1/* $NetBSD$ */ 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, "$NetBSD$"); 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/ufs/inode.h> 99#include <ufs/ufs/dir.h> 100#include <ufs/ufs/ufsmount.h> 101#include <ufs/ufs/ufs_extern.h> 102 103#include <ufs/lfs/lfs.h> 104#include <ufs/lfs/lfs_extern.h> 105 106#include <uvm/uvm.h> 107#include <uvm/uvm_extern.h> 108 109MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 110 111static void lfs_generic_callback(struct buf *, void (*)(struct buf *)); 112static void lfs_free_aiodone(struct buf *); 113static void lfs_super_aiodone(struct buf *); 114static void lfs_cluster_aiodone(struct buf *); 115static void lfs_cluster_callback(struct buf *); 116 117/* 118 * Determine if it's OK to start a partial in this segment, or if we need 119 * to go on to a new segment. 120 */ 121#define LFS_PARTIAL_FITS(fs) \ 122 ((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \ 123 (fs)->lfs_frag) 124 125/* 126 * Figure out whether we should do a checkpoint write or go ahead with 127 * an ordinary write. 128 */ 129#define LFS_SHOULD_CHECKPOINT(fs, flags) \ 130 ((flags & SEGM_CLEAN) == 0 && \ 131 ((fs->lfs_nactive > LFS_MAX_ACTIVE || \ 132 (flags & SEGM_CKP) || \ 133 fs->lfs_nclean < LFS_MAX_ACTIVE))) 134 135int lfs_match_fake(struct lfs *, struct buf *); 136void lfs_newseg(struct lfs *); 137/* XXX ondisk32 */ 138void lfs_shellsort(struct buf **, int32_t *, int, int); 139void lfs_supercallback(struct buf *); 140void lfs_updatemeta(struct segment *); 141void lfs_writesuper(struct lfs *, daddr_t); 142int lfs_writevnodes(struct lfs *fs, struct mount *mp, 143 struct segment *sp, int dirops); 144 145int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 146int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 147int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 148int lfs_dirvcount = 0; /* # active dirops */ 149 150/* Statistics Counters */ 151int lfs_dostats = 1; 152struct lfs_stats lfs_stats; 153 154/* op values to lfs_writevnodes */ 155#define VN_REG 0 156#define VN_DIROP 1 157#define VN_EMPTY 2 158#define VN_CLEAN 3 159 160/* 161 * XXX KS - Set modification time on the Ifile, so the cleaner can 162 * read the fs mod time off of it. We don't set IN_UPDATE here, 163 * since we don't really need this to be flushed to disk (and in any 164 * case that wouldn't happen to the Ifile until we checkpoint). 165 */ 166void 167lfs_imtime(struct lfs *fs) 168{ 169 struct timespec ts; 170 struct inode *ip; 171 172 ASSERT_MAYBE_SEGLOCK(fs); 173 vfs_timestamp(&ts); 174 ip = VTOI(fs->lfs_ivnode); 175 ip->i_ffs1_mtime = ts.tv_sec; 176 ip->i_ffs1_mtimensec = ts.tv_nsec; 177} 178 179/* 180 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 181 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 182 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 183 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 184 */ 185 186#define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 187 188int 189lfs_vflush(struct vnode *vp) 190{ 191 struct inode *ip; 192 struct lfs *fs; 193 struct segment *sp; 194 struct buf *bp, *nbp, *tbp, *tnbp; 195 int error; 196 int flushed; 197 int relock; 198 int loopcount; 199 200 ip = VTOI(vp); 201 fs = VFSTOUFS(vp->v_mount)->um_lfs; 202 relock = 0; 203 204 top: 205 KASSERT(mutex_owned(vp->v_interlock) == false); 206 KASSERT(mutex_owned(&lfs_lock) == false); 207 KASSERT(mutex_owned(&bufcache_lock) == false); 208 ASSERT_NO_SEGLOCK(fs); 209 if (ip->i_flag & IN_CLEANING) { 210 ivndebug(vp,"vflush/in_cleaning"); 211 mutex_enter(&lfs_lock); 212 LFS_CLR_UINO(ip, IN_CLEANING); 213 LFS_SET_UINO(ip, IN_MODIFIED); 214 mutex_exit(&lfs_lock); 215 216 /* 217 * Toss any cleaning buffers that have real counterparts 218 * to avoid losing new data. 219 */ 220 mutex_enter(vp->v_interlock); 221 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 222 nbp = LIST_NEXT(bp, b_vnbufs); 223 if (!LFS_IS_MALLOC_BUF(bp)) 224 continue; 225 /* 226 * Look for pages matching the range covered 227 * by cleaning blocks. It's okay if more dirty 228 * pages appear, so long as none disappear out 229 * from under us. 230 */ 231 if (bp->b_lblkno > 0 && vp->v_type == VREG && 232 vp != fs->lfs_ivnode) { 233 struct vm_page *pg; 234 voff_t off; 235 236 for (off = lblktosize(fs, bp->b_lblkno); 237 off < lblktosize(fs, bp->b_lblkno + 1); 238 off += PAGE_SIZE) { 239 pg = uvm_pagelookup(&vp->v_uobj, off); 240 if (pg == NULL) 241 continue; 242 if ((pg->flags & PG_CLEAN) == 0 || 243 pmap_is_modified(pg)) { 244 fs->lfs_avail += btofsb(fs, 245 bp->b_bcount); 246 wakeup(&fs->lfs_avail); 247 mutex_exit(vp->v_interlock); 248 lfs_freebuf(fs, bp); 249 mutex_enter(vp->v_interlock); 250 bp = NULL; 251 break; 252 } 253 } 254 } 255 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 256 tbp = tnbp) 257 { 258 tnbp = LIST_NEXT(tbp, b_vnbufs); 259 if (tbp->b_vp == bp->b_vp 260 && tbp->b_lblkno == bp->b_lblkno 261 && tbp != bp) 262 { 263 fs->lfs_avail += btofsb(fs, 264 bp->b_bcount); 265 wakeup(&fs->lfs_avail); 266 mutex_exit(vp->v_interlock); 267 lfs_freebuf(fs, bp); 268 mutex_enter(vp->v_interlock); 269 bp = NULL; 270 break; 271 } 272 } 273 } 274 } else { 275 mutex_enter(vp->v_interlock); 276 } 277 278 /* If the node is being written, wait until that is done */ 279 while (WRITEINPROG(vp)) { 280 ivndebug(vp,"vflush/writeinprog"); 281 cv_wait(&vp->v_cv, vp->v_interlock); 282 } 283 mutex_exit(vp->v_interlock); 284 285 /* Protect against VI_XLOCK deadlock in vinvalbuf() */ 286 lfs_seglock(fs, SEGM_SYNC | ((vp->v_iflag & VI_XLOCK) ? SEGM_RECLAIM : 0)); 287 if (vp->v_iflag & VI_XLOCK) { 288 fs->lfs_reclino = ip->i_number; 289 } 290 291 /* If we're supposed to flush a freed inode, just toss it */ 292 if (ip->i_lfs_iflags & LFSI_DELETED) { 293 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 294 ip->i_number)); 295 /* Drain v_numoutput */ 296 mutex_enter(vp->v_interlock); 297 while (vp->v_numoutput > 0) { 298 cv_wait(&vp->v_cv, vp->v_interlock); 299 } 300 KASSERT(vp->v_numoutput == 0); 301 mutex_exit(vp->v_interlock); 302 303 mutex_enter(&bufcache_lock); 304 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 305 nbp = LIST_NEXT(bp, b_vnbufs); 306 307 KASSERT((bp->b_flags & B_GATHERED) == 0); 308 if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */ 309 fs->lfs_avail += btofsb(fs, bp->b_bcount); 310 wakeup(&fs->lfs_avail); 311 } 312 /* Copied from lfs_writeseg */ 313 if (bp->b_iodone != NULL) { 314 mutex_exit(&bufcache_lock); 315 biodone(bp); 316 mutex_enter(&bufcache_lock); 317 } else { 318 bremfree(bp); 319 LFS_UNLOCK_BUF(bp); 320 mutex_enter(vp->v_interlock); 321 bp->b_flags &= ~(B_READ | B_GATHERED); 322 bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE; 323 bp->b_error = 0; 324 reassignbuf(bp, vp); 325 mutex_exit(vp->v_interlock); 326 brelse(bp, 0); 327 } 328 } 329 mutex_exit(&bufcache_lock); 330 LFS_CLR_UINO(ip, IN_CLEANING); 331 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 332 ip->i_flag &= ~IN_ALLMOD; 333 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 334 ip->i_number)); 335 lfs_segunlock(fs); 336 337 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 338 339 return 0; 340 } 341 342 fs->lfs_flushvp = vp; 343 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 344 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 345 fs->lfs_flushvp = NULL; 346 KASSERT(fs->lfs_flushvp_fakevref == 0); 347 lfs_segunlock(fs); 348 349 /* Make sure that any pending buffers get written */ 350 mutex_enter(vp->v_interlock); 351 while (vp->v_numoutput > 0) { 352 cv_wait(&vp->v_cv, vp->v_interlock); 353 } 354 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 355 KASSERT(vp->v_numoutput == 0); 356 mutex_exit(vp->v_interlock); 357 358 return error; 359 } 360 sp = fs->lfs_sp; 361 362 flushed = 0; 363 if (VPISEMPTY(vp)) { 364 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 365 ++flushed; 366 } else if ((ip->i_flag & IN_CLEANING) && 367 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 368 ivndebug(vp,"vflush/clean"); 369 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 370 ++flushed; 371 } else if (lfs_dostats) { 372 if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD)) 373 ++lfs_stats.vflush_invoked; 374 ivndebug(vp,"vflush"); 375 } 376 377#ifdef DIAGNOSTIC 378 if (vp->v_uflag & VU_DIROP) { 379 DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n")); 380 /* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */ 381 } 382#endif 383 384 do { 385 loopcount = 0; 386 do { 387 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 388 relock = lfs_writefile(fs, sp, vp); 389 if (relock && vp != fs->lfs_ivnode) { 390 /* 391 * Might have to wait for the 392 * cleaner to run; but we're 393 * still not done with this vnode. 394 * XXX we can do better than this. 395 */ 396 KDASSERT(ip->i_number != LFS_IFILE_INUM); 397 lfs_writeinode(fs, sp, ip); 398 mutex_enter(&lfs_lock); 399 LFS_SET_UINO(ip, IN_MODIFIED); 400 mutex_exit(&lfs_lock); 401 lfs_writeseg(fs, sp); 402 lfs_segunlock(fs); 403 lfs_segunlock_relock(fs); 404 goto top; 405 } 406 } 407 /* 408 * If we begin a new segment in the middle of writing 409 * the Ifile, it creates an inconsistent checkpoint, 410 * since the Ifile information for the new segment 411 * is not up-to-date. Take care of this here by 412 * sending the Ifile through again in case there 413 * are newly dirtied blocks. But wait, there's more! 414 * This second Ifile write could *also* cross a segment 415 * boundary, if the first one was large. The second 416 * one is guaranteed to be no more than 8 blocks, 417 * though (two segment blocks and supporting indirects) 418 * so the third write *will not* cross the boundary. 419 */ 420 if (vp == fs->lfs_ivnode) { 421 lfs_writefile(fs, sp, vp); 422 lfs_writefile(fs, sp, vp); 423 } 424#ifdef DEBUG 425 if (++loopcount > 2) 426 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 427#endif 428 } while (lfs_writeinode(fs, sp, ip)); 429 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 430 431 if (lfs_dostats) { 432 ++lfs_stats.nwrites; 433 if (sp->seg_flags & SEGM_SYNC) 434 ++lfs_stats.nsync_writes; 435 if (sp->seg_flags & SEGM_CKP) 436 ++lfs_stats.ncheckpoints; 437 } 438 /* 439 * If we were called from somewhere that has already held the seglock 440 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 441 * the write to complete because we are still locked. 442 * Since lfs_vflush() must return the vnode with no dirty buffers, 443 * we must explicitly wait, if that is the case. 444 * 445 * We compare the iocount against 1, not 0, because it is 446 * artificially incremented by lfs_seglock(). 447 */ 448 mutex_enter(&lfs_lock); 449 if (fs->lfs_seglock > 1) { 450 while (fs->lfs_iocount > 1) 451 (void)mtsleep(&fs->lfs_iocount, PRIBIO + 1, 452 "lfs_vflush", 0, &lfs_lock); 453 } 454 mutex_exit(&lfs_lock); 455 456 lfs_segunlock(fs); 457 458 /* Wait for these buffers to be recovered by aiodoned */ 459 mutex_enter(vp->v_interlock); 460 while (vp->v_numoutput > 0) { 461 cv_wait(&vp->v_cv, vp->v_interlock); 462 } 463 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 464 KASSERT(vp->v_numoutput == 0); 465 mutex_exit(vp->v_interlock); 466 467 fs->lfs_flushvp = NULL; 468 KASSERT(fs->lfs_flushvp_fakevref == 0); 469 470 return (0); 471} 472 473int 474lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 475{ 476 struct inode *ip; 477 struct vnode *vp; 478 int inodes_written = 0, only_cleaning; 479 int error = 0; 480 481 ASSERT_SEGLOCK(fs); 482 loop: 483 /* start at last (newest) vnode. */ 484 mutex_enter(&mntvnode_lock); 485 TAILQ_FOREACH_REVERSE(vp, &mp->mnt_vnodelist, vnodelst, v_mntvnodes) { 486 /* 487 * If the vnode that we are about to sync is no longer 488 * associated with this mount point, start over. 489 */ 490 if (vp->v_mount != mp) { 491 DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n")); 492 /* 493 * After this, pages might be busy 494 * due to our own previous putpages. 495 * Start actual segment write here to avoid deadlock. 496 * If we were just writing one segment and we've done 497 * that, break out. 498 */ 499 mutex_exit(&mntvnode_lock); 500 if (lfs_writeseg(fs, sp) && 501 (sp->seg_flags & SEGM_SINGLE) && 502 fs->lfs_curseg != fs->lfs_startseg) { 503 DLOG((DLOG_VNODE, "lfs_writevnodes: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset)); 504 break; 505 } 506 goto loop; 507 } 508 509 mutex_enter(vp->v_interlock); 510 if (vp->v_type == VNON || vismarker(vp) || 511 (vp->v_iflag & VI_CLEAN) != 0) { 512 mutex_exit(vp->v_interlock); 513 continue; 514 } 515 516 ip = VTOI(vp); 517 if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) || 518 (op != VN_DIROP && op != VN_CLEAN && 519 (vp->v_uflag & VU_DIROP))) { 520 mutex_exit(vp->v_interlock); 521 vndebug(vp,"dirop"); 522 continue; 523 } 524 525 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 526 mutex_exit(vp->v_interlock); 527 vndebug(vp,"empty"); 528 continue; 529 } 530 531 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM 532 && vp != fs->lfs_flushvp 533 && !(ip->i_flag & IN_CLEANING)) { 534 mutex_exit(vp->v_interlock); 535 vndebug(vp,"cleaning"); 536 continue; 537 } 538 539 mutex_exit(&mntvnode_lock); 540 if (lfs_vref(vp)) { 541 vndebug(vp,"vref"); 542 mutex_enter(&mntvnode_lock); 543 continue; 544 } 545 546 only_cleaning = 0; 547 /* 548 * Write the inode/file if dirty and it's not the IFILE. 549 */ 550 if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) { 551 only_cleaning = 552 ((ip->i_flag & IN_ALLMOD) == IN_CLEANING); 553 554 if (ip->i_number != LFS_IFILE_INUM) { 555 error = lfs_writefile(fs, sp, vp); 556 if (error) { 557 lfs_vunref(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 if (lfs_clean_vnhead && only_cleaning) 598 lfs_vunref_head(vp); 599 else 600 lfs_vunref(vp); 601 602 mutex_enter(&mntvnode_lock); 603 } 604 mutex_exit(&mntvnode_lock); 605 return error; 606} 607 608/* 609 * Do a checkpoint. 610 */ 611int 612lfs_segwrite(struct mount *mp, int flags) 613{ 614 struct buf *bp; 615 struct inode *ip; 616 struct lfs *fs; 617 struct segment *sp; 618 struct vnode *vp; 619 SEGUSE *segusep; 620 int do_ckp, did_ckp, error; 621 unsigned n, segleft, maxseg, sn, i, curseg; 622 int writer_set = 0; 623 int dirty; 624 int redo; 625 int um_error; 626 int loopcount; 627 628 fs = VFSTOUFS(mp)->um_lfs; 629 ASSERT_MAYBE_SEGLOCK(fs); 630 631 if (fs->lfs_ronly) 632 return EROFS; 633 634 lfs_imtime(fs); 635 636 /* 637 * Allocate a segment structure and enough space to hold pointers to 638 * the maximum possible number of buffers which can be described in a 639 * single summary block. 640 */ 641 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 642 643 /* We can't do a partial write and checkpoint at the same time. */ 644 if (do_ckp) 645 flags &= ~SEGM_SINGLE; 646 647 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 648 sp = fs->lfs_sp; 649 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 650 do_ckp = 1; 651 652 /* 653 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 654 * in which case we have to flush *all* buffers off of this vnode. 655 * We don't care about other nodes, but write any non-dirop nodes 656 * anyway in anticipation of another getnewvnode(). 657 * 658 * If we're cleaning we only write cleaning and ifile blocks, and 659 * no dirops, since otherwise we'd risk corruption in a crash. 660 */ 661 if (sp->seg_flags & SEGM_CLEAN) 662 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 663 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 664 do { 665 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 666 if ((sp->seg_flags & SEGM_SINGLE) && 667 fs->lfs_curseg != fs->lfs_startseg) { 668 DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset)); 669 break; 670 } 671 672 if (do_ckp || fs->lfs_dirops == 0) { 673 if (!writer_set) { 674 lfs_writer_enter(fs, "lfs writer"); 675 writer_set = 1; 676 } 677 error = lfs_writevnodes(fs, mp, sp, VN_DIROP); 678 if (um_error == 0) 679 um_error = error; 680 /* In case writevnodes errored out */ 681 lfs_flush_dirops(fs); 682 ((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT); 683 lfs_finalize_fs_seguse(fs); 684 } 685 if (do_ckp && um_error) { 686 lfs_segunlock_relock(fs); 687 sp = fs->lfs_sp; 688 } 689 } while (do_ckp && um_error != 0); 690 } 691 692 /* 693 * If we are doing a checkpoint, mark everything since the 694 * last checkpoint as no longer ACTIVE. 695 */ 696 if (do_ckp || fs->lfs_doifile) { 697 segleft = fs->lfs_nseg; 698 curseg = 0; 699 for (n = 0; n < fs->lfs_segtabsz; n++) { 700 dirty = 0; 701 if (bread(fs->lfs_ivnode, fs->lfs_cleansz + n, 702 fs->lfs_bsize, NOCRED, B_MODIFY, &bp)) 703 panic("lfs_segwrite: ifile read"); 704 segusep = (SEGUSE *)bp->b_data; 705 maxseg = min(segleft, fs->lfs_sepb); 706 for (i = 0; i < maxseg; i++) { 707 sn = curseg + i; 708 if (sn != dtosn(fs, fs->lfs_curseg) && 709 segusep->su_flags & SEGUSE_ACTIVE) { 710 segusep->su_flags &= ~SEGUSE_ACTIVE; 711 --fs->lfs_nactive; 712 ++dirty; 713 } 714 fs->lfs_suflags[fs->lfs_activesb][sn] = 715 segusep->su_flags; 716 if (fs->lfs_version > 1) 717 ++segusep; 718 else 719 segusep = (SEGUSE *) 720 ((SEGUSE_V1 *)segusep + 1); 721 } 722 723 if (dirty) 724 error = LFS_BWRITE_LOG(bp); /* Ifile */ 725 else 726 brelse(bp, 0); 727 segleft -= fs->lfs_sepb; 728 curseg += fs->lfs_sepb; 729 } 730 } 731 732 KASSERT(LFS_SEGLOCK_HELD(fs)); 733 734 did_ckp = 0; 735 if (do_ckp || fs->lfs_doifile) { 736 vp = fs->lfs_ivnode; 737 loopcount = 0; 738 do { 739#ifdef DEBUG 740 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 741#endif 742 mutex_enter(&lfs_lock); 743 fs->lfs_flags &= ~LFS_IFDIRTY; 744 mutex_exit(&lfs_lock); 745 746 ip = VTOI(vp); 747 748 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 749 /* 750 * Ifile has no pages, so we don't need 751 * to check error return here. 752 */ 753 lfs_writefile(fs, sp, vp); 754 /* 755 * Ensure the Ifile takes the current segment 756 * into account. See comment in lfs_vflush. 757 */ 758 lfs_writefile(fs, sp, vp); 759 lfs_writefile(fs, sp, vp); 760 } 761 762 if (ip->i_flag & IN_ALLMOD) 763 ++did_ckp; 764#if 0 765 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 766#else 767 redo = lfs_writeinode(fs, sp, ip); 768#endif 769 redo += lfs_writeseg(fs, sp); 770 mutex_enter(&lfs_lock); 771 redo += (fs->lfs_flags & LFS_IFDIRTY); 772 mutex_exit(&lfs_lock); 773#ifdef DEBUG 774 if (++loopcount > 2) 775 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 776 loopcount); 777#endif 778 } while (redo && do_ckp); 779 780 /* 781 * Unless we are unmounting, the Ifile may continue to have 782 * dirty blocks even after a checkpoint, due to changes to 783 * inodes' atime. If we're checkpointing, it's "impossible" 784 * for other parts of the Ifile to be dirty after the loop 785 * above, since we hold the segment lock. 786 */ 787 mutex_enter(vp->v_interlock); 788 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 789 LFS_CLR_UINO(ip, IN_ALLMOD); 790 } 791#ifdef DIAGNOSTIC 792 else if (do_ckp) { 793 int do_panic = 0; 794 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 795 if (bp->b_lblkno < fs->lfs_cleansz + 796 fs->lfs_segtabsz && 797 !(bp->b_flags & B_GATHERED)) { 798 printf("ifile lbn %ld still dirty (flags %lx)\n", 799 (long)bp->b_lblkno, 800 (long)bp->b_flags); 801 ++do_panic; 802 } 803 } 804 if (do_panic) 805 panic("dirty blocks"); 806 } 807#endif 808 mutex_exit(vp->v_interlock); 809 } else { 810 (void) lfs_writeseg(fs, sp); 811 } 812 813 /* Note Ifile no longer needs to be written */ 814 fs->lfs_doifile = 0; 815 if (writer_set) 816 lfs_writer_leave(fs); 817 818 /* 819 * If we didn't write the Ifile, we didn't really do anything. 820 * That means that (1) there is a checkpoint on disk and (2) 821 * nothing has changed since it was written. 822 * 823 * Take the flags off of the segment so that lfs_segunlock 824 * doesn't have to write the superblock either. 825 */ 826 if (do_ckp && !did_ckp) { 827 sp->seg_flags &= ~SEGM_CKP; 828 } 829 830 if (lfs_dostats) { 831 ++lfs_stats.nwrites; 832 if (sp->seg_flags & SEGM_SYNC) 833 ++lfs_stats.nsync_writes; 834 if (sp->seg_flags & SEGM_CKP) 835 ++lfs_stats.ncheckpoints; 836 } 837 lfs_segunlock(fs); 838 return (0); 839} 840 841/* 842 * Write the dirty blocks associated with a vnode. 843 */ 844int 845lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp) 846{ 847 struct finfo *fip; 848 struct inode *ip; 849 int i, frag; 850 int error; 851 852 ASSERT_SEGLOCK(fs); 853 error = 0; 854 ip = VTOI(vp); 855 856 fip = sp->fip; 857 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 858 859 if (vp->v_uflag & VU_DIROP) 860 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 861 862 if (sp->seg_flags & SEGM_CLEAN) { 863 lfs_gather(fs, sp, vp, lfs_match_fake); 864 /* 865 * For a file being flushed, we need to write *all* blocks. 866 * This means writing the cleaning blocks first, and then 867 * immediately following with any non-cleaning blocks. 868 * The same is true of the Ifile since checkpoints assume 869 * that all valid Ifile blocks are written. 870 */ 871 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 872 lfs_gather(fs, sp, vp, lfs_match_data); 873 /* 874 * Don't call VOP_PUTPAGES: if we're flushing, 875 * we've already done it, and the Ifile doesn't 876 * use the page cache. 877 */ 878 } 879 } else { 880 lfs_gather(fs, sp, vp, lfs_match_data); 881 /* 882 * If we're flushing, we've already called VOP_PUTPAGES 883 * so don't do it again. Otherwise, we want to write 884 * everything we've got. 885 */ 886 if (!IS_FLUSHING(fs, vp)) { 887 mutex_enter(vp->v_interlock); 888 error = VOP_PUTPAGES(vp, 0, 0, 889 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 890 } 891 } 892 893 /* 894 * It may not be necessary to write the meta-data blocks at this point, 895 * as the roll-forward recovery code should be able to reconstruct the 896 * list. 897 * 898 * We have to write them anyway, though, under two conditions: (1) the 899 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 900 * checkpointing. 901 * 902 * BUT if we are cleaning, we might have indirect blocks that refer to 903 * new blocks not being written yet, in addition to fragments being 904 * moved out of a cleaned segment. If that is the case, don't 905 * write the indirect blocks, or the finfo will have a small block 906 * in the middle of it! 907 * XXX in this case isn't the inode size wrong too? 908 */ 909 frag = 0; 910 if (sp->seg_flags & SEGM_CLEAN) { 911 for (i = 0; i < NDADDR; i++) 912 if (ip->i_lfs_fragsize[i] > 0 && 913 ip->i_lfs_fragsize[i] < fs->lfs_bsize) 914 ++frag; 915 } 916#ifdef DIAGNOSTIC 917 if (frag > 1) 918 panic("lfs_writefile: more than one fragment!"); 919#endif 920 if (IS_FLUSHING(fs, vp) || 921 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 922 lfs_gather(fs, sp, vp, lfs_match_indir); 923 lfs_gather(fs, sp, vp, lfs_match_dindir); 924 lfs_gather(fs, sp, vp, lfs_match_tindir); 925 } 926 fip = sp->fip; 927 lfs_release_finfo(fs); 928 929 return error; 930} 931 932/* 933 * Update segment accounting to reflect this inode's change of address. 934 */ 935static int 936lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 937{ 938 struct buf *bp; 939 daddr_t daddr; 940 IFILE *ifp; 941 SEGUSE *sup; 942 ino_t ino; 943 int redo_ifile, error; 944 u_int32_t sn; 945 946 redo_ifile = 0; 947 948 /* 949 * If updating the ifile, update the super-block. Update the disk 950 * address and access times for this inode in the ifile. 951 */ 952 ino = ip->i_number; 953 if (ino == LFS_IFILE_INUM) { 954 daddr = fs->lfs_idaddr; 955 fs->lfs_idaddr = dbtofsb(fs, ndaddr); 956 } else { 957 LFS_IENTRY(ifp, fs, ino, bp); 958 daddr = ifp->if_daddr; 959 ifp->if_daddr = dbtofsb(fs, ndaddr); 960 error = LFS_BWRITE_LOG(bp); /* Ifile */ 961 } 962 963 /* 964 * If this is the Ifile and lfs_offset is set to the first block 965 * in the segment, dirty the new segment's accounting block 966 * (XXX should already be dirty?) and tell the caller to do it again. 967 */ 968 if (ip->i_number == LFS_IFILE_INUM) { 969 sn = dtosn(fs, fs->lfs_offset); 970 if (sntod(fs, sn) + btofsb(fs, fs->lfs_sumsize) == 971 fs->lfs_offset) { 972 LFS_SEGENTRY(sup, fs, sn, bp); 973 KASSERT(bp->b_oflags & BO_DELWRI); 974 LFS_WRITESEGENTRY(sup, fs, sn, bp); 975 /* fs->lfs_flags |= LFS_IFDIRTY; */ 976 redo_ifile |= 1; 977 } 978 } 979 980 /* 981 * The inode's last address should not be in the current partial 982 * segment, except under exceptional circumstances (lfs_writevnodes 983 * had to start over, and in the meantime more blocks were written 984 * to a vnode). Both inodes will be accounted to this segment 985 * in lfs_writeseg so we need to subtract the earlier version 986 * here anyway. The segment count can temporarily dip below 987 * zero here; keep track of how many duplicates we have in 988 * "dupino" so we don't panic below. 989 */ 990 if (daddr >= fs->lfs_lastpseg && daddr <= fs->lfs_offset) { 991 ++sp->ndupino; 992 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 993 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 994 (long long)daddr, sp->ndupino)); 995 } 996 /* 997 * Account the inode: it no longer belongs to its former segment, 998 * though it will not belong to the new segment until that segment 999 * is actually written. 1000 */ 1001 if (daddr != LFS_UNUSED_DADDR) { 1002 u_int32_t oldsn = dtosn(fs, daddr); 1003#ifdef DIAGNOSTIC 1004 int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0; 1005#endif 1006 LFS_SEGENTRY(sup, fs, oldsn, bp); 1007#ifdef DIAGNOSTIC 1008 if (sup->su_nbytes + 1009 sizeof (struct ufs1_dinode) * ndupino 1010 < sizeof (struct ufs1_dinode)) { 1011 printf("lfs_writeinode: negative bytes " 1012 "(segment %" PRIu32 " short by %d, " 1013 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1014 ", daddr=%" PRId64 ", su_nbytes=%u, " 1015 "ndupino=%d)\n", 1016 dtosn(fs, daddr), 1017 (int)sizeof (struct ufs1_dinode) * 1018 (1 - sp->ndupino) - sup->su_nbytes, 1019 oldsn, sp->seg_number, daddr, 1020 (unsigned int)sup->su_nbytes, 1021 sp->ndupino); 1022 panic("lfs_writeinode: negative bytes"); 1023 sup->su_nbytes = sizeof (struct ufs1_dinode); 1024 } 1025#endif 1026 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1027 dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino)); 1028 sup->su_nbytes -= sizeof (struct ufs1_dinode); 1029 redo_ifile |= 1030 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1031 if (redo_ifile) { 1032 mutex_enter(&lfs_lock); 1033 fs->lfs_flags |= LFS_IFDIRTY; 1034 mutex_exit(&lfs_lock); 1035 /* Don't double-account */ 1036 fs->lfs_idaddr = 0x0; 1037 } 1038 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1039 } 1040 1041 return redo_ifile; 1042} 1043 1044int 1045lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1046{ 1047 struct buf *bp; 1048 struct ufs1_dinode *cdp; 1049 struct vnode *vp = ITOV(ip); 1050 daddr_t daddr; 1051 int32_t *daddrp; /* XXX ondisk32 */ 1052 int i, ndx; 1053 int redo_ifile = 0; 1054 int gotblk = 0; 1055 int count; 1056 1057 ASSERT_SEGLOCK(fs); 1058 if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP)) 1059 return (0); 1060 1061 /* Can't write ifile when writer is not set */ 1062 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1063 (sp->seg_flags & SEGM_CLEAN)); 1064 1065 /* 1066 * If this is the Ifile, see if writing it here will generate a 1067 * temporary misaccounting. If it will, do the accounting and write 1068 * the blocks, postponing the inode write until the accounting is 1069 * solid. 1070 */ 1071 count = 0; 1072 while (vp == fs->lfs_ivnode) { 1073 int redo = 0; 1074 1075 if (sp->idp == NULL && sp->ibp == NULL && 1076 (sp->seg_bytes_left < fs->lfs_ibsize || 1077 sp->sum_bytes_left < sizeof(int32_t))) { 1078 (void) lfs_writeseg(fs, sp); 1079 continue; 1080 } 1081 1082 /* Look for dirty Ifile blocks */ 1083 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1084 if (!(bp->b_flags & B_GATHERED)) { 1085 redo = 1; 1086 break; 1087 } 1088 } 1089 1090 if (redo == 0) 1091 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1092 if (redo == 0) 1093 break; 1094 1095 if (sp->idp) { 1096 sp->idp->di_inumber = 0; 1097 sp->idp = NULL; 1098 } 1099 ++count; 1100 if (count > 2) 1101 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1102 lfs_writefile(fs, sp, fs->lfs_ivnode); 1103 } 1104 1105 /* Allocate a new inode block if necessary. */ 1106 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1107 sp->ibp == NULL) { 1108 /* Allocate a new segment if necessary. */ 1109 if (sp->seg_bytes_left < fs->lfs_ibsize || 1110 sp->sum_bytes_left < sizeof(int32_t)) 1111 (void) lfs_writeseg(fs, sp); 1112 1113 /* Get next inode block. */ 1114 daddr = fs->lfs_offset; 1115 fs->lfs_offset += btofsb(fs, fs->lfs_ibsize); 1116 sp->ibp = *sp->cbpp++ = 1117 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1118 fsbtodb(fs, daddr), fs->lfs_ibsize, 0, 0); 1119 gotblk++; 1120 1121 /* Zero out inode numbers */ 1122 for (i = 0; i < INOPB(fs); ++i) 1123 ((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber = 1124 0; 1125 1126 ++sp->start_bpp; 1127 fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize); 1128 /* Set remaining space counters. */ 1129 sp->seg_bytes_left -= fs->lfs_ibsize; 1130 sp->sum_bytes_left -= sizeof(int32_t); 1131 ndx = fs->lfs_sumsize / sizeof(int32_t) - 1132 sp->ninodes / INOPB(fs) - 1; 1133 ((int32_t *)(sp->segsum))[ndx] = daddr; 1134 } 1135 1136 /* Check VU_DIROP in case there is a new file with no data blocks */ 1137 if (vp->v_uflag & VU_DIROP) 1138 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 1139 1140 /* Update the inode times and copy the inode onto the inode page. */ 1141 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1142 if (ip->i_number != LFS_IFILE_INUM) 1143 LFS_ITIMES(ip, NULL, NULL, NULL); 1144 1145 /* 1146 * If this is the Ifile, and we've already written the Ifile in this 1147 * partial segment, just overwrite it (it's not on disk yet) and 1148 * continue. 1149 * 1150 * XXX we know that the bp that we get the second time around has 1151 * already been gathered. 1152 */ 1153 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1154 *(sp->idp) = *ip->i_din.ffs1_din; 1155 ip->i_lfs_osize = ip->i_size; 1156 return 0; 1157 } 1158 1159 bp = sp->ibp; 1160 cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs)); 1161 *cdp = *ip->i_din.ffs1_din; 1162 1163 /* 1164 * This inode is on its way to disk; clear its VU_DIROP status when 1165 * the write is complete. 1166 */ 1167 if (vp->v_uflag & VU_DIROP) { 1168 if (!(sp->seg_flags & SEGM_CLEAN)) 1169 ip->i_flag |= IN_CDIROP; 1170 else { 1171 DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number)); 1172 } 1173 } 1174 1175 /* 1176 * If cleaning, link counts and directory file sizes cannot change, 1177 * since those would be directory operations---even if the file 1178 * we are writing is marked VU_DIROP we should write the old values. 1179 * If we're not cleaning, of course, update the values so we get 1180 * current values the next time we clean. 1181 */ 1182 if (sp->seg_flags & SEGM_CLEAN) { 1183 if (vp->v_uflag & VU_DIROP) { 1184 cdp->di_nlink = ip->i_lfs_odnlink; 1185 /* if (vp->v_type == VDIR) */ 1186 cdp->di_size = ip->i_lfs_osize; 1187 } 1188 } else { 1189 ip->i_lfs_odnlink = cdp->di_nlink; 1190 ip->i_lfs_osize = ip->i_size; 1191 } 1192 1193 1194 /* We can finish the segment accounting for truncations now */ 1195 lfs_finalize_ino_seguse(fs, ip); 1196 1197 /* 1198 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1199 * addresses to disk; possibly change the on-disk record of 1200 * the inode size, either by reverting to the previous size 1201 * (in the case of cleaning) or by verifying the inode's block 1202 * holdings (in the case of files being allocated as they are being 1203 * written). 1204 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1205 * XXX count on disk wrong by the same amount. We should be 1206 * XXX able to "borrow" from lfs_avail and return it after the 1207 * XXX Ifile is written. See also in lfs_writeseg. 1208 */ 1209 1210 /* Check file size based on highest allocated block */ 1211 if (((ip->i_ffs1_mode & IFMT) == IFREG || 1212 (ip->i_ffs1_mode & IFMT) == IFDIR) && 1213 ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) { 1214 cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift; 1215 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1216 PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size)); 1217 } 1218 if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) { 1219 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)" 1220 " at %x\n", ip->i_number, ip->i_lfs_effnblks, 1221 ip->i_ffs1_blocks, fs->lfs_offset)); 1222 for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR; 1223 daddrp++) { 1224 if (*daddrp == UNWRITTEN) { 1225 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1226 *daddrp = 0; 1227 } 1228 } 1229 } 1230 1231#ifdef DIAGNOSTIC 1232 /* 1233 * Check dinode held blocks against dinode size. 1234 * This should be identical to the check in lfs_vget(). 1235 */ 1236 for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift; 1237 i < NDADDR; i++) { 1238 KASSERT(i >= 0); 1239 if ((cdp->di_mode & IFMT) == IFLNK) 1240 continue; 1241 if (((cdp->di_mode & IFMT) == IFBLK || 1242 (cdp->di_mode & IFMT) == IFCHR) && i == 0) 1243 continue; 1244 if (cdp->di_db[i] != 0) { 1245# ifdef DEBUG 1246 lfs_dump_dinode(cdp); 1247# endif 1248 panic("writing inconsistent inode"); 1249 } 1250 } 1251#endif /* DIAGNOSTIC */ 1252 1253 if (ip->i_flag & IN_CLEANING) 1254 LFS_CLR_UINO(ip, IN_CLEANING); 1255 else { 1256 /* XXX IN_ALLMOD */ 1257 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1258 IN_UPDATE | IN_MODIFY); 1259 if (ip->i_lfs_effnblks == ip->i_ffs1_blocks) 1260 LFS_CLR_UINO(ip, IN_MODIFIED); 1261 else { 1262 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1263 "blks=%d, eff=%d\n", ip->i_number, 1264 ip->i_ffs1_blocks, ip->i_lfs_effnblks)); 1265 } 1266 } 1267 1268 if (ip->i_number == LFS_IFILE_INUM) { 1269 /* We know sp->idp == NULL */ 1270 sp->idp = ((struct ufs1_dinode *)bp->b_data) + 1271 (sp->ninodes % INOPB(fs)); 1272 1273 /* Not dirty any more */ 1274 mutex_enter(&lfs_lock); 1275 fs->lfs_flags &= ~LFS_IFDIRTY; 1276 mutex_exit(&lfs_lock); 1277 } 1278 1279 if (gotblk) { 1280 mutex_enter(&bufcache_lock); 1281 LFS_LOCK_BUF(bp); 1282 brelsel(bp, 0); 1283 mutex_exit(&bufcache_lock); 1284 } 1285 1286 /* Increment inode count in segment summary block. */ 1287 ++((SEGSUM *)(sp->segsum))->ss_ninos; 1288 1289 /* If this page is full, set flag to allocate a new page. */ 1290 if (++sp->ninodes % 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 1299int 1300lfs_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 /* 1308 * If full, finish this segment. We may be doing I/O, so 1309 * release and reacquire the splbio(). 1310 */ 1311#ifdef DIAGNOSTIC 1312 if (sp->vp == NULL) 1313 panic ("lfs_gatherblock: Null vp in segment"); 1314#endif 1315 fs = sp->fs; 1316 blksinblk = howmany(bp->b_bcount, fs->lfs_bsize); 1317 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1318 sp->seg_bytes_left < bp->b_bcount) { 1319 if (mptr) 1320 mutex_exit(mptr); 1321 lfs_updatemeta(sp); 1322 1323 vers = sp->fip->fi_version; 1324 (void) lfs_writeseg(fs, sp); 1325 1326 /* Add the current file to the segment summary. */ 1327 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1328 1329 if (mptr) 1330 mutex_enter(mptr); 1331 return (1); 1332 } 1333 1334 if (bp->b_flags & B_GATHERED) { 1335 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d," 1336 " lbn %" PRId64 "\n", 1337 sp->fip->fi_ino, bp->b_lblkno)); 1338 return (0); 1339 } 1340 1341 /* Insert into the buffer list, update the FINFO block. */ 1342 bp->b_flags |= B_GATHERED; 1343 1344 *sp->cbpp++ = bp; 1345 for (j = 0; j < blksinblk; j++) { 1346 sp->fip->fi_blocks[sp->fip->fi_nblocks++] = 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 1356int 1357lfs_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 1379loop: 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 */ 1388loop: 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%x", 1400 bp->b_lblkno, bp->b_flags, 1401 (unsigned)fs->lfs_offset); 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 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 ufs_bmaparray(). 1452 * 1453 * Account for this change in the segment table. 1454 * 1455 * called with sp == NULL by roll-forwarding code. 1456 */ 1457void 1458lfs_update_single(struct lfs *fs, struct segment *sp, 1459 struct vnode *vp, daddr_t lbn, int32_t ndaddr, int size) 1460{ 1461 SEGUSE *sup; 1462 struct buf *bp; 1463 struct indir a[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 = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1474 if (error) 1475 panic("lfs_updatemeta: ufs_bmaparray returned %d", error); 1476 1477 daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */ 1478 KASSERT(daddr <= LFS_MAX_DADDR); 1479 if (daddr > 0) 1480 daddr = dbtofsb(fs, daddr); 1481 1482 bb = numfrags(fs, size); 1483 switch (num) { 1484 case 0: 1485 ooff = ip->i_ffs1_db[lbn]; 1486 DEBUG_OOFF(0); 1487 if (ooff == UNWRITTEN) 1488 ip->i_ffs1_blocks += bb; 1489 else { 1490 /* possible fragment truncation or extension */ 1491 obb = btofsb(fs, ip->i_lfs_fragsize[lbn]); 1492 ip->i_ffs1_blocks += (bb - obb); 1493 } 1494 ip->i_ffs1_db[lbn] = ndaddr; 1495 break; 1496 case 1: 1497 ooff = ip->i_ffs1_ib[a[0].in_off]; 1498 DEBUG_OOFF(1); 1499 if (ooff == UNWRITTEN) 1500 ip->i_ffs1_blocks += bb; 1501 ip->i_ffs1_ib[a[0].in_off] = ndaddr; 1502 break; 1503 default: 1504 ap = &a[num - 1]; 1505 if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, 1506 B_MODIFY, &bp)) 1507 panic("lfs_updatemeta: bread bno %" PRId64, 1508 ap->in_lbn); 1509 1510 /* XXX ondisk32 */ 1511 ooff = ((int32_t *)bp->b_data)[ap->in_off]; 1512 DEBUG_OOFF(num); 1513 if (ooff == UNWRITTEN) 1514 ip->i_ffs1_blocks += bb; 1515 /* XXX ondisk32 */ 1516 ((int32_t *)bp->b_data)[ap->in_off] = ndaddr; 1517 (void) VOP_BWRITE(bp->b_vp, bp); 1518 } 1519 1520 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1521 1522 /* Update hiblk when extending the file */ 1523 if (lbn > ip->i_lfs_hiblk) 1524 ip->i_lfs_hiblk = lbn; 1525 1526 /* 1527 * Though we'd rather it couldn't, this *can* happen right now 1528 * if cleaning blocks and regular blocks coexist. 1529 */ 1530 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1531 1532 /* 1533 * Update segment usage information, based on old size 1534 * and location. 1535 */ 1536 if (daddr > 0) { 1537 u_int32_t oldsn = dtosn(fs, daddr); 1538#ifdef DIAGNOSTIC 1539 int ndupino; 1540 1541 if (sp && sp->seg_number == oldsn) { 1542 ndupino = sp->ndupino; 1543 } else { 1544 ndupino = 0; 1545 } 1546#endif 1547 KASSERT(oldsn < fs->lfs_nseg); 1548 if (lbn >= 0 && lbn < NDADDR) 1549 osize = ip->i_lfs_fragsize[lbn]; 1550 else 1551 osize = fs->lfs_bsize; 1552 LFS_SEGENTRY(sup, fs, oldsn, bp); 1553#ifdef DIAGNOSTIC 1554 if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino 1555 < osize) { 1556 printf("lfs_updatemeta: negative bytes " 1557 "(segment %" PRIu32 " short by %" PRId64 1558 ")\n", dtosn(fs, daddr), 1559 (int64_t)osize - 1560 (sizeof (struct ufs1_dinode) * ndupino + 1561 sup->su_nbytes)); 1562 printf("lfs_updatemeta: ino %llu, lbn %" PRId64 1563 ", addr = 0x%" PRIx64 "\n", 1564 (unsigned long long)ip->i_number, lbn, daddr); 1565 printf("lfs_updatemeta: ndupino=%d\n", ndupino); 1566 panic("lfs_updatemeta: negative bytes"); 1567 sup->su_nbytes = osize - 1568 sizeof (struct ufs1_dinode) * ndupino; 1569 } 1570#endif 1571 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1572 " db 0x%" PRIx64 "\n", 1573 dtosn(fs, daddr), osize, 1574 ip->i_number, lbn, daddr)); 1575 sup->su_nbytes -= osize; 1576 if (!(bp->b_flags & B_GATHERED)) { 1577 mutex_enter(&lfs_lock); 1578 fs->lfs_flags |= LFS_IFDIRTY; 1579 mutex_exit(&lfs_lock); 1580 } 1581 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1582 } 1583 /* 1584 * Now that this block has a new address, and its old 1585 * segment no longer owns it, we can forget about its 1586 * old size. 1587 */ 1588 if (lbn >= 0 && lbn < NDADDR) 1589 ip->i_lfs_fragsize[lbn] = size; 1590} 1591 1592/* 1593 * Update the metadata that points to the blocks listed in the FINFO 1594 * array. 1595 */ 1596void 1597lfs_updatemeta(struct segment *sp) 1598{ 1599 struct buf *sbp; 1600 struct lfs *fs; 1601 struct vnode *vp; 1602 daddr_t lbn; 1603 int i, nblocks, num; 1604 int bb; 1605 int bytesleft, size; 1606 1607 ASSERT_SEGLOCK(sp->fs); 1608 vp = sp->vp; 1609 nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1610 KASSERT(nblocks >= 0); 1611 KASSERT(vp != NULL); 1612 if (nblocks == 0) 1613 return; 1614 1615 /* 1616 * This count may be high due to oversize blocks from lfs_gop_write. 1617 * Correct for this. (XXX we should be able to keep track of these.) 1618 */ 1619 fs = sp->fs; 1620 for (i = 0; i < nblocks; i++) { 1621 if (sp->start_bpp[i] == NULL) { 1622 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1623 nblocks = i; 1624 break; 1625 } 1626 num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize); 1627 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1628 nblocks -= num - 1; 1629 } 1630 1631 KASSERT(vp->v_type == VREG || 1632 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1633 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1634 1635 /* 1636 * Sort the blocks. 1637 * 1638 * We have to sort even if the blocks come from the 1639 * cleaner, because there might be other pending blocks on the 1640 * same inode...and if we don't sort, and there are fragments 1641 * present, blocks may be written in the wrong place. 1642 */ 1643 lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize); 1644 1645 /* 1646 * Record the length of the last block in case it's a fragment. 1647 * If there are indirect blocks present, they sort last. An 1648 * indirect block will be lfs_bsize and its presence indicates 1649 * that you cannot have fragments. 1650 * 1651 * XXX This last is a lie. A cleaned fragment can coexist with 1652 * XXX a later indirect block. This will continue to be 1653 * XXX true until lfs_markv is fixed to do everything with 1654 * XXX fake blocks (including fake inodes and fake indirect blocks). 1655 */ 1656 sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1657 fs->lfs_bmask) + 1; 1658 1659 /* 1660 * Assign disk addresses, and update references to the logical 1661 * block and the segment usage information. 1662 */ 1663 for (i = nblocks; i--; ++sp->start_bpp) { 1664 sbp = *sp->start_bpp; 1665 lbn = *sp->start_lbp; 1666 KASSERT(sbp->b_lblkno == lbn); 1667 1668 sbp->b_blkno = fsbtodb(fs, fs->lfs_offset); 1669 1670 /* 1671 * If we write a frag in the wrong place, the cleaner won't 1672 * be able to correctly identify its size later, and the 1673 * segment will be uncleanable. (Even worse, it will assume 1674 * that the indirect block that actually ends the list 1675 * is of a smaller size!) 1676 */ 1677 if ((sbp->b_bcount & fs->lfs_bmask) && i != 0) 1678 panic("lfs_updatemeta: fragment is not last block"); 1679 1680 /* 1681 * For each subblock in this possibly oversized block, 1682 * update its address on disk. 1683 */ 1684 KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize); 1685 KASSERT(vp == sbp->b_vp); 1686 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1687 bytesleft -= fs->lfs_bsize) { 1688 size = MIN(bytesleft, fs->lfs_bsize); 1689 bb = numfrags(fs, size); 1690 lbn = *sp->start_lbp++; 1691 lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset, 1692 size); 1693 fs->lfs_offset += bb; 1694 } 1695 1696 } 1697 1698 /* This inode has been modified */ 1699 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 1700} 1701 1702/* 1703 * Move lfs_offset to a segment earlier than sn. 1704 */ 1705int 1706lfs_rewind(struct lfs *fs, int newsn) 1707{ 1708 int sn, osn, isdirty; 1709 struct buf *bp; 1710 SEGUSE *sup; 1711 1712 ASSERT_SEGLOCK(fs); 1713 1714 osn = dtosn(fs, fs->lfs_offset); 1715 if (osn < newsn) 1716 return 0; 1717 1718 /* lfs_avail eats the remaining space in this segment */ 1719 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg); 1720 1721 /* Find a low-numbered segment */ 1722 for (sn = 0; sn < fs->lfs_nseg; ++sn) { 1723 LFS_SEGENTRY(sup, fs, sn, bp); 1724 isdirty = sup->su_flags & SEGUSE_DIRTY; 1725 brelse(bp, 0); 1726 1727 if (!isdirty) 1728 break; 1729 } 1730 if (sn == fs->lfs_nseg) 1731 panic("lfs_rewind: no clean segments"); 1732 if (newsn >= 0 && sn >= newsn) 1733 return ENOENT; 1734 fs->lfs_nextseg = sn; 1735 lfs_newseg(fs); 1736 fs->lfs_offset = fs->lfs_curseg; 1737 1738 return 0; 1739} 1740 1741/* 1742 * Start a new partial segment. 1743 * 1744 * Return 1 when we entered to a new segment. 1745 * Otherwise, return 0. 1746 */ 1747int 1748lfs_initseg(struct lfs *fs) 1749{ 1750 struct segment *sp = fs->lfs_sp; 1751 SEGSUM *ssp; 1752 struct buf *sbp; /* buffer for SEGSUM */ 1753 int repeat = 0; /* return value */ 1754 1755 ASSERT_SEGLOCK(fs); 1756 /* Advance to the next segment. */ 1757 if (!LFS_PARTIAL_FITS(fs)) { 1758 SEGUSE *sup; 1759 struct buf *bp; 1760 1761 /* lfs_avail eats the remaining space */ 1762 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - 1763 fs->lfs_curseg); 1764 /* Wake up any cleaning procs waiting on this file system. */ 1765 lfs_wakeup_cleaner(fs); 1766 lfs_newseg(fs); 1767 repeat = 1; 1768 fs->lfs_offset = fs->lfs_curseg; 1769 1770 sp->seg_number = dtosn(fs, fs->lfs_curseg); 1771 sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg); 1772 1773 /* 1774 * If the segment contains a superblock, update the offset 1775 * and summary address to skip over it. 1776 */ 1777 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1778 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1779 fs->lfs_offset += btofsb(fs, LFS_SBPAD); 1780 sp->seg_bytes_left -= LFS_SBPAD; 1781 } 1782 brelse(bp, 0); 1783 /* Segment zero could also contain the labelpad */ 1784 if (fs->lfs_version > 1 && sp->seg_number == 0 && 1785 fs->lfs_start < btofsb(fs, LFS_LABELPAD)) { 1786 fs->lfs_offset += 1787 btofsb(fs, LFS_LABELPAD) - fs->lfs_start; 1788 sp->seg_bytes_left -= 1789 LFS_LABELPAD - fsbtob(fs, fs->lfs_start); 1790 } 1791 } else { 1792 sp->seg_number = dtosn(fs, fs->lfs_curseg); 1793 sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg - 1794 (fs->lfs_offset - fs->lfs_curseg)); 1795 } 1796 fs->lfs_lastpseg = fs->lfs_offset; 1797 1798 /* Record first address of this partial segment */ 1799 if (sp->seg_flags & SEGM_CLEAN) { 1800 fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset; 1801 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1802 /* "1" is the artificial inc in lfs_seglock */ 1803 mutex_enter(&lfs_lock); 1804 while (fs->lfs_iocount > 1) { 1805 mtsleep(&fs->lfs_iocount, PRIBIO + 1, 1806 "lfs_initseg", 0, &lfs_lock); 1807 } 1808 mutex_exit(&lfs_lock); 1809 fs->lfs_cleanind = 0; 1810 } 1811 } 1812 1813 sp->fs = fs; 1814 sp->ibp = NULL; 1815 sp->idp = NULL; 1816 sp->ninodes = 0; 1817 sp->ndupino = 0; 1818 1819 sp->cbpp = sp->bpp; 1820 1821 /* Get a new buffer for SEGSUM */ 1822 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1823 fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY); 1824 1825 /* ... and enter it into the buffer list. */ 1826 *sp->cbpp = sbp; 1827 sp->cbpp++; 1828 fs->lfs_offset += btofsb(fs, fs->lfs_sumsize); 1829 1830 sp->start_bpp = sp->cbpp; 1831 1832 /* Set point to SEGSUM, initialize it. */ 1833 ssp = sp->segsum = sbp->b_data; 1834 memset(ssp, 0, fs->lfs_sumsize); 1835 ssp->ss_next = fs->lfs_nextseg; 1836 ssp->ss_nfinfo = ssp->ss_ninos = 0; 1837 ssp->ss_magic = SS_MAGIC; 1838 1839 /* Set pointer to first FINFO, initialize it. */ 1840 sp->fip = (struct finfo *)((char *)sp->segsum + SEGSUM_SIZE(fs)); 1841 sp->fip->fi_nblocks = 0; 1842 sp->start_lbp = &sp->fip->fi_blocks[0]; 1843 sp->fip->fi_lastlength = 0; 1844 1845 sp->seg_bytes_left -= fs->lfs_sumsize; 1846 sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs); 1847 1848 return (repeat); 1849} 1850 1851/* 1852 * Remove SEGUSE_INVAL from all segments. 1853 */ 1854void 1855lfs_unset_inval_all(struct lfs *fs) 1856{ 1857 SEGUSE *sup; 1858 struct buf *bp; 1859 int i; 1860 1861 for (i = 0; i < fs->lfs_nseg; i++) { 1862 LFS_SEGENTRY(sup, fs, i, bp); 1863 if (sup->su_flags & SEGUSE_INVAL) { 1864 sup->su_flags &= ~SEGUSE_INVAL; 1865 LFS_WRITESEGENTRY(sup, fs, i, bp); 1866 } else 1867 brelse(bp, 0); 1868 } 1869} 1870 1871/* 1872 * Return the next segment to write. 1873 */ 1874void 1875lfs_newseg(struct lfs *fs) 1876{ 1877 CLEANERINFO *cip; 1878 SEGUSE *sup; 1879 struct buf *bp; 1880 int curseg, isdirty, sn, skip_inval; 1881 1882 ASSERT_SEGLOCK(fs); 1883 1884 /* Honor LFCNWRAPSTOP */ 1885 mutex_enter(&lfs_lock); 1886 while (fs->lfs_nextseg < fs->lfs_curseg && fs->lfs_nowrap) { 1887 if (fs->lfs_wrappass) { 1888 log(LOG_NOTICE, "%s: wrappass=%d\n", 1889 fs->lfs_fsmnt, fs->lfs_wrappass); 1890 fs->lfs_wrappass = 0; 1891 break; 1892 } 1893 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1894 wakeup(&fs->lfs_nowrap); 1895 log(LOG_NOTICE, "%s: waiting at log wrap\n", fs->lfs_fsmnt); 1896 mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1897 &lfs_lock); 1898 } 1899 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1900 mutex_exit(&lfs_lock); 1901 1902 LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp); 1903 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1904 dtosn(fs, fs->lfs_nextseg))); 1905 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1906 sup->su_nbytes = 0; 1907 sup->su_nsums = 0; 1908 sup->su_ninos = 0; 1909 LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp); 1910 1911 LFS_CLEANERINFO(cip, fs, bp); 1912 --cip->clean; 1913 ++cip->dirty; 1914 fs->lfs_nclean = cip->clean; 1915 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1916 1917 fs->lfs_lastseg = fs->lfs_curseg; 1918 fs->lfs_curseg = fs->lfs_nextseg; 1919 skip_inval = 1; 1920 for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) { 1921 sn = (sn + 1) % fs->lfs_nseg; 1922 1923 if (sn == curseg) { 1924 if (skip_inval) 1925 skip_inval = 0; 1926 else 1927 panic("lfs_nextseg: no clean segments"); 1928 } 1929 LFS_SEGENTRY(sup, fs, sn, bp); 1930 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1931 /* Check SEGUSE_EMPTY as we go along */ 1932 if (isdirty && sup->su_nbytes == 0 && 1933 !(sup->su_flags & SEGUSE_EMPTY)) 1934 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1935 else 1936 brelse(bp, 0); 1937 1938 if (!isdirty) 1939 break; 1940 } 1941 if (skip_inval == 0) 1942 lfs_unset_inval_all(fs); 1943 1944 ++fs->lfs_nactive; 1945 fs->lfs_nextseg = sntod(fs, sn); 1946 if (lfs_dostats) { 1947 ++lfs_stats.segsused; 1948 } 1949} 1950 1951static struct buf * 1952lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1953 int n) 1954{ 1955 struct lfs_cluster *cl; 1956 struct buf **bpp, *bp; 1957 1958 ASSERT_SEGLOCK(fs); 1959 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1960 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1961 memset(cl, 0, sizeof(*cl)); 1962 cl->fs = fs; 1963 cl->bpp = bpp; 1964 cl->bufcount = 0; 1965 cl->bufsize = 0; 1966 1967 /* If this segment is being written synchronously, note that */ 1968 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 1969 cl->flags |= LFS_CL_SYNC; 1970 cl->seg = fs->lfs_sp; 1971 ++cl->seg->seg_iocount; 1972 } 1973 1974 /* Get an empty buffer header, or maybe one with something on it */ 1975 bp = getiobuf(vp, true); 1976 bp->b_dev = NODEV; 1977 bp->b_blkno = bp->b_lblkno = addr; 1978 bp->b_iodone = lfs_cluster_callback; 1979 bp->b_private = cl; 1980 1981 return bp; 1982} 1983 1984int 1985lfs_writeseg(struct lfs *fs, struct segment *sp) 1986{ 1987 struct buf **bpp, *bp, *cbp, *newbp, *unbusybp; 1988 SEGUSE *sup; 1989 SEGSUM *ssp; 1990 int i; 1991 int do_again, nblocks, byteoffset; 1992 size_t el_size; 1993 struct lfs_cluster *cl; 1994 u_short ninos; 1995 struct vnode *devvp; 1996 char *p = NULL; 1997 struct vnode *vp; 1998 int32_t *daddrp; /* XXX ondisk32 */ 1999 int changed; 2000 u_int32_t sum; 2001#ifdef DEBUG 2002 FINFO *fip; 2003 int findex; 2004#endif 2005 2006 ASSERT_SEGLOCK(fs); 2007 2008 ssp = (SEGSUM *)sp->segsum; 2009 2010 /* 2011 * If there are no buffers other than the segment summary to write, 2012 * don't do anything. If we are the end of a dirop sequence, however, 2013 * write the empty segment summary anyway, to help out the 2014 * roll-forward agent. 2015 */ 2016 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 2017 if ((ssp->ss_flags & (SS_DIROP | SS_CONT)) != SS_DIROP) 2018 return 0; 2019 } 2020 2021 /* Note if partial segment is being written by the cleaner */ 2022 if (sp->seg_flags & SEGM_CLEAN) 2023 ssp->ss_flags |= SS_CLEAN; 2024 2025 /* Note if we are writing to reclaim */ 2026 if (sp->seg_flags & SEGM_RECLAIM) { 2027 ssp->ss_flags |= SS_RECLAIM; 2028 ssp->ss_reclino = fs->lfs_reclino; 2029 } 2030 2031 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2032 2033 /* Update the segment usage information. */ 2034 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2035 2036 /* Loop through all blocks, except the segment summary. */ 2037 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2038 if ((*bpp)->b_vp != devvp) { 2039 sup->su_nbytes += (*bpp)->b_bcount; 2040 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2041 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2042 sp->seg_number, (*bpp)->b_bcount, 2043 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2044 (*bpp)->b_blkno)); 2045 } 2046 } 2047 2048#ifdef DEBUG 2049 /* Check for zero-length and zero-version FINFO entries. */ 2050 fip = (struct finfo *)((char *)ssp + SEGSUM_SIZE(fs)); 2051 for (findex = 0; findex < ssp->ss_nfinfo; findex++) { 2052 KDASSERT(fip->fi_nblocks > 0); 2053 KDASSERT(fip->fi_version > 0); 2054 fip = (FINFO *)((char *)fip + FINFOSIZE + 2055 sizeof(int32_t) * fip->fi_nblocks); 2056 } 2057#endif /* DEBUG */ 2058 2059 ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs); 2060 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2061 sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode), 2062 ssp->ss_ninos)); 2063 sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode); 2064 /* sup->su_nbytes += fs->lfs_sumsize; */ 2065 if (fs->lfs_version == 1) 2066 sup->su_olastmod = time_second; 2067 else 2068 sup->su_lastmod = time_second; 2069 sup->su_ninos += ninos; 2070 ++sup->su_nsums; 2071 fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize); 2072 2073 do_again = !(bp->b_flags & B_GATHERED); 2074 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2075 2076 /* 2077 * Mark blocks B_BUSY, to prevent then from being changed between 2078 * the checksum computation and the actual write. 2079 * 2080 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2081 * there are any, replace them with copies that have UNASSIGNED 2082 * instead. 2083 */ 2084 mutex_enter(&bufcache_lock); 2085 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2086 ++bpp; 2087 bp = *bpp; 2088 if (bp->b_iodone != NULL) { /* UBC or malloced buffer */ 2089 bp->b_cflags |= BC_BUSY; 2090 continue; 2091 } 2092 2093 while (bp->b_cflags & BC_BUSY) { 2094 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2095 " data summary corruption for ino %d, lbn %" 2096 PRId64 "\n", 2097 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2098 bp->b_cflags |= BC_WANTED; 2099 cv_wait(&bp->b_busy, &bufcache_lock); 2100 } 2101 bp->b_cflags |= BC_BUSY; 2102 mutex_exit(&bufcache_lock); 2103 unbusybp = NULL; 2104 2105 /* 2106 * Check and replace indirect block UNWRITTEN bogosity. 2107 * XXX See comment in lfs_writefile. 2108 */ 2109 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2110 VTOI(bp->b_vp)->i_ffs1_blocks != 2111 VTOI(bp->b_vp)->i_lfs_effnblks) { 2112 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n", 2113 VTOI(bp->b_vp)->i_number, 2114 VTOI(bp->b_vp)->i_lfs_effnblks, 2115 VTOI(bp->b_vp)->i_ffs1_blocks)); 2116 /* Make a copy we'll make changes to */ 2117 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2118 bp->b_bcount, LFS_NB_IBLOCK); 2119 newbp->b_blkno = bp->b_blkno; 2120 memcpy(newbp->b_data, bp->b_data, 2121 newbp->b_bcount); 2122 2123 changed = 0; 2124 /* XXX ondisk32 */ 2125 for (daddrp = (int32_t *)(newbp->b_data); 2126 daddrp < (int32_t *)((char *)newbp->b_data + 2127 newbp->b_bcount); daddrp++) { 2128 if (*daddrp == UNWRITTEN) { 2129 ++changed; 2130 *daddrp = 0; 2131 } 2132 } 2133 /* 2134 * Get rid of the old buffer. Don't mark it clean, 2135 * though, if it still has dirty data on it. 2136 */ 2137 if (changed) { 2138 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2139 " bp = %p newbp = %p\n", changed, bp, 2140 newbp)); 2141 *bpp = newbp; 2142 bp->b_flags &= ~B_GATHERED; 2143 bp->b_error = 0; 2144 if (bp->b_iodone != NULL) { 2145 DLOG((DLOG_SEG, "lfs_writeseg: " 2146 "indir bp should not be B_CALL\n")); 2147 biodone(bp); 2148 bp = NULL; 2149 } else { 2150 /* Still on free list, leave it there */ 2151 unbusybp = bp; 2152 /* 2153 * We have to re-decrement lfs_avail 2154 * since this block is going to come 2155 * back around to us in the next 2156 * segment. 2157 */ 2158 fs->lfs_avail -= 2159 btofsb(fs, bp->b_bcount); 2160 } 2161 } else { 2162 lfs_freebuf(fs, newbp); 2163 } 2164 } 2165 mutex_enter(&bufcache_lock); 2166 if (unbusybp != NULL) { 2167 unbusybp->b_cflags &= ~BC_BUSY; 2168 if (unbusybp->b_cflags & BC_WANTED) 2169 cv_broadcast(&bp->b_busy); 2170 } 2171 } 2172 mutex_exit(&bufcache_lock); 2173 2174 /* 2175 * Compute checksum across data and then across summary; the first 2176 * block (the summary block) is skipped. Set the create time here 2177 * so that it's guaranteed to be later than the inode mod times. 2178 */ 2179 sum = 0; 2180 if (fs->lfs_version == 1) 2181 el_size = sizeof(u_long); 2182 else 2183 el_size = sizeof(u_int32_t); 2184 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2185 ++bpp; 2186 /* Loop through gop_write cluster blocks */ 2187 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2188 byteoffset += fs->lfs_bsize) { 2189#ifdef LFS_USE_B_INVAL 2190 if (((*bpp)->b_cflags & BC_INVAL) != 0 && 2191 (*bpp)->b_iodone != NULL) { 2192 if (copyin((void *)(*bpp)->b_saveaddr + 2193 byteoffset, dp, el_size)) { 2194 panic("lfs_writeseg: copyin failed [1]:" 2195 " ino %d blk %" PRId64, 2196 VTOI((*bpp)->b_vp)->i_number, 2197 (*bpp)->b_lblkno); 2198 } 2199 } else 2200#endif /* LFS_USE_B_INVAL */ 2201 { 2202 sum = lfs_cksum_part((char *) 2203 (*bpp)->b_data + byteoffset, el_size, sum); 2204 } 2205 } 2206 } 2207 if (fs->lfs_version == 1) 2208 ssp->ss_ocreate = time_second; 2209 else { 2210 ssp->ss_create = time_second; 2211 ssp->ss_serial = ++fs->lfs_serial; 2212 ssp->ss_ident = fs->lfs_ident; 2213 } 2214 ssp->ss_datasum = lfs_cksum_fold(sum); 2215 ssp->ss_sumsum = cksum(&ssp->ss_datasum, 2216 fs->lfs_sumsize - sizeof(ssp->ss_sumsum)); 2217 2218 mutex_enter(&lfs_lock); 2219 fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) + 2220 btofsb(fs, fs->lfs_sumsize)); 2221 fs->lfs_dmeta += (btofsb(fs, ninos * fs->lfs_ibsize) + 2222 btofsb(fs, fs->lfs_sumsize)); 2223 mutex_exit(&lfs_lock); 2224 2225 /* 2226 * When we simply write the blocks we lose a rotation for every block 2227 * written. To avoid this problem, we cluster the buffers into a 2228 * chunk and write the chunk. MAXPHYS is the largest size I/O 2229 * devices can handle, use that for the size of the chunks. 2230 * 2231 * Blocks that are already clusters (from GOP_WRITE), however, we 2232 * don't bother to copy into other clusters. 2233 */ 2234 2235#define CHUNKSIZE MAXPHYS 2236 2237 if (devvp == NULL) 2238 panic("devvp is NULL"); 2239 for (bpp = sp->bpp, i = nblocks; i;) { 2240 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2241 cl = cbp->b_private; 2242 2243 cbp->b_flags |= B_ASYNC; 2244 cbp->b_cflags |= BC_BUSY; 2245 cbp->b_bcount = 0; 2246 2247#if defined(DEBUG) && defined(DIAGNOSTIC) 2248 if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs)) 2249 / sizeof(int32_t)) { 2250 panic("lfs_writeseg: real bpp overwrite"); 2251 } 2252 if (bpp - sp->bpp > segsize(fs) / fs->lfs_fsize) { 2253 panic("lfs_writeseg: theoretical bpp overwrite"); 2254 } 2255#endif 2256 2257 /* 2258 * Construct the cluster. 2259 */ 2260 mutex_enter(&lfs_lock); 2261 ++fs->lfs_iocount; 2262 mutex_exit(&lfs_lock); 2263 while (i && cbp->b_bcount < CHUNKSIZE) { 2264 bp = *bpp; 2265 2266 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2267 break; 2268 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2269 break; 2270 2271 /* Clusters from GOP_WRITE are expedited */ 2272 if (bp->b_bcount > fs->lfs_bsize) { 2273 if (cbp->b_bcount > 0) 2274 /* Put in its own buffer */ 2275 break; 2276 else { 2277 cbp->b_data = bp->b_data; 2278 } 2279 } else if (cbp->b_bcount == 0) { 2280 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2281 LFS_NB_CLUSTER); 2282 cl->flags |= LFS_CL_MALLOC; 2283 } 2284#ifdef DIAGNOSTIC 2285 if (dtosn(fs, dbtofsb(fs, bp->b_blkno + 2286 btodb(bp->b_bcount - 1))) != 2287 sp->seg_number) { 2288 printf("blk size %d daddr %" PRIx64 2289 " not in seg %d\n", 2290 bp->b_bcount, bp->b_blkno, 2291 sp->seg_number); 2292 panic("segment overwrite"); 2293 } 2294#endif 2295 2296#ifdef LFS_USE_B_INVAL 2297 /* 2298 * Fake buffers from the cleaner are marked as B_INVAL. 2299 * We need to copy the data from user space rather than 2300 * from the buffer indicated. 2301 * XXX == what do I do on an error? 2302 */ 2303 if ((bp->b_cflags & BC_INVAL) != 0 && 2304 bp->b_iodone != NULL) { 2305 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2306 panic("lfs_writeseg: " 2307 "copyin failed [2]"); 2308 } else 2309#endif /* LFS_USE_B_INVAL */ 2310 if (cl->flags & LFS_CL_MALLOC) { 2311 /* copy data into our cluster. */ 2312 memcpy(p, bp->b_data, bp->b_bcount); 2313 p += bp->b_bcount; 2314 } 2315 2316 cbp->b_bcount += bp->b_bcount; 2317 cl->bufsize += bp->b_bcount; 2318 2319 bp->b_flags &= ~B_READ; 2320 bp->b_error = 0; 2321 cl->bpp[cl->bufcount++] = bp; 2322 2323 vp = bp->b_vp; 2324 mutex_enter(&bufcache_lock); 2325 mutex_enter(vp->v_interlock); 2326 bp->b_oflags &= ~(BO_DELWRI | BO_DONE); 2327 reassignbuf(bp, vp); 2328 vp->v_numoutput++; 2329 mutex_exit(vp->v_interlock); 2330 mutex_exit(&bufcache_lock); 2331 2332 bpp++; 2333 i--; 2334 } 2335 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2336 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2337 else 2338 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2339 mutex_enter(devvp->v_interlock); 2340 devvp->v_numoutput++; 2341 mutex_exit(devvp->v_interlock); 2342 VOP_STRATEGY(devvp, cbp); 2343 curlwp->l_ru.ru_oublock++; 2344 } 2345 2346 if (lfs_dostats) { 2347 ++lfs_stats.psegwrites; 2348 lfs_stats.blocktot += nblocks - 1; 2349 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2350 ++lfs_stats.psyncwrites; 2351 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2352 ++lfs_stats.pcleanwrites; 2353 lfs_stats.cleanblocks += nblocks - 1; 2354 } 2355 } 2356 2357 return (lfs_initseg(fs) || do_again); 2358} 2359 2360void 2361lfs_writesuper(struct lfs *fs, daddr_t daddr) 2362{ 2363 struct buf *bp; 2364 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2365 int s; 2366 2367 ASSERT_MAYBE_SEGLOCK(fs); 2368#ifdef DIAGNOSTIC 2369 KASSERT(fs->lfs_magic == LFS_MAGIC); 2370#endif 2371 /* 2372 * If we can write one superblock while another is in 2373 * progress, we risk not having a complete checkpoint if we crash. 2374 * So, block here if a superblock write is in progress. 2375 */ 2376 mutex_enter(&lfs_lock); 2377 s = splbio(); 2378 while (fs->lfs_sbactive) { 2379 mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2380 &lfs_lock); 2381 } 2382 fs->lfs_sbactive = daddr; 2383 splx(s); 2384 mutex_exit(&lfs_lock); 2385 2386 /* Set timestamp of this version of the superblock */ 2387 if (fs->lfs_version == 1) 2388 fs->lfs_otstamp = time_second; 2389 fs->lfs_tstamp = time_second; 2390 2391 /* Checksum the superblock and copy it into a buffer. */ 2392 fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs)); 2393 bp = lfs_newbuf(fs, devvp, 2394 fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2395 memset((char *)bp->b_data + sizeof(struct dlfs), 0, 2396 LFS_SBPAD - sizeof(struct dlfs)); 2397 *(struct dlfs *)bp->b_data = fs->lfs_dlfs; 2398 2399 bp->b_cflags |= BC_BUSY; 2400 bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC; 2401 bp->b_oflags &= ~(BO_DONE | BO_DELWRI); 2402 bp->b_error = 0; 2403 bp->b_iodone = lfs_supercallback; 2404 2405 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2406 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2407 else 2408 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2409 curlwp->l_ru.ru_oublock++; 2410 2411 mutex_enter(devvp->v_interlock); 2412 devvp->v_numoutput++; 2413 mutex_exit(devvp->v_interlock); 2414 2415 mutex_enter(&lfs_lock); 2416 ++fs->lfs_iocount; 2417 mutex_exit(&lfs_lock); 2418 VOP_STRATEGY(devvp, bp); 2419} 2420 2421/* 2422 * Logical block number match routines used when traversing the dirty block 2423 * chain. 2424 */ 2425int 2426lfs_match_fake(struct lfs *fs, struct buf *bp) 2427{ 2428 2429 ASSERT_SEGLOCK(fs); 2430 return LFS_IS_MALLOC_BUF(bp); 2431} 2432 2433#if 0 2434int 2435lfs_match_real(struct lfs *fs, struct buf *bp) 2436{ 2437 2438 ASSERT_SEGLOCK(fs); 2439 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2440} 2441#endif 2442 2443int 2444lfs_match_data(struct lfs *fs, struct buf *bp) 2445{ 2446 2447 ASSERT_SEGLOCK(fs); 2448 return (bp->b_lblkno >= 0); 2449} 2450 2451int 2452lfs_match_indir(struct lfs *fs, struct buf *bp) 2453{ 2454 daddr_t lbn; 2455 2456 ASSERT_SEGLOCK(fs); 2457 lbn = bp->b_lblkno; 2458 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0); 2459} 2460 2461int 2462lfs_match_dindir(struct lfs *fs, struct buf *bp) 2463{ 2464 daddr_t lbn; 2465 2466 ASSERT_SEGLOCK(fs); 2467 lbn = bp->b_lblkno; 2468 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1); 2469} 2470 2471int 2472lfs_match_tindir(struct lfs *fs, struct buf *bp) 2473{ 2474 daddr_t lbn; 2475 2476 ASSERT_SEGLOCK(fs); 2477 lbn = bp->b_lblkno; 2478 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2); 2479} 2480 2481static void 2482lfs_free_aiodone(struct buf *bp) 2483{ 2484 struct lfs *fs; 2485 2486 KERNEL_LOCK(1, curlwp); 2487 fs = bp->b_private; 2488 ASSERT_NO_SEGLOCK(fs); 2489 lfs_freebuf(fs, bp); 2490 KERNEL_UNLOCK_LAST(curlwp); 2491} 2492 2493static void 2494lfs_super_aiodone(struct buf *bp) 2495{ 2496 struct lfs *fs; 2497 2498 KERNEL_LOCK(1, curlwp); 2499 fs = bp->b_private; 2500 ASSERT_NO_SEGLOCK(fs); 2501 mutex_enter(&lfs_lock); 2502 fs->lfs_sbactive = 0; 2503 if (--fs->lfs_iocount <= 1) 2504 wakeup(&fs->lfs_iocount); 2505 wakeup(&fs->lfs_sbactive); 2506 mutex_exit(&lfs_lock); 2507 lfs_freebuf(fs, bp); 2508 KERNEL_UNLOCK_LAST(curlwp); 2509} 2510 2511static void 2512lfs_cluster_aiodone(struct buf *bp) 2513{ 2514 struct lfs_cluster *cl; 2515 struct lfs *fs; 2516 struct buf *tbp, *fbp; 2517 struct vnode *vp, *devvp, *ovp; 2518 struct inode *ip; 2519 int error; 2520 2521 KERNEL_LOCK(1, curlwp); 2522 2523 error = bp->b_error; 2524 cl = bp->b_private; 2525 fs = cl->fs; 2526 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2527 ASSERT_NO_SEGLOCK(fs); 2528 2529 /* Put the pages back, and release the buffer */ 2530 while (cl->bufcount--) { 2531 tbp = cl->bpp[cl->bufcount]; 2532 KASSERT(tbp->b_cflags & BC_BUSY); 2533 if (error) { 2534 tbp->b_error = error; 2535 } 2536 2537 /* 2538 * We're done with tbp. If it has not been re-dirtied since 2539 * the cluster was written, free it. Otherwise, keep it on 2540 * the locked list to be written again. 2541 */ 2542 vp = tbp->b_vp; 2543 2544 tbp->b_flags &= ~B_GATHERED; 2545 2546 LFS_BCLEAN_LOG(fs, tbp); 2547 2548 mutex_enter(&bufcache_lock); 2549 if (tbp->b_iodone == NULL) { 2550 KASSERT(tbp->b_flags & B_LOCKED); 2551 bremfree(tbp); 2552 if (vp) { 2553 mutex_enter(vp->v_interlock); 2554 reassignbuf(tbp, vp); 2555 mutex_exit(vp->v_interlock); 2556 } 2557 tbp->b_flags |= B_ASYNC; /* for biodone */ 2558 } 2559 2560 if (((tbp->b_flags | tbp->b_oflags) & 2561 (B_LOCKED | BO_DELWRI)) == B_LOCKED) 2562 LFS_UNLOCK_BUF(tbp); 2563 2564 if (tbp->b_oflags & BO_DONE) { 2565 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2566 cl->bufcount, (long)tbp->b_flags)); 2567 } 2568 2569 if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) { 2570 /* 2571 * A buffer from the page daemon. 2572 * We use the same iodone as it does, 2573 * so we must manually disassociate its 2574 * buffers from the vp. 2575 */ 2576 if ((ovp = tbp->b_vp) != NULL) { 2577 /* This is just silly */ 2578 mutex_enter(ovp->v_interlock); 2579 brelvp(tbp); 2580 mutex_exit(ovp->v_interlock); 2581 tbp->b_vp = vp; 2582 tbp->b_objlock = vp->v_interlock; 2583 } 2584 /* Put it back the way it was */ 2585 tbp->b_flags |= B_ASYNC; 2586 /* Master buffers have BC_AGE */ 2587 if (tbp->b_private == tbp) 2588 tbp->b_cflags |= BC_AGE; 2589 } 2590 mutex_exit(&bufcache_lock); 2591 2592 biodone(tbp); 2593 2594 /* 2595 * If this is the last block for this vnode, but 2596 * there are other blocks on its dirty list, 2597 * set IN_MODIFIED/IN_CLEANING depending on what 2598 * sort of block. Only do this for our mount point, 2599 * not for, e.g., inode blocks that are attached to 2600 * the devvp. 2601 * XXX KS - Shouldn't we set *both* if both types 2602 * of blocks are present (traverse the dirty list?) 2603 */ 2604 mutex_enter(vp->v_interlock); 2605 mutex_enter(&lfs_lock); 2606 if (vp != devvp && vp->v_numoutput == 0 && 2607 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2608 ip = VTOI(vp); 2609 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2610 ip->i_number)); 2611 if (LFS_IS_MALLOC_BUF(fbp)) 2612 LFS_SET_UINO(ip, IN_CLEANING); 2613 else 2614 LFS_SET_UINO(ip, IN_MODIFIED); 2615 } 2616 cv_broadcast(&vp->v_cv); 2617 mutex_exit(&lfs_lock); 2618 mutex_exit(vp->v_interlock); 2619 } 2620 2621 /* Fix up the cluster buffer, and release it */ 2622 if (cl->flags & LFS_CL_MALLOC) 2623 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2624 putiobuf(bp); 2625 2626 /* Note i/o done */ 2627 if (cl->flags & LFS_CL_SYNC) { 2628 if (--cl->seg->seg_iocount == 0) 2629 wakeup(&cl->seg->seg_iocount); 2630 } 2631 mutex_enter(&lfs_lock); 2632#ifdef DIAGNOSTIC 2633 if (fs->lfs_iocount == 0) 2634 panic("lfs_cluster_aiodone: zero iocount"); 2635#endif 2636 if (--fs->lfs_iocount <= 1) 2637 wakeup(&fs->lfs_iocount); 2638 mutex_exit(&lfs_lock); 2639 2640 KERNEL_UNLOCK_LAST(curlwp); 2641 2642 pool_put(&fs->lfs_bpppool, cl->bpp); 2643 cl->bpp = NULL; 2644 pool_put(&fs->lfs_clpool, cl); 2645} 2646 2647static void 2648lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2649{ 2650 /* reset b_iodone for when this is a single-buf i/o. */ 2651 bp->b_iodone = aiodone; 2652 2653 workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL); 2654} 2655 2656static void 2657lfs_cluster_callback(struct buf *bp) 2658{ 2659 2660 lfs_generic_callback(bp, lfs_cluster_aiodone); 2661} 2662 2663void 2664lfs_supercallback(struct buf *bp) 2665{ 2666 2667 lfs_generic_callback(bp, lfs_super_aiodone); 2668} 2669 2670/* 2671 * The only buffers that are going to hit these functions are the 2672 * segment write blocks, or the segment summaries, or the superblocks. 2673 * 2674 * All of the above are created by lfs_newbuf, and so do not need to be 2675 * released via brelse. 2676 */ 2677void 2678lfs_callback(struct buf *bp) 2679{ 2680 2681 lfs_generic_callback(bp, lfs_free_aiodone); 2682} 2683 2684/* 2685 * Shellsort (diminishing increment sort) from Data Structures and 2686 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2687 * see also Knuth Vol. 3, page 84. The increments are selected from 2688 * formula (8), page 95. Roughly O(N^3/2). 2689 */ 2690/* 2691 * This is our own private copy of shellsort because we want to sort 2692 * two parallel arrays (the array of buffer pointers and the array of 2693 * logical block numbers) simultaneously. Note that we cast the array 2694 * of logical block numbers to a unsigned in this routine so that the 2695 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2696 */ 2697 2698void 2699lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size) 2700{ 2701 static int __rsshell_increments[] = { 4, 1, 0 }; 2702 int incr, *incrp, t1, t2; 2703 struct buf *bp_temp; 2704 2705#ifdef DEBUG 2706 incr = 0; 2707 for (t1 = 0; t1 < nmemb; t1++) { 2708 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2709 if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) { 2710 /* dump before panic */ 2711 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2712 nmemb, size); 2713 incr = 0; 2714 for (t1 = 0; t1 < nmemb; t1++) { 2715 const struct buf *bp = bp_array[t1]; 2716 2717 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2718 PRIu64 "\n", t1, 2719 (uint64_t)bp->b_bcount, 2720 (uint64_t)bp->b_lblkno); 2721 printf("lbns:"); 2722 for (t2 = 0; t2 * size < bp->b_bcount; 2723 t2++) { 2724 printf(" %" PRId32, 2725 lb_array[incr++]); 2726 } 2727 printf("\n"); 2728 } 2729 panic("lfs_shellsort: inconsistent input"); 2730 } 2731 } 2732 } 2733#endif 2734 2735 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2736 for (t1 = incr; t1 < nmemb; ++t1) 2737 for (t2 = t1 - incr; t2 >= 0;) 2738 if ((u_int32_t)bp_array[t2]->b_lblkno > 2739 (u_int32_t)bp_array[t2 + incr]->b_lblkno) { 2740 bp_temp = bp_array[t2]; 2741 bp_array[t2] = bp_array[t2 + incr]; 2742 bp_array[t2 + incr] = bp_temp; 2743 t2 -= incr; 2744 } else 2745 break; 2746 2747 /* Reform the list of logical blocks */ 2748 incr = 0; 2749 for (t1 = 0; t1 < nmemb; t1++) { 2750 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2751 lb_array[incr++] = bp_array[t1]->b_lblkno + t2; 2752 } 2753 } 2754} 2755 2756/* 2757 * Call vget with LK_NOWAIT. If we are the one who holds VI_XLOCK, 2758 * however, we must press on. Just fake success in that case. 2759 */ 2760int 2761lfs_vref(struct vnode *vp) 2762{ 2763 struct lfs *fs; 2764 2765 KASSERT(mutex_owned(vp->v_interlock)); 2766 2767 fs = VTOI(vp)->i_lfs; 2768 2769 ASSERT_MAYBE_SEGLOCK(fs); 2770 2771 /* 2772 * If we return 1 here during a flush, we risk vinvalbuf() not 2773 * being able to flush all of the pages from this vnode, which 2774 * will cause it to panic. So, return 0 if a flush is in progress. 2775 */ 2776 if (IS_FLUSHING(VTOI(vp)->i_lfs, vp)) { 2777 ++fs->lfs_flushvp_fakevref; 2778 mutex_exit(vp->v_interlock); 2779 return 0; 2780 } 2781 2782 return vget(vp, LK_NOWAIT); 2783} 2784 2785/* 2786 * This is vrele except that we do not want to VOP_INACTIVE this vnode. We 2787 * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end. 2788 */ 2789void 2790lfs_vunref(struct vnode *vp) 2791{ 2792 struct lfs *fs; 2793 2794 fs = VTOI(vp)->i_lfs; 2795 ASSERT_MAYBE_SEGLOCK(fs); 2796 2797 /* 2798 * Analogous to lfs_vref, if the node is flushing, fake it. 2799 */ 2800 if (IS_FLUSHING(fs, vp) && fs->lfs_flushvp_fakevref) { 2801 --fs->lfs_flushvp_fakevref; 2802 return; 2803 } 2804 2805 /* does not call inactive */ 2806 mutex_enter(vp->v_interlock); 2807 vrelel(vp, 0); 2808} 2809 2810/* 2811 * We use this when we have vnodes that were loaded in solely for cleaning. 2812 * There is no reason to believe that these vnodes will be referenced again 2813 * soon, since the cleaning process is unrelated to normal filesystem 2814 * activity. Putting cleaned vnodes at the tail of the list has the effect 2815 * of flushing the vnode LRU. So, put vnodes that were loaded only for 2816 * cleaning at the head of the list, instead. 2817 */ 2818void 2819lfs_vunref_head(struct vnode *vp) 2820{ 2821 2822 ASSERT_SEGLOCK(VTOI(vp)->i_lfs); 2823 2824 /* does not call inactive, inserts non-held vnode at head of freelist */ 2825 mutex_enter(vp->v_interlock); 2826 vrelel(vp, 0); 2827} 2828 2829 2830/* 2831 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2832 * point at uninitialized space. 2833 */ 2834void 2835lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2836{ 2837 struct segment *sp = fs->lfs_sp; 2838 2839 KASSERT(vers > 0); 2840 2841 if (sp->seg_bytes_left < fs->lfs_bsize || 2842 sp->sum_bytes_left < sizeof(struct finfo)) 2843 (void) lfs_writeseg(fs, fs->lfs_sp); 2844 2845 sp->sum_bytes_left -= FINFOSIZE; 2846 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 2847 sp->fip->fi_nblocks = 0; 2848 sp->fip->fi_ino = ino; 2849 sp->fip->fi_version = vers; 2850} 2851 2852/* 2853 * Release the FINFO entry, either clearing out an unused entry or 2854 * advancing us to the next available entry. 2855 */ 2856void 2857lfs_release_finfo(struct lfs *fs) 2858{ 2859 struct segment *sp = fs->lfs_sp; 2860 2861 if (sp->fip->fi_nblocks != 0) { 2862 sp->fip = (FINFO*)((char *)sp->fip + FINFOSIZE + 2863 sizeof(int32_t) * sp->fip->fi_nblocks); 2864 sp->start_lbp = &sp->fip->fi_blocks[0]; 2865 } else { 2866 sp->sum_bytes_left += FINFOSIZE; 2867 --((SEGSUM *)(sp->segsum))->ss_nfinfo; 2868 } 2869} 2870