Message-Id: <01918C7B-9D9B-4BD8-8ED1-BA1CBF53CA95@dilger.ca>
Date:   Wed, 17 Feb 2021 15:48:39 -0700
From:   Andreas Dilger <adilger@...ger.ca>
To:     Theodore Ts'o <tytso@....edu>
Cc:     Daniel Rosenberg <drosen@...gle.com>,
        Eric Biggers <ebiggers@...nel.org>,
        Ext4 Developers List <linux-ext4@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        Gabriel Krisman Bertazi <krisman@...labora.com>,
        kernel-team@...roid.com, Paul Lawrence <paullawrence@...gle.com>
Subject: Re: [PATCH 1/2] ext4: Handle casefolding with encryption

On Feb 17, 2021, at 9:08 AM, Theodore Ts'o <tytso@....edu> wrote:
> 
> On Tue, Feb 16, 2021 at 08:01:11PM -0800, Daniel Rosenberg wrote:
>> I'm not sure what the conflict is, at least format-wise. Naturally,
>> there would need to be some work to reconcile the two patches, but my
>> patch only alters the format for directories which are encrypted and
>> casefolded, which always must have the additional hash field. In the
>> case of dirdata along with encryption and casefolding, couldn't we
>> have the dirdata simply follow after the existing data? Since we
>> always already know the length, it'd be unambiguous where that would
>> start. Casefolding can only be altered on an empty directory, and you
>> can only enable encryption for an empty directory, so I'm not too
>> concerned there. I feel like having it swapping between the different
>> methods makes it more prone to bugs, although it would be doable. I've
>> started rebasing the dirdata patch on my end to see how easy it is to
>> mix the two. At a glance, they touch a lot of the same areas in
>> similar ways, so it shouldn't be too hard. It's more of a question of
>> which way we want to resolve that, and which patch goes first.
>> 
>> I've been trying to figure out how many devices in the field are using
>> casefolded encryption, but haven't found out yet. The code is
>> definitely available though, so I would not be surprised if it's being
>> used, or is about to be.
> 
> The problem is in how the space after the filename in a directory is
> encoded.  The dirdata format is (mildly) expandable, supporting up to
> 4 different metadata chunks after the filename, using a very
> compatctly encoded TLV (or moral equivalent) scheme.  For directory
> inodes that have both the encyption and compression flags set, we have
> a single blob which gets used as the IV for the crypto.
> 
> So it's the difference between a simple blob that is only used for one
> thing in this particular case, and something which is the moral
> equivalent of simple ASN.1 or protobuf encoding.
> 
> Currently, datadata has defined uses for 2 of the 4 "chunks", which is
> used in Lustre servers.  The proposal which Andreas has suggested is
> if the dirdata feature is supported, then the 3rd dirdata chunk would
> be used for the case where we currently used by the
> encrypted-casefolded extension, and the 4th would get reserved for a
> to-be-defined extension mechanism.
> 
> If there ext4 encrypted/casefold is not yet in use, and we can get the
> changes out to all potential users before they release products out
> into the field, then one approach would be to only support
> encrypted/casefold when dirdata is also enabled.
> 
> If ext4 encrypted/casefold is in use, my suggestion is that we support
> both encrypted/casefold && !dirdata as you have currently implemented
> it, and encrypted/casefold && dirdata as Andreas has proposed.
> 
> IIRC, supporting that Andreas's scheme essentially means that we use
> the top four bits in the rec_len field to indicate which chunks are
> present, and then for each chunk which is present, there is a 1 byte
> length followed by payload.  So that means in the case where it's
> encrypted/casefold && dirdata, the required storage of the directory
> entry would take one additional byte, plus setting a bit indicating
> that the encrypted/casefold dirdata chunk was present.

I think your email already covers pretty much all of the points.

One small difference between current "raw" encrypted/casefold hash vs.
dirdata is that the former is 4-byte aligned within the dirent, while
dirdata is packed.  So in 3/4 cases dirdata would take the same amount
of space (the 1-byte length would use one of the 1-3 bytes of padding
vs. the raw format), since the next dirent needs to be aligned anyway.

The other implication here is that the 8-byte hash may need to be
copied out of the dirent into a local variable before use, due to
alignment issues, but I'm not sure if that is actually needed or not.

> So, no, they aren't incompatible ultimatly, but it might require a
> tiny bit more work to integrate the combined support for dirdata plus
> encrypted/casefold.  One way we can do this, if we have to support the
> current encrypted/casefold format because it's out there in deployed
> implementations already, is to integrate encrypted/casefold &&
> !dirdata first upstream, and then when we integrate dirdata into
> upstream, we'll have to add support for the encrypted/casefold &&
> dirdata case.  This means that we'll have two variants of the on-disk
> format to test and support, but I don't think it's the going to be
> that difficult.

It would be possible to detect if the encrypted/casefold+dirdata
variant is in use, because the dirdata variant would have the 0x40
bit set in the file_type byte.  It isn't possible to positively
identify the "raw" non-dirdata variant, but the assumption would be
if (rec_len >= round_up(name_len, 4) + 8) in an encrypted+casefold
directory that the "raw" hash must be present in the dirent.

Cheers, Andreas






Download attachment "signature.asc" of type "application/pgp-signature" (874 bytes)

Powered by blists - more mailing lists