CVE-2023-26489: wasmtime vulnerable to guest-controlled out-of-bounds read/write on x86_64

Wasmtime’s code generator, Cranelift, has a bug on x86_64 targets where address-mode computation mistakenly would calculate a 35-bit effective address instead of WebAssembly’s defined 33-bit effective address. This bug means that, with default codegen settings, a wasm-controlled load/store operation could read/write addresses up to 35 bits away from the base of linear memory. Wasmtime’s default sandbox settings provide up to 6G of protection from the base of linear memory to guarantee that any memory access in that range will be semantically correct. Due to this bug, however, addresses up to 0xffffffff * 8 + 0x7ffffffc = 36507222004 = ~34G bytes away from the base of linear memory are possible from guest code. This means that the virtual memory 6G away from the base of linear memory up to ~34G away can be read/written by a malicious module.

This out of bounds read/write is not semantically correct and poses a threat as an arbitrary read/write within ~34G of linear memory away from the base of a wasm module’s linear memory. A guest module can, without the knowledge of the embedder, read/write memory in this region. The memory may belong to other WebAssembly instances when using the pooling allocator, for example. The memory may also belong to the embedder, depending on address layout.

Embedders do not have a necessarily reliable means of detecting when this happens. Wasm loads/stores are allowed to cause machine segfaults meaning that an invalid read/write would be translated to a nominal WebAssembly trap. This means that a malicious module in the worst case silently reads/writes memory outside its bounds and in the “best” case looks like a normal “something trapped here” during its execution. This makes it difficult to retroactively determine whether this bug has been exploited on hosts. Affected embedders are recommended to analyze preexisting wasm modules to see if they’re affected by the incorrect codegen rules and possibly correlate that with an anomalous number of traps during historical execution to locate possibly suspicious modules.

The specific bug in Cranelift’s x86_64 backend is that a WebAssembly address which is left-shifted by a constant amount from 1 to 3 will get folded into x86_64’s addressing modes which perform shifts. For example (i32.load (i32.shl (local.get 0) (i32.const 3))) loads from the WebAssembly address $local0 << 3. When translated to Cranelift the $local0 << 3 computation, a 32-bit value, is zero-extended to a 64-bit value and then added to the base address of linear memory. Cranelift would generate an instruction of the form movl (%base, %local0, 8), %dst which calculates %base + %local0 << 3. The bug here, however, is that the address computation happens with 64-bit values, where the $local0 << 3 computation was supposed to be truncated to a 32-bit value. This means that %local0, which can use up to 32-bits for an address, gets 3 extra bits of address space to be accessible via this movl instruction.

The fix in Cranelift is to remove the erroneous lowering rules in the backend which handle these zero-extended expressions. The above example is then translated to movl %local0, %temp; shl $3, %temp; movl (%base, %temp), %dst which correctly truncates the intermediate computation of %local0 << 3 to 32-bits inside the %temp register which is then added to the %base value.