Re: [RFC PATCH 0/1] ovl: brtfs' temp_fsid doesn't work with ovl index=on

[Qu Wenruo <wqu@suse.com>]

在 2025/10/15 10:35, Anand Jain 写道:
> On 15-Oct-25 5:08 AM, Qu Wenruo wrote:
>>
>>
>> 在 2025/10/15 04:54, David Sterba 写道:
>>> On Mon, Oct 13, 2025 at 10:57:06PM -0300, André Almeida wrote:
>>>> Hi everyone,
>>>>
>>>> When using overlayfs with the mount option index=on, the first time 
>>>> a directory is
>>>> used as upper dir, overlayfs stores in a xattr "overlay.origin" the 
>>>> UUID of the
>>>> filesystem being used in the layers. If the upper dir is reused, 
>>>> overlayfs
>>>> refuses to mount for a different filesystem, by comparing the UUID 
>>>> with what's
>>>> stored at overlay.origin, and it fails with "failed to verify upper 
>>>> root origin"
>>>> on dmesg. Remounting with the very same fs is supported and works fine.
>>>>
>>>> However, btrfs mounts may have volatiles UUIDs. When mounting the 
>>>> exact same
>>>> disk image with btrfs, a random UUID is assigned for the following 
>>>> disks each
>>>> time they are mounted, stored at temp_fsid and used across the 
>>>> kernel as the
>>>> disk UUID. `btrfs filesystem show` presents that. Calling statfs() 
>>>> however shows
>>>> the original (and duplicated) UUID for all disks.
>>>>
>>>> This feature doesn't work well with overlayfs with index=on, as when 
>>>> the image
>>>> is mounted a second time, will get a different UUID and ovl will 
>>>> refuse to
>>>> mount, breaking the user expectation that using the same image 
>>>> should work. A
>>>> small script can be find in the end of this cover letter that 
>>>> illustrates this.
>>>>
>>>> >From this, I can think of some options:
>>>>
>>>> - Use statfs() internally to always get the fsid, that is 
>>>> persistent. The patch
>>>> here illustrates that approach, but doesn't fully implement it.
>>>> - Create a new sb op, called get_uuid() so the filesystem returns 
>>>> what's
>>>> appropriated.
>>>> - Have a workaround in ovl for btrfs.
>>>> - Document this as unsupported, and userland needs to erase 
>>>> overlay.origin each
>>>> time it wants to remount.
>>>> - If ovl detects that temp_fsid and index are being used at the same 
>>>> time,
>>>> refuses to mount.
>>>>
>>>> I'm not sure which one would be better here, so I would like to hear 
>>>> some ideas
>>>> on this.
>>>
>>> I haven't looked deeper if there's a workable solution, but the feature
>>> combination should be refused. I don't think this will affect many
>>> users.
>>>
>>
>> I believe the root problem is that we're not fully implementing the 
>> proper handling just like other single-device fses.
>>
>> We do not use on-disk flags which means at least one fsid is 
>> registered into btrfs, thus we have to use different temp-fsid.
>>
>> If fully single-device feature flag is properly implemented, we should 
>> be able to return the same uuid without extra hacks thus solve the 
>> problem.
> 
> I had looked into this some time ago. Some libs, like libblkid,
> don't handle multi-device filesystems or cloned devices with
> temp FSIDs very well. I'm aware of it.
> 
> I've been making some progress on fixing those cases, but it's
> a bit extensive since we first need enough test coverage,
> and recent reappear-device inline with that.
> 
> Let's see how we can support use cases with identical devices
> (where changing the UUID isn't an option) and keep things
> compatible with systemd and library tools.
> 

My current idea is to introduce a new ro compat flag, so that mounting 
that device will not go through the fsid lookup procedure completely.

But go through the common get_tree_bdev() routine, which will check if 
the fs is already mounted using bdev holder.
(And of course, no fsid recorded inside btrfs module)

This will make single device btrfs with that special flag to behave 
exactly like all the other filesystems.