Re: [PATCH 18/26] xfs: use merkle tree offset as attr hash

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, May 22, 2024 at 11:29:00AM -0700, Eric Biggers wrote:
> On Wed, May 22, 2024 at 07:37:23AM -0700, Christoph Hellwig wrote:
> > On Mon, May 20, 2024 at 09:02:59AM -0700, Darrick J. Wong wrote:
> > > On Mon, May 20, 2024 at 05:39:59AM -0700, Christoph Hellwig wrote:
> > > > On Fri, May 17, 2024 at 10:17:20AM -0700, Darrick J. Wong wrote:
> > > > > >   Note that the verity metadata *must* be encrypted when the file is,
> > > > > >   since it contains hashes of the plaintext data.
> > > > > 
> > > > > Refresh my memory of fscrypt -- does it encrypt directory names, xattr
> > > > > names, and xattr values too?  Or does it only do that to file data?
> > > > 
> > > > It does encrypt the file names in the directories, but nothing in
> > > > xattrs as far as I can tell.
> > > 
> > > Do we want that for user.* attrs?  That seems like quite an omission.
> > 
> > I'll let Eric answer that.  Btw, is the threat model for fscrypt written
> > down somewhere?
> 
> See https://www.kernel.org/doc/html/latest/filesystems/fscrypt.html?highlight=fscrypt#threat-model
> 
> As for why it stopped at file contents and names (and fsverity Merkle tree
> blocks which ext4 and f2fs encrypt in the same way as contents), it's just
> because it's very difficult to add more, and file contents and names alone were
> enough for parity with most other file-level encryption systems.  That's just
> the nature of file-level encryption.  For each additional type of data or
> metadata that's encrypted, there are a huge number of things that need to be
> resolved including algorithm selection, key derivation, IV selection, on-disk
> format, padding, UAPI for enabling the feature, userspace tool support including
> fsck and debugging tools, access semantics without the key, etc...
> 
> xattr encryption is definitely something that people have thought about, and it
> probably would be the next thing to consider after the existing contents and
> names.  Encrypting the exact file sizes is also something to consider.  But it's
> not something that someone has volunteered to do all the work for (yet).  If you
> restricted it to the contents of user xattrs only (not other xattrs, and not
> xattr names), it would be more feasible than trying to encrypt the names and
> values of all xattrs, though it would still be difficult.
> 
> Of course, generally speaking, when adding fscrypt support to a filesystem, it's
> going to be much easier to just target the existing feature set, and not try to
> include new, unproven features too.  (FWIW, this also applies to fsverity.)  If
> someone is interested in taking on an experimental project add xattr encryption
> support, I'd be glad to try to provide guidance, but it probably should be
> separated out from adding fscrypt support in the first place.

Does the fscrypt data unit size have to match i_blocksize?  Or the
fsverity merkle tree block size?

> > > > > And if we copy the ext4 method of putting the merkle data after eof and
> > > > > loading it into the pagecache, how much of the generic fs/verity cleanup
> > > > > patches do we really need?
> > > > 
> > > > We shouldn't need anything.  A bunch of cleanup
> > > 
> > > Should we do the read/drop_merkle_tree_block cleanup anyway?
> > 
> > To me the block based interface seems a lot cleaner, but Eric has some
> > reservations due to the added indirect call on the drop side.
> 
> The point of the block based interface is really so that filesystems could
> actually do the block-based caching.  If XFS isn't going to end up using that
> after all, I think we should just stay with ->read_merkle_tree_page for now.

TBH I'd rather keep the existing caching mechanism we have now and not
mess further with the pagecache.  But that really depends on who's going
to get this patchset across the finish line.

> > > One of the advantages of xfs caching merkle tree blocks ourselves
> > > is that we neither extend the usage of PageChecked when merkle blocksize
> > > == pagesize nor become subject to the 1-million merkle block limit when
> > > merkle blocksize < pagesize.  There's a tripping hazard if you mount a 4k
> > > merkle block filesystem on a computer with 64k pages -- now you can't
> > > open 6T verity files.
> > > 
> > > That said, it also sounds dumb to maintain a separate index for
> > > pagecache pages to track a single bit.
> > 
> > Yeah.  As I mentioned earlier I think fsverify really should enforce
> > a size limit.  Right now it will simply run out space eventually which
> > doesn't seem like a nice failure mode.
> 
> You could enforce the limit of 1 << 23 Merkle tree blocks regardless of whether
> the block size and page size are equal, if you want.  This probably would need
> to be specific to XFS, though, as it would be a breaking change for other
> filesystems.

Hmm.  If xfs stores the merkle data in posteof blocks, then I guess the
following constraint has to hold:

roundup_64(isize, rounding) + merkle_data_size < 2^63-1.

where rounding = max(file allocation unit, max folio size)

The "max folio size" part creates problems because the folio size
changes depending on arch and whether or not large folios are enabled,
and we cannot spill merkle tree data into an mmapped folio that crosses
EOF.

I guess we could hack around that by setting rounding = 1TB or
something, and then fail the begin_enable_verity call if that
relationship doesn't hold.

Then we'd store the merkle data at file position roundup_64(isize,
rounding).

> As for ext4 and f2fs failing with EFBIG in the middle of building the Merkle
> tree due to it exceeding s_maxbytes, yes that could be detected at the beginning
> of building the tree to get the error right away.  This just isn't done
> currently because the current way is simpler, and this case isn't really
> encountered in practice so there isn't too much reason to optimize it.

Yeah I don't think it matters if we run out of space or whatever during
construction of the merkle data since (if we switch to using posteof
space) we can just truncate at eof to get rid of the old tree, just like
everyone else.

--D

> - Eric
> 




[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux