Re: [PATCH] ext4: goto right label 'out_mmap_sem' in ext4_setattr()

[Baokun Li <libaokun1@huawei.com>]

Hi,

On 2025/2/13 20:51, Brian Foster wrote:
> On Thu, Feb 13, 2025 at 07:22:47PM +0800, libaokun@huaweicloud.com wrote:
>> From: Baokun Li <libaokun1@huawei.com>
>>
>> Otherwise, if ext4_inode_attach_jinode() fails, a hung task will
>> happen because filemap_invalidate_unlock() isn't called to unlock
>> mapping->invalidate_lock. Like this:
>>
>> EXT4-fs error (device sda) in ext4_setattr:5557: Out of memory
>> INFO: task fsstress:374 blocked for more than 122 seconds.
>>        Not tainted 6.14.0-rc1-next-20250206-xfstests-dirty #726
>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
>> task:fsstress state:D stack:0     pid:374   tgid:374   ppid:373
>>                                    task_flags:0x440140 flags:0x00000000
>> Call Trace:
>>   <TASK>
>>   __schedule+0x2c9/0x7f0
>>   schedule+0x27/0xa0
>>   schedule_preempt_disabled+0x15/0x30
>>   rwsem_down_read_slowpath+0x278/0x4c0
>>   down_read+0x59/0xb0
>>   page_cache_ra_unbounded+0x65/0x1b0
>>   filemap_get_pages+0x124/0x3e0
>>   filemap_read+0x114/0x3d0
>>   vfs_read+0x297/0x360
>>   ksys_read+0x6c/0xe0
>>   do_syscall_64+0x4b/0x110
>>   entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>
>> Fixes: c7fc0366c656 ("ext4: partial zero eof block on unaligned inode size extension")
>> Signed-off-by: Baokun Li <libaokun1@huawei.com>
>> ---
> First off, thank you for catching this. :)
Thanks for your review!
>
>>   fs/ext4/inode.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
>> index 3cc8da6357aa..04ffd802dbde 100644
>> --- a/fs/ext4/inode.c
>> +++ b/fs/ext4/inode.c
>> @@ -5452,7 +5452,7 @@ int ext4_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
>>   			    oldsize & (inode->i_sb->s_blocksize - 1)) {
>>   				error = ext4_inode_attach_jinode(inode);
>>   				if (error)
>> -					goto err_out;
>> +					goto out_mmap_sem;
>>   			}
> This looks reasonable to me, but I notice that the immediate previous
> error check looks like this:
>
> 		...
>                  rc = ext4_break_layouts(inode);
>                  if (rc) {
>                          filemap_invalidate_unlock(inode->i_mapping);
>                          goto err_out;
>                  }
> 		...
>
> ... and then the following after the broken logic uses out_mmap_sem.
> Could we be a little more consistent here one way or the other? The
> change looks functionally correct to me either way:
>
> Reviewed-by: Brian Foster <bfoster@redhat.com>
>
> Brian
Indeed, this is confusing.

The reason is that we don't want to call ext4_std_error() when
ext4_break_layouts() fails. So we first store the error in 'rc', and then
pass the error to 'error' at the end. (See b9c1c26739ec
("ext4: gracefully handle ext4_break_layouts() failure during truncate"))

However, because 'error' is not assigned, the goto out_mmap_sem label will
execute some code that shouldn't be executed. Therefore, in the error
handling of ext4_break_layouts(), we unlock and then goto err_out label.

While under normal error conditions, 'error' is assigned, and it should
enter the out_mmap_sem label. Therefore, in the error handling of
ext4_inode_attach_jinode(), we directly goto out_mmap_sem label.

The handling of 'rc' in this function is indeed very subtle.


Cheers,
Baokun
>>   
>>   			handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
>> -- 
>> 2.39.2
>>
>>