Skip to content

dtrugman/pfs

Repository files navigation

pfs

Production grade, very easy to use, procfs parsing library in C++. Used in production by S&P 500 tech companies and startups!

NEW Basic parsing of sysfs (Additional sysfs feature requests are welcome!)

Build

Build & Test

Run cmake . && make

Currently supported CMake configuration flags:

  • CMAKE_BUILD_TYPE=<Debug|Release>: Standard CMake flags to control build type (DEFAULT: Debug)
  • pfs_BUILD_SHARED_LIBS=<ON|OFF>: ON to compile a shared library. OFF to compile a static library (DEFAULT: Inherit BUILD_SHARE_LIBS, which is OFF by default))
  • pfs_BUILD_ASAN=<ON|OFF>: ON to enable address sanitizer (DEFAULT: OFF)
  • pfs_BUILD_SAMPLES=<ON|OFF>: ON to build the sample programs (DEFAULT: ON)
  • pfs_BUILD_TESTS=<ON|OFF>: ON to build the tests (DEFAULT: ON)

You can pass any number of those to the cmake command: cmake -D<CONFIG_FLAG>=<VALUE> .

NOTE: After running cmake for the first time, some values are cached in CMakeCache.txt and will not change when running cmake for a second time with different flags.

Build using clang

If you prefer using clang, just configure the compiler while running cmake:

CXX=<clang++> CC=<clang> cmake .

After that, just use make as always.

Integrate

  • Compile as a shared or static library.
  • Add the contents of /lib into your link directories
  • Add the contents of /include into your include directories. That's it, you are good to go.

Use CMake's find_package()

Option #1: make install (Preferred way)

After building the project, you can install it locally using make install. In your project's CMake file, you can then add the following snippet, and CMake will handle the rest:

find_package (pfs REQUIRED)
...
# Somewhere along the file you define your target
add_<library|executable> (<your-target> ...)
...
target_link_libraries (<your-target> pfs)

NOTE: CMake generates an install_manifest.txt file to track all the created files, this will help you uninstall the library if you need to do so.

Option #2: Without make install

Build the pfs project. No need to call make install. In your project's CMake file, you can then add the following snippet:

find_package (pfs REQUIRED)
...
# Somewhere along the file you define your target
add_<library|executable> (<your-target> ...)
...
target_link_libraries (<your-target> -L${pfs_LIBRARY_DIR} ${pfs_LIBRARIES})
target_include_directories (<your-target> [PUBLIC|PRIVATE] ${pfs_INCLUDE_DIRS})

Features

  • Parsing system-wide information from files directly under /procfs. See procfs.hpp for all the supported files.
  • Parsing per-task (processes and threads) information from files under /procfs/[task-id]/. See task.hpp for all the supported files.
  • Parsing network information from files under /procfs/net (which is an alias to /procfs/self/net nowadays)
  • NEW Parsing of basic disk information from sysfs/block (Additional sysfs feature requests are welcome!)

Requirements

  • The library requires C++11 or newer
  • The library aims to support Linux kernel versions >= 2.6.32.

Notes

General notes

  • All APIs and function calls might throw std::bad_alloc exceptions when allocations of standard containers such as std::string fail.
  • APIs are thread-safe. There are no internal states/members/caches that might be affected by simultaneous calls.
  • Objects do NOT handle data caching. All the APIs are pure getters that always(!) fetch the information from the filesystem.
  • The location of the procfs filesystem is configurable. Just create the procfs object with the right path for your machine.

Accessing inexisting tasks

If you call procfs().get_task(<id>) and that task doesn't really exist, the constructor will succeed.

Since tasks can die any time, instead of adding extra validation during construction, which might be confusing, the current design assumes the first call after the tasks died will fail.

Collecting thread information

There are two ways to collect information about a thread:

  1. /proc/<tid>
  2. /proc/<pid>/task/<tid> And the amazing fact is that they MIGHT provide different information.

For example, when accessing using the first path, the utime and stime values under stat represent the amount of time scheduled for the entire process(!), whereas when accessing using the second path, they represent the amount of time scheduled for that thread only.

How does that affect pfs?

  • When using procfs().get_task(<id>) you'll be accessing information using /proc/<tid>.
  • When using my_task = procfs().get_task(<pid>) and then my_task.get_task(<id>) OR my_task.get_tasks() you'll be accessing information through /proc/<pid>/task/<tid>

Samples

The directory sample contains a full blown application that calls all(!) the supported APIs and prints all the information gathered. When compiling the library, the sample applications is compiled as well.

You can find a basic implementations of netstat (see sample/tool_netstat.cpp) and lsmod (see sample/tool_lsmod.cpp) that you can easily reuse in your projects.

Anyway, here are some cool (and concise) examples:

Example 1: Find all the process that hold an open file descriptor to a specific file:

auto file = "/path/to/file";
auto pfs = pfs::procfs();
for (const auto& process : pfs.get_processes())
{
    for (const auto& thread : process.get_tasks())
    {
        for (const auto& fd : thread.get_fds())
        {
            if (fd.second.get_target() == file)
            {
                ... do something ...
            }
        }
    }
}

Note: This is pedantic implementation that takes into account the fact that a threads might not share the file descriptor with the process, see CLONE_FILES in clone(2)

Example 2: Iterate over all the IPv4 TCP sockets currently in listening state (in my current network namespace) and print their local port:

auto filter_listening = [](const net_socket& socket){
    return socket.socket_net_state == pfs::net_socket::net_state::listen;
}

for (auto& socket : pfs::procfs().get_net().get_tcp(filter_listening))
{
    std::cout << socket.local_port << std::endl;
}

Example 3: Find all the memory maps for task 1234 (This can be both a process or a thread) that start with an ELFs header

auto task = pfs::procfs().get_task(1234);
auto mem = task.get_mem();
for (auto& map : task.get_maps())
{
    if (!map.perm.can_read)
    {
        continue;
    }

    static const std::vector<uint8_t> ELF_HEADER = { 0x7F, 0x45, 0x4C, 0x46 };
    if (mem.read(map.start_address, ELF_HEADER.size()) == ELF_HEADER)
    {
        ... do something ...
    }
}

(You can either create pfs::procfs() every time or once and keep it, the overhead is really small)