executor_runner segfaults in portable kernels when a .pte tensor has neither a data buffer nor a memory plan (null data pointer) #23955

🐛 Describe the bug

A .pte can declare a tensor with data_buffer_idx == 0 and no allocation_info. getTensorDataPtr (runtime/executor/tensor_parser_exec_aten.cpp:289-291 at main 897d28fe) accepts this case as "pointer received at runtime" and returns nullptr. If that tensor is not a plan input, nothing requires a data pointer to be supplied before Method::execute(): execute checks only input_set_ (method.cpp:1792-1799). The portable kernels then load or store through the null pointer (kernels/portable/cpu/util/dtype_util.h:20-27, elementwise_util.h:191), because Tensor::const_data_ptr / mutable_data_ptr (runtime/core/portable_type/tensor.h:135-153) do not check for null.

Reproduced on 2026-10-18 at main 897d28fe3094f27898115ebd65fecafdef5970cb (Linux x86_64, clang 21.1.8) with the shipped examples/portable/executor_runner --model_path <file>.pte. Seven fuzzer-generated .pte files (704 to 764 bytes each; all seven fault inside add_out) were run through three builds:

                     ASan+UBSan -O1   Release -O3 -DNDEBUG   Release + logging
7 crafted .pte       exit 1           exit 139 (SIGSEGV)     exit 139
upstream add model   exit 0           exit 0                 exit 0

UBSan reports, for example:

kernels/portable/cpu/util/elementwise_util.h:191:11: runtime error: store to null pointer of type 'float'
kernels/portable/cpu/util/dtype_util.h:21:26: runtime error: load of null pointer of type 'const float'
  ... torch::executor::native::add_out(...)
  ... executorch::runtime::Method::execute_instruction() runtime/executor/method.cpp:1557:7
  ... executorch::runtime::Method::execute() runtime/executor/method.cpp:1844:21
  ... main examples/portable/executor_runner/executor_runner.cpp:856:28

I found #19269, which proposed a null check in the six Tensor data accessors to cover a large class of internal fuzzer crashes. It was closed without merging, and the accessors at 897d28fe still have no check. This report adds a file-only path to the same state: the .pte alone produces the null-data tensor, through the stock runner's default flow. Possible fixes are to reject such a tensor at load unless it is a plan input or output, to require output pointers to be set before execute(), or to add the accessor check from #19269. I am happy to send a PR for whichever you prefer.

Two smaller null dereferences on the same malformed-input path, also reproduced at 897d28fe with the shipped runner:

• Method::execute_instruction formats op->overload()->c_str() in the "KernelCall failed" ET_LOG (method.cpp:1565-1573) without the has_overload guard that populate_operator_name uses at load time (:756-768). A .pte whose operator has no overload string, plus any kernel check failure, segfaults in an -O3 -DNDEBUG runner built with EXECUTORCH_ENABLE_LOGGING=ON (exit 139; ASan: SEGV at address 0x4 in strlen via vsnprintf from method.cpp:1565). The same .pte without the kernel failure runs normally, and a kernel failure with the overload present exits cleanly with the logged error.
• get_named_data (extension/flat_tensor/flat_tensor_data_map.cpp:66) and FlatTensorDataMap::get_key (:263) call key()->size() / key()->c_str() without a null check. A .ptd with a NamedData entry that omits key segfaults at address 0 in executor_runner --model_path ... --data_path ... (ASan: flat_tensor_data_map.cpp:66:50: runtime error: member call on null pointer of type 'flatbuffers::Vector<char>').

I have a small patch for these two, returning InvalidExternalData for a missing key and using "" for a missing overload. I can attach the crafted files, the controls, and the generator for the last two on request. Every run above recorded the SHA-256 of the binary and the input before and after.

Versions

main 897d28fe3094f27898115ebd65fecafdef5970cb (2026-10-18), built from source; Fedora 43 x86_64, clang 21.1.8, cmake 3.31.11. Release builds use -DCMAKE_BUILD_TYPE=Release (-O3 -DNDEBUG). Release v1.5.1 has the same source at these sites but was not run. (The template's collect_env.py output is not applicable to a C++ source build; paste it if the maintainers ask.)

Disclosure

AI tooling was used and assisted in the discovery and authoring of this issue.

cc @larryliu0820 @JacobSzwejbka @lucylq