lfs_segment.c revision 1.281
1 /* $NetBSD: lfs_segment.c,v 1.281 2020/01/15 17:55:44 ad 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 1.281 2020/01/15 17:55:44 ad 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 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 kcondvar_t 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_state & IN_CLEANING) { 212 ivndebug(vp,"vflush/in_cleaning"); 213 mutex_enter(&lfs_lock); 214 LFS_CLR_UINO(ip, IN_CLEANING); 215 LFS_SET_UINO(ip, IN_MODIFIED); 216 mutex_exit(&lfs_lock); 217 218 /* 219 * Toss any cleaning buffers that have real counterparts 220 * to avoid losing new data. 221 */ 222 mutex_enter(vp->v_interlock); 223 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 224 nbp = LIST_NEXT(bp, b_vnbufs); 225 if (!LFS_IS_MALLOC_BUF(bp)) 226 continue; 227 /* 228 * Look for pages matching the range covered 229 * by cleaning blocks. It's okay if more dirty 230 * pages appear, so long as none disappear out 231 * from under us. 232 */ 233 if (bp->b_lblkno > 0 && vp->v_type == VREG && 234 vp != fs->lfs_ivnode) { 235 struct vm_page *pg; 236 voff_t off; 237 238 for (off = lfs_lblktosize(fs, bp->b_lblkno); 239 off < lfs_lblktosize(fs, bp->b_lblkno + 1); 240 off += PAGE_SIZE) { 241 pg = uvm_pagelookup(&vp->v_uobj, off); 242 if (pg == NULL) 243 continue; 244 if (uvm_pagegetdirty(pg) 245 == UVM_PAGE_STATUS_DIRTY || 246 pmap_is_modified(pg)) { 247 lfs_sb_addavail(fs, 248 lfs_btofsb(fs, 249 bp->b_bcount)); 250 wakeup(&fs->lfs_availsleep); 251 mutex_exit(vp->v_interlock); 252 lfs_freebuf(fs, bp); 253 mutex_enter(vp->v_interlock); 254 bp = NULL; 255 break; 256 } 257 } 258 } 259 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 260 tbp = tnbp) 261 { 262 tnbp = LIST_NEXT(tbp, b_vnbufs); 263 if (tbp->b_vp == bp->b_vp 264 && tbp->b_lblkno == bp->b_lblkno 265 && tbp != bp) 266 { 267 lfs_sb_addavail(fs, lfs_btofsb(fs, 268 bp->b_bcount)); 269 wakeup(&fs->lfs_availsleep); 270 mutex_exit(vp->v_interlock); 271 lfs_freebuf(fs, bp); 272 mutex_enter(vp->v_interlock); 273 bp = NULL; 274 break; 275 } 276 } 277 } 278 } else { 279 mutex_enter(vp->v_interlock); 280 } 281 282 /* If the node is being written, wait until that is done */ 283 while (WRITEINPROG(vp)) { 284 ivndebug(vp,"vflush/writeinprog"); 285 cv_wait(&vp->v_cv, vp->v_interlock); 286 } 287 error = vdead_check(vp, VDEAD_NOWAIT); 288 mutex_exit(vp->v_interlock); 289 290 /* Protect against deadlock in vinvalbuf() */ 291 lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0)); 292 if (error != 0) { 293 fs->lfs_reclino = ip->i_number; 294 } 295 296 /* If we're supposed to flush a freed inode, just toss it */ 297 if (ip->i_lfs_iflags & LFSI_DELETED) { 298 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 299 ip->i_number)); 300 /* Drain v_numoutput */ 301 mutex_enter(vp->v_interlock); 302 while (vp->v_numoutput > 0) { 303 cv_wait(&vp->v_cv, vp->v_interlock); 304 } 305 KASSERT(vp->v_numoutput == 0); 306 mutex_exit(vp->v_interlock); 307 308 mutex_enter(&bufcache_lock); 309 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 310 nbp = LIST_NEXT(bp, b_vnbufs); 311 312 KASSERT((bp->b_flags & B_GATHERED) == 0); 313 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 314 lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount)); 315 wakeup(&fs->lfs_availsleep); 316 } 317 /* Copied from lfs_writeseg */ 318 if (bp->b_iodone != NULL) { 319 mutex_exit(&bufcache_lock); 320 biodone(bp); 321 mutex_enter(&bufcache_lock); 322 } else { 323 bremfree(bp); 324 LFS_UNLOCK_BUF(bp); 325 mutex_enter(vp->v_interlock); 326 bp->b_flags &= ~(B_READ | B_GATHERED); 327 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 328 bp->b_error = 0; 329 reassignbuf(bp, vp); 330 mutex_exit(vp->v_interlock); 331 brelse(bp, 0); 332 } 333 } 334 mutex_exit(&bufcache_lock); 335 LFS_CLR_UINO(ip, IN_CLEANING); 336 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 337 ip->i_state &= ~IN_ALLMOD; 338 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 339 ip->i_number)); 340 lfs_segunlock(fs); 341 342 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 343 344 return 0; 345 } 346 347 fs->lfs_flushvp = vp; 348 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 349 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 350 fs->lfs_flushvp = NULL; 351 KASSERT(fs->lfs_flushvp_fakevref == 0); 352 lfs_segunlock(fs); 353 354 /* Make sure that any pending buffers get written */ 355 mutex_enter(vp->v_interlock); 356 while (vp->v_numoutput > 0) { 357 cv_wait(&vp->v_cv, vp->v_interlock); 358 } 359 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 360 KASSERT(vp->v_numoutput == 0); 361 mutex_exit(vp->v_interlock); 362 363 return error; 364 } 365 sp = fs->lfs_sp; 366 367 flushed = 0; 368 if (VPISEMPTY(vp)) { 369 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 370 ++flushed; 371 } else if ((ip->i_state & IN_CLEANING) && 372 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 373 ivndebug(vp,"vflush/clean"); 374 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 375 ++flushed; 376 } else if (lfs_dostats) { 377 if (!VPISEMPTY(vp) || (VTOI(vp)->i_state & IN_ALLMOD)) 378 ++lfs_stats.vflush_invoked; 379 ivndebug(vp,"vflush"); 380 } 381 382 #ifdef DIAGNOSTIC 383 if (vp->v_uflag & VU_DIROP) { 384 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 385 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 386 } 387 #endif 388 389 do { 390 #ifdef DEBUG 391 int loopcount = 0; 392 #endif 393 do { 394 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 395 relock = lfs_writefile(fs, sp, vp); 396 if (relock && vp != fs->lfs_ivnode) { 397 /* 398 * Might have to wait for the 399 * cleaner to run; but we're 400 * still not done with this vnode. 401 * XXX we can do better than this. 402 */ 403 KDASSERT(ip->i_number != LFS_IFILE_INUM); 404 lfs_writeinode(fs, sp, ip); 405 mutex_enter(&lfs_lock); 406 LFS_SET_UINO(ip, IN_MODIFIED); 407 mutex_exit(&lfs_lock); 408 lfs_writeseg(fs, sp); 409 lfs_segunlock(fs); 410 lfs_segunlock_relock(fs); 411 goto top; 412 } 413 } 414 /* 415 * If we begin a new segment in the middle of writing 416 * the Ifile, it creates an inconsistent checkpoint, 417 * since the Ifile information for the new segment 418 * is not up-to-date. Take care of this here by 419 * sending the Ifile through again in case there 420 * are newly dirtied blocks. But wait, there's more! 421 * This second Ifile write could *also* cross a segment 422 * boundary, if the first one was large. The second 423 * one is guaranteed to be no more than 8 blocks, 424 * though (two segment blocks and supporting indirects) 425 * so the third write *will not* cross the boundary. 426 */ 427 if (vp == fs->lfs_ivnode) { 428 lfs_writefile(fs, sp, vp); 429 lfs_writefile(fs, sp, vp); 430 } 431 #ifdef DEBUG 432 if (++loopcount > 2) 433 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 434 #endif 435 } while (lfs_writeinode(fs, sp, ip)); 436 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 437 438 if (lfs_dostats) { 439 ++lfs_stats.nwrites; 440 if (sp->seg_flags & SEGM_SYNC) 441 ++lfs_stats.nsync_writes; 442 if (sp->seg_flags & SEGM_CKP) 443 ++lfs_stats.ncheckpoints; 444 } 445 /* 446 * If we were called from somewhere that has already held the seglock 447 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 448 * the write to complete because we are still locked. 449 * Since lfs_vflush() must return the vnode with no dirty buffers, 450 * we must explicitly wait, if that is the case. 451 * 452 * We compare the iocount against 1, not 0, because it is 453 * artificially incremented by lfs_seglock(). 454 */ 455 mutex_enter(&lfs_lock); 456 if (fs->lfs_seglock > 1) { 457 while (fs->lfs_iocount > 1) 458 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 459 "lfs_vflush", 0, &lfs_lock); 460 } 461 mutex_exit(&lfs_lock); 462 463 lfs_segunlock(fs); 464 465 /* Wait for these buffers to be recovered by aiodoned */ 466 mutex_enter(vp->v_interlock); 467 while (vp->v_numoutput > 0) { 468 cv_wait(&vp->v_cv, vp->v_interlock); 469 } 470 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 471 KASSERT(vp->v_numoutput == 0); 472 mutex_exit(vp->v_interlock); 473 474 fs->lfs_flushvp = NULL; 475 KASSERT(fs->lfs_flushvp_fakevref == 0); 476 477 return (0); 478 } 479 480 struct lfs_writevnodes_ctx { 481 int op; 482 struct lfs *fs; 483 }; 484 static bool 485 lfs_writevnodes_selector(void *cl, struct vnode *vp) 486 { 487 struct lfs_writevnodes_ctx *c = cl; 488 struct inode *ip; 489 int op = c->op; 490 491 KASSERT(mutex_owned(vp->v_interlock)); 492 493 ip = VTOI(vp); 494 if (ip == NULL || vp->v_type == VNON) 495 return false; 496 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 497 (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) { 498 vndebug(vp, "dirop"); 499 return false; 500 } 501 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 502 vndebug(vp,"empty"); 503 return false; 504 } 505 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM && 506 vp != c->fs->lfs_flushvp && !(ip->i_state & IN_CLEANING)) { 507 vndebug(vp,"cleaning"); 508 return false; 509 } 510 mutex_enter(&lfs_lock); 511 if (vp == c->fs->lfs_unlockvp) { 512 mutex_exit(&lfs_lock); 513 return false; 514 } 515 mutex_exit(&lfs_lock); 516 517 return true; 518 } 519 520 int 521 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 522 { 523 struct inode *ip; 524 struct vnode *vp; 525 struct vnode_iterator *marker; 526 struct lfs_writevnodes_ctx ctx; 527 int inodes_written = 0; 528 int error = 0; 529 530 /* 531 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist. 532 * XXX The rationale is unclear, the initial commit had no information. 533 * XXX If the order really matters we have to sort the vnodes first. 534 */ 535 536 ASSERT_SEGLOCK(fs); 537 vfs_vnode_iterator_init(mp, &marker); 538 ctx.op = op; 539 ctx.fs = fs; 540 while ((vp = vfs_vnode_iterator_next(marker, 541 lfs_writevnodes_selector, &ctx)) != NULL) { 542 ip = VTOI(vp); 543 544 /* 545 * Write the inode/file if dirty and it's not the IFILE. 546 */ 547 if (((ip->i_state & IN_ALLMOD) || !VPISEMPTY(vp)) && 548 ip->i_number != LFS_IFILE_INUM) { 549 error = lfs_writefile(fs, sp, vp); 550 if (error) { 551 vrele(vp); 552 if (error == EAGAIN) { 553 /* 554 * This error from lfs_putpages 555 * indicates we need to drop 556 * the segment lock and start 557 * over after the cleaner has 558 * had a chance to run. 559 */ 560 lfs_writeinode(fs, sp, ip); 561 lfs_writeseg(fs, sp); 562 if (!VPISEMPTY(vp) && 563 !WRITEINPROG(vp) && 564 !(ip->i_state & IN_ALLMOD)) { 565 mutex_enter(&lfs_lock); 566 LFS_SET_UINO(ip, IN_MODIFIED); 567 mutex_exit(&lfs_lock); 568 } 569 break; 570 } 571 error = 0; /* XXX not quite right */ 572 continue; 573 } 574 575 if (!VPISEMPTY(vp)) { 576 if (WRITEINPROG(vp)) { 577 ivndebug(vp,"writevnodes/write2"); 578 } else if (!(ip->i_state & IN_ALLMOD)) { 579 mutex_enter(&lfs_lock); 580 LFS_SET_UINO(ip, IN_MODIFIED); 581 mutex_exit(&lfs_lock); 582 } 583 } 584 (void) lfs_writeinode(fs, sp, ip); 585 inodes_written++; 586 } 587 vrele(vp); 588 } 589 vfs_vnode_iterator_destroy(marker); 590 return error; 591 } 592 593 /* 594 * Do a checkpoint. 595 */ 596 int 597 lfs_segwrite(struct mount *mp, int flags) 598 { 599 struct buf *bp; 600 struct inode *ip; 601 struct lfs *fs; 602 struct segment *sp; 603 struct vnode *vp; 604 SEGUSE *segusep; 605 int do_ckp, did_ckp, error; 606 unsigned n, segleft, maxseg, sn, i, curseg; 607 int writer_set = 0; 608 int dirty; 609 int redo; 610 SEGSUM *ssp; 611 int um_error; 612 613 fs = VFSTOULFS(mp)->um_lfs; 614 ASSERT_MAYBE_SEGLOCK(fs); 615 616 if (fs->lfs_ronly) 617 return EROFS; 618 619 lfs_imtime(fs); 620 621 /* 622 * Allocate a segment structure and enough space to hold pointers to 623 * the maximum possible number of buffers which can be described in a 624 * single summary block. 625 */ 626 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 627 628 /* We can't do a partial write and checkpoint at the same time. */ 629 if (do_ckp) 630 flags &= ~SEGM_SINGLE; 631 632 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 633 sp = fs->lfs_sp; 634 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 635 do_ckp = 1; 636 637 /* 638 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 639 * in which case we have to flush *all* buffers off of this vnode. 640 * We don't care about other nodes, but write any non-dirop nodes 641 * anyway in anticipation of another getnewvnode(). 642 * 643 * If we're cleaning we only write cleaning and ifile blocks, and 644 * no dirops, since otherwise we'd risk corruption in a crash. 645 */ 646 if (sp->seg_flags & SEGM_CLEAN) 647 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 648 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 649 do { 650 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 651 if ((sp->seg_flags & SEGM_SINGLE) && 652 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 653 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 654 break; 655 } 656 657 if (do_ckp || fs->lfs_dirops == 0) { 658 if (!writer_set) { 659 lfs_writer_enter(fs, "lfs writer"); 660 writer_set = 1; 661 } 662 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 663 if (um_error == 0) 664 um_error = error; 665 /* 666 * In case writevnodes errored out 667 * XXX why are we always doing this and not 668 * just on error? 669 */ 670 lfs_flush_dirops(fs); 671 ssp = (SEGSUM *)(sp->segsum); 672 lfs_ss_setflags(fs, ssp, 673 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 674 lfs_finalize_fs_seguse(fs); 675 } 676 if (do_ckp && um_error) { 677 lfs_segunlock_relock(fs); 678 sp = fs->lfs_sp; 679 } 680 } while (do_ckp && um_error != 0); 681 } 682 683 /* 684 * If we are doing a checkpoint, mark everything since the 685 * last checkpoint as no longer ACTIVE. 686 */ 687 if (do_ckp || fs->lfs_doifile) { 688 segleft = lfs_sb_getnseg(fs); 689 curseg = 0; 690 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 691 int bread_error; 692 693 dirty = 0; 694 bread_error = bread(fs->lfs_ivnode, 695 lfs_sb_getcleansz(fs) + n, 696 lfs_sb_getbsize(fs), B_MODIFY, &bp); 697 if (bread_error) 698 panic("lfs_segwrite: ifile read: " 699 "seguse %u: error %d\n", 700 n, bread_error); 701 segusep = (SEGUSE *)bp->b_data; 702 maxseg = uimin(segleft, lfs_sb_getsepb(fs)); 703 for (i = 0; i < maxseg; i++) { 704 sn = curseg + i; 705 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 706 segusep->su_flags & SEGUSE_ACTIVE) { 707 segusep->su_flags &= ~SEGUSE_ACTIVE; 708 --fs->lfs_nactive; 709 ++dirty; 710 } 711 fs->lfs_suflags[fs->lfs_activesb][sn] = 712 segusep->su_flags; 713 if (lfs_sb_getversion(fs) > 1) 714 ++segusep; 715 else 716 segusep = (SEGUSE *) 717 ((SEGUSE_V1 *)segusep + 1); 718 } 719 720 if (dirty) 721 error = LFS_BWRITE_LOG(bp); /* Ifile */ 722 else 723 brelse(bp, 0); 724 segleft -= lfs_sb_getsepb(fs); 725 curseg += lfs_sb_getsepb(fs); 726 } 727 } 728 729 KASSERT(LFS_SEGLOCK_HELD(fs)); 730 731 did_ckp = 0; 732 if (do_ckp || fs->lfs_doifile) { 733 vp = fs->lfs_ivnode; 734 #ifdef DEBUG 735 int loopcount = 0; 736 #endif 737 do { 738 739 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 740 741 mutex_enter(&lfs_lock); 742 fs->lfs_flags &= ~LFS_IFDIRTY; 743 mutex_exit(&lfs_lock); 744 745 ip = VTOI(vp); 746 747 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 748 /* 749 * Ifile has no pages, so we don't need 750 * to check error return here. 751 */ 752 lfs_writefile(fs, sp, vp); 753 /* 754 * Ensure the Ifile takes the current segment 755 * into account. See comment in lfs_vflush. 756 */ 757 lfs_writefile(fs, sp, vp); 758 lfs_writefile(fs, sp, vp); 759 } 760 761 if (ip->i_state & IN_ALLMOD) 762 ++did_ckp; 763 #if 0 764 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 765 #else 766 redo = lfs_writeinode(fs, sp, ip); 767 #endif 768 redo += lfs_writeseg(fs, sp); 769 mutex_enter(&lfs_lock); 770 redo += (fs->lfs_flags & LFS_IFDIRTY); 771 mutex_exit(&lfs_lock); 772 #ifdef DEBUG 773 if (++loopcount > 2) 774 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 775 loopcount); 776 #endif 777 } while (redo && do_ckp); 778 779 /* 780 * Unless we are unmounting, the Ifile may continue to have 781 * dirty blocks even after a checkpoint, due to changes to 782 * inodes' atime. If we're checkpointing, it's "impossible" 783 * for other parts of the Ifile to be dirty after the loop 784 * above, since we hold the segment lock. 785 */ 786 mutex_enter(vp->v_interlock); 787 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 788 LFS_CLR_UINO(ip, IN_ALLMOD); 789 } 790 #ifdef DIAGNOSTIC 791 else if (do_ckp) { 792 int do_panic = 0; 793 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 794 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 795 lfs_sb_getsegtabsz(fs) && 796 !(bp->b_flags & B_GATHERED)) { 797 printf("ifile lbn %ld still dirty (flags %lx)\n", 798 (long)bp->b_lblkno, 799 (long)bp->b_flags); 800 ++do_panic; 801 } 802 } 803 if (do_panic) 804 panic("dirty blocks"); 805 } 806 #endif 807 mutex_exit(vp->v_interlock); 808 } else { 809 (void) lfs_writeseg(fs, sp); 810 } 811 812 /* Note Ifile no longer needs to be written */ 813 fs->lfs_doifile = 0; 814 if (writer_set) 815 lfs_writer_leave(fs); 816 817 /* 818 * If we didn't write the Ifile, we didn't really do anything. 819 * That means that (1) there is a checkpoint on disk and (2) 820 * nothing has changed since it was written. 821 * 822 * Take the flags off of the segment so that lfs_segunlock 823 * doesn't have to write the superblock either. 824 */ 825 if (do_ckp && !did_ckp) { 826 sp->seg_flags &= ~SEGM_CKP; 827 } 828 829 if (lfs_dostats) { 830 ++lfs_stats.nwrites; 831 if (sp->seg_flags & SEGM_SYNC) 832 ++lfs_stats.nsync_writes; 833 if (sp->seg_flags & SEGM_CKP) 834 ++lfs_stats.ncheckpoints; 835 } 836 lfs_segunlock(fs); 837 return (0); 838 } 839 840 /* 841 * Write the dirty blocks associated with a vnode. 842 */ 843 int 844 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 845 { 846 struct inode *ip; 847 int i, frag; 848 SEGSUM *ssp; 849 int error; 850 851 ASSERT_SEGLOCK(fs); 852 error = 0; 853 ip = VTOI(vp); 854 855 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 856 857 if (vp->v_uflag & VU_DIROP) { 858 ssp = (SEGSUM *)sp->segsum; 859 lfs_ss_setflags(fs, ssp, 860 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 861 } 862 863 if (sp->seg_flags & SEGM_CLEAN) { 864 lfs_gather(fs, sp, vp, lfs_match_fake); 865 /* 866 * For a file being flushed, we need to write *all* blocks. 867 * This means writing the cleaning blocks first, and then 868 * immediately following with any non-cleaning blocks. 869 * The same is true of the Ifile since checkpoints assume 870 * that all valid Ifile blocks are written. 871 */ 872 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 873 lfs_gather(fs, sp, vp, lfs_match_data); 874 /* 875 * Don't call VOP_PUTPAGES: if we're flushing, 876 * we've already done it, and the Ifile doesn't 877 * use the page cache. 878 */ 879 } 880 } else { 881 lfs_gather(fs, sp, vp, lfs_match_data); 882 /* 883 * If we're flushing, we've already called VOP_PUTPAGES 884 * so don't do it again. Otherwise, we want to write 885 * everything we've got. 886 */ 887 if (!IS_FLUSHING(fs, vp)) { 888 mutex_enter(vp->v_interlock); 889 error = VOP_PUTPAGES(vp, 0, 0, 890 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 891 } 892 } 893 894 /* 895 * It may not be necessary to write the meta-data blocks at this point, 896 * as the roll-forward recovery code should be able to reconstruct the 897 * list. 898 * 899 * We have to write them anyway, though, under two conditions: (1) the 900 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 901 * checkpointing. 902 * 903 * BUT if we are cleaning, we might have indirect blocks that refer to 904 * new blocks not being written yet, in addition to fragments being 905 * moved out of a cleaned segment. If that is the case, don't 906 * write the indirect blocks, or the finfo will have a small block 907 * in the middle of it! 908 * XXX in this case isn't the inode size wrong too? 909 */ 910 frag = 0; 911 if (sp->seg_flags & SEGM_CLEAN) { 912 for (i = 0; i < ULFS_NDADDR; i++) 913 if (ip->i_lfs_fragsize[i] > 0 && 914 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 915 ++frag; 916 } 917 KASSERTMSG((frag <= 1), 918 "lfs_writefile: more than one fragment! frag=%d", frag); 919 if (IS_FLUSHING(fs, vp) || 920 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 921 lfs_gather(fs, sp, vp, lfs_match_indir); 922 lfs_gather(fs, sp, vp, lfs_match_dindir); 923 lfs_gather(fs, sp, vp, lfs_match_tindir); 924 } 925 lfs_release_finfo(fs); 926 927 return error; 928 } 929 930 /* 931 * Update segment accounting to reflect this inode's change of address. 932 */ 933 static int 934 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 935 { 936 struct buf *bp; 937 daddr_t daddr; 938 IFILE *ifp; 939 SEGUSE *sup; 940 ino_t ino; 941 int redo_ifile; 942 u_int32_t sn; 943 944 redo_ifile = 0; 945 946 /* 947 * If updating the ifile, update the super-block. Update the disk 948 * address and access times for this inode in the ifile. 949 */ 950 ino = ip->i_number; 951 if (ino == LFS_IFILE_INUM) { 952 daddr = lfs_sb_getidaddr(fs); 953 lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr)); 954 } else { 955 LFS_IENTRY(ifp, fs, ino, bp); 956 daddr = lfs_if_getdaddr(fs, ifp); 957 lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr)); 958 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 959 } 960 961 /* 962 * If this is the Ifile and lfs_offset is set to the first block 963 * in the segment, dirty the new segment's accounting block 964 * (XXX should already be dirty?) and tell the caller to do it again. 965 */ 966 if (ip->i_number == LFS_IFILE_INUM) { 967 sn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 968 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) == 969 lfs_sb_getoffset(fs)) { 970 LFS_SEGENTRY(sup, fs, sn, bp); 971 KASSERT(bp->b_oflags & BO_DELWRI); 972 LFS_WRITESEGENTRY(sup, fs, sn, bp); 973 /* fs->lfs_flags |= LFS_IFDIRTY; */ 974 redo_ifile |= 1; 975 } 976 } 977 978 /* 979 * The inode's last address should not be in the current partial 980 * segment, except under exceptional circumstances (lfs_writevnodes 981 * had to start over, and in the meantime more blocks were written 982 * to a vnode). Both inodes will be accounted to this segment 983 * in lfs_writeseg so we need to subtract the earlier version 984 * here anyway. The segment count can temporarily dip below 985 * zero here; keep track of how many duplicates we have in 986 * "dupino" so we don't panic below. 987 */ 988 if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) { 989 ++sp->ndupino; 990 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 991 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 992 (long long)daddr, sp->ndupino)); 993 } 994 /* 995 * Account the inode: it no longer belongs to its former segment, 996 * though it will not belong to the new segment until that segment 997 * is actually written. 998 */ 999 if (daddr != LFS_UNUSED_DADDR) { 1000 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1001 int ndupino __diagused = 1002 (sp->seg_number == oldsn) ? sp->ndupino : 0; 1003 LFS_SEGENTRY(sup, fs, oldsn, bp); 1004 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) 1005 >= DINOSIZE(fs)), 1006 "lfs_writeinode: negative bytes " 1007 "(segment %" PRIu32 " short by %d, " 1008 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1009 ", daddr=%" PRId64 ", su_nbytes=%u, " 1010 "ndupino=%d)\n", 1011 lfs_dtosn(fs, daddr), 1012 (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes, 1013 oldsn, sp->seg_number, daddr, 1014 (unsigned int)sup->su_nbytes, 1015 sp->ndupino); 1016 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1017 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1018 sup->su_nbytes -= DINOSIZE(fs); 1019 redo_ifile |= 1020 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1021 if (redo_ifile) { 1022 mutex_enter(&lfs_lock); 1023 fs->lfs_flags |= LFS_IFDIRTY; 1024 mutex_exit(&lfs_lock); 1025 /* Don't double-account */ 1026 lfs_sb_setidaddr(fs, 0x0); 1027 } 1028 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1029 } 1030 1031 return redo_ifile; 1032 } 1033 1034 int 1035 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1036 { 1037 struct buf *bp; 1038 union lfs_dinode *cdp; 1039 struct vnode *vp = ITOV(ip); 1040 daddr_t daddr; 1041 IINFO *iip; 1042 int i; 1043 int redo_ifile = 0; 1044 int gotblk = 0; 1045 int count; 1046 SEGSUM *ssp; 1047 1048 ASSERT_SEGLOCK(fs); 1049 if (!(ip->i_state & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1050 return (0); 1051 1052 /* Can't write ifile when writer is not set */ 1053 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1054 (sp->seg_flags & SEGM_CLEAN)); 1055 1056 /* 1057 * If this is the Ifile, see if writing it here will generate a 1058 * temporary misaccounting. If it will, do the accounting and write 1059 * the blocks, postponing the inode write until the accounting is 1060 * solid. 1061 */ 1062 count = 0; 1063 while (vp == fs->lfs_ivnode) { 1064 int redo = 0; 1065 1066 if (sp->idp == NULL && sp->ibp == NULL && 1067 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1068 sp->sum_bytes_left < sizeof(int32_t))) { 1069 (void) lfs_writeseg(fs, sp); 1070 continue; 1071 } 1072 1073 /* Look for dirty Ifile blocks */ 1074 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1075 if (!(bp->b_flags & B_GATHERED)) { 1076 redo = 1; 1077 break; 1078 } 1079 } 1080 1081 if (redo == 0) 1082 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1083 if (redo == 0) 1084 break; 1085 1086 if (sp->idp) { 1087 lfs_dino_setinumber(fs, sp->idp, 0); 1088 sp->idp = NULL; 1089 } 1090 ++count; 1091 if (count > 2) 1092 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1093 lfs_writefile(fs, sp, fs->lfs_ivnode); 1094 } 1095 1096 /* Allocate a new inode block if necessary. */ 1097 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1098 sp->ibp == NULL) { 1099 /* Allocate a new segment if necessary. */ 1100 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1101 sp->sum_bytes_left < sizeof(int32_t)) 1102 (void) lfs_writeseg(fs, sp); 1103 1104 /* Get next inode block. */ 1105 daddr = lfs_sb_getoffset(fs); 1106 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1107 sp->ibp = *sp->cbpp++ = 1108 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1109 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1110 gotblk++; 1111 1112 /* Zero out inode numbers */ 1113 for (i = 0; i < LFS_INOPB(fs); ++i) { 1114 union lfs_dinode *tmpdi; 1115 1116 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1117 DINOSIZE(fs) * i); 1118 lfs_dino_setinumber(fs, tmpdi, 0); 1119 } 1120 1121 ++sp->start_bpp; 1122 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1123 /* Set remaining space counters. */ 1124 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1125 sp->sum_bytes_left -= sizeof(int32_t); 1126 1127 /* Store the address in the segment summary. */ 1128 iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs)); 1129 lfs_ii_setblock(fs, iip, daddr); 1130 } 1131 1132 /* Check VU_DIROP in case there is a new file with no data blocks */ 1133 if (vp->v_uflag & VU_DIROP) { 1134 ssp = (SEGSUM *)sp->segsum; 1135 lfs_ss_setflags(fs, ssp, 1136 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1137 } 1138 1139 /* Update the inode times and copy the inode onto the inode page. */ 1140 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1141 if (ip->i_number != LFS_IFILE_INUM) 1142 LFS_ITIMES(ip, NULL, NULL, NULL); 1143 1144 /* 1145 * If this is the Ifile, and we've already written the Ifile in this 1146 * partial segment, just overwrite it (it's not on disk yet) and 1147 * continue. 1148 * 1149 * XXX we know that the bp that we get the second time around has 1150 * already been gathered. 1151 */ 1152 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1153 lfs_copy_dinode(fs, sp->idp, ip->i_din); 1154 ip->i_lfs_osize = ip->i_size; 1155 return 0; 1156 } 1157 1158 bp = sp->ibp; 1159 cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs)); 1160 lfs_copy_dinode(fs, cdp, ip->i_din); 1161 1162 /* 1163 * This inode is on its way to disk; clear its VU_DIROP status when 1164 * the write is complete. 1165 */ 1166 if (vp->v_uflag & VU_DIROP) { 1167 if (!(sp->seg_flags & SEGM_CLEAN)) 1168 ip->i_state |= IN_CDIROP; 1169 else { 1170 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1171 } 1172 } 1173 1174 /* 1175 * If cleaning, link counts and directory file sizes cannot change, 1176 * since those would be directory operations---even if the file 1177 * we are writing is marked VU_DIROP we should write the old values. 1178 * If we're not cleaning, of course, update the values so we get 1179 * current values the next time we clean. 1180 */ 1181 if (sp->seg_flags & SEGM_CLEAN) { 1182 if (vp->v_uflag & VU_DIROP) { 1183 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1184 /* if (vp->v_type == VDIR) */ 1185 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1186 } 1187 } else { 1188 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1189 ip->i_lfs_osize = ip->i_size; 1190 } 1191 1192 1193 /* We can finish the segment accounting for truncations now */ 1194 lfs_finalize_ino_seguse(fs, ip); 1195 1196 /* 1197 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1198 * addresses to disk; possibly change the on-disk record of 1199 * the inode size, either by reverting to the previous size 1200 * (in the case of cleaning) or by verifying the inode's block 1201 * holdings (in the case of files being allocated as they are being 1202 * written). 1203 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1204 * XXX count on disk wrong by the same amount. We should be 1205 * XXX able to "borrow" from lfs_avail and return it after the 1206 * XXX Ifile is written. See also in lfs_writeseg. 1207 */ 1208 1209 /* Check file size based on highest allocated block */ 1210 if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG || 1211 (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) && 1212 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1213 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1214 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1215 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1216 } 1217 if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) { 1218 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1219 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1220 lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs))); 1221 for (i=0; i<ULFS_NDADDR; i++) { 1222 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1223 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1224 lfs_dino_setdb(fs, cdp, i, 0); 1225 } 1226 } 1227 for (i=0; i<ULFS_NIADDR; i++) { 1228 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1229 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1230 lfs_dino_setib(fs, cdp, i, 0); 1231 } 1232 } 1233 } 1234 1235 #ifdef DIAGNOSTIC 1236 /* 1237 * Check dinode held blocks against dinode size. 1238 * This should be identical to the check in lfs_vget(). 1239 */ 1240 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1241 i < ULFS_NDADDR; i++) { 1242 KASSERT(i >= 0); 1243 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1244 continue; 1245 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1246 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1247 continue; 1248 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1249 # ifdef DEBUG 1250 lfs_dump_dinode(fs, cdp); 1251 # endif 1252 panic("writing inconsistent inode"); 1253 } 1254 } 1255 #endif /* DIAGNOSTIC */ 1256 1257 if (ip->i_state & IN_CLEANING) 1258 LFS_CLR_UINO(ip, IN_CLEANING); 1259 else { 1260 /* XXX IN_ALLMOD */ 1261 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1262 IN_UPDATE | IN_MODIFY); 1263 if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din)) 1264 LFS_CLR_UINO(ip, IN_MODIFIED); 1265 else { 1266 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1267 "blks=%d, eff=%jd\n", ip->i_number, 1268 lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks)); 1269 } 1270 } 1271 1272 if (ip->i_number == LFS_IFILE_INUM) { 1273 /* We know sp->idp == NULL */ 1274 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1275 1276 /* Not dirty any more */ 1277 mutex_enter(&lfs_lock); 1278 fs->lfs_flags &= ~LFS_IFDIRTY; 1279 mutex_exit(&lfs_lock); 1280 } 1281 1282 if (gotblk) { 1283 mutex_enter(&bufcache_lock); 1284 LFS_LOCK_BUF(bp); 1285 brelsel(bp, 0); 1286 mutex_exit(&bufcache_lock); 1287 } 1288 1289 /* Increment inode count in segment summary block. */ 1290 1291 ssp = (SEGSUM *)sp->segsum; 1292 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1293 1294 /* If this page is full, set flag to allocate a new page. */ 1295 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1296 sp->ibp = NULL; 1297 1298 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1299 1300 KASSERT(redo_ifile == 0); 1301 return (redo_ifile); 1302 } 1303 1304 int 1305 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1306 { 1307 struct lfs *fs; 1308 int vers; 1309 int j, blksinblk; 1310 1311 ASSERT_SEGLOCK(sp->fs); 1312 KASSERTMSG((sp->vp != NULL), 1313 "lfs_gatherblock: Null vp in segment"); 1314 1315 /* If full, finish this segment. */ 1316 fs = sp->fs; 1317 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1318 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1319 sp->seg_bytes_left < bp->b_bcount) { 1320 if (mptr) 1321 mutex_exit(mptr); 1322 lfs_updatemeta(sp); 1323 1324 vers = lfs_fi_getversion(fs, sp->fip); 1325 (void) lfs_writeseg(fs, sp); 1326 1327 /* Add the current file to the segment summary. */ 1328 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1329 1330 if (mptr) 1331 mutex_enter(mptr); 1332 return (1); 1333 } 1334 1335 if (bp->b_flags & B_GATHERED) { 1336 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1337 " lbn %" PRId64 "\n", 1338 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1339 return (0); 1340 } 1341 1342 /* Insert into the buffer list, update the FINFO block. */ 1343 bp->b_flags |= B_GATHERED; 1344 1345 *sp->cbpp++ = bp; 1346 for (j = 0; j < blksinblk; j++) { 1347 unsigned bn; 1348 1349 bn = lfs_fi_getnblocks(fs, sp->fip); 1350 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1351 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1352 /* This block's accounting moves from lfs_favail to lfs_avail */ 1353 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1354 } 1355 1356 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1357 sp->seg_bytes_left -= bp->b_bcount; 1358 return (0); 1359 } 1360 1361 int 1362 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1363 int (*match)(struct lfs *, struct buf *)) 1364 { 1365 struct buf *bp, *nbp; 1366 int count = 0; 1367 1368 ASSERT_SEGLOCK(fs); 1369 if (vp->v_type == VBLK) 1370 return 0; 1371 KASSERT(sp->vp == NULL); 1372 sp->vp = vp; 1373 mutex_enter(&bufcache_lock); 1374 1375 #ifndef LFS_NO_BACKBUF_HACK 1376 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1377 # define BUF_OFFSET \ 1378 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1379 # define BACK_BUF(BP) \ 1380 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1381 # define BEG_OF_LIST \ 1382 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1383 1384 loop: 1385 /* Find last buffer. */ 1386 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1387 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1388 bp = LIST_NEXT(bp, b_vnbufs)) 1389 continue; 1390 1391 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1392 nbp = BACK_BUF(bp); 1393 #else /* LFS_NO_BACKBUF_HACK */ 1394 loop: 1395 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1396 nbp = LIST_NEXT(bp, b_vnbufs); 1397 #endif /* LFS_NO_BACKBUF_HACK */ 1398 if ((bp->b_cflags & BC_BUSY) != 0 || 1399 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1400 #ifdef DEBUG 1401 if (vp == fs->lfs_ivnode && 1402 (bp->b_cflags & BC_BUSY) != 0 && 1403 (bp->b_flags & B_GATHERED) == 0) 1404 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1405 PRId64 " busy (%x) at 0x%jx", 1406 bp->b_lblkno, bp->b_flags, 1407 (uintmax_t)lfs_sb_getoffset(fs)); 1408 #endif 1409 continue; 1410 } 1411 #ifdef DIAGNOSTIC 1412 # ifdef LFS_USE_BC_INVAL 1413 if ((bp->b_cflags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1414 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1415 " is BC_INVAL\n", bp->b_lblkno)); 1416 VOP_PRINT(bp->b_vp); 1417 } 1418 # endif /* LFS_USE_BC_INVAL */ 1419 if (!(bp->b_oflags & BO_DELWRI)) 1420 panic("lfs_gather: bp not BO_DELWRI"); 1421 if (!(bp->b_flags & B_LOCKED)) { 1422 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1423 " blk %" PRId64 " not B_LOCKED\n", 1424 bp->b_lblkno, 1425 LFS_DBTOFSB(fs, bp->b_blkno))); 1426 VOP_PRINT(bp->b_vp); 1427 panic("lfs_gather: bp not B_LOCKED"); 1428 } 1429 #endif 1430 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1431 goto loop; 1432 } 1433 count++; 1434 } 1435 mutex_exit(&bufcache_lock); 1436 lfs_updatemeta(sp); 1437 KASSERT(sp->vp == vp); 1438 sp->vp = NULL; 1439 return count; 1440 } 1441 1442 #if DEBUG 1443 # define DEBUG_OOFF(n) do { \ 1444 if (ooff == 0) { \ 1445 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1446 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1447 ", was 0x0 (or %" PRId64 ")\n", \ 1448 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1449 } \ 1450 } while (0) 1451 #else 1452 # define DEBUG_OOFF(n) 1453 #endif 1454 1455 /* 1456 * Change the given block's address to ndaddr, finding its previous 1457 * location using ulfs_bmaparray(). 1458 * 1459 * Account for this change in the segment table. 1460 * 1461 * called with sp == NULL by roll-forwarding code. 1462 */ 1463 void 1464 lfs_update_single(struct lfs *fs, struct segment *sp, 1465 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1466 { 1467 SEGUSE *sup; 1468 struct buf *bp; 1469 struct indir a[ULFS_NIADDR + 2], *ap; 1470 struct inode *ip; 1471 daddr_t daddr, ooff; 1472 int num, error; 1473 int bb, osize, obb; 1474 1475 ASSERT_SEGLOCK(fs); 1476 KASSERT(sp == NULL || sp->vp == vp); 1477 ip = VTOI(vp); 1478 1479 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1480 if (error) 1481 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1482 1483 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1484 if (daddr > 0) 1485 daddr = LFS_DBTOFSB(fs, daddr); 1486 1487 bb = lfs_numfrags(fs, size); 1488 switch (num) { 1489 case 0: 1490 ooff = lfs_dino_getdb(fs, ip->i_din, lbn); 1491 DEBUG_OOFF(0); 1492 if (ooff == UNWRITTEN) 1493 lfs_dino_setblocks(fs, ip->i_din, 1494 lfs_dino_getblocks(fs, ip->i_din) + bb); 1495 else { 1496 /* possible fragment truncation or extension */ 1497 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1498 lfs_dino_setblocks(fs, ip->i_din, 1499 lfs_dino_getblocks(fs, ip->i_din) + (bb-obb)); 1500 } 1501 lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr); 1502 break; 1503 case 1: 1504 ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off); 1505 DEBUG_OOFF(1); 1506 if (ooff == UNWRITTEN) 1507 lfs_dino_setblocks(fs, ip->i_din, 1508 lfs_dino_getblocks(fs, ip->i_din) + bb); 1509 lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr); 1510 break; 1511 default: 1512 ap = &a[num - 1]; 1513 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1514 B_MODIFY, &bp)) 1515 panic("lfs_updatemeta: bread bno %" PRId64, 1516 ap->in_lbn); 1517 1518 ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off); 1519 DEBUG_OOFF(num); 1520 if (ooff == UNWRITTEN) 1521 lfs_dino_setblocks(fs, ip->i_din, 1522 lfs_dino_getblocks(fs, ip->i_din) + bb); 1523 lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr); 1524 (void) VOP_BWRITE(bp->b_vp, bp); 1525 } 1526 1527 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1528 1529 /* Update hiblk when extending the file */ 1530 if (lbn > ip->i_lfs_hiblk) 1531 ip->i_lfs_hiblk = lbn; 1532 1533 /* 1534 * Though we'd rather it couldn't, this *can* happen right now 1535 * if cleaning blocks and regular blocks coexist. 1536 */ 1537 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1538 1539 /* 1540 * Update segment usage information, based on old size 1541 * and location. 1542 */ 1543 if (daddr > 0) { 1544 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1545 int ndupino __diagused = (sp && sp->seg_number == oldsn ? 1546 sp->ndupino : 0); 1547 1548 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1549 if (lbn >= 0 && lbn < ULFS_NDADDR) 1550 osize = ip->i_lfs_fragsize[lbn]; 1551 else 1552 osize = lfs_sb_getbsize(fs); 1553 LFS_SEGENTRY(sup, fs, oldsn, bp); 1554 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize), 1555 "lfs_updatemeta: negative bytes " 1556 "(segment %" PRIu32 " short by %" PRId64 1557 ")\n" 1558 "lfs_updatemeta: ino %llu, lbn %" PRId64 1559 ", addr = 0x%" PRIx64 "\n" 1560 "lfs_updatemeta: ndupino=%d", 1561 lfs_dtosn(fs, daddr), 1562 (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes), 1563 (unsigned long long)ip->i_number, lbn, daddr, 1564 ndupino); 1565 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1566 " db 0x%" PRIx64 "\n", 1567 lfs_dtosn(fs, daddr), osize, 1568 ip->i_number, lbn, daddr)); 1569 sup->su_nbytes -= osize; 1570 if (!(bp->b_flags & B_GATHERED)) { 1571 mutex_enter(&lfs_lock); 1572 fs->lfs_flags |= LFS_IFDIRTY; 1573 mutex_exit(&lfs_lock); 1574 } 1575 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1576 } 1577 /* 1578 * Now that this block has a new address, and its old 1579 * segment no longer owns it, we can forget about its 1580 * old size. 1581 */ 1582 if (lbn >= 0 && lbn < ULFS_NDADDR) 1583 ip->i_lfs_fragsize[lbn] = size; 1584 } 1585 1586 /* 1587 * Update the metadata that points to the blocks listed in the FINFO 1588 * array. 1589 */ 1590 void 1591 lfs_updatemeta(struct segment *sp) 1592 { 1593 struct buf *sbp; 1594 struct lfs *fs; 1595 struct vnode *vp; 1596 daddr_t lbn; 1597 int i, nblocks, num; 1598 int __diagused nblocks_orig; 1599 int bb; 1600 int bytesleft, size; 1601 unsigned lastlength; 1602 union lfs_blocks tmpptr; 1603 1604 fs = sp->fs; 1605 vp = sp->vp; 1606 ASSERT_SEGLOCK(fs); 1607 1608 /* 1609 * This used to be: 1610 * 1611 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1612 * 1613 * that is, it allowed for the possibility that start_lbp did 1614 * not point to the beginning of the finfo block pointer area. 1615 * This particular formulation is six kinds of painful in the 1616 * lfs64 world where we have two sizes of block pointer, so 1617 * unless/until everything can be cleaned up to not move 1618 * start_lbp around but instead use an offset, we do the 1619 * following: 1620 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1621 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1622 * type. (Ugh.) 1623 * 2. Cast it to void *, then assign it to a temporary 1624 * union lfs_blocks. 1625 * 3. Subtract start_lbp from that. 1626 * 4. Save the value of nblocks in blocks_orig so we can 1627 * assert below that it hasn't changed without repeating this 1628 * rubbish. 1629 * 1630 * XXX. 1631 */ 1632 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1633 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1634 nblocks_orig = nblocks; 1635 1636 KASSERT(nblocks >= 0); 1637 KASSERT(vp != NULL); 1638 if (nblocks == 0) 1639 return; 1640 1641 /* 1642 * This count may be high due to oversize blocks from lfs_gop_write. 1643 * Correct for this. (XXX we should be able to keep track of these.) 1644 */ 1645 for (i = 0; i < nblocks; i++) { 1646 if (sp->start_bpp[i] == NULL) { 1647 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1648 nblocks = i; 1649 break; 1650 } 1651 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1652 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1653 nblocks -= num - 1; 1654 } 1655 1656 #if 0 1657 /* pre-lfs64 assertion */ 1658 KASSERT(vp->v_type == VREG || 1659 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1660 #else 1661 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1662 #endif 1663 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1664 1665 /* 1666 * Sort the blocks. 1667 * 1668 * We have to sort even if the blocks come from the 1669 * cleaner, because there might be other pending blocks on the 1670 * same inode...and if we don't sort, and there are fragments 1671 * present, blocks may be written in the wrong place. 1672 */ 1673 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1674 1675 /* 1676 * Record the length of the last block in case it's a fragment. 1677 * If there are indirect blocks present, they sort last. An 1678 * indirect block will be lfs_bsize and its presence indicates 1679 * that you cannot have fragments. 1680 * 1681 * XXX This last is a lie. A cleaned fragment can coexist with 1682 * XXX a later indirect block. This will continue to be 1683 * XXX true until lfs_markv is fixed to do everything with 1684 * XXX fake blocks (including fake inodes and fake indirect blocks). 1685 */ 1686 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1687 lfs_sb_getbmask(fs)) + 1; 1688 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1689 1690 /* 1691 * Assign disk addresses, and update references to the logical 1692 * block and the segment usage information. 1693 */ 1694 for (i = nblocks; i--; ++sp->start_bpp) { 1695 sbp = *sp->start_bpp; 1696 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1697 KASSERT(sbp->b_lblkno == lbn); 1698 1699 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1700 1701 /* 1702 * If we write a frag in the wrong place, the cleaner won't 1703 * be able to correctly identify its size later, and the 1704 * segment will be uncleanable. (Even worse, it will assume 1705 * that the indirect block that actually ends the list 1706 * is of a smaller size!) 1707 */ 1708 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1709 panic("lfs_updatemeta: fragment is not last block"); 1710 1711 /* 1712 * For each subblock in this possibly oversized block, 1713 * update its address on disk. 1714 */ 1715 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1716 KASSERT(vp == sbp->b_vp); 1717 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1718 bytesleft -= lfs_sb_getbsize(fs)) { 1719 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1720 bb = lfs_numfrags(fs, size); 1721 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1722 lfs_blocks_inc(fs, &sp->start_lbp); 1723 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1724 size); 1725 lfs_sb_addoffset(fs, bb); 1726 } 1727 1728 } 1729 1730 /* This inode has been modified */ 1731 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1732 } 1733 1734 /* 1735 * Move lfs_offset to a segment earlier than newsn. 1736 */ 1737 int 1738 lfs_rewind(struct lfs *fs, int newsn) 1739 { 1740 int sn, osn, isdirty; 1741 struct buf *bp; 1742 SEGUSE *sup; 1743 1744 ASSERT_SEGLOCK(fs); 1745 1746 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1747 if (osn < newsn) 1748 return 0; 1749 1750 /* lfs_avail eats the remaining space in this segment */ 1751 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1752 1753 /* Find a low-numbered segment */ 1754 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1755 LFS_SEGENTRY(sup, fs, sn, bp); 1756 isdirty = sup->su_flags & SEGUSE_DIRTY; 1757 brelse(bp, 0); 1758 1759 if (!isdirty) 1760 break; 1761 } 1762 if (sn == lfs_sb_getnseg(fs)) 1763 panic("lfs_rewind: no clean segments"); 1764 if (newsn >= 0 && sn >= newsn) 1765 return ENOENT; 1766 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1767 lfs_newseg(fs); 1768 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1769 1770 return 0; 1771 } 1772 1773 /* 1774 * Start a new partial segment. 1775 * 1776 * Return 1 when we entered to a new segment. 1777 * Otherwise, return 0. 1778 */ 1779 int 1780 lfs_initseg(struct lfs *fs) 1781 { 1782 struct segment *sp = fs->lfs_sp; 1783 SEGSUM *ssp; 1784 struct buf *sbp; /* buffer for SEGSUM */ 1785 int repeat = 0; /* return value */ 1786 1787 ASSERT_SEGLOCK(fs); 1788 /* Advance to the next segment. */ 1789 if (!LFS_PARTIAL_FITS(fs)) { 1790 SEGUSE *sup; 1791 struct buf *bp; 1792 1793 /* lfs_avail eats the remaining space */ 1794 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1795 lfs_sb_getcurseg(fs))); 1796 /* Wake up any cleaning procs waiting on this file system. */ 1797 lfs_wakeup_cleaner(fs); 1798 lfs_newseg(fs); 1799 repeat = 1; 1800 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1801 1802 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1803 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1804 1805 /* 1806 * If the segment contains a superblock, update the offset 1807 * and summary address to skip over it. 1808 */ 1809 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1810 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1811 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1812 sp->seg_bytes_left -= LFS_SBPAD; 1813 } 1814 brelse(bp, 0); 1815 /* Segment zero could also contain the labelpad */ 1816 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1817 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1818 lfs_sb_addoffset(fs, 1819 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1820 sp->seg_bytes_left -= 1821 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1822 } 1823 } else { 1824 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1825 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1826 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1827 } 1828 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1829 1830 /* Record first address of this partial segment */ 1831 if (sp->seg_flags & SEGM_CLEAN) { 1832 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1833 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1834 /* "1" is the artificial inc in lfs_seglock */ 1835 mutex_enter(&lfs_lock); 1836 while (fs->lfs_iocount > 1) { 1837 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1838 "lfs_initseg", 0, &lfs_lock); 1839 } 1840 mutex_exit(&lfs_lock); 1841 fs->lfs_cleanind = 0; 1842 } 1843 } 1844 1845 sp->fs = fs; 1846 sp->ibp = NULL; 1847 sp->idp = NULL; 1848 sp->ninodes = 0; 1849 sp->ndupino = 0; 1850 1851 sp->cbpp = sp->bpp; 1852 1853 /* Get a new buffer for SEGSUM */ 1854 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1855 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1856 1857 /* ... and enter it into the buffer list. */ 1858 *sp->cbpp = sbp; 1859 sp->cbpp++; 1860 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1861 1862 sp->start_bpp = sp->cbpp; 1863 1864 /* Set point to SEGSUM, initialize it. */ 1865 ssp = sp->segsum = sbp->b_data; 1866 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1867 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1868 lfs_ss_setnfinfo(fs, ssp, 0); 1869 lfs_ss_setninos(fs, ssp, 0); 1870 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1871 1872 /* Set pointer to first FINFO, initialize it. */ 1873 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1874 lfs_fi_setnblocks(fs, sp->fip, 0); 1875 lfs_fi_setlastlength(fs, sp->fip, 0); 1876 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1877 1878 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1879 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1880 1881 return (repeat); 1882 } 1883 1884 /* 1885 * Remove SEGUSE_INVAL from all segments. 1886 */ 1887 void 1888 lfs_unset_inval_all(struct lfs *fs) 1889 { 1890 SEGUSE *sup; 1891 struct buf *bp; 1892 int i; 1893 1894 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1895 LFS_SEGENTRY(sup, fs, i, bp); 1896 if (sup->su_flags & SEGUSE_INVAL) { 1897 sup->su_flags &= ~SEGUSE_INVAL; 1898 LFS_WRITESEGENTRY(sup, fs, i, bp); 1899 } else 1900 brelse(bp, 0); 1901 } 1902 } 1903 1904 /* 1905 * Return the next segment to write. 1906 */ 1907 void 1908 lfs_newseg(struct lfs *fs) 1909 { 1910 CLEANERINFO *cip; 1911 SEGUSE *sup; 1912 struct buf *bp; 1913 int curseg, isdirty, sn, skip_inval; 1914 1915 ASSERT_SEGLOCK(fs); 1916 1917 /* Honor LFCNWRAPSTOP */ 1918 mutex_enter(&lfs_lock); 1919 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1920 if (fs->lfs_wrappass) { 1921 log(LOG_NOTICE, "%s: wrappass=%d\n", 1922 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1923 fs->lfs_wrappass = 0; 1924 break; 1925 } 1926 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1927 wakeup(&fs->lfs_nowrap); 1928 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1929 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1930 &lfs_lock); 1931 } 1932 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1933 mutex_exit(&lfs_lock); 1934 1935 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1936 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1937 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1938 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1939 sup->su_nbytes = 0; 1940 sup->su_nsums = 0; 1941 sup->su_ninos = 0; 1942 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1943 1944 LFS_CLEANERINFO(cip, fs, bp); 1945 lfs_ci_shiftcleantodirty(fs, cip, 1); 1946 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1947 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1948 1949 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1950 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1951 skip_inval = 1; 1952 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1953 sn = (sn + 1) % lfs_sb_getnseg(fs); 1954 1955 if (sn == curseg) { 1956 if (skip_inval) 1957 skip_inval = 0; 1958 else 1959 panic("lfs_nextseg: no clean segments"); 1960 } 1961 LFS_SEGENTRY(sup, fs, sn, bp); 1962 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1963 /* Check SEGUSE_EMPTY as we go along */ 1964 if (isdirty && sup->su_nbytes == 0 && 1965 !(sup->su_flags & SEGUSE_EMPTY)) 1966 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1967 else 1968 brelse(bp, 0); 1969 1970 if (!isdirty) 1971 break; 1972 } 1973 if (skip_inval == 0) 1974 lfs_unset_inval_all(fs); 1975 1976 ++fs->lfs_nactive; 1977 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1978 if (lfs_dostats) { 1979 ++lfs_stats.segsused; 1980 } 1981 } 1982 1983 static struct buf * 1984 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1985 int n) 1986 { 1987 struct lfs_cluster *cl; 1988 struct buf **bpp, *bp; 1989 1990 ASSERT_SEGLOCK(fs); 1991 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1992 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1993 memset(cl, 0, sizeof(*cl)); 1994 cl->fs = fs; 1995 cl->bpp = bpp; 1996 cl->bufcount = 0; 1997 cl->bufsize = 0; 1998 1999 /* If this segment is being written synchronously, note that */ 2000 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 2001 cl->flags |= LFS_CL_SYNC; 2002 cl->seg = fs->lfs_sp; 2003 ++cl->seg->seg_iocount; 2004 } 2005 2006 /* Get an empty buffer header, or maybe one with something on it */ 2007 bp = getiobuf(vp, true); 2008 bp->b_dev = NODEV; 2009 bp->b_blkno = bp->b_lblkno = addr; 2010 bp->b_iodone = lfs_cluster_callback; 2011 bp->b_private = cl; 2012 2013 return bp; 2014 } 2015 2016 int 2017 lfs_writeseg(struct lfs *fs, struct segment *sp) 2018 { 2019 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2020 SEGUSE *sup; 2021 SEGSUM *ssp; 2022 int i; 2023 int do_again, nblocks, byteoffset; 2024 size_t el_size; 2025 struct lfs_cluster *cl; 2026 u_short ninos; 2027 struct vnode *devvp; 2028 char *p = NULL; 2029 struct vnode *vp; 2030 unsigned ibindex, iblimit; 2031 int changed; 2032 u_int32_t sum; 2033 size_t sumstart; 2034 #ifdef DEBUG 2035 FINFO *fip; 2036 int findex; 2037 #endif 2038 2039 ASSERT_SEGLOCK(fs); 2040 2041 ssp = (SEGSUM *)sp->segsum; 2042 2043 /* 2044 * If there are no buffers other than the segment summary to write, 2045 * don't do anything. If we are the end of a dirop sequence, however, 2046 * write the empty segment summary anyway, to help out the 2047 * roll-forward agent. 2048 */ 2049 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2050 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2051 return 0; 2052 } 2053 2054 /* Note if partial segment is being written by the cleaner */ 2055 if (sp->seg_flags & SEGM_CLEAN) 2056 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2057 2058 /* Note if we are writing to reclaim */ 2059 if (sp->seg_flags & SEGM_RECLAIM) { 2060 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2061 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2062 } 2063 2064 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2065 2066 /* Update the segment usage information. */ 2067 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2068 2069 /* Loop through all blocks, except the segment summary. */ 2070 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2071 if ((*bpp)->b_vp != devvp) { 2072 sup->su_nbytes += (*bpp)->b_bcount; 2073 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2074 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2075 sp->seg_number, (*bpp)->b_bcount, 2076 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2077 (*bpp)->b_blkno)); 2078 } 2079 } 2080 2081 #ifdef DEBUG 2082 /* Check for zero-length and zero-version FINFO entries. */ 2083 fip = SEGSUM_FINFOBASE(fs, ssp); 2084 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2085 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2086 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2087 fip = NEXT_FINFO(fs, fip); 2088 } 2089 #endif /* DEBUG */ 2090 2091 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2092 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2093 sp->seg_number, 2094 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2095 lfs_ss_getninos(fs, ssp))); 2096 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2097 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2098 if (lfs_sb_getversion(fs) == 1) 2099 sup->su_olastmod = time_second; 2100 else 2101 sup->su_lastmod = time_second; 2102 sup->su_ninos += ninos; 2103 ++sup->su_nsums; 2104 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2105 2106 do_again = !(bp->b_flags & B_GATHERED); 2107 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2108 2109 /* 2110 * Mark blocks BC_BUSY, to prevent then from being changed between 2111 * the checksum computation and the actual write. 2112 * 2113 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2114 * there are any, replace them with copies that have UNASSIGNED 2115 * instead. 2116 */ 2117 mutex_enter(&bufcache_lock); 2118 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2119 ++bpp; 2120 bp = *bpp; 2121 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2122 bp->b_cflags |= BC_BUSY; 2123 continue; 2124 } 2125 2126 while (bp->b_cflags & BC_BUSY) { 2127 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2128 " data summary corruption for ino %d, lbn %" 2129 PRId64 "\n", 2130 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2131 bp->b_cflags |= BC_WANTED; 2132 cv_wait(&bp->b_busy, &bufcache_lock); 2133 } 2134 bp->b_cflags |= BC_BUSY; 2135 mutex_exit(&bufcache_lock); 2136 unbusybp = NULL; 2137 2138 /* 2139 * Check and replace indirect block UNWRITTEN bogosity. 2140 * XXX See comment in lfs_writefile. 2141 */ 2142 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2143 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) != 2144 VTOI(bp->b_vp)->i_lfs_effnblks) { 2145 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2146 VTOI(bp->b_vp)->i_number, 2147 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2148 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din))); 2149 /* Make a copy we'll make changes to */ 2150 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2151 bp->b_bcount, LFS_NB_IBLOCK); 2152 newbp->b_blkno = bp->b_blkno; 2153 memcpy(newbp->b_data, bp->b_data, 2154 newbp->b_bcount); 2155 2156 changed = 0; 2157 iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs); 2158 for (ibindex = 0; ibindex < iblimit; ibindex++) { 2159 if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) { 2160 ++changed; 2161 lfs_iblock_set(fs, newbp->b_data, 2162 ibindex, 0); 2163 } 2164 } 2165 /* 2166 * Get rid of the old buffer. Don't mark it clean, 2167 * though, if it still has dirty data on it. 2168 */ 2169 if (changed) { 2170 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2171 " bp = %p newbp = %p\n", changed, bp, 2172 newbp)); 2173 *bpp = newbp; 2174 bp->b_flags &= ~B_GATHERED; 2175 bp->b_error = 0; 2176 if (bp->b_iodone != NULL) { 2177 DLOG((DLOG_SEG, "lfs_writeseg: " 2178 "indir bp should not be B_CALL\n")); 2179 biodone(bp); 2180 bp = NULL; 2181 } else { 2182 /* Still on free list, leave it there */ 2183 unbusybp = bp; 2184 /* 2185 * We have to re-decrement lfs_avail 2186 * since this block is going to come 2187 * back around to us in the next 2188 * segment. 2189 */ 2190 lfs_sb_subavail(fs, 2191 lfs_btofsb(fs, bp->b_bcount)); 2192 } 2193 } else { 2194 lfs_freebuf(fs, newbp); 2195 } 2196 } 2197 mutex_enter(&bufcache_lock); 2198 if (unbusybp != NULL) { 2199 unbusybp->b_cflags &= ~BC_BUSY; 2200 if (unbusybp->b_cflags & BC_WANTED) 2201 cv_broadcast(&bp->b_busy); 2202 } 2203 } 2204 mutex_exit(&bufcache_lock); 2205 2206 /* 2207 * Compute checksum across data and then across summary; the first 2208 * block (the summary block) is skipped. Set the create time here 2209 * so that it's guaranteed to be later than the inode mod times. 2210 */ 2211 sum = 0; 2212 if (lfs_sb_getversion(fs) == 1) 2213 el_size = sizeof(u_long); 2214 else 2215 el_size = sizeof(u_int32_t); 2216 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2217 ++bpp; 2218 /* Loop through gop_write cluster blocks */ 2219 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2220 byteoffset += lfs_sb_getbsize(fs)) { 2221 #ifdef LFS_USE_BC_INVAL 2222 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2223 (*bpp)->b_iodone != NULL) { 2224 if (copyin((void *)(*bpp)->b_saveaddr + 2225 byteoffset, dp, el_size)) { 2226 panic("lfs_writeseg: copyin failed [1]:" 2227 " ino %" PRIu64 " blk %" PRId64, 2228 VTOI((*bpp)->b_vp)->i_number, 2229 (*bpp)->b_lblkno); 2230 } 2231 } else 2232 #endif /* LFS_USE_BC_INVAL */ 2233 { 2234 sum = lfs_cksum_part((char *) 2235 (*bpp)->b_data + byteoffset, el_size, sum); 2236 } 2237 } 2238 } 2239 if (lfs_sb_getversion(fs) == 1) 2240 lfs_ss_setocreate(fs, ssp, time_second); 2241 else { 2242 lfs_ss_setcreate(fs, ssp, time_second); 2243 lfs_sb_addserial(fs, 1); 2244 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2245 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2246 } 2247 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2248 sumstart = lfs_ss_getsumstart(fs); 2249 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2250 lfs_sb_getsumsize(fs) - sumstart)); 2251 2252 mutex_enter(&lfs_lock); 2253 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2254 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2255 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2256 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2257 mutex_exit(&lfs_lock); 2258 2259 /* 2260 * When we simply write the blocks we lose a rotation for every block 2261 * written. To avoid this problem, we cluster the buffers into a 2262 * chunk and write the chunk. MAXPHYS is the largest size I/O 2263 * devices can handle, use that for the size of the chunks. 2264 * 2265 * Blocks that are already clusters (from GOP_WRITE), however, we 2266 * don't bother to copy into other clusters. 2267 */ 2268 2269 #define CHUNKSIZE MAXPHYS 2270 2271 if (devvp == NULL) 2272 panic("devvp is NULL"); 2273 for (bpp = sp->bpp, i = nblocks; i;) { 2274 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2275 cl = cbp->b_private; 2276 2277 cbp->b_flags |= B_ASYNC; 2278 cbp->b_cflags |= BC_BUSY; 2279 cbp->b_bcount = 0; 2280 2281 KASSERTMSG((bpp - sp->bpp <= 2282 (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2283 / sizeof(int32_t)), 2284 "lfs_writeseg: real bpp overwrite"); 2285 KASSERTMSG((bpp - sp->bpp <= 2286 lfs_segsize(fs) / lfs_sb_getfsize(fs)), 2287 "lfs_writeseg: theoretical bpp overwrite"); 2288 2289 /* 2290 * Construct the cluster. 2291 */ 2292 mutex_enter(&lfs_lock); 2293 ++fs->lfs_iocount; 2294 mutex_exit(&lfs_lock); 2295 while (i && cbp->b_bcount < CHUNKSIZE) { 2296 bp = *bpp; 2297 2298 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2299 break; 2300 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2301 break; 2302 2303 /* Clusters from GOP_WRITE are expedited */ 2304 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2305 if (cbp->b_bcount > 0) 2306 /* Put in its own buffer */ 2307 break; 2308 else { 2309 cbp->b_data = bp->b_data; 2310 } 2311 } else if (cbp->b_bcount == 0) { 2312 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2313 LFS_NB_CLUSTER); 2314 cl->flags |= LFS_CL_MALLOC; 2315 } 2316 KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2317 btodb(bp->b_bcount - 1))) == 2318 sp->seg_number), 2319 "segment overwrite: blk size %d daddr %" PRIx64 2320 " not in seg %d\n", 2321 bp->b_bcount, bp->b_blkno, 2322 sp->seg_number); 2323 2324 #ifdef LFS_USE_BC_INVAL 2325 /* 2326 * Fake buffers from the cleaner are marked as BC_INVAL. 2327 * We need to copy the data from user space rather than 2328 * from the buffer indicated. 2329 * XXX == what do I do on an error? 2330 */ 2331 if ((bp->b_cflags & BC_INVAL) != 0 && 2332 bp->b_iodone != NULL) { 2333 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2334 panic("lfs_writeseg: " 2335 "copyin failed [2]"); 2336 } else 2337 #endif /* LFS_USE_BC_INVAL */ 2338 if (cl->flags & LFS_CL_MALLOC) { 2339 /* copy data into our cluster. */ 2340 memcpy(p, bp->b_data, bp->b_bcount); 2341 p += bp->b_bcount; 2342 } 2343 2344 cbp->b_bcount += bp->b_bcount; 2345 cl->bufsize += bp->b_bcount; 2346 2347 bp->b_flags &= ~B_READ; 2348 bp->b_error = 0; 2349 cl->bpp[cl->bufcount++] = bp; 2350 2351 vp = bp->b_vp; 2352 mutex_enter(&bufcache_lock); 2353 mutex_enter(vp->v_interlock); 2354 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2355 reassignbuf(bp, vp); 2356 vp->v_numoutput++; 2357 mutex_exit(vp->v_interlock); 2358 mutex_exit(&bufcache_lock); 2359 2360 bpp++; 2361 i--; 2362 } 2363 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2364 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2365 else 2366 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2367 mutex_enter(devvp->v_interlock); 2368 devvp->v_numoutput++; 2369 mutex_exit(devvp->v_interlock); 2370 VOP_STRATEGY(devvp, cbp); 2371 curlwp->l_ru.ru_oublock++; 2372 } 2373 2374 if (lfs_dostats) { 2375 ++lfs_stats.psegwrites; 2376 lfs_stats.blocktot += nblocks - 1; 2377 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2378 ++lfs_stats.psyncwrites; 2379 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2380 ++lfs_stats.pcleanwrites; 2381 lfs_stats.cleanblocks += nblocks - 1; 2382 } 2383 } 2384 2385 return (lfs_initseg(fs) || do_again); 2386 } 2387 2388 void 2389 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2390 { 2391 struct buf *bp; 2392 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2393 2394 ASSERT_MAYBE_SEGLOCK(fs); 2395 if (fs->lfs_is64) { 2396 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2397 } else { 2398 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2399 } 2400 /* 2401 * If we can write one superblock while another is in 2402 * progress, we risk not having a complete checkpoint if we crash. 2403 * So, block here if a superblock write is in progress. 2404 */ 2405 mutex_enter(&lfs_lock); 2406 while (fs->lfs_sbactive) { 2407 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2408 &lfs_lock); 2409 } 2410 fs->lfs_sbactive = daddr; 2411 mutex_exit(&lfs_lock); 2412 2413 /* Set timestamp of this version of the superblock */ 2414 if (lfs_sb_getversion(fs) == 1) 2415 lfs_sb_setotstamp(fs, time_second); 2416 lfs_sb_settstamp(fs, time_second); 2417 2418 /* The next chunk of code relies on this assumption */ 2419 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2420 2421 /* Checksum the superblock and copy it into a buffer. */ 2422 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2423 bp = lfs_newbuf(fs, devvp, 2424 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2425 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2426 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2427 LFS_SBPAD - sizeof(struct dlfs)); 2428 2429 bp->b_cflags |= BC_BUSY; 2430 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2431 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2432 bp->b_error = 0; 2433 bp->b_iodone = lfs_supercallback; 2434 2435 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2436 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2437 else 2438 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2439 curlwp->l_ru.ru_oublock++; 2440 2441 mutex_enter(devvp->v_interlock); 2442 devvp->v_numoutput++; 2443 mutex_exit(devvp->v_interlock); 2444 2445 mutex_enter(&lfs_lock); 2446 ++fs->lfs_iocount; 2447 mutex_exit(&lfs_lock); 2448 VOP_STRATEGY(devvp, bp); 2449 } 2450 2451 /* 2452 * Logical block number match routines used when traversing the dirty block 2453 * chain. 2454 */ 2455 int 2456 lfs_match_fake(struct lfs *fs, struct buf *bp) 2457 { 2458 2459 ASSERT_SEGLOCK(fs); 2460 return LFS_IS_MALLOC_BUF(bp); 2461 } 2462 2463 #if 0 2464 int 2465 lfs_match_real(struct lfs *fs, struct buf *bp) 2466 { 2467 2468 ASSERT_SEGLOCK(fs); 2469 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2470 } 2471 #endif 2472 2473 int 2474 lfs_match_data(struct lfs *fs, struct buf *bp) 2475 { 2476 2477 ASSERT_SEGLOCK(fs); 2478 return (bp->b_lblkno >= 0); 2479 } 2480 2481 int 2482 lfs_match_indir(struct lfs *fs, struct buf *bp) 2483 { 2484 daddr_t lbn; 2485 2486 ASSERT_SEGLOCK(fs); 2487 lbn = bp->b_lblkno; 2488 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2489 } 2490 2491 int 2492 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2493 { 2494 daddr_t lbn; 2495 2496 ASSERT_SEGLOCK(fs); 2497 lbn = bp->b_lblkno; 2498 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2499 } 2500 2501 int 2502 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2503 { 2504 daddr_t lbn; 2505 2506 ASSERT_SEGLOCK(fs); 2507 lbn = bp->b_lblkno; 2508 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2509 } 2510 2511 static void 2512 lfs_free_aiodone(struct buf *bp) 2513 { 2514 struct lfs *fs; 2515 2516 KERNEL_LOCK(1, curlwp); 2517 fs = bp->b_private; 2518 ASSERT_NO_SEGLOCK(fs); 2519 lfs_freebuf(fs, bp); 2520 KERNEL_UNLOCK_LAST(curlwp); 2521 } 2522 2523 static void 2524 lfs_super_aiodone(struct buf *bp) 2525 { 2526 struct lfs *fs; 2527 2528 KERNEL_LOCK(1, curlwp); 2529 fs = bp->b_private; 2530 ASSERT_NO_SEGLOCK(fs); 2531 mutex_enter(&lfs_lock); 2532 fs->lfs_sbactive = 0; 2533 if (--fs->lfs_iocount <= 1) 2534 wakeup(&fs->lfs_iocount); 2535 wakeup(&fs->lfs_sbactive); 2536 mutex_exit(&lfs_lock); 2537 lfs_freebuf(fs, bp); 2538 KERNEL_UNLOCK_LAST(curlwp); 2539 } 2540 2541 static void 2542 lfs_cluster_aiodone(struct buf *bp) 2543 { 2544 struct lfs_cluster *cl; 2545 struct lfs *fs; 2546 struct buf *tbp, *fbp; 2547 struct vnode *vp, *devvp, *ovp; 2548 struct inode *ip; 2549 int error; 2550 2551 KERNEL_LOCK(1, curlwp); 2552 2553 error = bp->b_error; 2554 cl = bp->b_private; 2555 fs = cl->fs; 2556 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2557 ASSERT_NO_SEGLOCK(fs); 2558 2559 /* Put the pages back, and release the buffer */ 2560 while (cl->bufcount--) { 2561 tbp = cl->bpp[cl->bufcount]; 2562 KASSERT(tbp->b_cflags & BC_BUSY); 2563 if (error) { 2564 tbp->b_error = error; 2565 } 2566 2567 /* 2568 * We're done with tbp. If it has not been re-dirtied since 2569 * the cluster was written, free it. Otherwise, keep it on 2570 * the locked list to be written again. 2571 */ 2572 vp = tbp->b_vp; 2573 2574 tbp->b_flags &= ~B_GATHERED; 2575 2576 #ifdef DEBUG 2577 if ((tbp)->b_vp == (fs)->lfs_ivnode) 2578 LFS_ENTER_LOG("clear", __FILE__, __LINE__, 2579 tbp->b_lblkno, tbp->b_flags, curproc->p_pid); 2580 #endif 2581 2582 mutex_enter(&bufcache_lock); 2583 if (tbp->b_iodone == NULL) { 2584 KASSERT(tbp->b_flags & B_LOCKED); 2585 bremfree(tbp); 2586 if (vp) { 2587 mutex_enter(vp->v_interlock); 2588 reassignbuf(tbp, vp); 2589 mutex_exit(vp->v_interlock); 2590 } 2591 tbp->b_flags |= B_ASYNC; /* for biodone */ 2592 } 2593 2594 if ((tbp->b_flags & B_LOCKED) && !(tbp->b_oflags & BO_DELWRI)) 2595 LFS_UNLOCK_BUF(tbp); 2596 2597 if (tbp->b_oflags & BO_DONE) { 2598 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2599 cl->bufcount, (long)tbp->b_flags)); 2600 } 2601 2602 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2603 /* 2604 * A buffer from the page daemon. 2605 * We use the same iodone as it does, 2606 * so we must manually disassociate its 2607 * buffers from the vp. 2608 */ 2609 if ((ovp = tbp->b_vp) != NULL) { 2610 /* This is just silly */ 2611 mutex_enter(ovp->v_interlock); 2612 brelvp(tbp); 2613 mutex_exit(ovp->v_interlock); 2614 tbp->b_vp = vp; 2615 tbp->b_objlock = vp->v_interlock; 2616 } 2617 /* Put it back the way it was */ 2618 tbp->b_flags |= B_ASYNC; 2619 /* Master buffers have BC_AGE */ 2620 if (tbp->b_private == tbp) 2621 tbp->b_cflags |= BC_AGE; 2622 } 2623 mutex_exit(&bufcache_lock); 2624 2625 biodone(tbp); 2626 2627 /* 2628 * If this is the last block for this vnode, but 2629 * there are other blocks on its dirty list, 2630 * set IN_MODIFIED/IN_CLEANING depending on what 2631 * sort of block. Only do this for our mount point, 2632 * not for, e.g., inode blocks that are attached to 2633 * the devvp. 2634 * XXX KS - Shouldn't we set *both* if both types 2635 * of blocks are present (traverse the dirty list?) 2636 */ 2637 mutex_enter(vp->v_interlock); 2638 mutex_enter(&lfs_lock); 2639 if (vp != devvp && vp->v_numoutput == 0 && 2640 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2641 ip = VTOI(vp); 2642 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2643 ip->i_number)); 2644 if (LFS_IS_MALLOC_BUF(fbp)) 2645 LFS_SET_UINO(ip, IN_CLEANING); 2646 else 2647 LFS_SET_UINO(ip, IN_MODIFIED); 2648 } 2649 cv_broadcast(&vp->v_cv); 2650 mutex_exit(&lfs_lock); 2651 mutex_exit(vp->v_interlock); 2652 } 2653 2654 /* Fix up the cluster buffer, and release it */ 2655 if (cl->flags & LFS_CL_MALLOC) 2656 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2657 putiobuf(bp); 2658 2659 /* Note i/o done */ 2660 if (cl->flags & LFS_CL_SYNC) { 2661 if (--cl->seg->seg_iocount == 0) 2662 wakeup(&cl->seg->seg_iocount); 2663 } 2664 mutex_enter(&lfs_lock); 2665 KASSERTMSG((fs->lfs_iocount != 0), 2666 "lfs_cluster_aiodone: zero iocount"); 2667 if (--fs->lfs_iocount <= 1) 2668 wakeup(&fs->lfs_iocount); 2669 mutex_exit(&lfs_lock); 2670 2671 KERNEL_UNLOCK_LAST(curlwp); 2672 2673 pool_put(&fs->lfs_bpppool, cl->bpp); 2674 cl->bpp = NULL; 2675 pool_put(&fs->lfs_clpool, cl); 2676 } 2677 2678 static void 2679 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2680 { 2681 /* reset b_iodone for when this is a single-buf i/o. */ 2682 bp->b_iodone = aiodone; 2683 2684 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2685 } 2686 2687 static void 2688 lfs_cluster_callback(struct buf *bp) 2689 { 2690 2691 lfs_generic_callback(bp, lfs_cluster_aiodone); 2692 } 2693 2694 void 2695 lfs_supercallback(struct buf *bp) 2696 { 2697 2698 lfs_generic_callback(bp, lfs_super_aiodone); 2699 } 2700 2701 /* 2702 * The only buffers that are going to hit these functions are the 2703 * segment write blocks, or the segment summaries, or the superblocks. 2704 * 2705 * All of the above are created by lfs_newbuf, and so do not need to be 2706 * released via brelse. 2707 */ 2708 void 2709 lfs_callback(struct buf *bp) 2710 { 2711 2712 lfs_generic_callback(bp, lfs_free_aiodone); 2713 } 2714 2715 /* 2716 * Shellsort (diminishing increment sort) from Data Structures and 2717 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2718 * see also Knuth Vol. 3, page 84. The increments are selected from 2719 * formula (8), page 95. Roughly O(N^3/2). 2720 */ 2721 /* 2722 * This is our own private copy of shellsort because we want to sort 2723 * two parallel arrays (the array of buffer pointers and the array of 2724 * logical block numbers) simultaneously. Note that we cast the array 2725 * of logical block numbers to a unsigned in this routine so that the 2726 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2727 */ 2728 2729 static void 2730 lfs_shellsort(struct lfs *fs, 2731 struct buf **bp_array, union lfs_blocks *lb_array, 2732 int nmemb, int size) 2733 { 2734 static int __rsshell_increments[] = { 4, 1, 0 }; 2735 int incr, *incrp, t1, t2; 2736 struct buf *bp_temp; 2737 2738 #ifdef DEBUG 2739 incr = 0; 2740 for (t1 = 0; t1 < nmemb; t1++) { 2741 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2742 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2743 /* dump before panic */ 2744 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2745 nmemb, size); 2746 incr = 0; 2747 for (t1 = 0; t1 < nmemb; t1++) { 2748 const struct buf *bp = bp_array[t1]; 2749 2750 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2751 PRIu64 "\n", t1, 2752 (uint64_t)bp->b_bcount, 2753 (uint64_t)bp->b_lblkno); 2754 printf("lbns:"); 2755 for (t2 = 0; t2 * size < bp->b_bcount; 2756 t2++) { 2757 printf(" %jd", 2758 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2759 } 2760 printf("\n"); 2761 } 2762 panic("lfs_shellsort: inconsistent input"); 2763 } 2764 } 2765 } 2766 #endif 2767 2768 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2769 for (t1 = incr; t1 < nmemb; ++t1) 2770 for (t2 = t1 - incr; t2 >= 0;) 2771 if ((u_int64_t)bp_array[t2]->b_lblkno > 2772 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2773 bp_temp = bp_array[t2]; 2774 bp_array[t2] = bp_array[t2 + incr]; 2775 bp_array[t2 + incr] = bp_temp; 2776 t2 -= incr; 2777 } else 2778 break; 2779 2780 /* Reform the list of logical blocks */ 2781 incr = 0; 2782 for (t1 = 0; t1 < nmemb; t1++) { 2783 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2784 lfs_blocks_set(fs, lb_array, incr++, 2785 bp_array[t1]->b_lblkno + t2); 2786 } 2787 } 2788 } 2789 2790 /* 2791 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2792 * point at uninitialized space. 2793 */ 2794 void 2795 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2796 { 2797 struct segment *sp = fs->lfs_sp; 2798 SEGSUM *ssp; 2799 2800 KASSERT(vers > 0); 2801 2802 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2803 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2804 (void) lfs_writeseg(fs, fs->lfs_sp); 2805 2806 sp->sum_bytes_left -= FINFOSIZE(fs); 2807 ssp = (SEGSUM *)sp->segsum; 2808 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2809 lfs_fi_setnblocks(fs, sp->fip, 0); 2810 lfs_fi_setino(fs, sp->fip, ino); 2811 lfs_fi_setversion(fs, sp->fip, vers); 2812 } 2813 2814 /* 2815 * Release the FINFO entry, either clearing out an unused entry or 2816 * advancing us to the next available entry. 2817 */ 2818 void 2819 lfs_release_finfo(struct lfs *fs) 2820 { 2821 struct segment *sp = fs->lfs_sp; 2822 SEGSUM *ssp; 2823 2824 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2825 sp->fip = NEXT_FINFO(fs, sp->fip); 2826 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2827 } else { 2828 /* XXX shouldn't this update sp->fip? */ 2829 sp->sum_bytes_left += FINFOSIZE(fs); 2830 ssp = (SEGSUM *)sp->segsum; 2831 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2832 } 2833 } 2834
Indexes created Mon Sep 22 13:09:51 GMT 2025