1 /* $NetBSD: lfs_segment.c,v 1.193 2006/10/12 01:32:51 christos 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 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 /* 39 * Copyright (c) 1991, 1993 40 * The Regents of the University of California. All rights reserved. 41 * 42 * Redistribution and use in source and binary forms, with or without 43 * modification, are permitted provided that the following conditions 44 * are met: 45 * 1. Redistributions of source code must retain the above copyright 46 * notice, this list of conditions and the following disclaimer. 47 * 2. Redistributions in binary form must reproduce the above copyright 48 * notice, this list of conditions and the following disclaimer in the 49 * documentation and/or other materials provided with the distribution. 50 * 3. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)lfs_segment.c 8.10 (Berkeley) 6/10/95 67 */ 68 69 #include <sys/cdefs.h> 70 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.193 2006/10/12 01:32:51 christos Exp $"); 71 72 #ifdef DEBUG 73 # define vndebug(vp, str) do { \ 74 if (VTOI(vp)->i_flag & IN_CLEANING) \ 75 DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \ 76 VTOI(vp)->i_number, (str), op)); \ 77 } while(0) 78 #else 79 # define vndebug(vp, str) 80 #endif 81 #define ivndebug(vp, str) \ 82 DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str))) 83 84 #if defined(_KERNEL_OPT) 85 #include "opt_ddb.h" 86 #endif 87 88 #include <sys/param.h> 89 #include <sys/systm.h> 90 #include <sys/namei.h> 91 #include <sys/kernel.h> 92 #include <sys/resourcevar.h> 93 #include <sys/file.h> 94 #include <sys/stat.h> 95 #include <sys/buf.h> 96 #include <sys/proc.h> 97 #include <sys/vnode.h> 98 #include <sys/mount.h> 99 #include <sys/kauth.h> 100 #include <sys/syslog.h> 101 102 #include <miscfs/specfs/specdev.h> 103 #include <miscfs/fifofs/fifo.h> 104 105 #include <ufs/ufs/inode.h> 106 #include <ufs/ufs/dir.h> 107 #include <ufs/ufs/ufsmount.h> 108 #include <ufs/ufs/ufs_extern.h> 109 110 #include <ufs/lfs/lfs.h> 111 #include <ufs/lfs/lfs_extern.h> 112 113 #include <uvm/uvm.h> 114 #include <uvm/uvm_extern.h> 115 116 MALLOC_DEFINE(M_SEGMENT, "LFS segment", "Segment for LFS"); 117 118 extern int count_lock_queue(void); 119 extern struct simplelock vnode_free_list_slock; /* XXX */ 120 extern struct simplelock bqueue_slock; /* XXX */ 121 122 static void lfs_generic_callback(struct buf *, void (*)(struct buf *)); 123 static void lfs_free_aiodone(struct buf *); 124 static void lfs_super_aiodone(struct buf *); 125 static void lfs_cluster_aiodone(struct buf *); 126 static void lfs_cluster_callback(struct buf *); 127 128 /* 129 * Determine if it's OK to start a partial in this segment, or if we need 130 * to go on to a new segment. 131 */ 132 #define LFS_PARTIAL_FITS(fs) \ 133 ((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \ 134 fragstofsb((fs), (fs)->lfs_frag)) 135 136 /* 137 * Figure out whether we should do a checkpoint write or go ahead with 138 * an ordinary write. 139 */ 140 #define LFS_SHOULD_CHECKPOINT(fs, flags) \ 141 ((flags & SEGM_CLEAN) == 0 && \ 142 ((fs->lfs_nactive > LFS_MAX_ACTIVE || \ 143 (flags & SEGM_CKP) || \ 144 fs->lfs_nclean < LFS_MAX_ACTIVE))) 145 146 int lfs_match_fake(struct lfs *, struct buf *); 147 void lfs_newseg(struct lfs *); 148 /* XXX ondisk32 */ 149 void lfs_shellsort(struct buf **, int32_t *, int, int); 150 void lfs_supercallback(struct buf *); 151 void lfs_updatemeta(struct segment *); 152 void lfs_writesuper(struct lfs *, daddr_t); 153 int lfs_writevnodes(struct lfs *fs, struct mount *mp, 154 struct segment *sp, int dirops); 155 156 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 157 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 158 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 159 int lfs_dirvcount = 0; /* # active dirops */ 160 161 /* Statistics Counters */ 162 int lfs_dostats = 1; 163 struct lfs_stats lfs_stats; 164 165 /* op values to lfs_writevnodes */ 166 #define VN_REG 0 167 #define VN_DIROP 1 168 #define VN_EMPTY 2 169 #define VN_CLEAN 3 170 171 /* 172 * XXX KS - Set modification time on the Ifile, so the cleaner can 173 * read the fs mod time off of it. We don't set IN_UPDATE here, 174 * since we don't really need this to be flushed to disk (and in any 175 * case that wouldn't happen to the Ifile until we checkpoint). 176 */ 177 void 178 lfs_imtime(struct lfs *fs) 179 { 180 struct timespec ts; 181 struct inode *ip; 182 183 ASSERT_MAYBE_SEGLOCK(fs); 184 vfs_timestamp(&ts); 185 ip = VTOI(fs->lfs_ivnode); 186 ip->i_ffs1_mtime = ts.tv_sec; 187 ip->i_ffs1_mtimensec = ts.tv_nsec; 188 } 189 190 /* 191 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 192 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 193 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 194 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 195 */ 196 197 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 198 199 int 200 lfs_vflush(struct vnode *vp) 201 { 202 struct inode *ip; 203 struct lfs *fs; 204 struct segment *sp; 205 struct buf *bp, *nbp, *tbp, *tnbp; 206 int error, s; 207 int flushed; 208 int relock; 209 int loopcount; 210 211 ip = VTOI(vp); 212 fs = VFSTOUFS(vp->v_mount)->um_lfs; 213 relock = 0; 214 215 top: 216 ASSERT_NO_SEGLOCK(fs); 217 if (ip->i_flag & IN_CLEANING) { 218 ivndebug(vp,"vflush/in_cleaning"); 219 LFS_CLR_UINO(ip, IN_CLEANING); 220 LFS_SET_UINO(ip, IN_MODIFIED); 221 222 /* 223 * Toss any cleaning buffers that have real counterparts 224 * to avoid losing new data. 225 */ 226 s = splbio(); 227 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 228 nbp = LIST_NEXT(bp, b_vnbufs); 229 if (!LFS_IS_MALLOC_BUF(bp)) 230 continue; 231 /* 232 * Look for pages matching the range covered 233 * by cleaning blocks. It's okay if more dirty 234 * pages appear, so long as none disappear out 235 * from under us. 236 */ 237 if (bp->b_lblkno > 0 && vp->v_type == VREG && 238 vp != fs->lfs_ivnode) { 239 struct vm_page *pg; 240 voff_t off; 241 242 simple_lock(&vp->v_interlock); 243 for (off = lblktosize(fs, bp->b_lblkno); 244 off < lblktosize(fs, bp->b_lblkno + 1); 245 off += PAGE_SIZE) { 246 pg = uvm_pagelookup(&vp->v_uobj, off); 247 if (pg == NULL) 248 continue; 249 if ((pg->flags & PG_CLEAN) == 0 || 250 pmap_is_modified(pg)) { 251 fs->lfs_avail += btofsb(fs, 252 bp->b_bcount); 253 wakeup(&fs->lfs_avail); 254 lfs_freebuf(fs, bp); 255 bp = NULL; 256 simple_unlock(&vp->v_interlock); 257 goto nextbp; 258 } 259 } 260 simple_unlock(&vp->v_interlock); 261 } 262 for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp; 263 tbp = tnbp) 264 { 265 tnbp = LIST_NEXT(tbp, b_vnbufs); 266 if (tbp->b_vp == bp->b_vp 267 && tbp->b_lblkno == bp->b_lblkno 268 && tbp != bp) 269 { 270 fs->lfs_avail += btofsb(fs, 271 bp->b_bcount); 272 wakeup(&fs->lfs_avail); 273 lfs_freebuf(fs, bp); 274 bp = NULL; 275 break; 276 } 277 } 278 nextbp: 279 ; 280 } 281 splx(s); 282 } 283 284 /* If the node is being written, wait until that is done */ 285 simple_lock(&vp->v_interlock); 286 s = splbio(); 287 if (WRITEINPROG(vp)) { 288 ivndebug(vp,"vflush/writeinprog"); 289 ltsleep(vp, (PRIBIO+1), "lfs_vw", 0, &vp->v_interlock); 290 } 291 splx(s); 292 simple_unlock(&vp->v_interlock); 293 294 /* Protect against VXLOCK deadlock in vinvalbuf() */ 295 lfs_seglock(fs, SEGM_SYNC); 296 297 /* If we're supposed to flush a freed inode, just toss it */ 298 if (ip->i_lfs_iflags & LFSI_DELETED) { 299 DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n", 300 ip->i_number)); 301 s = splbio(); 302 /* Drain v_numoutput */ 303 simple_lock(&global_v_numoutput_slock); 304 while (vp->v_numoutput > 0) { 305 vp->v_flag |= VBWAIT; 306 ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf4", 0, 307 &global_v_numoutput_slock); 308 } 309 simple_unlock(&global_v_numoutput_slock); 310 KASSERT(vp->v_numoutput == 0); 311 312 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 313 nbp = LIST_NEXT(bp, b_vnbufs); 314 315 KASSERT((bp->b_flags & B_GATHERED) == 0); 316 if (bp->b_flags & B_DELWRI) { /* XXX always true? */ 317 fs->lfs_avail += btofsb(fs, bp->b_bcount); 318 wakeup(&fs->lfs_avail); 319 } 320 /* Copied from lfs_writeseg */ 321 if (bp->b_flags & B_CALL) { 322 biodone(bp); 323 } else { 324 bremfree(bp); 325 LFS_UNLOCK_BUF(bp); 326 bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | 327 B_GATHERED); 328 bp->b_flags |= B_DONE; 329 reassignbuf(bp, vp); 330 brelse(bp); 331 } 332 } 333 splx(s); 334 LFS_CLR_UINO(ip, IN_CLEANING); 335 LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED); 336 ip->i_flag &= ~IN_ALLMOD; 337 DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n", 338 ip->i_number)); 339 lfs_segunlock(fs); 340 341 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 342 343 return 0; 344 } 345 346 fs->lfs_flushvp = vp; 347 if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) { 348 error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC); 349 fs->lfs_flushvp = NULL; 350 KASSERT(fs->lfs_flushvp_fakevref == 0); 351 lfs_segunlock(fs); 352 353 /* Make sure that any pending buffers get written */ 354 s = splbio(); 355 simple_lock(&global_v_numoutput_slock); 356 while (vp->v_numoutput > 0) { 357 vp->v_flag |= VBWAIT; 358 ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf3", 0, 359 &global_v_numoutput_slock); 360 } 361 simple_unlock(&global_v_numoutput_slock); 362 splx(s); 363 364 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 365 KASSERT(vp->v_numoutput == 0); 366 367 return error; 368 } 369 sp = fs->lfs_sp; 370 371 flushed = 0; 372 if (VPISEMPTY(vp)) { 373 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 374 ++flushed; 375 } else if ((ip->i_flag & IN_CLEANING) && 376 (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 377 ivndebug(vp,"vflush/clean"); 378 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 379 ++flushed; 380 } else if (lfs_dostats) { 381 if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD)) 382 ++lfs_stats.vflush_invoked; 383 ivndebug(vp,"vflush"); 384 } 385 386 #ifdef DIAGNOSTIC 387 if (vp->v_flag & VDIROP) { 388 DLOG((DLOG_VNODE, "lfs_vflush: flushing VDIROP\n")); 389 /* panic("lfs_vflush: VDIROP being flushed...this can\'t happen"); */ 390 } 391 if (vp->v_usecount < 0) { 392 printf("usecount=%ld\n", (long)vp->v_usecount); 393 panic("lfs_vflush: usecount<0"); 394 } 395 #endif 396 397 do { 398 loopcount = 0; 399 do { 400 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 401 relock = lfs_writefile(fs, sp, vp); 402 if (relock) { 403 /* 404 * Might have to wait for the 405 * cleaner to run; but we're 406 * still not done with this vnode. 407 */ 408 KDASSERT(ip->i_number != LFS_IFILE_INUM); 409 lfs_writeinode(fs, sp, ip); 410 LFS_SET_UINO(ip, IN_MODIFIED); 411 lfs_writeseg(fs, sp); 412 lfs_segunlock(fs); 413 lfs_segunlock_relock(fs); 414 goto top; 415 } 416 } 417 /* 418 * If we begin a new segment in the middle of writing 419 * the Ifile, it creates an inconsistent checkpoint, 420 * since the Ifile information for the new segment 421 * is not up-to-date. Take care of this here by 422 * sending the Ifile through again in case there 423 * are newly dirtied blocks. But wait, there's more! 424 * This second Ifile write could *also* cross a segment 425 * boundary, if the first one was large. The second 426 * one is guaranteed to be no more than 8 blocks, 427 * though (two segment blocks and supporting indirects) 428 * so the third write *will not* cross the boundary. 429 */ 430 if (vp == fs->lfs_ivnode) { 431 lfs_writefile(fs, sp, vp); 432 lfs_writefile(fs, sp, vp); 433 } 434 #ifdef DEBUG 435 if (++loopcount > 2) 436 log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount); 437 #endif 438 } while (lfs_writeinode(fs, sp, ip)); 439 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 440 441 if (lfs_dostats) { 442 ++lfs_stats.nwrites; 443 if (sp->seg_flags & SEGM_SYNC) 444 ++lfs_stats.nsync_writes; 445 if (sp->seg_flags & SEGM_CKP) 446 ++lfs_stats.ncheckpoints; 447 } 448 /* 449 * If we were called from somewhere that has already held the seglock 450 * (e.g., lfs_markv()), the lfs_segunlock will not wait for 451 * the write to complete because we are still locked. 452 * Since lfs_vflush() must return the vnode with no dirty buffers, 453 * we must explicitly wait, if that is the case. 454 * 455 * We compare the iocount against 1, not 0, because it is 456 * artificially incremented by lfs_seglock(). 457 */ 458 simple_lock(&fs->lfs_interlock); 459 if (fs->lfs_seglock > 1) { 460 while (fs->lfs_iocount > 1) 461 (void)ltsleep(&fs->lfs_iocount, PRIBIO + 1, 462 "lfs_vflush", 0, &fs->lfs_interlock); 463 } 464 simple_unlock(&fs->lfs_interlock); 465 466 lfs_segunlock(fs); 467 468 /* Wait for these buffers to be recovered by aiodoned */ 469 s = splbio(); 470 simple_lock(&global_v_numoutput_slock); 471 while (vp->v_numoutput > 0) { 472 vp->v_flag |= VBWAIT; 473 ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf2", 0, 474 &global_v_numoutput_slock); 475 } 476 simple_unlock(&global_v_numoutput_slock); 477 splx(s); 478 479 KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL); 480 KASSERT(vp->v_numoutput == 0); 481 482 fs->lfs_flushvp = NULL; 483 KASSERT(fs->lfs_flushvp_fakevref == 0); 484 485 return (0); 486 } 487 488 int 489 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op) 490 { 491 struct inode *ip; 492 struct vnode *vp, *nvp; 493 int inodes_written = 0, only_cleaning; 494 int error = 0; 495 496 ASSERT_SEGLOCK(fs); 497 #ifndef LFS_NO_BACKVP_HACK 498 /* BEGIN HACK */ 499 #define VN_OFFSET \ 500 (((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp) 501 #define BACK_VP(VP) \ 502 ((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET)) 503 #define BEG_OF_VLIST \ 504 ((struct vnode *)(((caddr_t)&LIST_FIRST(&mp->mnt_vnodelist)) \ 505 - VN_OFFSET)) 506 507 /* Find last vnode. */ 508 loop: for (vp = LIST_FIRST(&mp->mnt_vnodelist); 509 vp && LIST_NEXT(vp, v_mntvnodes) != NULL; 510 vp = LIST_NEXT(vp, v_mntvnodes)); 511 for (; vp && vp != BEG_OF_VLIST; vp = nvp) { 512 nvp = BACK_VP(vp); 513 #else 514 loop: 515 for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) { 516 nvp = LIST_NEXT(vp, v_mntvnodes); 517 #endif 518 /* 519 * If the vnode that we are about to sync is no longer 520 * associated with this mount point, start over. 521 */ 522 if (vp->v_mount != mp) { 523 DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n")); 524 /* 525 * After this, pages might be busy 526 * due to our own previous putpages. 527 * Start actual segment write here to avoid deadlock. 528 */ 529 (void)lfs_writeseg(fs, sp); 530 goto loop; 531 } 532 533 if (vp->v_type == VNON) { 534 continue; 535 } 536 537 ip = VTOI(vp); 538 if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) || 539 (op != VN_DIROP && op != VN_CLEAN && 540 (vp->v_flag & VDIROP))) { 541 vndebug(vp,"dirop"); 542 continue; 543 } 544 545 if (op == VN_EMPTY && !VPISEMPTY(vp)) { 546 vndebug(vp,"empty"); 547 continue; 548 } 549 550 if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM 551 && vp != fs->lfs_flushvp 552 && !(ip->i_flag & IN_CLEANING)) { 553 vndebug(vp,"cleaning"); 554 continue; 555 } 556 557 if (lfs_vref(vp)) { 558 vndebug(vp,"vref"); 559 continue; 560 } 561 562 only_cleaning = 0; 563 /* 564 * Write the inode/file if dirty and it's not the IFILE. 565 */ 566 if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) { 567 only_cleaning = 568 ((ip->i_flag & IN_ALLMOD) == IN_CLEANING); 569 570 if (ip->i_number != LFS_IFILE_INUM) { 571 error = lfs_writefile(fs, sp, vp); 572 if (error) { 573 lfs_vunref(vp); 574 if (error == EAGAIN) { 575 /* 576 * This error from lfs_putpages 577 * indicates we need to drop 578 * the segment lock and start 579 * over after the cleaner has 580 * had a chance to run. 581 */ 582 lfs_writeinode(fs, sp, ip); 583 lfs_writeseg(fs, sp); 584 if (!VPISEMPTY(vp) && 585 !WRITEINPROG(vp) && 586 !(ip->i_flag & IN_ALLMOD)) 587 LFS_SET_UINO(ip, IN_MODIFIED); 588 break; 589 } 590 error = 0; /* XXX not quite right */ 591 continue; 592 } 593 594 if (!VPISEMPTY(vp)) { 595 if (WRITEINPROG(vp)) { 596 ivndebug(vp,"writevnodes/write2"); 597 } else if (!(ip->i_flag & IN_ALLMOD)) { 598 LFS_SET_UINO(ip, IN_MODIFIED); 599 } 600 } 601 (void) lfs_writeinode(fs, sp, ip); 602 inodes_written++; 603 } 604 } 605 606 if (lfs_clean_vnhead && only_cleaning) 607 lfs_vunref_head(vp); 608 else 609 lfs_vunref(vp); 610 } 611 return error; 612 } 613 614 /* 615 * Do a checkpoint. 616 */ 617 int 618 lfs_segwrite(struct mount *mp, int flags) 619 { 620 struct buf *bp; 621 struct inode *ip; 622 struct lfs *fs; 623 struct segment *sp; 624 struct vnode *vp; 625 SEGUSE *segusep; 626 int do_ckp, did_ckp, error, s; 627 unsigned n, segleft, maxseg, sn, i, curseg; 628 int writer_set = 0; 629 int dirty; 630 int redo; 631 int um_error; 632 int loopcount; 633 634 fs = VFSTOUFS(mp)->um_lfs; 635 ASSERT_MAYBE_SEGLOCK(fs); 636 637 if (fs->lfs_ronly) 638 return EROFS; 639 640 lfs_imtime(fs); 641 642 /* 643 * Allocate a segment structure and enough space to hold pointers to 644 * the maximum possible number of buffers which can be described in a 645 * single summary block. 646 */ 647 do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags); 648 649 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 650 sp = fs->lfs_sp; 651 if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP)) 652 do_ckp = 1; 653 654 /* 655 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 656 * in which case we have to flush *all* buffers off of this vnode. 657 * We don't care about other nodes, but write any non-dirop nodes 658 * anyway in anticipation of another getnewvnode(). 659 * 660 * If we're cleaning we only write cleaning and ifile blocks, and 661 * no dirops, since otherwise we'd risk corruption in a crash. 662 */ 663 if (sp->seg_flags & SEGM_CLEAN) 664 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 665 else if (!(sp->seg_flags & SEGM_FORCE_CKP)) { 666 do { 667 um_error = lfs_writevnodes(fs, mp, sp, VN_REG); 668 if (!fs->lfs_dirops || !fs->lfs_flushvp) { 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, 698 fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &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 != 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); 723 segleft -= fs->lfs_sepb; 724 curseg += fs->lfs_sepb; 725 } 726 } 727 728 LOCK_ASSERT(LFS_SEGLOCK_HELD(fs)); 729 730 did_ckp = 0; 731 if (do_ckp || fs->lfs_doifile) { 732 vp = fs->lfs_ivnode; 733 vn_lock(vp, LK_EXCLUSIVE); 734 loopcount = 0; 735 do { 736 #ifdef DEBUG 737 LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid); 738 #endif 739 simple_lock(&fs->lfs_interlock); 740 fs->lfs_flags &= ~LFS_IFDIRTY; 741 simple_unlock(&fs->lfs_interlock); 742 743 ip = VTOI(vp); 744 745 if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) { 746 /* 747 * Ifile has no pages, so we don't need 748 * to check error return here. 749 */ 750 lfs_writefile(fs, sp, vp); 751 /* 752 * Ensure the Ifile takes the current segment 753 * into account. See comment in lfs_vflush. 754 */ 755 lfs_writefile(fs, sp, vp); 756 lfs_writefile(fs, sp, vp); 757 } 758 759 if (ip->i_flag & IN_ALLMOD) 760 ++did_ckp; 761 #if 0 762 redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0); 763 #else 764 redo = lfs_writeinode(fs, sp, ip); 765 #endif 766 redo += lfs_writeseg(fs, sp); 767 simple_lock(&fs->lfs_interlock); 768 redo += (fs->lfs_flags & LFS_IFDIRTY); 769 simple_unlock(&fs->lfs_interlock); 770 #ifdef DEBUG 771 if (++loopcount > 2) 772 log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n", 773 loopcount); 774 #endif 775 } while (redo && do_ckp); 776 777 /* 778 * Unless we are unmounting, the Ifile may continue to have 779 * dirty blocks even after a checkpoint, due to changes to 780 * inodes' atime. If we're checkpointing, it's "impossible" 781 * for other parts of the Ifile to be dirty after the loop 782 * above, since we hold the segment lock. 783 */ 784 s = splbio(); 785 if (LIST_EMPTY(&vp->v_dirtyblkhd)) { 786 LFS_CLR_UINO(ip, IN_ALLMOD); 787 } 788 #ifdef DIAGNOSTIC 789 else if (do_ckp) { 790 int do_panic = 0; 791 LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 792 if (bp->b_lblkno < fs->lfs_cleansz + 793 fs->lfs_segtabsz && 794 !(bp->b_flags & B_GATHERED)) { 795 printf("ifile lbn %ld still dirty (flags %lx)\n", 796 (long)bp->b_lblkno, 797 (long)bp->b_flags); 798 ++do_panic; 799 } 800 } 801 if (do_panic) 802 panic("dirty blocks"); 803 } 804 #endif 805 splx(s); 806 VOP_UNLOCK(vp, 0); 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 finfo *fip; 846 struct inode *ip; 847 int i, frag; 848 int error; 849 850 ASSERT_SEGLOCK(fs); 851 error = 0; 852 ip = VTOI(vp); 853 854 fip = sp->fip; 855 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 856 857 if (vp->v_flag & VDIROP) 858 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 859 860 if (sp->seg_flags & SEGM_CLEAN) { 861 lfs_gather(fs, sp, vp, lfs_match_fake); 862 /* 863 * For a file being flushed, we need to write *all* blocks. 864 * This means writing the cleaning blocks first, and then 865 * immediately following with any non-cleaning blocks. 866 * The same is true of the Ifile since checkpoints assume 867 * that all valid Ifile blocks are written. 868 */ 869 if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) { 870 lfs_gather(fs, sp, vp, lfs_match_data); 871 /* 872 * Don't call VOP_PUTPAGES: if we're flushing, 873 * we've already done it, and the Ifile doesn't 874 * use the page cache. 875 */ 876 } 877 } else { 878 lfs_gather(fs, sp, vp, lfs_match_data); 879 /* 880 * If we're flushing, we've already called VOP_PUTPAGES 881 * so don't do it again. Otherwise, we want to write 882 * everything we've got. 883 */ 884 if (!IS_FLUSHING(fs, vp)) { 885 simple_lock(&vp->v_interlock); 886 error = VOP_PUTPAGES(vp, 0, 0, 887 PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED); 888 } 889 } 890 891 /* 892 * It may not be necessary to write the meta-data blocks at this point, 893 * as the roll-forward recovery code should be able to reconstruct the 894 * list. 895 * 896 * We have to write them anyway, though, under two conditions: (1) the 897 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 898 * checkpointing. 899 * 900 * BUT if we are cleaning, we might have indirect blocks that refer to 901 * new blocks not being written yet, in addition to fragments being 902 * moved out of a cleaned segment. If that is the case, don't 903 * write the indirect blocks, or the finfo will have a small block 904 * in the middle of it! 905 * XXX in this case isn't the inode size wrong too? 906 */ 907 frag = 0; 908 if (sp->seg_flags & SEGM_CLEAN) { 909 for (i = 0; i < NDADDR; i++) 910 if (ip->i_lfs_fragsize[i] > 0 && 911 ip->i_lfs_fragsize[i] < fs->lfs_bsize) 912 ++frag; 913 } 914 #ifdef DIAGNOSTIC 915 if (frag > 1) 916 panic("lfs_writefile: more than one fragment!"); 917 #endif 918 if (IS_FLUSHING(fs, vp) || 919 (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) { 920 lfs_gather(fs, sp, vp, lfs_match_indir); 921 lfs_gather(fs, sp, vp, lfs_match_dindir); 922 lfs_gather(fs, sp, vp, lfs_match_tindir); 923 } 924 fip = sp->fip; 925 lfs_release_finfo(fs); 926 927 return error; 928 } 929 930 /* 931 * Update segment accounting to reflect this inode's change of address. 932 */ 933 static int 934 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr) 935 { 936 struct buf *bp; 937 daddr_t daddr; 938 IFILE *ifp; 939 SEGUSE *sup; 940 ino_t ino; 941 int redo_ifile, error; 942 u_int32_t sn; 943 944 redo_ifile = 0; 945 946 /* 947 * If updating the ifile, update the super-block. Update the disk 948 * address and access times for this inode in the ifile. 949 */ 950 ino = ip->i_number; 951 if (ino == LFS_IFILE_INUM) { 952 daddr = fs->lfs_idaddr; 953 fs->lfs_idaddr = dbtofsb(fs, ndaddr); 954 } else { 955 LFS_IENTRY(ifp, fs, ino, bp); 956 daddr = ifp->if_daddr; 957 ifp->if_daddr = dbtofsb(fs, ndaddr); 958 error = LFS_BWRITE_LOG(bp); /* Ifile */ 959 } 960 961 /* 962 * If this is the Ifile and lfs_offset is set to the first block 963 * in the segment, dirty the new segment's accounting block 964 * (XXX should already be dirty?) and tell the caller to do it again. 965 */ 966 if (ip->i_number == LFS_IFILE_INUM) { 967 sn = dtosn(fs, fs->lfs_offset); 968 if (sntod(fs, sn) + btofsb(fs, fs->lfs_sumsize) == 969 fs->lfs_offset) { 970 LFS_SEGENTRY(sup, fs, sn, bp); 971 KASSERT(bp->b_flags & B_DELWRI); 972 LFS_WRITESEGENTRY(sup, fs, sn, bp); 973 /* fs->lfs_flags |= LFS_IFDIRTY; */ 974 redo_ifile |= 1; 975 } 976 } 977 978 /* 979 * The inode's last address should not be in the current partial 980 * segment, except under exceptional circumstances (lfs_writevnodes 981 * had to start over, and in the meantime more blocks were written 982 * to a vnode). Both inodes will be accounted to this segment 983 * in lfs_writeseg so we need to subtract the earlier version 984 * here anyway. The segment count can temporarily dip below 985 * zero here; keep track of how many duplicates we have in 986 * "dupino" so we don't panic below. 987 */ 988 if (daddr >= fs->lfs_lastpseg && daddr <= fs->lfs_offset) { 989 ++sp->ndupino; 990 DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg " 991 "(ino %d daddr 0x%llx) ndupino=%d\n", ino, 992 (long long)daddr, sp->ndupino)); 993 } 994 /* 995 * Account the inode: it no longer belongs to its former segment, 996 * though it will not belong to the new segment until that segment 997 * is actually written. 998 */ 999 if (daddr != LFS_UNUSED_DADDR) { 1000 u_int32_t oldsn = dtosn(fs, daddr); 1001 #ifdef DIAGNOSTIC 1002 int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0; 1003 #endif 1004 LFS_SEGENTRY(sup, fs, oldsn, bp); 1005 #ifdef DIAGNOSTIC 1006 if (sup->su_nbytes + 1007 sizeof (struct ufs1_dinode) * ndupino 1008 < sizeof (struct ufs1_dinode)) { 1009 printf("lfs_writeinode: negative bytes " 1010 "(segment %" PRIu32 " short by %d, " 1011 "oldsn=%" PRIu32 ", cursn=%" PRIu32 1012 ", daddr=%" PRId64 ", su_nbytes=%u, " 1013 "ndupino=%d)\n", 1014 dtosn(fs, daddr), 1015 (int)sizeof (struct ufs1_dinode) * 1016 (1 - sp->ndupino) - sup->su_nbytes, 1017 oldsn, sp->seg_number, daddr, 1018 (unsigned int)sup->su_nbytes, 1019 sp->ndupino); 1020 panic("lfs_writeinode: negative bytes"); 1021 sup->su_nbytes = sizeof (struct ufs1_dinode); 1022 } 1023 #endif 1024 DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n", 1025 dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino)); 1026 sup->su_nbytes -= sizeof (struct ufs1_dinode); 1027 redo_ifile |= 1028 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 1029 if (redo_ifile) { 1030 simple_lock(&fs->lfs_interlock); 1031 fs->lfs_flags |= LFS_IFDIRTY; 1032 simple_unlock(&fs->lfs_interlock); 1033 /* Don't double-account */ 1034 fs->lfs_idaddr = 0x0; 1035 } 1036 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */ 1037 } 1038 1039 return redo_ifile; 1040 } 1041 1042 int 1043 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip) 1044 { 1045 struct buf *bp; 1046 struct ufs1_dinode *cdp; 1047 daddr_t daddr; 1048 int32_t *daddrp; /* XXX ondisk32 */ 1049 int i, ndx; 1050 int redo_ifile = 0; 1051 int gotblk = 0; 1052 int count; 1053 1054 ASSERT_SEGLOCK(fs); 1055 if (!(ip->i_flag & IN_ALLMOD)) 1056 return (0); 1057 1058 /* Can't write ifile when writer is not set */ 1059 KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 || 1060 (sp->seg_flags & SEGM_CLEAN)); 1061 1062 /* 1063 * If this is the Ifile, see if writing it here will generate a 1064 * temporary misaccounting. If it will, do the accounting and write 1065 * the blocks, postponing the inode write until the accounting is 1066 * solid. 1067 */ 1068 count = 0; 1069 while (ip->i_number == LFS_IFILE_INUM) { 1070 int redo = 0; 1071 1072 if (sp->idp == NULL && sp->ibp == NULL && 1073 (sp->seg_bytes_left < fs->lfs_ibsize || 1074 sp->sum_bytes_left < sizeof(int32_t))) { 1075 (void) lfs_writeseg(fs, sp); 1076 continue; 1077 } 1078 1079 /* Look for dirty Ifile blocks */ 1080 LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) { 1081 if (!(bp->b_flags & B_GATHERED)) { 1082 redo = 1; 1083 break; 1084 } 1085 } 1086 1087 if (redo == 0) 1088 redo = lfs_update_iaddr(fs, sp, ip, 0x0); 1089 if (redo == 0) 1090 break; 1091 1092 if (sp->idp) { 1093 sp->idp->di_inumber = 0; 1094 sp->idp = NULL; 1095 } 1096 ++count; 1097 if (count > 2) 1098 log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count); 1099 lfs_writefile(fs, sp, fs->lfs_ivnode); 1100 } 1101 1102 /* Allocate a new inode block if necessary. */ 1103 if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && 1104 sp->ibp == NULL) { 1105 /* Allocate a new segment if necessary. */ 1106 if (sp->seg_bytes_left < fs->lfs_ibsize || 1107 sp->sum_bytes_left < sizeof(int32_t)) 1108 (void) lfs_writeseg(fs, sp); 1109 1110 /* Get next inode block. */ 1111 daddr = fs->lfs_offset; 1112 fs->lfs_offset += btofsb(fs, fs->lfs_ibsize); 1113 sp->ibp = *sp->cbpp++ = 1114 getblk(VTOI(fs->lfs_ivnode)->i_devvp, 1115 fsbtodb(fs, daddr), fs->lfs_ibsize, 0, 0); 1116 gotblk++; 1117 1118 /* Zero out inode numbers */ 1119 for (i = 0; i < INOPB(fs); ++i) 1120 ((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber = 1121 0; 1122 1123 ++sp->start_bpp; 1124 fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize); 1125 /* Set remaining space counters. */ 1126 sp->seg_bytes_left -= fs->lfs_ibsize; 1127 sp->sum_bytes_left -= sizeof(int32_t); 1128 ndx = fs->lfs_sumsize / sizeof(int32_t) - 1129 sp->ninodes / INOPB(fs) - 1; 1130 ((int32_t *)(sp->segsum))[ndx] = daddr; 1131 } 1132 1133 /* Check VDIROP in case there is a new file with no data blocks */ 1134 if (ITOV(ip)->v_flag & VDIROP) 1135 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 1136 1137 /* Update the inode times and copy the inode onto the inode page. */ 1138 /* XXX kludge --- don't redirty the ifile just to put times on it */ 1139 if (ip->i_number != LFS_IFILE_INUM) 1140 LFS_ITIMES(ip, NULL, NULL, NULL); 1141 1142 /* 1143 * If this is the Ifile, and we've already written the Ifile in this 1144 * partial segment, just overwrite it (it's not on disk yet) and 1145 * continue. 1146 * 1147 * XXX we know that the bp that we get the second time around has 1148 * already been gathered. 1149 */ 1150 if (ip->i_number == LFS_IFILE_INUM && sp->idp) { 1151 *(sp->idp) = *ip->i_din.ffs1_din; 1152 ip->i_lfs_osize = ip->i_size; 1153 return 0; 1154 } 1155 1156 bp = sp->ibp; 1157 cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs)); 1158 *cdp = *ip->i_din.ffs1_din; 1159 1160 /* 1161 * If cleaning, link counts and directory file sizes cannot change, 1162 * since those would be directory operations---even if the file 1163 * we are writing is marked VDIROP we should write the old values. 1164 * If we're not cleaning, of course, update the values so we get 1165 * current values the next time we clean. 1166 */ 1167 if (sp->seg_flags & SEGM_CLEAN) { 1168 if (ITOV(ip)->v_flag & VDIROP) { 1169 cdp->di_nlink = ip->i_lfs_odnlink; 1170 /* if (ITOV(ip)->v_type == VDIR) */ 1171 cdp->di_size = ip->i_lfs_osize; 1172 } 1173 } else { 1174 ip->i_lfs_odnlink = cdp->di_nlink; 1175 ip->i_lfs_osize = ip->i_size; 1176 } 1177 1178 1179 /* We can finish the segment accounting for truncations now */ 1180 lfs_finalize_ino_seguse(fs, ip); 1181 1182 /* 1183 * If we are cleaning, ensure that we don't write UNWRITTEN disk 1184 * addresses to disk; possibly change the on-disk record of 1185 * the inode size, either by reverting to the previous size 1186 * (in the case of cleaning) or by verifying the inode's block 1187 * holdings (in the case of files being allocated as they are being 1188 * written). 1189 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail 1190 * XXX count on disk wrong by the same amount. We should be 1191 * XXX able to "borrow" from lfs_avail and return it after the 1192 * XXX Ifile is written. See also in lfs_writeseg. 1193 */ 1194 1195 /* Check file size based on highest allocated block */ 1196 if (((ip->i_ffs1_mode & IFMT) == IFREG || 1197 (ip->i_ffs1_mode & IFMT) == IFDIR) && 1198 ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) { 1199 cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift; 1200 DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %" 1201 PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size)); 1202 } 1203 if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) { 1204 DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)" 1205 " at %x\n", ip->i_number, ip->i_lfs_effnblks, 1206 ip->i_ffs1_blocks, fs->lfs_offset)); 1207 for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR; 1208 daddrp++) { 1209 if (*daddrp == UNWRITTEN) { 1210 DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n")); 1211 *daddrp = 0; 1212 } 1213 } 1214 } 1215 1216 #ifdef DIAGNOSTIC 1217 /* 1218 * Check dinode held blocks against dinode size. 1219 * This should be identical to the check in lfs_vget(). 1220 */ 1221 for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift; 1222 i < NDADDR; i++) { 1223 KASSERT(i >= 0); 1224 if ((cdp->di_mode & IFMT) == IFLNK) 1225 continue; 1226 if (((cdp->di_mode & IFMT) == IFBLK || 1227 (cdp->di_mode & IFMT) == IFCHR) && i == 0) 1228 continue; 1229 if (cdp->di_db[i] != 0) { 1230 # ifdef DEBUG 1231 lfs_dump_dinode(cdp); 1232 # endif 1233 panic("writing inconsistent inode"); 1234 } 1235 } 1236 #endif /* DIAGNOSTIC */ 1237 1238 if (ip->i_flag & IN_CLEANING) 1239 LFS_CLR_UINO(ip, IN_CLEANING); 1240 else { 1241 /* XXX IN_ALLMOD */ 1242 LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE | 1243 IN_UPDATE | IN_MODIFY); 1244 if (ip->i_lfs_effnblks == ip->i_ffs1_blocks) 1245 LFS_CLR_UINO(ip, IN_MODIFIED); 1246 else { 1247 DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real " 1248 "blks=%d, eff=%d\n", ip->i_number, 1249 ip->i_ffs1_blocks, ip->i_lfs_effnblks)); 1250 } 1251 } 1252 1253 if (ip->i_number == LFS_IFILE_INUM) { 1254 /* We know sp->idp == NULL */ 1255 sp->idp = ((struct ufs1_dinode *)bp->b_data) + 1256 (sp->ninodes % INOPB(fs)); 1257 1258 /* Not dirty any more */ 1259 simple_lock(&fs->lfs_interlock); 1260 fs->lfs_flags &= ~LFS_IFDIRTY; 1261 simple_unlock(&fs->lfs_interlock); 1262 } 1263 1264 if (gotblk) { 1265 LFS_LOCK_BUF(bp); 1266 brelse(bp); 1267 } 1268 1269 /* Increment inode count in segment summary block. */ 1270 ++((SEGSUM *)(sp->segsum))->ss_ninos; 1271 1272 /* If this page is full, set flag to allocate a new page. */ 1273 if (++sp->ninodes % INOPB(fs) == 0) 1274 sp->ibp = NULL; 1275 1276 redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno); 1277 1278 KASSERT(redo_ifile == 0); 1279 return (redo_ifile); 1280 } 1281 1282 int 1283 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr) 1284 { 1285 struct lfs *fs; 1286 int vers; 1287 int j, blksinblk; 1288 1289 ASSERT_SEGLOCK(sp->fs); 1290 /* 1291 * If full, finish this segment. We may be doing I/O, so 1292 * release and reacquire the splbio(). 1293 */ 1294 #ifdef DIAGNOSTIC 1295 if (sp->vp == NULL) 1296 panic ("lfs_gatherblock: Null vp in segment"); 1297 #endif 1298 fs = sp->fs; 1299 blksinblk = howmany(bp->b_bcount, fs->lfs_bsize); 1300 if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk || 1301 sp->seg_bytes_left < bp->b_bcount) { 1302 if (sptr) 1303 splx(*sptr); 1304 lfs_updatemeta(sp); 1305 1306 vers = sp->fip->fi_version; 1307 (void) lfs_writeseg(fs, sp); 1308 1309 /* Add the current file to the segment summary. */ 1310 lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers); 1311 1312 if (sptr) 1313 *sptr = splbio(); 1314 return (1); 1315 } 1316 1317 if (bp->b_flags & B_GATHERED) { 1318 DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d," 1319 " lbn %" PRId64 "\n", 1320 sp->fip->fi_ino, bp->b_lblkno)); 1321 return (0); 1322 } 1323 1324 /* Insert into the buffer list, update the FINFO block. */ 1325 bp->b_flags |= B_GATHERED; 1326 1327 *sp->cbpp++ = bp; 1328 for (j = 0; j < blksinblk; j++) { 1329 sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j; 1330 /* This block's accounting moves from lfs_favail to lfs_avail */ 1331 lfs_deregister_block(sp->vp, bp->b_lblkno + j); 1332 } 1333 1334 sp->sum_bytes_left -= sizeof(int32_t) * blksinblk; 1335 sp->seg_bytes_left -= bp->b_bcount; 1336 return (0); 1337 } 1338 1339 int 1340 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, 1341 int (*match)(struct lfs *, struct buf *)) 1342 { 1343 struct buf *bp, *nbp; 1344 int s, count = 0; 1345 1346 ASSERT_SEGLOCK(fs); 1347 if (vp->v_type == VBLK) 1348 return 0; 1349 KASSERT(sp->vp == NULL); 1350 sp->vp = vp; 1351 s = splbio(); 1352 1353 #ifndef LFS_NO_BACKBUF_HACK 1354 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 1355 # define BUF_OFFSET \ 1356 (((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp) 1357 # define BACK_BUF(BP) \ 1358 ((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET)) 1359 # define BEG_OF_LIST \ 1360 ((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET)) 1361 1362 loop: 1363 /* Find last buffer. */ 1364 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); 1365 bp && LIST_NEXT(bp, b_vnbufs) != NULL; 1366 bp = LIST_NEXT(bp, b_vnbufs)) 1367 /* nothing */; 1368 for (; bp && bp != BEG_OF_LIST; bp = nbp) { 1369 nbp = BACK_BUF(bp); 1370 #else /* LFS_NO_BACKBUF_HACK */ 1371 loop: 1372 for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 1373 nbp = LIST_NEXT(bp, b_vnbufs); 1374 #endif /* LFS_NO_BACKBUF_HACK */ 1375 if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) { 1376 #ifdef DEBUG 1377 if (vp == fs->lfs_ivnode && 1378 (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY) 1379 log(LOG_NOTICE, "lfs_gather: ifile lbn %" 1380 PRId64 " busy (%x) at 0x%x", 1381 bp->b_lblkno, bp->b_flags, 1382 (unsigned)fs->lfs_offset); 1383 #endif 1384 continue; 1385 } 1386 #ifdef DIAGNOSTIC 1387 # ifdef LFS_USE_B_INVAL 1388 if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) { 1389 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1390 " is B_INVAL\n", bp->b_lblkno)); 1391 VOP_PRINT(bp->b_vp); 1392 } 1393 # endif /* LFS_USE_B_INVAL */ 1394 if (!(bp->b_flags & B_DELWRI)) 1395 panic("lfs_gather: bp not B_DELWRI"); 1396 if (!(bp->b_flags & B_LOCKED)) { 1397 DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64 1398 " blk %" PRId64 " not B_LOCKED\n", 1399 bp->b_lblkno, 1400 dbtofsb(fs, bp->b_blkno))); 1401 VOP_PRINT(bp->b_vp); 1402 panic("lfs_gather: bp not B_LOCKED"); 1403 } 1404 #endif 1405 if (lfs_gatherblock(sp, bp, &s)) { 1406 goto loop; 1407 } 1408 count++; 1409 } 1410 splx(s); 1411 lfs_updatemeta(sp); 1412 KASSERT(sp->vp == vp); 1413 sp->vp = NULL; 1414 return count; 1415 } 1416 1417 #if DEBUG 1418 # define DEBUG_OOFF(n) do { \ 1419 if (ooff == 0) { \ 1420 DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \ 1421 "ino %d lbn %" PRId64 " at 0x%" PRIx32 \ 1422 ", was 0x0 (or %" PRId64 ")\n", \ 1423 (n), ip->i_number, lbn, ndaddr, daddr)); \ 1424 } \ 1425 } while (0) 1426 #else 1427 # define DEBUG_OOFF(n) 1428 #endif 1429 1430 /* 1431 * Change the given block's address to ndaddr, finding its previous 1432 * location using ufs_bmaparray(). 1433 * 1434 * Account for this change in the segment table. 1435 * 1436 * called with sp == NULL by roll-forwarding code. 1437 */ 1438 void 1439 lfs_update_single(struct lfs *fs, struct segment *sp __unused, 1440 struct vnode *vp, daddr_t lbn, int32_t ndaddr, int size) 1441 { 1442 SEGUSE *sup; 1443 struct buf *bp; 1444 struct indir a[NIADDR + 2], *ap; 1445 struct inode *ip; 1446 daddr_t daddr, ooff; 1447 int num, error; 1448 int bb, osize, obb; 1449 1450 ASSERT_SEGLOCK(fs); 1451 KASSERT(sp == NULL || sp->vp == vp); 1452 ip = VTOI(vp); 1453 1454 error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL); 1455 if (error) 1456 panic("lfs_updatemeta: ufs_bmaparray returned %d", error); 1457 1458 daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */ 1459 KASSERT(daddr <= LFS_MAX_DADDR); 1460 if (daddr > 0) 1461 daddr = dbtofsb(fs, daddr); 1462 1463 bb = fragstofsb(fs, numfrags(fs, size)); 1464 switch (num) { 1465 case 0: 1466 ooff = ip->i_ffs1_db[lbn]; 1467 DEBUG_OOFF(0); 1468 if (ooff == UNWRITTEN) 1469 ip->i_ffs1_blocks += bb; 1470 else { 1471 /* possible fragment truncation or extension */ 1472 obb = btofsb(fs, ip->i_lfs_fragsize[lbn]); 1473 ip->i_ffs1_blocks += (bb - obb); 1474 } 1475 ip->i_ffs1_db[lbn] = ndaddr; 1476 break; 1477 case 1: 1478 ooff = ip->i_ffs1_ib[a[0].in_off]; 1479 DEBUG_OOFF(1); 1480 if (ooff == UNWRITTEN) 1481 ip->i_ffs1_blocks += bb; 1482 ip->i_ffs1_ib[a[0].in_off] = ndaddr; 1483 break; 1484 default: 1485 ap = &a[num - 1]; 1486 if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp)) 1487 panic("lfs_updatemeta: bread bno %" PRId64, 1488 ap->in_lbn); 1489 1490 /* XXX ondisk32 */ 1491 ooff = ((int32_t *)bp->b_data)[ap->in_off]; 1492 DEBUG_OOFF(num); 1493 if (ooff == UNWRITTEN) 1494 ip->i_ffs1_blocks += bb; 1495 /* XXX ondisk32 */ 1496 ((int32_t *)bp->b_data)[ap->in_off] = ndaddr; 1497 (void) VOP_BWRITE(bp); 1498 } 1499 1500 KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr); 1501 1502 /* Update hiblk when extending the file */ 1503 if (lbn > ip->i_lfs_hiblk) 1504 ip->i_lfs_hiblk = lbn; 1505 1506 /* 1507 * Though we'd rather it couldn't, this *can* happen right now 1508 * if cleaning blocks and regular blocks coexist. 1509 */ 1510 /* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */ 1511 1512 /* 1513 * Update segment usage information, based on old size 1514 * and location. 1515 */ 1516 if (daddr > 0) { 1517 u_int32_t oldsn = dtosn(fs, daddr); 1518 #ifdef DIAGNOSTIC 1519 int ndupino; 1520 1521 if (sp && sp->seg_number == oldsn) { 1522 ndupino = sp->ndupino; 1523 } else { 1524 ndupino = 0; 1525 } 1526 #endif 1527 KASSERT(oldsn < fs->lfs_nseg); 1528 if (lbn >= 0 && lbn < NDADDR) 1529 osize = ip->i_lfs_fragsize[lbn]; 1530 else 1531 osize = fs->lfs_bsize; 1532 LFS_SEGENTRY(sup, fs, oldsn, bp); 1533 #ifdef DIAGNOSTIC 1534 if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino 1535 < osize) { 1536 printf("lfs_updatemeta: negative bytes " 1537 "(segment %" PRIu32 " short by %" PRId64 1538 ")\n", dtosn(fs, daddr), 1539 (int64_t)osize - 1540 (sizeof (struct ufs1_dinode) * ndupino + 1541 sup->su_nbytes)); 1542 printf("lfs_updatemeta: ino %llu, lbn %" PRId64 1543 ", addr = 0x%" PRIx64 "\n", 1544 (unsigned long long)ip->i_number, lbn, daddr); 1545 printf("lfs_updatemeta: ndupino=%d\n", ndupino); 1546 panic("lfs_updatemeta: negative bytes"); 1547 sup->su_nbytes = osize - 1548 sizeof (struct ufs1_dinode) * ndupino; 1549 } 1550 #endif 1551 DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64 1552 " db 0x%" PRIx64 "\n", 1553 dtosn(fs, daddr), osize, 1554 ip->i_number, lbn, daddr)); 1555 sup->su_nbytes -= osize; 1556 if (!(bp->b_flags & B_GATHERED)) { 1557 simple_lock(&fs->lfs_interlock); 1558 fs->lfs_flags |= LFS_IFDIRTY; 1559 simple_unlock(&fs->lfs_interlock); 1560 } 1561 LFS_WRITESEGENTRY(sup, fs, oldsn, bp); 1562 } 1563 /* 1564 * Now that this block has a new address, and its old 1565 * segment no longer owns it, we can forget about its 1566 * old size. 1567 */ 1568 if (lbn >= 0 && lbn < NDADDR) 1569 ip->i_lfs_fragsize[lbn] = size; 1570 } 1571 1572 /* 1573 * Update the metadata that points to the blocks listed in the FINFO 1574 * array. 1575 */ 1576 void 1577 lfs_updatemeta(struct segment *sp) 1578 { 1579 struct buf *sbp; 1580 struct lfs *fs; 1581 struct vnode *vp; 1582 daddr_t lbn; 1583 int i, nblocks, num; 1584 int bb; 1585 int bytesleft, size; 1586 1587 ASSERT_SEGLOCK(sp->fs); 1588 vp = sp->vp; 1589 nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 1590 KASSERT(nblocks >= 0); 1591 KASSERT(vp != NULL); 1592 if (nblocks == 0) 1593 return; 1594 1595 /* 1596 * This count may be high due to oversize blocks from lfs_gop_write. 1597 * Correct for this. (XXX we should be able to keep track of these.) 1598 */ 1599 fs = sp->fs; 1600 for (i = 0; i < nblocks; i++) { 1601 if (sp->start_bpp[i] == NULL) { 1602 DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks)); 1603 nblocks = i; 1604 break; 1605 } 1606 num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize); 1607 KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1); 1608 nblocks -= num - 1; 1609 } 1610 1611 KASSERT(vp->v_type == VREG || 1612 nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp); 1613 KASSERT(nblocks == sp->cbpp - sp->start_bpp); 1614 1615 /* 1616 * Sort the blocks. 1617 * 1618 * We have to sort even if the blocks come from the 1619 * cleaner, because there might be other pending blocks on the 1620 * same inode...and if we don't sort, and there are fragments 1621 * present, blocks may be written in the wrong place. 1622 */ 1623 lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize); 1624 1625 /* 1626 * Record the length of the last block in case it's a fragment. 1627 * If there are indirect blocks present, they sort last. An 1628 * indirect block will be lfs_bsize and its presence indicates 1629 * that you cannot have fragments. 1630 * 1631 * XXX This last is a lie. A cleaned fragment can coexist with 1632 * XXX a later indirect block. This will continue to be 1633 * XXX true until lfs_markv is fixed to do everything with 1634 * XXX fake blocks (including fake inodes and fake indirect blocks). 1635 */ 1636 sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) & 1637 fs->lfs_bmask) + 1; 1638 1639 /* 1640 * Assign disk addresses, and update references to the logical 1641 * block and the segment usage information. 1642 */ 1643 for (i = nblocks; i--; ++sp->start_bpp) { 1644 sbp = *sp->start_bpp; 1645 lbn = *sp->start_lbp; 1646 KASSERT(sbp->b_lblkno == lbn); 1647 1648 sbp->b_blkno = fsbtodb(fs, fs->lfs_offset); 1649 1650 /* 1651 * If we write a frag in the wrong place, the cleaner won't 1652 * be able to correctly identify its size later, and the 1653 * segment will be uncleanable. (Even worse, it will assume 1654 * that the indirect block that actually ends the list 1655 * is of a smaller size!) 1656 */ 1657 if ((sbp->b_bcount & fs->lfs_bmask) && i != 0) 1658 panic("lfs_updatemeta: fragment is not last block"); 1659 1660 /* 1661 * For each subblock in this possibly oversized block, 1662 * update its address on disk. 1663 */ 1664 KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize); 1665 KASSERT(vp == sbp->b_vp); 1666 for (bytesleft = sbp->b_bcount; bytesleft > 0; 1667 bytesleft -= fs->lfs_bsize) { 1668 size = MIN(bytesleft, fs->lfs_bsize); 1669 bb = fragstofsb(fs, numfrags(fs, size)); 1670 lbn = *sp->start_lbp++; 1671 lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset, 1672 size); 1673 fs->lfs_offset += bb; 1674 } 1675 1676 } 1677 } 1678 1679 /* 1680 * Move lfs_offset to a segment earlier than sn. 1681 */ 1682 int 1683 lfs_rewind(struct lfs *fs, int newsn) 1684 { 1685 int sn, osn, isdirty; 1686 struct buf *bp; 1687 SEGUSE *sup; 1688 1689 ASSERT_SEGLOCK(fs); 1690 1691 osn = dtosn(fs, fs->lfs_offset); 1692 if (osn < newsn) 1693 return 0; 1694 1695 /* lfs_avail eats the remaining space in this segment */ 1696 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg); 1697 1698 /* Find a low-numbered segment */ 1699 for (sn = 0; sn < fs->lfs_nseg; ++sn) { 1700 LFS_SEGENTRY(sup, fs, sn, bp); 1701 isdirty = sup->su_flags & SEGUSE_DIRTY; 1702 brelse(bp); 1703 1704 if (!isdirty) 1705 break; 1706 } 1707 if (sn == fs->lfs_nseg) 1708 panic("lfs_rewind: no clean segments"); 1709 if (newsn >= 0 && sn >= newsn) 1710 return ENOENT; 1711 fs->lfs_nextseg = sn; 1712 lfs_newseg(fs); 1713 fs->lfs_offset = fs->lfs_curseg; 1714 1715 return 0; 1716 } 1717 1718 /* 1719 * Start a new partial segment. 1720 * 1721 * Return 1 when we entered to a new segment. 1722 * Otherwise, return 0. 1723 */ 1724 int 1725 lfs_initseg(struct lfs *fs) 1726 { 1727 struct segment *sp = fs->lfs_sp; 1728 SEGSUM *ssp; 1729 struct buf *sbp; /* buffer for SEGSUM */ 1730 int repeat = 0; /* return value */ 1731 1732 ASSERT_SEGLOCK(fs); 1733 /* Advance to the next segment. */ 1734 if (!LFS_PARTIAL_FITS(fs)) { 1735 SEGUSE *sup; 1736 struct buf *bp; 1737 1738 /* lfs_avail eats the remaining space */ 1739 fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - 1740 fs->lfs_curseg); 1741 /* Wake up any cleaning procs waiting on this file system. */ 1742 lfs_wakeup_cleaner(fs); 1743 lfs_newseg(fs); 1744 repeat = 1; 1745 fs->lfs_offset = fs->lfs_curseg; 1746 1747 sp->seg_number = dtosn(fs, fs->lfs_curseg); 1748 sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg); 1749 1750 /* 1751 * If the segment contains a superblock, update the offset 1752 * and summary address to skip over it. 1753 */ 1754 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1755 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1756 fs->lfs_offset += btofsb(fs, LFS_SBPAD); 1757 sp->seg_bytes_left -= LFS_SBPAD; 1758 } 1759 brelse(bp); 1760 /* Segment zero could also contain the labelpad */ 1761 if (fs->lfs_version > 1 && sp->seg_number == 0 && 1762 fs->lfs_start < btofsb(fs, LFS_LABELPAD)) { 1763 fs->lfs_offset += 1764 btofsb(fs, LFS_LABELPAD) - fs->lfs_start; 1765 sp->seg_bytes_left -= 1766 LFS_LABELPAD - fsbtob(fs, fs->lfs_start); 1767 } 1768 } else { 1769 sp->seg_number = dtosn(fs, fs->lfs_curseg); 1770 sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg - 1771 (fs->lfs_offset - fs->lfs_curseg)); 1772 } 1773 fs->lfs_lastpseg = fs->lfs_offset; 1774 1775 /* Record first address of this partial segment */ 1776 if (sp->seg_flags & SEGM_CLEAN) { 1777 fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset; 1778 if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) { 1779 /* "1" is the artificial inc in lfs_seglock */ 1780 simple_lock(&fs->lfs_interlock); 1781 while (fs->lfs_iocount > 1) { 1782 ltsleep(&fs->lfs_iocount, PRIBIO + 1, 1783 "lfs_initseg", 0, &fs->lfs_interlock); 1784 } 1785 simple_unlock(&fs->lfs_interlock); 1786 fs->lfs_cleanind = 0; 1787 } 1788 } 1789 1790 sp->fs = fs; 1791 sp->ibp = NULL; 1792 sp->idp = NULL; 1793 sp->ninodes = 0; 1794 sp->ndupino = 0; 1795 1796 sp->cbpp = sp->bpp; 1797 1798 /* Get a new buffer for SEGSUM */ 1799 sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, 1800 fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY); 1801 1802 /* ... and enter it into the buffer list. */ 1803 *sp->cbpp = sbp; 1804 sp->cbpp++; 1805 fs->lfs_offset += btofsb(fs, fs->lfs_sumsize); 1806 1807 sp->start_bpp = sp->cbpp; 1808 1809 /* Set point to SEGSUM, initialize it. */ 1810 ssp = sp->segsum = sbp->b_data; 1811 memset(ssp, 0, fs->lfs_sumsize); 1812 ssp->ss_next = fs->lfs_nextseg; 1813 ssp->ss_nfinfo = ssp->ss_ninos = 0; 1814 ssp->ss_magic = SS_MAGIC; 1815 1816 /* Set pointer to first FINFO, initialize it. */ 1817 sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs)); 1818 sp->fip->fi_nblocks = 0; 1819 sp->start_lbp = &sp->fip->fi_blocks[0]; 1820 sp->fip->fi_lastlength = 0; 1821 1822 sp->seg_bytes_left -= fs->lfs_sumsize; 1823 sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs); 1824 1825 return (repeat); 1826 } 1827 1828 /* 1829 * Remove SEGUSE_INVAL from all segments. 1830 */ 1831 void 1832 lfs_unset_inval_all(struct lfs *fs) 1833 { 1834 SEGUSE *sup; 1835 struct buf *bp; 1836 int i; 1837 1838 for (i = 0; i < fs->lfs_nseg; i++) { 1839 LFS_SEGENTRY(sup, fs, i, bp); 1840 if (sup->su_flags & SEGUSE_INVAL) { 1841 sup->su_flags &= ~SEGUSE_INVAL; 1842 LFS_WRITESEGENTRY(sup, fs, i, bp); 1843 } else 1844 brelse(bp); 1845 } 1846 } 1847 1848 /* 1849 * Return the next segment to write. 1850 */ 1851 void 1852 lfs_newseg(struct lfs *fs) 1853 { 1854 CLEANERINFO *cip; 1855 SEGUSE *sup; 1856 struct buf *bp; 1857 int curseg, isdirty, sn, skip_inval; 1858 1859 ASSERT_SEGLOCK(fs); 1860 1861 /* Honor LFCNWRAPSTOP */ 1862 simple_lock(&fs->lfs_interlock); 1863 while (fs->lfs_nextseg < fs->lfs_curseg && fs->lfs_nowrap) { 1864 if (fs->lfs_wrappass) { 1865 log(LOG_NOTICE, "%s: wrappass=%d\n", 1866 fs->lfs_fsmnt, fs->lfs_wrappass); 1867 fs->lfs_wrappass = 0; 1868 break; 1869 } 1870 fs->lfs_wrapstatus = LFS_WRAP_WAITING; 1871 wakeup(&fs->lfs_nowrap); 1872 log(LOG_NOTICE, "%s: waiting at log wrap\n", fs->lfs_fsmnt); 1873 ltsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz, 1874 &fs->lfs_interlock); 1875 } 1876 fs->lfs_wrapstatus = LFS_WRAP_GOING; 1877 simple_unlock(&fs->lfs_interlock); 1878 1879 LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp); 1880 DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n", 1881 dtosn(fs, fs->lfs_nextseg))); 1882 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1883 sup->su_nbytes = 0; 1884 sup->su_nsums = 0; 1885 sup->su_ninos = 0; 1886 LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp); 1887 1888 LFS_CLEANERINFO(cip, fs, bp); 1889 --cip->clean; 1890 ++cip->dirty; 1891 fs->lfs_nclean = cip->clean; 1892 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1893 1894 fs->lfs_lastseg = fs->lfs_curseg; 1895 fs->lfs_curseg = fs->lfs_nextseg; 1896 skip_inval = 1; 1897 for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) { 1898 sn = (sn + 1) % fs->lfs_nseg; 1899 1900 if (sn == curseg) { 1901 if (skip_inval) 1902 skip_inval = 0; 1903 else 1904 panic("lfs_nextseg: no clean segments"); 1905 } 1906 LFS_SEGENTRY(sup, fs, sn, bp); 1907 isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0)); 1908 /* Check SEGUSE_EMPTY as we go along */ 1909 if (isdirty && sup->su_nbytes == 0 && 1910 !(sup->su_flags & SEGUSE_EMPTY)) 1911 LFS_WRITESEGENTRY(sup, fs, sn, bp); 1912 else 1913 brelse(bp); 1914 1915 if (!isdirty) 1916 break; 1917 } 1918 if (skip_inval == 0) 1919 lfs_unset_inval_all(fs); 1920 1921 ++fs->lfs_nactive; 1922 fs->lfs_nextseg = sntod(fs, sn); 1923 if (lfs_dostats) { 1924 ++lfs_stats.segsused; 1925 } 1926 } 1927 1928 static struct buf * 1929 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, 1930 int n __unused) 1931 { 1932 struct lfs_cluster *cl; 1933 struct buf **bpp, *bp; 1934 1935 ASSERT_SEGLOCK(fs); 1936 cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK); 1937 bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK); 1938 memset(cl, 0, sizeof(*cl)); 1939 cl->fs = fs; 1940 cl->bpp = bpp; 1941 cl->bufcount = 0; 1942 cl->bufsize = 0; 1943 1944 /* If this segment is being written synchronously, note that */ 1945 if (fs->lfs_sp->seg_flags & SEGM_SYNC) { 1946 cl->flags |= LFS_CL_SYNC; 1947 cl->seg = fs->lfs_sp; 1948 ++cl->seg->seg_iocount; 1949 } 1950 1951 /* Get an empty buffer header, or maybe one with something on it */ 1952 bp = getiobuf(); 1953 bp->b_flags = B_BUSY | B_CALL; 1954 bp->b_dev = NODEV; 1955 bp->b_blkno = bp->b_lblkno = addr; 1956 bp->b_iodone = lfs_cluster_callback; 1957 bp->b_private = cl; 1958 bp->b_vp = vp; 1959 1960 return bp; 1961 } 1962 1963 int 1964 lfs_writeseg(struct lfs *fs, struct segment *sp) 1965 { 1966 struct buf **bpp, *bp, *cbp, *newbp; 1967 SEGUSE *sup; 1968 SEGSUM *ssp; 1969 int i, s; 1970 int do_again, nblocks, byteoffset; 1971 size_t el_size; 1972 struct lfs_cluster *cl; 1973 u_short ninos; 1974 struct vnode *devvp; 1975 char *p = NULL; 1976 struct vnode *vp; 1977 int32_t *daddrp; /* XXX ondisk32 */ 1978 int changed; 1979 u_int32_t sum; 1980 #ifdef DEBUG 1981 FINFO *fip; 1982 int findex; 1983 #endif 1984 1985 ASSERT_SEGLOCK(fs); 1986 1987 ssp = (SEGSUM *)sp->segsum; 1988 1989 /* 1990 * If there are no buffers other than the segment summary to write, 1991 * don't do anything. If we are the end of a dirop sequence, however, 1992 * write the empty segment summary anyway, to help out the 1993 * roll-forward agent. 1994 */ 1995 if ((nblocks = sp->cbpp - sp->bpp) == 1) { 1996 if ((ssp->ss_flags & (SS_DIROP | SS_CONT)) != SS_DIROP) 1997 return 0; 1998 } 1999 2000 /* Note if partial segment is being written by the cleaner */ 2001 if (sp->seg_flags & SEGM_CLEAN) 2002 ssp->ss_flags |= SS_CLEAN; 2003 2004 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2005 2006 /* Update the segment usage information. */ 2007 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 2008 2009 /* Loop through all blocks, except the segment summary. */ 2010 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 2011 if ((*bpp)->b_vp != devvp) { 2012 sup->su_nbytes += (*bpp)->b_bcount; 2013 DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d" 2014 " lbn %" PRId64 " db 0x%" PRIx64 "\n", 2015 sp->seg_number, (*bpp)->b_bcount, 2016 VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno, 2017 (*bpp)->b_blkno)); 2018 } 2019 } 2020 2021 #ifdef DEBUG 2022 /* Check for zero-length and zero-version FINFO entries. */ 2023 fip = (struct finfo *)((caddr_t)ssp + SEGSUM_SIZE(fs)); 2024 for (findex = 0; findex < ssp->ss_nfinfo; findex++) { 2025 KDASSERT(fip->fi_nblocks > 0); 2026 KDASSERT(fip->fi_version > 0); 2027 fip = (FINFO *)((caddr_t)fip + FINFOSIZE + 2028 sizeof(int32_t) * fip->fi_nblocks); 2029 } 2030 #endif /* DEBUG */ 2031 2032 ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs); 2033 DLOG((DLOG_SU, "seg %d += %d for %d inodes\n", 2034 sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode), 2035 ssp->ss_ninos)); 2036 sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode); 2037 /* sup->su_nbytes += fs->lfs_sumsize; */ 2038 if (fs->lfs_version == 1) 2039 sup->su_olastmod = time_second; 2040 else 2041 sup->su_lastmod = time_second; 2042 sup->su_ninos += ninos; 2043 ++sup->su_nsums; 2044 fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize); 2045 2046 do_again = !(bp->b_flags & B_GATHERED); 2047 LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */ 2048 2049 /* 2050 * Mark blocks B_BUSY, to prevent then from being changed between 2051 * the checksum computation and the actual write. 2052 * 2053 * If we are cleaning, check indirect blocks for UNWRITTEN, and if 2054 * there are any, replace them with copies that have UNASSIGNED 2055 * instead. 2056 */ 2057 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 2058 ++bpp; 2059 bp = *bpp; 2060 if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */ 2061 bp->b_flags |= B_BUSY; 2062 continue; 2063 } 2064 2065 simple_lock(&bp->b_interlock); 2066 s = splbio(); 2067 while (bp->b_flags & B_BUSY) { 2068 DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential" 2069 " data summary corruption for ino %d, lbn %" 2070 PRId64 "\n", 2071 VTOI(bp->b_vp)->i_number, bp->b_lblkno)); 2072 bp->b_flags |= B_WANTED; 2073 ltsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0, 2074 &bp->b_interlock); 2075 splx(s); 2076 s = splbio(); 2077 } 2078 bp->b_flags |= B_BUSY; 2079 splx(s); 2080 simple_unlock(&bp->b_interlock); 2081 2082 /* 2083 * Check and replace indirect block UNWRITTEN bogosity. 2084 * XXX See comment in lfs_writefile. 2085 */ 2086 if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp && 2087 VTOI(bp->b_vp)->i_ffs1_blocks != 2088 VTOI(bp->b_vp)->i_lfs_effnblks) { 2089 DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n", 2090 VTOI(bp->b_vp)->i_number, 2091 VTOI(bp->b_vp)->i_lfs_effnblks, 2092 VTOI(bp->b_vp)->i_ffs1_blocks)); 2093 /* Make a copy we'll make changes to */ 2094 newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno, 2095 bp->b_bcount, LFS_NB_IBLOCK); 2096 newbp->b_blkno = bp->b_blkno; 2097 memcpy(newbp->b_data, bp->b_data, 2098 newbp->b_bcount); 2099 2100 changed = 0; 2101 /* XXX ondisk32 */ 2102 for (daddrp = (int32_t *)(newbp->b_data); 2103 daddrp < (int32_t *)(newbp->b_data + 2104 newbp->b_bcount); daddrp++) { 2105 if (*daddrp == UNWRITTEN) { 2106 ++changed; 2107 *daddrp = 0; 2108 } 2109 } 2110 /* 2111 * Get rid of the old buffer. Don't mark it clean, 2112 * though, if it still has dirty data on it. 2113 */ 2114 if (changed) { 2115 DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):" 2116 " bp = %p newbp = %p\n", changed, bp, 2117 newbp)); 2118 *bpp = newbp; 2119 bp->b_flags &= ~(B_ERROR | B_GATHERED); 2120 if (bp->b_flags & B_CALL) { 2121 DLOG((DLOG_SEG, "lfs_writeseg: " 2122 "indir bp should not be B_CALL\n")); 2123 s = splbio(); 2124 biodone(bp); 2125 splx(s); 2126 bp = NULL; 2127 } else { 2128 /* Still on free list, leave it there */ 2129 s = splbio(); 2130 bp->b_flags &= ~B_BUSY; 2131 if (bp->b_flags & B_WANTED) 2132 wakeup(bp); 2133 splx(s); 2134 /* 2135 * We have to re-decrement lfs_avail 2136 * since this block is going to come 2137 * back around to us in the next 2138 * segment. 2139 */ 2140 fs->lfs_avail -= 2141 btofsb(fs, bp->b_bcount); 2142 } 2143 } else { 2144 lfs_freebuf(fs, newbp); 2145 } 2146 } 2147 } 2148 /* 2149 * Compute checksum across data and then across summary; the first 2150 * block (the summary block) is skipped. Set the create time here 2151 * so that it's guaranteed to be later than the inode mod times. 2152 */ 2153 sum = 0; 2154 if (fs->lfs_version == 1) 2155 el_size = sizeof(u_long); 2156 else 2157 el_size = sizeof(u_int32_t); 2158 for (bpp = sp->bpp, i = nblocks - 1; i--; ) { 2159 ++bpp; 2160 /* Loop through gop_write cluster blocks */ 2161 for (byteoffset = 0; byteoffset < (*bpp)->b_bcount; 2162 byteoffset += fs->lfs_bsize) { 2163 #ifdef LFS_USE_B_INVAL 2164 if (((*bpp)->b_flags & (B_CALL | B_INVAL)) == 2165 (B_CALL | B_INVAL)) { 2166 if (copyin((caddr_t)(*bpp)->b_saveaddr + 2167 byteoffset, dp, el_size)) { 2168 panic("lfs_writeseg: copyin failed [1]:" 2169 " ino %d blk %" PRId64, 2170 VTOI((*bpp)->b_vp)->i_number, 2171 (*bpp)->b_lblkno); 2172 } 2173 } else 2174 #endif /* LFS_USE_B_INVAL */ 2175 { 2176 sum = lfs_cksum_part( 2177 (*bpp)->b_data + byteoffset, el_size, sum); 2178 } 2179 } 2180 } 2181 if (fs->lfs_version == 1) 2182 ssp->ss_ocreate = time_second; 2183 else { 2184 ssp->ss_create = time_second; 2185 ssp->ss_serial = ++fs->lfs_serial; 2186 ssp->ss_ident = fs->lfs_ident; 2187 } 2188 ssp->ss_datasum = lfs_cksum_fold(sum); 2189 ssp->ss_sumsum = cksum(&ssp->ss_datasum, 2190 fs->lfs_sumsize - sizeof(ssp->ss_sumsum)); 2191 2192 simple_lock(&fs->lfs_interlock); 2193 fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) + 2194 btofsb(fs, fs->lfs_sumsize)); 2195 fs->lfs_dmeta += (btofsb(fs, ninos * fs->lfs_ibsize) + 2196 btofsb(fs, fs->lfs_sumsize)); 2197 simple_unlock(&fs->lfs_interlock); 2198 2199 /* 2200 * When we simply write the blocks we lose a rotation for every block 2201 * written. To avoid this problem, we cluster the buffers into a 2202 * chunk and write the chunk. MAXPHYS is the largest size I/O 2203 * devices can handle, use that for the size of the chunks. 2204 * 2205 * Blocks that are already clusters (from GOP_WRITE), however, we 2206 * don't bother to copy into other clusters. 2207 */ 2208 2209 #define CHUNKSIZE MAXPHYS 2210 2211 if (devvp == NULL) 2212 panic("devvp is NULL"); 2213 for (bpp = sp->bpp, i = nblocks; i;) { 2214 cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i); 2215 cl = cbp->b_private; 2216 2217 cbp->b_flags |= B_ASYNC | B_BUSY; 2218 cbp->b_bcount = 0; 2219 2220 #if defined(DEBUG) && defined(DIAGNOSTIC) 2221 if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs)) 2222 / sizeof(int32_t)) { 2223 panic("lfs_writeseg: real bpp overwrite"); 2224 } 2225 if (bpp - sp->bpp > segsize(fs) / fs->lfs_fsize) { 2226 panic("lfs_writeseg: theoretical bpp overwrite"); 2227 } 2228 #endif 2229 2230 /* 2231 * Construct the cluster. 2232 */ 2233 simple_lock(&fs->lfs_interlock); 2234 ++fs->lfs_iocount; 2235 simple_unlock(&fs->lfs_interlock); 2236 while (i && cbp->b_bcount < CHUNKSIZE) { 2237 bp = *bpp; 2238 2239 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 2240 break; 2241 if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC)) 2242 break; 2243 2244 /* Clusters from GOP_WRITE are expedited */ 2245 if (bp->b_bcount > fs->lfs_bsize) { 2246 if (cbp->b_bcount > 0) 2247 /* Put in its own buffer */ 2248 break; 2249 else { 2250 cbp->b_data = bp->b_data; 2251 } 2252 } else if (cbp->b_bcount == 0) { 2253 p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE, 2254 LFS_NB_CLUSTER); 2255 cl->flags |= LFS_CL_MALLOC; 2256 } 2257 #ifdef DIAGNOSTIC 2258 if (dtosn(fs, dbtofsb(fs, bp->b_blkno + 2259 btodb(bp->b_bcount - 1))) != 2260 sp->seg_number) { 2261 printf("blk size %d daddr %" PRIx64 2262 " not in seg %d\n", 2263 bp->b_bcount, bp->b_blkno, 2264 sp->seg_number); 2265 panic("segment overwrite"); 2266 } 2267 #endif 2268 2269 #ifdef LFS_USE_B_INVAL 2270 /* 2271 * Fake buffers from the cleaner are marked as B_INVAL. 2272 * We need to copy the data from user space rather than 2273 * from the buffer indicated. 2274 * XXX == what do I do on an error? 2275 */ 2276 if ((bp->b_flags & (B_CALL|B_INVAL)) == 2277 (B_CALL|B_INVAL)) { 2278 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 2279 panic("lfs_writeseg: " 2280 "copyin failed [2]"); 2281 } else 2282 #endif /* LFS_USE_B_INVAL */ 2283 if (cl->flags & LFS_CL_MALLOC) { 2284 /* copy data into our cluster. */ 2285 memcpy(p, bp->b_data, bp->b_bcount); 2286 p += bp->b_bcount; 2287 } 2288 2289 cbp->b_bcount += bp->b_bcount; 2290 cl->bufsize += bp->b_bcount; 2291 2292 bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE); 2293 cl->bpp[cl->bufcount++] = bp; 2294 vp = bp->b_vp; 2295 s = splbio(); 2296 reassignbuf(bp, vp); 2297 V_INCR_NUMOUTPUT(vp); 2298 splx(s); 2299 2300 bpp++; 2301 i--; 2302 } 2303 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2304 BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL); 2305 else 2306 BIO_SETPRIO(cbp, BPRIO_TIMELIMITED); 2307 s = splbio(); 2308 V_INCR_NUMOUTPUT(devvp); 2309 splx(s); 2310 VOP_STRATEGY(devvp, cbp); 2311 curproc->p_stats->p_ru.ru_oublock++; 2312 } 2313 2314 if (lfs_dostats) { 2315 ++lfs_stats.psegwrites; 2316 lfs_stats.blocktot += nblocks - 1; 2317 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 2318 ++lfs_stats.psyncwrites; 2319 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 2320 ++lfs_stats.pcleanwrites; 2321 lfs_stats.cleanblocks += nblocks - 1; 2322 } 2323 } 2324 return (lfs_initseg(fs) || do_again); 2325 } 2326 2327 void 2328 lfs_writesuper(struct lfs *fs, daddr_t daddr) 2329 { 2330 struct buf *bp; 2331 int s; 2332 struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2333 2334 ASSERT_MAYBE_SEGLOCK(fs); 2335 #ifdef DIAGNOSTIC 2336 KASSERT(fs->lfs_magic == LFS_MAGIC); 2337 #endif 2338 /* 2339 * If we can write one superblock while another is in 2340 * progress, we risk not having a complete checkpoint if we crash. 2341 * So, block here if a superblock write is in progress. 2342 */ 2343 simple_lock(&fs->lfs_interlock); 2344 s = splbio(); 2345 while (fs->lfs_sbactive) { 2346 ltsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0, 2347 &fs->lfs_interlock); 2348 } 2349 fs->lfs_sbactive = daddr; 2350 splx(s); 2351 simple_unlock(&fs->lfs_interlock); 2352 2353 /* Set timestamp of this version of the superblock */ 2354 if (fs->lfs_version == 1) 2355 fs->lfs_otstamp = time_second; 2356 fs->lfs_tstamp = time_second; 2357 2358 /* Checksum the superblock and copy it into a buffer. */ 2359 fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs)); 2360 bp = lfs_newbuf(fs, devvp, 2361 fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK); 2362 memset(bp->b_data + sizeof(struct dlfs), 0, 2363 LFS_SBPAD - sizeof(struct dlfs)); 2364 *(struct dlfs *)bp->b_data = fs->lfs_dlfs; 2365 2366 bp->b_flags |= B_BUSY | B_CALL | B_ASYNC; 2367 bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI); 2368 bp->b_iodone = lfs_supercallback; 2369 2370 if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC) 2371 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 2372 else 2373 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 2374 curproc->p_stats->p_ru.ru_oublock++; 2375 s = splbio(); 2376 V_INCR_NUMOUTPUT(bp->b_vp); 2377 splx(s); 2378 simple_lock(&fs->lfs_interlock); 2379 ++fs->lfs_iocount; 2380 simple_unlock(&fs->lfs_interlock); 2381 VOP_STRATEGY(devvp, bp); 2382 } 2383 2384 /* 2385 * Logical block number match routines used when traversing the dirty block 2386 * chain. 2387 */ 2388 int 2389 lfs_match_fake(struct lfs *fs, struct buf *bp) 2390 { 2391 2392 ASSERT_SEGLOCK(fs); 2393 return LFS_IS_MALLOC_BUF(bp); 2394 } 2395 2396 #if 0 2397 int 2398 lfs_match_real(struct lfs *fs, struct buf *bp) 2399 { 2400 2401 ASSERT_SEGLOCK(fs); 2402 return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp)); 2403 } 2404 #endif 2405 2406 int 2407 lfs_match_data(struct lfs *fs, struct buf *bp) 2408 { 2409 2410 ASSERT_SEGLOCK(fs); 2411 return (bp->b_lblkno >= 0); 2412 } 2413 2414 int 2415 lfs_match_indir(struct lfs *fs, struct buf *bp) 2416 { 2417 daddr_t lbn; 2418 2419 ASSERT_SEGLOCK(fs); 2420 lbn = bp->b_lblkno; 2421 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0); 2422 } 2423 2424 int 2425 lfs_match_dindir(struct lfs *fs, struct buf *bp) 2426 { 2427 daddr_t lbn; 2428 2429 ASSERT_SEGLOCK(fs); 2430 lbn = bp->b_lblkno; 2431 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1); 2432 } 2433 2434 int 2435 lfs_match_tindir(struct lfs *fs, struct buf *bp) 2436 { 2437 daddr_t lbn; 2438 2439 ASSERT_SEGLOCK(fs); 2440 lbn = bp->b_lblkno; 2441 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2); 2442 } 2443 2444 static void 2445 lfs_free_aiodone(struct buf *bp) 2446 { 2447 struct lfs *fs; 2448 2449 fs = bp->b_private; 2450 ASSERT_NO_SEGLOCK(fs); 2451 lfs_freebuf(fs, bp); 2452 } 2453 2454 static void 2455 lfs_super_aiodone(struct buf *bp) 2456 { 2457 struct lfs *fs; 2458 2459 fs = bp->b_private; 2460 ASSERT_NO_SEGLOCK(fs); 2461 simple_lock(&fs->lfs_interlock); 2462 fs->lfs_sbactive = 0; 2463 if (--fs->lfs_iocount <= 1) 2464 wakeup(&fs->lfs_iocount); 2465 simple_unlock(&fs->lfs_interlock); 2466 wakeup(&fs->lfs_sbactive); 2467 lfs_freebuf(fs, bp); 2468 } 2469 2470 static void 2471 lfs_cluster_aiodone(struct buf *bp) 2472 { 2473 struct lfs_cluster *cl; 2474 struct lfs *fs; 2475 struct buf *tbp, *fbp; 2476 struct vnode *vp, *devvp; 2477 struct inode *ip; 2478 int s, error=0; 2479 2480 if (bp->b_flags & B_ERROR) 2481 error = bp->b_error; 2482 2483 cl = bp->b_private; 2484 fs = cl->fs; 2485 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 2486 ASSERT_NO_SEGLOCK(fs); 2487 2488 /* Put the pages back, and release the buffer */ 2489 while (cl->bufcount--) { 2490 tbp = cl->bpp[cl->bufcount]; 2491 KASSERT(tbp->b_flags & B_BUSY); 2492 if (error) { 2493 tbp->b_flags |= B_ERROR; 2494 tbp->b_error = error; 2495 } 2496 2497 /* 2498 * We're done with tbp. If it has not been re-dirtied since 2499 * the cluster was written, free it. Otherwise, keep it on 2500 * the locked list to be written again. 2501 */ 2502 vp = tbp->b_vp; 2503 2504 tbp->b_flags &= ~B_GATHERED; 2505 2506 LFS_BCLEAN_LOG(fs, tbp); 2507 2508 if (!(tbp->b_flags & B_CALL)) { 2509 KASSERT(tbp->b_flags & B_LOCKED); 2510 s = splbio(); 2511 simple_lock(&bqueue_slock); 2512 bremfree(tbp); 2513 simple_unlock(&bqueue_slock); 2514 if (vp) 2515 reassignbuf(tbp, vp); 2516 splx(s); 2517 tbp->b_flags |= B_ASYNC; /* for biodone */ 2518 } 2519 2520 if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED) 2521 LFS_UNLOCK_BUF(tbp); 2522 2523 if (tbp->b_flags & B_DONE) { 2524 DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n", 2525 cl->bufcount, (long)tbp->b_flags)); 2526 } 2527 2528 if ((tbp->b_flags & B_CALL) && !LFS_IS_MALLOC_BUF(tbp)) { 2529 /* 2530 * A buffer from the page daemon. 2531 * We use the same iodone as it does, 2532 * so we must manually disassociate its 2533 * buffers from the vp. 2534 */ 2535 if (tbp->b_vp) { 2536 /* This is just silly */ 2537 s = splbio(); 2538 brelvp(tbp); 2539 tbp->b_vp = vp; 2540 splx(s); 2541 } 2542 /* Put it back the way it was */ 2543 tbp->b_flags |= B_ASYNC; 2544 /* Master buffers have B_AGE */ 2545 if (tbp->b_private == tbp) 2546 tbp->b_flags |= B_AGE; 2547 } 2548 s = splbio(); 2549 biodone(tbp); 2550 2551 /* 2552 * If this is the last block for this vnode, but 2553 * there are other blocks on its dirty list, 2554 * set IN_MODIFIED/IN_CLEANING depending on what 2555 * sort of block. Only do this for our mount point, 2556 * not for, e.g., inode blocks that are attached to 2557 * the devvp. 2558 * XXX KS - Shouldn't we set *both* if both types 2559 * of blocks are present (traverse the dirty list?) 2560 */ 2561 simple_lock(&global_v_numoutput_slock); 2562 if (vp != devvp && vp->v_numoutput == 0 && 2563 (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) { 2564 ip = VTOI(vp); 2565 DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n", 2566 ip->i_number)); 2567 if (LFS_IS_MALLOC_BUF(fbp)) 2568 LFS_SET_UINO(ip, IN_CLEANING); 2569 else 2570 LFS_SET_UINO(ip, IN_MODIFIED); 2571 } 2572 simple_unlock(&global_v_numoutput_slock); 2573 splx(s); 2574 wakeup(vp); 2575 } 2576 2577 /* Fix up the cluster buffer, and release it */ 2578 if (cl->flags & LFS_CL_MALLOC) 2579 lfs_free(fs, bp->b_data, LFS_NB_CLUSTER); 2580 putiobuf(bp); 2581 2582 /* Note i/o done */ 2583 if (cl->flags & LFS_CL_SYNC) { 2584 if (--cl->seg->seg_iocount == 0) 2585 wakeup(&cl->seg->seg_iocount); 2586 } 2587 simple_lock(&fs->lfs_interlock); 2588 #ifdef DIAGNOSTIC 2589 if (fs->lfs_iocount == 0) 2590 panic("lfs_cluster_aiodone: zero iocount"); 2591 #endif 2592 if (--fs->lfs_iocount <= 1) 2593 wakeup(&fs->lfs_iocount); 2594 simple_unlock(&fs->lfs_interlock); 2595 2596 pool_put(&fs->lfs_bpppool, cl->bpp); 2597 cl->bpp = NULL; 2598 pool_put(&fs->lfs_clpool, cl); 2599 } 2600 2601 static void 2602 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *)) 2603 { 2604 /* reset b_iodone for when this is a single-buf i/o. */ 2605 bp->b_iodone = aiodone; 2606 2607 simple_lock(&uvm.aiodoned_lock); /* locks uvm.aio_done */ 2608 TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist); 2609 wakeup(&uvm.aiodoned); 2610 simple_unlock(&uvm.aiodoned_lock); 2611 } 2612 2613 static void 2614 lfs_cluster_callback(struct buf *bp) 2615 { 2616 2617 lfs_generic_callback(bp, lfs_cluster_aiodone); 2618 } 2619 2620 void 2621 lfs_supercallback(struct buf *bp) 2622 { 2623 2624 lfs_generic_callback(bp, lfs_super_aiodone); 2625 } 2626 2627 /* 2628 * The only buffers that are going to hit these functions are the 2629 * segment write blocks, or the segment summaries, or the superblocks. 2630 * 2631 * All of the above are created by lfs_newbuf, and so do not need to be 2632 * released via brelse. 2633 */ 2634 void 2635 lfs_callback(struct buf *bp) 2636 { 2637 2638 lfs_generic_callback(bp, lfs_free_aiodone); 2639 } 2640 2641 /* 2642 * Shellsort (diminishing increment sort) from Data Structures and 2643 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 2644 * see also Knuth Vol. 3, page 84. The increments are selected from 2645 * formula (8), page 95. Roughly O(N^3/2). 2646 */ 2647 /* 2648 * This is our own private copy of shellsort because we want to sort 2649 * two parallel arrays (the array of buffer pointers and the array of 2650 * logical block numbers) simultaneously. Note that we cast the array 2651 * of logical block numbers to a unsigned in this routine so that the 2652 * negative block numbers (meta data blocks) sort AFTER the data blocks. 2653 */ 2654 2655 void 2656 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size) 2657 { 2658 static int __rsshell_increments[] = { 4, 1, 0 }; 2659 int incr, *incrp, t1, t2; 2660 struct buf *bp_temp; 2661 2662 #ifdef DEBUG 2663 incr = 0; 2664 for (t1 = 0; t1 < nmemb; t1++) { 2665 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2666 if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) { 2667 /* dump before panic */ 2668 printf("lfs_shellsort: nmemb=%d, size=%d\n", 2669 nmemb, size); 2670 incr = 0; 2671 for (t1 = 0; t1 < nmemb; t1++) { 2672 const struct buf *bp = bp_array[t1]; 2673 2674 printf("bp[%d]: lbn=%" PRIu64 ", size=%" 2675 PRIu64 "\n", t1, 2676 (uint64_t)bp->b_bcount, 2677 (uint64_t)bp->b_lblkno); 2678 printf("lbns:"); 2679 for (t2 = 0; t2 * size < bp->b_bcount; 2680 t2++) { 2681 printf(" %" PRId32, 2682 lb_array[incr++]); 2683 } 2684 printf("\n"); 2685 } 2686 panic("lfs_shellsort: inconsistent input"); 2687 } 2688 } 2689 } 2690 #endif 2691 2692 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 2693 for (t1 = incr; t1 < nmemb; ++t1) 2694 for (t2 = t1 - incr; t2 >= 0;) 2695 if ((u_int32_t)bp_array[t2]->b_lblkno > 2696 (u_int32_t)bp_array[t2 + incr]->b_lblkno) { 2697 bp_temp = bp_array[t2]; 2698 bp_array[t2] = bp_array[t2 + incr]; 2699 bp_array[t2 + incr] = bp_temp; 2700 t2 -= incr; 2701 } else 2702 break; 2703 2704 /* Reform the list of logical blocks */ 2705 incr = 0; 2706 for (t1 = 0; t1 < nmemb; t1++) { 2707 for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) { 2708 lb_array[incr++] = bp_array[t1]->b_lblkno + t2; 2709 } 2710 } 2711 } 2712 2713 /* 2714 * Call vget with LK_NOWAIT. If we are the one who holds VXLOCK/VFREEING, 2715 * however, we must press on. Just fake success in that case. 2716 */ 2717 int 2718 lfs_vref(struct vnode *vp) 2719 { 2720 int error; 2721 struct lfs *fs; 2722 2723 fs = VTOI(vp)->i_lfs; 2724 2725 ASSERT_MAYBE_SEGLOCK(fs); 2726 2727 /* 2728 * If we return 1 here during a flush, we risk vinvalbuf() not 2729 * being able to flush all of the pages from this vnode, which 2730 * will cause it to panic. So, return 0 if a flush is in progress. 2731 */ 2732 error = vget(vp, LK_NOWAIT); 2733 if (error == EBUSY && IS_FLUSHING(VTOI(vp)->i_lfs, vp)) { 2734 ++fs->lfs_flushvp_fakevref; 2735 return 0; 2736 } 2737 return error; 2738 } 2739 2740 /* 2741 * This is vrele except that we do not want to VOP_INACTIVE this vnode. We 2742 * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end. 2743 */ 2744 void 2745 lfs_vunref(struct vnode *vp) 2746 { 2747 struct lfs *fs; 2748 2749 fs = VTOI(vp)->i_lfs; 2750 ASSERT_MAYBE_SEGLOCK(fs); 2751 2752 /* 2753 * Analogous to lfs_vref, if the node is flushing, fake it. 2754 */ 2755 if (IS_FLUSHING(fs, vp) && fs->lfs_flushvp_fakevref) { 2756 --fs->lfs_flushvp_fakevref; 2757 return; 2758 } 2759 2760 simple_lock(&vp->v_interlock); 2761 #ifdef DIAGNOSTIC 2762 if (vp->v_usecount <= 0) { 2763 printf("lfs_vunref: inum is %llu\n", (unsigned long long) 2764 VTOI(vp)->i_number); 2765 printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag); 2766 printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount); 2767 panic("lfs_vunref: v_usecount < 0"); 2768 } 2769 #endif 2770 vp->v_usecount--; 2771 if (vp->v_usecount > 0) { 2772 simple_unlock(&vp->v_interlock); 2773 return; 2774 } 2775 /* 2776 * insert at tail of LRU list 2777 */ 2778 simple_lock(&vnode_free_list_slock); 2779 if (vp->v_holdcnt > 0) 2780 TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist); 2781 else 2782 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 2783 simple_unlock(&vnode_free_list_slock); 2784 simple_unlock(&vp->v_interlock); 2785 } 2786 2787 /* 2788 * We use this when we have vnodes that were loaded in solely for cleaning. 2789 * There is no reason to believe that these vnodes will be referenced again 2790 * soon, since the cleaning process is unrelated to normal filesystem 2791 * activity. Putting cleaned vnodes at the tail of the list has the effect 2792 * of flushing the vnode LRU. So, put vnodes that were loaded only for 2793 * cleaning at the head of the list, instead. 2794 */ 2795 void 2796 lfs_vunref_head(struct vnode *vp) 2797 { 2798 2799 ASSERT_SEGLOCK(VTOI(vp)->i_lfs); 2800 simple_lock(&vp->v_interlock); 2801 #ifdef DIAGNOSTIC 2802 if (vp->v_usecount == 0) { 2803 panic("lfs_vunref: v_usecount<0"); 2804 } 2805 #endif 2806 vp->v_usecount--; 2807 if (vp->v_usecount > 0) { 2808 simple_unlock(&vp->v_interlock); 2809 return; 2810 } 2811 /* 2812 * insert at head of LRU list 2813 */ 2814 simple_lock(&vnode_free_list_slock); 2815 if (vp->v_holdcnt > 0) 2816 TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist); 2817 else 2818 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 2819 simple_unlock(&vnode_free_list_slock); 2820 simple_unlock(&vp->v_interlock); 2821 } 2822 2823 2824 /* 2825 * Set up an FINFO entry for a new file. The fip pointer is assumed to 2826 * point at uninitialized space. 2827 */ 2828 void 2829 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers) 2830 { 2831 struct segment *sp = fs->lfs_sp; 2832 2833 KASSERT(vers > 0); 2834 2835 if (sp->seg_bytes_left < fs->lfs_bsize || 2836 sp->sum_bytes_left < sizeof(struct finfo)) 2837 (void) lfs_writeseg(fs, fs->lfs_sp); 2838 2839 sp->sum_bytes_left -= FINFOSIZE; 2840 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 2841 sp->fip->fi_nblocks = 0; 2842 sp->fip->fi_ino = ino; 2843 sp->fip->fi_version = vers; 2844 } 2845 2846 /* 2847 * Release the FINFO entry, either clearing out an unused entry or 2848 * advancing us to the next available entry. 2849 */ 2850 void 2851 lfs_release_finfo(struct lfs *fs) 2852 { 2853 struct segment *sp = fs->lfs_sp; 2854 2855 if (sp->fip->fi_nblocks != 0) { 2856 sp->fip = (FINFO*)((caddr_t)sp->fip + FINFOSIZE + 2857 sizeof(int32_t) * sp->fip->fi_nblocks); 2858 sp->start_lbp = &sp->fip->fi_blocks[0]; 2859 } else { 2860 sp->sum_bytes_left += FINFOSIZE; 2861 --((SEGSUM *)(sp->segsum))->ss_nfinfo; 2862 } 2863 } 2864