ref: 6a51a03b801b21f42e2eb3dfa02c99e96c86b10c
dir: /ios.h/
// this flag controls when data actually moves out to the underlying I/O
// channel. memory streams are a special case of this where the data
// never moves out.
typedef enum {
bm_none,
bm_line,
bm_block,
bm_mem,
}bufmode_t;
typedef enum {
bst_none,
bst_rd,
bst_wr,
}bufstate_t;
#define IOS_INLSIZE 128
#define IOS_BUFSIZE 32768
typedef struct {
char *filename;
uint32_t lineno;
uint32_t colno;
}ios_loc_t;
typedef struct {
uint8_t *buf; // start of buffer
size_t maxsize; // space allocated to buffer
size_t size; // length of valid data in buf, >=ndirty
size_t bpos; // current position in buffer
size_t ndirty; // # bytes at &buf[0] that need to be written
off_t fpos; // cached file pos
bufmode_t bm;
int colnowait;
// the state only indicates where the underlying file position is relative
// to the buffer. reading: at the end. writing: at the beginning.
// in general, you can do any operation in any state.
bufstate_t state;
int fd;
uint8_t readonly:1;
uint8_t ownbuf:1;
uint8_t ownfd:1;
uint8_t _eof:1;
// this means you can read, seek back, then read the same data
// again any number of times. usually only true for files and strings.
uint8_t rereadable:1;
// this enables "stenciled writes". you can alternately write and
// seek without flushing in between. this performs read-before-write
// to populate the buffer, so "rereadable" capability is required.
// this is off by default.
//uint8_t stenciled:1;
// request durable writes (fsync)
// uint8_t durable:1;
ios_loc_t loc;
// todo: mutex
uint8_t local[IOS_INLSIZE];
}ios_t;
void *llt_memrchr(const void *s, int c, size_t n);
/* low-level interface functions */
size_t ios_read(ios_t *s, void *dest, size_t n);
size_t ios_write(ios_t *s, const void *data, size_t n);
off_t ios_seek(ios_t *s, off_t pos); // absolute seek
off_t ios_seek_end(ios_t *s);
off_t ios_skip(ios_t *s, off_t offs); // relative seek
off_t ios_pos(ios_t *s); // get current position
size_t ios_trunc(ios_t *s, size_t size);
int ios_eof(ios_t *s);
int ios_flush(ios_t *s);
void ios_close(ios_t *s);
uint8_t *ios_takebuf(ios_t *s, size_t *psize); // null-terminate and release buffer to caller
// set buffer space to use
int ios_setbuf(ios_t *s, uint8_t *buf, size_t size, int own);
int ios_bufmode(ios_t *s, bufmode_t mode);
void ios_set_readonly(ios_t *s);
size_t ios_copy(ios_t *to, ios_t *from, size_t nbytes);
size_t ios_copyall(ios_t *to, ios_t *from);
size_t ios_copyuntil(ios_t *to, ios_t *from, uint8_t delim);
// ensure at least n bytes are buffered if possible. returns # available.
size_t ios_readprep(ios_t *from, size_t n);
int ios_wait(ios_t *s, double ws);
/* stream creation */
ios_t *ios_file(ios_t *s, char *fname, int rd, int wr, int create, int trunc);
ios_t *ios_mem(ios_t *s, size_t initsize);
ios_t *ios_str(ios_t *s, char *str);
ios_t *ios_static_buffer(ios_t *s, const uint8_t *buf, size_t sz);
ios_t *ios_fd(ios_t *s, int fd, int isfile, int own);
// todo: ios_socket
extern ios_t *ios_stdin;
extern ios_t *ios_stdout;
extern ios_t *ios_stderr;
void ios_init_stdstreams(void);
/* high-level functions - output */
int ios_pututf8(ios_t *s, Rune r);
int ios_printf(ios_t *s, const char *format, ...);
int ios_vprintf(ios_t *s, const char *format, va_list args);
void hexdump(ios_t *dest, const uint8_t *buffer, size_t len, size_t startoffs);
/* high-level stream functions - input */
int ios_getutf8(ios_t *s, Rune *r);
// discard data buffered for reading
void ios_purge(ios_t *s);
/* stdio-style functions */
#define IOS_EOF (-1)
int ios_putc(ios_t *s, int c);
int ios_getc(ios_t *s);
int ios_peekc(ios_t *s);
#define ios_puts(s, str) ios_write(s, str, strlen(str))
/*
With memory streams, mixed reads and writes are equivalent to performing
sequences of *p++, as either an lvalue or rvalue. File streams behave
similarly, but other streams might not support this. Using unbuffered
mode makes this more predictable.
Note on "unget" functions:
There are two kinds of functions here: those that operate on sized
blocks of bytes and those that operate on logical units like "character"
or "integer". The "unget" functions only work on logical units. There
is no "unget n bytes". You can only do an unget after a matching get.
However, data pushed back by an unget is available to all read operations.
The reason for this is that unget is defined in terms of its effect on
the underlying buffer (namely, it rebuffers data as if it had been
buffered but not read yet). IOS reserves the right to perform large block
operations directly, bypassing the buffer. In such a case data was
never buffered, so "rebuffering" has no meaning (i.e. there is no
correspondence between the buffer and the physical stream).
Single-bit I/O is able to write partial bytes ONLY IF the stream supports
seeking. Also, line buffering is not well-defined in the context of
single-bit I/O, so it might not do what you expect.
implementation notes:
in order to know where we are in a file, we must ensure the buffer
is only populated from the underlying stream starting with p==buf.
to switch from writing to reading: flush, set p=buf, cnt=0
to switch from reading to writing: seek backwards cnt bytes, p=buf, cnt=0
when writing: buf starts at curr. physical stream pos, p - buf is how
many bytes we've written logically. cnt==0
dirty == (bitpos>0 && state==iost_wr), EXCEPT right after switching from
reading to writing, where we might be in the middle of a byte without
having changed it.
to write a bit: if !dirty, read up to maxsize-(p-buf) into buffer, then
seek back by the same amount (undo it). write onto those bits. now set
the dirty bit. in this state, we can bit-read up to the end of the byte,
then formally switch to the read state using flush.
design points:
- data-source independence, including memory streams
- expose buffer to user, allow user-owned buffers
- allow direct I/O, don't always go through buffer
- buffer-internal seeking. makes seeking back 1-2 bytes very fast,
and makes it possible for sockets where it otherwise wouldn't be
- tries to allow switching between reading and writing
- support 64-bit and large files
- efficient, low-latency buffering
- special support for utf8
- type-aware functions with byte-order swapping service
- position counter for meaningful data offsets with sockets
theory of operation:
the buffer is a view of part of a file/stream. you can seek, read, and
write around in it as much as you like, as if it were just a string.
we keep track of the part of the buffer that's invalid (written to).
we remember whether the position of the underlying stream is aligned
with the end of the buffer (reading mode) or the beginning (writing mode).
based on this info, we might have to seek back before doing a flush.
as optimizations, we do no writing if the buffer isn't "dirty", and we
do no reading if the data will only be overwritten.
*/