aboutsummaryrefslogtreecommitdiff
path: root/libbutl/fdstream.hxx
diff options
context:
space:
mode:
Diffstat (limited to 'libbutl/fdstream.hxx')
-rw-r--r--libbutl/fdstream.hxx917
1 files changed, 917 insertions, 0 deletions
diff --git a/libbutl/fdstream.hxx b/libbutl/fdstream.hxx
new file mode 100644
index 0000000..4dc89a3
--- /dev/null
+++ b/libbutl/fdstream.hxx
@@ -0,0 +1,917 @@
+// file : libbutl/fdstream.hxx -*- C++ -*-
+// license : MIT; see accompanying LICENSE file
+
+#pragma once
+
+#include <ios> // streamsize
+#include <vector>
+#include <string>
+#include <chrono>
+#include <istream>
+#include <ostream>
+#include <utility> // move(), pair
+#include <cstdint> // uint16_t, uint64_t
+#include <cstddef> // size_t
+
+#include <libbutl/path.hxx>
+#include <libbutl/filesystem.hxx> // permissions, entry_stat
+#include <libbutl/small-vector.hxx>
+#include <libbutl/bufstreambuf.hxx>
+
+#include <libbutl/export.hxx>
+
+namespace butl
+{
+ // RAII type for file descriptors. Note that failure to close the descriptor
+ // is silently ignored by both the destructor and reset().
+ //
+ // The descriptor can be negative. Such a descriptor is treated as unopened
+ // and is not closed.
+ //
+ struct nullfd_t
+ {
+ constexpr explicit nullfd_t (int) {}
+ constexpr operator int () const {return -1;}
+ };
+
+ constexpr nullfd_t nullfd (-1);
+
+ class LIBBUTL_SYMEXPORT auto_fd
+ {
+ public:
+ auto_fd (nullfd_t = nullfd) noexcept: fd_ (-1) {}
+
+ explicit
+ auto_fd (int fd) noexcept: fd_ (fd) {}
+
+ auto_fd (auto_fd&& fd) noexcept: fd_ (fd.release ()) {}
+ auto_fd& operator= (auto_fd&&) noexcept;
+
+ auto_fd (const auto_fd&) = delete;
+ auto_fd& operator= (const auto_fd&) = delete;
+
+ ~auto_fd () noexcept;
+
+ int
+ get () const noexcept {return fd_;}
+
+ void
+ reset (int fd = -1) noexcept;
+
+ int
+ release () noexcept
+ {
+ int r (fd_);
+ fd_ = -1;
+ return r;
+ }
+
+ // Close an open file descriptor. Throw ios::failure on the underlying OS
+ // error. Reset the descriptor to -1 whether the exception is thrown or
+ // not.
+ //
+ void
+ close ();
+
+ private:
+ int fd_;
+ };
+
+ inline bool
+ operator== (const auto_fd& x, const auto_fd& y)
+ {
+ return x.get () == y.get ();
+ }
+
+ inline bool
+ operator!= (const auto_fd& x, const auto_fd& y)
+ {
+ return !(x == y);
+ }
+
+ inline bool
+ operator== (const auto_fd& x, nullfd_t)
+ {
+ return x.get () == -1;
+ }
+
+ inline bool
+ operator!= (const auto_fd& x, nullfd_t y)
+ {
+ return !(x == y);
+ }
+
+ // An [io]fstream that can be initialized with a file descriptor in addition
+ // to a file name and that also by default enables exceptions on badbit and
+ // failbit. So instead of a dance like this:
+ //
+ // ifstream ifs;
+ // ifs.exceptions (ifstream::badbit | ifstream::failbit);
+ // ifs.open (path.string ());
+ //
+ // You can simply do:
+ //
+ // ifdstream ifs (path);
+ //
+ // Notes and limitations:
+ //
+ // - char only
+ // - input or output but not both (can use a union of two streams for that)
+ // - no support for put back
+ // - use of tell[gp]() and seek[gp]() is discouraged on Windows for
+ // fdstreams opened in the text mode (see fdstreambuf::seekoff()
+ // implementation for reasoning and consider using non-standard tellg()
+ // and seekg() in fdstreambuf, instead)
+ // - non-blocking file descriptor is supported only by showmanyc() function
+ // and only for pipes on Windows, in contrast to POSIX systems
+ // - throws ios::failure in case of open(), read(), write(), close(),
+ // seek[gp](), or tell[gp]() errors
+ // - exception mask has at least badbit
+ // - after catching an exception caused by badbit the stream is no longer
+ // usable
+ // - not movable, though can be easily supported (or not: there is no move
+ // constructor for istream/ostream in GCC 4.9)
+ // - passing to constructor auto_fd with a negative file descriptor is valid
+ // and results in the creation of an unopened object
+ //
+ class LIBBUTL_SYMEXPORT fdstreambuf: public bufstreambuf
+ {
+ public:
+ fdstreambuf () = default;
+
+ // Unless specified, the current read/write position is assumed to
+ // be 0 (note: not queried).
+ //
+ fdstreambuf (auto_fd&&, std::uint64_t pos = 0);
+
+ // Before we invented auto_fd into fdstreams we keept fdstreambuf opened
+ // on faulty close attempt. Now fdstreambuf is always closed by close()
+ // function. This semantics change seems to be the right one as there is
+ // no reason to expect fdclose() to succeed after it has already failed
+ // once.
+ //
+ void
+ close () {fd_.close ();}
+
+ auto_fd
+ release ();
+
+ void
+ open (auto_fd&&, std::uint64_t pos = 0);
+
+ bool
+ is_open () const {return fd_.get () >= 0;}
+
+ int
+ fd () const {return fd_.get ();}
+
+ // Set the file descriptor blocking mode returning the previous mode on
+ // success and throwing ios::failure otherwise (see fdmode() for details).
+ //
+ // Note that besides calling fdmode(fd()), this function also updating its
+ // internal state according to the new mode.
+ //
+ bool
+ blocking (bool);
+
+ public:
+ using base = bufstreambuf;
+
+ // basic_streambuf input interface.
+ //
+ public:
+ virtual std::streamsize
+ showmanyc ();
+
+ virtual int_type
+ underflow ();
+
+ // Direct access to the get area. Use with caution.
+ //
+ using base::gptr;
+ using base::egptr;
+ using base::gbump;
+
+ // Return the (logical) position of the next byte to be read.
+ //
+ using base::tellg;
+
+ // Seek to the (logical) position as if by reading the specified number of
+ // bytes from the beginning of the stream. Throw ios::failure on the
+ // underlying OS errors.
+ //
+ void
+ seekg (std::uint64_t);
+
+ private:
+ bool
+ load ();
+
+ // basic_streambuf output interface.
+ //
+ public:
+ virtual int_type
+ overflow (int_type);
+
+ virtual int
+ sync ();
+
+ virtual std::streamsize
+ xsputn (const char_type*, std::streamsize);
+
+ // Return the (logical) position of the next byte to be written.
+ //
+ using base::tellp;
+
+ // basic_streambuf positioning interface (both input/output).
+ //
+ public:
+ virtual pos_type
+ seekpos (pos_type, std::ios_base::openmode);
+
+ virtual pos_type
+ seekoff (off_type, std::ios_base::seekdir, std::ios_base::openmode);
+
+ private:
+ bool
+ save ();
+
+ private:
+ auto_fd fd_;
+ char buf_[8192];
+ bool non_blocking_ = false;
+ };
+
+ // File stream mode.
+ //
+ // The text/binary flags have the same semantics as those in std::fstream.
+ // Specifically, this is a noop for POSIX systems where the two modes are
+ // the same. On Windows, 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.
+ //
+ // The skip flag instructs the stream to skip to the end before closing the
+ // file descriptor. This is primarily useful when working with pipes where
+ // you may want not to "offend" the other end by closing your end before
+ // reading all the data.
+ //
+ // The blocking/non_blocking flags determine whether the IO operation should
+ // block or return control if currently there is no data to read or no room
+ // to write. Only the istream::readsome() function supports the semantics of
+ // non-blocking operations. In contrast to POSIX systems, we only support
+ // this for pipes on Windows, always assuming the blocking mode for other
+ // file descriptors. IO stream operations other than readsome() are illegal
+ // in the non-blocking mode and result in the badbit being set (note that
+ // it is not the more appropriate failbit for implementation reasons).
+ //
+ enum class fdstream_mode: std::uint16_t
+ {
+ text = 0x01,
+ binary = 0x02,
+ skip = 0x04,
+ blocking = 0x08,
+ non_blocking = 0x10
+ };
+
+ inline fdstream_mode operator& (fdstream_mode, fdstream_mode);
+ inline fdstream_mode operator| (fdstream_mode, fdstream_mode);
+ inline fdstream_mode operator&= (fdstream_mode&, fdstream_mode);
+ inline fdstream_mode operator|= (fdstream_mode&, fdstream_mode);
+
+ // Extended (compared to ios::openmode) file open flags.
+ //
+ enum class fdopen_mode: std::uint16_t
+ {
+ in = 0x01, // Open for reading.
+ out = 0x02, // Open for writing.
+ append = 0x04, // Seek to the end of file before each write.
+ truncate = 0x08, // Discard the file contents on open.
+ create = 0x10, // Create a file if not exists.
+ exclusive = 0x20, // Fail if the file exists and the create flag is set.
+ binary = 0x40, // Set binary translation mode.
+ at_end = 0x80, // Seek to the end of stream immediately after open.
+
+ none = 0 // Usefull when building the mode incrementally.
+ };
+
+ inline fdopen_mode operator& (fdopen_mode, fdopen_mode);
+ inline fdopen_mode operator| (fdopen_mode, fdopen_mode);
+ inline fdopen_mode operator&= (fdopen_mode&, fdopen_mode);
+ inline fdopen_mode operator|= (fdopen_mode&, fdopen_mode);
+
+ class LIBBUTL_SYMEXPORT fdstream_base
+ {
+ protected:
+ fdstream_base () = default;
+ fdstream_base (auto_fd&&, std::uint64_t pos);
+ fdstream_base (auto_fd&&, fdstream_mode, std::uint64_t pos);
+
+ public:
+ int
+ fd () const {return buf_.fd ();}
+
+ protected:
+ fdstreambuf buf_;
+ };
+
+ // iofdstream constructors and open() functions that take openmode as an
+ // argument mimic the corresponding iofstream functions in terms of the
+ // openmode mask interpretation. They throw std::invalid_argument for an
+ // invalid combination of flags (as per the standard). Note that the in and
+ // out flags are always added implicitly for ifdstream and ofdstream,
+ // respectively.
+ //
+ // iofdstream constructors and open() functions that take fdopen_mode as an
+ // argument interpret the mask literally just ignoring some flags which are
+ // meaningless in the absense of others (read more on that in the comment
+ // for fdopen()). Note that the in and out flags are always added implicitly
+ // for ifdstream and ofdstream, respectively.
+ //
+ // iofdstream constructors and open() functions that take file path as a
+ // const std::string& or const char* may throw the invalid_path exception.
+ //
+ // Passing auto_fd with a negative file descriptor is valid and results in
+ // the creation of an unopened object.
+ //
+ // Also note that open() and close() functions can be successfully called
+ // for an opened and unopened objects respectively. That is in contrast with
+ // iofstream that sets failbit in such cases.
+ //
+
+ // Note that ifdstream destructor will close an open file descriptor but
+ // will ignore any errors. To detect such errors, call close() explicitly.
+ //
+ // This is a sample usage of iofdstreams with process. Note that here it is
+ // expected that the child process reads from STDIN first and writes to
+ // STDOUT afterwards.
+ //
+ // try
+ // {
+ // process pr (args, -1, -1);
+ //
+ // try
+ // {
+ // // In case of exception, skip and close input after output.
+ // //
+ // ifdstream is (move (pr.in_ofd), fdstream_mode::skip);
+ // ofdstream os (move (pr.out_fd));
+ //
+ // // Write.
+ //
+ // os.close (); // Don't block the other end.
+ //
+ // // Read.
+ //
+ // is.close (); // Skip till end and close.
+ //
+ // if (pr.wait ())
+ // {
+ // return ...; // Good.
+ // }
+ //
+ // // Non-zero exit, diagnostics presumably issued, fall through.
+ // }
+ // catch (const failure&)
+ // {
+ // // IO failure, child exit status doesn't matter. Just wait for the
+ // // process completion and fall through.
+ // //
+ // // Note that this is optional if the process_error handler simply
+ // // falls through since process destructor will wait (but will ignore
+ // // any errors).
+ // //
+ // pr.wait ();
+ // }
+ //
+ // error << .... ;
+ //
+ // // Fall through.
+ // }
+ // catch (const process_error& e)
+ // {
+ // error << ... << e;
+ //
+ // if (e.child ())
+ // exit (1);
+ //
+ // // Fall through.
+ // }
+ //
+ // throw failed ();
+ //
+ class LIBBUTL_SYMEXPORT ifdstream: public fdstream_base, public std::istream
+ {
+ public:
+ // Create an unopened object.
+ //
+ explicit
+ ifdstream (iostate = badbit | failbit);
+
+ explicit
+ ifdstream (auto_fd&&,
+ iostate = badbit | failbit,
+ std::uint64_t pos = 0);
+
+ ifdstream (auto_fd&&,
+ fdstream_mode m,
+ iostate = badbit | failbit,
+ std::uint64_t pos = 0);
+
+ explicit
+ ifdstream (const char*,
+ iostate = badbit | failbit);
+
+ explicit
+ ifdstream (const std::string&,
+ iostate = badbit | failbit);
+
+ explicit
+ ifdstream (const path&,
+ iostate = badbit | failbit);
+
+ // @@ In some implementations (for example, MSVC), iostate and openmode
+ // (and/or their respective constants) are not distinct enough which
+ // causes overload resolution errors.
+ //
+ ifdstream (const char*,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ifdstream (const std::string&,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ifdstream (const path&,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ifdstream (const char*,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ifdstream (const std::string&,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ifdstream (const path&,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ~ifdstream () override;
+
+ void
+ open (const char*, openmode = in);
+
+ void
+ open (const std::string&, openmode = in);
+
+ void
+ open (const path&, openmode = in);
+
+ void
+ open (const char*, fdopen_mode);
+
+ void
+ open (const std::string&, fdopen_mode);
+
+ void
+ open (const path&, fdopen_mode);
+
+ void
+ open (auto_fd&& fd, std::uint64_t pos = 0)
+ {
+ buf_.open (std::move (fd), pos);
+ clear ();
+ }
+
+ void
+ open (auto_fd&& fd, fdstream_mode m, std::uint64_t pos = 0);
+
+ void close ();
+ auto_fd release (); // Note: no skipping.
+ bool is_open () const {return buf_.is_open ();}
+
+ // Read the textual stream. The stream is supposed not to contain the null
+ // character.
+ //
+ std::string
+ read_text ();
+
+ // Read the binary stream.
+ //
+ std::vector<char>
+ read_binary ();
+
+ private:
+ bool skip_ = false;
+ };
+
+ // Note that ofdstream requires that you explicitly call close() before
+ // destroying it. Or, more specifically, the ofdstream object should not be
+ // in the opened state by the time its destructor is called, unless it is in
+ // the "not good" state (good() == false) or the destructor is being called
+ // during the stack unwinding due to an exception being thrown
+ // (std::uncaught_exception() == true). This is enforced with assert() in
+ // the ofdstream destructor.
+ //
+ class LIBBUTL_SYMEXPORT ofdstream: public fdstream_base, public std::ostream
+ {
+ public:
+ // Create an unopened object.
+ //
+ explicit
+ ofdstream (iostate = badbit | failbit);
+
+ explicit
+ ofdstream (auto_fd&&,
+ iostate = badbit | failbit,
+ std::uint64_t pos = 0);
+
+ ofdstream (auto_fd&&,
+ fdstream_mode m,
+ iostate = badbit | failbit,
+ std::uint64_t pos = 0);
+
+ explicit
+ ofdstream (const char*,
+ iostate = badbit | failbit);
+
+ explicit
+ ofdstream (const std::string&,
+ iostate = badbit | failbit);
+
+ explicit
+ ofdstream (const path&,
+ iostate = badbit | failbit);
+
+ // @@ In some implementations (for example, MSVC), iostate and openmode
+ // (and/or their respective constants) are not distinct enough which
+ // causes overload resolution errors.
+ //
+ ofdstream (const char*,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ofdstream (const std::string&,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ofdstream (const path&,
+ openmode,
+ iostate /*= badbit | failbit*/);
+
+ ofdstream (const char*,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ofdstream (const std::string&,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ofdstream (const path&,
+ fdopen_mode,
+ iostate = badbit | failbit);
+
+ ~ofdstream () override;
+
+ void
+ open (const char*, openmode = out);
+
+ void
+ open (const std::string&, openmode = out);
+
+ void
+ open (const path&, openmode = out);
+
+ void
+ open (const char*, fdopen_mode);
+
+ void
+ open (const std::string&, fdopen_mode);
+
+ void
+ open (const path&, fdopen_mode);
+
+ void
+ open (auto_fd&& fd, std::uint64_t pos = 0)
+ {
+ buf_.open (std::move (fd), pos);
+ clear ();
+ }
+
+ void close () {if (is_open ()) flush (); buf_.close ();}
+ auto_fd release ();
+ bool is_open () const {return buf_.is_open ();}
+ };
+
+ // Open a file or, if the file name is `-`, stdin/stdout.
+ //
+ // In case of the stdin/stdout, these functions simply adjust the exception
+ // mask on std::cin/cout to match the i/ofdstreams argument.
+ //
+ // Return a reference to the opened i/ofdstream or cin/cout and, in the
+ // latter case, set the translated name in path_name to <stdin>/<stdout>
+ // (unless it is already present).
+ //
+ // Note that ofdstream::close() should be called explicitly unless stdout
+ // was opened (but harmless to call even if it was).
+ //
+ LIBBUTL_SYMEXPORT std::istream&
+ open_file_or_stdin (path_name&, ifdstream&);
+
+ LIBBUTL_SYMEXPORT std::ostream&
+ open_file_or_stdout (path_name&, ofdstream&);
+
+ // The std::getline() replacement that provides a workaround for libstdc++'s
+ // ios::failure ABI fiasco (#66145) by throwing ios::failure, as it is
+ // defined at libbutl build time (new ABI on recent distributions) rather
+ // than libstdc++ build time (still old ABI on most distributions).
+ //
+ // Notes:
+ //
+ // - This relies of ADL so if the stream is used via the std::istream
+ // interface, then std::getline() will still be used. To put it another
+ // way, this is "the best we can do" until GCC folks get their act
+ // together.
+ //
+ // - The fail and eof bits may be left cleared in the stream exception mask
+ // when the function throws because of badbit.
+ //
+ LIBBUTL_SYMEXPORT ifdstream&
+ getline (ifdstream&, std::string&, char delim = '\n');
+
+ // Open a file returning an auto_fd that holds its file descriptor on
+ // success and throwing ios::failure otherwise.
+ //
+ // The mode argument should have at least one of the in or out flags set.
+ // The append and truncate flags are meaningless in the absense of the out
+ // flag and are ignored without it. The exclusive flag is meaningless in the
+ // absense of the create flag and is ignored without it. Note also that if
+ // the exclusive flag is specified then a dangling symbolic link is treated
+ // as an existing file.
+ //
+ // The permissions argument is taken into account only if the file is
+ // created. Note also that permissions can be adjusted while being set in a
+ // way specific for the OS. On POSIX systems they are modified with the
+ // process' umask, so effective permissions are permissions & ~umask. On
+ // Windows permissions other than ru and wu are unlikelly to have effect.
+ //
+ // Also note that on POSIX the FD_CLOEXEC flag is set for the file descriptor
+ // to prevent its leakage into child processes. On Windows, for the same
+ // purpose, the _O_NOINHERIT flag is set. Note that the process class, that
+ // passes such a descriptor to the child, makes it inheritable for a while.
+ //
+ LIBBUTL_SYMEXPORT auto_fd
+ fdopen (const char*,
+ fdopen_mode,
+ permissions = permissions::ru | permissions::wu |
+ permissions::rg | permissions::wg |
+ permissions::ro | permissions::wo);
+
+ LIBBUTL_SYMEXPORT auto_fd
+ fdopen (const std::string&,
+ fdopen_mode,
+ permissions = permissions::ru | permissions::wu |
+ permissions::rg | permissions::wg |
+ permissions::ro | permissions::wo);
+
+ LIBBUTL_SYMEXPORT auto_fd
+ fdopen (const path&,
+ fdopen_mode,
+ permissions = permissions::ru | permissions::wu |
+ permissions::rg | permissions::wg |
+ permissions::ro | permissions::wo);
+
+ // Duplicate an open file descriptor. Throw ios::failure on the underlying
+ // OS error.
+ //
+ // Note that on POSIX the FD_CLOEXEC flag is set for the new descriptor if it
+ // is present for the source one. That's in contrast to POSIX dup() that
+ // doesn't copy file descriptor flags. Also note that duplicating descriptor
+ // and setting the flag is not an atomic operation generally, but it is in
+ // regards to child process spawning (to prevent file descriptor leakage into
+ // a child process).
+ //
+ // Note that on Windows the _O_NOINHERIT flag is set for the new descriptor
+ // if it is present for the source one. That's in contrast to Windows _dup()
+ // that doesn't copy the flag. Also note that duplicating descriptor and
+ // setting the flag is not an atomic operation generally, but it is in
+ // regards to child process spawning (to prevent file descriptor leakage into
+ // a child process).
+ //
+ LIBBUTL_SYMEXPORT auto_fd
+ fddup (int fd);
+
+ // Set the translation and/or blocking modes for the file descriptor. Throw
+ // invalid_argument for an invalid combination of flags. Return the previous
+ // mode on success, throw ios::failure otherwise.
+ //
+ // The text and binary flags are mutually exclusive on Windows. On POSIX
+ // system the two modes are the same and so no check is performed.
+ //
+ // The blocking and non-blocking flags are mutually exclusive. In contrast
+ // to POSIX systems, on Windows the non-blocking mode is only supported for
+ // pipes, with the blocking mode assumed for other file descriptors
+ // regardless of the flags.
+ //
+ // Note that on Wine currently pipes always behave as blocking regardless of
+ // the mode set.
+ //
+ LIBBUTL_SYMEXPORT fdstream_mode
+ fdmode (int, fdstream_mode);
+
+ // Portable functions for obtaining file descriptors of standard streams.
+ // Throw ios::failure on the underlying OS error.
+ //
+ // Note that you normally wouldn't want to close them using fddup() to
+ // convert them to auto_fd, for example:
+ //
+ // ifdstream is (fddup (stdin_fd ()));
+ //
+ LIBBUTL_SYMEXPORT int stdin_fd ();
+ LIBBUTL_SYMEXPORT int stdout_fd ();
+ LIBBUTL_SYMEXPORT int stderr_fd ();
+
+ // Convenience functions for setting the translation mode for standard
+ // streams.
+ //
+ LIBBUTL_SYMEXPORT fdstream_mode stdin_fdmode (fdstream_mode);
+ LIBBUTL_SYMEXPORT fdstream_mode stdout_fdmode (fdstream_mode);
+ LIBBUTL_SYMEXPORT fdstream_mode stderr_fdmode (fdstream_mode);
+
+ // Low-level, nothrow file descriptor API.
+ //
+
+ // Close the file descriptor. Return true on success, set errno and return
+ // false otherwise.
+ //
+ LIBBUTL_SYMEXPORT bool
+ fdclose (int) noexcept;
+
+ // Open the null device (e.g., /dev/null) that discards all data written to
+ // it and provides no data for read operations (i.e., yelds EOF on read).
+ // Return an auto_fd that holds its file descriptor on success and throwing
+ // ios::failure otherwise.
+ //
+ // On Windows the null device is NUL and writing anything substantial to it
+ // (like redirecting a process' output) is extremely slow, as in, an order
+ // of magnitude slower than writing to disk. If you are using the descriptor
+ // yourself this can be mitigated by setting the binary mode (already done
+ // by fdopen()) and using a buffer of around 64K. However, sometimes you
+ // have no control of how the descriptor will be used. For instance, it can
+ // be used to redirect a child's stdout and the way the child sets up its
+ // stdout is out of your control (on Windows). For such cases, there is an
+ // emulation via a temporary file. Mostly it functions as a proper null
+ // device with the file automatically removed once the descriptor is
+ // closed. One difference, however, would be if you were to both write to
+ // and read from the descriptor.
+ //
+ // Note that on POSIX the FD_CLOEXEC flag is set for the file descriptor to
+ // prevent its leakage into child processes. On Windows, for the same
+ // purpose, the _O_NOINHERIT flag is set.
+ //
+#ifndef _WIN32
+ LIBBUTL_SYMEXPORT auto_fd
+ fdopen_null ();
+#else
+ LIBBUTL_SYMEXPORT auto_fd
+ fdopen_null (bool temp = false);
+#endif
+
+ struct fdpipe
+ {
+ auto_fd in;
+ auto_fd out;
+
+ void
+ close ()
+ {
+ in.close ();
+ out.close ();
+ }
+ };
+
+ // Create a pipe. Throw ios::failure on the underlying OS error. By default
+ // both ends of the pipe are opened in the text mode. Pass the binary flag
+ // to instead open them in the binary mode. Passing a mode other than none
+ // or binary is illegal.
+ //
+ // Note that on Windows both ends of the created pipe are not inheritable.
+ // In particular, the process class that uses fdpipe underneath makes the
+ // appropriate end (the one being passed to the child) inheritable.
+ //
+ // Note that on POSIX the FD_CLOEXEC flag is set for both ends, so they get
+ // automatically closed by the child process to prevent undesired behaviors
+ // (such as child deadlock on read from a pipe due to the write-end leakage
+ // into the child process). Opening a pipe and setting the flag is not an
+ // atomic operation generally, but it is in regards to child process spawning
+ // (to prevent file descriptor leakage into child processes spawned from
+ // other threads). Also note that you don't need to reset the flag for a pipe
+ // end being passed to the process class ctor.
+ //
+ LIBBUTL_SYMEXPORT fdpipe
+ fdopen_pipe (fdopen_mode = fdopen_mode::none);
+
+ // Seeking.
+ //
+ enum class fdseek_mode {set, cur, end};
+
+ LIBBUTL_SYMEXPORT std::uint64_t
+ fdseek (int, std::int64_t, fdseek_mode);
+
+ // Truncate or expand the file to the specified size. Throw ios::failure on
+ // the underlying OS error.
+ //
+ LIBBUTL_SYMEXPORT void
+ fdtruncate (int, std::uint64_t);
+
+ // Return filesystem entry stat from file descriptor. Throw ios::failure on
+ // the underlying OS error.
+ //
+ // See also path_entry() in filesystem.
+ //
+ LIBBUTL_SYMEXPORT entry_stat
+ fdstat (int);
+
+ // Test whether a file descriptor refers to a terminal. Throw ios::failure
+ // on the underlying OS error.
+ //
+ LIBBUTL_SYMEXPORT bool
+ fdterm (int);
+
+ // Wait until one or more file descriptors becomes ready for input (reading)
+ // or output (writing). Return the pair of numbers of descriptors that are
+ // ready. Throw std::invalid_argument if anything is wrong with arguments
+ // (both sets are empty, invalid fd, etc). Throw ios::failure on the
+ // underlying OS error.
+ //
+ // Note that the function clears all the previously-ready entries on each
+ // call. Entries with nullfd are ignored.
+ //
+ // On Windows only pipes and only their input (read) ends are supported.
+ //
+ struct fdselect_state
+ {
+ int fd;
+ bool ready;
+
+ // Note: intentionally non-explicit to allow implicit initialization when
+ // pushing to fdselect_set.
+ //
+ fdselect_state (int fd): fd (fd), ready (false) {}
+ };
+
+ using fdselect_set = small_vector<fdselect_state, 4>;
+
+ LIBBUTL_SYMEXPORT std::pair<std::size_t, std::size_t>
+ fdselect (fdselect_set& ifds, fdselect_set& ofds);
+
+ inline std::size_t
+ ifdselect (fdselect_set& ifds)
+ {
+ fdselect_set ofds;
+ return fdselect (ifds, ofds).first;
+ }
+
+ inline std::size_t
+ ofdselect (fdselect_set& ofds)
+ {
+ fdselect_set ifds;
+ return fdselect (ifds, ofds).second;
+ }
+
+ // As above but wait up to the specified timeout returning a pair of zeroes
+ // if none of the descriptors became ready.
+ //
+ template <typename R, typename P>
+ std::pair<std::size_t, std::size_t>
+ fdselect (fdselect_set&, fdselect_set&, const std::chrono::duration<R, P>&);
+
+ template <typename R, typename P>
+ inline std::size_t
+ ifdselect (fdselect_set& ifds, const std::chrono::duration<R, P>& timeout)
+ {
+ fdselect_set ofds;
+ return fdselect (ifds, ofds, timeout).first;
+ }
+
+ template <typename R, typename P>
+ inline std::size_t
+ ofdselect (fdselect_set& ofds, const std::chrono::duration<R, P>& timeout)
+ {
+ fdselect_set ifds;
+ return fdselect (ifds, ofds, timeout).second;
+ }
+
+ // POSIX read() function wrapper. In particular, it supports the semantics
+ // of non-blocking read for pipes on Windows.
+ //
+ // Note that on Wine currently pipes always behave as blocking regardless of
+ // the mode.
+ //
+ LIBBUTL_SYMEXPORT std::streamsize
+ fdread (int, void*, std::size_t);
+}
+
+#include <libbutl/fdstream.ixx>