Re: [PATCH] btrfs: fix BUG_ON condition in btrfs_cancel_balance

[David Sterba]

On Tue, Aug 15, 2023 at 02:55:59AM -0400, xiaoshoukui wrote:
> Pausing and canceling balance can race to intterupt balance lead to BUG_ON 
> panic in btrfs_cancel_balance. The BUG_ON condition in btrfs_cancel_balance
> does not take this race scenario into account.

Seems that it's from times the balance was not cancellable the same way
as now. Also it's a good time to switch the BUG_ON to an assertion or
handle it properly.
> 
> However, the race condition has no other side effects. We can fix that.
> 
> Reproducing it with panic trace like this:
> kernel BUG at fs/btrfs/volumes.c:4618!
> RIP: 0010:btrfs_cancel_balance+0x5cf/0x6a0
> Call Trace:
>  <TASK>
>  ? do_nanosleep+0x60/0x120
>  ? hrtimer_nanosleep+0xb7/0x1a0
>  ? sched_core_clone_cookie+0x70/0x70
>  btrfs_ioctl_balance_ctl+0x55/0x70
>  btrfs_ioctl+0xa46/0xd20
>  __x64_sys_ioctl+0x7d/0xa0
>  do_syscall_64+0x38/0x80
>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 
> Race scenario as follows:
> > mutex_unlock(&fs_info->balance_mutex);
> > --------------------
> > .......issue pause and cancel req in another thread
> > --------------------
> > ret = __btrfs_balance(fs_info);
> > 
> > mutex_lock(&fs_info->balance_mutex);
> > if (ret == -ECANCELED && atomic_read(&fs_info->balance_pause_req)) {
> >         btrfs_info(fs_info, "balance: paused");
> >         btrfs_exclop_balance(fs_info, BTRFS_EXCLOP_BALANCE_PAUSED);
> > }
> 
> Signed-off-by: xiaoshoukui <xiaoshoukui@xxxxxxxxxxxxx>
> ---
>  fs/btrfs/volumes.c | 3 +--
>  1 file changed, 1 insertion(+), 2 deletions(-)
> 
> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> index 2ecb76cf3d91..886d667419ed 100644
> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -4638,8 +4638,7 @@ int btrfs_cancel_balance(struct btrfs_fs_info *fs_info)
>  		}
>  	}
>  
> -	BUG_ON(fs_info->balance_ctl ||
> -		test_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags));
> +	BUG_ON(test_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags));

I'll change to to ASSERT, this is really to verify that the state
tracking works properly.

>  	atomic_dec(&fs_info->balance_cancel_req);
>  	mutex_unlock(&fs_info->balance_mutex);
>  	return 0;
> -- 
> 2.34.1