[PATCH v2] btrfs: fix rw device counting in __btrfs_free_extra_devids

[David Sterba]

On Fri, Aug 13, 2021 at 01:31:25AM +0800, Desmond Cheong Zhi Xi wrote:
> On 12/8/21 11:50 pm, David Sterba wrote:
> > On Thu, Aug 12, 2021 at 11:43:16PM +0800, Desmond Cheong Zhi Xi wrote:
> >> On 12/8/21 6:38 pm, David Sterba wrote:
> >>> On Tue, Jul 27, 2021 at 03:13:03PM +0800, Desmond Cheong Zhi Xi wrote:
> >>>> --- a/fs/btrfs/volumes.c
> >>>> +++ b/fs/btrfs/volumes.c
> >>>> @@ -1078,6 +1078,7 @@ static void __btrfs_free_extra_devids(struct btrfs_fs_devices *fs_devices,
> >>>>    		if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) {
> >>>>    			list_del_init(&device->dev_alloc_list);
> >>>>    			clear_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
> >>>> +			fs_devices->rw_devices--;
> >>>>    		}
> >>>>    		list_del_init(&device->dev_list);
> >>>>    		fs_devices->num_devices--;
> >>>
> >>> I've hit a crash on master branch with stacktrace very similar to one
> >>> this bug was supposed to fix. It's a failed assertion on device close.
> >>> This patch was the last one to touch it and it matches some of the
> >>> keywords, namely the BTRFS_DEV_STATE_REPLACE_TGT bit that used to be in
> >>> the original patch but was not reinstated in your fix.
> >>>
> >>> I'm not sure how reproducible it is, right now I have only one instance
> >>> and am hunting another strange problem. They could be related.
> >>>
> >>> assertion failed: !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state), in fs/btrfs/volumes.c:1150
> >>>
> >>> https://susepaste.org/view/raw/18223056 full log with other stacktraces,
> >>> possibly relatedg
> >>>
> >>
> >> Looking at the logs, it seems that a dev_replace was started, then
> >> suspended. But it wasn't canceled or resumed before the fs devices were
> >> closed.
> >>
> >> I'll investigate further, just throwing some observations out there.
> > 
> > Thanks. I'm testing the patch revert, no crash after first loop, I'll
> > run a few more to be sure as it's not entirely reliable.
> > 
> > Sending the revert is option of last resort as we're approaching end of
> > 5.14 dev cycle and the crash prevents testing (unlike the fuzzer
> > warning).
> > 
> 
> I might be missing something, so any thoughts would be appreciated. But 
> I don't think the assertion in btrfs_close_one_device is correct.
> 
>  From what I see, this crash happens when close_ctree is called while a 
> dev_replace hasn't completed. In close_ctree, we suspend the 
> dev_replace, but keep the replace target around so that we can resume 
> the dev_replace procedure when we mount the root again. This is the call 
> trace:
> 
>    close_ctree():
>      btrfs_dev_replace_suspend_for_unmount();
>      btrfs_close_devices():
>        btrfs_close_fs_devices():
>          btrfs_close_one_device():
>            ASSERT(!test_bit(BTRFS_DEV_STATE_REPLACE_TGT, 
> &device->dev_state));
> 
> However, since the replace target sticks around, there is a device with 
> BTRFS_DEV_STATE_REPLACE_TGT set, and we fail the assertion in 
> btrfs_close_one_device.
> 
> Two options I can think of:
> 
> - We could remove the assertion.
> 
> - Or we could clear the BTRFS_DEV_STATE_REPLACE_TGT bit in 
> btrfs_dev_replace_suspend_for_unmount. This is fine since the bit is set 
> again in btrfs_init_dev_replace if the dev_replace->replace_state is 
> BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED. But this approach strikes me as 
> a little odd because the device is still the replace target when 
> mounting in the future.

The option #2 does not sound safe because the TGT bit is checked in
several places where device list is queried for various reasons, even
without a mounted filesystem.

Removing the assertion makes more sense but I'm still not convinced that
the this is expected/allowed state of a closed device.