Tampilkan postingan dengan label Security. Tampilkan semua postingan
Tampilkan postingan dengan label Security. Tampilkan semua postingan

Rabu, 27 Juli 2016

Protecting Android with more Linux kernel defenses


Posted by Jeff Vander Stoep, Android Security team



Android relies heavily on the Linux kernel for enforcement of its security
model. To better protect the kernel, we’ve enabled a number of mechanisms within
Android. At a high level these protections are grouped into two
categories—memory protections and attack surface reduction.


Memory protections



One of the major security features provided by the kernel is memory protection
for userspace processes in the form of address space separation. Unlike
userspace processes, the kernel’s various tasks live within one address space
and a vulnerability anywhere in the kernel can potentially impact unrelated
portions of the system’s memory. Kernel memory protections are designed to
maintain the integrity of the kernel in spite of vulnerabilities.


Mark memory as read-only/no-execute



This feature segments kernel memory into logical sections and sets restrictive
page access permissions on each section. Code is marked as read only + execute.
Data sections are marked as no-execute and further segmented into read-only and
read-write sections. This feature is enabled with config option
CONFIG_DEBUG_RODATA. It was put together by Kees Cook and is based on a subset
of Grsecurity’s KERNEXEC feature by Brad
Spengler and Qualcomm’s CONFIG_STRICT_MEMORY_RWX feature by Larry Bassel and
Laura Abbott. CONFIG_DEBUG_RODATA landed in the upstream kernel for arm/arm64
and has been backported to Android’s 3.18+ arm/href="https://android-review.googlesource.com/#/c/174947/">arm64 common
kernel.


Restrict kernel access to userspace



This feature improves protection of the kernel by preventing it from directly
accessing userspace memory. This can make a number of attacks more difficult
because attackers have significantly less control over kernel memory
that is executable, particularly with CONFIG_DEBUG_RODATA enabled. Similar
features were already in existence, the earliest being Grsecurity’s UDEREF. This
feature is enabled with config option CONFIG_CPU_SW_DOMAIN_PAN and was
implemented by Russell King for ARMv7 and backported to href="https://android-review.googlesource.com/#/q/topic:sw_PAN">Android’s
4.1 kernel by Kees Cook.


Improve protection against stack buffer overflows



Much like its predecessor, stack-protector, stack-protector-strong protects
against stack
buffer overflows
, but additionally provides coverage for href="https://outflux.net/blog/archives/2014/01/27/fstack-protector-strong/">more
array types, as the original only protected character arrays.
Stack-protector-strong was implemented by Han Shen and href="https://gcc.gnu.org/ml/gcc-patches/2012-06/msg00974.html">added to the gcc
4.9 compiler.



Attack surface reduction



Attack surface reduction attempts to expose fewer entry points to the kernel
without breaking legitimate functionality. Reducing attack surface can include
removing code, removing access to entry points, or selectively exposing
features.


Remove default access to debug features



The kernel’s perf system provides infrastructure for performance measurement and
can be used for analyzing both the kernel and userspace applications. Perf is a
valuable tool for developers, but adds unnecessary attack surface for the vast
majority of Android users. In Android Nougat, access to perf will be blocked by
default. Developers may still access perf by enabling developer settings and
using adb to set a property: “adb shell setprop security.perf_harden 0”.



The patchset for blocking access to perf may be broken down into kernel and
userspace sections. The href="https://android-review.googlesource.com/#/c/234573/">kernel patch is
by Ben Hutchings and is
derived from Grsecurity’s CONFIG_GRKERNSEC_PERF_HARDEN by Brad Spengler. The
userspace changes were href="https://android-review.googlesource.com/#/q/topic:perf_harden">contributed
by Daniel Micay. Thanks to href="https://conference.hitb.org/hitbsecconf2016ams/sessions/perf-from-profiling-to-kernel-exploiting/">Wish
Wu and others for responsibly disclosing security vulnerabilities in perf.


Restrict app access to ioctl commands



Much of Android security model is described and enforced by SELinux. The ioctl()
syscall represented a major gap in the granularity of enforcement via SELinux.
Ioctl command
whitelisting with SELinux
was added as a means to provide per-command
control over the ioctl syscall by SELinux.



Most of the kernel vulnerabilities reported on Android occur in drivers and are
reached using the ioctl syscall, for example href="https://source.android.com/security/bulletin/2016-03-01.html#elevation_of_privilege_vulnerability_in_mediatek_wi-fi_kernel_driver">CVE-2016-0820.
Some ioctl commands are needed by third-party applications, however most are not
and access can be restricted without breaking legitimate functionality. In
Android Nougat, only a small whitelist of socket ioctl commands are available to
applications. For select devices, applications’ access to GPU ioctls has been
similarly restricted.


Require seccomp-bpf



Seccomp provides an additional sandboxing mechanism allowing a process to
restrict the syscalls and syscall arguments available using a configurable
filter. Restricting the availability of syscalls can dramatically cut down on
the exposed attack surface of the kernel. Since seccomp was first introduced on
Nexus devices in Lollipop, its availability across the Android ecosystem has
steadily improved. With Android Nougat, seccomp support is a requirement for all
devices. On Android Nougat we are using seccomp on the mediaextractor and
mediacodec processes as part of the href="http://android-developers.blogspot.com/2016/05/hardening-media-stack.html">media
hardening effort.


Ongoing efforts



There are other projects underway aimed at protecting the kernel:


  • The href="http://kernsec.org/wiki/index.php/Kernel_Self_Protection_Project">Kernel
    Self Protection Project is developing runtime and compiler defenses for the
    upstream kernel.
  • Further sandbox tightening and attack surface reduction with SELinux is
    ongoing in AOSP.
  • href="https://www.chromium.org/chromium-os/developer-guide/chromium-os-sandboxing#h.l7ou90opzirq">Minijail
    provides a convenient mechanism for applying many containment and sandboxing
    features offered by the kernel, including seccomp filters and namespaces.
  • Projects like href="https://www.kernel.org/doc/Documentation/kasan.txt">kasan and href="https://www.kernel.org/doc/Documentation/kcov.txt">kcov help fuzzers
    discover the root cause of crashes and to intelligently construct test cases
    that increase code coverage—ultimately resulting in a more efficient bug hunting
    process.


Due to these efforts and others, we expect the security of the kernel to
continue improving. As always, we appreciate feedback on our work and welcome
suggestions for how we can improve Android. Contact us at href="mailto:security@android.com">security@android.com.



Selasa, 19 Juli 2016

Strictly Enforced Verified Boot with Error Correction

Posted by Sami Tolvanen, Software Engineer



Overview



Android uses multiple layers of protection to keep users safe. One of these
layers is verified
boot
, which improves security by using cryptographic integrity checking to
detect changes to the operating system. Android has href="https://g.co/ABH">alerted about system integrity since Marshmallow,
but starting with devices first shipping with Android 7.0, we require verified
boot to be strictly enforcing. This means that a device with a corrupt boot
image or verified partition will not boot or will boot in a limited capacity
with user consent. Such strict checking, though, means that non-malicious data
corruption, which previously would be less visible, could now start affecting
process functionality more.



By default, Android verifies large partitions using the dm-verity kernel driver,
which divides the partition into 4 KiB blocks and verifies each block when read,
against a signed hash tree. A detected single byte corruption will therefore
result in an entire block becoming inaccessible when dm-verity is in enforcing
mode, leading to the kernel returning EIO errors to userspace on verified
partition data access.



This post describes our work in improving dm-verity robustness by introducing
forward error correction (FEC), and explains how this allowed us to make the
operating system more resistant to data corruption. These improvements are
available to any device running Android 7.0 and this post reflects the default
implementation in AOSP that we ship on our Nexus devices.


Error-correcting codes



Using forward error correction, we can detect and correct errors in source data
by shipping redundant encoding data generated using an error-correcting code.
The exact number of errors that can be corrected depends on the code used and
the amount of space allocated for the encoding data.



href="https://en.wikipedia.org/wiki/Reed%E2%80%93Solomon_error_correction">Reed-Solomon
is one of the most commonly used error-correcting code families, and is readily
available in the Linux kernel, which makes it an obvious candidate for
dm-verity. These codes can correct up to ⌊t/2⌋ unknown errors and up to
t known errors, also called href="https://en.wikipedia.org/wiki/Erasure_code">erasures, when t
encoding symbols are added.



A typical RS(255, 223) code that generates 32 bytes of encoding data for every
223 bytes of source data can correct up to 16 unknown errors in each 255 byte
block. However, using this code results in ~15% space overhead, which is
unacceptable for mobile devices with limited storage. We can decrease the space
overhead by sacrificing error correction capabilities. An RS(255, 253) code can
correct only one unknown error, but also has an overhead of only 0.8%.




An additional complication is that block-based storage corruption often occurs
for an entire block and sometimes spans multiple consecutive blocks. Because
Reed-Solomon is only able to recover from a limited number of corrupted bytes
within relatively short encoded blocks, a naive implementation is not going to
be very effective without a huge space overhead.


Recovering from consecutive corrupted blocks



In the changes we made to href="https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=a739ff3f543afbb4a041c16cd0182c8e8d366e70">dm-verity
for Android 7.0, we used a technique called interleaving to allow us to recover
not only from a loss of an entire 4 KiB source block, but several consecutive
blocks, while significantly reducing the space overhead required to achieve
usable error correction capabilities compared to the naive implementation.



Efficient interleaving means mapping each byte in a block to a separate
Reed-Solomon code, with each code covering N bytes across the corresponding N
source blocks. A trivial interleaving where each code covers a consecutive
sequence of N blocks already makes it possible for us to recover from the
corruption of up to (255 - N) / 2 blocks, which for RS(255, 223) would
mean 64 KiB, for example.



An even better solution is to maximize the distance between the bytes covered by
the same code by spreading each code over the entire partition, thereby
increasing the maximum number of consecutive corrupted blocks an RS(255, N) code
can handle on a partition consisting of T blocks to ⌈T/N⌉ × (255 -
N) / 2
.




Interleaving with distance D and block size B.



An additional benefit of interleaving, when combined with the integrity
verification already performed by dm-verity, is that we can tell exactly where
the errors are in each code. Because each byte of the code covers a different
source block—and we can verify the integrity of each block using the existing
dm-verity metadata—we know which of the bytes contain errors. Being able to
pinpoint erasure locations allows us to effectively double our error correction
performance to at most ⌈T/N⌉ × (255 - N) consecutive blocks.



For a ~2 GiB partition with 524256 4 KiB blocks and RS(255, 253), the maximum
distance between the bytes of a single code is 2073 blocks. Because each code
can recover from two erasures, using this method of interleaving allows us to
recover from up to 4146 consecutive corrupted blocks (~16 MiB). Of course, if
the encoding data itself gets corrupted or we lose more than two of the blocks
covered by any single code, we cannot recover anymore.



While making error correction feasible for block-based storage, interleaving
does have the side effect of making decoding slower, because instead of reading
a single block, we need to read multiple blocks spread across the partition to
recover from an error. Fortunately, this is not a huge issue when combined with
dm-verity and solid-state storage as we only need to resort to decoding if a
block is actually corrupted, which still is rather rare, and random access reads
are relatively fast even if we have to correct errors.


Conclusion



Strictly enforced verified boot improves security, but can also reduce
reliability by increasing the impact of disk corruption that may occur on
devices due to software bugs or hardware issues.



The new error correction feature we developed for dm-verity makes it possible
for devices to recover from the loss of up to 16-24 MiB of consecutive blocks
anywhere on a typical 2-3 GiB system partition with only 0.8% space overhead and
no performance impact unless corruption is detected. This improves the security
and reliability of devices running Android 7.0.