page.title=Implementing dm-verity
@jd:body

<!--
    Copyright 2015 The Android Open Source Project

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.
-->
<div id="qv-wrapper">
  <div id="qv">
    <h2>In this document</h2>
    <ol id="auto-toc">
    </ol>
  </div>
</div>

<h2 id="operation">Operation</h2>

<p>dm-verity protection lives in the kernel. So if rooting software compromises the 
system before the kernel comes up, it will retain that access. To mitigate this 
risk, most manufacturers verify the kernel using a key burned into the device. 
That key is not changeable once the device leaves the factory.</p>

<p>Manufacturers use that key to verify the signature on the first-level 
bootloader, which in turn verifies the signature on subsequent levels, the 
application bootloader and eventually the kernel. Each manufacturer wishing to 
take advantage of <a href="verified-boot.html">verified boot</a> should have a
method for verifying the integrity of the kernel. Assuming the kernel has been
verified, the kernel can look at a block device and verify it as it is mounted.</p>

<p>One way of verifying a block device is to directly hash its contents and compare 
them to a stored value. However, attempting to verify an entire block device can 
take an extended period and consume much of a device's power. Devices would take 
long periods to boot and then be significantly drained prior to use.</p>

<p>Instead, dm-verity verifies blocks individually and only when each one is 
accessed. When read into memory, the block is hashed in parallel. The hash is 
then verified up the tree. And since reading the block is such an expensive 
operation, the latency introduced by this block-level verification is 
comparatively nominal.</p>

<p>If verification fails, the device generates an I/O error indicating the block 
cannot be read. It will appear as if the filesystem has been corrupted, as is 
expected.</p>

<p>Applications may choose to proceed without the resulting data, such as when 
those results are not required to the application's primary function. However, 
if the application cannot continue without the data, it will fail.</p>

<h2 id="implementation">Implementation</h2>

<h3 id="summary">Summary</h3>

<ol>
<li>Generate an ext4 system image.</li>
<li><a href="#hash-tree">Generate a hash tree</a> for that image.</li>
<li><a href="#mapping-table">Build a dm-verity table</a> for that hash tree.</li>
<li><a href="#signing">Sign that dm-verity table</a> to produce a table 
signature.</li>
<li><a href="#metadata">Bundle the table signature</a> and dm-verity table 
into verity metadata.</li>
<li>Concatenate the system image, the verity metadata, and the hash tree.</li>
</ol>

<p>See the <a href="http://www.chromium.org/chromium-os/chromiumos-design-docs/verified-boot">The Chromium Projects - Verified 
Boot</a> 
for a detailed description of the hash tree and dm-verity table.</p>

<h3 id="hash-tree">Generating the hash tree</h3>

<p>As described in the <a href="#introduction">Introduction</a>, the hash tree is 
integral to dm-verity. The 
<a href="https://code.google.com/p/cryptsetup/wiki/DMVerity">cryptsetup</a> tool will 
generate a hash tree for you. Alternatively, a compatible one is defined here:</p>

<pre>
&lt;your block device name&gt; &lt;your block device name&gt; &lt;block size&gt; &lt;block size&gt; &lt;image size in blocks&gt; &lt;image size in blocks + 8&gt; &lt;root hash&gt; &lt;salt&gt;
</pre>

<p>To form the hash, the system image is split at layer 0 into 4k blocks, each 
assigned a SHA256 hash. Layer 1 is formed by joining only those SHA256 hashes 
into 4k blocks, resulting in a much smaller image. Layer 2 is formed 
identically, with the SHA256 hashes of Layer 1.</p>

<p>This is done until the SHA256 hashes of the previous layer can fit in a single 
block. When get the SHA256 of that block, you have the root hash of the tree. </p>

<p>The size of the hash tree (and corresponding disk space usage) varies with the 
size of the verified partition. In practice, the size of hash trees tends to be 
small, often less than 30 MB.</p>

<p>If you have a block in a layer that isn't completely filled naturally by the 
hashes of the previous layer, you should pad it with zeroes to achieve the 
expected 4k. This allows you to know the hash tree hasn't been removed and is 
instead completed with blank data.</p>

<p>To generate the hash tree, concatenate the layer 2 hashes onto those for layer 
1, the layer 3 the hashes onto those of layer 2, and so on. Write all of this 
out to disk. Note that this doesn't reference layer 0 of the root hash.</p>

<p>To recap, the general algorithm to construct the hash tree is as follows:</p>

<ol>
<li>Choose a random salt (hexadecimal encoding).</li>
<li>Unsparse your system image into 4k blocks.</li>
<li>For each block, get its (salted) SHA256 hash.</li>
<li>Concatenate these hashes to form a level</li>
<li>Pad the level with 0s to a 4k block boundary.</li>
<li>Concatenate the level to your hash tree.</li>
<li>Repeat steps 2-6 using the previous level as the source for the next until 
you have only a single hash.</li>
</ol>

<p>The result of this is a single hash, which is your root hash. This and your salt 
are used during the construction of your dm-verity mapping hash table.</p>

<h3 id="mapping-table">Building the dm-verity mapping table</h3>

<p>Build the dm-verity mapping table, which identifies the block device (or target) 
for the kernel and the location of the hash tree (which is the same value.) This 
mapping is used for <code>fstab</code> generation and booting. The table also identifies 
the size of the blocks and the hash_start, or the offset in hash size blocks 
(length of layer 0).</p>

<p>See <a href="https://code.google.com/p/cryptsetup/wiki/DMVerity">cryptsetup</a> for a 
detailed description of the verity target mapping table fields.</p>

<h3 id="signing">Signing the dm-verity table</h3>

<p>Sign the dm-verity table to produce a table signature. When verifying a 
partition, the table signature is validated first. This is done against a key on 
your boot image in a fixed location. Keys are typically included in the 
manufacturers' build systems for automatic inclusion on devices in a fixed 
location.</p>

<p>To verify the partition with this signature and key combination:</p>

<ol>
<li>Add an RSA-2048 key in libmincrypt-compatible format to the /boot partition 
at /verity_key. Identify the location of the key used to verify the hash 
tree.</li>
<li>In the fstab for the relevant entry, add 'verify' to the fs_mgr flags.</li>
</ol>

<h3 id="metadata">Bundling the table signature into metadata</h3>

<p>Bundle the table signature and dm-verity table into verity metadata. The entire 
block of metadata is versioned so it may be extended, such as to add a second 
kind of signature or change some ordering.</p>

<p>As a sanity check, a magic number is associated with each set of table metadata 
that helps identify the table. Since the length is included in the ext4 system 
image header, this provides a way to search for the metadata without knowing the 
contents of the data itself.</p>

<p>This makes sure you haven't elected to verify an unverified partition. If so, 
the absence of this magic number will halt the verification process. This number 
resembles:<br/>
0xb001b001</p>

<p>The byte values in hex are:</p>

<ul>
<li>first byte = b0</li>
<li>second byte = 01</li>
<li>third byte = b0</li>
<li>fourth byte = 01</li>
</ul>

<p>The following diagram depicts the breakdown of the verity metadata:</p>

<pre>&lt;magic number&gt;|&lt;version&gt;|&lt;signature&gt;|&lt;table length&gt;|&lt;table&gt;|&lt;padding&gt;
\-------------------------------------------------------------------/
\----------------------------------------------------------/   |
                            |                                  |
                            |                                 32K
                       block content
</pre>

<p>And this table describes those metadata fields.</p>

<table>
<tr>
<th>Field</th>
<th>Purpose</th>
<th>Size</th>
<th>Value</th>
</tr>
<tr>
<td>magic number</td>
<td>used by fs_mgr as a sanity check</td>
<td>4 bytes</td>
<td>0xb001b001</td>
</tr>
<tr>
<td>version</td>
<td>used to version the metadata block</td>
<td>4 bytes</td>
<td>currently 0</td>
</tr>
<tr>
<td>signature</td>
<td>the signature of the table in PKCS1.5 padded form</td>
<td>256 bytes</td>
<td></td>
</tr>
<tr>
<td>table length</td>
<td>the length of the dm-verity table in bytes</td>
<td>4 bytes</td>
<td></td>
</tr>
<tr>
<td>table</td>
<td>the dm-verity table described earlier</td>
<td>`table length` bytes</td>
<td></td>
</tr>
<tr>
<td>padding</td>
<td>this structure is 0-padded to 32k in length</td>
<td></td>
<td>0</td>
</tr>
</table>