Re: [RFC][PATCH 0/76] vfs: 'views' for filesystems with more than one root

[Amir Goldstein]

On Wed, Jun 6, 2018 at 11:42 PM, Mark Fasheh <mfasheh@xxxxxxx> wrote:
> Hi Amir, thanks for the comments!
>

Hi Mark,

[...]
>
> Btw, sorry that the name is confusing. I've never been good at picking them.

Didn't say that it was confusing. It might very well be the perfect name...
if I only knew what sort of thing we are trying to name...

> That said, if you have a minute to check out the first patch or two you'll
> see that the patches are basically putting a struct in between the super
> block and the inode.
>
>
> One thing I'd like to politely suggest is that anyone now following this
> conversation to please read the at least the first patch. It's an easy read
> and I feel like the conversation overall would be much more clear if
> everyone understood what we're going for. I worry that I didn't do a
> particularly good job explaining the actual code changes.
>
> https://www.spinics.net/lists/linux-fsdevel/msg125492.html
>


I did look at the patches. They look simple and clean and they solve A problem.
All I'm saying is that we should look at the set of problems that we know of
before we design an abstraction layer.

>
> Regarding a layout of the problems, I have a more complete e-mail coming
> soon which should describe in detail the issues I've seen with respect to
> how the kernel is exporting ino/dev pairs (outside of statx). fs_view alone
> is not enough to solve that problem. I'd be happy to CC you on that one if
> you'd like.
>

Sure.

[...]

>>
>> And what is the SUSE way?
>
> At SUSE, we carry a version of this patch:
>
> https://marc.info/?l=linux-btrfs&m=130532890824992&w=2
>
> Essentially a callback which was rejected for various reasons.
>

Don't see a patch here. Wrong link?

> The fs_view work was intended to replace that patch with an upstream
> solution.
>
>

[...]

>>
>> FYI, the Overlayfs file/inode mapping is about to change with many
>> VFS hacks queued for removal, so stay tuned.
>>
>> [...]
>
> Actually, would you mind giving me a pointer to this work? I'd be very
> interested to see what exactly might be changing.
>

Mostly, less VFS code is going to be exposed to underlying inode:

https://marc.info/?l=linux-fsdevel&m=152760014530531&w=2

[...]

>> I have an interest of solving another problem.
>> In VFS operations where only inode is available, I would like to be able to
>> report fsnotify events (e.g. fsnotify_move()) only in directories under a
>> certain subtree root. That could be achieved either by bind mount the subtree
>> root and passing vfsmount into vfs_rename() or by defining an fs_view on the
>> subtree and mounting that fs_view.
>
> I'm not sure whether fs_view works for this. Taking a quick look at
> fsnotify, the state is already on the inode? If there's a globabl fsnotify
> state that needs to be per subtree than yes we could move that to the
> fs_view and you'd simply deref it from the inode struct.
>

That was my thinking. I have patches to attach an fsnotify mark
to super block. If fs_view could have a root that is different than
super block's root and if file system can guaranty that objects
cannot be moved outside of fs_view root, then fsnotify mark
could be attached to fs_view.

Thanks,
Amir.