elemental-ircd/doc/technical/linebuf.txt

140 lines
4.5 KiB
Plaintext

linebuf - a dbuf replacement for the New World Order(tm)
By Adrian Chadd <adrian@creative.net.au>
$Id: linebuf.txt 6 2005-09-10 01:02:21Z nenolod $
History
-------
I could probably learn the dbuf history, but basically its evil. The
general idea is that a dbuf holds incoming and outgoing data streams.
The trouble is that well.. it was evil. You can check it out by getting
the old src/dbuf.c and include/dbuf.h files if you really want.
Replacement
-----------
The linebuf system is a replacement for the dbuf code. The general idea here
is that the data should be buffered in "lines" rather than just linearly
like in the dbuf code. This lends to easier manipulation at a later date
(think flushing data lines to a socket, and even "sharing" linebufs to
reduce the copying required for one to many delivery.)
The linebuf system is broken into two structures, the buf_head and buf_line .
buf_head contains the buffer information (queue head/tail, length, allocated
length and the write offset for flushing), and buf_line contains the
line buffer information (buffer and various flags.)
linebuf->terminated is *only* set when a CR/LF combination is received.
linebuf->overflow is set if we get more data than we should, and we simply
truncate the incoming data.
linebuf->flushing is set when we are currently writing the buffer. We should
_NEVER_ be appending to a buffer which we're flushing!
When you get a buffer through linebuf_get() or write one through
linebuf_flush(), it will *always* be terminated with a CR/LF (and a NUL if
its a linebuf_get()).
Linebuf manipulation
--------------------
To use a linebuf, you simply stick a buf_head_t in your structure somewhere.
You then use the following routines:
int
linebuf_parse(buf_head_t *bufhead, char *buf, int len)
Parse the given buf. This routine does some complex manipulation:
- if there is an incomplete buffer at the tail, buf is parsed to try and
fill that incomplete buffer
- a buffer is completed by a CR/LF/CRLF/LFCR. It accepts everything purely
because I wanted to be "liberal in what you accept" ..
- If a buffer is terminated, the linebuf is flagged terminated
- If more data is trying to be squeezed into the buffer than space LEFT
in the buffer, we skip to the next "CRLF", and tag the buffer terminated
_and_ overflowed.
- We treat multiple runs of CR/LF/CRLF/LFCR as a single CRLF. This is just
a little extra goody to stop people sending hundreds of "CRLF"s and creating
unnecessary buffers.
- The number of lines parsed is returned (so you can implement per-line flood
protection ..)
void
linebuf_put(buf_head_t *bufhead, char *buf, int len)
Parse the given buf, ASSUMING it is a single buffer line. This is useful
for server-generated messages where you know you have a single line, and
you don't want to go through the overhead of parsing the data just for
this.
int
linebuf_get(buf_head_t *bufhead, char *buf, int maxlen)
Get a single line from the buffer. This removes data from the head of the
buffer. If the first buffer is empty or is not terminated, 0 is returned
which indicates that there is no data to parse. Terminated buffers are
returned (CR/LF/NUL), and the length INCLUDING the CR/LF/NUL is returned.
The buffer is copied and the linebuf is then deallocated.
int
linebuf_flush(int fd, buf_head_t *bufhead)
Attempt to flush some data to the given socket. bufhead->writeofs tracks
where in the head buffer we currently are. If the buffer is not terminated,
-1 is returned with errno == EWOULDBLOCK to simulate a "retry me" condition.
(See TODO..)
linebuf_flush() returns whatever write() returns, and sets (ie doesn't touch
after write()) errno accordingly.
int
linebuf_len(buf_head_t *bufhead)
Return the length of the buffer, in bytes. This should be used when calculating
how big a buffer is for statistics.
int
linebuf_alloclen(buf_head_t *bufhead)
Return how big the *allocated* space is. This is much more suitable for
anti-flood checking, as someone might be sending a whole bunch of 1-byte
linebufs which might not trigger a recvq / sendq limit but might chew up
way too much memory.
Notes
-----
* Remember that the trailing NUL isn't covered in the string length.
Limitations
-----------
* all the buffers are a fixed size - here they are current 513 bytes
(510 bytes + CR/LF/NUL)
TODO
----
* linebuf_flush() should be changed a little so if the buffer isn't
terminated, we *dont* retry flushing a buffer until we get more data.
* Implement a reference-friendly linebuf to reduce copies ..