xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 200d779b75dbeafa7bc01fd0f60bc61185f6967b)
1 /*	$NetBSD: lfs_segment.c,v 1.259 2015/09/01 06:08:37 dholland Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*
32  * Copyright (c) 1991, 1993
33  *	The Regents of the University of California.  All rights reserved.
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. Neither the name of the University nor the names of its contributors
44  *    may be used to endorse or promote products derived from this software
45  *    without specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
60  */
61 
62 #include <sys/cdefs.h>
63 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.259 2015/09/01 06:08:37 dholland Exp $");
64 
65 #ifdef DEBUG
66 # define vndebug(vp, str) do {						\
67 	if (VTOI(vp)->i_flag & IN_CLEANING)				\
68 		DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \
69 		     VTOI(vp)->i_number, (str), op));			\
70 } while(0)
71 #else
72 # define vndebug(vp, str)
73 #endif
74 #define ivndebug(vp, str) \
75 	DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str)))
76 
77 #if defined(_KERNEL_OPT)
78 #include "opt_ddb.h"
79 #endif
80 
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/namei.h>
84 #include <sys/kernel.h>
85 #include <sys/resourcevar.h>
86 #include <sys/file.h>
87 #include <sys/stat.h>
88 #include <sys/buf.h>
89 #include <sys/proc.h>
90 #include <sys/vnode.h>
91 #include <sys/mount.h>
92 #include <sys/kauth.h>
93 #include <sys/syslog.h>
94 
95 #include <miscfs/specfs/specdev.h>
96 #include <miscfs/fifofs/fifo.h>
97 
98 #include <ufs/lfs/ulfs_inode.h>
99 #include <ufs/lfs/ulfsmount.h>
100 #include <ufs/lfs/ulfs_extern.h>
101 
102 #include <ufs/lfs/lfs.h>
103 #include <ufs/lfs/lfs_accessors.h>
104 #include <ufs/lfs/lfs_kernel.h>
105 #include <ufs/lfs/lfs_extern.h>
106 
107 #include <uvm/uvm.h>
108 #include <uvm/uvm_extern.h>
109 
110 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
111 
112 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
113 static void lfs_free_aiodone(struct buf *);
114 static void lfs_super_aiodone(struct buf *);
115 static void lfs_cluster_aiodone(struct buf *);
116 static void lfs_cluster_callback(struct buf *);
117 
118 /*
119  * Determine if it's OK to start a partial in this segment, or if we need
120  * to go on to a new segment.
121  */
122 #define	LFS_PARTIAL_FITS(fs) \
123 	(lfs_sb_getfsbpseg(fs) - \
124 	    (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)) > \
125 	lfs_sb_getfrag(fs))
126 
127 /*
128  * Figure out whether we should do a checkpoint write or go ahead with
129  * an ordinary write.
130  */
131 #define LFS_SHOULD_CHECKPOINT(fs, flags) \
132         ((flags & SEGM_CLEAN) == 0 &&					\
133 	  ((fs->lfs_nactive > LFS_MAX_ACTIVE ||				\
134 	    (flags & SEGM_CKP) ||					\
135 	    lfs_sb_getnclean(fs) < LFS_MAX_ACTIVE)))
136 
137 int	 lfs_match_fake(struct lfs *, struct buf *);
138 void	 lfs_newseg(struct lfs *);
139 void	 lfs_supercallback(struct buf *);
140 void	 lfs_updatemeta(struct segment *);
141 void	 lfs_writesuper(struct lfs *, daddr_t);
142 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
143 	    struct segment *sp, int dirops);
144 
145 static void lfs_shellsort(struct lfs *, struct buf **, union lfs_blocks *,
146 			  int, int);
147 
148 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
149 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
150 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
151 int	lfs_dirvcount = 0;		/* # active dirops */
152 
153 /* Statistics Counters */
154 int lfs_dostats = 1;
155 struct lfs_stats lfs_stats;
156 
157 /* op values to lfs_writevnodes */
158 #define	VN_REG		0
159 #define	VN_DIROP	1
160 #define	VN_EMPTY	2
161 #define VN_CLEAN	3
162 
163 /*
164  * XXX KS - Set modification time on the Ifile, so the cleaner can
165  * read the fs mod time off of it.  We don't set IN_UPDATE here,
166  * since we don't really need this to be flushed to disk (and in any
167  * case that wouldn't happen to the Ifile until we checkpoint).
168  */
169 void
170 lfs_imtime(struct lfs *fs)
171 {
172 	struct timespec ts;
173 	struct inode *ip;
174 
175 	ASSERT_MAYBE_SEGLOCK(fs);
176 	vfs_timestamp(&ts);
177 	ip = VTOI(fs->lfs_ivnode);
178 	lfs_dino_setmtime(fs, ip->i_din, ts.tv_sec);
179 	lfs_dino_setmtimensec(fs, ip->i_din, ts.tv_nsec);
180 }
181 
182 /*
183  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
184  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
185  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
186  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
187  */
188 
189 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
190 
191 int
192 lfs_vflush(struct vnode *vp)
193 {
194 	struct inode *ip;
195 	struct lfs *fs;
196 	struct segment *sp;
197 	struct buf *bp, *nbp, *tbp, *tnbp;
198 	int error;
199 	int flushed;
200 	int relock;
201 
202 	ip = VTOI(vp);
203 	fs = VFSTOULFS(vp->v_mount)->um_lfs;
204 	relock = 0;
205 
206     top:
207 	KASSERT(mutex_owned(vp->v_interlock) == false);
208 	KASSERT(mutex_owned(&lfs_lock) == false);
209 	KASSERT(mutex_owned(&bufcache_lock) == false);
210 	ASSERT_NO_SEGLOCK(fs);
211 	if (ip->i_flag & IN_CLEANING) {
212 		ivndebug(vp,"vflush/in_cleaning");
213 		mutex_enter(&lfs_lock);
214 		LFS_CLR_UINO(ip, IN_CLEANING);
215 		LFS_SET_UINO(ip, IN_MODIFIED);
216 		mutex_exit(&lfs_lock);
217 
218 		/*
219 		 * Toss any cleaning buffers that have real counterparts
220 		 * to avoid losing new data.
221 		 */
222 		mutex_enter(vp->v_interlock);
223 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
224 			nbp = LIST_NEXT(bp, b_vnbufs);
225 			if (!LFS_IS_MALLOC_BUF(bp))
226 				continue;
227 			/*
228 			 * Look for pages matching the range covered
229 			 * by cleaning blocks.  It's okay if more dirty
230 			 * pages appear, so long as none disappear out
231 			 * from under us.
232 			 */
233 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
234 			    vp != fs->lfs_ivnode) {
235 				struct vm_page *pg;
236 				voff_t off;
237 
238 				for (off = lfs_lblktosize(fs, bp->b_lblkno);
239 				     off < lfs_lblktosize(fs, bp->b_lblkno + 1);
240 				     off += PAGE_SIZE) {
241 					pg = uvm_pagelookup(&vp->v_uobj, off);
242 					if (pg == NULL)
243 						continue;
244 					if ((pg->flags & PG_CLEAN) == 0 ||
245 					    pmap_is_modified(pg)) {
246 						lfs_sb_addavail(fs,
247 							lfs_btofsb(fs,
248 								bp->b_bcount));
249 						wakeup(&fs->lfs_availsleep);
250 						mutex_exit(vp->v_interlock);
251 						lfs_freebuf(fs, bp);
252 						mutex_enter(vp->v_interlock);
253 						bp = NULL;
254 						break;
255 					}
256 				}
257 			}
258 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
259 			    tbp = tnbp)
260 			{
261 				tnbp = LIST_NEXT(tbp, b_vnbufs);
262 				if (tbp->b_vp == bp->b_vp
263 				   && tbp->b_lblkno == bp->b_lblkno
264 				   && tbp != bp)
265 				{
266 					lfs_sb_addavail(fs, lfs_btofsb(fs,
267 						bp->b_bcount));
268 					wakeup(&fs->lfs_availsleep);
269 					mutex_exit(vp->v_interlock);
270 					lfs_freebuf(fs, bp);
271 					mutex_enter(vp->v_interlock);
272 					bp = NULL;
273 					break;
274 				}
275 			}
276 		}
277 	} else {
278 		mutex_enter(vp->v_interlock);
279 	}
280 
281 	/* If the node is being written, wait until that is done */
282 	while (WRITEINPROG(vp)) {
283 		ivndebug(vp,"vflush/writeinprog");
284 		cv_wait(&vp->v_cv, vp->v_interlock);
285 	}
286 	error = vdead_check(vp, VDEAD_NOWAIT);
287 	mutex_exit(vp->v_interlock);
288 
289 	/* Protect against deadlock in vinvalbuf() */
290 	lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0));
291 	if (error != 0) {
292 		fs->lfs_reclino = ip->i_number;
293 	}
294 
295 	/* If we're supposed to flush a freed inode, just toss it */
296 	if (ip->i_lfs_iflags & LFSI_DELETED) {
297 		DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n",
298 		      ip->i_number));
299 		/* Drain v_numoutput */
300 		mutex_enter(vp->v_interlock);
301 		while (vp->v_numoutput > 0) {
302 			cv_wait(&vp->v_cv, vp->v_interlock);
303 		}
304 		KASSERT(vp->v_numoutput == 0);
305 		mutex_exit(vp->v_interlock);
306 
307 		mutex_enter(&bufcache_lock);
308 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
309 			nbp = LIST_NEXT(bp, b_vnbufs);
310 
311 			KASSERT((bp->b_flags & B_GATHERED) == 0);
312 			if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */
313 				lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount));
314 				wakeup(&fs->lfs_availsleep);
315 			}
316 			/* Copied from lfs_writeseg */
317 			if (bp->b_iodone != NULL) {
318 				mutex_exit(&bufcache_lock);
319 				biodone(bp);
320 				mutex_enter(&bufcache_lock);
321 			} else {
322 				bremfree(bp);
323 				LFS_UNLOCK_BUF(bp);
324 				mutex_enter(vp->v_interlock);
325 				bp->b_flags &= ~(B_READ | B_GATHERED);
326 				bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE;
327 				bp->b_error = 0;
328 				reassignbuf(bp, vp);
329 				mutex_exit(vp->v_interlock);
330 				brelse(bp, 0);
331 			}
332 		}
333 		mutex_exit(&bufcache_lock);
334 		LFS_CLR_UINO(ip, IN_CLEANING);
335 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
336 		ip->i_flag &= ~IN_ALLMOD;
337 		DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n",
338 		      ip->i_number));
339 		lfs_segunlock(fs);
340 
341 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
342 
343 		return 0;
344 	}
345 
346 	fs->lfs_flushvp = vp;
347 	if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) {
348 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
349 		fs->lfs_flushvp = NULL;
350 		KASSERT(fs->lfs_flushvp_fakevref == 0);
351 		lfs_segunlock(fs);
352 
353 		/* Make sure that any pending buffers get written */
354 		mutex_enter(vp->v_interlock);
355 		while (vp->v_numoutput > 0) {
356 			cv_wait(&vp->v_cv, vp->v_interlock);
357 		}
358 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
359 		KASSERT(vp->v_numoutput == 0);
360 		mutex_exit(vp->v_interlock);
361 
362 		return error;
363 	}
364 	sp = fs->lfs_sp;
365 
366 	flushed = 0;
367 	if (VPISEMPTY(vp)) {
368 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
369 		++flushed;
370 	} else if ((ip->i_flag & IN_CLEANING) &&
371 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
372 		ivndebug(vp,"vflush/clean");
373 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
374 		++flushed;
375 	} else if (lfs_dostats) {
376 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
377 			++lfs_stats.vflush_invoked;
378 		ivndebug(vp,"vflush");
379 	}
380 
381 #ifdef DIAGNOSTIC
382 	if (vp->v_uflag & VU_DIROP) {
383 		DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n"));
384 		/* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */
385 	}
386 #endif
387 
388 	do {
389 #ifdef DEBUG
390 		int loopcount = 0;
391 #endif
392 		do {
393 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
394 				relock = lfs_writefile(fs, sp, vp);
395 				if (relock && vp != fs->lfs_ivnode) {
396 					/*
397 					 * Might have to wait for the
398 					 * cleaner to run; but we're
399 					 * still not done with this vnode.
400 					 * XXX we can do better than this.
401 					 */
402 					KDASSERT(ip->i_number != LFS_IFILE_INUM);
403 					lfs_writeinode(fs, sp, ip);
404 					mutex_enter(&lfs_lock);
405 					LFS_SET_UINO(ip, IN_MODIFIED);
406 					mutex_exit(&lfs_lock);
407 					lfs_writeseg(fs, sp);
408 					lfs_segunlock(fs);
409 					lfs_segunlock_relock(fs);
410 					goto top;
411 				}
412 			}
413 			/*
414 			 * If we begin a new segment in the middle of writing
415 			 * the Ifile, it creates an inconsistent checkpoint,
416 			 * since the Ifile information for the new segment
417 			 * is not up-to-date.  Take care of this here by
418 			 * sending the Ifile through again in case there
419 			 * are newly dirtied blocks.  But wait, there's more!
420 			 * This second Ifile write could *also* cross a segment
421 			 * boundary, if the first one was large.  The second
422 			 * one is guaranteed to be no more than 8 blocks,
423 			 * though (two segment blocks and supporting indirects)
424 			 * so the third write *will not* cross the boundary.
425 			 */
426 			if (vp == fs->lfs_ivnode) {
427 				lfs_writefile(fs, sp, vp);
428 				lfs_writefile(fs, sp, vp);
429 			}
430 #ifdef DEBUG
431 			if (++loopcount > 2)
432 				log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount);
433 #endif
434 		} while (lfs_writeinode(fs, sp, ip));
435 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
436 
437 	if (lfs_dostats) {
438 		++lfs_stats.nwrites;
439 		if (sp->seg_flags & SEGM_SYNC)
440 			++lfs_stats.nsync_writes;
441 		if (sp->seg_flags & SEGM_CKP)
442 			++lfs_stats.ncheckpoints;
443 	}
444 	/*
445 	 * If we were called from somewhere that has already held the seglock
446 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
447 	 * the write to complete because we are still locked.
448 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
449 	 * we must explicitly wait, if that is the case.
450 	 *
451 	 * We compare the iocount against 1, not 0, because it is
452 	 * artificially incremented by lfs_seglock().
453 	 */
454 	mutex_enter(&lfs_lock);
455 	if (fs->lfs_seglock > 1) {
456 		while (fs->lfs_iocount > 1)
457 			(void)mtsleep(&fs->lfs_iocount, PRIBIO + 1,
458 				     "lfs_vflush", 0, &lfs_lock);
459 	}
460 	mutex_exit(&lfs_lock);
461 
462 	lfs_segunlock(fs);
463 
464 	/* Wait for these buffers to be recovered by aiodoned */
465 	mutex_enter(vp->v_interlock);
466 	while (vp->v_numoutput > 0) {
467 		cv_wait(&vp->v_cv, vp->v_interlock);
468 	}
469 	KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
470 	KASSERT(vp->v_numoutput == 0);
471 	mutex_exit(vp->v_interlock);
472 
473 	fs->lfs_flushvp = NULL;
474 	KASSERT(fs->lfs_flushvp_fakevref == 0);
475 
476 	return (0);
477 }
478 
479 struct lfs_writevnodes_ctx {
480 	int op;
481 	struct lfs *fs;
482 };
483 static bool
484 lfs_writevnodes_selector(void *cl, struct vnode *vp)
485 {
486 	struct lfs_writevnodes_ctx *c = cl;
487 	struct inode *ip = VTOI(vp);
488 	int op = c->op;
489 
490 	if (ip == NULL || vp->v_type == VNON)
491 		return false;
492 	if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) ||
493 	    (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) {
494 		vndebug(vp, "dirop");
495 		return false;
496 	}
497 	if (op == VN_EMPTY && !VPISEMPTY(vp)) {
498 		vndebug(vp,"empty");
499 		return false;;
500 	}
501 	if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM &&
502 	    vp != c->fs->lfs_flushvp && !(ip->i_flag & IN_CLEANING)) {
503 		vndebug(vp,"cleaning");
504 		return false;
505 	}
506 	mutex_enter(&lfs_lock);
507 	if (vp == c->fs->lfs_unlockvp) {
508 		mutex_exit(&lfs_lock);
509 		return false;
510 	}
511 	mutex_exit(&lfs_lock);
512 
513 	return true;
514 }
515 
516 int
517 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
518 {
519 	struct inode *ip;
520 	struct vnode *vp;
521 	struct vnode_iterator *marker;
522 	struct lfs_writevnodes_ctx ctx;
523 	int inodes_written = 0;
524 	int error = 0;
525 
526 	/*
527 	 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist.
528 	 * XXX The rationale is unclear, the initial commit had no information.
529 	 * XXX If the order really matters we have to sort the vnodes first.
530 	*/
531 
532 	ASSERT_SEGLOCK(fs);
533 	vfs_vnode_iterator_init(mp, &marker);
534 	ctx.op = op;
535 	ctx.fs = fs;
536 	while ((vp = vfs_vnode_iterator_next(marker,
537 	    lfs_writevnodes_selector, &ctx)) != NULL) {
538 		ip = VTOI(vp);
539 
540 		/*
541 		 * Write the inode/file if dirty and it's not the IFILE.
542 		 */
543 		if (((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) &&
544 		    ip->i_number != LFS_IFILE_INUM) {
545 			error = lfs_writefile(fs, sp, vp);
546 			if (error) {
547 				vrele(vp);
548 				if (error == EAGAIN) {
549 					/*
550 					 * This error from lfs_putpages
551 					 * indicates we need to drop
552 					 * the segment lock and start
553 					 * over after the cleaner has
554 					 * had a chance to run.
555 					 */
556 					lfs_writeinode(fs, sp, ip);
557 					lfs_writeseg(fs, sp);
558 					if (!VPISEMPTY(vp) &&
559 					    !WRITEINPROG(vp) &&
560 					    !(ip->i_flag & IN_ALLMOD)) {
561 						mutex_enter(&lfs_lock);
562 						LFS_SET_UINO(ip, IN_MODIFIED);
563 						mutex_exit(&lfs_lock);
564 					}
565 					break;
566 				}
567 				error = 0; /* XXX not quite right */
568 				continue;
569 			}
570 
571 			if (!VPISEMPTY(vp)) {
572 				if (WRITEINPROG(vp)) {
573 					ivndebug(vp,"writevnodes/write2");
574 				} else if (!(ip->i_flag & IN_ALLMOD)) {
575 					mutex_enter(&lfs_lock);
576 					LFS_SET_UINO(ip, IN_MODIFIED);
577 					mutex_exit(&lfs_lock);
578 				}
579 			}
580 			(void) lfs_writeinode(fs, sp, ip);
581 			inodes_written++;
582 		}
583 		vrele(vp);
584 	}
585 	vfs_vnode_iterator_destroy(marker);
586 	return error;
587 }
588 
589 /*
590  * Do a checkpoint.
591  */
592 int
593 lfs_segwrite(struct mount *mp, int flags)
594 {
595 	struct buf *bp;
596 	struct inode *ip;
597 	struct lfs *fs;
598 	struct segment *sp;
599 	struct vnode *vp;
600 	SEGUSE *segusep;
601 	int do_ckp, did_ckp, error;
602 	unsigned n, segleft, maxseg, sn, i, curseg;
603 	int writer_set = 0;
604 	int dirty;
605 	int redo;
606 	SEGSUM *ssp;
607 	int um_error;
608 
609 	fs = VFSTOULFS(mp)->um_lfs;
610 	ASSERT_MAYBE_SEGLOCK(fs);
611 
612 	if (fs->lfs_ronly)
613 		return EROFS;
614 
615 	lfs_imtime(fs);
616 
617 	/*
618 	 * Allocate a segment structure and enough space to hold pointers to
619 	 * the maximum possible number of buffers which can be described in a
620 	 * single summary block.
621 	 */
622 	do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags);
623 
624 	/* We can't do a partial write and checkpoint at the same time. */
625 	if (do_ckp)
626 		flags &= ~SEGM_SINGLE;
627 
628 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
629 	sp = fs->lfs_sp;
630 	if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP))
631 		do_ckp = 1;
632 
633 	/*
634 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
635 	 * in which case we have to flush *all* buffers off of this vnode.
636 	 * We don't care about other nodes, but write any non-dirop nodes
637 	 * anyway in anticipation of another getnewvnode().
638 	 *
639 	 * If we're cleaning we only write cleaning and ifile blocks, and
640 	 * no dirops, since otherwise we'd risk corruption in a crash.
641 	 */
642 	if (sp->seg_flags & SEGM_CLEAN)
643 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
644 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
645 		do {
646 			um_error = lfs_writevnodes(fs, mp, sp, VN_REG);
647 			if ((sp->seg_flags & SEGM_SINGLE) &&
648 			    lfs_sb_getcurseg(fs) != fs->lfs_startseg) {
649 				DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs)));
650 				break;
651 			}
652 
653 			if (do_ckp || fs->lfs_dirops == 0) {
654 				if (!writer_set) {
655 					lfs_writer_enter(fs, "lfs writer");
656 					writer_set = 1;
657 				}
658 				error = lfs_writevnodes(fs, mp, sp, VN_DIROP);
659 				if (um_error == 0)
660 					um_error = error;
661 				/* In case writevnodes errored out */
662 				lfs_flush_dirops(fs);
663 				ssp = (SEGSUM *)(sp->segsum);
664 				lfs_ss_setflags(fs, ssp,
665 						lfs_ss_getflags(fs, ssp) & ~(SS_CONT));
666 				lfs_finalize_fs_seguse(fs);
667 			}
668 			if (do_ckp && um_error) {
669 				lfs_segunlock_relock(fs);
670 				sp = fs->lfs_sp;
671 			}
672 		} while (do_ckp && um_error != 0);
673 	}
674 
675 	/*
676 	 * If we are doing a checkpoint, mark everything since the
677 	 * last checkpoint as no longer ACTIVE.
678 	 */
679 	if (do_ckp || fs->lfs_doifile) {
680 		segleft = lfs_sb_getnseg(fs);
681 		curseg = 0;
682 		for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) {
683 			dirty = 0;
684 			if (bread(fs->lfs_ivnode, lfs_sb_getcleansz(fs) + n,
685 			    lfs_sb_getbsize(fs), B_MODIFY, &bp))
686 				panic("lfs_segwrite: ifile read");
687 			segusep = (SEGUSE *)bp->b_data;
688 			maxseg = min(segleft, lfs_sb_getsepb(fs));
689 			for (i = 0; i < maxseg; i++) {
690 				sn = curseg + i;
691 				if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) &&
692 				    segusep->su_flags & SEGUSE_ACTIVE) {
693 					segusep->su_flags &= ~SEGUSE_ACTIVE;
694 					--fs->lfs_nactive;
695 					++dirty;
696 				}
697 				fs->lfs_suflags[fs->lfs_activesb][sn] =
698 					segusep->su_flags;
699 				if (lfs_sb_getversion(fs) > 1)
700 					++segusep;
701 				else
702 					segusep = (SEGUSE *)
703 						((SEGUSE_V1 *)segusep + 1);
704 			}
705 
706 			if (dirty)
707 				error = LFS_BWRITE_LOG(bp); /* Ifile */
708 			else
709 				brelse(bp, 0);
710 			segleft -= lfs_sb_getsepb(fs);
711 			curseg += lfs_sb_getsepb(fs);
712 		}
713 	}
714 
715 	KASSERT(LFS_SEGLOCK_HELD(fs));
716 
717 	did_ckp = 0;
718 	if (do_ckp || fs->lfs_doifile) {
719 		vp = fs->lfs_ivnode;
720 #ifdef DEBUG
721 		int loopcount = 0;
722 #endif
723 		do {
724 #ifdef DEBUG
725 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid);
726 #endif
727 			mutex_enter(&lfs_lock);
728 			fs->lfs_flags &= ~LFS_IFDIRTY;
729 			mutex_exit(&lfs_lock);
730 
731 			ip = VTOI(vp);
732 
733 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
734 				/*
735 				 * Ifile has no pages, so we don't need
736 				 * to check error return here.
737 				 */
738 				lfs_writefile(fs, sp, vp);
739 				/*
740 				 * Ensure the Ifile takes the current segment
741 				 * into account.  See comment in lfs_vflush.
742 				 */
743 				lfs_writefile(fs, sp, vp);
744 				lfs_writefile(fs, sp, vp);
745 			}
746 
747 			if (ip->i_flag & IN_ALLMOD)
748 				++did_ckp;
749 #if 0
750 			redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0);
751 #else
752 			redo = lfs_writeinode(fs, sp, ip);
753 #endif
754 			redo += lfs_writeseg(fs, sp);
755 			mutex_enter(&lfs_lock);
756 			redo += (fs->lfs_flags & LFS_IFDIRTY);
757 			mutex_exit(&lfs_lock);
758 #ifdef DEBUG
759 			if (++loopcount > 2)
760 				log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n",
761 					loopcount);
762 #endif
763 		} while (redo && do_ckp);
764 
765 		/*
766 		 * Unless we are unmounting, the Ifile may continue to have
767 		 * dirty blocks even after a checkpoint, due to changes to
768 		 * inodes' atime.  If we're checkpointing, it's "impossible"
769 		 * for other parts of the Ifile to be dirty after the loop
770 		 * above, since we hold the segment lock.
771 		 */
772 		mutex_enter(vp->v_interlock);
773 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
774 			LFS_CLR_UINO(ip, IN_ALLMOD);
775 		}
776 #ifdef DIAGNOSTIC
777 		else if (do_ckp) {
778 			int do_panic = 0;
779 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
780 				if (bp->b_lblkno < lfs_sb_getcleansz(fs) +
781 				    lfs_sb_getsegtabsz(fs) &&
782 				    !(bp->b_flags & B_GATHERED)) {
783 					printf("ifile lbn %ld still dirty (flags %lx)\n",
784 						(long)bp->b_lblkno,
785 						(long)bp->b_flags);
786 					++do_panic;
787 				}
788 			}
789 			if (do_panic)
790 				panic("dirty blocks");
791 		}
792 #endif
793 		mutex_exit(vp->v_interlock);
794 	} else {
795 		(void) lfs_writeseg(fs, sp);
796 	}
797 
798 	/* Note Ifile no longer needs to be written */
799 	fs->lfs_doifile = 0;
800 	if (writer_set)
801 		lfs_writer_leave(fs);
802 
803 	/*
804 	 * If we didn't write the Ifile, we didn't really do anything.
805 	 * That means that (1) there is a checkpoint on disk and (2)
806 	 * nothing has changed since it was written.
807 	 *
808 	 * Take the flags off of the segment so that lfs_segunlock
809 	 * doesn't have to write the superblock either.
810 	 */
811 	if (do_ckp && !did_ckp) {
812 		sp->seg_flags &= ~SEGM_CKP;
813 	}
814 
815 	if (lfs_dostats) {
816 		++lfs_stats.nwrites;
817 		if (sp->seg_flags & SEGM_SYNC)
818 			++lfs_stats.nsync_writes;
819 		if (sp->seg_flags & SEGM_CKP)
820 			++lfs_stats.ncheckpoints;
821 	}
822 	lfs_segunlock(fs);
823 	return (0);
824 }
825 
826 /*
827  * Write the dirty blocks associated with a vnode.
828  */
829 int
830 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
831 {
832 	struct inode *ip;
833 	int i, frag;
834 	SEGSUM *ssp;
835 	int error;
836 
837 	ASSERT_SEGLOCK(fs);
838 	error = 0;
839 	ip = VTOI(vp);
840 
841 	lfs_acquire_finfo(fs, ip->i_number, ip->i_gen);
842 
843 	if (vp->v_uflag & VU_DIROP) {
844 		ssp = (SEGSUM *)sp->segsum;
845 		lfs_ss_setflags(fs, ssp,
846 				lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT));
847 	}
848 
849 	if (sp->seg_flags & SEGM_CLEAN) {
850 		lfs_gather(fs, sp, vp, lfs_match_fake);
851 		/*
852 		 * For a file being flushed, we need to write *all* blocks.
853 		 * This means writing the cleaning blocks first, and then
854 		 * immediately following with any non-cleaning blocks.
855 		 * The same is true of the Ifile since checkpoints assume
856 		 * that all valid Ifile blocks are written.
857 		 */
858 		if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) {
859 			lfs_gather(fs, sp, vp, lfs_match_data);
860 			/*
861 			 * Don't call VOP_PUTPAGES: if we're flushing,
862 			 * we've already done it, and the Ifile doesn't
863 			 * use the page cache.
864 			 */
865 		}
866 	} else {
867 		lfs_gather(fs, sp, vp, lfs_match_data);
868 		/*
869 		 * If we're flushing, we've already called VOP_PUTPAGES
870 		 * so don't do it again.  Otherwise, we want to write
871 		 * everything we've got.
872 		 */
873 		if (!IS_FLUSHING(fs, vp)) {
874 			mutex_enter(vp->v_interlock);
875 			error = VOP_PUTPAGES(vp, 0, 0,
876 				PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
877 		}
878 	}
879 
880 	/*
881 	 * It may not be necessary to write the meta-data blocks at this point,
882 	 * as the roll-forward recovery code should be able to reconstruct the
883 	 * list.
884 	 *
885 	 * We have to write them anyway, though, under two conditions: (1) the
886 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
887 	 * checkpointing.
888 	 *
889 	 * BUT if we are cleaning, we might have indirect blocks that refer to
890 	 * new blocks not being written yet, in addition to fragments being
891 	 * moved out of a cleaned segment.  If that is the case, don't
892 	 * write the indirect blocks, or the finfo will have a small block
893 	 * in the middle of it!
894 	 * XXX in this case isn't the inode size wrong too?
895 	 */
896 	frag = 0;
897 	if (sp->seg_flags & SEGM_CLEAN) {
898 		for (i = 0; i < ULFS_NDADDR; i++)
899 			if (ip->i_lfs_fragsize[i] > 0 &&
900 			    ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs))
901 				++frag;
902 	}
903 #ifdef DIAGNOSTIC
904 	if (frag > 1)
905 		panic("lfs_writefile: more than one fragment!");
906 #endif
907 	if (IS_FLUSHING(fs, vp) ||
908 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
909 		lfs_gather(fs, sp, vp, lfs_match_indir);
910 		lfs_gather(fs, sp, vp, lfs_match_dindir);
911 		lfs_gather(fs, sp, vp, lfs_match_tindir);
912 	}
913 	lfs_release_finfo(fs);
914 
915 	return error;
916 }
917 
918 /*
919  * Update segment accounting to reflect this inode's change of address.
920  */
921 static int
922 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr)
923 {
924 	struct buf *bp;
925 	daddr_t daddr;
926 	IFILE *ifp;
927 	SEGUSE *sup;
928 	ino_t ino;
929 	int redo_ifile;
930 	u_int32_t sn;
931 
932 	redo_ifile = 0;
933 
934 	/*
935 	 * If updating the ifile, update the super-block.  Update the disk
936 	 * address and access times for this inode in the ifile.
937 	 */
938 	ino = ip->i_number;
939 	if (ino == LFS_IFILE_INUM) {
940 		daddr = lfs_sb_getidaddr(fs);
941 		lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr));
942 	} else {
943 		LFS_IENTRY(ifp, fs, ino, bp);
944 		daddr = lfs_if_getdaddr(fs, ifp);
945 		lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr));
946 		(void)LFS_BWRITE_LOG(bp); /* Ifile */
947 	}
948 
949 	/*
950 	 * If this is the Ifile and lfs_offset is set to the first block
951 	 * in the segment, dirty the new segment's accounting block
952 	 * (XXX should already be dirty?) and tell the caller to do it again.
953 	 */
954 	if (ip->i_number == LFS_IFILE_INUM) {
955 		sn = lfs_dtosn(fs, lfs_sb_getoffset(fs));
956 		if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) ==
957 		    lfs_sb_getoffset(fs)) {
958 			LFS_SEGENTRY(sup, fs, sn, bp);
959 			KASSERT(bp->b_oflags & BO_DELWRI);
960 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
961 			/* fs->lfs_flags |= LFS_IFDIRTY; */
962 			redo_ifile |= 1;
963 		}
964 	}
965 
966 	/*
967 	 * The inode's last address should not be in the current partial
968 	 * segment, except under exceptional circumstances (lfs_writevnodes
969 	 * had to start over, and in the meantime more blocks were written
970 	 * to a vnode).	 Both inodes will be accounted to this segment
971 	 * in lfs_writeseg so we need to subtract the earlier version
972 	 * here anyway.	 The segment count can temporarily dip below
973 	 * zero here; keep track of how many duplicates we have in
974 	 * "dupino" so we don't panic below.
975 	 */
976 	if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) {
977 		++sp->ndupino;
978 		DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg "
979 		      "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
980 		      (long long)daddr, sp->ndupino));
981 	}
982 	/*
983 	 * Account the inode: it no longer belongs to its former segment,
984 	 * though it will not belong to the new segment until that segment
985 	 * is actually written.
986 	 */
987 	if (daddr != LFS_UNUSED_DADDR) {
988 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
989 #ifdef DIAGNOSTIC
990 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
991 #endif
992 		LFS_SEGENTRY(sup, fs, oldsn, bp);
993 #ifdef DIAGNOSTIC
994 		if (sup->su_nbytes + DINOSIZE(fs) * ndupino < DINOSIZE(fs)) {
995 			printf("lfs_writeinode: negative bytes "
996 			       "(segment %" PRIu32 " short by %d, "
997 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
998 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
999 			       "ndupino=%d)\n",
1000 			       lfs_dtosn(fs, daddr),
1001 			       (int)DINOSIZE(fs) *
1002 				   (1 - sp->ndupino) - sup->su_nbytes,
1003 			       oldsn, sp->seg_number, daddr,
1004 			       (unsigned int)sup->su_nbytes,
1005 			       sp->ndupino);
1006 			panic("lfs_writeinode: negative bytes");
1007 			sup->su_nbytes = DINOSIZE(fs);
1008 		}
1009 #endif
1010 		DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n",
1011 		      lfs_dtosn(fs, daddr), DINOSIZE(fs), ino));
1012 		sup->su_nbytes -= DINOSIZE(fs);
1013 		redo_ifile |=
1014 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1015 		if (redo_ifile) {
1016 			mutex_enter(&lfs_lock);
1017 			fs->lfs_flags |= LFS_IFDIRTY;
1018 			mutex_exit(&lfs_lock);
1019 			/* Don't double-account */
1020 			lfs_sb_setidaddr(fs, 0x0);
1021 		}
1022 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1023 	}
1024 
1025 	return redo_ifile;
1026 }
1027 
1028 int
1029 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
1030 {
1031 	struct buf *bp;
1032 	union lfs_dinode *cdp;
1033 	struct vnode *vp = ITOV(ip);
1034 	daddr_t daddr;
1035 	int i, ndx;
1036 	int redo_ifile = 0;
1037 	int gotblk = 0;
1038 	int count;
1039 	SEGSUM *ssp;
1040 
1041 	ASSERT_SEGLOCK(fs);
1042 	if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP))
1043 		return (0);
1044 
1045 	/* Can't write ifile when writer is not set */
1046 	KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 ||
1047 		(sp->seg_flags & SEGM_CLEAN));
1048 
1049 	/*
1050 	 * If this is the Ifile, see if writing it here will generate a
1051 	 * temporary misaccounting.  If it will, do the accounting and write
1052 	 * the blocks, postponing the inode write until the accounting is
1053 	 * solid.
1054 	 */
1055 	count = 0;
1056 	while (vp == fs->lfs_ivnode) {
1057 		int redo = 0;
1058 
1059 		if (sp->idp == NULL && sp->ibp == NULL &&
1060 		    (sp->seg_bytes_left < lfs_sb_getibsize(fs) ||
1061 		     sp->sum_bytes_left < sizeof(int32_t))) {
1062 			(void) lfs_writeseg(fs, sp);
1063 			continue;
1064 		}
1065 
1066 		/* Look for dirty Ifile blocks */
1067 		LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) {
1068 			if (!(bp->b_flags & B_GATHERED)) {
1069 				redo = 1;
1070 				break;
1071 			}
1072 		}
1073 
1074 		if (redo == 0)
1075 			redo = lfs_update_iaddr(fs, sp, ip, 0x0);
1076 		if (redo == 0)
1077 			break;
1078 
1079 		if (sp->idp) {
1080 			lfs_dino_setinumber(fs, sp->idp, 0);
1081 			sp->idp = NULL;
1082 		}
1083 		++count;
1084 		if (count > 2)
1085 			log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count);
1086 		lfs_writefile(fs, sp, fs->lfs_ivnode);
1087 	}
1088 
1089 	/* Allocate a new inode block if necessary. */
1090 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
1091 	    sp->ibp == NULL) {
1092 		/* Allocate a new segment if necessary. */
1093 		if (sp->seg_bytes_left < lfs_sb_getibsize(fs) ||
1094 		    sp->sum_bytes_left < sizeof(int32_t))
1095 			(void) lfs_writeseg(fs, sp);
1096 
1097 		/* Get next inode block. */
1098 		daddr = lfs_sb_getoffset(fs);
1099 		lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs)));
1100 		sp->ibp = *sp->cbpp++ =
1101 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1102 			    LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0);
1103 		gotblk++;
1104 
1105 		/* Zero out inode numbers */
1106 		for (i = 0; i < LFS_INOPB(fs); ++i) {
1107 			union lfs_dinode *tmpdi;
1108 
1109 			tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data +
1110 						     DINOSIZE(fs) * i);
1111 			lfs_dino_setinumber(fs, tmpdi, 0);
1112 		}
1113 
1114 		++sp->start_bpp;
1115 		lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs)));
1116 		/* Set remaining space counters. */
1117 		sp->seg_bytes_left -= lfs_sb_getibsize(fs);
1118 		sp->sum_bytes_left -= sizeof(int32_t);
1119 		ndx = lfs_sb_getsumsize(fs) / sizeof(int32_t) -
1120 			sp->ninodes / LFS_INOPB(fs) - 1;
1121 		/* XXX ondisk32 */
1122 		((int32_t *)(sp->segsum))[ndx] = daddr;
1123 	}
1124 
1125 	/* Check VU_DIROP in case there is a new file with no data blocks */
1126 	if (vp->v_uflag & VU_DIROP) {
1127 		ssp = (SEGSUM *)sp->segsum;
1128 		lfs_ss_setflags(fs, ssp,
1129 				lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT));
1130 	}
1131 
1132 	/* Update the inode times and copy the inode onto the inode page. */
1133 	/* XXX kludge --- don't redirty the ifile just to put times on it */
1134 	if (ip->i_number != LFS_IFILE_INUM)
1135 		LFS_ITIMES(ip, NULL, NULL, NULL);
1136 
1137 	/*
1138 	 * If this is the Ifile, and we've already written the Ifile in this
1139 	 * partial segment, just overwrite it (it's not on disk yet) and
1140 	 * continue.
1141 	 *
1142 	 * XXX we know that the bp that we get the second time around has
1143 	 * already been gathered.
1144 	 */
1145 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
1146 		lfs_copy_dinode(fs, sp->idp, ip->i_din);
1147 		ip->i_lfs_osize = ip->i_size;
1148 		return 0;
1149 	}
1150 
1151 	bp = sp->ibp;
1152 	cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs));
1153 	lfs_copy_dinode(fs, cdp, ip->i_din);
1154 
1155 	/*
1156 	 * This inode is on its way to disk; clear its VU_DIROP status when
1157 	 * the write is complete.
1158 	 */
1159 	if (vp->v_uflag & VU_DIROP) {
1160 		if (!(sp->seg_flags & SEGM_CLEAN))
1161 			ip->i_flag |= IN_CDIROP;
1162 		else {
1163 			DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number));
1164 		}
1165 	}
1166 
1167 	/*
1168 	 * If cleaning, link counts and directory file sizes cannot change,
1169 	 * since those would be directory operations---even if the file
1170 	 * we are writing is marked VU_DIROP we should write the old values.
1171 	 * If we're not cleaning, of course, update the values so we get
1172 	 * current values the next time we clean.
1173 	 */
1174 	if (sp->seg_flags & SEGM_CLEAN) {
1175 		if (vp->v_uflag & VU_DIROP) {
1176 			lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink);
1177 			/* if (vp->v_type == VDIR) */
1178 			lfs_dino_setsize(fs, cdp, ip->i_lfs_osize);
1179 		}
1180 	} else {
1181 		ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp);
1182 		ip->i_lfs_osize = ip->i_size;
1183 	}
1184 
1185 
1186 	/* We can finish the segment accounting for truncations now */
1187 	lfs_finalize_ino_seguse(fs, ip);
1188 
1189 	/*
1190 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
1191 	 * addresses to disk; possibly change the on-disk record of
1192 	 * the inode size, either by reverting to the previous size
1193 	 * (in the case of cleaning) or by verifying the inode's block
1194 	 * holdings (in the case of files being allocated as they are being
1195 	 * written).
1196 	 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail
1197 	 * XXX count on disk wrong by the same amount.	We should be
1198 	 * XXX able to "borrow" from lfs_avail and return it after the
1199 	 * XXX Ifile is written.  See also in lfs_writeseg.
1200 	 */
1201 
1202 	/* Check file size based on highest allocated block */
1203 	if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG ||
1204 	     (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) &&
1205 	    ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) {
1206 		lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs));
1207 		DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %"
1208 		      PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp)));
1209 	}
1210 	if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) {
1211 		DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)"
1212 		      " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks,
1213 		      lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs)));
1214 		for (i=0; i<ULFS_NDADDR; i++) {
1215 			if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) {
1216 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1217 				lfs_dino_setdb(fs, cdp, i, 0);
1218 			}
1219 		}
1220 		for (i=0; i<ULFS_NIADDR; i++) {
1221 			if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) {
1222 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1223 				lfs_dino_setib(fs, cdp, i, 0);
1224 			}
1225 		}
1226 	}
1227 
1228 #ifdef DIAGNOSTIC
1229 	/*
1230 	 * Check dinode held blocks against dinode size.
1231 	 * This should be identical to the check in lfs_vget().
1232 	 */
1233 	for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs);
1234 	     i < ULFS_NDADDR; i++) {
1235 		KASSERT(i >= 0);
1236 		if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK)
1237 			continue;
1238 		if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK ||
1239 		     (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0)
1240 			continue;
1241 		if (lfs_dino_getdb(fs, cdp, i) != 0) {
1242 # ifdef DEBUG
1243 			lfs_dump_dinode(fs, cdp);
1244 # endif
1245 			panic("writing inconsistent inode");
1246 		}
1247 	}
1248 #endif /* DIAGNOSTIC */
1249 
1250 	if (ip->i_flag & IN_CLEANING)
1251 		LFS_CLR_UINO(ip, IN_CLEANING);
1252 	else {
1253 		/* XXX IN_ALLMOD */
1254 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
1255 			     IN_UPDATE | IN_MODIFY);
1256 		if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din))
1257 			LFS_CLR_UINO(ip, IN_MODIFIED);
1258 		else {
1259 			DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real "
1260 			    "blks=%d, eff=%jd\n", ip->i_number,
1261 			    lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks));
1262 		}
1263 	}
1264 
1265 	if (ip->i_number == LFS_IFILE_INUM) {
1266 		/* We know sp->idp == NULL */
1267 		sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs));
1268 
1269 		/* Not dirty any more */
1270 		mutex_enter(&lfs_lock);
1271 		fs->lfs_flags &= ~LFS_IFDIRTY;
1272 		mutex_exit(&lfs_lock);
1273 	}
1274 
1275 	if (gotblk) {
1276 		mutex_enter(&bufcache_lock);
1277 		LFS_LOCK_BUF(bp);
1278 		brelsel(bp, 0);
1279 		mutex_exit(&bufcache_lock);
1280 	}
1281 
1282 	/* Increment inode count in segment summary block. */
1283 
1284 	ssp = (SEGSUM *)sp->segsum;
1285 	lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1);
1286 
1287 	/* If this page is full, set flag to allocate a new page. */
1288 	if (++sp->ninodes % LFS_INOPB(fs) == 0)
1289 		sp->ibp = NULL;
1290 
1291 	redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno);
1292 
1293 	KASSERT(redo_ifile == 0);
1294 	return (redo_ifile);
1295 }
1296 
1297 int
1298 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr)
1299 {
1300 	struct lfs *fs;
1301 	int vers;
1302 	int j, blksinblk;
1303 
1304 	ASSERT_SEGLOCK(sp->fs);
1305 	/*
1306 	 * If full, finish this segment.  We may be doing I/O, so
1307 	 * release and reacquire the splbio().
1308 	 */
1309 #ifdef DIAGNOSTIC
1310 	if (sp->vp == NULL)
1311 		panic ("lfs_gatherblock: Null vp in segment");
1312 #endif
1313 	fs = sp->fs;
1314 	blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs));
1315 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1316 	    sp->seg_bytes_left < bp->b_bcount) {
1317 		if (mptr)
1318 			mutex_exit(mptr);
1319 		lfs_updatemeta(sp);
1320 
1321 		vers = lfs_fi_getversion(fs, sp->fip);
1322 		(void) lfs_writeseg(fs, sp);
1323 
1324 		/* Add the current file to the segment summary. */
1325 		lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers);
1326 
1327 		if (mptr)
1328 			mutex_enter(mptr);
1329 		return (1);
1330 	}
1331 
1332 	if (bp->b_flags & B_GATHERED) {
1333 		DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju,"
1334 		      " lbn %" PRId64 "\n",
1335 		      (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno));
1336 		return (0);
1337 	}
1338 
1339 	/* Insert into the buffer list, update the FINFO block. */
1340 	bp->b_flags |= B_GATHERED;
1341 
1342 	*sp->cbpp++ = bp;
1343 	for (j = 0; j < blksinblk; j++) {
1344 		unsigned bn;
1345 
1346 		bn = lfs_fi_getnblocks(fs, sp->fip);
1347 		lfs_fi_setnblocks(fs, sp->fip, bn+1);
1348 		lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j);
1349 		/* This block's accounting moves from lfs_favail to lfs_avail */
1350 		lfs_deregister_block(sp->vp, bp->b_lblkno + j);
1351 	}
1352 
1353 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1354 	sp->seg_bytes_left -= bp->b_bcount;
1355 	return (0);
1356 }
1357 
1358 int
1359 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1360     int (*match)(struct lfs *, struct buf *))
1361 {
1362 	struct buf *bp, *nbp;
1363 	int count = 0;
1364 
1365 	ASSERT_SEGLOCK(fs);
1366 	if (vp->v_type == VBLK)
1367 		return 0;
1368 	KASSERT(sp->vp == NULL);
1369 	sp->vp = vp;
1370 	mutex_enter(&bufcache_lock);
1371 
1372 #ifndef LFS_NO_BACKBUF_HACK
1373 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1374 # define	BUF_OFFSET	\
1375 	(((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp)
1376 # define	BACK_BUF(BP)	\
1377 	((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1378 # define	BEG_OF_LIST	\
1379 	((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1380 
1381 loop:
1382 	/* Find last buffer. */
1383 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1384 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1385 	     bp = LIST_NEXT(bp, b_vnbufs))
1386 		/* nothing */;
1387 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1388 		nbp = BACK_BUF(bp);
1389 #else /* LFS_NO_BACKBUF_HACK */
1390 loop:
1391 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1392 		nbp = LIST_NEXT(bp, b_vnbufs);
1393 #endif /* LFS_NO_BACKBUF_HACK */
1394 		if ((bp->b_cflags & BC_BUSY) != 0 ||
1395 		    (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) {
1396 #ifdef DEBUG
1397 			if (vp == fs->lfs_ivnode &&
1398 			    (bp->b_cflags & BC_BUSY) != 0 &&
1399 			    (bp->b_flags & B_GATHERED) == 0)
1400 				log(LOG_NOTICE, "lfs_gather: ifile lbn %"
1401 				      PRId64 " busy (%x) at 0x%jx",
1402 				      bp->b_lblkno, bp->b_flags,
1403 				      (uintmax_t)lfs_sb_getoffset(fs));
1404 #endif
1405 			continue;
1406 		}
1407 #ifdef DIAGNOSTIC
1408 # ifdef LFS_USE_B_INVAL
1409 		if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) {
1410 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1411 			      " is BC_INVAL\n", bp->b_lblkno));
1412 			VOP_PRINT(bp->b_vp);
1413 		}
1414 # endif /* LFS_USE_B_INVAL */
1415 		if (!(bp->b_oflags & BO_DELWRI))
1416 			panic("lfs_gather: bp not BO_DELWRI");
1417 		if (!(bp->b_flags & B_LOCKED)) {
1418 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1419 			      " blk %" PRId64 " not B_LOCKED\n",
1420 			      bp->b_lblkno,
1421 			      LFS_DBTOFSB(fs, bp->b_blkno)));
1422 			VOP_PRINT(bp->b_vp);
1423 			panic("lfs_gather: bp not B_LOCKED");
1424 		}
1425 #endif
1426 		if (lfs_gatherblock(sp, bp, &bufcache_lock)) {
1427 			goto loop;
1428 		}
1429 		count++;
1430 	}
1431 	mutex_exit(&bufcache_lock);
1432 	lfs_updatemeta(sp);
1433 	KASSERT(sp->vp == vp);
1434 	sp->vp = NULL;
1435 	return count;
1436 }
1437 
1438 #if DEBUG
1439 # define DEBUG_OOFF(n) do {						\
1440 	if (ooff == 0) {						\
1441 		DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \
1442 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1443 			", was 0x0 (or %" PRId64 ")\n",			\
1444 			(n), ip->i_number, lbn, ndaddr, daddr));	\
1445 	}								\
1446 } while (0)
1447 #else
1448 # define DEBUG_OOFF(n)
1449 #endif
1450 
1451 /*
1452  * Change the given block's address to ndaddr, finding its previous
1453  * location using ulfs_bmaparray().
1454  *
1455  * Account for this change in the segment table.
1456  *
1457  * called with sp == NULL by roll-forwarding code.
1458  */
1459 void
1460 lfs_update_single(struct lfs *fs, struct segment *sp,
1461     struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size)
1462 {
1463 	SEGUSE *sup;
1464 	struct buf *bp;
1465 	struct indir a[ULFS_NIADDR + 2], *ap;
1466 	struct inode *ip;
1467 	daddr_t daddr, ooff;
1468 	int num, error;
1469 	int bb, osize, obb;
1470 
1471 	ASSERT_SEGLOCK(fs);
1472 	KASSERT(sp == NULL || sp->vp == vp);
1473 	ip = VTOI(vp);
1474 
1475 	error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1476 	if (error)
1477 		panic("lfs_updatemeta: ulfs_bmaparray returned %d", error);
1478 
1479 	KASSERT(daddr <= LFS_MAX_DADDR(fs));
1480 	if (daddr > 0)
1481 		daddr = LFS_DBTOFSB(fs, daddr);
1482 
1483 	bb = lfs_numfrags(fs, size);
1484 	switch (num) {
1485 	    case 0:
1486 		    ooff = lfs_dino_getdb(fs, ip->i_din, lbn);
1487 		    DEBUG_OOFF(0);
1488 		    if (ooff == UNWRITTEN)
1489 			    lfs_dino_setblocks(fs, ip->i_din,
1490 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1491 		    else {
1492 			    /* possible fragment truncation or extension */
1493 			    obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]);
1494 			    lfs_dino_setblocks(fs, ip->i_din,
1495 				lfs_dino_getblocks(fs, ip->i_din) + (bb-obb));
1496 		    }
1497 		    lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr);
1498 		    break;
1499 	    case 1:
1500 		    ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off);
1501 		    DEBUG_OOFF(1);
1502 		    if (ooff == UNWRITTEN)
1503 			    lfs_dino_setblocks(fs, ip->i_din,
1504 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1505 		    lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr);
1506 		    break;
1507 	    default:
1508 		    ap = &a[num - 1];
1509 		    if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs),
1510 			B_MODIFY, &bp))
1511 			    panic("lfs_updatemeta: bread bno %" PRId64,
1512 				  ap->in_lbn);
1513 
1514 		    ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off);
1515 		    DEBUG_OOFF(num);
1516 		    if (ooff == UNWRITTEN)
1517 			    lfs_dino_setblocks(fs, ip->i_din,
1518 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1519 		    lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr);
1520 		    (void) VOP_BWRITE(bp->b_vp, bp);
1521 	}
1522 
1523 	KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr);
1524 
1525 	/* Update hiblk when extending the file */
1526 	if (lbn > ip->i_lfs_hiblk)
1527 		ip->i_lfs_hiblk = lbn;
1528 
1529 	/*
1530 	 * Though we'd rather it couldn't, this *can* happen right now
1531 	 * if cleaning blocks and regular blocks coexist.
1532 	 */
1533 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1534 
1535 	/*
1536 	 * Update segment usage information, based on old size
1537 	 * and location.
1538 	 */
1539 	if (daddr > 0) {
1540 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
1541 #ifdef DIAGNOSTIC
1542 		int ndupino;
1543 
1544 		if (sp && sp->seg_number == oldsn) {
1545 			ndupino = sp->ndupino;
1546 		} else {
1547 			ndupino = 0;
1548 		}
1549 #endif
1550 		KASSERT(oldsn < lfs_sb_getnseg(fs));
1551 		if (lbn >= 0 && lbn < ULFS_NDADDR)
1552 			osize = ip->i_lfs_fragsize[lbn];
1553 		else
1554 			osize = lfs_sb_getbsize(fs);
1555 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1556 #ifdef DIAGNOSTIC
1557 		if (sup->su_nbytes + DINOSIZE(fs) * ndupino < osize) {
1558 			printf("lfs_updatemeta: negative bytes "
1559 			       "(segment %" PRIu32 " short by %" PRId64
1560 			       ")\n", lfs_dtosn(fs, daddr),
1561 			       (int64_t)osize -
1562 			       (DINOSIZE(fs) * ndupino + sup->su_nbytes));
1563 			printf("lfs_updatemeta: ino %llu, lbn %" PRId64
1564 			       ", addr = 0x%" PRIx64 "\n",
1565 			       (unsigned long long)ip->i_number, lbn, daddr);
1566 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1567 			panic("lfs_updatemeta: negative bytes");
1568 			sup->su_nbytes = osize -
1569 			    DINOSIZE(fs) * ndupino;
1570 		}
1571 #endif
1572 		DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1573 		      " db 0x%" PRIx64 "\n",
1574 		      lfs_dtosn(fs, daddr), osize,
1575 		      ip->i_number, lbn, daddr));
1576 		sup->su_nbytes -= osize;
1577 		if (!(bp->b_flags & B_GATHERED)) {
1578 			mutex_enter(&lfs_lock);
1579 			fs->lfs_flags |= LFS_IFDIRTY;
1580 			mutex_exit(&lfs_lock);
1581 		}
1582 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1583 	}
1584 	/*
1585 	 * Now that this block has a new address, and its old
1586 	 * segment no longer owns it, we can forget about its
1587 	 * old size.
1588 	 */
1589 	if (lbn >= 0 && lbn < ULFS_NDADDR)
1590 		ip->i_lfs_fragsize[lbn] = size;
1591 }
1592 
1593 /*
1594  * Update the metadata that points to the blocks listed in the FINFO
1595  * array.
1596  */
1597 void
1598 lfs_updatemeta(struct segment *sp)
1599 {
1600 	struct buf *sbp;
1601 	struct lfs *fs;
1602 	struct vnode *vp;
1603 	daddr_t lbn;
1604 	int i, nblocks, num;
1605 	int __diagused nblocks_orig;
1606 	int bb;
1607 	int bytesleft, size;
1608 	unsigned lastlength;
1609 	union lfs_blocks tmpptr;
1610 
1611 	fs = sp->fs;
1612 	vp = sp->vp;
1613 	ASSERT_SEGLOCK(fs);
1614 
1615 	/*
1616 	 * This used to be:
1617 	 *
1618 	 *  nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1619 	 *
1620 	 * that is, it allowed for the possibility that start_lbp did
1621 	 * not point to the beginning of the finfo block pointer area.
1622 	 * This particular formulation is six kinds of painful in the
1623 	 * lfs64 world where we have two sizes of block pointer, so
1624 	 * unless/until everything can be cleaned up to not move
1625 	 * start_lbp around but instead use an offset, we do the
1626 	 * following:
1627 	 *    1. Get NEXT_FINFO(sp->fip). This is the same pointer as
1628 	 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong
1629 	 * type. (Ugh.)
1630 	 *    2. Cast it to void *, then assign it to a temporary
1631 	 * union lfs_blocks.
1632 	 *    3. Subtract start_lbp from that.
1633 	 *    4. Save the value of nblocks in blocks_orig so we can
1634 	 * assert below that it hasn't changed without repeating this
1635 	 * rubbish.
1636 	 *
1637 	 * XXX.
1638 	 */
1639 	lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip));
1640 	nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp);
1641 	nblocks_orig = nblocks;
1642 
1643 	KASSERT(nblocks >= 0);
1644 	KASSERT(vp != NULL);
1645 	if (nblocks == 0)
1646 		return;
1647 
1648 	/*
1649 	 * This count may be high due to oversize blocks from lfs_gop_write.
1650 	 * Correct for this. (XXX we should be able to keep track of these.)
1651 	 */
1652 	for (i = 0; i < nblocks; i++) {
1653 		if (sp->start_bpp[i] == NULL) {
1654 			DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks));
1655 			nblocks = i;
1656 			break;
1657 		}
1658 		num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs));
1659 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1660 		nblocks -= num - 1;
1661 	}
1662 
1663 #if 0
1664 	/* pre-lfs64 assertion */
1665 	KASSERT(vp->v_type == VREG ||
1666 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1667 #else
1668 	KASSERT(vp->v_type == VREG || nblocks == nblocks_orig);
1669 #endif
1670 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1671 
1672 	/*
1673 	 * Sort the blocks.
1674 	 *
1675 	 * We have to sort even if the blocks come from the
1676 	 * cleaner, because there might be other pending blocks on the
1677 	 * same inode...and if we don't sort, and there are fragments
1678 	 * present, blocks may be written in the wrong place.
1679 	 */
1680 	lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs));
1681 
1682 	/*
1683 	 * Record the length of the last block in case it's a fragment.
1684 	 * If there are indirect blocks present, they sort last.  An
1685 	 * indirect block will be lfs_bsize and its presence indicates
1686 	 * that you cannot have fragments.
1687 	 *
1688 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1689 	 * XXX a later indirect block.	This will continue to be
1690 	 * XXX true until lfs_markv is fixed to do everything with
1691 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1692 	 */
1693 	lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1694 		lfs_sb_getbmask(fs)) + 1;
1695 	lfs_fi_setlastlength(fs, sp->fip, lastlength);
1696 
1697 	/*
1698 	 * Assign disk addresses, and update references to the logical
1699 	 * block and the segment usage information.
1700 	 */
1701 	for (i = nblocks; i--; ++sp->start_bpp) {
1702 		sbp = *sp->start_bpp;
1703 		lbn = lfs_blocks_get(fs, &sp->start_lbp, 0);
1704 		KASSERT(sbp->b_lblkno == lbn);
1705 
1706 		sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs));
1707 
1708 		/*
1709 		 * If we write a frag in the wrong place, the cleaner won't
1710 		 * be able to correctly identify its size later, and the
1711 		 * segment will be uncleanable.	 (Even worse, it will assume
1712 		 * that the indirect block that actually ends the list
1713 		 * is of a smaller size!)
1714 		 */
1715 		if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0)
1716 			panic("lfs_updatemeta: fragment is not last block");
1717 
1718 		/*
1719 		 * For each subblock in this possibly oversized block,
1720 		 * update its address on disk.
1721 		 */
1722 		KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs));
1723 		KASSERT(vp == sbp->b_vp);
1724 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1725 		     bytesleft -= lfs_sb_getbsize(fs)) {
1726 			size = MIN(bytesleft, lfs_sb_getbsize(fs));
1727 			bb = lfs_numfrags(fs, size);
1728 			lbn = lfs_blocks_get(fs, &sp->start_lbp, 0);
1729 			lfs_blocks_inc(fs, &sp->start_lbp);
1730 			lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs),
1731 			    size);
1732 			lfs_sb_addoffset(fs, bb);
1733 		}
1734 
1735 	}
1736 
1737 	/* This inode has been modified */
1738 	LFS_SET_UINO(VTOI(vp), IN_MODIFIED);
1739 }
1740 
1741 /*
1742  * Move lfs_offset to a segment earlier than newsn.
1743  */
1744 int
1745 lfs_rewind(struct lfs *fs, int newsn)
1746 {
1747 	int sn, osn, isdirty;
1748 	struct buf *bp;
1749 	SEGUSE *sup;
1750 
1751 	ASSERT_SEGLOCK(fs);
1752 
1753 	osn = lfs_dtosn(fs, lfs_sb_getoffset(fs));
1754 	if (osn < newsn)
1755 		return 0;
1756 
1757 	/* lfs_avail eats the remaining space in this segment */
1758 	lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)));
1759 
1760 	/* Find a low-numbered segment */
1761 	for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) {
1762 		LFS_SEGENTRY(sup, fs, sn, bp);
1763 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1764 		brelse(bp, 0);
1765 
1766 		if (!isdirty)
1767 			break;
1768 	}
1769 	if (sn == lfs_sb_getnseg(fs))
1770 		panic("lfs_rewind: no clean segments");
1771 	if (newsn >= 0 && sn >= newsn)
1772 		return ENOENT;
1773 	lfs_sb_setnextseg(fs, lfs_sntod(fs, sn));
1774 	lfs_newseg(fs);
1775 	lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs));
1776 
1777 	return 0;
1778 }
1779 
1780 /*
1781  * Start a new partial segment.
1782  *
1783  * Return 1 when we entered to a new segment.
1784  * Otherwise, return 0.
1785  */
1786 int
1787 lfs_initseg(struct lfs *fs)
1788 {
1789 	struct segment *sp = fs->lfs_sp;
1790 	SEGSUM *ssp;
1791 	struct buf *sbp;	/* buffer for SEGSUM */
1792 	int repeat = 0;		/* return value */
1793 
1794 	ASSERT_SEGLOCK(fs);
1795 	/* Advance to the next segment. */
1796 	if (!LFS_PARTIAL_FITS(fs)) {
1797 		SEGUSE *sup;
1798 		struct buf *bp;
1799 
1800 		/* lfs_avail eats the remaining space */
1801 		lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) -
1802 						   lfs_sb_getcurseg(fs)));
1803 		/* Wake up any cleaning procs waiting on this file system. */
1804 		lfs_wakeup_cleaner(fs);
1805 		lfs_newseg(fs);
1806 		repeat = 1;
1807 		lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs));
1808 
1809 		sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs));
1810 		sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs));
1811 
1812 		/*
1813 		 * If the segment contains a superblock, update the offset
1814 		 * and summary address to skip over it.
1815 		 */
1816 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1817 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1818 			lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD));
1819 			sp->seg_bytes_left -= LFS_SBPAD;
1820 		}
1821 		brelse(bp, 0);
1822 		/* Segment zero could also contain the labelpad */
1823 		if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 &&
1824 		    lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) {
1825 			lfs_sb_addoffset(fs,
1826 			    lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs));
1827 			sp->seg_bytes_left -=
1828 			    LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs));
1829 		}
1830 	} else {
1831 		sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs));
1832 		sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) -
1833 				      (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)));
1834 	}
1835 	lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs));
1836 
1837 	/* Record first address of this partial segment */
1838 	if (sp->seg_flags & SEGM_CLEAN) {
1839 		fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs);
1840 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1841 			/* "1" is the artificial inc in lfs_seglock */
1842 			mutex_enter(&lfs_lock);
1843 			while (fs->lfs_iocount > 1) {
1844 				mtsleep(&fs->lfs_iocount, PRIBIO + 1,
1845 				    "lfs_initseg", 0, &lfs_lock);
1846 			}
1847 			mutex_exit(&lfs_lock);
1848 			fs->lfs_cleanind = 0;
1849 		}
1850 	}
1851 
1852 	sp->fs = fs;
1853 	sp->ibp = NULL;
1854 	sp->idp = NULL;
1855 	sp->ninodes = 0;
1856 	sp->ndupino = 0;
1857 
1858 	sp->cbpp = sp->bpp;
1859 
1860 	/* Get a new buffer for SEGSUM */
1861 	sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1862 	    LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY);
1863 
1864 	/* ... and enter it into the buffer list. */
1865 	*sp->cbpp = sbp;
1866 	sp->cbpp++;
1867 	lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs)));
1868 
1869 	sp->start_bpp = sp->cbpp;
1870 
1871 	/* Set point to SEGSUM, initialize it. */
1872 	ssp = sp->segsum = sbp->b_data;
1873 	memset(ssp, 0, lfs_sb_getsumsize(fs));
1874 	lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs));
1875 	lfs_ss_setnfinfo(fs, ssp, 0);
1876 	lfs_ss_setninos(fs, ssp, 0);
1877 	lfs_ss_setmagic(fs, ssp, SS_MAGIC);
1878 
1879 	/* Set pointer to first FINFO, initialize it. */
1880 	sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum);
1881 	lfs_fi_setnblocks(fs, sp->fip, 0);
1882 	lfs_fi_setlastlength(fs, sp->fip, 0);
1883 	lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip);
1884 
1885 	sp->seg_bytes_left -= lfs_sb_getsumsize(fs);
1886 	sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs);
1887 
1888 	return (repeat);
1889 }
1890 
1891 /*
1892  * Remove SEGUSE_INVAL from all segments.
1893  */
1894 void
1895 lfs_unset_inval_all(struct lfs *fs)
1896 {
1897 	SEGUSE *sup;
1898 	struct buf *bp;
1899 	int i;
1900 
1901 	for (i = 0; i < lfs_sb_getnseg(fs); i++) {
1902 		LFS_SEGENTRY(sup, fs, i, bp);
1903 		if (sup->su_flags & SEGUSE_INVAL) {
1904 			sup->su_flags &= ~SEGUSE_INVAL;
1905 			LFS_WRITESEGENTRY(sup, fs, i, bp);
1906 		} else
1907 			brelse(bp, 0);
1908 	}
1909 }
1910 
1911 /*
1912  * Return the next segment to write.
1913  */
1914 void
1915 lfs_newseg(struct lfs *fs)
1916 {
1917 	CLEANERINFO *cip;
1918 	SEGUSE *sup;
1919 	struct buf *bp;
1920 	int curseg, isdirty, sn, skip_inval;
1921 
1922 	ASSERT_SEGLOCK(fs);
1923 
1924 	/* Honor LFCNWRAPSTOP */
1925 	mutex_enter(&lfs_lock);
1926 	while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) {
1927 		if (fs->lfs_wrappass) {
1928 			log(LOG_NOTICE, "%s: wrappass=%d\n",
1929 				lfs_sb_getfsmnt(fs), fs->lfs_wrappass);
1930 			fs->lfs_wrappass = 0;
1931 			break;
1932 		}
1933 		fs->lfs_wrapstatus = LFS_WRAP_WAITING;
1934 		wakeup(&fs->lfs_nowrap);
1935 		log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs));
1936 		mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz,
1937 			&lfs_lock);
1938 	}
1939 	fs->lfs_wrapstatus = LFS_WRAP_GOING;
1940 	mutex_exit(&lfs_lock);
1941 
1942 	LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp);
1943 	DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n",
1944 	      lfs_dtosn(fs, lfs_sb_getnextseg(fs))));
1945 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1946 	sup->su_nbytes = 0;
1947 	sup->su_nsums = 0;
1948 	sup->su_ninos = 0;
1949 	LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp);
1950 
1951 	LFS_CLEANERINFO(cip, fs, bp);
1952 	lfs_ci_shiftcleantodirty(fs, cip, 1);
1953 	lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip));
1954 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1955 
1956 	lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs));
1957 	lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs));
1958 	skip_inval = 1;
1959 	for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) {
1960 		sn = (sn + 1) % lfs_sb_getnseg(fs);
1961 
1962 		if (sn == curseg) {
1963 			if (skip_inval)
1964 				skip_inval = 0;
1965 			else
1966 				panic("lfs_nextseg: no clean segments");
1967 		}
1968 		LFS_SEGENTRY(sup, fs, sn, bp);
1969 		isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0));
1970 		/* Check SEGUSE_EMPTY as we go along */
1971 		if (isdirty && sup->su_nbytes == 0 &&
1972 		    !(sup->su_flags & SEGUSE_EMPTY))
1973 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1974 		else
1975 			brelse(bp, 0);
1976 
1977 		if (!isdirty)
1978 			break;
1979 	}
1980 	if (skip_inval == 0)
1981 		lfs_unset_inval_all(fs);
1982 
1983 	++fs->lfs_nactive;
1984 	lfs_sb_setnextseg(fs, lfs_sntod(fs, sn));
1985 	if (lfs_dostats) {
1986 		++lfs_stats.segsused;
1987 	}
1988 }
1989 
1990 static struct buf *
1991 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr,
1992     int n)
1993 {
1994 	struct lfs_cluster *cl;
1995 	struct buf **bpp, *bp;
1996 
1997 	ASSERT_SEGLOCK(fs);
1998 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1999 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
2000 	memset(cl, 0, sizeof(*cl));
2001 	cl->fs = fs;
2002 	cl->bpp = bpp;
2003 	cl->bufcount = 0;
2004 	cl->bufsize = 0;
2005 
2006 	/* If this segment is being written synchronously, note that */
2007 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
2008 		cl->flags |= LFS_CL_SYNC;
2009 		cl->seg = fs->lfs_sp;
2010 		++cl->seg->seg_iocount;
2011 	}
2012 
2013 	/* Get an empty buffer header, or maybe one with something on it */
2014 	bp = getiobuf(vp, true);
2015 	bp->b_dev = NODEV;
2016 	bp->b_blkno = bp->b_lblkno = addr;
2017 	bp->b_iodone = lfs_cluster_callback;
2018 	bp->b_private = cl;
2019 
2020 	return bp;
2021 }
2022 
2023 int
2024 lfs_writeseg(struct lfs *fs, struct segment *sp)
2025 {
2026 	struct buf **bpp, *bp, *cbp, *newbp, *unbusybp;
2027 	SEGUSE *sup;
2028 	SEGSUM *ssp;
2029 	int i;
2030 	int do_again, nblocks, byteoffset;
2031 	size_t el_size;
2032 	struct lfs_cluster *cl;
2033 	u_short ninos;
2034 	struct vnode *devvp;
2035 	char *p = NULL;
2036 	struct vnode *vp;
2037 	int32_t *daddrp;	/* XXX ondisk32 */
2038 	int changed;
2039 	u_int32_t sum;
2040 	size_t sumstart;
2041 #ifdef DEBUG
2042 	FINFO *fip;
2043 	int findex;
2044 #endif
2045 
2046 	ASSERT_SEGLOCK(fs);
2047 
2048 	ssp = (SEGSUM *)sp->segsum;
2049 
2050 	/*
2051 	 * If there are no buffers other than the segment summary to write,
2052 	 * don't do anything.  If we are the end of a dirop sequence, however,
2053 	 * write the empty segment summary anyway, to help out the
2054 	 * roll-forward agent.
2055 	 */
2056 	if ((nblocks = sp->cbpp - sp->bpp) == 1) {
2057 		if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP)
2058 			return 0;
2059 	}
2060 
2061 	/* Note if partial segment is being written by the cleaner */
2062 	if (sp->seg_flags & SEGM_CLEAN)
2063 		lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN);
2064 
2065 	/* Note if we are writing to reclaim */
2066 	if (sp->seg_flags & SEGM_RECLAIM) {
2067 		lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM);
2068 		lfs_ss_setreclino(fs, ssp, fs->lfs_reclino);
2069 	}
2070 
2071 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2072 
2073 	/* Update the segment usage information. */
2074 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
2075 
2076 	/* Loop through all blocks, except the segment summary. */
2077 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
2078 		if ((*bpp)->b_vp != devvp) {
2079 			sup->su_nbytes += (*bpp)->b_bcount;
2080 			DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d"
2081 			      " lbn %" PRId64 " db 0x%" PRIx64 "\n",
2082 			      sp->seg_number, (*bpp)->b_bcount,
2083 			      VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
2084 			      (*bpp)->b_blkno));
2085 		}
2086 	}
2087 
2088 #ifdef DEBUG
2089 	/* Check for zero-length and zero-version FINFO entries. */
2090 	fip = SEGSUM_FINFOBASE(fs, ssp);
2091 	for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) {
2092 		KDASSERT(lfs_fi_getnblocks(fs, fip) > 0);
2093 		KDASSERT(lfs_fi_getversion(fs, fip) > 0);
2094 		fip = NEXT_FINFO(fs, fip);
2095 	}
2096 #endif /* DEBUG */
2097 
2098 	ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs);
2099 	DLOG((DLOG_SU, "seg %d += %d for %d inodes\n",
2100 	      sp->seg_number,
2101 	      lfs_ss_getninos(fs, ssp) * DINOSIZE(fs),
2102 	      lfs_ss_getninos(fs, ssp)));
2103 	sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs);
2104 	/* sup->su_nbytes += lfs_sb_getsumsize(fs); */
2105 	if (lfs_sb_getversion(fs) == 1)
2106 		sup->su_olastmod = time_second;
2107 	else
2108 		sup->su_lastmod = time_second;
2109 	sup->su_ninos += ninos;
2110 	++sup->su_nsums;
2111 	lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs)));
2112 
2113 	do_again = !(bp->b_flags & B_GATHERED);
2114 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
2115 
2116 	/*
2117 	 * Mark blocks B_BUSY, to prevent then from being changed between
2118 	 * the checksum computation and the actual write.
2119 	 *
2120 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
2121 	 * there are any, replace them with copies that have UNASSIGNED
2122 	 * instead.
2123 	 */
2124 	mutex_enter(&bufcache_lock);
2125 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
2126 		++bpp;
2127 		bp = *bpp;
2128 		if (bp->b_iodone != NULL) {	 /* UBC or malloced buffer */
2129 			bp->b_cflags |= BC_BUSY;
2130 			continue;
2131 		}
2132 
2133 		while (bp->b_cflags & BC_BUSY) {
2134 			DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential"
2135 			      " data summary corruption for ino %d, lbn %"
2136 			      PRId64 "\n",
2137 			      VTOI(bp->b_vp)->i_number, bp->b_lblkno));
2138 			bp->b_cflags |= BC_WANTED;
2139 			cv_wait(&bp->b_busy, &bufcache_lock);
2140 		}
2141 		bp->b_cflags |= BC_BUSY;
2142 		mutex_exit(&bufcache_lock);
2143 		unbusybp = NULL;
2144 
2145 		/*
2146 		 * Check and replace indirect block UNWRITTEN bogosity.
2147 		 * XXX See comment in lfs_writefile.
2148 		 */
2149 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
2150 		   lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) !=
2151 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
2152 			DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n",
2153 			      VTOI(bp->b_vp)->i_number,
2154 			      (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks,
2155 			      lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din)));
2156 			/* Make a copy we'll make changes to */
2157 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
2158 					   bp->b_bcount, LFS_NB_IBLOCK);
2159 			newbp->b_blkno = bp->b_blkno;
2160 			memcpy(newbp->b_data, bp->b_data,
2161 			       newbp->b_bcount);
2162 
2163 			changed = 0;
2164 			/* XXX ondisk32 */
2165 			for (daddrp = (int32_t *)(newbp->b_data);
2166 			     daddrp < (int32_t *)((char *)newbp->b_data +
2167 						  newbp->b_bcount); daddrp++) {
2168 				if (*daddrp == UNWRITTEN) {
2169 					++changed;
2170 					*daddrp = 0;
2171 				}
2172 			}
2173 			/*
2174 			 * Get rid of the old buffer.  Don't mark it clean,
2175 			 * though, if it still has dirty data on it.
2176 			 */
2177 			if (changed) {
2178 				DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):"
2179 				      " bp = %p newbp = %p\n", changed, bp,
2180 				      newbp));
2181 				*bpp = newbp;
2182 				bp->b_flags &= ~B_GATHERED;
2183 				bp->b_error = 0;
2184 				if (bp->b_iodone != NULL) {
2185 					DLOG((DLOG_SEG, "lfs_writeseg: "
2186 					      "indir bp should not be B_CALL\n"));
2187 					biodone(bp);
2188 					bp = NULL;
2189 				} else {
2190 					/* Still on free list, leave it there */
2191 					unbusybp = bp;
2192 					/*
2193 					 * We have to re-decrement lfs_avail
2194 					 * since this block is going to come
2195 					 * back around to us in the next
2196 					 * segment.
2197 					 */
2198 					lfs_sb_subavail(fs,
2199 					    lfs_btofsb(fs, bp->b_bcount));
2200 				}
2201 			} else {
2202 				lfs_freebuf(fs, newbp);
2203 			}
2204 		}
2205 		mutex_enter(&bufcache_lock);
2206 		if (unbusybp != NULL) {
2207 			unbusybp->b_cflags &= ~BC_BUSY;
2208 			if (unbusybp->b_cflags & BC_WANTED)
2209 				cv_broadcast(&bp->b_busy);
2210 		}
2211 	}
2212 	mutex_exit(&bufcache_lock);
2213 
2214 	/*
2215 	 * Compute checksum across data and then across summary; the first
2216 	 * block (the summary block) is skipped.  Set the create time here
2217 	 * so that it's guaranteed to be later than the inode mod times.
2218 	 */
2219 	sum = 0;
2220 	if (lfs_sb_getversion(fs) == 1)
2221 		el_size = sizeof(u_long);
2222 	else
2223 		el_size = sizeof(u_int32_t);
2224 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
2225 		++bpp;
2226 		/* Loop through gop_write cluster blocks */
2227 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
2228 		     byteoffset += lfs_sb_getbsize(fs)) {
2229 #ifdef LFS_USE_B_INVAL
2230 			if (((*bpp)->b_cflags & BC_INVAL) != 0 &&
2231 			    (*bpp)->b_iodone != NULL) {
2232 				if (copyin((void *)(*bpp)->b_saveaddr +
2233 					   byteoffset, dp, el_size)) {
2234 					panic("lfs_writeseg: copyin failed [1]:"
2235 						" ino %d blk %" PRId64,
2236 						VTOI((*bpp)->b_vp)->i_number,
2237 						(*bpp)->b_lblkno);
2238 				}
2239 			} else
2240 #endif /* LFS_USE_B_INVAL */
2241 			{
2242 				sum = lfs_cksum_part((char *)
2243 				    (*bpp)->b_data + byteoffset, el_size, sum);
2244 			}
2245 		}
2246 	}
2247 	if (lfs_sb_getversion(fs) == 1)
2248 		lfs_ss_setocreate(fs, ssp, time_second);
2249 	else {
2250 		lfs_ss_setcreate(fs, ssp, time_second);
2251 		lfs_sb_addserial(fs, 1);
2252 		lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs));
2253 		lfs_ss_setident(fs, ssp, lfs_sb_getident(fs));
2254 	}
2255 	lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum));
2256 	sumstart = lfs_ss_getsumstart(fs);
2257 	lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart,
2258 	    lfs_sb_getsumsize(fs) - sumstart));
2259 
2260 	mutex_enter(&lfs_lock);
2261 	lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) +
2262 			  lfs_btofsb(fs, lfs_sb_getsumsize(fs))));
2263 	lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) +
2264 			  lfs_btofsb(fs, lfs_sb_getsumsize(fs))));
2265 	mutex_exit(&lfs_lock);
2266 
2267 	/*
2268 	 * When we simply write the blocks we lose a rotation for every block
2269 	 * written.  To avoid this problem, we cluster the buffers into a
2270 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
2271 	 * devices can handle, use that for the size of the chunks.
2272 	 *
2273 	 * Blocks that are already clusters (from GOP_WRITE), however, we
2274 	 * don't bother to copy into other clusters.
2275 	 */
2276 
2277 #define CHUNKSIZE MAXPHYS
2278 
2279 	if (devvp == NULL)
2280 		panic("devvp is NULL");
2281 	for (bpp = sp->bpp, i = nblocks; i;) {
2282 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
2283 		cl = cbp->b_private;
2284 
2285 		cbp->b_flags |= B_ASYNC;
2286 		cbp->b_cflags |= BC_BUSY;
2287 		cbp->b_bcount = 0;
2288 
2289 #if defined(DEBUG) && defined(DIAGNOSTIC)
2290 		if (bpp - sp->bpp > (lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs))
2291 		    / sizeof(int32_t)) {
2292 			panic("lfs_writeseg: real bpp overwrite");
2293 		}
2294 		if (bpp - sp->bpp > lfs_segsize(fs) / lfs_sb_getfsize(fs)) {
2295 			panic("lfs_writeseg: theoretical bpp overwrite");
2296 		}
2297 #endif
2298 
2299 		/*
2300 		 * Construct the cluster.
2301 		 */
2302 		mutex_enter(&lfs_lock);
2303 		++fs->lfs_iocount;
2304 		mutex_exit(&lfs_lock);
2305 		while (i && cbp->b_bcount < CHUNKSIZE) {
2306 			bp = *bpp;
2307 
2308 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
2309 				break;
2310 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
2311 				break;
2312 
2313 			/* Clusters from GOP_WRITE are expedited */
2314 			if (bp->b_bcount > lfs_sb_getbsize(fs)) {
2315 				if (cbp->b_bcount > 0)
2316 					/* Put in its own buffer */
2317 					break;
2318 				else {
2319 					cbp->b_data = bp->b_data;
2320 				}
2321 			} else if (cbp->b_bcount == 0) {
2322 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
2323 							     LFS_NB_CLUSTER);
2324 				cl->flags |= LFS_CL_MALLOC;
2325 			}
2326 #ifdef DIAGNOSTIC
2327 			if (lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno +
2328 					      btodb(bp->b_bcount - 1))) !=
2329 			    sp->seg_number) {
2330 				printf("blk size %d daddr %" PRIx64
2331 				    " not in seg %d\n",
2332 				    bp->b_bcount, bp->b_blkno,
2333 				    sp->seg_number);
2334 				panic("segment overwrite");
2335 			}
2336 #endif
2337 
2338 #ifdef LFS_USE_B_INVAL
2339 			/*
2340 			 * Fake buffers from the cleaner are marked as B_INVAL.
2341 			 * We need to copy the data from user space rather than
2342 			 * from the buffer indicated.
2343 			 * XXX == what do I do on an error?
2344 			 */
2345 			if ((bp->b_cflags & BC_INVAL) != 0 &&
2346 			    bp->b_iodone != NULL) {
2347 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
2348 					panic("lfs_writeseg: "
2349 					    "copyin failed [2]");
2350 			} else
2351 #endif /* LFS_USE_B_INVAL */
2352 			if (cl->flags & LFS_CL_MALLOC) {
2353 				/* copy data into our cluster. */
2354 				memcpy(p, bp->b_data, bp->b_bcount);
2355 				p += bp->b_bcount;
2356 			}
2357 
2358 			cbp->b_bcount += bp->b_bcount;
2359 			cl->bufsize += bp->b_bcount;
2360 
2361 			bp->b_flags &= ~B_READ;
2362 			bp->b_error = 0;
2363 			cl->bpp[cl->bufcount++] = bp;
2364 
2365 			vp = bp->b_vp;
2366 			mutex_enter(&bufcache_lock);
2367 			mutex_enter(vp->v_interlock);
2368 			bp->b_oflags &= ~(BO_DELWRI | BO_DONE);
2369 			reassignbuf(bp, vp);
2370 			vp->v_numoutput++;
2371 			mutex_exit(vp->v_interlock);
2372 			mutex_exit(&bufcache_lock);
2373 
2374 			bpp++;
2375 			i--;
2376 		}
2377 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2378 			BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL);
2379 		else
2380 			BIO_SETPRIO(cbp, BPRIO_TIMELIMITED);
2381 		mutex_enter(devvp->v_interlock);
2382 		devvp->v_numoutput++;
2383 		mutex_exit(devvp->v_interlock);
2384 		VOP_STRATEGY(devvp, cbp);
2385 		curlwp->l_ru.ru_oublock++;
2386 	}
2387 
2388 	if (lfs_dostats) {
2389 		++lfs_stats.psegwrites;
2390 		lfs_stats.blocktot += nblocks - 1;
2391 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2392 			++lfs_stats.psyncwrites;
2393 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2394 			++lfs_stats.pcleanwrites;
2395 			lfs_stats.cleanblocks += nblocks - 1;
2396 		}
2397 	}
2398 
2399 	return (lfs_initseg(fs) || do_again);
2400 }
2401 
2402 void
2403 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2404 {
2405 	struct buf *bp;
2406 	struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2407 	int s;
2408 
2409 	ASSERT_MAYBE_SEGLOCK(fs);
2410 #ifdef DIAGNOSTIC
2411 	if (fs->lfs_is64) {
2412 		KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC);
2413 	} else {
2414 		KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC);
2415 	}
2416 #endif
2417 	/*
2418 	 * If we can write one superblock while another is in
2419 	 * progress, we risk not having a complete checkpoint if we crash.
2420 	 * So, block here if a superblock write is in progress.
2421 	 */
2422 	mutex_enter(&lfs_lock);
2423 	s = splbio();
2424 	while (fs->lfs_sbactive) {
2425 		mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0,
2426 			&lfs_lock);
2427 	}
2428 	fs->lfs_sbactive = daddr;
2429 	splx(s);
2430 	mutex_exit(&lfs_lock);
2431 
2432 	/* Set timestamp of this version of the superblock */
2433 	if (lfs_sb_getversion(fs) == 1)
2434 		lfs_sb_setotstamp(fs, time_second);
2435 	lfs_sb_settstamp(fs, time_second);
2436 
2437 	/* The next chunk of code relies on this assumption */
2438 	CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64));
2439 
2440 	/* Checksum the superblock and copy it into a buffer. */
2441 	lfs_sb_setcksum(fs, lfs_sb_cksum(fs));
2442 	bp = lfs_newbuf(fs, devvp,
2443 	    LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2444 	memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs));
2445 	memset((char *)bp->b_data + sizeof(struct dlfs), 0,
2446 	    LFS_SBPAD - sizeof(struct dlfs));
2447 
2448 	bp->b_cflags |= BC_BUSY;
2449 	bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC;
2450 	bp->b_oflags &= ~(BO_DONE | BO_DELWRI);
2451 	bp->b_error = 0;
2452 	bp->b_iodone = lfs_supercallback;
2453 
2454 	if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC)
2455 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
2456 	else
2457 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
2458 	curlwp->l_ru.ru_oublock++;
2459 
2460 	mutex_enter(devvp->v_interlock);
2461 	devvp->v_numoutput++;
2462 	mutex_exit(devvp->v_interlock);
2463 
2464 	mutex_enter(&lfs_lock);
2465 	++fs->lfs_iocount;
2466 	mutex_exit(&lfs_lock);
2467 	VOP_STRATEGY(devvp, bp);
2468 }
2469 
2470 /*
2471  * Logical block number match routines used when traversing the dirty block
2472  * chain.
2473  */
2474 int
2475 lfs_match_fake(struct lfs *fs, struct buf *bp)
2476 {
2477 
2478 	ASSERT_SEGLOCK(fs);
2479 	return LFS_IS_MALLOC_BUF(bp);
2480 }
2481 
2482 #if 0
2483 int
2484 lfs_match_real(struct lfs *fs, struct buf *bp)
2485 {
2486 
2487 	ASSERT_SEGLOCK(fs);
2488 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2489 }
2490 #endif
2491 
2492 int
2493 lfs_match_data(struct lfs *fs, struct buf *bp)
2494 {
2495 
2496 	ASSERT_SEGLOCK(fs);
2497 	return (bp->b_lblkno >= 0);
2498 }
2499 
2500 int
2501 lfs_match_indir(struct lfs *fs, struct buf *bp)
2502 {
2503 	daddr_t lbn;
2504 
2505 	ASSERT_SEGLOCK(fs);
2506 	lbn = bp->b_lblkno;
2507 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0);
2508 }
2509 
2510 int
2511 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2512 {
2513 	daddr_t lbn;
2514 
2515 	ASSERT_SEGLOCK(fs);
2516 	lbn = bp->b_lblkno;
2517 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1);
2518 }
2519 
2520 int
2521 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2522 {
2523 	daddr_t lbn;
2524 
2525 	ASSERT_SEGLOCK(fs);
2526 	lbn = bp->b_lblkno;
2527 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2);
2528 }
2529 
2530 static void
2531 lfs_free_aiodone(struct buf *bp)
2532 {
2533 	struct lfs *fs;
2534 
2535 	KERNEL_LOCK(1, curlwp);
2536 	fs = bp->b_private;
2537 	ASSERT_NO_SEGLOCK(fs);
2538 	lfs_freebuf(fs, bp);
2539 	KERNEL_UNLOCK_LAST(curlwp);
2540 }
2541 
2542 static void
2543 lfs_super_aiodone(struct buf *bp)
2544 {
2545 	struct lfs *fs;
2546 
2547 	KERNEL_LOCK(1, curlwp);
2548 	fs = bp->b_private;
2549 	ASSERT_NO_SEGLOCK(fs);
2550 	mutex_enter(&lfs_lock);
2551 	fs->lfs_sbactive = 0;
2552 	if (--fs->lfs_iocount <= 1)
2553 		wakeup(&fs->lfs_iocount);
2554 	wakeup(&fs->lfs_sbactive);
2555 	mutex_exit(&lfs_lock);
2556 	lfs_freebuf(fs, bp);
2557 	KERNEL_UNLOCK_LAST(curlwp);
2558 }
2559 
2560 static void
2561 lfs_cluster_aiodone(struct buf *bp)
2562 {
2563 	struct lfs_cluster *cl;
2564 	struct lfs *fs;
2565 	struct buf *tbp, *fbp;
2566 	struct vnode *vp, *devvp, *ovp;
2567 	struct inode *ip;
2568 	int error;
2569 
2570 	KERNEL_LOCK(1, curlwp);
2571 
2572 	error = bp->b_error;
2573 	cl = bp->b_private;
2574 	fs = cl->fs;
2575 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2576 	ASSERT_NO_SEGLOCK(fs);
2577 
2578 	/* Put the pages back, and release the buffer */
2579 	while (cl->bufcount--) {
2580 		tbp = cl->bpp[cl->bufcount];
2581 		KASSERT(tbp->b_cflags & BC_BUSY);
2582 		if (error) {
2583 			tbp->b_error = error;
2584 		}
2585 
2586 		/*
2587 		 * We're done with tbp.	 If it has not been re-dirtied since
2588 		 * the cluster was written, free it.  Otherwise, keep it on
2589 		 * the locked list to be written again.
2590 		 */
2591 		vp = tbp->b_vp;
2592 
2593 		tbp->b_flags &= ~B_GATHERED;
2594 
2595 		LFS_BCLEAN_LOG(fs, tbp);
2596 
2597 		mutex_enter(&bufcache_lock);
2598 		if (tbp->b_iodone == NULL) {
2599 			KASSERT(tbp->b_flags & B_LOCKED);
2600 			bremfree(tbp);
2601 			if (vp) {
2602 				mutex_enter(vp->v_interlock);
2603 				reassignbuf(tbp, vp);
2604 				mutex_exit(vp->v_interlock);
2605 			}
2606 			tbp->b_flags |= B_ASYNC; /* for biodone */
2607 		}
2608 
2609 		if (((tbp->b_flags | tbp->b_oflags) &
2610 		    (B_LOCKED | BO_DELWRI)) == B_LOCKED)
2611 			LFS_UNLOCK_BUF(tbp);
2612 
2613 		if (tbp->b_oflags & BO_DONE) {
2614 			DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n",
2615 				cl->bufcount, (long)tbp->b_flags));
2616 		}
2617 
2618 		if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) {
2619 			/*
2620 			 * A buffer from the page daemon.
2621 			 * We use the same iodone as it does,
2622 			 * so we must manually disassociate its
2623 			 * buffers from the vp.
2624 			 */
2625 			if ((ovp = tbp->b_vp) != NULL) {
2626 				/* This is just silly */
2627 				mutex_enter(ovp->v_interlock);
2628 				brelvp(tbp);
2629 				mutex_exit(ovp->v_interlock);
2630 				tbp->b_vp = vp;
2631 				tbp->b_objlock = vp->v_interlock;
2632 			}
2633 			/* Put it back the way it was */
2634 			tbp->b_flags |= B_ASYNC;
2635 			/* Master buffers have BC_AGE */
2636 			if (tbp->b_private == tbp)
2637 				tbp->b_cflags |= BC_AGE;
2638 		}
2639 		mutex_exit(&bufcache_lock);
2640 
2641 		biodone(tbp);
2642 
2643 		/*
2644 		 * If this is the last block for this vnode, but
2645 		 * there are other blocks on its dirty list,
2646 		 * set IN_MODIFIED/IN_CLEANING depending on what
2647 		 * sort of block.  Only do this for our mount point,
2648 		 * not for, e.g., inode blocks that are attached to
2649 		 * the devvp.
2650 		 * XXX KS - Shouldn't we set *both* if both types
2651 		 * of blocks are present (traverse the dirty list?)
2652 		 */
2653 		mutex_enter(vp->v_interlock);
2654 		mutex_enter(&lfs_lock);
2655 		if (vp != devvp && vp->v_numoutput == 0 &&
2656 		    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2657 			ip = VTOI(vp);
2658 			DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n",
2659 			       ip->i_number));
2660 			if (LFS_IS_MALLOC_BUF(fbp))
2661 				LFS_SET_UINO(ip, IN_CLEANING);
2662 			else
2663 				LFS_SET_UINO(ip, IN_MODIFIED);
2664 		}
2665 		cv_broadcast(&vp->v_cv);
2666 		mutex_exit(&lfs_lock);
2667 		mutex_exit(vp->v_interlock);
2668 	}
2669 
2670 	/* Fix up the cluster buffer, and release it */
2671 	if (cl->flags & LFS_CL_MALLOC)
2672 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2673 	putiobuf(bp);
2674 
2675 	/* Note i/o done */
2676 	if (cl->flags & LFS_CL_SYNC) {
2677 		if (--cl->seg->seg_iocount == 0)
2678 			wakeup(&cl->seg->seg_iocount);
2679 	}
2680 	mutex_enter(&lfs_lock);
2681 #ifdef DIAGNOSTIC
2682 	if (fs->lfs_iocount == 0)
2683 		panic("lfs_cluster_aiodone: zero iocount");
2684 #endif
2685 	if (--fs->lfs_iocount <= 1)
2686 		wakeup(&fs->lfs_iocount);
2687 	mutex_exit(&lfs_lock);
2688 
2689 	KERNEL_UNLOCK_LAST(curlwp);
2690 
2691 	pool_put(&fs->lfs_bpppool, cl->bpp);
2692 	cl->bpp = NULL;
2693 	pool_put(&fs->lfs_clpool, cl);
2694 }
2695 
2696 static void
2697 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2698 {
2699 	/* reset b_iodone for when this is a single-buf i/o. */
2700 	bp->b_iodone = aiodone;
2701 
2702 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
2703 }
2704 
2705 static void
2706 lfs_cluster_callback(struct buf *bp)
2707 {
2708 
2709 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2710 }
2711 
2712 void
2713 lfs_supercallback(struct buf *bp)
2714 {
2715 
2716 	lfs_generic_callback(bp, lfs_super_aiodone);
2717 }
2718 
2719 /*
2720  * The only buffers that are going to hit these functions are the
2721  * segment write blocks, or the segment summaries, or the superblocks.
2722  *
2723  * All of the above are created by lfs_newbuf, and so do not need to be
2724  * released via brelse.
2725  */
2726 void
2727 lfs_callback(struct buf *bp)
2728 {
2729 
2730 	lfs_generic_callback(bp, lfs_free_aiodone);
2731 }
2732 
2733 /*
2734  * Shellsort (diminishing increment sort) from Data Structures and
2735  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2736  * see also Knuth Vol. 3, page 84.  The increments are selected from
2737  * formula (8), page 95.  Roughly O(N^3/2).
2738  */
2739 /*
2740  * This is our own private copy of shellsort because we want to sort
2741  * two parallel arrays (the array of buffer pointers and the array of
2742  * logical block numbers) simultaneously.  Note that we cast the array
2743  * of logical block numbers to a unsigned in this routine so that the
2744  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2745  */
2746 
2747 static void
2748 lfs_shellsort(struct lfs *fs,
2749 	      struct buf **bp_array, union lfs_blocks *lb_array,
2750 	      int nmemb, int size)
2751 {
2752 	static int __rsshell_increments[] = { 4, 1, 0 };
2753 	int incr, *incrp, t1, t2;
2754 	struct buf *bp_temp;
2755 
2756 #ifdef DEBUG
2757 	incr = 0;
2758 	for (t1 = 0; t1 < nmemb; t1++) {
2759 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2760 			if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) {
2761 				/* dump before panic */
2762 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2763 				    nmemb, size);
2764 				incr = 0;
2765 				for (t1 = 0; t1 < nmemb; t1++) {
2766 					const struct buf *bp = bp_array[t1];
2767 
2768 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2769 					    PRIu64 "\n", t1,
2770 					    (uint64_t)bp->b_bcount,
2771 					    (uint64_t)bp->b_lblkno);
2772 					printf("lbns:");
2773 					for (t2 = 0; t2 * size < bp->b_bcount;
2774 					    t2++) {
2775 						printf(" %jd",
2776 						    (intmax_t)lfs_blocks_get(fs, lb_array, incr++));
2777 					}
2778 					printf("\n");
2779 				}
2780 				panic("lfs_shellsort: inconsistent input");
2781 			}
2782 		}
2783 	}
2784 #endif
2785 
2786 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2787 		for (t1 = incr; t1 < nmemb; ++t1)
2788 			for (t2 = t1 - incr; t2 >= 0;)
2789 				if ((u_int64_t)bp_array[t2]->b_lblkno >
2790 				    (u_int64_t)bp_array[t2 + incr]->b_lblkno) {
2791 					bp_temp = bp_array[t2];
2792 					bp_array[t2] = bp_array[t2 + incr];
2793 					bp_array[t2 + incr] = bp_temp;
2794 					t2 -= incr;
2795 				} else
2796 					break;
2797 
2798 	/* Reform the list of logical blocks */
2799 	incr = 0;
2800 	for (t1 = 0; t1 < nmemb; t1++) {
2801 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2802 			lfs_blocks_set(fs, lb_array, incr++,
2803 				       bp_array[t1]->b_lblkno + t2);
2804 		}
2805 	}
2806 }
2807 
2808 /*
2809  * Set up an FINFO entry for a new file.  The fip pointer is assumed to
2810  * point at uninitialized space.
2811  */
2812 void
2813 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers)
2814 {
2815 	struct segment *sp = fs->lfs_sp;
2816 	SEGSUM *ssp;
2817 
2818 	KASSERT(vers > 0);
2819 
2820 	if (sp->seg_bytes_left < lfs_sb_getbsize(fs) ||
2821 	    sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs))
2822 		(void) lfs_writeseg(fs, fs->lfs_sp);
2823 
2824 	sp->sum_bytes_left -= FINFOSIZE(fs);
2825 	ssp = (SEGSUM *)sp->segsum;
2826 	lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1);
2827 	lfs_fi_setnblocks(fs, sp->fip, 0);
2828 	lfs_fi_setino(fs, sp->fip, ino);
2829 	lfs_fi_setversion(fs, sp->fip, vers);
2830 }
2831 
2832 /*
2833  * Release the FINFO entry, either clearing out an unused entry or
2834  * advancing us to the next available entry.
2835  */
2836 void
2837 lfs_release_finfo(struct lfs *fs)
2838 {
2839 	struct segment *sp = fs->lfs_sp;
2840 	SEGSUM *ssp;
2841 
2842 	if (lfs_fi_getnblocks(fs, sp->fip) != 0) {
2843 		sp->fip = NEXT_FINFO(fs, sp->fip);
2844 		lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip);
2845 	} else {
2846 		/* XXX shouldn't this update sp->fip? */
2847 		sp->sum_bytes_left += FINFOSIZE(fs);
2848 		ssp = (SEGSUM *)sp->segsum;
2849 		lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1);
2850 	}
2851 }
2852