Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10 beta 3/9/83; site uthub.UUCP
Path: utzoo!utcsrgv!utai!uthub!thomson
From: thomson@uthub.UUCP (Brian Thomson)
Newsgroups: net.bugs.4bsd,net.unix-wizards
Subject: ptcread() throughput sextupling
Message-ID: <168@uthub.UUCP>
Date: Tue, 19-Jun-84 10:42:42 EDT
Article-I.D.: uthub.168
Posted: Tue Jun 19 10:42:42 1984
Date-Received: Tue, 19-Jun-84 14:26:50 EDT
Organization: CSRG, University of Toronto
Lines: 110

Index: /sys/sys/tty_pty.c 4.2BSD Fix

Description:
	Oink oink!
	That is the sound that data makes as it travels through
	4.2BSD's pseudo-tty driver.  Even in the high-volume
	direction (slave to controller) there is a great deal
	of code executed per-character.  On our otherwise idle 750 I
	measured the maximum pty throughput at 5K chars/sec.;
	after applying the following mods it reached 30K chars/sec.
	If your machine is often accessed through rlogin(1c) this
	can mean considerable savings in system-state CPU time.

Repeat-By:
	Run this program and use iostat(1) to see what your character
	rate is.

	#include <sys/types.h>
	char buf[1024];
	int wsize = 1024;

	main()
	{
		int csock, dsock, i;

		for(i=0 ; i<wsize; i++)
			buf[i] = '0';
		csock = getpty(&dsock);
		if(csock == -1) {
			perror("ptty");
			exit(1);
		}
		if(fork() == 0) {
			/* Child, writes on slave. */
			close(csock);
			while(write(dsock, buf, wsize) != -1)
				;
		} else {
			/* parent reads from controller */
			close(dsock);
			while(read(csock, buf, wsize)  != -1)
				;
		}
		exit(0);
	}
	getpty(ip)
		int *ip;
	{
		static char name[] = "/dev/ptyp0";
		int i;
		int res;

		for(i=0; i<16; i++) {
			name[9] = i+'0';
			res = open(name, 2);
			if(res != -1) {
				name[5] = 't';
				*ip = open(name, 2);
				if(*ip != -1)
					return(res);
				name[5] = 'p';
				close(res);
			}
		}
		return(-1);
	}


Fix:
	*WARNING*
	This modification introduces the following bug:
	
      *	A read from the controller side into an invalid or write-protected *
      *	user buffer (failing with an EFAULT error) may consume data that   *
      *	should have been read but was not delivered to the user.	   *

	I consider this a minor nit, and am willing to put up with it
	considering the payoff.

	First, remove the #ifdef notdef around the q_to_b() function
	in file sys/tty_subr.c.  
	Then, edit routine ptcread() in file sys/tty_pty.c as follows:

	1) add the local definition
		char ptc_rdbuf[BUFSIZ];
	   after the other local defs in ptcread()

	2) change the loop

	    while (tp->t_outq.c_cc && uio->uio_resid > 0)
		    if (ureadc(getc(&tp->t_outq), uio) < 0) {
			    error = EFAULT;
			    break;
		    }

	   to be

	    while(uio->uio_resid > 0 && error == 0) {
		    int cc = q_to_b(&tp->t_outq, ptc_rdbuf, MIN(uio->uio_resid, BUFSIZ));
		    if(cc == 0)
			    break;
		    error = uiomove(ptc_rdbuf, cc, UIO_READ, uio);
	    }

	We have been using this code for 4 months.
	You may not experience quite as great an improvement as we did,
	because we also improved the performance of q_to_b().
-- 
		    Brian Thomson,	    CSRG Univ. of Toronto
		    {linus,ihnp4,uw-beaver,floyd,utzoo}!utcsrgv!uthub!thomson