/libnvram-hypr

NVRAM emulator

Primary LanguageCMIT LicenseMIT

Introduction

This is a library that emulates the behavior of the NVRAM peripheral, by storing key-value pairs into a tmpfs mounted at MOUNT_POINT (default: /firmadyne/libnvram).

Upon first access of a NVRAM-related function, the emulated peripheral is initialized as follows:

  1. The NVRAM filesystem is mounted
  2. Built-in default values are loaded from the NVRAM_DEFAULTS macro
  3. A built-in list of system-specific default value sources, defined by the NVRAM_DEFAULTS_PATH macro, is iteratively checked and loaded. This includes native functions (NATIVE() macro), text files on the filesystem (PATH() macro), and loaded symbols of type char *[] (TABLE() macro).
  4. System-specific override values are loaded from keys located at OVERRIDE_POINT (default: /firmadyne/libnvram.override)

Due to differences in C runtime libraries on various firmware, this library is compiled as a dynamically-linked shared library. Since ELF dynamic loaders on Linux systems support lazy linking and global symbol resolution scope, resolution of external symbols used by this library are effectively delayed until after the calling process has already loaded the system C runtime library. As a result, this library behaves like a statically-linked shared library, while dynamically using functions loaded by the calling process, including those from the standard C runtime library.

Although compilation as a statically-linked shared library might appear to preferable, in practice we've had better results using our approach discussed above. Different C runtime libraries may have incompatible implementations of features such as thread-local storage. Additionally, many C runtime libraries are not built with position-independent code (PIC) to support static compilation.

Usage

Build the library, and copy it into the firmware image:

cp libnvram.so /firmadyne/libnvram.so

Create the NVRAM storage directories on the filesystem:

mkdir -p /firmadyne/libnvram/ mkdir -p /firmadyne/libnvram.override/

This will be automatically loaded by the instrumented kernel through LD_PRELOAD by modifications to init/main.c.

Notes

This library is alpha quality. In particular, memory safety of string operations has not been checked, there may be concurrency bugs, and it uses some nasty hacks. Additionally, various firmware have different calling semantics for NVRAM-related functions, which may not be supported by this library, or may be mutually incompatible with other devices.

Furthermore, this library is missing graceful fallbacks for library functions not available in certain C runtime libraries. This is because C runtime libraries only maintain source-level compatibility with one another, not binary-level compatibility. As a result, fallbacks need to be manually implemented by declaring library functions to be weak symbols and checking for initialization before calling them. Currently, uses of ftok() are implemented in this manner, meaning that semaphores are not used if this function is unavailable. This is not desirable, but preferred over a NULL pointer dereference.

As another example, stat() is not necessarily available at the binary level in all C runtime libraries. GNU glibc only exports stat() as a private weak symbol, linking instead to __xstat(), which is part of the Linux Standard Base (LSB) specification. However, uClibc instead links to __xstat_conv(), and does not provide a __xstat() function. Furthermore, most libraries will prefer stat64() over stat(), depending on whether Large File Support (LFS) is enabled, which results in additional indirection. Additionally, making a direct syscall to sys_stat() can be unpredictable due to changes in the system call interface. As an alternative, use access() or fopen().

Unfortunately, any of these failures typically cause the init process to crash during system startup, potentially with a NULL pointer dereference. This is because some firmware assume that NVRAM-related functions will never return NULL, making it difficult to debug. Resolving this typically requires manual reverse engineering to identify the crashing function, and modifying the NVRAM library by defining a default value, adding a source of default values, or writing a shim function to call into this library to perform a NVRAM-related operation.

Pull requests are greatly appreciated!