OSDN Git Service

ext4: fix deadlock when writing in ENOSPC conditions
authorJan Kara <jack@suse.cz>
Wed, 18 Dec 2013 05:44:44 +0000 (00:44 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 9 Jan 2014 20:24:21 +0000 (12:24 -0800)
commit 34cf865d54813aab3497838132fb1bbd293f4054 upstream.

Akira-san has been reporting rare deadlocks of his machine when running
xfstests test 269 on ext4 filesystem. The problem turned out to be in
ext4_da_reserve_metadata() and ext4_da_reserve_space() which called
ext4_should_retry_alloc() while holding i_data_sem. Since
ext4_should_retry_alloc() can force a transaction commit, this is a
lock ordering violation and leads to deadlocks.

Fix the problem by just removing the retry loops. These functions should
just report ENOSPC to the caller (e.g. ext4_da_write_begin()) and that
function must take care of retrying after dropping all necessary locks.

Reported-and-tested-by: Akira Fujita <a-fujita@rs.jp.nec.com>
Reviewed-by: Zheng Liu <wenqing.lz@taobao.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/ext4/inode.c

index 904ca1a..cb2bdc7 100644 (file)
@@ -1263,7 +1263,6 @@ static int ext4_journalled_write_end(struct file *file,
  */
 static int ext4_da_reserve_metadata(struct inode *inode, ext4_lblk_t lblock)
 {
-       int retries = 0;
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
        struct ext4_inode_info *ei = EXT4_I(inode);
        unsigned int md_needed;
@@ -1275,7 +1274,6 @@ static int ext4_da_reserve_metadata(struct inode *inode, ext4_lblk_t lblock)
         * in order to allocate nrblocks
         * worse case is one extent per block
         */
-repeat:
        spin_lock(&ei->i_block_reservation_lock);
        /*
         * ext4_calc_metadata_amount() has side effects, which we have
@@ -1295,10 +1293,6 @@ repeat:
                ei->i_da_metadata_calc_len = save_len;
                ei->i_da_metadata_calc_last_lblock = save_last_lblock;
                spin_unlock(&ei->i_block_reservation_lock);
-               if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
-                       cond_resched();
-                       goto repeat;
-               }
                return -ENOSPC;
        }
        ei->i_reserved_meta_blocks += md_needed;
@@ -1312,7 +1306,6 @@ repeat:
  */
 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
 {
-       int retries = 0;
        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
        struct ext4_inode_info *ei = EXT4_I(inode);
        unsigned int md_needed;
@@ -1334,7 +1327,6 @@ static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
         * in order to allocate nrblocks
         * worse case is one extent per block
         */
-repeat:
        spin_lock(&ei->i_block_reservation_lock);
        /*
         * ext4_calc_metadata_amount() has side effects, which we have
@@ -1354,10 +1346,6 @@ repeat:
                ei->i_da_metadata_calc_len = save_len;
                ei->i_da_metadata_calc_last_lblock = save_last_lblock;
                spin_unlock(&ei->i_block_reservation_lock);
-               if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
-                       cond_resched();
-                       goto repeat;
-               }
                dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
                return -ENOSPC;
        }