.\"-
.\" Copyright (c) 2021 The FreeBSD Foundation
.\"
.\" This documentation was written by Mark Johnston under sponsorship from
.\" the FreeBSD Foundation.
.\"
.\" Redistribution and use in source and binary forms, with or without
.\" modification, are permitted provided that the following conditions
.\" are met:
.\" 1. Redistributions of source code must retain the above copyright
.\"    notice, this list of conditions and the following disclaimer.
.\" 2. Redistributions in binary form must reproduce the above copyright
.\"    notice, this list of conditions and the following disclaimer in the
.\"    documentation and/or other materials provided with the distribution.
.\"
.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
.\" ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
.\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
.\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
.\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
.\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
.\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
.\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
.\" SUCH DAMAGE.
.\"
.Dd December 6, 2023
.Dt KMSAN 9
.Os
.Sh NAME
.Nm KMSAN
.Nd Kernel Memory SANitizer
.Sh SYNOPSIS
The
.Pa GENERIC-KMSAN
kernel configuration can be used to compile a KMSAN-enabled kernel using
.Pa GENERIC
as a base configuration.
Alternately, to compile KMSAN into the kernel, place the following line in your
kernel configuration file:
.Bd -ragged -offset indent
.Cd "options KMSAN"
.Ed
.Pp
.In sys/msan.h
.Ft void
.Fn kmsan_mark "const void *addr" "size_t size" "uint8_t code"
.Ft void
.Fn kmsan_orig "const void *addr" "size_t size" "int type" "uintptr_t pc"
.Ft void
.Fn kmsan_check "const void *addr" "size_t size" "const char *descr"
.Ft void
.Fn kmsan_check_bio "const struct bio *" "const char *descr"
.Ft void
.Fn kmsan_check_ccb "const union ccb *" "const char *descr"
.Ft void
.Fn kmsan_check_mbuf "const struct mbuf *" "const char *descr"
.Ft void
.Fn kmsan_check_uio "const struct uio *" "const char *descr"
.Sh DESCRIPTION
.Nm
is a subsystem which leverages compiler instrumentation to detect uses of
uninitialized memory in the kernel.
Currently it is implemented only on the amd64 platform.
.Pp
When
.Nm
is compiled into the kernel, the compiler is configured to emit function
calls preceding memory accesses.
The functions are implemented by the
.Nm
runtime component and use hidden, byte-granular shadow state to determine
whether the source operand has been initialized.
When uninitialized memory is used as a source operand in certain operations,
such as control flow expressions or memory accesses, the runtime reports
an error.
Otherwise, the shadow state is propagated to destination operand.
For example, a
variable assignment or a
.Fn memcpy
call which copies uninitialized memory will cause the destination buffer or
variable to be marked uninitialized.
.Pp
To report an error, the
.Nm
runtime will either trigger a kernel panic or print a message to the console,
depending on the value of the
.Sy debug.kmsan.panic_on_violation
sysctl.
In both cases, a stack trace and information about the origin of the
uninitialized memory is included.
.Pp
In addition to compiler-detected uses of uninitialized memory,
various kernel I/O
.Dq exit points ,
such as
.Xr copyout 9 ,
perform validation of the input's shadow state and will raise an error if
any uninitialized bytes are detected.
.Pp
The
.Nm
option imposes a significant performance penalty.
Kernel code typically runs two or three times slower, and each byte mapped in
the kernel map requires two bytes of shadow state.
As a result,
.Nm
should be used only for kernel testing and development.
It is not recommended to enable
.Nm
in systems with less than 8GB of physical RAM.
.Pp
The sanitizer in a KMSAN-configured kernel can be disabled by setting the loader
tunable
.Sy debug.kmsan.disable=1 .
.Sh FUNCTIONS
The
.Fn kmsan_mark
and
.Fn kmsan_orig
functions update
.Nm
shadow state.
.Fn kmsan_mark
marks an address range as valid or invalid according to the value of the
.Va code
parameter.
The valid values for this parameter are
.Dv KMSAN_STATE_INITED
and
.Dv KMSAN_STATE_UNINIT ,
which mark the range as initialized and uninitialized, respectively.
For example, when a piece of memory is freed to a kernel allocator, it will
typically have been marked initialized; before the memory is reused for a new
allocation, the allocator should mark it as uninitialized.
As another example, writes to host memory performed by devices, e.g., via DMA,
are not intercepted by the sanitizer; to avoid false positives, drivers should
mark device-written memory as initialized.
For many drivers this is handled internally by the
.Xr busdma 9
subsystem.
.Pp
The
.Fn kmsan_orig
function updates
.Dq origin
shadow state.
In particular, it associates a given uninitialized buffer with a memory type
and code address.
This is used by the
.Nm
runtime to track the source of uninitialized memory and is only for debugging
purposes.
See
.Sx IMPLEMENTATION NOTES
for more details.
.Pp
The
.Fn kmsan_check
function and its sub-typed siblings validate the shadow state of the region(s)
of kernel memory passed as input parameters.
If any byte of the input is marked as uninitialized, the runtime will generate
a report.
These functions are useful during debugging, as they can be strategically
inserted into code paths to narrow down the source of uninitialized memory.
They are also used to perform validation in various kernel I/O paths, helping
ensure that, for example, packets transmitted over a network do not contain
uninitialized kernel memory.
.Fn kmsan_check
and related functions also take a
.Fa descr
parameter which is inserted into any reports raised by the check.
.Sh IMPLEMENTATION NOTES
.Ss Shadow Maps
The
.Nm
runtime makes use of two shadows of the kernel map.
Each address in the kernel map has a linear mapping to addresses in the
two shadows.
The first, simply called the shadow map, tracks the state of the corresponding
kernel memory.
A non-zero byte in the shadow map indicates that the corresponding byte of
kernel memory is uninitialized.
The
.Nm
instrumentation automatically propagates shadow state as the contents of kernel
memory are transformed and copied.
.Pp
The second shadow is called the origin map, and exists only to help debug
reports from the sanitizer.
To avoid false positives,
.Nm
does not raise reports for certain operations on uninitialized memory, such
as copying or arithmetic.
Thus, operations on uninitialized state which raise a report may be far removed
from the source of the bug, complicating debugging.
The origin map contains information which can help pinpoint the root cause of
a particular
.Nm
report; when generating a report, the runtime uses state from the origin map
to provide extra details.
.Pp
Unlike the shadow map, the origin map is not byte-granular, but consists of 4-byte
.Dq cells .
Each cell describes the corresponding four bytes of mapped kernel memory and
holds a type and compressed code address.
When kernel memory is allocated for some purpose, its origin is initialized
either by the compiler instrumentation or by runtime hooks in the allocator.
The type indicates the specific allocator, e.g.,
.Xr uma 9 ,
and the address provides the location in the kernel code where the memory was
allocated.
.Ss Assembly Code
When
.Nm
is configured, the compiler will only emit instrumentation for C code.
Files containing assembly code are left un-instrumented.
In some cases this is handled by the sanitizer runtime, which defines
wrappers for subroutines implemented in assembly.
These wrappers are referred to as interceptors and handle updating
shadow state to reflect the operations performed by the original
subroutines.
In other cases, C code which calls assembly code or is called from
assembly code may need to use
.Fn kmsan_mark
to manually update shadow state.
This is typically only necessary in machine-dependent code.
.Pp
Inline assembly is instrumented by the compiler to update shadow state
based on the output operands of the code, and thus does not usually
require any special handling to avoid false positives.
.Ss Interrupts and Exceptions
In addition to the shadow maps, the sanitizer requires some thread-local
storage (TLS) to track initialization and origin state for function
parameters and return values.
The sanitizer instrumentation will automatically fetch, update and
verify this state.
In particular, this storage block has a layout defined by the sanitizer
ABI.
.Pp
Most kernel code runs in a context where interrupts or exceptions may
redirect the CPU to begin execution of unrelated code.
To ensure that thread-local sanitizer state remains consistent, the
runtime maintains a stack of TLS blocks for each thread.
When machine-dependent interrupt and exception handlers begin execution,
they push a new entry onto the stack before calling into any C code, and
pop the stack before resuming execution of the interrupted code.
These operations are performed by the
.Fn kmsan_intr_enter
and
.Fn kmsan_intr_leave
functions in the sanitizer runtime.
.Sh EXAMPLES
The following contrived example demonstrates some of the types of bugs
that are automatically detected by
.Nm :
.Bd -literal -offset indent
int
f(size_t osz)
{
	struct {
		uint32_t bar;
		uint16_t baz;
		/* A 2-byte hole is here. */
	} foo;
	char *buf;
	size_t sz;
	int error;

	/*
	 * This will raise a report since "sz" is uninitialized
	 * here.  If it is initialized, and "osz" was left uninitialized
	 * by the caller, a report would also be raised.
	 */
	if (sz < osz)
		return (1);

	buf = malloc(32, M_TEMP, M_WAITOK);

	/*
	 * This will raise a report since "buf" has not been
	 * initialized and contains whatever data is left over from the
	 * previous use of that memory.
	 */
	for (i = 0; i < 32; i++)
		if (buf[i] != '\0')
			foo.bar++;
	foo.baz = 0;

	/*
	 * This will raise a report since the pad bytes in "foo" have
	 * not been initialized, e.g., by memset(), and this call will
	 * thus copy uninitialized kernel stack memory into userspace.
	 */
	copyout(&foo, uaddr, sizeof(foo));

	/*
	 * This line itself will not raise a report, but may trigger
	 * a report in the caller depending on how the return value is
	 * used.
	 */
	return (error);
}
.Ed
.Sh SEE ALSO
.Xr build 7 ,
.Xr busdma 9 ,
.Xr copyout 9 ,
.Xr KASAN 9 ,
.Xr uio 9 ,
.Xr uma 9
.Rs
.%A Evgeniy Stepanov
.%A Konstantin Serebryany
.%T MemorySanitizer: fast detector of uninitialized memory use in C++
.%J 2015 IEEE/ACM International Symposium on Code Generation and Optimization (CGO)
.%D 2015
.Re
.Sh HISTORY
.Nm
was ported from
.Nx
and first appeared in
.Fx 14.0 .
.Sh BUGS
Accesses to kernel memory outside of the kernel map are ignored by the
.Nm
runtime.
In particular, memory accesses via the direct map are not validated.
When memory is copied from outside the kernel map into the kernel map,
that region of the kernel map is marked as initialized.
When
.Nm
is configured, kernel memory allocators are configured to use the kernel map,
and filesystems are configured to always map data buffers into the kernel map,
so usage of the direct map is minimized.
However, some uses of the direct map remain.
This is a conservative policy which aims to avoid false positives, but it will
mask bug in some kernel subsystems.
.Pp
On amd64, global variables and the physical page array
.Va vm_page_array
are not sanitized.
This is intentional, as it reduces memory usage by avoiding creating
shadows of large regions of the kernel map.
However, this can allow bugs to go undetected by
.Nm .
.Pp
Some kernel memory allocators provide type-stable objects, and code which uses
them frequently depends on object data being preserved across allocations.
Such allocations cannot be sanitized by
.Nm .
However, in some cases it may be possible to use
.Fn kmsan_mark
to manually annotate fields which are known to contain invalid data upon
allocation.