lfs_segment.c revision 1.263.2.1
1 /* $NetBSD: lfs_segment.c,v 1.263.2.1 2017/03/20 06:57:54 pgoyette 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 1.263.2.1 2017/03/20 06:57:54 pgoyette Exp $"); 64 65 #ifdef DEBUG 66 # define vndebug(vp, str) do { \ 67 if (VTOI(vp)->i_flag & 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 95 #include <miscfs/specfs/specdev.h> 96 #include <miscfs/fifofs/fifo.h> 97 98 #include <ufs/lfs/ulfs_inode.h> 99 #include <ufs/lfs/ulfsmount.h> 100 #include <ufs/lfs/ulfs_extern.h> 101 102 #include <ufs/lfs/lfs.h> 103 #include <ufs/lfs/lfs_accessors.h> 104 #include <ufs/lfs/lfs_kernel.h> 105 #include <ufs/lfs/lfs_extern.h> 106 107 #include <uvm/uvm.h> 108 #include <uvm/uvm_extern.h> 109 110 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 111 112 static void lfs_generic_callback(struct buf *, void (*)(struct buf *)); 113 static void lfs_free_aiodone(struct buf *); 114 static void lfs_super_aiodone(struct buf *); 115 static void lfs_cluster_aiodone(struct buf *); 116 static void lfs_cluster_callback(struct buf *); 117 118 /* 119 * Determine if it's OK to start a partial in this segment, or if we need 120 * to go on to a new segment. 121 */ 122 #define LFS_PARTIAL_FITS(fs) \ 123 (lfs_sb_getfsbpseg(fs) - \ 124 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)) > \ 125 lfs_sb_getfrag(fs)) 126 127 /* 128 * Figure out whether we should do a checkpoint write or go ahead with 129 * an ordinary write. 130 */ 131 #define LFS_SHOULD_CHECKPOINT(fs, flags) \ 132 ((flags & SEGM_CLEAN) == 0 && \ 133 ((fs->lfs_nactive > LFS_MAX_ACTIVE || \ 134 (flags & SEGM_CKP) || \ 135 lfs_sb_getnclean(fs) < LFS_MAX_ACTIVE))) 136 137 int lfs_match_fake(struct lfs *, struct buf *); 138 void lfs_newseg(struct lfs *); 139 void lfs_supercallback(struct buf *); 140 void lfs_updatemeta(struct segment *); 141 void lfs_writesuper(struct lfs *, daddr_t); 142 int lfs_writevnodes(struct lfs *fs, struct mount *mp, 143 struct segment *sp, int dirops); 144 145 static void lfs_shellsort(struct lfs *, struct buf **, union lfs_blocks *, 146 int, int); 147 148 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 149 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 150 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 151 int lfs_dirvcount = 0; /* # active dirops */ 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_flag & 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 ((pg->flags & PG_CLEAN) == 0 || 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_flag &= ~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_flag & 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_flag & 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 KDASSERT(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 = VTOI(vp); 488 int op = c->op; 489 490 if (ip == NULL || vp->v_type == VNON) 491 return false; 492 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 493 (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) { 494 vndebug(vp, "dirop"); 495 return false; 496 } 497 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 498 vndebug(vp,"empty"); 499 return false;; 500 } 501 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM && 502 vp != c->fs->lfs_flushvp && !(ip->i_flag & IN_CLEANING)) { 503 vndebug(vp,"cleaning"); 504 return false; 505 } 506 mutex_enter(&lfs_lock); 507 if (vp == c->fs->lfs_unlockvp) { 508 mutex_exit(&lfs_lock); 509 return false; 510 } 511 mutex_exit(&lfs_lock); 512 513 return true; 514 } 515 516 int 517 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 518 { 519 struct inode *ip; 520 struct vnode *vp; 521 struct vnode_iterator *marker; 522 struct lfs_writevnodes_ctx ctx; 523 int inodes_written = 0; 524 int error = 0; 525 526 /* 527 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist. 528 * XXX The rationale is unclear, the initial commit had no information. 529 * XXX If the order really matters we have to sort the vnodes first. 530 */ 531 532 ASSERT_SEGLOCK(fs); 533 vfs_vnode_iterator_init(mp, &marker); 534 ctx.op = op; 535 ctx.fs = fs; 536 while ((vp = vfs_vnode_iterator_next(marker, 537 lfs_writevnodes_selector, &ctx)) != NULL) { 538 ip = VTOI(vp); 539 540 /* 541 * Write the inode/file if dirty and it's not the IFILE. 542 */ 543 if (((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) && 544 ip->i_number != LFS_IFILE_INUM) { 545 error = lfs_writefile(fs, sp, vp); 546 if (error) { 547 vrele(vp); 548 if (error == EAGAIN) { 549 /* 550 * This error from lfs_putpages 551 * indicates we need to drop 552 * the segment lock and start 553 * over after the cleaner has 554 * had a chance to run. 555 */ 556 lfs_writeinode(fs, sp, ip); 557 lfs_writeseg(fs, sp); 558 if (!VPISEMPTY(vp) && 559 !WRITEINPROG(vp) && 560 !(ip->i_flag & IN_ALLMOD)) { 561 mutex_enter(&lfs_lock); 562 LFS_SET_UINO(ip, IN_MODIFIED); 563 mutex_exit(&lfs_lock); 564 } 565 break; 566 } 567 error = 0; /* XXX not quite right */ 568 continue; 569 } 570 571 if (!VPISEMPTY(vp)) { 572 if (WRITEINPROG(vp)) { 573 ivndebug(vp,"writevnodes/write2"); 574 } else if (!(ip->i_flag & IN_ALLMOD)) { 575 mutex_enter(&lfs_lock); 576 LFS_SET_UINO(ip, IN_MODIFIED); 577 mutex_exit(&lfs_lock); 578 } 579 } 580 (void) lfs_writeinode(fs, sp, ip); 581 inodes_written++; 582 } 583 vrele(vp); 584 } 585 vfs_vnode_iterator_destroy(marker); 586 return error; 587 } 588 589 /* 590 * Do a checkpoint. 591 */ 592 int 593 lfs_segwrite(struct mount *mp, int flags) 594 { 595 struct buf *bp; 596 struct inode *ip; 597 struct lfs *fs; 598 struct segment *sp; 599 struct vnode *vp; 600 SEGUSE *segusep; 601 int do_ckp, did_ckp, error; 602 unsigned n, segleft, maxseg, sn, i, curseg; 603 int writer_set = 0; 604 int dirty; 605 int redo; 606 SEGSUM *ssp; 607 int um_error; 608 609 fs = VFSTOULFS(mp)->um_lfs; 610 ASSERT_MAYBE_SEGLOCK(fs); 611 612 if (fs->lfs_ronly) 613 return EROFS; 614 615 lfs_imtime(fs); 616 617 /* 618 * Allocate a segment structure and enough space to hold pointers to 619 * the maximum possible number of buffers which can be described in a 620 * single summary block. 621 */ 622 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 623 624 /* We can't do a partial write and checkpoint at the same time. */ 625 if (do_ckp) 626 flags &= ~SEGM_SINGLE; 627 628 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 629 sp = fs->lfs_sp; 630 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 631 do_ckp = 1; 632 633 /* 634 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 635 * in which case we have to flush *all* buffers off of this vnode. 636 * We don't care about other nodes, but write any non-dirop nodes 637 * anyway in anticipation of another getnewvnode(). 638 * 639 * If we're cleaning we only write cleaning and ifile blocks, and 640 * no dirops, since otherwise we'd risk corruption in a crash. 641 */ 642 if (sp->seg_flags & SEGM_CLEAN) 643 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 644 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 645 do { 646 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 647 if ((sp->seg_flags & SEGM_SINGLE) && 648 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 649 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 650 break; 651 } 652 653 if (do_ckp || fs->lfs_dirops == 0) { 654 if (!writer_set) { 655 lfs_writer_enter(fs, "lfs writer"); 656 writer_set = 1; 657 } 658 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 659 if (um_error == 0) 660 um_error = error; 661 /* In case writevnodes errored out */ 662 lfs_flush_dirops(fs); 663 ssp = (SEGSUM *)(sp->segsum); 664 lfs_ss_setflags(fs, ssp, 665 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 666 lfs_finalize_fs_seguse(fs); 667 } 668 if (do_ckp && um_error) { 669 lfs_segunlock_relock(fs); 670 sp = fs->lfs_sp; 671 } 672 } while (do_ckp && um_error != 0); 673 } 674 675 /* 676 * If we are doing a checkpoint, mark everything since the 677 * last checkpoint as no longer ACTIVE. 678 */ 679 if (do_ckp || fs->lfs_doifile) { 680 segleft = lfs_sb_getnseg(fs); 681 curseg = 0; 682 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 683 int bread_error; 684 685 dirty = 0; 686 bread_error = bread(fs->lfs_ivnode, 687 lfs_sb_getcleansz(fs) + n, 688 lfs_sb_getbsize(fs), B_MODIFY, &bp); 689 if (bread_error) 690 panic("lfs_segwrite: ifile read: " 691 "seguse %u: error %d\n", 692 n, bread_error); 693 segusep = (SEGUSE *)bp->b_data; 694 maxseg = min(segleft, lfs_sb_getsepb(fs)); 695 for (i = 0; i < maxseg; i++) { 696 sn = curseg + i; 697 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 698 segusep->su_flags & SEGUSE_ACTIVE) { 699 segusep->su_flags &= ~SEGUSE_ACTIVE; 700 --fs->lfs_nactive; 701 ++dirty; 702 } 703 fs->lfs_suflags[fs->lfs_activesb][sn] = 704 segusep->su_flags; 705 if (lfs_sb_getversion(fs) > 1) 706 ++segusep; 707 else 708 segusep = (SEGUSE *) 709 ((SEGUSE_V1 *)segusep + 1); 710 } 711 712 if (dirty) 713 error = LFS_BWRITE_LOG(bp); /* Ifile */ 714 else 715 brelse(bp, 0); 716 segleft -= lfs_sb_getsepb(fs); 717 curseg += lfs_sb_getsepb(fs); 718 } 719 } 720 721 KASSERT(LFS_SEGLOCK_HELD(fs)); 722 723 did_ckp = 0; 724 if (do_ckp || fs->lfs_doifile) { 725 vp = fs->lfs_ivnode; 726 #ifdef DEBUG 727 int loopcount = 0; 728 #endif 729 do { 730 #ifdef DEBUG 731 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 732 #endif 733 mutex_enter(&lfs_lock); 734 fs->lfs_flags &= ~LFS_IFDIRTY; 735 mutex_exit(&lfs_lock); 736 737 ip = VTOI(vp); 738 739 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 740 /* 741 * Ifile has no pages, so we don't need 742 * to check error return here. 743 */ 744 lfs_writefile(fs, sp, vp); 745 /* 746 * Ensure the Ifile takes the current segment 747 * into account. See comment in lfs_vflush. 748 */ 749 lfs_writefile(fs, sp, vp); 750 lfs_writefile(fs, sp, vp); 751 } 752 753 if (ip->i_flag & IN_ALLMOD) 754 ++did_ckp; 755 #if 0 756 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 757 #else 758 redo = lfs_writeinode(fs, sp, ip); 759 #endif 760 redo += lfs_writeseg(fs, sp); 761 mutex_enter(&lfs_lock); 762 redo += (fs->lfs_flags & LFS_IFDIRTY); 763 mutex_exit(&lfs_lock); 764 #ifdef DEBUG 765 if (++loopcount > 2) 766 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 767 loopcount); 768 #endif 769 } while (redo && do_ckp); 770 771 /* 772 * Unless we are unmounting, the Ifile may continue to have 773 * dirty blocks even after a checkpoint, due to changes to 774 * inodes' atime. If we're checkpointing, it's "impossible" 775 * for other parts of the Ifile to be dirty after the loop 776 * above, since we hold the segment lock. 777 */ 778 mutex_enter(vp->v_interlock); 779 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 780 LFS_CLR_UINO(ip, IN_ALLMOD); 781 } 782 #ifdef DIAGNOSTIC 783 else if (do_ckp) { 784 int do_panic = 0; 785 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 786 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 787 lfs_sb_getsegtabsz(fs) && 788 !(bp->b_flags & B_GATHERED)) { 789 printf("ifile lbn %ld still dirty (flags %lx)\n", 790 (long)bp->b_lblkno, 791 (long)bp->b_flags); 792 ++do_panic; 793 } 794 } 795 if (do_panic) 796 panic("dirty blocks"); 797 } 798 #endif 799 mutex_exit(vp->v_interlock); 800 } else { 801 (void) lfs_writeseg(fs, sp); 802 } 803 804 /* Note Ifile no longer needs to be written */ 805 fs->lfs_doifile = 0; 806 if (writer_set) 807 lfs_writer_leave(fs); 808 809 /* 810 * If we didn't write the Ifile, we didn't really do anything. 811 * That means that (1) there is a checkpoint on disk and (2) 812 * nothing has changed since it was written. 813 * 814 * Take the flags off of the segment so that lfs_segunlock 815 * doesn't have to write the superblock either. 816 */ 817 if (do_ckp && !did_ckp) { 818 sp->seg_flags &= ~SEGM_CKP; 819 } 820 821 if (lfs_dostats) { 822 ++lfs_stats.nwrites; 823 if (sp->seg_flags & SEGM_SYNC) 824 ++lfs_stats.nsync_writes; 825 if (sp->seg_flags & SEGM_CKP) 826 ++lfs_stats.ncheckpoints; 827 } 828 lfs_segunlock(fs); 829 return (0); 830 } 831 832 /* 833 * Write the dirty blocks associated with a vnode. 834 */ 835 int 836 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 837 { 838 struct inode *ip; 839 int i, frag; 840 SEGSUM *ssp; 841 int error; 842 843 ASSERT_SEGLOCK(fs); 844 error = 0; 845 ip = VTOI(vp); 846 847 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 848 849 if (vp->v_uflag & VU_DIROP) { 850 ssp = (SEGSUM *)sp->segsum; 851 lfs_ss_setflags(fs, ssp, 852 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 853 } 854 855 if (sp->seg_flags & SEGM_CLEAN) { 856 lfs_gather(fs, sp, vp, lfs_match_fake); 857 /* 858 * For a file being flushed, we need to write *all* blocks. 859 * This means writing the cleaning blocks first, and then 860 * immediately following with any non-cleaning blocks. 861 * The same is true of the Ifile since checkpoints assume 862 * that all valid Ifile blocks are written. 863 */ 864 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 865 lfs_gather(fs, sp, vp, lfs_match_data); 866 /* 867 * Don't call VOP_PUTPAGES: if we're flushing, 868 * we've already done it, and the Ifile doesn't 869 * use the page cache. 870 */ 871 } 872 } else { 873 lfs_gather(fs, sp, vp, lfs_match_data); 874 /* 875 * If we're flushing, we've already called VOP_PUTPAGES 876 * so don't do it again. Otherwise, we want to write 877 * everything we've got. 878 */ 879 if (!IS_FLUSHING(fs, vp)) { 880 mutex_enter(vp->v_interlock); 881 error = VOP_PUTPAGES(vp, 0, 0, 882 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 883 } 884 } 885 886 /* 887 * It may not be necessary to write the meta-data blocks at this point, 888 * as the roll-forward recovery code should be able to reconstruct the 889 * list. 890 * 891 * We have to write them anyway, though, under two conditions: (1) the 892 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 893 * checkpointing. 894 * 895 * BUT if we are cleaning, we might have indirect blocks that refer to 896 * new blocks not being written yet, in addition to fragments being 897 * moved out of a cleaned segment. If that is the case, don't 898 * write the indirect blocks, or the finfo will have a small block 899 * in the middle of it! 900 * XXX in this case isn't the inode size wrong too? 901 */ 902 frag = 0; 903 if (sp->seg_flags & SEGM_CLEAN) { 904 for (i = 0; i < ULFS_NDADDR; i++) 905 if (ip->i_lfs_fragsize[i] > 0 && 906 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 907 ++frag; 908 } 909 KASSERTMSG((frag <= 1), 910 "lfs_writefile: more than one fragment! frag=%d", frag); 911 if (IS_FLUSHING(fs, vp) || 912 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 913 lfs_gather(fs, sp, vp, lfs_match_indir); 914 lfs_gather(fs, sp, vp, lfs_match_dindir); 915 lfs_gather(fs, sp, vp, lfs_match_tindir); 916 } 917 lfs_release_finfo(fs); 918 919 return error; 920 } 921 922 /* 923 * Update segment accounting to reflect this inode's change of address. 924 */ 925 static int 926 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 927 { 928 struct buf *bp; 929 daddr_t daddr; 930 IFILE *ifp; 931 SEGUSE *sup; 932 ino_t ino; 933 int redo_ifile; 934 u_int32_t sn; 935 936 redo_ifile = 0; 937 938 /* 939 * If updating the ifile, update the super-block. Update the disk 940 * address and access times for this inode in the ifile. 941 */ 942 ino = ip->i_number; 943 if (ino == LFS_IFILE_INUM) { 944 daddr = lfs_sb_getidaddr(fs); 945 lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr)); 946 } else { 947 LFS_IENTRY(ifp, fs, ino, bp); 948 daddr = lfs_if_getdaddr(fs, ifp); 949 lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr)); 950 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 951 } 952 953 /* 954 * If this is the Ifile and lfs_offset is set to the first block 955 * in the segment, dirty the new segment's accounting block 956 * (XXX should already be dirty?) and tell the caller to do it again. 957 */ 958 if (ip->i_number == LFS_IFILE_INUM) { 959 sn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 960 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) == 961 lfs_sb_getoffset(fs)) { 962 LFS_SEGENTRY(sup, fs, sn, bp); 963 KASSERT(bp->b_oflags & BO_DELWRI); 964 LFS_WRITESEGENTRY(sup, fs, sn, bp); 965 /* fs->lfs_flags |= LFS_IFDIRTY; */ 966 redo_ifile |= 1; 967 } 968 } 969 970 /* 971 * The inode's last address should not be in the current partial 972 * segment, except under exceptional circumstances (lfs_writevnodes 973 * had to start over, and in the meantime more blocks were written 974 * to a vnode). Both inodes will be accounted to this segment 975 * in lfs_writeseg so we need to subtract the earlier version 976 * here anyway. The segment count can temporarily dip below 977 * zero here; keep track of how many duplicates we have in 978 * "dupino" so we don't panic below. 979 */ 980 if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) { 981 ++sp->ndupino; 982 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 983 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 984 (long long)daddr, sp->ndupino)); 985 } 986 /* 987 * Account the inode: it no longer belongs to its former segment, 988 * though it will not belong to the new segment until that segment 989 * is actually written. 990 */ 991 if (daddr != LFS_UNUSED_DADDR) { 992 u_int32_t oldsn = lfs_dtosn(fs, daddr); 993 int ndupino __diagused = 994 (sp->seg_number == oldsn) ? sp->ndupino : 0; 995 LFS_SEGENTRY(sup, fs, oldsn, bp); 996 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) 997 >= DINOSIZE(fs)), 998 "lfs_writeinode: negative bytes " 999 "(segment %" PRIu32 " short by %d, " 1000 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1001 ", daddr=%" PRId64 ", su_nbytes=%u, " 1002 "ndupino=%d)\n", 1003 lfs_dtosn(fs, daddr), 1004 (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes, 1005 oldsn, sp->seg_number, daddr, 1006 (unsigned int)sup->su_nbytes, 1007 sp->ndupino); 1008 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1009 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1010 sup->su_nbytes -= DINOSIZE(fs); 1011 redo_ifile |= 1012 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1013 if (redo_ifile) { 1014 mutex_enter(&lfs_lock); 1015 fs->lfs_flags |= LFS_IFDIRTY; 1016 mutex_exit(&lfs_lock); 1017 /* Don't double-account */ 1018 lfs_sb_setidaddr(fs, 0x0); 1019 } 1020 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1021 } 1022 1023 return redo_ifile; 1024 } 1025 1026 int 1027 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1028 { 1029 struct buf *bp; 1030 union lfs_dinode *cdp; 1031 struct vnode *vp = ITOV(ip); 1032 daddr_t daddr; 1033 IINFO *iip; 1034 int i; 1035 int redo_ifile = 0; 1036 int gotblk = 0; 1037 int count; 1038 SEGSUM *ssp; 1039 1040 ASSERT_SEGLOCK(fs); 1041 if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1042 return (0); 1043 1044 /* Can't write ifile when writer is not set */ 1045 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1046 (sp->seg_flags & SEGM_CLEAN)); 1047 1048 /* 1049 * If this is the Ifile, see if writing it here will generate a 1050 * temporary misaccounting. If it will, do the accounting and write 1051 * the blocks, postponing the inode write until the accounting is 1052 * solid. 1053 */ 1054 count = 0; 1055 while (vp == fs->lfs_ivnode) { 1056 int redo = 0; 1057 1058 if (sp->idp == NULL && sp->ibp == NULL && 1059 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1060 sp->sum_bytes_left < sizeof(int32_t))) { 1061 (void) lfs_writeseg(fs, sp); 1062 continue; 1063 } 1064 1065 /* Look for dirty Ifile blocks */ 1066 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1067 if (!(bp->b_flags & B_GATHERED)) { 1068 redo = 1; 1069 break; 1070 } 1071 } 1072 1073 if (redo == 0) 1074 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1075 if (redo == 0) 1076 break; 1077 1078 if (sp->idp) { 1079 lfs_dino_setinumber(fs, sp->idp, 0); 1080 sp->idp = NULL; 1081 } 1082 ++count; 1083 if (count > 2) 1084 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1085 lfs_writefile(fs, sp, fs->lfs_ivnode); 1086 } 1087 1088 /* Allocate a new inode block if necessary. */ 1089 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1090 sp->ibp == NULL) { 1091 /* Allocate a new segment if necessary. */ 1092 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1093 sp->sum_bytes_left < sizeof(int32_t)) 1094 (void) lfs_writeseg(fs, sp); 1095 1096 /* Get next inode block. */ 1097 daddr = lfs_sb_getoffset(fs); 1098 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1099 sp->ibp = *sp->cbpp++ = 1100 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1101 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1102 gotblk++; 1103 1104 /* Zero out inode numbers */ 1105 for (i = 0; i < LFS_INOPB(fs); ++i) { 1106 union lfs_dinode *tmpdi; 1107 1108 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1109 DINOSIZE(fs) * i); 1110 lfs_dino_setinumber(fs, tmpdi, 0); 1111 } 1112 1113 ++sp->start_bpp; 1114 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1115 /* Set remaining space counters. */ 1116 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1117 sp->sum_bytes_left -= sizeof(int32_t); 1118 1119 /* Store the address in the segment summary. */ 1120 iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs)); 1121 lfs_ii_setblock(fs, iip, daddr); 1122 } 1123 1124 /* Check VU_DIROP in case there is a new file with no data blocks */ 1125 if (vp->v_uflag & VU_DIROP) { 1126 ssp = (SEGSUM *)sp->segsum; 1127 lfs_ss_setflags(fs, ssp, 1128 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1129 } 1130 1131 /* Update the inode times and copy the inode onto the inode page. */ 1132 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1133 if (ip->i_number != LFS_IFILE_INUM) 1134 LFS_ITIMES(ip, NULL, NULL, NULL); 1135 1136 /* 1137 * If this is the Ifile, and we've already written the Ifile in this 1138 * partial segment, just overwrite it (it's not on disk yet) and 1139 * continue. 1140 * 1141 * XXX we know that the bp that we get the second time around has 1142 * already been gathered. 1143 */ 1144 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1145 lfs_copy_dinode(fs, sp->idp, ip->i_din); 1146 ip->i_lfs_osize = ip->i_size; 1147 return 0; 1148 } 1149 1150 bp = sp->ibp; 1151 cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs)); 1152 lfs_copy_dinode(fs, cdp, ip->i_din); 1153 1154 /* 1155 * This inode is on its way to disk; clear its VU_DIROP status when 1156 * the write is complete. 1157 */ 1158 if (vp->v_uflag & VU_DIROP) { 1159 if (!(sp->seg_flags & SEGM_CLEAN)) 1160 ip->i_flag |= IN_CDIROP; 1161 else { 1162 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1163 } 1164 } 1165 1166 /* 1167 * If cleaning, link counts and directory file sizes cannot change, 1168 * since those would be directory operations---even if the file 1169 * we are writing is marked VU_DIROP we should write the old values. 1170 * If we're not cleaning, of course, update the values so we get 1171 * current values the next time we clean. 1172 */ 1173 if (sp->seg_flags & SEGM_CLEAN) { 1174 if (vp->v_uflag & VU_DIROP) { 1175 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1176 /* if (vp->v_type == VDIR) */ 1177 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1178 } 1179 } else { 1180 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1181 ip->i_lfs_osize = ip->i_size; 1182 } 1183 1184 1185 /* We can finish the segment accounting for truncations now */ 1186 lfs_finalize_ino_seguse(fs, ip); 1187 1188 /* 1189 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1190 * addresses to disk; possibly change the on-disk record of 1191 * the inode size, either by reverting to the previous size 1192 * (in the case of cleaning) or by verifying the inode's block 1193 * holdings (in the case of files being allocated as they are being 1194 * written). 1195 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1196 * XXX count on disk wrong by the same amount. We should be 1197 * XXX able to "borrow" from lfs_avail and return it after the 1198 * XXX Ifile is written. See also in lfs_writeseg. 1199 */ 1200 1201 /* Check file size based on highest allocated block */ 1202 if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG || 1203 (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) && 1204 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1205 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1206 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1207 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1208 } 1209 if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) { 1210 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1211 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1212 lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs))); 1213 for (i=0; i<ULFS_NDADDR; i++) { 1214 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1215 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1216 lfs_dino_setdb(fs, cdp, i, 0); 1217 } 1218 } 1219 for (i=0; i<ULFS_NIADDR; i++) { 1220 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1221 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1222 lfs_dino_setib(fs, cdp, i, 0); 1223 } 1224 } 1225 } 1226 1227 #ifdef DIAGNOSTIC 1228 /* 1229 * Check dinode held blocks against dinode size. 1230 * This should be identical to the check in lfs_vget(). 1231 */ 1232 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1233 i < ULFS_NDADDR; i++) { 1234 KASSERT(i >= 0); 1235 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1236 continue; 1237 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1238 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1239 continue; 1240 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1241 # ifdef DEBUG 1242 lfs_dump_dinode(fs, cdp); 1243 # endif 1244 panic("writing inconsistent inode"); 1245 } 1246 } 1247 #endif /* DIAGNOSTIC */ 1248 1249 if (ip->i_flag & IN_CLEANING) 1250 LFS_CLR_UINO(ip, IN_CLEANING); 1251 else { 1252 /* XXX IN_ALLMOD */ 1253 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1254 IN_UPDATE | IN_MODIFY); 1255 if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din)) 1256 LFS_CLR_UINO(ip, IN_MODIFIED); 1257 else { 1258 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1259 "blks=%d, eff=%jd\n", ip->i_number, 1260 lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks)); 1261 } 1262 } 1263 1264 if (ip->i_number == LFS_IFILE_INUM) { 1265 /* We know sp->idp == NULL */ 1266 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1267 1268 /* Not dirty any more */ 1269 mutex_enter(&lfs_lock); 1270 fs->lfs_flags &= ~LFS_IFDIRTY; 1271 mutex_exit(&lfs_lock); 1272 } 1273 1274 if (gotblk) { 1275 mutex_enter(&bufcache_lock); 1276 LFS_LOCK_BUF(bp); 1277 brelsel(bp, 0); 1278 mutex_exit(&bufcache_lock); 1279 } 1280 1281 /* Increment inode count in segment summary block. */ 1282 1283 ssp = (SEGSUM *)sp->segsum; 1284 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1285 1286 /* If this page is full, set flag to allocate a new page. */ 1287 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1288 sp->ibp = NULL; 1289 1290 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1291 1292 KASSERT(redo_ifile == 0); 1293 return (redo_ifile); 1294 } 1295 1296 int 1297 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1298 { 1299 struct lfs *fs; 1300 int vers; 1301 int j, blksinblk; 1302 1303 ASSERT_SEGLOCK(sp->fs); 1304 /* 1305 * If full, finish this segment. We may be doing I/O, so 1306 * release and reacquire the splbio(). 1307 */ 1308 KASSERTMSG((sp->vp != NULL), 1309 "lfs_gatherblock: Null vp in segment"); 1310 fs = sp->fs; 1311 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1312 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1313 sp->seg_bytes_left < bp->b_bcount) { 1314 if (mptr) 1315 mutex_exit(mptr); 1316 lfs_updatemeta(sp); 1317 1318 vers = lfs_fi_getversion(fs, sp->fip); 1319 (void) lfs_writeseg(fs, sp); 1320 1321 /* Add the current file to the segment summary. */ 1322 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1323 1324 if (mptr) 1325 mutex_enter(mptr); 1326 return (1); 1327 } 1328 1329 if (bp->b_flags & B_GATHERED) { 1330 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1331 " lbn %" PRId64 "\n", 1332 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1333 return (0); 1334 } 1335 1336 /* Insert into the buffer list, update the FINFO block. */ 1337 bp->b_flags |= B_GATHERED; 1338 1339 *sp->cbpp++ = bp; 1340 for (j = 0; j < blksinblk; j++) { 1341 unsigned bn; 1342 1343 bn = lfs_fi_getnblocks(fs, sp->fip); 1344 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1345 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1346 /* This block's accounting moves from lfs_favail to lfs_avail */ 1347 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1348 } 1349 1350 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1351 sp->seg_bytes_left -= bp->b_bcount; 1352 return (0); 1353 } 1354 1355 int 1356 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1357 int (*match)(struct lfs *, struct buf *)) 1358 { 1359 struct buf *bp, *nbp; 1360 int count = 0; 1361 1362 ASSERT_SEGLOCK(fs); 1363 if (vp->v_type == VBLK) 1364 return 0; 1365 KASSERT(sp->vp == NULL); 1366 sp->vp = vp; 1367 mutex_enter(&bufcache_lock); 1368 1369 #ifndef LFS_NO_BACKBUF_HACK 1370 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1371 # define BUF_OFFSET \ 1372 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1373 # define BACK_BUF(BP) \ 1374 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1375 # define BEG_OF_LIST \ 1376 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1377 1378 loop: 1379 /* Find last buffer. */ 1380 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1381 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1382 bp = LIST_NEXT(bp, b_vnbufs)) 1383 /* nothing */; 1384 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1385 nbp = BACK_BUF(bp); 1386 #else /* LFS_NO_BACKBUF_HACK */ 1387 loop: 1388 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1389 nbp = LIST_NEXT(bp, b_vnbufs); 1390 #endif /* LFS_NO_BACKBUF_HACK */ 1391 if ((bp->b_cflags & BC_BUSY) != 0 || 1392 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1393 #ifdef DEBUG 1394 if (vp == fs->lfs_ivnode && 1395 (bp->b_cflags & BC_BUSY) != 0 && 1396 (bp->b_flags & B_GATHERED) == 0) 1397 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1398 PRId64 " busy (%x) at 0x%jx", 1399 bp->b_lblkno, bp->b_flags, 1400 (uintmax_t)lfs_sb_getoffset(fs)); 1401 #endif 1402 continue; 1403 } 1404 #ifdef DIAGNOSTIC 1405 # ifdef LFS_USE_B_INVAL 1406 if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1407 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1408 " is BC_INVAL\n", bp->b_lblkno)); 1409 VOP_PRINT(bp->b_vp); 1410 } 1411 # endif /* LFS_USE_B_INVAL */ 1412 if (!(bp->b_oflags & BO_DELWRI)) 1413 panic("lfs_gather: bp not BO_DELWRI"); 1414 if (!(bp->b_flags & B_LOCKED)) { 1415 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1416 " blk %" PRId64 " not B_LOCKED\n", 1417 bp->b_lblkno, 1418 LFS_DBTOFSB(fs, bp->b_blkno))); 1419 VOP_PRINT(bp->b_vp); 1420 panic("lfs_gather: bp not B_LOCKED"); 1421 } 1422 #endif 1423 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1424 goto loop; 1425 } 1426 count++; 1427 } 1428 mutex_exit(&bufcache_lock); 1429 lfs_updatemeta(sp); 1430 KASSERT(sp->vp == vp); 1431 sp->vp = NULL; 1432 return count; 1433 } 1434 1435 #if DEBUG 1436 # define DEBUG_OOFF(n) do { \ 1437 if (ooff == 0) { \ 1438 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1439 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1440 ", was 0x0 (or %" PRId64 ")\n", \ 1441 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1442 } \ 1443 } while (0) 1444 #else 1445 # define DEBUG_OOFF(n) 1446 #endif 1447 1448 /* 1449 * Change the given block's address to ndaddr, finding its previous 1450 * location using ulfs_bmaparray(). 1451 * 1452 * Account for this change in the segment table. 1453 * 1454 * called with sp == NULL by roll-forwarding code. 1455 */ 1456 void 1457 lfs_update_single(struct lfs *fs, struct segment *sp, 1458 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1459 { 1460 SEGUSE *sup; 1461 struct buf *bp; 1462 struct indir a[ULFS_NIADDR + 2], *ap; 1463 struct inode *ip; 1464 daddr_t daddr, ooff; 1465 int num, error; 1466 int bb, osize, obb; 1467 1468 ASSERT_SEGLOCK(fs); 1469 KASSERT(sp == NULL || sp->vp == vp); 1470 ip = VTOI(vp); 1471 1472 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1473 if (error) 1474 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1475 1476 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1477 if (daddr > 0) 1478 daddr = LFS_DBTOFSB(fs, daddr); 1479 1480 bb = lfs_numfrags(fs, size); 1481 switch (num) { 1482 case 0: 1483 ooff = lfs_dino_getdb(fs, ip->i_din, lbn); 1484 DEBUG_OOFF(0); 1485 if (ooff == UNWRITTEN) 1486 lfs_dino_setblocks(fs, ip->i_din, 1487 lfs_dino_getblocks(fs, ip->i_din) + bb); 1488 else { 1489 /* possible fragment truncation or extension */ 1490 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1491 lfs_dino_setblocks(fs, ip->i_din, 1492 lfs_dino_getblocks(fs, ip->i_din) + (bb-obb)); 1493 } 1494 lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr); 1495 break; 1496 case 1: 1497 ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off); 1498 DEBUG_OOFF(1); 1499 if (ooff == UNWRITTEN) 1500 lfs_dino_setblocks(fs, ip->i_din, 1501 lfs_dino_getblocks(fs, ip->i_din) + bb); 1502 lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr); 1503 break; 1504 default: 1505 ap = &a[num - 1]; 1506 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1507 B_MODIFY, &bp)) 1508 panic("lfs_updatemeta: bread bno %" PRId64, 1509 ap->in_lbn); 1510 1511 ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off); 1512 DEBUG_OOFF(num); 1513 if (ooff == UNWRITTEN) 1514 lfs_dino_setblocks(fs, ip->i_din, 1515 lfs_dino_getblocks(fs, ip->i_din) + bb); 1516 lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr); 1517 (void) VOP_BWRITE(bp->b_vp, bp); 1518 } 1519 1520 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1521 1522 /* Update hiblk when extending the file */ 1523 if (lbn > ip->i_lfs_hiblk) 1524 ip->i_lfs_hiblk = lbn; 1525 1526 /* 1527 * Though we'd rather it couldn't, this *can* happen right now 1528 * if cleaning blocks and regular blocks coexist. 1529 */ 1530 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1531 1532 /* 1533 * Update segment usage information, based on old size 1534 * and location. 1535 */ 1536 if (daddr > 0) { 1537 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1538 int ndupino __diagused = (sp && sp->seg_number == oldsn ? 1539 sp->ndupino : 0); 1540 1541 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1542 if (lbn >= 0 && lbn < ULFS_NDADDR) 1543 osize = ip->i_lfs_fragsize[lbn]; 1544 else 1545 osize = lfs_sb_getbsize(fs); 1546 LFS_SEGENTRY(sup, fs, oldsn, bp); 1547 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize), 1548 "lfs_updatemeta: negative bytes " 1549 "(segment %" PRIu32 " short by %" PRId64 1550 ")\n" 1551 "lfs_updatemeta: ino %llu, lbn %" PRId64 1552 ", addr = 0x%" PRIx64 "\n" 1553 "lfs_updatemeta: ndupino=%d", 1554 lfs_dtosn(fs, daddr), 1555 (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes), 1556 (unsigned long long)ip->i_number, lbn, daddr, 1557 ndupino); 1558 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1559 " db 0x%" PRIx64 "\n", 1560 lfs_dtosn(fs, daddr), osize, 1561 ip->i_number, lbn, daddr)); 1562 sup->su_nbytes -= osize; 1563 if (!(bp->b_flags & B_GATHERED)) { 1564 mutex_enter(&lfs_lock); 1565 fs->lfs_flags |= LFS_IFDIRTY; 1566 mutex_exit(&lfs_lock); 1567 } 1568 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1569 } 1570 /* 1571 * Now that this block has a new address, and its old 1572 * segment no longer owns it, we can forget about its 1573 * old size. 1574 */ 1575 if (lbn >= 0 && lbn < ULFS_NDADDR) 1576 ip->i_lfs_fragsize[lbn] = size; 1577 } 1578 1579 /* 1580 * Update the metadata that points to the blocks listed in the FINFO 1581 * array. 1582 */ 1583 void 1584 lfs_updatemeta(struct segment *sp) 1585 { 1586 struct buf *sbp; 1587 struct lfs *fs; 1588 struct vnode *vp; 1589 daddr_t lbn; 1590 int i, nblocks, num; 1591 int __diagused nblocks_orig; 1592 int bb; 1593 int bytesleft, size; 1594 unsigned lastlength; 1595 union lfs_blocks tmpptr; 1596 1597 fs = sp->fs; 1598 vp = sp->vp; 1599 ASSERT_SEGLOCK(fs); 1600 1601 /* 1602 * This used to be: 1603 * 1604 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1605 * 1606 * that is, it allowed for the possibility that start_lbp did 1607 * not point to the beginning of the finfo block pointer area. 1608 * This particular formulation is six kinds of painful in the 1609 * lfs64 world where we have two sizes of block pointer, so 1610 * unless/until everything can be cleaned up to not move 1611 * start_lbp around but instead use an offset, we do the 1612 * following: 1613 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1614 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1615 * type. (Ugh.) 1616 * 2. Cast it to void *, then assign it to a temporary 1617 * union lfs_blocks. 1618 * 3. Subtract start_lbp from that. 1619 * 4. Save the value of nblocks in blocks_orig so we can 1620 * assert below that it hasn't changed without repeating this 1621 * rubbish. 1622 * 1623 * XXX. 1624 */ 1625 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1626 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1627 nblocks_orig = nblocks; 1628 1629 KASSERT(nblocks >= 0); 1630 KASSERT(vp != NULL); 1631 if (nblocks == 0) 1632 return; 1633 1634 /* 1635 * This count may be high due to oversize blocks from lfs_gop_write. 1636 * Correct for this. (XXX we should be able to keep track of these.) 1637 */ 1638 for (i = 0; i < nblocks; i++) { 1639 if (sp->start_bpp[i] == NULL) { 1640 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1641 nblocks = i; 1642 break; 1643 } 1644 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1645 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1646 nblocks -= num - 1; 1647 } 1648 1649 #if 0 1650 /* pre-lfs64 assertion */ 1651 KASSERT(vp->v_type == VREG || 1652 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1653 #else 1654 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1655 #endif 1656 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1657 1658 /* 1659 * Sort the blocks. 1660 * 1661 * We have to sort even if the blocks come from the 1662 * cleaner, because there might be other pending blocks on the 1663 * same inode...and if we don't sort, and there are fragments 1664 * present, blocks may be written in the wrong place. 1665 */ 1666 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1667 1668 /* 1669 * Record the length of the last block in case it's a fragment. 1670 * If there are indirect blocks present, they sort last. An 1671 * indirect block will be lfs_bsize and its presence indicates 1672 * that you cannot have fragments. 1673 * 1674 * XXX This last is a lie. A cleaned fragment can coexist with 1675 * XXX a later indirect block. This will continue to be 1676 * XXX true until lfs_markv is fixed to do everything with 1677 * XXX fake blocks (including fake inodes and fake indirect blocks). 1678 */ 1679 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1680 lfs_sb_getbmask(fs)) + 1; 1681 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1682 1683 /* 1684 * Assign disk addresses, and update references to the logical 1685 * block and the segment usage information. 1686 */ 1687 for (i = nblocks; i--; ++sp->start_bpp) { 1688 sbp = *sp->start_bpp; 1689 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1690 KASSERT(sbp->b_lblkno == lbn); 1691 1692 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1693 1694 /* 1695 * If we write a frag in the wrong place, the cleaner won't 1696 * be able to correctly identify its size later, and the 1697 * segment will be uncleanable. (Even worse, it will assume 1698 * that the indirect block that actually ends the list 1699 * is of a smaller size!) 1700 */ 1701 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1702 panic("lfs_updatemeta: fragment is not last block"); 1703 1704 /* 1705 * For each subblock in this possibly oversized block, 1706 * update its address on disk. 1707 */ 1708 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1709 KASSERT(vp == sbp->b_vp); 1710 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1711 bytesleft -= lfs_sb_getbsize(fs)) { 1712 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1713 bb = lfs_numfrags(fs, size); 1714 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1715 lfs_blocks_inc(fs, &sp->start_lbp); 1716 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1717 size); 1718 lfs_sb_addoffset(fs, bb); 1719 } 1720 1721 } 1722 1723 /* This inode has been modified */ 1724 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1725 } 1726 1727 /* 1728 * Move lfs_offset to a segment earlier than newsn. 1729 */ 1730 int 1731 lfs_rewind(struct lfs *fs, int newsn) 1732 { 1733 int sn, osn, isdirty; 1734 struct buf *bp; 1735 SEGUSE *sup; 1736 1737 ASSERT_SEGLOCK(fs); 1738 1739 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1740 if (osn < newsn) 1741 return 0; 1742 1743 /* lfs_avail eats the remaining space in this segment */ 1744 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1745 1746 /* Find a low-numbered segment */ 1747 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1748 LFS_SEGENTRY(sup, fs, sn, bp); 1749 isdirty = sup->su_flags & SEGUSE_DIRTY; 1750 brelse(bp, 0); 1751 1752 if (!isdirty) 1753 break; 1754 } 1755 if (sn == lfs_sb_getnseg(fs)) 1756 panic("lfs_rewind: no clean segments"); 1757 if (newsn >= 0 && sn >= newsn) 1758 return ENOENT; 1759 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1760 lfs_newseg(fs); 1761 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1762 1763 return 0; 1764 } 1765 1766 /* 1767 * Start a new partial segment. 1768 * 1769 * Return 1 when we entered to a new segment. 1770 * Otherwise, return 0. 1771 */ 1772 int 1773 lfs_initseg(struct lfs *fs) 1774 { 1775 struct segment *sp = fs->lfs_sp; 1776 SEGSUM *ssp; 1777 struct buf *sbp; /* buffer for SEGSUM */ 1778 int repeat = 0; /* return value */ 1779 1780 ASSERT_SEGLOCK(fs); 1781 /* Advance to the next segment. */ 1782 if (!LFS_PARTIAL_FITS(fs)) { 1783 SEGUSE *sup; 1784 struct buf *bp; 1785 1786 /* lfs_avail eats the remaining space */ 1787 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1788 lfs_sb_getcurseg(fs))); 1789 /* Wake up any cleaning procs waiting on this file system. */ 1790 lfs_wakeup_cleaner(fs); 1791 lfs_newseg(fs); 1792 repeat = 1; 1793 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1794 1795 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1796 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1797 1798 /* 1799 * If the segment contains a superblock, update the offset 1800 * and summary address to skip over it. 1801 */ 1802 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1803 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1804 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1805 sp->seg_bytes_left -= LFS_SBPAD; 1806 } 1807 brelse(bp, 0); 1808 /* Segment zero could also contain the labelpad */ 1809 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1810 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1811 lfs_sb_addoffset(fs, 1812 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1813 sp->seg_bytes_left -= 1814 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1815 } 1816 } else { 1817 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1818 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1819 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1820 } 1821 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1822 1823 /* Record first address of this partial segment */ 1824 if (sp->seg_flags & SEGM_CLEAN) { 1825 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1826 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1827 /* "1" is the artificial inc in lfs_seglock */ 1828 mutex_enter(&lfs_lock); 1829 while (fs->lfs_iocount > 1) { 1830 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1831 "lfs_initseg", 0, &lfs_lock); 1832 } 1833 mutex_exit(&lfs_lock); 1834 fs->lfs_cleanind = 0; 1835 } 1836 } 1837 1838 sp->fs = fs; 1839 sp->ibp = NULL; 1840 sp->idp = NULL; 1841 sp->ninodes = 0; 1842 sp->ndupino = 0; 1843 1844 sp->cbpp = sp->bpp; 1845 1846 /* Get a new buffer for SEGSUM */ 1847 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1848 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1849 1850 /* ... and enter it into the buffer list. */ 1851 *sp->cbpp = sbp; 1852 sp->cbpp++; 1853 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1854 1855 sp->start_bpp = sp->cbpp; 1856 1857 /* Set point to SEGSUM, initialize it. */ 1858 ssp = sp->segsum = sbp->b_data; 1859 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1860 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1861 lfs_ss_setnfinfo(fs, ssp, 0); 1862 lfs_ss_setninos(fs, ssp, 0); 1863 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1864 1865 /* Set pointer to first FINFO, initialize it. */ 1866 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1867 lfs_fi_setnblocks(fs, sp->fip, 0); 1868 lfs_fi_setlastlength(fs, sp->fip, 0); 1869 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1870 1871 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1872 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1873 1874 return (repeat); 1875 } 1876 1877 /* 1878 * Remove SEGUSE_INVAL from all segments. 1879 */ 1880 void 1881 lfs_unset_inval_all(struct lfs *fs) 1882 { 1883 SEGUSE *sup; 1884 struct buf *bp; 1885 int i; 1886 1887 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1888 LFS_SEGENTRY(sup, fs, i, bp); 1889 if (sup->su_flags & SEGUSE_INVAL) { 1890 sup->su_flags &= ~SEGUSE_INVAL; 1891 LFS_WRITESEGENTRY(sup, fs, i, bp); 1892 } else 1893 brelse(bp, 0); 1894 } 1895 } 1896 1897 /* 1898 * Return the next segment to write. 1899 */ 1900 void 1901 lfs_newseg(struct lfs *fs) 1902 { 1903 CLEANERINFO *cip; 1904 SEGUSE *sup; 1905 struct buf *bp; 1906 int curseg, isdirty, sn, skip_inval; 1907 1908 ASSERT_SEGLOCK(fs); 1909 1910 /* Honor LFCNWRAPSTOP */ 1911 mutex_enter(&lfs_lock); 1912 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1913 if (fs->lfs_wrappass) { 1914 log(LOG_NOTICE, "%s: wrappass=%d\n", 1915 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1916 fs->lfs_wrappass = 0; 1917 break; 1918 } 1919 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1920 wakeup(&fs->lfs_nowrap); 1921 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1922 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1923 &lfs_lock); 1924 } 1925 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1926 mutex_exit(&lfs_lock); 1927 1928 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1929 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1930 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1931 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1932 sup->su_nbytes = 0; 1933 sup->su_nsums = 0; 1934 sup->su_ninos = 0; 1935 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1936 1937 LFS_CLEANERINFO(cip, fs, bp); 1938 lfs_ci_shiftcleantodirty(fs, cip, 1); 1939 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1940 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1941 1942 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1943 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1944 skip_inval = 1; 1945 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1946 sn = (sn + 1) % lfs_sb_getnseg(fs); 1947 1948 if (sn == curseg) { 1949 if (skip_inval) 1950 skip_inval = 0; 1951 else 1952 panic("lfs_nextseg: no clean segments"); 1953 } 1954 LFS_SEGENTRY(sup, fs, sn, bp); 1955 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1956 /* Check SEGUSE_EMPTY as we go along */ 1957 if (isdirty && sup->su_nbytes == 0 && 1958 !(sup->su_flags & SEGUSE_EMPTY)) 1959 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1960 else 1961 brelse(bp, 0); 1962 1963 if (!isdirty) 1964 break; 1965 } 1966 if (skip_inval == 0) 1967 lfs_unset_inval_all(fs); 1968 1969 ++fs->lfs_nactive; 1970 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1971 if (lfs_dostats) { 1972 ++lfs_stats.segsused; 1973 } 1974 } 1975 1976 static struct buf * 1977 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1978 int n) 1979 { 1980 struct lfs_cluster *cl; 1981 struct buf **bpp, *bp; 1982 1983 ASSERT_SEGLOCK(fs); 1984 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1985 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1986 memset(cl, 0, sizeof(*cl)); 1987 cl->fs = fs; 1988 cl->bpp = bpp; 1989 cl->bufcount = 0; 1990 cl->bufsize = 0; 1991 1992 /* If this segment is being written synchronously, note that */ 1993 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 1994 cl->flags |= LFS_CL_SYNC; 1995 cl->seg = fs->lfs_sp; 1996 ++cl->seg->seg_iocount; 1997 } 1998 1999 /* Get an empty buffer header, or maybe one with something on it */ 2000 bp = getiobuf(vp, true); 2001 bp->b_dev = NODEV; 2002 bp->b_blkno = bp->b_lblkno = addr; 2003 bp->b_iodone = lfs_cluster_callback; 2004 bp->b_private = cl; 2005 2006 return bp; 2007 } 2008 2009 int 2010 lfs_writeseg(struct lfs *fs, struct segment *sp) 2011 { 2012 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2013 SEGUSE *sup; 2014 SEGSUM *ssp; 2015 int i; 2016 int do_again, nblocks, byteoffset; 2017 size_t el_size; 2018 struct lfs_cluster *cl; 2019 u_short ninos; 2020 struct vnode *devvp; 2021 char *p = NULL; 2022 struct vnode *vp; 2023 unsigned ibindex, iblimit; 2024 int changed; 2025 u_int32_t sum; 2026 size_t sumstart; 2027 #ifdef DEBUG 2028 FINFO *fip; 2029 int findex; 2030 #endif 2031 2032 ASSERT_SEGLOCK(fs); 2033 2034 ssp = (SEGSUM *)sp->segsum; 2035 2036 /* 2037 * If there are no buffers other than the segment summary to write, 2038 * don't do anything. If we are the end of a dirop sequence, however, 2039 * write the empty segment summary anyway, to help out the 2040 * roll-forward agent. 2041 */ 2042 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2043 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2044 return 0; 2045 } 2046 2047 /* Note if partial segment is being written by the cleaner */ 2048 if (sp->seg_flags & SEGM_CLEAN) 2049 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2050 2051 /* Note if we are writing to reclaim */ 2052 if (sp->seg_flags & SEGM_RECLAIM) { 2053 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2054 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2055 } 2056 2057 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2058 2059 /* Update the segment usage information. */ 2060 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2061 2062 /* Loop through all blocks, except the segment summary. */ 2063 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2064 if ((*bpp)->b_vp != devvp) { 2065 sup->su_nbytes += (*bpp)->b_bcount; 2066 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2067 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2068 sp->seg_number, (*bpp)->b_bcount, 2069 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2070 (*bpp)->b_blkno)); 2071 } 2072 } 2073 2074 #ifdef DEBUG 2075 /* Check for zero-length and zero-version FINFO entries. */ 2076 fip = SEGSUM_FINFOBASE(fs, ssp); 2077 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2078 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2079 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2080 fip = NEXT_FINFO(fs, fip); 2081 } 2082 #endif /* DEBUG */ 2083 2084 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2085 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2086 sp->seg_number, 2087 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2088 lfs_ss_getninos(fs, ssp))); 2089 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2090 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2091 if (lfs_sb_getversion(fs) == 1) 2092 sup->su_olastmod = time_second; 2093 else 2094 sup->su_lastmod = time_second; 2095 sup->su_ninos += ninos; 2096 ++sup->su_nsums; 2097 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2098 2099 do_again = !(bp->b_flags & B_GATHERED); 2100 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2101 2102 /* 2103 * Mark blocks B_BUSY, to prevent then from being changed between 2104 * the checksum computation and the actual write. 2105 * 2106 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2107 * there are any, replace them with copies that have UNASSIGNED 2108 * instead. 2109 */ 2110 mutex_enter(&bufcache_lock); 2111 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2112 ++bpp; 2113 bp = *bpp; 2114 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2115 bp->b_cflags |= BC_BUSY; 2116 continue; 2117 } 2118 2119 while (bp->b_cflags & BC_BUSY) { 2120 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2121 " data summary corruption for ino %d, lbn %" 2122 PRId64 "\n", 2123 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2124 bp->b_cflags |= BC_WANTED; 2125 cv_wait(&bp->b_busy, &bufcache_lock); 2126 } 2127 bp->b_cflags |= BC_BUSY; 2128 mutex_exit(&bufcache_lock); 2129 unbusybp = NULL; 2130 2131 /* 2132 * Check and replace indirect block UNWRITTEN bogosity. 2133 * XXX See comment in lfs_writefile. 2134 */ 2135 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2136 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) != 2137 VTOI(bp->b_vp)->i_lfs_effnblks) { 2138 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2139 VTOI(bp->b_vp)->i_number, 2140 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2141 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din))); 2142 /* Make a copy we'll make changes to */ 2143 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2144 bp->b_bcount, LFS_NB_IBLOCK); 2145 newbp->b_blkno = bp->b_blkno; 2146 memcpy(newbp->b_data, bp->b_data, 2147 newbp->b_bcount); 2148 2149 changed = 0; 2150 iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs); 2151 for (ibindex = 0; ibindex < iblimit; ibindex++) { 2152 if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) { 2153 ++changed; 2154 lfs_iblock_set(fs, newbp->b_data, 2155 ibindex, 0); 2156 } 2157 } 2158 /* 2159 * Get rid of the old buffer. Don't mark it clean, 2160 * though, if it still has dirty data on it. 2161 */ 2162 if (changed) { 2163 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2164 " bp = %p newbp = %p\n", changed, bp, 2165 newbp)); 2166 *bpp = newbp; 2167 bp->b_flags &= ~B_GATHERED; 2168 bp->b_error = 0; 2169 if (bp->b_iodone != NULL) { 2170 DLOG((DLOG_SEG, "lfs_writeseg: " 2171 "indir bp should not be B_CALL\n")); 2172 biodone(bp); 2173 bp = NULL; 2174 } else { 2175 /* Still on free list, leave it there */ 2176 unbusybp = bp; 2177 /* 2178 * We have to re-decrement lfs_avail 2179 * since this block is going to come 2180 * back around to us in the next 2181 * segment. 2182 */ 2183 lfs_sb_subavail(fs, 2184 lfs_btofsb(fs, bp->b_bcount)); 2185 } 2186 } else { 2187 lfs_freebuf(fs, newbp); 2188 } 2189 } 2190 mutex_enter(&bufcache_lock); 2191 if (unbusybp != NULL) { 2192 unbusybp->b_cflags &= ~BC_BUSY; 2193 if (unbusybp->b_cflags & BC_WANTED) 2194 cv_broadcast(&bp->b_busy); 2195 } 2196 } 2197 mutex_exit(&bufcache_lock); 2198 2199 /* 2200 * Compute checksum across data and then across summary; the first 2201 * block (the summary block) is skipped. Set the create time here 2202 * so that it's guaranteed to be later than the inode mod times. 2203 */ 2204 sum = 0; 2205 if (lfs_sb_getversion(fs) == 1) 2206 el_size = sizeof(u_long); 2207 else 2208 el_size = sizeof(u_int32_t); 2209 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2210 ++bpp; 2211 /* Loop through gop_write cluster blocks */ 2212 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2213 byteoffset += lfs_sb_getbsize(fs)) { 2214 #ifdef LFS_USE_B_INVAL 2215 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2216 (*bpp)->b_iodone != NULL) { 2217 if (copyin((void *)(*bpp)->b_saveaddr + 2218 byteoffset, dp, el_size)) { 2219 panic("lfs_writeseg: copyin failed [1]:" 2220 " ino %" PRIu64 " blk %" PRId64, 2221 VTOI((*bpp)->b_vp)->i_number, 2222 (*bpp)->b_lblkno); 2223 } 2224 } else 2225 #endif /* LFS_USE_B_INVAL */ 2226 { 2227 sum = lfs_cksum_part((char *) 2228 (*bpp)->b_data + byteoffset, el_size, sum); 2229 } 2230 } 2231 } 2232 if (lfs_sb_getversion(fs) == 1) 2233 lfs_ss_setocreate(fs, ssp, time_second); 2234 else { 2235 lfs_ss_setcreate(fs, ssp, time_second); 2236 lfs_sb_addserial(fs, 1); 2237 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2238 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2239 } 2240 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2241 sumstart = lfs_ss_getsumstart(fs); 2242 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2243 lfs_sb_getsumsize(fs) - sumstart)); 2244 2245 mutex_enter(&lfs_lock); 2246 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2247 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2248 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2249 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2250 mutex_exit(&lfs_lock); 2251 2252 /* 2253 * When we simply write the blocks we lose a rotation for every block 2254 * written. To avoid this problem, we cluster the buffers into a 2255 * chunk and write the chunk. MAXPHYS is the largest size I/O 2256 * devices can handle, use that for the size of the chunks. 2257 * 2258 * Blocks that are already clusters (from GOP_WRITE), however, we 2259 * don't bother to copy into other clusters. 2260 */ 2261 2262 #define CHUNKSIZE MAXPHYS 2263 2264 if (devvp == NULL) 2265 panic("devvp is NULL"); 2266 for (bpp = sp->bpp, i = nblocks; i;) { 2267 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2268 cl = cbp->b_private; 2269 2270 cbp->b_flags |= B_ASYNC; 2271 cbp->b_cflags |= BC_BUSY; 2272 cbp->b_bcount = 0; 2273 2274 KASSERTMSG((bpp - sp->bpp <= 2275 (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2276 / sizeof(int32_t)), 2277 "lfs_writeseg: real bpp overwrite"); 2278 KASSERTMSG((bpp - sp->bpp <= 2279 lfs_segsize(fs) / lfs_sb_getfsize(fs)), 2280 "lfs_writeseg: theoretical bpp overwrite"); 2281 2282 /* 2283 * Construct the cluster. 2284 */ 2285 mutex_enter(&lfs_lock); 2286 ++fs->lfs_iocount; 2287 mutex_exit(&lfs_lock); 2288 while (i && cbp->b_bcount < CHUNKSIZE) { 2289 bp = *bpp; 2290 2291 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2292 break; 2293 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2294 break; 2295 2296 /* Clusters from GOP_WRITE are expedited */ 2297 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2298 if (cbp->b_bcount > 0) 2299 /* Put in its own buffer */ 2300 break; 2301 else { 2302 cbp->b_data = bp->b_data; 2303 } 2304 } else if (cbp->b_bcount == 0) { 2305 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2306 LFS_NB_CLUSTER); 2307 cl->flags |= LFS_CL_MALLOC; 2308 } 2309 KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2310 btodb(bp->b_bcount - 1))) == 2311 sp->seg_number), 2312 "segment overwrite: blk size %d daddr %" PRIx64 2313 " not in seg %d\n", 2314 bp->b_bcount, bp->b_blkno, 2315 sp->seg_number); 2316 2317 #ifdef LFS_USE_B_INVAL 2318 /* 2319 * Fake buffers from the cleaner are marked as B_INVAL. 2320 * We need to copy the data from user space rather than 2321 * from the buffer indicated. 2322 * XXX == what do I do on an error? 2323 */ 2324 if ((bp->b_cflags & BC_INVAL) != 0 && 2325 bp->b_iodone != NULL) { 2326 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2327 panic("lfs_writeseg: " 2328 "copyin failed [2]"); 2329 } else 2330 #endif /* LFS_USE_B_INVAL */ 2331 if (cl->flags & LFS_CL_MALLOC) { 2332 /* copy data into our cluster. */ 2333 memcpy(p, bp->b_data, bp->b_bcount); 2334 p += bp->b_bcount; 2335 } 2336 2337 cbp->b_bcount += bp->b_bcount; 2338 cl->bufsize += bp->b_bcount; 2339 2340 bp->b_flags &= ~B_READ; 2341 bp->b_error = 0; 2342 cl->bpp[cl->bufcount++] = bp; 2343 2344 vp = bp->b_vp; 2345 mutex_enter(&bufcache_lock); 2346 mutex_enter(vp->v_interlock); 2347 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2348 reassignbuf(bp, vp); 2349 vp->v_numoutput++; 2350 mutex_exit(vp->v_interlock); 2351 mutex_exit(&bufcache_lock); 2352 2353 bpp++; 2354 i--; 2355 } 2356 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2357 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2358 else 2359 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2360 mutex_enter(devvp->v_interlock); 2361 devvp->v_numoutput++; 2362 mutex_exit(devvp->v_interlock); 2363 VOP_STRATEGY(devvp, cbp); 2364 curlwp->l_ru.ru_oublock++; 2365 } 2366 2367 if (lfs_dostats) { 2368 ++lfs_stats.psegwrites; 2369 lfs_stats.blocktot += nblocks - 1; 2370 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2371 ++lfs_stats.psyncwrites; 2372 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2373 ++lfs_stats.pcleanwrites; 2374 lfs_stats.cleanblocks += nblocks - 1; 2375 } 2376 } 2377 2378 return (lfs_initseg(fs) || do_again); 2379 } 2380 2381 void 2382 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2383 { 2384 struct buf *bp; 2385 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2386 int s; 2387 2388 ASSERT_MAYBE_SEGLOCK(fs); 2389 if (fs->lfs_is64) { 2390 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2391 } else { 2392 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2393 } 2394 /* 2395 * If we can write one superblock while another is in 2396 * progress, we risk not having a complete checkpoint if we crash. 2397 * So, block here if a superblock write is in progress. 2398 */ 2399 mutex_enter(&lfs_lock); 2400 s = splbio(); 2401 while (fs->lfs_sbactive) { 2402 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2403 &lfs_lock); 2404 } 2405 fs->lfs_sbactive = daddr; 2406 splx(s); 2407 mutex_exit(&lfs_lock); 2408 2409 /* Set timestamp of this version of the superblock */ 2410 if (lfs_sb_getversion(fs) == 1) 2411 lfs_sb_setotstamp(fs, time_second); 2412 lfs_sb_settstamp(fs, time_second); 2413 2414 /* The next chunk of code relies on this assumption */ 2415 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2416 2417 /* Checksum the superblock and copy it into a buffer. */ 2418 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2419 bp = lfs_newbuf(fs, devvp, 2420 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2421 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2422 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2423 LFS_SBPAD - sizeof(struct dlfs)); 2424 2425 bp->b_cflags |= BC_BUSY; 2426 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2427 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2428 bp->b_error = 0; 2429 bp->b_iodone = lfs_supercallback; 2430 2431 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2432 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2433 else 2434 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2435 curlwp->l_ru.ru_oublock++; 2436 2437 mutex_enter(devvp->v_interlock); 2438 devvp->v_numoutput++; 2439 mutex_exit(devvp->v_interlock); 2440 2441 mutex_enter(&lfs_lock); 2442 ++fs->lfs_iocount; 2443 mutex_exit(&lfs_lock); 2444 VOP_STRATEGY(devvp, bp); 2445 } 2446 2447 /* 2448 * Logical block number match routines used when traversing the dirty block 2449 * chain. 2450 */ 2451 int 2452 lfs_match_fake(struct lfs *fs, struct buf *bp) 2453 { 2454 2455 ASSERT_SEGLOCK(fs); 2456 return LFS_IS_MALLOC_BUF(bp); 2457 } 2458 2459 #if 0 2460 int 2461 lfs_match_real(struct lfs *fs, struct buf *bp) 2462 { 2463 2464 ASSERT_SEGLOCK(fs); 2465 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2466 } 2467 #endif 2468 2469 int 2470 lfs_match_data(struct lfs *fs, struct buf *bp) 2471 { 2472 2473 ASSERT_SEGLOCK(fs); 2474 return (bp->b_lblkno >= 0); 2475 } 2476 2477 int 2478 lfs_match_indir(struct lfs *fs, struct buf *bp) 2479 { 2480 daddr_t lbn; 2481 2482 ASSERT_SEGLOCK(fs); 2483 lbn = bp->b_lblkno; 2484 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2485 } 2486 2487 int 2488 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2489 { 2490 daddr_t lbn; 2491 2492 ASSERT_SEGLOCK(fs); 2493 lbn = bp->b_lblkno; 2494 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2495 } 2496 2497 int 2498 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2499 { 2500 daddr_t lbn; 2501 2502 ASSERT_SEGLOCK(fs); 2503 lbn = bp->b_lblkno; 2504 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2505 } 2506 2507 static void 2508 lfs_free_aiodone(struct buf *bp) 2509 { 2510 struct lfs *fs; 2511 2512 KERNEL_LOCK(1, curlwp); 2513 fs = bp->b_private; 2514 ASSERT_NO_SEGLOCK(fs); 2515 lfs_freebuf(fs, bp); 2516 KERNEL_UNLOCK_LAST(curlwp); 2517 } 2518 2519 static void 2520 lfs_super_aiodone(struct buf *bp) 2521 { 2522 struct lfs *fs; 2523 2524 KERNEL_LOCK(1, curlwp); 2525 fs = bp->b_private; 2526 ASSERT_NO_SEGLOCK(fs); 2527 mutex_enter(&lfs_lock); 2528 fs->lfs_sbactive = 0; 2529 if (--fs->lfs_iocount <= 1) 2530 wakeup(&fs->lfs_iocount); 2531 wakeup(&fs->lfs_sbactive); 2532 mutex_exit(&lfs_lock); 2533 lfs_freebuf(fs, bp); 2534 KERNEL_UNLOCK_LAST(curlwp); 2535 } 2536 2537 static void 2538 lfs_cluster_aiodone(struct buf *bp) 2539 { 2540 struct lfs_cluster *cl; 2541 struct lfs *fs; 2542 struct buf *tbp, *fbp; 2543 struct vnode *vp, *devvp, *ovp; 2544 struct inode *ip; 2545 int error; 2546 2547 KERNEL_LOCK(1, curlwp); 2548 2549 error = bp->b_error; 2550 cl = bp->b_private; 2551 fs = cl->fs; 2552 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2553 ASSERT_NO_SEGLOCK(fs); 2554 2555 /* Put the pages back, and release the buffer */ 2556 while (cl->bufcount--) { 2557 tbp = cl->bpp[cl->bufcount]; 2558 KASSERT(tbp->b_cflags & BC_BUSY); 2559 if (error) { 2560 tbp->b_error = error; 2561 } 2562 2563 /* 2564 * We're done with tbp. If it has not been re-dirtied since 2565 * the cluster was written, free it. Otherwise, keep it on 2566 * the locked list to be written again. 2567 */ 2568 vp = tbp->b_vp; 2569 2570 tbp->b_flags &= ~B_GATHERED; 2571 2572 LFS_BCLEAN_LOG(fs, tbp); 2573 2574 mutex_enter(&bufcache_lock); 2575 if (tbp->b_iodone == NULL) { 2576 KASSERT(tbp->b_flags & B_LOCKED); 2577 bremfree(tbp); 2578 if (vp) { 2579 mutex_enter(vp->v_interlock); 2580 reassignbuf(tbp, vp); 2581 mutex_exit(vp->v_interlock); 2582 } 2583 tbp->b_flags |= B_ASYNC; /* for biodone */ 2584 } 2585 2586 if (((tbp->b_flags | tbp->b_oflags) & 2587 (B_LOCKED | BO_DELWRI)) == B_LOCKED) 2588 LFS_UNLOCK_BUF(tbp); 2589 2590 if (tbp->b_oflags & BO_DONE) { 2591 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2592 cl->bufcount, (long)tbp->b_flags)); 2593 } 2594 2595 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2596 /* 2597 * A buffer from the page daemon. 2598 * We use the same iodone as it does, 2599 * so we must manually disassociate its 2600 * buffers from the vp. 2601 */ 2602 if ((ovp = tbp->b_vp) != NULL) { 2603 /* This is just silly */ 2604 mutex_enter(ovp->v_interlock); 2605 brelvp(tbp); 2606 mutex_exit(ovp->v_interlock); 2607 tbp->b_vp = vp; 2608 tbp->b_objlock = vp->v_interlock; 2609 } 2610 /* Put it back the way it was */ 2611 tbp->b_flags |= B_ASYNC; 2612 /* Master buffers have BC_AGE */ 2613 if (tbp->b_private == tbp) 2614 tbp->b_cflags |= BC_AGE; 2615 } 2616 mutex_exit(&bufcache_lock); 2617 2618 biodone(tbp); 2619 2620 /* 2621 * If this is the last block for this vnode, but 2622 * there are other blocks on its dirty list, 2623 * set IN_MODIFIED/IN_CLEANING depending on what 2624 * sort of block. Only do this for our mount point, 2625 * not for, e.g., inode blocks that are attached to 2626 * the devvp. 2627 * XXX KS - Shouldn't we set *both* if both types 2628 * of blocks are present (traverse the dirty list?) 2629 */ 2630 mutex_enter(vp->v_interlock); 2631 mutex_enter(&lfs_lock); 2632 if (vp != devvp && vp->v_numoutput == 0 && 2633 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2634 ip = VTOI(vp); 2635 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2636 ip->i_number)); 2637 if (LFS_IS_MALLOC_BUF(fbp)) 2638 LFS_SET_UINO(ip, IN_CLEANING); 2639 else 2640 LFS_SET_UINO(ip, IN_MODIFIED); 2641 } 2642 cv_broadcast(&vp->v_cv); 2643 mutex_exit(&lfs_lock); 2644 mutex_exit(vp->v_interlock); 2645 } 2646 2647 /* Fix up the cluster buffer, and release it */ 2648 if (cl->flags & LFS_CL_MALLOC) 2649 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2650 putiobuf(bp); 2651 2652 /* Note i/o done */ 2653 if (cl->flags & LFS_CL_SYNC) { 2654 if (--cl->seg->seg_iocount == 0) 2655 wakeup(&cl->seg->seg_iocount); 2656 } 2657 mutex_enter(&lfs_lock); 2658 KASSERTMSG((fs->lfs_iocount != 0), 2659 "lfs_cluster_aiodone: zero iocount"); 2660 if (--fs->lfs_iocount <= 1) 2661 wakeup(&fs->lfs_iocount); 2662 mutex_exit(&lfs_lock); 2663 2664 KERNEL_UNLOCK_LAST(curlwp); 2665 2666 pool_put(&fs->lfs_bpppool, cl->bpp); 2667 cl->bpp = NULL; 2668 pool_put(&fs->lfs_clpool, cl); 2669 } 2670 2671 static void 2672 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2673 { 2674 /* reset b_iodone for when this is a single-buf i/o. */ 2675 bp->b_iodone = aiodone; 2676 2677 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2678 } 2679 2680 static void 2681 lfs_cluster_callback(struct buf *bp) 2682 { 2683 2684 lfs_generic_callback(bp, lfs_cluster_aiodone); 2685 } 2686 2687 void 2688 lfs_supercallback(struct buf *bp) 2689 { 2690 2691 lfs_generic_callback(bp, lfs_super_aiodone); 2692 } 2693 2694 /* 2695 * The only buffers that are going to hit these functions are the 2696 * segment write blocks, or the segment summaries, or the superblocks. 2697 * 2698 * All of the above are created by lfs_newbuf, and so do not need to be 2699 * released via brelse. 2700 */ 2701 void 2702 lfs_callback(struct buf *bp) 2703 { 2704 2705 lfs_generic_callback(bp, lfs_free_aiodone); 2706 } 2707 2708 /* 2709 * Shellsort (diminishing increment sort) from Data Structures and 2710 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2711 * see also Knuth Vol. 3, page 84. The increments are selected from 2712 * formula (8), page 95. Roughly O(N^3/2). 2713 */ 2714 /* 2715 * This is our own private copy of shellsort because we want to sort 2716 * two parallel arrays (the array of buffer pointers and the array of 2717 * logical block numbers) simultaneously. Note that we cast the array 2718 * of logical block numbers to a unsigned in this routine so that the 2719 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2720 */ 2721 2722 static void 2723 lfs_shellsort(struct lfs *fs, 2724 struct buf **bp_array, union lfs_blocks *lb_array, 2725 int nmemb, int size) 2726 { 2727 static int __rsshell_increments[] = { 4, 1, 0 }; 2728 int incr, *incrp, t1, t2; 2729 struct buf *bp_temp; 2730 2731 #ifdef DEBUG 2732 incr = 0; 2733 for (t1 = 0; t1 < nmemb; t1++) { 2734 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2735 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2736 /* dump before panic */ 2737 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2738 nmemb, size); 2739 incr = 0; 2740 for (t1 = 0; t1 < nmemb; t1++) { 2741 const struct buf *bp = bp_array[t1]; 2742 2743 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2744 PRIu64 "\n", t1, 2745 (uint64_t)bp->b_bcount, 2746 (uint64_t)bp->b_lblkno); 2747 printf("lbns:"); 2748 for (t2 = 0; t2 * size < bp->b_bcount; 2749 t2++) { 2750 printf(" %jd", 2751 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2752 } 2753 printf("\n"); 2754 } 2755 panic("lfs_shellsort: inconsistent input"); 2756 } 2757 } 2758 } 2759 #endif 2760 2761 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2762 for (t1 = incr; t1 < nmemb; ++t1) 2763 for (t2 = t1 - incr; t2 >= 0;) 2764 if ((u_int64_t)bp_array[t2]->b_lblkno > 2765 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2766 bp_temp = bp_array[t2]; 2767 bp_array[t2] = bp_array[t2 + incr]; 2768 bp_array[t2 + incr] = bp_temp; 2769 t2 -= incr; 2770 } else 2771 break; 2772 2773 /* Reform the list of logical blocks */ 2774 incr = 0; 2775 for (t1 = 0; t1 < nmemb; t1++) { 2776 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2777 lfs_blocks_set(fs, lb_array, incr++, 2778 bp_array[t1]->b_lblkno + t2); 2779 } 2780 } 2781 } 2782 2783 /* 2784 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2785 * point at uninitialized space. 2786 */ 2787 void 2788 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2789 { 2790 struct segment *sp = fs->lfs_sp; 2791 SEGSUM *ssp; 2792 2793 KASSERT(vers > 0); 2794 2795 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2796 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2797 (void) lfs_writeseg(fs, fs->lfs_sp); 2798 2799 sp->sum_bytes_left -= FINFOSIZE(fs); 2800 ssp = (SEGSUM *)sp->segsum; 2801 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2802 lfs_fi_setnblocks(fs, sp->fip, 0); 2803 lfs_fi_setino(fs, sp->fip, ino); 2804 lfs_fi_setversion(fs, sp->fip, vers); 2805 } 2806 2807 /* 2808 * Release the FINFO entry, either clearing out an unused entry or 2809 * advancing us to the next available entry. 2810 */ 2811 void 2812 lfs_release_finfo(struct lfs *fs) 2813 { 2814 struct segment *sp = fs->lfs_sp; 2815 SEGSUM *ssp; 2816 2817 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2818 sp->fip = NEXT_FINFO(fs, sp->fip); 2819 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2820 } else { 2821 /* XXX shouldn't this update sp->fip? */ 2822 sp->sum_bytes_left += FINFOSIZE(fs); 2823 ssp = (SEGSUM *)sp->segsum; 2824 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2825 } 2826 } 2827
Indexes created Fri Oct 03 02:09:57 GMT 2025