Re: [PATCH v6 2/3] btrfs: initial fsverity support

[Boris Burkov <boris@bur.io>]

On Wed, Sep 15, 2021 at 01:45:23PM -0700, Eric Biggers wrote:
> On Tue, Sep 14, 2021 at 11:34:29AM -0700, Boris Burkov wrote:
> > > Okay, so it is used.  (Due to the macro, it didn't show up when grepping.)
> > > 
> > > Doesn't it defeat the purpose of a ro_compat inode flag if the whole filesystem
> > > is marked with a ro_compat feature flag, though?  I thought that the point of
> > > the ro_compat inode flag is to allow old kernels to mount the filesystem
> > > read-write, with only verity files being forced to read-only.  That would be
> > > more flexible than ext4's implementation of fs-verity which forces the whole
> > > filesystem to read-only.  But it seems you're forcing the whole filesystem to
> > > read-only anyway?
> > > 
> > > - Eric
> > 
> > I was thinking of it in terms of "RO compat is the goal" and having new
> > inode flags totally broke that and was treated as a corruption of the
> > inode regardless of the fs being ro/rw. I think a check on a live fs
> > would just flip the fs ro, which was the goal anyway, but a check that
> > happened during mount would fail the mount, even for a read-only fs. 
> > 
> > Making it fully per file would be pretty cool! The only thing
> > really missing as far as I can tell is a way to mark a file read only
> > with the same semantics fsverity uses from within btrfs.
> 
> I don't understand.  Why are you bothering with the ro_compat inode flag at all
> if it doesn't actually work?
> 
> - Eric

Sorry I explained that really badly.

My first try was ro-compat bit only, that failed because btrfs couldn't
add an inode flag in a ro-compat way before my changes, as it could
fail to mount.

To fix that, I had to work on the inode flag compatibility, which
evolved into this notion of inode ro-compat flags, which does work as
expected: if you see a file with an unknown ro-compat flag it's an error
if you aren't read-only. Read-only mount will never fail.

I think changing the semantics of the ro-compat inodes from:
"an unknown ro inode flag -> fs ro" to
"an unknown ro inode flag -> file ro"
could be a big win. I don't think there is a rush to do that, though? If
I add it now, on top of the existing code, then you might go back to a
kernel that can only mount the fs read-only or you might go back to one
which is clever enough to only force the file read-only.

Hope I'm making a bit more sense, now.