1 /* $NetBSD: lfs_segment.c,v 1.241 2015/06/07 13:39:48 hannken 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@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 1.241 2015/06/07 13:39:48 hannken Exp $"); 64 65 #define _VFS_VNODE_PRIVATE /* XXX: check for VI_MARKER, this has to go */ 66 67 #ifdef DEBUG 68 # define vndebug(vp, str) do { \ 69 if (VTOI(vp)->i_flag & IN_CLEANING) \ 70 DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \ 71 VTOI(vp)->i_number, (str), op)); \ 72 } while(0) 73 #else 74 # define vndebug(vp, str) 75 #endif 76 #define ivndebug(vp, str) \ 77 DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str))) 78 79 #if defined(_KERNEL_OPT) 80 #include "opt_ddb.h" 81 #endif 82 83 #include <sys/param.h> 84 #include <sys/systm.h> 85 #include <sys/namei.h> 86 #include <sys/kernel.h> 87 #include <sys/resourcevar.h> 88 #include <sys/file.h> 89 #include <sys/stat.h> 90 #include <sys/buf.h> 91 #include <sys/proc.h> 92 #include <sys/vnode.h> 93 #include <sys/mount.h> 94 #include <sys/kauth.h> 95 #include <sys/syslog.h> 96 97 #include <miscfs/specfs/specdev.h> 98 #include <miscfs/fifofs/fifo.h> 99 100 #include <ufs/lfs/ulfs_inode.h> 101 #include <ufs/lfs/ulfsmount.h> 102 #include <ufs/lfs/ulfs_extern.h> 103 104 #include <ufs/lfs/lfs.h> 105 #include <ufs/lfs/lfs_kernel.h> 106 #include <ufs/lfs/lfs_extern.h> 107 108 #include <uvm/uvm.h> 109 #include <uvm/uvm_extern.h> 110 111 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 112 113 static void lfs_generic_callback(struct buf *, void (*)(struct buf *)); 114 static void lfs_free_aiodone(struct buf *); 115 static void lfs_super_aiodone(struct buf *); 116 static void lfs_cluster_aiodone(struct buf *); 117 static void lfs_cluster_callback(struct buf *); 118 119 /* 120 * Determine if it's OK to start a partial in this segment, or if we need 121 * to go on to a new segment. 122 */ 123 #define LFS_PARTIAL_FITS(fs) \ 124 ((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \ 125 (fs)->lfs_frag) 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 fs->lfs_nclean < LFS_MAX_ACTIVE))) 136 137 int lfs_match_fake(struct lfs *, struct buf *); 138 void lfs_newseg(struct lfs *); 139 /* XXX ondisk32 */ 140 void lfs_shellsort(struct buf **, int32_t *, int, int); 141 void lfs_supercallback(struct buf *); 142 void lfs_updatemeta(struct segment *); 143 void lfs_writesuper(struct lfs *, daddr_t); 144 int lfs_writevnodes(struct lfs *fs, struct mount *mp, 145 struct segment *sp, int dirops); 146 147 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 148 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 149 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 150 int lfs_dirvcount = 0; /* # active dirops */ 151 152 /* Statistics Counters */ 153 int lfs_dostats = 1; 154 struct lfs_stats lfs_stats; 155 156 /* op values to lfs_writevnodes */ 157 #define VN_REG 0 158 #define VN_DIROP 1 159 #define VN_EMPTY 2 160 #define VN_CLEAN 3 161 162 /* 163 * XXX KS - Set modification time on the Ifile, so the cleaner can 164 * read the fs mod time off of it. We don't set IN_UPDATE here, 165 * since we don't really need this to be flushed to disk (and in any 166 * case that wouldn't happen to the Ifile until we checkpoint). 167 */ 168 void 169 lfs_imtime(struct lfs *fs) 170 { 171 struct timespec ts; 172 struct inode *ip; 173 174 ASSERT_MAYBE_SEGLOCK(fs); 175 vfs_timestamp(&ts); 176 ip = VTOI(fs->lfs_ivnode); 177 ip->i_ffs1_mtime = ts.tv_sec; 178 ip->i_ffs1_mtimensec = ts.tv_nsec; 179 } 180 181 /* 182 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 183 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 184 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 185 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 186 */ 187 188 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 189 190 int 191 lfs_vflush(struct vnode *vp) 192 { 193 struct inode *ip; 194 struct lfs *fs; 195 struct segment *sp; 196 struct buf *bp, *nbp, *tbp, *tnbp; 197 int error; 198 int flushed; 199 int relock; 200 201 ip = VTOI(vp); 202 fs = VFSTOULFS(vp->v_mount)->um_lfs; 203 relock = 0; 204 205 top: 206 KASSERT(mutex_owned(vp->v_interlock) == false); 207 KASSERT(mutex_owned(&lfs_lock) == false); 208 KASSERT(mutex_owned(&bufcache_lock) == false); 209 ASSERT_NO_SEGLOCK(fs); 210 if (ip->i_flag & IN_CLEANING) { 211 ivndebug(vp,"vflush/in_cleaning"); 212 mutex_enter(&lfs_lock); 213 LFS_CLR_UINO(ip, IN_CLEANING); 214 LFS_SET_UINO(ip, IN_MODIFIED); 215 mutex_exit(&lfs_lock); 216 217 /* 218 * Toss any cleaning buffers that have real counterparts 219 * to avoid losing new data. 220 */ 221 mutex_enter(vp->v_interlock); 222 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 223 nbp = LIST_NEXT(bp, b_vnbufs); 224 if (!LFS_IS_MALLOC_BUF(bp)) 225 continue; 226 /* 227 * Look for pages matching the range covered 228 * by cleaning blocks. It's okay if more dirty 229 * pages appear, so long as none disappear out 230 * from under us. 231 */ 232 if (bp->b_lblkno > 0 && vp->v_type == VREG && 233 vp != fs->lfs_ivnode) { 234 struct vm_page *pg; 235 voff_t off; 236 237 for (off = lfs_lblktosize(fs, bp->b_lblkno); 238 off < lfs_lblktosize(fs, bp->b_lblkno + 1); 239 off += PAGE_SIZE) { 240 pg = uvm_pagelookup(&vp->v_uobj, off); 241 if (pg == NULL) 242 continue; 243 if ((pg->flags & PG_CLEAN) == 0 || 244 pmap_is_modified(pg)) { 245 fs->lfs_avail += lfs_btofsb(fs, 246 bp->b_bcount); 247 wakeup(&fs->lfs_avail); 248 mutex_exit(vp->v_interlock); 249 lfs_freebuf(fs, bp); 250 mutex_enter(vp->v_interlock); 251 bp = NULL; 252 break; 253 } 254 } 255 } 256 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 257 tbp = tnbp) 258 { 259 tnbp = LIST_NEXT(tbp, b_vnbufs); 260 if (tbp->b_vp == bp->b_vp 261 && tbp->b_lblkno == bp->b_lblkno 262 && tbp != bp) 263 { 264 fs->lfs_avail += lfs_btofsb(fs, 265 bp->b_bcount); 266 wakeup(&fs->lfs_avail); 267 mutex_exit(vp->v_interlock); 268 lfs_freebuf(fs, bp); 269 mutex_enter(vp->v_interlock); 270 bp = NULL; 271 break; 272 } 273 } 274 } 275 } else { 276 mutex_enter(vp->v_interlock); 277 } 278 279 /* If the node is being written, wait until that is done */ 280 while (WRITEINPROG(vp)) { 281 ivndebug(vp,"vflush/writeinprog"); 282 cv_wait(&vp->v_cv, vp->v_interlock); 283 } 284 error = vdead_check(vp, VDEAD_NOWAIT); 285 mutex_exit(vp->v_interlock); 286 287 /* Protect against deadlock in vinvalbuf() */ 288 lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0)); 289 if (error != 0) { 290 fs->lfs_reclino = ip->i_number; 291 } 292 293 /* If we're supposed to flush a freed inode, just toss it */ 294 if (ip->i_lfs_iflags & LFSI_DELETED) { 295 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 296 ip->i_number)); 297 /* Drain v_numoutput */ 298 mutex_enter(vp->v_interlock); 299 while (vp->v_numoutput > 0) { 300 cv_wait(&vp->v_cv, vp->v_interlock); 301 } 302 KASSERT(vp->v_numoutput == 0); 303 mutex_exit(vp->v_interlock); 304 305 mutex_enter(&bufcache_lock); 306 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 307 nbp = LIST_NEXT(bp, b_vnbufs); 308 309 KASSERT((bp->b_flags & B_GATHERED) == 0); 310 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 311 fs->lfs_avail += lfs_btofsb(fs, bp->b_bcount); 312 wakeup(&fs->lfs_avail); 313 } 314 /* Copied from lfs_writeseg */ 315 if (bp->b_iodone != NULL) { 316 mutex_exit(&bufcache_lock); 317 biodone(bp); 318 mutex_enter(&bufcache_lock); 319 } else { 320 bremfree(bp); 321 LFS_UNLOCK_BUF(bp); 322 mutex_enter(vp->v_interlock); 323 bp->b_flags &= ~(B_READ | B_GATHERED); 324 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 325 bp->b_error = 0; 326 reassignbuf(bp, vp); 327 mutex_exit(vp->v_interlock); 328 brelse(bp, 0); 329 } 330 } 331 mutex_exit(&bufcache_lock); 332 LFS_CLR_UINO(ip, IN_CLEANING); 333 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 334 ip->i_flag &= ~IN_ALLMOD; 335 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 336 ip->i_number)); 337 lfs_segunlock(fs); 338 339 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 340 341 return 0; 342 } 343 344 fs->lfs_flushvp = vp; 345 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 346 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 347 fs->lfs_flushvp = NULL; 348 KASSERT(fs->lfs_flushvp_fakevref == 0); 349 lfs_segunlock(fs); 350 351 /* Make sure that any pending buffers get written */ 352 mutex_enter(vp->v_interlock); 353 while (vp->v_numoutput > 0) { 354 cv_wait(&vp->v_cv, vp->v_interlock); 355 } 356 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 357 KASSERT(vp->v_numoutput == 0); 358 mutex_exit(vp->v_interlock); 359 360 return error; 361 } 362 sp = fs->lfs_sp; 363 364 flushed = 0; 365 if (VPISEMPTY(vp)) { 366 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 367 ++flushed; 368 } else if ((ip->i_flag & IN_CLEANING) && 369 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 370 ivndebug(vp,"vflush/clean"); 371 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 372 ++flushed; 373 } else if (lfs_dostats) { 374 if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD)) 375 ++lfs_stats.vflush_invoked; 376 ivndebug(vp,"vflush"); 377 } 378 379 #ifdef DIAGNOSTIC 380 if (vp->v_uflag & VU_DIROP) { 381 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 382 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 383 } 384 #endif 385 386 do { 387 #ifdef DEBUG 388 int loopcount = 0; 389 #endif 390 do { 391 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 392 relock = lfs_writefile(fs, sp, vp); 393 if (relock && vp != fs->lfs_ivnode) { 394 /* 395 * Might have to wait for the 396 * cleaner to run; but we're 397 * still not done with this vnode. 398 * XXX we can do better than this. 399 */ 400 KDASSERT(ip->i_number != LFS_IFILE_INUM); 401 lfs_writeinode(fs, sp, ip); 402 mutex_enter(&lfs_lock); 403 LFS_SET_UINO(ip, IN_MODIFIED); 404 mutex_exit(&lfs_lock); 405 lfs_writeseg(fs, sp); 406 lfs_segunlock(fs); 407 lfs_segunlock_relock(fs); 408 goto top; 409 } 410 } 411 /* 412 * If we begin a new segment in the middle of writing 413 * the Ifile, it creates an inconsistent checkpoint, 414 * since the Ifile information for the new segment 415 * is not up-to-date. Take care of this here by 416 * sending the Ifile through again in case there 417 * are newly dirtied blocks. But wait, there's more! 418 * This second Ifile write could *also* cross a segment 419 * boundary, if the first one was large. The second 420 * one is guaranteed to be no more than 8 blocks, 421 * though (two segment blocks and supporting indirects) 422 * so the third write *will not* cross the boundary. 423 */ 424 if (vp == fs->lfs_ivnode) { 425 lfs_writefile(fs, sp, vp); 426 lfs_writefile(fs, sp, vp); 427 } 428 #ifdef DEBUG 429 if (++loopcount > 2) 430 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 431 #endif 432 } while (lfs_writeinode(fs, sp, ip)); 433 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 434 435 if (lfs_dostats) { 436 ++lfs_stats.nwrites; 437 if (sp->seg_flags & SEGM_SYNC) 438 ++lfs_stats.nsync_writes; 439 if (sp->seg_flags & SEGM_CKP) 440 ++lfs_stats.ncheckpoints; 441 } 442 /* 443 * If we were called from somewhere that has already held the seglock 444 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 445 * the write to complete because we are still locked. 446 * Since lfs_vflush() must return the vnode with no dirty buffers, 447 * we must explicitly wait, if that is the case. 448 * 449 * We compare the iocount against 1, not 0, because it is 450 * artificially incremented by lfs_seglock(). 451 */ 452 mutex_enter(&lfs_lock); 453 if (fs->lfs_seglock > 1) { 454 while (fs->lfs_iocount > 1) 455 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 456 "lfs_vflush", 0, &lfs_lock); 457 } 458 mutex_exit(&lfs_lock); 459 460 lfs_segunlock(fs); 461 462 /* Wait for these buffers to be recovered by aiodoned */ 463 mutex_enter(vp->v_interlock); 464 while (vp->v_numoutput > 0) { 465 cv_wait(&vp->v_cv, vp->v_interlock); 466 } 467 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 468 KASSERT(vp->v_numoutput == 0); 469 mutex_exit(vp->v_interlock); 470 471 fs->lfs_flushvp = NULL; 472 KASSERT(fs->lfs_flushvp_fakevref == 0); 473 474 return (0); 475 } 476 477 int 478 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 479 { 480 struct inode *ip; 481 struct vnode *vp; 482 int inodes_written = 0; 483 int error = 0; 484 485 ASSERT_SEGLOCK(fs); 486 loop: 487 /* start at last (newest) vnode. */ 488 mutex_enter(&mntvnode_lock); 489 TAILQ_FOREACH_REVERSE(vp, &mp->mnt_vnodelist, vnodelst, v_mntvnodes) { 490 /* 491 * If the vnode that we are about to sync is no longer 492 * associated with this mount point, start over. 493 */ 494 if (vp->v_mount != mp) { 495 DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n")); 496 /* 497 * After this, pages might be busy 498 * due to our own previous putpages. 499 * Start actual segment write here to avoid deadlock. 500 * If we were just writing one segment and we've done 501 * that, break out. 502 */ 503 mutex_exit(&mntvnode_lock); 504 if (lfs_writeseg(fs, sp) && 505 (sp->seg_flags & SEGM_SINGLE) && 506 fs->lfs_curseg != fs->lfs_startseg) { 507 DLOG((DLOG_VNODE, "lfs_writevnodes: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset)); 508 break; 509 } 510 goto loop; 511 } 512 513 mutex_enter(vp->v_interlock); 514 if (vp->v_type == VNON || (vp->v_iflag & VI_MARKER) || 515 vdead_check(vp, VDEAD_NOWAIT) != 0) { 516 mutex_exit(vp->v_interlock); 517 continue; 518 } 519 520 ip = VTOI(vp); 521 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 522 (op != VN_DIROP && op != VN_CLEAN && 523 (vp->v_uflag & VU_DIROP))) { 524 mutex_exit(vp->v_interlock); 525 vndebug(vp,"dirop"); 526 continue; 527 } 528 529 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 530 mutex_exit(vp->v_interlock); 531 vndebug(vp,"empty"); 532 continue; 533 } 534 535 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM 536 && vp != fs->lfs_flushvp 537 && !(ip->i_flag & IN_CLEANING)) { 538 mutex_exit(vp->v_interlock); 539 vndebug(vp,"cleaning"); 540 continue; 541 } 542 543 mutex_exit(&mntvnode_lock); 544 if (vget(vp, LK_NOWAIT, false /* !wait */)) { 545 vndebug(vp,"vget"); 546 mutex_enter(&mntvnode_lock); 547 continue; 548 } 549 550 /* 551 * Write the inode/file if dirty and it's not the IFILE. 552 */ 553 if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) { 554 if (ip->i_number != LFS_IFILE_INUM) { 555 error = lfs_writefile(fs, sp, vp); 556 if (error) { 557 vrele(vp); 558 if (error == EAGAIN) { 559 /* 560 * This error from lfs_putpages 561 * indicates we need to drop 562 * the segment lock and start 563 * over after the cleaner has 564 * had a chance to run. 565 */ 566 lfs_writeinode(fs, sp, ip); 567 lfs_writeseg(fs, sp); 568 if (!VPISEMPTY(vp) && 569 !WRITEINPROG(vp) && 570 !(ip->i_flag & IN_ALLMOD)) { 571 mutex_enter(&lfs_lock); 572 LFS_SET_UINO(ip, IN_MODIFIED); 573 mutex_exit(&lfs_lock); 574 } 575 mutex_enter(&mntvnode_lock); 576 break; 577 } 578 error = 0; /* XXX not quite right */ 579 mutex_enter(&mntvnode_lock); 580 continue; 581 } 582 583 if (!VPISEMPTY(vp)) { 584 if (WRITEINPROG(vp)) { 585 ivndebug(vp,"writevnodes/write2"); 586 } else if (!(ip->i_flag & IN_ALLMOD)) { 587 mutex_enter(&lfs_lock); 588 LFS_SET_UINO(ip, IN_MODIFIED); 589 mutex_exit(&lfs_lock); 590 } 591 } 592 (void) lfs_writeinode(fs, sp, ip); 593 inodes_written++; 594 } 595 } 596 597 vrele(vp); 598 599 mutex_enter(&mntvnode_lock); 600 } 601 mutex_exit(&mntvnode_lock); 602 return error; 603 } 604 605 /* 606 * Do a checkpoint. 607 */ 608 int 609 lfs_segwrite(struct mount *mp, int flags) 610 { 611 struct buf *bp; 612 struct inode *ip; 613 struct lfs *fs; 614 struct segment *sp; 615 struct vnode *vp; 616 SEGUSE *segusep; 617 int do_ckp, did_ckp, error; 618 unsigned n, segleft, maxseg, sn, i, curseg; 619 int writer_set = 0; 620 int dirty; 621 int redo; 622 int um_error; 623 624 fs = VFSTOULFS(mp)->um_lfs; 625 ASSERT_MAYBE_SEGLOCK(fs); 626 627 if (fs->lfs_ronly) 628 return EROFS; 629 630 lfs_imtime(fs); 631 632 /* 633 * Allocate a segment structure and enough space to hold pointers to 634 * the maximum possible number of buffers which can be described in a 635 * single summary block. 636 */ 637 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 638 639 /* We can't do a partial write and checkpoint at the same time. */ 640 if (do_ckp) 641 flags &= ~SEGM_SINGLE; 642 643 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 644 sp = fs->lfs_sp; 645 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 646 do_ckp = 1; 647 648 /* 649 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 650 * in which case we have to flush *all* buffers off of this vnode. 651 * We don't care about other nodes, but write any non-dirop nodes 652 * anyway in anticipation of another getnewvnode(). 653 * 654 * If we're cleaning we only write cleaning and ifile blocks, and 655 * no dirops, since otherwise we'd risk corruption in a crash. 656 */ 657 if (sp->seg_flags & SEGM_CLEAN) 658 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 659 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 660 do { 661 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 662 if ((sp->seg_flags & SEGM_SINGLE) && 663 fs->lfs_curseg != fs->lfs_startseg) { 664 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset)); 665 break; 666 } 667 668 if (do_ckp || fs->lfs_dirops == 0) { 669 if (!writer_set) { 670 lfs_writer_enter(fs, "lfs writer"); 671 writer_set = 1; 672 } 673 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 674 if (um_error == 0) 675 um_error = error; 676 /* In case writevnodes errored out */ 677 lfs_flush_dirops(fs); 678 ((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT); 679 lfs_finalize_fs_seguse(fs); 680 } 681 if (do_ckp && um_error) { 682 lfs_segunlock_relock(fs); 683 sp = fs->lfs_sp; 684 } 685 } while (do_ckp && um_error != 0); 686 } 687 688 /* 689 * If we are doing a checkpoint, mark everything since the 690 * last checkpoint as no longer ACTIVE. 691 */ 692 if (do_ckp || fs->lfs_doifile) { 693 segleft = fs->lfs_nseg; 694 curseg = 0; 695 for (n = 0; n < fs->lfs_segtabsz; n++) { 696 dirty = 0; 697 if (bread(fs->lfs_ivnode, fs->lfs_cleansz + n, 698 fs->lfs_bsize, B_MODIFY, &bp)) 699 panic("lfs_segwrite: ifile read"); 700 segusep = (SEGUSE *)bp->b_data; 701 maxseg = min(segleft, fs->lfs_sepb); 702 for (i = 0; i < maxseg; i++) { 703 sn = curseg + i; 704 if (sn != lfs_dtosn(fs, fs->lfs_curseg) && 705 segusep->su_flags & SEGUSE_ACTIVE) { 706 segusep->su_flags &= ~SEGUSE_ACTIVE; 707 --fs->lfs_nactive; 708 ++dirty; 709 } 710 fs->lfs_suflags[fs->lfs_activesb][sn] = 711 segusep->su_flags; 712 if (fs->lfs_version > 1) 713 ++segusep; 714 else 715 segusep = (SEGUSE *) 716 ((SEGUSE_V1 *)segusep + 1); 717 } 718 719 if (dirty) 720 error = LFS_BWRITE_LOG(bp); /* Ifile */ 721 else 722 brelse(bp, 0); 723 segleft -= fs->lfs_sepb; 724 curseg += fs->lfs_sepb; 725 } 726 } 727 728 KASSERT(LFS_SEGLOCK_HELD(fs)); 729 730 did_ckp = 0; 731 if (do_ckp || fs->lfs_doifile) { 732 vp = fs->lfs_ivnode; 733 #ifdef DEBUG 734 int loopcount = 0; 735 #endif 736 do { 737 #ifdef DEBUG 738 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 739 #endif 740 mutex_enter(&lfs_lock); 741 fs->lfs_flags &= ~LFS_IFDIRTY; 742 mutex_exit(&lfs_lock); 743 744 ip = VTOI(vp); 745 746 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 747 /* 748 * Ifile has no pages, so we don't need 749 * to check error return here. 750 */ 751 lfs_writefile(fs, sp, vp); 752 /* 753 * Ensure the Ifile takes the current segment 754 * into account. See comment in lfs_vflush. 755 */ 756 lfs_writefile(fs, sp, vp); 757 lfs_writefile(fs, sp, vp); 758 } 759 760 if (ip->i_flag & IN_ALLMOD) 761 ++did_ckp; 762 #if 0 763 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 764 #else 765 redo = lfs_writeinode(fs, sp, ip); 766 #endif 767 redo += lfs_writeseg(fs, sp); 768 mutex_enter(&lfs_lock); 769 redo += (fs->lfs_flags & LFS_IFDIRTY); 770 mutex_exit(&lfs_lock); 771 #ifdef DEBUG 772 if (++loopcount > 2) 773 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 774 loopcount); 775 #endif 776 } while (redo && do_ckp); 777 778 /* 779 * Unless we are unmounting, the Ifile may continue to have 780 * dirty blocks even after a checkpoint, due to changes to 781 * inodes' atime. If we're checkpointing, it's "impossible" 782 * for other parts of the Ifile to be dirty after the loop 783 * above, since we hold the segment lock. 784 */ 785 mutex_enter(vp->v_interlock); 786 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 787 LFS_CLR_UINO(ip, IN_ALLMOD); 788 } 789 #ifdef DIAGNOSTIC 790 else if (do_ckp) { 791 int do_panic = 0; 792 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 793 if (bp->b_lblkno < fs->lfs_cleansz + 794 fs->lfs_segtabsz && 795 !(bp->b_flags & B_GATHERED)) { 796 printf("ifile lbn %ld still dirty (flags %lx)\n", 797 (long)bp->b_lblkno, 798 (long)bp->b_flags); 799 ++do_panic; 800 } 801 } 802 if (do_panic) 803 panic("dirty blocks"); 804 } 805 #endif 806 mutex_exit(vp->v_interlock); 807 } else { 808 (void) lfs_writeseg(fs, sp); 809 } 810 811 /* Note Ifile no longer needs to be written */ 812 fs->lfs_doifile = 0; 813 if (writer_set) 814 lfs_writer_leave(fs); 815 816 /* 817 * If we didn't write the Ifile, we didn't really do anything. 818 * That means that (1) there is a checkpoint on disk and (2) 819 * nothing has changed since it was written. 820 * 821 * Take the flags off of the segment so that lfs_segunlock 822 * doesn't have to write the superblock either. 823 */ 824 if (do_ckp && !did_ckp) { 825 sp->seg_flags &= ~SEGM_CKP; 826 } 827 828 if (lfs_dostats) { 829 ++lfs_stats.nwrites; 830 if (sp->seg_flags & SEGM_SYNC) 831 ++lfs_stats.nsync_writes; 832 if (sp->seg_flags & SEGM_CKP) 833 ++lfs_stats.ncheckpoints; 834 } 835 lfs_segunlock(fs); 836 return (0); 837 } 838 839 /* 840 * Write the dirty blocks associated with a vnode. 841 */ 842 int 843 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 844 { 845 struct inode *ip; 846 int i, frag; 847 int error; 848 849 ASSERT_SEGLOCK(fs); 850 error = 0; 851 ip = VTOI(vp); 852 853 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 854 855 if (vp->v_uflag & VU_DIROP) 856 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 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] < fs->lfs_bsize) 910 ++frag; 911 } 912 #ifdef DIAGNOSTIC 913 if (frag > 1) 914 panic("lfs_writefile: more than one fragment!"); 915 #endif 916 if (IS_FLUSHING(fs, vp) || 917 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 918 lfs_gather(fs, sp, vp, lfs_match_indir); 919 lfs_gather(fs, sp, vp, lfs_match_dindir); 920 lfs_gather(fs, sp, vp, lfs_match_tindir); 921 } 922 lfs_release_finfo(fs); 923 924 return error; 925 } 926 927 /* 928 * Update segment accounting to reflect this inode's change of address. 929 */ 930 static int 931 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 932 { 933 struct buf *bp; 934 daddr_t daddr; 935 IFILE *ifp; 936 SEGUSE *sup; 937 ino_t ino; 938 int redo_ifile; 939 u_int32_t sn; 940 941 redo_ifile = 0; 942 943 /* 944 * If updating the ifile, update the super-block. Update the disk 945 * address and access times for this inode in the ifile. 946 */ 947 ino = ip->i_number; 948 if (ino == LFS_IFILE_INUM) { 949 daddr = fs->lfs_idaddr; 950 fs->lfs_idaddr = LFS_DBTOFSB(fs, ndaddr); 951 } else { 952 LFS_IENTRY(ifp, fs, ino, bp); 953 daddr = ifp->if_daddr; 954 ifp->if_daddr = LFS_DBTOFSB(fs, ndaddr); 955 (void)LFS_BWRITE_LOG(bp); /* Ifile */ 956 } 957 958 /* 959 * If this is the Ifile and lfs_offset is set to the first block 960 * in the segment, dirty the new segment's accounting block 961 * (XXX should already be dirty?) and tell the caller to do it again. 962 */ 963 if (ip->i_number == LFS_IFILE_INUM) { 964 sn = lfs_dtosn(fs, fs->lfs_offset); 965 if (lfs_sntod(fs, sn) + lfs_btofsb(fs, fs->lfs_sumsize) == 966 fs->lfs_offset) { 967 LFS_SEGENTRY(sup, fs, sn, bp); 968 KASSERT(bp->b_oflags & BO_DELWRI); 969 LFS_WRITESEGENTRY(sup, fs, sn, bp); 970 /* fs->lfs_flags |= LFS_IFDIRTY; */ 971 redo_ifile |= 1; 972 } 973 } 974 975 /* 976 * The inode's last address should not be in the current partial 977 * segment, except under exceptional circumstances (lfs_writevnodes 978 * had to start over, and in the meantime more blocks were written 979 * to a vnode). Both inodes will be accounted to this segment 980 * in lfs_writeseg so we need to subtract the earlier version 981 * here anyway. The segment count can temporarily dip below 982 * zero here; keep track of how many duplicates we have in 983 * "dupino" so we don't panic below. 984 */ 985 if (daddr >= fs->lfs_lastpseg && daddr <= fs->lfs_offset) { 986 ++sp->ndupino; 987 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 988 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 989 (long long)daddr, sp->ndupino)); 990 } 991 /* 992 * Account the inode: it no longer belongs to its former segment, 993 * though it will not belong to the new segment until that segment 994 * is actually written. 995 */ 996 if (daddr != LFS_UNUSED_DADDR) { 997 u_int32_t oldsn = lfs_dtosn(fs, daddr); 998 #ifdef DIAGNOSTIC 999 int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0; 1000 #endif 1001 LFS_SEGENTRY(sup, fs, oldsn, bp); 1002 #ifdef DIAGNOSTIC 1003 if (sup->su_nbytes + 1004 sizeof (struct ulfs1_dinode) * ndupino 1005 < sizeof (struct ulfs1_dinode)) { 1006 printf("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)sizeof (struct ulfs1_dinode) * 1013 (1 - sp->ndupino) - sup->su_nbytes, 1014 oldsn, sp->seg_number, daddr, 1015 (unsigned int)sup->su_nbytes, 1016 sp->ndupino); 1017 panic("lfs_writeinode: negative bytes"); 1018 sup->su_nbytes = sizeof (struct ulfs1_dinode); 1019 } 1020 #endif 1021 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1022 lfs_dtosn(fs, daddr), sizeof (struct ulfs1_dinode), ino)); 1023 sup->su_nbytes -= sizeof (struct ulfs1_dinode); 1024 redo_ifile |= 1025 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1026 if (redo_ifile) { 1027 mutex_enter(&lfs_lock); 1028 fs->lfs_flags |= LFS_IFDIRTY; 1029 mutex_exit(&lfs_lock); 1030 /* Don't double-account */ 1031 fs->lfs_idaddr = 0x0; 1032 } 1033 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1034 } 1035 1036 return redo_ifile; 1037 } 1038 1039 int 1040 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1041 { 1042 struct buf *bp; 1043 struct ulfs1_dinode *cdp; 1044 struct vnode *vp = ITOV(ip); 1045 daddr_t daddr; 1046 int32_t *daddrp; /* XXX ondisk32 */ 1047 int i, ndx; 1048 int redo_ifile = 0; 1049 int gotblk = 0; 1050 int count; 1051 1052 ASSERT_SEGLOCK(fs); 1053 if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1054 return (0); 1055 1056 /* Can't write ifile when writer is not set */ 1057 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1058 (sp->seg_flags & SEGM_CLEAN)); 1059 1060 /* 1061 * If this is the Ifile, see if writing it here will generate a 1062 * temporary misaccounting. If it will, do the accounting and write 1063 * the blocks, postponing the inode write until the accounting is 1064 * solid. 1065 */ 1066 count = 0; 1067 while (vp == fs->lfs_ivnode) { 1068 int redo = 0; 1069 1070 if (sp->idp == NULL && sp->ibp == NULL && 1071 (sp->seg_bytes_left < fs->lfs_ibsize || 1072 sp->sum_bytes_left < sizeof(int32_t))) { 1073 (void) lfs_writeseg(fs, sp); 1074 continue; 1075 } 1076 1077 /* Look for dirty Ifile blocks */ 1078 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1079 if (!(bp->b_flags & B_GATHERED)) { 1080 redo = 1; 1081 break; 1082 } 1083 } 1084 1085 if (redo == 0) 1086 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1087 if (redo == 0) 1088 break; 1089 1090 if (sp->idp) { 1091 sp->idp->di_inumber = 0; 1092 sp->idp = NULL; 1093 } 1094 ++count; 1095 if (count > 2) 1096 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1097 lfs_writefile(fs, sp, fs->lfs_ivnode); 1098 } 1099 1100 /* Allocate a new inode block if necessary. */ 1101 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1102 sp->ibp == NULL) { 1103 /* Allocate a new segment if necessary. */ 1104 if (sp->seg_bytes_left < fs->lfs_ibsize || 1105 sp->sum_bytes_left < sizeof(int32_t)) 1106 (void) lfs_writeseg(fs, sp); 1107 1108 /* Get next inode block. */ 1109 daddr = fs->lfs_offset; 1110 fs->lfs_offset += lfs_btofsb(fs, fs->lfs_ibsize); 1111 sp->ibp = *sp->cbpp++ = 1112 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1113 LFS_FSBTODB(fs, daddr), fs->lfs_ibsize, 0, 0); 1114 gotblk++; 1115 1116 /* Zero out inode numbers */ 1117 for (i = 0; i < LFS_INOPB(fs); ++i) 1118 ((struct ulfs1_dinode *)sp->ibp->b_data)[i].di_inumber = 1119 0; 1120 1121 ++sp->start_bpp; 1122 fs->lfs_avail -= lfs_btofsb(fs, fs->lfs_ibsize); 1123 /* Set remaining space counters. */ 1124 sp->seg_bytes_left -= fs->lfs_ibsize; 1125 sp->sum_bytes_left -= sizeof(int32_t); 1126 ndx = fs->lfs_sumsize / sizeof(int32_t) - 1127 sp->ninodes / LFS_INOPB(fs) - 1; 1128 ((int32_t *)(sp->segsum))[ndx] = daddr; 1129 } 1130 1131 /* Check VU_DIROP in case there is a new file with no data blocks */ 1132 if (vp->v_uflag & VU_DIROP) 1133 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 1134 1135 /* Update the inode times and copy the inode onto the inode page. */ 1136 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1137 if (ip->i_number != LFS_IFILE_INUM) 1138 LFS_ITIMES(ip, NULL, NULL, NULL); 1139 1140 /* 1141 * If this is the Ifile, and we've already written the Ifile in this 1142 * partial segment, just overwrite it (it's not on disk yet) and 1143 * continue. 1144 * 1145 * XXX we know that the bp that we get the second time around has 1146 * already been gathered. 1147 */ 1148 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1149 *(sp->idp) = *ip->i_din.ffs1_din; 1150 ip->i_lfs_osize = ip->i_size; 1151 return 0; 1152 } 1153 1154 bp = sp->ibp; 1155 cdp = ((struct ulfs1_dinode *)bp->b_data) + (sp->ninodes % LFS_INOPB(fs)); 1156 *cdp = *ip->i_din.ffs1_din; 1157 1158 /* 1159 * This inode is on its way to disk; clear its VU_DIROP status when 1160 * the write is complete. 1161 */ 1162 if (vp->v_uflag & VU_DIROP) { 1163 if (!(sp->seg_flags & SEGM_CLEAN)) 1164 ip->i_flag |= IN_CDIROP; 1165 else { 1166 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1167 } 1168 } 1169 1170 /* 1171 * If cleaning, link counts and directory file sizes cannot change, 1172 * since those would be directory operations---even if the file 1173 * we are writing is marked VU_DIROP we should write the old values. 1174 * If we're not cleaning, of course, update the values so we get 1175 * current values the next time we clean. 1176 */ 1177 if (sp->seg_flags & SEGM_CLEAN) { 1178 if (vp->v_uflag & VU_DIROP) { 1179 cdp->di_nlink = ip->i_lfs_odnlink; 1180 /* if (vp->v_type == VDIR) */ 1181 cdp->di_size = ip->i_lfs_osize; 1182 } 1183 } else { 1184 ip->i_lfs_odnlink = cdp->di_nlink; 1185 ip->i_lfs_osize = ip->i_size; 1186 } 1187 1188 1189 /* We can finish the segment accounting for truncations now */ 1190 lfs_finalize_ino_seguse(fs, ip); 1191 1192 /* 1193 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1194 * addresses to disk; possibly change the on-disk record of 1195 * the inode size, either by reverting to the previous size 1196 * (in the case of cleaning) or by verifying the inode's block 1197 * holdings (in the case of files being allocated as they are being 1198 * written). 1199 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1200 * XXX count on disk wrong by the same amount. We should be 1201 * XXX able to "borrow" from lfs_avail and return it after the 1202 * XXX Ifile is written. See also in lfs_writeseg. 1203 */ 1204 1205 /* Check file size based on highest allocated block */ 1206 if (((ip->i_ffs1_mode & LFS_IFMT) == LFS_IFREG || 1207 (ip->i_ffs1_mode & LFS_IFMT) == LFS_IFDIR) && 1208 ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) { 1209 cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift; 1210 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1211 PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size)); 1212 } 1213 if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) { 1214 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)" 1215 " at %x\n", ip->i_number, ip->i_lfs_effnblks, 1216 ip->i_ffs1_blocks, fs->lfs_offset)); 1217 for (daddrp = cdp->di_db; daddrp < cdp->di_ib + ULFS_NIADDR; 1218 daddrp++) { 1219 if (*daddrp == UNWRITTEN) { 1220 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1221 *daddrp = 0; 1222 } 1223 } 1224 } 1225 1226 #ifdef DIAGNOSTIC 1227 /* 1228 * Check dinode held blocks against dinode size. 1229 * This should be identical to the check in lfs_vget(). 1230 */ 1231 for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift; 1232 i < ULFS_NDADDR; i++) { 1233 KASSERT(i >= 0); 1234 if ((cdp->di_mode & LFS_IFMT) == LFS_IFLNK) 1235 continue; 1236 if (((cdp->di_mode & LFS_IFMT) == LFS_IFBLK || 1237 (cdp->di_mode & LFS_IFMT) == LFS_IFCHR) && i == 0) 1238 continue; 1239 if (cdp->di_db[i] != 0) { 1240 # ifdef DEBUG 1241 lfs_dump_dinode(cdp); 1242 # endif 1243 panic("writing inconsistent inode"); 1244 } 1245 } 1246 #endif /* DIAGNOSTIC */ 1247 1248 if (ip->i_flag & IN_CLEANING) 1249 LFS_CLR_UINO(ip, IN_CLEANING); 1250 else { 1251 /* XXX IN_ALLMOD */ 1252 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1253 IN_UPDATE | IN_MODIFY); 1254 if (ip->i_lfs_effnblks == ip->i_ffs1_blocks) 1255 LFS_CLR_UINO(ip, IN_MODIFIED); 1256 else { 1257 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1258 "blks=%d, eff=%d\n", ip->i_number, 1259 ip->i_ffs1_blocks, ip->i_lfs_effnblks)); 1260 } 1261 } 1262 1263 if (ip->i_number == LFS_IFILE_INUM) { 1264 /* We know sp->idp == NULL */ 1265 sp->idp = ((struct ulfs1_dinode *)bp->b_data) + 1266 (sp->ninodes % LFS_INOPB(fs)); 1267 1268 /* Not dirty any more */ 1269 mutex_enter(&lfs_lock); 1270 fs->lfs_flags &= ~LFS_IFDIRTY; 1271 mutex_exit(&lfs_lock); 1272 } 1273 1274 if (gotblk) { 1275 mutex_enter(&bufcache_lock); 1276 LFS_LOCK_BUF(bp); 1277 brelsel(bp, 0); 1278 mutex_exit(&bufcache_lock); 1279 } 1280 1281 /* Increment inode count in segment summary block. */ 1282 ++((SEGSUM *)(sp->segsum))->ss_ninos; 1283 1284 /* If this page is full, set flag to allocate a new page. */ 1285 if (++sp->ninodes % LFS_INOPB(fs) == 0) 1286 sp->ibp = NULL; 1287 1288 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1289 1290 KASSERT(redo_ifile == 0); 1291 return (redo_ifile); 1292 } 1293 1294 int 1295 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr) 1296 { 1297 struct lfs *fs; 1298 int vers; 1299 int j, blksinblk; 1300 1301 ASSERT_SEGLOCK(sp->fs); 1302 /* 1303 * If full, finish this segment. We may be doing I/O, so 1304 * release and reacquire the splbio(). 1305 */ 1306 #ifdef DIAGNOSTIC 1307 if (sp->vp == NULL) 1308 panic ("lfs_gatherblock: Null vp in segment"); 1309 #endif 1310 fs = sp->fs; 1311 blksinblk = howmany(bp->b_bcount, fs->lfs_bsize); 1312 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1313 sp->seg_bytes_left < bp->b_bcount) { 1314 if (mptr) 1315 mutex_exit(mptr); 1316 lfs_updatemeta(sp); 1317 1318 vers = sp->fip->fi_version; 1319 (void) lfs_writeseg(fs, sp); 1320 1321 /* Add the current file to the segment summary. */ 1322 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1323 1324 if (mptr) 1325 mutex_enter(mptr); 1326 return (1); 1327 } 1328 1329 if (bp->b_flags & B_GATHERED) { 1330 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d," 1331 " lbn %" PRId64 "\n", 1332 sp->fip->fi_ino, bp->b_lblkno)); 1333 return (0); 1334 } 1335 1336 /* Insert into the buffer list, update the FINFO block. */ 1337 bp->b_flags |= B_GATHERED; 1338 1339 *sp->cbpp++ = bp; 1340 for (j = 0; j < blksinblk; j++) { 1341 sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j; 1342 /* This block's accounting moves from lfs_favail to lfs_avail */ 1343 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1344 } 1345 1346 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1347 sp->seg_bytes_left -= bp->b_bcount; 1348 return (0); 1349 } 1350 1351 int 1352 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1353 int (*match)(struct lfs *, struct buf *)) 1354 { 1355 struct buf *bp, *nbp; 1356 int count = 0; 1357 1358 ASSERT_SEGLOCK(fs); 1359 if (vp->v_type == VBLK) 1360 return 0; 1361 KASSERT(sp->vp == NULL); 1362 sp->vp = vp; 1363 mutex_enter(&bufcache_lock); 1364 1365 #ifndef LFS_NO_BACKBUF_HACK 1366 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1367 # define BUF_OFFSET \ 1368 (((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp) 1369 # define BACK_BUF(BP) \ 1370 ((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1371 # define BEG_OF_LIST \ 1372 ((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1373 1374 loop: 1375 /* Find last buffer. */ 1376 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1377 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1378 bp = LIST_NEXT(bp, b_vnbufs)) 1379 /* nothing */; 1380 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1381 nbp = BACK_BUF(bp); 1382 #else /* LFS_NO_BACKBUF_HACK */ 1383 loop: 1384 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1385 nbp = LIST_NEXT(bp, b_vnbufs); 1386 #endif /* LFS_NO_BACKBUF_HACK */ 1387 if ((bp->b_cflags & BC_BUSY) != 0 || 1388 (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) { 1389 #ifdef DEBUG 1390 if (vp == fs->lfs_ivnode && 1391 (bp->b_cflags & BC_BUSY) != 0 && 1392 (bp->b_flags & B_GATHERED) == 0) 1393 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1394 PRId64 " busy (%x) at 0x%x", 1395 bp->b_lblkno, bp->b_flags, 1396 (unsigned)fs->lfs_offset); 1397 #endif 1398 continue; 1399 } 1400 #ifdef DIAGNOSTIC 1401 # ifdef LFS_USE_B_INVAL 1402 if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) { 1403 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1404 " is BC_INVAL\n", bp->b_lblkno)); 1405 VOP_PRINT(bp->b_vp); 1406 } 1407 # endif /* LFS_USE_B_INVAL */ 1408 if (!(bp->b_oflags & BO_DELWRI)) 1409 panic("lfs_gather: bp not BO_DELWRI"); 1410 if (!(bp->b_flags & B_LOCKED)) { 1411 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1412 " blk %" PRId64 " not B_LOCKED\n", 1413 bp->b_lblkno, 1414 LFS_DBTOFSB(fs, bp->b_blkno))); 1415 VOP_PRINT(bp->b_vp); 1416 panic("lfs_gather: bp not B_LOCKED"); 1417 } 1418 #endif 1419 if (lfs_gatherblock(sp, bp, &bufcache_lock)) { 1420 goto loop; 1421 } 1422 count++; 1423 } 1424 mutex_exit(&bufcache_lock); 1425 lfs_updatemeta(sp); 1426 KASSERT(sp->vp == vp); 1427 sp->vp = NULL; 1428 return count; 1429 } 1430 1431 #if DEBUG 1432 # define DEBUG_OOFF(n) do { \ 1433 if (ooff == 0) { \ 1434 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1435 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1436 ", was 0x0 (or %" PRId64 ")\n", \ 1437 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1438 } \ 1439 } while (0) 1440 #else 1441 # define DEBUG_OOFF(n) 1442 #endif 1443 1444 /* 1445 * Change the given block's address to ndaddr, finding its previous 1446 * location using ulfs_bmaparray(). 1447 * 1448 * Account for this change in the segment table. 1449 * 1450 * called with sp == NULL by roll-forwarding code. 1451 */ 1452 void 1453 lfs_update_single(struct lfs *fs, struct segment *sp, 1454 struct vnode *vp, daddr_t lbn, int32_t ndaddr, int size) 1455 { 1456 SEGUSE *sup; 1457 struct buf *bp; 1458 struct indir a[ULFS_NIADDR + 2], *ap; 1459 struct inode *ip; 1460 daddr_t daddr, ooff; 1461 int num, error; 1462 int bb, osize, obb; 1463 1464 ASSERT_SEGLOCK(fs); 1465 KASSERT(sp == NULL || sp->vp == vp); 1466 ip = VTOI(vp); 1467 1468 error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1469 if (error) 1470 panic("lfs_updatemeta: ulfs_bmaparray returned %d", error); 1471 1472 daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */ 1473 KASSERT(daddr <= LFS_MAX_DADDR); 1474 if (daddr > 0) 1475 daddr = LFS_DBTOFSB(fs, daddr); 1476 1477 bb = lfs_numfrags(fs, size); 1478 switch (num) { 1479 case 0: 1480 ooff = ip->i_ffs1_db[lbn]; 1481 DEBUG_OOFF(0); 1482 if (ooff == UNWRITTEN) 1483 ip->i_ffs1_blocks += bb; 1484 else { 1485 /* possible fragment truncation or extension */ 1486 obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]); 1487 ip->i_ffs1_blocks += (bb - obb); 1488 } 1489 ip->i_ffs1_db[lbn] = ndaddr; 1490 break; 1491 case 1: 1492 ooff = ip->i_ffs1_ib[a[0].in_off]; 1493 DEBUG_OOFF(1); 1494 if (ooff == UNWRITTEN) 1495 ip->i_ffs1_blocks += bb; 1496 ip->i_ffs1_ib[a[0].in_off] = ndaddr; 1497 break; 1498 default: 1499 ap = &a[num - 1]; 1500 if (bread(vp, ap->in_lbn, fs->lfs_bsize, 1501 B_MODIFY, &bp)) 1502 panic("lfs_updatemeta: bread bno %" PRId64, 1503 ap->in_lbn); 1504 1505 /* XXX ondisk32 */ 1506 ooff = ((int32_t *)bp->b_data)[ap->in_off]; 1507 DEBUG_OOFF(num); 1508 if (ooff == UNWRITTEN) 1509 ip->i_ffs1_blocks += bb; 1510 /* XXX ondisk32 */ 1511 ((int32_t *)bp->b_data)[ap->in_off] = ndaddr; 1512 (void) VOP_BWRITE(bp->b_vp, bp); 1513 } 1514 1515 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1516 1517 /* Update hiblk when extending the file */ 1518 if (lbn > ip->i_lfs_hiblk) 1519 ip->i_lfs_hiblk = lbn; 1520 1521 /* 1522 * Though we'd rather it couldn't, this *can* happen right now 1523 * if cleaning blocks and regular blocks coexist. 1524 */ 1525 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1526 1527 /* 1528 * Update segment usage information, based on old size 1529 * and location. 1530 */ 1531 if (daddr > 0) { 1532 u_int32_t oldsn = lfs_dtosn(fs, daddr); 1533 #ifdef DIAGNOSTIC 1534 int ndupino; 1535 1536 if (sp && sp->seg_number == oldsn) { 1537 ndupino = sp->ndupino; 1538 } else { 1539 ndupino = 0; 1540 } 1541 #endif 1542 KASSERT(oldsn < fs->lfs_nseg); 1543 if (lbn >= 0 && lbn < ULFS_NDADDR) 1544 osize = ip->i_lfs_fragsize[lbn]; 1545 else 1546 osize = fs->lfs_bsize; 1547 LFS_SEGENTRY(sup, fs, oldsn, bp); 1548 #ifdef DIAGNOSTIC 1549 if (sup->su_nbytes + sizeof (struct ulfs1_dinode) * ndupino 1550 < osize) { 1551 printf("lfs_updatemeta: negative bytes " 1552 "(segment %" PRIu32 " short by %" PRId64 1553 ")\n", lfs_dtosn(fs, daddr), 1554 (int64_t)osize - 1555 (sizeof (struct ulfs1_dinode) * ndupino + 1556 sup->su_nbytes)); 1557 printf("lfs_updatemeta: ino %llu, lbn %" PRId64 1558 ", addr = 0x%" PRIx64 "\n", 1559 (unsigned long long)ip->i_number, lbn, daddr); 1560 printf("lfs_updatemeta: ndupino=%d\n", ndupino); 1561 panic("lfs_updatemeta: negative bytes"); 1562 sup->su_nbytes = osize - 1563 sizeof (struct ulfs1_dinode) * ndupino; 1564 } 1565 #endif 1566 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1567 " db 0x%" PRIx64 "\n", 1568 lfs_dtosn(fs, daddr), osize, 1569 ip->i_number, lbn, daddr)); 1570 sup->su_nbytes -= osize; 1571 if (!(bp->b_flags & B_GATHERED)) { 1572 mutex_enter(&lfs_lock); 1573 fs->lfs_flags |= LFS_IFDIRTY; 1574 mutex_exit(&lfs_lock); 1575 } 1576 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1577 } 1578 /* 1579 * Now that this block has a new address, and its old 1580 * segment no longer owns it, we can forget about its 1581 * old size. 1582 */ 1583 if (lbn >= 0 && lbn < ULFS_NDADDR) 1584 ip->i_lfs_fragsize[lbn] = size; 1585 } 1586 1587 /* 1588 * Update the metadata that points to the blocks listed in the FINFO 1589 * array. 1590 */ 1591 void 1592 lfs_updatemeta(struct segment *sp) 1593 { 1594 struct buf *sbp; 1595 struct lfs *fs; 1596 struct vnode *vp; 1597 daddr_t lbn; 1598 int i, nblocks, num; 1599 int bb; 1600 int bytesleft, size; 1601 1602 ASSERT_SEGLOCK(sp->fs); 1603 vp = sp->vp; 1604 nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1605 KASSERT(nblocks >= 0); 1606 KASSERT(vp != NULL); 1607 if (nblocks == 0) 1608 return; 1609 1610 /* 1611 * This count may be high due to oversize blocks from lfs_gop_write. 1612 * Correct for this. (XXX we should be able to keep track of these.) 1613 */ 1614 fs = sp->fs; 1615 for (i = 0; i < nblocks; i++) { 1616 if (sp->start_bpp[i] == NULL) { 1617 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1618 nblocks = i; 1619 break; 1620 } 1621 num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize); 1622 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1623 nblocks -= num - 1; 1624 } 1625 1626 KASSERT(vp->v_type == VREG || 1627 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1628 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1629 1630 /* 1631 * Sort the blocks. 1632 * 1633 * We have to sort even if the blocks come from the 1634 * cleaner, because there might be other pending blocks on the 1635 * same inode...and if we don't sort, and there are fragments 1636 * present, blocks may be written in the wrong place. 1637 */ 1638 lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize); 1639 1640 /* 1641 * Record the length of the last block in case it's a fragment. 1642 * If there are indirect blocks present, they sort last. An 1643 * indirect block will be lfs_bsize and its presence indicates 1644 * that you cannot have fragments. 1645 * 1646 * XXX This last is a lie. A cleaned fragment can coexist with 1647 * XXX a later indirect block. This will continue to be 1648 * XXX true until lfs_markv is fixed to do everything with 1649 * XXX fake blocks (including fake inodes and fake indirect blocks). 1650 */ 1651 sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1652 fs->lfs_bmask) + 1; 1653 1654 /* 1655 * Assign disk addresses, and update references to the logical 1656 * block and the segment usage information. 1657 */ 1658 for (i = nblocks; i--; ++sp->start_bpp) { 1659 sbp = *sp->start_bpp; 1660 lbn = *sp->start_lbp; 1661 KASSERT(sbp->b_lblkno == lbn); 1662 1663 sbp->b_blkno = LFS_FSBTODB(fs, fs->lfs_offset); 1664 1665 /* 1666 * If we write a frag in the wrong place, the cleaner won't 1667 * be able to correctly identify its size later, and the 1668 * segment will be uncleanable. (Even worse, it will assume 1669 * that the indirect block that actually ends the list 1670 * is of a smaller size!) 1671 */ 1672 if ((sbp->b_bcount & fs->lfs_bmask) && i != 0) 1673 panic("lfs_updatemeta: fragment is not last block"); 1674 1675 /* 1676 * For each subblock in this possibly oversized block, 1677 * update its address on disk. 1678 */ 1679 KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize); 1680 KASSERT(vp == sbp->b_vp); 1681 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1682 bytesleft -= fs->lfs_bsize) { 1683 size = MIN(bytesleft, fs->lfs_bsize); 1684 bb = lfs_numfrags(fs, size); 1685 lbn = *sp->start_lbp++; 1686 lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset, 1687 size); 1688 fs->lfs_offset += bb; 1689 } 1690 1691 } 1692 1693 /* This inode has been modified */ 1694 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1695 } 1696 1697 /* 1698 * Move lfs_offset to a segment earlier than sn. 1699 */ 1700 int 1701 lfs_rewind(struct lfs *fs, int newsn) 1702 { 1703 int sn, osn, isdirty; 1704 struct buf *bp; 1705 SEGUSE *sup; 1706 1707 ASSERT_SEGLOCK(fs); 1708 1709 osn = lfs_dtosn(fs, fs->lfs_offset); 1710 if (osn < newsn) 1711 return 0; 1712 1713 /* lfs_avail eats the remaining space in this segment */ 1714 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg); 1715 1716 /* Find a low-numbered segment */ 1717 for (sn = 0; sn < fs->lfs_nseg; ++sn) { 1718 LFS_SEGENTRY(sup, fs, sn, bp); 1719 isdirty = sup->su_flags & SEGUSE_DIRTY; 1720 brelse(bp, 0); 1721 1722 if (!isdirty) 1723 break; 1724 } 1725 if (sn == fs->lfs_nseg) 1726 panic("lfs_rewind: no clean segments"); 1727 if (newsn >= 0 && sn >= newsn) 1728 return ENOENT; 1729 fs->lfs_nextseg = sn; 1730 lfs_newseg(fs); 1731 fs->lfs_offset = fs->lfs_curseg; 1732 1733 return 0; 1734 } 1735 1736 /* 1737 * Start a new partial segment. 1738 * 1739 * Return 1 when we entered to a new segment. 1740 * Otherwise, return 0. 1741 */ 1742 int 1743 lfs_initseg(struct lfs *fs) 1744 { 1745 struct segment *sp = fs->lfs_sp; 1746 SEGSUM *ssp; 1747 struct buf *sbp; /* buffer for SEGSUM */ 1748 int repeat = 0; /* return value */ 1749 1750 ASSERT_SEGLOCK(fs); 1751 /* Advance to the next segment. */ 1752 if (!LFS_PARTIAL_FITS(fs)) { 1753 SEGUSE *sup; 1754 struct buf *bp; 1755 1756 /* lfs_avail eats the remaining space */ 1757 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - 1758 fs->lfs_curseg); 1759 /* Wake up any cleaning procs waiting on this file system. */ 1760 lfs_wakeup_cleaner(fs); 1761 lfs_newseg(fs); 1762 repeat = 1; 1763 fs->lfs_offset = fs->lfs_curseg; 1764 1765 sp->seg_number = lfs_dtosn(fs, fs->lfs_curseg); 1766 sp->seg_bytes_left = lfs_fsbtob(fs, fs->lfs_fsbpseg); 1767 1768 /* 1769 * If the segment contains a superblock, update the offset 1770 * and summary address to skip over it. 1771 */ 1772 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1773 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1774 fs->lfs_offset += lfs_btofsb(fs, LFS_SBPAD); 1775 sp->seg_bytes_left -= LFS_SBPAD; 1776 } 1777 brelse(bp, 0); 1778 /* Segment zero could also contain the labelpad */ 1779 if (fs->lfs_version > 1 && sp->seg_number == 0 && 1780 fs->lfs_s0addr < lfs_btofsb(fs, LFS_LABELPAD)) { 1781 fs->lfs_offset += 1782 lfs_btofsb(fs, LFS_LABELPAD) - fs->lfs_s0addr; 1783 sp->seg_bytes_left -= 1784 LFS_LABELPAD - lfs_fsbtob(fs, fs->lfs_s0addr); 1785 } 1786 } else { 1787 sp->seg_number = lfs_dtosn(fs, fs->lfs_curseg); 1788 sp->seg_bytes_left = lfs_fsbtob(fs, fs->lfs_fsbpseg - 1789 (fs->lfs_offset - fs->lfs_curseg)); 1790 } 1791 fs->lfs_lastpseg = fs->lfs_offset; 1792 1793 /* Record first address of this partial segment */ 1794 if (sp->seg_flags & SEGM_CLEAN) { 1795 fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset; 1796 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1797 /* "1" is the artificial inc in lfs_seglock */ 1798 mutex_enter(&lfs_lock); 1799 while (fs->lfs_iocount > 1) { 1800 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1801 "lfs_initseg", 0, &lfs_lock); 1802 } 1803 mutex_exit(&lfs_lock); 1804 fs->lfs_cleanind = 0; 1805 } 1806 } 1807 1808 sp->fs = fs; 1809 sp->ibp = NULL; 1810 sp->idp = NULL; 1811 sp->ninodes = 0; 1812 sp->ndupino = 0; 1813 1814 sp->cbpp = sp->bpp; 1815 1816 /* Get a new buffer for SEGSUM */ 1817 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1818 LFS_FSBTODB(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY); 1819 1820 /* ... and enter it into the buffer list. */ 1821 *sp->cbpp = sbp; 1822 sp->cbpp++; 1823 fs->lfs_offset += lfs_btofsb(fs, fs->lfs_sumsize); 1824 1825 sp->start_bpp = sp->cbpp; 1826 1827 /* Set point to SEGSUM, initialize it. */ 1828 ssp = sp->segsum = sbp->b_data; 1829 memset(ssp, 0, fs->lfs_sumsize); 1830 ssp->ss_next = fs->lfs_nextseg; 1831 ssp->ss_nfinfo = ssp->ss_ninos = 0; 1832 ssp->ss_magic = SS_MAGIC; 1833 1834 /* Set pointer to first FINFO, initialize it. */ 1835 sp->fip = (struct finfo *)((char *)sp->segsum + SEGSUM_SIZE(fs)); 1836 sp->fip->fi_nblocks = 0; 1837 sp->start_lbp = &sp->fip->fi_blocks[0]; 1838 sp->fip->fi_lastlength = 0; 1839 1840 sp->seg_bytes_left -= fs->lfs_sumsize; 1841 sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs); 1842 1843 return (repeat); 1844 } 1845 1846 /* 1847 * Remove SEGUSE_INVAL from all segments. 1848 */ 1849 void 1850 lfs_unset_inval_all(struct lfs *fs) 1851 { 1852 SEGUSE *sup; 1853 struct buf *bp; 1854 int i; 1855 1856 for (i = 0; i < fs->lfs_nseg; i++) { 1857 LFS_SEGENTRY(sup, fs, i, bp); 1858 if (sup->su_flags & SEGUSE_INVAL) { 1859 sup->su_flags &= ~SEGUSE_INVAL; 1860 LFS_WRITESEGENTRY(sup, fs, i, bp); 1861 } else 1862 brelse(bp, 0); 1863 } 1864 } 1865 1866 /* 1867 * Return the next segment to write. 1868 */ 1869 void 1870 lfs_newseg(struct lfs *fs) 1871 { 1872 CLEANERINFO *cip; 1873 SEGUSE *sup; 1874 struct buf *bp; 1875 int curseg, isdirty, sn, skip_inval; 1876 1877 ASSERT_SEGLOCK(fs); 1878 1879 /* Honor LFCNWRAPSTOP */ 1880 mutex_enter(&lfs_lock); 1881 while (fs->lfs_nextseg < fs->lfs_curseg && fs->lfs_nowrap) { 1882 if (fs->lfs_wrappass) { 1883 log(LOG_NOTICE, "%s: wrappass=%d\n", 1884 fs->lfs_fsmnt, fs->lfs_wrappass); 1885 fs->lfs_wrappass = 0; 1886 break; 1887 } 1888 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1889 wakeup(&fs->lfs_nowrap); 1890 log(LOG_NOTICE, "%s: waiting at log wrap\n", fs->lfs_fsmnt); 1891 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1892 &lfs_lock); 1893 } 1894 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1895 mutex_exit(&lfs_lock); 1896 1897 LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, fs->lfs_nextseg), bp); 1898 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1899 lfs_dtosn(fs, fs->lfs_nextseg))); 1900 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1901 sup->su_nbytes = 0; 1902 sup->su_nsums = 0; 1903 sup->su_ninos = 0; 1904 LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, fs->lfs_nextseg), bp); 1905 1906 LFS_CLEANERINFO(cip, fs, bp); 1907 --cip->clean; 1908 ++cip->dirty; 1909 fs->lfs_nclean = cip->clean; 1910 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1911 1912 fs->lfs_lastseg = fs->lfs_curseg; 1913 fs->lfs_curseg = fs->lfs_nextseg; 1914 skip_inval = 1; 1915 for (sn = curseg = lfs_dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) { 1916 sn = (sn + 1) % fs->lfs_nseg; 1917 1918 if (sn == curseg) { 1919 if (skip_inval) 1920 skip_inval = 0; 1921 else 1922 panic("lfs_nextseg: no clean segments"); 1923 } 1924 LFS_SEGENTRY(sup, fs, sn, bp); 1925 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1926 /* Check SEGUSE_EMPTY as we go along */ 1927 if (isdirty && sup->su_nbytes == 0 && 1928 !(sup->su_flags & SEGUSE_EMPTY)) 1929 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1930 else 1931 brelse(bp, 0); 1932 1933 if (!isdirty) 1934 break; 1935 } 1936 if (skip_inval == 0) 1937 lfs_unset_inval_all(fs); 1938 1939 ++fs->lfs_nactive; 1940 fs->lfs_nextseg = lfs_sntod(fs, sn); 1941 if (lfs_dostats) { 1942 ++lfs_stats.segsused; 1943 } 1944 } 1945 1946 static struct buf * 1947 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1948 int n) 1949 { 1950 struct lfs_cluster *cl; 1951 struct buf **bpp, *bp; 1952 1953 ASSERT_SEGLOCK(fs); 1954 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1955 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1956 memset(cl, 0, sizeof(*cl)); 1957 cl->fs = fs; 1958 cl->bpp = bpp; 1959 cl->bufcount = 0; 1960 cl->bufsize = 0; 1961 1962 /* If this segment is being written synchronously, note that */ 1963 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 1964 cl->flags |= LFS_CL_SYNC; 1965 cl->seg = fs->lfs_sp; 1966 ++cl->seg->seg_iocount; 1967 } 1968 1969 /* Get an empty buffer header, or maybe one with something on it */ 1970 bp = getiobuf(vp, true); 1971 bp->b_dev = NODEV; 1972 bp->b_blkno = bp->b_lblkno = addr; 1973 bp->b_iodone = lfs_cluster_callback; 1974 bp->b_private = cl; 1975 1976 return bp; 1977 } 1978 1979 int 1980 lfs_writeseg(struct lfs *fs, struct segment *sp) 1981 { 1982 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 1983 SEGUSE *sup; 1984 SEGSUM *ssp; 1985 int i; 1986 int do_again, nblocks, byteoffset; 1987 size_t el_size; 1988 struct lfs_cluster *cl; 1989 u_short ninos; 1990 struct vnode *devvp; 1991 char *p = NULL; 1992 struct vnode *vp; 1993 int32_t *daddrp; /* XXX ondisk32 */ 1994 int changed; 1995 u_int32_t sum; 1996 #ifdef DEBUG 1997 FINFO *fip; 1998 int findex; 1999 #endif 2000 2001 ASSERT_SEGLOCK(fs); 2002 2003 ssp = (SEGSUM *)sp->segsum; 2004 2005 /* 2006 * If there are no buffers other than the segment summary to write, 2007 * don't do anything. If we are the end of a dirop sequence, however, 2008 * write the empty segment summary anyway, to help out the 2009 * roll-forward agent. 2010 */ 2011 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2012 if ((ssp->ss_flags & (SS_DIROP | SS_CONT)) != SS_DIROP) 2013 return 0; 2014 } 2015 2016 /* Note if partial segment is being written by the cleaner */ 2017 if (sp->seg_flags & SEGM_CLEAN) 2018 ssp->ss_flags |= SS_CLEAN; 2019 2020 /* Note if we are writing to reclaim */ 2021 if (sp->seg_flags & SEGM_RECLAIM) { 2022 ssp->ss_flags |= SS_RECLAIM; 2023 ssp->ss_reclino = fs->lfs_reclino; 2024 } 2025 2026 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2027 2028 /* Update the segment usage information. */ 2029 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2030 2031 /* Loop through all blocks, except the segment summary. */ 2032 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2033 if ((*bpp)->b_vp != devvp) { 2034 sup->su_nbytes += (*bpp)->b_bcount; 2035 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2036 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2037 sp->seg_number, (*bpp)->b_bcount, 2038 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2039 (*bpp)->b_blkno)); 2040 } 2041 } 2042 2043 #ifdef DEBUG 2044 /* Check for zero-length and zero-version FINFO entries. */ 2045 fip = (struct finfo *)((char *)ssp + SEGSUM_SIZE(fs)); 2046 for (findex = 0; findex < ssp->ss_nfinfo; findex++) { 2047 KDASSERT(fip->fi_nblocks > 0); 2048 KDASSERT(fip->fi_version > 0); 2049 fip = (FINFO *)((char *)fip + FINFOSIZE + 2050 sizeof(int32_t) * fip->fi_nblocks); 2051 } 2052 #endif /* DEBUG */ 2053 2054 ninos = (ssp->ss_ninos + LFS_INOPB(fs) - 1) / LFS_INOPB(fs); 2055 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2056 sp->seg_number, ssp->ss_ninos * sizeof (struct ulfs1_dinode), 2057 ssp->ss_ninos)); 2058 sup->su_nbytes += ssp->ss_ninos * sizeof (struct ulfs1_dinode); 2059 /* sup->su_nbytes += fs->lfs_sumsize; */ 2060 if (fs->lfs_version == 1) 2061 sup->su_olastmod = time_second; 2062 else 2063 sup->su_lastmod = time_second; 2064 sup->su_ninos += ninos; 2065 ++sup->su_nsums; 2066 fs->lfs_avail -= lfs_btofsb(fs, fs->lfs_sumsize); 2067 2068 do_again = !(bp->b_flags & B_GATHERED); 2069 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2070 2071 /* 2072 * Mark blocks B_BUSY, to prevent then from being changed between 2073 * the checksum computation and the actual write. 2074 * 2075 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2076 * there are any, replace them with copies that have UNASSIGNED 2077 * instead. 2078 */ 2079 mutex_enter(&bufcache_lock); 2080 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2081 ++bpp; 2082 bp = *bpp; 2083 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2084 bp->b_cflags |= BC_BUSY; 2085 continue; 2086 } 2087 2088 while (bp->b_cflags & BC_BUSY) { 2089 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2090 " data summary corruption for ino %d, lbn %" 2091 PRId64 "\n", 2092 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2093 bp->b_cflags |= BC_WANTED; 2094 cv_wait(&bp->b_busy, &bufcache_lock); 2095 } 2096 bp->b_cflags |= BC_BUSY; 2097 mutex_exit(&bufcache_lock); 2098 unbusybp = NULL; 2099 2100 /* 2101 * Check and replace indirect block UNWRITTEN bogosity. 2102 * XXX See comment in lfs_writefile. 2103 */ 2104 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2105 VTOI(bp->b_vp)->i_ffs1_blocks != 2106 VTOI(bp->b_vp)->i_lfs_effnblks) { 2107 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n", 2108 VTOI(bp->b_vp)->i_number, 2109 VTOI(bp->b_vp)->i_lfs_effnblks, 2110 VTOI(bp->b_vp)->i_ffs1_blocks)); 2111 /* Make a copy we'll make changes to */ 2112 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2113 bp->b_bcount, LFS_NB_IBLOCK); 2114 newbp->b_blkno = bp->b_blkno; 2115 memcpy(newbp->b_data, bp->b_data, 2116 newbp->b_bcount); 2117 2118 changed = 0; 2119 /* XXX ondisk32 */ 2120 for (daddrp = (int32_t *)(newbp->b_data); 2121 daddrp < (int32_t *)((char *)newbp->b_data + 2122 newbp->b_bcount); daddrp++) { 2123 if (*daddrp == UNWRITTEN) { 2124 ++changed; 2125 *daddrp = 0; 2126 } 2127 } 2128 /* 2129 * Get rid of the old buffer. Don't mark it clean, 2130 * though, if it still has dirty data on it. 2131 */ 2132 if (changed) { 2133 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2134 " bp = %p newbp = %p\n", changed, bp, 2135 newbp)); 2136 *bpp = newbp; 2137 bp->b_flags &= ~B_GATHERED; 2138 bp->b_error = 0; 2139 if (bp->b_iodone != NULL) { 2140 DLOG((DLOG_SEG, "lfs_writeseg: " 2141 "indir bp should not be B_CALL\n")); 2142 biodone(bp); 2143 bp = NULL; 2144 } else { 2145 /* Still on free list, leave it there */ 2146 unbusybp = bp; 2147 /* 2148 * We have to re-decrement lfs_avail 2149 * since this block is going to come 2150 * back around to us in the next 2151 * segment. 2152 */ 2153 fs->lfs_avail -= 2154 lfs_btofsb(fs, bp->b_bcount); 2155 } 2156 } else { 2157 lfs_freebuf(fs, newbp); 2158 } 2159 } 2160 mutex_enter(&bufcache_lock); 2161 if (unbusybp != NULL) { 2162 unbusybp->b_cflags &= ~BC_BUSY; 2163 if (unbusybp->b_cflags & BC_WANTED) 2164 cv_broadcast(&bp->b_busy); 2165 } 2166 } 2167 mutex_exit(&bufcache_lock); 2168 2169 /* 2170 * Compute checksum across data and then across summary; the first 2171 * block (the summary block) is skipped. Set the create time here 2172 * so that it's guaranteed to be later than the inode mod times. 2173 */ 2174 sum = 0; 2175 if (fs->lfs_version == 1) 2176 el_size = sizeof(u_long); 2177 else 2178 el_size = sizeof(u_int32_t); 2179 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2180 ++bpp; 2181 /* Loop through gop_write cluster blocks */ 2182 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2183 byteoffset += fs->lfs_bsize) { 2184 #ifdef LFS_USE_B_INVAL 2185 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2186 (*bpp)->b_iodone != NULL) { 2187 if (copyin((void *)(*bpp)->b_saveaddr + 2188 byteoffset, dp, el_size)) { 2189 panic("lfs_writeseg: copyin failed [1]:" 2190 " ino %d blk %" PRId64, 2191 VTOI((*bpp)->b_vp)->i_number, 2192 (*bpp)->b_lblkno); 2193 } 2194 } else 2195 #endif /* LFS_USE_B_INVAL */ 2196 { 2197 sum = lfs_cksum_part((char *) 2198 (*bpp)->b_data + byteoffset, el_size, sum); 2199 } 2200 } 2201 } 2202 if (fs->lfs_version == 1) 2203 ssp->ss_ocreate = time_second; 2204 else { 2205 ssp->ss_create = time_second; 2206 ssp->ss_serial = ++fs->lfs_serial; 2207 ssp->ss_ident = fs->lfs_ident; 2208 } 2209 ssp->ss_datasum = lfs_cksum_fold(sum); 2210 ssp->ss_sumsum = cksum(&ssp->ss_datasum, 2211 fs->lfs_sumsize - sizeof(ssp->ss_sumsum)); 2212 2213 mutex_enter(&lfs_lock); 2214 fs->lfs_bfree -= (lfs_btofsb(fs, ninos * fs->lfs_ibsize) + 2215 lfs_btofsb(fs, fs->lfs_sumsize)); 2216 fs->lfs_dmeta += (lfs_btofsb(fs, ninos * fs->lfs_ibsize) + 2217 lfs_btofsb(fs, fs->lfs_sumsize)); 2218 mutex_exit(&lfs_lock); 2219 2220 /* 2221 * When we simply write the blocks we lose a rotation for every block 2222 * written. To avoid this problem, we cluster the buffers into a 2223 * chunk and write the chunk. MAXPHYS is the largest size I/O 2224 * devices can handle, use that for the size of the chunks. 2225 * 2226 * Blocks that are already clusters (from GOP_WRITE), however, we 2227 * don't bother to copy into other clusters. 2228 */ 2229 2230 #define CHUNKSIZE MAXPHYS 2231 2232 if (devvp == NULL) 2233 panic("devvp is NULL"); 2234 for (bpp = sp->bpp, i = nblocks; i;) { 2235 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2236 cl = cbp->b_private; 2237 2238 cbp->b_flags |= B_ASYNC; 2239 cbp->b_cflags |= BC_BUSY; 2240 cbp->b_bcount = 0; 2241 2242 #if defined(DEBUG) && defined(DIAGNOSTIC) 2243 if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs)) 2244 / sizeof(int32_t)) { 2245 panic("lfs_writeseg: real bpp overwrite"); 2246 } 2247 if (bpp - sp->bpp > lfs_segsize(fs) / fs->lfs_fsize) { 2248 panic("lfs_writeseg: theoretical bpp overwrite"); 2249 } 2250 #endif 2251 2252 /* 2253 * Construct the cluster. 2254 */ 2255 mutex_enter(&lfs_lock); 2256 ++fs->lfs_iocount; 2257 mutex_exit(&lfs_lock); 2258 while (i && cbp->b_bcount < CHUNKSIZE) { 2259 bp = *bpp; 2260 2261 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2262 break; 2263 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2264 break; 2265 2266 /* Clusters from GOP_WRITE are expedited */ 2267 if (bp->b_bcount > fs->lfs_bsize) { 2268 if (cbp->b_bcount > 0) 2269 /* Put in its own buffer */ 2270 break; 2271 else { 2272 cbp->b_data = bp->b_data; 2273 } 2274 } else if (cbp->b_bcount == 0) { 2275 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2276 LFS_NB_CLUSTER); 2277 cl->flags |= LFS_CL_MALLOC; 2278 } 2279 #ifdef DIAGNOSTIC 2280 if (lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno + 2281 btodb(bp->b_bcount - 1))) != 2282 sp->seg_number) { 2283 printf("blk size %d daddr %" PRIx64 2284 " not in seg %d\n", 2285 bp->b_bcount, bp->b_blkno, 2286 sp->seg_number); 2287 panic("segment overwrite"); 2288 } 2289 #endif 2290 2291 #ifdef LFS_USE_B_INVAL 2292 /* 2293 * Fake buffers from the cleaner are marked as B_INVAL. 2294 * We need to copy the data from user space rather than 2295 * from the buffer indicated. 2296 * XXX == what do I do on an error? 2297 */ 2298 if ((bp->b_cflags & BC_INVAL) != 0 && 2299 bp->b_iodone != NULL) { 2300 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2301 panic("lfs_writeseg: " 2302 "copyin failed [2]"); 2303 } else 2304 #endif /* LFS_USE_B_INVAL */ 2305 if (cl->flags & LFS_CL_MALLOC) { 2306 /* copy data into our cluster. */ 2307 memcpy(p, bp->b_data, bp->b_bcount); 2308 p += bp->b_bcount; 2309 } 2310 2311 cbp->b_bcount += bp->b_bcount; 2312 cl->bufsize += bp->b_bcount; 2313 2314 bp->b_flags &= ~B_READ; 2315 bp->b_error = 0; 2316 cl->bpp[cl->bufcount++] = bp; 2317 2318 vp = bp->b_vp; 2319 mutex_enter(&bufcache_lock); 2320 mutex_enter(vp->v_interlock); 2321 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2322 reassignbuf(bp, vp); 2323 vp->v_numoutput++; 2324 mutex_exit(vp->v_interlock); 2325 mutex_exit(&bufcache_lock); 2326 2327 bpp++; 2328 i--; 2329 } 2330 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2331 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2332 else 2333 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2334 mutex_enter(devvp->v_interlock); 2335 devvp->v_numoutput++; 2336 mutex_exit(devvp->v_interlock); 2337 VOP_STRATEGY(devvp, cbp); 2338 curlwp->l_ru.ru_oublock++; 2339 } 2340 2341 if (lfs_dostats) { 2342 ++lfs_stats.psegwrites; 2343 lfs_stats.blocktot += nblocks - 1; 2344 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2345 ++lfs_stats.psyncwrites; 2346 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2347 ++lfs_stats.pcleanwrites; 2348 lfs_stats.cleanblocks += nblocks - 1; 2349 } 2350 } 2351 2352 return (lfs_initseg(fs) || do_again); 2353 } 2354 2355 void 2356 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2357 { 2358 struct buf *bp; 2359 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2360 int s; 2361 2362 ASSERT_MAYBE_SEGLOCK(fs); 2363 #ifdef DIAGNOSTIC 2364 KASSERT(fs->lfs_magic == LFS_MAGIC); 2365 #endif 2366 /* 2367 * If we can write one superblock while another is in 2368 * progress, we risk not having a complete checkpoint if we crash. 2369 * So, block here if a superblock write is in progress. 2370 */ 2371 mutex_enter(&lfs_lock); 2372 s = splbio(); 2373 while (fs->lfs_sbactive) { 2374 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2375 &lfs_lock); 2376 } 2377 fs->lfs_sbactive = daddr; 2378 splx(s); 2379 mutex_exit(&lfs_lock); 2380 2381 /* Set timestamp of this version of the superblock */ 2382 if (fs->lfs_version == 1) 2383 fs->lfs_otstamp = time_second; 2384 fs->lfs_tstamp = time_second; 2385 2386 /* Checksum the superblock and copy it into a buffer. */ 2387 fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs)); 2388 bp = lfs_newbuf(fs, devvp, 2389 LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2390 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2391 LFS_SBPAD - sizeof(struct dlfs)); 2392 *(struct dlfs *)bp->b_data = fs->lfs_dlfs; 2393 2394 bp->b_cflags |= BC_BUSY; 2395 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2396 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2397 bp->b_error = 0; 2398 bp->b_iodone = lfs_supercallback; 2399 2400 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2401 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2402 else 2403 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2404 curlwp->l_ru.ru_oublock++; 2405 2406 mutex_enter(devvp->v_interlock); 2407 devvp->v_numoutput++; 2408 mutex_exit(devvp->v_interlock); 2409 2410 mutex_enter(&lfs_lock); 2411 ++fs->lfs_iocount; 2412 mutex_exit(&lfs_lock); 2413 VOP_STRATEGY(devvp, bp); 2414 } 2415 2416 /* 2417 * Logical block number match routines used when traversing the dirty block 2418 * chain. 2419 */ 2420 int 2421 lfs_match_fake(struct lfs *fs, struct buf *bp) 2422 { 2423 2424 ASSERT_SEGLOCK(fs); 2425 return LFS_IS_MALLOC_BUF(bp); 2426 } 2427 2428 #if 0 2429 int 2430 lfs_match_real(struct lfs *fs, struct buf *bp) 2431 { 2432 2433 ASSERT_SEGLOCK(fs); 2434 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2435 } 2436 #endif 2437 2438 int 2439 lfs_match_data(struct lfs *fs, struct buf *bp) 2440 { 2441 2442 ASSERT_SEGLOCK(fs); 2443 return (bp->b_lblkno >= 0); 2444 } 2445 2446 int 2447 lfs_match_indir(struct lfs *fs, struct buf *bp) 2448 { 2449 daddr_t lbn; 2450 2451 ASSERT_SEGLOCK(fs); 2452 lbn = bp->b_lblkno; 2453 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0); 2454 } 2455 2456 int 2457 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2458 { 2459 daddr_t lbn; 2460 2461 ASSERT_SEGLOCK(fs); 2462 lbn = bp->b_lblkno; 2463 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1); 2464 } 2465 2466 int 2467 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2468 { 2469 daddr_t lbn; 2470 2471 ASSERT_SEGLOCK(fs); 2472 lbn = bp->b_lblkno; 2473 return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2); 2474 } 2475 2476 static void 2477 lfs_free_aiodone(struct buf *bp) 2478 { 2479 struct lfs *fs; 2480 2481 KERNEL_LOCK(1, curlwp); 2482 fs = bp->b_private; 2483 ASSERT_NO_SEGLOCK(fs); 2484 lfs_freebuf(fs, bp); 2485 KERNEL_UNLOCK_LAST(curlwp); 2486 } 2487 2488 static void 2489 lfs_super_aiodone(struct buf *bp) 2490 { 2491 struct lfs *fs; 2492 2493 KERNEL_LOCK(1, curlwp); 2494 fs = bp->b_private; 2495 ASSERT_NO_SEGLOCK(fs); 2496 mutex_enter(&lfs_lock); 2497 fs->lfs_sbactive = 0; 2498 if (--fs->lfs_iocount <= 1) 2499 wakeup(&fs->lfs_iocount); 2500 wakeup(&fs->lfs_sbactive); 2501 mutex_exit(&lfs_lock); 2502 lfs_freebuf(fs, bp); 2503 KERNEL_UNLOCK_LAST(curlwp); 2504 } 2505 2506 static void 2507 lfs_cluster_aiodone(struct buf *bp) 2508 { 2509 struct lfs_cluster *cl; 2510 struct lfs *fs; 2511 struct buf *tbp, *fbp; 2512 struct vnode *vp, *devvp, *ovp; 2513 struct inode *ip; 2514 int error; 2515 2516 KERNEL_LOCK(1, curlwp); 2517 2518 error = bp->b_error; 2519 cl = bp->b_private; 2520 fs = cl->fs; 2521 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2522 ASSERT_NO_SEGLOCK(fs); 2523 2524 /* Put the pages back, and release the buffer */ 2525 while (cl->bufcount--) { 2526 tbp = cl->bpp[cl->bufcount]; 2527 KASSERT(tbp->b_cflags & BC_BUSY); 2528 if (error) { 2529 tbp->b_error = error; 2530 } 2531 2532 /* 2533 * We're done with tbp. If it has not been re-dirtied since 2534 * the cluster was written, free it. Otherwise, keep it on 2535 * the locked list to be written again. 2536 */ 2537 vp = tbp->b_vp; 2538 2539 tbp->b_flags &= ~B_GATHERED; 2540 2541 LFS_BCLEAN_LOG(fs, tbp); 2542 2543 mutex_enter(&bufcache_lock); 2544 if (tbp->b_iodone == NULL) { 2545 KASSERT(tbp->b_flags & B_LOCKED); 2546 bremfree(tbp); 2547 if (vp) { 2548 mutex_enter(vp->v_interlock); 2549 reassignbuf(tbp, vp); 2550 mutex_exit(vp->v_interlock); 2551 } 2552 tbp->b_flags |= B_ASYNC; /* for biodone */ 2553 } 2554 2555 if (((tbp->b_flags | tbp->b_oflags) & 2556 (B_LOCKED | BO_DELWRI)) == B_LOCKED) 2557 LFS_UNLOCK_BUF(tbp); 2558 2559 if (tbp->b_oflags & BO_DONE) { 2560 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2561 cl->bufcount, (long)tbp->b_flags)); 2562 } 2563 2564 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2565 /* 2566 * A buffer from the page daemon. 2567 * We use the same iodone as it does, 2568 * so we must manually disassociate its 2569 * buffers from the vp. 2570 */ 2571 if ((ovp = tbp->b_vp) != NULL) { 2572 /* This is just silly */ 2573 mutex_enter(ovp->v_interlock); 2574 brelvp(tbp); 2575 mutex_exit(ovp->v_interlock); 2576 tbp->b_vp = vp; 2577 tbp->b_objlock = vp->v_interlock; 2578 } 2579 /* Put it back the way it was */ 2580 tbp->b_flags |= B_ASYNC; 2581 /* Master buffers have BC_AGE */ 2582 if (tbp->b_private == tbp) 2583 tbp->b_cflags |= BC_AGE; 2584 } 2585 mutex_exit(&bufcache_lock); 2586 2587 biodone(tbp); 2588 2589 /* 2590 * If this is the last block for this vnode, but 2591 * there are other blocks on its dirty list, 2592 * set IN_MODIFIED/IN_CLEANING depending on what 2593 * sort of block. Only do this for our mount point, 2594 * not for, e.g., inode blocks that are attached to 2595 * the devvp. 2596 * XXX KS - Shouldn't we set *both* if both types 2597 * of blocks are present (traverse the dirty list?) 2598 */ 2599 mutex_enter(vp->v_interlock); 2600 mutex_enter(&lfs_lock); 2601 if (vp != devvp && vp->v_numoutput == 0 && 2602 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2603 ip = VTOI(vp); 2604 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2605 ip->i_number)); 2606 if (LFS_IS_MALLOC_BUF(fbp)) 2607 LFS_SET_UINO(ip, IN_CLEANING); 2608 else 2609 LFS_SET_UINO(ip, IN_MODIFIED); 2610 } 2611 cv_broadcast(&vp->v_cv); 2612 mutex_exit(&lfs_lock); 2613 mutex_exit(vp->v_interlock); 2614 } 2615 2616 /* Fix up the cluster buffer, and release it */ 2617 if (cl->flags & LFS_CL_MALLOC) 2618 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2619 putiobuf(bp); 2620 2621 /* Note i/o done */ 2622 if (cl->flags & LFS_CL_SYNC) { 2623 if (--cl->seg->seg_iocount == 0) 2624 wakeup(&cl->seg->seg_iocount); 2625 } 2626 mutex_enter(&lfs_lock); 2627 #ifdef DIAGNOSTIC 2628 if (fs->lfs_iocount == 0) 2629 panic("lfs_cluster_aiodone: zero iocount"); 2630 #endif 2631 if (--fs->lfs_iocount <= 1) 2632 wakeup(&fs->lfs_iocount); 2633 mutex_exit(&lfs_lock); 2634 2635 KERNEL_UNLOCK_LAST(curlwp); 2636 2637 pool_put(&fs->lfs_bpppool, cl->bpp); 2638 cl->bpp = NULL; 2639 pool_put(&fs->lfs_clpool, cl); 2640 } 2641 2642 static void 2643 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2644 { 2645 /* reset b_iodone for when this is a single-buf i/o. */ 2646 bp->b_iodone = aiodone; 2647 2648 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2649 } 2650 2651 static void 2652 lfs_cluster_callback(struct buf *bp) 2653 { 2654 2655 lfs_generic_callback(bp, lfs_cluster_aiodone); 2656 } 2657 2658 void 2659 lfs_supercallback(struct buf *bp) 2660 { 2661 2662 lfs_generic_callback(bp, lfs_super_aiodone); 2663 } 2664 2665 /* 2666 * The only buffers that are going to hit these functions are the 2667 * segment write blocks, or the segment summaries, or the superblocks. 2668 * 2669 * All of the above are created by lfs_newbuf, and so do not need to be 2670 * released via brelse. 2671 */ 2672 void 2673 lfs_callback(struct buf *bp) 2674 { 2675 2676 lfs_generic_callback(bp, lfs_free_aiodone); 2677 } 2678 2679 /* 2680 * Shellsort (diminishing increment sort) from Data Structures and 2681 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2682 * see also Knuth Vol. 3, page 84. The increments are selected from 2683 * formula (8), page 95. Roughly O(N^3/2). 2684 */ 2685 /* 2686 * This is our own private copy of shellsort because we want to sort 2687 * two parallel arrays (the array of buffer pointers and the array of 2688 * logical block numbers) simultaneously. Note that we cast the array 2689 * of logical block numbers to a unsigned in this routine so that the 2690 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2691 */ 2692 2693 void 2694 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size) 2695 { 2696 static int __rsshell_increments[] = { 4, 1, 0 }; 2697 int incr, *incrp, t1, t2; 2698 struct buf *bp_temp; 2699 2700 #ifdef DEBUG 2701 incr = 0; 2702 for (t1 = 0; t1 < nmemb; t1++) { 2703 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2704 if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) { 2705 /* dump before panic */ 2706 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2707 nmemb, size); 2708 incr = 0; 2709 for (t1 = 0; t1 < nmemb; t1++) { 2710 const struct buf *bp = bp_array[t1]; 2711 2712 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2713 PRIu64 "\n", t1, 2714 (uint64_t)bp->b_bcount, 2715 (uint64_t)bp->b_lblkno); 2716 printf("lbns:"); 2717 for (t2 = 0; t2 * size < bp->b_bcount; 2718 t2++) { 2719 printf(" %" PRId32, 2720 lb_array[incr++]); 2721 } 2722 printf("\n"); 2723 } 2724 panic("lfs_shellsort: inconsistent input"); 2725 } 2726 } 2727 } 2728 #endif 2729 2730 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2731 for (t1 = incr; t1 < nmemb; ++t1) 2732 for (t2 = t1 - incr; t2 >= 0;) 2733 if ((u_int32_t)bp_array[t2]->b_lblkno > 2734 (u_int32_t)bp_array[t2 + incr]->b_lblkno) { 2735 bp_temp = bp_array[t2]; 2736 bp_array[t2] = bp_array[t2 + incr]; 2737 bp_array[t2 + incr] = bp_temp; 2738 t2 -= incr; 2739 } else 2740 break; 2741 2742 /* Reform the list of logical blocks */ 2743 incr = 0; 2744 for (t1 = 0; t1 < nmemb; t1++) { 2745 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2746 lb_array[incr++] = bp_array[t1]->b_lblkno + t2; 2747 } 2748 } 2749 } 2750 2751 /* 2752 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2753 * point at uninitialized space. 2754 */ 2755 void 2756 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2757 { 2758 struct segment *sp = fs->lfs_sp; 2759 2760 KASSERT(vers > 0); 2761 2762 if (sp->seg_bytes_left < fs->lfs_bsize || 2763 sp->sum_bytes_left < sizeof(struct finfo)) 2764 (void) lfs_writeseg(fs, fs->lfs_sp); 2765 2766 sp->sum_bytes_left -= FINFOSIZE; 2767 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 2768 sp->fip->fi_nblocks = 0; 2769 sp->fip->fi_ino = ino; 2770 sp->fip->fi_version = vers; 2771 } 2772 2773 /* 2774 * Release the FINFO entry, either clearing out an unused entry or 2775 * advancing us to the next available entry. 2776 */ 2777 void 2778 lfs_release_finfo(struct lfs *fs) 2779 { 2780 struct segment *sp = fs->lfs_sp; 2781 2782 if (sp->fip->fi_nblocks != 0) { 2783 sp->fip = (FINFO*)((char *)sp->fip + FINFOSIZE + 2784 sizeof(int32_t) * sp->fip->fi_nblocks); 2785 sp->start_lbp = &sp->fip->fi_blocks[0]; 2786 } else { 2787 sp->sum_bytes_left += FINFOSIZE; 2788 --((SEGSUM *)(sp->segsum))->ss_nfinfo; 2789 } 2790 } 2791