linux-next/Documentation/security/self-protection.txt

# Kernel Self-Protection

Kernel self-protection is the design and implementation of systems and
structures within the Linux kernel to protect against security flaws in
the kernel itself. This covers a wide range of issues, including removing
entire classes of bugs, blocking security flaw exploitation methods,
and actively detecting attack attempts. Not all topics are explored in
this document, but it should serve as a reasonable starting point and
answer any frequently asked questions. (Patches welcome, of course!)

In the worst-case scenario, we assume an unprivileged local attacker
has arbitrary read and write access to the kernel's memory. In many
cases, bugs being exploited will not provide this level of access,
but with systems in place that defend against the worst case we'll
cover the more limited cases as well. A higher bar, and one that should
still be kept in mind, is protecting the kernel against a _privileged_
local attacker, since the root user has access to a vastly increased
attack surface. (Especially when they have the ability to load arbitrary
kernel modules.)

The goals for successful self-protection systems would be that they
are effective, on by default, require no opt-in by developers, have no
performance impact, do not impede kernel debugging, and have tests. It
is uncommon that all these goals can be met, but it is worth explicitly
mentioning them, since these aspects need to be explored, dealt with,
and/or accepted.


## Attack Surface Reduction

The most fundamental defense against security exploits is to reduce the
areas of the kernel that can be used to redirect execution. This ranges
from limiting the exposed APIs available to userspace, making in-kernel
APIs hard to use incorrectly, minimizing the areas of writable kernel
memory, etc.

### Strict kernel memory permissions

When all of kernel memory is writable, it becomes trivial for attacks
to redirect execution flow. To reduce the availability of these targets
the kernel needs to protect its memory with a tight set of permissions.

#### Executable code and read-only data must not be writable

Any areas of the kernel with executable memory must not be writable.
While this obviously includes the kernel text itself, we must consider
all additional places too: kernel modules, JIT memory, etc. (There are
temporary exceptions to this rule to support things like instruction
alternatives, breakpoints, kprobes, etc. If these must exist in a
kernel, they are implemented in a way where the memory is temporarily
made writable during the update, and then returned to the original
permissions.)

In support of this are (the poorly named) CONFIG_DEBUG_RODATA and
CONFIG_DEBUG_SET_MODULE_RONX, which seek to make sure that code is not
writable, data is not executable, and read-only data is neither writable
nor executable.

#### Function pointers and sensitive variables must not be writable

Vast areas of kernel memory contain function pointers that are looked
up by the kernel and used to continue execution (e.g. descriptor/vector
tables, file/network/etc operation structures, etc). The number of these
variables must be reduced to an absolute minimum.

Many such variables can be made read-only by setting them "const"
so that they live in the .rodata section instead of the .data section
of the kernel, gaining the protection of the kernel's strict memory
permissions as described above.

For variables that are initialized once at __init time, these can
be marked with the (new and under development) __ro_after_init
attribute.

What remains are variables that are updated rarely (e.g. GDT). These
will need another infrastructure (similar to the temporary exceptions
made to kernel code mentioned above) that allow them to spend the rest
of their lifetime read-only. (For example, when being updated, only the
CPU thread performing the update would be given uninterruptible write
access to the memory.)

#### Segregation of kernel memory from userspace memory

The kernel must never execute userspace memory. The kernel must also never
access userspace memory without explicit expectation to do so. These
rules can be enforced either by support of hardware-based restrictions
(x86's SMEP/SMAP, ARM's PXN/PAN) or via emulation (ARM's Memory Domains).
By blocking userspace memory in this way, execution and data parsing
cannot be passed to trivially-controlled userspace memory, forcing
attacks to operate entirely in kernel memory.

### Reduced access to syscalls

One trivial way to eliminate many syscalls for 64-bit systems is building
without CONFIG_COMPAT. However, this is rarely a feasible scenario.

The "seccomp" system provides an opt-in feature made available to
userspace, which provides a way to reduce the number of kernel entry
points available to a running process. This limits the breadth of kernel
code that can be reached, possibly reducing the availability of a given
bug to an attack.

An area of improvement would be creating viable ways to keep access to
things like compat, user namespaces, BPF creation, and perf limited only
to trusted processes. This would keep the scope of kernel entry points
restricted to the more regular set of normally available to unprivileged
userspace.

### Restricting access to kernel modules

The kernel should never allow an unprivileged user the ability to
load specific kernel modules, since that would provide a facility to
unexpectedly extend the available attack surface. (The on-demand loading
of modules via their predefined subsystems, e.g. MODULE_ALIAS_*, is
considered "expected" here, though additional consideration should be
given even to these.) For example, loading a filesystem module via an
unprivileged socket API is nonsense: only the root or physically local
user should trigger filesystem module loading. (And even this can be up
for debate in some scenarios.)

To protect against even privileged users, systems may need to either
disable module loading entirely (e.g. monolithic kernel builds or
modules_disabled sysctl), or provide signed modules (e.g.
CONFIG_MODULE_SIG_FORCE, or dm-crypt with LoadPin), to keep from having
root load arbitrary kernel code via the module loader interface.


## Memory integrity

There are many memory structures in the kernel that are regularly abused
to gain execution control during an attack, By far the most commonly
understood is that of the stack buffer overflow in which the return
address stored on the stack is overwritten. Many other examples of this
kind of attack exist, and protections exist to defend against them.

### Stack buffer overflow

The classic stack buffer overflow involves writing past the expected end
of a variable stored on the stack, ultimately writing a controlled value
to the stack frame's stored return address. The most widely used defense
is the presence of a stack canary between the stack variables and the
return address (CONFIG_CC_STACKPROTECTOR), which is verified just before
the function returns. Other defenses include things like shadow stacks.

### Stack depth overflow

A less well understood attack is using a bug that triggers the
kernel to consume stack memory with deep function calls or large stack
allocations. With this attack it is possible to write beyond the end of
the kernel's preallocated stack space and into sensitive structures. Two
important changes need to be made for better protections: moving the
sensitive thread_info structure elsewhere, and adding a faulting memory
hole at the bottom of the stack to catch these overflows.

### Heap memory integrity

The structures used to track heap free lists can be sanity-checked during
allocation and freeing to make sure they aren't being used to manipulate
other memory areas.

### Counter integrity

Many places in the kernel use atomic counters to track object references
or perform similar lifetime management. When these counters can be made
to wrap (over or under) this traditionally exposes a use-after-free
flaw. By trapping atomic wrapping, this class of bug vanishes.

### Size calculation overflow detection

Similar to counter overflow, integer overflows (usually size calculations)
need to be detected at runtime to kill this class of bug, which
traditionally leads to being able to write past the end of kernel buffers.


## Statistical defenses

While many protections can be considered deterministic (e.g. read-only
memory cannot be written to), some protections provide only statistical
defense, in that an attack must gather enough information about a
running system to overcome the defense. While not perfect, these do
provide meaningful defenses.

### Canaries, blinding, and other secrets

It should be noted that things like the stack canary discussed earlier
are technically statistical defenses, since they rely on a secret value,
and such values may become discoverable through an information exposure
flaw.

Blinding literal values for things like JITs, where the executable
contents may be partially under the control of userspace, need a similar
secret value.

It is critical that the secret values used must be separate (e.g.
different canary per stack) and high entropy (e.g. is the RNG actually
working?) in order to maximize their success.

### Kernel Address Space Layout Randomization (KASLR)

Since the location of kernel memory is almost always instrumental in
mounting a successful attack, making the location non-deterministic
raises the difficulty of an exploit. (Note that this in turn makes
the value of information exposures higher, since they may be used to
discover desired memory locations.)

#### Text and module base

By relocating the physical and virtual base address of the kernel at
boot-time (CONFIG_RANDOMIZE_BASE), attacks needing kernel code will be
frustrated. Additionally, offsetting the module loading base address
means that even systems that load the same set of modules in the same
order every boot will not share a common base address with the rest of
the kernel text.

#### Stack base

If the base address of the kernel stack is not the same between processes,
or even not the same between syscalls, targets on or beyond the stack
become more difficult to locate.

#### Dynamic memory base

Much of the kernel's dynamic memory (e.g. kmalloc, vmalloc, etc) ends up
being relatively deterministic in layout due to the order of early-boot
initializations. If the base address of these areas is not the same
between boots, targeting them is frustrated, requiring an information
exposure specific to the region.

#### Structure layout

By performing a per-build randomization of the layout of sensitive
structures, attacks must either be tuned to known kernel builds or expose
enough kernel memory to determine structure layouts before manipulating
them.


## Preventing Information Exposures

Since the locations of sensitive structures are the primary target for
attacks, it is important to defend against exposure of both kernel memory
addresses and kernel memory contents (since they may contain kernel
addresses or other sensitive things like canary values).

### Unique identifiers

Kernel memory addresses must never be used as identifiers exposed to
userspace. Instead, use an atomic counter, an idr, or similar unique
identifier.

### Memory initialization

Memory copied to userspace must always be fully initialized. If not
explicitly memset(), this will require changes to the compiler to make
sure structure holes are cleared.

### Memory poisoning

When releasing memory, it is best to poison the contents (clear stack on
syscall return, wipe heap memory on a free), to avoid reuse attacks that
rely on the old contents of memory. This frustrates many uninitialized
variable attacks, stack content exposures, heap content exposures, and
use-after-free attacks.

### Destination tracking

To help kill classes of bugs that result in kernel addresses being
written to userspace, the destination of writes needs to be tracked. If
the buffer is destined for userspace (e.g. seq_file backed /proc files),
it should automatically censor sensitive values.