lfs_segment.c revision 1.292
1 /* $NetBSD: lfs_segment.c,v 1.292 2025/09/17 04:01:53 perseant 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 1.292 2025/09/17 04:01:53 perseant Exp $"); 64 65 #ifdef DEBUG 66 # define vndebug(vp, str) do { \ 67 if (VTOI(vp)->i_state & IN_CLEANING) \ 68 DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \ 69 VTOI(vp)->i_number, (str), op)); \ 70 } while(0) 71 #else 72 # define vndebug(vp, str) 73 #endif 74 #define ivndebug(vp, str) \ 75 DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str))) 76 77 #if defined(_KERNEL_OPT) 78 #include "opt_ddb.h" 79 #endif 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/namei.h> 84 #include <sys/kernel.h> 85 #include <sys/resourcevar.h> 86 #include <sys/file.h> 87 #include <sys/stat.h> 88 #include <sys/buf.h> 89 #include <sys/proc.h> 90 #include <sys/vnode.h> 91 #include <sys/mount.h> 92 #include <sys/kauth.h> 93 #include <sys/syslog.h> 94 #include <sys/workqueue.h> 95 96 #include <miscfs/specfs/specdev.h> 97 #include <miscfs/fifofs/fifo.h> 98 99 #include <ufs/lfs/ulfs_inode.h> 100 #include <ufs/lfs/ulfsmount.h> 101 #include <ufs/lfs/ulfs_extern.h> 102 103 #include <ufs/lfs/lfs.h> 104 #include <ufs/lfs/lfs_accessors.h> 105 #include <ufs/lfs/lfs_kernel.h> 106 #include <ufs/lfs/lfs_extern.h> 107 108 #include <uvm/uvm_extern.h> 109 #include <uvm/uvm_page.h> 110 111 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 112 113 static void lfs_super_aiodone(struct buf *); 114 static void lfs_cluster_aiodone(struct buf *); 115 116 /* 117 * Determine if it's OK to start a partial in this segment, or if we need 118 * to go on to a new segment. 119 */ 120 #define LFS_PARTIAL_FITS(fs) \ 121 (lfs_sb_getfsbpseg(fs) - \ 122 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)) > \ 123 lfs_sb_getfrag(fs)) 124 125 /* 126 * Figure out whether we should do a checkpoint write or go ahead with 127 * an ordinary write. 128 */ 129 #define LFS_SHOULD_CHECKPOINT(fs, flags) \ 130 ((flags & SEGM_CLEAN) == 0 && \ 131 ((fs->lfs_nactive > LFS_MAX_ACTIVE || \ 132 (flags & SEGM_CKP) || \ 133 lfs_sb_getnclean(fs) < LFS_MAX_ACTIVE))) 134 135 int lfs_match_fake(struct lfs *, struct buf *); 136 void lfs_newseg(struct lfs *); 137 void lfs_updatemeta(struct segment *); 138 void lfs_writesuper(struct lfs *, daddr_t); 139 int lfs_writevnodes(struct lfs *fs, struct mount *mp, 140 struct segment *sp, int dirops); 141 142 static void lfs_shellsort(struct lfs *, struct buf **, union lfs_blocks *, 143 int, int); 144 145 kcondvar_t lfs_allclean_wakeup; /* Cleaner wakeup address. */ 146 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 147 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 148 int lfs_dirvcount = 0; /* # active dirops */ 149 150 extern struct workqueue *lfs_cluster_wq; 151 extern struct workqueue *lfs_super_wq; 152 153 /* Statistics Counters */ 154 int lfs_dostats = 1; 155 struct lfs_stats lfs_stats; 156 157 /* op values to lfs_writevnodes */ 158 #define VN_REG 0 159 #define VN_DIROP 1 160 #define VN_EMPTY 2 161 #define VN_CLEAN 3 162 163 /* 164 * XXX KS - Set modification time on the Ifile, so the cleaner can 165 * read the fs mod time off of it. We don't set IN_UPDATE here, 166 * since we don't really need this to be flushed to disk (and in any 167 * case that wouldn't happen to the Ifile until we checkpoint). 168 */ 169 void 170 lfs_imtime(struct lfs *fs) 171 { 172 struct timespec ts; 173 struct inode *ip; 174 175 ASSERT_MAYBE_SEGLOCK(fs); 176 vfs_timestamp(&ts); 177 ip = VTOI(fs->lfs_ivnode); 178 lfs_dino_setmtime(fs, ip->i_din, ts.tv_sec); 179 lfs_dino_setmtimensec(fs, ip->i_din, ts.tv_nsec); 180 } 181 182 /* 183 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 184 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 185 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 186 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 187 */ 188 189 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 190 191 int 192 lfs_vflush(struct vnode *vp) 193 { 194 struct inode *ip; 195 struct lfs *fs; 196 struct segment *sp; 197 struct buf *bp, *nbp, *tbp, *tnbp; 198 int error; 199 int flushed; 200 int relock; 201 202 ip = VTOI(vp); 203 fs = VFSTOULFS(vp->v_mount)->um_lfs; 204 relock = 0; 205 206 top: 207 KASSERT(mutex_owned(vp->v_interlock) == false); 208 KASSERT(mutex_owned(&lfs_lock) == false); 209 KASSERT(mutex_owned(&bufcache_lock) == false); 210 ASSERT_NO_SEGLOCK(fs); 211 if (ip->i_state & IN_CLEANING) { 212 ivndebug(vp,"vflush/in_cleaning"); 213 mutex_enter(&lfs_lock); 214 LFS_CLR_UINO(ip, IN_CLEANING); 215 LFS_SET_UINO(ip, IN_MODIFIED); 216 mutex_exit(&lfs_lock); 217 218 /* 219 * Toss any cleaning buffers that have real counterparts 220 * to avoid losing new data. 221 */ 222 mutex_enter(vp->v_interlock); 223 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 224 nbp = LIST_NEXT(bp, b_vnbufs); 225 if (!LFS_IS_MALLOC_BUF(bp)) 226 continue; 227 /* 228 * Look for pages matching the range covered 229 * by cleaning blocks. It's okay if more dirty 230 * pages appear, so long as none disappear out 231 * from under us. 232 */ 233 if (bp->b_lblkno > 0 && vp->v_type == VREG && 234 vp != fs->lfs_ivnode) { 235 struct vm_page *pg; 236 voff_t off; 237 238 for (off = lfs_lblktosize(fs, bp->b_lblkno); 239 off < lfs_lblktosize(fs, bp->b_lblkno + 1); 240 off += PAGE_SIZE) { 241 pg = uvm_pagelookup(&vp->v_uobj, off); 242 if (pg == NULL) 243 continue; 244 if (uvm_pagegetdirty(pg) 245 == UVM_PAGE_STATUS_DIRTY || 246 pmap_is_modified(pg)) { 247 lfs_sb_addavail(fs, 248 lfs_btofsb(fs, 249 bp->b_bcount)); 250 wakeup(&fs->lfs_availsleep); 251 mutex_exit(vp->v_interlock); 252 lfs_freebuf(fs, bp); 253 mutex_enter(vp->v_interlock); 254 bp = NULL; 255 break; 256 } 257 } 258 } 259 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 260 tbp = tnbp) 261 { 262 tnbp = LIST_NEXT(tbp, b_vnbufs); 263 if (tbp->b_vp == bp->b_vp 264 && tbp->b_lblkno == bp->b_lblkno 265 && tbp != bp) 266 { 267 lfs_sb_addavail(fs, lfs_btofsb(fs, 268 bp->b_bcount)); 269 wakeup(&fs->lfs_availsleep); 270 mutex_exit(vp->v_interlock); 271 lfs_freebuf(fs, bp); 272 mutex_enter(vp->v_interlock); 273 bp = NULL; 274 break; 275 } 276 } 277 } 278 } else { 279 mutex_enter(vp->v_interlock); 280 } 281 282 /* If the node is being written, wait until that is done */ 283 while (WRITEINPROG(vp)) { 284 ivndebug(vp,"vflush/writeinprog"); 285 cv_wait(&vp->v_cv, vp->v_interlock); 286 } 287 error = vdead_check(vp, VDEAD_NOWAIT); 288 mutex_exit(vp->v_interlock); 289 290 /* Protect against deadlock in vinvalbuf() */ 291 lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0)); 292 if (error != 0) { 293 fs->lfs_reclino = ip->i_number; 294 } 295 296 /* If we're supposed to flush a freed inode, just toss it */ 297 if (ip->i_lfs_iflags & LFSI_DELETED) { 298 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 299 ip->i_number)); 300 /* Drain v_numoutput */ 301 mutex_enter(vp->v_interlock); 302 while (vp->v_numoutput > 0) { 303 cv_wait(&vp->v_cv, vp->v_interlock); 304 } 305 KASSERT(vp->v_numoutput == 0); 306 mutex_exit(vp->v_interlock); 307 308 mutex_enter(&bufcache_lock); 309 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 310 nbp = LIST_NEXT(bp, b_vnbufs); 311 312 KASSERT((bp->b_flags & B_GATHERED) == 0); 313 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 314 lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount)); 315 wakeup(&fs->lfs_availsleep); 316 } 317 /* Copied from lfs_writeseg */ 318 if (bp->b_iodone != NULL) { 319 mutex_exit(&bufcache_lock); 320 biodone(bp); 321 mutex_enter(&bufcache_lock); 322 } else { 323 bremfree(bp); 324 LFS_UNLOCK_BUF(bp); 325 mutex_enter(vp->v_interlock); 326 bp->b_flags &= ~(B_READ | B_GATHERED); 327 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 328 bp->b_error = 0; 329 reassignbuf(bp, vp); 330 mutex_exit(vp->v_interlock); 331 brelse(bp, 0); 332 } 333 } 334 mutex_exit(&bufcache_lock); 335 LFS_CLR_UINO(ip, IN_CLEANING); 336 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 337 ip->i_state &= ~IN_ALLMOD; 338 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 339 ip->i_number)); 340 lfs_segunlock(fs); 341 342 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 343 344 return 0; 345 } 346 347 fs->lfs_flushvp = vp; 348 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 349 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 350 fs->lfs_flushvp = NULL; 351 KASSERT(fs->lfs_flushvp_fakevref == 0); 352 lfs_segunlock(fs); 353 354 /* Make sure that any pending buffers get written */ 355 mutex_enter(vp->v_interlock); 356 while (vp->v_numoutput > 0) { 357 cv_wait(&vp->v_cv, vp->v_interlock); 358 } 359 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 360 KASSERT(vp->v_numoutput == 0); 361 mutex_exit(vp->v_interlock); 362 363 return error; 364 } 365 sp = fs->lfs_sp; 366 367 flushed = 0; 368 if (VPISEMPTY(vp)) { 369 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 370 ++flushed; 371 } else if ((ip->i_state & IN_CLEANING) && 372 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 373 ivndebug(vp,"vflush/clean"); 374 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 375 ++flushed; 376 } else if (lfs_dostats) { 377 if (!VPISEMPTY(vp) || (VTOI(vp)->i_state & IN_ALLMOD)) 378 ++lfs_stats.vflush_invoked; 379 ivndebug(vp,"vflush"); 380 } 381 382 #ifdef DIAGNOSTIC 383 if (vp->v_uflag & VU_DIROP) { 384 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 385 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 386 } 387 #endif 388 389 do { 390 #ifdef DEBUG 391 int loopcount = 0; 392 #endif 393 do { 394 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 395 relock = lfs_writefile(fs, sp, vp); 396 if (relock && vp != fs->lfs_ivnode) { 397 /* 398 * Might have to wait for the 399 * cleaner to run; but we're 400 * still not done with this vnode. 401 * XXX we can do better than this. 402 */ 403 KASSERT(ip->i_number != LFS_IFILE_INUM); 404 lfs_writeinode(fs, sp, ip); 405 mutex_enter(&lfs_lock); 406 LFS_SET_UINO(ip, IN_MODIFIED); 407 mutex_exit(&lfs_lock); 408 lfs_writeseg(fs, sp); 409 lfs_segunlock(fs); 410 lfs_segunlock_relock(fs); 411 goto top; 412 } 413 } 414 /* 415 * If we begin a new segment in the middle of writing 416 * the Ifile, it creates an inconsistent checkpoint, 417 * since the Ifile information for the new segment 418 * is not up-to-date. Take care of this here by 419 * sending the Ifile through again in case there 420 * are newly dirtied blocks. But wait, there's more! 421 * This second Ifile write could *also* cross a segment 422 * boundary, if the first one was large. The second 423 * one is guaranteed to be no more than 8 blocks, 424 * though (two segment blocks and supporting indirects) 425 * so the third write *will not* cross the boundary. 426 */ 427 if (vp == fs->lfs_ivnode) { 428 lfs_writefile(fs, sp, vp); 429 lfs_writefile(fs, sp, vp); 430 } 431 #ifdef DEBUG 432 if (++loopcount > 2) 433 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 434 #endif 435 } while (lfs_writeinode(fs, sp, ip)); 436 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 437 438 if (lfs_dostats) { 439 ++lfs_stats.nwrites; 440 if (sp->seg_flags & SEGM_SYNC) 441 ++lfs_stats.nsync_writes; 442 if (sp->seg_flags & SEGM_CKP) 443 ++lfs_stats.ncheckpoints; 444 } 445 /* 446 * If we were called from somewhere that has already held the seglock 447 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 448 * the write to complete because we are still locked. 449 * Since lfs_vflush() must return the vnode with no dirty buffers, 450 * we must explicitly wait, if that is the case. 451 * 452 * We compare the iocount against 1, not 0, because it is 453 * artificially incremented by lfs_seglock(). 454 */ 455 mutex_enter(&lfs_lock); 456 if (fs->lfs_seglock > 1) { 457 while (fs->lfs_iocount > 1) 458 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 459 "lfs_vflush", 0, &lfs_lock); 460 } 461 mutex_exit(&lfs_lock); 462 463 lfs_segunlock(fs); 464 465 /* Wait for these buffers to be recovered by aiodoned */ 466 mutex_enter(vp->v_interlock); 467 while (vp->v_numoutput > 0) { 468 cv_wait(&vp->v_cv, vp->v_interlock); 469 } 470 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 471 KASSERT(vp->v_numoutput == 0); 472 mutex_exit(vp->v_interlock); 473 474 fs->lfs_flushvp = NULL; 475 KASSERT(fs->lfs_flushvp_fakevref == 0); 476 477 return (0); 478 } 479 480 struct lfs_writevnodes_ctx { 481 int op; 482 struct lfs *fs; 483 }; 484 static bool 485 lfs_writevnodes_selector(void *cl, struct vnode *vp) 486 { 487 struct lfs_writevnodes_ctx *c = cl; 488 struct inode *ip; 489 int op = c->op; 490 491 KASSERT(mutex_owned(vp->v_interlock)); 492 493 ip = VTOI(vp); 494 if (ip == NULL || vp->v_type == VNON || ip->i_nlink <= 0) 495 return false; 496 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 497 (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) { 498 vndebug(vp, "dirop"); 499 return false; 500 } 501 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 502 vndebug(vp,"empty"); 503 return false; 504 } 505 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM && 506 vp != c->fs->lfs_flushvp && !(ip->i_state & IN_CLEANING)) { 507 vndebug(vp,"cleaning"); 508 return false; 509 } 510 mutex_enter(&lfs_lock); 511 if (vp == c->fs->lfs_unlockvp) { 512 mutex_exit(&lfs_lock); 513 return false; 514 } 515 mutex_exit(&lfs_lock); 516 517 return true; 518 } 519 520 int 521 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 522 { 523 struct inode *ip; 524 struct vnode *vp; 525 struct vnode_iterator *marker; 526 struct lfs_writevnodes_ctx ctx; 527 int inodes_written = 0; 528 int error = 0; 529 530 /* 531 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist. 532 * XXX The rationale is unclear, the initial commit had no information. 533 * XXX If the order really matters we have to sort the vnodes first. 534 */ 535 536 ASSERT_SEGLOCK(fs); 537 vfs_vnode_iterator_init(mp, &marker); 538 ctx.op = op; 539 ctx.fs = fs; 540 while ((vp = vfs_vnode_iterator_next(marker, 541 lfs_writevnodes_selector, &ctx)) != NULL) { 542 ip = VTOI(vp); 543 544 /* 545 * Write the inode/file if dirty and it's not the IFILE. 546 */ 547 if (((ip->i_state & IN_ALLMOD) || !VPISEMPTY(vp)) && 548 ip->i_number != LFS_IFILE_INUM) { 549 error = lfs_writefile(fs, sp, vp); 550 if (error) { 551 vrele(vp); 552 if (error == EAGAIN) { 553 /* 554 * This error from lfs_putpages 555 * indicates we need to drop 556 * the segment lock and start 557 * over after the cleaner has 558 * had a chance to run. 559 */ 560 lfs_writeinode(fs, sp, ip); 561 lfs_writeseg(fs, sp); 562 if (!VPISEMPTY(vp) && 563 !WRITEINPROG(vp) && 564 !(ip->i_state & IN_ALLMOD)) { 565 mutex_enter(&lfs_lock); 566 LFS_SET_UINO(ip, IN_MODIFIED); 567 mutex_exit(&lfs_lock); 568 } 569 break; 570 } 571 error = 0; /* XXX not quite right */ 572 continue; 573 } 574 575 if (!VPISEMPTY(vp)) { 576 if (WRITEINPROG(vp)) { 577 ivndebug(vp,"writevnodes/write2"); 578 } else if (!(ip->i_state & IN_ALLMOD)) { 579 mutex_enter(&lfs_lock); 580 LFS_SET_UINO(ip, IN_MODIFIED); 581 mutex_exit(&lfs_lock); 582 } 583 } 584 (void) lfs_writeinode(fs, sp, ip); 585 inodes_written++; 586 } 587 vrele(vp); 588 } 589 vfs_vnode_iterator_destroy(marker); 590 return error; 591 } 592 593 /* 594 * Do a checkpoint. 595 */ 596 int 597 lfs_segwrite(struct mount *mp, int flags) 598 { 599 struct buf *bp; 600 struct inode *ip; 601 struct lfs *fs; 602 struct segment *sp; 603 struct vnode *vp; 604 SEGUSE *segusep; 605 int do_ckp, did_ckp, error; 606 unsigned n, segleft, maxseg, sn, i, curseg; 607 int writer_set = 0; 608 int dirty; 609 int redo; 610 SEGSUM *ssp; 611 int um_error; 612 613 fs = VFSTOULFS(mp)->um_lfs; 614 ASSERT_MAYBE_SEGLOCK(fs); 615 616 if (fs->lfs_ronly) 617 return EROFS; 618 619 lfs_imtime(fs); 620 621 /* 622 * Allocate a segment structure and enough space to hold pointers to 623 * the maximum possible number of buffers which can be described in a 624 * single summary block. 625 */ 626 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 627 628 /* 629 * If we know we're gonna need the writer lock, take it now to 630 * preserve the lock order lfs_writer -> lfs_seglock. 631 */ 632 if (do_ckp) { 633 lfs_writer_enter(fs, "ckpwriter"); 634 writer_set = 1; 635 } 636 637 /* We can't do a partial write and checkpoint at the same time. */ 638 if (do_ckp) 639 flags &= ~SEGM_SINGLE; 640 641 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 642 sp = fs->lfs_sp; 643 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 644 do_ckp = 1; 645 646 /* 647 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 648 * in which case we have to flush *all* buffers off of this vnode. 649 * We don't care about other nodes, but write any non-dirop nodes 650 * anyway in anticipation of another getnewvnode(). 651 * 652 * If we're cleaning we only write cleaning and ifile blocks, and 653 * no dirops, since otherwise we'd risk corruption in a crash. 654 */ 655 if (sp->seg_flags & SEGM_CLEAN) 656 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 657 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 658 do { 659 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 660 if ((sp->seg_flags & SEGM_SINGLE) && 661 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 662 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 663 break; 664 } 665 666 if (do_ckp || 667 (writer_set = lfs_writer_tryenter(fs)) != 0) { 668 KASSERT(writer_set); 669 KASSERT(fs->lfs_writer); 670 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 671 if (um_error == 0) 672 um_error = error; 673 /* 674 * In case writevnodes errored out 675 * XXX why are we always doing this and not 676 * just on error? 677 */ 678 lfs_flush_dirops(fs); 679 ssp = (SEGSUM *)(sp->segsum); 680 lfs_ss_setflags(fs, ssp, 681 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 682 lfs_finalize_fs_seguse(fs); 683 } 684 if (do_ckp && um_error) { 685 lfs_segunlock_relock(fs); 686 sp = fs->lfs_sp; 687 } 688 } while (do_ckp && um_error != 0); 689 } 690 691 /* 692 * If we are doing a checkpoint, mark everything since the 693 * last checkpoint as no longer ACTIVE. 694 */ 695 if (do_ckp || fs->lfs_doifile) { 696 segleft = lfs_sb_getnseg(fs); 697 curseg = 0; 698 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 699 int bread_error; 700 701 dirty = 0; 702 bread_error = bread(fs->lfs_ivnode, 703 lfs_sb_getcleansz(fs) + n, 704 lfs_sb_getbsize(fs), B_MODIFY, &bp); 705 if (bread_error) 706 panic("lfs_segwrite: ifile read: " 707 "seguse %u: error %d\n", 708 n, bread_error); 709 segusep = (SEGUSE *)bp->b_data; 710 maxseg = uimin(segleft, lfs_sb_getsepb(fs)); 711 for (i = 0; i < maxseg; i++) { 712 sn = curseg + i; 713 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 714 segusep->su_flags & SEGUSE_ACTIVE) { 715 segusep->su_flags &= ~SEGUSE_ACTIVE; 716 --fs->lfs_nactive; 717 ++dirty; 718 } 719 fs->lfs_suflags[fs->lfs_activesb][sn] = 720 segusep->su_flags; 721 if (lfs_sb_getversion(fs) > 1) 722 ++segusep; 723 else 724 segusep = (SEGUSE *) 725 ((SEGUSE_V1 *)segusep + 1); 726 } 727 728 if (dirty) 729 error = LFS_BWRITE_LOG(bp); /* Ifile */ 730 else 731 brelse(bp, 0); 732 segleft -= lfs_sb_getsepb(fs); 733 curseg += lfs_sb_getsepb(fs); 734 } 735 } 736 737 KASSERT(LFS_SEGLOCK_HELD(fs)); 738 739 did_ckp = 0; 740 if (do_ckp || fs->lfs_doifile) { 741 vp = fs->lfs_ivnode; 742 #ifdef DEBUG 743 int loopcount = 0; 744 #endif 745 do { 746 747 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 748 749 mutex_enter(&lfs_lock); 750 fs->lfs_flags &= ~LFS_IFDIRTY; 751 mutex_exit(&lfs_lock); 752 753 ip = VTOI(vp); 754 755 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 756 /* 757 * Ifile has no pages, so we don't need 758 * to check error return here. 759 */ 760 lfs_writefile(fs, sp, vp); 761 /* 762 * Ensure the Ifile takes the current segment 763 * into account. See comment in lfs_vflush. 764 */ 765 lfs_writefile(fs, sp, vp); 766 lfs_writefile(fs, sp, vp); 767 } 768 769 if (ip->i_state & IN_ALLMOD) 770 ++did_ckp; 771 #if 0 772 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 773 #else 774 redo = lfs_writeinode(fs, sp, ip); 775 #endif 776 redo += lfs_writeseg(fs, sp); 777 mutex_enter(&lfs_lock); 778 redo += (fs->lfs_flags & LFS_IFDIRTY); 779 mutex_exit(&lfs_lock); 780 #ifdef DEBUG 781 if (++loopcount > 2) 782 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 783 loopcount); 784 #endif 785 } while (redo && do_ckp); 786 787 /* 788 * Unless we are unmounting, the Ifile may continue to have 789 * dirty blocks even after a checkpoint, due to changes to 790 * inodes' atime. If we're checkpointing, it's "impossible" 791 * for other parts of the Ifile to be dirty after the loop 792 * above, since we hold the segment lock. 793 */ 794 mutex_enter(vp->v_interlock); 795 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 796 LFS_CLR_UINO(ip, IN_ALLMOD); 797 } 798 #ifdef DIAGNOSTIC 799 else if (do_ckp) { 800 int do_panic = 0; 801 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 802 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 803 lfs_sb_getsegtabsz(fs) && 804 !(bp->b_flags & B_GATHERED)) { 805 printf("ifile lbn %ld still dirty (flags %lx)\n", 806 (long)bp->b_lblkno, 807 (long)bp->b_flags); 808 ++do_panic; 809 } 810 } 811 if (do_panic) 812 panic("dirty blocks"); 813 } 814 #endif 815 mutex_exit(vp->v_interlock); 816 } else { 817 (void) lfs_writeseg(fs, sp); 818 } 819 820 /* Note Ifile no longer needs to be written */ 821 fs->lfs_doifile = 0; 822 if (writer_set) 823 lfs_writer_leave(fs); 824 825 /* 826 * If we didn't write the Ifile, we didn't really do anything. 827 * That means that (1) there is a checkpoint on disk and (2) 828 * nothing has changed since it was written. 829 * 830 * Take the flags off of the segment so that lfs_segunlock 831 * doesn't have to write the superblock either. 832 */ 833 if (do_ckp && !did_ckp) { 834 sp->seg_flags &= ~SEGM_CKP; 835 } 836 837 if (lfs_dostats) { 838 ++lfs_stats.nwrites; 839 if (sp->seg_flags & SEGM_SYNC) 840 ++lfs_stats.nsync_writes; 841 if (sp->seg_flags & SEGM_CKP) 842 ++lfs_stats.ncheckpoints; 843 } 844 lfs_segunlock(fs); 845 return (0); 846 } 847 848 /* 849 * Write the dirty blocks associated with a vnode. 850 */ 851 int 852 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 853 { 854 struct inode *ip; 855 int i, frag; 856 SEGSUM *ssp; 857 int error; 858 859 ASSERT_SEGLOCK(fs); 860 error = 0; 861 ip = VTOI(vp); 862 863 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 864 865 if (vp->v_uflag & VU_DIROP) { 866 ssp = (SEGSUM *)sp->segsum; 867 lfs_ss_setflags(fs, ssp, 868 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 869 } 870 871 if (sp->seg_flags & SEGM_CLEAN) { 872 lfs_gather(fs, sp, vp, lfs_match_fake); 873 /* 874 * For a file being flushed, we need to write *all* blocks. 875 * This means writing the cleaning blocks first, and then 876 * immediately following with any non-cleaning blocks. 877 * The same is true of the Ifile since checkpoints assume 878 * that all valid Ifile blocks are written. 879 */ 880 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 881 lfs_gather(fs, sp, vp, lfs_match_data); 882 /* 883 * Don't call VOP_PUTPAGES: if we're flushing, 884 * we've already done it, and the Ifile doesn't 885 * use the page cache. 886 */ 887 } 888 } else { 889 lfs_gather(fs, sp, vp, lfs_match_data); 890 /* 891 * If we're flushing, we've already called VOP_PUTPAGES 892 * so don't do it again. Otherwise, we want to write 893 * everything we've got. 894 */ 895 if (!IS_FLUSHING(fs, vp)) { 896 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER); 897 error = VOP_PUTPAGES(vp, 0, 0, 898 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 899 } 900 } 901 902 /* 903 * It may not be necessary to write the meta-data blocks at this point, 904 * as the roll-forward recovery code should be able to reconstruct the 905 * list. 906 * 907 * We have to write them anyway, though, under two conditions: (1) the 908 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 909 * checkpointing. 910 * 911 * BUT if we are cleaning, we might have indirect blocks that refer to 912 * new blocks not being written yet, in addition to fragments being 913 * moved out of a cleaned segment. If that is the case, don't 914 * write the indirect blocks, or the finfo will have a small block 915 * in the middle of it! 916 * XXX in this case isn't the inode size wrong too? 917 */ 918 frag = 0; 919 if (sp->seg_flags & SEGM_CLEAN) { 920 for (i = 0; i < ULFS_NDADDR; i++) 921 if (ip->i_lfs_fragsize[i] > 0 && 922 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 923 ++frag; 924 } 925 KASSERTMSG((frag <= 1), 926 "lfs_writefile: more than one fragment! frag=%d", frag); 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_IS_BAD(daddr)) { 1008 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1009 int ndupino __diagused = 1010 (sp->seg_number == oldsn) ? sp->ndupino : 0; 1011 LFS_SEGENTRY(sup, fs, oldsn, bp); 1012 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) 1013 >= DINOSIZE(fs)), 1014 "lfs_writeinode: negative bytes " 1015 "(segment %" PRIu32 " short by %d, " 1016 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1017 ", daddr=%" PRId64 ", su_nbytes=%u, " 1018 "ndupino=%d)\n", 1019 lfs_dtosn(fs, daddr), 1020 (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes, 1021 oldsn, sp->seg_number, daddr, 1022 (unsigned int)sup->su_nbytes, 1023 sp->ndupino); 1024 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1025 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1026 sup->su_nbytes -= DINOSIZE(fs); 1027 redo_ifile |= 1028 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1029 if (redo_ifile) { 1030 mutex_enter(&lfs_lock); 1031 fs->lfs_flags |= LFS_IFDIRTY; 1032 mutex_exit(&lfs_lock); 1033 /* Don't double-account */ 1034 lfs_sb_setidaddr(fs, 0x0); 1035 } 1036 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1037 } 1038 1039 return redo_ifile; 1040 } 1041 1042 int 1043 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1044 { 1045 struct buf *bp; 1046 union lfs_dinode *cdp; 1047 struct vnode *vp = ITOV(ip); 1048 daddr_t daddr; 1049 IINFO *iip; 1050 int i; 1051 int redo_ifile = 0; 1052 int gotblk = 0; 1053 int count; 1054 SEGSUM *ssp; 1055 1056 ASSERT_SEGLOCK(fs); 1057 if (!(ip->i_state & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1058 return (0); 1059 1060 /* Can't write ifile when writer is not set */ 1061 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1062 (sp->seg_flags & SEGM_CLEAN)); 1063 1064 /* 1065 * If this is the Ifile, see if writing it here will generate a 1066 * temporary misaccounting. If it will, do the accounting and write 1067 * the blocks, postponing the inode write until the accounting is 1068 * solid. 1069 */ 1070 count = 0; 1071 while (vp == fs->lfs_ivnode) { 1072 int redo = 0; 1073 1074 if (sp->idp == NULL && sp->ibp == NULL && 1075 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1076 sp->sum_bytes_left < sizeof(int32_t))) { 1077 (void) lfs_writeseg(fs, sp); 1078 continue; 1079 } 1080 1081 /* Look for dirty Ifile blocks */ 1082 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1083 if (!(bp->b_flags & B_GATHERED)) { 1084 redo = 1; 1085 break; 1086 } 1087 } 1088 1089 if (redo == 0) 1090 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1091 if (redo == 0) 1092 break; 1093 1094 if (sp->idp) { 1095 lfs_dino_setinumber(fs, sp->idp, 0); 1096 sp->idp = NULL; 1097 } 1098 ++count; 1099 if (count > 2) 1100 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1101 lfs_writefile(fs, sp, fs->lfs_ivnode); 1102 } 1103 1104 /* Allocate a new inode block if necessary. */ 1105 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1106 sp->ibp == NULL) { 1107 /* Allocate a new segment if necessary. */ 1108 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1109 sp->sum_bytes_left < sizeof(int32_t)) 1110 (void) lfs_writeseg(fs, sp); 1111 1112 /* Get next inode block. */ 1113 daddr = lfs_sb_getoffset(fs); 1114 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1115 sp->ibp = *sp->cbpp++ = 1116 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1117 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1118 gotblk++; 1119 1120 /* Zero out inode numbers */ 1121 for (i = 0; i < LFS_INOPB(fs); ++i) { 1122 union lfs_dinode *tmpdi; 1123 1124 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1125 DINOSIZE(fs) * i); 1126 lfs_dino_setinumber(fs, tmpdi, 0); 1127 } 1128 1129 ++sp->start_bpp; 1130 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1131 /* Set remaining space counters. */ 1132 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1133 sp->sum_bytes_left -= sizeof(int32_t); 1134 1135 /* Store the address in the segment summary. */ 1136 iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs)); 1137 lfs_ii_setblock(fs, iip, daddr); 1138 } 1139 1140 /* Check VU_DIROP in case there is a new file with no data blocks */ 1141 if (vp->v_uflag & VU_DIROP) { 1142 ssp = (SEGSUM *)sp->segsum; 1143 lfs_ss_setflags(fs, ssp, 1144 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1145 } 1146 1147 /* Update the inode times and copy the inode onto the inode page. */ 1148 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1149 if (ip->i_number != LFS_IFILE_INUM) 1150 LFS_ITIMES(ip, NULL, NULL, NULL); 1151 1152 /* 1153 * If this is the Ifile, and we've already written the Ifile in this 1154 * partial segment, just overwrite it (it's not on disk yet) and 1155 * continue. 1156 * 1157 * XXX we know that the bp that we get the second time around has 1158 * already been gathered. 1159 */ 1160 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1161 lfs_copy_dinode(fs, sp->idp, ip->i_din); 1162 ip->i_lfs_osize = ip->i_size; 1163 return 0; 1164 } 1165 1166 bp = sp->ibp; 1167 cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs)); 1168 lfs_copy_dinode(fs, cdp, ip->i_din); 1169 1170 /* 1171 * This inode is on its way to disk; clear its VU_DIROP status when 1172 * the write is complete. 1173 */ 1174 if (vp->v_uflag & VU_DIROP) { 1175 if (!(sp->seg_flags & SEGM_CLEAN)) 1176 ip->i_state |= IN_CDIROP; 1177 else { 1178 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1179 } 1180 } 1181 1182 /* 1183 * If cleaning, link counts and directory file sizes cannot change, 1184 * since those would be directory operations---even if the file 1185 * we are writing is marked VU_DIROP we should write the old values. 1186 * If we're not cleaning, of course, update the values so we get 1187 * current values the next time we clean. 1188 */ 1189 if (sp->seg_flags & SEGM_CLEAN) { 1190 if (vp->v_uflag & VU_DIROP) { 1191 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1192 /* if (vp->v_type == VDIR) */ 1193 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1194 } 1195 } else { 1196 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1197 ip->i_lfs_osize = ip->i_size; 1198 } 1199 1200 1201 /* We can finish the segment accounting for truncations now */ 1202 lfs_finalize_ino_seguse(fs, ip); 1203 1204 /* 1205 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1206 * addresses to disk; possibly change the on-disk record of 1207 * the inode size, either by reverting to the previous size 1208 * (in the case of cleaning) or by verifying the inode's block 1209 * holdings (in the case of files being allocated as they are being 1210 * written). 1211 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1212 * XXX count on disk wrong by the same amount. We should be 1213 * XXX able to "borrow" from lfs_avail and return it after the 1214 * XXX Ifile is written. See also in lfs_writeseg. 1215 */ 1216 1217 /* Check file size based on highest allocated block */ 1218 if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG || 1219 (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) && 1220 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1221 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1222 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1223 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1224 } 1225 if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) { 1226 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1227 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1228 lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs))); 1229 for (i=0; i<ULFS_NDADDR; i++) { 1230 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1231 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1232 lfs_dino_setdb(fs, cdp, i, 0); 1233 } 1234 } 1235 for (i=0; i<ULFS_NIADDR; i++) { 1236 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1237 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1238 lfs_dino_setib(fs, cdp, i, 0); 1239 } 1240 } 1241 } 1242 1243 #ifdef DIAGNOSTIC 1244 /* 1245 * Check dinode held blocks against dinode size. 1246 * This should be identical to the check in lfs_vget(). 1247 */ 1248 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1249 i < ULFS_NDADDR; i++) { 1250 KASSERT(i >= 0); 1251 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1252 continue; 1253 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1254 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1255 continue; 1256 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1257 # ifdef DEBUG 1258 lfs_dump_dinode(fs, cdp); 1259 # endif 1260 panic("writing inconsistent inode"); 1261 } 1262 } 1263 #endif /* DIAGNOSTIC */ 1264 1265 if (ip->i_state & IN_CLEANING) 1266 LFS_CLR_UINO(ip, IN_CLEANING); 1267 else { 1268 /* XXX IN_ALLMOD */ 1269 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1270 IN_UPDATE | IN_MODIFY); 1271 if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din)) 1272 LFS_CLR_UINO(ip, IN_MODIFIED); 1273 else { 1274 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1275 "blks=%d, eff=%jd\n", ip->i_number, 1276 lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks)); 1277 } 1278 } 1279 1280 if (ip->i_number == LFS_IFILE_INUM) { 1281 /* We know sp->idp == NULL */ 1282 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1283 1284 /* Not dirty any more */ 1285 mutex_enter(&lfs_lock); 1286 fs->lfs_flags &= ~LFS_IFDIRTY; 1287 mutex_exit(&lfs_lock); 1288 } 1289 1290 if (gotblk) { 1291 mutex_enter(&bufcache_lock); 1292 LFS_LOCK_BUF(bp); 1293 brelsel(bp, 0); 1294 mutex_exit(&bufcache_lock); 1295 } 1296 1297 /* Increment inode count in segment summary block. */ 1298 1299 ssp = (SEGSUM *)sp->segsum; 1300 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1301 1302 /* If this page is full, set flag to allocate a new page. */ 1303 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1304 sp->ibp = NULL; 1305 1306 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1307 1308 KASSERT(redo_ifile == 0); 1309 return (redo_ifile); 1310 } 1311 1312 int 1313 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1314 { 1315 struct lfs *fs; 1316 int vers; 1317 int j, blksinblk; 1318 1319 ASSERT_SEGLOCK(sp->fs); 1320 KASSERTMSG((sp->vp != NULL), 1321 "lfs_gatherblock: Null vp in segment"); 1322 1323 /* If full, finish this segment. */ 1324 fs = sp->fs; 1325 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1326 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1327 sp->seg_bytes_left < bp->b_bcount) { 1328 if (mptr) 1329 mutex_exit(mptr); 1330 lfs_updatemeta(sp); 1331 1332 vers = lfs_fi_getversion(fs, sp->fip); 1333 (void) lfs_writeseg(fs, sp); 1334 1335 /* Add the current file to the segment summary. */ 1336 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1337 1338 if (mptr) 1339 mutex_enter(mptr); 1340 return (1); 1341 } 1342 1343 if (bp->b_flags & B_GATHERED) { 1344 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1345 " lbn %" PRId64 "\n", 1346 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1347 return (0); 1348 } 1349 1350 /* Insert into the buffer list, update the FINFO block. */ 1351 bp->b_flags |= B_GATHERED; 1352 1353 *sp->cbpp++ = bp; 1354 for (j = 0; j < blksinblk; j++) { 1355 unsigned bn; 1356 1357 bn = lfs_fi_getnblocks(fs, sp->fip); 1358 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1359 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1360 /* This block's accounting moves from lfs_favail to lfs_avail */ 1361 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1362 } 1363 1364 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1365 sp->seg_bytes_left -= bp->b_bcount; 1366 return (0); 1367 } 1368 1369 int 1370 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1371 int (*match)(struct lfs *, struct buf *)) 1372 { 1373 struct buf *bp, *nbp; 1374 int count = 0; 1375 1376 ASSERT_SEGLOCK(fs); 1377 if (vp->v_type == VBLK) 1378 return 0; 1379 KASSERT(sp->vp == NULL); 1380 sp->vp = vp; 1381 mutex_enter(&bufcache_lock); 1382 1383 #ifndef LFS_NO_BACKBUF_HACK 1384 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1385 # define BUF_OFFSET \ 1386 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1387 # define BACK_BUF(BP) \ 1388 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1389 # define BEG_OF_LIST \ 1390 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1391 1392 loop: 1393 /* Find last buffer. */ 1394 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1395 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1396 bp = LIST_NEXT(bp, b_vnbufs)) 1397 continue; 1398 1399 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1400 nbp = BACK_BUF(bp); 1401 #else /* LFS_NO_BACKBUF_HACK */ 1402 loop: 1403 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1404 nbp = LIST_NEXT(bp, b_vnbufs); 1405 #endif /* LFS_NO_BACKBUF_HACK */ 1406 if ((bp->b_cflags & BC_BUSY) != 0 || 1407 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1408 #ifdef DEBUG 1409 if (vp == fs->lfs_ivnode && 1410 (bp->b_cflags & BC_BUSY) != 0 && 1411 (bp->b_flags & B_GATHERED) == 0) 1412 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1413 PRId64 " busy (%x) at 0x%jx", 1414 bp->b_lblkno, bp->b_flags, 1415 (uintmax_t)lfs_sb_getoffset(fs)); 1416 #endif 1417 continue; 1418 } 1419 #ifdef DIAGNOSTIC 1420 # ifdef LFS_USE_BC_INVAL 1421 if ((bp->b_cflags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1422 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1423 " is BC_INVAL\n", bp->b_lblkno)); 1424 VOP_PRINT(bp->b_vp); 1425 } 1426 # endif /* LFS_USE_BC_INVAL */ 1427 if (!(bp->b_oflags & BO_DELWRI)) 1428 panic("lfs_gather: bp not BO_DELWRI"); 1429 if (!(bp->b_flags & B_LOCKED)) { 1430 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1431 " blk %" PRId64 " not B_LOCKED\n", 1432 bp->b_lblkno, 1433 LFS_DBTOFSB(fs, bp->b_blkno))); 1434 VOP_PRINT(bp->b_vp); 1435 panic("lfs_gather: bp not B_LOCKED"); 1436 } 1437 #endif 1438 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1439 goto loop; 1440 } 1441 count++; 1442 } 1443 mutex_exit(&bufcache_lock); 1444 lfs_updatemeta(sp); 1445 KASSERT(sp->vp == vp); 1446 sp->vp = NULL; 1447 return count; 1448 } 1449 1450 #if DEBUG 1451 # define DEBUG_OOFF(n) do { \ 1452 if (ooff == 0) { \ 1453 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1454 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1455 ", was 0x0 (or %" PRId64 ")\n", \ 1456 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1457 } \ 1458 } while (0) 1459 #else 1460 # define DEBUG_OOFF(n) 1461 #endif 1462 1463 /* 1464 * Change the given block's address to ndaddr, finding its previous 1465 * location using ulfs_bmaparray(). 1466 * 1467 * Account for this change in the segment table. 1468 * 1469 * called with sp == NULL by roll-forwarding code. 1470 */ 1471 #define NOT_ON_DISK(daddr) ((daddr) == 0 || (daddr) == UNASSIGNED || (daddr) == UNWRITTEN) 1472 1473 void 1474 lfs_update_single(struct lfs *fs, struct segment *sp, 1475 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1476 { 1477 SEGUSE *sup; 1478 struct buf *bp; 1479 struct indir a[ULFS_NIADDR + 2], *ap; 1480 struct inode *ip; 1481 daddr_t daddr, ooff; 1482 int num, error; 1483 int bb, osize, obb; 1484 1485 ASSERT_SEGLOCK(fs); 1486 KASSERT(sp == NULL || sp->vp == vp); 1487 ip = VTOI(vp); 1488 1489 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1490 if (error) 1491 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1492 1493 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1494 if (daddr > 0) 1495 daddr = LFS_DBTOFSB(fs, daddr); 1496 1497 bb = lfs_numfrags(fs, size); 1498 switch (num) { 1499 case 0: 1500 ooff = lfs_dino_getdb(fs, ip->i_din, lbn); 1501 DEBUG_OOFF(0); 1502 if (NOT_ON_DISK(ooff)) 1503 lfs_dino_setblocks(fs, ip->i_din, 1504 lfs_dino_getblocks(fs, ip->i_din) + bb); 1505 else { 1506 /* possible fragment truncation or extension */ 1507 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1508 lfs_dino_setblocks(fs, ip->i_din, 1509 lfs_dino_getblocks(fs, ip->i_din) + (bb-obb)); 1510 } 1511 lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr); 1512 break; 1513 case 1: 1514 ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off); 1515 DEBUG_OOFF(1); 1516 if (NOT_ON_DISK(ooff)) 1517 lfs_dino_setblocks(fs, ip->i_din, 1518 lfs_dino_getblocks(fs, ip->i_din) + bb); 1519 lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr); 1520 break; 1521 default: 1522 ap = &a[num - 1]; 1523 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1524 B_MODIFY, &bp)) 1525 panic("lfs_updatemeta: bread bno %" PRId64, 1526 ap->in_lbn); 1527 1528 ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off); 1529 DEBUG_OOFF(num); 1530 if (NOT_ON_DISK(ooff)) 1531 lfs_dino_setblocks(fs, ip->i_din, 1532 lfs_dino_getblocks(fs, ip->i_din) + bb); 1533 lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr); 1534 (void) VOP_BWRITE(bp->b_vp, bp); 1535 } 1536 1537 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1538 1539 /* Update hiblk when extending the file */ 1540 if (lbn > ip->i_lfs_hiblk) 1541 ip->i_lfs_hiblk = lbn; 1542 1543 /* 1544 * Though we'd rather it couldn't, this *can* happen right now 1545 * if cleaning blocks and regular blocks coexist. 1546 */ 1547 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1548 1549 /* 1550 * Update segment usage information, based on old size 1551 * and location. 1552 */ 1553 if (daddr > 0) { 1554 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1555 int ndupino __diagused = (sp && sp->seg_number == oldsn ? 1556 sp->ndupino : 0); 1557 1558 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1559 if (lbn >= 0 && lbn < ULFS_NDADDR) 1560 osize = ip->i_lfs_fragsize[lbn]; 1561 else 1562 osize = lfs_sb_getbsize(fs); 1563 LFS_SEGENTRY(sup, fs, oldsn, bp); 1564 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize), 1565 "lfs_updatemeta: negative bytes " 1566 "(segment %" PRIu32 " short by %" PRId64 1567 ")\n" 1568 "lfs_updatemeta: ino %llu, lbn %" PRId64 1569 ", addr = 0x%" PRIx64 "\n" 1570 "lfs_updatemeta: ndupino=%d", 1571 lfs_dtosn(fs, daddr), 1572 (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes), 1573 (unsigned long long)ip->i_number, lbn, daddr, 1574 ndupino); 1575 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1576 " db 0x%" PRIx64 "\n", 1577 lfs_dtosn(fs, daddr), osize, 1578 ip->i_number, lbn, daddr)); 1579 sup->su_nbytes -= osize; 1580 if (!(bp->b_flags & B_GATHERED)) { 1581 mutex_enter(&lfs_lock); 1582 fs->lfs_flags |= LFS_IFDIRTY; 1583 mutex_exit(&lfs_lock); 1584 } 1585 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1586 } 1587 /* 1588 * Now that this block has a new address, and its old 1589 * segment no longer owns it, we can forget about its 1590 * old size. 1591 */ 1592 if (lbn >= 0 && lbn < ULFS_NDADDR) 1593 ip->i_lfs_fragsize[lbn] = size; 1594 } 1595 1596 /* 1597 * Update the metadata that points to the blocks listed in the FINFO 1598 * array. 1599 */ 1600 void 1601 lfs_updatemeta(struct segment *sp) 1602 { 1603 struct buf *sbp; 1604 struct lfs *fs; 1605 struct vnode *vp; 1606 daddr_t lbn; 1607 int i, nblocks, num; 1608 int __diagused nblocks_orig; 1609 int bb; 1610 int bytesleft, size; 1611 unsigned lastlength; 1612 union lfs_blocks tmpptr; 1613 1614 fs = sp->fs; 1615 vp = sp->vp; 1616 ASSERT_SEGLOCK(fs); 1617 1618 /* 1619 * This used to be: 1620 * 1621 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1622 * 1623 * that is, it allowed for the possibility that start_lbp did 1624 * not point to the beginning of the finfo block pointer area. 1625 * This particular formulation is six kinds of painful in the 1626 * lfs64 world where we have two sizes of block pointer, so 1627 * unless/until everything can be cleaned up to not move 1628 * start_lbp around but instead use an offset, we do the 1629 * following: 1630 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1631 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1632 * type. (Ugh.) 1633 * 2. Cast it to void *, then assign it to a temporary 1634 * union lfs_blocks. 1635 * 3. Subtract start_lbp from that. 1636 * 4. Save the value of nblocks in blocks_orig so we can 1637 * assert below that it hasn't changed without repeating this 1638 * rubbish. 1639 * 1640 * XXX. 1641 */ 1642 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1643 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1644 nblocks_orig = nblocks; 1645 1646 KASSERT(nblocks >= 0); 1647 KASSERT(vp != NULL); 1648 if (nblocks == 0) 1649 return; 1650 1651 /* 1652 * This count may be high due to oversize blocks from lfs_gop_write. 1653 * Correct for this. (XXX we should be able to keep track of these.) 1654 */ 1655 for (i = 0; i < nblocks; i++) { 1656 if (sp->start_bpp[i] == NULL) { 1657 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1658 nblocks = i; 1659 break; 1660 } 1661 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1662 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1663 nblocks -= num - 1; 1664 } 1665 1666 #if 0 1667 /* pre-lfs64 assertion */ 1668 KASSERT(vp->v_type == VREG || 1669 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1670 #else 1671 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1672 #endif 1673 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1674 1675 /* 1676 * Sort the blocks. 1677 * 1678 * We have to sort even if the blocks come from the 1679 * cleaner, because there might be other pending blocks on the 1680 * same inode...and if we don't sort, and there are fragments 1681 * present, blocks may be written in the wrong place. 1682 */ 1683 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1684 1685 /* 1686 * Record the length of the last block in case it's a fragment. 1687 * If there are indirect blocks present, they sort last. An 1688 * indirect block will be lfs_bsize and its presence indicates 1689 * that you cannot have fragments. 1690 * 1691 * XXX This last is a lie. A cleaned fragment can coexist with 1692 * XXX a later indirect block. This will continue to be 1693 * XXX true until lfs_markv is fixed to do everything with 1694 * XXX fake blocks (including fake inodes and fake indirect blocks). 1695 */ 1696 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1697 lfs_sb_getbmask(fs)) + 1; 1698 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1699 1700 /* 1701 * Assign disk addresses, and update references to the logical 1702 * block and the segment usage information. 1703 */ 1704 for (i = nblocks; i--; ++sp->start_bpp) { 1705 sbp = *sp->start_bpp; 1706 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1707 KASSERT(sbp->b_lblkno == lbn); 1708 1709 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1710 1711 /* 1712 * If we write a frag in the wrong place, the cleaner won't 1713 * be able to correctly identify its size later, and the 1714 * segment will be uncleanable. (Even worse, it will assume 1715 * that the indirect block that actually ends the list 1716 * is of a smaller size!) 1717 */ 1718 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1719 panic("lfs_updatemeta: fragment is not last block"); 1720 1721 /* 1722 * For each subblock in this possibly oversized block, 1723 * update its address on disk. 1724 */ 1725 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1726 KASSERT(vp == sbp->b_vp); 1727 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1728 bytesleft -= lfs_sb_getbsize(fs)) { 1729 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1730 bb = lfs_numfrags(fs, size); 1731 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1732 lfs_blocks_inc(fs, &sp->start_lbp); 1733 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1734 size); 1735 lfs_sb_addoffset(fs, bb); 1736 } 1737 1738 } 1739 1740 /* This inode has been modified */ 1741 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1742 } 1743 1744 /* 1745 * Move lfs_offset to a segment earlier than newsn. 1746 */ 1747 int 1748 lfs_rewind(struct lfs *fs, int newsn) 1749 { 1750 int sn, osn, isdirty; 1751 struct buf *bp; 1752 SEGUSE *sup; 1753 1754 ASSERT_SEGLOCK(fs); 1755 1756 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1757 if (osn < newsn) 1758 return 0; 1759 1760 /* lfs_avail eats the remaining space in this segment */ 1761 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1762 1763 /* Find a low-numbered segment */ 1764 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1765 LFS_SEGENTRY(sup, fs, sn, bp); 1766 isdirty = sup->su_flags & SEGUSE_DIRTY; 1767 brelse(bp, 0); 1768 1769 if (!isdirty) 1770 break; 1771 } 1772 if (sn == lfs_sb_getnseg(fs)) 1773 panic("lfs_rewind: no clean segments"); 1774 if (newsn >= 0 && sn >= newsn) 1775 return ENOENT; 1776 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1777 lfs_newseg(fs); 1778 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1779 1780 return 0; 1781 } 1782 1783 /* 1784 * Start a new partial segment. 1785 * 1786 * Return 1 when we entered to a new segment. 1787 * Otherwise, return 0. 1788 */ 1789 int 1790 lfs_initseg(struct lfs *fs, uint16_t flags) 1791 { 1792 struct segment *sp = fs->lfs_sp; 1793 SEGSUM *ssp; 1794 struct buf *sbp; /* buffer for SEGSUM */ 1795 int repeat = 0; /* return value */ 1796 1797 ASSERT_SEGLOCK(fs); 1798 /* Advance to the next segment. */ 1799 if (!LFS_PARTIAL_FITS(fs)) { 1800 SEGUSE *sup; 1801 struct buf *bp; 1802 1803 /* lfs_avail eats the remaining space */ 1804 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1805 lfs_sb_getcurseg(fs))); 1806 /* Wake up any cleaning procs waiting on this file system. */ 1807 lfs_wakeup_cleaner(fs); 1808 lfs_newseg(fs); 1809 repeat = 1; 1810 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1811 1812 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1813 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1814 1815 /* 1816 * If the segment contains a superblock, update the offset 1817 * and summary address to skip over it. 1818 */ 1819 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1820 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1821 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1822 sp->seg_bytes_left -= LFS_SBPAD; 1823 } 1824 brelse(bp, 0); 1825 /* Segment zero could also contain the labelpad */ 1826 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1827 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1828 lfs_sb_addoffset(fs, 1829 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1830 sp->seg_bytes_left -= 1831 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1832 } 1833 } else { 1834 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1835 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1836 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1837 } 1838 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1839 1840 /* Record first address of this partial segment */ 1841 if (sp->seg_flags & SEGM_CLEAN) { 1842 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1843 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1844 /* "1" is the artificial inc in lfs_seglock */ 1845 mutex_enter(&lfs_lock); 1846 while (fs->lfs_iocount > 1) { 1847 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1848 "lfs_initseg", 0, &lfs_lock); 1849 } 1850 mutex_exit(&lfs_lock); 1851 fs->lfs_cleanind = 0; 1852 } 1853 } 1854 1855 sp->fs = fs; 1856 sp->ibp = NULL; 1857 sp->idp = NULL; 1858 sp->ninodes = 0; 1859 sp->ndupino = 0; 1860 1861 sp->cbpp = sp->bpp; 1862 1863 /* Get a new buffer for SEGSUM */ 1864 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1865 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1866 1867 /* ... and enter it into the buffer list. */ 1868 *sp->cbpp = sbp; 1869 sp->cbpp++; 1870 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1871 1872 sp->start_bpp = sp->cbpp; 1873 1874 /* Set point to SEGSUM, initialize it. */ 1875 ssp = sp->segsum = sbp->b_data; 1876 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1877 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1878 lfs_ss_setnfinfo(fs, ssp, 0); 1879 lfs_ss_setninos(fs, ssp, 0); 1880 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1881 lfs_ss_setflags(fs, ssp, flags); 1882 1883 /* Set pointer to first FINFO, initialize it. */ 1884 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1885 lfs_fi_setnblocks(fs, sp->fip, 0); 1886 lfs_fi_setlastlength(fs, sp->fip, 0); 1887 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1888 1889 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1890 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1891 1892 return (repeat); 1893 } 1894 1895 /* 1896 * Remove SEGUSE_INVAL from all segments. 1897 */ 1898 void 1899 lfs_unset_inval_all(struct lfs *fs) 1900 { 1901 SEGUSE *sup; 1902 struct buf *bp; 1903 int i; 1904 1905 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1906 LFS_SEGENTRY(sup, fs, i, bp); 1907 if (sup->su_flags & SEGUSE_INVAL) { 1908 sup->su_flags &= ~SEGUSE_INVAL; 1909 LFS_WRITESEGENTRY(sup, fs, i, bp); 1910 } else 1911 brelse(bp, 0); 1912 } 1913 } 1914 1915 /* 1916 * Return the next segment to write. 1917 */ 1918 void 1919 lfs_newseg(struct lfs *fs) 1920 { 1921 CLEANERINFO *cip; 1922 SEGUSE *sup; 1923 struct buf *bp; 1924 int curseg, isdirty, sn, skip_inval; 1925 1926 ASSERT_SEGLOCK(fs); 1927 1928 /* Honor LFCNWRAPSTOP */ 1929 mutex_enter(&lfs_lock); 1930 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1931 if (fs->lfs_wrappass) { 1932 log(LOG_NOTICE, "%s: wrappass=%d\n", 1933 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1934 fs->lfs_wrappass = 0; 1935 break; 1936 } 1937 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1938 wakeup(&fs->lfs_nowrap); 1939 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1940 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1941 &lfs_lock); 1942 } 1943 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1944 mutex_exit(&lfs_lock); 1945 1946 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1947 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1948 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1949 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1950 sup->su_nbytes = 0; 1951 sup->su_nsums = 0; 1952 sup->su_ninos = 0; 1953 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1954 1955 LFS_CLEANERINFO(cip, fs, bp); 1956 lfs_ci_shiftcleantodirty(fs, cip, 1); 1957 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1958 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1959 1960 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1961 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1962 skip_inval = 1; 1963 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1964 sn = (sn + 1) % lfs_sb_getnseg(fs); 1965 1966 if (sn == curseg) { 1967 if (skip_inval) 1968 skip_inval = 0; 1969 else 1970 panic("lfs_nextseg: no clean segments"); 1971 } 1972 LFS_SEGENTRY(sup, fs, sn, bp); 1973 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1974 /* Check SEGUSE_EMPTY as we go along */ 1975 if (isdirty && sup->su_nbytes == 0 && 1976 !(sup->su_flags & SEGUSE_EMPTY)) 1977 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1978 else 1979 brelse(bp, 0); 1980 1981 if (!isdirty) 1982 break; 1983 } 1984 if (skip_inval == 0) 1985 lfs_unset_inval_all(fs); 1986 1987 ++fs->lfs_nactive; 1988 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1989 if (lfs_dostats) { 1990 ++lfs_stats.segsused; 1991 } 1992 } 1993 1994 static struct buf * 1995 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1996 int n) 1997 { 1998 struct lfs_cluster *cl; 1999 struct buf **bpp, *bp; 2000 2001 ASSERT_SEGLOCK(fs); 2002 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 2003 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 2004 memset(cl, 0, sizeof(*cl)); 2005 cl->fs = fs; 2006 cl->bpp = bpp; 2007 cl->bufcount = 0; 2008 cl->bufsize = 0; 2009 2010 /* If this segment is being written synchronously, note that */ 2011 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 2012 cl->flags |= LFS_CL_SYNC; 2013 cl->seg = fs->lfs_sp; 2014 ++cl->seg->seg_iocount; 2015 } 2016 2017 /* Get an empty buffer header, or maybe one with something on it */ 2018 bp = getiobuf(vp, true); 2019 bp->b_dev = NODEV; 2020 bp->b_blkno = bp->b_lblkno = addr; 2021 bp->b_iodone = lfs_cluster_aiodone; 2022 bp->b_private = cl; 2023 2024 return bp; 2025 } 2026 2027 int 2028 lfs_writeseg(struct lfs *fs, struct segment *sp) 2029 { 2030 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2031 SEGUSE *sup; 2032 SEGSUM *ssp; 2033 int i; 2034 int do_again, nblocks, byteoffset; 2035 size_t el_size; 2036 struct lfs_cluster *cl; 2037 u_short ninos; 2038 struct vnode *devvp; 2039 char *p = NULL; 2040 struct vnode *vp; 2041 unsigned ibindex, iblimit; 2042 int changed; 2043 u_int32_t sum; 2044 size_t sumstart; 2045 uint16_t oflags; 2046 #ifdef DEBUG 2047 FINFO *fip; 2048 int findex; 2049 #endif 2050 2051 ASSERT_SEGLOCK(fs); 2052 2053 ssp = (SEGSUM *)sp->segsum; 2054 2055 /* 2056 * If there are no buffers other than the segment summary to write, 2057 * don't do anything. If we are the end of a dirop sequence, however, 2058 * write the empty segment summary anyway, to help out the 2059 * roll-forward agent. 2060 */ 2061 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2062 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2063 return 0; 2064 } 2065 2066 /* Note if partial segment is being written by the cleaner */ 2067 if (sp->seg_flags & SEGM_CLEAN) 2068 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2069 2070 /* Note if we are writing to reclaim */ 2071 if (sp->seg_flags & SEGM_RECLAIM) { 2072 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2073 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2074 } 2075 /* Save flags to attach to a new partial segment, if we need one */ 2076 oflags = lfs_ss_getflags(fs, ssp); 2077 2078 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2079 2080 /* Update the segment usage information. */ 2081 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2082 2083 /* Loop through all blocks, except the segment summary. */ 2084 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2085 if ((*bpp)->b_vp != devvp) { 2086 sup->su_nbytes += (*bpp)->b_bcount; 2087 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2088 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2089 sp->seg_number, (*bpp)->b_bcount, 2090 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2091 (*bpp)->b_blkno)); 2092 } 2093 } 2094 2095 #ifdef DEBUG 2096 /* Check for zero-length and zero-version FINFO entries. */ 2097 fip = SEGSUM_FINFOBASE(fs, ssp); 2098 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2099 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2100 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2101 fip = NEXT_FINFO(fs, fip); 2102 } 2103 #endif /* DEBUG */ 2104 2105 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2106 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2107 sp->seg_number, 2108 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2109 lfs_ss_getninos(fs, ssp))); 2110 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2111 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2112 if (lfs_sb_getversion(fs) == 1) 2113 sup->su_olastmod = time_second; 2114 else 2115 sup->su_lastmod = time_second; 2116 sup->su_ninos += ninos; 2117 ++sup->su_nsums; 2118 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2119 2120 do_again = !(bp->b_flags & B_GATHERED); 2121 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2122 2123 /* 2124 * Mark blocks BC_BUSY, to prevent then from being changed between 2125 * the checksum computation and the actual write. 2126 * 2127 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2128 * there are any, replace them with copies that have UNASSIGNED 2129 * instead. 2130 */ 2131 mutex_enter(&bufcache_lock); 2132 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2133 ++bpp; 2134 bp = *bpp; 2135 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2136 bp->b_cflags |= BC_BUSY; 2137 continue; 2138 } 2139 2140 while (bp->b_cflags & BC_BUSY) { 2141 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2142 " data summary corruption for ino %d, lbn %" 2143 PRId64 "\n", 2144 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2145 bp->b_cflags |= BC_WANTED; 2146 cv_wait(&bp->b_busy, &bufcache_lock); 2147 } 2148 bp->b_cflags |= BC_BUSY; 2149 mutex_exit(&bufcache_lock); 2150 unbusybp = NULL; 2151 2152 /* 2153 * Check and replace indirect block UNWRITTEN bogosity. 2154 * XXX See comment in lfs_writefile. 2155 */ 2156 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2157 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) != 2158 VTOI(bp->b_vp)->i_lfs_effnblks) { 2159 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2160 VTOI(bp->b_vp)->i_number, 2161 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2162 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din))); 2163 /* Make a copy we'll make changes to */ 2164 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2165 bp->b_bcount, LFS_NB_IBLOCK); 2166 newbp->b_blkno = bp->b_blkno; 2167 memcpy(newbp->b_data, bp->b_data, 2168 newbp->b_bcount); 2169 2170 changed = 0; 2171 iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs); 2172 for (ibindex = 0; ibindex < iblimit; ibindex++) { 2173 if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) { 2174 ++changed; 2175 lfs_iblock_set(fs, newbp->b_data, 2176 ibindex, 0); 2177 } 2178 } 2179 /* 2180 * Get rid of the old buffer. Don't mark it clean, 2181 * though, if it still has dirty data on it. 2182 */ 2183 if (changed) { 2184 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2185 " bp = %p newbp = %p\n", changed, bp, 2186 newbp)); 2187 *bpp = newbp; 2188 bp->b_flags &= ~B_GATHERED; 2189 bp->b_error = 0; 2190 if (bp->b_iodone != NULL) { 2191 DLOG((DLOG_SEG, "lfs_writeseg: " 2192 "indir bp should not be B_CALL\n")); 2193 biodone(bp); 2194 bp = NULL; 2195 } else { 2196 /* Still on free list, leave it there */ 2197 unbusybp = bp; 2198 /* 2199 * We have to re-decrement lfs_avail 2200 * since this block is going to come 2201 * back around to us in the next 2202 * segment. 2203 */ 2204 lfs_sb_subavail(fs, 2205 lfs_btofsb(fs, bp->b_bcount)); 2206 } 2207 } else { 2208 lfs_freebuf(fs, newbp); 2209 } 2210 } 2211 mutex_enter(&bufcache_lock); 2212 if (unbusybp != NULL) { 2213 unbusybp->b_cflags &= ~BC_BUSY; 2214 if (unbusybp->b_cflags & BC_WANTED) 2215 cv_broadcast(&bp->b_busy); 2216 } 2217 } 2218 mutex_exit(&bufcache_lock); 2219 2220 /* 2221 * Compute checksum across data and then across summary; the first 2222 * block (the summary block) is skipped. Set the create time here 2223 * so that it's guaranteed to be later than the inode mod times. 2224 */ 2225 sum = 0; 2226 if (lfs_sb_getversion(fs) == 1) 2227 el_size = sizeof(u_long); 2228 else 2229 el_size = sizeof(u_int32_t); 2230 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2231 ++bpp; 2232 /* Loop through gop_write cluster blocks */ 2233 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2234 byteoffset += lfs_sb_getbsize(fs)) { 2235 #ifdef LFS_USE_BC_INVAL 2236 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2237 (*bpp)->b_iodone != NULL) { 2238 if (copyin((void *)(*bpp)->b_saveaddr + 2239 byteoffset, dp, el_size)) { 2240 panic("lfs_writeseg: copyin failed [1]:" 2241 " ino %" PRIu64 " blk %" PRId64, 2242 VTOI((*bpp)->b_vp)->i_number, 2243 (*bpp)->b_lblkno); 2244 } 2245 } else 2246 #endif /* LFS_USE_BC_INVAL */ 2247 { 2248 sum = lfs_cksum_part((char *) 2249 (*bpp)->b_data + byteoffset, el_size, sum); 2250 } 2251 } 2252 } 2253 if (lfs_sb_getversion(fs) == 1) 2254 lfs_ss_setocreate(fs, ssp, time_second); 2255 else { 2256 lfs_ss_setcreate(fs, ssp, time_second); 2257 lfs_sb_addserial(fs, 1); 2258 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2259 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2260 } 2261 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2262 sumstart = lfs_ss_getsumstart(fs); 2263 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2264 lfs_sb_getsumsize(fs) - sumstart)); 2265 2266 mutex_enter(&lfs_lock); 2267 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2268 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2269 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2270 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2271 mutex_exit(&lfs_lock); 2272 2273 /* 2274 * When we simply write the blocks we lose a rotation for every block 2275 * written. To avoid this problem, we cluster the buffers into a 2276 * chunk and write the chunk. MAXPHYS is the largest size I/O 2277 * devices can handle, use that for the size of the chunks. 2278 * 2279 * Blocks that are already clusters (from GOP_WRITE), however, we 2280 * don't bother to copy into other clusters. 2281 */ 2282 2283 #define CHUNKSIZE MAXPHYS 2284 2285 if (devvp == NULL) 2286 panic("devvp is NULL"); 2287 for (bpp = sp->bpp, i = nblocks; i;) { 2288 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2289 cl = cbp->b_private; 2290 2291 cbp->b_flags |= B_ASYNC; 2292 cbp->b_cflags |= BC_BUSY; 2293 cbp->b_bcount = 0; 2294 2295 KASSERTMSG((bpp - sp->bpp <= 2296 (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2297 / sizeof(int32_t)), 2298 "lfs_writeseg: real bpp overwrite"); 2299 KASSERTMSG((bpp - sp->bpp <= 2300 lfs_segsize(fs) / lfs_sb_getfsize(fs)), 2301 "lfs_writeseg: theoretical bpp overwrite"); 2302 2303 /* 2304 * Construct the cluster. 2305 */ 2306 mutex_enter(&lfs_lock); 2307 ++fs->lfs_iocount; 2308 mutex_exit(&lfs_lock); 2309 while (i && cbp->b_bcount < CHUNKSIZE) { 2310 bp = *bpp; 2311 2312 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2313 break; 2314 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2315 break; 2316 2317 /* Clusters from GOP_WRITE are expedited */ 2318 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2319 if (cbp->b_bcount > 0) 2320 /* Put in its own buffer */ 2321 break; 2322 else { 2323 cbp->b_data = bp->b_data; 2324 } 2325 } else if (cbp->b_bcount == 0) { 2326 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2327 LFS_NB_CLUSTER); 2328 cl->flags |= LFS_CL_MALLOC; 2329 } 2330 KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2331 btodb(bp->b_bcount - 1))) == 2332 sp->seg_number), 2333 "segment overwrite: blk size %d daddr %" PRIx64 2334 " not in seg %d\n", 2335 bp->b_bcount, bp->b_blkno, 2336 sp->seg_number); 2337 2338 #ifdef LFS_USE_BC_INVAL 2339 /* 2340 * Fake buffers from the cleaner are marked as BC_INVAL. 2341 * We need to copy the data from user space rather than 2342 * from the buffer indicated. 2343 * XXX == what do I do on an error? 2344 */ 2345 if ((bp->b_cflags & BC_INVAL) != 0 && 2346 bp->b_iodone != NULL) { 2347 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2348 panic("lfs_writeseg: " 2349 "copyin failed [2]"); 2350 } else 2351 #endif /* LFS_USE_BC_INVAL */ 2352 if (cl->flags & LFS_CL_MALLOC) { 2353 /* copy data into our cluster. */ 2354 memcpy(p, bp->b_data, bp->b_bcount); 2355 p += bp->b_bcount; 2356 } 2357 2358 cbp->b_bcount += bp->b_bcount; 2359 cl->bufsize += bp->b_bcount; 2360 2361 bp->b_flags &= ~B_READ; 2362 bp->b_error = 0; 2363 cl->bpp[cl->bufcount++] = bp; 2364 2365 vp = bp->b_vp; 2366 mutex_enter(&bufcache_lock); 2367 mutex_enter(vp->v_interlock); 2368 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2369 reassignbuf(bp, vp); 2370 vp->v_numoutput++; 2371 mutex_exit(vp->v_interlock); 2372 mutex_exit(&bufcache_lock); 2373 2374 bpp++; 2375 i--; 2376 } 2377 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2378 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2379 else 2380 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2381 mutex_enter(devvp->v_interlock); 2382 devvp->v_numoutput++; 2383 mutex_exit(devvp->v_interlock); 2384 VOP_STRATEGY(devvp, cbp); 2385 curlwp->l_ru.ru_oublock++; 2386 } 2387 2388 if (lfs_dostats) { 2389 ++lfs_stats.psegwrites; 2390 lfs_stats.blocktot += nblocks - 1; 2391 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2392 ++lfs_stats.psyncwrites; 2393 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2394 ++lfs_stats.pcleanwrites; 2395 lfs_stats.cleanblocks += nblocks - 1; 2396 } 2397 } 2398 2399 return (lfs_initseg(fs, oflags) || do_again); 2400 } 2401 2402 void 2403 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2404 { 2405 struct buf *bp; 2406 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2407 2408 ASSERT_MAYBE_SEGLOCK(fs); 2409 if (fs->lfs_is64) { 2410 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2411 } else { 2412 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2413 } 2414 /* 2415 * If we can write one superblock while another is in 2416 * progress, we risk not having a complete checkpoint if we crash. 2417 * So, block here if a superblock write is in progress. 2418 */ 2419 mutex_enter(&lfs_lock); 2420 while (fs->lfs_sbactive) { 2421 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2422 &lfs_lock); 2423 } 2424 fs->lfs_sbactive = daddr; 2425 mutex_exit(&lfs_lock); 2426 2427 /* Set timestamp of this version of the superblock */ 2428 if (lfs_sb_getversion(fs) == 1) 2429 lfs_sb_setotstamp(fs, time_second); 2430 lfs_sb_settstamp(fs, time_second); 2431 2432 /* The next chunk of code relies on this assumption */ 2433 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2434 2435 /* Checksum the superblock and copy it into a buffer. */ 2436 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2437 bp = lfs_newbuf(fs, devvp, 2438 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2439 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2440 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2441 LFS_SBPAD - sizeof(struct dlfs)); 2442 2443 bp->b_cflags |= BC_BUSY; 2444 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2445 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2446 bp->b_error = 0; 2447 bp->b_iodone = lfs_super_aiodone; 2448 2449 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2450 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2451 else 2452 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2453 curlwp->l_ru.ru_oublock++; 2454 2455 mutex_enter(devvp->v_interlock); 2456 devvp->v_numoutput++; 2457 mutex_exit(devvp->v_interlock); 2458 2459 mutex_enter(&lfs_lock); 2460 ++fs->lfs_iocount; 2461 mutex_exit(&lfs_lock); 2462 VOP_STRATEGY(devvp, bp); 2463 } 2464 2465 /* 2466 * Logical block number match routines used when traversing the dirty block 2467 * chain. 2468 */ 2469 int 2470 lfs_match_fake(struct lfs *fs, struct buf *bp) 2471 { 2472 2473 ASSERT_SEGLOCK(fs); 2474 return LFS_IS_MALLOC_BUF(bp); 2475 } 2476 2477 #if 0 2478 int 2479 lfs_match_real(struct lfs *fs, struct buf *bp) 2480 { 2481 2482 ASSERT_SEGLOCK(fs); 2483 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2484 } 2485 #endif 2486 2487 int 2488 lfs_match_data(struct lfs *fs, struct buf *bp) 2489 { 2490 2491 ASSERT_SEGLOCK(fs); 2492 return (bp->b_lblkno >= 0); 2493 } 2494 2495 int 2496 lfs_match_indir(struct lfs *fs, struct buf *bp) 2497 { 2498 daddr_t lbn; 2499 2500 ASSERT_SEGLOCK(fs); 2501 lbn = bp->b_lblkno; 2502 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2503 } 2504 2505 int 2506 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2507 { 2508 daddr_t lbn; 2509 2510 ASSERT_SEGLOCK(fs); 2511 lbn = bp->b_lblkno; 2512 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2513 } 2514 2515 int 2516 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2517 { 2518 daddr_t lbn; 2519 2520 ASSERT_SEGLOCK(fs); 2521 lbn = bp->b_lblkno; 2522 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2523 } 2524 2525 void 2526 lfs_free_aiodone(struct buf *bp) 2527 { 2528 struct lfs *fs; 2529 2530 fs = bp->b_private; 2531 ASSERT_NO_SEGLOCK(fs); 2532 lfs_freebuf(fs, bp); 2533 } 2534 2535 static void 2536 lfs_super_aiodone(struct buf *bp) 2537 { 2538 workqueue_enqueue(lfs_super_wq, (struct work *)bp, NULL); 2539 } 2540 2541 void 2542 lfs_super_work(struct work *wk, void *arg) 2543 { 2544 struct buf *bp = (struct buf *)wk; 2545 struct lfs *fs; 2546 2547 fs = bp->b_private; 2548 ASSERT_NO_SEGLOCK(fs); 2549 mutex_enter(&lfs_lock); 2550 fs->lfs_sbactive = 0; 2551 if (--fs->lfs_iocount <= 1) 2552 wakeup(&fs->lfs_iocount); 2553 wakeup(&fs->lfs_sbactive); 2554 mutex_exit(&lfs_lock); 2555 lfs_freebuf(fs, bp); 2556 } 2557 2558 static void 2559 lfs_cluster_aiodone(struct buf *bp) 2560 { 2561 workqueue_enqueue(lfs_cluster_wq, (struct work *)bp, NULL); 2562 } 2563 2564 void 2565 lfs_cluster_work(struct work *wk, void *arg) 2566 { 2567 struct buf *bp = (struct buf *)wk; 2568 struct lfs_cluster *cl; 2569 struct lfs *fs; 2570 struct buf *tbp, *fbp; 2571 struct vnode *vp, *devvp, *ovp; 2572 struct inode *ip; 2573 int error; 2574 2575 error = bp->b_error; 2576 cl = bp->b_private; 2577 fs = cl->fs; 2578 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2579 ASSERT_NO_SEGLOCK(fs); 2580 2581 /* Put the pages back, and release the buffer */ 2582 while (cl->bufcount--) { 2583 tbp = cl->bpp[cl->bufcount]; 2584 KASSERT(tbp->b_cflags & BC_BUSY); 2585 if (error) { 2586 tbp->b_error = error; 2587 } 2588 2589 /* 2590 * We're done with tbp. If it has not been re-dirtied since 2591 * the cluster was written, free it. Otherwise, keep it on 2592 * the locked list to be written again. 2593 */ 2594 vp = tbp->b_vp; 2595 2596 tbp->b_flags &= ~B_GATHERED; 2597 2598 #ifdef DEBUG 2599 if ((tbp)->b_vp == (fs)->lfs_ivnode) 2600 LFS_ENTER_LOG("clear", __FILE__, __LINE__, 2601 tbp->b_lblkno, tbp->b_flags, curproc->p_pid); 2602 #endif 2603 2604 mutex_enter(&bufcache_lock); 2605 if (tbp->b_iodone == NULL) { 2606 KASSERT(tbp->b_flags & B_LOCKED); 2607 bremfree(tbp); 2608 if (vp) { 2609 mutex_enter(vp->v_interlock); 2610 reassignbuf(tbp, vp); 2611 mutex_exit(vp->v_interlock); 2612 } 2613 tbp->b_flags |= B_ASYNC; /* for biodone */ 2614 } 2615 2616 if ((tbp->b_flags & B_LOCKED) && !(tbp->b_oflags & BO_DELWRI)) 2617 LFS_UNLOCK_BUF(tbp); 2618 2619 if (tbp->b_oflags & BO_DONE) { 2620 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2621 cl->bufcount, (long)tbp->b_flags)); 2622 } 2623 2624 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2625 /* 2626 * A buffer from the page daemon. 2627 * We use the same iodone as it does, 2628 * so we must manually disassociate its 2629 * buffers from the vp. 2630 */ 2631 if ((ovp = tbp->b_vp) != NULL) { 2632 /* This is just silly */ 2633 mutex_enter(ovp->v_interlock); 2634 brelvp(tbp); 2635 mutex_exit(ovp->v_interlock); 2636 tbp->b_vp = vp; 2637 tbp->b_objlock = vp->v_interlock; 2638 } 2639 /* Put it back the way it was */ 2640 tbp->b_flags |= B_ASYNC; 2641 /* Master buffers have BC_AGE */ 2642 if (tbp->b_private == tbp) 2643 tbp->b_cflags |= BC_AGE; 2644 } 2645 mutex_exit(&bufcache_lock); 2646 2647 biodone(tbp); 2648 2649 /* 2650 * If this is the last block for this vnode, but 2651 * there are other blocks on its dirty list, 2652 * set IN_MODIFIED/IN_CLEANING depending on what 2653 * sort of block. Only do this for our mount point, 2654 * not for, e.g., inode blocks that are attached to 2655 * the devvp. 2656 * XXX KS - Shouldn't we set *both* if both types 2657 * of blocks are present (traverse the dirty list?) 2658 */ 2659 mutex_enter(vp->v_interlock); 2660 mutex_enter(&lfs_lock); 2661 if (vp != devvp && vp->v_numoutput == 0 && 2662 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2663 ip = VTOI(vp); 2664 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2665 ip->i_number)); 2666 if (LFS_IS_MALLOC_BUF(fbp)) 2667 LFS_SET_UINO(ip, IN_CLEANING); 2668 else 2669 LFS_SET_UINO(ip, IN_MODIFIED); 2670 } 2671 cv_broadcast(&vp->v_cv); 2672 mutex_exit(&lfs_lock); 2673 mutex_exit(vp->v_interlock); 2674 } 2675 2676 /* Fix up the cluster buffer, and release it */ 2677 if (cl->flags & LFS_CL_MALLOC) 2678 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2679 putiobuf(bp); 2680 2681 /* Note i/o done */ 2682 if (cl->flags & LFS_CL_SYNC) { 2683 if (--cl->seg->seg_iocount == 0) 2684 wakeup(&cl->seg->seg_iocount); 2685 } 2686 mutex_enter(&lfs_lock); 2687 KASSERTMSG((fs->lfs_iocount != 0), 2688 "lfs_cluster_aiodone: zero iocount"); 2689 if (--fs->lfs_iocount <= 1) 2690 wakeup(&fs->lfs_iocount); 2691 mutex_exit(&lfs_lock); 2692 2693 pool_put(&fs->lfs_bpppool, cl->bpp); 2694 cl->bpp = NULL; 2695 pool_put(&fs->lfs_clpool, cl); 2696 } 2697 2698 /* 2699 * Shellsort (diminishing increment sort) from Data Structures and 2700 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2701 * see also Knuth Vol. 3, page 84. The increments are selected from 2702 * formula (8), page 95. Roughly O(N^3/2). 2703 */ 2704 /* 2705 * This is our own private copy of shellsort because we want to sort 2706 * two parallel arrays (the array of buffer pointers and the array of 2707 * logical block numbers) simultaneously. Note that we cast the array 2708 * of logical block numbers to a unsigned in this routine so that the 2709 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2710 */ 2711 2712 static void 2713 lfs_shellsort(struct lfs *fs, 2714 struct buf **bp_array, union lfs_blocks *lb_array, 2715 int nmemb, int size) 2716 { 2717 static int __rsshell_increments[] = { 4, 1, 0 }; 2718 int incr, *incrp, t1, t2; 2719 struct buf *bp_temp; 2720 2721 #ifdef DEBUG 2722 incr = 0; 2723 for (t1 = 0; t1 < nmemb; t1++) { 2724 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2725 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2726 /* dump before panic */ 2727 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2728 nmemb, size); 2729 incr = 0; 2730 for (t1 = 0; t1 < nmemb; t1++) { 2731 const struct buf *bp = bp_array[t1]; 2732 2733 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2734 PRIu64 "\n", t1, 2735 (uint64_t)bp->b_bcount, 2736 (uint64_t)bp->b_lblkno); 2737 printf("lbns:"); 2738 for (t2 = 0; t2 * size < bp->b_bcount; 2739 t2++) { 2740 printf(" %jd", 2741 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2742 } 2743 printf("\n"); 2744 } 2745 panic("lfs_shellsort: inconsistent input"); 2746 } 2747 } 2748 } 2749 #endif 2750 2751 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2752 for (t1 = incr; t1 < nmemb; ++t1) 2753 for (t2 = t1 - incr; t2 >= 0;) 2754 if ((u_int64_t)bp_array[t2]->b_lblkno > 2755 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2756 bp_temp = bp_array[t2]; 2757 bp_array[t2] = bp_array[t2 + incr]; 2758 bp_array[t2 + incr] = bp_temp; 2759 t2 -= incr; 2760 } else 2761 break; 2762 2763 /* Reform the list of logical blocks */ 2764 incr = 0; 2765 for (t1 = 0; t1 < nmemb; t1++) { 2766 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2767 lfs_blocks_set(fs, lb_array, incr++, 2768 bp_array[t1]->b_lblkno + t2); 2769 } 2770 } 2771 } 2772 2773 /* 2774 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2775 * point at uninitialized space. 2776 */ 2777 void 2778 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2779 { 2780 struct segment *sp = fs->lfs_sp; 2781 SEGSUM *ssp; 2782 2783 KASSERT(vers > 0); 2784 2785 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2786 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2787 (void) lfs_writeseg(fs, fs->lfs_sp); 2788 2789 sp->sum_bytes_left -= FINFOSIZE(fs); 2790 ssp = (SEGSUM *)sp->segsum; 2791 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2792 lfs_fi_setnblocks(fs, sp->fip, 0); 2793 lfs_fi_setino(fs, sp->fip, ino); 2794 lfs_fi_setversion(fs, sp->fip, vers); 2795 } 2796 2797 /* 2798 * Release the FINFO entry, either clearing out an unused entry or 2799 * advancing us to the next available entry. 2800 */ 2801 void 2802 lfs_release_finfo(struct lfs *fs) 2803 { 2804 struct segment *sp = fs->lfs_sp; 2805 SEGSUM *ssp; 2806 2807 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2808 sp->fip = NEXT_FINFO(fs, sp->fip); 2809 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2810 } else { 2811 /* XXX shouldn't this update sp->fip? */ 2812 sp->sum_bytes_left += FINFOSIZE(fs); 2813 ssp = (SEGSUM *)sp->segsum; 2814 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2815 } 2816 } 2817
Indexes created Tue Sep 23 02:09:52 GMT 2025