Probe UEFI compatibility without assuming a file object

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Loading a BSD-flavored static archive fails with AttributeError: 'ArchiveFileData' object has no attribute 'fileno'. Detection offers every stream to every default backend, and the loader promises a backend only read and seek, but UefiFirmware._to_bytes calls fileno() and catches only the io.UnsupportedOperation an io.IOBase would raise.

The probe has to answer for such a stream, so it now reads it instead, taking the mmap shortcut only when the stream spans the file behind the descriptor. It also returns bytes rather than a memoryview for a BytesIO — firmware unpacked by the CARTFile backend arrives that way, and uefi_firmware joins a compressed section's preamble to its body before decompressing it.

The regression test loads a real firmware volume and a real BSD archive, both fixtures added by angr/binaries#176, so the checks here stay red until that merges. Such an archive still does not load afterwards, because its symbol index is handed to the loader as if it were an object file; and #716 registers STM32Backend, whose is_compatible calls tell() and will need the same treatment.

Validation: #720 (comment)

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Validation record for head 200cf5d9f6bbc111e9116ccf272a3e39855dd8ba against baseline b58ea02a446106647cdaae32bdf91b7062404cc1. Commands run from the cle checkout root on CPython 3.12.13, with uefi-firmware 1.16 and arpy 1.1.1, and with a binaries checkout on the branch of angr/binaries#176. The Test jobs on this PR stay red until that merges, because CI checks out binaries master.

• Focused: pytest tests/test_uefi_firmware.py — 27 passed
• Regression: the same command with only cle/backends/uefi_firmware.py reverted to the baseline — 7 failed, 20 passed
• Baseline failures, all from the fileno() call in _to_bytes unless noted: AttributeError: 'ArchiveFileData' object has no attribute 'fileno' from is_compatible on an archive member and again from cle.Loader on the archive itself; AttributeError: 'MinimalStream' object has no attribute 'fileno' for a bare read/seek stream, both from is_compatible and from loading the volume that way; TypeError: unsupported operand type(s) for +: 'memoryview' and 'memoryview' at uefi_firmware/uefi.py:644 when the volume is given as a BytesIO; ValueError: cannot mmap an empty file from mmap.mmap for the empty file; and, for a stream over one archive member, the whole container returned rather than the member's bytes
• Pre-commit: all configured hooks, --all-files — passed, no file rewritten
• Merge-base comparison, the one the hosted Lint and Typecheck jobs make over changed files: pylint cle/backends/uefi_firmware.py 10.00 -> 10.00, tests/test_uefi_firmware.py 10.00 as a new file; pyright badness cle/backends/uefi_firmware.py 0.050 -> 0.048, tests/test_uefi_firmware.py 0.0 -> 0.0
• is_compatible was also offered 401 inputs derived from the volume — truncations at one percent steps, 200 single-bit flips of its first 128 KiB, and random and zero buffers — and returned without raising for every one
• At the previous head 2369d83, which carried this same production change under a test that assembled its own inputs: full suite pytest tests/ — 229 passed, 9 skipped; pylint cle/backends/uefi_firmware.py 10.00 -> 10.00; pyright badness 0.050 -> 0.048

Head 200cf5d repairs the Typecheck job, which head e030880 failed for a reason unrelated to the missing fixtures: the test read user_interface_name straight off child_objects, which is typed list[Backend], while that attribute lives on UefiModuleMixin. Every child a volume builds is a UefiPE or a UefiTE, so the test now asserts that and reads the name through the narrowed type, which also makes it say what a child of a firmware volume is.

Inputs, both added by angr/binaries#176: tests/aarch64/edk2_armvirtqemu.fd, Debian's edk2 build of the ArmVirtQemu platform, a single firmware volume whose 93 AArch64 PE drivers sit behind an LZMA-compressed section, so reaching them by name is what says the volume was parsed rather than only recognized; and tests/aarch64/bsd_symdef_archive.a, a BSD-flavored static archive whose first member is a __.SYMDEF SORTED symbol index and whose second is a Mach-O object, which gives _to_bytes a stream sharing a descriptor with a larger file. The failure was originally found in a corpus sweep, on real static archives produced by a BSD or Apple ar.

Caveats: the two assertions on _to_bytes call the private helper directly, because no public path can distinguish a stream that shares a descriptor with a larger file; everything else goes through cle.Loader or is_compatible. The mmap shortcut and the fallback are both covered on every platform CI runs, since the same volume is loaded as a path, a file object, a BytesIO and a bare stream.

[angr-bot]

Corpus decompilation diffs can be found at angr/dec-snapshots@master...angr/cle_720