offset, must cow */
+               goto out;
+       }
+
+       fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
+       found_type = btrfs_file_extent_type(leaf, fi);
+       if (found_type != BTRFS_FILE_EXTENT_REG &&
+           found_type != BTRFS_FILE_EXTENT_PREALLOC) {
+               /* not a regular extent, must cow */
+               goto out;
+       }
+       disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
+       backref_offset = btrfs_file_extent_offset(leaf, fi);
+
+       extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
+       if (extent_end < offset + len) {
+               /* extent doesn't include our full range, must cow */
+               goto out;
+       }
+
+       if (btrfs_extent_readonly(root, disk_bytenr))
+               goto out;
+
+       /*
+        * look for other files referencing this extent, if we
+        * find any we must cow
+        */
+       if (btrfs_cross_ref_exist(trans, root, inode->i_ino,
+                                 key.offset - backref_offset, disk_bytenr))
+               goto out;
+
+       /*
+        * adjust disk_bytenr and num_bytes to cover just the bytes
+        * in this extent we are about to write.  If there
+        * are any csums in that range we have to cow in order
+        * to keep the csums correct
+        */
+       disk_bytenr += backref_offset;
+       disk_bytenr += offset - key.offset;
+       num_bytes = min(offset + len, extent_end) - offset;
+       if (csum_exist_in_range(root, disk_bytenr, num_bytes))
+                               goto out;
+       /*
+        * all of the above have passed, it is safe to overwrite this extent
+        * without cow
+        */
+       ret = 1;
+out:
+       btrfs_free_path(path);
+       return ret;
+}
+
 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
                                   struct buffer_head *bh_result, int create)
 {
@@ -5262,6 +5362,7 @@ static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
        struct btrfs_root *root = BTRFS_I(inode)->root;
        u64 start = iblock << inode->i_blkbits;
        u64 len = bh_result->b_size;
+       struct btrfs_trans_handle *trans;
 
        em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
        if (IS_ERR(em))
@@ -5306,42 +5407,65 @@ static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
         * just use the extent.
         *
         */
-       if (!create)
+       if (!create) {
+               len = em->len - (start - em->start);
                goto map;
+       }
 
        if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
            ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
             em->block_start != EXTENT_MAP_HOLE)) {
-               u64 block_start;
                int type;
                int ret;
+               u64 block_start;
 
                if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
                        type = BTRFS_ORDERED_PREALLOC;
                else
                        type = BTRFS_ORDERED_NOCOW;
-               len = min(len, em->block_len - (start - em->start));
+               len = min(len, em->len - (start - em->start));
                block_start = em->block_start + (start - em->start);
-               ret = btrfs_add_ordered_extent_dio(inode, start,
-                                                  start, len, len, type);
-               if (ret) {
-                       free_extent_map(em);
-                       return ret;
+
+               /*
+                * we're not going to log anything, but we do need
+                * to make sure the current transaction stays open
+                * while we look for nocow cross refs
+                */
+               trans = btrfs_join_transaction(root, 0);
+               if (!trans)
+                       goto must_cow;
+
+               if (can_nocow_odirect(trans, inode, start, len) == 1) {
+                       ret = btrfs_add_ordered_extent_dio(inode, start,
+                                          block_start, len, len, type);
+                       btrfs_end_transaction(trans, root);
+                       if (ret) {
+                               free_extent_map(em);
+                               return ret;
+                       }
+                       goto unlock;
                }
-       } else {
-               free_extent_map(em);
-               em = btrfs_new_extent_direct(inode, start, len);
-               if (IS_ERR(em))
-                       return PTR_ERR(em);
-               len = min(len, em->block_len);
+               btrfs_end_transaction(trans, root);
        }
+must_cow:
+       /*
+        * this will cow the extent, reset the len in case we changed
+        * it above
+        */
+       len = bh_result->b_size;
+       free_extent_map(em);
+       em = btrfs_new_extent_direct(inode, start, len);
+       if (IS_ERR(em))
+               return PTR_ERR(em);
+       len = min(len, em->len - (start - em->start));
+unlock:
        clear_extent_bit(&BTRFS_I(inode)->io_tree, start, start + len - 1,
                          EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DIRTY, 1,
                          0, NULL, GFP_NOFS);
 map:
        bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
                inode->i_blkbits;
-       bh_result->b_size = em->len - (start - em->start);
+       bh_result->b_size = len;
        bh_result->b_bdev = em->bdev;
        set_buffer_mapped(bh_result);
        if (create && !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
@@ -5547,7 +5671,7 @@ static void btrfs_submit_direct(int rw, struct bio *bio, struct inode *inode,
                bvec++;
        } while (bvec <= (bio->bi_io_vec + bio->bi_vcnt - 1));
 
-       dip->disk_bytenr = bio->bi_sector << 9;
+       dip->disk_bytenr = (u64)bio->bi_sector << 9;
        bio->bi_private = dip;
 
        if (write)