xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: lfs_segment.c,v 1.206 2007/10/10 20:42:35 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the NetBSD
21  *	Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 /*
39  * Copyright (c) 1991, 1993
40  *	The Regents of the University of California.  All rights reserved.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
67  */
68 
69 #include <sys/cdefs.h>
70 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.206 2007/10/10 20:42:35 ad Exp $");
71 
72 #ifdef DEBUG
73 # define vndebug(vp, str) do {						\
74 	if (VTOI(vp)->i_flag & IN_CLEANING)				\
75 		DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \
76 		     VTOI(vp)->i_number, (str), op));			\
77 } while(0)
78 #else
79 # define vndebug(vp, str)
80 #endif
81 #define ivndebug(vp, str) \
82 	DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str)))
83 
84 #if defined(_KERNEL_OPT)
85 #include "opt_ddb.h"
86 #endif
87 
88 #include <sys/param.h>
89 #include <sys/systm.h>
90 #include <sys/namei.h>
91 #include <sys/kernel.h>
92 #include <sys/resourcevar.h>
93 #include <sys/file.h>
94 #include <sys/stat.h>
95 #include <sys/buf.h>
96 #include <sys/proc.h>
97 #include <sys/vnode.h>
98 #include <sys/mount.h>
99 #include <sys/kauth.h>
100 #include <sys/syslog.h>
101 
102 #include <miscfs/specfs/specdev.h>
103 #include <miscfs/fifofs/fifo.h>
104 
105 #include <ufs/ufs/inode.h>
106 #include <ufs/ufs/dir.h>
107 #include <ufs/ufs/ufsmount.h>
108 #include <ufs/ufs/ufs_extern.h>
109 
110 #include <ufs/lfs/lfs.h>
111 #include <ufs/lfs/lfs_extern.h>
112 
113 #include <uvm/uvm.h>
114 #include <uvm/uvm_extern.h>
115 
116 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
117 
118 extern int count_lock_queue(void);
119 extern struct simplelock vnode_free_list_slock;		/* XXX */
120 extern struct simplelock bqueue_slock;			/* XXX */
121 
122 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
123 static void lfs_free_aiodone(struct buf *);
124 static void lfs_super_aiodone(struct buf *);
125 static void lfs_cluster_aiodone(struct buf *);
126 static void lfs_cluster_callback(struct buf *);
127 
128 /*
129  * Determine if it's OK to start a partial in this segment, or if we need
130  * to go on to a new segment.
131  */
132 #define	LFS_PARTIAL_FITS(fs) \
133 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
134 	fragstofsb((fs), (fs)->lfs_frag))
135 
136 /*
137  * Figure out whether we should do a checkpoint write or go ahead with
138  * an ordinary write.
139  */
140 #define LFS_SHOULD_CHECKPOINT(fs, flags) \
141         ((flags & SEGM_CLEAN) == 0 &&					\
142 	  ((fs->lfs_nactive > LFS_MAX_ACTIVE ||				\
143 	    (flags & SEGM_CKP) ||					\
144 	    fs->lfs_nclean < LFS_MAX_ACTIVE)))
145 
146 int	 lfs_match_fake(struct lfs *, struct buf *);
147 void	 lfs_newseg(struct lfs *);
148 /* XXX ondisk32 */
149 void	 lfs_shellsort(struct buf **, int32_t *, int, int);
150 void	 lfs_supercallback(struct buf *);
151 void	 lfs_updatemeta(struct segment *);
152 void	 lfs_writesuper(struct lfs *, daddr_t);
153 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
154 	    struct segment *sp, int dirops);
155 
156 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
157 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
158 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
159 int	lfs_dirvcount = 0;		/* # active dirops */
160 
161 /* Statistics Counters */
162 int lfs_dostats = 1;
163 struct lfs_stats lfs_stats;
164 
165 /* op values to lfs_writevnodes */
166 #define	VN_REG		0
167 #define	VN_DIROP	1
168 #define	VN_EMPTY	2
169 #define VN_CLEAN	3
170 
171 /*
172  * XXX KS - Set modification time on the Ifile, so the cleaner can
173  * read the fs mod time off of it.  We don't set IN_UPDATE here,
174  * since we don't really need this to be flushed to disk (and in any
175  * case that wouldn't happen to the Ifile until we checkpoint).
176  */
177 void
178 lfs_imtime(struct lfs *fs)
179 {
180 	struct timespec ts;
181 	struct inode *ip;
182 
183 	ASSERT_MAYBE_SEGLOCK(fs);
184 	vfs_timestamp(&ts);
185 	ip = VTOI(fs->lfs_ivnode);
186 	ip->i_ffs1_mtime = ts.tv_sec;
187 	ip->i_ffs1_mtimensec = ts.tv_nsec;
188 }
189 
190 /*
191  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
192  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
193  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
194  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
195  */
196 
197 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
198 
199 int
200 lfs_vflush(struct vnode *vp)
201 {
202 	struct inode *ip;
203 	struct lfs *fs;
204 	struct segment *sp;
205 	struct buf *bp, *nbp, *tbp, *tnbp;
206 	int error, s;
207 	int flushed;
208 	int relock;
209 	int loopcount;
210 
211 	ip = VTOI(vp);
212 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
213 	relock = 0;
214 
215     top:
216 	ASSERT_NO_SEGLOCK(fs);
217 	if (ip->i_flag & IN_CLEANING) {
218 		ivndebug(vp,"vflush/in_cleaning");
219 		LFS_CLR_UINO(ip, IN_CLEANING);
220 		LFS_SET_UINO(ip, IN_MODIFIED);
221 
222 		/*
223 		 * Toss any cleaning buffers that have real counterparts
224 		 * to avoid losing new data.
225 		 */
226 		s = splbio();
227 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
228 			nbp = LIST_NEXT(bp, b_vnbufs);
229 			if (!LFS_IS_MALLOC_BUF(bp))
230 				continue;
231 			/*
232 			 * Look for pages matching the range covered
233 			 * by cleaning blocks.  It's okay if more dirty
234 			 * pages appear, so long as none disappear out
235 			 * from under us.
236 			 */
237 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
238 			    vp != fs->lfs_ivnode) {
239 				struct vm_page *pg;
240 				voff_t off;
241 
242 				simple_lock(&vp->v_interlock);
243 				for (off = lblktosize(fs, bp->b_lblkno);
244 				     off < lblktosize(fs, bp->b_lblkno + 1);
245 				     off += PAGE_SIZE) {
246 					pg = uvm_pagelookup(&vp->v_uobj, off);
247 					if (pg == NULL)
248 						continue;
249 					if ((pg->flags & PG_CLEAN) == 0 ||
250 					    pmap_is_modified(pg)) {
251 						fs->lfs_avail += btofsb(fs,
252 							bp->b_bcount);
253 						wakeup(&fs->lfs_avail);
254 						lfs_freebuf(fs, bp);
255 						bp = NULL;
256 						simple_unlock(&vp->v_interlock);
257 						goto nextbp;
258 					}
259 				}
260 				simple_unlock(&vp->v_interlock);
261 			}
262 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
263 			    tbp = tnbp)
264 			{
265 				tnbp = LIST_NEXT(tbp, b_vnbufs);
266 				if (tbp->b_vp == bp->b_vp
267 				   && tbp->b_lblkno == bp->b_lblkno
268 				   && tbp != bp)
269 				{
270 					fs->lfs_avail += btofsb(fs,
271 						bp->b_bcount);
272 					wakeup(&fs->lfs_avail);
273 					lfs_freebuf(fs, bp);
274 					bp = NULL;
275 					break;
276 				}
277 			}
278 		    nextbp:
279 			;
280 		}
281 		splx(s);
282 	}
283 
284 	/* If the node is being written, wait until that is done */
285 	simple_lock(&vp->v_interlock);
286 	s = splbio();
287 	if (WRITEINPROG(vp)) {
288 		ivndebug(vp,"vflush/writeinprog");
289 		ltsleep(vp, (PRIBIO+1), "lfs_vw", 0, &vp->v_interlock);
290 	}
291 	splx(s);
292 	simple_unlock(&vp->v_interlock);
293 
294 	/* Protect against VI_XLOCK deadlock in vinvalbuf() */
295 	lfs_seglock(fs, SEGM_SYNC);
296 
297 	/* If we're supposed to flush a freed inode, just toss it */
298 	if (ip->i_lfs_iflags & LFSI_DELETED) {
299 		DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n",
300 		      ip->i_number));
301 		s = splbio();
302 		/* Drain v_numoutput */
303 		simple_lock(&global_v_numoutput_slock);
304 		while (vp->v_numoutput > 0) {
305 			vp->v_iflag |= VI_BWAIT;
306 			ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf4", 0,
307 				&global_v_numoutput_slock);
308 		}
309 		simple_unlock(&global_v_numoutput_slock);
310 		KASSERT(vp->v_numoutput == 0);
311 
312 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
313 			nbp = LIST_NEXT(bp, b_vnbufs);
314 
315 			KASSERT((bp->b_flags & B_GATHERED) == 0);
316 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
317 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
318 				wakeup(&fs->lfs_avail);
319 			}
320 			/* Copied from lfs_writeseg */
321 			if (bp->b_flags & B_CALL) {
322 				biodone(bp);
323 			} else {
324 				bremfree(bp);
325 				LFS_UNLOCK_BUF(bp);
326 				bp->b_flags &=
327 				    ~(B_READ | B_DELWRI | B_GATHERED);
328 				bp->b_flags |= B_DONE;
329 				bp->b_error = 0;
330 				reassignbuf(bp, vp);
331 				brelse(bp, 0);
332 			}
333 		}
334 		splx(s);
335 		LFS_CLR_UINO(ip, IN_CLEANING);
336 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
337 		ip->i_flag &= ~IN_ALLMOD;
338 		DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n",
339 		      ip->i_number));
340 		lfs_segunlock(fs);
341 
342 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
343 
344 		return 0;
345 	}
346 
347 	fs->lfs_flushvp = vp;
348 	if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) {
349 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
350 		fs->lfs_flushvp = NULL;
351 		KASSERT(fs->lfs_flushvp_fakevref == 0);
352 		lfs_segunlock(fs);
353 
354 		/* Make sure that any pending buffers get written */
355 		s = splbio();
356 		simple_lock(&global_v_numoutput_slock);
357 		while (vp->v_numoutput > 0) {
358 			vp->v_iflag |= VI_BWAIT;
359 			ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf3", 0,
360 				&global_v_numoutput_slock);
361 		}
362 		simple_unlock(&global_v_numoutput_slock);
363 		splx(s);
364 
365 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
366 		KASSERT(vp->v_numoutput == 0);
367 
368 		return error;
369 	}
370 	sp = fs->lfs_sp;
371 
372 	flushed = 0;
373 	if (VPISEMPTY(vp)) {
374 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
375 		++flushed;
376 	} else if ((ip->i_flag & IN_CLEANING) &&
377 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
378 		ivndebug(vp,"vflush/clean");
379 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
380 		++flushed;
381 	} else if (lfs_dostats) {
382 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
383 			++lfs_stats.vflush_invoked;
384 		ivndebug(vp,"vflush");
385 	}
386 
387 #ifdef DIAGNOSTIC
388 	if (vp->v_uflag & VU_DIROP) {
389 		DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n"));
390 		/* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */
391 	}
392 	if (vp->v_usecount < 0) {
393 		printf("usecount=%ld\n", (long)vp->v_usecount);
394 		panic("lfs_vflush: usecount<0");
395 	}
396 #endif
397 
398 	do {
399 		loopcount = 0;
400 		do {
401 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
402 				relock = lfs_writefile(fs, sp, vp);
403 				if (relock) {
404 					/*
405 					 * Might have to wait for the
406 					 * cleaner to run; but we're
407 					 * still not done with this vnode.
408 					 */
409 					KDASSERT(ip->i_number != LFS_IFILE_INUM);
410 					lfs_writeinode(fs, sp, ip);
411 					LFS_SET_UINO(ip, IN_MODIFIED);
412 					lfs_writeseg(fs, sp);
413 					lfs_segunlock(fs);
414 					lfs_segunlock_relock(fs);
415 					goto top;
416 				}
417 			}
418 			/*
419 			 * If we begin a new segment in the middle of writing
420 			 * the Ifile, it creates an inconsistent checkpoint,
421 			 * since the Ifile information for the new segment
422 			 * is not up-to-date.  Take care of this here by
423 			 * sending the Ifile through again in case there
424 			 * are newly dirtied blocks.  But wait, there's more!
425 			 * This second Ifile write could *also* cross a segment
426 			 * boundary, if the first one was large.  The second
427 			 * one is guaranteed to be no more than 8 blocks,
428 			 * though (two segment blocks and supporting indirects)
429 			 * so the third write *will not* cross the boundary.
430 			 */
431 			if (vp == fs->lfs_ivnode) {
432 				lfs_writefile(fs, sp, vp);
433 				lfs_writefile(fs, sp, vp);
434 			}
435 #ifdef DEBUG
436 			if (++loopcount > 2)
437 				log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount);
438 #endif
439 		} while (lfs_writeinode(fs, sp, ip));
440 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
441 
442 	if (lfs_dostats) {
443 		++lfs_stats.nwrites;
444 		if (sp->seg_flags & SEGM_SYNC)
445 			++lfs_stats.nsync_writes;
446 		if (sp->seg_flags & SEGM_CKP)
447 			++lfs_stats.ncheckpoints;
448 	}
449 	/*
450 	 * If we were called from somewhere that has already held the seglock
451 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
452 	 * the write to complete because we are still locked.
453 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
454 	 * we must explicitly wait, if that is the case.
455 	 *
456 	 * We compare the iocount against 1, not 0, because it is
457 	 * artificially incremented by lfs_seglock().
458 	 */
459 	simple_lock(&fs->lfs_interlock);
460 	if (fs->lfs_seglock > 1) {
461 		while (fs->lfs_iocount > 1)
462 			(void)ltsleep(&fs->lfs_iocount, PRIBIO + 1,
463 				     "lfs_vflush", 0, &fs->lfs_interlock);
464 	}
465 	simple_unlock(&fs->lfs_interlock);
466 
467 	lfs_segunlock(fs);
468 
469 	/* Wait for these buffers to be recovered by aiodoned */
470 	s = splbio();
471 	simple_lock(&global_v_numoutput_slock);
472 	while (vp->v_numoutput > 0) {
473 		vp->v_iflag |= VI_BWAIT;
474 		ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf2", 0,
475 			&global_v_numoutput_slock);
476 	}
477 	simple_unlock(&global_v_numoutput_slock);
478 	splx(s);
479 
480 	KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
481 	KASSERT(vp->v_numoutput == 0);
482 
483 	fs->lfs_flushvp = NULL;
484 	KASSERT(fs->lfs_flushvp_fakevref == 0);
485 
486 	return (0);
487 }
488 
489 int
490 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
491 {
492 	struct inode *ip;
493 	struct vnode *vp;
494 	int inodes_written = 0, only_cleaning;
495 	int error = 0;
496 
497 	ASSERT_SEGLOCK(fs);
498  loop:
499 	/* start at last (newest) vnode. */
500 	TAILQ_FOREACH_REVERSE(vp, &mp->mnt_vnodelist, vnodelst, v_mntvnodes) {
501 		/*
502 		 * If the vnode that we are about to sync is no longer
503 		 * associated with this mount point, start over.
504 		 */
505 		if (vp->v_mount != mp) {
506 			DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n"));
507 			/*
508 			 * After this, pages might be busy
509 			 * due to our own previous putpages.
510 			 * Start actual segment write here to avoid deadlock.
511 			 */
512 			(void)lfs_writeseg(fs, sp);
513 			goto loop;
514 		}
515 
516 		if (vp->v_type == VNON) {
517 			continue;
518 		}
519 
520 		ip = VTOI(vp);
521 		if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) ||
522 		    (op != VN_DIROP && op != VN_CLEAN &&
523 		    (vp->v_uflag & VU_DIROP))) {
524 			vndebug(vp,"dirop");
525 			continue;
526 		}
527 
528 		if (op == VN_EMPTY && !VPISEMPTY(vp)) {
529 			vndebug(vp,"empty");
530 			continue;
531 		}
532 
533 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
534 		   && vp != fs->lfs_flushvp
535 		   && !(ip->i_flag & IN_CLEANING)) {
536 			vndebug(vp,"cleaning");
537 			continue;
538 		}
539 
540 		if (lfs_vref(vp)) {
541 			vndebug(vp,"vref");
542 			continue;
543 		}
544 
545 		only_cleaning = 0;
546 		/*
547 		 * Write the inode/file if dirty and it's not the IFILE.
548 		 */
549 		if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) {
550 			only_cleaning =
551 			    ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
552 
553 			if (ip->i_number != LFS_IFILE_INUM) {
554 				error = lfs_writefile(fs, sp, vp);
555 				if (error) {
556 					lfs_vunref(vp);
557 					if (error == EAGAIN) {
558 						/*
559 						 * This error from lfs_putpages
560 						 * indicates we need to drop
561 						 * the segment lock and start
562 						 * over after the cleaner has
563 						 * had a chance to run.
564 						 */
565 						lfs_writeinode(fs, sp, ip);
566 						lfs_writeseg(fs, sp);
567 						if (!VPISEMPTY(vp) &&
568 						    !WRITEINPROG(vp) &&
569 						    !(ip->i_flag & IN_ALLMOD))
570 							LFS_SET_UINO(ip, IN_MODIFIED);
571 						break;
572 					}
573 					error = 0; /* XXX not quite right */
574 					continue;
575 				}
576 
577 				if (!VPISEMPTY(vp)) {
578 					if (WRITEINPROG(vp)) {
579 						ivndebug(vp,"writevnodes/write2");
580 					} else if (!(ip->i_flag & IN_ALLMOD)) {
581 						LFS_SET_UINO(ip, IN_MODIFIED);
582 					}
583 				}
584 				(void) lfs_writeinode(fs, sp, ip);
585 				inodes_written++;
586 			}
587 		}
588 
589 		if (lfs_clean_vnhead && only_cleaning)
590 			lfs_vunref_head(vp);
591 		else
592 			lfs_vunref(vp);
593 	}
594 	return error;
595 }
596 
597 /*
598  * Do a checkpoint.
599  */
600 int
601 lfs_segwrite(struct mount *mp, int flags)
602 {
603 	struct buf *bp;
604 	struct inode *ip;
605 	struct lfs *fs;
606 	struct segment *sp;
607 	struct vnode *vp;
608 	SEGUSE *segusep;
609 	int do_ckp, did_ckp, error, s;
610 	unsigned n, segleft, maxseg, sn, i, curseg;
611 	int writer_set = 0;
612 	int dirty;
613 	int redo;
614 	int um_error;
615 	int loopcount;
616 
617 	fs = VFSTOUFS(mp)->um_lfs;
618 	ASSERT_MAYBE_SEGLOCK(fs);
619 
620 	if (fs->lfs_ronly)
621 		return EROFS;
622 
623 	lfs_imtime(fs);
624 
625 	/*
626 	 * Allocate a segment structure and enough space to hold pointers to
627 	 * the maximum possible number of buffers which can be described in a
628 	 * single summary block.
629 	 */
630 	do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags);
631 
632 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
633 	sp = fs->lfs_sp;
634 	if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP))
635 		do_ckp = 1;
636 
637 	/*
638 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
639 	 * in which case we have to flush *all* buffers off of this vnode.
640 	 * We don't care about other nodes, but write any non-dirop nodes
641 	 * anyway in anticipation of another getnewvnode().
642 	 *
643 	 * If we're cleaning we only write cleaning and ifile blocks, and
644 	 * no dirops, since otherwise we'd risk corruption in a crash.
645 	 */
646 	if (sp->seg_flags & SEGM_CLEAN)
647 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
648 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
649 		do {
650 			um_error = lfs_writevnodes(fs, mp, sp, VN_REG);
651 
652 			if (do_ckp || fs->lfs_dirops == 0) {
653 				if (!writer_set) {
654 					lfs_writer_enter(fs, "lfs writer");
655 					writer_set = 1;
656 				}
657 				error = lfs_writevnodes(fs, mp, sp, VN_DIROP);
658 				if (um_error == 0)
659 					um_error = error;
660 				/* In case writevnodes errored out */
661 				lfs_flush_dirops(fs);
662 				((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
663 				lfs_finalize_fs_seguse(fs);
664 			}
665 			if (do_ckp && um_error) {
666 				lfs_segunlock_relock(fs);
667 				sp = fs->lfs_sp;
668 			}
669 		} while (do_ckp && um_error != 0);
670 	}
671 
672 	/*
673 	 * If we are doing a checkpoint, mark everything since the
674 	 * last checkpoint as no longer ACTIVE.
675 	 */
676 	if (do_ckp || fs->lfs_doifile) {
677 		segleft = fs->lfs_nseg;
678 		curseg = 0;
679 		for (n = 0; n < fs->lfs_segtabsz; n++) {
680 			dirty = 0;
681 			if (bread(fs->lfs_ivnode,
682 			    fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &bp))
683 				panic("lfs_segwrite: ifile read");
684 			segusep = (SEGUSE *)bp->b_data;
685 			maxseg = min(segleft, fs->lfs_sepb);
686 			for (i = 0; i < maxseg; i++) {
687 				sn = curseg + i;
688 				if (sn != dtosn(fs, fs->lfs_curseg) &&
689 				    segusep->su_flags & SEGUSE_ACTIVE) {
690 					segusep->su_flags &= ~SEGUSE_ACTIVE;
691 					--fs->lfs_nactive;
692 					++dirty;
693 				}
694 				fs->lfs_suflags[fs->lfs_activesb][sn] =
695 					segusep->su_flags;
696 				if (fs->lfs_version > 1)
697 					++segusep;
698 				else
699 					segusep = (SEGUSE *)
700 						((SEGUSE_V1 *)segusep + 1);
701 			}
702 
703 			if (dirty)
704 				error = LFS_BWRITE_LOG(bp); /* Ifile */
705 			else
706 				brelse(bp, 0);
707 			segleft -= fs->lfs_sepb;
708 			curseg += fs->lfs_sepb;
709 		}
710 	}
711 
712 	LOCK_ASSERT(LFS_SEGLOCK_HELD(fs));
713 
714 	did_ckp = 0;
715 	if (do_ckp || fs->lfs_doifile) {
716 		vp = fs->lfs_ivnode;
717 		vn_lock(vp, LK_EXCLUSIVE);
718 		loopcount = 0;
719 		do {
720 #ifdef DEBUG
721 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid);
722 #endif
723 			simple_lock(&fs->lfs_interlock);
724 			fs->lfs_flags &= ~LFS_IFDIRTY;
725 			simple_unlock(&fs->lfs_interlock);
726 
727 			ip = VTOI(vp);
728 
729 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
730 				/*
731 				 * Ifile has no pages, so we don't need
732 				 * to check error return here.
733 				 */
734 				lfs_writefile(fs, sp, vp);
735 				/*
736 				 * Ensure the Ifile takes the current segment
737 				 * into account.  See comment in lfs_vflush.
738 				 */
739 				lfs_writefile(fs, sp, vp);
740 				lfs_writefile(fs, sp, vp);
741 			}
742 
743 			if (ip->i_flag & IN_ALLMOD)
744 				++did_ckp;
745 #if 0
746 			redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0);
747 #else
748 			redo = lfs_writeinode(fs, sp, ip);
749 #endif
750 			redo += lfs_writeseg(fs, sp);
751 			simple_lock(&fs->lfs_interlock);
752 			redo += (fs->lfs_flags & LFS_IFDIRTY);
753 			simple_unlock(&fs->lfs_interlock);
754 #ifdef DEBUG
755 			if (++loopcount > 2)
756 				log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n",
757 					loopcount);
758 #endif
759 		} while (redo && do_ckp);
760 
761 		/*
762 		 * Unless we are unmounting, the Ifile may continue to have
763 		 * dirty blocks even after a checkpoint, due to changes to
764 		 * inodes' atime.  If we're checkpointing, it's "impossible"
765 		 * for other parts of the Ifile to be dirty after the loop
766 		 * above, since we hold the segment lock.
767 		 */
768 		s = splbio();
769 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
770 			LFS_CLR_UINO(ip, IN_ALLMOD);
771 		}
772 #ifdef DIAGNOSTIC
773 		else if (do_ckp) {
774 			int do_panic = 0;
775 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
776 				if (bp->b_lblkno < fs->lfs_cleansz +
777 				    fs->lfs_segtabsz &&
778 				    !(bp->b_flags & B_GATHERED)) {
779 					printf("ifile lbn %ld still dirty (flags %lx)\n",
780 						(long)bp->b_lblkno,
781 						(long)bp->b_flags);
782 					++do_panic;
783 				}
784 			}
785 			if (do_panic)
786 				panic("dirty blocks");
787 		}
788 #endif
789 		splx(s);
790 		VOP_UNLOCK(vp, 0);
791 	} else {
792 		(void) lfs_writeseg(fs, sp);
793 	}
794 
795 	/* Note Ifile no longer needs to be written */
796 	fs->lfs_doifile = 0;
797 	if (writer_set)
798 		lfs_writer_leave(fs);
799 
800 	/*
801 	 * If we didn't write the Ifile, we didn't really do anything.
802 	 * That means that (1) there is a checkpoint on disk and (2)
803 	 * nothing has changed since it was written.
804 	 *
805 	 * Take the flags off of the segment so that lfs_segunlock
806 	 * doesn't have to write the superblock either.
807 	 */
808 	if (do_ckp && !did_ckp) {
809 		sp->seg_flags &= ~SEGM_CKP;
810 	}
811 
812 	if (lfs_dostats) {
813 		++lfs_stats.nwrites;
814 		if (sp->seg_flags & SEGM_SYNC)
815 			++lfs_stats.nsync_writes;
816 		if (sp->seg_flags & SEGM_CKP)
817 			++lfs_stats.ncheckpoints;
818 	}
819 	lfs_segunlock(fs);
820 	return (0);
821 }
822 
823 /*
824  * Write the dirty blocks associated with a vnode.
825  */
826 int
827 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
828 {
829 	struct finfo *fip;
830 	struct inode *ip;
831 	int i, frag;
832 	int error;
833 
834 	ASSERT_SEGLOCK(fs);
835 	error = 0;
836 	ip = VTOI(vp);
837 
838 	fip = sp->fip;
839 	lfs_acquire_finfo(fs, ip->i_number, ip->i_gen);
840 
841 	if (vp->v_uflag & VU_DIROP)
842 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
843 
844 	if (sp->seg_flags & SEGM_CLEAN) {
845 		lfs_gather(fs, sp, vp, lfs_match_fake);
846 		/*
847 		 * For a file being flushed, we need to write *all* blocks.
848 		 * This means writing the cleaning blocks first, and then
849 		 * immediately following with any non-cleaning blocks.
850 		 * The same is true of the Ifile since checkpoints assume
851 		 * that all valid Ifile blocks are written.
852 		 */
853 		if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) {
854 			lfs_gather(fs, sp, vp, lfs_match_data);
855 			/*
856 			 * Don't call VOP_PUTPAGES: if we're flushing,
857 			 * we've already done it, and the Ifile doesn't
858 			 * use the page cache.
859 			 */
860 		}
861 	} else {
862 		lfs_gather(fs, sp, vp, lfs_match_data);
863 		/*
864 		 * If we're flushing, we've already called VOP_PUTPAGES
865 		 * so don't do it again.  Otherwise, we want to write
866 		 * everything we've got.
867 		 */
868 		if (!IS_FLUSHING(fs, vp)) {
869 			simple_lock(&vp->v_interlock);
870 			error = VOP_PUTPAGES(vp, 0, 0,
871 				PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
872 		}
873 	}
874 
875 	/*
876 	 * It may not be necessary to write the meta-data blocks at this point,
877 	 * as the roll-forward recovery code should be able to reconstruct the
878 	 * list.
879 	 *
880 	 * We have to write them anyway, though, under two conditions: (1) the
881 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
882 	 * checkpointing.
883 	 *
884 	 * BUT if we are cleaning, we might have indirect blocks that refer to
885 	 * new blocks not being written yet, in addition to fragments being
886 	 * moved out of a cleaned segment.  If that is the case, don't
887 	 * write the indirect blocks, or the finfo will have a small block
888 	 * in the middle of it!
889 	 * XXX in this case isn't the inode size wrong too?
890 	 */
891 	frag = 0;
892 	if (sp->seg_flags & SEGM_CLEAN) {
893 		for (i = 0; i < NDADDR; i++)
894 			if (ip->i_lfs_fragsize[i] > 0 &&
895 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
896 				++frag;
897 	}
898 #ifdef DIAGNOSTIC
899 	if (frag > 1)
900 		panic("lfs_writefile: more than one fragment!");
901 #endif
902 	if (IS_FLUSHING(fs, vp) ||
903 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
904 		lfs_gather(fs, sp, vp, lfs_match_indir);
905 		lfs_gather(fs, sp, vp, lfs_match_dindir);
906 		lfs_gather(fs, sp, vp, lfs_match_tindir);
907 	}
908 	fip = sp->fip;
909 	lfs_release_finfo(fs);
910 
911 	return error;
912 }
913 
914 /*
915  * Update segment accounting to reflect this inode's change of address.
916  */
917 static int
918 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr)
919 {
920 	struct buf *bp;
921 	daddr_t daddr;
922 	IFILE *ifp;
923 	SEGUSE *sup;
924 	ino_t ino;
925 	int redo_ifile, error;
926 	u_int32_t sn;
927 
928 	redo_ifile = 0;
929 
930 	/*
931 	 * If updating the ifile, update the super-block.  Update the disk
932 	 * address and access times for this inode in the ifile.
933 	 */
934 	ino = ip->i_number;
935 	if (ino == LFS_IFILE_INUM) {
936 		daddr = fs->lfs_idaddr;
937 		fs->lfs_idaddr = dbtofsb(fs, ndaddr);
938 	} else {
939 		LFS_IENTRY(ifp, fs, ino, bp);
940 		daddr = ifp->if_daddr;
941 		ifp->if_daddr = dbtofsb(fs, ndaddr);
942 		error = LFS_BWRITE_LOG(bp); /* Ifile */
943 	}
944 
945 	/*
946 	 * If this is the Ifile and lfs_offset is set to the first block
947 	 * in the segment, dirty the new segment's accounting block
948 	 * (XXX should already be dirty?) and tell the caller to do it again.
949 	 */
950 	if (ip->i_number == LFS_IFILE_INUM) {
951 		sn = dtosn(fs, fs->lfs_offset);
952 		if (sntod(fs, sn) + btofsb(fs, fs->lfs_sumsize) ==
953 		    fs->lfs_offset) {
954 			LFS_SEGENTRY(sup, fs, sn, bp);
955 			KASSERT(bp->b_flags & B_DELWRI);
956 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
957 			/* fs->lfs_flags |= LFS_IFDIRTY; */
958 			redo_ifile |= 1;
959 		}
960 	}
961 
962 	/*
963 	 * The inode's last address should not be in the current partial
964 	 * segment, except under exceptional circumstances (lfs_writevnodes
965 	 * had to start over, and in the meantime more blocks were written
966 	 * to a vnode).	 Both inodes will be accounted to this segment
967 	 * in lfs_writeseg so we need to subtract the earlier version
968 	 * here anyway.	 The segment count can temporarily dip below
969 	 * zero here; keep track of how many duplicates we have in
970 	 * "dupino" so we don't panic below.
971 	 */
972 	if (daddr >= fs->lfs_lastpseg && daddr <= fs->lfs_offset) {
973 		++sp->ndupino;
974 		DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg "
975 		      "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
976 		      (long long)daddr, sp->ndupino));
977 	}
978 	/*
979 	 * Account the inode: it no longer belongs to its former segment,
980 	 * though it will not belong to the new segment until that segment
981 	 * is actually written.
982 	 */
983 	if (daddr != LFS_UNUSED_DADDR) {
984 		u_int32_t oldsn = dtosn(fs, daddr);
985 #ifdef DIAGNOSTIC
986 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
987 #endif
988 		LFS_SEGENTRY(sup, fs, oldsn, bp);
989 #ifdef DIAGNOSTIC
990 		if (sup->su_nbytes +
991 		    sizeof (struct ufs1_dinode) * ndupino
992 		      < sizeof (struct ufs1_dinode)) {
993 			printf("lfs_writeinode: negative bytes "
994 			       "(segment %" PRIu32 " short by %d, "
995 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
996 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
997 			       "ndupino=%d)\n",
998 			       dtosn(fs, daddr),
999 			       (int)sizeof (struct ufs1_dinode) *
1000 				   (1 - sp->ndupino) - sup->su_nbytes,
1001 			       oldsn, sp->seg_number, daddr,
1002 			       (unsigned int)sup->su_nbytes,
1003 			       sp->ndupino);
1004 			panic("lfs_writeinode: negative bytes");
1005 			sup->su_nbytes = sizeof (struct ufs1_dinode);
1006 		}
1007 #endif
1008 		DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n",
1009 		      dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino));
1010 		sup->su_nbytes -= sizeof (struct ufs1_dinode);
1011 		redo_ifile |=
1012 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1013 		if (redo_ifile) {
1014 			simple_lock(&fs->lfs_interlock);
1015 			fs->lfs_flags |= LFS_IFDIRTY;
1016 			simple_unlock(&fs->lfs_interlock);
1017 			/* Don't double-account */
1018 			fs->lfs_idaddr = 0x0;
1019 		}
1020 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1021 	}
1022 
1023 	return redo_ifile;
1024 }
1025 
1026 int
1027 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
1028 {
1029 	struct buf *bp;
1030 	struct ufs1_dinode *cdp;
1031 	daddr_t daddr;
1032 	int32_t *daddrp;	/* XXX ondisk32 */
1033 	int i, ndx;
1034 	int redo_ifile = 0;
1035 	int gotblk = 0;
1036 	int count;
1037 
1038 	ASSERT_SEGLOCK(fs);
1039 	if (!(ip->i_flag & IN_ALLMOD))
1040 		return (0);
1041 
1042 	/* Can't write ifile when writer is not set */
1043 	KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 ||
1044 		(sp->seg_flags & SEGM_CLEAN));
1045 
1046 	/*
1047 	 * If this is the Ifile, see if writing it here will generate a
1048 	 * temporary misaccounting.  If it will, do the accounting and write
1049 	 * the blocks, postponing the inode write until the accounting is
1050 	 * solid.
1051 	 */
1052 	count = 0;
1053 	while (ip->i_number == LFS_IFILE_INUM) {
1054 		int redo = 0;
1055 
1056 		if (sp->idp == NULL && sp->ibp == NULL &&
1057 		    (sp->seg_bytes_left < fs->lfs_ibsize ||
1058 		     sp->sum_bytes_left < sizeof(int32_t))) {
1059 			(void) lfs_writeseg(fs, sp);
1060 			continue;
1061 		}
1062 
1063 		/* Look for dirty Ifile blocks */
1064 		LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) {
1065 			if (!(bp->b_flags & B_GATHERED)) {
1066 				redo = 1;
1067 				break;
1068 			}
1069 		}
1070 
1071 		if (redo == 0)
1072 			redo = lfs_update_iaddr(fs, sp, ip, 0x0);
1073 		if (redo == 0)
1074 			break;
1075 
1076 		if (sp->idp) {
1077 			sp->idp->di_inumber = 0;
1078 			sp->idp = NULL;
1079 		}
1080 		++count;
1081 		if (count > 2)
1082 			log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count);
1083 		lfs_writefile(fs, sp, fs->lfs_ivnode);
1084 	}
1085 
1086 	/* Allocate a new inode block if necessary. */
1087 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
1088 	    sp->ibp == NULL) {
1089 		/* Allocate a new segment if necessary. */
1090 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
1091 		    sp->sum_bytes_left < sizeof(int32_t))
1092 			(void) lfs_writeseg(fs, sp);
1093 
1094 		/* Get next inode block. */
1095 		daddr = fs->lfs_offset;
1096 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
1097 		sp->ibp = *sp->cbpp++ =
1098 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1099 			    fsbtodb(fs, daddr), fs->lfs_ibsize, 0, 0);
1100 		gotblk++;
1101 
1102 		/* Zero out inode numbers */
1103 		for (i = 0; i < INOPB(fs); ++i)
1104 			((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber =
1105 			    0;
1106 
1107 		++sp->start_bpp;
1108 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
1109 		/* Set remaining space counters. */
1110 		sp->seg_bytes_left -= fs->lfs_ibsize;
1111 		sp->sum_bytes_left -= sizeof(int32_t);
1112 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
1113 			sp->ninodes / INOPB(fs) - 1;
1114 		((int32_t *)(sp->segsum))[ndx] = daddr;
1115 	}
1116 
1117 	/* Check VU_DIROP in case there is a new file with no data blocks */
1118 	if (ITOV(ip)->v_uflag & VU_DIROP)
1119 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
1120 
1121 	/* Update the inode times and copy the inode onto the inode page. */
1122 	/* XXX kludge --- don't redirty the ifile just to put times on it */
1123 	if (ip->i_number != LFS_IFILE_INUM)
1124 		LFS_ITIMES(ip, NULL, NULL, NULL);
1125 
1126 	/*
1127 	 * If this is the Ifile, and we've already written the Ifile in this
1128 	 * partial segment, just overwrite it (it's not on disk yet) and
1129 	 * continue.
1130 	 *
1131 	 * XXX we know that the bp that we get the second time around has
1132 	 * already been gathered.
1133 	 */
1134 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
1135 		*(sp->idp) = *ip->i_din.ffs1_din;
1136 		ip->i_lfs_osize = ip->i_size;
1137 		return 0;
1138 	}
1139 
1140 	bp = sp->ibp;
1141 	cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
1142 	*cdp = *ip->i_din.ffs1_din;
1143 
1144 	/*
1145 	 * If cleaning, link counts and directory file sizes cannot change,
1146 	 * since those would be directory operations---even if the file
1147 	 * we are writing is marked VU_DIROP we should write the old values.
1148 	 * If we're not cleaning, of course, update the values so we get
1149 	 * current values the next time we clean.
1150 	 */
1151 	if (sp->seg_flags & SEGM_CLEAN) {
1152 		if (ITOV(ip)->v_uflag & VU_DIROP) {
1153 			cdp->di_nlink = ip->i_lfs_odnlink;
1154 			/* if (ITOV(ip)->v_type == VDIR) */
1155 			cdp->di_size = ip->i_lfs_osize;
1156 		}
1157 	} else {
1158 		ip->i_lfs_odnlink = cdp->di_nlink;
1159 		ip->i_lfs_osize = ip->i_size;
1160 	}
1161 
1162 
1163 	/* We can finish the segment accounting for truncations now */
1164 	lfs_finalize_ino_seguse(fs, ip);
1165 
1166 	/*
1167 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
1168 	 * addresses to disk; possibly change the on-disk record of
1169 	 * the inode size, either by reverting to the previous size
1170 	 * (in the case of cleaning) or by verifying the inode's block
1171 	 * holdings (in the case of files being allocated as they are being
1172 	 * written).
1173 	 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail
1174 	 * XXX count on disk wrong by the same amount.	We should be
1175 	 * XXX able to "borrow" from lfs_avail and return it after the
1176 	 * XXX Ifile is written.  See also in lfs_writeseg.
1177 	 */
1178 
1179 	/* Check file size based on highest allocated block */
1180 	if (((ip->i_ffs1_mode & IFMT) == IFREG ||
1181 	     (ip->i_ffs1_mode & IFMT) == IFDIR) &&
1182 	    ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) {
1183 		cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift;
1184 		DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %"
1185 		      PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size));
1186 	}
1187 	if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) {
1188 		DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)"
1189 		      " at %x\n", ip->i_number, ip->i_lfs_effnblks,
1190 		      ip->i_ffs1_blocks, fs->lfs_offset));
1191 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
1192 		     daddrp++) {
1193 			if (*daddrp == UNWRITTEN) {
1194 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1195 				*daddrp = 0;
1196 			}
1197 		}
1198 	}
1199 
1200 #ifdef DIAGNOSTIC
1201 	/*
1202 	 * Check dinode held blocks against dinode size.
1203 	 * This should be identical to the check in lfs_vget().
1204 	 */
1205 	for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift;
1206 	     i < NDADDR; i++) {
1207 		KASSERT(i >= 0);
1208 		if ((cdp->di_mode & IFMT) == IFLNK)
1209 			continue;
1210 		if (((cdp->di_mode & IFMT) == IFBLK ||
1211 		     (cdp->di_mode & IFMT) == IFCHR) && i == 0)
1212 			continue;
1213 		if (cdp->di_db[i] != 0) {
1214 # ifdef DEBUG
1215 			lfs_dump_dinode(cdp);
1216 # endif
1217 			panic("writing inconsistent inode");
1218 		}
1219 	}
1220 #endif /* DIAGNOSTIC */
1221 
1222 	if (ip->i_flag & IN_CLEANING)
1223 		LFS_CLR_UINO(ip, IN_CLEANING);
1224 	else {
1225 		/* XXX IN_ALLMOD */
1226 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
1227 			     IN_UPDATE | IN_MODIFY);
1228 		if (ip->i_lfs_effnblks == ip->i_ffs1_blocks)
1229 			LFS_CLR_UINO(ip, IN_MODIFIED);
1230 		else {
1231 			DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real "
1232 			    "blks=%d, eff=%d\n", ip->i_number,
1233 			    ip->i_ffs1_blocks, ip->i_lfs_effnblks));
1234 		}
1235 	}
1236 
1237 	if (ip->i_number == LFS_IFILE_INUM) {
1238 		/* We know sp->idp == NULL */
1239 		sp->idp = ((struct ufs1_dinode *)bp->b_data) +
1240 			(sp->ninodes % INOPB(fs));
1241 
1242 		/* Not dirty any more */
1243 		simple_lock(&fs->lfs_interlock);
1244 		fs->lfs_flags &= ~LFS_IFDIRTY;
1245 		simple_unlock(&fs->lfs_interlock);
1246 	}
1247 
1248 	if (gotblk) {
1249 		LFS_LOCK_BUF(bp);
1250 		brelse(bp, 0);
1251 	}
1252 
1253 	/* Increment inode count in segment summary block. */
1254 	++((SEGSUM *)(sp->segsum))->ss_ninos;
1255 
1256 	/* If this page is full, set flag to allocate a new page. */
1257 	if (++sp->ninodes % INOPB(fs) == 0)
1258 		sp->ibp = NULL;
1259 
1260 	redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno);
1261 
1262 	KASSERT(redo_ifile == 0);
1263 	return (redo_ifile);
1264 }
1265 
1266 int
1267 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1268 {
1269 	struct lfs *fs;
1270 	int vers;
1271 	int j, blksinblk;
1272 
1273 	ASSERT_SEGLOCK(sp->fs);
1274 	/*
1275 	 * If full, finish this segment.  We may be doing I/O, so
1276 	 * release and reacquire the splbio().
1277 	 */
1278 #ifdef DIAGNOSTIC
1279 	if (sp->vp == NULL)
1280 		panic ("lfs_gatherblock: Null vp in segment");
1281 #endif
1282 	fs = sp->fs;
1283 	blksinblk = howmany(bp->b_bcount, fs->lfs_bsize);
1284 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1285 	    sp->seg_bytes_left < bp->b_bcount) {
1286 		if (sptr)
1287 			splx(*sptr);
1288 		lfs_updatemeta(sp);
1289 
1290 		vers = sp->fip->fi_version;
1291 		(void) lfs_writeseg(fs, sp);
1292 
1293 		/* Add the current file to the segment summary. */
1294 		lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers);
1295 
1296 		if (sptr)
1297 			*sptr = splbio();
1298 		return (1);
1299 	}
1300 
1301 	if (bp->b_flags & B_GATHERED) {
1302 		DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d,"
1303 		      " lbn %" PRId64 "\n",
1304 		      sp->fip->fi_ino, bp->b_lblkno));
1305 		return (0);
1306 	}
1307 
1308 	/* Insert into the buffer list, update the FINFO block. */
1309 	bp->b_flags |= B_GATHERED;
1310 
1311 	*sp->cbpp++ = bp;
1312 	for (j = 0; j < blksinblk; j++) {
1313 		sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j;
1314 		/* This block's accounting moves from lfs_favail to lfs_avail */
1315 		lfs_deregister_block(sp->vp, bp->b_lblkno + j);
1316 	}
1317 
1318 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1319 	sp->seg_bytes_left -= bp->b_bcount;
1320 	return (0);
1321 }
1322 
1323 int
1324 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1325     int (*match)(struct lfs *, struct buf *))
1326 {
1327 	struct buf *bp, *nbp;
1328 	int s, count = 0;
1329 
1330 	ASSERT_SEGLOCK(fs);
1331 	if (vp->v_type == VBLK)
1332 		return 0;
1333 	KASSERT(sp->vp == NULL);
1334 	sp->vp = vp;
1335 	s = splbio();
1336 
1337 #ifndef LFS_NO_BACKBUF_HACK
1338 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1339 # define	BUF_OFFSET	\
1340 	(((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp)
1341 # define	BACK_BUF(BP)	\
1342 	((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1343 # define	BEG_OF_LIST	\
1344 	((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1345 
1346 loop:
1347 	/* Find last buffer. */
1348 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1349 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1350 	     bp = LIST_NEXT(bp, b_vnbufs))
1351 		/* nothing */;
1352 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1353 		nbp = BACK_BUF(bp);
1354 #else /* LFS_NO_BACKBUF_HACK */
1355 loop:
1356 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1357 		nbp = LIST_NEXT(bp, b_vnbufs);
1358 #endif /* LFS_NO_BACKBUF_HACK */
1359 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1360 #ifdef DEBUG
1361 			if (vp == fs->lfs_ivnode &&
1362 			    (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1363 				log(LOG_NOTICE, "lfs_gather: ifile lbn %"
1364 				      PRId64 " busy (%x) at 0x%x",
1365 				      bp->b_lblkno, bp->b_flags,
1366 				      (unsigned)fs->lfs_offset);
1367 #endif
1368 			continue;
1369 		}
1370 #ifdef DIAGNOSTIC
1371 # ifdef LFS_USE_B_INVAL
1372 		if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1373 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1374 			      " is B_INVAL\n", bp->b_lblkno));
1375 			VOP_PRINT(bp->b_vp);
1376 		}
1377 # endif /* LFS_USE_B_INVAL */
1378 		if (!(bp->b_flags & B_DELWRI))
1379 			panic("lfs_gather: bp not B_DELWRI");
1380 		if (!(bp->b_flags & B_LOCKED)) {
1381 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1382 			      " blk %" PRId64 " not B_LOCKED\n",
1383 			      bp->b_lblkno,
1384 			      dbtofsb(fs, bp->b_blkno)));
1385 			VOP_PRINT(bp->b_vp);
1386 			panic("lfs_gather: bp not B_LOCKED");
1387 		}
1388 #endif
1389 		if (lfs_gatherblock(sp, bp, &s)) {
1390 			goto loop;
1391 		}
1392 		count++;
1393 	}
1394 	splx(s);
1395 	lfs_updatemeta(sp);
1396 	KASSERT(sp->vp == vp);
1397 	sp->vp = NULL;
1398 	return count;
1399 }
1400 
1401 #if DEBUG
1402 # define DEBUG_OOFF(n) do {						\
1403 	if (ooff == 0) {						\
1404 		DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \
1405 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1406 			", was 0x0 (or %" PRId64 ")\n",			\
1407 			(n), ip->i_number, lbn, ndaddr, daddr));	\
1408 	}								\
1409 } while (0)
1410 #else
1411 # define DEBUG_OOFF(n)
1412 #endif
1413 
1414 /*
1415  * Change the given block's address to ndaddr, finding its previous
1416  * location using ufs_bmaparray().
1417  *
1418  * Account for this change in the segment table.
1419  *
1420  * called with sp == NULL by roll-forwarding code.
1421  */
1422 void
1423 lfs_update_single(struct lfs *fs, struct segment *sp,
1424     struct vnode *vp, daddr_t lbn, int32_t ndaddr, int size)
1425 {
1426 	SEGUSE *sup;
1427 	struct buf *bp;
1428 	struct indir a[NIADDR + 2], *ap;
1429 	struct inode *ip;
1430 	daddr_t daddr, ooff;
1431 	int num, error;
1432 	int bb, osize, obb;
1433 
1434 	ASSERT_SEGLOCK(fs);
1435 	KASSERT(sp == NULL || sp->vp == vp);
1436 	ip = VTOI(vp);
1437 
1438 	error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1439 	if (error)
1440 		panic("lfs_updatemeta: ufs_bmaparray returned %d", error);
1441 
1442 	daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */
1443 	KASSERT(daddr <= LFS_MAX_DADDR);
1444 	if (daddr > 0)
1445 		daddr = dbtofsb(fs, daddr);
1446 
1447 	bb = fragstofsb(fs, numfrags(fs, size));
1448 	switch (num) {
1449 	    case 0:
1450 		    ooff = ip->i_ffs1_db[lbn];
1451 		    DEBUG_OOFF(0);
1452 		    if (ooff == UNWRITTEN)
1453 			    ip->i_ffs1_blocks += bb;
1454 		    else {
1455 			    /* possible fragment truncation or extension */
1456 			    obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1457 			    ip->i_ffs1_blocks += (bb - obb);
1458 		    }
1459 		    ip->i_ffs1_db[lbn] = ndaddr;
1460 		    break;
1461 	    case 1:
1462 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1463 		    DEBUG_OOFF(1);
1464 		    if (ooff == UNWRITTEN)
1465 			    ip->i_ffs1_blocks += bb;
1466 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1467 		    break;
1468 	    default:
1469 		    ap = &a[num - 1];
1470 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1471 			    panic("lfs_updatemeta: bread bno %" PRId64,
1472 				  ap->in_lbn);
1473 
1474 		    /* XXX ondisk32 */
1475 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1476 		    DEBUG_OOFF(num);
1477 		    if (ooff == UNWRITTEN)
1478 			    ip->i_ffs1_blocks += bb;
1479 		    /* XXX ondisk32 */
1480 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1481 		    (void) VOP_BWRITE(bp);
1482 	}
1483 
1484 	KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr);
1485 
1486 	/* Update hiblk when extending the file */
1487 	if (lbn > ip->i_lfs_hiblk)
1488 		ip->i_lfs_hiblk = lbn;
1489 
1490 	/*
1491 	 * Though we'd rather it couldn't, this *can* happen right now
1492 	 * if cleaning blocks and regular blocks coexist.
1493 	 */
1494 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1495 
1496 	/*
1497 	 * Update segment usage information, based on old size
1498 	 * and location.
1499 	 */
1500 	if (daddr > 0) {
1501 		u_int32_t oldsn = dtosn(fs, daddr);
1502 #ifdef DIAGNOSTIC
1503 		int ndupino;
1504 
1505 		if (sp && sp->seg_number == oldsn) {
1506 			ndupino = sp->ndupino;
1507 		} else {
1508 			ndupino = 0;
1509 		}
1510 #endif
1511 		KASSERT(oldsn < fs->lfs_nseg);
1512 		if (lbn >= 0 && lbn < NDADDR)
1513 			osize = ip->i_lfs_fragsize[lbn];
1514 		else
1515 			osize = fs->lfs_bsize;
1516 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1517 #ifdef DIAGNOSTIC
1518 		if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino
1519 		    < osize) {
1520 			printf("lfs_updatemeta: negative bytes "
1521 			       "(segment %" PRIu32 " short by %" PRId64
1522 			       ")\n", dtosn(fs, daddr),
1523 			       (int64_t)osize -
1524 			       (sizeof (struct ufs1_dinode) * ndupino +
1525 				sup->su_nbytes));
1526 			printf("lfs_updatemeta: ino %llu, lbn %" PRId64
1527 			       ", addr = 0x%" PRIx64 "\n",
1528 			       (unsigned long long)ip->i_number, lbn, daddr);
1529 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1530 			panic("lfs_updatemeta: negative bytes");
1531 			sup->su_nbytes = osize -
1532 			    sizeof (struct ufs1_dinode) * ndupino;
1533 		}
1534 #endif
1535 		DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1536 		      " db 0x%" PRIx64 "\n",
1537 		      dtosn(fs, daddr), osize,
1538 		      ip->i_number, lbn, daddr));
1539 		sup->su_nbytes -= osize;
1540 		if (!(bp->b_flags & B_GATHERED)) {
1541 			simple_lock(&fs->lfs_interlock);
1542 			fs->lfs_flags |= LFS_IFDIRTY;
1543 			simple_unlock(&fs->lfs_interlock);
1544 		}
1545 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1546 	}
1547 	/*
1548 	 * Now that this block has a new address, and its old
1549 	 * segment no longer owns it, we can forget about its
1550 	 * old size.
1551 	 */
1552 	if (lbn >= 0 && lbn < NDADDR)
1553 		ip->i_lfs_fragsize[lbn] = size;
1554 }
1555 
1556 /*
1557  * Update the metadata that points to the blocks listed in the FINFO
1558  * array.
1559  */
1560 void
1561 lfs_updatemeta(struct segment *sp)
1562 {
1563 	struct buf *sbp;
1564 	struct lfs *fs;
1565 	struct vnode *vp;
1566 	daddr_t lbn;
1567 	int i, nblocks, num;
1568 	int bb;
1569 	int bytesleft, size;
1570 
1571 	ASSERT_SEGLOCK(sp->fs);
1572 	vp = sp->vp;
1573 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1574 	KASSERT(nblocks >= 0);
1575 	KASSERT(vp != NULL);
1576 	if (nblocks == 0)
1577 		return;
1578 
1579 	/*
1580 	 * This count may be high due to oversize blocks from lfs_gop_write.
1581 	 * Correct for this. (XXX we should be able to keep track of these.)
1582 	 */
1583 	fs = sp->fs;
1584 	for (i = 0; i < nblocks; i++) {
1585 		if (sp->start_bpp[i] == NULL) {
1586 			DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks));
1587 			nblocks = i;
1588 			break;
1589 		}
1590 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1591 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1592 		nblocks -= num - 1;
1593 	}
1594 
1595 	KASSERT(vp->v_type == VREG ||
1596 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1597 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1598 
1599 	/*
1600 	 * Sort the blocks.
1601 	 *
1602 	 * We have to sort even if the blocks come from the
1603 	 * cleaner, because there might be other pending blocks on the
1604 	 * same inode...and if we don't sort, and there are fragments
1605 	 * present, blocks may be written in the wrong place.
1606 	 */
1607 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1608 
1609 	/*
1610 	 * Record the length of the last block in case it's a fragment.
1611 	 * If there are indirect blocks present, they sort last.  An
1612 	 * indirect block will be lfs_bsize and its presence indicates
1613 	 * that you cannot have fragments.
1614 	 *
1615 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1616 	 * XXX a later indirect block.	This will continue to be
1617 	 * XXX true until lfs_markv is fixed to do everything with
1618 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1619 	 */
1620 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1621 		fs->lfs_bmask) + 1;
1622 
1623 	/*
1624 	 * Assign disk addresses, and update references to the logical
1625 	 * block and the segment usage information.
1626 	 */
1627 	for (i = nblocks; i--; ++sp->start_bpp) {
1628 		sbp = *sp->start_bpp;
1629 		lbn = *sp->start_lbp;
1630 		KASSERT(sbp->b_lblkno == lbn);
1631 
1632 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1633 
1634 		/*
1635 		 * If we write a frag in the wrong place, the cleaner won't
1636 		 * be able to correctly identify its size later, and the
1637 		 * segment will be uncleanable.	 (Even worse, it will assume
1638 		 * that the indirect block that actually ends the list
1639 		 * is of a smaller size!)
1640 		 */
1641 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1642 			panic("lfs_updatemeta: fragment is not last block");
1643 
1644 		/*
1645 		 * For each subblock in this possibly oversized block,
1646 		 * update its address on disk.
1647 		 */
1648 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1649 		KASSERT(vp == sbp->b_vp);
1650 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1651 		     bytesleft -= fs->lfs_bsize) {
1652 			size = MIN(bytesleft, fs->lfs_bsize);
1653 			bb = fragstofsb(fs, numfrags(fs, size));
1654 			lbn = *sp->start_lbp++;
1655 			lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset,
1656 			    size);
1657 			fs->lfs_offset += bb;
1658 		}
1659 
1660 	}
1661 
1662 	/* This inode has been modified */
1663 	LFS_SET_UINO(VTOI(vp), IN_MODIFIED);
1664 }
1665 
1666 /*
1667  * Move lfs_offset to a segment earlier than sn.
1668  */
1669 int
1670 lfs_rewind(struct lfs *fs, int newsn)
1671 {
1672 	int sn, osn, isdirty;
1673 	struct buf *bp;
1674 	SEGUSE *sup;
1675 
1676 	ASSERT_SEGLOCK(fs);
1677 
1678 	osn = dtosn(fs, fs->lfs_offset);
1679 	if (osn < newsn)
1680 		return 0;
1681 
1682 	/* lfs_avail eats the remaining space in this segment */
1683 	fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg);
1684 
1685 	/* Find a low-numbered segment */
1686 	for (sn = 0; sn < fs->lfs_nseg; ++sn) {
1687 		LFS_SEGENTRY(sup, fs, sn, bp);
1688 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1689 		brelse(bp, 0);
1690 
1691 		if (!isdirty)
1692 			break;
1693 	}
1694 	if (sn == fs->lfs_nseg)
1695 		panic("lfs_rewind: no clean segments");
1696 	if (newsn >= 0 && sn >= newsn)
1697 		return ENOENT;
1698 	fs->lfs_nextseg = sn;
1699 	lfs_newseg(fs);
1700 	fs->lfs_offset = fs->lfs_curseg;
1701 
1702 	return 0;
1703 }
1704 
1705 /*
1706  * Start a new partial segment.
1707  *
1708  * Return 1 when we entered to a new segment.
1709  * Otherwise, return 0.
1710  */
1711 int
1712 lfs_initseg(struct lfs *fs)
1713 {
1714 	struct segment *sp = fs->lfs_sp;
1715 	SEGSUM *ssp;
1716 	struct buf *sbp;	/* buffer for SEGSUM */
1717 	int repeat = 0;		/* return value */
1718 
1719 	ASSERT_SEGLOCK(fs);
1720 	/* Advance to the next segment. */
1721 	if (!LFS_PARTIAL_FITS(fs)) {
1722 		SEGUSE *sup;
1723 		struct buf *bp;
1724 
1725 		/* lfs_avail eats the remaining space */
1726 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1727 						   fs->lfs_curseg);
1728 		/* Wake up any cleaning procs waiting on this file system. */
1729 		lfs_wakeup_cleaner(fs);
1730 		lfs_newseg(fs);
1731 		repeat = 1;
1732 		fs->lfs_offset = fs->lfs_curseg;
1733 
1734 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1735 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1736 
1737 		/*
1738 		 * If the segment contains a superblock, update the offset
1739 		 * and summary address to skip over it.
1740 		 */
1741 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1742 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1743 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1744 			sp->seg_bytes_left -= LFS_SBPAD;
1745 		}
1746 		brelse(bp, 0);
1747 		/* Segment zero could also contain the labelpad */
1748 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1749 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1750 			fs->lfs_offset +=
1751 			    btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1752 			sp->seg_bytes_left -=
1753 			    LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1754 		}
1755 	} else {
1756 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1757 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1758 				      (fs->lfs_offset - fs->lfs_curseg));
1759 	}
1760 	fs->lfs_lastpseg = fs->lfs_offset;
1761 
1762 	/* Record first address of this partial segment */
1763 	if (sp->seg_flags & SEGM_CLEAN) {
1764 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1765 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1766 			/* "1" is the artificial inc in lfs_seglock */
1767 			simple_lock(&fs->lfs_interlock);
1768 			while (fs->lfs_iocount > 1) {
1769 				ltsleep(&fs->lfs_iocount, PRIBIO + 1,
1770 				    "lfs_initseg", 0, &fs->lfs_interlock);
1771 			}
1772 			simple_unlock(&fs->lfs_interlock);
1773 			fs->lfs_cleanind = 0;
1774 		}
1775 	}
1776 
1777 	sp->fs = fs;
1778 	sp->ibp = NULL;
1779 	sp->idp = NULL;
1780 	sp->ninodes = 0;
1781 	sp->ndupino = 0;
1782 
1783 	sp->cbpp = sp->bpp;
1784 
1785 	/* Get a new buffer for SEGSUM */
1786 	sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1787 	    fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1788 
1789 	/* ... and enter it into the buffer list. */
1790 	*sp->cbpp = sbp;
1791 	sp->cbpp++;
1792 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1793 
1794 	sp->start_bpp = sp->cbpp;
1795 
1796 	/* Set point to SEGSUM, initialize it. */
1797 	ssp = sp->segsum = sbp->b_data;
1798 	memset(ssp, 0, fs->lfs_sumsize);
1799 	ssp->ss_next = fs->lfs_nextseg;
1800 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1801 	ssp->ss_magic = SS_MAGIC;
1802 
1803 	/* Set pointer to first FINFO, initialize it. */
1804 	sp->fip = (struct finfo *)((char *)sp->segsum + SEGSUM_SIZE(fs));
1805 	sp->fip->fi_nblocks = 0;
1806 	sp->start_lbp = &sp->fip->fi_blocks[0];
1807 	sp->fip->fi_lastlength = 0;
1808 
1809 	sp->seg_bytes_left -= fs->lfs_sumsize;
1810 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1811 
1812 	return (repeat);
1813 }
1814 
1815 /*
1816  * Remove SEGUSE_INVAL from all segments.
1817  */
1818 void
1819 lfs_unset_inval_all(struct lfs *fs)
1820 {
1821 	SEGUSE *sup;
1822 	struct buf *bp;
1823 	int i;
1824 
1825 	for (i = 0; i < fs->lfs_nseg; i++) {
1826 		LFS_SEGENTRY(sup, fs, i, bp);
1827 		if (sup->su_flags & SEGUSE_INVAL) {
1828 			sup->su_flags &= ~SEGUSE_INVAL;
1829 			LFS_WRITESEGENTRY(sup, fs, i, bp);
1830 		} else
1831 			brelse(bp, 0);
1832 	}
1833 }
1834 
1835 /*
1836  * Return the next segment to write.
1837  */
1838 void
1839 lfs_newseg(struct lfs *fs)
1840 {
1841 	CLEANERINFO *cip;
1842 	SEGUSE *sup;
1843 	struct buf *bp;
1844 	int curseg, isdirty, sn, skip_inval;
1845 
1846 	ASSERT_SEGLOCK(fs);
1847 
1848 	/* Honor LFCNWRAPSTOP */
1849 	simple_lock(&fs->lfs_interlock);
1850 	while (fs->lfs_nextseg < fs->lfs_curseg && fs->lfs_nowrap) {
1851 		if (fs->lfs_wrappass) {
1852 			log(LOG_NOTICE, "%s: wrappass=%d\n",
1853 				fs->lfs_fsmnt, fs->lfs_wrappass);
1854 			fs->lfs_wrappass = 0;
1855 			break;
1856 		}
1857 		fs->lfs_wrapstatus = LFS_WRAP_WAITING;
1858 		wakeup(&fs->lfs_nowrap);
1859 		log(LOG_NOTICE, "%s: waiting at log wrap\n", fs->lfs_fsmnt);
1860 		ltsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz,
1861 			&fs->lfs_interlock);
1862 	}
1863 	fs->lfs_wrapstatus = LFS_WRAP_GOING;
1864 	simple_unlock(&fs->lfs_interlock);
1865 
1866 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1867 	DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n",
1868 	      dtosn(fs, fs->lfs_nextseg)));
1869 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1870 	sup->su_nbytes = 0;
1871 	sup->su_nsums = 0;
1872 	sup->su_ninos = 0;
1873 	LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1874 
1875 	LFS_CLEANERINFO(cip, fs, bp);
1876 	--cip->clean;
1877 	++cip->dirty;
1878 	fs->lfs_nclean = cip->clean;
1879 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1880 
1881 	fs->lfs_lastseg = fs->lfs_curseg;
1882 	fs->lfs_curseg = fs->lfs_nextseg;
1883 	skip_inval = 1;
1884 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1885 		sn = (sn + 1) % fs->lfs_nseg;
1886 
1887 		if (sn == curseg) {
1888 			if (skip_inval)
1889 				skip_inval = 0;
1890 			else
1891 				panic("lfs_nextseg: no clean segments");
1892 		}
1893 		LFS_SEGENTRY(sup, fs, sn, bp);
1894 		isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0));
1895 		/* Check SEGUSE_EMPTY as we go along */
1896 		if (isdirty && sup->su_nbytes == 0 &&
1897 		    !(sup->su_flags & SEGUSE_EMPTY))
1898 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1899 		else
1900 			brelse(bp, 0);
1901 
1902 		if (!isdirty)
1903 			break;
1904 	}
1905 	if (skip_inval == 0)
1906 		lfs_unset_inval_all(fs);
1907 
1908 	++fs->lfs_nactive;
1909 	fs->lfs_nextseg = sntod(fs, sn);
1910 	if (lfs_dostats) {
1911 		++lfs_stats.segsused;
1912 	}
1913 }
1914 
1915 static struct buf *
1916 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr,
1917     int n)
1918 {
1919 	struct lfs_cluster *cl;
1920 	struct buf **bpp, *bp;
1921 
1922 	ASSERT_SEGLOCK(fs);
1923 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1924 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1925 	memset(cl, 0, sizeof(*cl));
1926 	cl->fs = fs;
1927 	cl->bpp = bpp;
1928 	cl->bufcount = 0;
1929 	cl->bufsize = 0;
1930 
1931 	/* If this segment is being written synchronously, note that */
1932 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1933 		cl->flags |= LFS_CL_SYNC;
1934 		cl->seg = fs->lfs_sp;
1935 		++cl->seg->seg_iocount;
1936 	}
1937 
1938 	/* Get an empty buffer header, or maybe one with something on it */
1939 	bp = getiobuf();
1940 	bp->b_flags = B_BUSY | B_CALL;
1941 	bp->b_dev = NODEV;
1942 	bp->b_blkno = bp->b_lblkno = addr;
1943 	bp->b_iodone = lfs_cluster_callback;
1944 	bp->b_private = cl;
1945 	bp->b_vp = vp;
1946 
1947 	return bp;
1948 }
1949 
1950 int
1951 lfs_writeseg(struct lfs *fs, struct segment *sp)
1952 {
1953 	struct buf **bpp, *bp, *cbp, *newbp;
1954 	SEGUSE *sup;
1955 	SEGSUM *ssp;
1956 	int i, s;
1957 	int do_again, nblocks, byteoffset;
1958 	size_t el_size;
1959 	struct lfs_cluster *cl;
1960 	u_short ninos;
1961 	struct vnode *devvp;
1962 	char *p = NULL;
1963 	struct vnode *vp;
1964 	int32_t *daddrp;	/* XXX ondisk32 */
1965 	int changed;
1966 	u_int32_t sum;
1967 #ifdef DEBUG
1968 	FINFO *fip;
1969 	int findex;
1970 #endif
1971 
1972 	ASSERT_SEGLOCK(fs);
1973 
1974 	ssp = (SEGSUM *)sp->segsum;
1975 
1976 	/*
1977 	 * If there are no buffers other than the segment summary to write,
1978 	 * don't do anything.  If we are the end of a dirop sequence, however,
1979 	 * write the empty segment summary anyway, to help out the
1980 	 * roll-forward agent.
1981 	 */
1982 	if ((nblocks = sp->cbpp - sp->bpp) == 1) {
1983 		if ((ssp->ss_flags & (SS_DIROP | SS_CONT)) != SS_DIROP)
1984 			return 0;
1985 	}
1986 
1987 	/* Note if partial segment is being written by the cleaner */
1988 	if (sp->seg_flags & SEGM_CLEAN)
1989 		ssp->ss_flags |= SS_CLEAN;
1990 
1991 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1992 
1993 	/* Update the segment usage information. */
1994 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1995 
1996 	/* Loop through all blocks, except the segment summary. */
1997 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1998 		if ((*bpp)->b_vp != devvp) {
1999 			sup->su_nbytes += (*bpp)->b_bcount;
2000 			DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d"
2001 			      " lbn %" PRId64 " db 0x%" PRIx64 "\n",
2002 			      sp->seg_number, (*bpp)->b_bcount,
2003 			      VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
2004 			      (*bpp)->b_blkno));
2005 		}
2006 	}
2007 
2008 #ifdef DEBUG
2009 	/* Check for zero-length and zero-version FINFO entries. */
2010 	fip = (struct finfo *)((char *)ssp + SEGSUM_SIZE(fs));
2011 	for (findex = 0; findex < ssp->ss_nfinfo; findex++) {
2012 		KDASSERT(fip->fi_nblocks > 0);
2013 		KDASSERT(fip->fi_version > 0);
2014 		fip = (FINFO *)((char *)fip + FINFOSIZE +
2015 			sizeof(int32_t) * fip->fi_nblocks);
2016 	}
2017 #endif /* DEBUG */
2018 
2019 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
2020 	DLOG((DLOG_SU, "seg %d += %d for %d inodes\n",
2021 	      sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode),
2022 	      ssp->ss_ninos));
2023 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode);
2024 	/* sup->su_nbytes += fs->lfs_sumsize; */
2025 	if (fs->lfs_version == 1)
2026 		sup->su_olastmod = time_second;
2027 	else
2028 		sup->su_lastmod = time_second;
2029 	sup->su_ninos += ninos;
2030 	++sup->su_nsums;
2031 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
2032 
2033 	do_again = !(bp->b_flags & B_GATHERED);
2034 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
2035 
2036 	/*
2037 	 * Mark blocks B_BUSY, to prevent then from being changed between
2038 	 * the checksum computation and the actual write.
2039 	 *
2040 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
2041 	 * there are any, replace them with copies that have UNASSIGNED
2042 	 * instead.
2043 	 */
2044 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
2045 		++bpp;
2046 		bp = *bpp;
2047 		if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */
2048 			bp->b_flags |= B_BUSY;
2049 			continue;
2050 		}
2051 
2052 		simple_lock(&bp->b_interlock);
2053 		s = splbio();
2054 		while (bp->b_flags & B_BUSY) {
2055 			DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential"
2056 			      " data summary corruption for ino %d, lbn %"
2057 			      PRId64 "\n",
2058 			      VTOI(bp->b_vp)->i_number, bp->b_lblkno));
2059 			bp->b_flags |= B_WANTED;
2060 			ltsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0,
2061 				&bp->b_interlock);
2062 			splx(s);
2063 			s = splbio();
2064 		}
2065 		bp->b_flags |= B_BUSY;
2066 		splx(s);
2067 		simple_unlock(&bp->b_interlock);
2068 
2069 		/*
2070 		 * Check and replace indirect block UNWRITTEN bogosity.
2071 		 * XXX See comment in lfs_writefile.
2072 		 */
2073 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
2074 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
2075 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
2076 			DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n",
2077 			      VTOI(bp->b_vp)->i_number,
2078 			      VTOI(bp->b_vp)->i_lfs_effnblks,
2079 			      VTOI(bp->b_vp)->i_ffs1_blocks));
2080 			/* Make a copy we'll make changes to */
2081 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
2082 					   bp->b_bcount, LFS_NB_IBLOCK);
2083 			newbp->b_blkno = bp->b_blkno;
2084 			memcpy(newbp->b_data, bp->b_data,
2085 			       newbp->b_bcount);
2086 
2087 			changed = 0;
2088 			/* XXX ondisk32 */
2089 			for (daddrp = (int32_t *)(newbp->b_data);
2090 			     daddrp < (int32_t *)((char *)newbp->b_data +
2091 						  newbp->b_bcount); daddrp++) {
2092 				if (*daddrp == UNWRITTEN) {
2093 					++changed;
2094 					*daddrp = 0;
2095 				}
2096 			}
2097 			/*
2098 			 * Get rid of the old buffer.  Don't mark it clean,
2099 			 * though, if it still has dirty data on it.
2100 			 */
2101 			if (changed) {
2102 				DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):"
2103 				      " bp = %p newbp = %p\n", changed, bp,
2104 				      newbp));
2105 				*bpp = newbp;
2106 				bp->b_flags &= ~B_GATHERED;
2107 				bp->b_error = 0;
2108 				if (bp->b_flags & B_CALL) {
2109 					DLOG((DLOG_SEG, "lfs_writeseg: "
2110 					      "indir bp should not be B_CALL\n"));
2111 					s = splbio();
2112 					biodone(bp);
2113 					splx(s);
2114 					bp = NULL;
2115 				} else {
2116 					/* Still on free list, leave it there */
2117 					s = splbio();
2118 					bp->b_flags &= ~B_BUSY;
2119 					if (bp->b_flags & B_WANTED)
2120 						wakeup(bp);
2121 					splx(s);
2122 					/*
2123 					 * We have to re-decrement lfs_avail
2124 					 * since this block is going to come
2125 					 * back around to us in the next
2126 					 * segment.
2127 					 */
2128 					fs->lfs_avail -=
2129 					    btofsb(fs, bp->b_bcount);
2130 				}
2131 			} else {
2132 				lfs_freebuf(fs, newbp);
2133 			}
2134 		}
2135 	}
2136 	/*
2137 	 * Compute checksum across data and then across summary; the first
2138 	 * block (the summary block) is skipped.  Set the create time here
2139 	 * so that it's guaranteed to be later than the inode mod times.
2140 	 */
2141 	sum = 0;
2142 	if (fs->lfs_version == 1)
2143 		el_size = sizeof(u_long);
2144 	else
2145 		el_size = sizeof(u_int32_t);
2146 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
2147 		++bpp;
2148 		/* Loop through gop_write cluster blocks */
2149 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
2150 		     byteoffset += fs->lfs_bsize) {
2151 #ifdef LFS_USE_B_INVAL
2152 			if (((*bpp)->b_flags & (B_CALL | B_INVAL)) ==
2153 			    (B_CALL | B_INVAL)) {
2154 				if (copyin((void *)(*bpp)->b_saveaddr +
2155 					   byteoffset, dp, el_size)) {
2156 					panic("lfs_writeseg: copyin failed [1]:"
2157 						" ino %d blk %" PRId64,
2158 						VTOI((*bpp)->b_vp)->i_number,
2159 						(*bpp)->b_lblkno);
2160 				}
2161 			} else
2162 #endif /* LFS_USE_B_INVAL */
2163 			{
2164 				sum = lfs_cksum_part((char *)
2165 				    (*bpp)->b_data + byteoffset, el_size, sum);
2166 			}
2167 		}
2168 	}
2169 	if (fs->lfs_version == 1)
2170 		ssp->ss_ocreate = time_second;
2171 	else {
2172 		ssp->ss_create = time_second;
2173 		ssp->ss_serial = ++fs->lfs_serial;
2174 		ssp->ss_ident  = fs->lfs_ident;
2175 	}
2176 	ssp->ss_datasum = lfs_cksum_fold(sum);
2177 	ssp->ss_sumsum = cksum(&ssp->ss_datasum,
2178 	    fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
2179 
2180 	simple_lock(&fs->lfs_interlock);
2181 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
2182 			  btofsb(fs, fs->lfs_sumsize));
2183 	fs->lfs_dmeta += (btofsb(fs, ninos * fs->lfs_ibsize) +
2184 			  btofsb(fs, fs->lfs_sumsize));
2185 	simple_unlock(&fs->lfs_interlock);
2186 
2187 	/*
2188 	 * When we simply write the blocks we lose a rotation for every block
2189 	 * written.  To avoid this problem, we cluster the buffers into a
2190 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
2191 	 * devices can handle, use that for the size of the chunks.
2192 	 *
2193 	 * Blocks that are already clusters (from GOP_WRITE), however, we
2194 	 * don't bother to copy into other clusters.
2195 	 */
2196 
2197 #define CHUNKSIZE MAXPHYS
2198 
2199 	if (devvp == NULL)
2200 		panic("devvp is NULL");
2201 	for (bpp = sp->bpp, i = nblocks; i;) {
2202 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
2203 		cl = cbp->b_private;
2204 
2205 		cbp->b_flags |= B_ASYNC | B_BUSY;
2206 		cbp->b_bcount = 0;
2207 
2208 #if defined(DEBUG) && defined(DIAGNOSTIC)
2209 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
2210 		    / sizeof(int32_t)) {
2211 			panic("lfs_writeseg: real bpp overwrite");
2212 		}
2213 		if (bpp - sp->bpp > segsize(fs) / fs->lfs_fsize) {
2214 			panic("lfs_writeseg: theoretical bpp overwrite");
2215 		}
2216 #endif
2217 
2218 		/*
2219 		 * Construct the cluster.
2220 		 */
2221 		simple_lock(&fs->lfs_interlock);
2222 		++fs->lfs_iocount;
2223 		simple_unlock(&fs->lfs_interlock);
2224 		while (i && cbp->b_bcount < CHUNKSIZE) {
2225 			bp = *bpp;
2226 
2227 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
2228 				break;
2229 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
2230 				break;
2231 
2232 			/* Clusters from GOP_WRITE are expedited */
2233 			if (bp->b_bcount > fs->lfs_bsize) {
2234 				if (cbp->b_bcount > 0)
2235 					/* Put in its own buffer */
2236 					break;
2237 				else {
2238 					cbp->b_data = bp->b_data;
2239 				}
2240 			} else if (cbp->b_bcount == 0) {
2241 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
2242 							     LFS_NB_CLUSTER);
2243 				cl->flags |= LFS_CL_MALLOC;
2244 			}
2245 #ifdef DIAGNOSTIC
2246 			if (dtosn(fs, dbtofsb(fs, bp->b_blkno +
2247 					      btodb(bp->b_bcount - 1))) !=
2248 			    sp->seg_number) {
2249 				printf("blk size %d daddr %" PRIx64
2250 				    " not in seg %d\n",
2251 				    bp->b_bcount, bp->b_blkno,
2252 				    sp->seg_number);
2253 				panic("segment overwrite");
2254 			}
2255 #endif
2256 
2257 #ifdef LFS_USE_B_INVAL
2258 			/*
2259 			 * Fake buffers from the cleaner are marked as B_INVAL.
2260 			 * We need to copy the data from user space rather than
2261 			 * from the buffer indicated.
2262 			 * XXX == what do I do on an error?
2263 			 */
2264 			if ((bp->b_flags & (B_CALL|B_INVAL)) ==
2265 			    (B_CALL|B_INVAL)) {
2266 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
2267 					panic("lfs_writeseg: "
2268 					    "copyin failed [2]");
2269 			} else
2270 #endif /* LFS_USE_B_INVAL */
2271 			if (cl->flags & LFS_CL_MALLOC) {
2272 				/* copy data into our cluster. */
2273 				memcpy(p, bp->b_data, bp->b_bcount);
2274 				p += bp->b_bcount;
2275 			}
2276 
2277 			cbp->b_bcount += bp->b_bcount;
2278 			cl->bufsize += bp->b_bcount;
2279 
2280 			bp->b_flags &= ~(B_READ | B_DELWRI | B_DONE);
2281 			bp->b_error = 0;
2282 			cl->bpp[cl->bufcount++] = bp;
2283 			vp = bp->b_vp;
2284 			s = splbio();
2285 			reassignbuf(bp, vp);
2286 			V_INCR_NUMOUTPUT(vp);
2287 			splx(s);
2288 
2289 			bpp++;
2290 			i--;
2291 		}
2292 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2293 			BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL);
2294 		else
2295 			BIO_SETPRIO(cbp, BPRIO_TIMELIMITED);
2296 		s = splbio();
2297 		V_INCR_NUMOUTPUT(devvp);
2298 		splx(s);
2299 		VOP_STRATEGY(devvp, cbp);
2300 		curproc->p_stats->p_ru.ru_oublock++;
2301 	}
2302 
2303 	if (lfs_dostats) {
2304 		++lfs_stats.psegwrites;
2305 		lfs_stats.blocktot += nblocks - 1;
2306 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2307 			++lfs_stats.psyncwrites;
2308 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2309 			++lfs_stats.pcleanwrites;
2310 			lfs_stats.cleanblocks += nblocks - 1;
2311 		}
2312 	}
2313 
2314 	return (lfs_initseg(fs) || do_again);
2315 }
2316 
2317 void
2318 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2319 {
2320 	struct buf *bp;
2321 	int s;
2322 	struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2323 
2324 	ASSERT_MAYBE_SEGLOCK(fs);
2325 #ifdef DIAGNOSTIC
2326 	KASSERT(fs->lfs_magic == LFS_MAGIC);
2327 #endif
2328 	/*
2329 	 * If we can write one superblock while another is in
2330 	 * progress, we risk not having a complete checkpoint if we crash.
2331 	 * So, block here if a superblock write is in progress.
2332 	 */
2333 	simple_lock(&fs->lfs_interlock);
2334 	s = splbio();
2335 	while (fs->lfs_sbactive) {
2336 		ltsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0,
2337 			&fs->lfs_interlock);
2338 	}
2339 	fs->lfs_sbactive = daddr;
2340 	splx(s);
2341 	simple_unlock(&fs->lfs_interlock);
2342 
2343 	/* Set timestamp of this version of the superblock */
2344 	if (fs->lfs_version == 1)
2345 		fs->lfs_otstamp = time_second;
2346 	fs->lfs_tstamp = time_second;
2347 
2348 	/* Checksum the superblock and copy it into a buffer. */
2349 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2350 	bp = lfs_newbuf(fs, devvp,
2351 	    fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2352 	memset((char *)bp->b_data + sizeof(struct dlfs), 0,
2353 	    LFS_SBPAD - sizeof(struct dlfs));
2354 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2355 
2356 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2357 	bp->b_flags &= ~(B_DONE | B_READ | B_DELWRI);
2358 	bp->b_error = 0;
2359 	bp->b_iodone = lfs_supercallback;
2360 
2361 	if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC)
2362 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
2363 	else
2364 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
2365 	curproc->p_stats->p_ru.ru_oublock++;
2366 	s = splbio();
2367 	V_INCR_NUMOUTPUT(bp->b_vp);
2368 	splx(s);
2369 	simple_lock(&fs->lfs_interlock);
2370 	++fs->lfs_iocount;
2371 	simple_unlock(&fs->lfs_interlock);
2372 	VOP_STRATEGY(devvp, bp);
2373 }
2374 
2375 /*
2376  * Logical block number match routines used when traversing the dirty block
2377  * chain.
2378  */
2379 int
2380 lfs_match_fake(struct lfs *fs, struct buf *bp)
2381 {
2382 
2383 	ASSERT_SEGLOCK(fs);
2384 	return LFS_IS_MALLOC_BUF(bp);
2385 }
2386 
2387 #if 0
2388 int
2389 lfs_match_real(struct lfs *fs, struct buf *bp)
2390 {
2391 
2392 	ASSERT_SEGLOCK(fs);
2393 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2394 }
2395 #endif
2396 
2397 int
2398 lfs_match_data(struct lfs *fs, struct buf *bp)
2399 {
2400 
2401 	ASSERT_SEGLOCK(fs);
2402 	return (bp->b_lblkno >= 0);
2403 }
2404 
2405 int
2406 lfs_match_indir(struct lfs *fs, struct buf *bp)
2407 {
2408 	daddr_t lbn;
2409 
2410 	ASSERT_SEGLOCK(fs);
2411 	lbn = bp->b_lblkno;
2412 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2413 }
2414 
2415 int
2416 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2417 {
2418 	daddr_t lbn;
2419 
2420 	ASSERT_SEGLOCK(fs);
2421 	lbn = bp->b_lblkno;
2422 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2423 }
2424 
2425 int
2426 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2427 {
2428 	daddr_t lbn;
2429 
2430 	ASSERT_SEGLOCK(fs);
2431 	lbn = bp->b_lblkno;
2432 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2433 }
2434 
2435 static void
2436 lfs_free_aiodone(struct buf *bp)
2437 {
2438 	struct lfs *fs;
2439 
2440 	fs = bp->b_private;
2441 	ASSERT_NO_SEGLOCK(fs);
2442 	lfs_freebuf(fs, bp);
2443 }
2444 
2445 static void
2446 lfs_super_aiodone(struct buf *bp)
2447 {
2448 	struct lfs *fs;
2449 
2450 	fs = bp->b_private;
2451 	ASSERT_NO_SEGLOCK(fs);
2452 	simple_lock(&fs->lfs_interlock);
2453 	fs->lfs_sbactive = 0;
2454 	if (--fs->lfs_iocount <= 1)
2455 		wakeup(&fs->lfs_iocount);
2456 	simple_unlock(&fs->lfs_interlock);
2457 	wakeup(&fs->lfs_sbactive);
2458 	lfs_freebuf(fs, bp);
2459 }
2460 
2461 static void
2462 lfs_cluster_aiodone(struct buf *bp)
2463 {
2464 	struct lfs_cluster *cl;
2465 	struct lfs *fs;
2466 	struct buf *tbp, *fbp;
2467 	struct vnode *vp, *devvp;
2468 	struct inode *ip;
2469 	int s, error;
2470 
2471 	error = bp->b_error;
2472 	cl = bp->b_private;
2473 	fs = cl->fs;
2474 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2475 	ASSERT_NO_SEGLOCK(fs);
2476 
2477 	/* Put the pages back, and release the buffer */
2478 	while (cl->bufcount--) {
2479 		tbp = cl->bpp[cl->bufcount];
2480 		KASSERT(tbp->b_flags & B_BUSY);
2481 		if (error) {
2482 			tbp->b_error = error;
2483 		}
2484 
2485 		/*
2486 		 * We're done with tbp.	 If it has not been re-dirtied since
2487 		 * the cluster was written, free it.  Otherwise, keep it on
2488 		 * the locked list to be written again.
2489 		 */
2490 		vp = tbp->b_vp;
2491 
2492 		tbp->b_flags &= ~B_GATHERED;
2493 
2494 		LFS_BCLEAN_LOG(fs, tbp);
2495 
2496 		if (!(tbp->b_flags & B_CALL)) {
2497 			KASSERT(tbp->b_flags & B_LOCKED);
2498 			s = splbio();
2499 			simple_lock(&bqueue_slock);
2500 			bremfree(tbp);
2501 			simple_unlock(&bqueue_slock);
2502 			if (vp)
2503 				reassignbuf(tbp, vp);
2504 			splx(s);
2505 			tbp->b_flags |= B_ASYNC; /* for biodone */
2506 		}
2507 
2508 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2509 			LFS_UNLOCK_BUF(tbp);
2510 
2511 		if (tbp->b_flags & B_DONE) {
2512 			DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n",
2513 				cl->bufcount, (long)tbp->b_flags));
2514 		}
2515 
2516 		if ((tbp->b_flags & B_CALL) && !LFS_IS_MALLOC_BUF(tbp)) {
2517 			/*
2518 			 * A buffer from the page daemon.
2519 			 * We use the same iodone as it does,
2520 			 * so we must manually disassociate its
2521 			 * buffers from the vp.
2522 			 */
2523 			if (tbp->b_vp) {
2524 				/* This is just silly */
2525 				s = splbio();
2526 				brelvp(tbp);
2527 				tbp->b_vp = vp;
2528 				splx(s);
2529 			}
2530 			/* Put it back the way it was */
2531 			tbp->b_flags |= B_ASYNC;
2532 			/* Master buffers have B_AGE */
2533 			if (tbp->b_private == tbp)
2534 				tbp->b_flags |= B_AGE;
2535 		}
2536 		s = splbio();
2537 		biodone(tbp);
2538 
2539 		/*
2540 		 * If this is the last block for this vnode, but
2541 		 * there are other blocks on its dirty list,
2542 		 * set IN_MODIFIED/IN_CLEANING depending on what
2543 		 * sort of block.  Only do this for our mount point,
2544 		 * not for, e.g., inode blocks that are attached to
2545 		 * the devvp.
2546 		 * XXX KS - Shouldn't we set *both* if both types
2547 		 * of blocks are present (traverse the dirty list?)
2548 		 */
2549 		simple_lock(&global_v_numoutput_slock);
2550 		if (vp != devvp && vp->v_numoutput == 0 &&
2551 		    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2552 			ip = VTOI(vp);
2553 			DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n",
2554 			       ip->i_number));
2555 			if (LFS_IS_MALLOC_BUF(fbp))
2556 				LFS_SET_UINO(ip, IN_CLEANING);
2557 			else
2558 				LFS_SET_UINO(ip, IN_MODIFIED);
2559 		}
2560 		simple_unlock(&global_v_numoutput_slock);
2561 		splx(s);
2562 		wakeup(vp);
2563 	}
2564 
2565 	/* Fix up the cluster buffer, and release it */
2566 	if (cl->flags & LFS_CL_MALLOC)
2567 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2568 	putiobuf(bp);
2569 
2570 	/* Note i/o done */
2571 	if (cl->flags & LFS_CL_SYNC) {
2572 		if (--cl->seg->seg_iocount == 0)
2573 			wakeup(&cl->seg->seg_iocount);
2574 	}
2575 	simple_lock(&fs->lfs_interlock);
2576 #ifdef DIAGNOSTIC
2577 	if (fs->lfs_iocount == 0)
2578 		panic("lfs_cluster_aiodone: zero iocount");
2579 #endif
2580 	if (--fs->lfs_iocount <= 1)
2581 		wakeup(&fs->lfs_iocount);
2582 	simple_unlock(&fs->lfs_interlock);
2583 
2584 	pool_put(&fs->lfs_bpppool, cl->bpp);
2585 	cl->bpp = NULL;
2586 	pool_put(&fs->lfs_clpool, cl);
2587 }
2588 
2589 static void
2590 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2591 {
2592 	/* reset b_iodone for when this is a single-buf i/o. */
2593 	bp->b_iodone = aiodone;
2594 
2595 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
2596 }
2597 
2598 static void
2599 lfs_cluster_callback(struct buf *bp)
2600 {
2601 
2602 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2603 }
2604 
2605 void
2606 lfs_supercallback(struct buf *bp)
2607 {
2608 
2609 	lfs_generic_callback(bp, lfs_super_aiodone);
2610 }
2611 
2612 /*
2613  * The only buffers that are going to hit these functions are the
2614  * segment write blocks, or the segment summaries, or the superblocks.
2615  *
2616  * All of the above are created by lfs_newbuf, and so do not need to be
2617  * released via brelse.
2618  */
2619 void
2620 lfs_callback(struct buf *bp)
2621 {
2622 
2623 	lfs_generic_callback(bp, lfs_free_aiodone);
2624 }
2625 
2626 /*
2627  * Shellsort (diminishing increment sort) from Data Structures and
2628  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2629  * see also Knuth Vol. 3, page 84.  The increments are selected from
2630  * formula (8), page 95.  Roughly O(N^3/2).
2631  */
2632 /*
2633  * This is our own private copy of shellsort because we want to sort
2634  * two parallel arrays (the array of buffer pointers and the array of
2635  * logical block numbers) simultaneously.  Note that we cast the array
2636  * of logical block numbers to a unsigned in this routine so that the
2637  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2638  */
2639 
2640 void
2641 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2642 {
2643 	static int __rsshell_increments[] = { 4, 1, 0 };
2644 	int incr, *incrp, t1, t2;
2645 	struct buf *bp_temp;
2646 
2647 #ifdef DEBUG
2648 	incr = 0;
2649 	for (t1 = 0; t1 < nmemb; t1++) {
2650 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2651 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2652 				/* dump before panic */
2653 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2654 				    nmemb, size);
2655 				incr = 0;
2656 				for (t1 = 0; t1 < nmemb; t1++) {
2657 					const struct buf *bp = bp_array[t1];
2658 
2659 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2660 					    PRIu64 "\n", t1,
2661 					    (uint64_t)bp->b_bcount,
2662 					    (uint64_t)bp->b_lblkno);
2663 					printf("lbns:");
2664 					for (t2 = 0; t2 * size < bp->b_bcount;
2665 					    t2++) {
2666 						printf(" %" PRId32,
2667 						    lb_array[incr++]);
2668 					}
2669 					printf("\n");
2670 				}
2671 				panic("lfs_shellsort: inconsistent input");
2672 			}
2673 		}
2674 	}
2675 #endif
2676 
2677 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2678 		for (t1 = incr; t1 < nmemb; ++t1)
2679 			for (t2 = t1 - incr; t2 >= 0;)
2680 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2681 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2682 					bp_temp = bp_array[t2];
2683 					bp_array[t2] = bp_array[t2 + incr];
2684 					bp_array[t2 + incr] = bp_temp;
2685 					t2 -= incr;
2686 				} else
2687 					break;
2688 
2689 	/* Reform the list of logical blocks */
2690 	incr = 0;
2691 	for (t1 = 0; t1 < nmemb; t1++) {
2692 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2693 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2694 		}
2695 	}
2696 }
2697 
2698 /*
2699  * Call vget with LK_NOWAIT.  If we are the one who holds VI_XLOCK/VI_FREEING,
2700  * however, we must press on.  Just fake success in that case.
2701  */
2702 int
2703 lfs_vref(struct vnode *vp)
2704 {
2705 	int error;
2706 	struct lfs *fs;
2707 
2708 	fs = VTOI(vp)->i_lfs;
2709 
2710 	ASSERT_MAYBE_SEGLOCK(fs);
2711 
2712 	/*
2713 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2714 	 * being able to flush all of the pages from this vnode, which
2715 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2716 	 */
2717 	error = vget(vp, LK_NOWAIT);
2718 	if (error == EBUSY && IS_FLUSHING(VTOI(vp)->i_lfs, vp)) {
2719 		++fs->lfs_flushvp_fakevref;
2720 		return 0;
2721 	}
2722 	return error;
2723 }
2724 
2725 /*
2726  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2727  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2728  */
2729 void
2730 lfs_vunref(struct vnode *vp)
2731 {
2732 	struct lfs *fs;
2733 
2734 	fs = VTOI(vp)->i_lfs;
2735 	ASSERT_MAYBE_SEGLOCK(fs);
2736 
2737 	/*
2738 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2739 	 */
2740 	if (IS_FLUSHING(fs, vp) && fs->lfs_flushvp_fakevref) {
2741 		--fs->lfs_flushvp_fakevref;
2742 		return;
2743 	}
2744 
2745 	/* does not call inactive */
2746 	vrele2(vp, 0);
2747 }
2748 
2749 /*
2750  * We use this when we have vnodes that were loaded in solely for cleaning.
2751  * There is no reason to believe that these vnodes will be referenced again
2752  * soon, since the cleaning process is unrelated to normal filesystem
2753  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2754  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2755  * cleaning at the head of the list, instead.
2756  */
2757 void
2758 lfs_vunref_head(struct vnode *vp)
2759 {
2760 
2761 	ASSERT_SEGLOCK(VTOI(vp)->i_lfs);
2762 
2763 	/* does not call inactive, inserts non-held vnode at head of freelist */
2764 	vrele2(vp, 1);
2765 }
2766 
2767 
2768 /*
2769  * Set up an FINFO entry for a new file.  The fip pointer is assumed to
2770  * point at uninitialized space.
2771  */
2772 void
2773 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers)
2774 {
2775 	struct segment *sp = fs->lfs_sp;
2776 
2777 	KASSERT(vers > 0);
2778 
2779 	if (sp->seg_bytes_left < fs->lfs_bsize ||
2780 	    sp->sum_bytes_left < sizeof(struct finfo))
2781 		(void) lfs_writeseg(fs, fs->lfs_sp);
2782 
2783 	sp->sum_bytes_left -= FINFOSIZE;
2784 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
2785 	sp->fip->fi_nblocks = 0;
2786 	sp->fip->fi_ino = ino;
2787 	sp->fip->fi_version = vers;
2788 }
2789 
2790 /*
2791  * Release the FINFO entry, either clearing out an unused entry or
2792  * advancing us to the next available entry.
2793  */
2794 void
2795 lfs_release_finfo(struct lfs *fs)
2796 {
2797 	struct segment *sp = fs->lfs_sp;
2798 
2799 	if (sp->fip->fi_nblocks != 0) {
2800 		sp->fip = (FINFO*)((char *)sp->fip + FINFOSIZE +
2801 			sizeof(int32_t) * sp->fip->fi_nblocks);
2802 		sp->start_lbp = &sp->fip->fi_blocks[0];
2803 	} else {
2804 		sp->sum_bytes_left += FINFOSIZE;
2805 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
2806 	}
2807 }
2808