1 /* $NetBSD: lfs_segment.c,v 1.31 1999/10/01 22:07:42 mycroft Exp $ */ 2 3 /*- 4 * Copyright (c) 1999 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. All advertising materials mentioning features or use of this software 51 * must display the following acknowledgement: 52 * This product includes software developed by the University of 53 * California, Berkeley and its contributors. 54 * 4. Neither the name of the University nor the names of its contributors 55 * may be used to endorse or promote products derived from this software 56 * without specific prior written permission. 57 * 58 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 59 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 60 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 61 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 62 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 63 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 64 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 65 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 66 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 67 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 68 * SUCH DAMAGE. 69 * 70 * @(#)lfs_segment.c 8.10 (Berkeley) 6/10/95 71 */ 72 73 #define ivndebug(vp,str) printf("ino %d: %s\n",VTOI(vp)->i_number,(str)) 74 75 #include "opt_ddb.h" 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/namei.h> 79 #include <sys/kernel.h> 80 #include <sys/resourcevar.h> 81 #include <sys/file.h> 82 #include <sys/stat.h> 83 #include <sys/buf.h> 84 #include <sys/proc.h> 85 #include <sys/conf.h> 86 #include <sys/vnode.h> 87 #include <sys/malloc.h> 88 #include <sys/mount.h> 89 90 #include <miscfs/specfs/specdev.h> 91 #include <miscfs/fifofs/fifo.h> 92 93 #include <ufs/ufs/quota.h> 94 #include <ufs/ufs/inode.h> 95 #include <ufs/ufs/dir.h> 96 #include <ufs/ufs/ufsmount.h> 97 #include <ufs/ufs/ufs_extern.h> 98 99 #include <ufs/lfs/lfs.h> 100 #include <ufs/lfs/lfs_extern.h> 101 102 extern int count_lock_queue __P((void)); 103 extern struct simplelock vnode_free_list_slock; /* XXX */ 104 extern TAILQ_HEAD(freelst, vnode) vnode_free_list; /* XXX */ 105 106 /* 107 * Determine if it's OK to start a partial in this segment, or if we need 108 * to go on to a new segment. 109 */ 110 #define LFS_PARTIAL_FITS(fs) \ 111 ((fs)->lfs_dbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \ 112 1 << (fs)->lfs_fsbtodb) 113 114 void lfs_callback __P((struct buf *)); 115 int lfs_gather __P((struct lfs *, struct segment *, 116 struct vnode *, int (*) __P((struct lfs *, struct buf *)))); 117 int lfs_gatherblock __P((struct segment *, struct buf *, int *)); 118 void lfs_iset __P((struct inode *, ufs_daddr_t, time_t)); 119 int lfs_match_fake __P((struct lfs *, struct buf *)); 120 int lfs_match_data __P((struct lfs *, struct buf *)); 121 int lfs_match_dindir __P((struct lfs *, struct buf *)); 122 int lfs_match_indir __P((struct lfs *, struct buf *)); 123 int lfs_match_tindir __P((struct lfs *, struct buf *)); 124 void lfs_newseg __P((struct lfs *)); 125 void lfs_shellsort __P((struct buf **, ufs_daddr_t *, register int)); 126 void lfs_supercallback __P((struct buf *)); 127 void lfs_updatemeta __P((struct segment *)); 128 int lfs_vref __P((struct vnode *)); 129 void lfs_vunref __P((struct vnode *)); 130 void lfs_writefile __P((struct lfs *, struct segment *, struct vnode *)); 131 int lfs_writeinode __P((struct lfs *, struct segment *, struct inode *)); 132 int lfs_writeseg __P((struct lfs *, struct segment *)); 133 void lfs_writesuper __P((struct lfs *, daddr_t)); 134 int lfs_writevnodes __P((struct lfs *fs, struct mount *mp, 135 struct segment *sp, int dirops)); 136 137 int lfs_allclean_wakeup; /* Cleaner wakeup address. */ 138 int lfs_writeindir = 1; /* whether to flush indir on non-ckp */ 139 int lfs_clean_vnhead = 0; /* Allow freeing to head of vn list */ 140 141 /* Statistics Counters */ 142 int lfs_dostats = 1; 143 struct lfs_stats lfs_stats; 144 145 /* op values to lfs_writevnodes */ 146 #define VN_REG 0 147 #define VN_DIROP 1 148 #define VN_EMPTY 2 149 #define VN_CLEAN 3 150 151 #define LFS_MAX_ACTIVE 10 152 153 /* 154 * XXX KS - Set modification time on the Ifile, so the cleaner can 155 * read the fs mod time off of it. We don't set IN_UPDATE here, 156 * since we don't really need this to be flushed to disk (and in any 157 * case that wouldn't happen to the Ifile until we checkpoint). 158 */ 159 void 160 lfs_imtime(fs) 161 struct lfs *fs; 162 { 163 struct timespec ts; 164 struct inode *ip; 165 166 TIMEVAL_TO_TIMESPEC(&time, &ts); 167 ip = VTOI(fs->lfs_ivnode); 168 ip->i_ffs_mtime = ts.tv_sec; 169 ip->i_ffs_mtimensec = ts.tv_nsec; 170 } 171 172 /* 173 * Ifile and meta data blocks are not marked busy, so segment writes MUST be 174 * single threaded. Currently, there are two paths into lfs_segwrite, sync() 175 * and getnewbuf(). They both mark the file system busy. Lfs_vflush() 176 * explicitly marks the file system busy. So lfs_segwrite is safe. I think. 177 */ 178 179 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp) 180 #define IS_FLUSHING(fs,vp) ((fs)->lfs_flushvp == (vp)) 181 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL 182 183 int 184 lfs_vflush(vp) 185 struct vnode *vp; 186 { 187 struct inode *ip; 188 struct lfs *fs; 189 struct segment *sp; 190 struct buf *bp, *nbp; 191 int error, s; 192 193 ip = VTOI(vp); 194 fs = VFSTOUFS(vp->v_mount)->um_lfs; 195 196 if(ip->i_flag & IN_CLEANING) { 197 #ifdef DEBUG_LFS 198 ivndebug(vp,"vflush/in_cleaning"); 199 #endif 200 ip->i_flag &= ~IN_CLEANING; 201 if(ip->i_flag & IN_MODIFIED) { 202 fs->lfs_uinodes--; 203 } else 204 ip->i_flag |= IN_MODIFIED; 205 } 206 207 /* If the node is being written, wait until that is done */ 208 if(WRITEINPROG(vp)) { 209 #ifdef DEBUG_LFS 210 ivndebug(vp,"vflush/writeinprog"); 211 #endif 212 tsleep(vp, PRIBIO+1, "lfs_vw", 0); 213 } 214 215 /* Protect against VXLOCK deadlock in vinvalbuf() */ 216 lfs_seglock(fs, SEGM_SYNC); 217 218 /* If we're supposed to flush a freed inode, just toss it */ 219 /* XXX - seglock, so these buffers can't be gathered, right? */ 220 if(ip->i_ffs_mode == 0) { 221 printf("lfs_vflush: ino %d is freed, not flushing\n", 222 ip->i_number); 223 s = splbio(); 224 for(bp=vp->v_dirtyblkhd.lh_first; bp; bp=nbp) { 225 nbp = bp->b_vnbufs.le_next; 226 /* Copied from lfs_writeseg */ 227 if (bp->b_flags & B_CALL) { 228 /* if B_CALL, it was created with newbuf */ 229 lfs_freebuf(bp); 230 } else { 231 bremfree(bp); 232 bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | 233 B_LOCKED | B_GATHERED); 234 bp->b_flags |= B_DONE; 235 reassignbuf(bp, vp); 236 brelse(bp); 237 } 238 } 239 splx(s); 240 if(ip->i_flag & IN_CLEANING) 241 fs->lfs_uinodes--; 242 if(ip->i_flag & IN_MODIFIED) 243 fs->lfs_uinodes--; 244 ip->i_flag &= ~(IN_MODIFIED|IN_UPDATE|IN_ACCESS|IN_CHANGE|IN_CLEANING); 245 printf("lfs_vflush: done not flushing ino %d\n", 246 ip->i_number); 247 lfs_segunlock(fs); 248 return 0; 249 } 250 251 SET_FLUSHING(fs,vp); 252 if (fs->lfs_nactive > LFS_MAX_ACTIVE) { 253 error = lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP); 254 CLR_FLUSHING(fs,vp); 255 lfs_segunlock(fs); 256 return error; 257 } 258 sp = fs->lfs_sp; 259 260 if (vp->v_dirtyblkhd.lh_first == NULL) { 261 lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY); 262 } else if((ip->i_flag & IN_CLEANING) && (fs->lfs_sp->seg_flags & SEGM_CLEAN)) { 263 #ifdef DEBUG_LFS 264 ivndebug(vp,"vflush/clean"); 265 #endif 266 lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN); 267 } 268 else if(lfs_dostats) { 269 if(vp->v_dirtyblkhd.lh_first || (VTOI(vp)->i_flag & (IN_MODIFIED|IN_UPDATE|IN_ACCESS|IN_CHANGE|IN_CLEANING))) 270 ++lfs_stats.vflush_invoked; 271 #ifdef DEBUG_LFS 272 ivndebug(vp,"vflush"); 273 #endif 274 } 275 276 #ifdef DIAGNOSTIC 277 /* XXX KS This actually can happen right now, though it shouldn't(?) */ 278 if(vp->v_flag & VDIROP) { 279 printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n"); 280 /* panic("VDIROP being flushed...this can\'t happen"); */ 281 } 282 if(vp->v_usecount<0) { 283 printf("usecount=%ld\n",vp->v_usecount); 284 panic("lfs_vflush: usecount<0"); 285 } 286 #endif 287 288 do { 289 do { 290 if (vp->v_dirtyblkhd.lh_first != NULL) 291 lfs_writefile(fs, sp, vp); 292 } while (lfs_writeinode(fs, sp, ip)); 293 } while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM); 294 295 if(lfs_dostats) { 296 ++lfs_stats.nwrites; 297 if (sp->seg_flags & SEGM_SYNC) 298 ++lfs_stats.nsync_writes; 299 if (sp->seg_flags & SEGM_CKP) 300 ++lfs_stats.ncheckpoints; 301 } 302 lfs_segunlock(fs); 303 304 CLR_FLUSHING(fs,vp); 305 return (0); 306 } 307 308 #ifdef DEBUG_LFS_VERBOSE 309 # define vndebug(vp,str) if(VTOI(vp)->i_flag & IN_CLEANING) printf("not writing ino %d because %s (op %d)\n",VTOI(vp)->i_number,(str),op) 310 #else 311 # define vndebug(vp,str) 312 #endif 313 314 int 315 lfs_writevnodes(fs, mp, sp, op) 316 struct lfs *fs; 317 struct mount *mp; 318 struct segment *sp; 319 int op; 320 { 321 struct inode *ip; 322 struct vnode *vp; 323 int inodes_written=0, only_cleaning; 324 325 #ifndef LFS_NO_BACKVP_HACK 326 /* BEGIN HACK */ 327 #define VN_OFFSET (((caddr_t)&vp->v_mntvnodes.le_next) - (caddr_t)vp) 328 #define BACK_VP(VP) ((struct vnode *)(((caddr_t)VP->v_mntvnodes.le_prev) - VN_OFFSET)) 329 #define BEG_OF_VLIST ((struct vnode *)(((caddr_t)&mp->mnt_vnodelist.lh_first) - VN_OFFSET)) 330 331 /* Find last vnode. */ 332 loop: for (vp = mp->mnt_vnodelist.lh_first; 333 vp && vp->v_mntvnodes.le_next != NULL; 334 vp = vp->v_mntvnodes.le_next); 335 for (; vp && vp != BEG_OF_VLIST; vp = BACK_VP(vp)) { 336 #else 337 loop: 338 for (vp = mp->mnt_vnodelist.lh_first; 339 vp != NULL; 340 vp = vp->v_mntvnodes.le_next) { 341 #endif 342 /* 343 * If the vnode that we are about to sync is no longer 344 * associated with this mount point, start over. 345 */ 346 if (vp->v_mount != mp) 347 goto loop; 348 349 ip = VTOI(vp); 350 if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) || 351 (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) { 352 vndebug(vp,"dirop"); 353 continue; 354 } 355 356 if (op == VN_EMPTY && vp->v_dirtyblkhd.lh_first) { 357 vndebug(vp,"empty"); 358 continue; 359 } 360 361 if (vp->v_type == VNON) { 362 continue; 363 } 364 365 if(op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM 366 && !(ip->i_flag & IN_CLEANING)) { 367 vndebug(vp,"cleaning"); 368 continue; 369 } 370 371 if (lfs_vref(vp)) { 372 vndebug(vp,"vref"); 373 continue; 374 } 375 376 #if 0 /* XXX KS - if we skip the ifile, things could go badly for us. */ 377 if(WRITEINPROG(vp)) { 378 lfs_vunref(vp); 379 #ifdef DEBUG_LFS 380 ivndebug(vp,"writevnodes/writeinprog"); 381 #endif 382 continue; 383 } 384 #endif 385 only_cleaning = 0; 386 /* 387 * Write the inode/file if dirty and it's not the 388 * the IFILE. 389 */ 390 if ((ip->i_flag & 391 (IN_ACCESS | IN_CHANGE | IN_MODIFIED | IN_UPDATE | IN_CLEANING) || 392 vp->v_dirtyblkhd.lh_first != NULL)) 393 { 394 only_cleaning = ((ip->i_flag & (IN_ACCESS|IN_CHANGE|IN_MODIFIED|IN_UPDATE|IN_CLEANING))==IN_CLEANING); 395 396 if(ip->i_number != LFS_IFILE_INUM 397 && vp->v_dirtyblkhd.lh_first != NULL) 398 { 399 lfs_writefile(fs, sp, vp); 400 } 401 if(vp->v_dirtyblkhd.lh_first != NULL) { 402 if(WRITEINPROG(vp)) { 403 #ifdef DEBUG_LFS 404 ivndebug(vp,"writevnodes/write2"); 405 #endif 406 } else if(!(ip->i_flag & (IN_ACCESS|IN_CHANGE|IN_MODIFIED|IN_UPDATE|IN_CLEANING))) { 407 #ifdef DEBUG_LFS 408 printf("<%d>",ip->i_number); 409 #endif 410 ip->i_flag |= IN_MODIFIED; 411 ++fs->lfs_uinodes; 412 } 413 } 414 (void) lfs_writeinode(fs, sp, ip); 415 inodes_written++; 416 } 417 418 if(vp->v_flag & VDIROP) { 419 --fs->lfs_dirvcount; 420 vp->v_flag &= ~VDIROP; 421 wakeup(&fs->lfs_dirvcount); 422 lfs_vunref(vp); 423 } 424 425 if(lfs_clean_vnhead && only_cleaning) 426 lfs_vunref_head(vp); 427 else 428 lfs_vunref(vp); 429 } 430 return inodes_written; 431 } 432 433 int 434 lfs_segwrite(mp, flags) 435 struct mount *mp; 436 int flags; /* Do a checkpoint. */ 437 { 438 struct buf *bp; 439 struct inode *ip; 440 struct lfs *fs; 441 struct segment *sp; 442 struct vnode *vp; 443 SEGUSE *segusep; 444 ufs_daddr_t ibno; 445 int do_ckp, error, i; 446 int writer_set = 0; 447 int need_unlock = 0; 448 449 fs = VFSTOUFS(mp)->um_lfs; 450 451 lfs_imtime(fs); 452 453 /* 454 * If we are not the cleaner, and we have fewer than MIN_FREE_SEGS 455 * clean segments, wait until cleaner writes. 456 */ 457 if(!(flags & SEGM_CLEAN) 458 && (!fs->lfs_seglock || !(fs->lfs_sp->seg_flags & SEGM_CLEAN))) 459 { 460 do { 461 if (fs->lfs_nclean <= MIN_FREE_SEGS 462 || fs->lfs_avail <= 0) 463 { 464 wakeup(&lfs_allclean_wakeup); 465 wakeup(&fs->lfs_nextseg); 466 error = tsleep(&fs->lfs_avail, PRIBIO + 1, 467 "lfs_avail", 0); 468 if (error) { 469 return (error); 470 } 471 } 472 } while (fs->lfs_nclean <= MIN_FREE_SEGS || fs->lfs_avail <= 0); 473 } 474 475 /* 476 * Allocate a segment structure and enough space to hold pointers to 477 * the maximum possible number of buffers which can be described in a 478 * single summary block. 479 */ 480 do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE; 481 lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0)); 482 sp = fs->lfs_sp; 483 484 /* 485 * If lfs_flushvp is non-NULL, we are called from lfs_vflush, 486 * in which case we have to flush *all* buffers off of this vnode. 487 */ 488 if((sp->seg_flags & SEGM_CLEAN) && !(fs->lfs_flushvp)) 489 lfs_writevnodes(fs, mp, sp, VN_CLEAN); 490 else { 491 lfs_writevnodes(fs, mp, sp, VN_REG); 492 /* 493 * XXX KS - If we're cleaning, we can't wait for dirops, 494 * because they might be waiting on us. The downside of this 495 * is that, if we write anything besides cleaning blocks 496 * while cleaning, the checkpoint is not completely 497 * consistent. 498 */ 499 if(!(sp->seg_flags & SEGM_CLEAN)) { 500 while(fs->lfs_dirops) 501 if((error = tsleep(&fs->lfs_writer, PRIBIO + 1, 502 "lfs writer", 0))) 503 { 504 free(sp->bpp, M_SEGMENT); 505 free(sp, M_SEGMENT); 506 return (error); 507 } 508 fs->lfs_writer++; 509 writer_set=1; 510 lfs_writevnodes(fs, mp, sp, VN_DIROP); 511 ((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT); 512 } 513 } 514 515 /* 516 * If we are doing a checkpoint, mark everything since the 517 * last checkpoint as no longer ACTIVE. 518 */ 519 if (do_ckp) { 520 for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz; 521 --ibno >= fs->lfs_cleansz; ) { 522 if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp)) 523 524 panic("lfs_segwrite: ifile read"); 525 segusep = (SEGUSE *)bp->b_data; 526 for (i = fs->lfs_sepb; i--; segusep++) 527 segusep->su_flags &= ~SEGUSE_ACTIVE; 528 529 /* But the current segment is still ACTIVE */ 530 if (fs->lfs_curseg/fs->lfs_sepb==(ibno-fs->lfs_cleansz)) 531 ((SEGUSE *)(bp->b_data))[fs->lfs_curseg%fs->lfs_sepb].su_flags |= SEGUSE_ACTIVE; 532 error = VOP_BWRITE(bp); 533 } 534 } 535 536 if (do_ckp || fs->lfs_doifile) { 537 redo: 538 vp = fs->lfs_ivnode; 539 /* 540 * Depending on the circumstances of our calling, the ifile 541 * inode might be locked. If it is, and if it is locked by 542 * us, we should VREF instead of vget here. 543 */ 544 need_unlock = 0; 545 if(VOP_ISLOCKED(vp) 546 && vp->v_lock.lk_lockholder == curproc->p_pid) { 547 VREF(vp); 548 } else { 549 while (vget(vp, LK_EXCLUSIVE)) 550 continue; 551 need_unlock = 1; 552 } 553 ip = VTOI(vp); 554 if (vp->v_dirtyblkhd.lh_first != NULL) 555 lfs_writefile(fs, sp, vp); 556 (void)lfs_writeinode(fs, sp, ip); 557 558 /* Only vput if we used vget() above. */ 559 if(need_unlock) 560 vput(vp); 561 else 562 vrele(vp); 563 564 if (lfs_writeseg(fs, sp) && do_ckp) 565 goto redo; 566 } else { 567 (void) lfs_writeseg(fs, sp); 568 } 569 570 /* 571 * If the I/O count is non-zero, sleep until it reaches zero. 572 * At the moment, the user's process hangs around so we can 573 * sleep. 574 */ 575 fs->lfs_doifile = 0; 576 if(writer_set && --fs->lfs_writer==0) 577 wakeup(&fs->lfs_dirops); 578 579 if(lfs_dostats) { 580 ++lfs_stats.nwrites; 581 if (sp->seg_flags & SEGM_SYNC) 582 ++lfs_stats.nsync_writes; 583 if (sp->seg_flags & SEGM_CKP) 584 ++lfs_stats.ncheckpoints; 585 } 586 lfs_segunlock(fs); 587 return (0); 588 } 589 590 /* 591 * Write the dirty blocks associated with a vnode. 592 */ 593 void 594 lfs_writefile(fs, sp, vp) 595 struct lfs *fs; 596 struct segment *sp; 597 struct vnode *vp; 598 { 599 struct buf *bp; 600 struct finfo *fip; 601 IFILE *ifp; 602 603 604 if (sp->seg_bytes_left < fs->lfs_bsize || 605 sp->sum_bytes_left < sizeof(struct finfo)) 606 (void) lfs_writeseg(fs, sp); 607 608 sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t); 609 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 610 611 if(vp->v_flag & VDIROP) 612 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 613 614 fip = sp->fip; 615 fip->fi_nblocks = 0; 616 fip->fi_ino = VTOI(vp)->i_number; 617 LFS_IENTRY(ifp, fs, fip->fi_ino, bp); 618 fip->fi_version = ifp->if_version; 619 brelse(bp); 620 621 /* 622 * It may not be necessary to write the meta-data blocks at this point, 623 * as the roll-forward recovery code should be able to reconstruct the 624 * list. 625 * 626 * We have to write them anyway, though, under two conditions: (1) the 627 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are 628 * checkpointing. 629 */ 630 if((sp->seg_flags & SEGM_CLEAN) 631 && VTOI(vp)->i_number != LFS_IFILE_INUM 632 && !IS_FLUSHING(fs,vp)) 633 { 634 lfs_gather(fs, sp, vp, lfs_match_fake); 635 } else 636 lfs_gather(fs, sp, vp, lfs_match_data); 637 638 if(lfs_writeindir 639 || IS_FLUSHING(fs,vp) 640 || (sp->seg_flags & SEGM_CKP)) 641 { 642 lfs_gather(fs, sp, vp, lfs_match_indir); 643 lfs_gather(fs, sp, vp, lfs_match_dindir); 644 /* XXX KS - when is TRIPLE not true? */ /* #ifdef TRIPLE */ 645 lfs_gather(fs, sp, vp, lfs_match_tindir); 646 /* #endif */ 647 } 648 fip = sp->fip; 649 if (fip->fi_nblocks != 0) { 650 sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) + 651 sizeof(ufs_daddr_t) * (fip->fi_nblocks-1)); 652 sp->start_lbp = &sp->fip->fi_blocks[0]; 653 } else { 654 sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t); 655 --((SEGSUM *)(sp->segsum))->ss_nfinfo; 656 } 657 } 658 659 int 660 lfs_writeinode(fs, sp, ip) 661 struct lfs *fs; 662 struct segment *sp; 663 struct inode *ip; 664 { 665 struct buf *bp, *ibp; 666 IFILE *ifp; 667 SEGUSE *sup; 668 ufs_daddr_t daddr; 669 ino_t ino; 670 int error, i, ndx; 671 int redo_ifile = 0; 672 struct timespec ts; 673 int gotblk=0; 674 675 if (!(ip->i_flag & (IN_ACCESS | IN_CHANGE | IN_MODIFIED | IN_UPDATE | IN_CLEANING))) 676 return(0); 677 678 /* Allocate a new inode block if necessary. */ 679 if ((ip->i_number != LFS_IFILE_INUM || sp->idp==NULL) && sp->ibp == NULL) { 680 /* Allocate a new segment if necessary. */ 681 if (sp->seg_bytes_left < fs->lfs_bsize || 682 sp->sum_bytes_left < sizeof(ufs_daddr_t)) 683 (void) lfs_writeseg(fs, sp); 684 685 /* Get next inode block. */ 686 daddr = fs->lfs_offset; 687 fs->lfs_offset += fsbtodb(fs, 1); 688 sp->ibp = *sp->cbpp++ = 689 getblk(VTOI(fs->lfs_ivnode)->i_devvp, daddr, fs->lfs_bsize, 0, 0); 690 gotblk++; 691 692 /* Zero out inode numbers */ 693 for (i = 0; i < INOPB(fs); ++i) 694 ((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0; 695 696 ++sp->start_bpp; 697 fs->lfs_avail -= fsbtodb(fs, 1); 698 /* Set remaining space counters. */ 699 sp->seg_bytes_left -= fs->lfs_bsize; 700 sp->sum_bytes_left -= sizeof(ufs_daddr_t); 701 ndx = LFS_SUMMARY_SIZE / sizeof(ufs_daddr_t) - 702 sp->ninodes / INOPB(fs) - 1; 703 ((ufs_daddr_t *)(sp->segsum))[ndx] = daddr; 704 } 705 706 /* Update the inode times and copy the inode onto the inode page. */ 707 if (ip->i_flag & (IN_CLEANING|IN_MODIFIED)) 708 --fs->lfs_uinodes; 709 TIMEVAL_TO_TIMESPEC(&time, &ts); 710 LFS_ITIMES(ip, &ts, &ts, &ts); 711 712 if(ip->i_flag & IN_CLEANING) 713 ip->i_flag &= ~IN_CLEANING; 714 else 715 ip->i_flag &= ~(IN_ACCESS|IN_CHANGE|IN_MODIFIED|IN_UPDATE); 716 717 /* 718 * If this is the Ifile, and we've already written the Ifile in this 719 * partial segment, just overwrite it (it's not on disk yet) and 720 * continue. 721 * 722 * XXX we know that the bp that we get the second time around has 723 * already been gathered. 724 */ 725 if(ip->i_number == LFS_IFILE_INUM && sp->idp) { 726 *(sp->idp) = ip->i_din.ffs_din; 727 return 0; 728 } 729 730 bp = sp->ibp; 731 ((struct dinode *)bp->b_data)[sp->ninodes % INOPB(fs)] = 732 ip->i_din.ffs_din; 733 734 if(ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */ 735 sp->idp = ((struct dinode *)bp->b_data)+(sp->ninodes % INOPB(fs)); 736 if(gotblk) { 737 bp->b_flags |= B_LOCKED; 738 brelse(bp); 739 } 740 741 /* Increment inode count in segment summary block. */ 742 ++((SEGSUM *)(sp->segsum))->ss_ninos; 743 744 /* If this page is full, set flag to allocate a new page. */ 745 if (++sp->ninodes % INOPB(fs) == 0) 746 sp->ibp = NULL; 747 748 /* 749 * If updating the ifile, update the super-block. Update the disk 750 * address and access times for this inode in the ifile. 751 */ 752 ino = ip->i_number; 753 if (ino == LFS_IFILE_INUM) { 754 daddr = fs->lfs_idaddr; 755 fs->lfs_idaddr = bp->b_blkno; 756 } else { 757 LFS_IENTRY(ifp, fs, ino, ibp); 758 daddr = ifp->if_daddr; 759 ifp->if_daddr = bp->b_blkno; 760 #ifdef LFS_DEBUG_NEXTFREE 761 if(ino > 3 && ifp->if_nextfree) { 762 vprint("lfs_writeinode",ITOV(ip)); 763 printf("lfs_writeinode: updating free ino %d\n", 764 ip->i_number); 765 } 766 #endif 767 error = VOP_BWRITE(ibp); 768 } 769 770 /* 771 * No need to update segment usage if there was no former inode address 772 * or if the last inode address is in the current partial segment. 773 */ 774 if (daddr != LFS_UNUSED_DADDR && 775 !(daddr >= fs->lfs_lastpseg && daddr <= bp->b_blkno)) { 776 LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp); 777 #ifdef DIAGNOSTIC 778 if (sup->su_nbytes < DINODE_SIZE) { 779 /* XXX -- Change to a panic. */ 780 printf("lfs_writeinode: negative bytes (segment %d short by %d)\n", 781 datosn(fs, daddr), (int)DINODE_SIZE - sup->su_nbytes); 782 panic("lfs_writeinode: negative bytes"); 783 sup->su_nbytes = DINODE_SIZE; 784 } 785 #endif 786 sup->su_nbytes -= DINODE_SIZE; 787 redo_ifile = 788 (ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED)); 789 error = VOP_BWRITE(bp); 790 } 791 return (redo_ifile); 792 } 793 794 int 795 lfs_gatherblock(sp, bp, sptr) 796 struct segment *sp; 797 struct buf *bp; 798 int *sptr; 799 { 800 struct lfs *fs; 801 int version; 802 803 /* 804 * If full, finish this segment. We may be doing I/O, so 805 * release and reacquire the splbio(). 806 */ 807 #ifdef DIAGNOSTIC 808 if (sp->vp == NULL) 809 panic ("lfs_gatherblock: Null vp in segment"); 810 #endif 811 fs = sp->fs; 812 if (sp->sum_bytes_left < sizeof(ufs_daddr_t) || 813 sp->seg_bytes_left < bp->b_bcount) { 814 if (sptr) 815 splx(*sptr); 816 lfs_updatemeta(sp); 817 818 version = sp->fip->fi_version; 819 (void) lfs_writeseg(fs, sp); 820 821 sp->fip->fi_version = version; 822 sp->fip->fi_ino = VTOI(sp->vp)->i_number; 823 /* Add the current file to the segment summary. */ 824 ++((SEGSUM *)(sp->segsum))->ss_nfinfo; 825 sp->sum_bytes_left -= 826 sizeof(struct finfo) - sizeof(ufs_daddr_t); 827 828 if (sptr) 829 *sptr = splbio(); 830 return(1); 831 } 832 833 #ifdef DEBUG 834 if(bp->b_flags & B_GATHERED) { 835 printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n", 836 sp->fip->fi_ino, bp->b_lblkno); 837 return(0); 838 } 839 #endif 840 /* Insert into the buffer list, update the FINFO block. */ 841 bp->b_flags |= B_GATHERED; 842 *sp->cbpp++ = bp; 843 sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno; 844 845 sp->sum_bytes_left -= sizeof(ufs_daddr_t); 846 sp->seg_bytes_left -= bp->b_bcount; 847 return(0); 848 } 849 850 int 851 lfs_gather(fs, sp, vp, match) 852 struct lfs *fs; 853 struct segment *sp; 854 struct vnode *vp; 855 int (*match) __P((struct lfs *, struct buf *)); 856 { 857 struct buf *bp; 858 int s, count=0; 859 860 sp->vp = vp; 861 s = splbio(); 862 863 #ifndef LFS_NO_BACKBUF_HACK 864 loop: for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = bp->b_vnbufs.le_next) { 865 #else /* LFS_NO_BACKBUF_HACK */ 866 /* This is a hack to see if ordering the blocks in LFS makes a difference. */ 867 # define BUF_OFFSET (((void *)&bp->b_vnbufs.le_next) - (void *)bp) 868 # define BACK_BUF(BP) ((struct buf *)(((void *)BP->b_vnbufs.le_prev) - BUF_OFFSET)) 869 # define BEG_OF_LIST ((struct buf *)(((void *)&vp->v_dirtyblkhd.lh_first) - BUF_OFFSET)) 870 /* Find last buffer. */ 871 loop: for (bp = vp->v_dirtyblkhd.lh_first; bp && bp->b_vnbufs.le_next != NULL; 872 bp = bp->b_vnbufs.le_next); 873 for (; bp && bp != BEG_OF_LIST; bp = BACK_BUF(bp)) { 874 #endif /* LFS_NO_BACKBUF_HACK */ 875 if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) 876 continue; 877 if(vp->v_type == VBLK) { 878 /* For block devices, just write the blocks. */ 879 /* XXX Do we really need to even do this? */ 880 #ifdef DEBUG_LFS 881 if(count==0) 882 printf("BLK("); 883 printf("."); 884 #endif 885 /* Get the block before bwrite, so we don't corrupt the free list */ 886 bp->b_flags |= B_BUSY; 887 bremfree(bp); 888 bwrite(bp); 889 } else { 890 #ifdef DIAGNOSTIC 891 if (!(bp->b_flags & B_DELWRI)) 892 panic("lfs_gather: bp not B_DELWRI"); 893 if (!(bp->b_flags & B_LOCKED)) { 894 printf("lfs_gather: lbn %d blk %d not B_LOCKED\n", bp->b_lblkno, bp->b_blkno); 895 VOP_PRINT(bp->b_vp); 896 panic("lfs_gather: bp not B_LOCKED"); 897 } 898 #endif 899 if (lfs_gatherblock(sp, bp, &s)) { 900 goto loop; 901 } 902 } 903 count++; 904 } 905 splx(s); 906 #ifdef DEBUG_LFS 907 if(vp->v_type == VBLK && count) 908 printf(")\n"); 909 #endif 910 lfs_updatemeta(sp); 911 sp->vp = NULL; 912 return count; 913 } 914 915 /* 916 * Update the metadata that points to the blocks listed in the FINFO 917 * array. 918 */ 919 void 920 lfs_updatemeta(sp) 921 struct segment *sp; 922 { 923 SEGUSE *sup; 924 struct buf *bp; 925 struct lfs *fs; 926 struct vnode *vp; 927 struct indir a[NIADDR + 2], *ap; 928 struct inode *ip; 929 ufs_daddr_t daddr, lbn, off; 930 int error, i, nblocks, num; 931 932 vp = sp->vp; 933 nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp; 934 if (nblocks < 0) 935 panic("This is a bad thing\n"); 936 if (vp == NULL || nblocks == 0) 937 return; 938 939 /* Sort the blocks. */ 940 /* 941 * XXX KS - We have to sort even if the blocks come from the 942 * cleaner, because there might be other pending blocks on the 943 * same inode...and if we don't sort, and there are fragments 944 * present, blocks may be written in the wrong place. 945 */ 946 /* if (!(sp->seg_flags & SEGM_CLEAN)) */ 947 lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks); 948 949 /* 950 * Record the length of the last block in case it's a fragment. 951 * If there are indirect blocks present, they sort last. An 952 * indirect block will be lfs_bsize and its presence indicates 953 * that you cannot have fragments. 954 */ 955 sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount; 956 957 /* 958 * Assign disk addresses, and update references to the logical 959 * block and the segment usage information. 960 */ 961 fs = sp->fs; 962 for (i = nblocks; i--; ++sp->start_bpp) { 963 lbn = *sp->start_lbp++; 964 965 (*sp->start_bpp)->b_blkno = off = fs->lfs_offset; 966 if((*sp->start_bpp)->b_blkno == (*sp->start_bpp)->b_lblkno) { 967 printf("lfs_updatemeta: ino %d blk %d has same lbn and daddr\n", VTOI(vp)->i_number, off); 968 } 969 fs->lfs_offset += 970 fragstodb(fs, numfrags(fs, (*sp->start_bpp)->b_bcount)); 971 error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL); 972 if (error) 973 panic("lfs_updatemeta: ufs_bmaparray %d", error); 974 ip = VTOI(vp); 975 switch (num) { 976 case 0: 977 ip->i_ffs_db[lbn] = off; 978 break; 979 case 1: 980 ip->i_ffs_ib[a[0].in_off] = off; 981 break; 982 default: 983 ap = &a[num - 1]; 984 if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp)) 985 panic("lfs_updatemeta: bread bno %d", 986 ap->in_lbn); 987 /* 988 * Bread may create a new (indirect) block which needs 989 * to get counted for the inode. 990 */ 991 if (/* bp->b_blkno == -1 && */ 992 !(bp->b_flags & (B_DELWRI|B_DONE))) { 993 ip->i_ffs_blocks += fsbtodb(fs, 1); 994 fs->lfs_bfree -= fragstodb(fs, fs->lfs_frag); 995 } 996 ((ufs_daddr_t *)bp->b_data)[ap->in_off] = off; 997 VOP_BWRITE(bp); 998 } 999 /* Update segment usage information. */ 1000 if (daddr != UNASSIGNED && !(daddr >= fs->lfs_lastpseg && daddr <= off)) { 1001 LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp); 1002 #ifdef DIAGNOSTIC 1003 if (sup->su_nbytes < (*sp->start_bpp)->b_bcount) { 1004 /* XXX -- Change to a panic. */ 1005 printf("lfs_updatemeta: negative bytes (segment %d short by %ld)\n", 1006 datosn(fs, daddr), (*sp->start_bpp)->b_bcount - sup->su_nbytes); 1007 printf("lfs_updatemeta: ino %d, lbn %d, addr = %x\n", 1008 VTOI(sp->vp)->i_number, (*sp->start_bpp)->b_lblkno, daddr); 1009 panic("lfs_updatemeta: negative bytes"); 1010 sup->su_nbytes = (*sp->start_bpp)->b_bcount; 1011 } 1012 #endif 1013 sup->su_nbytes -= (*sp->start_bpp)->b_bcount; 1014 error = VOP_BWRITE(bp); 1015 } 1016 } 1017 } 1018 1019 /* 1020 * Start a new segment. 1021 */ 1022 int 1023 lfs_initseg(fs) 1024 struct lfs *fs; 1025 { 1026 struct segment *sp; 1027 SEGUSE *sup; 1028 SEGSUM *ssp; 1029 struct buf *bp; 1030 int repeat; 1031 1032 sp = fs->lfs_sp; 1033 1034 repeat = 0; 1035 /* Advance to the next segment. */ 1036 if (!LFS_PARTIAL_FITS(fs)) { 1037 /* Wake up any cleaning procs waiting on this file system. */ 1038 wakeup(&lfs_allclean_wakeup); 1039 wakeup(&fs->lfs_nextseg); 1040 lfs_newseg(fs); 1041 repeat = 1; 1042 fs->lfs_offset = fs->lfs_curseg; 1043 sp->seg_number = datosn(fs, fs->lfs_curseg); 1044 sp->seg_bytes_left = fs->lfs_dbpseg * DEV_BSIZE; 1045 /* 1046 * If the segment contains a superblock, update the offset 1047 * and summary address to skip over it. 1048 */ 1049 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1050 if (sup->su_flags & SEGUSE_SUPERBLOCK) { 1051 fs->lfs_offset += LFS_SBPAD / DEV_BSIZE; 1052 sp->seg_bytes_left -= LFS_SBPAD; 1053 } 1054 brelse(bp); 1055 } else { 1056 sp->seg_number = datosn(fs, fs->lfs_curseg); 1057 sp->seg_bytes_left = (fs->lfs_dbpseg - 1058 (fs->lfs_offset - fs->lfs_curseg)) * DEV_BSIZE; 1059 } 1060 fs->lfs_lastpseg = fs->lfs_offset; 1061 1062 sp->fs = fs; 1063 sp->ibp = NULL; 1064 sp->idp = NULL; 1065 sp->ninodes = 0; 1066 1067 /* Get a new buffer for SEGSUM and enter it into the buffer list. */ 1068 sp->cbpp = sp->bpp; 1069 *sp->cbpp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, 1070 fs->lfs_offset, LFS_SUMMARY_SIZE); 1071 sp->segsum = (*sp->cbpp)->b_data; 1072 bzero(sp->segsum, LFS_SUMMARY_SIZE); 1073 sp->start_bpp = ++sp->cbpp; 1074 fs->lfs_offset += LFS_SUMMARY_SIZE / DEV_BSIZE; 1075 1076 /* Set point to SEGSUM, initialize it. */ 1077 ssp = sp->segsum; 1078 ssp->ss_next = fs->lfs_nextseg; 1079 ssp->ss_nfinfo = ssp->ss_ninos = 0; 1080 ssp->ss_magic = SS_MAGIC; 1081 1082 /* Set pointer to first FINFO, initialize it. */ 1083 sp->fip = (struct finfo *)((caddr_t)sp->segsum + sizeof(SEGSUM)); 1084 sp->fip->fi_nblocks = 0; 1085 sp->start_lbp = &sp->fip->fi_blocks[0]; 1086 sp->fip->fi_lastlength = 0; 1087 1088 sp->seg_bytes_left -= LFS_SUMMARY_SIZE; 1089 sp->sum_bytes_left = LFS_SUMMARY_SIZE - sizeof(SEGSUM); 1090 1091 return(repeat); 1092 } 1093 1094 /* 1095 * Return the next segment to write. 1096 */ 1097 void 1098 lfs_newseg(fs) 1099 struct lfs *fs; 1100 { 1101 CLEANERINFO *cip; 1102 SEGUSE *sup; 1103 struct buf *bp; 1104 int curseg, isdirty, sn; 1105 1106 LFS_SEGENTRY(sup, fs, datosn(fs, fs->lfs_nextseg), bp); 1107 sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE; 1108 sup->su_nbytes = 0; 1109 sup->su_nsums = 0; 1110 sup->su_ninos = 0; 1111 (void) VOP_BWRITE(bp); 1112 1113 LFS_CLEANERINFO(cip, fs, bp); 1114 --cip->clean; 1115 ++cip->dirty; 1116 fs->lfs_nclean = cip->clean; 1117 (void) VOP_BWRITE(bp); 1118 1119 fs->lfs_lastseg = fs->lfs_curseg; 1120 fs->lfs_curseg = fs->lfs_nextseg; 1121 for (sn = curseg = datosn(fs, fs->lfs_curseg);;) { 1122 sn = (sn + 1) % fs->lfs_nseg; 1123 if (sn == curseg) 1124 panic("lfs_nextseg: no clean segments"); 1125 LFS_SEGENTRY(sup, fs, sn, bp); 1126 isdirty = sup->su_flags & SEGUSE_DIRTY; 1127 brelse(bp); 1128 if (!isdirty) 1129 break; 1130 } 1131 1132 ++fs->lfs_nactive; 1133 fs->lfs_nextseg = sntoda(fs, sn); 1134 if(lfs_dostats) { 1135 ++lfs_stats.segsused; 1136 } 1137 } 1138 1139 int 1140 lfs_writeseg(fs, sp) 1141 struct lfs *fs; 1142 struct segment *sp; 1143 { 1144 extern int locked_queue_count; 1145 extern long locked_queue_bytes; 1146 struct buf **bpp, *bp, *cbp; 1147 SEGUSE *sup; 1148 SEGSUM *ssp; 1149 dev_t i_dev; 1150 u_long *datap, *dp; 1151 int do_again, i, nblocks, s; 1152 #ifdef LFS_TRACK_IOS 1153 int j; 1154 #endif 1155 int (*strategy)__P((void *)); 1156 struct vop_strategy_args vop_strategy_a; 1157 u_short ninos; 1158 struct vnode *devvp; 1159 char *p; 1160 struct vnode *vn; 1161 struct inode *ip; 1162 #if defined(DEBUG) && defined(LFS_PROPELLER) 1163 static int propeller; 1164 char propstring[4] = "-\\|/"; 1165 1166 printf("%c\b",propstring[propeller++]); 1167 if(propeller==4) 1168 propeller = 0; 1169 #endif 1170 1171 /* 1172 * If there are no buffers other than the segment summary to write 1173 * and it is not a checkpoint, don't do anything. On a checkpoint, 1174 * even if there aren't any buffers, you need to write the superblock. 1175 */ 1176 if ((nblocks = sp->cbpp - sp->bpp) == 1) 1177 return (0); 1178 1179 #ifdef DEBUG_LFS 1180 lfs_check_bpp(fs,sp,__FILE__,__LINE__); 1181 #endif 1182 i_dev = VTOI(fs->lfs_ivnode)->i_dev; 1183 devvp = VTOI(fs->lfs_ivnode)->i_devvp; 1184 1185 /* Update the segment usage information. */ 1186 LFS_SEGENTRY(sup, fs, sp->seg_number, bp); 1187 1188 /* Loop through all blocks, except the segment summary. */ 1189 for (bpp = sp->bpp; ++bpp < sp->cbpp; ) { 1190 if((*bpp)->b_vp != devvp) 1191 sup->su_nbytes += (*bpp)->b_bcount; 1192 } 1193 1194 ssp = (SEGSUM *)sp->segsum; 1195 1196 ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs); 1197 sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE; 1198 /* sup->su_nbytes += LFS_SUMMARY_SIZE; */ 1199 sup->su_lastmod = time.tv_sec; 1200 sup->su_ninos += ninos; 1201 ++sup->su_nsums; 1202 1203 do_again = !(bp->b_flags & B_GATHERED); 1204 (void)VOP_BWRITE(bp); 1205 /* 1206 * Compute checksum across data and then across summary; the first 1207 * block (the summary block) is skipped. Set the create time here 1208 * so that it's guaranteed to be later than the inode mod times. 1209 * 1210 * XXX 1211 * Fix this to do it inline, instead of malloc/copy. 1212 */ 1213 datap = dp = malloc(nblocks * sizeof(u_long), M_SEGMENT, M_WAITOK); 1214 for (bpp = sp->bpp, i = nblocks - 1; i--;) { 1215 if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) { 1216 if (copyin((*bpp)->b_saveaddr, dp++, sizeof(u_long))) 1217 panic("lfs_writeseg: copyin failed [1]: ino %d blk %d", VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno); 1218 } else { 1219 if( !((*bpp)->b_flags & B_CALL) ) { 1220 /* 1221 * Before we record data for a checksm, 1222 * make sure the data won't change in between 1223 * the checksum calculation and the write, 1224 * by marking the buffer B_BUSY. It will 1225 * be freed later by brelse(). 1226 */ 1227 again: 1228 s = splbio(); 1229 if((*bpp)->b_flags & B_BUSY) { 1230 #ifdef DEBUG 1231 printf("lfs_writeseg: avoiding potential data summary corruption for ino %d, lbn %d\n", 1232 VTOI((*bpp)->b_vp)->i_number, 1233 bp->b_lblkno); 1234 #endif 1235 (*bpp)->b_flags |= B_WANTED; 1236 tsleep((*bpp), (PRIBIO + 1), 1237 "lfs_writeseg", 0); 1238 splx(s); 1239 goto again; 1240 } 1241 (*bpp)->b_flags |= B_BUSY; 1242 splx(s); 1243 } 1244 *dp++ = ((u_long *)(*bpp)->b_data)[0]; 1245 } 1246 } 1247 ssp->ss_create = time.tv_sec; 1248 ssp->ss_datasum = cksum(datap, (nblocks - 1) * sizeof(u_long)); 1249 ssp->ss_sumsum = 1250 cksum(&ssp->ss_datasum, LFS_SUMMARY_SIZE - sizeof(ssp->ss_sumsum)); 1251 free(datap, M_SEGMENT); 1252 #ifdef DIAGNOSTIC 1253 if (fs->lfs_bfree < fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE) 1254 panic("lfs_writeseg: No diskspace for summary"); 1255 #endif 1256 fs->lfs_bfree -= (fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE); 1257 1258 strategy = devvp->v_op[VOFFSET(vop_strategy)]; 1259 1260 /* 1261 * When we simply write the blocks we lose a rotation for every block 1262 * written. To avoid this problem, we allocate memory in chunks, copy 1263 * the buffers into the chunk and write the chunk. CHUNKSIZE is the 1264 * largest size I/O devices can handle. 1265 * When the data is copied to the chunk, turn off the the B_LOCKED bit 1266 * and brelse the buffer (which will move them to the LRU list). Add 1267 * the B_CALL flag to the buffer header so we can count I/O's for the 1268 * checkpoints and so we can release the allocated memory. 1269 * 1270 * XXX 1271 * This should be removed if the new virtual memory system allows us to 1272 * easily make the buffers contiguous in kernel memory and if that's 1273 * fast enough. 1274 */ 1275 1276 #define CHUNKSIZE MAXPHYS 1277 1278 if(devvp==NULL) 1279 panic("devvp is NULL"); 1280 for (bpp = sp->bpp,i = nblocks; i;) { 1281 cbp = lfs_newbuf(devvp, (*bpp)->b_blkno, CHUNKSIZE); 1282 cbp->b_dev = i_dev; 1283 cbp->b_flags |= B_ASYNC | B_BUSY; 1284 cbp->b_bcount = 0; 1285 1286 #ifdef DIAGNOSTIC 1287 if(datosn(fs,(*bpp)->b_blkno + ((*bpp)->b_bcount - 1)/DEV_BSIZE) != datosn(fs,cbp->b_blkno)) { 1288 panic("lfs_writeseg: Segment overwrite"); 1289 } 1290 #endif 1291 1292 if(fs->lfs_iocount >= LFS_THROTTLE) { 1293 tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs throttle", 0); 1294 } 1295 s = splbio(); 1296 ++fs->lfs_iocount; 1297 #ifdef LFS_TRACK_IOS 1298 for(j=0;j<LFS_THROTTLE;j++) { 1299 if(fs->lfs_pending[j]==LFS_UNUSED_DADDR) { 1300 fs->lfs_pending[j] = cbp->b_blkno; 1301 break; 1302 } 1303 } 1304 #endif /* LFS_TRACK_IOS */ 1305 for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) { 1306 bp = *bpp; 1307 1308 if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount)) 1309 break; 1310 1311 /* 1312 * Fake buffers from the cleaner are marked as B_INVAL. 1313 * We need to copy the data from user space rather than 1314 * from the buffer indicated. 1315 * XXX == what do I do on an error? 1316 */ 1317 if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) { 1318 if (copyin(bp->b_saveaddr, p, bp->b_bcount)) 1319 panic("lfs_writeseg: copyin failed [2]"); 1320 } else 1321 bcopy(bp->b_data, p, bp->b_bcount); 1322 p += bp->b_bcount; 1323 cbp->b_bcount += bp->b_bcount; 1324 if (bp->b_flags & B_LOCKED) { 1325 --locked_queue_count; 1326 locked_queue_bytes -= bp->b_bufsize; 1327 } 1328 bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | 1329 B_LOCKED | B_GATHERED); 1330 vn = bp->b_vp; 1331 if (bp->b_flags & B_CALL) { 1332 /* if B_CALL, it was created with newbuf */ 1333 lfs_freebuf(bp); 1334 } else { 1335 bremfree(bp); 1336 bp->b_flags |= B_DONE; 1337 if(vn) 1338 reassignbuf(bp, vn); 1339 brelse(bp); 1340 } 1341 if(bp->b_flags & B_NEEDCOMMIT) { /* XXX */ 1342 bp->b_flags &= ~B_NEEDCOMMIT; 1343 wakeup(bp); 1344 } 1345 1346 bpp++; 1347 1348 /* 1349 * If this is the last block for this vnode, but 1350 * there are other blocks on its dirty list, 1351 * set IN_MODIFIED/IN_CLEANING depending on what 1352 * sort of block. Only do this for our mount point, 1353 * not for, e.g., inode blocks that are attached to 1354 * the devvp. 1355 */ 1356 if(i>1 && vn && *bpp && (*bpp)->b_vp != vn 1357 && (*bpp)->b_vp && (bp=vn->v_dirtyblkhd.lh_first)!=NULL && 1358 vn->v_mount == fs->lfs_ivnode->v_mount) 1359 { 1360 ip = VTOI(vn); 1361 #ifdef DEBUG_LFS 1362 printf("lfs_writeseg: marking ino %d\n",ip->i_number); 1363 #endif 1364 if(!(ip->i_flag & (IN_CLEANING|IN_MODIFIED))) { 1365 fs->lfs_uinodes++; 1366 if(bp->b_flags & B_CALL) 1367 ip->i_flag |= IN_CLEANING; 1368 else 1369 ip->i_flag |= IN_MODIFIED; 1370 } 1371 } 1372 /* if(vn->v_dirtyblkhd.lh_first == NULL) */ 1373 wakeup(vn); 1374 } 1375 ++cbp->b_vp->v_numoutput; 1376 splx(s); 1377 /* 1378 * XXXX This is a gross and disgusting hack. Since these 1379 * buffers are physically addressed, they hang off the 1380 * device vnode (devvp). As a result, they have no way 1381 * of getting to the LFS superblock or lfs structure to 1382 * keep track of the number of I/O's pending. So, I am 1383 * going to stuff the fs into the saveaddr field of 1384 * the buffer (yuk). 1385 */ 1386 cbp->b_saveaddr = (caddr_t)fs; 1387 vop_strategy_a.a_desc = VDESC(vop_strategy); 1388 vop_strategy_a.a_bp = cbp; 1389 (strategy)(&vop_strategy_a); 1390 } 1391 /* 1392 * XXX 1393 * Vinvalbuf can move locked buffers off the locked queue 1394 * and we have no way of knowing about this. So, after 1395 * doing a big write, we recalculate how many buffers are 1396 * really still left on the locked queue. 1397 */ 1398 lfs_countlocked(&locked_queue_count,&locked_queue_bytes); 1399 wakeup(&locked_queue_count); 1400 if(lfs_dostats) { 1401 ++lfs_stats.psegwrites; 1402 lfs_stats.blocktot += nblocks - 1; 1403 if (fs->lfs_sp->seg_flags & SEGM_SYNC) 1404 ++lfs_stats.psyncwrites; 1405 if (fs->lfs_sp->seg_flags & SEGM_CLEAN) { 1406 ++lfs_stats.pcleanwrites; 1407 lfs_stats.cleanblocks += nblocks - 1; 1408 } 1409 } 1410 return (lfs_initseg(fs) || do_again); 1411 } 1412 1413 void 1414 lfs_writesuper(fs, daddr) 1415 struct lfs *fs; 1416 daddr_t daddr; 1417 { 1418 struct buf *bp; 1419 dev_t i_dev; 1420 int (*strategy) __P((void *)); 1421 int s; 1422 struct vop_strategy_args vop_strategy_a; 1423 1424 #ifdef LFS_CANNOT_ROLLFW 1425 /* 1426 * If we can write one superblock while another is in 1427 * progress, we risk not having a complete checkpoint if we crash. 1428 * So, block here if a superblock write is in progress. 1429 * 1430 * XXX - should be a proper lock, not this hack 1431 */ 1432 while(fs->lfs_sbactive) { 1433 tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0); 1434 } 1435 fs->lfs_sbactive = daddr; 1436 #endif 1437 i_dev = VTOI(fs->lfs_ivnode)->i_dev; 1438 strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)]; 1439 1440 /* Set timestamp of this version of the superblock */ 1441 fs->lfs_tstamp = time.tv_sec; 1442 1443 /* Checksum the superblock and copy it into a buffer. */ 1444 fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs)); 1445 bp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, daddr, LFS_SBPAD); 1446 *(struct dlfs *)bp->b_data = fs->lfs_dlfs; 1447 1448 bp->b_dev = i_dev; 1449 bp->b_flags |= B_BUSY | B_CALL | B_ASYNC; 1450 bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI); 1451 bp->b_iodone = lfs_supercallback; 1452 /* XXX KS - same nasty hack as above */ 1453 bp->b_saveaddr = (caddr_t)fs; 1454 1455 vop_strategy_a.a_desc = VDESC(vop_strategy); 1456 vop_strategy_a.a_bp = bp; 1457 s = splbio(); 1458 ++bp->b_vp->v_numoutput; 1459 splx(s); 1460 (strategy)(&vop_strategy_a); 1461 } 1462 1463 /* 1464 * Logical block number match routines used when traversing the dirty block 1465 * chain. 1466 */ 1467 int 1468 lfs_match_fake(fs, bp) 1469 struct lfs *fs; 1470 struct buf *bp; 1471 { 1472 return (bp->b_flags & B_CALL); 1473 } 1474 1475 int 1476 lfs_match_data(fs, bp) 1477 struct lfs *fs; 1478 struct buf *bp; 1479 { 1480 return (bp->b_lblkno >= 0); 1481 } 1482 1483 int 1484 lfs_match_indir(fs, bp) 1485 struct lfs *fs; 1486 struct buf *bp; 1487 { 1488 int lbn; 1489 1490 lbn = bp->b_lblkno; 1491 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0); 1492 } 1493 1494 int 1495 lfs_match_dindir(fs, bp) 1496 struct lfs *fs; 1497 struct buf *bp; 1498 { 1499 int lbn; 1500 1501 lbn = bp->b_lblkno; 1502 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1); 1503 } 1504 1505 int 1506 lfs_match_tindir(fs, bp) 1507 struct lfs *fs; 1508 struct buf *bp; 1509 { 1510 int lbn; 1511 1512 lbn = bp->b_lblkno; 1513 return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2); 1514 } 1515 1516 /* 1517 * XXX - The only buffers that are going to hit these functions are the 1518 * segment write blocks, or the segment summaries, or the superblocks. 1519 * 1520 * All of the above are created by lfs_newbuf, and so do not need to be 1521 * released via brelse. 1522 */ 1523 void 1524 lfs_callback(bp) 1525 struct buf *bp; 1526 { 1527 struct lfs *fs; 1528 #ifdef LFS_TRACK_IOS 1529 int j; 1530 #endif 1531 1532 fs = (struct lfs *)bp->b_saveaddr; 1533 #ifdef DIAGNOSTIC 1534 if (fs->lfs_iocount == 0) 1535 panic("lfs_callback: zero iocount\n"); 1536 #endif 1537 if (--fs->lfs_iocount < LFS_THROTTLE) 1538 wakeup(&fs->lfs_iocount); 1539 #ifdef LFS_TRACK_IOS 1540 for(j=0;j<LFS_THROTTLE;j++) { 1541 if(fs->lfs_pending[j]==bp->b_blkno) { 1542 fs->lfs_pending[j] = LFS_UNUSED_DADDR; 1543 wakeup(&(fs->lfs_pending[j])); 1544 break; 1545 } 1546 } 1547 #endif /* LFS_TRACK_IOS */ 1548 1549 lfs_freebuf(bp); 1550 } 1551 1552 void 1553 lfs_supercallback(bp) 1554 struct buf *bp; 1555 { 1556 #ifdef LFS_CANNOT_ROLLFW 1557 struct lfs *fs; 1558 1559 fs = (struct lfs *)bp->b_saveaddr; 1560 fs->lfs_sbactive=NULL; 1561 wakeup(&fs->lfs_sbactive); 1562 #endif 1563 lfs_freebuf(bp); 1564 } 1565 1566 /* 1567 * Shellsort (diminishing increment sort) from Data Structures and 1568 * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290; 1569 * see also Knuth Vol. 3, page 84. The increments are selected from 1570 * formula (8), page 95. Roughly O(N^3/2). 1571 */ 1572 /* 1573 * This is our own private copy of shellsort because we want to sort 1574 * two parallel arrays (the array of buffer pointers and the array of 1575 * logical block numbers) simultaneously. Note that we cast the array 1576 * of logical block numbers to a unsigned in this routine so that the 1577 * negative block numbers (meta data blocks) sort AFTER the data blocks. 1578 */ 1579 1580 void 1581 lfs_shellsort(bp_array, lb_array, nmemb) 1582 struct buf **bp_array; 1583 ufs_daddr_t *lb_array; 1584 register int nmemb; 1585 { 1586 static int __rsshell_increments[] = { 4, 1, 0 }; 1587 register int incr, *incrp, t1, t2; 1588 struct buf *bp_temp; 1589 u_long lb_temp; 1590 1591 for (incrp = __rsshell_increments; (incr = *incrp++) != 0;) 1592 for (t1 = incr; t1 < nmemb; ++t1) 1593 for (t2 = t1 - incr; t2 >= 0;) 1594 if (lb_array[t2] > lb_array[t2 + incr]) { 1595 lb_temp = lb_array[t2]; 1596 lb_array[t2] = lb_array[t2 + incr]; 1597 lb_array[t2 + incr] = lb_temp; 1598 bp_temp = bp_array[t2]; 1599 bp_array[t2] = bp_array[t2 + incr]; 1600 bp_array[t2 + incr] = bp_temp; 1601 t2 -= incr; 1602 } else 1603 break; 1604 } 1605 1606 /* 1607 * Check VXLOCK. Return 1 if the vnode is locked. Otherwise, vget it. 1608 */ 1609 int 1610 lfs_vref(vp) 1611 register struct vnode *vp; 1612 { 1613 /* 1614 * If we return 1 here during a flush, we risk vinvalbuf() not 1615 * being able to flush all of the pages from this vnode, which 1616 * will cause it to panic. So, return 0 if a flush is in progress. 1617 */ 1618 if (vp->v_flag & VXLOCK) { 1619 if(IS_FLUSHING(VTOI(vp)->i_lfs,vp)) { 1620 return 0; 1621 } 1622 return(1); 1623 } 1624 return (vget(vp, 0)); 1625 } 1626 1627 /* 1628 * This is vrele except that we do not want to VOP_INACTIVE this vnode. We 1629 * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end. 1630 */ 1631 void 1632 lfs_vunref(vp) 1633 register struct vnode *vp; 1634 { 1635 /* 1636 * Analogous to lfs_vref, if the node is flushing, fake it. 1637 */ 1638 if((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) { 1639 return; 1640 } 1641 1642 simple_lock(&vp->v_interlock); 1643 #ifdef DIAGNOSTIC 1644 if(vp->v_usecount<=0) { 1645 printf("lfs_vunref: flags are 0x%lx\n", vp->v_flag); 1646 printf("lfs_vunref: usecount = %ld\n", vp->v_usecount); 1647 panic("lfs_vunref: v_usecount<0"); 1648 } 1649 #endif 1650 vp->v_usecount--; 1651 if (vp->v_usecount > 0) { 1652 simple_unlock(&vp->v_interlock); 1653 return; 1654 } 1655 #ifdef DIAGNOSTIC 1656 if(VOP_ISLOCKED(vp)) 1657 panic("lfs_vunref: vnode locked"); 1658 #endif 1659 /* 1660 * insert at tail of LRU list 1661 */ 1662 simple_lock(&vnode_free_list_slock); 1663 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); 1664 simple_unlock(&vnode_free_list_slock); 1665 simple_unlock(&vp->v_interlock); 1666 } 1667 1668 /* 1669 * We use this when we have vnodes that were loaded in solely for cleaning. 1670 * There is no reason to believe that these vnodes will be referenced again 1671 * soon, since the cleaning process is unrelated to normal filesystem 1672 * activity. Putting cleaned vnodes at the tail of the list has the effect 1673 * of flushing the vnode LRU. So, put vnodes that were loaded only for 1674 * cleaning at the head of the list, instead. 1675 */ 1676 void 1677 lfs_vunref_head(vp) 1678 register struct vnode *vp; 1679 { 1680 simple_lock(&vp->v_interlock); 1681 #ifdef DIAGNOSTIC 1682 if(vp->v_usecount==0) { 1683 panic("lfs_vunref: v_usecount<0"); 1684 } 1685 #endif 1686 vp->v_usecount--; 1687 if (vp->v_usecount > 0) { 1688 simple_unlock(&vp->v_interlock); 1689 return; 1690 } 1691 #ifdef DIAGNOSTIC 1692 if(VOP_ISLOCKED(vp)) 1693 panic("lfs_vunref_head: vnode locked"); 1694 #endif 1695 /* 1696 * insert at head of LRU list 1697 */ 1698 simple_lock(&vnode_free_list_slock); 1699 TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); 1700 simple_unlock(&vnode_free_list_slock); 1701 simple_unlock(&vp->v_interlock); 1702 } 1703 1704