ext4: Fix circular locking dependency with migrate and rm.
authorAneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Sun, 10 Feb 2008 06:20:05 +0000 (01:20 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Sun, 10 Feb 2008 06:20:05 +0000 (01:20 -0500)
In order to prevent a circular locking dependency when an unlink
operation is racing with an ext4 migration, we delay taking i_data_sem
until just before switch the inode format, and use i_mutex to prevent
writes and truncates during the first part of the migration operation.

Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/migrate.c

index 9ee1f7cfb2c541b9908fe8821a43e708efaea87d..8c6c685b9d22e1d399d34066e5e6a497150c81a9 100644 (file)
@@ -61,10 +61,9 @@ static int finish_range(handle_t *handle, struct inode *inode,
                retval = ext4_journal_restart(handle, needed);
                if (retval)
                        goto err_out;
-       }
-       if (needed) {
+       } else if (needed) {
                retval = ext4_journal_extend(handle, needed);
-               if (retval != 0) {
+               if (retval) {
                        /*
                         * IF not able to extend the journal restart the journal
                         */
@@ -220,6 +219,26 @@ static int update_tind_extent_range(handle_t *handle, struct inode *inode,
 
 }
 
+static int extend_credit_for_blkdel(handle_t *handle, struct inode *inode)
+{
+       int retval = 0, needed;
+
+       if (handle->h_buffer_credits > EXT4_RESERVE_TRANS_BLOCKS)
+               return 0;
+       /*
+        * We are freeing a blocks. During this we touch
+        * superblock, group descriptor and block bitmap.
+        * So allocate a credit of 3. We may update
+        * quota (user and group).
+        */
+       needed = 3 + 2*EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb);
+
+       if (ext4_journal_extend(handle, needed) != 0)
+               retval = ext4_journal_restart(handle, needed);
+
+       return retval;
+}
+
 static int free_dind_blocks(handle_t *handle,
                                struct inode *inode, __le32 i_data)
 {
@@ -234,11 +253,14 @@ static int free_dind_blocks(handle_t *handle,
 
        tmp_idata = (__le32 *)bh->b_data;
        for (i = 0; i < max_entries; i++) {
-               if (tmp_idata[i])
+               if (tmp_idata[i]) {
+                       extend_credit_for_blkdel(handle, inode);
                        ext4_free_blocks(handle, inode,
                                        le32_to_cpu(tmp_idata[i]), 1, 1);
+               }
        }
        put_bh(bh);
+       extend_credit_for_blkdel(handle, inode);
        ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
        return 0;
 }
@@ -267,29 +289,32 @@ static int free_tind_blocks(handle_t *handle,
                }
        }
        put_bh(bh);
+       extend_credit_for_blkdel(handle, inode);
        ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
        return 0;
 }
 
-static int free_ind_block(handle_t *handle, struct inode *inode)
+static int free_ind_block(handle_t *handle, struct inode *inode, __le32 *i_data)
 {
        int retval;
-       struct ext4_inode_info *ei = EXT4_I(inode);
 
-       if (ei->i_data[EXT4_IND_BLOCK])
+       /* ei->i_data[EXT4_IND_BLOCK] */
+       if (i_data[0]) {
+               extend_credit_for_blkdel(handle, inode);
                ext4_free_blocks(handle, inode,
-                               le32_to_cpu(ei->i_data[EXT4_IND_BLOCK]), 1, 1);
+                               le32_to_cpu(i_data[0]), 1, 1);
+       }
 
-       if (ei->i_data[EXT4_DIND_BLOCK]) {
-               retval = free_dind_blocks(handle, inode,
-                                               ei->i_data[EXT4_DIND_BLOCK]);
+       /* ei->i_data[EXT4_DIND_BLOCK] */
+       if (i_data[1]) {
+               retval = free_dind_blocks(handle, inode, i_data[1]);
                if (retval)
                        return retval;
        }
 
-       if (ei->i_data[EXT4_TIND_BLOCK]) {
-               retval = free_tind_blocks(handle, inode,
-                                               ei->i_data[EXT4_TIND_BLOCK]);
+       /* ei->i_data[EXT4_TIND_BLOCK] */
+       if (i_data[2]) {
+               retval = free_tind_blocks(handle, inode, i_data[2]);
                if (retval)
                        return retval;
        }
@@ -297,15 +322,13 @@ static int free_ind_block(handle_t *handle, struct inode *inode)
 }
 
 static int ext4_ext_swap_inode_data(handle_t *handle, struct inode *inode,
-                               struct inode *tmp_inode, int retval)
+                               struct inode *tmp_inode)
 {
+       int retval;
+       __le32  i_data[3];
        struct ext4_inode_info *ei = EXT4_I(inode);
        struct ext4_inode_info *tmp_ei = EXT4_I(tmp_inode);
 
-       retval = free_ind_block(handle, inode);
-       if (retval)
-               goto err_out;
-
        /*
         * One credit accounted for writing the
         * i_data field of the original inode
@@ -317,6 +340,11 @@ static int ext4_ext_swap_inode_data(handle_t *handle, struct inode *inode,
                        goto err_out;
        }
 
+       i_data[0] = ei->i_data[EXT4_IND_BLOCK];
+       i_data[1] = ei->i_data[EXT4_DIND_BLOCK];
+       i_data[2] = ei->i_data[EXT4_TIND_BLOCK];
+
+       down_write(&EXT4_I(inode)->i_data_sem);
        /*
         * We have the extent map build with the tmp inode.
         * Now copy the i_data across
@@ -336,8 +364,15 @@ static int ext4_ext_swap_inode_data(handle_t *handle, struct inode *inode,
        spin_lock(&inode->i_lock);
        inode->i_blocks += tmp_inode->i_blocks;
        spin_unlock(&inode->i_lock);
+       up_write(&EXT4_I(inode)->i_data_sem);
 
+       /*
+        * We mark the inode dirty after, because we decrement the
+        * i_blocks when freeing the indirect meta-data blocks
+        */
+       retval = free_ind_block(handle, inode, i_data);
        ext4_mark_inode_dirty(handle, inode);
+
 err_out:
        return retval;
 }
@@ -365,6 +400,7 @@ static int free_ext_idx(handle_t *handle, struct inode *inode,
                }
        }
        put_bh(bh);
+       extend_credit_for_blkdel(handle, inode);
        ext4_free_blocks(handle, inode, block, 1, 1);
        return retval;
 }
@@ -420,7 +456,6 @@ int ext4_ext_migrate(struct inode *inode, struct file *filp,
                 */
                return retval;
 
-       down_write(&EXT4_I(inode)->i_data_sem);
        handle = ext4_journal_start(inode,
                                        EXT4_DATA_TRANS_BLOCKS(inode->i_sb) +
                                        EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
@@ -454,13 +489,6 @@ int ext4_ext_migrate(struct inode *inode, struct file *filp,
        ext4_orphan_add(handle, tmp_inode);
        ext4_journal_stop(handle);
 
-       ei = EXT4_I(inode);
-       i_data = ei->i_data;
-       memset(&lb, 0, sizeof(lb));
-
-       /* 32 bit block address 4 bytes */
-       max_entries = inode->i_sb->s_blocksize >> 2;
-
        /*
         * start with one credit accounted for
         * superblock modification.
@@ -469,7 +497,20 @@ int ext4_ext_migrate(struct inode *inode, struct file *filp,
         * trascation that created the inode. Later as and
         * when we add extents we extent the journal
         */
+       /*
+        * inode_mutex prevent write and truncate on the file. Read still goes
+        * through. We take i_data_sem in ext4_ext_swap_inode_data before we
+        * switch the inode format to prevent read.
+        */
+       mutex_lock(&(inode->i_mutex));
        handle = ext4_journal_start(inode, 1);
+
+       ei = EXT4_I(inode);
+       i_data = ei->i_data;
+       memset(&lb, 0, sizeof(lb));
+
+       /* 32 bit block address 4 bytes */
+       max_entries = inode->i_sb->s_blocksize >> 2;
        for (i = 0; i < EXT4_NDIR_BLOCKS; i++, blk_count++) {
                if (i_data[i]) {
                        retval = update_extent_range(handle, tmp_inode,
@@ -507,19 +548,6 @@ int ext4_ext_migrate(struct inode *inode, struct file *filp,
         */
        retval = finish_range(handle, tmp_inode, &lb);
 err_out:
-       /*
-        * We are either freeing extent information or indirect
-        * blocks. During this we touch superblock, group descriptor
-        * and block bitmap. Later we mark the tmp_inode dirty
-        * via ext4_ext_tree_init. So allocate a credit of 4
-        * We may update quota (user and group).
-        *
-        * FIXME!! we may be touching bitmaps in different block groups.
-        */
-       if (ext4_journal_extend(handle,
-                       4 + 2*EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb)) != 0)
-               ext4_journal_restart(handle,
-                               4 + 2*EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb));
        if (retval)
                /*
                 * Failure case delete the extent information with the
@@ -528,7 +556,11 @@ err_out:
                free_ext_block(handle, tmp_inode);
        else
                retval = ext4_ext_swap_inode_data(handle, inode,
-                                                       tmp_inode, retval);
+                                                       tmp_inode);
+
+       /* We mark the tmp_inode dirty via ext4_ext_tree_init. */
+       if (ext4_journal_extend(handle, 1) != 0)
+               ext4_journal_restart(handle, 1);
 
        /*
         * Mark the tmp_inode as of size zero
@@ -556,8 +588,7 @@ err_out:
        tmp_inode->i_nlink = 0;
 
        ext4_journal_stop(handle);
-
-       up_write(&EXT4_I(inode)->i_data_sem);
+       mutex_unlock(&(inode->i_mutex));
 
        if (tmp_inode)
                iput(tmp_inode);