Re: [PATCH v2] btrfs: fix BUG trying to resume balance without resume flag

[Anand Jain <anand.jain@oracle.com>]

On 05/16/2018 10:35 PM, David Sterba wrote:
> On Mon, Apr 30, 2018 at 05:48:45PM +0800, Anand Jain wrote:
>> We set the BTRFS_BALANCE_RESUME flag in the btrfs_recover_balance(),
>> which is not called during the remount. So when resuming from the
>> paused balance we hit BUG.
>>
>>   kernel: kernel BUG at fs/btrfs/volumes.c:3890!
>>   ::
>>   kernel:  balance_kthread+0x51/0x60 [btrfs]
>>   kernel:  kthread+0x111/0x130
>>   ::
>>   kernel: RIP: btrfs_balance+0x12e1/0x1570 [btrfs] RSP: ffffba7d0090bde8
>>
>> Reproducer:
>>    On a mounted BTRFS.
>>
>>    btrfs balance start --full-balance /btrfs
>>    btrfs balance pause /btrfs
>>    mount -o remount,ro /dev/sdb /btrfs
>>    mount -o remount,rw /dev/sdb /btrfs
>>
>> To fix this set the BTRFS_BALANCE_RESUME flag in btrfs_resume_balance_async()
>> instead of btrfs_recover_balance().
>>
>> Signed-off-by: Anand Jain <anand.jain@oracle.com>
>> ---
>> v1->v2: btrfs_resume_balance_async() can be called only from remount or
>> mount, we don't need to hold fs_info->balance_lock.
> 
> For mount it's ok, there's nothing that can run in parallel, but remount
> ro->rw that resumes the balance can be run with the ioctl that checks
> balance status in parallel, at any time. As the fs_info::balance_ctl is
> set up, btrfs_ioctl_balance_progress will proceed and return the current
> status.

  You are right.

>> Strictly speaking we should rather keep the balance at the paused state
>> unless it is resumed by the user again, that means neither mount nor
>> remount-rw should resume the balance automatically, former case needs
>> writing balance status to the disk. Which needs compatibility
>> verification.  So for now just avoid BUG.
>>
>>   fs/btrfs/volumes.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
>> index 3e6983a169c4..64bcaf25908b 100644
>> --- a/fs/btrfs/volumes.c
>> +++ b/fs/btrfs/volumes.c
>> @@ -4115,6 +4115,8 @@ int btrfs_resume_balance_async(struct btrfs_fs_info *fs_info)
>>   		return 0;
>>   	}
>>   
>> +	fs_info->balance_ctl->b_flags |= BTRFS_BALANCE_RESUME;
> 
> Though there's no update operation on flags, I think it's still better
> to add the locks that set the resume status. And a comment why it's
> here.

ok.

>> +
>>   	tsk = kthread_run(balance_kthread, fs_info, "btrfs-balance");
>>   	return PTR_ERR_OR_ZERO(tsk);
>>   }
>> @@ -4156,7 +4158,6 @@ int btrfs_recover_balance(struct btrfs_fs_info *fs_info)
>>   
>>   	bctl->fs_info = fs_info;
>>   	bctl->flags = btrfs_balance_flags(leaf, item);
>> -	bctl->flags |= BTRFS_BALANCE_RESUME;
> 
> Also, it does not hurt to leave this here, as it logically matches what
> the function does: we found the balance item, thus we set the status
> accordingly.

ok.

> Please update and resend, I'd like to push this to 4.17-rc as it's a
> crash fix with a reproducer. Thanks.

Sent v3 with the above changes.

Thanks, Anand
> --
> To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>