The Kernel Memory Sanitizer (KMSAN)

KMSAN is a dynamic error detector aimed at finding uses of uninitialized values. It is based on compiler instrumentation, and is quite similar to the userspace MemorySanitizer tool.

An important note is that KMSAN is not intended for production use, because it drastically increases kernel memory footprint and slows the whole system down.

Usage

Building the kernel

In order to build a kernel with KMSAN you will need a fresh Clang (14.0.6+). Please refer to LLVM documentation for the instructions on how to build Clang.

Now configure and build the kernel with CONFIG_KMSAN enabled.

Example report

Here is an example of a KMSAN report:

=====================================================
BUG: KMSAN: uninit-value in test_uninit_kmsan_check_memory+0x1be/0x380 [kmsan_test]
 test_uninit_kmsan_check_memory+0x1be/0x380 mm/kmsan/kmsan_test.c:273
 kunit_run_case_internal lib/kunit/test.c:333
 kunit_try_run_case+0x206/0x420 lib/kunit/test.c:374
 kunit_generic_run_threadfn_adapter+0x6d/0xc0 lib/kunit/try-catch.c:28
 kthread+0x721/0x850 kernel/kthread.c:327
 ret_from_fork+0x1f/0x30 ??:?

Uninit was stored to memory at:
 do_uninit_local_array+0xfa/0x110 mm/kmsan/kmsan_test.c:260
 test_uninit_kmsan_check_memory+0x1a2/0x380 mm/kmsan/kmsan_test.c:271
 kunit_run_case_internal lib/kunit/test.c:333
 kunit_try_run_case+0x206/0x420 lib/kunit/test.c:374
 kunit_generic_run_threadfn_adapter+0x6d/0xc0 lib/kunit/try-catch.c:28
 kthread+0x721/0x850 kernel/kthread.c:327
 ret_from_fork+0x1f/0x30 ??:?

Local variable uninit created at:
 do_uninit_local_array+0x4a/0x110 mm/kmsan/kmsan_test.c:256
 test_uninit_kmsan_check_memory+0x1a2/0x380 mm/kmsan/kmsan_test.c:271

Bytes 4-7 of 8 are uninitialized
Memory access of size 8 starts at ffff888083fe3da0

CPU: 0 PID: 6731 Comm: kunit_try_catch Tainted: G    B       E     5.16.0-rc3+ #104
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 04/01/2014
=====================================================

The report says that the local variable uninit was created uninitialized in do_uninit_local_array(). The third stack trace corresponds to the place where this variable was created.

The first stack trace shows where the uninit value was used (in test_uninit_kmsan_check_memory()). The tool shows the bytes which were left uninitialized in the local variable, as well as the stack where the value was copied to another memory location before use.

A use of uninitialized value v is reported by KMSAN in the following cases:

  • in a condition, e.g. if (v) { ... };

  • in an indexing or pointer dereferencing, e.g. array[v] or *v;

  • when it is copied to userspace or hardware, e.g. copy_to_user(..., &v, ...);

  • when it is passed as an argument to a function, and CONFIG_KMSAN_CHECK_PARAM_RETVAL is enabled (see below).

The mentioned cases (apart from copying data to userspace or hardware, which is a security issue) are considered undefined behavior from the C11 Standard point of view.

Disabling the instrumentation

A function can be marked with __no_kmsan_checks. Doing so makes KMSAN ignore uninitialized values in that function and mark its output as initialized. As a result, the user will not get KMSAN reports related to that function.

Another function attribute supported by KMSAN is __no_sanitize_memory. Applying this attribute to a function will result in KMSAN not instrumenting it, which can be helpful if we do not want the compiler to interfere with some low-level code (e.g. that marked with noinstr which implicitly adds __no_sanitize_memory).

This however comes at a cost: stack allocations from such functions will have incorrect shadow/origin values, likely leading to false positives. Functions called from non-instrumented code may also receive incorrect metadata for their parameters.

As a rule of thumb, avoid using __no_sanitize_memory explicitly.

It is also possible to disable KMSAN for a single file (e.g. main.o):

KMSAN_SANITIZE_main.o := n

or for the whole directory:

KMSAN_SANITIZE := n

in the Makefile. Think of this as applying __no_sanitize_memory to every function in the file or directory. Most users won’t need KMSAN_SANITIZE, unless their code gets broken by KMSAN (e.g. runs at early boot time).

Support

In order for KMSAN to work the kernel must be built with Clang, which so far is the only compiler that has KMSAN support. The kernel instrumentation pass is based on the userspace MemorySanitizer tool.

The runtime library only supports x86_64 at the moment.

How KMSAN works

KMSAN shadow memory

KMSAN associates a metadata byte (also called shadow byte) with every byte of kernel memory. A bit in the shadow byte is set iff the corresponding bit of the kernel memory byte is uninitialized. Marking the memory uninitialized (i.e. setting its shadow bytes to 0xff) is called poisoning, marking it initialized (setting the shadow bytes to 0x00) is called unpoisoning.

When a new variable is allocated on the stack, it is poisoned by default by instrumentation code inserted by the compiler (unless it is a stack variable that is immediately initialized). Any new heap allocation done without __GFP_ZERO is also poisoned.

Compiler instrumentation also tracks the shadow values as they are used along the code. When needed, instrumentation code invokes the runtime library in mm/kmsan/ to persist shadow values.

The shadow value of a basic or compound type is an array of bytes of the same length. When a constant value is written into memory, that memory is unpoisoned. When a value is read from memory, its shadow memory is also obtained and propagated into all the operations which use that value. For every instruction that takes one or more values the compiler generates code that calculates the shadow of the result depending on those values and their shadows.

Example:

int a = 0xff;  // i.e. 0x000000ff
int b;
int c = a | b;

In this case the shadow of a is 0, shadow of b is 0xffffffff, shadow of c is 0xffffff00. This means that the upper three bytes of c are uninitialized, while the lower byte is initialized.

Origin tracking

Every four bytes of kernel memory also have a so-called origin mapped to them. This origin describes the point in program execution at which the uninitialized value was created. Every origin is associated with either the full allocation stack (for heap-allocated memory), or the function containing the uninitialized variable (for locals).

When an uninitialized variable is allocated on stack or heap, a new origin value is created, and that variable’s origin is filled with that value. When a value is read from memory, its origin is also read and kept together with the shadow. For every instruction that takes one or more values, the origin of the result is one of the origins corresponding to any of the uninitialized inputs. If a poisoned value is written into memory, its origin is written to the corresponding storage as well.

Example 1:

int a = 42;
int b;
int c = a + b;

In this case the origin of b is generated upon function entry, and is stored to the origin of c right before the addition result is written into memory.

Several variables may share the same origin address, if they are stored in the same four-byte chunk. In this case every write to either variable updates the origin for all of them. We have to sacrifice precision in this case, because storing origins for individual bits (and even bytes) would be too costly.

Example 2:

int combine(short a, short b) {
  union ret_t {
    int i;
    short s[2];
  } ret;
  ret.s[0] = a;
  ret.s[1] = b;
  return ret.i;
}

If a is initialized and b is not, the shadow of the result would be 0xffff0000, and the origin of the result would be the origin of b. ret.s[0] would have the same origin, but it will never be used, because that variable is initialized.

If both function arguments are uninitialized, only the origin of the second argument is preserved.

Origin chaining

To ease debugging, KMSAN creates a new origin for every store of an uninitialized value to memory. The new origin references both its creation stack and the previous origin the value had. This may cause increased memory consumption, so we limit the length of origin chains in the runtime.

Clang instrumentation API

Clang instrumentation pass inserts calls to functions defined in mm/kmsan/nstrumentation.c into the kernel code.

Shadow manipulation

For every memory access the compiler emits a call to a function that returns a pair of pointers to the shadow and origin addresses of the given memory:

typedef struct {
  void *shadow, *origin;
} shadow_origin_ptr_t

shadow_origin_ptr_t __msan_metadata_ptr_for_load_{1,2,4,8}(void *addr)
shadow_origin_ptr_t __msan_metadata_ptr_for_store_{1,2,4,8}(void *addr)
shadow_origin_ptr_t __msan_metadata_ptr_for_load_n(void *addr, uintptr_t size)
shadow_origin_ptr_t __msan_metadata_ptr_for_store_n(void *addr, uintptr_t size)

The function name depends on the memory access size.

The compiler makes sure that for every loaded value its shadow and origin values are read from memory. When a value is stored to memory, its shadow and origin are also stored using the metadata pointers.

Handling locals

A special function is used to create a new origin value for a local variable and set the origin of that variable to that value:

void __msan_poison_alloca(void *addr, uintptr_t size, char *descr)

Access to per-task data

At the beginning of every instrumented function KMSAN inserts a call to __msan_get_context_state():

kmsan_context_state *__msan_get_context_state(void)

kmsan_context_state is declared in include/linux/kmsan.h:

struct kmsan_context_state {
  char param_tls[KMSAN_PARAM_SIZE];
  char retval_tls[KMSAN_RETVAL_SIZE];
  char va_arg_tls[KMSAN_PARAM_SIZE];
  char va_arg_origin_tls[KMSAN_PARAM_SIZE];
  u64 va_arg_overflow_size_tls;
  char param_origin_tls[KMSAN_PARAM_SIZE];
  depot_stack_handle_t retval_origin_tls;
};

This structure is used by KMSAN to pass parameter shadows and origins between instrumented functions (unless the parameters are checked immediately by CONFIG_KMSAN_CHECK_PARAM_RETVAL).

Passing uninitialized values to functions

Clang’s MemorySanitizer instrumentation has an option, -fsanitize-memory-param-retval, which makes the compiler check function parameters passed by value, as well as function return values.

The option is controlled by CONFIG_KMSAN_CHECK_PARAM_RETVAL, which is enabled by default to let KMSAN report uninitialized values earlier. Please refer to the LKML discussion for more details.

Because of the way the checks are implemented in LLVM (they are only applied to parameters marked as noundef), not all parameters are guaranteed to be checked, so we cannot give up the metadata storage in kmsan_context_state.

String functions

The compiler replaces calls to memcpy()/memmove()/memset() with the following functions. These functions are also called when data structures are initialized or copied, making sure shadow and origin values are copied alongside with the data:

void *__msan_memcpy(void *dst, void *src, uintptr_t n)
void *__msan_memmove(void *dst, void *src, uintptr_t n)
void *__msan_memset(void *dst, int c, uintptr_t n)

Error reporting

For each use of a value the compiler emits a shadow check that calls __msan_warning() in the case that value is poisoned:

void __msan_warning(u32 origin)

__msan_warning() causes KMSAN runtime to print an error report.

Inline assembly instrumentation

KMSAN instruments every inline assembly output with a call to:

void __msan_instrument_asm_store(void *addr, uintptr_t size)

, which unpoisons the memory region.

This approach may mask certain errors, but it also helps to avoid a lot of false positives in bitwise operations, atomics etc.

Sometimes the pointers passed into inline assembly do not point to valid memory. In such cases they are ignored at runtime.

Runtime library

The code is located in mm/kmsan/.

Per-task KMSAN state

Every task_struct has an associated KMSAN task state that holds the KMSAN context (see above) and a per-task flag disallowing KMSAN reports:

struct kmsan_context {
  ...
  bool allow_reporting;
  struct kmsan_context_state cstate;
  ...
}

struct task_struct {
  ...
  struct kmsan_context kmsan;
  ...
}

KMSAN contexts

When running in a kernel task context, KMSAN uses current->kmsan.cstate to hold the metadata for function parameters and return values.

But in the case the kernel is running in the interrupt, softirq or NMI context, where current is unavailable, KMSAN switches to per-cpu interrupt state:

DEFINE_PER_CPU(struct kmsan_ctx, kmsan_percpu_ctx);

Metadata allocation

There are several places in the kernel for which the metadata is stored.

1. Each struct page instance contains two pointers to its shadow and origin pages:

struct page {
  ...
  struct page *shadow, *origin;
  ...
};

At boot-time, the kernel allocates shadow and origin pages for every available kernel page. This is done quite late, when the kernel address space is already fragmented, so normal data pages may arbitrarily interleave with the metadata pages.

This means that in general for two contiguous memory pages their shadow/origin pages may not be contiguous. Consequently, if a memory access crosses the boundary of a memory block, accesses to shadow/origin memory may potentially corrupt other pages or read incorrect values from them.

In practice, contiguous memory pages returned by the same alloc_pages() call will have contiguous metadata, whereas if these pages belong to two different allocations their metadata pages can be fragmented.

For the kernel data (.data, .bss etc.) and percpu memory regions there also are no guarantees on metadata contiguity.

In the case __msan_metadata_ptr_for_XXX_YYY() hits the border between two pages with non-contiguous metadata, it returns pointers to fake shadow/origin regions:

char dummy_load_page[PAGE_SIZE] __attribute__((aligned(PAGE_SIZE)));
char dummy_store_page[PAGE_SIZE] __attribute__((aligned(PAGE_SIZE)));

dummy_load_page is zero-initialized, so reads from it always yield zeroes. All stores to dummy_store_page are ignored.

2. For vmalloc memory and modules, there is a direct mapping between the memory range, its shadow and origin. KMSAN reduces the vmalloc area by 3/4, making only the first quarter available to vmalloc(). The second quarter of the vmalloc area contains shadow memory for the first quarter, the third one holds the origins. A small part of the fourth quarter contains shadow and origins for the kernel modules. Please refer to arch/x86/include/asm/pgtable_64_types.h for more details.

When an array of pages is mapped into a contiguous virtual memory space, their shadow and origin pages are similarly mapped into contiguous regions.

References

E. Stepanov, K. Serebryany. MemorySanitizer: fast detector of uninitialized memory use in C++. In Proceedings of CGO 2015.