Mitigations(7) - SerenityOS man pages

#Name

Mitigations - Security mitigations implemented by SerenityOS

#Description

The SerenityOS developers have put substantial effort into integrating various mitigation technologies into the system in order to enhance its security. The goal of this document is to collect and describe the mitigations in one centralized place.

#List of Mitigations

#SMEP (Supervisor Mode Execution Protection)

Supervisor Mode Execution Protection is an Intel CPU feature which prevents execution of userspace code with kernel privileges.

It was enabled in the following commit:

commit 8602fa5b49aa4e2b039764a14698f0baa3ad0532
Author: Andreas Kling <awesomekling@gmail.com>
Date:   Wed Jan 1 01:56:58 2020 +0100

Kernel: Enable x86 SMEP (Supervisor Mode Execution Protection)

#SMAP (Supervisor Mode Access Prevention)

Supervisor Mode Access Prevention complements SMEP by also guarding read/write access to userspace memory while executing in kernel mode.

It was enabled in the following commit:

commit 9eef39d68a99c5e29099ae4eb4a56934b35eecde
Author: Andreas Kling <awesomekling@gmail.com>
Date:   Sun Jan 5 18:00:15 2020 +0100

Kernel: Start implementing x86 SMAP support

#UMIP (User Mode Instruction Prevention)

User Mode Instruction Prevention is an x86 CPU security feature which prevents execution of specific privileged instructions in user mode (SGDT, SIDT, SLDT, SMSW, STR). These instructions let user mode code query the addresses of various kernel structures (the GDT, LDT, IDT, etc), meaning that they leak kernel addresses that can be exploited to defeat ASLR.

It was enabled in the following commit:

commit 9c0836ce97ae36165abd8eb5241bb5239af3a756
Author: Andreas Kling <awesomekling@gmail.com>
Date:   Wed Jan 1 13:02:32 2020 +0100

Kernel: Enable x86 UMIP (User Mode Instruction Prevention) if supported

#Pledge

pledge is a mitigation which originated from OpenBSD. It allows a program to voluntarily restrict its access to system calls and kernel facilities.

It was first added in the following commit, and the majority of programs were enlightened later:

commit 41c504a33becea8aa9b437cd3c0dc312b2bf1fe9
Author: Andreas Kling <awesomekling@gmail.com>
Date:   Sat Jan 11 20:45:51 2020 +0100

Kernel: Add pledge() syscall :^)

#Unveil

unveil is a mitigation originating from OpenBSD. It allows a program to voluntarily restrict its access to the filesystem.

It was first added in the following commit, and the majority of programs were enlightened later:

commit 0569123ad7cb9c54df724c2bb85933ea3cf97134
Author: Andreas Kling <kling@serenityos.org>
Date:   Mon Jan 20 22:12:04 2020 +0100

Kernel: Add a basic implementation of unveil()

#Jails

jails are mitigation originating from FreeBSD. It allows a program to be placed inside a lightweight OS-level virtualization environment.

Current restrictions on jailed processes (configurable when creating a Jail):

Special restrictions on filesystem also apply:

It was first added in the following commit, for kernel support, and the following commits added basic userspace utilities:

commit 5e062414c11df31ed595c363990005eef00fa263
Author: Liav A <liavalb@gmail.com>
Date:   Wed Nov 2 22:26:02 2022 +0200

Kernel: Add support for jails

...

#Readonly atexit

Readonly atexit is a mitigation originating from OpenBSD. Thanks to it, an attacker can no longer use the atexit region to escalate from arbitrary-write to code-execution.

It was first added in the following commit, and was later improved to incur no additional cost during program initialization and finalization:

commit 553361d83f7bc6499dc4821eff9b23a6549bd99c
Author: Andreas Kling <kling@serenityos.org>
Date:   Sat Jan 30 10:34:41 2021 +0100

LibC: Protect the atexit() handler list when not writing to it

Remap the list of atexit handlers as read-only while we're not actively
writing to it. This prevents an attacker from using a memory write
primitive to gain code execution via the atexit list.

This is based on a technique used in OpenBSD. :^)

#Syscall call-from verification

Syscall call-from verification is a mitigation which originated from OpenBSD. In short the kernel checks that all syscalls originate from the address of the system's libc. This makes attacks on OpenBSD more difficult as they random-relink their libc on boot, which makes finding syscall stubs in libc difficult for attackers. On serenity it is mostly just an inconvenience, as there currently is no libc random-relinking.

It was first enabled in the following commit:

commit 823186031d9250217f9a51829d34a96b74113334
Author Andreas Kling <kling@serenityos.org>
Date:  Tue Feb 2 19:56:11 2021 +0100

Kernel: Add a way to specify which memory regions can make syscalls

#Immutable memory mappings

Immutable memory mappings is a mitigation which originated from OpenBSD. In short the annotation of a particular Kernel Region as immutable implies that that these virtual memory mappings are locked to their last state (in regard to protection bits, etc), and they cannot be unmapped by a process until that process gets finalized.

It was first enabled in the following commit:

commit 8585b2dc23ec206777a4cfbd558766d90fc577e7
Author: Liav A <liavalb@gmail.com>
Date:   Thu Dec 15 21:08:57 2022 +0200

Kernel/Memory: Add option to annotate region mapping as immutable

We add this basic functionality to the Kernel so Userspace can request a
particular virtual memory mapping to be immutable. This will be useful
later on in the DynamicLoader code.

The annotation of a particular Kernel Region as immutable implies that
the following restrictions apply, so these features are prohibited:
- Changing the region's protection bits
- Unmapping the region
- Annotating the region with other virtual memory flags
- Applying further memory advises on the region
- Changing the region name
- Re-mapping the region

#Post-init read-only memory

Post-init read-only memory is a mitigation which originated from the Linux Kernel. It tracks data that is initialized during kernel boot and never changed again. Post kernel initialization, the memory is marked read-only to protect it from potentially being modified by exploits.

It was first enabled in the following commit and other kernel data structures were enlightened later:

commit d8013c60bb52756788e747183572067d6e3f204a
Author: Andreas Kling <kling@serenityos.org>
Date:   Sun Feb 14 17:35:07 2021 +0100

Kernel: Add mechanism to make some memory read-only after init finishes

#KUBSAN (Kernel Undefined Behavior Sanitizer)

UndefinedBehaviorSANitizer is a dynamic analysis tool, implemented in GCC, which instruments generated code to flag undefined behavior at runtime. It can find various issues, including integer overflows, out-of-bounds array accesses, type corruption, and more.

It was first enabled in the following commit:

commit d44be968938ecf95023351a358c43c4957638d87
Author: Andreas Kling <kling@serenityos.org>
Date:   Fri Feb 5 19:44:26 2021 +0100

Kernel: KUBSAN! (Kernel Undefined Behavior SANitizer) :^)

#Kernel unmap-after-init

Unmap-after-init allows the kernel to remove functions which contain potentially dangerous ROP gadgets from kernel memory after we've booted up and they are no longer needed. Notably the write_cr4(..) function used to control processor features like the SMEP/SMAP bits in the CR4 register, and the write_cr0(..) function used to control processor features like write protection, etc.

With this mitigation it is now more difficult to craft a kernel exploit to do something like disabling SMEP / SMAP.

It was first enabled in the following commit:

commit 6136faa4ebf6a878606f33bc03c5e62de9d5e662
Author: Andreas Kling <kling@serenityos.org>
Date:   Fri Feb 19 18:21:54 2021 +0100

Kernel: Add .unmap_after_init section for code we don't need after init

#Relocation Read-Only (RELRO)

RELRO is a mitigation in the linker and loader that hardens the data sections of an ELF binary.

When enabled, it segregates function pointers resolved by the dynamic loader into a separate section of the runtime executable memory, and allows the loader to make that memory read-only before passing control to the main executable.

This prevents attackers from overwriting the Global Offset Table (GOT).

It was first enabled for executables in the following commit:

commit fa4c249425a65076ca04a3cb0c173d49472796fb
Author: Andreas Kling <kling@serenityos.org>
Date:   Thu Feb 18 18:43:20 2021 +0100

LibELF+Userland: Enable RELRO for all userland executables :^)

Shared libraries were enabled in a follow-up commit:

commit 713b3b36be4f659e58e253b6c830509898dbd2fa
Author: Andreas Kling <kling@serenityos.org>
Date:   Thu Feb 18 22:49:58 2021 +0100

DynamicLoader+Userland: Enable RELRO for shared libraries as well :^)

#-fstack-clash-protection

The GCC compiler option -fstack-clash-protection is a mitigation which helps prevent stack clash style attacks by generating code that probes the stack in page-sized increments to ensure a fault is provoked. This prevents attackers from using a large stack allocation to "jump over" the stack guard page into adjacent memory.

It was first enabled in the following commit:

commit 7142562310e631156d1f64aff22f068ae2c48a5e
Author: Andreas Kling <kling@serenityos.org>
Date:   Fri Feb 19 09:11:02 2021 +0100

Everywhere: Build with -fstack-clash-protection

#-fstack-protector / -fstack-protector-strong

The GCC compiler provides a few variants of the -fstack-protector option mitigation. This family of flags enables buffer overflow protection to mitigate stack-smashing attacks.

The compiler implements the mitigation by storing a canary value randomized on program startup into the preamble of all functions. Code is then generated to validate that stack canary on function return and crash if the value has been changed (and hence a stack corruption has been detected.)

-fstack-protector was first enabled in the following commit:

commit 842716a0b5eceb8db31416cd643720c1037032b2
Author: Andreas Kling <awesomekling@gmail.com>
Date:   Fri Dec 20 20:51:50 2019 +0100

Kernel+LibC: Build with basic -fstack-protector support

It was later re-enabled and refined to -fstack-protector-strong in the following commits:

commit fd08c93ef57f71360d74b035214c71d7f7bfc5b8
Author: Brian Gianforcaro <b.gianfo@gmail.com>
Date:   Sat Jan 2 04:27:35 2021 -0800

LibC: Randomize the stack check cookie value on initialization

commit 79328b2aba6192caf28f560881e56ff23fcb7294
Author: Brian Gianforcaro <b.gianfo@gmail.com>
Date:   Sat Jan 2 03:02:42 2021 -0800

Kernel: Enable -fstack-protector-strong (again)

commit 06da50afc71a5ab2bc63de54c66930a2dbe379cd
Author: Brian Gianforcaro <b.gianfo@gmail.com>
Date:   Fri Jan 1 15:27:42 2021 -0800

Build + LibC: Enable -fstack-protector-strong in user space

#Protected Kernel Process Data

The kernel applies a exploit mitigation technique where vulnerable data related to the state of a process is separated out into it's own region in memory which is always remapped as read-only after it's initialized or updated. This means that an attacker needs more than an arbitrary kernel write primitive to be able to elevate a process to root for example.

It was first enabled in the following commit:

commit cbcf891040e9921ff628fdda668c9738f358a178
Author: Andreas Kling <kling@serenityos.org>
Date:   Wed Mar 10 19:59:46 2021 +0100

Kernel: Move select Process members into protected memory

#-fzero-call-used-regs

GCC-11 added a new option -fzero-call-used-regs which causes the compiler to zero function arguments before return of a function. The goal being to reduce the possible attack surface by disarming ROP gadgets that might be potentially useful to attackers, and reducing the risk of information leaks via stale register data.

It was first enabled when compiling the Kernel in the following commit:

commit 204d5ff8f86547a8b100cf26a958aaabf49211f2
Author: Brian Gianforcaro <bgianf@serenityos.org>
Date:   Fri Jul 23 00:42:54 2021 -0700

Kernel: Reduce useful ROP gadgets by zeroing used function registers

#Linking with "separate-code"

The linker is passed the separate-code option, so it won't combine read-only data and executable code. This reduces the total amount of executable pages in the system.

It was first enabled in the following commit:

commit fac0bbe739154abb416526bdc983487c05ba0c81
Author: Andreas Kling <kling@serenityos.org>
Date:   Tue Aug 31 16:08:11 2021 +0200

Build: Pass "-z separate-code" to linker

#KASLR (Kernel Address Space Layout Randomization)

The location of the kernel code is randomized at boot time, this ensures that attackers can not use a hardcoded kernel addresses when attempting ROP, instead they must first find an additional information leak to expose the KASLR offset.

It was first enabled in the following commit:

commit 1ad0e05ea1d3491e4724669d6f00f5668d8e0aa1
Author: Idan Horowitz <idan.horowitz@gmail.com>
Date:   Mon Mar 21 22:59:48 2022 +0200

Kernel: Add an extremely primitive version of KASLR

#Kernel -ftrivial-auto-var-init

As of GCC 12, both Clang and GCC now support the -ftrivial-auto-var-init compiler flag. The flag will cause the compiler to automatically initialize all variables to a pattern based on it's type. The goal being here is to eradicate an entire bug class of issues that can originate from uninitialized variables.

It was first enabled for the SerenityOS Kernel in the following commit:

From 458244c0c1c8f077030fa0d8964fad8d75c60d4a Mon Sep 17 00:00:00 2001
From: Brian Gianforcaro <bgianf@serenityos.org>
Date: Fri, 24 Jun 2022 00:34:38 -0700

Kernel: Enable -ftrivial-auto-var-init as a security mitigation

#See also