lkml.org 
[lkml]   [2021]   [Aug]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2] btrfs: fix rw device counting in __btrfs_free_extra_devids
From
Date
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:
>>>> When removing a writeable device in __btrfs_free_extra_devids, the rw
>>>> device count should be decremented.
>>>>
>>>> This error was caught by Syzbot which reported a warning in
>>>> close_fs_devices because fs_devices->rw_devices was not 0 after
>>>> closing all devices. Here is the call trace that was observed:
>>>>
>>>> btrfs_mount_root():
>>>> btrfs_scan_one_device():
>>>> device_list_add(); <---------------- device added
>>>> btrfs_open_devices():
>>>> open_fs_devices():
>>>> btrfs_open_one_device(); <-------- writable device opened,
>>>> rw device count ++
>>>> btrfs_fill_super():
>>>> open_ctree():
>>>> btrfs_free_extra_devids():
>>>> __btrfs_free_extra_devids(); <--- writable device removed,
>>>> rw device count not decremented
>>>> fail_tree_roots:
>>>> btrfs_close_devices():
>>>> close_fs_devices(); <------- rw device count off by 1
>>>>
>>>> As a note, prior to commit cf89af146b7e ("btrfs: dev-replace: fail
>>>> mount if we don't have replace item with target device"), rw_devices
>>>> was decremented on removing a writable device in
>>>> __btrfs_free_extra_devids only if the BTRFS_DEV_STATE_REPLACE_TGT bit
>>>> was not set for the device. However, this check does not need to be
>>>> reinstated as it is now redundant and incorrect.
>>>>
>>>> In __btrfs_free_extra_devids, we skip removing the device if it is the
>>>> target for replacement. This is done by checking whether device->devid
>>>> == BTRFS_DEV_REPLACE_DEVID. Since BTRFS_DEV_STATE_REPLACE_TGT is set
>>>> only on the device with devid BTRFS_DEV_REPLACE_DEVID, no devices
>>>> should have the BTRFS_DEV_STATE_REPLACE_TGT bit set after the check,
>>>> and so it's redundant to test for that bit.
>>>>
>>>> Additionally, following commit 82372bc816d7 ("Btrfs: make
>>>> the logic of source device removing more clear"), rw_devices is
>>>> incremented whenever a writeable device is added to the alloc
>>>> list (including the target device in btrfs_dev_replace_finishing), so
>>>> all removals of writable devices from the alloc list should also be
>>>> accompanied by a decrement to rw_devices.
>>>>
>>>> Fixes: cf89af146b7e ("btrfs: dev-replace: fail mount if we don't have replace item with target device")
>>>> Reported-by: syzbot+a70e2ad0879f160b9217@syzkaller.appspotmail.com
>>>> Tested-by: syzbot+a70e2ad0879f160b9217@syzkaller.appspotmail.com
>>>> Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
>>>> Reviewed-by: Anand Jain <anand.jain@oracle.com>
>>>> ---
>>>> fs/btrfs/volumes.c | 1 +
>>>> 1 file changed, 1 insertion(+)
>>>>
>>>> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
>>>> index 807502cd6510..916c25371658 100644
>>>> --- 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.

Thoughts?

\
 
 \ /
  Last update: 2021-08-12 19:32    [W:0.057 / U:0.548 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site