/* +------------------------------------+ * | Inspire Internet Relay Chat Daemon | * +------------------------------------+ * * InspIRCd: (C) 2002-2008 InspIRCd Development Team * See: http://www.inspircd.org/wiki/index.php/Credits * * This program is free but copyrighted software; see * the file COPYING for details. * * --------------------------------------------------- */ /** Right, Line Buffers. Time for an explanation as to how sendqs work. By the way, ircd people, * before you jump up and down screaming, this is not anything like Adrian Chadd's linebuffer * stuff done for hybrid(and never really completed.) Rather, this is (I think) where linebuffer * was heading, and should have been. * * Enough introduction, actual explanation starts here. * In IRCd, we have a sendq. Traditionally, a sendq has been a big string. Stuff is tacked onto * the end, and when we can, we send the user some data off the start of it. A circular buffer. * * This model works okay, and is quite simplistic to code, but has a drawback in that most IRC * messages are multicast: topic changes, joins, parts, channel messages, and so on. * This means that for each of these messages, we must do O(n^n) amount of work: bytes ^ recipients * of writes will be made to sendqs, and this is a slow and expensive operation. * * The solution comes with the use of this linebuffer class below, which is managed entirely by * the user class (though it *may* be possible to add a server to server implementation of this later, * but that's nowhere near so needed, and nowhere near so trivial, thanks to the inherited nature * of buffered socket, but I digress). * * What this class does, in a nutshell: * When we need to send a message to a user, we create a LineBuffer object. It has a reference count, and * we copy the string we need to send into the LineBuffer object also. * We then tack a pointer to this LineBuffer into an std::list stored in the User class. * When the user writes data, a ptr is advanced depending how much of that line they wrote. If they wrote all * of the line, the pointer is popped off the std::list, the ptr is reset, and the buffer's refcount is * decremented - and if it reaches 0, the linebuffer is destroyed as it has fulfilled it's purpose. * * Effectively, this means that multicast writes become O(n) + time taken to copy message once, or just about. * * We gain efficiency, and much, much better RAM usage. */ static unsigned int totalbuffers = 0; class LineBuffer { private: std::string msg; unsigned long refcount; // Don't let it be copied. LineBuffer(const LineBuffer &) { } public: ~LineBuffer() { totalbuffers--; printf("Destroying LineBuffer with %u bytes, total buffers is %u\n", msg.length(), totalbuffers); msg.resize(0); } LineBuffer(std::string &m) { if (m.length() > MAXBUF - 2) /* MAXBUF has a value of 514, to account for line terminators */ { // Trim the message to fit, 510 characters max. m = m.substr(0, MAXBUF - 4); // MAXBUF is 514, we need 510. } // Add line terminator m.append("\r\n"); // And copy msg = m; refcount = 0; totalbuffers++; printf("Creating LineBuffer with %u bytes, total buffers is %u\n", msg.length(), totalbuffers); } std::string &GetMessage() { return msg; } unsigned long GetMessageLength() { return msg.length(); } // To be used after creation, when we know how many recipients we actually have. void SetRefcount(unsigned long r) { refcount = r; } unsigned long DecrementCount() { if (refcount == 0) { throw "decrementing a refcount when nobody is using it is weird and wrong"; } refcount--; return refcount; } // There is no increment method as it isn't really needed. };