[PATCH] udf: Avoid double brelse() in udf_rename()

[PATCH] udf: Avoid double brelse() in udf_rename()

[Shigeru Yoshida]

syzbot reported a warning like below [1]:

VFS: brelse: Trying to free free buffer
WARNING: CPU: 2 PID: 7301 at fs/buffer.c:1145 __brelse+0x67/0xa0
...
Call Trace:
 <TASK>
 invalidate_bh_lru+0x99/0x150
 smp_call_function_many_cond+0xe2a/0x10c0
 ? generic_remap_file_range_prep+0x50/0x50
 ? __brelse+0xa0/0xa0
 ? __mutex_lock+0x21c/0x12d0
 ? smp_call_on_cpu+0x250/0x250
 ? rcu_read_lock_sched_held+0xb/0x60
 ? lock_release+0x587/0x810
 ? __brelse+0xa0/0xa0
 ? generic_remap_file_range_prep+0x50/0x50
 on_each_cpu_cond_mask+0x3c/0x80
 blkdev_flush_mapping+0x13a/0x2f0
 blkdev_put_whole+0xd3/0xf0
 blkdev_put+0x222/0x760
 deactivate_locked_super+0x96/0x160
 deactivate_super+0xda/0x100
 cleanup_mnt+0x222/0x3d0
 task_work_run+0x149/0x240
 ? task_work_cancel+0x30/0x30
 do_exit+0xb29/0x2a40
 ? reacquire_held_locks+0x4a0/0x4a0
 ? do_raw_spin_lock+0x12a/0x2b0
 ? mm_update_next_owner+0x7c0/0x7c0
 ? rwlock_bug.part.0+0x90/0x90
 ? zap_other_threads+0x234/0x2d0
 do_group_exit+0xd0/0x2a0
 __x64_sys_exit_group+0x3a/0x50
 do_syscall_64+0x34/0xb0
 entry_SYSCALL_64_after_hwframe+0x63/0xcd

The cause of the issue is that brelse() is called on both ofibh.sbh
and ofibh.ebh by udf_find_entry() when it returns NULL.  However,
brelse() is called by udf_rename(), too.  So, b_count on buffer_head
becomes unbalanced.

This patch fixes the issue by not calling brelse() by udf_rename()
when udf_find_entry() returns NULL.

Link: https://syzkaller.appspot.com/bug?id=8297f45698159c6bca8a1f87dc983667c1a1c851 [1]
Reported-by: syzbot+7902cd7684bc35306224@syzkaller.appspotmail.com
Signed-off-by: Shigeru Yoshida <syoshida@redhat.com>
---
 fs/udf/namei.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/fs/udf/namei.c b/fs/udf/namei.c
index fb4c30e05245..d6081538bfc0 100644
--- a/fs/udf/namei.c
+++ b/fs/udf/namei.c
@@ -1091,8 +1091,9 @@ static int udf_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
 		return -EINVAL;
 
 	ofi = udf_find_entry(old_dir, &old_dentry->d_name, &ofibh, &ocfi);
-	if (IS_ERR(ofi)) {
-		retval = PTR_ERR(ofi);
+	if (!ofi || IS_ERR(ofi)) {
+		if (IS_ERR(ofi))
+			retval = PTR_ERR(ofi);
 		goto end_rename;
 	}
 
@@ -1101,8 +1102,7 @@ static int udf_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
 
 	brelse(ofibh.sbh);
 	tloc = lelb_to_cpu(ocfi.icb.extLocation);
-	if (!ofi || udf_get_lb_pblock(old_dir->i_sb, &tloc, 0)
-	    != old_inode->i_ino)
+	if (udf_get_lb_pblock(old_dir->i_sb, &tloc, 0) != old_inode->i_ino)
 		goto end_rename;
 
 	nfi = udf_find_entry(new_dir, &new_dentry->d_name, &nfibh, &ncfi);
-- 
2.37.3

Re: [PATCH] udf: Avoid double brelse() in udf_rename()

[Jan Kara]

On Sun 23-10-22 18:57:41, Shigeru Yoshida wrote:
> syzbot reported a warning like below [1]:
> 
> VFS: brelse: Trying to free free buffer
> WARNING: CPU: 2 PID: 7301 at fs/buffer.c:1145 __brelse+0x67/0xa0
> ...
> Call Trace:
>  <TASK>
>  invalidate_bh_lru+0x99/0x150
>  smp_call_function_many_cond+0xe2a/0x10c0
>  ? generic_remap_file_range_prep+0x50/0x50
>  ? __brelse+0xa0/0xa0
>  ? __mutex_lock+0x21c/0x12d0
>  ? smp_call_on_cpu+0x250/0x250
>  ? rcu_read_lock_sched_held+0xb/0x60
>  ? lock_release+0x587/0x810
>  ? __brelse+0xa0/0xa0
>  ? generic_remap_file_range_prep+0x50/0x50
>  on_each_cpu_cond_mask+0x3c/0x80
>  blkdev_flush_mapping+0x13a/0x2f0
>  blkdev_put_whole+0xd3/0xf0
>  blkdev_put+0x222/0x760
>  deactivate_locked_super+0x96/0x160
>  deactivate_super+0xda/0x100
>  cleanup_mnt+0x222/0x3d0
>  task_work_run+0x149/0x240
>  ? task_work_cancel+0x30/0x30
>  do_exit+0xb29/0x2a40
>  ? reacquire_held_locks+0x4a0/0x4a0
>  ? do_raw_spin_lock+0x12a/0x2b0
>  ? mm_update_next_owner+0x7c0/0x7c0
>  ? rwlock_bug.part.0+0x90/0x90
>  ? zap_other_threads+0x234/0x2d0
>  do_group_exit+0xd0/0x2a0
>  __x64_sys_exit_group+0x3a/0x50
>  do_syscall_64+0x34/0xb0
>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 
> The cause of the issue is that brelse() is called on both ofibh.sbh
> and ofibh.ebh by udf_find_entry() when it returns NULL.  However,
> brelse() is called by udf_rename(), too.  So, b_count on buffer_head
> becomes unbalanced.
> 
> This patch fixes the issue by not calling brelse() by udf_rename()
> when udf_find_entry() returns NULL.
> 
> Link: https://syzkaller.appspot.com/bug?id=8297f45698159c6bca8a1f87dc983667c1a1c851 [1]
> Reported-by: syzbot+7902cd7684bc35306224@syzkaller.appspotmail.com
> Signed-off-by: Shigeru Yoshida <syoshida@redhat.com>

Thanks! I've added the fix to my tree.

								Honza

> ---
>  fs/udf/namei.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/udf/namei.c b/fs/udf/namei.c
> index fb4c30e05245..d6081538bfc0 100644
> --- a/fs/udf/namei.c
> +++ b/fs/udf/namei.c
> @@ -1091,8 +1091,9 @@ static int udf_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
>  		return -EINVAL;
>  
>  	ofi = udf_find_entry(old_dir, &old_dentry->d_name, &ofibh, &ocfi);
> -	if (IS_ERR(ofi)) {
> -		retval = PTR_ERR(ofi);
> +	if (!ofi || IS_ERR(ofi)) {
> +		if (IS_ERR(ofi))
> +			retval = PTR_ERR(ofi);
>  		goto end_rename;
>  	}
>  
> @@ -1101,8 +1102,7 @@ static int udf_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
>  
>  	brelse(ofibh.sbh);
>  	tloc = lelb_to_cpu(ocfi.icb.extLocation);
> -	if (!ofi || udf_get_lb_pblock(old_dir->i_sb, &tloc, 0)
> -	    != old_inode->i_ino)
> +	if (udf_get_lb_pblock(old_dir->i_sb, &tloc, 0) != old_inode->i_ino)
>  		goto end_rename;
>  
>  	nfi = udf_find_entry(new_dir, &new_dentry->d_name, &nfibh, &ncfi);
> -- 
> 2.37.3
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR