pytorch - 💡(How to fix) Fix [JIT] Out-of-bounds read in MemoryReadAdapter::read() via crafted TorchScript model

// caffe2/serialize/in_memory_adapter.h, line ~20
size_t read(uint64_t pos, void* buf, size_t n, const char* what = "") const override {
  (void)what;
  memcpy(buf, (int8_t*)(data_) + pos, n);  // no bounds check
  return n;
}

---

torch::jit::load()
  → ScriptModuleDeserializer::deserialize()
    → PyTorchStreamReader::getRecord()
      → mz_zip_reader_extract_to_mem()
        → MemoryReadAdapter::read(pos, n)  // pos+n >> size_
          → memcpy OOB read → SEGV

---

UndefinedBehaviorSanitizer: SEGV on unknown address 0x00001ea55000
The signal is caused by a READ memory access.
  #0 __memcpy_avx512_unaligned_erms
  #1 caffe2::serialize::MemoryReadAdapter::read()
  #2 mz_zip_reader_extract_to_mem
  #3 PyTorchStreamReader::getRecord()
  ...
  #9 torch::jit::load()

---

import torch
torch.jit.load("poc-047-memcpy-oob.pt")
# → Segmentation fault

---

size_t read(uint64_t pos, void* buf, size_t n, const char* what = "") const override {
  (void)what;
  if (pos >= static_cast<uint64_t>(size_) ||
      n > static_cast<uint64_t>(size_) - pos) {
    CAFFE_THROW("Read past end of buffer: pos=", pos, " n=", n, " size=", size_);
  }
  memcpy(buf, (int8_t*)(data_) + pos, n);
  return n;
}