lfs_segment.c revision 1.268
1 /* $NetBSD: lfs_segment.c,v 1.268 2017/04/06 03:15:03 maya 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 1.268 2017/04/06 03:15:03 maya 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; 488 int op = c->op; 489 490 KASSERT(mutex_owned(vp->v_interlock)); 491 492 ip = VTOI(vp); 493 if (ip == NULL || vp->v_type == VNON) 494 return false; 495 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 496 (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) { 497 vndebug(vp, "dirop"); 498 return false; 499 } 500 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 501 vndebug(vp,"empty"); 502 return false;; 503 } 504 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM && 505 vp != c->fs->lfs_flushvp && !(ip->i_flag & IN_CLEANING)) { 506 vndebug(vp,"cleaning"); 507 return false; 508 } 509 mutex_enter(&lfs_lock); 510 if (vp == c->fs->lfs_unlockvp) { 511 mutex_exit(&lfs_lock); 512 return false; 513 } 514 mutex_exit(&lfs_lock); 515 516 return true; 517 } 518 519 int 520 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 521 { 522 struct inode *ip; 523 struct vnode *vp; 524 struct vnode_iterator *marker; 525 struct lfs_writevnodes_ctx ctx; 526 int inodes_written = 0; 527 int error = 0; 528 529 /* 530 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist. 531 * XXX The rationale is unclear, the initial commit had no information. 532 * XXX If the order really matters we have to sort the vnodes first. 533 */ 534 535 ASSERT_SEGLOCK(fs); 536 vfs_vnode_iterator_init(mp, &marker); 537 ctx.op = op; 538 ctx.fs = fs; 539 while ((vp = vfs_vnode_iterator_next(marker, 540 lfs_writevnodes_selector, &ctx)) != NULL) { 541 ip = VTOI(vp); 542 543 /* 544 * Write the inode/file if dirty and it's not the IFILE. 545 */ 546 if (((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) && 547 ip->i_number != LFS_IFILE_INUM) { 548 error = lfs_writefile(fs, sp, vp); 549 if (error) { 550 vrele(vp); 551 if (error == EAGAIN) { 552 /* 553 * This error from lfs_putpages 554 * indicates we need to drop 555 * the segment lock and start 556 * over after the cleaner has 557 * had a chance to run. 558 */ 559 lfs_writeinode(fs, sp, ip); 560 lfs_writeseg(fs, sp); 561 if (!VPISEMPTY(vp) && 562 !WRITEINPROG(vp) && 563 !(ip->i_flag & IN_ALLMOD)) { 564 mutex_enter(&lfs_lock); 565 LFS_SET_UINO(ip, IN_MODIFIED); 566 mutex_exit(&lfs_lock); 567 } 568 break; 569 } 570 error = 0; /* XXX not quite right */ 571 continue; 572 } 573 574 if (!VPISEMPTY(vp)) { 575 if (WRITEINPROG(vp)) { 576 ivndebug(vp,"writevnodes/write2"); 577 } else if (!(ip->i_flag & IN_ALLMOD)) { 578 mutex_enter(&lfs_lock); 579 LFS_SET_UINO(ip, IN_MODIFIED); 580 mutex_exit(&lfs_lock); 581 } 582 } 583 (void) lfs_writeinode(fs, sp, ip); 584 inodes_written++; 585 } 586 vrele(vp); 587 } 588 vfs_vnode_iterator_destroy(marker); 589 return error; 590 } 591 592 /* 593 * Do a checkpoint. 594 */ 595 int 596 lfs_segwrite(struct mount *mp, int flags) 597 { 598 struct buf *bp; 599 struct inode *ip; 600 struct lfs *fs; 601 struct segment *sp; 602 struct vnode *vp; 603 SEGUSE *segusep; 604 int do_ckp, did_ckp, error; 605 unsigned n, segleft, maxseg, sn, i, curseg; 606 int writer_set = 0; 607 int dirty; 608 int redo; 609 SEGSUM *ssp; 610 int um_error; 611 612 fs = VFSTOULFS(mp)->um_lfs; 613 ASSERT_MAYBE_SEGLOCK(fs); 614 615 if (fs->lfs_ronly) 616 return EROFS; 617 618 lfs_imtime(fs); 619 620 /* 621 * Allocate a segment structure and enough space to hold pointers to 622 * the maximum possible number of buffers which can be described in a 623 * single summary block. 624 */ 625 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 626 627 /* We can't do a partial write and checkpoint at the same time. */ 628 if (do_ckp) 629 flags &= ~SEGM_SINGLE; 630 631 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 632 sp = fs->lfs_sp; 633 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 634 do_ckp = 1; 635 636 /* 637 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 638 * in which case we have to flush *all* buffers off of this vnode. 639 * We don't care about other nodes, but write any non-dirop nodes 640 * anyway in anticipation of another getnewvnode(). 641 * 642 * If we're cleaning we only write cleaning and ifile blocks, and 643 * no dirops, since otherwise we'd risk corruption in a crash. 644 */ 645 if (sp->seg_flags & SEGM_CLEAN) 646 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 647 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 648 do { 649 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 650 if ((sp->seg_flags & SEGM_SINGLE) && 651 lfs_sb_getcurseg(fs) != fs->lfs_startseg) { 652 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs))); 653 break; 654 } 655 656 if (do_ckp || fs->lfs_dirops == 0) { 657 if (!writer_set) { 658 lfs_writer_enter(fs, "lfs writer"); 659 writer_set = 1; 660 } 661 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 662 if (um_error == 0) 663 um_error = error; 664 /* In case writevnodes errored out */ 665 lfs_flush_dirops(fs); 666 ssp = (SEGSUM *)(sp->segsum); 667 lfs_ss_setflags(fs, ssp, 668 lfs_ss_getflags(fs, ssp) & ~(SS_CONT)); 669 lfs_finalize_fs_seguse(fs); 670 } 671 if (do_ckp && um_error) { 672 lfs_segunlock_relock(fs); 673 sp = fs->lfs_sp; 674 } 675 } while (do_ckp && um_error != 0); 676 } 677 678 /* 679 * If we are doing a checkpoint, mark everything since the 680 * last checkpoint as no longer ACTIVE. 681 */ 682 if (do_ckp || fs->lfs_doifile) { 683 segleft = lfs_sb_getnseg(fs); 684 curseg = 0; 685 for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) { 686 int bread_error; 687 688 dirty = 0; 689 bread_error = bread(fs->lfs_ivnode, 690 lfs_sb_getcleansz(fs) + n, 691 lfs_sb_getbsize(fs), B_MODIFY, &bp); 692 if (bread_error) 693 panic("lfs_segwrite: ifile read: " 694 "seguse %u: error %d\n", 695 n, bread_error); 696 segusep = (SEGUSE *)bp->b_data; 697 maxseg = min(segleft, lfs_sb_getsepb(fs)); 698 for (i = 0; i < maxseg; i++) { 699 sn = curseg + i; 700 if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) && 701 segusep->su_flags & SEGUSE_ACTIVE) { 702 segusep->su_flags &= ~SEGUSE_ACTIVE; 703 --fs->lfs_nactive; 704 ++dirty; 705 } 706 fs->lfs_suflags[fs->lfs_activesb][sn] = 707 segusep->su_flags; 708 if (lfs_sb_getversion(fs) > 1) 709 ++segusep; 710 else 711 segusep = (SEGUSE *) 712 ((SEGUSE_V1 *)segusep + 1); 713 } 714 715 if (dirty) 716 error = LFS_BWRITE_LOG(bp); /* Ifile */ 717 else 718 brelse(bp, 0); 719 segleft -= lfs_sb_getsepb(fs); 720 curseg += lfs_sb_getsepb(fs); 721 } 722 } 723 724 KASSERT(LFS_SEGLOCK_HELD(fs)); 725 726 did_ckp = 0; 727 if (do_ckp || fs->lfs_doifile) { 728 vp = fs->lfs_ivnode; 729 #ifdef DEBUG 730 int loopcount = 0; 731 #endif 732 do { 733 734 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 735 736 mutex_enter(&lfs_lock); 737 fs->lfs_flags &= ~LFS_IFDIRTY; 738 mutex_exit(&lfs_lock); 739 740 ip = VTOI(vp); 741 742 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 743 /* 744 * Ifile has no pages, so we don't need 745 * to check error return here. 746 */ 747 lfs_writefile(fs, sp, vp); 748 /* 749 * Ensure the Ifile takes the current segment 750 * into account. See comment in lfs_vflush. 751 */ 752 lfs_writefile(fs, sp, vp); 753 lfs_writefile(fs, sp, vp); 754 } 755 756 if (ip->i_flag & IN_ALLMOD) 757 ++did_ckp; 758 #if 0 759 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 760 #else 761 redo = lfs_writeinode(fs, sp, ip); 762 #endif 763 redo += lfs_writeseg(fs, sp); 764 mutex_enter(&lfs_lock); 765 redo += (fs->lfs_flags & LFS_IFDIRTY); 766 mutex_exit(&lfs_lock); 767 #ifdef DEBUG 768 if (++loopcount > 2) 769 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 770 loopcount); 771 #endif 772 } while (redo && do_ckp); 773 774 /* 775 * Unless we are unmounting, the Ifile may continue to have 776 * dirty blocks even after a checkpoint, due to changes to 777 * inodes' atime. If we're checkpointing, it's "impossible" 778 * for other parts of the Ifile to be dirty after the loop 779 * above, since we hold the segment lock. 780 */ 781 mutex_enter(vp->v_interlock); 782 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 783 LFS_CLR_UINO(ip, IN_ALLMOD); 784 } 785 #ifdef DIAGNOSTIC 786 else if (do_ckp) { 787 int do_panic = 0; 788 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 789 if (bp->b_lblkno < lfs_sb_getcleansz(fs) + 790 lfs_sb_getsegtabsz(fs) && 791 !(bp->b_flags & B_GATHERED)) { 792 printf("ifile lbn %ld still dirty (flags %lx)\n", 793 (long)bp->b_lblkno, 794 (long)bp->b_flags); 795 ++do_panic; 796 } 797 } 798 if (do_panic) 799 panic("dirty blocks"); 800 } 801 #endif 802 mutex_exit(vp->v_interlock); 803 } else { 804 (void) lfs_writeseg(fs, sp); 805 } 806 807 /* Note Ifile no longer needs to be written */ 808 fs->lfs_doifile = 0; 809 if (writer_set) 810 lfs_writer_leave(fs); 811 812 /* 813 * If we didn't write the Ifile, we didn't really do anything. 814 * That means that (1) there is a checkpoint on disk and (2) 815 * nothing has changed since it was written. 816 * 817 * Take the flags off of the segment so that lfs_segunlock 818 * doesn't have to write the superblock either. 819 */ 820 if (do_ckp && !did_ckp) { 821 sp->seg_flags &= ~SEGM_CKP; 822 } 823 824 if (lfs_dostats) { 825 ++lfs_stats.nwrites; 826 if (sp->seg_flags & SEGM_SYNC) 827 ++lfs_stats.nsync_writes; 828 if (sp->seg_flags & SEGM_CKP) 829 ++lfs_stats.ncheckpoints; 830 } 831 lfs_segunlock(fs); 832 return (0); 833 } 834 835 /* 836 * Write the dirty blocks associated with a vnode. 837 */ 838 int 839 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 840 { 841 struct inode *ip; 842 int i, frag; 843 SEGSUM *ssp; 844 int error; 845 846 ASSERT_SEGLOCK(fs); 847 error = 0; 848 ip = VTOI(vp); 849 850 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 851 852 if (vp->v_uflag & VU_DIROP) { 853 ssp = (SEGSUM *)sp->segsum; 854 lfs_ss_setflags(fs, ssp, 855 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 856 } 857 858 if (sp->seg_flags & SEGM_CLEAN) { 859 lfs_gather(fs, sp, vp, lfs_match_fake); 860 /* 861 * For a file being flushed, we need to write *all* blocks. 862 * This means writing the cleaning blocks first, and then 863 * immediately following with any non-cleaning blocks. 864 * The same is true of the Ifile since checkpoints assume 865 * that all valid Ifile blocks are written. 866 */ 867 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 868 lfs_gather(fs, sp, vp, lfs_match_data); 869 /* 870 * Don't call VOP_PUTPAGES: if we're flushing, 871 * we've already done it, and the Ifile doesn't 872 * use the page cache. 873 */ 874 } 875 } else { 876 lfs_gather(fs, sp, vp, lfs_match_data); 877 /* 878 * If we're flushing, we've already called VOP_PUTPAGES 879 * so don't do it again. Otherwise, we want to write 880 * everything we've got. 881 */ 882 if (!IS_FLUSHING(fs, vp)) { 883 mutex_enter(vp->v_interlock); 884 error = VOP_PUTPAGES(vp, 0, 0, 885 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 886 } 887 } 888 889 /* 890 * It may not be necessary to write the meta-data blocks at this point, 891 * as the roll-forward recovery code should be able to reconstruct the 892 * list. 893 * 894 * We have to write them anyway, though, under two conditions: (1) the 895 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 896 * checkpointing. 897 * 898 * BUT if we are cleaning, we might have indirect blocks that refer to 899 * new blocks not being written yet, in addition to fragments being 900 * moved out of a cleaned segment. If that is the case, don't 901 * write the indirect blocks, or the finfo will have a small block 902 * in the middle of it! 903 * XXX in this case isn't the inode size wrong too? 904 */ 905 frag = 0; 906 if (sp->seg_flags & SEGM_CLEAN) { 907 for (i = 0; i < ULFS_NDADDR; i++) 908 if (ip->i_lfs_fragsize[i] > 0 && 909 ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs)) 910 ++frag; 911 } 912 KASSERTMSG((frag <= 1), 913 "lfs_writefile: more than one fragment! frag=%d", frag); 914 if (IS_FLUSHING(fs, vp) || 915 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 916 lfs_gather(fs, sp, vp, lfs_match_indir); 917 lfs_gather(fs, sp, vp, lfs_match_dindir); 918 lfs_gather(fs, sp, vp, lfs_match_tindir); 919 } 920 lfs_release_finfo(fs); 921 922 return error; 923 } 924 925 /* 926 * Update segment accounting to reflect this inode's change of address. 927 */ 928 static int 929 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 930 { 931 struct buf *bp; 932 daddr_t daddr; 933 IFILE *ifp; 934 SEGUSE *sup; 935 ino_t ino; 936 int redo_ifile; 937 u_int32_t sn; 938 939 redo_ifile = 0; 940 941 /* 942 * If updating the ifile, update the super-block. Update the disk 943 * address and access times for this inode in the ifile. 944 */ 945 ino = ip->i_number; 946 if (ino == LFS_IFILE_INUM) { 947 daddr = lfs_sb_getidaddr(fs); 948 lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr)); 949 } else { 950 LFS_IENTRY(ifp, fs, ino, bp); 951 daddr = lfs_if_getdaddr(fs, ifp); 952 lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr)); 953 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 954 } 955 956 /* 957 * If this is the Ifile and lfs_offset is set to the first block 958 * in the segment, dirty the new segment's accounting block 959 * (XXX should already be dirty?) and tell the caller to do it again. 960 */ 961 if (ip->i_number == LFS_IFILE_INUM) { 962 sn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 963 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) == 964 lfs_sb_getoffset(fs)) { 965 LFS_SEGENTRY(sup, fs, sn, bp); 966 KASSERT(bp->b_oflags & BO_DELWRI); 967 LFS_WRITESEGENTRY(sup, fs, sn, bp); 968 /* fs->lfs_flags |= LFS_IFDIRTY; */ 969 redo_ifile |= 1; 970 } 971 } 972 973 /* 974 * The inode's last address should not be in the current partial 975 * segment, except under exceptional circumstances (lfs_writevnodes 976 * had to start over, and in the meantime more blocks were written 977 * to a vnode). Both inodes will be accounted to this segment 978 * in lfs_writeseg so we need to subtract the earlier version 979 * here anyway. The segment count can temporarily dip below 980 * zero here; keep track of how many duplicates we have in 981 * "dupino" so we don't panic below. 982 */ 983 if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) { 984 ++sp->ndupino; 985 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 986 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 987 (long long)daddr, sp->ndupino)); 988 } 989 /* 990 * Account the inode: it no longer belongs to its former segment, 991 * though it will not belong to the new segment until that segment 992 * is actually written. 993 */ 994 if (daddr != LFS_UNUSED_DADDR) { 995 u_int32_t oldsn = lfs_dtosn(fs, daddr); 996 int ndupino __diagused = 997 (sp->seg_number == oldsn) ? sp->ndupino : 0; 998 LFS_SEGENTRY(sup, fs, oldsn, bp); 999 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) 1000 >= DINOSIZE(fs)), 1001 "lfs_writeinode: negative bytes " 1002 "(segment %" PRIu32 " short by %d, " 1003 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1004 ", daddr=%" PRId64 ", su_nbytes=%u, " 1005 "ndupino=%d)\n", 1006 lfs_dtosn(fs, daddr), 1007 (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes, 1008 oldsn, sp->seg_number, daddr, 1009 (unsigned int)sup->su_nbytes, 1010 sp->ndupino); 1011 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1012 lfs_dtosn(fs, daddr), DINOSIZE(fs), ino)); 1013 sup->su_nbytes -= DINOSIZE(fs); 1014 redo_ifile |= 1015 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1016 if (redo_ifile) { 1017 mutex_enter(&lfs_lock); 1018 fs->lfs_flags |= LFS_IFDIRTY; 1019 mutex_exit(&lfs_lock); 1020 /* Don't double-account */ 1021 lfs_sb_setidaddr(fs, 0x0); 1022 } 1023 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1024 } 1025 1026 return redo_ifile; 1027 } 1028 1029 int 1030 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1031 { 1032 struct buf *bp; 1033 union lfs_dinode *cdp; 1034 struct vnode *vp = ITOV(ip); 1035 daddr_t daddr; 1036 IINFO *iip; 1037 int i; 1038 int redo_ifile = 0; 1039 int gotblk = 0; 1040 int count; 1041 SEGSUM *ssp; 1042 1043 ASSERT_SEGLOCK(fs); 1044 if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1045 return (0); 1046 1047 /* Can't write ifile when writer is not set */ 1048 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1049 (sp->seg_flags & SEGM_CLEAN)); 1050 1051 /* 1052 * If this is the Ifile, see if writing it here will generate a 1053 * temporary misaccounting. If it will, do the accounting and write 1054 * the blocks, postponing the inode write until the accounting is 1055 * solid. 1056 */ 1057 count = 0; 1058 while (vp == fs->lfs_ivnode) { 1059 int redo = 0; 1060 1061 if (sp->idp == NULL && sp->ibp == NULL && 1062 (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1063 sp->sum_bytes_left < sizeof(int32_t))) { 1064 (void) lfs_writeseg(fs, sp); 1065 continue; 1066 } 1067 1068 /* Look for dirty Ifile blocks */ 1069 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1070 if (!(bp->b_flags & B_GATHERED)) { 1071 redo = 1; 1072 break; 1073 } 1074 } 1075 1076 if (redo == 0) 1077 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1078 if (redo == 0) 1079 break; 1080 1081 if (sp->idp) { 1082 lfs_dino_setinumber(fs, sp->idp, 0); 1083 sp->idp = NULL; 1084 } 1085 ++count; 1086 if (count > 2) 1087 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1088 lfs_writefile(fs, sp, fs->lfs_ivnode); 1089 } 1090 1091 /* Allocate a new inode block if necessary. */ 1092 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1093 sp->ibp == NULL) { 1094 /* Allocate a new segment if necessary. */ 1095 if (sp->seg_bytes_left < lfs_sb_getibsize(fs) || 1096 sp->sum_bytes_left < sizeof(int32_t)) 1097 (void) lfs_writeseg(fs, sp); 1098 1099 /* Get next inode block. */ 1100 daddr = lfs_sb_getoffset(fs); 1101 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1102 sp->ibp = *sp->cbpp++ = 1103 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1104 LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0); 1105 gotblk++; 1106 1107 /* Zero out inode numbers */ 1108 for (i = 0; i < LFS_INOPB(fs); ++i) { 1109 union lfs_dinode *tmpdi; 1110 1111 tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data + 1112 DINOSIZE(fs) * i); 1113 lfs_dino_setinumber(fs, tmpdi, 0); 1114 } 1115 1116 ++sp->start_bpp; 1117 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs))); 1118 /* Set remaining space counters. */ 1119 sp->seg_bytes_left -= lfs_sb_getibsize(fs); 1120 sp->sum_bytes_left -= sizeof(int32_t); 1121 1122 /* Store the address in the segment summary. */ 1123 iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs)); 1124 lfs_ii_setblock(fs, iip, daddr); 1125 } 1126 1127 /* Check VU_DIROP in case there is a new file with no data blocks */ 1128 if (vp->v_uflag & VU_DIROP) { 1129 ssp = (SEGSUM *)sp->segsum; 1130 lfs_ss_setflags(fs, ssp, 1131 lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT)); 1132 } 1133 1134 /* Update the inode times and copy the inode onto the inode page. */ 1135 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1136 if (ip->i_number != LFS_IFILE_INUM) 1137 LFS_ITIMES(ip, NULL, NULL, NULL); 1138 1139 /* 1140 * If this is the Ifile, and we've already written the Ifile in this 1141 * partial segment, just overwrite it (it's not on disk yet) and 1142 * continue. 1143 * 1144 * XXX we know that the bp that we get the second time around has 1145 * already been gathered. 1146 */ 1147 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1148 lfs_copy_dinode(fs, sp->idp, ip->i_din); 1149 ip->i_lfs_osize = ip->i_size; 1150 return 0; 1151 } 1152 1153 bp = sp->ibp; 1154 cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs)); 1155 lfs_copy_dinode(fs, cdp, ip->i_din); 1156 1157 /* 1158 * This inode is on its way to disk; clear its VU_DIROP status when 1159 * the write is complete. 1160 */ 1161 if (vp->v_uflag & VU_DIROP) { 1162 if (!(sp->seg_flags & SEGM_CLEAN)) 1163 ip->i_flag |= IN_CDIROP; 1164 else { 1165 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1166 } 1167 } 1168 1169 /* 1170 * If cleaning, link counts and directory file sizes cannot change, 1171 * since those would be directory operations---even if the file 1172 * we are writing is marked VU_DIROP we should write the old values. 1173 * If we're not cleaning, of course, update the values so we get 1174 * current values the next time we clean. 1175 */ 1176 if (sp->seg_flags & SEGM_CLEAN) { 1177 if (vp->v_uflag & VU_DIROP) { 1178 lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink); 1179 /* if (vp->v_type == VDIR) */ 1180 lfs_dino_setsize(fs, cdp, ip->i_lfs_osize); 1181 } 1182 } else { 1183 ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp); 1184 ip->i_lfs_osize = ip->i_size; 1185 } 1186 1187 1188 /* We can finish the segment accounting for truncations now */ 1189 lfs_finalize_ino_seguse(fs, ip); 1190 1191 /* 1192 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1193 * addresses to disk; possibly change the on-disk record of 1194 * the inode size, either by reverting to the previous size 1195 * (in the case of cleaning) or by verifying the inode's block 1196 * holdings (in the case of files being allocated as they are being 1197 * written). 1198 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1199 * XXX count on disk wrong by the same amount. We should be 1200 * XXX able to "borrow" from lfs_avail and return it after the 1201 * XXX Ifile is written. See also in lfs_writeseg. 1202 */ 1203 1204 /* Check file size based on highest allocated block */ 1205 if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG || 1206 (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) && 1207 ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) { 1208 lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs)); 1209 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1210 PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp))); 1211 } 1212 if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) { 1213 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)" 1214 " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks, 1215 lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs))); 1216 for (i=0; i<ULFS_NDADDR; i++) { 1217 if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) { 1218 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1219 lfs_dino_setdb(fs, cdp, i, 0); 1220 } 1221 } 1222 for (i=0; i<ULFS_NIADDR; i++) { 1223 if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) { 1224 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1225 lfs_dino_setib(fs, cdp, i, 0); 1226 } 1227 } 1228 } 1229 1230 #ifdef DIAGNOSTIC 1231 /* 1232 * Check dinode held blocks against dinode size. 1233 * This should be identical to the check in lfs_vget(). 1234 */ 1235 for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs); 1236 i < ULFS_NDADDR; i++) { 1237 KASSERT(i >= 0); 1238 if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK) 1239 continue; 1240 if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK || 1241 (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0) 1242 continue; 1243 if (lfs_dino_getdb(fs, cdp, i) != 0) { 1244 # ifdef DEBUG 1245 lfs_dump_dinode(fs, cdp); 1246 # endif 1247 panic("writing inconsistent inode"); 1248 } 1249 } 1250 #endif /* DIAGNOSTIC */ 1251 1252 if (ip->i_flag & IN_CLEANING) 1253 LFS_CLR_UINO(ip, IN_CLEANING); 1254 else { 1255 /* XXX IN_ALLMOD */ 1256 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1257 IN_UPDATE | IN_MODIFY); 1258 if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din)) 1259 LFS_CLR_UINO(ip, IN_MODIFIED); 1260 else { 1261 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1262 "blks=%d, eff=%jd\n", ip->i_number, 1263 lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks)); 1264 } 1265 } 1266 1267 if (ip->i_number == LFS_IFILE_INUM) { 1268 /* We know sp->idp == NULL */ 1269 sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs)); 1270 1271 /* Not dirty any more */ 1272 mutex_enter(&lfs_lock); 1273 fs->lfs_flags &= ~LFS_IFDIRTY; 1274 mutex_exit(&lfs_lock); 1275 } 1276 1277 if (gotblk) { 1278 mutex_enter(&bufcache_lock); 1279 LFS_LOCK_BUF(bp); 1280 brelsel(bp, 0); 1281 mutex_exit(&bufcache_lock); 1282 } 1283 1284 /* Increment inode count in segment summary block. */ 1285 1286 ssp = (SEGSUM *)sp->segsum; 1287 lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1); 1288 1289 /* If this page is full, set flag to allocate a new page. */ 1290 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1291 sp->ibp = NULL; 1292 1293 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1294 1295 KASSERT(redo_ifile == 0); 1296 return (redo_ifile); 1297 } 1298 1299 int 1300 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1301 { 1302 struct lfs *fs; 1303 int vers; 1304 int j, blksinblk; 1305 1306 ASSERT_SEGLOCK(sp->fs); 1307 KASSERTMSG((sp->vp != NULL), 1308 "lfs_gatherblock: Null vp in segment"); 1309 1310 /* If full, finish this segment. */ 1311 fs = sp->fs; 1312 blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs)); 1313 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1314 sp->seg_bytes_left < bp->b_bcount) { 1315 if (mptr) 1316 mutex_exit(mptr); 1317 lfs_updatemeta(sp); 1318 1319 vers = lfs_fi_getversion(fs, sp->fip); 1320 (void) lfs_writeseg(fs, sp); 1321 1322 /* Add the current file to the segment summary. */ 1323 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1324 1325 if (mptr) 1326 mutex_enter(mptr); 1327 return (1); 1328 } 1329 1330 if (bp->b_flags & B_GATHERED) { 1331 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju," 1332 " lbn %" PRId64 "\n", 1333 (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno)); 1334 return (0); 1335 } 1336 1337 /* Insert into the buffer list, update the FINFO block. */ 1338 bp->b_flags |= B_GATHERED; 1339 1340 *sp->cbpp++ = bp; 1341 for (j = 0; j < blksinblk; j++) { 1342 unsigned bn; 1343 1344 bn = lfs_fi_getnblocks(fs, sp->fip); 1345 lfs_fi_setnblocks(fs, sp->fip, bn+1); 1346 lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j); 1347 /* This block's accounting moves from lfs_favail to lfs_avail */ 1348 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1349 } 1350 1351 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1352 sp->seg_bytes_left -= bp->b_bcount; 1353 return (0); 1354 } 1355 1356 int 1357 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1358 int (*match)(struct lfs *, struct buf *)) 1359 { 1360 struct buf *bp, *nbp; 1361 int count = 0; 1362 1363 ASSERT_SEGLOCK(fs); 1364 if (vp->v_type == VBLK) 1365 return 0; 1366 KASSERT(sp->vp == NULL); 1367 sp->vp = vp; 1368 mutex_enter(&bufcache_lock); 1369 1370 #ifndef LFS_NO_BACKBUF_HACK 1371 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1372 # define BUF_OFFSET \ 1373 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1374 # define BACK_BUF(BP) \ 1375 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1376 # define BEG_OF_LIST \ 1377 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1378 1379 loop: 1380 /* Find last buffer. */ 1381 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1382 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1383 bp = LIST_NEXT(bp, b_vnbufs)) 1384 /* nothing */; 1385 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1386 nbp = BACK_BUF(bp); 1387 #else /* LFS_NO_BACKBUF_HACK */ 1388 loop: 1389 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1390 nbp = LIST_NEXT(bp, b_vnbufs); 1391 #endif /* LFS_NO_BACKBUF_HACK */ 1392 if ((bp->b_cflags & BC_BUSY) != 0 || 1393 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1394 #ifdef DEBUG 1395 if (vp == fs->lfs_ivnode && 1396 (bp->b_cflags & BC_BUSY) != 0 && 1397 (bp->b_flags & B_GATHERED) == 0) 1398 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1399 PRId64 " busy (%x) at 0x%jx", 1400 bp->b_lblkno, bp->b_flags, 1401 (uintmax_t)lfs_sb_getoffset(fs)); 1402 #endif 1403 continue; 1404 } 1405 #ifdef DIAGNOSTIC 1406 # ifdef LFS_USE_B_INVAL 1407 if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1408 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1409 " is BC_INVAL\n", bp->b_lblkno)); 1410 VOP_PRINT(bp->b_vp); 1411 } 1412 # endif /* LFS_USE_B_INVAL */ 1413 if (!(bp->b_oflags & BO_DELWRI)) 1414 panic("lfs_gather: bp not BO_DELWRI"); 1415 if (!(bp->b_flags & B_LOCKED)) { 1416 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1417 " blk %" PRId64 " not B_LOCKED\n", 1418 bp->b_lblkno, 1419 LFS_DBTOFSB(fs, bp->b_blkno))); 1420 VOP_PRINT(bp->b_vp); 1421 panic("lfs_gather: bp not B_LOCKED"); 1422 } 1423 #endif 1424 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1425 goto loop; 1426 } 1427 count++; 1428 } 1429 mutex_exit(&bufcache_lock); 1430 lfs_updatemeta(sp); 1431 KASSERT(sp->vp == vp); 1432 sp->vp = NULL; 1433 return count; 1434 } 1435 1436 #if DEBUG 1437 # define DEBUG_OOFF(n) do { \ 1438 if (ooff == 0) { \ 1439 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1440 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1441 ", was 0x0 (or %" PRId64 ")\n", \ 1442 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1443 } \ 1444 } while (0) 1445 #else 1446 # define DEBUG_OOFF(n) 1447 #endif 1448 1449 /* 1450 * Change the given block's address to ndaddr, finding its previous 1451 * location using ulfs_bmaparray(). 1452 * 1453 * Account for this change in the segment table. 1454 * 1455 * called with sp == NULL by roll-forwarding code. 1456 */ 1457 void 1458 lfs_update_single(struct lfs *fs, struct segment *sp, 1459 struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size) 1460 { 1461 SEGUSE *sup; 1462 struct buf *bp; 1463 struct indir a[ULFS_NIADDR + 2], *ap; 1464 struct inode *ip; 1465 daddr_t daddr, ooff; 1466 int num, error; 1467 int bb, osize, obb; 1468 1469 ASSERT_SEGLOCK(fs); 1470 KASSERT(sp == NULL || sp->vp == vp); 1471 ip = VTOI(vp); 1472 1473 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1474 if (error) 1475 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1476 1477 KASSERT(daddr <= LFS_MAX_DADDR(fs)); 1478 if (daddr > 0) 1479 daddr = LFS_DBTOFSB(fs, daddr); 1480 1481 bb = lfs_numfrags(fs, size); 1482 switch (num) { 1483 case 0: 1484 ooff = lfs_dino_getdb(fs, ip->i_din, lbn); 1485 DEBUG_OOFF(0); 1486 if (ooff == UNWRITTEN) 1487 lfs_dino_setblocks(fs, ip->i_din, 1488 lfs_dino_getblocks(fs, ip->i_din) + bb); 1489 else { 1490 /* possible fragment truncation or extension */ 1491 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1492 lfs_dino_setblocks(fs, ip->i_din, 1493 lfs_dino_getblocks(fs, ip->i_din) + (bb-obb)); 1494 } 1495 lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr); 1496 break; 1497 case 1: 1498 ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off); 1499 DEBUG_OOFF(1); 1500 if (ooff == UNWRITTEN) 1501 lfs_dino_setblocks(fs, ip->i_din, 1502 lfs_dino_getblocks(fs, ip->i_din) + bb); 1503 lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr); 1504 break; 1505 default: 1506 ap = &a[num - 1]; 1507 if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs), 1508 B_MODIFY, &bp)) 1509 panic("lfs_updatemeta: bread bno %" PRId64, 1510 ap->in_lbn); 1511 1512 ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off); 1513 DEBUG_OOFF(num); 1514 if (ooff == UNWRITTEN) 1515 lfs_dino_setblocks(fs, ip->i_din, 1516 lfs_dino_getblocks(fs, ip->i_din) + bb); 1517 lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr); 1518 (void) VOP_BWRITE(bp->b_vp, bp); 1519 } 1520 1521 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1522 1523 /* Update hiblk when extending the file */ 1524 if (lbn > ip->i_lfs_hiblk) 1525 ip->i_lfs_hiblk = lbn; 1526 1527 /* 1528 * Though we'd rather it couldn't, this *can* happen right now 1529 * if cleaning blocks and regular blocks coexist. 1530 */ 1531 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1532 1533 /* 1534 * Update segment usage information, based on old size 1535 * and location. 1536 */ 1537 if (daddr > 0) { 1538 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1539 int ndupino __diagused = (sp && sp->seg_number == oldsn ? 1540 sp->ndupino : 0); 1541 1542 KASSERT(oldsn < lfs_sb_getnseg(fs)); 1543 if (lbn >= 0 && lbn < ULFS_NDADDR) 1544 osize = ip->i_lfs_fragsize[lbn]; 1545 else 1546 osize = lfs_sb_getbsize(fs); 1547 LFS_SEGENTRY(sup, fs, oldsn, bp); 1548 KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize), 1549 "lfs_updatemeta: negative bytes " 1550 "(segment %" PRIu32 " short by %" PRId64 1551 ")\n" 1552 "lfs_updatemeta: ino %llu, lbn %" PRId64 1553 ", addr = 0x%" PRIx64 "\n" 1554 "lfs_updatemeta: ndupino=%d", 1555 lfs_dtosn(fs, daddr), 1556 (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes), 1557 (unsigned long long)ip->i_number, lbn, daddr, 1558 ndupino); 1559 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1560 " db 0x%" PRIx64 "\n", 1561 lfs_dtosn(fs, daddr), osize, 1562 ip->i_number, lbn, daddr)); 1563 sup->su_nbytes -= osize; 1564 if (!(bp->b_flags & B_GATHERED)) { 1565 mutex_enter(&lfs_lock); 1566 fs->lfs_flags |= LFS_IFDIRTY; 1567 mutex_exit(&lfs_lock); 1568 } 1569 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1570 } 1571 /* 1572 * Now that this block has a new address, and its old 1573 * segment no longer owns it, we can forget about its 1574 * old size. 1575 */ 1576 if (lbn >= 0 && lbn < ULFS_NDADDR) 1577 ip->i_lfs_fragsize[lbn] = size; 1578 } 1579 1580 /* 1581 * Update the metadata that points to the blocks listed in the FINFO 1582 * array. 1583 */ 1584 void 1585 lfs_updatemeta(struct segment *sp) 1586 { 1587 struct buf *sbp; 1588 struct lfs *fs; 1589 struct vnode *vp; 1590 daddr_t lbn; 1591 int i, nblocks, num; 1592 int __diagused nblocks_orig; 1593 int bb; 1594 int bytesleft, size; 1595 unsigned lastlength; 1596 union lfs_blocks tmpptr; 1597 1598 fs = sp->fs; 1599 vp = sp->vp; 1600 ASSERT_SEGLOCK(fs); 1601 1602 /* 1603 * This used to be: 1604 * 1605 * nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1606 * 1607 * that is, it allowed for the possibility that start_lbp did 1608 * not point to the beginning of the finfo block pointer area. 1609 * This particular formulation is six kinds of painful in the 1610 * lfs64 world where we have two sizes of block pointer, so 1611 * unless/until everything can be cleaned up to not move 1612 * start_lbp around but instead use an offset, we do the 1613 * following: 1614 * 1. Get NEXT_FINFO(sp->fip). This is the same pointer as 1615 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong 1616 * type. (Ugh.) 1617 * 2. Cast it to void *, then assign it to a temporary 1618 * union lfs_blocks. 1619 * 3. Subtract start_lbp from that. 1620 * 4. Save the value of nblocks in blocks_orig so we can 1621 * assert below that it hasn't changed without repeating this 1622 * rubbish. 1623 * 1624 * XXX. 1625 */ 1626 lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip)); 1627 nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp); 1628 nblocks_orig = nblocks; 1629 1630 KASSERT(nblocks >= 0); 1631 KASSERT(vp != NULL); 1632 if (nblocks == 0) 1633 return; 1634 1635 /* 1636 * This count may be high due to oversize blocks from lfs_gop_write. 1637 * Correct for this. (XXX we should be able to keep track of these.) 1638 */ 1639 for (i = 0; i < nblocks; i++) { 1640 if (sp->start_bpp[i] == NULL) { 1641 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1642 nblocks = i; 1643 break; 1644 } 1645 num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs)); 1646 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1647 nblocks -= num - 1; 1648 } 1649 1650 #if 0 1651 /* pre-lfs64 assertion */ 1652 KASSERT(vp->v_type == VREG || 1653 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1654 #else 1655 KASSERT(vp->v_type == VREG || nblocks == nblocks_orig); 1656 #endif 1657 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1658 1659 /* 1660 * Sort the blocks. 1661 * 1662 * We have to sort even if the blocks come from the 1663 * cleaner, because there might be other pending blocks on the 1664 * same inode...and if we don't sort, and there are fragments 1665 * present, blocks may be written in the wrong place. 1666 */ 1667 lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs)); 1668 1669 /* 1670 * Record the length of the last block in case it's a fragment. 1671 * If there are indirect blocks present, they sort last. An 1672 * indirect block will be lfs_bsize and its presence indicates 1673 * that you cannot have fragments. 1674 * 1675 * XXX This last is a lie. A cleaned fragment can coexist with 1676 * XXX a later indirect block. This will continue to be 1677 * XXX true until lfs_markv is fixed to do everything with 1678 * XXX fake blocks (including fake inodes and fake indirect blocks). 1679 */ 1680 lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1681 lfs_sb_getbmask(fs)) + 1; 1682 lfs_fi_setlastlength(fs, sp->fip, lastlength); 1683 1684 /* 1685 * Assign disk addresses, and update references to the logical 1686 * block and the segment usage information. 1687 */ 1688 for (i = nblocks; i--; ++sp->start_bpp) { 1689 sbp = *sp->start_bpp; 1690 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1691 KASSERT(sbp->b_lblkno == lbn); 1692 1693 sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs)); 1694 1695 /* 1696 * If we write a frag in the wrong place, the cleaner won't 1697 * be able to correctly identify its size later, and the 1698 * segment will be uncleanable. (Even worse, it will assume 1699 * that the indirect block that actually ends the list 1700 * is of a smaller size!) 1701 */ 1702 if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0) 1703 panic("lfs_updatemeta: fragment is not last block"); 1704 1705 /* 1706 * For each subblock in this possibly oversized block, 1707 * update its address on disk. 1708 */ 1709 KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs)); 1710 KASSERT(vp == sbp->b_vp); 1711 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1712 bytesleft -= lfs_sb_getbsize(fs)) { 1713 size = MIN(bytesleft, lfs_sb_getbsize(fs)); 1714 bb = lfs_numfrags(fs, size); 1715 lbn = lfs_blocks_get(fs, &sp->start_lbp, 0); 1716 lfs_blocks_inc(fs, &sp->start_lbp); 1717 lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs), 1718 size); 1719 lfs_sb_addoffset(fs, bb); 1720 } 1721 1722 } 1723 1724 /* This inode has been modified */ 1725 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1726 } 1727 1728 /* 1729 * Move lfs_offset to a segment earlier than newsn. 1730 */ 1731 int 1732 lfs_rewind(struct lfs *fs, int newsn) 1733 { 1734 int sn, osn, isdirty; 1735 struct buf *bp; 1736 SEGUSE *sup; 1737 1738 ASSERT_SEGLOCK(fs); 1739 1740 osn = lfs_dtosn(fs, lfs_sb_getoffset(fs)); 1741 if (osn < newsn) 1742 return 0; 1743 1744 /* lfs_avail eats the remaining space in this segment */ 1745 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1746 1747 /* Find a low-numbered segment */ 1748 for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) { 1749 LFS_SEGENTRY(sup, fs, sn, bp); 1750 isdirty = sup->su_flags & SEGUSE_DIRTY; 1751 brelse(bp, 0); 1752 1753 if (!isdirty) 1754 break; 1755 } 1756 if (sn == lfs_sb_getnseg(fs)) 1757 panic("lfs_rewind: no clean segments"); 1758 if (newsn >= 0 && sn >= newsn) 1759 return ENOENT; 1760 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1761 lfs_newseg(fs); 1762 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1763 1764 return 0; 1765 } 1766 1767 /* 1768 * Start a new partial segment. 1769 * 1770 * Return 1 when we entered to a new segment. 1771 * Otherwise, return 0. 1772 */ 1773 int 1774 lfs_initseg(struct lfs *fs) 1775 { 1776 struct segment *sp = fs->lfs_sp; 1777 SEGSUM *ssp; 1778 struct buf *sbp; /* buffer for SEGSUM */ 1779 int repeat = 0; /* return value */ 1780 1781 ASSERT_SEGLOCK(fs); 1782 /* Advance to the next segment. */ 1783 if (!LFS_PARTIAL_FITS(fs)) { 1784 SEGUSE *sup; 1785 struct buf *bp; 1786 1787 /* lfs_avail eats the remaining space */ 1788 lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - 1789 lfs_sb_getcurseg(fs))); 1790 /* Wake up any cleaning procs waiting on this file system. */ 1791 lfs_wakeup_cleaner(fs); 1792 lfs_newseg(fs); 1793 repeat = 1; 1794 lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs)); 1795 1796 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1797 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs)); 1798 1799 /* 1800 * If the segment contains a superblock, update the offset 1801 * and summary address to skip over it. 1802 */ 1803 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1804 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1805 lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD)); 1806 sp->seg_bytes_left -= LFS_SBPAD; 1807 } 1808 brelse(bp, 0); 1809 /* Segment zero could also contain the labelpad */ 1810 if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 && 1811 lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) { 1812 lfs_sb_addoffset(fs, 1813 lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs)); 1814 sp->seg_bytes_left -= 1815 LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs)); 1816 } 1817 } else { 1818 sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs)); 1819 sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) - 1820 (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs))); 1821 } 1822 lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs)); 1823 1824 /* Record first address of this partial segment */ 1825 if (sp->seg_flags & SEGM_CLEAN) { 1826 fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs); 1827 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1828 /* "1" is the artificial inc in lfs_seglock */ 1829 mutex_enter(&lfs_lock); 1830 while (fs->lfs_iocount > 1) { 1831 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1832 "lfs_initseg", 0, &lfs_lock); 1833 } 1834 mutex_exit(&lfs_lock); 1835 fs->lfs_cleanind = 0; 1836 } 1837 } 1838 1839 sp->fs = fs; 1840 sp->ibp = NULL; 1841 sp->idp = NULL; 1842 sp->ninodes = 0; 1843 sp->ndupino = 0; 1844 1845 sp->cbpp = sp->bpp; 1846 1847 /* Get a new buffer for SEGSUM */ 1848 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1849 LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY); 1850 1851 /* ... and enter it into the buffer list. */ 1852 *sp->cbpp = sbp; 1853 sp->cbpp++; 1854 lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 1855 1856 sp->start_bpp = sp->cbpp; 1857 1858 /* Set point to SEGSUM, initialize it. */ 1859 ssp = sp->segsum = sbp->b_data; 1860 memset(ssp, 0, lfs_sb_getsumsize(fs)); 1861 lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs)); 1862 lfs_ss_setnfinfo(fs, ssp, 0); 1863 lfs_ss_setninos(fs, ssp, 0); 1864 lfs_ss_setmagic(fs, ssp, SS_MAGIC); 1865 1866 /* Set pointer to first FINFO, initialize it. */ 1867 sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum); 1868 lfs_fi_setnblocks(fs, sp->fip, 0); 1869 lfs_fi_setlastlength(fs, sp->fip, 0); 1870 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 1871 1872 sp->seg_bytes_left -= lfs_sb_getsumsize(fs); 1873 sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs); 1874 1875 return (repeat); 1876 } 1877 1878 /* 1879 * Remove SEGUSE_INVAL from all segments. 1880 */ 1881 void 1882 lfs_unset_inval_all(struct lfs *fs) 1883 { 1884 SEGUSE *sup; 1885 struct buf *bp; 1886 int i; 1887 1888 for (i = 0; i < lfs_sb_getnseg(fs); i++) { 1889 LFS_SEGENTRY(sup, fs, i, bp); 1890 if (sup->su_flags & SEGUSE_INVAL) { 1891 sup->su_flags &= ~SEGUSE_INVAL; 1892 LFS_WRITESEGENTRY(sup, fs, i, bp); 1893 } else 1894 brelse(bp, 0); 1895 } 1896 } 1897 1898 /* 1899 * Return the next segment to write. 1900 */ 1901 void 1902 lfs_newseg(struct lfs *fs) 1903 { 1904 CLEANERINFO *cip; 1905 SEGUSE *sup; 1906 struct buf *bp; 1907 int curseg, isdirty, sn, skip_inval; 1908 1909 ASSERT_SEGLOCK(fs); 1910 1911 /* Honor LFCNWRAPSTOP */ 1912 mutex_enter(&lfs_lock); 1913 while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) { 1914 if (fs->lfs_wrappass) { 1915 log(LOG_NOTICE, "%s: wrappass=%d\n", 1916 lfs_sb_getfsmnt(fs), fs->lfs_wrappass); 1917 fs->lfs_wrappass = 0; 1918 break; 1919 } 1920 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1921 wakeup(&fs->lfs_nowrap); 1922 log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs)); 1923 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1924 &lfs_lock); 1925 } 1926 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1927 mutex_exit(&lfs_lock); 1928 1929 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1930 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1931 lfs_dtosn(fs, lfs_sb_getnextseg(fs)))); 1932 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1933 sup->su_nbytes = 0; 1934 sup->su_nsums = 0; 1935 sup->su_ninos = 0; 1936 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp); 1937 1938 LFS_CLEANERINFO(cip, fs, bp); 1939 lfs_ci_shiftcleantodirty(fs, cip, 1); 1940 lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip)); 1941 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1942 1943 lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs)); 1944 lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs)); 1945 skip_inval = 1; 1946 for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) { 1947 sn = (sn + 1) % lfs_sb_getnseg(fs); 1948 1949 if (sn == curseg) { 1950 if (skip_inval) 1951 skip_inval = 0; 1952 else 1953 panic("lfs_nextseg: no clean segments"); 1954 } 1955 LFS_SEGENTRY(sup, fs, sn, bp); 1956 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1957 /* Check SEGUSE_EMPTY as we go along */ 1958 if (isdirty && sup->su_nbytes == 0 && 1959 !(sup->su_flags & SEGUSE_EMPTY)) 1960 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1961 else 1962 brelse(bp, 0); 1963 1964 if (!isdirty) 1965 break; 1966 } 1967 if (skip_inval == 0) 1968 lfs_unset_inval_all(fs); 1969 1970 ++fs->lfs_nactive; 1971 lfs_sb_setnextseg(fs, lfs_sntod(fs, sn)); 1972 if (lfs_dostats) { 1973 ++lfs_stats.segsused; 1974 } 1975 } 1976 1977 static struct buf * 1978 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1979 int n) 1980 { 1981 struct lfs_cluster *cl; 1982 struct buf **bpp, *bp; 1983 1984 ASSERT_SEGLOCK(fs); 1985 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1986 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1987 memset(cl, 0, sizeof(*cl)); 1988 cl->fs = fs; 1989 cl->bpp = bpp; 1990 cl->bufcount = 0; 1991 cl->bufsize = 0; 1992 1993 /* If this segment is being written synchronously, note that */ 1994 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 1995 cl->flags |= LFS_CL_SYNC; 1996 cl->seg = fs->lfs_sp; 1997 ++cl->seg->seg_iocount; 1998 } 1999 2000 /* Get an empty buffer header, or maybe one with something on it */ 2001 bp = getiobuf(vp, true); 2002 bp->b_dev = NODEV; 2003 bp->b_blkno = bp->b_lblkno = addr; 2004 bp->b_iodone = lfs_cluster_callback; 2005 bp->b_private = cl; 2006 2007 return bp; 2008 } 2009 2010 int 2011 lfs_writeseg(struct lfs *fs, struct segment *sp) 2012 { 2013 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 2014 SEGUSE *sup; 2015 SEGSUM *ssp; 2016 int i; 2017 int do_again, nblocks, byteoffset; 2018 size_t el_size; 2019 struct lfs_cluster *cl; 2020 u_short ninos; 2021 struct vnode *devvp; 2022 char *p = NULL; 2023 struct vnode *vp; 2024 unsigned ibindex, iblimit; 2025 int changed; 2026 u_int32_t sum; 2027 size_t sumstart; 2028 #ifdef DEBUG 2029 FINFO *fip; 2030 int findex; 2031 #endif 2032 2033 ASSERT_SEGLOCK(fs); 2034 2035 ssp = (SEGSUM *)sp->segsum; 2036 2037 /* 2038 * If there are no buffers other than the segment summary to write, 2039 * don't do anything. If we are the end of a dirop sequence, however, 2040 * write the empty segment summary anyway, to help out the 2041 * roll-forward agent. 2042 */ 2043 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2044 if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP) 2045 return 0; 2046 } 2047 2048 /* Note if partial segment is being written by the cleaner */ 2049 if (sp->seg_flags & SEGM_CLEAN) 2050 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN); 2051 2052 /* Note if we are writing to reclaim */ 2053 if (sp->seg_flags & SEGM_RECLAIM) { 2054 lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM); 2055 lfs_ss_setreclino(fs, ssp, fs->lfs_reclino); 2056 } 2057 2058 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2059 2060 /* Update the segment usage information. */ 2061 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2062 2063 /* Loop through all blocks, except the segment summary. */ 2064 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2065 if ((*bpp)->b_vp != devvp) { 2066 sup->su_nbytes += (*bpp)->b_bcount; 2067 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2068 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2069 sp->seg_number, (*bpp)->b_bcount, 2070 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2071 (*bpp)->b_blkno)); 2072 } 2073 } 2074 2075 #ifdef DEBUG 2076 /* Check for zero-length and zero-version FINFO entries. */ 2077 fip = SEGSUM_FINFOBASE(fs, ssp); 2078 for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) { 2079 KDASSERT(lfs_fi_getnblocks(fs, fip) > 0); 2080 KDASSERT(lfs_fi_getversion(fs, fip) > 0); 2081 fip = NEXT_FINFO(fs, fip); 2082 } 2083 #endif /* DEBUG */ 2084 2085 ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2086 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2087 sp->seg_number, 2088 lfs_ss_getninos(fs, ssp) * DINOSIZE(fs), 2089 lfs_ss_getninos(fs, ssp))); 2090 sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs); 2091 /* sup->su_nbytes += lfs_sb_getsumsize(fs); */ 2092 if (lfs_sb_getversion(fs) == 1) 2093 sup->su_olastmod = time_second; 2094 else 2095 sup->su_lastmod = time_second; 2096 sup->su_ninos += ninos; 2097 ++sup->su_nsums; 2098 lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs))); 2099 2100 do_again = !(bp->b_flags & B_GATHERED); 2101 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2102 2103 /* 2104 * Mark blocks B_BUSY, to prevent then from being changed between 2105 * the checksum computation and the actual write. 2106 * 2107 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2108 * there are any, replace them with copies that have UNASSIGNED 2109 * instead. 2110 */ 2111 mutex_enter(&bufcache_lock); 2112 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2113 ++bpp; 2114 bp = *bpp; 2115 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2116 bp->b_cflags |= BC_BUSY; 2117 continue; 2118 } 2119 2120 while (bp->b_cflags & BC_BUSY) { 2121 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2122 " data summary corruption for ino %d, lbn %" 2123 PRId64 "\n", 2124 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2125 bp->b_cflags |= BC_WANTED; 2126 cv_wait(&bp->b_busy, &bufcache_lock); 2127 } 2128 bp->b_cflags |= BC_BUSY; 2129 mutex_exit(&bufcache_lock); 2130 unbusybp = NULL; 2131 2132 /* 2133 * Check and replace indirect block UNWRITTEN bogosity. 2134 * XXX See comment in lfs_writefile. 2135 */ 2136 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2137 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) != 2138 VTOI(bp->b_vp)->i_lfs_effnblks) { 2139 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n", 2140 VTOI(bp->b_vp)->i_number, 2141 (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks, 2142 lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din))); 2143 /* Make a copy we'll make changes to */ 2144 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2145 bp->b_bcount, LFS_NB_IBLOCK); 2146 newbp->b_blkno = bp->b_blkno; 2147 memcpy(newbp->b_data, bp->b_data, 2148 newbp->b_bcount); 2149 2150 changed = 0; 2151 iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs); 2152 for (ibindex = 0; ibindex < iblimit; ibindex++) { 2153 if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) { 2154 ++changed; 2155 lfs_iblock_set(fs, newbp->b_data, 2156 ibindex, 0); 2157 } 2158 } 2159 /* 2160 * Get rid of the old buffer. Don't mark it clean, 2161 * though, if it still has dirty data on it. 2162 */ 2163 if (changed) { 2164 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2165 " bp = %p newbp = %p\n", changed, bp, 2166 newbp)); 2167 *bpp = newbp; 2168 bp->b_flags &= ~B_GATHERED; 2169 bp->b_error = 0; 2170 if (bp->b_iodone != NULL) { 2171 DLOG((DLOG_SEG, "lfs_writeseg: " 2172 "indir bp should not be B_CALL\n")); 2173 biodone(bp); 2174 bp = NULL; 2175 } else { 2176 /* Still on free list, leave it there */ 2177 unbusybp = bp; 2178 /* 2179 * We have to re-decrement lfs_avail 2180 * since this block is going to come 2181 * back around to us in the next 2182 * segment. 2183 */ 2184 lfs_sb_subavail(fs, 2185 lfs_btofsb(fs, bp->b_bcount)); 2186 } 2187 } else { 2188 lfs_freebuf(fs, newbp); 2189 } 2190 } 2191 mutex_enter(&bufcache_lock); 2192 if (unbusybp != NULL) { 2193 unbusybp->b_cflags &= ~BC_BUSY; 2194 if (unbusybp->b_cflags & BC_WANTED) 2195 cv_broadcast(&bp->b_busy); 2196 } 2197 } 2198 mutex_exit(&bufcache_lock); 2199 2200 /* 2201 * Compute checksum across data and then across summary; the first 2202 * block (the summary block) is skipped. Set the create time here 2203 * so that it's guaranteed to be later than the inode mod times. 2204 */ 2205 sum = 0; 2206 if (lfs_sb_getversion(fs) == 1) 2207 el_size = sizeof(u_long); 2208 else 2209 el_size = sizeof(u_int32_t); 2210 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2211 ++bpp; 2212 /* Loop through gop_write cluster blocks */ 2213 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2214 byteoffset += lfs_sb_getbsize(fs)) { 2215 #ifdef LFS_USE_B_INVAL 2216 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2217 (*bpp)->b_iodone != NULL) { 2218 if (copyin((void *)(*bpp)->b_saveaddr + 2219 byteoffset, dp, el_size)) { 2220 panic("lfs_writeseg: copyin failed [1]:" 2221 " ino %" PRIu64 " blk %" PRId64, 2222 VTOI((*bpp)->b_vp)->i_number, 2223 (*bpp)->b_lblkno); 2224 } 2225 } else 2226 #endif /* LFS_USE_B_INVAL */ 2227 { 2228 sum = lfs_cksum_part((char *) 2229 (*bpp)->b_data + byteoffset, el_size, sum); 2230 } 2231 } 2232 } 2233 if (lfs_sb_getversion(fs) == 1) 2234 lfs_ss_setocreate(fs, ssp, time_second); 2235 else { 2236 lfs_ss_setcreate(fs, ssp, time_second); 2237 lfs_sb_addserial(fs, 1); 2238 lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs)); 2239 lfs_ss_setident(fs, ssp, lfs_sb_getident(fs)); 2240 } 2241 lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum)); 2242 sumstart = lfs_ss_getsumstart(fs); 2243 lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart, 2244 lfs_sb_getsumsize(fs) - sumstart)); 2245 2246 mutex_enter(&lfs_lock); 2247 lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2248 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2249 lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) + 2250 lfs_btofsb(fs, lfs_sb_getsumsize(fs)))); 2251 mutex_exit(&lfs_lock); 2252 2253 /* 2254 * When we simply write the blocks we lose a rotation for every block 2255 * written. To avoid this problem, we cluster the buffers into a 2256 * chunk and write the chunk. MAXPHYS is the largest size I/O 2257 * devices can handle, use that for the size of the chunks. 2258 * 2259 * Blocks that are already clusters (from GOP_WRITE), however, we 2260 * don't bother to copy into other clusters. 2261 */ 2262 2263 #define CHUNKSIZE MAXPHYS 2264 2265 if (devvp == NULL) 2266 panic("devvp is NULL"); 2267 for (bpp = sp->bpp, i = nblocks; i;) { 2268 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2269 cl = cbp->b_private; 2270 2271 cbp->b_flags |= B_ASYNC; 2272 cbp->b_cflags |= BC_BUSY; 2273 cbp->b_bcount = 0; 2274 2275 KASSERTMSG((bpp - sp->bpp <= 2276 (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs)) 2277 / sizeof(int32_t)), 2278 "lfs_writeseg: real bpp overwrite"); 2279 KASSERTMSG((bpp - sp->bpp <= 2280 lfs_segsize(fs) / lfs_sb_getfsize(fs)), 2281 "lfs_writeseg: theoretical bpp overwrite"); 2282 2283 /* 2284 * Construct the cluster. 2285 */ 2286 mutex_enter(&lfs_lock); 2287 ++fs->lfs_iocount; 2288 mutex_exit(&lfs_lock); 2289 while (i && cbp->b_bcount < CHUNKSIZE) { 2290 bp = *bpp; 2291 2292 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2293 break; 2294 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2295 break; 2296 2297 /* Clusters from GOP_WRITE are expedited */ 2298 if (bp->b_bcount > lfs_sb_getbsize(fs)) { 2299 if (cbp->b_bcount > 0) 2300 /* Put in its own buffer */ 2301 break; 2302 else { 2303 cbp->b_data = bp->b_data; 2304 } 2305 } else if (cbp->b_bcount == 0) { 2306 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2307 LFS_NB_CLUSTER); 2308 cl->flags |= LFS_CL_MALLOC; 2309 } 2310 KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2311 btodb(bp->b_bcount - 1))) == 2312 sp->seg_number), 2313 "segment overwrite: blk size %d daddr %" PRIx64 2314 " not in seg %d\n", 2315 bp->b_bcount, bp->b_blkno, 2316 sp->seg_number); 2317 2318 #ifdef LFS_USE_B_INVAL 2319 /* 2320 * Fake buffers from the cleaner are marked as B_INVAL. 2321 * We need to copy the data from user space rather than 2322 * from the buffer indicated. 2323 * XXX == what do I do on an error? 2324 */ 2325 if ((bp->b_cflags & BC_INVAL) != 0 && 2326 bp->b_iodone != NULL) { 2327 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2328 panic("lfs_writeseg: " 2329 "copyin failed [2]"); 2330 } else 2331 #endif /* LFS_USE_B_INVAL */ 2332 if (cl->flags & LFS_CL_MALLOC) { 2333 /* copy data into our cluster. */ 2334 memcpy(p, bp->b_data, bp->b_bcount); 2335 p += bp->b_bcount; 2336 } 2337 2338 cbp->b_bcount += bp->b_bcount; 2339 cl->bufsize += bp->b_bcount; 2340 2341 bp->b_flags &= ~B_READ; 2342 bp->b_error = 0; 2343 cl->bpp[cl->bufcount++] = bp; 2344 2345 vp = bp->b_vp; 2346 mutex_enter(&bufcache_lock); 2347 mutex_enter(vp->v_interlock); 2348 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2349 reassignbuf(bp, vp); 2350 vp->v_numoutput++; 2351 mutex_exit(vp->v_interlock); 2352 mutex_exit(&bufcache_lock); 2353 2354 bpp++; 2355 i--; 2356 } 2357 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2358 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2359 else 2360 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2361 mutex_enter(devvp->v_interlock); 2362 devvp->v_numoutput++; 2363 mutex_exit(devvp->v_interlock); 2364 VOP_STRATEGY(devvp, cbp); 2365 curlwp->l_ru.ru_oublock++; 2366 } 2367 2368 if (lfs_dostats) { 2369 ++lfs_stats.psegwrites; 2370 lfs_stats.blocktot += nblocks - 1; 2371 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2372 ++lfs_stats.psyncwrites; 2373 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2374 ++lfs_stats.pcleanwrites; 2375 lfs_stats.cleanblocks += nblocks - 1; 2376 } 2377 } 2378 2379 return (lfs_initseg(fs) || do_again); 2380 } 2381 2382 void 2383 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2384 { 2385 struct buf *bp; 2386 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2387 int s; 2388 2389 ASSERT_MAYBE_SEGLOCK(fs); 2390 if (fs->lfs_is64) { 2391 KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC); 2392 } else { 2393 KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC); 2394 } 2395 /* 2396 * If we can write one superblock while another is in 2397 * progress, we risk not having a complete checkpoint if we crash. 2398 * So, block here if a superblock write is in progress. 2399 */ 2400 mutex_enter(&lfs_lock); 2401 s = splbio(); 2402 while (fs->lfs_sbactive) { 2403 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2404 &lfs_lock); 2405 } 2406 fs->lfs_sbactive = daddr; 2407 splx(s); 2408 mutex_exit(&lfs_lock); 2409 2410 /* Set timestamp of this version of the superblock */ 2411 if (lfs_sb_getversion(fs) == 1) 2412 lfs_sb_setotstamp(fs, time_second); 2413 lfs_sb_settstamp(fs, time_second); 2414 2415 /* The next chunk of code relies on this assumption */ 2416 CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64)); 2417 2418 /* Checksum the superblock and copy it into a buffer. */ 2419 lfs_sb_setcksum(fs, lfs_sb_cksum(fs)); 2420 bp = lfs_newbuf(fs, devvp, 2421 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2422 memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs)); 2423 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2424 LFS_SBPAD - sizeof(struct dlfs)); 2425 2426 bp->b_cflags |= BC_BUSY; 2427 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2428 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2429 bp->b_error = 0; 2430 bp->b_iodone = lfs_supercallback; 2431 2432 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2433 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2434 else 2435 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2436 curlwp->l_ru.ru_oublock++; 2437 2438 mutex_enter(devvp->v_interlock); 2439 devvp->v_numoutput++; 2440 mutex_exit(devvp->v_interlock); 2441 2442 mutex_enter(&lfs_lock); 2443 ++fs->lfs_iocount; 2444 mutex_exit(&lfs_lock); 2445 VOP_STRATEGY(devvp, bp); 2446 } 2447 2448 /* 2449 * Logical block number match routines used when traversing the dirty block 2450 * chain. 2451 */ 2452 int 2453 lfs_match_fake(struct lfs *fs, struct buf *bp) 2454 { 2455 2456 ASSERT_SEGLOCK(fs); 2457 return LFS_IS_MALLOC_BUF(bp); 2458 } 2459 2460 #if 0 2461 int 2462 lfs_match_real(struct lfs *fs, struct buf *bp) 2463 { 2464 2465 ASSERT_SEGLOCK(fs); 2466 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2467 } 2468 #endif 2469 2470 int 2471 lfs_match_data(struct lfs *fs, struct buf *bp) 2472 { 2473 2474 ASSERT_SEGLOCK(fs); 2475 return (bp->b_lblkno >= 0); 2476 } 2477 2478 int 2479 lfs_match_indir(struct lfs *fs, struct buf *bp) 2480 { 2481 daddr_t lbn; 2482 2483 ASSERT_SEGLOCK(fs); 2484 lbn = bp->b_lblkno; 2485 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2486 } 2487 2488 int 2489 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2490 { 2491 daddr_t lbn; 2492 2493 ASSERT_SEGLOCK(fs); 2494 lbn = bp->b_lblkno; 2495 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2496 } 2497 2498 int 2499 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2500 { 2501 daddr_t lbn; 2502 2503 ASSERT_SEGLOCK(fs); 2504 lbn = bp->b_lblkno; 2505 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2506 } 2507 2508 static void 2509 lfs_free_aiodone(struct buf *bp) 2510 { 2511 struct lfs *fs; 2512 2513 KERNEL_LOCK(1, curlwp); 2514 fs = bp->b_private; 2515 ASSERT_NO_SEGLOCK(fs); 2516 lfs_freebuf(fs, bp); 2517 KERNEL_UNLOCK_LAST(curlwp); 2518 } 2519 2520 static void 2521 lfs_super_aiodone(struct buf *bp) 2522 { 2523 struct lfs *fs; 2524 2525 KERNEL_LOCK(1, curlwp); 2526 fs = bp->b_private; 2527 ASSERT_NO_SEGLOCK(fs); 2528 mutex_enter(&lfs_lock); 2529 fs->lfs_sbactive = 0; 2530 if (--fs->lfs_iocount <= 1) 2531 wakeup(&fs->lfs_iocount); 2532 wakeup(&fs->lfs_sbactive); 2533 mutex_exit(&lfs_lock); 2534 lfs_freebuf(fs, bp); 2535 KERNEL_UNLOCK_LAST(curlwp); 2536 } 2537 2538 static void 2539 lfs_cluster_aiodone(struct buf *bp) 2540 { 2541 struct lfs_cluster *cl; 2542 struct lfs *fs; 2543 struct buf *tbp, *fbp; 2544 struct vnode *vp, *devvp, *ovp; 2545 struct inode *ip; 2546 int error; 2547 2548 KERNEL_LOCK(1, curlwp); 2549 2550 error = bp->b_error; 2551 cl = bp->b_private; 2552 fs = cl->fs; 2553 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2554 ASSERT_NO_SEGLOCK(fs); 2555 2556 /* Put the pages back, and release the buffer */ 2557 while (cl->bufcount--) { 2558 tbp = cl->bpp[cl->bufcount]; 2559 KASSERT(tbp->b_cflags & BC_BUSY); 2560 if (error) { 2561 tbp->b_error = error; 2562 } 2563 2564 /* 2565 * We're done with tbp. If it has not been re-dirtied since 2566 * the cluster was written, free it. Otherwise, keep it on 2567 * the locked list to be written again. 2568 */ 2569 vp = tbp->b_vp; 2570 2571 tbp->b_flags &= ~B_GATHERED; 2572 2573 #ifdef DEBUG 2574 if ((tbp)->b_vp == (fs)->lfs_ivnode) 2575 LFS_ENTER_LOG("clear", __FILE__, __LINE__, 2576 tbp->b_lblkno, tbp->b_flags, curproc->p_pid); 2577 #endif 2578 2579 mutex_enter(&bufcache_lock); 2580 if (tbp->b_iodone == NULL) { 2581 KASSERT(tbp->b_flags & B_LOCKED); 2582 bremfree(tbp); 2583 if (vp) { 2584 mutex_enter(vp->v_interlock); 2585 reassignbuf(tbp, vp); 2586 mutex_exit(vp->v_interlock); 2587 } 2588 tbp->b_flags |= B_ASYNC; /* for biodone */ 2589 } 2590 2591 if (((tbp->b_flags | tbp->b_oflags) & 2592 (B_LOCKED | BO_DELWRI)) == B_LOCKED) 2593 LFS_UNLOCK_BUF(tbp); 2594 2595 if (tbp->b_oflags & BO_DONE) { 2596 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2597 cl->bufcount, (long)tbp->b_flags)); 2598 } 2599 2600 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2601 /* 2602 * A buffer from the page daemon. 2603 * We use the same iodone as it does, 2604 * so we must manually disassociate its 2605 * buffers from the vp. 2606 */ 2607 if ((ovp = tbp->b_vp) != NULL) { 2608 /* This is just silly */ 2609 mutex_enter(ovp->v_interlock); 2610 brelvp(tbp); 2611 mutex_exit(ovp->v_interlock); 2612 tbp->b_vp = vp; 2613 tbp->b_objlock = vp->v_interlock; 2614 } 2615 /* Put it back the way it was */ 2616 tbp->b_flags |= B_ASYNC; 2617 /* Master buffers have BC_AGE */ 2618 if (tbp->b_private == tbp) 2619 tbp->b_cflags |= BC_AGE; 2620 } 2621 mutex_exit(&bufcache_lock); 2622 2623 biodone(tbp); 2624 2625 /* 2626 * If this is the last block for this vnode, but 2627 * there are other blocks on its dirty list, 2628 * set IN_MODIFIED/IN_CLEANING depending on what 2629 * sort of block. Only do this for our mount point, 2630 * not for, e.g., inode blocks that are attached to 2631 * the devvp. 2632 * XXX KS - Shouldn't we set *both* if both types 2633 * of blocks are present (traverse the dirty list?) 2634 */ 2635 mutex_enter(vp->v_interlock); 2636 mutex_enter(&lfs_lock); 2637 if (vp != devvp && vp->v_numoutput == 0 && 2638 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2639 ip = VTOI(vp); 2640 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2641 ip->i_number)); 2642 if (LFS_IS_MALLOC_BUF(fbp)) 2643 LFS_SET_UINO(ip, IN_CLEANING); 2644 else 2645 LFS_SET_UINO(ip, IN_MODIFIED); 2646 } 2647 cv_broadcast(&vp->v_cv); 2648 mutex_exit(&lfs_lock); 2649 mutex_exit(vp->v_interlock); 2650 } 2651 2652 /* Fix up the cluster buffer, and release it */ 2653 if (cl->flags & LFS_CL_MALLOC) 2654 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2655 putiobuf(bp); 2656 2657 /* Note i/o done */ 2658 if (cl->flags & LFS_CL_SYNC) { 2659 if (--cl->seg->seg_iocount == 0) 2660 wakeup(&cl->seg->seg_iocount); 2661 } 2662 mutex_enter(&lfs_lock); 2663 KASSERTMSG((fs->lfs_iocount != 0), 2664 "lfs_cluster_aiodone: zero iocount"); 2665 if (--fs->lfs_iocount <= 1) 2666 wakeup(&fs->lfs_iocount); 2667 mutex_exit(&lfs_lock); 2668 2669 KERNEL_UNLOCK_LAST(curlwp); 2670 2671 pool_put(&fs->lfs_bpppool, cl->bpp); 2672 cl->bpp = NULL; 2673 pool_put(&fs->lfs_clpool, cl); 2674 } 2675 2676 static void 2677 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2678 { 2679 /* reset b_iodone for when this is a single-buf i/o. */ 2680 bp->b_iodone = aiodone; 2681 2682 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2683 } 2684 2685 static void 2686 lfs_cluster_callback(struct buf *bp) 2687 { 2688 2689 lfs_generic_callback(bp, lfs_cluster_aiodone); 2690 } 2691 2692 void 2693 lfs_supercallback(struct buf *bp) 2694 { 2695 2696 lfs_generic_callback(bp, lfs_super_aiodone); 2697 } 2698 2699 /* 2700 * The only buffers that are going to hit these functions are the 2701 * segment write blocks, or the segment summaries, or the superblocks. 2702 * 2703 * All of the above are created by lfs_newbuf, and so do not need to be 2704 * released via brelse. 2705 */ 2706 void 2707 lfs_callback(struct buf *bp) 2708 { 2709 2710 lfs_generic_callback(bp, lfs_free_aiodone); 2711 } 2712 2713 /* 2714 * Shellsort (diminishing increment sort) from Data Structures and 2715 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2716 * see also Knuth Vol. 3, page 84. The increments are selected from 2717 * formula (8), page 95. Roughly O(N^3/2). 2718 */ 2719 /* 2720 * This is our own private copy of shellsort because we want to sort 2721 * two parallel arrays (the array of buffer pointers and the array of 2722 * logical block numbers) simultaneously. Note that we cast the array 2723 * of logical block numbers to a unsigned in this routine so that the 2724 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2725 */ 2726 2727 static void 2728 lfs_shellsort(struct lfs *fs, 2729 struct buf **bp_array, union lfs_blocks *lb_array, 2730 int nmemb, int size) 2731 { 2732 static int __rsshell_increments[] = { 4, 1, 0 }; 2733 int incr, *incrp, t1, t2; 2734 struct buf *bp_temp; 2735 2736 #ifdef DEBUG 2737 incr = 0; 2738 for (t1 = 0; t1 < nmemb; t1++) { 2739 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2740 if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) { 2741 /* dump before panic */ 2742 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2743 nmemb, size); 2744 incr = 0; 2745 for (t1 = 0; t1 < nmemb; t1++) { 2746 const struct buf *bp = bp_array[t1]; 2747 2748 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2749 PRIu64 "\n", t1, 2750 (uint64_t)bp->b_bcount, 2751 (uint64_t)bp->b_lblkno); 2752 printf("lbns:"); 2753 for (t2 = 0; t2 * size < bp->b_bcount; 2754 t2++) { 2755 printf(" %jd", 2756 (intmax_t)lfs_blocks_get(fs, lb_array, incr++)); 2757 } 2758 printf("\n"); 2759 } 2760 panic("lfs_shellsort: inconsistent input"); 2761 } 2762 } 2763 } 2764 #endif 2765 2766 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2767 for (t1 = incr; t1 < nmemb; ++t1) 2768 for (t2 = t1 - incr; t2 >= 0;) 2769 if ((u_int64_t)bp_array[t2]->b_lblkno > 2770 (u_int64_t)bp_array[t2 + incr]->b_lblkno) { 2771 bp_temp = bp_array[t2]; 2772 bp_array[t2] = bp_array[t2 + incr]; 2773 bp_array[t2 + incr] = bp_temp; 2774 t2 -= incr; 2775 } else 2776 break; 2777 2778 /* Reform the list of logical blocks */ 2779 incr = 0; 2780 for (t1 = 0; t1 < nmemb; t1++) { 2781 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2782 lfs_blocks_set(fs, lb_array, incr++, 2783 bp_array[t1]->b_lblkno + t2); 2784 } 2785 } 2786 } 2787 2788 /* 2789 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2790 * point at uninitialized space. 2791 */ 2792 void 2793 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2794 { 2795 struct segment *sp = fs->lfs_sp; 2796 SEGSUM *ssp; 2797 2798 KASSERT(vers > 0); 2799 2800 if (sp->seg_bytes_left < lfs_sb_getbsize(fs) || 2801 sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs)) 2802 (void) lfs_writeseg(fs, fs->lfs_sp); 2803 2804 sp->sum_bytes_left -= FINFOSIZE(fs); 2805 ssp = (SEGSUM *)sp->segsum; 2806 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1); 2807 lfs_fi_setnblocks(fs, sp->fip, 0); 2808 lfs_fi_setino(fs, sp->fip, ino); 2809 lfs_fi_setversion(fs, sp->fip, vers); 2810 } 2811 2812 /* 2813 * Release the FINFO entry, either clearing out an unused entry or 2814 * advancing us to the next available entry. 2815 */ 2816 void 2817 lfs_release_finfo(struct lfs *fs) 2818 { 2819 struct segment *sp = fs->lfs_sp; 2820 SEGSUM *ssp; 2821 2822 if (lfs_fi_getnblocks(fs, sp->fip) != 0) { 2823 sp->fip = NEXT_FINFO(fs, sp->fip); 2824 lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip); 2825 } else { 2826 /* XXX shouldn't this update sp->fip? */ 2827 sp->sum_bytes_left += FINFOSIZE(fs); 2828 ssp = (SEGSUM *)sp->segsum; 2829 lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1); 2830 } 2831 } 2832
Indexes created Mon Sep 22 05:09:51 GMT 2025