1/* $NetBSD: lfs_cleanerd.c,v 1.28 2012/01/02 21:35:18 perseant Exp $ */ 2 3/*- 4 * Copyright (c) 2005 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Konrad E. Schroder <perseant@hhhh.org>. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32/* 33 * The cleaner daemon for the NetBSD Log-structured File System. 34 * Only tested for use with version 2 LFSs. 35 */ 36 37#include <sys/syslog.h> 38#include <sys/param.h> 39#include <sys/mount.h> 40#include <sys/stat.h> 41#include <ufs/ufs/inode.h> 42#include <ufs/lfs/lfs.h> 43 44#include <assert.h> 45#include <err.h> 46#include <errno.h> 47#include <fcntl.h> 48#include <semaphore.h> 49#include <stdio.h> 50#include <stdlib.h> 51#include <string.h> 52#include <unistd.h> 53#include <time.h> 54#include <util.h> 55 56#include "bufcache.h" 57#include "vnode.h" 58#include "lfs_user.h" 59#include "fdfs.h" 60#include "cleaner.h" 61#include "kernelops.h" 62#include "mount_lfs.h" 63 64/* 65 * Global variables. 66 */ 67/* XXX these top few should really be fs-specific */ 68int use_fs_idle; /* Use fs idle rather than cpu idle time */ 69int use_bytes; /* Use bytes written rather than segments cleaned */ 70int load_threshold; /* How idle is idle (CPU idle) */ 71int atatime; /* How many segments (bytes) to clean at a time */ 72 73int nfss; /* Number of filesystems monitored by this cleanerd */ 74struct clfs **fsp; /* Array of extended filesystem structures */ 75int segwait_timeout; /* Time to wait in lfs_segwait() */ 76int do_quit; /* Quit after one cleaning loop */ 77int do_coalesce; /* Coalesce filesystem */ 78int do_small; /* Use small writes through markv */ 79char *copylog_filename; /* File to use for fs debugging analysis */ 80int inval_segment; /* Segment to invalidate */ 81int stat_report; /* Report statistics for this period of cycles */ 82int debug; /* Turn on debugging */ 83struct cleaner_stats { 84 double util_tot; 85 double util_sos; 86 off_t bytes_read; 87 off_t bytes_written; 88 off_t segs_cleaned; 89 off_t segs_empty; 90 off_t segs_error; 91} cleaner_stats; 92 93extern u_int32_t cksum(void *, size_t); 94extern u_int32_t lfs_sb_cksum(struct dlfs *); 95extern u_int32_t lfs_cksum_part(void *, size_t, u_int32_t); 96extern int ufs_getlbns(struct lfs *, struct uvnode *, daddr_t, struct indir *, int *); 97 98/* Compat */ 99void pwarn(const char *unused, ...) { /* Does nothing */ }; 100 101/* 102 * Log a message if debugging is turned on. 103 */ 104void 105dlog(const char *fmt, ...) 106{ 107 va_list ap; 108 109 if (debug == 0) 110 return; 111 112 va_start(ap, fmt); 113 vsyslog(LOG_DEBUG, fmt, ap); 114 va_end(ap); 115} 116 117/* 118 * Remove the specified filesystem from the list, due to its having 119 * become unmounted or other error condition. 120 */ 121void 122handle_error(struct clfs **cfsp, int n) 123{ 124 syslog(LOG_NOTICE, "%s: detaching cleaner", cfsp[n]->lfs_fsmnt); 125 free(cfsp[n]); 126 if (n != nfss - 1) 127 cfsp[n] = cfsp[nfss - 1]; 128 --nfss; 129} 130 131/* 132 * Reinitialize a filesystem if, e.g., its size changed. 133 */ 134int 135reinit_fs(struct clfs *fs) 136{ 137 char fsname[MNAMELEN]; 138 139 strncpy(fsname, (char *)fs->lfs_fsmnt, MNAMELEN); 140 kops.ko_close(fs->clfs_ifilefd); 141 kops.ko_close(fs->clfs_devfd); 142 fd_reclaim(fs->clfs_devvp); 143 fd_reclaim(fs->lfs_ivnode); 144 free(fs->clfs_dev); 145 free(fs->clfs_segtab); 146 free(fs->clfs_segtabp); 147 148 return init_fs(fs, fsname); 149} 150 151#ifdef REPAIR_ZERO_FINFO 152/* 153 * Use fsck's lfs routines to load the Ifile from an unmounted fs. 154 * We interpret "fsname" as the name of the raw disk device. 155 */ 156int 157init_unmounted_fs(struct clfs *fs, char *fsname) 158{ 159 struct lfs *disc_fs; 160 int i; 161 162 fs->clfs_dev = fsname; 163 if ((fs->clfs_devfd = kops.ko_open(fs->clfs_dev, O_RDWR)) < 0) { 164 syslog(LOG_ERR, "couldn't open device %s read/write", 165 fs->clfs_dev); 166 return -1; 167 } 168 169 disc_fs = lfs_init(fs->clfs_devfd, 0, 0, 0, 0); 170 171 fs->lfs_dlfs = disc_fs->lfs_dlfs; /* Structure copy */ 172 strncpy(fs->lfs_fsmnt, fsname, MNAMELEN); 173 fs->lfs_ivnode = (struct uvnode *)disc_fs->lfs_ivnode; 174 fs->clfs_devvp = fd_vget(fs->clfs_devfd, fs->lfs_fsize, fs->lfs_ssize, 175 atatime); 176 177 /* Allocate and clear segtab */ 178 fs->clfs_segtab = (struct clfs_seguse *)malloc(fs->lfs_nseg * 179 sizeof(*fs->clfs_segtab)); 180 fs->clfs_segtabp = (struct clfs_seguse **)malloc(fs->lfs_nseg * 181 sizeof(*fs->clfs_segtabp)); 182 for (i = 0; i < fs->lfs_nseg; i++) { 183 fs->clfs_segtabp[i] = &(fs->clfs_segtab[i]); 184 fs->clfs_segtab[i].flags = 0x0; 185 } 186 syslog(LOG_NOTICE, "%s: unmounted cleaner starting", fsname); 187 188 return 0; 189} 190#endif 191 192/* 193 * Set up the file descriptors, including the Ifile descriptor. 194 * If we can't get the Ifile, this is not an LFS (or the kernel is 195 * too old to support the fcntl). 196 * XXX Merge this and init_unmounted_fs, switching on whether 197 * XXX "fsname" is a dir or a char special device. Should 198 * XXX also be able to read unmounted devices out of fstab, the way 199 * XXX fsck does. 200 */ 201int 202init_fs(struct clfs *fs, char *fsname) 203{ 204 struct statvfs sf; 205 int rootfd; 206 int i; 207 void *sbuf; 208 char *bn; 209 210 /* 211 * Get the raw device from the block device. 212 * XXX this is ugly. Is there a way to discover the raw device 213 * XXX for a given mount point? 214 */ 215 if (kops.ko_statvfs(fsname, &sf, ST_WAIT) < 0) 216 return -1; 217 fs->clfs_dev = malloc(strlen(sf.f_mntfromname) + 2); 218 if (fs->clfs_dev == NULL) { 219 syslog(LOG_ERR, "couldn't malloc device name string: %m"); 220 return -1; 221 } 222 bn = strrchr(sf.f_mntfromname, '/'); 223 bn = bn ? bn+1 : sf.f_mntfromname; 224 strlcpy(fs->clfs_dev, sf.f_mntfromname, bn - sf.f_mntfromname + 1); 225 strcat(fs->clfs_dev, "r"); 226 strcat(fs->clfs_dev, bn); 227 if ((fs->clfs_devfd = kops.ko_open(fs->clfs_dev, O_RDONLY, 0)) < 0) { 228 syslog(LOG_ERR, "couldn't open device %s for reading", 229 fs->clfs_dev); 230 return -1; 231 } 232 233 /* Find the Ifile and open it */ 234 if ((rootfd = kops.ko_open(fsname, O_RDONLY, 0)) < 0) 235 return -2; 236 if (kops.ko_fcntl(rootfd, LFCNIFILEFH, &fs->clfs_ifilefh) < 0) 237 return -3; 238 if ((fs->clfs_ifilefd = kops.ko_fhopen(&fs->clfs_ifilefh, 239 sizeof(fs->clfs_ifilefh), O_RDONLY)) < 0) 240 return -4; 241 kops.ko_close(rootfd); 242 243 sbuf = malloc(LFS_SBPAD); 244 if (sbuf == NULL) { 245 syslog(LOG_ERR, "couldn't malloc superblock buffer"); 246 return -1; 247 } 248 249 /* Load in the superblock */ 250 if (kops.ko_pread(fs->clfs_devfd, sbuf, LFS_SBPAD, LFS_LABELPAD) < 0) { 251 free(sbuf); 252 return -1; 253 } 254 255 memcpy(&(fs->lfs_dlfs), sbuf, sizeof(struct dlfs)); 256 free(sbuf); 257 258 /* If this is not a version 2 filesystem, complain and exit */ 259 if (fs->lfs_version != 2) { 260 syslog(LOG_ERR, "%s: not a version 2 LFS", fsname); 261 return -1; 262 } 263 264 /* Assume fsname is the mounted name */ 265 strncpy((char *)fs->lfs_fsmnt, fsname, MNAMELEN); 266 267 /* Set up vnodes for Ifile and raw device */ 268 fs->lfs_ivnode = fd_vget(fs->clfs_ifilefd, fs->lfs_bsize, 0, 0); 269 fs->clfs_devvp = fd_vget(fs->clfs_devfd, fs->lfs_fsize, fs->lfs_ssize, 270 atatime); 271 272 /* Allocate and clear segtab */ 273 fs->clfs_segtab = (struct clfs_seguse *)malloc(fs->lfs_nseg * 274 sizeof(*fs->clfs_segtab)); 275 fs->clfs_segtabp = (struct clfs_seguse **)malloc(fs->lfs_nseg * 276 sizeof(*fs->clfs_segtabp)); 277 if (fs->clfs_segtab == NULL || fs->clfs_segtabp == NULL) { 278 syslog(LOG_ERR, "%s: couldn't malloc segment table: %m", 279 fs->clfs_dev); 280 return -1; 281 } 282 283 for (i = 0; i < fs->lfs_nseg; i++) { 284 fs->clfs_segtabp[i] = &(fs->clfs_segtab[i]); 285 fs->clfs_segtab[i].flags = 0x0; 286 } 287 288 syslog(LOG_NOTICE, "%s: attaching cleaner", fsname); 289 return 0; 290} 291 292/* 293 * Invalidate all the currently held Ifile blocks so they will be 294 * reread when we clean. Check the size while we're at it, and 295 * resize the buffer cache if necessary. 296 */ 297void 298reload_ifile(struct clfs *fs) 299{ 300 struct ubuf *bp; 301 struct stat st; 302 int ohashmax; 303 extern int hashmax; 304 305 while ((bp = LIST_FIRST(&fs->lfs_ivnode->v_dirtyblkhd)) != NULL) { 306 bremfree(bp); 307 buf_destroy(bp); 308 } 309 while ((bp = LIST_FIRST(&fs->lfs_ivnode->v_cleanblkhd)) != NULL) { 310 bremfree(bp); 311 buf_destroy(bp); 312 } 313 314 /* If Ifile is larger than buffer cache, rehash */ 315 fstat(fs->clfs_ifilefd, &st); 316 if (st.st_size / fs->lfs_bsize > hashmax) { 317 ohashmax = hashmax; 318 bufrehash(st.st_size / fs->lfs_bsize); 319 dlog("%s: resized buffer hash from %d to %d", 320 fs->lfs_fsmnt, ohashmax, hashmax); 321 } 322} 323 324/* 325 * Get IFILE entry for the given inode, store in ifpp. The buffer 326 * which contains that data is returned in bpp, and must be brelse()d 327 * by the caller. 328 */ 329void 330lfs_ientry(IFILE **ifpp, struct clfs *fs, ino_t ino, struct ubuf **bpp) 331{ 332 int error; 333 334 error = bread(fs->lfs_ivnode, ino / fs->lfs_ifpb + fs->lfs_cleansz + 335 fs->lfs_segtabsz, fs->lfs_bsize, NOCRED, 0, bpp); 336 if (error) 337 syslog(LOG_ERR, "%s: ientry failed for ino %d", 338 fs->lfs_fsmnt, (int)ino); 339 *ifpp = (IFILE *)(*bpp)->b_data + ino % fs->lfs_ifpb; 340 return; 341} 342 343#ifdef TEST_PATTERN 344/* 345 * Check ROOTINO for file data. The assumption is that we are running 346 * the "twofiles" test with the rest of the filesystem empty. Files 347 * created by "twofiles" match the test pattern, but ROOTINO and the 348 * executable itself (assumed to be inode 3) should not match. 349 */ 350static void 351check_test_pattern(BLOCK_INFO *bip) 352{ 353 int j; 354 unsigned char *cp = bip->bi_bp; 355 356 /* Check inode sanity */ 357 if (bip->bi_lbn == LFS_UNUSED_LBN) { 358 assert(((struct ufs1_dinode *)bip->bi_bp)->di_inumber == 359 bip->bi_inode); 360 } 361 362 /* These can have the test pattern and it's all good */ 363 if (bip->bi_inode > 3) 364 return; 365 366 for (j = 0; j < bip->bi_size; j++) { 367 if (cp[j] != (j & 0xff)) 368 break; 369 } 370 assert(j < bip->bi_size); 371} 372#endif /* TEST_PATTERN */ 373 374/* 375 * Parse the partial segment at daddr, adding its information to 376 * bip. Return the address of the next partial segment to read. 377 */ 378int32_t 379parse_pseg(struct clfs *fs, daddr_t daddr, BLOCK_INFO **bipp, int *bic) 380{ 381 SEGSUM *ssp; 382 IFILE *ifp; 383 BLOCK_INFO *bip, *nbip; 384 int32_t *iaddrp, idaddr, odaddr; 385 FINFO *fip; 386 struct ubuf *ifbp; 387 struct ufs1_dinode *dip; 388 u_int32_t ck, vers; 389 int fic, inoc, obic; 390 int i; 391 char *cp; 392 393 odaddr = daddr; 394 obic = *bic; 395 bip = *bipp; 396 397 /* 398 * Retrieve the segment header, set up the SEGSUM pointer 399 * as well as the first FINFO and inode address pointer. 400 */ 401 cp = fd_ptrget(fs->clfs_devvp, daddr); 402 ssp = (SEGSUM *)cp; 403 iaddrp = ((int32_t *)(cp + fs->lfs_ibsize)) - 1; 404 fip = (FINFO *)(cp + sizeof(SEGSUM)); 405 406 /* 407 * Check segment header magic and checksum 408 */ 409 if (ssp->ss_magic != SS_MAGIC) { 410 syslog(LOG_WARNING, "%s: sumsum magic number bad at 0x%x:" 411 " read 0x%x, expected 0x%x", fs->lfs_fsmnt, 412 (int32_t)daddr, ssp->ss_magic, SS_MAGIC); 413 return 0x0; 414 } 415 ck = cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum)); 416 if (ck != ssp->ss_sumsum) { 417 syslog(LOG_WARNING, "%s: sumsum checksum mismatch at 0x%x:" 418 " read 0x%x, computed 0x%x", fs->lfs_fsmnt, 419 (int32_t)daddr, ssp->ss_sumsum, ck); 420 return 0x0; 421 } 422 423 /* Initialize data sum */ 424 ck = 0; 425 426 /* Point daddr at next block after segment summary */ 427 ++daddr; 428 429 /* 430 * Loop over file info and inode pointers. We always move daddr 431 * forward here because we are also computing the data checksum 432 * as we go. 433 */ 434 fic = inoc = 0; 435 while (fic < ssp->ss_nfinfo || inoc < ssp->ss_ninos) { 436 /* 437 * We must have either a file block or an inode block. 438 * If we don't have either one, it's an error. 439 */ 440 if (fic >= ssp->ss_nfinfo && *iaddrp != daddr) { 441 syslog(LOG_WARNING, "%s: bad pseg at %x (seg %d)", 442 fs->lfs_fsmnt, odaddr, dtosn(fs, odaddr)); 443 *bipp = bip; 444 return 0x0; 445 } 446 447 /* 448 * Note each inode from the inode blocks 449 */ 450 if (inoc < ssp->ss_ninos && *iaddrp == daddr) { 451 cp = fd_ptrget(fs->clfs_devvp, daddr); 452 ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck); 453 dip = (struct ufs1_dinode *)cp; 454 for (i = 0; i < fs->lfs_inopb; i++) { 455 if (dip[i].di_inumber == 0) 456 break; 457 458 /* 459 * Check currency before adding it 460 */ 461#ifndef REPAIR_ZERO_FINFO 462 lfs_ientry(&ifp, fs, dip[i].di_inumber, &ifbp); 463 idaddr = ifp->if_daddr; 464 brelse(ifbp, 0); 465 if (idaddr != daddr) 466#endif 467 continue; 468 469 /* 470 * A current inode. Add it. 471 */ 472 ++*bic; 473 nbip = (BLOCK_INFO *)realloc(bip, *bic * 474 sizeof(*bip)); 475 if (nbip) 476 bip = nbip; 477 else { 478 --*bic; 479 *bipp = bip; 480 return 0x0; 481 } 482 bip[*bic - 1].bi_inode = dip[i].di_inumber; 483 bip[*bic - 1].bi_lbn = LFS_UNUSED_LBN; 484 bip[*bic - 1].bi_daddr = daddr; 485 bip[*bic - 1].bi_segcreate = ssp->ss_create; 486 bip[*bic - 1].bi_version = dip[i].di_gen; 487 bip[*bic - 1].bi_bp = &(dip[i]); 488 bip[*bic - 1].bi_size = DINODE1_SIZE; 489 } 490 inoc += i; 491 daddr += btofsb(fs, fs->lfs_ibsize); 492 --iaddrp; 493 continue; 494 } 495 496 /* 497 * Note each file block from the finfo blocks 498 */ 499 if (fic >= ssp->ss_nfinfo) 500 continue; 501 502 /* Count this finfo, whether or not we use it */ 503 ++fic; 504 505 /* 506 * If this finfo has nblocks==0, it was written wrong. 507 * Kernels with this problem always wrote this zero-sized 508 * finfo last, so just ignore it. 509 */ 510 if (fip->fi_nblocks == 0) { 511#ifdef REPAIR_ZERO_FINFO 512 struct ubuf *nbp; 513 SEGSUM *nssp; 514 515 syslog(LOG_WARNING, "fixing short FINFO at %x (seg %d)", 516 odaddr, dtosn(fs, odaddr)); 517 bread(fs->clfs_devvp, odaddr, fs->lfs_fsize, 518 NOCRED, 0, &nbp); 519 nssp = (SEGSUM *)nbp->b_data; 520 --nssp->ss_nfinfo; 521 nssp->ss_sumsum = cksum(&nssp->ss_datasum, 522 fs->lfs_sumsize - sizeof(nssp->ss_sumsum)); 523 bwrite(nbp); 524#endif 525 syslog(LOG_WARNING, "zero-length FINFO at %x (seg %d)", 526 odaddr, dtosn(fs, odaddr)); 527 continue; 528 } 529 530 /* 531 * Check currency before adding blocks 532 */ 533#ifdef REPAIR_ZERO_FINFO 534 vers = -1; 535#else 536 lfs_ientry(&ifp, fs, fip->fi_ino, &ifbp); 537 vers = ifp->if_version; 538 brelse(ifbp, 0); 539#endif 540 if (vers != fip->fi_version) { 541 size_t size; 542 543 /* Read all the blocks from the data summary */ 544 for (i = 0; i < fip->fi_nblocks; i++) { 545 size = (i == fip->fi_nblocks - 1) ? 546 fip->fi_lastlength : fs->lfs_bsize; 547 cp = fd_ptrget(fs->clfs_devvp, daddr); 548 ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck); 549 daddr += btofsb(fs, size); 550 } 551 fip = (FINFO *)(fip->fi_blocks + fip->fi_nblocks); 552 continue; 553 } 554 555 /* Add all the blocks from the finfos (current or not) */ 556 nbip = (BLOCK_INFO *)realloc(bip, (*bic + fip->fi_nblocks) * 557 sizeof(*bip)); 558 if (nbip) 559 bip = nbip; 560 else { 561 *bipp = bip; 562 return 0x0; 563 } 564 565 for (i = 0; i < fip->fi_nblocks; i++) { 566 bip[*bic + i].bi_inode = fip->fi_ino; 567 bip[*bic + i].bi_lbn = fip->fi_blocks[i]; 568 bip[*bic + i].bi_daddr = daddr; 569 bip[*bic + i].bi_segcreate = ssp->ss_create; 570 bip[*bic + i].bi_version = fip->fi_version; 571 bip[*bic + i].bi_size = (i == fip->fi_nblocks - 1) ? 572 fip->fi_lastlength : fs->lfs_bsize; 573 cp = fd_ptrget(fs->clfs_devvp, daddr); 574 ck = lfs_cksum_part(cp, sizeof(u_int32_t), ck); 575 bip[*bic + i].bi_bp = cp; 576 daddr += btofsb(fs, bip[*bic + i].bi_size); 577 578#ifdef TEST_PATTERN 579 check_test_pattern(bip + *bic + i); /* XXXDEBUG */ 580#endif 581 } 582 *bic += fip->fi_nblocks; 583 fip = (FINFO *)(fip->fi_blocks + fip->fi_nblocks); 584 } 585 586#ifndef REPAIR_ZERO_FINFO 587 if (ssp->ss_datasum != ck) { 588 syslog(LOG_WARNING, "%s: data checksum bad at 0x%x:" 589 " read 0x%x, computed 0x%x", fs->lfs_fsmnt, odaddr, 590 ssp->ss_datasum, ck); 591 *bic = obic; 592 return 0x0; 593 } 594#endif 595 596 *bipp = bip; 597 return daddr; 598} 599 600static void 601log_segment_read(struct clfs *fs, int sn) 602{ 603 FILE *fp; 604 char *cp; 605 606 /* 607 * Write the segment read, and its contents, into a log file in 608 * the current directory. We don't need to log the location of 609 * the segment, since that can be inferred from the segments up 610 * to this point (ss_nextseg field of the previously written segment). 611 * 612 * We can use this info later to reconstruct the filesystem at any 613 * given point in time for analysis, by replaying the log forward 614 * indexed by the segment serial numbers; but it is not suitable 615 * for everyday use since the copylog will be simply enormous. 616 */ 617 cp = fd_ptrget(fs->clfs_devvp, sntod(fs, sn)); 618 619 fp = fopen(copylog_filename, "ab"); 620 if (fp != NULL) { 621 if (fwrite(cp, (size_t)fs->lfs_ssize, 1, fp) != 1) { 622 perror("writing segment to copy log"); 623 } 624 } 625 fclose(fp); 626} 627 628/* 629 * Read a segment to populate the BLOCK_INFO structures. 630 * Return the number of partial segments read and parsed. 631 */ 632int 633load_segment(struct clfs *fs, int sn, BLOCK_INFO **bipp, int *bic) 634{ 635 int32_t daddr; 636 int i, npseg; 637 638 daddr = sntod(fs, sn); 639 if (daddr < btofsb(fs, LFS_LABELPAD)) 640 daddr = btofsb(fs, LFS_LABELPAD); 641 for (i = 0; i < LFS_MAXNUMSB; i++) { 642 if (fs->lfs_sboffs[i] == daddr) { 643 daddr += btofsb(fs, LFS_SBPAD); 644 break; 645 } 646 } 647 648 /* Preload the segment buffer */ 649 if (fd_preload(fs->clfs_devvp, sntod(fs, sn)) < 0) 650 return -1; 651 652 if (copylog_filename) 653 log_segment_read(fs, sn); 654 655 /* Note bytes read for stats */ 656 cleaner_stats.segs_cleaned++; 657 cleaner_stats.bytes_read += fs->lfs_ssize; 658 ++fs->clfs_nactive; 659 660 npseg = 0; 661 while(dtosn(fs, daddr) == sn && 662 dtosn(fs, daddr + btofsb(fs, fs->lfs_bsize)) == sn) { 663 daddr = parse_pseg(fs, daddr, bipp, bic); 664 if (daddr == 0x0) { 665 ++cleaner_stats.segs_error; 666 break; 667 } 668 ++npseg; 669 } 670 671 return npseg; 672} 673 674void 675calc_cb(struct clfs *fs, int sn, struct clfs_seguse *t) 676{ 677 time_t now; 678 int64_t age, benefit, cost; 679 680 time(&now); 681 age = (now < t->lastmod ? 0 : now - t->lastmod); 682 683 /* Under no circumstances clean active or already-clean segments */ 684 if ((t->flags & SEGUSE_ACTIVE) || !(t->flags & SEGUSE_DIRTY)) { 685 t->priority = 0; 686 return; 687 } 688 689 /* 690 * If the segment is empty, there is no reason to clean it. 691 * Clear its error condition, if any, since we are never going to 692 * try to parse this one. 693 */ 694 if (t->nbytes == 0) { 695 t->flags &= ~SEGUSE_ERROR; /* Strip error once empty */ 696 t->priority = 0; 697 return; 698 } 699 700 if (t->flags & SEGUSE_ERROR) { /* No good if not already empty */ 701 /* No benefit */ 702 t->priority = 0; 703 return; 704 } 705 706 if (t->nbytes > fs->lfs_ssize) { 707 /* Another type of error */ 708 syslog(LOG_WARNING, "segment %d: bad seguse count %d", 709 sn, t->nbytes); 710 t->flags |= SEGUSE_ERROR; 711 t->priority = 0; 712 return; 713 } 714 715 /* 716 * The non-degenerate case. Use Rosenblum's cost-benefit algorithm. 717 * Calculate the benefit from cleaning this segment (one segment, 718 * minus fragmentation, dirty blocks and a segment summary block) 719 * and weigh that against the cost (bytes read plus bytes written). 720 * We count the summary headers as "dirty" to avoid cleaning very 721 * old and very full segments. 722 */ 723 benefit = (int64_t)fs->lfs_ssize - t->nbytes - 724 (t->nsums + 1) * fs->lfs_fsize; 725 if (fs->lfs_bsize > fs->lfs_fsize) /* fragmentation */ 726 benefit -= (fs->lfs_bsize / 2); 727 if (benefit <= 0) { 728 t->priority = 0; 729 return; 730 } 731 732 cost = fs->lfs_ssize + t->nbytes; 733 t->priority = (256 * benefit * age) / cost; 734 735 return; 736} 737 738/* 739 * Comparator for BLOCK_INFO structures. Anything not in one of the segments 740 * we're looking at sorts higher; after that we sort first by inode number 741 * and then by block number (unsigned, i.e., negative sorts higher) *but* 742 * sort inodes before data blocks. 743 */ 744static int 745bi_comparator(const void *va, const void *vb) 746{ 747 const BLOCK_INFO *a, *b; 748 749 a = (const BLOCK_INFO *)va; 750 b = (const BLOCK_INFO *)vb; 751 752 /* Check for out-of-place block */ 753 if (a->bi_segcreate == a->bi_daddr && 754 b->bi_segcreate != b->bi_daddr) 755 return -1; 756 if (a->bi_segcreate != a->bi_daddr && 757 b->bi_segcreate == b->bi_daddr) 758 return 1; 759 if (a->bi_size <= 0 && b->bi_size > 0) 760 return 1; 761 if (b->bi_size <= 0 && a->bi_size > 0) 762 return -1; 763 764 /* Check inode number */ 765 if (a->bi_inode != b->bi_inode) 766 return a->bi_inode - b->bi_inode; 767 768 /* Check lbn */ 769 if (a->bi_lbn == LFS_UNUSED_LBN) /* Inodes sort lower than blocks */ 770 return -1; 771 if (b->bi_lbn == LFS_UNUSED_LBN) 772 return 1; 773 if ((u_int32_t)a->bi_lbn > (u_int32_t)b->bi_lbn) 774 return 1; 775 else 776 return -1; 777 778 return 0; 779} 780 781/* 782 * Comparator for sort_segments: cost-benefit equation. 783 */ 784static int 785cb_comparator(const void *va, const void *vb) 786{ 787 const struct clfs_seguse *a, *b; 788 789 a = *(const struct clfs_seguse * const *)va; 790 b = *(const struct clfs_seguse * const *)vb; 791 return a->priority > b->priority ? -1 : 1; 792} 793 794void 795toss_old_blocks(struct clfs *fs, BLOCK_INFO **bipp, int *bic, int *sizep) 796{ 797 int i, r; 798 BLOCK_INFO *bip = *bipp; 799 struct lfs_fcntl_markv /* { 800 BLOCK_INFO *blkiov; 801 int blkcnt; 802 } */ lim; 803 804 if (bic == 0 || bip == NULL) 805 return; 806 807 /* 808 * Kludge: Store the disk address in segcreate so we know which 809 * ones to toss. 810 */ 811 for (i = 0; i < *bic; i++) 812 bip[i].bi_segcreate = bip[i].bi_daddr; 813 814 /* Sort the blocks */ 815 heapsort(bip, *bic, sizeof(BLOCK_INFO), bi_comparator); 816 817 /* Use bmapv to locate the blocks */ 818 lim.blkiov = bip; 819 lim.blkcnt = *bic; 820 if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNBMAPV, &lim)) < 0) { 821 syslog(LOG_WARNING, "%s: bmapv returned %d (%m)", 822 fs->lfs_fsmnt, r); 823 return; 824 } 825 826 /* Toss blocks not in this segment */ 827 heapsort(bip, *bic, sizeof(BLOCK_INFO), bi_comparator); 828 829 /* Get rid of stale blocks */ 830 if (sizep) 831 *sizep = 0; 832 for (i = 0; i < *bic; i++) { 833 if (bip[i].bi_segcreate != bip[i].bi_daddr) 834 break; 835 if (sizep) 836 *sizep += bip[i].bi_size; 837 } 838 *bic = i; /* XXX should we shrink bip? */ 839 *bipp = bip; 840 841 return; 842} 843 844/* 845 * Clean a segment and mark it invalid. 846 */ 847int 848invalidate_segment(struct clfs *fs, int sn) 849{ 850 BLOCK_INFO *bip; 851 int i, r, bic; 852 off_t nb; 853 double util; 854 struct lfs_fcntl_markv /* { 855 BLOCK_INFO *blkiov; 856 int blkcnt; 857 } */ lim; 858 859 dlog("%s: inval seg %d", fs->lfs_fsmnt, sn); 860 861 bip = NULL; 862 bic = 0; 863 fs->clfs_nactive = 0; 864 if (load_segment(fs, sn, &bip, &bic) <= 0) 865 return -1; 866 toss_old_blocks(fs, &bip, &bic, NULL); 867 868 /* Record statistics */ 869 for (i = nb = 0; i < bic; i++) 870 nb += bip[i].bi_size; 871 util = ((double)nb) / (fs->clfs_nactive * fs->lfs_ssize); 872 cleaner_stats.util_tot += util; 873 cleaner_stats.util_sos += util * util; 874 cleaner_stats.bytes_written += nb; 875 876 /* 877 * Use markv to move the blocks. 878 */ 879 lim.blkiov = bip; 880 lim.blkcnt = bic; 881 if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNMARKV, &lim)) < 0) { 882 syslog(LOG_WARNING, "%s: markv returned %d (%m) " 883 "for seg %d", fs->lfs_fsmnt, r, sn); 884 return r; 885 } 886 887 /* 888 * Finally call invalidate to invalidate the segment. 889 */ 890 if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNINVAL, &sn)) < 0) { 891 syslog(LOG_WARNING, "%s: inval returned %d (%m) " 892 "for seg %d", fs->lfs_fsmnt, r, sn); 893 return r; 894 } 895 896 return 0; 897} 898 899/* 900 * Check to see if the given ino/lbn pair is represented in the BLOCK_INFO 901 * array we are sending to the kernel, or if the kernel will have to add it. 902 * The kernel will only add each such pair once, though, so keep track of 903 * previous requests in a separate "extra" BLOCK_INFO array. Returns 1 904 * if the block needs to be added, 0 if it is already represented. 905 */ 906static int 907check_or_add(ino_t ino, int32_t lbn, BLOCK_INFO *bip, int bic, BLOCK_INFO **ebipp, int *ebicp) 908{ 909 BLOCK_INFO *t, *ebip = *ebipp; 910 int ebic = *ebicp; 911 int k; 912 913 for (k = 0; k < bic; k++) { 914 if (bip[k].bi_inode != ino) 915 break; 916 if (bip[k].bi_lbn == lbn) { 917 return 0; 918 } 919 } 920 921 /* Look on the list of extra blocks, too */ 922 for (k = 0; k < ebic; k++) { 923 if (ebip[k].bi_inode == ino && ebip[k].bi_lbn == lbn) { 924 return 0; 925 } 926 } 927 928 ++ebic; 929 t = realloc(ebip, ebic * sizeof(BLOCK_INFO)); 930 if (t == NULL) 931 return 1; /* Note *ebicp is unchanged */ 932 933 ebip = t; 934 ebip[ebic - 1].bi_inode = ino; 935 ebip[ebic - 1].bi_lbn = lbn; 936 937 *ebipp = ebip; 938 *ebicp = ebic; 939 return 1; 940} 941 942/* 943 * Look for indirect blocks we will have to write which are not 944 * contained in this collection of blocks. This constitutes 945 * a hidden cleaning cost, since we are unaware of it until we 946 * have already read the segments. Return the total cost, and fill 947 * in *ifc with the part of that cost due to rewriting the Ifile. 948 */ 949static off_t 950check_hidden_cost(struct clfs *fs, BLOCK_INFO *bip, int bic, off_t *ifc) 951{ 952 int start; 953 struct indir in[NIADDR + 1]; 954 int num; 955 int i, j, ebic; 956 BLOCK_INFO *ebip; 957 int32_t lbn; 958 959 start = 0; 960 ebip = NULL; 961 ebic = 0; 962 for (i = 0; i < bic; i++) { 963 if (i == 0 || bip[i].bi_inode != bip[start].bi_inode) { 964 start = i; 965 /* 966 * Look for IFILE blocks, unless this is the Ifile. 967 */ 968 if (bip[i].bi_inode != fs->lfs_ifile) { 969 lbn = fs->lfs_cleansz + bip[i].bi_inode / 970 fs->lfs_ifpb; 971 *ifc += check_or_add(fs->lfs_ifile, lbn, 972 bip, bic, &ebip, &ebic); 973 } 974 } 975 if (bip[i].bi_lbn == LFS_UNUSED_LBN) 976 continue; 977 if (bip[i].bi_lbn < NDADDR) 978 continue; 979 980 ufs_getlbns((struct lfs *)fs, NULL, (daddr_t)bip[i].bi_lbn, in, &num); 981 for (j = 0; j < num; j++) { 982 check_or_add(bip[i].bi_inode, in[j].in_lbn, 983 bip + start, bic - start, &ebip, &ebic); 984 } 985 } 986 return ebic; 987} 988 989/* 990 * Select segments to clean, add blocks from these segments to a cleaning 991 * list, and send this list through lfs_markv() to move them to new 992 * locations on disk. 993 */ 994int 995clean_fs(struct clfs *fs, CLEANERINFO *cip) 996{ 997 int i, j, ngood, sn, bic, r, npos; 998 int bytes, totbytes; 999 struct ubuf *bp; 1000 SEGUSE *sup; 1001 static BLOCK_INFO *bip; 1002 struct lfs_fcntl_markv /* { 1003 BLOCK_INFO *blkiov; 1004 int blkcnt; 1005 } */ lim; 1006 int mc; 1007 BLOCK_INFO *mbip; 1008 int inc; 1009 off_t nb; 1010 off_t goal; 1011 off_t extra, if_extra; 1012 double util; 1013 1014 /* Read the segment table into our private structure */ 1015 npos = 0; 1016 for (i = 0; i < fs->lfs_nseg; i+= fs->lfs_sepb) { 1017 bread(fs->lfs_ivnode, fs->lfs_cleansz + i / fs->lfs_sepb, 1018 fs->lfs_bsize, NOCRED, 0, &bp); 1019 for (j = 0; j < fs->lfs_sepb && i + j < fs->lfs_nseg; j++) { 1020 sup = ((SEGUSE *)bp->b_data) + j; 1021 fs->clfs_segtab[i + j].nbytes = sup->su_nbytes; 1022 fs->clfs_segtab[i + j].nsums = sup->su_nsums; 1023 fs->clfs_segtab[i + j].lastmod = sup->su_lastmod; 1024 /* Keep error status but renew other flags */ 1025 fs->clfs_segtab[i + j].flags &= SEGUSE_ERROR; 1026 fs->clfs_segtab[i + j].flags |= sup->su_flags; 1027 1028 /* Compute cost-benefit coefficient */ 1029 calc_cb(fs, i + j, fs->clfs_segtab + i + j); 1030 if (fs->clfs_segtab[i + j].priority > 0) 1031 ++npos; 1032 } 1033 brelse(bp, 0); 1034 } 1035 1036 /* Sort segments based on cleanliness, fulness, and condition */ 1037 heapsort(fs->clfs_segtabp, fs->lfs_nseg, sizeof(struct clfs_seguse *), 1038 cb_comparator); 1039 1040 /* If no segment is cleanable, just return */ 1041 if (fs->clfs_segtabp[0]->priority == 0) { 1042 dlog("%s: no segment cleanable", fs->lfs_fsmnt); 1043 return 0; 1044 } 1045 1046 /* Load some segments' blocks into bip */ 1047 bic = 0; 1048 fs->clfs_nactive = 0; 1049 ngood = 0; 1050 if (use_bytes) { 1051 /* Set attainable goal */ 1052 goal = fs->lfs_ssize * atatime; 1053 if (goal > (cip->clean - 1) * fs->lfs_ssize / 2) 1054 goal = MAX((cip->clean - 1) * fs->lfs_ssize, 1055 fs->lfs_ssize) / 2; 1056 1057 dlog("%s: cleaning with goal %" PRId64 1058 " bytes (%d segs clean, %d cleanable)", 1059 fs->lfs_fsmnt, goal, cip->clean, npos); 1060 syslog(LOG_INFO, "%s: cleaning with goal %" PRId64 1061 " bytes (%d segs clean, %d cleanable)", 1062 fs->lfs_fsmnt, goal, cip->clean, npos); 1063 totbytes = 0; 1064 for (i = 0; i < fs->lfs_nseg && totbytes < goal; i++) { 1065 if (fs->clfs_segtabp[i]->priority == 0) 1066 break; 1067 /* Upper bound on number of segments at once */ 1068 if (ngood * fs->lfs_ssize > 4 * goal) 1069 break; 1070 sn = (fs->clfs_segtabp[i] - fs->clfs_segtab); 1071 dlog("%s: add seg %d prio %" PRIu64 1072 " containing %ld bytes", 1073 fs->lfs_fsmnt, sn, fs->clfs_segtabp[i]->priority, 1074 fs->clfs_segtabp[i]->nbytes); 1075 if ((r = load_segment(fs, sn, &bip, &bic)) > 0) { 1076 ++ngood; 1077 toss_old_blocks(fs, &bip, &bic, &bytes); 1078 totbytes += bytes; 1079 } else if (r == 0) 1080 fd_release(fs->clfs_devvp); 1081 else 1082 break; 1083 } 1084 } else { 1085 /* Set attainable goal */ 1086 goal = atatime; 1087 if (goal > cip->clean - 1) 1088 goal = MAX(cip->clean - 1, 1); 1089 1090 dlog("%s: cleaning with goal %d segments (%d clean, %d cleanable)", 1091 fs->lfs_fsmnt, (int)goal, cip->clean, npos); 1092 for (i = 0; i < fs->lfs_nseg && ngood < goal; i++) { 1093 if (fs->clfs_segtabp[i]->priority == 0) 1094 break; 1095 sn = (fs->clfs_segtabp[i] - fs->clfs_segtab); 1096 dlog("%s: add seg %d prio %" PRIu64, 1097 fs->lfs_fsmnt, sn, fs->clfs_segtabp[i]->priority); 1098 if ((r = load_segment(fs, sn, &bip, &bic)) > 0) 1099 ++ngood; 1100 else if (r == 0) 1101 fd_release(fs->clfs_devvp); 1102 else 1103 break; 1104 } 1105 toss_old_blocks(fs, &bip, &bic, NULL); 1106 } 1107 1108 /* If there is nothing to do, try again later. */ 1109 if (bic == 0) { 1110 dlog("%s: no blocks to clean in %d cleanable segments", 1111 fs->lfs_fsmnt, (int)ngood); 1112 fd_release_all(fs->clfs_devvp); 1113 return 0; 1114 } 1115 1116 /* Record statistics */ 1117 for (i = nb = 0; i < bic; i++) 1118 nb += bip[i].bi_size; 1119 util = ((double)nb) / (fs->clfs_nactive * fs->lfs_ssize); 1120 cleaner_stats.util_tot += util; 1121 cleaner_stats.util_sos += util * util; 1122 cleaner_stats.bytes_written += nb; 1123 1124 /* 1125 * Check out our blocks to see if there are hidden cleaning costs. 1126 * If there are, we might be cleaning ourselves deeper into a hole 1127 * rather than doing anything useful. 1128 * XXX do something about this. 1129 */ 1130 if_extra = 0; 1131 extra = fs->lfs_bsize * (off_t)check_hidden_cost(fs, bip, bic, &if_extra); 1132 if_extra *= fs->lfs_bsize; 1133 1134 /* 1135 * Use markv to move the blocks. 1136 */ 1137 if (do_small) 1138 inc = MAXPHYS / fs->lfs_bsize - 1; 1139 else 1140 inc = LFS_MARKV_MAXBLKCNT / 2; 1141 for (mc = 0, mbip = bip; mc < bic; mc += inc, mbip += inc) { 1142 lim.blkiov = mbip; 1143 lim.blkcnt = (bic - mc > inc ? inc : bic - mc); 1144#ifdef TEST_PATTERN 1145 dlog("checking blocks %d-%d", mc, mc + lim.blkcnt - 1); 1146 for (i = 0; i < lim.blkcnt; i++) { 1147 check_test_pattern(mbip + i); 1148 } 1149#endif /* TEST_PATTERN */ 1150 dlog("sending blocks %d-%d", mc, mc + lim.blkcnt - 1); 1151 if ((r = kops.ko_fcntl(fs->clfs_ifilefd, LFCNMARKV, &lim))<0) { 1152 int oerrno = errno; 1153 syslog(LOG_WARNING, "%s: markv returned %d (errno %d, %m)", 1154 fs->lfs_fsmnt, r, errno); 1155 if (oerrno != EAGAIN && oerrno != ESHUTDOWN) { 1156 syslog(LOG_DEBUG, "%s: errno %d, returning", 1157 fs->lfs_fsmnt, oerrno); 1158 fd_release_all(fs->clfs_devvp); 1159 return r; 1160 } 1161 if (oerrno == ESHUTDOWN) { 1162 syslog(LOG_NOTICE, "%s: filesystem unmounted", 1163 fs->lfs_fsmnt); 1164 fd_release_all(fs->clfs_devvp); 1165 return r; 1166 } 1167 } 1168 } 1169 1170 /* 1171 * Report progress (or lack thereof) 1172 */ 1173 syslog(LOG_INFO, "%s: wrote %" PRId64 " dirty + %" 1174 PRId64 " supporting indirect + %" 1175 PRId64 " supporting Ifile = %" 1176 PRId64 " bytes to clean %d segs (%" PRId64 "%% recovery)", 1177 fs->lfs_fsmnt, (int64_t)nb, (int64_t)(extra - if_extra), 1178 (int64_t)if_extra, (int64_t)(nb + extra), ngood, 1179 (ngood ? (int64_t)(100 - (100 * (nb + extra)) / 1180 (ngood * fs->lfs_ssize)) : 1181 (int64_t)0)); 1182 if (nb + extra >= ngood * fs->lfs_ssize) 1183 syslog(LOG_WARNING, "%s: cleaner not making forward progress", 1184 fs->lfs_fsmnt); 1185 1186 /* 1187 * Finally call reclaim to prompt cleaning of the segments. 1188 */ 1189 kops.ko_fcntl(fs->clfs_ifilefd, LFCNRECLAIM, NULL); 1190 1191 fd_release_all(fs->clfs_devvp); 1192 return 0; 1193} 1194 1195/* 1196 * Read the cleanerinfo block and apply cleaning policy to determine whether 1197 * the given filesystem needs to be cleaned. Returns 1 if it does, 0 if it 1198 * does not, or -1 on error. 1199 */ 1200int 1201needs_cleaning(struct clfs *fs, CLEANERINFO *cip) 1202{ 1203 struct ubuf *bp; 1204 struct stat st; 1205 daddr_t fsb_per_seg, max_free_segs; 1206 time_t now; 1207 double loadavg; 1208 1209 /* If this fs is "on hold", don't clean it. */ 1210 if (fs->clfs_onhold) 1211 return 0; 1212 1213 /* 1214 * Read the cleanerinfo block from the Ifile. We don't want 1215 * the cached information, so invalidate the buffer before 1216 * handing it back. 1217 */ 1218 if (bread(fs->lfs_ivnode, 0, fs->lfs_bsize, NOCRED, 0, &bp)) { 1219 syslog(LOG_ERR, "%s: can't read inode", fs->lfs_fsmnt); 1220 return -1; 1221 } 1222 *cip = *(CLEANERINFO *)bp->b_data; /* Structure copy */ 1223 brelse(bp, B_INVAL); 1224 cleaner_stats.bytes_read += fs->lfs_bsize; 1225 1226 /* 1227 * If the number of segments changed under us, reinit. 1228 * We don't have to start over from scratch, however, 1229 * since we don't hold any buffers. 1230 */ 1231 if (fs->lfs_nseg != cip->clean + cip->dirty) { 1232 if (reinit_fs(fs) < 0) { 1233 /* The normal case for unmount */ 1234 syslog(LOG_NOTICE, "%s: filesystem unmounted", fs->lfs_fsmnt); 1235 return -1; 1236 } 1237 syslog(LOG_NOTICE, "%s: nsegs changed", fs->lfs_fsmnt); 1238 } 1239 1240 /* Compute theoretical "free segments" maximum based on usage */ 1241 fsb_per_seg = segtod(fs, 1); 1242 max_free_segs = MAX(cip->bfree, 0) / fsb_per_seg + fs->lfs_minfreeseg; 1243 1244 dlog("%s: bfree = %d, avail = %d, clean = %d/%d", 1245 fs->lfs_fsmnt, cip->bfree, cip->avail, cip->clean, fs->lfs_nseg); 1246 1247 /* If the writer is waiting on us, clean it */ 1248 if (cip->clean <= fs->lfs_minfreeseg || 1249 (cip->flags & LFS_CLEANER_MUST_CLEAN)) 1250 return 1; 1251 1252 /* If there are enough segments, don't clean it */ 1253 if (cip->bfree - cip->avail <= fsb_per_seg && 1254 cip->avail > fsb_per_seg) 1255 return 0; 1256 1257 /* If we are in dire straits, clean it */ 1258 if (cip->bfree - cip->avail > fsb_per_seg && 1259 cip->avail <= fsb_per_seg) 1260 return 1; 1261 1262 /* If under busy threshold, clean regardless of load */ 1263 if (cip->clean < max_free_segs * BUSY_LIM) 1264 return 1; 1265 1266 /* Check busy status; clean if idle and under idle limit */ 1267 if (use_fs_idle) { 1268 /* Filesystem idle */ 1269 time(&now); 1270 if (fstat(fs->clfs_ifilefd, &st) < 0) { 1271 syslog(LOG_ERR, "%s: failed to stat ifile", 1272 fs->lfs_fsmnt); 1273 return -1; 1274 } 1275 if (now - st.st_mtime > segwait_timeout && 1276 cip->clean < max_free_segs * IDLE_LIM) 1277 return 1; 1278 } else { 1279 /* CPU idle - use one-minute load avg */ 1280 if (getloadavg(&loadavg, 1) == -1) { 1281 syslog(LOG_ERR, "%s: failed to get load avg", 1282 fs->lfs_fsmnt); 1283 return -1; 1284 } 1285 if (loadavg < load_threshold && 1286 cip->clean < max_free_segs * IDLE_LIM) 1287 return 1; 1288 } 1289 1290 return 0; 1291} 1292 1293/* 1294 * Report statistics. If the signal was SIGUSR2, clear the statistics too. 1295 * If the signal was SIGINT, exit. 1296 */ 1297static void 1298sig_report(int sig) 1299{ 1300 double avg = 0.0, stddev; 1301 1302 avg = cleaner_stats.util_tot / MAX(cleaner_stats.segs_cleaned, 1.0); 1303 stddev = cleaner_stats.util_sos / MAX(cleaner_stats.segs_cleaned - 1304 avg * avg, 1.0); 1305 syslog(LOG_INFO, "bytes read: %" PRId64, cleaner_stats.bytes_read); 1306 syslog(LOG_INFO, "bytes written: %" PRId64, cleaner_stats.bytes_written); 1307 syslog(LOG_INFO, "segments cleaned: %" PRId64, cleaner_stats.segs_cleaned); 1308#if 0 1309 /* "Empty segments" is meaningless, since the kernel handles those */ 1310 syslog(LOG_INFO, "empty segments: %" PRId64, cleaner_stats.segs_empty); 1311#endif 1312 syslog(LOG_INFO, "error segments: %" PRId64, cleaner_stats.segs_error); 1313 syslog(LOG_INFO, "utilization total: %g", cleaner_stats.util_tot); 1314 syslog(LOG_INFO, "utilization sos: %g", cleaner_stats.util_sos); 1315 syslog(LOG_INFO, "utilization avg: %4.2f", avg); 1316 syslog(LOG_INFO, "utilization sdev: %9.6f", stddev); 1317 1318 if (debug) 1319 bufstats(); 1320 1321 if (sig == SIGUSR2) 1322 memset(&cleaner_stats, 0, sizeof(cleaner_stats)); 1323 if (sig == SIGINT) 1324 exit(0); 1325} 1326 1327static void 1328sig_exit(int sig) 1329{ 1330 exit(0); 1331} 1332 1333static void 1334usage(void) 1335{ 1336 errx(1, "usage: lfs_cleanerd [-bcdfmqs] [-i segnum] [-l load] " 1337 "[-n nsegs] [-r report_freq] [-t timeout] fs_name ..."); 1338} 1339 1340#ifndef LFS_CLEANER_AS_LIB 1341/* 1342 * Main. 1343 */ 1344int 1345main(int argc, char **argv) 1346{ 1347 1348 return lfs_cleaner_main(argc, argv); 1349} 1350#endif 1351 1352int 1353lfs_cleaner_main(int argc, char **argv) 1354{ 1355 int i, opt, error, r, loopcount, nodetach; 1356 struct timeval tv; 1357 sem_t *semaddr = NULL; 1358 CLEANERINFO ci; 1359#ifndef USE_CLIENT_SERVER 1360 char *cp, *pidname; 1361#endif 1362 1363 /* 1364 * Set up defaults 1365 */ 1366 atatime = 1; 1367 segwait_timeout = 300; /* Five minutes */ 1368 load_threshold = 0.2; 1369 stat_report = 0; 1370 inval_segment = -1; 1371 copylog_filename = NULL; 1372 nodetach = 0; 1373 1374 /* 1375 * Parse command-line arguments 1376 */ 1377 while ((opt = getopt(argc, argv, "bC:cdDfi:l:mn:qr:sS:t:")) != -1) { 1378 switch (opt) { 1379 case 'b': /* Use bytes written, not segments read */ 1380 use_bytes = 1; 1381 break; 1382 case 'C': /* copy log */ 1383 copylog_filename = optarg; 1384 break; 1385 case 'c': /* Coalesce files */ 1386 do_coalesce++; 1387 break; 1388 case 'd': /* Debug mode. */ 1389 nodetach++; 1390 debug++; 1391 break; 1392 case 'D': /* stay-on-foreground */ 1393 nodetach++; 1394 break; 1395 case 'f': /* Use fs idle time rather than cpu idle */ 1396 use_fs_idle = 1; 1397 break; 1398 case 'i': /* Invalidate this segment */ 1399 inval_segment = atoi(optarg); 1400 break; 1401 case 'l': /* Load below which to clean */ 1402 load_threshold = atof(optarg); 1403 break; 1404 case 'm': /* [compat only] */ 1405 break; 1406 case 'n': /* How many segs to clean at once */ 1407 atatime = atoi(optarg); 1408 break; 1409 case 'q': /* Quit after one run */ 1410 do_quit = 1; 1411 break; 1412 case 'r': /* Report every stat_report segments */ 1413 stat_report = atoi(optarg); 1414 break; 1415 case 's': /* Small writes */ 1416 do_small = 1; 1417 break; 1418 case 'S': /* semaphore */ 1419#ifndef LFS_CLEANER_AS_LIB 1420 usage(); 1421 /*NOTREACHED*/ 1422#endif 1423 semaddr = (void*)(uintptr_t)strtoull(optarg,NULL,0); 1424 break; 1425 case 't': /* timeout */ 1426 segwait_timeout = atoi(optarg); 1427 break; 1428 default: 1429 usage(); 1430 /* NOTREACHED */ 1431 } 1432 } 1433 argc -= optind; 1434 argv += optind; 1435 1436 if (argc < 1) 1437 usage(); 1438 if (inval_segment >= 0 && argc != 1) { 1439 errx(1, "lfs_cleanerd: may only specify one filesystem when " 1440 "using -i flag"); 1441 } 1442 1443 if (do_coalesce) { 1444 errx(1, "lfs_cleanerd: -c disabled due to reports of file " 1445 "corruption; you may re-enable it by rebuilding the " 1446 "cleaner"); 1447 } 1448 1449 /* 1450 * Set up daemon mode or foreground mode 1451 */ 1452 if (nodetach) { 1453 openlog("lfs_cleanerd", LOG_NDELAY | LOG_PID | LOG_PERROR, 1454 LOG_DAEMON); 1455 signal(SIGINT, sig_report); 1456 } else { 1457 if (daemon(0, 0) == -1) 1458 err(1, "lfs_cleanerd: couldn't become a daemon!"); 1459 openlog("lfs_cleanerd", LOG_NDELAY | LOG_PID, LOG_DAEMON); 1460 signal(SIGINT, sig_exit); 1461 } 1462 1463 /* 1464 * Look for an already-running master daemon. If there is one, 1465 * send it our filesystems to add to its list and exit. 1466 * If there is none, become the master. 1467 */ 1468#ifdef USE_CLIENT_SERVER 1469 try_to_become_master(argc, argv); 1470#else 1471 /* XXX think about this */ 1472 asprintf(&pidname, "lfs_cleanerd:m:%s", argv[0]); 1473 if (pidname == NULL) { 1474 syslog(LOG_ERR, "malloc failed: %m"); 1475 exit(1); 1476 } 1477 for (cp = pidname; cp != NULL; cp = strchr(cp, '/')) 1478 *cp = '|'; 1479 pidfile(pidname); 1480#endif 1481 1482 /* 1483 * Signals mean daemon should report its statistics 1484 */ 1485 memset(&cleaner_stats, 0, sizeof(cleaner_stats)); 1486 signal(SIGUSR1, sig_report); 1487 signal(SIGUSR2, sig_report); 1488 1489 /* 1490 * Start up buffer cache. We only use this for the Ifile, 1491 * and we will resize it if necessary, so it can start small. 1492 */ 1493 bufinit(4); 1494 1495#ifdef REPAIR_ZERO_FINFO 1496 { 1497 BLOCK_INFO *bip = NULL; 1498 int bic = 0; 1499 1500 nfss = 1; 1501 fsp = (struct clfs **)malloc(sizeof(*fsp)); 1502 fsp[0] = (struct clfs *)calloc(1, sizeof(**fsp)); 1503 1504 if (init_unmounted_fs(fsp[0], argv[0]) < 0) { 1505 err(1, "init_unmounted_fs"); 1506 } 1507 dlog("Filesystem has %d segments", fsp[0]->lfs_nseg); 1508 for (i = 0; i < fsp[0]->lfs_nseg; i++) { 1509 load_segment(fsp[0], i, &bip, &bic); 1510 bic = 0; 1511 } 1512 exit(0); 1513 } 1514#endif 1515 1516 /* 1517 * Initialize cleaning structures, open devices, etc. 1518 */ 1519 nfss = argc; 1520 fsp = (struct clfs **)malloc(nfss * sizeof(*fsp)); 1521 if (fsp == NULL) { 1522 syslog(LOG_ERR, "couldn't allocate fs table: %m"); 1523 exit(1); 1524 } 1525 for (i = 0; i < nfss; i++) { 1526 fsp[i] = (struct clfs *)calloc(1, sizeof(**fsp)); 1527 if ((r = init_fs(fsp[i], argv[i])) < 0) { 1528 syslog(LOG_ERR, "%s: couldn't init: error code %d", 1529 argv[i], r); 1530 handle_error(fsp, i); 1531 --i; /* Do the new #i over again */ 1532 } 1533 } 1534 1535 /* 1536 * If asked to coalesce, do so and exit. 1537 */ 1538 if (do_coalesce) { 1539 for (i = 0; i < nfss; i++) 1540 clean_all_inodes(fsp[i]); 1541 exit(0); 1542 } 1543 1544 /* 1545 * If asked to invalidate a segment, do that and exit. 1546 */ 1547 if (inval_segment >= 0) { 1548 invalidate_segment(fsp[0], inval_segment); 1549 exit(0); 1550 } 1551 1552 /* 1553 * Main cleaning loop. 1554 */ 1555 loopcount = 0; 1556#ifdef LFS_CLEANER_AS_LIB 1557 if (semaddr) 1558 sem_post(semaddr); 1559#endif 1560 error = 0; 1561 while (nfss > 0) { 1562 int cleaned_one; 1563 do { 1564#ifdef USE_CLIENT_SERVER 1565 check_control_socket(); 1566#endif 1567 cleaned_one = 0; 1568 for (i = 0; i < nfss; i++) { 1569 if ((error = needs_cleaning(fsp[i], &ci)) < 0) { 1570 syslog(LOG_DEBUG, "%s: needs_cleaning returned %d", 1571 getprogname(), error); 1572 handle_error(fsp, i); 1573 continue; 1574 } 1575 if (error == 0) /* No need to clean */ 1576 continue; 1577 1578 reload_ifile(fsp[i]); 1579 if ((error = clean_fs(fsp[i], &ci)) < 0) { 1580 syslog(LOG_DEBUG, "%s: clean_fs returned %d", 1581 getprogname(), error); 1582 handle_error(fsp, i); 1583 continue; 1584 } 1585 ++cleaned_one; 1586 } 1587 ++loopcount; 1588 if (stat_report && loopcount % stat_report == 0) 1589 sig_report(0); 1590 if (do_quit) 1591 exit(0); 1592 } while(cleaned_one); 1593 tv.tv_sec = segwait_timeout; 1594 tv.tv_usec = 0; 1595 /* XXX: why couldn't others work if fsp socket is shutdown? */ 1596 error = kops.ko_fcntl(fsp[0]->clfs_ifilefd,LFCNSEGWAITALL,&tv); 1597 if (error) { 1598 if (errno == ESHUTDOWN) { 1599 for (i = 0; i < nfss; i++) { 1600 syslog(LOG_INFO, "%s: shutdown", 1601 getprogname()); 1602 handle_error(fsp, i); 1603 assert(nfss == 0); 1604 } 1605 } else { 1606#ifdef LFS_CLEANER_AS_LIB 1607 error = ESHUTDOWN; 1608 break; 1609#else 1610 err(1, "LFCNSEGWAITALL"); 1611#endif 1612 } 1613 } 1614 } 1615 1616 /* NOTREACHED */ 1617 return error; 1618} 1619