Home | History | Annotate | Line # | Download | only in dump
rcache.c revision 1.23.8.1
      1 /*	$NetBSD: rcache.c,v 1.23.8.1 2016/03/08 10:03:57 snj 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  *
     20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     30  * POSSIBILITY OF SUCH DAMAGE.
     31  */
     32 
     33 #include <sys/cdefs.h>
     34 #ifndef lint
     35 __RCSID("$NetBSD: rcache.c,v 1.23.8.1 2016/03/08 10:03:57 snj Exp $");
     36 #endif /* not lint */
     37 
     38 #include <sys/types.h>
     39 #include <sys/uio.h>
     40 #include <sys/mman.h>
     41 #include <sys/param.h>
     42 #include <sys/sysctl.h>
     43 #include <ufs/ufs/dinode.h>
     44 
     45 #include <stdio.h>
     46 #include <stdlib.h>
     47 #include <unistd.h>
     48 #include <fcntl.h>
     49 #include <errno.h>
     50 #include <string.h>
     51 
     52 #include "dump.h"
     53 
     54 /*-----------------------------------------------------------------------*/
     55 #define MAXCACHEBUFS	512	/* max 512 buffers */
     56 #define MAXMEMPART	6	/* max 15% of the user mem */
     57 
     58 /*-----------------------------------------------------------------------*/
     59 union cdesc {
     60 	volatile size_t cd_count;
     61 	struct {
     62 		volatile daddr_t blkstart;
     63 		volatile daddr_t blkend;	/* start + nblksread */
     64 		volatile daddr_t blocksRead;
     65 		volatile size_t time;
     66 #ifdef DIAGNOSTICS
     67 		volatile pid_t owner;
     68 #endif
     69 	} desc;
     70 #define cd_blkstart	desc.blkstart
     71 #define cd_blkend	desc.blkend
     72 #define cd_blocksRead	desc.blocksRead
     73 #define cd_time		desc.time
     74 #define cd_owner	desc.owner
     75 };
     76 
     77 static int findlru(void);
     78 
     79 static void *shareBuffer = NULL;
     80 static union cdesc *cheader;
     81 static union cdesc *cdesc;
     82 static char *cdata;
     83 static int cachebufs;
     84 static int nblksread;
     85 
     86 #ifdef STATS
     87 static int nreads;
     88 static int nphysread;
     89 static int64_t readsize;
     90 static int64_t physreadsize;
     91 #endif
     92 
     93 #define	CSIZE		(nblksread << dev_bshift)	/* cache buf size */
     94 #define	CDATA(desc)	(cdata + ((desc) - cdesc) * CSIZE)
     95 
     96 void
     97 initcache(int cachesize, int readblksize)
     98 {
     99 	size_t len;
    100 	size_t sharedSize;
    101 
    102 	if (readblksize == -1) { /* use kern.maxphys */
    103 		int kern_maxphys;
    104 		int mib[2] = { CTL_KERN, KERN_MAXPHYS };
    105 
    106 		len = sizeof(kern_maxphys);
    107 		if (sysctl(mib, 2, &kern_maxphys, &len, NULL, 0) < 0) {
    108 			msg("sysctl(kern.maxphys) failed: %s\n",
    109 			    strerror(errno));
    110 			return;
    111 		}
    112 		readblksize = kern_maxphys;
    113 	}
    114 
    115 	/* Convert read block size in terms of filesystem block size */
    116 	nblksread = howmany(readblksize, ufsib->ufs_bsize);
    117 
    118 	/* Then, convert it in terms of device block size */
    119 	nblksread <<= ufsib->ufs_bshift - dev_bshift;
    120 
    121 	if (cachesize == -1) {	/* Compute from memory available */
    122 		uint64_t usermem, cachetmp;
    123 		int mib[2] = { CTL_HW, HW_USERMEM64 };
    124 
    125 		len = sizeof(usermem);
    126 		if (sysctl(mib, 2, &usermem, &len, NULL, 0) < 0) {
    127 			msg("sysctl(hw.usermem) failed: %s\n",
    128 			    strerror(errno));
    129 			return;
    130 		}
    131 		cachetmp = (usermem / MAXMEMPART) / CSIZE;
    132 		/* for those with TB of RAM */
    133 		cachebufs = (cachetmp > INT_MAX) ? INT_MAX : cachetmp;
    134 	} else {		/* User specified */
    135 		cachebufs = cachesize;
    136 	}
    137 
    138 	if (cachebufs) {	/* Don't allocate if zero --> no caching */
    139 		if (cachebufs > MAXCACHEBUFS)
    140 			cachebufs = MAXCACHEBUFS;
    141 
    142 		sharedSize = sizeof(union cdesc) +
    143 	   	    sizeof(union cdesc) * cachebufs +
    144 	   	    cachebufs * CSIZE;
    145 #ifdef STATS
    146 		fprintf(stderr, "Using %d buffers (%d bytes)\n", cachebufs,
    147 	   	    sharedSize);
    148 #endif
    149 		shareBuffer = mmap(NULL, sharedSize, PROT_READ | PROT_WRITE,
    150 	   	    MAP_ANON | MAP_SHARED, -1, 0);
    151 		if (shareBuffer == MAP_FAILED) {
    152 			msg("can't mmap shared memory for buffer: %s\n",
    153 			    strerror(errno));
    154 			return;
    155 		}
    156 		cheader = shareBuffer;
    157 		cdesc = (union cdesc *) (((char *) shareBuffer) +
    158 		    sizeof(union cdesc));
    159 		cdata = ((char *) shareBuffer) + sizeof(union cdesc) +
    160 	   	    sizeof(union cdesc) * cachebufs;
    161 
    162 		memset(shareBuffer, '\0', sharedSize);
    163 	}
    164 }
    165 
    166 /*
    167  * Find the cache buffer descriptor that shows the minimal access time
    168  */
    169 static int
    170 findlru(void)
    171 {
    172 	int	i;
    173 	size_t	minTime = cdesc[0].cd_time;
    174 	int	minIdx = 0;
    175 
    176 	for (i = 0; i < cachebufs; i++) {
    177 		if (cdesc[i].cd_time < minTime) {
    178 			minIdx = i;
    179 			minTime = cdesc[i].cd_time;
    180 		}
    181 	}
    182 
    183 	return minIdx;
    184 }
    185 
    186 /*
    187  * Read data directly from disk, with smart error handling.
    188  * Try to recover from hard errors by reading in sector sized pieces.
    189  * Error recovery is attempted at most BREADEMAX times before seeking
    190  * consent from the operator to continue.
    191  */
    192 
    193 static int breaderrors = 0;
    194 #define BREADEMAX 32
    195 
    196 void
    197 rawread(daddr_t blkno, char *buf, int size)
    198 {
    199 	int cnt, i;
    200 
    201 #ifdef STATS
    202 	nphysread++;
    203 	physreadsize += size;
    204 #endif
    205 
    206 loop:
    207 	if (lseek(diskfd, ((off_t) blkno << dev_bshift), SEEK_SET) == -1) {
    208 		msg("rawread: lseek fails\n");
    209 		goto err;
    210 	}
    211 	if ((cnt = read(diskfd, buf, size)) == size)
    212 		return;
    213 	if (blkno + (size >> dev_bshift) > ufsib->ufs_dsize) {
    214 		/*
    215 		 * Trying to read the final fragment.
    216 		 *
    217 		 * NB - dump only works in TP_BSIZE blocks, hence
    218 		 * rounds `dev_bsize' fragments up to TP_BSIZE pieces.
    219 		 * It should be smarter about not actually trying to
    220 		 * read more than it can get, but for the time being
    221 		 * we punt and scale back the read only when it gets
    222 		 * us into trouble. (mkm 9/25/83)
    223 		 */
    224 		size -= dev_bsize;
    225 		goto loop;
    226 	}
    227 	if (cnt == -1)
    228 		msg("read error from %s: %s: [block %lld]: count=%d\n",
    229 		    disk, strerror(errno), (long long)blkno, size);
    230 	else
    231 		msg("short read error from %s: [block %lld]: "
    232 		    "count=%d, got=%d\n",
    233 		    disk, (long long)blkno, size, cnt);
    234 err:
    235 	if (++breaderrors > BREADEMAX) {
    236 		msg("More than %d block read errors from %s\n",
    237 		    BREADEMAX, disk);
    238 		broadcast("DUMP IS AILING!\n");
    239 		msg("This is an unrecoverable error.\n");
    240 		if (!query("Do you want to attempt to continue?")) {
    241 			dumpabort(0);
    242 			/*NOTREACHED*/
    243 		} else
    244 			breaderrors = 0;
    245 	}
    246 	/*
    247 	 * Zero buffer, then try to read each sector of buffer separately.
    248 	 */
    249 	memset(buf, 0, size);
    250 	for (i = 0; i < size; i += dev_bsize, buf += dev_bsize, blkno++) {
    251 		if (lseek(diskfd, ((off_t)blkno << dev_bshift),
    252 		    SEEK_SET) == -1) {
    253 			msg("rawread: lseek2 fails: %s!\n",
    254 			    strerror(errno));
    255 			continue;
    256 		}
    257 		if ((cnt = read(diskfd, buf, (int)dev_bsize)) == dev_bsize)
    258 			continue;
    259 		if (cnt == -1) {
    260 			msg("read error from %s: %s: [sector %lld]: "
    261 			    "count=%ld\n", disk, strerror(errno),
    262 			    (long long)blkno, dev_bsize);
    263 			continue;
    264 		}
    265 		msg("short read error from %s: [sector %lld]: "
    266 		    "count=%ld, got=%d\n",
    267 		    disk, (long long)blkno, dev_bsize, cnt);
    268 	}
    269 }
    270 
    271 void
    272 bread(daddr_t blkno, char *buf, int size)
    273 {
    274 	int	osize = size, idx;
    275 	daddr_t oblkno = blkno;
    276 	char   *obuf = buf;
    277 	daddr_t numBlocks = howmany(size, dev_bsize);
    278 
    279 #ifdef STATS
    280 	nreads++;
    281 	readsize += size;
    282 #endif
    283 
    284 	if (!shareBuffer) {
    285 		rawread(blkno, buf, size);
    286 		return;
    287 	}
    288 
    289 	if (flock(diskfd, LOCK_EX)) {
    290 		msg("flock(LOCK_EX) failed: %s\n",
    291 		    strerror(errno));
    292 		rawread(blkno, buf, size);
    293 		return;
    294 	}
    295 
    296 retry:
    297 	idx = 0;
    298 	while (size > 0) {
    299 		int	i;
    300 
    301 		for (i = 0; i < cachebufs; i++) {
    302 			union cdesc *curr = &cdesc[(i + idx) % cachebufs];
    303 
    304 #ifdef DIAGNOSTICS
    305 			if (curr->cd_owner) {
    306 				fprintf(stderr, "Owner is set (%d, me=%d), can"
    307 				    "not happen.\n", curr->cd_owner, getpid());
    308 			}
    309 #endif
    310 
    311 			if (curr->cd_blkend == 0)
    312 				continue;
    313 			/*
    314 			 * If we find a bit of the read in the buffers,
    315 			 * now compute how many blocks we can copy,
    316 			 * copy them out, adjust blkno, buf and size,
    317 			 * and restart
    318 			 */
    319 			if (curr->cd_blkstart <= blkno &&
    320 			    blkno < curr->cd_blkend) {
    321 				/* Number of data blocks to be copied */
    322 				int toCopy = MIN(size,
    323 				    (curr->cd_blkend - blkno) << dev_bshift);
    324 #ifdef DIAGNOSTICS
    325 				if (toCopy <= 0 || toCopy > CSIZE) {
    326 					fprintf(stderr, "toCopy %d !\n",
    327 					    toCopy);
    328 					dumpabort(0);
    329 				}
    330 				if (CDATA(curr) +
    331 				    ((blkno - curr->cd_blkstart) <<
    332 				    dev_bshift) < CDATA(curr) ||
    333 			   	    CDATA(curr) +
    334 				    ((blkno - curr->cd_blkstart) <<
    335 			   	    dev_bshift) > CDATA(curr) + CSIZE) {
    336 					fprintf(stderr, "%p < %p !!!\n",
    337 				   	   CDATA(curr) + ((blkno -
    338 					   curr->cd_blkstart) << dev_bshift),
    339 					   CDATA(curr));
    340 					fprintf(stderr,
    341 					    "cdesc[i].cd_blkstart %lld "
    342 					    "blkno %lld dev_bsize %ld\n",
    343 				   	    (long long)curr->cd_blkstart,
    344 					    (long long)blkno,
    345 					    dev_bsize);
    346 					dumpabort(0);
    347 				}
    348 #endif
    349 				memcpy(buf, CDATA(curr) +
    350 				    ((blkno - curr->cd_blkstart) <<
    351 				    dev_bshift),
    352 			   	    toCopy);
    353 
    354 				buf 	+= toCopy;
    355 				size 	-= toCopy;
    356 				blkno 	+= howmany(toCopy, dev_bsize);
    357 				numBlocks -= howmany(toCopy, dev_bsize);
    358 
    359 				curr->cd_time = cheader->cd_count++;
    360 
    361 				/*
    362 				 * If all data of a cache block have been
    363 				 * read, chances are good no more reads
    364 				 * will occur, so expire the cache immediately
    365 				 */
    366 
    367 				curr->cd_blocksRead +=
    368 				    howmany(toCopy, dev_bsize);
    369 				if (curr->cd_blocksRead >= nblksread)
    370 					curr->cd_time = 0;
    371 
    372 				goto retry;
    373 			}
    374 		}
    375 
    376 		/* No more to do? */
    377 		if (size == 0)
    378 			break;
    379 
    380 		/*
    381 		 * This does actually not happen if fs blocks are not greater
    382 		 * than nblksread.
    383 		 */
    384 		if (numBlocks > nblksread || blkno >= ufsib->ufs_dsize) {
    385 			rawread(oblkno, obuf, osize);
    386 			break;
    387 		} else {
    388 			ssize_t	rsize;
    389 			daddr_t	blockBlkNo;
    390 
    391 			blockBlkNo = (blkno / nblksread) * nblksread;
    392 			idx = findlru();
    393 			rsize = MIN(nblksread,
    394 			    ufsib->ufs_dsize - blockBlkNo) << dev_bshift;
    395 
    396 #ifdef DIAGNOSTICS
    397 			if (cdesc[idx].cd_owner)
    398 				fprintf(stderr, "Owner is set (%d, me=%d), can"
    399 				    "not happen(2).\n", cdesc[idx].cd_owner,
    400 				    getpid());
    401 			cdesc[idx].cd_owner = getpid();
    402 #endif
    403 			cdesc[idx].cd_time = cheader->cd_count++;
    404 			cdesc[idx].cd_blkstart = blockBlkNo;
    405 			cdesc[idx].cd_blkend = 0;
    406 			cdesc[idx].cd_blocksRead = 0;
    407 
    408 			if (lseek(diskfd, ((off_t) blockBlkNo << dev_bshift),
    409 			    SEEK_SET) == -1) {
    410 				msg("readBlocks: lseek fails: %s\n",
    411 				    strerror(errno));
    412 				rsize = -1;
    413 			} else {
    414 				rsize = read(diskfd,
    415 				    CDATA(&cdesc[idx]), rsize);
    416 				if (rsize < 0) {
    417 					msg("readBlocks: read fails: %s\n",
    418 					    strerror(errno));
    419 				}
    420 			}
    421 
    422 			/* On errors, panic, punt, try to read without
    423 			 * cache and let raw read routine do the rest.
    424 			 */
    425 
    426 			if (rsize <= 0) {
    427 				rawread(oblkno, obuf, osize);
    428 #ifdef DIAGNOSTICS
    429 				if (cdesc[idx].cd_owner != getpid())
    430 					fprintf(stderr, "Owner changed from "
    431 					    "%d to %d, can't happen\n",
    432 					    getpid(), cdesc[idx].cd_owner);
    433 				cdesc[idx].cd_owner = 0;
    434 #endif
    435 				break;
    436 			}
    437 
    438 			/* On short read, just note the fact and go on */
    439 			cdesc[idx].cd_blkend = blockBlkNo + rsize / dev_bsize;
    440 
    441 #ifdef STATS
    442 			nphysread++;
    443 			physreadsize += rsize;
    444 #endif
    445 #ifdef DIAGNOSTICS
    446 			if (cdesc[idx].cd_owner != getpid())
    447 				fprintf(stderr, "Owner changed from "
    448 				    "%d to %d, can't happen\n",
    449 				    getpid(), cdesc[idx].cd_owner);
    450 			cdesc[idx].cd_owner = 0;
    451 #endif
    452 			/*
    453 			 * We swapped some of data in, let the loop fetch
    454 			 * them from cache
    455 			 */
    456 		}
    457 	}
    458 
    459 	if (flock(diskfd, LOCK_UN))
    460 		msg("flock(LOCK_UN) failed: %s\n",
    461 		    strerror(errno));
    462 }
    463 
    464 void
    465 printcachestats(void)
    466 {
    467 
    468 #ifdef STATS
    469 	fprintf(stderr, "Pid %d: %d reads (%u bytes) "
    470 	    "%d physical reads (%u bytes) %d%% hits, %d%% overhead\n",
    471 	    getpid(), nreads, (u_int) readsize, nphysread,
    472 	    (u_int) physreadsize, (nreads - nphysread) * 100 / nreads,
    473 	    (int) (((physreadsize - readsize) * 100) / readsize));
    474 #endif
    475 }
    476