RE: [RFC][PATCH 5/5] shmem: Add fsverity support

[Roberto Sassu]

> From: Eric Biggers [mailto:ebiggers@xxxxxxxxxx]
> Sent: Monday, November 15, 2021 8:05 PM
> On Mon, Nov 15, 2021 at 08:49:41AM +0000, Roberto Sassu wrote:
> > > From: Eric Biggers [mailto:ebiggers@xxxxxxxxxx]
> > > Sent: Friday, November 12, 2021 8:12 PM
> > > On Fri, Nov 12, 2021 at 01:44:11PM +0100, Roberto Sassu wrote:
> > > > Make the necessary modifications to support fsverity in tmpfs.
> > > >
> > > > First, implement the fsverity operations (in a similar way of f2fs). These
> > > > operations make use of shmem_read_mapping_page() instead of
> > > > read_mapping_page() to handle the case where the page has been
> swapped
> > > out.
> > > > The fsverity descriptor is placed at the end of the file and its location
> > > > is stored in an xattr.
> > > >
> > > > Second, implement the ioctl operations to enable, measure and read
> fsverity
> > > > metadata.
> > > >
> > > > Lastly, add calls to fsverity functions, to ensure that fsverity-relevant
> > > > operations are checked and handled by fsverity (file open, attr set, inode
> > > > evict).
> > > >
> > > > Fsverity support can be enabled through the kernel configuration and
> > > > remains enabled by default for every tmpfs filesystem instantiated (there
> > > > should be no overhead, unless fsverity is enabled for a file).
> > > >
> > > > Signed-off-by: Roberto Sassu <roberto.sassu@xxxxxxxxxx>
> > >
> > > I don't see how this makes sense at all.  The point of fs-verity is to avoid
> > > having to hash the whole file when verifying it.  However, obviously the
> whole
> > > file still has to be hashed to build the Merkle tree in the first place.  That
> > > makes sense for a persistent filesystem where a file can be written once and
> > > verified many times.  I don't see how it makes sense for tmpfs, where files
> have
> > > to be re-created on every boot.  You might as well just hash the whole file.
> >
> > The point of adding fsverity support for tmpfs was to being able to do
> > integrity enforcement with just one mechanism, given that I was
> > planning to do integrity verification with reference values loaded
> > to the kernel with DIGLIM [1].
> >
> > With an LSM such as IPE [2], integrity verification would consist in
> > querying the fsverity digest with DIGLIM and allowing the operation
> > if the digest was found. With fsverity support in tmpfs, this can be
> > done from the very beginning of the boot process.
> >
> > Using regular file digests would be also possible but this requires
> > loading with DIGLIM both fsverity and non-fsverity reference values.
> > It would also require two separate mechanisms for calculating
> > the file digest depending on the filesystem. It could be done, but
> > I thought it was easier to add support for fsverity in tmpfs.
> >
> > > Also, you didn't implement actually verifying the data (by calling
> > > fsverity_verify_page()), so this patch doesn't really do anything anyway.
> >
> > Yes, at the end I didn't add it. Probably the only place where
> > calling fsverity_verify_page() would make sense is when a page
> > is swapped in (assuming that the swap device is untrusted).
> >
> > I tried to add a call in shmem_swapin_page() but fsverity complained
> > due to the fact that the page was already up to date, and also
> > rejected the page. I will check it better.
> >
> 
> It sounds like you really only care about calculating fs-verity file digests.
> That's just an algorithm for hashing a file, so it could just be implemented in
> generic code that operates on any file on any filesystem, like how IMA
> implemennts full file hashing for any file.  There isn't a need for any special
> filesystem support to do this.

Initially I thought the same. Then, I realized that fsverity is much more
than that. Fsverity could be seen as a sort of property enforcer, it provides
a property associated to the file (the fsverity digest) and ensures that
the property remains the same while the system is running. In addition,
it takes advantage of the page cache to avoid remeasuring an up to date
page.

This remove some burden from LSMs. IPE would have just to compare
the fsverity digest with that in the policy (or just query it with DIGLIM).
Not taking into consideration the specific filesystem, not having to
fall back to the new fsverity measurement function, and avoiding to
preserve the fsverity property by itself, would make the LSM
implementation very simple.

Roberto

HUAWEI TECHNOLOGIES Duesseldorf GmbH, HRB 56063
Managing Director: Li Peng, Zhong Ronghua