Home | History | Annotate | Line # | Download | only in dump
rcache.c revision 1.18
      1 /*	$NetBSD: rcache.c,v 1.18 2003/02/06 23:00:08 enami Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1999 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Martin J. Laubach <mjl (at) emsi.priv.at> and
      9  *    Manuel Bouyer <Manuel.Bouyer (at) lip6.fr>.
     10  *
     11  * Redistribution and use in source and binary forms, with or without
     12  * modification, are permitted provided that the following conditions
     13  * are met:
     14  * 1. Redistributions of source code must retain the above copyright
     15  *    notice, this list of conditions and the following disclaimer.
     16  * 2. Redistributions in binary form must reproduce the above copyright
     17  *    notice, this list of conditions and the following disclaimer in the
     18  *    documentation and/or other materials provided with the distribution.
     19  * 3. All advertising materials mentioning features or use of this software
     20  *    must display the following acknowledgement:
     21  *	This product includes software developed by the NetBSD
     22  *	Foundation, Inc. and its contributors.
     23  * 4. Neither the name of The NetBSD Foundation nor the names of its
     24  *    contributors may be used to endorse or promote products derived
     25  *    from this software without specific prior written permission.
     26  *
     27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     37  * POSSIBILITY OF SUCH DAMAGE.
     38  */
     39 
     40 #include <sys/cdefs.h>
     41 #ifndef lint
     42 __RCSID("$NetBSD: rcache.c,v 1.18 2003/02/06 23:00:08 enami Exp $");
     43 #endif /* not lint */
     44 
     45 #include <sys/types.h>
     46 #include <sys/uio.h>
     47 #include <sys/mman.h>
     48 #include <sys/param.h>
     49 #include <sys/sysctl.h>
     50 #include <ufs/ufs/dinode.h>
     51 
     52 #include <stdio.h>
     53 #include <stdlib.h>
     54 #include <unistd.h>
     55 #include <fcntl.h>
     56 #include <errno.h>
     57 #include <string.h>
     58 
     59 #include "dump.h"
     60 
     61 /*-----------------------------------------------------------------------*/
     62 #define MAXCACHEBUFS	512	/* max 512 buffers */
     63 #define MAXMEMPART	6	/* max 15% of the user mem */
     64 
     65 /*-----------------------------------------------------------------------*/
     66 union cdesc {
     67 	volatile size_t cd_count;
     68 	struct {
     69 		volatile daddr_t blkstart;
     70 		volatile daddr_t blkend;	/* start + nblksread */
     71 		volatile daddr_t blocksRead;
     72 		volatile size_t time;
     73 #ifdef DIAGNOSTICS
     74 		volatile pid_t owner;
     75 #endif
     76 	} desc;
     77 #define cd_blkstart	desc.blkstart
     78 #define cd_blkend	desc.blkend
     79 #define cd_blocksRead	desc.blocksRead
     80 #define cd_time		desc.time
     81 #define cd_owner	desc.owner
     82 };
     83 
     84 static int findlru(void);
     85 
     86 static void *shareBuffer = NULL;
     87 static union cdesc *cheader;
     88 static union cdesc *cdesc;
     89 static char *cdata;
     90 static int cachebufs;
     91 static int nblksread;
     92 
     93 #ifdef STATS
     94 static int nreads;
     95 static int nphysread;
     96 static int64_t readsize;
     97 static int64_t physreadsize;
     98 #endif
     99 
    100 #define	CSIZE		(nblksread << dev_bshift)	/* cache buf size */
    101 #define	CDATA(desc)	(cdata + ((desc) - cdesc) * CSIZE)
    102 
    103 void
    104 initcache(int cachesize, int readblksize)
    105 {
    106 	size_t len;
    107 	size_t sharedSize;
    108 
    109 	/* Convert read block size in terms of filesystem block size */
    110 	nblksread = howmany(readblksize, ufsib->ufs_bsize);
    111 
    112 	/* Then, convert it in terms of device block size */
    113 	nblksread <<= ufsib->ufs_bshift - dev_bshift;
    114 
    115 	if (cachesize == -1) {	/* Compute from memory available */
    116 		int usermem;
    117 		int mib[2] = { CTL_HW, HW_USERMEM };
    118 
    119 		len = sizeof(usermem);
    120 		if (sysctl(mib, 2, &usermem, &len, NULL, 0) < 0) {
    121 			msg("sysctl(hw.usermem) failed: %s\n",
    122 			    strerror(errno));
    123 			return;
    124 		}
    125 		cachebufs = (usermem / MAXMEMPART) / CSIZE;
    126 	} else {		/* User specified */
    127 		cachebufs = cachesize;
    128 	}
    129 
    130 	if (cachebufs) {	/* Don't allocate if zero --> no caching */
    131 		if (cachebufs > MAXCACHEBUFS)
    132 			cachebufs = MAXCACHEBUFS;
    133 
    134 		sharedSize = sizeof(union cdesc) +
    135 	   	    sizeof(union cdesc) * cachebufs +
    136 	   	    cachebufs * CSIZE;
    137 #ifdef STATS
    138 		fprintf(stderr, "Using %d buffers (%d bytes)\n", cachebufs,
    139 	   	    sharedSize);
    140 #endif
    141 		shareBuffer = mmap(NULL, sharedSize, PROT_READ | PROT_WRITE,
    142 	   	    MAP_ANON | MAP_SHARED, -1, 0);
    143 		if (shareBuffer == MAP_FAILED) {
    144 			msg("can't mmap shared memory for buffer: %s\n",
    145 			    strerror(errno));
    146 			return;
    147 		}
    148 		cheader = shareBuffer;
    149 		cdesc = (union cdesc *) (((char *) shareBuffer) +
    150 		    sizeof(union cdesc));
    151 		cdata = ((char *) shareBuffer) + sizeof(union cdesc) +
    152 	   	    sizeof(union cdesc) * cachebufs;
    153 
    154 		memset(shareBuffer, '\0', sharedSize);
    155 	}
    156 }
    157 
    158 /*
    159  * Find the cache buffer descriptor that shows the minimal access time
    160  */
    161 static int
    162 findlru(void)
    163 {
    164 	int	i;
    165 	size_t	minTime = cdesc[0].cd_time;
    166 	int	minIdx = 0;
    167 
    168 	for (i = 0; i < cachebufs; i++) {
    169 		if (cdesc[i].cd_time < minTime) {
    170 			minIdx = i;
    171 			minTime = cdesc[i].cd_time;
    172 		}
    173 	}
    174 
    175 	return minIdx;
    176 }
    177 
    178 /*
    179  * Read data directly from disk, with smart error handling.
    180  * Try to recover from hard errors by reading in sector sized pieces.
    181  * Error recovery is attempted at most BREADEMAX times before seeking
    182  * consent from the operator to continue.
    183  */
    184 
    185 static int breaderrors = 0;
    186 #define BREADEMAX 32
    187 
    188 void
    189 rawread(daddr_t blkno, char *buf, int size)
    190 {
    191 	int cnt, i;
    192 
    193 #ifdef STATS
    194 	nphysread++;
    195 	physreadsize += size;
    196 #endif
    197 
    198 loop:
    199 	if (lseek(diskfd, ((off_t) blkno << dev_bshift), SEEK_SET) == -1) {
    200 		msg("rawread: lseek fails\n");
    201 		goto err;
    202 	}
    203 	if ((cnt = read(diskfd, buf, size)) == size)
    204 		return;
    205 	if (blkno + (size >> dev_bshift) > ufsib->ufs_dsize) {
    206 		/*
    207 		 * Trying to read the final fragment.
    208 		 *
    209 		 * NB - dump only works in TP_BSIZE blocks, hence
    210 		 * rounds `dev_bsize' fragments up to TP_BSIZE pieces.
    211 		 * It should be smarter about not actually trying to
    212 		 * read more than it can get, but for the time being
    213 		 * we punt and scale back the read only when it gets
    214 		 * us into trouble. (mkm 9/25/83)
    215 		 */
    216 		size -= dev_bsize;
    217 		goto loop;
    218 	}
    219 	if (cnt == -1)
    220 		msg("read error from %s: %s: [block %lld]: count=%d\n",
    221 		    disk, strerror(errno), (long long)blkno, size);
    222 	else
    223 		msg("short read error from %s: [block %lld]: "
    224 		    "count=%d, got=%d\n",
    225 		    disk, (long long)blkno, size, cnt);
    226 err:
    227 	if (++breaderrors > BREADEMAX) {
    228 		msg("More than %d block read errors from %s\n",
    229 		    BREADEMAX, disk);
    230 		broadcast("DUMP IS AILING!\n");
    231 		msg("This is an unrecoverable error.\n");
    232 		if (!query("Do you want to attempt to continue?")) {
    233 			dumpabort(0);
    234 			/*NOTREACHED*/
    235 		} else
    236 			breaderrors = 0;
    237 	}
    238 	/*
    239 	 * Zero buffer, then try to read each sector of buffer separately.
    240 	 */
    241 	memset(buf, 0, size);
    242 	for (i = 0; i < size; i += dev_bsize, buf += dev_bsize, blkno++) {
    243 		if (lseek(diskfd, ((off_t)blkno << dev_bshift),
    244 		    SEEK_SET) == -1) {
    245 			msg("rawread: lseek2 fails: %s!\n",
    246 			    strerror(errno));
    247 			continue;
    248 		}
    249 		if ((cnt = read(diskfd, buf, (int)dev_bsize)) == dev_bsize)
    250 			continue;
    251 		if (cnt == -1) {
    252 			msg("read error from %s: %s: [sector %lld]: "
    253 			    "count=%ld: %s\n", disk, strerror(errno),
    254 			    (long long)blkno,
    255 			    dev_bsize, strerror(errno));
    256 			continue;
    257 		}
    258 		msg("short read error from %s: [sector %lld]: "
    259 		    "count=%ld, got=%d\n",
    260 		    disk, (long long)blkno, dev_bsize, cnt);
    261 	}
    262 }
    263 
    264 void
    265 bread(daddr_t blkno, char *buf, int size)
    266 {
    267 	int	osize = size, idx;
    268 	daddr_t oblkno = blkno;
    269 	char   *obuf = buf;
    270 	daddr_t numBlocks = howmany(size, dev_bsize);
    271 
    272 #ifdef STATS
    273 	nreads++;
    274 	readsize += size;
    275 #endif
    276 
    277 	if (!shareBuffer) {
    278 		rawread(blkno, buf, size);
    279 		return;
    280 	}
    281 
    282 	if (flock(diskfd, LOCK_EX)) {
    283 		msg("flock(LOCK_EX) failed: %s\n",
    284 		    strerror(errno));
    285 		rawread(blkno, buf, size);
    286 		return;
    287 	}
    288 
    289 retry:
    290 	idx = 0;
    291 	while (size > 0) {
    292 		int	i;
    293 
    294 		for (i = 0; i < cachebufs; i++) {
    295 			union cdesc *curr = &cdesc[(i + idx) % cachebufs];
    296 
    297 #ifdef DIAGNOSTICS
    298 			if (curr->cd_owner) {
    299 				fprintf(stderr, "Owner is set (%d, me=%d), can"
    300 				    "not happen.\n", curr->cd_owner, getpid());
    301 			}
    302 #endif
    303 
    304 			if (curr->cd_blkend == 0)
    305 				continue;
    306 			/*
    307 			 * If we find a bit of the read in the buffers,
    308 			 * now compute how many blocks we can copy,
    309 			 * copy them out, adjust blkno, buf and size,
    310 			 * and restart
    311 			 */
    312 			if (curr->cd_blkstart <= blkno &&
    313 			    blkno < curr->cd_blkend) {
    314 				/* Number of data blocks to be copied */
    315 				int toCopy = MIN(size,
    316 				    (curr->cd_blkend - blkno) << dev_bshift);
    317 #ifdef DIAGNOSTICS
    318 				if (toCopy <= 0 || toCopy > CSIZE) {
    319 					fprintf(stderr, "toCopy %d !\n",
    320 					    toCopy);
    321 					dumpabort(0);
    322 				}
    323 				if (CDATA(curr) +
    324 				    ((blkno - curr->cd_blkstart) <<
    325 				    dev_bshift) < CDATA(curr) ||
    326 			   	    CDATA(curr) +
    327 				    ((blkno - curr->cd_blkstart) <<
    328 			   	    dev_bshift) > CDATA(curr) + CSIZE) {
    329 					fprintf(stderr, "%p < %p !!!\n",
    330 				   	   CDATA(curr) + ((blkno -
    331 					   curr->cd_blkstart) << dev_bshift),
    332 					   CDATA(curr));
    333 					fprintf(stderr,
    334 					    "cdesc[i].cd_blkstart %lld "
    335 					    "blkno %lld dev_bsize %ld\n",
    336 				   	    (long long)curr->cd_blkstart,
    337 					    (long long)blkno,
    338 					    dev_bsize);
    339 					dumpabort(0);
    340 				}
    341 #endif
    342 				memcpy(buf, CDATA(curr) +
    343 				    ((blkno - curr->cd_blkstart) <<
    344 				    dev_bshift),
    345 			   	    toCopy);
    346 
    347 				buf 	+= toCopy;
    348 				size 	-= toCopy;
    349 				blkno 	+= howmany(toCopy, dev_bsize);
    350 				numBlocks -= howmany(toCopy, dev_bsize);
    351 
    352 				curr->cd_time = cheader->cd_count++;
    353 
    354 				/*
    355 				 * If all data of a cache block have been
    356 				 * read, chances are good no more reads
    357 				 * will occur, so expire the cache immediately
    358 				 */
    359 
    360 				curr->cd_blocksRead +=
    361 				    howmany(toCopy, dev_bsize);
    362 				if (curr->cd_blocksRead >= nblksread)
    363 					curr->cd_time = 0;
    364 
    365 				goto retry;
    366 			}
    367 		}
    368 
    369 		/* No more to do? */
    370 		if (size == 0)
    371 			break;
    372 
    373 		/*
    374 		 * This does actually not happen if fs blocks are not greater
    375 		 * than nblksread.
    376 		 */
    377 		if (numBlocks > nblksread || blkno >= ufsib->ufs_dsize) {
    378 			rawread(oblkno, obuf, osize);
    379 			break;
    380 		} else {
    381 			ssize_t	rsize;
    382 			daddr_t	blockBlkNo;
    383 
    384 			blockBlkNo = (blkno / nblksread) * nblksread;
    385 			idx = findlru();
    386 			rsize = MIN(nblksread,
    387 			    ufsib->ufs_dsize - blockBlkNo) << dev_bshift;
    388 
    389 #ifdef DIAGNOSTICS
    390 			if (cdesc[idx].cd_owner)
    391 				fprintf(stderr, "Owner is set (%d, me=%d), can"
    392 				    "not happen(2).\n", cdesc[idx].cd_owner,
    393 				    getpid());
    394 			cdesc[idx].cd_owner = getpid();
    395 #endif
    396 			cdesc[idx].cd_time = cheader->cd_count++;
    397 			cdesc[idx].cd_blkstart = blockBlkNo;
    398 			cdesc[idx].cd_blkend = 0;
    399 			cdesc[idx].cd_blocksRead = 0;
    400 
    401 			if (lseek(diskfd, ((off_t) blockBlkNo << dev_bshift),
    402 			    SEEK_SET) == -1) {
    403 				msg("readBlocks: lseek fails: %s\n",
    404 				    strerror(errno));
    405 				rsize = -1;
    406 			} else {
    407 				rsize = read(diskfd,
    408 				    CDATA(&cdesc[idx]), rsize);
    409 				if (rsize < 0) {
    410 					msg("readBlocks: read fails: %s\n",
    411 					    strerror(errno));
    412 				}
    413 			}
    414 
    415 			/* On errors, panic, punt, try to read without
    416 			 * cache and let raw read routine do the rest.
    417 			 */
    418 
    419 			if (rsize <= 0) {
    420 				rawread(oblkno, obuf, osize);
    421 #ifdef DIAGNOSTICS
    422 				if (cdesc[idx].cd_owner != getpid())
    423 					fprintf(stderr, "Owner changed from "
    424 					    "%d to %d, can't happen\n",
    425 					    getpid(), cdesc[idx].cd_owner);
    426 				cdesc[idx].cd_owner = 0;
    427 #endif
    428 				break;
    429 			}
    430 
    431 			/* On short read, just note the fact and go on */
    432 			cdesc[idx].cd_blkend = blockBlkNo + rsize / dev_bsize;
    433 
    434 #ifdef STATS
    435 			nphysread++;
    436 			physreadsize += rsize;
    437 #endif
    438 #ifdef DIAGNOSTICS
    439 			if (cdesc[idx].cd_owner != getpid())
    440 				fprintf(stderr, "Owner changed from "
    441 				    "%d to %d, can't happen\n",
    442 				    getpid(), cdesc[idx].cd_owner);
    443 			cdesc[idx].cd_owner = 0;
    444 #endif
    445 			/*
    446 			 * We swapped some of data in, let the loop fetch
    447 			 * them from cache
    448 			 */
    449 		}
    450 	}
    451 
    452 	if (flock(diskfd, LOCK_UN))
    453 		msg("flock(LOCK_UN) failed: %s\n",
    454 		    strerror(errno));
    455 }
    456 
    457 void
    458 printcachestats(void)
    459 {
    460 
    461 #ifdef STATS
    462 	fprintf(stderr, "Pid %d: %d reads (%u bytes) "
    463 	    "%d physical reads (%u bytes) %d%% hits, %d%% overhead\n",
    464 	    getpid(), nreads, (u_int) readsize, nphysread,
    465 	    (u_int) physreadsize, (nreads - nphysread) * 100 / nreads,
    466 	    (int) (((physreadsize - readsize) * 100) / readsize));
    467 #endif
    468 }
    469