P1759R1
Native handle from file streams

1. Revision History

1.1. Revision 1 (Draft)

• Make native_handle_type be standard layout

• Add precondition (is_open() == true) to .native_handle()

• Add feature test macro __cpp_lib_fstream_native_handle

• Fix errors with opening the file with POSIX APIs in Motivation (see, we need this paper, fstreams are easier to open correctly!)

• Add additional motivating use case in vectored/scatter-gather IO

• Regular -> regular

Incorporate LEWGI feedback from Cologne (July 2019):

• Move to a member function and member typedef

• Make native_handle return value not be mandated to be unique

• Add note about how the presence of the members is required, and not implementation-defined (like for thread)

LEWGI Polls from Cologne:

Attendance: 21

We should promise more committee time pursuing P1759, knowing that our time is scarce and this will leave less time for other work.

Knowing what we know now, we should promise more committee time to pursuing a unifying native handle type for the standard library, knowing that our time is scarce and this will leave less time for other work.

SA: Implementers had concerns with inability to add new native handle types without breaking ABI.

Member function (MF) vs free function (FF) addded to stream classes.

SFF: Member function bad, free function good

The native handle type and native handle member function should always exist (they may return an invalid handle).

An implementer had concerns how this paper wouldn’t be useful if native handles were as underspecified as they are with thread.

Forward P1759, with a member function API that is always defined, to LEWG for C++23, knowing that our time is scarce and this will leave less time for other work.

1.2. Revision 0

2. Overview

This paper proposes adding a new typedef to standard file streams: native_handle_type. This type is an alias to whatever type the platform uses for its file descriptors: int on POSIX, HANDLE (void*) on Windows, and something else on other platforms. This type is a non-owning handle and is to be small, Regular, standard layout, and trivially copyable.

Alongside this, this paper proposes adding a concrete member function: .native_handle(), returning a native_handle_type, to the following class templates:

• basic_filebuf

• basic_ifstream

• basic_ofstream

• basic_fstream

3. Motivation

For some operations, using OS/platform-specific file APIs is necessary. If this is the case, they are unable to use iostreams without reopening the file with the platform-specific APIs.

For example, if someone wanted to query the time a file was last modified on POSIX, they’d use ::fstat, which takes a file descriptor:

int fd = ::open("~/foo.txt", O_RDONLY);
::stat s{};
int err = ::fstat(fd, &s);
std::chrono::sys_seconds last_modified = std::chrono::seconds(s.st_mtime.tv_sec);

Note: The Filesystem TS introduced the file_status structure and status function retuning one. This doesn’t solve our problem, because std::filesystem::status takes a path, not a native file descriptor (using paths is potentially racy), and std::filesystem::file_status only contains member functions type() and permissions(), not one for last time of modification. Extending this structure is out of scope for this proposal, and not feasible for every single possible operation the user may wish to do with OS APIs.

If the user needs to do a single operation not supported by the standard library, they have to make choice between only using OS APIs, or reopening the file every time necessary, likely forgetting to close the file, or running into buffering or synchronization issues.

// Writing the latest modification date to a file
std::chrono::sys_seconds last_modified(int fd) {
    // See above
}

// Today’s code

// Option #1:
// Use iostreams by reopening the file
{
    int fd = ::open("~/foo.txt", O_RDONLY); // CreateFile on Windows
    auto lm = last_modified(fd);

    ::close(fd); // CloseFile on Windows
    // Hope the path still points to the file!
    // Need to allocate 
    std::ofstream of("~/foo.txt");
    of << std::chrono::format("%c", lm) << '\n';
    // Need to flush
}

// Option #2:
// Abstain from using iostreams altogether
{
    int fd = ::open("~/foo.txt", O_RDWR);
    auto lm = last_modified(fd);

    // Using ::write() is clunky;
    // skipping error handling for brevity
    auto str = std::chrono::format("%c", lm);
    str.push_back('\n');
    ::write(fd, str.data(), str.size());
    // Remember to close!
    // Hope format or push_back doesn’t throw
    ::close(fd);
}

// This proposal
// No need to use platform-specific APIs to open the file
{
    std::ofstream of("~/foo.txt");
    auto lm = last_modified(of.native_handle());
    of << std::chrono::format("%c", lm) << '\n';
    // RAII does ownership handling for us
}

The utility of getting a file descriptor (or other native file handle) is not limited to getting the last modification date. Other examples include, but are definitely not limited to:

• file locking (fcntl() + F_SETLK on POSIX, LockFile on Windows)

• getting file status flags (fcntl() + F_GETFL on POSIX, GetFileInformationByHandle on Windows)

• vectored/scatter-gather IO (vread()/vwrite())

• non-blocking IO (fcntl() + O_NONBLOCK/F_SETSIG on POSIX)

Basically, this paper would make standard file streams interoperable with operating system interfaces, making iostreams more useful in that regard.

An alternative would be adding a lot of this functionality to fstream and filesystem. The problem is, that some of this behavior is inherently platform-specific. For example, getting the inode of a file is something that only makes sense on POSIX, so cannot be made part of the fstream interface, and is only accessible through the native file descriptor.

Facilities replacing iostreams, although desirable, are not going to be available in the standard in the near future. The author, alongside many others, would thus find this functionality useful.

4. Scope

This paper does not propose constructing a file stream or stream buffer from a native file handle. The author is worried of ownership and implementation issues possibly associated with this design.

// NOT PROPOSED
#include <fstream>
#include <fcntl.h>

auto fd = ::open(/* ... */);
auto f = std::fstream{fd};

This paper also does not touch anything related to FILE*.

5. Design Discussion

See polls in §1 Revision History for more details

5.1. Type of native_handle_type

In this paper, the definition for native_handle_type is much more strict than in thread. For reference, this is the wording from 32.2.3 Native handles [thread.req.native], from [N4800]:

Several classes described in this Clause have members native_handle_type and native_handle. The presence of these members and their semantics is implementation-defined. [ Note: These members allow implementations to provide access to implementation details. Their names are specified to facilitate portable compile-time detection. Actual use of these members is inherently non-portable. — end note ]

During the review of R0 of this paper in Cologne by LEWGI, an implementor mentioned how having the same specification here would make this paper effectively useless. Having the presence of a member be implementation-defined without a feature detect macro was deemed as bad design, which should not be replicated in this paper.

The proposed alternative in this paper, as directed by LEWGI, is allowing a conforming implementation to return an invalid native file handle, if it cannot be retrieved.

5.2. Member function vs free function

R0 of this paper had a free function std::native_file_handle() instead of a member function, due to possible ABI-related concerns. This turned out not to be an issue, so the design was changed to be a member function, for consistency with thread.

5.3. Precondition

The member function .native_handle(), as specified in this paper, has a precondition of .is_open() == true. The precondition is specified with "Expects", so breaking it would be UB, and in practice enforced with an assert. An alternative to this would be throwing if the file is not open, or returning some unspecified invalid handle.

6. Impact On the Standard and Existing Code

This proposal is a pure library extension, requiring no changes to the core language. It would cause no existing conforming code to break.

7. Implementation

Implementing this paper should be a relatively trivial task.

Although all implementations surveyed (libstdc++, libc++ and MSVC) use FILE* instead of native file descriptors in their basic_filebuf implementations, these platforms provide facilites to get a native handle from a FILE*; fileno on POSIX, and _fileno + _get_osfhandle on Windows. The following reference implementations use these.

For libstdc++ on Linux:

template <class CharT, class Traits>
class basic_filebuf {
    // ...
    using native_handle_type = int;
    // ...
    native_handle_type native_handle() {
        assert(is_open());
        // _M_file (__basic_file<char>) has a member function for this purpose
        return _M_file.fd();
        // ::fileno(file.file()) could also be used
    }
    // ...
}

// (i|o)fstream implementation is trivial with rdbuf()

For libc++ on Linux:

template <class CharT, class Traits>
class basic_filebuf {
    // ...
    using native_handle_type = int;
    // ...
    native_handle_type native_handle() {
        assert(is_open());
        // __file_ is a FILE*
        return ::fileno(__file_)
    }
    // ...
}

// (i|o)fstream implementation is trivial with rdbuf()

For MSVC:

template <class CharT, class Traits>
class basic_filebuf {
    // ...
    using native_handle_type = HANDLE;
    // ...
    native_handle_type native_handle() {
        assert(is_open());
        // _Myfile is a FILE*
        auto cfile = ::_fileno(_Myfile);
        // _get_osfhandle returns intptr_t, HANDLE is a void*
        return static_cast<HANDLE>(::_get_osfhandle(cfile));
    }
    // ...
}

// (i|o)fstream implementation is trivial with rdbuf()

8. Prior Art

[Boost.IOStreams] provides file_descriptor, file_descriptor_source, and file_descriptor_sink, which, when used in conjunction with stream_buffer, are std::basic_streambufs using a file descriptor. These classes can be constructed from a path or a native handle (int or HANDLE) and can also return it with member function handle().

The Networking TS [N4734] has members native_handle_type and .native_handle() in numerous places, including std::net::socket. It specifies (in [socket.reqmts.native]) the presence of these members in a similar fashion to thread, as in making their presence implementation-defined. It does, however, recommend POSIX-based systems to use int for this purpose.

Niall Douglas’s [P1031] also defined a structure native_handle_type with an extensive interface and a member union with an int and a HANDLE, with a constructor taking either one of these.

8.1. Discussion

There has been some discussion over the years about various things relating to this issue, but as far as the author is aware, no concrete proposal has ever been submitted.

There have been a number of threads on std-discussion and std-proposals: [std-proposals-native-handle], [std-discussion-fd-io], [std-proposals-native-raw-io], [std-proposals-fd-access]. The last one of these lead to a draft paper, that was never submitted: [access-file-descriptors].

The consensus that the author took from these discussions is, that native handle support for iostreams would be very much welcome.

An objection was raised by Billy O’Neal to being able to retrieve a native file handle from a standard file stream:

[This] also would need to mandate that the C++ streams be implemented directly such that there was a 1:1 native handle relationship, which may not be the case. For instance, a valid implementation of C++ iostreams would be on top of cstdio, which would not have any kind of native handle to expose.

– Billy O’Neal: [std-proposals-billy-oneal]

Every implementation surveyed did implement basic_filebuf on top of C stdio, but these platforms also provide functionality for getting a file descriptor out of a FILE*. On every platform, file I/O is ultimately implemented on top of native APIs, so not providing access to a file descriptor from a FILE* would be rather unfortunate. Should such a platform exist, they probably don’t have a conforming C++ implementation anyway. See §7 Implementation for more.

Additionally, as directed by LEWGI, .native_handle() can just return an invalid handle, if the implementation really can’t get a valid one corresponding to the file.

8.2. Existing precendent for presence of native_handle

9. Technical Specifications

The wording is based on [N4800].

__cpp_lib_filesystem	201703L	<filesystem>
__cpp_lib_fstream_native_handle	TBD	<fstream>
__cpp_lib_gcd_lcm	201606L	<numeric>

Note to editor: Replace ? with the appropriate section number.

? Native handles [file.native]

Several classes described in this section have a member native_handle_type.

The type native_handle_type serves as a type representing a platform-specific handle to a file. It satisfies the requirements of regular, and is trivially copyable and standard layout.

[ Note: For operating systems based on POSIX, native_handle_type should be int. For Windows-based operating systems, it should be HANDLE. — end note ]

namespace std {
    template<class charT, class traits = char_traits<charT>>
    class basic_filebuf : public basic_streambuf<charT, traits> {
    public:
        // Note: Add after other member typedefs
        using native_handle_type = implementation-defined; // see [file.native]

        // Note: Add as the last [filebuf.members]
        native_handle_type native_handle();

        // ...
    }
}

Modify paragraph § 1 of Class template basic_filebuf [filebuf]:

The class basic_filebuf<charT, traits> associates both the input sequence and the output sequence with a file. The file has an associated native_handle_type. Whether the associated native_handle_type is unique for each instance of a basic_filebuf, is implementation-defined. [ Note: This differs from thread native handles [thread.req.native], the presence of which is implementation-defined. — end note ]

native_handle_type native_handle();

Expects: is_open() is true.

Throws: Nothing.

Returns: The native_handle_type associated with the underlying file of *this.

namespace std {
    template<class charT, class traits = char_traits<charT>>
    class basic_ifstream : public basic_istream<charT, traits> {
    public:
        // Note: Add after other member typedefs
        using native_handle_type =
            typename basic_filebuf<charT, traits>::native_handle_type;

        // Note: Add as the last [ifstream.members]
        native_handle_type native_handle();

        // ...
    }
}

native_handle_type native_handle();

Effects: Equivalent to: return rdbuf()->native_handle();.

namespace std {
    template<class charT, class traits = char_traits<charT>>
    class basic_ofstream : public basic_ostream<charT, traits> {
    public:
        // Note: Add after other member typedefs
        using native_handle_type =
            typename basic_filebuf<charT, traits>::native_handle_type;

        // Note: Add as the last [ofstream.members]
        native_handle_type native_handle();

        // ...
    }
}

native_handle_type native_handle();

Effects: Equivalent to: return rdbuf()->native_handle();.

namespace std {
    template<class charT, class traits = char_traits<charT>>
    class basic_fstream : public basic_iostream<charT, traits> {
    public:
        // Note: Add after other member typedefs
        using native_handle_type =
            typename basic_filebuf<charT, traits>::native_handle_type;

        // Note: Add as the last [fstream.members]
        native_handle_type native_handle();

        // ...
    }
}

native_handle_type native_handle();

Effects: Equivalent to: return rdbuf()->native_handle();.

10. Acknowledgements

Thanks to Niall Douglas for feedback, encouragement and ambitious suggestions for this paper.

Thanks to the rest of the co-authors of [P1750] for the idea after cutting this functionality out, especially to Jeff Garland for providing a heads-up about a possible ABI-break that I totally would’ve missed, even though it ended up being a non-issue.