Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Andreas Dilger <adilger@dilger.ca>
Mime-Version: 1.0 (1.0)
Subject: Re: algorithm for half-md4 used in htree directories
Date:   Sun, 3 Oct 2021 10:43:17 -0600
Message-Id: <3A493D20-568A-4D63-A575-5DEEBFAAF41A@dilger.ca>
References: <CAF1vpkgPAy3FJ9mN22OVQ41jQAYoRdoCdqzYwRYYPXD4uucdpg@mail.gmail.com>
Cc:     linux-ext4@vger.kernel.org
In-Reply-To: <CAF1vpkgPAy3FJ9mN22OVQ41jQAYoRdoCdqzYwRYYPXD4uucdpg@mail.gmail.com>
To:     Avi Deitcher <avi@deitcher.net>
Precedence: bulk

On Oct 3, 2021, at 06:47, Avi Deitcher <avi@deitcher.net> wrote:
>=20
> =EF=BB=BFI can narrow down the question further. In my live sample, one of=
 the
> entries in the tree is for a directory named "dir155".
>=20
> If I run "dx_hash dir155", I get:
>=20
> # debugfs -R "dx_hash dir155" /var/lib/file.img
> debugfs 1.46.2 (28-Feb-2021)
> Hash of dir155 is 0x16279534 (minor 0x0)
>=20
> If I look in the tree with "htree_dump", I get:
>=20
> # debugfs -R "htree_dump /testdir" /var/lib/file.img
> debugfs 1.46.2 (28-Feb-2021)
> ....
> Entry #0: Hash 0x00000000, block 1
> Reading directory block 1, phys 6459
> 168 0x00d11d98-b9b6b16b (16) dir155   332 0x009edafe-77de7d72 (16) dir319
>=20
> That hash for dir155 does not match what dx_hash gave. If I try to
> take the code from fs/ext4/hash.c and build a small program to
> calculate the hash, I get:
>=20
> $ ./md4 dir155
> MD4: d90278a1 25182ac7 a02e56be c3f30f04
> hash: 0x25182ac6
> minor: 0xa02e56be
>=20
> Clearly that isn't what is in the tree. What basic am I missing?

One important factor is that the directory hash has an initial seed
to prevent pathological cases where the user can construct thousands
of directory entries that have a hash collision.

Looking at the code explains this in the comment for __ext4fs_dirhash().
The seed itself comes from sbi->s_hash_seed and is stored in the
per-directory hinfo.seed to be used when counting the filename hash.
In theory there could be a per-directory hash, but it appears to be a
constant for the whole filesystem. =20

Cheers, Andreas

>=20
>> On Fri, Oct 1, 2021 at 2:49 PM Avi Deitcher <avi@deitcher.net> wrote:
>>=20
>> Hi,
>>=20
>> I have been trying to understand the algorithm used for the "half-md4"
>> in htree-structured directories. Going through the code (and trying
>> not to get into reverse engineering), it looks like it is part of md4
>> but not entirely? Yet any subset I take doesn't quite line up with
>> what I see in an actual sample.
>>=20
>> What is the algorithm it is using to turn an entry of, e.g., "file125"
>> into the appropriate hash. I did run a live sample, and try to get
>> some form of correlation between the actual md4 hash (16 bytes) of the
>> above to the actual entry (4 bytes) shown by debugfs, without much
>> luck.
>>=20
>> What basic thing am I missing?
>>=20
>> Separately, how does the seed play into it?
>>=20
>> Thanks
>> Avi
>=20
>=20
>=20
> --=20
> Avi Deitcher
> avi@deitcher.net
> Follow me http://twitter.com/avideitcher
> Read me http://blog.atomicinc.com