lfs_segment.c revision 1.256
1 /* $NetBSD: lfs_segment.c,v 1.256 2015/08/12 18:28:01 dholland Exp $ */ 2 3 /*- 4 * Copyright (c) 1999, 2000, 2001, 2002, 2003 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 (at) 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 * Copyright (c) 1991, 1993 33 * The Regents of the University of California. All rights reserved. 34 * 35 * Redistribution and use in source and binary forms, with or without 36 * modification, are permitted provided that the following conditions 37 * are met: 38 * 1. Redistributions of source code must retain the above copyright 39 * notice, this list of conditions and the following disclaimer. 40 * 2. Redistributions in binary form must reproduce the above copyright 41 * notice, this list of conditions and the following disclaimer in the 42 * documentation and/or other materials provided with the distribution. 43 * 3. Neither the name of the University nor the names of its contributors 44 * may be used to endorse or promote products derived from this software 45 * without specific prior written permission. 46 * 47 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 48 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 49 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 50 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 51 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 52 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 53 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 54 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 55 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 56 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 57 * SUCH DAMAGE. 58 * 59 * @(#)lfs_segment.c 8.10 (Berkeley) 6/10/95 60 */ 61 62 #include <sys/cdefs.h> 63 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.256 2015/08/12 18:28:01 dholland Exp $"); 64 65 #define _VFS_VNODE_PRIVATE /* XXX: check for VI_MARKER, this has to go */ 66 67 #ifdef DEBUG 68 # define vndebug(vp, str) do { \ 69 if (VTOI(vp)->i_flag & IN_CLEANING) \ 70 DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \ 71 VTOI(vp)->i_number, (str), op)); \ 72 } while(0) 73 #else 74 # define vndebug(vp, str) 75 #endif 76 #define ivndebug(vp, str) \ 77 DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str))) 78 79 #if defined(_KERNEL_OPT) 80 #include "opt_ddb.h" 81 #endif 82 83 #include <sys/param.h> 84 #include <sys/systm.h> 85 #include <sys/namei.h> 86 #include <sys/kernel.h> 87 #include <sys/resourcevar.h> 88 #include <sys/file.h> 89 #include <sys/stat.h> 90 #include <sys/buf.h> 91 #include <sys/proc.h> 92 #include <sys/vnode.h> 93 #include <sys/mount.h> 94 #include <sys/kauth.h> 95 #include <sys/syslog.h> 96 97 #include <miscfs/specfs/specdev.h> 98 #include <miscfs/fifofs/fifo.h> 99 100 #include <ufs/lfs/ulfs_inode.h> 101 #include <ufs/lfs/ulfsmount.h> 102 #include <ufs/lfs/ulfs_extern.h> 103 104 #include <ufs/lfs/lfs.h> 105 #include <ufs/lfs/lfs_accessors.h> 106 #include <ufs/lfs/lfs_kernel.h> 107 #include <ufs/lfs/lfs_extern.h> 108 109 #include <uvm/uvm.h> 110 #include <uvm/uvm_extern.h> 111 112 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 113 114 static void lfs_generic_callback(struct buf *, void (*)(struct buf *)); 115 static void lfs_free_aiodone(struct buf *); 116 static void lfs_super_aiodone(struct buf *); 117 static void lfs_cluster_aiodone(struct buf *); 118 static void lfs_cluster_callback(struct buf *); 119 120 /* 121 * Determine if it's OK to start a partial in this segment, or if we need 122 * to go on to a new segment. 123 */ 124 #define LFS_PARTIAL_FITS(fs) \ 125 (lfs_sb_getfsbpseg(fs) - \ 126 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)) > \ 127 lfs_sb_getfrag(fs)) 128 129 /* 130 * Figure out whether we should do a checkpoint write or go ahead with 131 * an ordinary write. 132 */ 133 #define LFS_SHOULD_CHECKPOINT(fs, flags) \ 134 ((flags & SEGM_CLEAN) == 0 && \ 135 ((fs->lfs_nactive > LFS_MAX_ACTIVE || \ 136 (flags & SEGM_CKP) || \ 137 lfs_sb_getnclean(fs) < LFS_MAX_ACTIVE))) 138 139 int lfs_match_fake(struct lfs *, struct buf *); 140 void lfs_newseg(struct lfs *); 141 void lfs_supercallback(struct buf *); 142 void lfs_updatemeta(struct segment *); 143 void lfs_writesuper(struct lfs *, daddr_t); 144 int lfs_writevnodes(struct lfs *fs, struct mount *mp, 145 struct segment *sp, int dirops); 146 147 static void lfs_shellsort(struct lfs *, struct buf **, union lfs_blocks *, 148 int, int); 149 150 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 151 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 152 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 153 int lfs_dirvcount = 0; /* # active dirops */ 154 155 /* Statistics Counters */ 156 int lfs_dostats = 1; 157 struct lfs_stats lfs_stats; 158 159 /* op values to lfs_writevnodes */ 160 #define VN_REG 0 161 #define VN_DIROP 1 162 #define VN_EMPTY 2 163 #define VN_CLEAN 3 164 165 /* 166 * XXX KS - Set modification time on the Ifile, so the cleaner can 167 * read the fs mod time off of it. We don't set IN_UPDATE here, 168 * since we don't really need this to be flushed to disk (and in any 169 * case that wouldn't happen to the Ifile until we checkpoint). 170 */ 171 void 172 lfs_imtime(struct lfs *fs) 173 { 174 struct timespec ts; 175 struct inode *ip; 176 177 ASSERT_MAYBE_SEGLOCK(fs); 178 vfs_timestamp(&ts); 179 ip = VTOI(fs->lfs_ivnode); 180 ip->i_ffs1_mtime = ts.tv_sec; 181 ip->i_ffs1_mtimensec = ts.tv_nsec; 182 } 183 184 /* 185 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 186 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 187 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 188 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 189 */ 190 191 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 192 193 int 194 lfs_vflush(struct vnode *vp) 195 { 196 struct inode *ip; 197 struct lfs *fs; 198 struct segment *sp; 199 struct buf *bp, *nbp, *tbp, *tnbp; 200 int error; 201 int flushed; 202 int relock; 203 204 ip = VTOI(vp); 205 fs = VFSTOULFS(vp->v_mount)->um_lfs; 206 relock = 0; 207 208 top: 209 KASSERT(mutex_owned(vp->v_interlock) == false); 210 KASSERT(mutex_owned(&lfs_lock) == false); 211 KASSERT(mutex_owned(&bufcache_lock) == false); 212 ASSERT_NO_SEGLOCK(fs); 213 if (ip->i_flag & IN_CLEANING) { 214 ivndebug(vp,"vflush/in_cleaning"); 215 mutex_enter(&lfs_lock); 216 LFS_CLR_UINO(ip, IN_CLEANING); 217 LFS_SET_UINO(ip, IN_MODIFIED); 218 mutex_exit(&lfs_lock); 219 220 /* 221 * Toss any cleaning buffers that have real counterparts 222 * to avoid losing new data. 223 */ 224 mutex_enter(vp->v_interlock); 225 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 226 nbp = LIST_NEXT(bp, b_vnbufs); 227 if (!LFS_IS_MALLOC_BUF(bp)) 228 continue; 229 /* 230 * Look for pages matching the range covered 231 * by cleaning blocks. It's okay if more dirty 232 * pages appear, so long as none disappear out 233 * from under us. 234 */ 235 if (bp->b_lblkno > 0 && vp->v_type == VREG && 236 vp != fs->lfs_ivnode) { 237 struct vm_page *pg; 238 voff_t off; 239 240 for (off = lfs_lblktosize(fs, bp->b_lblkno); 241 off < lfs_lblktosize(fs, bp->b_lblkno + 1); 242 off += PAGE_SIZE) { 243 pg = uvm_pagelookup(&vp->v_uobj, off); 244 if (pg == NULL) 245 continue; 246 if ((pg->flags & PG_CLEAN) == 0 || 247 pmap_is_modified(pg)) { 248 lfs_sb_addavail(fs, 249 lfs_btofsb(fs, 250 bp->b_bcount)); 251 wakeup(&fs->lfs_availsleep); 252 mutex_exit(vp->v_interlock); 253 lfs_freebuf(fs, bp); 254 mutex_enter(vp->v_interlock); 255 bp = NULL; 256 break; 257 } 258 } 259 } 260 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 261 tbp = tnbp) 262 { 263 tnbp = LIST_NEXT(tbp, b_vnbufs); 264 if (tbp->b_vp == bp->b_vp 265 && tbp->b_lblkno == bp->b_lblkno 266 && tbp != bp) 267 { 268 lfs_sb_addavail(fs, lfs_btofsb(fs, 269 bp->b_bcount)); 270 wakeup(&fs->lfs_availsleep); 271 mutex_exit(vp->v_interlock); 272 lfs_freebuf(fs, bp); 273 mutex_enter(vp->v_interlock); 274 bp = NULL; 275 break; 276 } 277 } 278 } 279 } else { 280 mutex_enter(vp->v_interlock); 281 } 282 283 /* If the node is being written, wait until that is done */ 284 while (WRITEINPROG(vp)) { 285 ivndebug(vp,"vflush/writeinprog"); 286 cv_wait(&vp->v_cv, vp->v_interlock); 287 } 288 error = vdead_check(vp, VDEAD_NOWAIT); 289 mutex_exit(vp->v_interlock); 290 291 /* Protect against deadlock in vinvalbuf() */ 292 lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0)); 293 if (error != 0) { 294 fs->lfs_reclino = ip->i_number; 295 } 296 297 /* If we're supposed to flush a freed inode, just toss it */ 298 if (ip->i_lfs_iflags & LFSI_DELETED) { 299 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 300 ip->i_number)); 301 /* Drain v_numoutput */ 302 mutex_enter(vp->v_interlock); 303 while (vp->v_numoutput > 0) { 304 cv_wait(&vp->v_cv, vp->v_interlock); 305 } 306 KASSERT(vp->v_numoutput == 0); 307 mutex_exit(vp->v_interlock); 308 309 mutex_enter(&bufcache_lock); 310 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 311 nbp = LIST_NEXT(bp, b_vnbufs); 312 313 KASSERT((bp->b_flags & B_GATHERED) == 0); 314 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 315 lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount)); 316 wakeup(&fs->lfs_availsleep); 317 } 318 /* Copied from lfs_writeseg */ 319 if (bp->b_iodone != NULL) { 320 mutex_exit(&bufcache_lock); 321 biodone(bp); 322 mutex_enter(&bufcache_lock); 323 } else { 324 bremfree(bp); 325 LFS_UNLOCK_BUF(bp); 326 mutex_enter(vp->v_interlock); 327 bp->b_flags &= ~(B_READ | B_GATHERED); 328 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 329 bp->b_error = 0; 330 reassignbuf(bp, vp); 331 mutex_exit(vp->v_interlock); 332 brelse(bp, 0); 333 } 334 } 335 mutex_exit(&bufcache_lock); 336 LFS_CLR_UINO(ip, IN_CLEANING); 337 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 338 ip->i_flag &= ~IN_ALLMOD; 339 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 340 ip->i_number)); 341 lfs_segunlock(fs); 342 343 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 344 345 return 0; 346 } 347 348 fs->lfs_flushvp = vp; 349 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 350 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 351 fs->lfs_flushvp = NULL; 352 KASSERT(fs->lfs_flushvp_fakevref == 0); 353 lfs_segunlock(fs); 354 355 /* Make sure that any pending buffers get written */ 356 mutex_enter(vp->v_interlock); 357 while (vp->v_numoutput > 0) { 358 cv_wait(&vp->v_cv, vp->v_interlock); 359 } 360 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 361 KASSERT(vp->v_numoutput == 0); 362 mutex_exit(vp->v_interlock); 363 364 return error; 365 } 366 sp = fs->lfs_sp; 367 368 flushed = 0; 369 if (VPISEMPTY(vp)) { 370 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 371 ++flushed; 372 } else if ((ip->i_flag & IN_CLEANING) && 373 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 374 ivndebug(vp,"vflush/clean"); 375 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 376 ++flushed; 377 } else if (lfs_dostats) { 378 if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD)) 379 ++lfs_stats.vflush_invoked; 380 ivndebug(vp,"vflush"); 381 } 382 383 #ifdef DIAGNOSTIC 384 if (vp->v_uflag & VU_DIROP) { 385 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 386 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 387 } 388 #endif 389 390 do { 391 #ifdef DEBUG 392 int loopcount = 0; 393 #endif 394 do { 395 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 396 relock = lfs_writefile(fs, sp, vp); 397 if (relock && vp != fs->lfs_ivnode) { 398 /* 399 * Might have to wait for the 400 * cleaner to run; but we're 401 * still not done with this vnode. 402 * XXX we can do better than this. 403 */ 404 KDASSERT(ip->i_number != LFS_IFILE_INUM); 405 lfs_writeinode(fs, sp, ip); 406 mutex_enter(&lfs_lock); 407 LFS_SET_UINO(ip, IN_MODIFIED); 408 mutex_exit(&lfs_lock); 409 lfs_writeseg(fs, sp); 410 lfs_segunlock(fs); 411 lfs_segunlock_relock(fs); 412 goto top; 413 } 414 } 415 /* 416 * If we begin a new segment in the middle of writing 417 * the Ifile, it creates an inconsistent checkpoint, 418 * since the Ifile information for the new segment 419 * is not up-to-date. Take care of this here by 420 * sending the Ifile through again in case there 421 * are newly dirtied blocks. But wait, there's more! 422 * This second Ifile write could *also* cross a segment 423 * boundary, if the first one was large. The second 424 * one is guaranteed to be no more than 8 blocks, 425 * though (two segment blocks and supporting indirects) 426 * so the third write *will not* cross the boundary. 427 */ 428 if (vp == fs->lfs_ivnode) { 429 lfs_writefile(fs, sp, vp); 430 lfs_writefile(fs, sp, vp); 431 } 432 #ifdef DEBUG 433 if (++loopcount > 2) 434 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 435 #endif 436 } while (lfs_writeinode(fs, sp, ip)); 437 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 438 439 if (lfs_dostats) { 440 ++lfs_stats.nwrites; 441 if (sp->seg_flags & SEGM_SYNC) 442 ++lfs_stats.nsync_writes; 443 if (sp->seg_flags & SEGM_CKP) 444 ++lfs_stats.ncheckpoints; 445 } 446 /* 447 * If we were called from somewhere that has already held the seglock 448 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 449 * the write to complete because we are still locked. 450 * Since lfs_vflush() must return the vnode with no dirty buffers, 451 * we must explicitly wait, if that is the case. 452 * 453 * We compare the iocount against 1, not 0, because it is 454 * artificially incremented by lfs_seglock(). 455 */ 456 mutex_enter(&lfs_lock); 457 if (fs->lfs_seglock > 1) { 458 while (fs->lfs_iocount > 1) 459 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 460 "lfs_vflush", 0, &lfs_lock); 461 } 462 mutex_exit(&lfs_lock); 463 464 lfs_segunlock(fs); 465 466 /* Wait for these buffers to be recovered by aiodoned */ 467 mutex_enter(vp->v_interlock); 468 while (vp->v_numoutput > 0) { 469 cv_wait(&vp->v_cv, vp->v_interlock); 470 } 471 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 472 KASSERT(vp->v_numoutput == 0); 473 mutex_exit(vp->v_interlock); 474 475 fs->lfs_flushvp = NULL; 476 KASSERT(fs->lfs_flushvp_fakevref == 0); 477 478 return (0); 479 } 480 481 int 482 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 483 { 484 struct inode *ip; 485 struct vnode *vp; 486 int inodes_written = 0; 487 int error = 0; 488 489 ASSERT_SEGLOCK(fs); 490 loop: 491 /* start at last (newest) vnode. */ 492 mutex_enter(&mntvnode_lock); 493 TAILQ_FOREACH_REVERSE(vp, &mp->mnt_vnodelist, vnodelst, v_mntvnodes) { 494 /* 495 * If the vnode that we are about to sync is no longer 496 * associated with this mount point, start over. 497 */ 498 if (vp->v_mount != mp) { 499 DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n")); 500 /* 501 * After this, pages might be busy 502 * due to our own previous putpages. 503 * Start actual segment write here to avoid deadlock. 504 * If we were just writing one segment and we've done 505 * that, break out. 506 */ 507 mutex_exit(&mntvnode_lock); 508 if (lfs_writeseg(fs, sp) && 509 (sp->seg_flags & SEGM_SINGLE) && 510 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 511 DLOG((DLOG_VNODE, "lfs_writevnodes: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 512 break; 513 } 514 goto loop; 515 } 516 517 mutex_enter(vp->v_interlock); 518 if (vp->v_type == VNON || (vp->v_iflag & VI_MARKER) || 519 vdead_check(vp, VDEAD_NOWAIT) != 0) { 520 mutex_exit(vp->v_interlock); 521 continue; 522 } 523 524 ip = VTOI(vp); 525 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 526 (op != VN_DIROP && op != VN_CLEAN && 527 (vp->v_uflag & VU_DIROP))) { 528 mutex_exit(vp->v_interlock); 529 vndebug(vp,"dirop"); 530 continue; 531 } 532 533 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 534 mutex_exit(vp->v_interlock); 535 vndebug(vp,"empty"); 536 continue; 537 } 538 539 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM 540 && vp != fs->lfs_flushvp 541 && !(ip->i_flag & IN_CLEANING)) { 542 mutex_exit(vp->v_interlock); 543 vndebug(vp,"cleaning"); 544 continue; 545 } 546 547 mutex_exit(&mntvnode_lock); 548 if (vget(vp, LK_NOWAIT, false /* !wait */)) { 549 vndebug(vp,"vget"); 550 mutex_enter(&mntvnode_lock); 551 continue; 552 } 553 554 /* 555 * Write the inode/file if dirty and it's not the IFILE. 556 */ 557 if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) { 558 if (ip->i_number != LFS_IFILE_INUM) { 559 error = lfs_writefile(fs, sp, vp); 560 if (error) { 561 vrele(vp); 562 if (error == EAGAIN) { 563 /* 564 * This error from lfs_putpages 565 * indicates we need to drop 566 * the segment lock and start 567 * over after the cleaner has 568 * had a chance to run. 569 */ 570 lfs_writeinode(fs, sp, ip); 571 lfs_writeseg(fs, sp); 572 if (!VPISEMPTY(vp) && 573 !WRITEINPROG(vp) && 574 !(ip->i_flag & IN_ALLMOD)) { 575 mutex_enter(&lfs_lock); 576 LFS_SET_UINO(ip, IN_MODIFIED); 577 mutex_exit(&lfs_lock); 578 } 579 mutex_enter(&mntvnode_lock); 580 break; 581 } 582 error = 0; /* XXX not quite right */ 583 mutex_enter(&mntvnode_lock); 584 continue; 585 } 586 587 if (!VPISEMPTY(vp)) { 588 if (WRITEINPROG(vp)) { 589 ivndebug(vp,"writevnodes/write2"); 590 } else if (!(ip->i_flag & IN_ALLMOD)) { 591 mutex_enter(&lfs_lock); 592 LFS_SET_UINO(ip, IN_MODIFIED); 593 mutex_exit(&lfs_lock); 594 } 595 } 596 (void) lfs_writeinode(fs, sp, ip); 597 inodes_written++; 598 } 599 } 600 601 vrele(vp); 602 603 mutex_enter(&mntvnode_lock); 604 } 605 mutex_exit(&mntvnode_lock); 606 return error; 607 } 608 609 /* 610 * Do a checkpoint. 611 */ 612 int 613 lfs_segwrite(struct mount *mp, int flags) 614 { 615 struct buf *bp; 616 struct inode *ip; 617 struct lfs *fs; 618 struct segment *sp; 619 struct vnode *vp; 620 SEGUSE *segusep; 621 int do_ckp, did_ckp, error; 622 unsigned n, segleft, maxseg, sn, i, curseg; 623 int writer_set = 0; 624 int dirty; 625 int redo; 626 SEGSUM *ssp; 627 int um_error; 628 629 fs = VFSTOULFS(mp)->um_lfs; 630 ASSERT_MAYBE_SEGLOCK(fs); 631 632 if (fs->lfs_ronly) 633 return EROFS; 634 635 lfs_imtime(fs); 636 637 /* 638 * Allocate a segment structure and enough space to hold pointers to 639 * the maximum possible number of buffers which can be described in a 640 * single summary block. 641 */ 642 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 643 644 /* We can't do a partial write and checkpoint at the same time. */ 645 if (do_ckp) 646 flags &= ~SEGM_SINGLE; 647 648 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 649 sp = fs->lfs_sp; 650 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 651 do_ckp = 1; 652 653 /* 654 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 655 * in which case we have to flush *all* buffers off of this vnode. 656 * We don't care about other nodes, but write any non-dirop nodes 657 * anyway in anticipation of another getnewvnode(). 658 * 659 * If we're cleaning we only write cleaning and ifile blocks, and 660 * no dirops, since otherwise we'd risk corruption in a crash. 661 */ 662 if (sp->seg_flags & SEGM_CLEAN) 663 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 664 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 665 do { 666 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 667 if ((sp->seg_flags & SEGM_SINGLE) && 668 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 669 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 670 break; 671 } 672 673 if (do_ckp || fs->lfs_dirops == 0) { 674 if (!writer_set) { 675 lfs_writer_enter(fs, "lfs writer"); 676 writer_set = 1; 677 } 678 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 679 if (um_error == 0) 680 um_error = error; 681 /* In case writevnodes errored out */ 682 lfs_flush_dirops(fs); 683 ssp = (SEGSUM *)(sp->segsum); 684 lfs_ss_setflags(fs, ssp, 685 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 686 lfs_finalize_fs_seguse(fs); 687 } 688 if (do_ckp && um_error) { 689 lfs_segunlock_relock(fs); 690 sp = fs->lfs_sp; 691 } 692 } while (do_ckp && um_error != 0); 693 } 694 695 /* 696 * If we are doing a checkpoint, mark everything since the 697 * last checkpoint as no longer ACTIVE. 698 */ 699 if (do_ckp || fs->lfs_doifile) { 700 segleft = lfs_sb_getnseg(fs); 701 curseg = 0; 702 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 703 dirty = 0; 704 if (bread(fs->lfs_ivnode, lfs_sb_getcleansz(fs) + n, 705 lfs_sb_getbsize(fs), B_MODIFY, &bp)) 706 panic("lfs_segwrite: ifile read"); 707 segusep = (SEGUSE *)bp->b_data; 708 maxseg = min(segleft, lfs_sb_getsepb(fs)); 709 for (i = 0; i < maxseg; i++) { 710 sn = curseg + i; 711 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 712 segusep->su_flags & SEGUSE_ACTIVE) { 713 segusep->su_flags &= ~SEGUSE_ACTIVE; 714 --fs->lfs_nactive; 715 ++dirty; 716 } 717 fs->lfs_suflags[fs->lfs_activesb][sn] = 718 segusep->su_flags; 719 if (lfs_sb_getversion(fs) > 1) 720 ++segusep; 721 else 722 segusep = (SEGUSE *) 723 ((SEGUSE_V1 *)segusep + 1); 724 } 725 726 if (dirty) 727 error = LFS_BWRITE_LOG(bp); /* Ifile */ 728 else 729 brelse(bp, 0); 730 segleft -= lfs_sb_getsepb(fs); 731 curseg += lfs_sb_getsepb(fs); 732 } 733 } 734 735 KASSERT(LFS_SEGLOCK_HELD(fs)); 736 737 did_ckp = 0; 738 if (do_ckp || fs->lfs_doifile) { 739 vp = fs->lfs_ivnode; 740 #ifdef DEBUG 741 int loopcount = 0; 742 #endif 743 do { 744 #ifdef DEBUG 745 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 746 #endif 747 mutex_enter(&lfs_lock); 748 fs->lfs_flags &= ~LFS_IFDIRTY; 749 mutex_exit(&lfs_lock); 750 751 ip = VTOI(vp); 752 753 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 754 /* 755 * Ifile has no pages, so we don't need 756 * to check error return here. 757 */ 758 lfs_writefile(fs, sp, vp); 759 /* 760 * Ensure the Ifile takes the current segment 761 * into account. See comment in lfs_vflush. 762 */ 763 lfs_writefile(fs, sp, vp); 764 lfs_writefile(fs, sp, vp); 765 } 766 767 if (ip->i_flag & IN_ALLMOD) 768 ++did_ckp; 769 #if 0 770 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 771 #else 772 redo = lfs_writeinode(fs, sp, ip); 773 #endif 774 redo += lfs_writeseg(fs, sp); 775 mutex_enter(&lfs_lock); 776 redo += (fs->lfs_flags & LFS_IFDIRTY); 777 mutex_exit(&lfs_lock); 778 #ifdef DEBUG 779 if (++loopcount > 2) 780 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 781 loopcount); 782 #endif 783 } while (redo && do_ckp); 784 785 /* 786 * Unless we are unmounting, the Ifile may continue to have 787 * dirty blocks even after a checkpoint, due to changes to 788 * inodes' atime. If we're checkpointing, it's "impossible" 789 * for other parts of the Ifile to be dirty after the loop 790 * above, since we hold the segment lock. 791 */ 792 mutex_enter(vp->v_interlock); 793 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 794 LFS_CLR_UINO(ip, IN_ALLMOD); 795 } 796 #ifdef DIAGNOSTIC 797 else if (do_ckp) { 798 int do_panic = 0; 799 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 800 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 801 lfs_sb_getsegtabsz(fs) && 802 !(bp->b_flags & B_GATHERED)) { 803 printf("ifile lbn %ld still dirty (flags %lx)\n", 804 (long)bp->b_lblkno, 805 (long)bp->b_flags); 806 ++do_panic; 807 } 808 } 809 if (do_panic) 810 panic("dirty blocks"); 811 } 812 #endif 813 mutex_exit(vp->v_interlock); 814 } else { 815 (void) lfs_writeseg(fs, sp); 816 } 817 818 /* Note Ifile no longer needs to be written */ 819 fs->lfs_doifile = 0; 820 if (writer_set) 821 lfs_writer_leave(fs); 822 823 /* 824 * If we didn't write the Ifile, we didn't really do anything. 825 * That means that (1) there is a checkpoint on disk and (2) 826 * nothing has changed since it was written. 827 * 828 * Take the flags off of the segment so that lfs_segunlock 829 * doesn't have to write the superblock either. 830 */ 831 if (do_ckp && !did_ckp) { 832 sp->seg_flags &= ~SEGM_CKP; 833 } 834 835 if (lfs_dostats) { 836 ++lfs_stats.nwrites; 837 if (sp->seg_flags & SEGM_SYNC) 838 ++lfs_stats.nsync_writes; 839 if (sp->seg_flags & SEGM_CKP) 840 ++lfs_stats.ncheckpoints; 841 } 842 lfs_segunlock(fs); 843 return (0); 844 } 845 846 /* 847 * Write the dirty blocks associated with a vnode. 848 */ 849 int 850 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 851 { 852 struct inode *ip; 853 int i, frag; 854 SEGSUM *ssp; 855 int error; 856 857 ASSERT_SEGLOCK(fs); 858 error = 0; 859 ip = VTOI(vp); 860 861 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 862 863 if (vp->v_uflag & VU_DIROP) { 864 ssp = (SEGSUM *)sp->segsum; 865 lfs_ss_setflags(fs, ssp, 866 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 867 } 868 869 if (sp->seg_flags & SEGM_CLEAN) { 870 lfs_gather(fs, sp, vp, lfs_match_fake); 871 /* 872 * For a file being flushed, we need to write *all* blocks. 873 * This means writing the cleaning blocks first, and then 874 * immediately following with any non-cleaning blocks. 875 * The same is true of the Ifile since checkpoints assume 876 * that all valid Ifile blocks are written. 877 */ 878 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 879 lfs_gather(fs, sp, vp, lfs_match_data); 880 /* 881 * Don't call VOP_PUTPAGES: if we're flushing, 882 * we've already done it, and the Ifile doesn't 883 * use the page cache. 884 */ 885 } 886 } else { 887 lfs_gather(fs, sp, vp, lfs_match_data); 888 /* 889 * If we're flushing, we've already called VOP_PUTPAGES 890 * so don't do it again. Otherwise, we want to write 891 * everything we've got. 892 */ 893 if (!IS_FLUSHING(fs, vp)) { 894 mutex_enter(vp->v_interlock); 895 error = VOP_PUTPAGES(vp, 0, 0, 896 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 897 } 898 } 899 900 /* 901 * It may not be necessary to write the meta-data blocks at this point, 902 * as the roll-forward recovery code should be able to reconstruct the 903 * list. 904 * 905 * We have to write them anyway, though, under two conditions: (1) the 906 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 907 * checkpointing. 908 * 909 * BUT if we are cleaning, we might have indirect blocks that refer to 910 * new blocks not being written yet, in addition to fragments being 911 * moved out of a cleaned segment. If that is the case, don't 912 * write the indirect blocks, or the finfo will have a small block 913 * in the middle of it! 914 * XXX in this case isn't the inode size wrong too? 915 */ 916 frag = 0; 917 if (sp->seg_flags & SEGM_CLEAN) { 918 for (i = 0; i < ULFS_NDADDR; i++) 919 if (ip->i_lfs_fragsize[i] > 0 && 920 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 921 ++frag; 922 } 923 #ifdef DIAGNOSTIC 924 if (frag > 1) 925 panic("lfs_writefile: more than one fragment!"); 926 #endif 927 if (IS_FLUSHING(fs, vp) || 928 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 929 lfs_gather(fs, sp, vp, lfs_match_indir); 930 lfs_gather(fs, sp, vp, lfs_match_dindir); 931 lfs_gather(fs, sp, vp, lfs_match_tindir); 932 } 933 lfs_release_finfo(fs); 934 935 return error; 936 } 937 938 /* 939 * Update segment accounting to reflect this inode's change of address. 940 */ 941 static int 942 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 943 { 944 struct buf *bp; 945 daddr_t daddr; 946 IFILE *ifp; 947 SEGUSE *sup; 948 ino_t ino; 949 int redo_ifile; 950 u_int32_t sn; 951 952 redo_ifile = 0; 953 954 /* 955 * If updating the ifile, update the super-block. Update the disk 956 * address and access times for this inode in the ifile. 957 */ 958 ino = ip->i_number; 959 if (ino == LFS_IFILE_INUM) { 960 daddr = lfs_sb_getidaddr(fs); 961 lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr)); 962 } else { 963 LFS_IENTRY(ifp, fs, ino, bp); 964 daddr = lfs_if_getdaddr(fs, ifp); 965 lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr)); 966 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 967 } 968 969 /* 970 * If this is the Ifile and lfs_offset is set to the first block 971 * in the segment, dirty the new segment's accounting block 972 * (XXX should already be dirty?) and tell the caller to do it again. 973 */ 974 if (ip->i_number == LFS_IFILE_INUM) { 975 sn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 976 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) == 977 lfs_sb_getoffset(fs)) { 978 LFS_SEGENTRY(sup, fs, sn, bp); 979 KASSERT(bp->b_oflags & BO_DELWRI); 980 LFS_WRITESEGENTRY(sup, fs, sn, bp); 981 /* fs->lfs_flags |= LFS_IFDIRTY; */ 982 redo_ifile |= 1; 983 } 984 } 985 986 /* 987 * The inode's last address should not be in the current partial 988 * segment, except under exceptional circumstances (lfs_writevnodes 989 * had to start over, and in the meantime more blocks were written 990 * to a vnode). Both inodes will be accounted to this segment 991 * in lfs_writeseg so we need to subtract the earlier version 992 * here anyway. The segment count can temporarily dip below 993 * zero here; keep track of how many duplicates we have in 994 * "dupino" so we don't panic below. 995 */ 996 if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) { 997 ++sp->ndupino; 998 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 999 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 1000 (long long)daddr, sp->ndupino)); 1001 } 1002 /* 1003 * Account the inode: it no longer belongs to its former segment, 1004 * though it will not belong to the new segment until that segment 1005 * is actually written. 1006 */ 1007 if (daddr != LFS_UNUSED_DADDR) { 1008 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1009 #ifdef DIAGNOSTIC 1010 int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0; 1011 #endif 1012 LFS_SEGENTRY(sup, fs, oldsn, bp); 1013 #ifdef DIAGNOSTIC 1014 if (sup->su_nbytes + DINOSIZE(fs) * ndupino < DINOSIZE(fs)) { 1015 printf("lfs_writeinode: negative bytes " 1016 "(segment %" PRIu32 " short by %d, " 1017 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1018 ", daddr=%" PRId64 ", su_nbytes=%u, " 1019 "ndupino=%d)\n", 1020 lfs_dtosn(fs, daddr), 1021 (int)DINOSIZE(fs) * 1022 (1 - sp->ndupino) - sup->su_nbytes, 1023 oldsn, sp->seg_number, daddr, 1024 (unsigned int)sup->su_nbytes, 1025 sp->ndupino); 1026 panic("lfs_writeinode: negative bytes"); 1027 sup->su_nbytes = DINOSIZE(fs); 1028 } 1029 #endif 1030 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1031 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1032 sup->su_nbytes -= DINOSIZE(fs); 1033 redo_ifile |= 1034 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1035 if (redo_ifile) { 1036 mutex_enter(&lfs_lock); 1037 fs->lfs_flags |= LFS_IFDIRTY; 1038 mutex_exit(&lfs_lock); 1039 /* Don't double-account */ 1040 lfs_sb_setidaddr(fs, 0x0); 1041 } 1042 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1043 } 1044 1045 return redo_ifile; 1046 } 1047 1048 int 1049 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1050 { 1051 struct buf *bp; 1052 union lfs_dinode *cdp; 1053 struct vnode *vp = ITOV(ip); 1054 daddr_t daddr; 1055 int i, ndx; 1056 int redo_ifile = 0; 1057 int gotblk = 0; 1058 int count; 1059 SEGSUM *ssp; 1060 1061 ASSERT_SEGLOCK(fs); 1062 if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1063 return (0); 1064 1065 /* Can't write ifile when writer is not set */ 1066 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1067 (sp->seg_flags & SEGM_CLEAN)); 1068 1069 /* 1070 * If this is the Ifile, see if writing it here will generate a 1071 * temporary misaccounting. If it will, do the accounting and write 1072 * the blocks, postponing the inode write until the accounting is 1073 * solid. 1074 */ 1075 count = 0; 1076 while (vp == fs->lfs_ivnode) { 1077 int redo = 0; 1078 1079 if (sp->idp == NULL && sp->ibp == NULL && 1080 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1081 sp->sum_bytes_left < sizeof(int32_t))) { 1082 (void) lfs_writeseg(fs, sp); 1083 continue; 1084 } 1085 1086 /* Look for dirty Ifile blocks */ 1087 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1088 if (!(bp->b_flags & B_GATHERED)) { 1089 redo = 1; 1090 break; 1091 } 1092 } 1093 1094 if (redo == 0) 1095 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1096 if (redo == 0) 1097 break; 1098 1099 if (sp->idp) { 1100 lfs_dino_setinumber(fs, sp->idp, 0); 1101 sp->idp = NULL; 1102 } 1103 ++count; 1104 if (count > 2) 1105 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1106 lfs_writefile(fs, sp, fs->lfs_ivnode); 1107 } 1108 1109 /* Allocate a new inode block if necessary. */ 1110 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1111 sp->ibp == NULL) { 1112 /* Allocate a new segment if necessary. */ 1113 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1114 sp->sum_bytes_left < sizeof(int32_t)) 1115 (void) lfs_writeseg(fs, sp); 1116 1117 /* Get next inode block. */ 1118 daddr = lfs_sb_getoffset(fs); 1119 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1120 sp->ibp = *sp->cbpp++ = 1121 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1122 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1123 gotblk++; 1124 1125 /* Zero out inode numbers */ 1126 for (i = 0; i < LFS_INOPB(fs); ++i) { 1127 union lfs_dinode *tmpdi; 1128 1129 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1130 DINOSIZE(fs) * i); 1131 lfs_dino_setinumber(fs, tmpdi, 0); 1132 } 1133 1134 ++sp->start_bpp; 1135 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1136 /* Set remaining space counters. */ 1137 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1138 sp->sum_bytes_left -= sizeof(int32_t); 1139 ndx = lfs_sb_getsumsize(fs) / sizeof(int32_t) - 1140 sp->ninodes / LFS_INOPB(fs) - 1; 1141 /* XXX ondisk32 */ 1142 ((int32_t *)(sp->segsum))[ndx] = daddr; 1143 } 1144 1145 /* Check VU_DIROP in case there is a new file with no data blocks */ 1146 if (vp->v_uflag & VU_DIROP) { 1147 ssp = (SEGSUM *)sp->segsum; 1148 lfs_ss_setflags(fs, ssp, 1149 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1150 } 1151 1152 /* Update the inode times and copy the inode onto the inode page. */ 1153 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1154 if (ip->i_number != LFS_IFILE_INUM) 1155 LFS_ITIMES(ip, NULL, NULL, NULL); 1156 1157 /* 1158 * If this is the Ifile, and we've already written the Ifile in this 1159 * partial segment, just overwrite it (it's not on disk yet) and 1160 * continue. 1161 * 1162 * XXX we know that the bp that we get the second time around has 1163 * already been gathered. 1164 */ 1165 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1166 if (fs->lfs_is64) { 1167 sp->idp->u_64 = *ip->i_din.ffs2_din; 1168 } else { 1169 sp->idp->u_32 = *ip->i_din.ffs1_din; 1170 } 1171 ip->i_lfs_osize = ip->i_size; 1172 return 0; 1173 } 1174 1175 bp = sp->ibp; 1176 cdp = (union lfs_dinode *)((char *)bp->b_data + DINOSIZE(fs) * (sp->ninodes % LFS_INOPB(fs))); 1177 if (fs->lfs_is64) { 1178 cdp->u_64 = *ip->i_din.ffs2_din; 1179 } else { 1180 cdp->u_32 = *ip->i_din.ffs1_din; 1181 } 1182 1183 /* 1184 * This inode is on its way to disk; clear its VU_DIROP status when 1185 * the write is complete. 1186 */ 1187 if (vp->v_uflag & VU_DIROP) { 1188 if (!(sp->seg_flags & SEGM_CLEAN)) 1189 ip->i_flag |= IN_CDIROP; 1190 else { 1191 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1192 } 1193 } 1194 1195 /* 1196 * If cleaning, link counts and directory file sizes cannot change, 1197 * since those would be directory operations---even if the file 1198 * we are writing is marked VU_DIROP we should write the old values. 1199 * If we're not cleaning, of course, update the values so we get 1200 * current values the next time we clean. 1201 */ 1202 if (sp->seg_flags & SEGM_CLEAN) { 1203 if (vp->v_uflag & VU_DIROP) { 1204 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1205 /* if (vp->v_type == VDIR) */ 1206 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1207 } 1208 } else { 1209 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1210 ip->i_lfs_osize = ip->i_size; 1211 } 1212 1213 1214 /* We can finish the segment accounting for truncations now */ 1215 lfs_finalize_ino_seguse(fs, ip); 1216 1217 /* 1218 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1219 * addresses to disk; possibly change the on-disk record of 1220 * the inode size, either by reverting to the previous size 1221 * (in the case of cleaning) or by verifying the inode's block 1222 * holdings (in the case of files being allocated as they are being 1223 * written). 1224 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1225 * XXX count on disk wrong by the same amount. We should be 1226 * XXX able to "borrow" from lfs_avail and return it after the 1227 * XXX Ifile is written. See also in lfs_writeseg. 1228 */ 1229 1230 /* Check file size based on highest allocated block */ 1231 if (((ip->i_ffs1_mode & LFS_IFMT) == LFS_IFREG || 1232 (ip->i_ffs1_mode & LFS_IFMT) == LFS_IFDIR) && 1233 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1234 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1235 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1236 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1237 } 1238 if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) { 1239 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1240 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1241 ip->i_ffs1_blocks, (uintmax_t)lfs_sb_getoffset(fs))); 1242 for (i=0; i<ULFS_NDADDR; i++) { 1243 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1244 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1245 lfs_dino_setdb(fs, cdp, i, 0); 1246 } 1247 } 1248 for (i=0; i<ULFS_NIADDR; i++) { 1249 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1250 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1251 lfs_dino_setib(fs, cdp, i, 0); 1252 } 1253 } 1254 } 1255 1256 #ifdef DIAGNOSTIC 1257 /* 1258 * Check dinode held blocks against dinode size. 1259 * This should be identical to the check in lfs_vget(). 1260 */ 1261 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1262 i < ULFS_NDADDR; i++) { 1263 KASSERT(i >= 0); 1264 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1265 continue; 1266 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1267 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1268 continue; 1269 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1270 # ifdef DEBUG 1271 lfs_dump_dinode(fs, cdp); 1272 # endif 1273 panic("writing inconsistent inode"); 1274 } 1275 } 1276 #endif /* DIAGNOSTIC */ 1277 1278 if (ip->i_flag & IN_CLEANING) 1279 LFS_CLR_UINO(ip, IN_CLEANING); 1280 else { 1281 /* XXX IN_ALLMOD */ 1282 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1283 IN_UPDATE | IN_MODIFY); 1284 if (ip->i_lfs_effnblks == ip->i_ffs1_blocks) 1285 LFS_CLR_UINO(ip, IN_MODIFIED); 1286 else { 1287 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1288 "blks=%d, eff=%jd\n", ip->i_number, 1289 ip->i_ffs1_blocks, (intmax_t)ip->i_lfs_effnblks)); 1290 } 1291 } 1292 1293 if (ip->i_number == LFS_IFILE_INUM) { 1294 /* We know sp->idp == NULL */ 1295 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1296 1297 /* Not dirty any more */ 1298 mutex_enter(&lfs_lock); 1299 fs->lfs_flags &= ~LFS_IFDIRTY; 1300 mutex_exit(&lfs_lock); 1301 } 1302 1303 if (gotblk) { 1304 mutex_enter(&bufcache_lock); 1305 LFS_LOCK_BUF(bp); 1306 brelsel(bp, 0); 1307 mutex_exit(&bufcache_lock); 1308 } 1309 1310 /* Increment inode count in segment summary block. */ 1311 1312 ssp = (SEGSUM *)sp->segsum; 1313 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1314 1315 /* If this page is full, set flag to allocate a new page. */ 1316 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1317 sp->ibp = NULL; 1318 1319 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1320 1321 KASSERT(redo_ifile == 0); 1322 return (redo_ifile); 1323 } 1324 1325 int 1326 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1327 { 1328 struct lfs *fs; 1329 int vers; 1330 int j, blksinblk; 1331 1332 ASSERT_SEGLOCK(sp->fs); 1333 /* 1334 * If full, finish this segment. We may be doing I/O, so 1335 * release and reacquire the splbio(). 1336 */ 1337 #ifdef DIAGNOSTIC 1338 if (sp->vp == NULL) 1339 panic ("lfs_gatherblock: Null vp in segment"); 1340 #endif 1341 fs = sp->fs; 1342 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1343 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1344 sp->seg_bytes_left < bp->b_bcount) { 1345 if (mptr) 1346 mutex_exit(mptr); 1347 lfs_updatemeta(sp); 1348 1349 vers = lfs_fi_getversion(fs, sp->fip); 1350 (void) lfs_writeseg(fs, sp); 1351 1352 /* Add the current file to the segment summary. */ 1353 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1354 1355 if (mptr) 1356 mutex_enter(mptr); 1357 return (1); 1358 } 1359 1360 if (bp->b_flags & B_GATHERED) { 1361 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1362 " lbn %" PRId64 "\n", 1363 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1364 return (0); 1365 } 1366 1367 /* Insert into the buffer list, update the FINFO block. */ 1368 bp->b_flags |= B_GATHERED; 1369 1370 *sp->cbpp++ = bp; 1371 for (j = 0; j < blksinblk; j++) { 1372 unsigned bn; 1373 1374 bn = lfs_fi_getnblocks(fs, sp->fip); 1375 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1376 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1377 /* This block's accounting moves from lfs_favail to lfs_avail */ 1378 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1379 } 1380 1381 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1382 sp->seg_bytes_left -= bp->b_bcount; 1383 return (0); 1384 } 1385 1386 int 1387 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1388 int (*match)(struct lfs *, struct buf *)) 1389 { 1390 struct buf *bp, *nbp; 1391 int count = 0; 1392 1393 ASSERT_SEGLOCK(fs); 1394 if (vp->v_type == VBLK) 1395 return 0; 1396 KASSERT(sp->vp == NULL); 1397 sp->vp = vp; 1398 mutex_enter(&bufcache_lock); 1399 1400 #ifndef LFS_NO_BACKBUF_HACK 1401 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1402 # define BUF_OFFSET \ 1403 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1404 # define BACK_BUF(BP) \ 1405 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1406 # define BEG_OF_LIST \ 1407 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1408 1409 loop: 1410 /* Find last buffer. */ 1411 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1412 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1413 bp = LIST_NEXT(bp, b_vnbufs)) 1414 /* nothing */; 1415 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1416 nbp = BACK_BUF(bp); 1417 #else /* LFS_NO_BACKBUF_HACK */ 1418 loop: 1419 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1420 nbp = LIST_NEXT(bp, b_vnbufs); 1421 #endif /* LFS_NO_BACKBUF_HACK */ 1422 if ((bp->b_cflags & BC_BUSY) != 0 || 1423 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1424 #ifdef DEBUG 1425 if (vp == fs->lfs_ivnode && 1426 (bp->b_cflags & BC_BUSY) != 0 && 1427 (bp->b_flags & B_GATHERED) == 0) 1428 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1429 PRId64 " busy (%x) at 0x%jx", 1430 bp->b_lblkno, bp->b_flags, 1431 (uintmax_t)lfs_sb_getoffset(fs)); 1432 #endif 1433 continue; 1434 } 1435 #ifdef DIAGNOSTIC 1436 # ifdef LFS_USE_B_INVAL 1437 if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1438 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1439 " is BC_INVAL\n", bp->b_lblkno)); 1440 VOP_PRINT(bp->b_vp); 1441 } 1442 # endif /* LFS_USE_B_INVAL */ 1443 if (!(bp->b_oflags & BO_DELWRI)) 1444 panic("lfs_gather: bp not BO_DELWRI"); 1445 if (!(bp->b_flags & B_LOCKED)) { 1446 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1447 " blk %" PRId64 " not B_LOCKED\n", 1448 bp->b_lblkno, 1449 LFS_DBTOFSB(fs, bp->b_blkno))); 1450 VOP_PRINT(bp->b_vp); 1451 panic("lfs_gather: bp not B_LOCKED"); 1452 } 1453 #endif 1454 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1455 goto loop; 1456 } 1457 count++; 1458 } 1459 mutex_exit(&bufcache_lock); 1460 lfs_updatemeta(sp); 1461 KASSERT(sp->vp == vp); 1462 sp->vp = NULL; 1463 return count; 1464 } 1465 1466 #if DEBUG 1467 # define DEBUG_OOFF(n) do { \ 1468 if (ooff == 0) { \ 1469 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1470 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1471 ", was 0x0 (or %" PRId64 ")\n", \ 1472 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1473 } \ 1474 } while (0) 1475 #else 1476 # define DEBUG_OOFF(n) 1477 #endif 1478 1479 /* 1480 * Change the given block's address to ndaddr, finding its previous 1481 * location using ulfs_bmaparray(). 1482 * 1483 * Account for this change in the segment table. 1484 * 1485 * called with sp == NULL by roll-forwarding code. 1486 */ 1487 void 1488 lfs_update_single(struct lfs *fs, struct segment *sp, 1489 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1490 { 1491 SEGUSE *sup; 1492 struct buf *bp; 1493 struct indir a[ULFS_NIADDR + 2], *ap; 1494 struct inode *ip; 1495 daddr_t daddr, ooff; 1496 int num, error; 1497 int bb, osize, obb; 1498 1499 ASSERT_SEGLOCK(fs); 1500 KASSERT(sp == NULL || sp->vp == vp); 1501 ip = VTOI(vp); 1502 1503 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1504 if (error) 1505 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1506 1507 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1508 if (daddr > 0) 1509 daddr = LFS_DBTOFSB(fs, daddr); 1510 1511 bb = lfs_numfrags(fs, size); 1512 switch (num) { 1513 case 0: 1514 ooff = ip->i_ffs1_db[lbn]; 1515 DEBUG_OOFF(0); 1516 if (ooff == UNWRITTEN) 1517 ip->i_ffs1_blocks += bb; 1518 else { 1519 /* possible fragment truncation or extension */ 1520 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1521 ip->i_ffs1_blocks += (bb - obb); 1522 } 1523 ip->i_ffs1_db[lbn] = ndaddr; 1524 break; 1525 case 1: 1526 ooff = ip->i_ffs1_ib[a[0].in_off]; 1527 DEBUG_OOFF(1); 1528 if (ooff == UNWRITTEN) 1529 ip->i_ffs1_blocks += bb; 1530 ip->i_ffs1_ib[a[0].in_off] = ndaddr; 1531 break; 1532 default: 1533 ap = &a[num - 1]; 1534 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1535 B_MODIFY, &bp)) 1536 panic("lfs_updatemeta: bread bno %" PRId64, 1537 ap->in_lbn); 1538 1539 /* XXX ondisk32 */ 1540 ooff = ((int32_t *)bp->b_data)[ap->in_off]; 1541 DEBUG_OOFF(num); 1542 if (ooff == UNWRITTEN) 1543 ip->i_ffs1_blocks += bb; 1544 /* XXX ondisk32 */ 1545 ((int32_t *)bp->b_data)[ap->in_off] = ndaddr; 1546 (void) VOP_BWRITE(bp->b_vp, bp); 1547 } 1548 1549 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1550 1551 /* Update hiblk when extending the file */ 1552 if (lbn > ip->i_lfs_hiblk) 1553 ip->i_lfs_hiblk = lbn; 1554 1555 /* 1556 * Though we'd rather it couldn't, this *can* happen right now 1557 * if cleaning blocks and regular blocks coexist. 1558 */ 1559 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1560 1561 /* 1562 * Update segment usage information, based on old size 1563 * and location. 1564 */ 1565 if (daddr > 0) { 1566 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1567 #ifdef DIAGNOSTIC 1568 int ndupino; 1569 1570 if (sp && sp->seg_number == oldsn) { 1571 ndupino = sp->ndupino; 1572 } else { 1573 ndupino = 0; 1574 } 1575 #endif 1576 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1577 if (lbn >= 0 && lbn < ULFS_NDADDR) 1578 osize = ip->i_lfs_fragsize[lbn]; 1579 else 1580 osize = lfs_sb_getbsize(fs); 1581 LFS_SEGENTRY(sup, fs, oldsn, bp); 1582 #ifdef DIAGNOSTIC 1583 if (sup->su_nbytes + DINOSIZE(fs) * ndupino < osize) { 1584 printf("lfs_updatemeta: negative bytes " 1585 "(segment %" PRIu32 " short by %" PRId64 1586 ")\n", lfs_dtosn(fs, daddr), 1587 (int64_t)osize - 1588 (DINOSIZE(fs) * ndupino + sup->su_nbytes)); 1589 printf("lfs_updatemeta: ino %llu, lbn %" PRId64 1590 ", addr = 0x%" PRIx64 "\n", 1591 (unsigned long long)ip->i_number, lbn, daddr); 1592 printf("lfs_updatemeta: ndupino=%d\n", ndupino); 1593 panic("lfs_updatemeta: negative bytes"); 1594 sup->su_nbytes = osize - 1595 DINOSIZE(fs) * ndupino; 1596 } 1597 #endif 1598 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1599 " db 0x%" PRIx64 "\n", 1600 lfs_dtosn(fs, daddr), osize, 1601 ip->i_number, lbn, daddr)); 1602 sup->su_nbytes -= osize; 1603 if (!(bp->b_flags & B_GATHERED)) { 1604 mutex_enter(&lfs_lock); 1605 fs->lfs_flags |= LFS_IFDIRTY; 1606 mutex_exit(&lfs_lock); 1607 } 1608 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1609 } 1610 /* 1611 * Now that this block has a new address, and its old 1612 * segment no longer owns it, we can forget about its 1613 * old size. 1614 */ 1615 if (lbn >= 0 && lbn < ULFS_NDADDR) 1616 ip->i_lfs_fragsize[lbn] = size; 1617 } 1618 1619 /* 1620 * Update the metadata that points to the blocks listed in the FINFO 1621 * array. 1622 */ 1623 void 1624 lfs_updatemeta(struct segment *sp) 1625 { 1626 struct buf *sbp; 1627 struct lfs *fs; 1628 struct vnode *vp; 1629 daddr_t lbn; 1630 int i, nblocks, num; 1631 int __diagused nblocks_orig; 1632 int bb; 1633 int bytesleft, size; 1634 unsigned lastlength; 1635 union lfs_blocks tmpptr; 1636 1637 fs = sp->fs; 1638 vp = sp->vp; 1639 ASSERT_SEGLOCK(fs); 1640 1641 /* 1642 * This used to be: 1643 * 1644 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1645 * 1646 * that is, it allowed for the possibility that start_lbp did 1647 * not point to the beginning of the finfo block pointer area. 1648 * This particular formulation is six kinds of painful in the 1649 * lfs64 world where we have two sizes of block pointer, so 1650 * unless/until everything can be cleaned up to not move 1651 * start_lbp around but instead use an offset, we do the 1652 * following: 1653 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1654 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1655 * type. (Ugh.) 1656 * 2. Cast it to void *, then assign it to a temporary 1657 * union lfs_blocks. 1658 * 3. Subtract start_lbp from that. 1659 * 4. Save the value of nblocks in blocks_orig so we can 1660 * assert below that it hasn't changed without repeating this 1661 * rubbish. 1662 * 1663 * XXX. 1664 */ 1665 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1666 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1667 nblocks_orig = nblocks; 1668 1669 KASSERT(nblocks >= 0); 1670 KASSERT(vp != NULL); 1671 if (nblocks == 0) 1672 return; 1673 1674 /* 1675 * This count may be high due to oversize blocks from lfs_gop_write. 1676 * Correct for this. (XXX we should be able to keep track of these.) 1677 */ 1678 for (i = 0; i < nblocks; i++) { 1679 if (sp->start_bpp[i] == NULL) { 1680 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1681 nblocks = i; 1682 break; 1683 } 1684 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1685 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1686 nblocks -= num - 1; 1687 } 1688 1689 #if 0 1690 /* pre-lfs64 assertion */ 1691 KASSERT(vp->v_type == VREG || 1692 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1693 #else 1694 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1695 #endif 1696 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1697 1698 /* 1699 * Sort the blocks. 1700 * 1701 * We have to sort even if the blocks come from the 1702 * cleaner, because there might be other pending blocks on the 1703 * same inode...and if we don't sort, and there are fragments 1704 * present, blocks may be written in the wrong place. 1705 */ 1706 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1707 1708 /* 1709 * Record the length of the last block in case it's a fragment. 1710 * If there are indirect blocks present, they sort last. An 1711 * indirect block will be lfs_bsize and its presence indicates 1712 * that you cannot have fragments. 1713 * 1714 * XXX This last is a lie. A cleaned fragment can coexist with 1715 * XXX a later indirect block. This will continue to be 1716 * XXX true until lfs_markv is fixed to do everything with 1717 * XXX fake blocks (including fake inodes and fake indirect blocks). 1718 */ 1719 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1720 lfs_sb_getbmask(fs)) + 1; 1721 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1722 1723 /* 1724 * Assign disk addresses, and update references to the logical 1725 * block and the segment usage information. 1726 */ 1727 for (i = nblocks; i--; ++sp->start_bpp) { 1728 sbp = *sp->start_bpp; 1729 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1730 KASSERT(sbp->b_lblkno == lbn); 1731 1732 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1733 1734 /* 1735 * If we write a frag in the wrong place, the cleaner won't 1736 * be able to correctly identify its size later, and the 1737 * segment will be uncleanable. (Even worse, it will assume 1738 * that the indirect block that actually ends the list 1739 * is of a smaller size!) 1740 */ 1741 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1742 panic("lfs_updatemeta: fragment is not last block"); 1743 1744 /* 1745 * For each subblock in this possibly oversized block, 1746 * update its address on disk. 1747 */ 1748 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1749 KASSERT(vp == sbp->b_vp); 1750 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1751 bytesleft -= lfs_sb_getbsize(fs)) { 1752 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1753 bb = lfs_numfrags(fs, size); 1754 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1755 lfs_blocks_inc(fs, &sp->start_lbp); 1756 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1757 size); 1758 lfs_sb_addoffset(fs, bb); 1759 } 1760 1761 } 1762 1763 /* This inode has been modified */ 1764 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1765 } 1766 1767 /* 1768 * Move lfs_offset to a segment earlier than newsn. 1769 */ 1770 int 1771 lfs_rewind(struct lfs *fs, int newsn) 1772 { 1773 int sn, osn, isdirty; 1774 struct buf *bp; 1775 SEGUSE *sup; 1776 1777 ASSERT_SEGLOCK(fs); 1778 1779 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1780 if (osn < newsn) 1781 return 0; 1782 1783 /* lfs_avail eats the remaining space in this segment */ 1784 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1785 1786 /* Find a low-numbered segment */ 1787 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1788 LFS_SEGENTRY(sup, fs, sn, bp); 1789 isdirty = sup->su_flags & SEGUSE_DIRTY; 1790 brelse(bp, 0); 1791 1792 if (!isdirty) 1793 break; 1794 } 1795 if (sn == lfs_sb_getnseg(fs)) 1796 panic("lfs_rewind: no clean segments"); 1797 if (newsn >= 0 && sn >= newsn) 1798 return ENOENT; 1799 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1800 lfs_newseg(fs); 1801 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1802 1803 return 0; 1804 } 1805 1806 /* 1807 * Start a new partial segment. 1808 * 1809 * Return 1 when we entered to a new segment. 1810 * Otherwise, return 0. 1811 */ 1812 int 1813 lfs_initseg(struct lfs *fs) 1814 { 1815 struct segment *sp = fs->lfs_sp; 1816 SEGSUM *ssp; 1817 struct buf *sbp; /* buffer for SEGSUM */ 1818 int repeat = 0; /* return value */ 1819 1820 ASSERT_SEGLOCK(fs); 1821 /* Advance to the next segment. */ 1822 if (!LFS_PARTIAL_FITS(fs)) { 1823 SEGUSE *sup; 1824 struct buf *bp; 1825 1826 /* lfs_avail eats the remaining space */ 1827 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1828 lfs_sb_getcurseg(fs))); 1829 /* Wake up any cleaning procs waiting on this file system. */ 1830 lfs_wakeup_cleaner(fs); 1831 lfs_newseg(fs); 1832 repeat = 1; 1833 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1834 1835 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1836 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1837 1838 /* 1839 * If the segment contains a superblock, update the offset 1840 * and summary address to skip over it. 1841 */ 1842 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1843 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1844 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1845 sp->seg_bytes_left -= LFS_SBPAD; 1846 } 1847 brelse(bp, 0); 1848 /* Segment zero could also contain the labelpad */ 1849 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1850 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1851 lfs_sb_addoffset(fs, 1852 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1853 sp->seg_bytes_left -= 1854 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1855 } 1856 } else { 1857 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1858 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1859 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1860 } 1861 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1862 1863 /* Record first address of this partial segment */ 1864 if (sp->seg_flags & SEGM_CLEAN) { 1865 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1866 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1867 /* "1" is the artificial inc in lfs_seglock */ 1868 mutex_enter(&lfs_lock); 1869 while (fs->lfs_iocount > 1) { 1870 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1871 "lfs_initseg", 0, &lfs_lock); 1872 } 1873 mutex_exit(&lfs_lock); 1874 fs->lfs_cleanind = 0; 1875 } 1876 } 1877 1878 sp->fs = fs; 1879 sp->ibp = NULL; 1880 sp->idp = NULL; 1881 sp->ninodes = 0; 1882 sp->ndupino = 0; 1883 1884 sp->cbpp = sp->bpp; 1885 1886 /* Get a new buffer for SEGSUM */ 1887 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1888 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1889 1890 /* ... and enter it into the buffer list. */ 1891 *sp->cbpp = sbp; 1892 sp->cbpp++; 1893 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1894 1895 sp->start_bpp = sp->cbpp; 1896 1897 /* Set point to SEGSUM, initialize it. */ 1898 ssp = sp->segsum = sbp->b_data; 1899 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1900 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1901 lfs_ss_setnfinfo(fs, ssp, 0); 1902 lfs_ss_setninos(fs, ssp, 0); 1903 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1904 1905 /* Set pointer to first FINFO, initialize it. */ 1906 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1907 lfs_fi_setnblocks(fs, sp->fip, 0); 1908 lfs_fi_setlastlength(fs, sp->fip, 0); 1909 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1910 1911 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1912 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1913 1914 return (repeat); 1915 } 1916 1917 /* 1918 * Remove SEGUSE_INVAL from all segments. 1919 */ 1920 void 1921 lfs_unset_inval_all(struct lfs *fs) 1922 { 1923 SEGUSE *sup; 1924 struct buf *bp; 1925 int i; 1926 1927 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1928 LFS_SEGENTRY(sup, fs, i, bp); 1929 if (sup->su_flags & SEGUSE_INVAL) { 1930 sup->su_flags &= ~SEGUSE_INVAL; 1931 LFS_WRITESEGENTRY(sup, fs, i, bp); 1932 } else 1933 brelse(bp, 0); 1934 } 1935 } 1936 1937 /* 1938 * Return the next segment to write. 1939 */ 1940 void 1941 lfs_newseg(struct lfs *fs) 1942 { 1943 CLEANERINFO *cip; 1944 SEGUSE *sup; 1945 struct buf *bp; 1946 int curseg, isdirty, sn, skip_inval; 1947 1948 ASSERT_SEGLOCK(fs); 1949 1950 /* Honor LFCNWRAPSTOP */ 1951 mutex_enter(&lfs_lock); 1952 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1953 if (fs->lfs_wrappass) { 1954 log(LOG_NOTICE, "%s: wrappass=%d\n", 1955 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1956 fs->lfs_wrappass = 0; 1957 break; 1958 } 1959 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1960 wakeup(&fs->lfs_nowrap); 1961 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1962 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1963 &lfs_lock); 1964 } 1965 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1966 mutex_exit(&lfs_lock); 1967 1968 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1969 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1970 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1971 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1972 sup->su_nbytes = 0; 1973 sup->su_nsums = 0; 1974 sup->su_ninos = 0; 1975 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1976 1977 LFS_CLEANERINFO(cip, fs, bp); 1978 lfs_ci_shiftcleantodirty(fs, cip, 1); 1979 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1980 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1981 1982 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1983 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1984 skip_inval = 1; 1985 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1986 sn = (sn + 1) % lfs_sb_getnseg(fs); 1987 1988 if (sn == curseg) { 1989 if (skip_inval) 1990 skip_inval = 0; 1991 else 1992 panic("lfs_nextseg: no clean segments"); 1993 } 1994 LFS_SEGENTRY(sup, fs, sn, bp); 1995 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1996 /* Check SEGUSE_EMPTY as we go along */ 1997 if (isdirty && sup->su_nbytes == 0 && 1998 !(sup->su_flags & SEGUSE_EMPTY)) 1999 LFS_WRITESEGENTRY(sup, fs, sn, bp); 2000 else 2001 brelse(bp, 0); 2002 2003 if (!isdirty) 2004 break; 2005 } 2006 if (skip_inval == 0) 2007 lfs_unset_inval_all(fs); 2008 2009 ++fs->lfs_nactive; 2010 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 2011 if (lfs_dostats) { 2012 ++lfs_stats.segsused; 2013 } 2014 } 2015 2016 static struct buf * 2017 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 2018 int n) 2019 { 2020 struct lfs_cluster *cl; 2021 struct buf **bpp, *bp; 2022 2023 ASSERT_SEGLOCK(fs); 2024 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 2025 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 2026 memset(cl, 0, sizeof(*cl)); 2027 cl->fs = fs; 2028 cl->bpp = bpp; 2029 cl->bufcount = 0; 2030 cl->bufsize = 0; 2031 2032 /* If this segment is being written synchronously, note that */ 2033 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 2034 cl->flags |= LFS_CL_SYNC; 2035 cl->seg = fs->lfs_sp; 2036 ++cl->seg->seg_iocount; 2037 } 2038 2039 /* Get an empty buffer header, or maybe one with something on it */ 2040 bp = getiobuf(vp, true); 2041 bp->b_dev = NODEV; 2042 bp->b_blkno = bp->b_lblkno = addr; 2043 bp->b_iodone = lfs_cluster_callback; 2044 bp->b_private = cl; 2045 2046 return bp; 2047 } 2048 2049 int 2050 lfs_writeseg(struct lfs *fs, struct segment *sp) 2051 { 2052 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2053 SEGUSE *sup; 2054 SEGSUM *ssp; 2055 int i; 2056 int do_again, nblocks, byteoffset; 2057 size_t el_size; 2058 struct lfs_cluster *cl; 2059 u_short ninos; 2060 struct vnode *devvp; 2061 char *p = NULL; 2062 struct vnode *vp; 2063 int32_t *daddrp; /* XXX ondisk32 */ 2064 int changed; 2065 u_int32_t sum; 2066 size_t sumstart; 2067 #ifdef DEBUG 2068 FINFO *fip; 2069 int findex; 2070 #endif 2071 2072 ASSERT_SEGLOCK(fs); 2073 2074 ssp = (SEGSUM *)sp->segsum; 2075 2076 /* 2077 * If there are no buffers other than the segment summary to write, 2078 * don't do anything. If we are the end of a dirop sequence, however, 2079 * write the empty segment summary anyway, to help out the 2080 * roll-forward agent. 2081 */ 2082 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2083 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2084 return 0; 2085 } 2086 2087 /* Note if partial segment is being written by the cleaner */ 2088 if (sp->seg_flags & SEGM_CLEAN) 2089 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2090 2091 /* Note if we are writing to reclaim */ 2092 if (sp->seg_flags & SEGM_RECLAIM) { 2093 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2094 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2095 } 2096 2097 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2098 2099 /* Update the segment usage information. */ 2100 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2101 2102 /* Loop through all blocks, except the segment summary. */ 2103 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2104 if ((*bpp)->b_vp != devvp) { 2105 sup->su_nbytes += (*bpp)->b_bcount; 2106 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2107 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2108 sp->seg_number, (*bpp)->b_bcount, 2109 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2110 (*bpp)->b_blkno)); 2111 } 2112 } 2113 2114 #ifdef DEBUG 2115 /* Check for zero-length and zero-version FINFO entries. */ 2116 fip = SEGSUM_FINFOBASE(fs, ssp); 2117 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2118 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2119 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2120 fip = NEXT_FINFO(fs, fip); 2121 } 2122 #endif /* DEBUG */ 2123 2124 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2125 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2126 sp->seg_number, 2127 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2128 lfs_ss_getninos(fs, ssp))); 2129 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2130 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2131 if (lfs_sb_getversion(fs) == 1) 2132 sup->su_olastmod = time_second; 2133 else 2134 sup->su_lastmod = time_second; 2135 sup->su_ninos += ninos; 2136 ++sup->su_nsums; 2137 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2138 2139 do_again = !(bp->b_flags & B_GATHERED); 2140 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2141 2142 /* 2143 * Mark blocks B_BUSY, to prevent then from being changed between 2144 * the checksum computation and the actual write. 2145 * 2146 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2147 * there are any, replace them with copies that have UNASSIGNED 2148 * instead. 2149 */ 2150 mutex_enter(&bufcache_lock); 2151 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2152 ++bpp; 2153 bp = *bpp; 2154 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2155 bp->b_cflags |= BC_BUSY; 2156 continue; 2157 } 2158 2159 while (bp->b_cflags & BC_BUSY) { 2160 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2161 " data summary corruption for ino %d, lbn %" 2162 PRId64 "\n", 2163 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2164 bp->b_cflags |= BC_WANTED; 2165 cv_wait(&bp->b_busy, &bufcache_lock); 2166 } 2167 bp->b_cflags |= BC_BUSY; 2168 mutex_exit(&bufcache_lock); 2169 unbusybp = NULL; 2170 2171 /* 2172 * Check and replace indirect block UNWRITTEN bogosity. 2173 * XXX See comment in lfs_writefile. 2174 */ 2175 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2176 VTOI(bp->b_vp)->i_ffs1_blocks != 2177 VTOI(bp->b_vp)->i_lfs_effnblks) { 2178 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2179 VTOI(bp->b_vp)->i_number, 2180 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2181 VTOI(bp->b_vp)->i_ffs1_blocks)); 2182 /* Make a copy we'll make changes to */ 2183 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2184 bp->b_bcount, LFS_NB_IBLOCK); 2185 newbp->b_blkno = bp->b_blkno; 2186 memcpy(newbp->b_data, bp->b_data, 2187 newbp->b_bcount); 2188 2189 changed = 0; 2190 /* XXX ondisk32 */ 2191 for (daddrp = (int32_t *)(newbp->b_data); 2192 daddrp < (int32_t *)((char *)newbp->b_data + 2193 newbp->b_bcount); daddrp++) { 2194 if (*daddrp == UNWRITTEN) { 2195 ++changed; 2196 *daddrp = 0; 2197 } 2198 } 2199 /* 2200 * Get rid of the old buffer. Don't mark it clean, 2201 * though, if it still has dirty data on it. 2202 */ 2203 if (changed) { 2204 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2205 " bp = %p newbp = %p\n", changed, bp, 2206 newbp)); 2207 *bpp = newbp; 2208 bp->b_flags &= ~B_GATHERED; 2209 bp->b_error = 0; 2210 if (bp->b_iodone != NULL) { 2211 DLOG((DLOG_SEG, "lfs_writeseg: " 2212 "indir bp should not be B_CALL\n")); 2213 biodone(bp); 2214 bp = NULL; 2215 } else { 2216 /* Still on free list, leave it there */ 2217 unbusybp = bp; 2218 /* 2219 * We have to re-decrement lfs_avail 2220 * since this block is going to come 2221 * back around to us in the next 2222 * segment. 2223 */ 2224 lfs_sb_subavail(fs, 2225 lfs_btofsb(fs, bp->b_bcount)); 2226 } 2227 } else { 2228 lfs_freebuf(fs, newbp); 2229 } 2230 } 2231 mutex_enter(&bufcache_lock); 2232 if (unbusybp != NULL) { 2233 unbusybp->b_cflags &= ~BC_BUSY; 2234 if (unbusybp->b_cflags & BC_WANTED) 2235 cv_broadcast(&bp->b_busy); 2236 } 2237 } 2238 mutex_exit(&bufcache_lock); 2239 2240 /* 2241 * Compute checksum across data and then across summary; the first 2242 * block (the summary block) is skipped. Set the create time here 2243 * so that it's guaranteed to be later than the inode mod times. 2244 */ 2245 sum = 0; 2246 if (lfs_sb_getversion(fs) == 1) 2247 el_size = sizeof(u_long); 2248 else 2249 el_size = sizeof(u_int32_t); 2250 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2251 ++bpp; 2252 /* Loop through gop_write cluster blocks */ 2253 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2254 byteoffset += lfs_sb_getbsize(fs)) { 2255 #ifdef LFS_USE_B_INVAL 2256 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2257 (*bpp)->b_iodone != NULL) { 2258 if (copyin((void *)(*bpp)->b_saveaddr + 2259 byteoffset, dp, el_size)) { 2260 panic("lfs_writeseg: copyin failed [1]:" 2261 " ino %d blk %" PRId64, 2262 VTOI((*bpp)->b_vp)->i_number, 2263 (*bpp)->b_lblkno); 2264 } 2265 } else 2266 #endif /* LFS_USE_B_INVAL */ 2267 { 2268 sum = lfs_cksum_part((char *) 2269 (*bpp)->b_data + byteoffset, el_size, sum); 2270 } 2271 } 2272 } 2273 if (lfs_sb_getversion(fs) == 1) 2274 lfs_ss_setocreate(fs, ssp, time_second); 2275 else { 2276 lfs_ss_setcreate(fs, ssp, time_second); 2277 lfs_sb_addserial(fs, 1); 2278 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2279 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2280 } 2281 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2282 sumstart = lfs_ss_getsumstart(fs); 2283 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2284 lfs_sb_getsumsize(fs) - sumstart)); 2285 2286 mutex_enter(&lfs_lock); 2287 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2288 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2289 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2290 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2291 mutex_exit(&lfs_lock); 2292 2293 /* 2294 * When we simply write the blocks we lose a rotation for every block 2295 * written. To avoid this problem, we cluster the buffers into a 2296 * chunk and write the chunk. MAXPHYS is the largest size I/O 2297 * devices can handle, use that for the size of the chunks. 2298 * 2299 * Blocks that are already clusters (from GOP_WRITE), however, we 2300 * don't bother to copy into other clusters. 2301 */ 2302 2303 #define CHUNKSIZE MAXPHYS 2304 2305 if (devvp == NULL) 2306 panic("devvp is NULL"); 2307 for (bpp = sp->bpp, i = nblocks; i;) { 2308 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2309 cl = cbp->b_private; 2310 2311 cbp->b_flags |= B_ASYNC; 2312 cbp->b_cflags |= BC_BUSY; 2313 cbp->b_bcount = 0; 2314 2315 #if defined(DEBUG) && defined(DIAGNOSTIC) 2316 if (bpp - sp->bpp > (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2317 / sizeof(int32_t)) { 2318 panic("lfs_writeseg: real bpp overwrite"); 2319 } 2320 if (bpp - sp->bpp > lfs_segsize(fs) / lfs_sb_getfsize(fs)) { 2321 panic("lfs_writeseg: theoretical bpp overwrite"); 2322 } 2323 #endif 2324 2325 /* 2326 * Construct the cluster. 2327 */ 2328 mutex_enter(&lfs_lock); 2329 ++fs->lfs_iocount; 2330 mutex_exit(&lfs_lock); 2331 while (i && cbp->b_bcount < CHUNKSIZE) { 2332 bp = *bpp; 2333 2334 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2335 break; 2336 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2337 break; 2338 2339 /* Clusters from GOP_WRITE are expedited */ 2340 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2341 if (cbp->b_bcount > 0) 2342 /* Put in its own buffer */ 2343 break; 2344 else { 2345 cbp->b_data = bp->b_data; 2346 } 2347 } else if (cbp->b_bcount == 0) { 2348 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2349 LFS_NB_CLUSTER); 2350 cl->flags |= LFS_CL_MALLOC; 2351 } 2352 #ifdef DIAGNOSTIC 2353 if (lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2354 btodb(bp->b_bcount - 1))) != 2355 sp->seg_number) { 2356 printf("blk size %d daddr %" PRIx64 2357 " not in seg %d\n", 2358 bp->b_bcount, bp->b_blkno, 2359 sp->seg_number); 2360 panic("segment overwrite"); 2361 } 2362 #endif 2363 2364 #ifdef LFS_USE_B_INVAL 2365 /* 2366 * Fake buffers from the cleaner are marked as B_INVAL. 2367 * We need to copy the data from user space rather than 2368 * from the buffer indicated. 2369 * XXX == what do I do on an error? 2370 */ 2371 if ((bp->b_cflags & BC_INVAL) != 0 && 2372 bp->b_iodone != NULL) { 2373 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2374 panic("lfs_writeseg: " 2375 "copyin failed [2]"); 2376 } else 2377 #endif /* LFS_USE_B_INVAL */ 2378 if (cl->flags & LFS_CL_MALLOC) { 2379 /* copy data into our cluster. */ 2380 memcpy(p, bp->b_data, bp->b_bcount); 2381 p += bp->b_bcount; 2382 } 2383 2384 cbp->b_bcount += bp->b_bcount; 2385 cl->bufsize += bp->b_bcount; 2386 2387 bp->b_flags &= ~B_READ; 2388 bp->b_error = 0; 2389 cl->bpp[cl->bufcount++] = bp; 2390 2391 vp = bp->b_vp; 2392 mutex_enter(&bufcache_lock); 2393 mutex_enter(vp->v_interlock); 2394 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2395 reassignbuf(bp, vp); 2396 vp->v_numoutput++; 2397 mutex_exit(vp->v_interlock); 2398 mutex_exit(&bufcache_lock); 2399 2400 bpp++; 2401 i--; 2402 } 2403 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2404 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2405 else 2406 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2407 mutex_enter(devvp->v_interlock); 2408 devvp->v_numoutput++; 2409 mutex_exit(devvp->v_interlock); 2410 VOP_STRATEGY(devvp, cbp); 2411 curlwp->l_ru.ru_oublock++; 2412 } 2413 2414 if (lfs_dostats) { 2415 ++lfs_stats.psegwrites; 2416 lfs_stats.blocktot += nblocks - 1; 2417 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2418 ++lfs_stats.psyncwrites; 2419 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2420 ++lfs_stats.pcleanwrites; 2421 lfs_stats.cleanblocks += nblocks - 1; 2422 } 2423 } 2424 2425 return (lfs_initseg(fs) || do_again); 2426 } 2427 2428 void 2429 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2430 { 2431 struct buf *bp; 2432 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2433 int s; 2434 2435 ASSERT_MAYBE_SEGLOCK(fs); 2436 #ifdef DIAGNOSTIC 2437 if (fs->lfs_is64) { 2438 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2439 } else { 2440 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2441 } 2442 #endif 2443 /* 2444 * If we can write one superblock while another is in 2445 * progress, we risk not having a complete checkpoint if we crash. 2446 * So, block here if a superblock write is in progress. 2447 */ 2448 mutex_enter(&lfs_lock); 2449 s = splbio(); 2450 while (fs->lfs_sbactive) { 2451 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2452 &lfs_lock); 2453 } 2454 fs->lfs_sbactive = daddr; 2455 splx(s); 2456 mutex_exit(&lfs_lock); 2457 2458 /* Set timestamp of this version of the superblock */ 2459 if (lfs_sb_getversion(fs) == 1) 2460 lfs_sb_setotstamp(fs, time_second); 2461 lfs_sb_settstamp(fs, time_second); 2462 2463 /* The next chunk of code relies on this assumption */ 2464 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2465 2466 /* Checksum the superblock and copy it into a buffer. */ 2467 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2468 bp = lfs_newbuf(fs, devvp, 2469 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2470 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2471 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2472 LFS_SBPAD - sizeof(struct dlfs)); 2473 2474 bp->b_cflags |= BC_BUSY; 2475 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2476 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2477 bp->b_error = 0; 2478 bp->b_iodone = lfs_supercallback; 2479 2480 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2481 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2482 else 2483 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2484 curlwp->l_ru.ru_oublock++; 2485 2486 mutex_enter(devvp->v_interlock); 2487 devvp->v_numoutput++; 2488 mutex_exit(devvp->v_interlock); 2489 2490 mutex_enter(&lfs_lock); 2491 ++fs->lfs_iocount; 2492 mutex_exit(&lfs_lock); 2493 VOP_STRATEGY(devvp, bp); 2494 } 2495 2496 /* 2497 * Logical block number match routines used when traversing the dirty block 2498 * chain. 2499 */ 2500 int 2501 lfs_match_fake(struct lfs *fs, struct buf *bp) 2502 { 2503 2504 ASSERT_SEGLOCK(fs); 2505 return LFS_IS_MALLOC_BUF(bp); 2506 } 2507 2508 #if 0 2509 int 2510 lfs_match_real(struct lfs *fs, struct buf *bp) 2511 { 2512 2513 ASSERT_SEGLOCK(fs); 2514 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2515 } 2516 #endif 2517 2518 int 2519 lfs_match_data(struct lfs *fs, struct buf *bp) 2520 { 2521 2522 ASSERT_SEGLOCK(fs); 2523 return (bp->b_lblkno >= 0); 2524 } 2525 2526 int 2527 lfs_match_indir(struct lfs *fs, struct buf *bp) 2528 { 2529 daddr_t lbn; 2530 2531 ASSERT_SEGLOCK(fs); 2532 lbn = bp->b_lblkno; 2533 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2534 } 2535 2536 int 2537 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2538 { 2539 daddr_t lbn; 2540 2541 ASSERT_SEGLOCK(fs); 2542 lbn = bp->b_lblkno; 2543 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2544 } 2545 2546 int 2547 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2548 { 2549 daddr_t lbn; 2550 2551 ASSERT_SEGLOCK(fs); 2552 lbn = bp->b_lblkno; 2553 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2554 } 2555 2556 static void 2557 lfs_free_aiodone(struct buf *bp) 2558 { 2559 struct lfs *fs; 2560 2561 KERNEL_LOCK(1, curlwp); 2562 fs = bp->b_private; 2563 ASSERT_NO_SEGLOCK(fs); 2564 lfs_freebuf(fs, bp); 2565 KERNEL_UNLOCK_LAST(curlwp); 2566 } 2567 2568 static void 2569 lfs_super_aiodone(struct buf *bp) 2570 { 2571 struct lfs *fs; 2572 2573 KERNEL_LOCK(1, curlwp); 2574 fs = bp->b_private; 2575 ASSERT_NO_SEGLOCK(fs); 2576 mutex_enter(&lfs_lock); 2577 fs->lfs_sbactive = 0; 2578 if (--fs->lfs_iocount <= 1) 2579 wakeup(&fs->lfs_iocount); 2580 wakeup(&fs->lfs_sbactive); 2581 mutex_exit(&lfs_lock); 2582 lfs_freebuf(fs, bp); 2583 KERNEL_UNLOCK_LAST(curlwp); 2584 } 2585 2586 static void 2587 lfs_cluster_aiodone(struct buf *bp) 2588 { 2589 struct lfs_cluster *cl; 2590 struct lfs *fs; 2591 struct buf *tbp, *fbp; 2592 struct vnode *vp, *devvp, *ovp; 2593 struct inode *ip; 2594 int error; 2595 2596 KERNEL_LOCK(1, curlwp); 2597 2598 error = bp->b_error; 2599 cl = bp->b_private; 2600 fs = cl->fs; 2601 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2602 ASSERT_NO_SEGLOCK(fs); 2603 2604 /* Put the pages back, and release the buffer */ 2605 while (cl->bufcount--) { 2606 tbp = cl->bpp[cl->bufcount]; 2607 KASSERT(tbp->b_cflags & BC_BUSY); 2608 if (error) { 2609 tbp->b_error = error; 2610 } 2611 2612 /* 2613 * We're done with tbp. If it has not been re-dirtied since 2614 * the cluster was written, free it. Otherwise, keep it on 2615 * the locked list to be written again. 2616 */ 2617 vp = tbp->b_vp; 2618 2619 tbp->b_flags &= ~B_GATHERED; 2620 2621 LFS_BCLEAN_LOG(fs, tbp); 2622 2623 mutex_enter(&bufcache_lock); 2624 if (tbp->b_iodone == NULL) { 2625 KASSERT(tbp->b_flags & B_LOCKED); 2626 bremfree(tbp); 2627 if (vp) { 2628 mutex_enter(vp->v_interlock); 2629 reassignbuf(tbp, vp); 2630 mutex_exit(vp->v_interlock); 2631 } 2632 tbp->b_flags |= B_ASYNC; /* for biodone */ 2633 } 2634 2635 if (((tbp->b_flags | tbp->b_oflags) & 2636 (B_LOCKED | BO_DELWRI)) == B_LOCKED) 2637 LFS_UNLOCK_BUF(tbp); 2638 2639 if (tbp->b_oflags & BO_DONE) { 2640 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2641 cl->bufcount, (long)tbp->b_flags)); 2642 } 2643 2644 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2645 /* 2646 * A buffer from the page daemon. 2647 * We use the same iodone as it does, 2648 * so we must manually disassociate its 2649 * buffers from the vp. 2650 */ 2651 if ((ovp = tbp->b_vp) != NULL) { 2652 /* This is just silly */ 2653 mutex_enter(ovp->v_interlock); 2654 brelvp(tbp); 2655 mutex_exit(ovp->v_interlock); 2656 tbp->b_vp = vp; 2657 tbp->b_objlock = vp->v_interlock; 2658 } 2659 /* Put it back the way it was */ 2660 tbp->b_flags |= B_ASYNC; 2661 /* Master buffers have BC_AGE */ 2662 if (tbp->b_private == tbp) 2663 tbp->b_cflags |= BC_AGE; 2664 } 2665 mutex_exit(&bufcache_lock); 2666 2667 biodone(tbp); 2668 2669 /* 2670 * If this is the last block for this vnode, but 2671 * there are other blocks on its dirty list, 2672 * set IN_MODIFIED/IN_CLEANING depending on what 2673 * sort of block. Only do this for our mount point, 2674 * not for, e.g., inode blocks that are attached to 2675 * the devvp. 2676 * XXX KS - Shouldn't we set *both* if both types 2677 * of blocks are present (traverse the dirty list?) 2678 */ 2679 mutex_enter(vp->v_interlock); 2680 mutex_enter(&lfs_lock); 2681 if (vp != devvp && vp->v_numoutput == 0 && 2682 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2683 ip = VTOI(vp); 2684 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2685 ip->i_number)); 2686 if (LFS_IS_MALLOC_BUF(fbp)) 2687 LFS_SET_UINO(ip, IN_CLEANING); 2688 else 2689 LFS_SET_UINO(ip, IN_MODIFIED); 2690 } 2691 cv_broadcast(&vp->v_cv); 2692 mutex_exit(&lfs_lock); 2693 mutex_exit(vp->v_interlock); 2694 } 2695 2696 /* Fix up the cluster buffer, and release it */ 2697 if (cl->flags & LFS_CL_MALLOC) 2698 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2699 putiobuf(bp); 2700 2701 /* Note i/o done */ 2702 if (cl->flags & LFS_CL_SYNC) { 2703 if (--cl->seg->seg_iocount == 0) 2704 wakeup(&cl->seg->seg_iocount); 2705 } 2706 mutex_enter(&lfs_lock); 2707 #ifdef DIAGNOSTIC 2708 if (fs->lfs_iocount == 0) 2709 panic("lfs_cluster_aiodone: zero iocount"); 2710 #endif 2711 if (--fs->lfs_iocount <= 1) 2712 wakeup(&fs->lfs_iocount); 2713 mutex_exit(&lfs_lock); 2714 2715 KERNEL_UNLOCK_LAST(curlwp); 2716 2717 pool_put(&fs->lfs_bpppool, cl->bpp); 2718 cl->bpp = NULL; 2719 pool_put(&fs->lfs_clpool, cl); 2720 } 2721 2722 static void 2723 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2724 { 2725 /* reset b_iodone for when this is a single-buf i/o. */ 2726 bp->b_iodone = aiodone; 2727 2728 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2729 } 2730 2731 static void 2732 lfs_cluster_callback(struct buf *bp) 2733 { 2734 2735 lfs_generic_callback(bp, lfs_cluster_aiodone); 2736 } 2737 2738 void 2739 lfs_supercallback(struct buf *bp) 2740 { 2741 2742 lfs_generic_callback(bp, lfs_super_aiodone); 2743 } 2744 2745 /* 2746 * The only buffers that are going to hit these functions are the 2747 * segment write blocks, or the segment summaries, or the superblocks. 2748 * 2749 * All of the above are created by lfs_newbuf, and so do not need to be 2750 * released via brelse. 2751 */ 2752 void 2753 lfs_callback(struct buf *bp) 2754 { 2755 2756 lfs_generic_callback(bp, lfs_free_aiodone); 2757 } 2758 2759 /* 2760 * Shellsort (diminishing increment sort) from Data Structures and 2761 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2762 * see also Knuth Vol. 3, page 84. The increments are selected from 2763 * formula (8), page 95. Roughly O(N^3/2). 2764 */ 2765 /* 2766 * This is our own private copy of shellsort because we want to sort 2767 * two parallel arrays (the array of buffer pointers and the array of 2768 * logical block numbers) simultaneously. Note that we cast the array 2769 * of logical block numbers to a unsigned in this routine so that the 2770 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2771 */ 2772 2773 static void 2774 lfs_shellsort(struct lfs *fs, 2775 struct buf **bp_array, union lfs_blocks *lb_array, 2776 int nmemb, int size) 2777 { 2778 static int __rsshell_increments[] = { 4, 1, 0 }; 2779 int incr, *incrp, t1, t2; 2780 struct buf *bp_temp; 2781 2782 #ifdef DEBUG 2783 incr = 0; 2784 for (t1 = 0; t1 < nmemb; t1++) { 2785 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2786 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2787 /* dump before panic */ 2788 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2789 nmemb, size); 2790 incr = 0; 2791 for (t1 = 0; t1 < nmemb; t1++) { 2792 const struct buf *bp = bp_array[t1]; 2793 2794 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2795 PRIu64 "\n", t1, 2796 (uint64_t)bp->b_bcount, 2797 (uint64_t)bp->b_lblkno); 2798 printf("lbns:"); 2799 for (t2 = 0; t2 * size < bp->b_bcount; 2800 t2++) { 2801 printf(" %jd", 2802 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2803 } 2804 printf("\n"); 2805 } 2806 panic("lfs_shellsort: inconsistent input"); 2807 } 2808 } 2809 } 2810 #endif 2811 2812 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2813 for (t1 = incr; t1 < nmemb; ++t1) 2814 for (t2 = t1 - incr; t2 >= 0;) 2815 if ((u_int64_t)bp_array[t2]->b_lblkno > 2816 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2817 bp_temp = bp_array[t2]; 2818 bp_array[t2] = bp_array[t2 + incr]; 2819 bp_array[t2 + incr] = bp_temp; 2820 t2 -= incr; 2821 } else 2822 break; 2823 2824 /* Reform the list of logical blocks */ 2825 incr = 0; 2826 for (t1 = 0; t1 < nmemb; t1++) { 2827 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2828 lfs_blocks_set(fs, lb_array, incr++, 2829 bp_array[t1]->b_lblkno + t2); 2830 } 2831 } 2832 } 2833 2834 /* 2835 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2836 * point at uninitialized space. 2837 */ 2838 void 2839 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2840 { 2841 struct segment *sp = fs->lfs_sp; 2842 SEGSUM *ssp; 2843 2844 KASSERT(vers > 0); 2845 2846 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2847 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2848 (void) lfs_writeseg(fs, fs->lfs_sp); 2849 2850 sp->sum_bytes_left -= FINFOSIZE(fs); 2851 ssp = (SEGSUM *)sp->segsum; 2852 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2853 lfs_fi_setnblocks(fs, sp->fip, 0); 2854 lfs_fi_setino(fs, sp->fip, ino); 2855 lfs_fi_setversion(fs, sp->fip, vers); 2856 } 2857 2858 /* 2859 * Release the FINFO entry, either clearing out an unused entry or 2860 * advancing us to the next available entry. 2861 */ 2862 void 2863 lfs_release_finfo(struct lfs *fs) 2864 { 2865 struct segment *sp = fs->lfs_sp; 2866 SEGSUM *ssp; 2867 2868 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2869 sp->fip = NEXT_FINFO(fs, sp->fip); 2870 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2871 } else { 2872 /* XXX shouldn't this update sp->fip? */ 2873 sp->sum_bytes_left += FINFOSIZE(fs); 2874 ssp = (SEGSUM *)sp->segsum; 2875 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2876 } 2877 } 2878
Indexes created Tue Sep 23 02:09:52 GMT 2025