lfs_segment.c revision 1.291
1 /* $NetBSD: lfs_segment.c,v 1.291 2025/09/17 03:50:38 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 1.291 2025/09/17 03:50:38 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_wq; 151 152 /* Statistics Counters */ 153 int lfs_dostats = 1; 154 struct lfs_stats lfs_stats; 155 156 /* op values to lfs_writevnodes */ 157 #define VN_REG 0 158 #define VN_DIROP 1 159 #define VN_EMPTY 2 160 #define VN_CLEAN 3 161 162 /* 163 * XXX KS - Set modification time on the Ifile, so the cleaner can 164 * read the fs mod time off of it. We don't set IN_UPDATE here, 165 * since we don't really need this to be flushed to disk (and in any 166 * case that wouldn't happen to the Ifile until we checkpoint). 167 */ 168 void 169 lfs_imtime(struct lfs *fs) 170 { 171 struct timespec ts; 172 struct inode *ip; 173 174 ASSERT_MAYBE_SEGLOCK(fs); 175 vfs_timestamp(&ts); 176 ip = VTOI(fs->lfs_ivnode); 177 lfs_dino_setmtime(fs, ip->i_din, ts.tv_sec); 178 lfs_dino_setmtimensec(fs, ip->i_din, ts.tv_nsec); 179 } 180 181 /* 182 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 183 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 184 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 185 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 186 */ 187 188 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 189 190 int 191 lfs_vflush(struct vnode *vp) 192 { 193 struct inode *ip; 194 struct lfs *fs; 195 struct segment *sp; 196 struct buf *bp, *nbp, *tbp, *tnbp; 197 int error; 198 int flushed; 199 int relock; 200 201 ip = VTOI(vp); 202 fs = VFSTOULFS(vp->v_mount)->um_lfs; 203 relock = 0; 204 205 top: 206 KASSERT(mutex_owned(vp->v_interlock) == false); 207 KASSERT(mutex_owned(&lfs_lock) == false); 208 KASSERT(mutex_owned(&bufcache_lock) == false); 209 ASSERT_NO_SEGLOCK(fs); 210 if (ip->i_state & IN_CLEANING) { 211 ivndebug(vp,"vflush/in_cleaning"); 212 mutex_enter(&lfs_lock); 213 LFS_CLR_UINO(ip, IN_CLEANING); 214 LFS_SET_UINO(ip, IN_MODIFIED); 215 mutex_exit(&lfs_lock); 216 217 /* 218 * Toss any cleaning buffers that have real counterparts 219 * to avoid losing new data. 220 */ 221 mutex_enter(vp->v_interlock); 222 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 223 nbp = LIST_NEXT(bp, b_vnbufs); 224 if (!LFS_IS_MALLOC_BUF(bp)) 225 continue; 226 /* 227 * Look for pages matching the range covered 228 * by cleaning blocks. It's okay if more dirty 229 * pages appear, so long as none disappear out 230 * from under us. 231 */ 232 if (bp->b_lblkno > 0 && vp->v_type == VREG && 233 vp != fs->lfs_ivnode) { 234 struct vm_page *pg; 235 voff_t off; 236 237 for (off = lfs_lblktosize(fs, bp->b_lblkno); 238 off < lfs_lblktosize(fs, bp->b_lblkno + 1); 239 off += PAGE_SIZE) { 240 pg = uvm_pagelookup(&vp->v_uobj, off); 241 if (pg == NULL) 242 continue; 243 if (uvm_pagegetdirty(pg) 244 == UVM_PAGE_STATUS_DIRTY || 245 pmap_is_modified(pg)) { 246 lfs_sb_addavail(fs, 247 lfs_btofsb(fs, 248 bp->b_bcount)); 249 wakeup(&fs->lfs_availsleep); 250 mutex_exit(vp->v_interlock); 251 lfs_freebuf(fs, bp); 252 mutex_enter(vp->v_interlock); 253 bp = NULL; 254 break; 255 } 256 } 257 } 258 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 259 tbp = tnbp) 260 { 261 tnbp = LIST_NEXT(tbp, b_vnbufs); 262 if (tbp->b_vp == bp->b_vp 263 && tbp->b_lblkno == bp->b_lblkno 264 && tbp != bp) 265 { 266 lfs_sb_addavail(fs, lfs_btofsb(fs, 267 bp->b_bcount)); 268 wakeup(&fs->lfs_availsleep); 269 mutex_exit(vp->v_interlock); 270 lfs_freebuf(fs, bp); 271 mutex_enter(vp->v_interlock); 272 bp = NULL; 273 break; 274 } 275 } 276 } 277 } else { 278 mutex_enter(vp->v_interlock); 279 } 280 281 /* If the node is being written, wait until that is done */ 282 while (WRITEINPROG(vp)) { 283 ivndebug(vp,"vflush/writeinprog"); 284 cv_wait(&vp->v_cv, vp->v_interlock); 285 } 286 error = vdead_check(vp, VDEAD_NOWAIT); 287 mutex_exit(vp->v_interlock); 288 289 /* Protect against deadlock in vinvalbuf() */ 290 lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0)); 291 if (error != 0) { 292 fs->lfs_reclino = ip->i_number; 293 } 294 295 /* If we're supposed to flush a freed inode, just toss it */ 296 if (ip->i_lfs_iflags & LFSI_DELETED) { 297 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 298 ip->i_number)); 299 /* Drain v_numoutput */ 300 mutex_enter(vp->v_interlock); 301 while (vp->v_numoutput > 0) { 302 cv_wait(&vp->v_cv, vp->v_interlock); 303 } 304 KASSERT(vp->v_numoutput == 0); 305 mutex_exit(vp->v_interlock); 306 307 mutex_enter(&bufcache_lock); 308 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 309 nbp = LIST_NEXT(bp, b_vnbufs); 310 311 KASSERT((bp->b_flags & B_GATHERED) == 0); 312 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 313 lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount)); 314 wakeup(&fs->lfs_availsleep); 315 } 316 /* Copied from lfs_writeseg */ 317 if (bp->b_iodone != NULL) { 318 mutex_exit(&bufcache_lock); 319 biodone(bp); 320 mutex_enter(&bufcache_lock); 321 } else { 322 bremfree(bp); 323 LFS_UNLOCK_BUF(bp); 324 mutex_enter(vp->v_interlock); 325 bp->b_flags &= ~(B_READ | B_GATHERED); 326 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 327 bp->b_error = 0; 328 reassignbuf(bp, vp); 329 mutex_exit(vp->v_interlock); 330 brelse(bp, 0); 331 } 332 } 333 mutex_exit(&bufcache_lock); 334 LFS_CLR_UINO(ip, IN_CLEANING); 335 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 336 ip->i_state &= ~IN_ALLMOD; 337 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 338 ip->i_number)); 339 lfs_segunlock(fs); 340 341 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 342 343 return 0; 344 } 345 346 fs->lfs_flushvp = vp; 347 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 348 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 349 fs->lfs_flushvp = NULL; 350 KASSERT(fs->lfs_flushvp_fakevref == 0); 351 lfs_segunlock(fs); 352 353 /* Make sure that any pending buffers get written */ 354 mutex_enter(vp->v_interlock); 355 while (vp->v_numoutput > 0) { 356 cv_wait(&vp->v_cv, vp->v_interlock); 357 } 358 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 359 KASSERT(vp->v_numoutput == 0); 360 mutex_exit(vp->v_interlock); 361 362 return error; 363 } 364 sp = fs->lfs_sp; 365 366 flushed = 0; 367 if (VPISEMPTY(vp)) { 368 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 369 ++flushed; 370 } else if ((ip->i_state & IN_CLEANING) && 371 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 372 ivndebug(vp,"vflush/clean"); 373 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 374 ++flushed; 375 } else if (lfs_dostats) { 376 if (!VPISEMPTY(vp) || (VTOI(vp)->i_state & IN_ALLMOD)) 377 ++lfs_stats.vflush_invoked; 378 ivndebug(vp,"vflush"); 379 } 380 381 #ifdef DIAGNOSTIC 382 if (vp->v_uflag & VU_DIROP) { 383 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 384 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 385 } 386 #endif 387 388 do { 389 #ifdef DEBUG 390 int loopcount = 0; 391 #endif 392 do { 393 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 394 relock = lfs_writefile(fs, sp, vp); 395 if (relock && vp != fs->lfs_ivnode) { 396 /* 397 * Might have to wait for the 398 * cleaner to run; but we're 399 * still not done with this vnode. 400 * XXX we can do better than this. 401 */ 402 KASSERT(ip->i_number != LFS_IFILE_INUM); 403 lfs_writeinode(fs, sp, ip); 404 mutex_enter(&lfs_lock); 405 LFS_SET_UINO(ip, IN_MODIFIED); 406 mutex_exit(&lfs_lock); 407 lfs_writeseg(fs, sp); 408 lfs_segunlock(fs); 409 lfs_segunlock_relock(fs); 410 goto top; 411 } 412 } 413 /* 414 * If we begin a new segment in the middle of writing 415 * the Ifile, it creates an inconsistent checkpoint, 416 * since the Ifile information for the new segment 417 * is not up-to-date. Take care of this here by 418 * sending the Ifile through again in case there 419 * are newly dirtied blocks. But wait, there's more! 420 * This second Ifile write could *also* cross a segment 421 * boundary, if the first one was large. The second 422 * one is guaranteed to be no more than 8 blocks, 423 * though (two segment blocks and supporting indirects) 424 * so the third write *will not* cross the boundary. 425 */ 426 if (vp == fs->lfs_ivnode) { 427 lfs_writefile(fs, sp, vp); 428 lfs_writefile(fs, sp, vp); 429 } 430 #ifdef DEBUG 431 if (++loopcount > 2) 432 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 433 #endif 434 } while (lfs_writeinode(fs, sp, ip)); 435 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 436 437 if (lfs_dostats) { 438 ++lfs_stats.nwrites; 439 if (sp->seg_flags & SEGM_SYNC) 440 ++lfs_stats.nsync_writes; 441 if (sp->seg_flags & SEGM_CKP) 442 ++lfs_stats.ncheckpoints; 443 } 444 /* 445 * If we were called from somewhere that has already held the seglock 446 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 447 * the write to complete because we are still locked. 448 * Since lfs_vflush() must return the vnode with no dirty buffers, 449 * we must explicitly wait, if that is the case. 450 * 451 * We compare the iocount against 1, not 0, because it is 452 * artificially incremented by lfs_seglock(). 453 */ 454 mutex_enter(&lfs_lock); 455 if (fs->lfs_seglock > 1) { 456 while (fs->lfs_iocount > 1) 457 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 458 "lfs_vflush", 0, &lfs_lock); 459 } 460 mutex_exit(&lfs_lock); 461 462 lfs_segunlock(fs); 463 464 /* Wait for these buffers to be recovered by aiodoned */ 465 mutex_enter(vp->v_interlock); 466 while (vp->v_numoutput > 0) { 467 cv_wait(&vp->v_cv, vp->v_interlock); 468 } 469 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 470 KASSERT(vp->v_numoutput == 0); 471 mutex_exit(vp->v_interlock); 472 473 fs->lfs_flushvp = NULL; 474 KASSERT(fs->lfs_flushvp_fakevref == 0); 475 476 return (0); 477 } 478 479 struct lfs_writevnodes_ctx { 480 int op; 481 struct lfs *fs; 482 }; 483 static bool 484 lfs_writevnodes_selector(void *cl, struct vnode *vp) 485 { 486 struct lfs_writevnodes_ctx *c = cl; 487 struct inode *ip; 488 int op = c->op; 489 490 KASSERT(mutex_owned(vp->v_interlock)); 491 492 ip = VTOI(vp); 493 if (ip == NULL || vp->v_type == VNON || ip->i_nlink <= 0) 494 return false; 495 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 496 (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) { 497 vndebug(vp, "dirop"); 498 return false; 499 } 500 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 501 vndebug(vp,"empty"); 502 return false; 503 } 504 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM && 505 vp != c->fs->lfs_flushvp && !(ip->i_state & IN_CLEANING)) { 506 vndebug(vp,"cleaning"); 507 return false; 508 } 509 mutex_enter(&lfs_lock); 510 if (vp == c->fs->lfs_unlockvp) { 511 mutex_exit(&lfs_lock); 512 return false; 513 } 514 mutex_exit(&lfs_lock); 515 516 return true; 517 } 518 519 int 520 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 521 { 522 struct inode *ip; 523 struct vnode *vp; 524 struct vnode_iterator *marker; 525 struct lfs_writevnodes_ctx ctx; 526 int inodes_written = 0; 527 int error = 0; 528 529 /* 530 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist. 531 * XXX The rationale is unclear, the initial commit had no information. 532 * XXX If the order really matters we have to sort the vnodes first. 533 */ 534 535 ASSERT_SEGLOCK(fs); 536 vfs_vnode_iterator_init(mp, &marker); 537 ctx.op = op; 538 ctx.fs = fs; 539 while ((vp = vfs_vnode_iterator_next(marker, 540 lfs_writevnodes_selector, &ctx)) != NULL) { 541 ip = VTOI(vp); 542 543 /* 544 * Write the inode/file if dirty and it's not the IFILE. 545 */ 546 if (((ip->i_state & IN_ALLMOD) || !VPISEMPTY(vp)) && 547 ip->i_number != LFS_IFILE_INUM) { 548 error = lfs_writefile(fs, sp, vp); 549 if (error) { 550 vrele(vp); 551 if (error == EAGAIN) { 552 /* 553 * This error from lfs_putpages 554 * indicates we need to drop 555 * the segment lock and start 556 * over after the cleaner has 557 * had a chance to run. 558 */ 559 lfs_writeinode(fs, sp, ip); 560 lfs_writeseg(fs, sp); 561 if (!VPISEMPTY(vp) && 562 !WRITEINPROG(vp) && 563 !(ip->i_state & IN_ALLMOD)) { 564 mutex_enter(&lfs_lock); 565 LFS_SET_UINO(ip, IN_MODIFIED); 566 mutex_exit(&lfs_lock); 567 } 568 break; 569 } 570 error = 0; /* XXX not quite right */ 571 continue; 572 } 573 574 if (!VPISEMPTY(vp)) { 575 if (WRITEINPROG(vp)) { 576 ivndebug(vp,"writevnodes/write2"); 577 } else if (!(ip->i_state & IN_ALLMOD)) { 578 mutex_enter(&lfs_lock); 579 LFS_SET_UINO(ip, IN_MODIFIED); 580 mutex_exit(&lfs_lock); 581 } 582 } 583 (void) lfs_writeinode(fs, sp, ip); 584 inodes_written++; 585 } 586 vrele(vp); 587 } 588 vfs_vnode_iterator_destroy(marker); 589 return error; 590 } 591 592 /* 593 * Do a checkpoint. 594 */ 595 int 596 lfs_segwrite(struct mount *mp, int flags) 597 { 598 struct buf *bp; 599 struct inode *ip; 600 struct lfs *fs; 601 struct segment *sp; 602 struct vnode *vp; 603 SEGUSE *segusep; 604 int do_ckp, did_ckp, error; 605 unsigned n, segleft, maxseg, sn, i, curseg; 606 int writer_set = 0; 607 int dirty; 608 int redo; 609 SEGSUM *ssp; 610 int um_error; 611 612 fs = VFSTOULFS(mp)->um_lfs; 613 ASSERT_MAYBE_SEGLOCK(fs); 614 615 if (fs->lfs_ronly) 616 return EROFS; 617 618 lfs_imtime(fs); 619 620 /* 621 * Allocate a segment structure and enough space to hold pointers to 622 * the maximum possible number of buffers which can be described in a 623 * single summary block. 624 */ 625 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 626 627 /* 628 * If we know we're gonna need the writer lock, take it now to 629 * preserve the lock order lfs_writer -> lfs_seglock. 630 */ 631 if (do_ckp) { 632 lfs_writer_enter(fs, "ckpwriter"); 633 writer_set = 1; 634 } 635 636 /* We can't do a partial write and checkpoint at the same time. */ 637 if (do_ckp) 638 flags &= ~SEGM_SINGLE; 639 640 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 641 sp = fs->lfs_sp; 642 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 643 do_ckp = 1; 644 645 /* 646 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 647 * in which case we have to flush *all* buffers off of this vnode. 648 * We don't care about other nodes, but write any non-dirop nodes 649 * anyway in anticipation of another getnewvnode(). 650 * 651 * If we're cleaning we only write cleaning and ifile blocks, and 652 * no dirops, since otherwise we'd risk corruption in a crash. 653 */ 654 if (sp->seg_flags & SEGM_CLEAN) 655 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 656 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 657 do { 658 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 659 if ((sp->seg_flags & SEGM_SINGLE) && 660 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 661 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 662 break; 663 } 664 665 if (do_ckp || 666 (writer_set = lfs_writer_tryenter(fs)) != 0) { 667 KASSERT(writer_set); 668 KASSERT(fs->lfs_writer); 669 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 670 if (um_error == 0) 671 um_error = error; 672 /* 673 * In case writevnodes errored out 674 * XXX why are we always doing this and not 675 * just on error? 676 */ 677 lfs_flush_dirops(fs); 678 ssp = (SEGSUM *)(sp->segsum); 679 lfs_ss_setflags(fs, ssp, 680 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 681 lfs_finalize_fs_seguse(fs); 682 } 683 if (do_ckp && um_error) { 684 lfs_segunlock_relock(fs); 685 sp = fs->lfs_sp; 686 } 687 } while (do_ckp && um_error != 0); 688 } 689 690 /* 691 * If we are doing a checkpoint, mark everything since the 692 * last checkpoint as no longer ACTIVE. 693 */ 694 if (do_ckp || fs->lfs_doifile) { 695 segleft = lfs_sb_getnseg(fs); 696 curseg = 0; 697 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 698 int bread_error; 699 700 dirty = 0; 701 bread_error = bread(fs->lfs_ivnode, 702 lfs_sb_getcleansz(fs) + n, 703 lfs_sb_getbsize(fs), B_MODIFY, &bp); 704 if (bread_error) 705 panic("lfs_segwrite: ifile read: " 706 "seguse %u: error %d\n", 707 n, bread_error); 708 segusep = (SEGUSE *)bp->b_data; 709 maxseg = uimin(segleft, lfs_sb_getsepb(fs)); 710 for (i = 0; i < maxseg; i++) { 711 sn = curseg + i; 712 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 713 segusep->su_flags & SEGUSE_ACTIVE) { 714 segusep->su_flags &= ~SEGUSE_ACTIVE; 715 --fs->lfs_nactive; 716 ++dirty; 717 } 718 fs->lfs_suflags[fs->lfs_activesb][sn] = 719 segusep->su_flags; 720 if (lfs_sb_getversion(fs) > 1) 721 ++segusep; 722 else 723 segusep = (SEGUSE *) 724 ((SEGUSE_V1 *)segusep + 1); 725 } 726 727 if (dirty) 728 error = LFS_BWRITE_LOG(bp); /* Ifile */ 729 else 730 brelse(bp, 0); 731 segleft -= lfs_sb_getsepb(fs); 732 curseg += lfs_sb_getsepb(fs); 733 } 734 } 735 736 KASSERT(LFS_SEGLOCK_HELD(fs)); 737 738 did_ckp = 0; 739 if (do_ckp || fs->lfs_doifile) { 740 vp = fs->lfs_ivnode; 741 #ifdef DEBUG 742 int loopcount = 0; 743 #endif 744 do { 745 746 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 747 748 mutex_enter(&lfs_lock); 749 fs->lfs_flags &= ~LFS_IFDIRTY; 750 mutex_exit(&lfs_lock); 751 752 ip = VTOI(vp); 753 754 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 755 /* 756 * Ifile has no pages, so we don't need 757 * to check error return here. 758 */ 759 lfs_writefile(fs, sp, vp); 760 /* 761 * Ensure the Ifile takes the current segment 762 * into account. See comment in lfs_vflush. 763 */ 764 lfs_writefile(fs, sp, vp); 765 lfs_writefile(fs, sp, vp); 766 } 767 768 if (ip->i_state & IN_ALLMOD) 769 ++did_ckp; 770 #if 0 771 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 772 #else 773 redo = lfs_writeinode(fs, sp, ip); 774 #endif 775 redo += lfs_writeseg(fs, sp); 776 mutex_enter(&lfs_lock); 777 redo += (fs->lfs_flags & LFS_IFDIRTY); 778 mutex_exit(&lfs_lock); 779 #ifdef DEBUG 780 if (++loopcount > 2) 781 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 782 loopcount); 783 #endif 784 } while (redo && do_ckp); 785 786 /* 787 * Unless we are unmounting, the Ifile may continue to have 788 * dirty blocks even after a checkpoint, due to changes to 789 * inodes' atime. If we're checkpointing, it's "impossible" 790 * for other parts of the Ifile to be dirty after the loop 791 * above, since we hold the segment lock. 792 */ 793 mutex_enter(vp->v_interlock); 794 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 795 LFS_CLR_UINO(ip, IN_ALLMOD); 796 } 797 #ifdef DIAGNOSTIC 798 else if (do_ckp) { 799 int do_panic = 0; 800 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 801 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 802 lfs_sb_getsegtabsz(fs) && 803 !(bp->b_flags & B_GATHERED)) { 804 printf("ifile lbn %ld still dirty (flags %lx)\n", 805 (long)bp->b_lblkno, 806 (long)bp->b_flags); 807 ++do_panic; 808 } 809 } 810 if (do_panic) 811 panic("dirty blocks"); 812 } 813 #endif 814 mutex_exit(vp->v_interlock); 815 } else { 816 (void) lfs_writeseg(fs, sp); 817 } 818 819 /* Note Ifile no longer needs to be written */ 820 fs->lfs_doifile = 0; 821 if (writer_set) 822 lfs_writer_leave(fs); 823 824 /* 825 * If we didn't write the Ifile, we didn't really do anything. 826 * That means that (1) there is a checkpoint on disk and (2) 827 * nothing has changed since it was written. 828 * 829 * Take the flags off of the segment so that lfs_segunlock 830 * doesn't have to write the superblock either. 831 */ 832 if (do_ckp && !did_ckp) { 833 sp->seg_flags &= ~SEGM_CKP; 834 } 835 836 if (lfs_dostats) { 837 ++lfs_stats.nwrites; 838 if (sp->seg_flags & SEGM_SYNC) 839 ++lfs_stats.nsync_writes; 840 if (sp->seg_flags & SEGM_CKP) 841 ++lfs_stats.ncheckpoints; 842 } 843 lfs_segunlock(fs); 844 return (0); 845 } 846 847 /* 848 * Write the dirty blocks associated with a vnode. 849 */ 850 int 851 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 852 { 853 struct inode *ip; 854 int i, frag; 855 SEGSUM *ssp; 856 int error; 857 858 ASSERT_SEGLOCK(fs); 859 error = 0; 860 ip = VTOI(vp); 861 862 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 863 864 if (vp->v_uflag & VU_DIROP) { 865 ssp = (SEGSUM *)sp->segsum; 866 lfs_ss_setflags(fs, ssp, 867 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 868 } 869 870 if (sp->seg_flags & SEGM_CLEAN) { 871 lfs_gather(fs, sp, vp, lfs_match_fake); 872 /* 873 * For a file being flushed, we need to write *all* blocks. 874 * This means writing the cleaning blocks first, and then 875 * immediately following with any non-cleaning blocks. 876 * The same is true of the Ifile since checkpoints assume 877 * that all valid Ifile blocks are written. 878 */ 879 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 880 lfs_gather(fs, sp, vp, lfs_match_data); 881 /* 882 * Don't call VOP_PUTPAGES: if we're flushing, 883 * we've already done it, and the Ifile doesn't 884 * use the page cache. 885 */ 886 } 887 } else { 888 lfs_gather(fs, sp, vp, lfs_match_data); 889 /* 890 * If we're flushing, we've already called VOP_PUTPAGES 891 * so don't do it again. Otherwise, we want to write 892 * everything we've got. 893 */ 894 if (!IS_FLUSHING(fs, vp)) { 895 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER); 896 error = VOP_PUTPAGES(vp, 0, 0, 897 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 898 } 899 } 900 901 /* 902 * It may not be necessary to write the meta-data blocks at this point, 903 * as the roll-forward recovery code should be able to reconstruct the 904 * list. 905 * 906 * We have to write them anyway, though, under two conditions: (1) the 907 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 908 * checkpointing. 909 * 910 * BUT if we are cleaning, we might have indirect blocks that refer to 911 * new blocks not being written yet, in addition to fragments being 912 * moved out of a cleaned segment. If that is the case, don't 913 * write the indirect blocks, or the finfo will have a small block 914 * in the middle of it! 915 * XXX in this case isn't the inode size wrong too? 916 */ 917 frag = 0; 918 if (sp->seg_flags & SEGM_CLEAN) { 919 for (i = 0; i < ULFS_NDADDR; i++) 920 if (ip->i_lfs_fragsize[i] > 0 && 921 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 922 ++frag; 923 } 924 KASSERTMSG((frag <= 1), 925 "lfs_writefile: more than one fragment! frag=%d", frag); 926 if (IS_FLUSHING(fs, vp) || 927 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 928 lfs_gather(fs, sp, vp, lfs_match_indir); 929 lfs_gather(fs, sp, vp, lfs_match_dindir); 930 lfs_gather(fs, sp, vp, lfs_match_tindir); 931 } 932 lfs_release_finfo(fs); 933 934 return error; 935 } 936 937 /* 938 * Update segment accounting to reflect this inode's change of address. 939 */ 940 static int 941 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 942 { 943 struct buf *bp; 944 daddr_t daddr; 945 IFILE *ifp; 946 SEGUSE *sup; 947 ino_t ino; 948 int redo_ifile; 949 u_int32_t sn; 950 951 redo_ifile = 0; 952 953 /* 954 * If updating the ifile, update the super-block. Update the disk 955 * address and access times for this inode in the ifile. 956 */ 957 ino = ip->i_number; 958 if (ino == LFS_IFILE_INUM) { 959 daddr = lfs_sb_getidaddr(fs); 960 lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr)); 961 } else { 962 LFS_IENTRY(ifp, fs, ino, bp); 963 daddr = lfs_if_getdaddr(fs, ifp); 964 lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr)); 965 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 966 } 967 968 /* 969 * If this is the Ifile and lfs_offset is set to the first block 970 * in the segment, dirty the new segment's accounting block 971 * (XXX should already be dirty?) and tell the caller to do it again. 972 */ 973 if (ip->i_number == LFS_IFILE_INUM) { 974 sn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 975 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) == 976 lfs_sb_getoffset(fs)) { 977 LFS_SEGENTRY(sup, fs, sn, bp); 978 KASSERT(bp->b_oflags & BO_DELWRI); 979 LFS_WRITESEGENTRY(sup, fs, sn, bp); 980 /* fs->lfs_flags |= LFS_IFDIRTY; */ 981 redo_ifile |= 1; 982 } 983 } 984 985 /* 986 * The inode's last address should not be in the current partial 987 * segment, except under exceptional circumstances (lfs_writevnodes 988 * had to start over, and in the meantime more blocks were written 989 * to a vnode). Both inodes will be accounted to this segment 990 * in lfs_writeseg so we need to subtract the earlier version 991 * here anyway. The segment count can temporarily dip below 992 * zero here; keep track of how many duplicates we have in 993 * "dupino" so we don't panic below. 994 */ 995 if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) { 996 ++sp->ndupino; 997 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 998 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 999 (long long)daddr, sp->ndupino)); 1000 } 1001 /* 1002 * Account the inode: it no longer belongs to its former segment, 1003 * though it will not belong to the new segment until that segment 1004 * is actually written. 1005 */ 1006 if (!DADDR_IS_BAD(daddr)) { 1007 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1008 int ndupino __diagused = 1009 (sp->seg_number == oldsn) ? sp->ndupino : 0; 1010 LFS_SEGENTRY(sup, fs, oldsn, bp); 1011 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) 1012 >= DINOSIZE(fs)), 1013 "lfs_writeinode: negative bytes " 1014 "(segment %" PRIu32 " short by %d, " 1015 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1016 ", daddr=%" PRId64 ", su_nbytes=%u, " 1017 "ndupino=%d)\n", 1018 lfs_dtosn(fs, daddr), 1019 (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes, 1020 oldsn, sp->seg_number, daddr, 1021 (unsigned int)sup->su_nbytes, 1022 sp->ndupino); 1023 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1024 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1025 sup->su_nbytes -= DINOSIZE(fs); 1026 redo_ifile |= 1027 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1028 if (redo_ifile) { 1029 mutex_enter(&lfs_lock); 1030 fs->lfs_flags |= LFS_IFDIRTY; 1031 mutex_exit(&lfs_lock); 1032 /* Don't double-account */ 1033 lfs_sb_setidaddr(fs, 0x0); 1034 } 1035 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1036 } 1037 1038 return redo_ifile; 1039 } 1040 1041 int 1042 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1043 { 1044 struct buf *bp; 1045 union lfs_dinode *cdp; 1046 struct vnode *vp = ITOV(ip); 1047 daddr_t daddr; 1048 IINFO *iip; 1049 int i; 1050 int redo_ifile = 0; 1051 int gotblk = 0; 1052 int count; 1053 SEGSUM *ssp; 1054 1055 ASSERT_SEGLOCK(fs); 1056 if (!(ip->i_state & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1057 return (0); 1058 1059 /* Can't write ifile when writer is not set */ 1060 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1061 (sp->seg_flags & SEGM_CLEAN)); 1062 1063 /* 1064 * If this is the Ifile, see if writing it here will generate a 1065 * temporary misaccounting. If it will, do the accounting and write 1066 * the blocks, postponing the inode write until the accounting is 1067 * solid. 1068 */ 1069 count = 0; 1070 while (vp == fs->lfs_ivnode) { 1071 int redo = 0; 1072 1073 if (sp->idp == NULL && sp->ibp == NULL && 1074 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1075 sp->sum_bytes_left < sizeof(int32_t))) { 1076 (void) lfs_writeseg(fs, sp); 1077 continue; 1078 } 1079 1080 /* Look for dirty Ifile blocks */ 1081 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1082 if (!(bp->b_flags & B_GATHERED)) { 1083 redo = 1; 1084 break; 1085 } 1086 } 1087 1088 if (redo == 0) 1089 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1090 if (redo == 0) 1091 break; 1092 1093 if (sp->idp) { 1094 lfs_dino_setinumber(fs, sp->idp, 0); 1095 sp->idp = NULL; 1096 } 1097 ++count; 1098 if (count > 2) 1099 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1100 lfs_writefile(fs, sp, fs->lfs_ivnode); 1101 } 1102 1103 /* Allocate a new inode block if necessary. */ 1104 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1105 sp->ibp == NULL) { 1106 /* Allocate a new segment if necessary. */ 1107 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1108 sp->sum_bytes_left < sizeof(int32_t)) 1109 (void) lfs_writeseg(fs, sp); 1110 1111 /* Get next inode block. */ 1112 daddr = lfs_sb_getoffset(fs); 1113 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1114 sp->ibp = *sp->cbpp++ = 1115 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1116 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1117 gotblk++; 1118 1119 /* Zero out inode numbers */ 1120 for (i = 0; i < LFS_INOPB(fs); ++i) { 1121 union lfs_dinode *tmpdi; 1122 1123 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1124 DINOSIZE(fs) * i); 1125 lfs_dino_setinumber(fs, tmpdi, 0); 1126 } 1127 1128 ++sp->start_bpp; 1129 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1130 /* Set remaining space counters. */ 1131 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1132 sp->sum_bytes_left -= sizeof(int32_t); 1133 1134 /* Store the address in the segment summary. */ 1135 iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs)); 1136 lfs_ii_setblock(fs, iip, daddr); 1137 } 1138 1139 /* Check VU_DIROP in case there is a new file with no data blocks */ 1140 if (vp->v_uflag & VU_DIROP) { 1141 ssp = (SEGSUM *)sp->segsum; 1142 lfs_ss_setflags(fs, ssp, 1143 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1144 } 1145 1146 /* Update the inode times and copy the inode onto the inode page. */ 1147 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1148 if (ip->i_number != LFS_IFILE_INUM) 1149 LFS_ITIMES(ip, NULL, NULL, NULL); 1150 1151 /* 1152 * If this is the Ifile, and we've already written the Ifile in this 1153 * partial segment, just overwrite it (it's not on disk yet) and 1154 * continue. 1155 * 1156 * XXX we know that the bp that we get the second time around has 1157 * already been gathered. 1158 */ 1159 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1160 lfs_copy_dinode(fs, sp->idp, ip->i_din); 1161 ip->i_lfs_osize = ip->i_size; 1162 return 0; 1163 } 1164 1165 bp = sp->ibp; 1166 cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs)); 1167 lfs_copy_dinode(fs, cdp, ip->i_din); 1168 1169 /* 1170 * This inode is on its way to disk; clear its VU_DIROP status when 1171 * the write is complete. 1172 */ 1173 if (vp->v_uflag & VU_DIROP) { 1174 if (!(sp->seg_flags & SEGM_CLEAN)) 1175 ip->i_state |= IN_CDIROP; 1176 else { 1177 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1178 } 1179 } 1180 1181 /* 1182 * If cleaning, link counts and directory file sizes cannot change, 1183 * since those would be directory operations---even if the file 1184 * we are writing is marked VU_DIROP we should write the old values. 1185 * If we're not cleaning, of course, update the values so we get 1186 * current values the next time we clean. 1187 */ 1188 if (sp->seg_flags & SEGM_CLEAN) { 1189 if (vp->v_uflag & VU_DIROP) { 1190 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1191 /* if (vp->v_type == VDIR) */ 1192 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1193 } 1194 } else { 1195 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1196 ip->i_lfs_osize = ip->i_size; 1197 } 1198 1199 1200 /* We can finish the segment accounting for truncations now */ 1201 lfs_finalize_ino_seguse(fs, ip); 1202 1203 /* 1204 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1205 * addresses to disk; possibly change the on-disk record of 1206 * the inode size, either by reverting to the previous size 1207 * (in the case of cleaning) or by verifying the inode's block 1208 * holdings (in the case of files being allocated as they are being 1209 * written). 1210 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1211 * XXX count on disk wrong by the same amount. We should be 1212 * XXX able to "borrow" from lfs_avail and return it after the 1213 * XXX Ifile is written. See also in lfs_writeseg. 1214 */ 1215 1216 /* Check file size based on highest allocated block */ 1217 if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG || 1218 (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) && 1219 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1220 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1221 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1222 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1223 } 1224 if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) { 1225 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1226 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1227 lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs))); 1228 for (i=0; i<ULFS_NDADDR; i++) { 1229 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1230 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1231 lfs_dino_setdb(fs, cdp, i, 0); 1232 } 1233 } 1234 for (i=0; i<ULFS_NIADDR; i++) { 1235 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1236 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1237 lfs_dino_setib(fs, cdp, i, 0); 1238 } 1239 } 1240 } 1241 1242 #ifdef DIAGNOSTIC 1243 /* 1244 * Check dinode held blocks against dinode size. 1245 * This should be identical to the check in lfs_vget(). 1246 */ 1247 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1248 i < ULFS_NDADDR; i++) { 1249 KASSERT(i >= 0); 1250 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1251 continue; 1252 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1253 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1254 continue; 1255 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1256 # ifdef DEBUG 1257 lfs_dump_dinode(fs, cdp); 1258 # endif 1259 panic("writing inconsistent inode"); 1260 } 1261 } 1262 #endif /* DIAGNOSTIC */ 1263 1264 if (ip->i_state & IN_CLEANING) 1265 LFS_CLR_UINO(ip, IN_CLEANING); 1266 else { 1267 /* XXX IN_ALLMOD */ 1268 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1269 IN_UPDATE | IN_MODIFY); 1270 if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din)) 1271 LFS_CLR_UINO(ip, IN_MODIFIED); 1272 else { 1273 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1274 "blks=%d, eff=%jd\n", ip->i_number, 1275 lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks)); 1276 } 1277 } 1278 1279 if (ip->i_number == LFS_IFILE_INUM) { 1280 /* We know sp->idp == NULL */ 1281 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1282 1283 /* Not dirty any more */ 1284 mutex_enter(&lfs_lock); 1285 fs->lfs_flags &= ~LFS_IFDIRTY; 1286 mutex_exit(&lfs_lock); 1287 } 1288 1289 if (gotblk) { 1290 mutex_enter(&bufcache_lock); 1291 LFS_LOCK_BUF(bp); 1292 brelsel(bp, 0); 1293 mutex_exit(&bufcache_lock); 1294 } 1295 1296 /* Increment inode count in segment summary block. */ 1297 1298 ssp = (SEGSUM *)sp->segsum; 1299 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1300 1301 /* If this page is full, set flag to allocate a new page. */ 1302 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1303 sp->ibp = NULL; 1304 1305 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1306 1307 KASSERT(redo_ifile == 0); 1308 return (redo_ifile); 1309 } 1310 1311 int 1312 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1313 { 1314 struct lfs *fs; 1315 int vers; 1316 int j, blksinblk; 1317 1318 ASSERT_SEGLOCK(sp->fs); 1319 KASSERTMSG((sp->vp != NULL), 1320 "lfs_gatherblock: Null vp in segment"); 1321 1322 /* If full, finish this segment. */ 1323 fs = sp->fs; 1324 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1325 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1326 sp->seg_bytes_left < bp->b_bcount) { 1327 if (mptr) 1328 mutex_exit(mptr); 1329 lfs_updatemeta(sp); 1330 1331 vers = lfs_fi_getversion(fs, sp->fip); 1332 (void) lfs_writeseg(fs, sp); 1333 1334 /* Add the current file to the segment summary. */ 1335 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1336 1337 if (mptr) 1338 mutex_enter(mptr); 1339 return (1); 1340 } 1341 1342 if (bp->b_flags & B_GATHERED) { 1343 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1344 " lbn %" PRId64 "\n", 1345 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1346 return (0); 1347 } 1348 1349 /* Insert into the buffer list, update the FINFO block. */ 1350 bp->b_flags |= B_GATHERED; 1351 1352 *sp->cbpp++ = bp; 1353 for (j = 0; j < blksinblk; j++) { 1354 unsigned bn; 1355 1356 bn = lfs_fi_getnblocks(fs, sp->fip); 1357 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1358 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1359 /* This block's accounting moves from lfs_favail to lfs_avail */ 1360 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1361 } 1362 1363 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1364 sp->seg_bytes_left -= bp->b_bcount; 1365 return (0); 1366 } 1367 1368 int 1369 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1370 int (*match)(struct lfs *, struct buf *)) 1371 { 1372 struct buf *bp, *nbp; 1373 int count = 0; 1374 1375 ASSERT_SEGLOCK(fs); 1376 if (vp->v_type == VBLK) 1377 return 0; 1378 KASSERT(sp->vp == NULL); 1379 sp->vp = vp; 1380 mutex_enter(&bufcache_lock); 1381 1382 #ifndef LFS_NO_BACKBUF_HACK 1383 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1384 # define BUF_OFFSET \ 1385 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1386 # define BACK_BUF(BP) \ 1387 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1388 # define BEG_OF_LIST \ 1389 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1390 1391 loop: 1392 /* Find last buffer. */ 1393 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1394 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1395 bp = LIST_NEXT(bp, b_vnbufs)) 1396 continue; 1397 1398 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1399 nbp = BACK_BUF(bp); 1400 #else /* LFS_NO_BACKBUF_HACK */ 1401 loop: 1402 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1403 nbp = LIST_NEXT(bp, b_vnbufs); 1404 #endif /* LFS_NO_BACKBUF_HACK */ 1405 if ((bp->b_cflags & BC_BUSY) != 0 || 1406 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1407 #ifdef DEBUG 1408 if (vp == fs->lfs_ivnode && 1409 (bp->b_cflags & BC_BUSY) != 0 && 1410 (bp->b_flags & B_GATHERED) == 0) 1411 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1412 PRId64 " busy (%x) at 0x%jx", 1413 bp->b_lblkno, bp->b_flags, 1414 (uintmax_t)lfs_sb_getoffset(fs)); 1415 #endif 1416 continue; 1417 } 1418 #ifdef DIAGNOSTIC 1419 # ifdef LFS_USE_BC_INVAL 1420 if ((bp->b_cflags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1421 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1422 " is BC_INVAL\n", bp->b_lblkno)); 1423 VOP_PRINT(bp->b_vp); 1424 } 1425 # endif /* LFS_USE_BC_INVAL */ 1426 if (!(bp->b_oflags & BO_DELWRI)) 1427 panic("lfs_gather: bp not BO_DELWRI"); 1428 if (!(bp->b_flags & B_LOCKED)) { 1429 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1430 " blk %" PRId64 " not B_LOCKED\n", 1431 bp->b_lblkno, 1432 LFS_DBTOFSB(fs, bp->b_blkno))); 1433 VOP_PRINT(bp->b_vp); 1434 panic("lfs_gather: bp not B_LOCKED"); 1435 } 1436 #endif 1437 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1438 goto loop; 1439 } 1440 count++; 1441 } 1442 mutex_exit(&bufcache_lock); 1443 lfs_updatemeta(sp); 1444 KASSERT(sp->vp == vp); 1445 sp->vp = NULL; 1446 return count; 1447 } 1448 1449 #if DEBUG 1450 # define DEBUG_OOFF(n) do { \ 1451 if (ooff == 0) { \ 1452 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1453 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1454 ", was 0x0 (or %" PRId64 ")\n", \ 1455 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1456 } \ 1457 } while (0) 1458 #else 1459 # define DEBUG_OOFF(n) 1460 #endif 1461 1462 /* 1463 * Change the given block's address to ndaddr, finding its previous 1464 * location using ulfs_bmaparray(). 1465 * 1466 * Account for this change in the segment table. 1467 * 1468 * called with sp == NULL by roll-forwarding code. 1469 */ 1470 #define NOT_ON_DISK(daddr) ((daddr) == 0 || (daddr) == UNASSIGNED || (daddr) == UNWRITTEN) 1471 1472 void 1473 lfs_update_single(struct lfs *fs, struct segment *sp, 1474 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1475 { 1476 SEGUSE *sup; 1477 struct buf *bp; 1478 struct indir a[ULFS_NIADDR + 2], *ap; 1479 struct inode *ip; 1480 daddr_t daddr, ooff; 1481 int num, error; 1482 int bb, osize, obb; 1483 1484 ASSERT_SEGLOCK(fs); 1485 KASSERT(sp == NULL || sp->vp == vp); 1486 ip = VTOI(vp); 1487 1488 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1489 if (error) 1490 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1491 1492 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1493 if (daddr > 0) 1494 daddr = LFS_DBTOFSB(fs, daddr); 1495 1496 bb = lfs_numfrags(fs, size); 1497 switch (num) { 1498 case 0: 1499 ooff = lfs_dino_getdb(fs, ip->i_din, lbn); 1500 DEBUG_OOFF(0); 1501 if (NOT_ON_DISK(ooff)) 1502 lfs_dino_setblocks(fs, ip->i_din, 1503 lfs_dino_getblocks(fs, ip->i_din) + bb); 1504 else { 1505 /* possible fragment truncation or extension */ 1506 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1507 lfs_dino_setblocks(fs, ip->i_din, 1508 lfs_dino_getblocks(fs, ip->i_din) + (bb-obb)); 1509 } 1510 lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr); 1511 break; 1512 case 1: 1513 ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off); 1514 DEBUG_OOFF(1); 1515 if (NOT_ON_DISK(ooff)) 1516 lfs_dino_setblocks(fs, ip->i_din, 1517 lfs_dino_getblocks(fs, ip->i_din) + bb); 1518 lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr); 1519 break; 1520 default: 1521 ap = &a[num - 1]; 1522 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1523 B_MODIFY, &bp)) 1524 panic("lfs_updatemeta: bread bno %" PRId64, 1525 ap->in_lbn); 1526 1527 ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off); 1528 DEBUG_OOFF(num); 1529 if (NOT_ON_DISK(ooff)) 1530 lfs_dino_setblocks(fs, ip->i_din, 1531 lfs_dino_getblocks(fs, ip->i_din) + bb); 1532 lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr); 1533 (void) VOP_BWRITE(bp->b_vp, bp); 1534 } 1535 1536 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1537 1538 /* Update hiblk when extending the file */ 1539 if (lbn > ip->i_lfs_hiblk) 1540 ip->i_lfs_hiblk = lbn; 1541 1542 /* 1543 * Though we'd rather it couldn't, this *can* happen right now 1544 * if cleaning blocks and regular blocks coexist. 1545 */ 1546 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1547 1548 /* 1549 * Update segment usage information, based on old size 1550 * and location. 1551 */ 1552 if (daddr > 0) { 1553 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1554 int ndupino __diagused = (sp && sp->seg_number == oldsn ? 1555 sp->ndupino : 0); 1556 1557 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1558 if (lbn >= 0 && lbn < ULFS_NDADDR) 1559 osize = ip->i_lfs_fragsize[lbn]; 1560 else 1561 osize = lfs_sb_getbsize(fs); 1562 LFS_SEGENTRY(sup, fs, oldsn, bp); 1563 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize), 1564 "lfs_updatemeta: negative bytes " 1565 "(segment %" PRIu32 " short by %" PRId64 1566 ")\n" 1567 "lfs_updatemeta: ino %llu, lbn %" PRId64 1568 ", addr = 0x%" PRIx64 "\n" 1569 "lfs_updatemeta: ndupino=%d", 1570 lfs_dtosn(fs, daddr), 1571 (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes), 1572 (unsigned long long)ip->i_number, lbn, daddr, 1573 ndupino); 1574 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1575 " db 0x%" PRIx64 "\n", 1576 lfs_dtosn(fs, daddr), osize, 1577 ip->i_number, lbn, daddr)); 1578 sup->su_nbytes -= osize; 1579 if (!(bp->b_flags & B_GATHERED)) { 1580 mutex_enter(&lfs_lock); 1581 fs->lfs_flags |= LFS_IFDIRTY; 1582 mutex_exit(&lfs_lock); 1583 } 1584 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1585 } 1586 /* 1587 * Now that this block has a new address, and its old 1588 * segment no longer owns it, we can forget about its 1589 * old size. 1590 */ 1591 if (lbn >= 0 && lbn < ULFS_NDADDR) 1592 ip->i_lfs_fragsize[lbn] = size; 1593 } 1594 1595 /* 1596 * Update the metadata that points to the blocks listed in the FINFO 1597 * array. 1598 */ 1599 void 1600 lfs_updatemeta(struct segment *sp) 1601 { 1602 struct buf *sbp; 1603 struct lfs *fs; 1604 struct vnode *vp; 1605 daddr_t lbn; 1606 int i, nblocks, num; 1607 int __diagused nblocks_orig; 1608 int bb; 1609 int bytesleft, size; 1610 unsigned lastlength; 1611 union lfs_blocks tmpptr; 1612 1613 fs = sp->fs; 1614 vp = sp->vp; 1615 ASSERT_SEGLOCK(fs); 1616 1617 /* 1618 * This used to be: 1619 * 1620 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1621 * 1622 * that is, it allowed for the possibility that start_lbp did 1623 * not point to the beginning of the finfo block pointer area. 1624 * This particular formulation is six kinds of painful in the 1625 * lfs64 world where we have two sizes of block pointer, so 1626 * unless/until everything can be cleaned up to not move 1627 * start_lbp around but instead use an offset, we do the 1628 * following: 1629 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1630 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1631 * type. (Ugh.) 1632 * 2. Cast it to void *, then assign it to a temporary 1633 * union lfs_blocks. 1634 * 3. Subtract start_lbp from that. 1635 * 4. Save the value of nblocks in blocks_orig so we can 1636 * assert below that it hasn't changed without repeating this 1637 * rubbish. 1638 * 1639 * XXX. 1640 */ 1641 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1642 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1643 nblocks_orig = nblocks; 1644 1645 KASSERT(nblocks >= 0); 1646 KASSERT(vp != NULL); 1647 if (nblocks == 0) 1648 return; 1649 1650 /* 1651 * This count may be high due to oversize blocks from lfs_gop_write. 1652 * Correct for this. (XXX we should be able to keep track of these.) 1653 */ 1654 for (i = 0; i < nblocks; i++) { 1655 if (sp->start_bpp[i] == NULL) { 1656 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1657 nblocks = i; 1658 break; 1659 } 1660 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1661 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1662 nblocks -= num - 1; 1663 } 1664 1665 #if 0 1666 /* pre-lfs64 assertion */ 1667 KASSERT(vp->v_type == VREG || 1668 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1669 #else 1670 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1671 #endif 1672 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1673 1674 /* 1675 * Sort the blocks. 1676 * 1677 * We have to sort even if the blocks come from the 1678 * cleaner, because there might be other pending blocks on the 1679 * same inode...and if we don't sort, and there are fragments 1680 * present, blocks may be written in the wrong place. 1681 */ 1682 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1683 1684 /* 1685 * Record the length of the last block in case it's a fragment. 1686 * If there are indirect blocks present, they sort last. An 1687 * indirect block will be lfs_bsize and its presence indicates 1688 * that you cannot have fragments. 1689 * 1690 * XXX This last is a lie. A cleaned fragment can coexist with 1691 * XXX a later indirect block. This will continue to be 1692 * XXX true until lfs_markv is fixed to do everything with 1693 * XXX fake blocks (including fake inodes and fake indirect blocks). 1694 */ 1695 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1696 lfs_sb_getbmask(fs)) + 1; 1697 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1698 1699 /* 1700 * Assign disk addresses, and update references to the logical 1701 * block and the segment usage information. 1702 */ 1703 for (i = nblocks; i--; ++sp->start_bpp) { 1704 sbp = *sp->start_bpp; 1705 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1706 KASSERT(sbp->b_lblkno == lbn); 1707 1708 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1709 1710 /* 1711 * If we write a frag in the wrong place, the cleaner won't 1712 * be able to correctly identify its size later, and the 1713 * segment will be uncleanable. (Even worse, it will assume 1714 * that the indirect block that actually ends the list 1715 * is of a smaller size!) 1716 */ 1717 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1718 panic("lfs_updatemeta: fragment is not last block"); 1719 1720 /* 1721 * For each subblock in this possibly oversized block, 1722 * update its address on disk. 1723 */ 1724 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1725 KASSERT(vp == sbp->b_vp); 1726 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1727 bytesleft -= lfs_sb_getbsize(fs)) { 1728 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1729 bb = lfs_numfrags(fs, size); 1730 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1731 lfs_blocks_inc(fs, &sp->start_lbp); 1732 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1733 size); 1734 lfs_sb_addoffset(fs, bb); 1735 } 1736 1737 } 1738 1739 /* This inode has been modified */ 1740 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1741 } 1742 1743 /* 1744 * Move lfs_offset to a segment earlier than newsn. 1745 */ 1746 int 1747 lfs_rewind(struct lfs *fs, int newsn) 1748 { 1749 int sn, osn, isdirty; 1750 struct buf *bp; 1751 SEGUSE *sup; 1752 1753 ASSERT_SEGLOCK(fs); 1754 1755 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1756 if (osn < newsn) 1757 return 0; 1758 1759 /* lfs_avail eats the remaining space in this segment */ 1760 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1761 1762 /* Find a low-numbered segment */ 1763 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1764 LFS_SEGENTRY(sup, fs, sn, bp); 1765 isdirty = sup->su_flags & SEGUSE_DIRTY; 1766 brelse(bp, 0); 1767 1768 if (!isdirty) 1769 break; 1770 } 1771 if (sn == lfs_sb_getnseg(fs)) 1772 panic("lfs_rewind: no clean segments"); 1773 if (newsn >= 0 && sn >= newsn) 1774 return ENOENT; 1775 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1776 lfs_newseg(fs); 1777 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1778 1779 return 0; 1780 } 1781 1782 /* 1783 * Start a new partial segment. 1784 * 1785 * Return 1 when we entered to a new segment. 1786 * Otherwise, return 0. 1787 */ 1788 int 1789 lfs_initseg(struct lfs *fs, uint16_t flags) 1790 { 1791 struct segment *sp = fs->lfs_sp; 1792 SEGSUM *ssp; 1793 struct buf *sbp; /* buffer for SEGSUM */ 1794 int repeat = 0; /* return value */ 1795 1796 ASSERT_SEGLOCK(fs); 1797 /* Advance to the next segment. */ 1798 if (!LFS_PARTIAL_FITS(fs)) { 1799 SEGUSE *sup; 1800 struct buf *bp; 1801 1802 /* lfs_avail eats the remaining space */ 1803 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1804 lfs_sb_getcurseg(fs))); 1805 /* Wake up any cleaning procs waiting on this file system. */ 1806 lfs_wakeup_cleaner(fs); 1807 lfs_newseg(fs); 1808 repeat = 1; 1809 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1810 1811 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1812 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1813 1814 /* 1815 * If the segment contains a superblock, update the offset 1816 * and summary address to skip over it. 1817 */ 1818 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1819 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1820 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1821 sp->seg_bytes_left -= LFS_SBPAD; 1822 } 1823 brelse(bp, 0); 1824 /* Segment zero could also contain the labelpad */ 1825 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1826 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1827 lfs_sb_addoffset(fs, 1828 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1829 sp->seg_bytes_left -= 1830 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1831 } 1832 } else { 1833 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1834 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1835 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1836 } 1837 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1838 1839 /* Record first address of this partial segment */ 1840 if (sp->seg_flags & SEGM_CLEAN) { 1841 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1842 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1843 /* "1" is the artificial inc in lfs_seglock */ 1844 mutex_enter(&lfs_lock); 1845 while (fs->lfs_iocount > 1) { 1846 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1847 "lfs_initseg", 0, &lfs_lock); 1848 } 1849 mutex_exit(&lfs_lock); 1850 fs->lfs_cleanind = 0; 1851 } 1852 } 1853 1854 sp->fs = fs; 1855 sp->ibp = NULL; 1856 sp->idp = NULL; 1857 sp->ninodes = 0; 1858 sp->ndupino = 0; 1859 1860 sp->cbpp = sp->bpp; 1861 1862 /* Get a new buffer for SEGSUM */ 1863 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1864 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1865 1866 /* ... and enter it into the buffer list. */ 1867 *sp->cbpp = sbp; 1868 sp->cbpp++; 1869 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1870 1871 sp->start_bpp = sp->cbpp; 1872 1873 /* Set point to SEGSUM, initialize it. */ 1874 ssp = sp->segsum = sbp->b_data; 1875 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1876 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1877 lfs_ss_setnfinfo(fs, ssp, 0); 1878 lfs_ss_setninos(fs, ssp, 0); 1879 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1880 lfs_ss_setflags(fs, ssp, flags); 1881 1882 /* Set pointer to first FINFO, initialize it. */ 1883 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1884 lfs_fi_setnblocks(fs, sp->fip, 0); 1885 lfs_fi_setlastlength(fs, sp->fip, 0); 1886 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1887 1888 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1889 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1890 1891 return (repeat); 1892 } 1893 1894 /* 1895 * Remove SEGUSE_INVAL from all segments. 1896 */ 1897 void 1898 lfs_unset_inval_all(struct lfs *fs) 1899 { 1900 SEGUSE *sup; 1901 struct buf *bp; 1902 int i; 1903 1904 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1905 LFS_SEGENTRY(sup, fs, i, bp); 1906 if (sup->su_flags & SEGUSE_INVAL) { 1907 sup->su_flags &= ~SEGUSE_INVAL; 1908 LFS_WRITESEGENTRY(sup, fs, i, bp); 1909 } else 1910 brelse(bp, 0); 1911 } 1912 } 1913 1914 /* 1915 * Return the next segment to write. 1916 */ 1917 void 1918 lfs_newseg(struct lfs *fs) 1919 { 1920 CLEANERINFO *cip; 1921 SEGUSE *sup; 1922 struct buf *bp; 1923 int curseg, isdirty, sn, skip_inval; 1924 1925 ASSERT_SEGLOCK(fs); 1926 1927 /* Honor LFCNWRAPSTOP */ 1928 mutex_enter(&lfs_lock); 1929 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1930 if (fs->lfs_wrappass) { 1931 log(LOG_NOTICE, "%s: wrappass=%d\n", 1932 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1933 fs->lfs_wrappass = 0; 1934 break; 1935 } 1936 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1937 wakeup(&fs->lfs_nowrap); 1938 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1939 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1940 &lfs_lock); 1941 } 1942 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1943 mutex_exit(&lfs_lock); 1944 1945 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1946 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1947 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1948 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1949 sup->su_nbytes = 0; 1950 sup->su_nsums = 0; 1951 sup->su_ninos = 0; 1952 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1953 1954 LFS_CLEANERINFO(cip, fs, bp); 1955 lfs_ci_shiftcleantodirty(fs, cip, 1); 1956 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1957 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1958 1959 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1960 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1961 skip_inval = 1; 1962 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1963 sn = (sn + 1) % lfs_sb_getnseg(fs); 1964 1965 if (sn == curseg) { 1966 if (skip_inval) 1967 skip_inval = 0; 1968 else 1969 panic("lfs_nextseg: no clean segments"); 1970 } 1971 LFS_SEGENTRY(sup, fs, sn, bp); 1972 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1973 /* Check SEGUSE_EMPTY as we go along */ 1974 if (isdirty && sup->su_nbytes == 0 && 1975 !(sup->su_flags & SEGUSE_EMPTY)) 1976 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1977 else 1978 brelse(bp, 0); 1979 1980 if (!isdirty) 1981 break; 1982 } 1983 if (skip_inval == 0) 1984 lfs_unset_inval_all(fs); 1985 1986 ++fs->lfs_nactive; 1987 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1988 if (lfs_dostats) { 1989 ++lfs_stats.segsused; 1990 } 1991 } 1992 1993 static struct buf * 1994 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1995 int n) 1996 { 1997 struct lfs_cluster *cl; 1998 struct buf **bpp, *bp; 1999 2000 ASSERT_SEGLOCK(fs); 2001 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 2002 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 2003 memset(cl, 0, sizeof(*cl)); 2004 cl->fs = fs; 2005 cl->bpp = bpp; 2006 cl->bufcount = 0; 2007 cl->bufsize = 0; 2008 2009 /* If this segment is being written synchronously, note that */ 2010 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 2011 cl->flags |= LFS_CL_SYNC; 2012 cl->seg = fs->lfs_sp; 2013 ++cl->seg->seg_iocount; 2014 } 2015 2016 /* Get an empty buffer header, or maybe one with something on it */ 2017 bp = getiobuf(vp, true); 2018 bp->b_dev = NODEV; 2019 bp->b_blkno = bp->b_lblkno = addr; 2020 bp->b_iodone = lfs_cluster_aiodone; 2021 bp->b_private = cl; 2022 2023 return bp; 2024 } 2025 2026 int 2027 lfs_writeseg(struct lfs *fs, struct segment *sp) 2028 { 2029 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2030 SEGUSE *sup; 2031 SEGSUM *ssp; 2032 int i; 2033 int do_again, nblocks, byteoffset; 2034 size_t el_size; 2035 struct lfs_cluster *cl; 2036 u_short ninos; 2037 struct vnode *devvp; 2038 char *p = NULL; 2039 struct vnode *vp; 2040 unsigned ibindex, iblimit; 2041 int changed; 2042 u_int32_t sum; 2043 size_t sumstart; 2044 uint16_t oflags; 2045 #ifdef DEBUG 2046 FINFO *fip; 2047 int findex; 2048 #endif 2049 2050 ASSERT_SEGLOCK(fs); 2051 2052 ssp = (SEGSUM *)sp->segsum; 2053 2054 /* 2055 * If there are no buffers other than the segment summary to write, 2056 * don't do anything. If we are the end of a dirop sequence, however, 2057 * write the empty segment summary anyway, to help out the 2058 * roll-forward agent. 2059 */ 2060 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2061 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2062 return 0; 2063 } 2064 2065 /* Note if partial segment is being written by the cleaner */ 2066 if (sp->seg_flags & SEGM_CLEAN) 2067 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2068 2069 /* Note if we are writing to reclaim */ 2070 if (sp->seg_flags & SEGM_RECLAIM) { 2071 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2072 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2073 } 2074 /* Save flags to attach to a new partial segment, if we need one */ 2075 oflags = lfs_ss_getflags(fs, ssp); 2076 2077 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2078 2079 /* Update the segment usage information. */ 2080 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2081 2082 /* Loop through all blocks, except the segment summary. */ 2083 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2084 if ((*bpp)->b_vp != devvp) { 2085 sup->su_nbytes += (*bpp)->b_bcount; 2086 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2087 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2088 sp->seg_number, (*bpp)->b_bcount, 2089 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2090 (*bpp)->b_blkno)); 2091 } 2092 } 2093 2094 #ifdef DEBUG 2095 /* Check for zero-length and zero-version FINFO entries. */ 2096 fip = SEGSUM_FINFOBASE(fs, ssp); 2097 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2098 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2099 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2100 fip = NEXT_FINFO(fs, fip); 2101 } 2102 #endif /* DEBUG */ 2103 2104 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2105 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2106 sp->seg_number, 2107 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2108 lfs_ss_getninos(fs, ssp))); 2109 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2110 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2111 if (lfs_sb_getversion(fs) == 1) 2112 sup->su_olastmod = time_second; 2113 else 2114 sup->su_lastmod = time_second; 2115 sup->su_ninos += ninos; 2116 ++sup->su_nsums; 2117 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2118 2119 do_again = !(bp->b_flags & B_GATHERED); 2120 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2121 2122 /* 2123 * Mark blocks BC_BUSY, to prevent then from being changed between 2124 * the checksum computation and the actual write. 2125 * 2126 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2127 * there are any, replace them with copies that have UNASSIGNED 2128 * instead. 2129 */ 2130 mutex_enter(&bufcache_lock); 2131 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2132 ++bpp; 2133 bp = *bpp; 2134 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2135 bp->b_cflags |= BC_BUSY; 2136 continue; 2137 } 2138 2139 while (bp->b_cflags & BC_BUSY) { 2140 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2141 " data summary corruption for ino %d, lbn %" 2142 PRId64 "\n", 2143 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2144 bp->b_cflags |= BC_WANTED; 2145 cv_wait(&bp->b_busy, &bufcache_lock); 2146 } 2147 bp->b_cflags |= BC_BUSY; 2148 mutex_exit(&bufcache_lock); 2149 unbusybp = NULL; 2150 2151 /* 2152 * Check and replace indirect block UNWRITTEN bogosity. 2153 * XXX See comment in lfs_writefile. 2154 */ 2155 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2156 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) != 2157 VTOI(bp->b_vp)->i_lfs_effnblks) { 2158 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2159 VTOI(bp->b_vp)->i_number, 2160 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2161 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din))); 2162 /* Make a copy we'll make changes to */ 2163 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2164 bp->b_bcount, LFS_NB_IBLOCK); 2165 newbp->b_blkno = bp->b_blkno; 2166 memcpy(newbp->b_data, bp->b_data, 2167 newbp->b_bcount); 2168 2169 changed = 0; 2170 iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs); 2171 for (ibindex = 0; ibindex < iblimit; ibindex++) { 2172 if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) { 2173 ++changed; 2174 lfs_iblock_set(fs, newbp->b_data, 2175 ibindex, 0); 2176 } 2177 } 2178 /* 2179 * Get rid of the old buffer. Don't mark it clean, 2180 * though, if it still has dirty data on it. 2181 */ 2182 if (changed) { 2183 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2184 " bp = %p newbp = %p\n", changed, bp, 2185 newbp)); 2186 *bpp = newbp; 2187 bp->b_flags &= ~B_GATHERED; 2188 bp->b_error = 0; 2189 if (bp->b_iodone != NULL) { 2190 DLOG((DLOG_SEG, "lfs_writeseg: " 2191 "indir bp should not be B_CALL\n")); 2192 biodone(bp); 2193 bp = NULL; 2194 } else { 2195 /* Still on free list, leave it there */ 2196 unbusybp = bp; 2197 /* 2198 * We have to re-decrement lfs_avail 2199 * since this block is going to come 2200 * back around to us in the next 2201 * segment. 2202 */ 2203 lfs_sb_subavail(fs, 2204 lfs_btofsb(fs, bp->b_bcount)); 2205 } 2206 } else { 2207 lfs_freebuf(fs, newbp); 2208 } 2209 } 2210 mutex_enter(&bufcache_lock); 2211 if (unbusybp != NULL) { 2212 unbusybp->b_cflags &= ~BC_BUSY; 2213 if (unbusybp->b_cflags & BC_WANTED) 2214 cv_broadcast(&bp->b_busy); 2215 } 2216 } 2217 mutex_exit(&bufcache_lock); 2218 2219 /* 2220 * Compute checksum across data and then across summary; the first 2221 * block (the summary block) is skipped. Set the create time here 2222 * so that it's guaranteed to be later than the inode mod times. 2223 */ 2224 sum = 0; 2225 if (lfs_sb_getversion(fs) == 1) 2226 el_size = sizeof(u_long); 2227 else 2228 el_size = sizeof(u_int32_t); 2229 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2230 ++bpp; 2231 /* Loop through gop_write cluster blocks */ 2232 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2233 byteoffset += lfs_sb_getbsize(fs)) { 2234 #ifdef LFS_USE_BC_INVAL 2235 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2236 (*bpp)->b_iodone != NULL) { 2237 if (copyin((void *)(*bpp)->b_saveaddr + 2238 byteoffset, dp, el_size)) { 2239 panic("lfs_writeseg: copyin failed [1]:" 2240 " ino %" PRIu64 " blk %" PRId64, 2241 VTOI((*bpp)->b_vp)->i_number, 2242 (*bpp)->b_lblkno); 2243 } 2244 } else 2245 #endif /* LFS_USE_BC_INVAL */ 2246 { 2247 sum = lfs_cksum_part((char *) 2248 (*bpp)->b_data + byteoffset, el_size, sum); 2249 } 2250 } 2251 } 2252 if (lfs_sb_getversion(fs) == 1) 2253 lfs_ss_setocreate(fs, ssp, time_second); 2254 else { 2255 lfs_ss_setcreate(fs, ssp, time_second); 2256 lfs_sb_addserial(fs, 1); 2257 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2258 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2259 } 2260 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2261 sumstart = lfs_ss_getsumstart(fs); 2262 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2263 lfs_sb_getsumsize(fs) - sumstart)); 2264 2265 mutex_enter(&lfs_lock); 2266 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2267 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2268 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2269 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2270 mutex_exit(&lfs_lock); 2271 2272 /* 2273 * When we simply write the blocks we lose a rotation for every block 2274 * written. To avoid this problem, we cluster the buffers into a 2275 * chunk and write the chunk. MAXPHYS is the largest size I/O 2276 * devices can handle, use that for the size of the chunks. 2277 * 2278 * Blocks that are already clusters (from GOP_WRITE), however, we 2279 * don't bother to copy into other clusters. 2280 */ 2281 2282 #define CHUNKSIZE MAXPHYS 2283 2284 if (devvp == NULL) 2285 panic("devvp is NULL"); 2286 for (bpp = sp->bpp, i = nblocks; i;) { 2287 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2288 cl = cbp->b_private; 2289 2290 cbp->b_flags |= B_ASYNC; 2291 cbp->b_cflags |= BC_BUSY; 2292 cbp->b_bcount = 0; 2293 2294 KASSERTMSG((bpp - sp->bpp <= 2295 (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2296 / sizeof(int32_t)), 2297 "lfs_writeseg: real bpp overwrite"); 2298 KASSERTMSG((bpp - sp->bpp <= 2299 lfs_segsize(fs) / lfs_sb_getfsize(fs)), 2300 "lfs_writeseg: theoretical bpp overwrite"); 2301 2302 /* 2303 * Construct the cluster. 2304 */ 2305 mutex_enter(&lfs_lock); 2306 ++fs->lfs_iocount; 2307 mutex_exit(&lfs_lock); 2308 while (i && cbp->b_bcount < CHUNKSIZE) { 2309 bp = *bpp; 2310 2311 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2312 break; 2313 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2314 break; 2315 2316 /* Clusters from GOP_WRITE are expedited */ 2317 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2318 if (cbp->b_bcount > 0) 2319 /* Put in its own buffer */ 2320 break; 2321 else { 2322 cbp->b_data = bp->b_data; 2323 } 2324 } else if (cbp->b_bcount == 0) { 2325 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2326 LFS_NB_CLUSTER); 2327 cl->flags |= LFS_CL_MALLOC; 2328 } 2329 KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2330 btodb(bp->b_bcount - 1))) == 2331 sp->seg_number), 2332 "segment overwrite: blk size %d daddr %" PRIx64 2333 " not in seg %d\n", 2334 bp->b_bcount, bp->b_blkno, 2335 sp->seg_number); 2336 2337 #ifdef LFS_USE_BC_INVAL 2338 /* 2339 * Fake buffers from the cleaner are marked as BC_INVAL. 2340 * We need to copy the data from user space rather than 2341 * from the buffer indicated. 2342 * XXX == what do I do on an error? 2343 */ 2344 if ((bp->b_cflags & BC_INVAL) != 0 && 2345 bp->b_iodone != NULL) { 2346 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2347 panic("lfs_writeseg: " 2348 "copyin failed [2]"); 2349 } else 2350 #endif /* LFS_USE_BC_INVAL */ 2351 if (cl->flags & LFS_CL_MALLOC) { 2352 /* copy data into our cluster. */ 2353 memcpy(p, bp->b_data, bp->b_bcount); 2354 p += bp->b_bcount; 2355 } 2356 2357 cbp->b_bcount += bp->b_bcount; 2358 cl->bufsize += bp->b_bcount; 2359 2360 bp->b_flags &= ~B_READ; 2361 bp->b_error = 0; 2362 cl->bpp[cl->bufcount++] = bp; 2363 2364 vp = bp->b_vp; 2365 mutex_enter(&bufcache_lock); 2366 mutex_enter(vp->v_interlock); 2367 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2368 reassignbuf(bp, vp); 2369 vp->v_numoutput++; 2370 mutex_exit(vp->v_interlock); 2371 mutex_exit(&bufcache_lock); 2372 2373 bpp++; 2374 i--; 2375 } 2376 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2377 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2378 else 2379 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2380 mutex_enter(devvp->v_interlock); 2381 devvp->v_numoutput++; 2382 mutex_exit(devvp->v_interlock); 2383 VOP_STRATEGY(devvp, cbp); 2384 curlwp->l_ru.ru_oublock++; 2385 } 2386 2387 if (lfs_dostats) { 2388 ++lfs_stats.psegwrites; 2389 lfs_stats.blocktot += nblocks - 1; 2390 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2391 ++lfs_stats.psyncwrites; 2392 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2393 ++lfs_stats.pcleanwrites; 2394 lfs_stats.cleanblocks += nblocks - 1; 2395 } 2396 } 2397 2398 return (lfs_initseg(fs, oflags) || do_again); 2399 } 2400 2401 void 2402 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2403 { 2404 struct buf *bp; 2405 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2406 2407 ASSERT_MAYBE_SEGLOCK(fs); 2408 if (fs->lfs_is64) { 2409 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2410 } else { 2411 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2412 } 2413 /* 2414 * If we can write one superblock while another is in 2415 * progress, we risk not having a complete checkpoint if we crash. 2416 * So, block here if a superblock write is in progress. 2417 */ 2418 mutex_enter(&lfs_lock); 2419 while (fs->lfs_sbactive) { 2420 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2421 &lfs_lock); 2422 } 2423 fs->lfs_sbactive = daddr; 2424 mutex_exit(&lfs_lock); 2425 2426 /* Set timestamp of this version of the superblock */ 2427 if (lfs_sb_getversion(fs) == 1) 2428 lfs_sb_setotstamp(fs, time_second); 2429 lfs_sb_settstamp(fs, time_second); 2430 2431 /* The next chunk of code relies on this assumption */ 2432 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2433 2434 /* Checksum the superblock and copy it into a buffer. */ 2435 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2436 bp = lfs_newbuf(fs, devvp, 2437 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2438 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2439 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2440 LFS_SBPAD - sizeof(struct dlfs)); 2441 2442 bp->b_cflags |= BC_BUSY; 2443 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2444 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2445 bp->b_error = 0; 2446 bp->b_iodone = lfs_super_aiodone; 2447 2448 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2449 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2450 else 2451 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2452 curlwp->l_ru.ru_oublock++; 2453 2454 mutex_enter(devvp->v_interlock); 2455 devvp->v_numoutput++; 2456 mutex_exit(devvp->v_interlock); 2457 2458 mutex_enter(&lfs_lock); 2459 ++fs->lfs_iocount; 2460 mutex_exit(&lfs_lock); 2461 VOP_STRATEGY(devvp, bp); 2462 } 2463 2464 /* 2465 * Logical block number match routines used when traversing the dirty block 2466 * chain. 2467 */ 2468 int 2469 lfs_match_fake(struct lfs *fs, struct buf *bp) 2470 { 2471 2472 ASSERT_SEGLOCK(fs); 2473 return LFS_IS_MALLOC_BUF(bp); 2474 } 2475 2476 #if 0 2477 int 2478 lfs_match_real(struct lfs *fs, struct buf *bp) 2479 { 2480 2481 ASSERT_SEGLOCK(fs); 2482 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2483 } 2484 #endif 2485 2486 int 2487 lfs_match_data(struct lfs *fs, struct buf *bp) 2488 { 2489 2490 ASSERT_SEGLOCK(fs); 2491 return (bp->b_lblkno >= 0); 2492 } 2493 2494 int 2495 lfs_match_indir(struct lfs *fs, struct buf *bp) 2496 { 2497 daddr_t lbn; 2498 2499 ASSERT_SEGLOCK(fs); 2500 lbn = bp->b_lblkno; 2501 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2502 } 2503 2504 int 2505 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2506 { 2507 daddr_t lbn; 2508 2509 ASSERT_SEGLOCK(fs); 2510 lbn = bp->b_lblkno; 2511 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2512 } 2513 2514 int 2515 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2516 { 2517 daddr_t lbn; 2518 2519 ASSERT_SEGLOCK(fs); 2520 lbn = bp->b_lblkno; 2521 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2522 } 2523 2524 void 2525 lfs_free_aiodone(struct buf *bp) 2526 { 2527 struct lfs *fs; 2528 2529 KERNEL_LOCK(1, curlwp); 2530 fs = bp->b_private; 2531 ASSERT_NO_SEGLOCK(fs); 2532 lfs_freebuf(fs, bp); 2533 KERNEL_UNLOCK_ONE(curlwp); 2534 } 2535 2536 static void 2537 lfs_super_aiodone(struct buf *bp) 2538 { 2539 struct lfs *fs; 2540 2541 KERNEL_LOCK(1, curlwp); 2542 fs = bp->b_private; 2543 ASSERT_NO_SEGLOCK(fs); 2544 mutex_enter(&lfs_lock); 2545 fs->lfs_sbactive = 0; 2546 if (--fs->lfs_iocount <= 1) 2547 wakeup(&fs->lfs_iocount); 2548 wakeup(&fs->lfs_sbactive); 2549 mutex_exit(&lfs_lock); 2550 lfs_freebuf(fs, bp); 2551 KERNEL_UNLOCK_ONE(curlwp); 2552 } 2553 2554 static void 2555 lfs_cluster_aiodone(struct buf *bp) 2556 { 2557 workqueue_enqueue(lfs_wq, (struct work *)bp, NULL); 2558 } 2559 2560 void 2561 lfs_cluster_wq(struct work *wk, void *arg) 2562 { 2563 struct buf *bp = (struct buf *)wk; 2564 struct lfs_cluster *cl; 2565 struct lfs *fs; 2566 struct buf *tbp, *fbp; 2567 struct vnode *vp, *devvp, *ovp; 2568 struct inode *ip; 2569 int error; 2570 2571 error = bp->b_error; 2572 cl = bp->b_private; 2573 fs = cl->fs; 2574 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2575 ASSERT_NO_SEGLOCK(fs); 2576 2577 /* Put the pages back, and release the buffer */ 2578 while (cl->bufcount--) { 2579 tbp = cl->bpp[cl->bufcount]; 2580 KASSERT(tbp->b_cflags & BC_BUSY); 2581 if (error) { 2582 tbp->b_error = error; 2583 } 2584 2585 /* 2586 * We're done with tbp. If it has not been re-dirtied since 2587 * the cluster was written, free it. Otherwise, keep it on 2588 * the locked list to be written again. 2589 */ 2590 vp = tbp->b_vp; 2591 2592 tbp->b_flags &= ~B_GATHERED; 2593 2594 #ifdef DEBUG 2595 if ((tbp)->b_vp == (fs)->lfs_ivnode) 2596 LFS_ENTER_LOG("clear", __FILE__, __LINE__, 2597 tbp->b_lblkno, tbp->b_flags, curproc->p_pid); 2598 #endif 2599 2600 mutex_enter(&bufcache_lock); 2601 if (tbp->b_iodone == NULL) { 2602 KASSERT(tbp->b_flags & B_LOCKED); 2603 bremfree(tbp); 2604 if (vp) { 2605 mutex_enter(vp->v_interlock); 2606 reassignbuf(tbp, vp); 2607 mutex_exit(vp->v_interlock); 2608 } 2609 tbp->b_flags |= B_ASYNC; /* for biodone */ 2610 } 2611 2612 if ((tbp->b_flags & B_LOCKED) && !(tbp->b_oflags & BO_DELWRI)) 2613 LFS_UNLOCK_BUF(tbp); 2614 2615 if (tbp->b_oflags & BO_DONE) { 2616 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2617 cl->bufcount, (long)tbp->b_flags)); 2618 } 2619 2620 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2621 /* 2622 * A buffer from the page daemon. 2623 * We use the same iodone as it does, 2624 * so we must manually disassociate its 2625 * buffers from the vp. 2626 */ 2627 if ((ovp = tbp->b_vp) != NULL) { 2628 /* This is just silly */ 2629 mutex_enter(ovp->v_interlock); 2630 brelvp(tbp); 2631 mutex_exit(ovp->v_interlock); 2632 tbp->b_vp = vp; 2633 tbp->b_objlock = vp->v_interlock; 2634 } 2635 /* Put it back the way it was */ 2636 tbp->b_flags |= B_ASYNC; 2637 /* Master buffers have BC_AGE */ 2638 if (tbp->b_private == tbp) 2639 tbp->b_cflags |= BC_AGE; 2640 } 2641 mutex_exit(&bufcache_lock); 2642 2643 biodone(tbp); 2644 2645 /* 2646 * If this is the last block for this vnode, but 2647 * there are other blocks on its dirty list, 2648 * set IN_MODIFIED/IN_CLEANING depending on what 2649 * sort of block. Only do this for our mount point, 2650 * not for, e.g., inode blocks that are attached to 2651 * the devvp. 2652 * XXX KS - Shouldn't we set *both* if both types 2653 * of blocks are present (traverse the dirty list?) 2654 */ 2655 mutex_enter(vp->v_interlock); 2656 mutex_enter(&lfs_lock); 2657 if (vp != devvp && vp->v_numoutput == 0 && 2658 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2659 ip = VTOI(vp); 2660 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2661 ip->i_number)); 2662 if (LFS_IS_MALLOC_BUF(fbp)) 2663 LFS_SET_UINO(ip, IN_CLEANING); 2664 else 2665 LFS_SET_UINO(ip, IN_MODIFIED); 2666 } 2667 cv_broadcast(&vp->v_cv); 2668 mutex_exit(&lfs_lock); 2669 mutex_exit(vp->v_interlock); 2670 } 2671 2672 /* Fix up the cluster buffer, and release it */ 2673 if (cl->flags & LFS_CL_MALLOC) 2674 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2675 putiobuf(bp); 2676 2677 /* Note i/o done */ 2678 if (cl->flags & LFS_CL_SYNC) { 2679 if (--cl->seg->seg_iocount == 0) 2680 wakeup(&cl->seg->seg_iocount); 2681 } 2682 mutex_enter(&lfs_lock); 2683 KASSERTMSG((fs->lfs_iocount != 0), 2684 "lfs_cluster_aiodone: zero iocount"); 2685 if (--fs->lfs_iocount <= 1) 2686 wakeup(&fs->lfs_iocount); 2687 mutex_exit(&lfs_lock); 2688 2689 pool_put(&fs->lfs_bpppool, cl->bpp); 2690 cl->bpp = NULL; 2691 pool_put(&fs->lfs_clpool, cl); 2692 } 2693 2694 /* 2695 * Shellsort (diminishing increment sort) from Data Structures and 2696 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2697 * see also Knuth Vol. 3, page 84. The increments are selected from 2698 * formula (8), page 95. Roughly O(N^3/2). 2699 */ 2700 /* 2701 * This is our own private copy of shellsort because we want to sort 2702 * two parallel arrays (the array of buffer pointers and the array of 2703 * logical block numbers) simultaneously. Note that we cast the array 2704 * of logical block numbers to a unsigned in this routine so that the 2705 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2706 */ 2707 2708 static void 2709 lfs_shellsort(struct lfs *fs, 2710 struct buf **bp_array, union lfs_blocks *lb_array, 2711 int nmemb, int size) 2712 { 2713 static int __rsshell_increments[] = { 4, 1, 0 }; 2714 int incr, *incrp, t1, t2; 2715 struct buf *bp_temp; 2716 2717 #ifdef DEBUG 2718 incr = 0; 2719 for (t1 = 0; t1 < nmemb; t1++) { 2720 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2721 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2722 /* dump before panic */ 2723 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2724 nmemb, size); 2725 incr = 0; 2726 for (t1 = 0; t1 < nmemb; t1++) { 2727 const struct buf *bp = bp_array[t1]; 2728 2729 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2730 PRIu64 "\n", t1, 2731 (uint64_t)bp->b_bcount, 2732 (uint64_t)bp->b_lblkno); 2733 printf("lbns:"); 2734 for (t2 = 0; t2 * size < bp->b_bcount; 2735 t2++) { 2736 printf(" %jd", 2737 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2738 } 2739 printf("\n"); 2740 } 2741 panic("lfs_shellsort: inconsistent input"); 2742 } 2743 } 2744 } 2745 #endif 2746 2747 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2748 for (t1 = incr; t1 < nmemb; ++t1) 2749 for (t2 = t1 - incr; t2 >= 0;) 2750 if ((u_int64_t)bp_array[t2]->b_lblkno > 2751 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2752 bp_temp = bp_array[t2]; 2753 bp_array[t2] = bp_array[t2 + incr]; 2754 bp_array[t2 + incr] = bp_temp; 2755 t2 -= incr; 2756 } else 2757 break; 2758 2759 /* Reform the list of logical blocks */ 2760 incr = 0; 2761 for (t1 = 0; t1 < nmemb; t1++) { 2762 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2763 lfs_blocks_set(fs, lb_array, incr++, 2764 bp_array[t1]->b_lblkno + t2); 2765 } 2766 } 2767 } 2768 2769 /* 2770 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2771 * point at uninitialized space. 2772 */ 2773 void 2774 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2775 { 2776 struct segment *sp = fs->lfs_sp; 2777 SEGSUM *ssp; 2778 2779 KASSERT(vers > 0); 2780 2781 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2782 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2783 (void) lfs_writeseg(fs, fs->lfs_sp); 2784 2785 sp->sum_bytes_left -= FINFOSIZE(fs); 2786 ssp = (SEGSUM *)sp->segsum; 2787 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2788 lfs_fi_setnblocks(fs, sp->fip, 0); 2789 lfs_fi_setino(fs, sp->fip, ino); 2790 lfs_fi_setversion(fs, sp->fip, vers); 2791 } 2792 2793 /* 2794 * Release the FINFO entry, either clearing out an unused entry or 2795 * advancing us to the next available entry. 2796 */ 2797 void 2798 lfs_release_finfo(struct lfs *fs) 2799 { 2800 struct segment *sp = fs->lfs_sp; 2801 SEGSUM *ssp; 2802 2803 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2804 sp->fip = NEXT_FINFO(fs, sp->fip); 2805 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2806 } else { 2807 /* XXX shouldn't this update sp->fip? */ 2808 sp->sum_bytes_left += FINFOSIZE(fs); 2809 ssp = (SEGSUM *)sp->segsum; 2810 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2811 } 2812 } 2813
Indexes created Mon Sep 22 13:09:51 GMT 2025