Address Sanitizer (ASAN) is a dynamic (run-time) sanitizer that checks for certain invalid memory accesses in C++ code - for example, buffer overruns, use-after-free, (stack) use-after-return/scope, and use of uninitialized globals. The sanitizer adds compiler-generated instrumentation before every data memory access and the runtime checks the validity of each access before it proceeds.
ASAN works by constructing a ‘shadow map’, a map with one byte per 8 bytes of kernel address space. Each shadow map byte tracks the validity of the kernel address it corresponds to - zero represents ‘valid for access’, non-zero bytes represent various sub-byte tracking or invalid states.
Allocators (for example the PMM) can invoke
asan_poison_shadow() to mark regions of memory in the physmap as “poisoned”, disallowing any data accesses to the region. They can also invoke
asan_unpoison_shadow() to mark regions of memory in the physmap as “unpoisoned”, allowing any data accesses to the region.
Poisoned Memory: kasan allows memory to be marked as either poisoned or unpoisoned. Memory accesses to poisoned memory result in kernel panics. Poisoning could be used in memory allocators to mark memory boundaries and to detect use-after-frees.
Redzone: memory allocators could add a small buffer before/after their allocations and poison it to detect buffer overflows. These buffers are called redzones.
Quarantine: Given that only memory accesses are checked, and memory can be reused (and thus, unpoisoned), increasing the time a memory region is poisoned allows more bugs to be detected. KASAN provides a way for allocators to hold off memory reuse, called quarantine. Instead of freeing memory right away, allocators can push pointers to a queue and free them in FIFO order.
Kernel ASAN is similar to userspace ASAN but has unique bootstrap and memory allocation requirements.
Note that any function that performs memory accesses outside of the kernel virtual address space has to be annotated with NO_ASAN, otherwise those accesses will result in a system crash.
When kASAN is enabled, all compiled kernel code is instrumented; so we need a valid shadow map very early in boot, before C code is called. Currently the x86-64 kernel has 512 GB of virtual addres space; KASAN requires 64 GB of shadow memory to track this entire region, corresponding to 1 byte per 8 bytes.
We create a shadow map at [-128GB ; -64GB) to cover all kernel virtual address space, and point every page of the shadow map to a single read-only zero page. One page table and one page directory are reused for all entries in the MMU, to save memory.
The shadow map is placed at [-128GB, -64GB) to avoid any potential mappings at the highest parts of the kernel address space. Currently the reallocated kernel is present there and it is foreseeable that other structures may be placed there for convenient access.
The map in x86_64 looks like this:
64 entries in pdp_hi (1GB each) point all to the same page directory (kasan_shadow_tables[512..1023]), with RW, NX and global permissions.
The kasan page directory has 512 entries pointing all to the same page table (kasan_shadow_tables), with RW, NX and global permissions.
The kasan page table has 512 entries pointing all to the same zero page (kasan_zero_page), with RO, NX and global permissions.
With this structure, all poison checks inside the kernel address space will succeed, as all shadow map memory is marked as unpoisoned.
In order to allow memory poisoning / tracking validity of kernel memory, asan needs to have writable pages backing portions of the shadow that cover the kernel physical map. These writable pages replace zero page mappings in parts of the shadow map that asan instruments.
During late boot, after PMM is initialized, we allocate a shadow page for every 8 pages of address space to instrument that contain at least one page of real memory. We do not consider MMIO regions, device memory, the ISA hole, etc. as real memory. We then replace the early boot zero page mappings with mappings to the newly allocated shadow pages. All the remaining early boot mappings remain the same.
We register the entire physmap and the kernel data/rodata/bss (sections with global variables) for instrumentation with asan during boot.
kasan for arm64 does nothing to setup the shadow during early boot. This means that fuchsia doesn't support kasan with inline instrumentation on arm64, as there is no shadow setup. Instead, the asan functions check for a global variable to signal when the shadow has been initialized.
After the VM subsystem is enabled, asan creates a vmar covering the entire shadow and maps it to the shadow address range. This vmar will be mostly unmapped. As the kernel creates memory mappings, asan will map memory pages to the corresponding parts of the shadow vmar, allowing reads and writes for poisoning and checks.
The two main differences with x86-64 are that we don't setup an early boot shadow, and that we use the memory subsystem to handle the shadow memory instead of doing manual page table manipulations.
This version of asan exposes an interface for callers to poison and check the validity of memory via the following functions:
Memory allocators should use
asan_poison_shadow to mark regions of memory as invalid, specifying different poison values for different types of memory. Allocators can use
asan_unpoison_shadow to mark regions of memory as valid for accesses.
The kernel heap is instrumented to poison metadata and free memory and unpoison allocations.
The kernel heap adds a ‘right-side redzone’ after every allocation and poisons it, to detect accesses past the end of a buffer. Heap metadata (before each allocation) is also poisoned and serves as a ‘left-side redzone’.
After an allocation is freed, it is kept in a ‘quarantine’ to delay its reuse. This improves detection of use-after-free errors.
The quarantine is implemented as a free-running circular queue, which stores up to kQuarantineElements pointers and frees them in FIFO order.
free, the memory to be deallocated is added to the quarantine, and once the queue is full, the oldest element is actually freed.
TODO(fxbug.dev/30033): kQuarantineElements is 65,536; this means that in the worst case it increases kernel heap memory usage by 256 MB (4K * 65536), which is the same as compiler-rt's default quarantine size. We could consider dynamically tuning this or having the quarantine not release memory until there is memory pressure.
Global variables are instrumented for out-of-bounds accesses.
When the kernel is compiled with kasan and global checking is enabled, a redzone is added to the right of every global object. Out-of-bounds accesses that hit the redzone are errors and are reported via the same mechanism as other out-of-bounds accesses.
ASAN can also instrument globals for initialization order bugs; we do not support that feature yet.