зеркало из https://github.com/github/ruby.git
138 строки
3.7 KiB
C
138 строки
3.7 KiB
C
#ifndef INTERNAL_IO_H /*-*-C-*-vi:se ft=c:*/
|
|
#define INTERNAL_IO_H
|
|
/**
|
|
* @author Ruby developers <ruby-core@ruby-lang.org>
|
|
* @copyright This file is a part of the programming language Ruby.
|
|
* Permission is hereby granted, to either redistribute and/or
|
|
* modify this file, provided that the conditions mentioned in the
|
|
* file COPYING are met. Consult the file for details.
|
|
* @brief Internal header for IO.
|
|
*/
|
|
#include "ruby/ruby.h" /* for VALUE */
|
|
|
|
#define HAVE_RB_IO_T
|
|
struct rb_io;
|
|
|
|
#include "ruby/io.h" /* for rb_io_t */
|
|
|
|
/** Ruby's IO, metadata and buffers. */
|
|
typedef struct rb_io {
|
|
|
|
/** The IO's Ruby level counterpart. */
|
|
VALUE self;
|
|
|
|
/** stdio ptr for read/write, if available. */
|
|
FILE *stdio_file;
|
|
|
|
/** file descriptor. */
|
|
int fd;
|
|
|
|
/** mode flags: FMODE_XXXs */
|
|
int mode;
|
|
|
|
/** child's pid (for pipes) */
|
|
rb_pid_t pid;
|
|
|
|
/** number of lines read */
|
|
int lineno;
|
|
|
|
/** pathname for file */
|
|
VALUE pathv;
|
|
|
|
/** finalize proc */
|
|
void (*finalize)(struct rb_io*,int);
|
|
|
|
/** Write buffer. */
|
|
rb_io_buffer_t wbuf;
|
|
|
|
/**
|
|
* (Byte) read buffer. Note also that there is a field called
|
|
* ::rb_io_t::cbuf, which also concerns read IO.
|
|
*/
|
|
rb_io_buffer_t rbuf;
|
|
|
|
/**
|
|
* Duplex IO object, if set.
|
|
*
|
|
* @see rb_io_set_write_io()
|
|
*/
|
|
VALUE tied_io_for_writing;
|
|
|
|
struct rb_io_encoding encs; /**< Decomposed encoding flags. */
|
|
|
|
/** Encoding converter used when reading from this IO. */
|
|
rb_econv_t *readconv;
|
|
|
|
/**
|
|
* rb_io_ungetc() destination. This buffer is read before checking
|
|
* ::rb_io_t::rbuf
|
|
*/
|
|
rb_io_buffer_t cbuf;
|
|
|
|
/** Encoding converter used when writing to this IO. */
|
|
rb_econv_t *writeconv;
|
|
|
|
/**
|
|
* This is, when set, an instance of ::rb_cString which holds the "common"
|
|
* encoding. Write conversion can convert strings twice... In case
|
|
* conversion from encoding X to encoding Y does not exist, Ruby finds an
|
|
* encoding Z that bridges the two, so that X to Z to Y conversion happens.
|
|
*/
|
|
VALUE writeconv_asciicompat;
|
|
|
|
/** Whether ::rb_io_t::writeconv is already set up. */
|
|
int writeconv_initialized;
|
|
|
|
/**
|
|
* Value of ::rb_io_t::rb_io_enc_t::ecflags stored right before
|
|
* initialising ::rb_io_t::writeconv.
|
|
*/
|
|
int writeconv_pre_ecflags;
|
|
|
|
/**
|
|
* Value of ::rb_io_t::rb_io_enc_t::ecopts stored right before initialising
|
|
* ::rb_io_t::writeconv.
|
|
*/
|
|
VALUE writeconv_pre_ecopts;
|
|
|
|
/**
|
|
* This is a Ruby level mutex. It avoids multiple threads to write to an
|
|
* IO at once; helps for instance rb_io_puts() to ensure newlines right
|
|
* next to its arguments.
|
|
*
|
|
* This of course doesn't help inter-process IO interleaves, though.
|
|
*/
|
|
VALUE write_lock;
|
|
|
|
/**
|
|
* The timeout associated with this IO when performing blocking operations.
|
|
*/
|
|
VALUE timeout;
|
|
} rb_io_t;
|
|
|
|
/* io.c */
|
|
void ruby_set_inplace_mode(const char *);
|
|
void rb_stdio_set_default_encoding(void);
|
|
VALUE rb_io_flush_raw(VALUE, int);
|
|
size_t rb_io_memsize(const rb_io_t *);
|
|
int rb_stderr_tty_p(void);
|
|
void rb_io_fptr_finalize_internal(void *ptr);
|
|
#ifdef rb_io_fptr_finalize
|
|
# undef rb_io_fptr_finalize
|
|
#endif
|
|
#define rb_io_fptr_finalize rb_io_fptr_finalize_internal
|
|
VALUE rb_io_popen(VALUE pname, VALUE pmode, VALUE env, VALUE opt);
|
|
|
|
VALUE rb_io_prep_stdin(void);
|
|
VALUE rb_io_prep_stdout(void);
|
|
VALUE rb_io_prep_stderr(void);
|
|
|
|
RUBY_SYMBOL_EXPORT_BEGIN
|
|
/* io.c (export) */
|
|
void rb_maygvl_fd_fix_cloexec(int fd);
|
|
int rb_gc_for_fd(int err);
|
|
void rb_write_error_str(VALUE mesg);
|
|
RUBY_SYMBOL_EXPORT_END
|
|
|
|
#endif /* INTERNAL_IO_H */
|