path: root/butl/process

// file      : butl/process -*- C++ -*-
// copyright : Copyright (c) 2014-2016 Code Synthesis Ltd
// license   : MIT; see accompanying LICENSE file

#ifndef BUTL_PROCESS
#define BUTL_PROCESS

#ifndef _WIN32
#  include <sys/types.h> // pid_t
#endif

#include <cassert>
#include <cstdint> // uint32_t
#include <system_error>

#include <butl/path>
#include <butl/export>

namespace butl
{
  struct process_error: std::system_error
  {
    bool
    child () const {return child_;}

  public:

#ifndef _WIN32
    process_error (int e, bool child)
        : system_error (e, std::system_category ()), child_ (child) {}
#else
    process_error (int e, bool child = false)
        : system_error (e, std::system_category ()), child_ (child) {}

    process_error (const std::string& d, int e = ECHILD)
        : system_error (e, std::system_category (), d), child_ (false) {}
#endif

  private:
    bool child_;
  };

  // A process executable has three paths: initial, recall, and effective.
  // Initial is the original "command" that you specify in argv[0] and on
  // POSIX that's what ends up in the child's argv[0]. But not on Windows. On
  // Windows the command is first searched for in the parent executable's
  // directory and if found then that's what should end up in child's argv[0].
  // So this is the recall path. It is called recall because this is what the
  // caller of the parent process will be able to execute if you printed the
  // command line. Finally, effective is the actual path to the executable
  // that will include the directory part if found in PATH, the .exe extension
  // if one is missing, etc.
  //
  // As an example, let's say we run foo\foo.exe that itself spawns bar which
  // is found as foo\bar.exe. The paths will then be:
  //
  // initial:   bar
  // recall:    foo\bar
  // effective: foo\bar.exe
  //
  // In most cases, at least on POSIX, all three paths will be the same. As an
  // optimization, if the recall path is empty, then it means it is the same
  // as initial. Similarly, if the effective path is empty then, it is the
  // same as recall (and if that is empty, as initial).
  //
  // Note that the call to path_search() below adjust args[0] to point to the
  // recall path which brings up lifetime issues. To address this this class
  // also implements an RAII-based auto-restore of args[0] to its initial
  // value.
  //
  class process_path
  {
  public:
    const char* initial = nullptr;
    path recall;
    path effect;

    // Handle empty recall/effect.
    //
    const char* recall_string () const;
    const char* effect_string () const;

    bool empty () const
    {
      return initial == nullptr && recall.empty () && effect.empty ();
    }

    // Moveable-only type.
    //
    process_path (process_path&&);
    process_path& operator= (process_path&&);

    process_path (const process_path&) = delete;
    process_path& operator= (const process_path&) = delete;

    process_path () = default;
    process_path (const char* i, path&& r, path&& e);
    ~process_path ();

  private:
    friend class process;
    const char** args0_ = nullptr;
  };

  class LIBBUTL_EXPORT process
  {
  public:
    // Start another process using the specified command line. The default
    // values to the in, out and err arguments indicate that the child process
    // should inherit the parent process stdin, stdout, and stderr,
    // respectively. If -1 is passed instead, then the corresponding child
    // process descriptor is connected (via a pipe) to out_fd for stdin,
    // in_ofd for stdout, and in_efd for stderr (see data members below). If
    // -2 is passed, then the corresponding child process descriptor is
    // replaced with the null device descriptor (e.g., /dev/null). This
    // results in the child process not being able to read anything from stdin
    // (gets immediate EOF) and all data written to stdout/stderr being
    // discarded.
    //
    // On Windows parent process pipe descriptors are set to text mode to be
    // consistent with the default (text) mode of standard file descriptors of
    // the child process. When reading in the text mode the sequence of 0xD,
    // 0xA characters is translated into the single OxA character and 0x1A is
    // interpreted as EOF. When writing in the text mode the OxA character is
    // translated into the 0xD, 0xA sequence. Use the _setmode() function to
    // change the mode, if required.
    //
    // Instead of passing -1, -2 or the default value, you can also pass your
    // own descriptors. Note, however, that in this case they are not closed by
    // the parent. So you should do this yourself, if required.  For example,
    // to redirect the child process stdout to stderr, you can do:
    //
    // process p (..., 0, 2);
    //
    // Throw process_error if anything goes wrong. Note that some of the
    // exceptions (e.g., if exec() failed) can be thrown in the child
    // version of us.
    //
    // Note that the versions without the the process_path argument may
    // temporarily change args[0] (see path_search() for details).
    //
    process (const char* args[], int in = 0, int out = 1, int err = 2);

    process (const process_path&, const char* args[],
             int in = 0, int out = 1, int err = 2);

    // The "piping" constructor, for example:
    //
    // process lhs (..., 0, -1); // Redirect stdout to a pipe.
    // process rhs (..., lhs);   // Redirect stdin to lhs's pipe.
    //
    // rhs.wait (); // Wait for last first.
    // lhs.wait ();
    //
    process (const char* args[], process& in, int out = 1, int err = 2);

    process (const process_path&, const char* args[],
             process& in, int out = 1, int err = 2);

    // Versions of the above constructors that allow us to change the
    // current working directory of the child process. NULL and empty
    // cwd arguments are ignored.
    //
    process (const char* cwd, const char* [], int = 0, int = 1, int = 2);

    process (const char* cwd,
             const process_path&, const char* [],
             int = 0, int = 1, int = 2);

    process (const char* cwd, const char* [], process&, int = 1, int = 2);

    process (const char* cwd,
             const process_path&, const char* [],
             process&, int = 1, int = 2);

    // Wait for the process to terminate. Return true if the process
    // terminated normally and with the zero exit status. Unless ignore_error
    // is true, throw process_error if anything goes wrong. This function can
    // be called multiple times with subsequent calls simply returning the
    // status.
    //
    bool
    wait (bool ignore_errors = false);

    // Return true if the process has already terminated in which case
    // the argument is set to the result of wait().
    //
    bool
    try_wait (bool&);

    // Note that the destructor will wait for the process but will ignore
    // any errors and the exit status.
    //
    ~process () {if (handle != 0) wait (true);}

    // Moveable-only type.
    //
    process (process&&);
    process& operator= (process&&);

    process (const process&) = delete;
    process& operator= (const process&) = delete;

    // Create an empty or "already terminated" process. That is, handle is 0
    // and exit status is 0.
    //
    process ();

    // Resolve process' paths based on the initial path in args0. If recall
    // differs from initial, adjust args0 to point to the recall path. If
    // resolution fails, throw process_error. Normally, you will use this
    // function like this:
    //
    // const char* args[] = {"foo", ..., nullptr};
    //
    // process_path pp (process::path_search (args[0]))
    //
    // ... // E.g., print args[0].
    //
    // process p (pp, args);
    //
    // You can also specify the fallback directory which will be tried last.
    // This, for example, can be used to implement the Windows "search in the
    // parent executable's directory" semantics across platforms.
    //
    static process_path
    path_search (const char*& args0, const dir_path& fallback = dir_path ());

    // This version is primarily useful when you want to pre-search the
    // executable before creating the args[] array. In this case you will
    // use the recall path for args[0].
    //
    // The init argument determines whether to initialize the initial path to
    // the shallow copy of file. If it is true, then initial is the same as
    // file and recall is either empty or contain a different path. If it is
    // false then initial contains a shallow copy of recall, and recall is
    // either a different path or a deep copy of file. Normally you don't care
    // about initial once you got recall and the main reason to pass true to
    // this argument is to save a copy (since initial and recall are usually
    // the same).
    //
    static process_path
    path_search (const char* file, bool init, const dir_path& = dir_path ());

    static process_path
    path_search (const std::string&, bool, const dir_path& = dir_path ());

    static process_path
    path_search (const path&, bool, const dir_path& = dir_path ());

    // As above but if not found return empty process_path instead of
    // throwing.
    //
    static process_path
    try_path_search (const char*, bool, const dir_path& = dir_path ());

    static process_path
    try_path_search (const std::string&, bool, const dir_path& = dir_path ());

    static process_path
    try_path_search (const path&, bool, const dir_path& = dir_path ());

  public:
#ifndef _WIN32
    using handle_type = pid_t;
    using id_type = pid_t;
    using status_type = int;
#else
    using handle_type = void*;         // Win32 HANDLE
    using id_type = std::uint32_t;     // Win32 DWORD
    using status_type = std::uint32_t; // Win32 DWORD
#endif

    static id_type
    current_id ();

  public:
    handle_type handle;
    status_type status;

    int out_fd; // Write to this fd to send to the new process' stdin.
    int in_ofd; // Read from this fd to receive from the new process' stdout.
    int in_efd; // Read from this fd to receive from the new process' stderr.
  };
}

#include <butl/process.ixx>

#endif // BUTL_PROCESS