Return to BSD News archive
Newsgroups: comp.unix.bsd Path: sserve!manuel!munnari.oz.au!mips!kpc!amdcad!BitBlocks.com!bvs From: bvs@BitBlocks.COM (Bakul Shah) Subject: New /sys/i386/i386/in_cksum.c Message-ID: <1992Jul25.231505.9325@BitBlocks.COM> Organization: Bit Blocks, Inc. Date: Sat, 25 Jul 1992 23:15:05 GMT The enclosed in_cksum.c is about 5 times faster for 128 byte mbufs (and upto 10 times faster for large mbufs). However, its effect on the overall throughput is rather small; which is not too surprising. The best ttcp performance I get with the new in_cksum is about 264KB/s (SunOS3.5/Sun 3 -> 25Mhz i486 + WE8013EPC) & 242KB/s (pc->sun), up from 256KB/s & 238KB/s. This in_cksum is derived from the CCI version of the same from the tahoe release. All I had to do was replace the carry related macros with i386 specific ones (+ I made some cosmetic changes). Thanks to gcc I didn't have to guess which registers contained the checksum and the mbuf data ptr! Use at your own risk. If you do decide to use it, make sure you move the original in_cksum.c out of the way or else this shar won't overwrite it. And now for some extra credit :-), for those of you who don't have a quayle, here is a quick crash course on configuring and creating new kernels. cd /sys/i386/conf # this is is where config info is kept cp GENERICISA FOO # GENERICISA is a confiuration script # that describes a particular config. # Most likely you are running a kernel # that was built out of this script. # Copy it to a new file so that you can # create your own configuration. vi FOO # Now remove devices you don't have, # especially the network devices and # make any other changes you wish; # such as adding option MFS for a memory # file system for /tmp etc. You can also # add psuedo-device ddb (the debugger). # Change`ident' entry to FOO or what is your config # file name and `cpu' to "i486" if you have a 486. # You may also wish to look at some other # all UPPERCASE files in the same dir. You shouldn't # change any other files unless you really know # what you are doing. config FOO # this will create /sys/compile/FOO dir # and put a bunch of files in there including # a Makefile. cd /sys/compile/FOO # If you are running the original you have some work to do #vi Makefile # If you are running the original # 0.1 release you will have to manually add # vers.o at the end of the line that starts # with SYSTEM_LD (or you can just get the # new Makefile.i386 patch from agate). #cp ../GENERICAISA/vers.c . # since this file is missing #cc -c vers.c # since the Makefile won't do it. make # this will make a new kernel Now you can copy it to / but make sure you have fixit.fs or dist.fs floppy handy so that you can recover. mv /386bsd /386bsd.alt # save the old kernel cp 386bsd /386bsd fastboot # that is it! A few seconds later you should # be running on your new kernel. -- Bakul Shah <bvs@bitblocks.com> #!/bin/sh # This is a shell archive (produced by shar 3.49) # To extract the files from this archive, save it to a file, remove # everything above the "!/bin/sh" line above, and type "sh file_name". # # made 07/25/1992 22:47 UTC by bvs@bitblocks.com # Source directory /kabul/src/sys.386bsd # # existing files will NOT be overwritten unless -c is specified # # This shar contains: # length mode name # ------ ---------- ------------------------------------------ # 5048 -rw-rw-r-- i386/i386/in_cksum.c # # ============= i386/i386/in_cksum.c ============== if test ! -d 'i386'; then echo 'x - creating directory i386' mkdir 'i386' fi if test ! -d 'i386/i386'; then echo 'x - creating directory i386/i386' mkdir 'i386/i386' fi if test -f 'i386/i386/in_cksum.c' -a X"$1" != X"-c"; then echo 'x - skipping i386/i386/in_cksum.c (File already exists)' else echo 'x - extracting i386/i386/in_cksum.c (Text)' sed 's/^X//' << 'SHAR_EOF' > 'i386/i386/in_cksum.c' && /*- X * Copyright (c) 1990 The Regents of the University of California. X * All rights reserved. X * X * Redistribution and use in source and binary forms, with or without X * modification, are permitted provided that the following conditions X * are met: X * 1. Redistributions of source code must retain the above copyright X * notice, this list of conditions and the following disclaimer. X * 2. Redistributions in binary form must reproduce the above copyright X * notice, this list of conditions and the following disclaimer in the X * documentation and/or other materials provided with the distribution. X * 3. All advertising materials mentioning features or use of this software X * must display the following acknowledgement: X * This product includes software developed by the University of X * California, Berkeley and its contributors. X * 4. Neither the name of the University nor the names of its contributors X * may be used to endorse or promote products derived from this software X * without specific prior written permission. X * X * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND X * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE X * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE X * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE X * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL X * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS X * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) X * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT X * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY X * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF X * SUCH DAMAGE. X * X * from tahoe: in_cksum.c 1.2 86/01/05 X * @(#)in_cksum.c 1.3 (Berkeley) 1/19/91 X * X * 920724 i386 changes by Bakul Shah <bvs@bitblocks.com> X */ X #include "param.h" #include "sys/mbuf.h" X /* X * Checksum routine for Internet Protocol family headers. X * X * This routine is very heavily used in the network X * code and should be modified for each CPU to be as fast as possible. X * X * This implementation is 386 version. X */ X #undef ADDCARRY #define ADDCARRY(x) if ((x) > 0xffff) (x) -= 0xffff #define REDUCE {sum = (sum & 0xffff) + (sum >> 16); ADDCARRY(sum);} X /* X * Thanks to gcc we don't have to guess X * which registers contain sum & w. X */ #define CLC asm("clc") #define ADD(n) asm("adcl " #n "(%2), %0": "=r"(sum): "0"(sum), "r"(w)) #define MOP asm("adcl $0, %0": "=r"(sum): "0"(sum)) X in_cksum(m, len) X register struct mbuf *m; X register int len; { X register u_short *w; X register unsigned sum = 0; X register int mlen = 0; X int byte_swapped = 0; X union { char c[2]; u_short s; } su; X X for (;m && len; m = m->m_next) { X if (m->m_len == 0) X continue; X w = mtod(m, u_short *); X if (mlen == -1) { X /* X * The first byte of this mbuf is the continuation X * of a word spanning between this mbuf and the X * last mbuf. X */ X X /* su.c[0] is already saved when scanning previous X * mbuf. sum was REDUCEd when we found mlen == -1 X */ X su.c[1] = *(u_char *)w; X sum += su.s; X w = (u_short *)((char *)w + 1); X mlen = m->m_len - 1; X len--; X } else X mlen = m->m_len; X if (len < mlen) X mlen = len; X len -= mlen; X /* X * Force to long boundary so we do longword aligned X * memory operations X */ X if (3 & (int) w) { X REDUCE; X if ((1 & (int) w) && (mlen > 0)) { X sum <<= 8; X su.c[0] = *(char *)w; X w = (u_short *)((char *)w + 1); X mlen--; X byte_swapped = 1; X } X if ((2 & (int) w) && (mlen >= 2)) { X sum += *w++; X mlen -= 2; X } X } X /* X * Do as much of the checksum as possible 32 bits at at time. X * In fact, this loop is unrolled to make overhead from X * branches &c small. X */ X while ((mlen -= 32) >= 0) { X /* X * Clear the carry flag, add with carry 16 words X * and fold-in last carry by adding a 0 with carry. X */ X CLC; X ADD(0); ADD(4); ADD(8); ADD(12); X ADD(16); ADD(20); ADD(24); ADD(28); X MOP; w += 16; X } X mlen += 32; X while ((mlen -= 8) >= 0) { X CLC; X ADD(0); ADD(4); X MOP; X w += 4; X } X mlen += 8; X if (mlen == 0 && byte_swapped == 0) X continue; /* worth 1% maybe ?? */ X REDUCE; X while ((mlen -= 2) >= 0) { X sum += *w++; X } X if (byte_swapped) { X sum <<= 8; X byte_swapped = 0; X if (mlen == -1) { X su.c[1] = *(char *)w; X sum += su.s; X mlen = 0; X } else X mlen = -1; X } else if (mlen == -1) X /* X * This mbuf has odd number of bytes. X * There could be a word split betwen X * this mbuf and the next mbuf. X * Save the last byte (to prepend to next mbuf). X */ X su.c[0] = *(char *)w; X } X X if (len) X printf("cksum: out of data\n"); X if (mlen == -1) { X /* The last mbuf has odd # of bytes. Follow the X standard (the odd byte is shifted left by 8 bits) */ X su.c[1] = 0; X sum += su.s; X } X REDUCE; X return (~sum & 0xffff); } SHAR_EOF chmod 0664 i386/i386/in_cksum.c || echo 'restore of i386/i386/in_cksum.c failed' Wc_c="`wc -c < 'i386/i386/in_cksum.c'`" test 5048 -eq "$Wc_c" || echo 'i386/i386/in_cksum.c: original size 5048, current size' "$Wc_c" fi exit 0 shar: saving i386/i386/in_cksum.c (Text)