1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * %sccs.include.redist.c% 6 * 7 * @(#)lfs_segment.c 8.3 (Berkeley) 09/23/93 8 */ 9 10 #include <sys/param.h> 11 #include <sys/systm.h> 12 #include <sys/namei.h> 13 #include <sys/kernel.h> 14 #include <sys/resourcevar.h> 15 #include <sys/file.h> 16 #include <sys/stat.h> 17 #include <sys/buf.h> 18 #include <sys/proc.h> 19 #include <sys/conf.h> 20 #include <sys/vnode.h> 21 #include <sys/malloc.h> 22 #include <sys/mount.h> 23 24 #include <miscfs/specfs/specdev.h> 25 #include <miscfs/fifofs/fifo.h> 26 27 #include <ufs/ufs/quota.h> 28 #include <ufs/ufs/inode.h> 29 #include <ufs/ufs/dir.h> 30 #include <ufs/ufs/ufsmount.h> 31 #include <ufs/ufs/ufs_extern.h> 32 33 #include <ufs/lfs/lfs.h> 34 #include <ufs/lfs/lfs_extern.h> 35 36 extern int count_lock_queue __P((void)); 37 38 #define MAX_ACTIVE 10 39 /* 40 * Determine if it's OK to start a partial in this segment, or if we need 41 * to go on to a new segment. 42 */ 43 #define LFS_PARTIAL_FITS(fs) \ 44 ((fs)->lfs_dbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \ 45 1 << (fs)->lfs_fsbtodb) 46 47 void lfs_callback __P((struct buf *)); 48 void lfs_gather __P((struct lfs *, struct segment *, 49 struct vnode *, int (*) __P((struct lfs *, struct buf *)))); 50 int lfs_gatherblock __P((struct segment *, struct buf *, int *)); 51 void lfs_iset __P((struct inode *, daddr_t, time_t)); 52 int lfs_match_data __P((struct lfs *, struct buf *)); 53 int lfs_match_dindir __P((struct lfs *, struct buf *)); 54 int lfs_match_indir __P((struct lfs *, struct buf *)); 55 int lfs_match_tindir __P((struct lfs *, struct buf *)); 56 void lfs_newseg __P((struct lfs *)); 57 void lfs_shellsort __P((struct buf **, daddr_t *, register int)); 58 void lfs_supercallback __P((struct buf *)); 59 void lfs_updatemeta __P((struct segment *)); 60 int lfs_vref __P((struct vnode *)); 61 void lfs_vunref __P((struct vnode *)); 62 void lfs_writefile __P((struct lfs *, struct segment *, struct vnode *)); 63 int lfs_writeinode __P((struct lfs *, struct segment *, struct inode *)); 64 int lfs_writeseg __P((struct lfs *, struct segment *)); 65 void lfs_writesuper __P((struct lfs *)); 66 void lfs_writevnodes __P((struct lfs *fs, struct mount *mp, 67 struct segment *sp, int dirops)); 68 69 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 70 71 /* Statistics Counters */ 72 #define DOSTATS 73 struct lfs_stats lfs_stats; 74 75 /* op values to lfs_writevnodes */ 76 #define VN_REG 0 77 #define VN_DIROP 1 78 #define VN_EMPTY 2 79 80 /* 81 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 82 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 83 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 84 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 85 */ 86 87 int 88 lfs_vflush(vp) 89 struct vnode *vp; 90 { 91 struct inode *ip; 92 struct lfs *fs; 93 struct segment *sp; 94 int error, s; 95 96 fs = VFSTOUFS(vp->v_mount)->um_lfs; 97 if (fs->lfs_nactive > MAX_ACTIVE) 98 return(lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP)); 99 lfs_seglock(fs, SEGM_SYNC); 100 sp = fs->lfs_sp; 101 102 103 ip = VTOI(vp); 104 if (vp->v_dirtyblkhd.le_next == NULL) 105 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 106 107 do { 108 do { 109 if (vp->v_dirtyblkhd.le_next != NULL) 110 lfs_writefile(fs, sp, vp); 111 } while (lfs_writeinode(fs, sp, ip)); 112 113 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 114 115 #ifdef DOSTATS 116 ++lfs_stats.nwrites; 117 if (sp->seg_flags & SEGM_SYNC) 118 ++lfs_stats.nsync_writes; 119 if (sp->seg_flags & SEGM_CKP) 120 ++lfs_stats.ncheckpoints; 121 #endif 122 lfs_segunlock(fs); 123 return (0); 124 } 125 126 void 127 lfs_writevnodes(fs, mp, sp, op) 128 struct lfs *fs; 129 struct mount *mp; 130 struct segment *sp; 131 int op; 132 { 133 struct inode *ip; 134 struct vnode *vp; 135 int error, s, active; 136 137 loop: for (vp = mp->mnt_mounth; vp; vp = vp->v_mountf) { 138 /* 139 * If the vnode that we are about to sync is no longer 140 * associated with this mount point, start over. 141 */ 142 if (vp->v_mount != mp) 143 goto loop; 144 145 /* XXX ignore dirops for now 146 if (op == VN_DIROP && !(vp->v_flag & VDIROP) || 147 op != VN_DIROP && (vp->v_flag & VDIROP)) 148 continue; 149 */ 150 151 if (op == VN_EMPTY && vp->v_dirtyblkhd.le_next) 152 continue; 153 154 if (vp->v_type == VNON) 155 continue; 156 157 if (lfs_vref(vp)) 158 continue; 159 160 /* 161 * Write the inode/file if dirty and it's not the 162 * the IFILE. 163 */ 164 ip = VTOI(vp); 165 if ((ip->i_flag & 166 (IN_ACCESS | IN_CHANGE | IN_MODIFIED | IN_UPDATE) || 167 vp->v_dirtyblkhd.le_next != NULL) && 168 ip->i_number != LFS_IFILE_INUM) { 169 if (vp->v_dirtyblkhd.le_next != NULL) 170 lfs_writefile(fs, sp, vp); 171 (void) lfs_writeinode(fs, sp, ip); 172 } 173 vp->v_flag &= ~VDIROP; 174 lfs_vunref(vp); 175 } 176 } 177 178 int 179 lfs_segwrite(mp, flags) 180 struct mount *mp; 181 int flags; /* Do a checkpoint. */ 182 { 183 struct buf *bp; 184 struct inode *ip; 185 struct lfs *fs; 186 struct segment *sp; 187 struct vnode *vp; 188 SEGUSE *segusep; 189 daddr_t ibno; 190 CLEANERINFO *cip; 191 int clean, error, i, s; 192 int do_ckp; 193 194 fs = VFSTOUFS(mp)->um_lfs; 195 196 /* 197 * If we have fewer than 2 clean segments, wait until cleaner 198 * writes. 199 */ 200 do { 201 LFS_CLEANERINFO(cip, fs, bp); 202 clean = cip->clean; 203 brelse(bp); 204 if (clean <= 2) { 205 printf ("segs clean: %d\n", clean); 206 wakeup(&lfs_allclean_wakeup); 207 if (error = tsleep(&fs->lfs_avail, PRIBIO + 1, 208 "lfs writer", 0)) 209 return (error); 210 } 211 } while (clean <= 2 ); 212 213 /* 214 * Allocate a segment structure and enough space to hold pointers to 215 * the maximum possible number of buffers which can be described in a 216 * single summary block. 217 */ 218 do_ckp = flags & SEGM_CKP || fs->lfs_nactive > MAX_ACTIVE; 219 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 220 sp = fs->lfs_sp; 221 222 lfs_writevnodes(fs, mp, sp, VN_REG); 223 224 /* XXX ignore ordering of dirops for now */ 225 /* XXX 226 fs->lfs_writer = 1; 227 if (fs->lfs_dirops && (error = 228 tsleep(&fs->lfs_writer, PRIBIO + 1, "lfs writer", 0))) { 229 free(sp->bpp, M_SEGMENT); 230 free(sp, M_SEGMENT); 231 fs->lfs_writer = 0; 232 return (error); 233 } 234 235 lfs_writevnodes(fs, mp, sp, VN_DIROP); 236 */ 237 238 /* 239 * If we are doing a checkpoint, mark everything since the 240 * last checkpoint as no longer ACTIVE. 241 */ 242 if (do_ckp) 243 for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz; 244 --ibno >= fs->lfs_cleansz; ) { 245 if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, 246 NOCRED, &bp)) 247 248 panic("lfs: ifile read"); 249 segusep = (SEGUSE *)bp->b_data; 250 for (i = fs->lfs_sepb; i--; segusep++) 251 segusep->su_flags &= ~SEGUSE_ACTIVE; 252 253 error = VOP_BWRITE(bp); 254 } 255 256 if (do_ckp || fs->lfs_doifile) { 257 redo: 258 vp = fs->lfs_ivnode; 259 while (vget(vp)); 260 ip = VTOI(vp); 261 if (vp->v_dirtyblkhd.le_next != NULL) 262 lfs_writefile(fs, sp, vp); 263 (void)lfs_writeinode(fs, sp, ip); 264 vput(vp); 265 if (lfs_writeseg(fs, sp) && do_ckp) 266 goto redo; 267 } else 268 (void) lfs_writeseg(fs, sp); 269 270 /* 271 * If the I/O count is non-zero, sleep until it reaches zero. At the 272 * moment, the user's process hangs around so we can sleep. 273 */ 274 /* XXX ignore dirops for now 275 fs->lfs_writer = 0; 276 fs->lfs_doifile = 0; 277 wakeup(&fs->lfs_dirops); 278 */ 279 280 #ifdef DOSTATS 281 ++lfs_stats.nwrites; 282 if (sp->seg_flags & SEGM_SYNC) 283 ++lfs_stats.nsync_writes; 284 if (sp->seg_flags & SEGM_CKP) 285 ++lfs_stats.ncheckpoints; 286 #endif 287 lfs_segunlock(fs); 288 return (0); 289 } 290 291 /* 292 * Write the dirty blocks associated with a vnode. 293 */ 294 void 295 lfs_writefile(fs, sp, vp) 296 struct lfs *fs; 297 struct segment *sp; 298 struct vnode *vp; 299 { 300 struct buf *bp; 301 struct finfo *fip; 302 IFILE *ifp; 303 304 if (sp->seg_bytes_left < fs->lfs_bsize || 305 sp->sum_bytes_left < sizeof(struct finfo)) 306 (void) lfs_writeseg(fs, sp); 307 308 sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(daddr_t); 309 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 310 311 fip = sp->fip; 312 fip->fi_nblocks = 0; 313 fip->fi_ino = VTOI(vp)->i_number; 314 LFS_IENTRY(ifp, fs, fip->fi_ino, bp); 315 fip->fi_version = ifp->if_version; 316 brelse(bp); 317 318 /* 319 * It may not be necessary to write the meta-data blocks at this point, 320 * as the roll-forward recovery code should be able to reconstruct the 321 * list. 322 */ 323 lfs_gather(fs, sp, vp, lfs_match_data); 324 lfs_gather(fs, sp, vp, lfs_match_indir); 325 lfs_gather(fs, sp, vp, lfs_match_dindir); 326 #ifdef TRIPLE 327 lfs_gather(fs, sp, vp, lfs_match_tindir); 328 #endif 329 330 fip = sp->fip; 331 if (fip->fi_nblocks != 0) { 332 sp->fip = 333 (struct finfo *)((caddr_t)fip + sizeof(struct finfo) + 334 sizeof(daddr_t) * (fip->fi_nblocks - 1)); 335 sp->start_lbp = &sp->fip->fi_blocks[0]; 336 } else { 337 sp->sum_bytes_left += sizeof(struct finfo) - sizeof(daddr_t); 338 --((SEGSUM *)(sp->segsum))->ss_nfinfo; 339 } 340 } 341 342 int 343 lfs_writeinode(fs, sp, ip) 344 struct lfs *fs; 345 struct segment *sp; 346 struct inode *ip; 347 { 348 struct buf *bp, *ibp; 349 IFILE *ifp; 350 SEGUSE *sup; 351 daddr_t daddr; 352 ino_t ino; 353 int error, i, ndx; 354 int redo_ifile = 0; 355 356 if (!(ip->i_flag & (IN_ACCESS | IN_CHANGE | IN_MODIFIED | IN_UPDATE))) 357 return(0); 358 359 /* Allocate a new inode block if necessary. */ 360 if (sp->ibp == NULL) { 361 /* Allocate a new segment if necessary. */ 362 if (sp->seg_bytes_left < fs->lfs_bsize || 363 sp->sum_bytes_left < sizeof(daddr_t)) 364 (void) lfs_writeseg(fs, sp); 365 366 /* Get next inode block. */ 367 daddr = fs->lfs_offset; 368 fs->lfs_offset += fsbtodb(fs, 1); 369 sp->ibp = *sp->cbpp++ = 370 lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, daddr, 371 fs->lfs_bsize); 372 /* Zero out inode numbers */ 373 for (i = 0; i < INOPB(fs); ++i) 374 ((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0; 375 ++sp->start_bpp; 376 fs->lfs_avail -= fsbtodb(fs, 1); 377 /* Set remaining space counters. */ 378 sp->seg_bytes_left -= fs->lfs_bsize; 379 sp->sum_bytes_left -= sizeof(daddr_t); 380 ndx = LFS_SUMMARY_SIZE / sizeof(daddr_t) - 381 sp->ninodes / INOPB(fs) - 1; 382 ((daddr_t *)(sp->segsum))[ndx] = daddr; 383 } 384 385 /* Update the inode times and copy the inode onto the inode page. */ 386 if (ip->i_flag & IN_MODIFIED) 387 --fs->lfs_uinodes; 388 ITIMES(ip, &time, &time); 389 ip->i_flag &= ~(IN_ACCESS | IN_CHANGE | IN_MODIFIED | IN_UPDATE); 390 bp = sp->ibp; 391 ((struct dinode *)bp->b_data)[sp->ninodes % INOPB(fs)] = ip->i_din; 392 /* Increment inode count in segment summary block. */ 393 ++((SEGSUM *)(sp->segsum))->ss_ninos; 394 395 /* If this page is full, set flag to allocate a new page. */ 396 if (++sp->ninodes % INOPB(fs) == 0) 397 sp->ibp = NULL; 398 399 /* 400 * If updating the ifile, update the super-block. Update the disk 401 * address and access times for this inode in the ifile. 402 */ 403 ino = ip->i_number; 404 if (ino == LFS_IFILE_INUM) { 405 daddr = fs->lfs_idaddr; 406 fs->lfs_idaddr = bp->b_blkno; 407 } else { 408 LFS_IENTRY(ifp, fs, ino, ibp); 409 daddr = ifp->if_daddr; 410 ifp->if_daddr = bp->b_blkno; 411 error = VOP_BWRITE(ibp); 412 } 413 414 /* 415 * No need to update segment usage if there was no former inode address 416 * or if the last inode address is in the current partial segment. 417 */ 418 if (daddr != LFS_UNUSED_DADDR && 419 !(daddr >= fs->lfs_lastpseg && daddr <= bp->b_blkno)) { 420 LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp); 421 #ifdef DIAGNOSTIC 422 if (sup->su_nbytes < sizeof(struct dinode)) { 423 /* XXX -- Change to a panic. */ 424 printf("lfs: negative bytes (segment %d)\n", 425 datosn(fs, daddr)); 426 panic("negative bytes"); 427 } 428 #endif 429 sup->su_nbytes -= sizeof(struct dinode); 430 redo_ifile = 431 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 432 error = VOP_BWRITE(bp); 433 } 434 return (redo_ifile); 435 } 436 437 int 438 lfs_gatherblock(sp, bp, sptr) 439 struct segment *sp; 440 struct buf *bp; 441 int *sptr; 442 { 443 struct lfs *fs; 444 int version; 445 446 /* 447 * If full, finish this segment. We may be doing I/O, so 448 * release and reacquire the splbio(). 449 */ 450 #ifdef DIAGNOSTIC 451 if (sp->vp == NULL) 452 panic ("lfs_gatherblock: Null vp in segment"); 453 #endif 454 fs = sp->fs; 455 if (sp->sum_bytes_left < sizeof(daddr_t) || 456 sp->seg_bytes_left < fs->lfs_bsize) { 457 if (sptr) 458 splx(*sptr); 459 lfs_updatemeta(sp); 460 461 version = sp->fip->fi_version; 462 (void) lfs_writeseg(fs, sp); 463 464 sp->fip->fi_version = version; 465 sp->fip->fi_ino = VTOI(sp->vp)->i_number; 466 /* Add the current file to the segment summary. */ 467 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 468 sp->sum_bytes_left -= 469 sizeof(struct finfo) - sizeof(daddr_t); 470 471 if (sptr) 472 *sptr = splbio(); 473 return(1); 474 } 475 476 /* Insert into the buffer list, update the FINFO block. */ 477 bp->b_flags |= B_GATHERED; 478 *sp->cbpp++ = bp; 479 sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno; 480 481 sp->sum_bytes_left -= sizeof(daddr_t); 482 sp->seg_bytes_left -= fs->lfs_bsize; 483 return(0); 484 } 485 486 void 487 lfs_gather(fs, sp, vp, match) 488 struct lfs *fs; 489 struct segment *sp; 490 struct vnode *vp; 491 int (*match) __P((struct lfs *, struct buf *)); 492 { 493 struct buf *bp; 494 int s; 495 496 sp->vp = vp; 497 s = splbio(); 498 loop: for (bp = vp->v_dirtyblkhd.le_next; bp; bp = bp->b_vnbufs.qe_next) { 499 if (bp->b_flags & B_BUSY || !match(fs, bp) || 500 bp->b_flags & B_GATHERED) 501 continue; 502 #ifdef DIAGNOSTIC 503 if (!(bp->b_flags & B_DELWRI)) 504 panic("lfs_gather: bp not B_DELWRI"); 505 if (!(bp->b_flags & B_LOCKED)) 506 panic("lfs_gather: bp not B_LOCKED"); 507 #endif 508 if (lfs_gatherblock(sp, bp, &s)) 509 goto loop; 510 } 511 splx(s); 512 lfs_updatemeta(sp); 513 sp->vp = NULL; 514 } 515 516 517 /* 518 * Update the metadata that points to the blocks listed in the FINFO 519 * array. 520 */ 521 void 522 lfs_updatemeta(sp) 523 struct segment *sp; 524 { 525 SEGUSE *sup; 526 struct buf *bp; 527 struct lfs *fs; 528 struct vnode *vp; 529 struct indir a[NIADDR + 2], *ap; 530 struct inode *ip; 531 daddr_t daddr, lbn, off; 532 int db_per_fsb, error, i, nblocks, num; 533 534 vp = sp->vp; 535 nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 536 if (vp == NULL || nblocks == 0) 537 return; 538 539 /* Sort the blocks. */ 540 if (!(sp->seg_flags & SEGM_CLEAN)) 541 lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks); 542 543 /* 544 * Assign disk addresses, and update references to the logical 545 * block and the segment usage information. 546 */ 547 fs = sp->fs; 548 db_per_fsb = fsbtodb(fs, 1); 549 for (i = nblocks; i--; ++sp->start_bpp) { 550 lbn = *sp->start_lbp++; 551 (*sp->start_bpp)->b_blkno = off = fs->lfs_offset; 552 fs->lfs_offset += db_per_fsb; 553 554 if (error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL)) 555 panic("lfs_updatemeta: ufs_bmaparray %d", error); 556 ip = VTOI(vp); 557 switch (num) { 558 case 0: 559 ip->i_db[lbn] = off; 560 break; 561 case 1: 562 ip->i_ib[a[0].in_off] = off; 563 break; 564 default: 565 ap = &a[num - 1]; 566 if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp)) 567 panic("lfs_updatemeta: bread bno %d", 568 ap->in_lbn); 569 /* 570 * Bread may create a new indirect block which needs 571 * to get counted for the inode. 572 */ 573 if (bp->b_blkno == -1 && !(bp->b_flags & B_CACHE)) { 574 printf ("Updatemeta allocating indirect block: shouldn't happen\n"); 575 ip->i_blocks += btodb(fs->lfs_bsize); 576 fs->lfs_bfree -= btodb(fs->lfs_bsize); 577 } 578 ((daddr_t *)bp->b_data)[ap->in_off] = off; 579 VOP_BWRITE(bp); 580 } 581 582 /* Update segment usage information. */ 583 if (daddr != UNASSIGNED && 584 !(daddr >= fs->lfs_lastpseg && daddr <= off)) { 585 LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp); 586 #ifdef DIAGNOSTIC 587 if (sup->su_nbytes < fs->lfs_bsize) { 588 /* XXX -- Change to a panic. */ 589 printf("lfs: negative bytes (segment %d)\n", 590 datosn(fs, daddr)); 591 panic ("Negative Bytes"); 592 } 593 #endif 594 sup->su_nbytes -= fs->lfs_bsize; 595 error = VOP_BWRITE(bp); 596 } 597 } 598 } 599 600 /* 601 * Start a new segment. 602 */ 603 int 604 lfs_initseg(fs) 605 struct lfs *fs; 606 { 607 struct segment *sp; 608 SEGUSE *sup; 609 SEGSUM *ssp; 610 struct buf *bp; 611 daddr_t lbn, *lbnp; 612 int repeat; 613 614 sp = fs->lfs_sp; 615 616 repeat = 0; 617 /* Advance to the next segment. */ 618 if (!LFS_PARTIAL_FITS(fs)) { 619 /* Wake up any cleaning procs waiting on this file system. */ 620 wakeup(&lfs_allclean_wakeup); 621 622 lfs_newseg(fs); 623 repeat = 1; 624 fs->lfs_offset = fs->lfs_curseg; 625 sp->seg_number = datosn(fs, fs->lfs_curseg); 626 sp->seg_bytes_left = fs->lfs_dbpseg * DEV_BSIZE; 627 628 /* 629 * If the segment contains a superblock, update the offset 630 * and summary address to skip over it. 631 */ 632 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 633 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 634 fs->lfs_offset += LFS_SBPAD / DEV_BSIZE; 635 sp->seg_bytes_left -= LFS_SBPAD; 636 } 637 brelse(bp); 638 } else { 639 sp->seg_number = datosn(fs, fs->lfs_curseg); 640 sp->seg_bytes_left = (fs->lfs_dbpseg - 641 (fs->lfs_offset - fs->lfs_curseg)) * DEV_BSIZE; 642 } 643 fs->lfs_lastpseg = fs->lfs_offset; 644 645 sp->fs = fs; 646 sp->ibp = NULL; 647 sp->ninodes = 0; 648 649 /* Get a new buffer for SEGSUM and enter it into the buffer list. */ 650 sp->cbpp = sp->bpp; 651 *sp->cbpp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, fs->lfs_offset, 652 LFS_SUMMARY_SIZE); 653 sp->segsum = (*sp->cbpp)->b_data; 654 bzero(sp->segsum, LFS_SUMMARY_SIZE); 655 sp->start_bpp = ++sp->cbpp; 656 fs->lfs_offset += LFS_SUMMARY_SIZE / DEV_BSIZE; 657 658 /* Set point to SEGSUM, initialize it. */ 659 ssp = sp->segsum; 660 ssp->ss_next = fs->lfs_nextseg; 661 ssp->ss_nfinfo = ssp->ss_ninos = 0; 662 663 /* Set pointer to first FINFO, initialize it. */ 664 sp->fip = (struct finfo *)(sp->segsum + sizeof(SEGSUM)); 665 sp->fip->fi_nblocks = 0; 666 sp->start_lbp = &sp->fip->fi_blocks[0]; 667 668 sp->seg_bytes_left -= LFS_SUMMARY_SIZE; 669 sp->sum_bytes_left = LFS_SUMMARY_SIZE - sizeof(SEGSUM); 670 671 return(repeat); 672 } 673 674 /* 675 * Return the next segment to write. 676 */ 677 void 678 lfs_newseg(fs) 679 struct lfs *fs; 680 { 681 CLEANERINFO *cip; 682 SEGUSE *sup; 683 struct buf *bp; 684 int curseg, error, isdirty, sn; 685 686 LFS_SEGENTRY(sup, fs, datosn(fs, fs->lfs_nextseg), bp); 687 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 688 sup->su_nbytes = 0; 689 sup->su_nsums = 0; 690 sup->su_ninos = 0; 691 (void) VOP_BWRITE(bp); 692 693 LFS_CLEANERINFO(cip, fs, bp); 694 --cip->clean; 695 ++cip->dirty; 696 (void) VOP_BWRITE(bp); 697 698 fs->lfs_lastseg = fs->lfs_curseg; 699 fs->lfs_curseg = fs->lfs_nextseg; 700 for (sn = curseg = datosn(fs, fs->lfs_curseg);;) { 701 sn = (sn + 1) % fs->lfs_nseg; 702 if (sn == curseg) 703 panic("lfs_nextseg: no clean segments"); 704 LFS_SEGENTRY(sup, fs, sn, bp); 705 isdirty = sup->su_flags & SEGUSE_DIRTY; 706 brelse(bp); 707 if (!isdirty) 708 break; 709 } 710 711 ++fs->lfs_nactive; 712 fs->lfs_nextseg = sntoda(fs, sn); 713 #ifdef DOSTATS 714 ++lfs_stats.segsused; 715 #endif 716 } 717 718 int 719 lfs_writeseg(fs, sp) 720 struct lfs *fs; 721 struct segment *sp; 722 { 723 extern int locked_queue_count; 724 struct buf **bpp, *bp, *cbp; 725 SEGUSE *sup; 726 SEGSUM *ssp; 727 dev_t i_dev; 728 size_t size; 729 u_long *datap, *dp; 730 int ch_per_blk, do_again, error, i, nblocks, num, s; 731 int (*strategy)__P((struct vop_strategy_args *)); 732 struct vop_strategy_args vop_strategy_a; 733 u_short ninos; 734 char *p; 735 736 /* 737 * If there are no buffers other than the segment summary to write 738 * and it is not a checkpoint, don't do anything. On a checkpoint, 739 * even if there aren't any buffers, you need to write the superblock. 740 */ 741 if ((nblocks = sp->cbpp - sp->bpp) == 1) 742 return (0); 743 744 ssp = (SEGSUM *)sp->segsum; 745 746 /* Update the segment usage information. */ 747 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 748 ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs); 749 sup->su_nbytes += nblocks - 1 - ninos << fs->lfs_bshift; 750 sup->su_nbytes += ssp->ss_ninos * sizeof(struct dinode); 751 sup->su_nbytes += LFS_SUMMARY_SIZE; 752 sup->su_lastmod = time.tv_sec; 753 sup->su_ninos += ninos; 754 ++sup->su_nsums; 755 do_again = !(bp->b_flags & B_GATHERED); 756 (void)VOP_BWRITE(bp); 757 /* 758 * Compute checksum across data and then across summary; the first 759 * block (the summary block) is skipped. Set the create time here 760 * so that it's guaranteed to be later than the inode mod times. 761 * 762 * XXX 763 * Fix this to do it inline, instead of malloc/copy. 764 */ 765 datap = dp = malloc(nblocks * sizeof(u_long), M_SEGMENT, M_WAITOK); 766 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 767 if ((*++bpp)->b_flags & B_INVAL) { 768 if (copyin((*bpp)->b_saveaddr, dp++, sizeof(u_long))) 769 panic("lfs_writeseg: copyin failed"); 770 } else 771 *dp++ = ((u_long *)(*bpp)->b_data)[0]; 772 } 773 ssp->ss_create = time.tv_sec; 774 ssp->ss_datasum = cksum(datap, (nblocks - 1) * sizeof(u_long)); 775 ssp->ss_sumsum = 776 cksum(&ssp->ss_datasum, LFS_SUMMARY_SIZE - sizeof(ssp->ss_sumsum)); 777 free(datap, M_SEGMENT); 778 #ifdef DIAGNOSTIC 779 if (fs->lfs_bfree < fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE) 780 panic("lfs_writeseg: No diskspace for summary"); 781 #endif 782 fs->lfs_bfree -= (fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE); 783 784 i_dev = VTOI(fs->lfs_ivnode)->i_dev; 785 strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)]; 786 787 /* 788 * When we simply write the blocks we lose a rotation for every block 789 * written. To avoid this problem, we allocate memory in chunks, copy 790 * the buffers into the chunk and write the chunk. MAXPHYS is the 791 * largest size I/O devices can handle. 792 * When the data is copied to the chunk, turn off the the B_LOCKED bit 793 * and brelse the buffer (which will move them to the LRU list). Add 794 * the B_CALL flag to the buffer header so we can count I/O's for the 795 * checkpoints and so we can release the allocated memory. 796 * 797 * XXX 798 * This should be removed if the new virtual memory system allows us to 799 * easily make the buffers contiguous in kernel memory and if that's 800 * fast enough. 801 */ 802 ch_per_blk = MAXPHYS / fs->lfs_bsize; 803 for (bpp = sp->bpp, i = nblocks; i;) { 804 num = ch_per_blk; 805 if (num > i) 806 num = i; 807 i -= num; 808 size = num * fs->lfs_bsize; 809 810 cbp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, 811 (*bpp)->b_blkno, size); 812 cbp->b_dev = i_dev; 813 cbp->b_flags |= B_ASYNC | B_BUSY; 814 815 s = splbio(); 816 ++fs->lfs_iocount; 817 for (p = cbp->b_data; num--;) { 818 bp = *bpp++; 819 /* 820 * Fake buffers from the cleaner are marked as B_INVAL. 821 * We need to copy the data from user space rather than 822 * from the buffer indicated. 823 * XXX == what do I do on an error? 824 */ 825 if (bp->b_flags & B_INVAL) { 826 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 827 panic("lfs_writeseg: copyin failed"); 828 } else 829 bcopy(bp->b_data, p, bp->b_bcount); 830 p += bp->b_bcount; 831 if (bp->b_flags & B_LOCKED) 832 --locked_queue_count; 833 bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | 834 B_LOCKED | B_GATHERED); 835 if (bp->b_flags & B_CALL) { 836 /* if B_CALL, it was created with newbuf */ 837 brelvp(bp); 838 if (!(bp->b_flags & B_INVAL)) 839 free(bp->b_data, M_SEGMENT); 840 free(bp, M_SEGMENT); 841 } else { 842 bremfree(bp); 843 bp->b_flags |= B_DONE; 844 reassignbuf(bp, bp->b_vp); 845 brelse(bp); 846 } 847 } 848 ++cbp->b_vp->v_numoutput; 849 splx(s); 850 cbp->b_bcount = p - (char *)cbp->b_data; 851 /* 852 * XXXX This is a gross and disgusting hack. Since these 853 * buffers are physically addressed, they hang off the 854 * device vnode (devvp). As a result, they have no way 855 * of getting to the LFS superblock or lfs structure to 856 * keep track of the number of I/O's pending. So, I am 857 * going to stuff the fs into the saveaddr field of 858 * the buffer (yuk). 859 */ 860 cbp->b_saveaddr = (caddr_t)fs; 861 vop_strategy_a.a_desc = VDESC(vop_strategy); 862 vop_strategy_a.a_bp = cbp; 863 (strategy)(&vop_strategy_a); 864 } 865 /* 866 * XXX 867 * Vinvalbuf can move locked buffers off the locked queue 868 * and we have no way of knowing about this. So, after 869 * doing a big write, we recalculate how many bufers are 870 * really still left on the locked queue. 871 */ 872 locked_queue_count = count_lock_queue(); 873 wakeup(&locked_queue_count); 874 #ifdef DOSTATS 875 ++lfs_stats.psegwrites; 876 lfs_stats.blocktot += nblocks - 1; 877 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 878 ++lfs_stats.psyncwrites; 879 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 880 ++lfs_stats.pcleanwrites; 881 lfs_stats.cleanblocks += nblocks - 1; 882 } 883 #endif 884 return (lfs_initseg(fs) || do_again); 885 } 886 887 void 888 lfs_writesuper(fs) 889 struct lfs *fs; 890 { 891 struct buf *bp; 892 dev_t i_dev; 893 int (*strategy) __P((struct vop_strategy_args *)); 894 int s; 895 struct vop_strategy_args vop_strategy_a; 896 897 i_dev = VTOI(fs->lfs_ivnode)->i_dev; 898 strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)]; 899 900 /* Checksum the superblock and copy it into a buffer. */ 901 fs->lfs_cksum = cksum(fs, sizeof(struct lfs) - sizeof(fs->lfs_cksum)); 902 bp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, fs->lfs_sboffs[0], 903 LFS_SBPAD); 904 *(struct lfs *)bp->b_data = *fs; 905 906 /* XXX Toggle between first two superblocks; for now just write first */ 907 bp->b_dev = i_dev; 908 bp->b_flags |= B_BUSY | B_CALL | B_ASYNC; 909 bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI); 910 bp->b_iodone = lfs_supercallback; 911 vop_strategy_a.a_desc = VDESC(vop_strategy); 912 vop_strategy_a.a_bp = bp; 913 s = splbio(); 914 ++bp->b_vp->v_numoutput; 915 splx(s); 916 (strategy)(&vop_strategy_a); 917 } 918 919 /* 920 * Logical block number match routines used when traversing the dirty block 921 * chain. 922 */ 923 int 924 lfs_match_data(fs, bp) 925 struct lfs *fs; 926 struct buf *bp; 927 { 928 return (bp->b_lblkno >= 0); 929 } 930 931 int 932 lfs_match_indir(fs, bp) 933 struct lfs *fs; 934 struct buf *bp; 935 { 936 int lbn; 937 938 lbn = bp->b_lblkno; 939 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0); 940 } 941 942 int 943 lfs_match_dindir(fs, bp) 944 struct lfs *fs; 945 struct buf *bp; 946 { 947 int lbn; 948 949 lbn = bp->b_lblkno; 950 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1); 951 } 952 953 int 954 lfs_match_tindir(fs, bp) 955 struct lfs *fs; 956 struct buf *bp; 957 { 958 int lbn; 959 960 lbn = bp->b_lblkno; 961 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2); 962 } 963 964 /* 965 * Allocate a new buffer header. 966 */ 967 struct buf * 968 lfs_newbuf(vp, daddr, size) 969 struct vnode *vp; 970 daddr_t daddr; 971 size_t size; 972 { 973 struct buf *bp; 974 size_t nbytes; 975 976 nbytes = roundup(size, DEV_BSIZE); 977 bp = malloc(sizeof(struct buf), M_SEGMENT, M_WAITOK); 978 bzero(bp, sizeof(struct buf)); 979 if (nbytes) 980 bp->b_data = malloc(nbytes, M_SEGMENT, M_WAITOK); 981 bgetvp(vp, bp); 982 bp->b_bufsize = size; 983 bp->b_bcount = size; 984 bp->b_lblkno = daddr; 985 bp->b_blkno = daddr; 986 bp->b_error = 0; 987 bp->b_resid = 0; 988 bp->b_iodone = lfs_callback; 989 bp->b_flags |= B_BUSY | B_CALL | B_NOCACHE; 990 return (bp); 991 } 992 993 void 994 lfs_callback(bp) 995 struct buf *bp; 996 { 997 struct lfs *fs; 998 999 fs = (struct lfs *)bp->b_saveaddr; 1000 #ifdef DIAGNOSTIC 1001 if (fs->lfs_iocount == 0) 1002 panic("lfs_callback: zero iocount\n"); 1003 #endif 1004 if (--fs->lfs_iocount == 0) 1005 wakeup(&fs->lfs_iocount); 1006 1007 brelvp(bp); 1008 free(bp->b_data, M_SEGMENT); 1009 free(bp, M_SEGMENT); 1010 } 1011 1012 void 1013 lfs_supercallback(bp) 1014 struct buf *bp; 1015 { 1016 brelvp(bp); 1017 free(bp->b_data, M_SEGMENT); 1018 free(bp, M_SEGMENT); 1019 } 1020 1021 /* 1022 * Shellsort (diminishing increment sort) from Data Structures and 1023 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 1024 * see also Knuth Vol. 3, page 84. The increments are selected from 1025 * formula (8), page 95. Roughly O(N^3/2). 1026 */ 1027 /* 1028 * This is our own private copy of shellsort because we want to sort 1029 * two parallel arrays (the array of buffer pointers and the array of 1030 * logical block numbers) simultaneously. Note that we cast the array 1031 * of logical block numbers to a unsigned in this routine so that the 1032 * negative block numbers (meta data blocks) sort AFTER the data blocks. 1033 */ 1034 void 1035 lfs_shellsort(bp_array, lb_array, nmemb) 1036 struct buf **bp_array; 1037 daddr_t *lb_array; 1038 register int nmemb; 1039 { 1040 static int __rsshell_increments[] = { 4, 1, 0 }; 1041 register int incr, *incrp, t1, t2; 1042 struct buf *bp_temp; 1043 u_long lb_temp; 1044 1045 for (incrp = __rsshell_increments; incr = *incrp++;) 1046 for (t1 = incr; t1 < nmemb; ++t1) 1047 for (t2 = t1 - incr; t2 >= 0;) 1048 if (lb_array[t2] > lb_array[t2 + incr]) { 1049 lb_temp = lb_array[t2]; 1050 lb_array[t2] = lb_array[t2 + incr]; 1051 lb_array[t2 + incr] = lb_temp; 1052 bp_temp = bp_array[t2]; 1053 bp_array[t2] = bp_array[t2 + incr]; 1054 bp_array[t2 + incr] = bp_temp; 1055 t2 -= incr; 1056 } else 1057 break; 1058 } 1059 1060 /* 1061 * Check VXLOCK. Return 1 if the vnode is locked. Otherwise, bump the 1062 * ref count, removing the vnode from the free list if it is on it. 1063 */ 1064 lfs_vref(vp) 1065 register struct vnode *vp; 1066 { 1067 register struct vnode *vq; 1068 extern struct vnode *vfreeh; 1069 extern struct vnode **vfreet; 1070 1071 if (vp->v_flag & VXLOCK) 1072 return(1); 1073 1074 if (vp->v_usecount == 0) { 1075 if (vq = vp->v_freef) 1076 vq->v_freeb = vp->v_freeb; 1077 else 1078 vfreet = vp->v_freeb; 1079 *vp->v_freeb = vq; 1080 vp->v_freef = NULL; 1081 vp->v_freeb = NULL; 1082 } 1083 VREF(vp); 1084 return (0); 1085 } 1086 1087 void 1088 lfs_vunref(vp) 1089 register struct vnode *vp; 1090 { 1091 extern struct vnode *vfreeh; 1092 extern struct vnode **vfreet; 1093 1094 --vp->v_usecount; 1095 1096 /* 1097 * return to free list 1098 */ 1099 if (vp->v_usecount == 0) { 1100 *vfreet = vp; 1101 vp->v_freeb = vfreet; 1102 vp->v_freef = NULL; 1103 vfreet = &vp->v_freef; 1104 } 1105 return; 1106 } 1107