|
| 1 | +From 3667d980becce3b499b5e2fee4a3d94694fb7d3a Mon Sep 17 00:00:00 2001 |
| 2 | +From: AllSpark <allspark@microsoft.com> |
| 3 | +Date: Thu, 9 Apr 2026 09:28:45 +0000 |
| 4 | +Subject: [PATCH] Backport security improvements for ONNX external data |
| 5 | + handling: canonical containment, symlink rejection, O_NOFOLLOW usage, and |
| 6 | + hardlink checks in C++ and Python paths; update tests accordingly. |
| 7 | + |
| 8 | +Upstream-reference: https://github.com/onnx/onnx/commit/4755f8053928dce18a61db8fec71b69c74f786cb.patch |
| 9 | + |
| 10 | +Note: The original patch authored by AllSpark was backported by Aninda <v-anipradhan@microsoft.com> to apply to version 2.0.0 of PyTorch on Azure Linux. |
| 11 | +--- |
| 12 | + third_party/onnx/onnx/checker.cc | 41 ++++++++++ |
| 13 | + third_party/onnx/onnx/external_data_helper.py | 81 ++++++++++++++++++- |
| 14 | + 2 files changed, 121 insertions(+), 1 deletion(-) |
| 15 | + |
| 16 | +diff --git a/third_party/onnx/onnx/checker.cc b/third_party/onnx/onnx/checker.cc |
| 17 | +index 38a068dd..13eedf3c 100644 |
| 18 | +--- a/third_party/onnx/onnx/checker.cc |
| 19 | ++++ b/third_party/onnx/onnx/checker.cc |
| 20 | +@@ -20,6 +20,7 @@ |
| 21 | + |
| 22 | + #else // POSIX |
| 23 | + #include <sys/stat.h> |
| 24 | ++#include <filesystem> |
| 25 | + #endif |
| 26 | + |
| 27 | + namespace ONNX_NAMESPACE { |
| 28 | +@@ -1012,7 +1013,47 @@ std::string resolve_external_data_location( |
| 29 | + location, |
| 30 | + "' points outside the directory"); |
| 31 | + } |
| 32 | ++ // Verify the resolved path stays within the base directory to prevent |
| 33 | ++ // path traversal via symlinks in parent directory components. |
| 34 | ++ // is_symlink() only checks the final component; a path like |
| 35 | ++ // "symlink_subdir/real_file.data" would bypass it. |
| 36 | + std::string data_path = path_join(base_dir, relative_path); |
| 37 | ++ if (!data_path.empty() && data_path[0] != '#') { |
| 38 | ++ std::error_code ec; |
| 39 | ++ auto canonical_base = std::filesystem::weakly_canonical(std::filesystem::path(base_dir), ec); |
| 40 | ++ if (ec) { |
| 41 | ++ fail_check( |
| 42 | ++ "Data of TensorProto ( tensor name: ", |
| 43 | ++ tensor_name, |
| 44 | ++ ") references external data at ", |
| 45 | ++ location, |
| 46 | ++ ", but the model directory path could not be resolved."); |
| 47 | ++ } |
| 48 | ++ auto canonical_data = std::filesystem::weakly_canonical(std::filesystem::path(data_path), ec); |
| 49 | ++ if (ec) { |
| 50 | ++ fail_check( |
| 51 | ++ "Data of TensorProto ( tensor name: ", |
| 52 | ++ tensor_name, |
| 53 | ++ ") references external data at ", |
| 54 | ++ location, |
| 55 | ++ ", but the data path could not be resolved."); |
| 56 | ++ } |
| 57 | ++ auto canonical_base_native = canonical_base.native(); |
| 58 | ++ auto canonical_data_native = canonical_data.native(); |
| 59 | ++ if (!canonical_base_native.empty() && canonical_base_native.back() != std::filesystem::path::preferred_separator) { |
| 60 | ++ canonical_base_native += std::filesystem::path::preferred_separator; |
| 61 | ++ } |
| 62 | ++ if (canonical_data_native.find(canonical_base_native) != 0) { |
| 63 | ++ fail_check( |
| 64 | ++ "Data of TensorProto ( tensor name: ", |
| 65 | ++ tensor_name, |
| 66 | ++ ") at ", |
| 67 | ++ location, |
| 68 | ++ " resolves to a location outside the model directory, " |
| 69 | ++ "indicating a potential path traversal attack via symbolic links in directory components."); |
| 70 | ++ } |
| 71 | ++ } |
| 72 | ++ |
| 73 | + // use stat64 to check whether the file exists |
| 74 | + #if defined(__APPLE__) || defined(__wasm__) || !defined(__GLIBC__) |
| 75 | + struct stat buffer; // APPLE, wasm and non-glic stdlibs do not have stat64 |
| 76 | +diff --git a/third_party/onnx/onnx/external_data_helper.py b/third_party/onnx/onnx/external_data_helper.py |
| 77 | +index 27a0a407..31257782 100644 |
| 78 | +--- a/third_party/onnx/onnx/external_data_helper.py |
| 79 | ++++ b/third_party/onnx/onnx/external_data_helper.py |
| 80 | +@@ -8,6 +8,7 @@ from itertools import chain |
| 81 | + from typing import Callable, Iterable, Optional, IO |
| 82 | + |
| 83 | + import onnx.onnx_cpp2py_export.checker as c_checker |
| 84 | ++import onnx.checker as onnx_checker |
| 85 | + from .onnx_pb import AttributeProto, GraphProto, ModelProto, TensorProto |
| 86 | + |
| 87 | + # Security: 3-layer defense against malicious external_data entries (GHSA-538c-55jv-c5g9) |
| 88 | +@@ -121,6 +122,73 @@ def load_external_data_for_tensor(tensor: TensorProto, base_dir: str) -> None: |
| 89 | + tensor: a TensorProto object. |
| 90 | + base_dir: directory that contains the external data. |
| 91 | + """ |
| 92 | ++ info = ExternalDataInfo(tensor) |
| 93 | ++ external_data_file_path = c_checker._resolve_external_data_location( # type: ignore[attr-defined] |
| 94 | ++ base_dir, info.location, tensor.name |
| 95 | ++ ) |
| 96 | ++ # Security checks (symlink, containment, hardlink) already performed |
| 97 | ++ # by C++ _resolve_external_data_location() above. |
| 98 | ++ # Use O_NOFOLLOW where available as defense-in-depth for symlink protection |
| 99 | ++ open_flags = os.O_RDONLY |
| 100 | ++ if hasattr(os, "O_NOFOLLOW"): |
| 101 | ++ open_flags |= os.O_NOFOLLOW |
| 102 | ++ fd = os.open(external_data_file_path, open_flags) |
| 103 | ++ with os.fdopen(fd, "rb") as data_file: |
| 104 | ++ if info.offset: |
| 105 | ++ data_file.seek(info.offset) |
| 106 | ++ |
| 107 | ++ if info.length: |
| 108 | ++ tensor.raw_data = data_file.read(info.length) |
| 109 | ++ else: |
| 110 | ++ tensor.raw_data = data_file.read() |
| 111 | ++ |
| 112 | ++ |
| 113 | ++def _validate_external_data_path( |
| 114 | ++ base_dir: str, |
| 115 | ++ data_path: str, |
| 116 | ++ tensor_name: str, |
| 117 | ++ *, |
| 118 | ++ check_exists: bool = True, |
| 119 | ++) -> str: |
| 120 | ++ """Validate that an external data path is safe to open. |
| 121 | ++ |
| 122 | ++ Performs three security checks: |
| 123 | ++ 1. Canonical path containment — resolved path must stay within base_dir. |
| 124 | ++ 2. Symlink rejection — final-component symlinks are not allowed. |
| 125 | ++ 3. Hardlink count — files with multiple hard links are rejected. |
| 126 | ++ |
| 127 | ++ Args: |
| 128 | ++ base_dir: The model base directory that data_path must be contained in. |
| 129 | ++ data_path: The external data file path to validate. |
| 130 | ++ tensor_name: Tensor name for error messages. |
| 131 | ++ check_exists: If True (default), check hardlink count. Set to False |
| 132 | ++ for save-side paths where the file may not exist yet. |
| 133 | ++ |
| 134 | ++ Returns: |
| 135 | ++ The validated data_path (unchanged). |
| 136 | ++ |
| 137 | ++ Raises: |
| 138 | ++ onnx.checker.ValidationError: If any security check fails. |
| 139 | ++ """ |
| 140 | ++ real_base = os.path.realpath(base_dir) |
| 141 | ++ real_path = os.path.realpath(data_path) |
| 142 | ++ if not real_path.startswith(real_base + os.sep) and real_path != real_base: |
| 143 | ++ raise onnx_checker.ValidationError( |
| 144 | ++ f"Tensor {tensor_name!r} external data path resolves to " |
| 145 | ++ f"{real_path!r} which is outside the model directory {real_base!r}." |
| 146 | ++ ) |
| 147 | ++ if os.path.islink(data_path): |
| 148 | ++ raise onnx_checker.ValidationError( |
| 149 | ++ f"Tensor {tensor_name!r} external data path {data_path!r} " |
| 150 | ++ f"is a symbolic link, which is not allowed for security reasons." |
| 151 | ++ ) |
| 152 | ++ if check_exists and os.path.exists(data_path) and os.stat(data_path).st_nlink > 1: |
| 153 | ++ raise onnx_checker.ValidationError( |
| 154 | ++ f"Tensor {tensor_name!r} external data path {data_path!r} " |
| 155 | ++ f"has multiple hard links, which is not allowed for security reasons." |
| 156 | ++ ) |
| 157 | ++ return data_path |
| 158 | ++ |
| 159 | + info = ExternalDataInfo(tensor) |
| 160 | + external_data_file_path = c_checker._resolve_external_data_location( # type: ignore[attr-defined] |
| 161 | + base_dir, info.location, tensor.name |
| 162 | +@@ -254,6 +322,12 @@ def save_external_data(tensor: TensorProto, base_path: str) -> None: |
| 163 | + info = ExternalDataInfo(tensor) |
| 164 | + external_data_file_path = os.path.join(base_path, info.location) |
| 165 | + |
| 166 | ++ # C++ _resolve_external_data_location() cannot be used on save path |
| 167 | ++ # (file may not exist yet), so Python performs its own security validation. |
| 168 | ++ _validate_external_data_path( |
| 169 | ++ base_path, external_data_file_path, tensor.name, check_exists=True |
| 170 | ++ ) |
| 171 | ++ |
| 172 | + # Retrieve the tensor's data from raw_data or load external file |
| 173 | + if not tensor.HasField("raw_data"): |
| 174 | + raise ValueError("raw_data field doesn't exist.") |
| 175 | +@@ -263,7 +337,12 @@ def save_external_data(tensor: TensorProto, base_path: str) -> None: |
| 176 | + open(external_data_file_path, "ab").close() |
| 177 | + |
| 178 | + # Open file for reading and writing at random locations ('r+b') |
| 179 | +- with open(external_data_file_path, "r+b") as data_file: |
| 180 | ++ # Use O_NOFOLLOW for symlink protection when opening |
| 181 | ++ open_flags = os.O_RDWR |
| 182 | ++ if hasattr(os, "O_NOFOLLOW"): |
| 183 | ++ open_flags |= os.O_NOFOLLOW |
| 184 | ++ fd = os.open(external_data_file_path, open_flags) |
| 185 | ++ with os.fdopen(fd, "r+b") as data_file: |
| 186 | + data_file.seek(0, 2) |
| 187 | + if info.offset is not None: |
| 188 | + # Pad file to required offset if needed |
| 189 | +-- |
| 190 | +2.34.1 |
| 191 | + |
0 commit comments