On Thu, Feb 20, 2020 at 1:59 PM David Howells [off-list ref] wrote:

Jann Horn [off-list ref] wrote:

quoted

Ewww. So basically, having one static set of .fsinfo_attributes is not
sufficiently flexible for everyone, but instead of allowing the
filesystem to dynamically provide a list of supported attributes, you
just duplicate the super_operations? Seems to me like it'd be cleaner
to add a function pointer to the super_operations that can dynamically
fill out the supported fsinfo attributes.

It seems to me like the current API is going to be a dead end if you
ever want to have decent passthrough of these things for e.g. FUSE, or
overlayfs, or VirtFS?

Ummm...

Would it be sufficient to have a function that returns a list of attributes?
Or does it need to be able to call to vfs_do_fsinfo() if it supports an
attribute?

There are two things I want to be able to do:

 (1) Do the buffer wrangling in the core - which means the core needs to see
     the type of the attribute.  That's fine if, say, afs_fsinfo() can call
     vfs_do_fsinfo() with the definition for any attribute it wants to handle
     and, say, return -ENOPKG otherways so that the core can then fall back to
     its private list.

 (2) Be able to retrieve the list of attributes and/or query an attribute.
     Now, I can probably manage this even through the same interface.  If,
     say, seeing FSINFO_ATTR_FSINFO_ATTRIBUTES causes the handler to simply
     append on the IDs of its own supported attributes (a helper can be
     provided for that).

     If it sees FSINFO_ATR_FSINFO_ATTRIBUTE_INFO, it can just look to see if
     it has the attribute with the ID matching Nth and return that, else
     ENOPKG - again a helper could be provided.

Chaining through overlayfs gets tricky.  You end up with multiple contributory
filesystems with different properties - and any one of those layers could
perhaps be another overlay.  Overlayfs would probably needs to integrate the
info and derive the lowest common set.

Hm - I guess just returning a list of attributes ought to be fine?
Then AFS can just return one of its two statically-allocated attribute
lists there, and a filesystem with more complicated circumstances
(like FUSE or overlayfs or whatever) can compute a heap-allocated list
on mount that is freed when the superblock goes away, or something
like that?