xref: /minix3/sys/ufs/lfs/lfs_segment.c (revision 0b98e8aad89f2bd4ba80b523d73cf29e9dd82ce1)
1 /*	$NetBSD: lfs_segment.c,v 1.233 2013/10/29 09:53:51 hannken 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 1.233 2013/10/29 09:53:51 hannken 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_kernel.h>
104 #include <ufs/lfs/lfs_extern.h>
105 
106 #include <uvm/uvm.h>
107 #include <uvm/uvm_extern.h>
108 
109 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
110 
111 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
112 static void lfs_free_aiodone(struct buf *);
113 static void lfs_super_aiodone(struct buf *);
114 static void lfs_cluster_aiodone(struct buf *);
115 static void lfs_cluster_callback(struct buf *);
116 
117 /*
118  * Determine if it's OK to start a partial in this segment, or if we need
119  * to go on to a new segment.
120  */
121 #define	LFS_PARTIAL_FITS(fs) \
122 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
123 	(fs)->lfs_frag)
124 
125 /*
126  * Figure out whether we should do a checkpoint write or go ahead with
127  * an ordinary write.
128  */
129 #define LFS_SHOULD_CHECKPOINT(fs, flags) \
130         ((flags & SEGM_CLEAN) == 0 &&					\
131 	  ((fs->lfs_nactive > LFS_MAX_ACTIVE ||				\
132 	    (flags & SEGM_CKP) ||					\
133 	    fs->lfs_nclean < LFS_MAX_ACTIVE)))
134 
135 int	 lfs_match_fake(struct lfs *, struct buf *);
136 void	 lfs_newseg(struct lfs *);
137 /* XXX ondisk32 */
138 void	 lfs_shellsort(struct buf **, int32_t *, int, int);
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 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
146 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
147 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
148 int	lfs_dirvcount = 0;		/* # active dirops */
149 
150 /* Statistics Counters */
151 int lfs_dostats = 1;
152 struct lfs_stats lfs_stats;
153 
154 /* op values to lfs_writevnodes */
155 #define	VN_REG		0
156 #define	VN_DIROP	1
157 #define	VN_EMPTY	2
158 #define VN_CLEAN	3
159 
160 /*
161  * XXX KS - Set modification time on the Ifile, so the cleaner can
162  * read the fs mod time off of it.  We don't set IN_UPDATE here,
163  * since we don't really need this to be flushed to disk (and in any
164  * case that wouldn't happen to the Ifile until we checkpoint).
165  */
166 void
167 lfs_imtime(struct lfs *fs)
168 {
169 	struct timespec ts;
170 	struct inode *ip;
171 
172 	ASSERT_MAYBE_SEGLOCK(fs);
173 	vfs_timestamp(&ts);
174 	ip = VTOI(fs->lfs_ivnode);
175 	ip->i_ffs1_mtime = ts.tv_sec;
176 	ip->i_ffs1_mtimensec = ts.tv_nsec;
177 }
178 
179 /*
180  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
181  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
182  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
183  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
184  */
185 
186 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
187 
188 int
189 lfs_vflush(struct vnode *vp)
190 {
191 	struct inode *ip;
192 	struct lfs *fs;
193 	struct segment *sp;
194 	struct buf *bp, *nbp, *tbp, *tnbp;
195 	int error;
196 	int flushed;
197 	int relock;
198 
199 	ip = VTOI(vp);
200 	fs = VFSTOULFS(vp->v_mount)->um_lfs;
201 	relock = 0;
202 
203     top:
204 	KASSERT(mutex_owned(vp->v_interlock) == false);
205 	KASSERT(mutex_owned(&lfs_lock) == false);
206 	KASSERT(mutex_owned(&bufcache_lock) == false);
207 	ASSERT_NO_SEGLOCK(fs);
208 	if (ip->i_flag & IN_CLEANING) {
209 		ivndebug(vp,"vflush/in_cleaning");
210 		mutex_enter(&lfs_lock);
211 		LFS_CLR_UINO(ip, IN_CLEANING);
212 		LFS_SET_UINO(ip, IN_MODIFIED);
213 		mutex_exit(&lfs_lock);
214 
215 		/*
216 		 * Toss any cleaning buffers that have real counterparts
217 		 * to avoid losing new data.
218 		 */
219 		mutex_enter(vp->v_interlock);
220 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
221 			nbp = LIST_NEXT(bp, b_vnbufs);
222 			if (!LFS_IS_MALLOC_BUF(bp))
223 				continue;
224 			/*
225 			 * Look for pages matching the range covered
226 			 * by cleaning blocks.  It's okay if more dirty
227 			 * pages appear, so long as none disappear out
228 			 * from under us.
229 			 */
230 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
231 			    vp != fs->lfs_ivnode) {
232 				struct vm_page *pg;
233 				voff_t off;
234 
235 				for (off = lfs_lblktosize(fs, bp->b_lblkno);
236 				     off < lfs_lblktosize(fs, bp->b_lblkno + 1);
237 				     off += PAGE_SIZE) {
238 					pg = uvm_pagelookup(&vp->v_uobj, off);
239 					if (pg == NULL)
240 						continue;
241 					if ((pg->flags & PG_CLEAN) == 0 ||
242 					    pmap_is_modified(pg)) {
243 						fs->lfs_avail += lfs_btofsb(fs,
244 							bp->b_bcount);
245 						wakeup(&fs->lfs_avail);
246 						mutex_exit(vp->v_interlock);
247 						lfs_freebuf(fs, bp);
248 						mutex_enter(vp->v_interlock);
249 						bp = NULL;
250 						break;
251 					}
252 				}
253 			}
254 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
255 			    tbp = tnbp)
256 			{
257 				tnbp = LIST_NEXT(tbp, b_vnbufs);
258 				if (tbp->b_vp == bp->b_vp
259 				   && tbp->b_lblkno == bp->b_lblkno
260 				   && tbp != bp)
261 				{
262 					fs->lfs_avail += lfs_btofsb(fs,
263 						bp->b_bcount);
264 					wakeup(&fs->lfs_avail);
265 					mutex_exit(vp->v_interlock);
266 					lfs_freebuf(fs, bp);
267 					mutex_enter(vp->v_interlock);
268 					bp = NULL;
269 					break;
270 				}
271 			}
272 		}
273 	} else {
274 		mutex_enter(vp->v_interlock);
275 	}
276 
277 	/* If the node is being written, wait until that is done */
278 	while (WRITEINPROG(vp)) {
279 		ivndebug(vp,"vflush/writeinprog");
280 		cv_wait(&vp->v_cv, vp->v_interlock);
281 	}
282 	mutex_exit(vp->v_interlock);
283 
284 	/* Protect against VI_XLOCK deadlock in vinvalbuf() */
285 	lfs_seglock(fs, SEGM_SYNC | ((vp->v_iflag & VI_XLOCK) ? SEGM_RECLAIM : 0));
286 	if (vp->v_iflag & VI_XLOCK) {
287 		fs->lfs_reclino = ip->i_number;
288 	}
289 
290 	/* If we're supposed to flush a freed inode, just toss it */
291 	if (ip->i_lfs_iflags & LFSI_DELETED) {
292 		DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n",
293 		      ip->i_number));
294 		/* Drain v_numoutput */
295 		mutex_enter(vp->v_interlock);
296 		while (vp->v_numoutput > 0) {
297 			cv_wait(&vp->v_cv, vp->v_interlock);
298 		}
299 		KASSERT(vp->v_numoutput == 0);
300 		mutex_exit(vp->v_interlock);
301 
302 		mutex_enter(&bufcache_lock);
303 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
304 			nbp = LIST_NEXT(bp, b_vnbufs);
305 
306 			KASSERT((bp->b_flags & B_GATHERED) == 0);
307 			if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */
308 				fs->lfs_avail += lfs_btofsb(fs, bp->b_bcount);
309 				wakeup(&fs->lfs_avail);
310 			}
311 			/* Copied from lfs_writeseg */
312 			if (bp->b_iodone != NULL) {
313 				mutex_exit(&bufcache_lock);
314 				biodone(bp);
315 				mutex_enter(&bufcache_lock);
316 			} else {
317 				bremfree(bp);
318 				LFS_UNLOCK_BUF(bp);
319 				mutex_enter(vp->v_interlock);
320 				bp->b_flags &= ~(B_READ | B_GATHERED);
321 				bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE;
322 				bp->b_error = 0;
323 				reassignbuf(bp, vp);
324 				mutex_exit(vp->v_interlock);
325 				brelse(bp, 0);
326 			}
327 		}
328 		mutex_exit(&bufcache_lock);
329 		LFS_CLR_UINO(ip, IN_CLEANING);
330 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
331 		ip->i_flag &= ~IN_ALLMOD;
332 		DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n",
333 		      ip->i_number));
334 		lfs_segunlock(fs);
335 
336 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
337 
338 		return 0;
339 	}
340 
341 	fs->lfs_flushvp = vp;
342 	if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) {
343 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
344 		fs->lfs_flushvp = NULL;
345 		KASSERT(fs->lfs_flushvp_fakevref == 0);
346 		lfs_segunlock(fs);
347 
348 		/* Make sure that any pending buffers get written */
349 		mutex_enter(vp->v_interlock);
350 		while (vp->v_numoutput > 0) {
351 			cv_wait(&vp->v_cv, vp->v_interlock);
352 		}
353 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
354 		KASSERT(vp->v_numoutput == 0);
355 		mutex_exit(vp->v_interlock);
356 
357 		return error;
358 	}
359 	sp = fs->lfs_sp;
360 
361 	flushed = 0;
362 	if (VPISEMPTY(vp)) {
363 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
364 		++flushed;
365 	} else if ((ip->i_flag & IN_CLEANING) &&
366 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
367 		ivndebug(vp,"vflush/clean");
368 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
369 		++flushed;
370 	} else if (lfs_dostats) {
371 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
372 			++lfs_stats.vflush_invoked;
373 		ivndebug(vp,"vflush");
374 	}
375 
376 #ifdef DIAGNOSTIC
377 	if (vp->v_uflag & VU_DIROP) {
378 		DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n"));
379 		/* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */
380 	}
381 #endif
382 
383 	do {
384 #ifdef DEBUG
385 		int loopcount = 0;
386 #endif
387 		do {
388 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
389 				relock = lfs_writefile(fs, sp, vp);
390 				if (relock && vp != fs->lfs_ivnode) {
391 					/*
392 					 * Might have to wait for the
393 					 * cleaner to run; but we're
394 					 * still not done with this vnode.
395 					 * XXX we can do better than this.
396 					 */
397 					KDASSERT(ip->i_number != LFS_IFILE_INUM);
398 					lfs_writeinode(fs, sp, ip);
399 					mutex_enter(&lfs_lock);
400 					LFS_SET_UINO(ip, IN_MODIFIED);
401 					mutex_exit(&lfs_lock);
402 					lfs_writeseg(fs, sp);
403 					lfs_segunlock(fs);
404 					lfs_segunlock_relock(fs);
405 					goto top;
406 				}
407 			}
408 			/*
409 			 * If we begin a new segment in the middle of writing
410 			 * the Ifile, it creates an inconsistent checkpoint,
411 			 * since the Ifile information for the new segment
412 			 * is not up-to-date.  Take care of this here by
413 			 * sending the Ifile through again in case there
414 			 * are newly dirtied blocks.  But wait, there's more!
415 			 * This second Ifile write could *also* cross a segment
416 			 * boundary, if the first one was large.  The second
417 			 * one is guaranteed to be no more than 8 blocks,
418 			 * though (two segment blocks and supporting indirects)
419 			 * so the third write *will not* cross the boundary.
420 			 */
421 			if (vp == fs->lfs_ivnode) {
422 				lfs_writefile(fs, sp, vp);
423 				lfs_writefile(fs, sp, vp);
424 			}
425 #ifdef DEBUG
426 			if (++loopcount > 2)
427 				log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount);
428 #endif
429 		} while (lfs_writeinode(fs, sp, ip));
430 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
431 
432 	if (lfs_dostats) {
433 		++lfs_stats.nwrites;
434 		if (sp->seg_flags & SEGM_SYNC)
435 			++lfs_stats.nsync_writes;
436 		if (sp->seg_flags & SEGM_CKP)
437 			++lfs_stats.ncheckpoints;
438 	}
439 	/*
440 	 * If we were called from somewhere that has already held the seglock
441 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
442 	 * the write to complete because we are still locked.
443 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
444 	 * we must explicitly wait, if that is the case.
445 	 *
446 	 * We compare the iocount against 1, not 0, because it is
447 	 * artificially incremented by lfs_seglock().
448 	 */
449 	mutex_enter(&lfs_lock);
450 	if (fs->lfs_seglock > 1) {
451 		while (fs->lfs_iocount > 1)
452 			(void)mtsleep(&fs->lfs_iocount, PRIBIO + 1,
453 				     "lfs_vflush", 0, &lfs_lock);
454 	}
455 	mutex_exit(&lfs_lock);
456 
457 	lfs_segunlock(fs);
458 
459 	/* Wait for these buffers to be recovered by aiodoned */
460 	mutex_enter(vp->v_interlock);
461 	while (vp->v_numoutput > 0) {
462 		cv_wait(&vp->v_cv, vp->v_interlock);
463 	}
464 	KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
465 	KASSERT(vp->v_numoutput == 0);
466 	mutex_exit(vp->v_interlock);
467 
468 	fs->lfs_flushvp = NULL;
469 	KASSERT(fs->lfs_flushvp_fakevref == 0);
470 
471 	return (0);
472 }
473 
474 int
475 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
476 {
477 	struct inode *ip;
478 	struct vnode *vp;
479 	int inodes_written = 0, only_cleaning;
480 	int error = 0;
481 
482 	ASSERT_SEGLOCK(fs);
483  loop:
484 	/* start at last (newest) vnode. */
485 	mutex_enter(&mntvnode_lock);
486 	TAILQ_FOREACH_REVERSE(vp, &mp->mnt_vnodelist, vnodelst, v_mntvnodes) {
487 		/*
488 		 * If the vnode that we are about to sync is no longer
489 		 * associated with this mount point, start over.
490 		 */
491 		if (vp->v_mount != mp) {
492 			DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n"));
493 			/*
494 			 * After this, pages might be busy
495 			 * due to our own previous putpages.
496 			 * Start actual segment write here to avoid deadlock.
497 			 * If we were just writing one segment and we've done
498 			 * that, break out.
499 			 */
500 			mutex_exit(&mntvnode_lock);
501 			if (lfs_writeseg(fs, sp) &&
502 			    (sp->seg_flags & SEGM_SINGLE) &&
503 			    fs->lfs_curseg != fs->lfs_startseg) {
504 				DLOG((DLOG_VNODE, "lfs_writevnodes: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset));
505 				break;
506 			}
507 			goto loop;
508 		}
509 
510 		mutex_enter(vp->v_interlock);
511 		if (vp->v_type == VNON || vismarker(vp) ||
512 		    (vp->v_iflag & VI_CLEAN) != 0) {
513 			mutex_exit(vp->v_interlock);
514 			continue;
515 		}
516 
517 		ip = VTOI(vp);
518 		if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) ||
519 		    (op != VN_DIROP && op != VN_CLEAN &&
520 		    (vp->v_uflag & VU_DIROP))) {
521 			mutex_exit(vp->v_interlock);
522 			vndebug(vp,"dirop");
523 			continue;
524 		}
525 
526 		if (op == VN_EMPTY && !VPISEMPTY(vp)) {
527 			mutex_exit(vp->v_interlock);
528 			vndebug(vp,"empty");
529 			continue;
530 		}
531 
532 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
533 		   && vp != fs->lfs_flushvp
534 		   && !(ip->i_flag & IN_CLEANING)) {
535 			mutex_exit(vp->v_interlock);
536 			vndebug(vp,"cleaning");
537 			continue;
538 		}
539 
540 		mutex_exit(&mntvnode_lock);
541 		if (lfs_vref(vp)) {
542 			vndebug(vp,"vref");
543 			mutex_enter(&mntvnode_lock);
544 			continue;
545 		}
546 
547 		only_cleaning = 0;
548 		/*
549 		 * Write the inode/file if dirty and it's not the IFILE.
550 		 */
551 		if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) {
552 			only_cleaning =
553 			    ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
554 
555 			if (ip->i_number != LFS_IFILE_INUM) {
556 				error = lfs_writefile(fs, sp, vp);
557 				if (error) {
558 					lfs_vunref(vp);
559 					if (error == EAGAIN) {
560 						/*
561 						 * This error from lfs_putpages
562 						 * indicates we need to drop
563 						 * the segment lock and start
564 						 * over after the cleaner has
565 						 * had a chance to run.
566 						 */
567 						lfs_writeinode(fs, sp, ip);
568 						lfs_writeseg(fs, sp);
569 						if (!VPISEMPTY(vp) &&
570 						    !WRITEINPROG(vp) &&
571 						    !(ip->i_flag & IN_ALLMOD)) {
572 							mutex_enter(&lfs_lock);
573 							LFS_SET_UINO(ip, IN_MODIFIED);
574 							mutex_exit(&lfs_lock);
575 						}
576 						mutex_enter(&mntvnode_lock);
577 						break;
578 					}
579 					error = 0; /* XXX not quite right */
580 					mutex_enter(&mntvnode_lock);
581 					continue;
582 				}
583 
584 				if (!VPISEMPTY(vp)) {
585 					if (WRITEINPROG(vp)) {
586 						ivndebug(vp,"writevnodes/write2");
587 					} else if (!(ip->i_flag & IN_ALLMOD)) {
588 						mutex_enter(&lfs_lock);
589 						LFS_SET_UINO(ip, IN_MODIFIED);
590 						mutex_exit(&lfs_lock);
591 					}
592 				}
593 				(void) lfs_writeinode(fs, sp, ip);
594 				inodes_written++;
595 			}
596 		}
597 
598 		if (lfs_clean_vnhead && only_cleaning)
599 			lfs_vunref_head(vp);
600 		else
601 			lfs_vunref(vp);
602 
603 		mutex_enter(&mntvnode_lock);
604 	}
605 	mutex_exit(&mntvnode_lock);
606 	return error;
607 }
608 
609 /*
610  * Do a checkpoint.
611  */
612 int
613 lfs_segwrite(struct mount *mp, int flags)
614 {
615 	struct buf *bp;
616 	struct inode *ip;
617 	struct lfs *fs;
618 	struct segment *sp;
619 	struct vnode *vp;
620 	SEGUSE *segusep;
621 	int do_ckp, did_ckp, error;
622 	unsigned n, segleft, maxseg, sn, i, curseg;
623 	int writer_set = 0;
624 	int dirty;
625 	int redo;
626 	int um_error;
627 
628 	fs = VFSTOULFS(mp)->um_lfs;
629 	ASSERT_MAYBE_SEGLOCK(fs);
630 
631 	if (fs->lfs_ronly)
632 		return EROFS;
633 
634 	lfs_imtime(fs);
635 
636 	/*
637 	 * Allocate a segment structure and enough space to hold pointers to
638 	 * the maximum possible number of buffers which can be described in a
639 	 * single summary block.
640 	 */
641 	do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags);
642 
643 	/* We can't do a partial write and checkpoint at the same time. */
644 	if (do_ckp)
645 		flags &= ~SEGM_SINGLE;
646 
647 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
648 	sp = fs->lfs_sp;
649 	if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP))
650 		do_ckp = 1;
651 
652 	/*
653 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
654 	 * in which case we have to flush *all* buffers off of this vnode.
655 	 * We don't care about other nodes, but write any non-dirop nodes
656 	 * anyway in anticipation of another getnewvnode().
657 	 *
658 	 * If we're cleaning we only write cleaning and ifile blocks, and
659 	 * no dirops, since otherwise we'd risk corruption in a crash.
660 	 */
661 	if (sp->seg_flags & SEGM_CLEAN)
662 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
663 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
664 		do {
665 			um_error = lfs_writevnodes(fs, mp, sp, VN_REG);
666 			if ((sp->seg_flags & SEGM_SINGLE) &&
667 			    fs->lfs_curseg != fs->lfs_startseg) {
668 				DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%x\n", fs->lfs_offset));
669 				break;
670 			}
671 
672 			if (do_ckp || fs->lfs_dirops == 0) {
673 				if (!writer_set) {
674 					lfs_writer_enter(fs, "lfs writer");
675 					writer_set = 1;
676 				}
677 				error = lfs_writevnodes(fs, mp, sp, VN_DIROP);
678 				if (um_error == 0)
679 					um_error = error;
680 				/* In case writevnodes errored out */
681 				lfs_flush_dirops(fs);
682 				((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
683 				lfs_finalize_fs_seguse(fs);
684 			}
685 			if (do_ckp && um_error) {
686 				lfs_segunlock_relock(fs);
687 				sp = fs->lfs_sp;
688 			}
689 		} while (do_ckp && um_error != 0);
690 	}
691 
692 	/*
693 	 * If we are doing a checkpoint, mark everything since the
694 	 * last checkpoint as no longer ACTIVE.
695 	 */
696 	if (do_ckp || fs->lfs_doifile) {
697 		segleft = fs->lfs_nseg;
698 		curseg = 0;
699 		for (n = 0; n < fs->lfs_segtabsz; n++) {
700 			dirty = 0;
701 			if (bread(fs->lfs_ivnode, fs->lfs_cleansz + n,
702 			    fs->lfs_bsize, NOCRED, B_MODIFY, &bp))
703 				panic("lfs_segwrite: ifile read");
704 			segusep = (SEGUSE *)bp->b_data;
705 			maxseg = min(segleft, fs->lfs_sepb);
706 			for (i = 0; i < maxseg; i++) {
707 				sn = curseg + i;
708 				if (sn != lfs_dtosn(fs, fs->lfs_curseg) &&
709 				    segusep->su_flags & SEGUSE_ACTIVE) {
710 					segusep->su_flags &= ~SEGUSE_ACTIVE;
711 					--fs->lfs_nactive;
712 					++dirty;
713 				}
714 				fs->lfs_suflags[fs->lfs_activesb][sn] =
715 					segusep->su_flags;
716 				if (fs->lfs_version > 1)
717 					++segusep;
718 				else
719 					segusep = (SEGUSE *)
720 						((SEGUSE_V1 *)segusep + 1);
721 			}
722 
723 			if (dirty)
724 				error = LFS_BWRITE_LOG(bp); /* Ifile */
725 			else
726 				brelse(bp, 0);
727 			segleft -= fs->lfs_sepb;
728 			curseg += fs->lfs_sepb;
729 		}
730 	}
731 
732 	KASSERT(LFS_SEGLOCK_HELD(fs));
733 
734 	did_ckp = 0;
735 	if (do_ckp || fs->lfs_doifile) {
736 		vp = fs->lfs_ivnode;
737 #ifdef DEBUG
738 		int loopcount = 0;
739 #endif
740 		do {
741 #ifdef DEBUG
742 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid);
743 #endif
744 			mutex_enter(&lfs_lock);
745 			fs->lfs_flags &= ~LFS_IFDIRTY;
746 			mutex_exit(&lfs_lock);
747 
748 			ip = VTOI(vp);
749 
750 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
751 				/*
752 				 * Ifile has no pages, so we don't need
753 				 * to check error return here.
754 				 */
755 				lfs_writefile(fs, sp, vp);
756 				/*
757 				 * Ensure the Ifile takes the current segment
758 				 * into account.  See comment in lfs_vflush.
759 				 */
760 				lfs_writefile(fs, sp, vp);
761 				lfs_writefile(fs, sp, vp);
762 			}
763 
764 			if (ip->i_flag & IN_ALLMOD)
765 				++did_ckp;
766 #if 0
767 			redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0);
768 #else
769 			redo = lfs_writeinode(fs, sp, ip);
770 #endif
771 			redo += lfs_writeseg(fs, sp);
772 			mutex_enter(&lfs_lock);
773 			redo += (fs->lfs_flags & LFS_IFDIRTY);
774 			mutex_exit(&lfs_lock);
775 #ifdef DEBUG
776 			if (++loopcount > 2)
777 				log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n",
778 					loopcount);
779 #endif
780 		} while (redo && do_ckp);
781 
782 		/*
783 		 * Unless we are unmounting, the Ifile may continue to have
784 		 * dirty blocks even after a checkpoint, due to changes to
785 		 * inodes' atime.  If we're checkpointing, it's "impossible"
786 		 * for other parts of the Ifile to be dirty after the loop
787 		 * above, since we hold the segment lock.
788 		 */
789 		mutex_enter(vp->v_interlock);
790 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
791 			LFS_CLR_UINO(ip, IN_ALLMOD);
792 		}
793 #ifdef DIAGNOSTIC
794 		else if (do_ckp) {
795 			int do_panic = 0;
796 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
797 				if (bp->b_lblkno < fs->lfs_cleansz +
798 				    fs->lfs_segtabsz &&
799 				    !(bp->b_flags & B_GATHERED)) {
800 					printf("ifile lbn %ld still dirty (flags %lx)\n",
801 						(long)bp->b_lblkno,
802 						(long)bp->b_flags);
803 					++do_panic;
804 				}
805 			}
806 			if (do_panic)
807 				panic("dirty blocks");
808 		}
809 #endif
810 		mutex_exit(vp->v_interlock);
811 	} else {
812 		(void) lfs_writeseg(fs, sp);
813 	}
814 
815 	/* Note Ifile no longer needs to be written */
816 	fs->lfs_doifile = 0;
817 	if (writer_set)
818 		lfs_writer_leave(fs);
819 
820 	/*
821 	 * If we didn't write the Ifile, we didn't really do anything.
822 	 * That means that (1) there is a checkpoint on disk and (2)
823 	 * nothing has changed since it was written.
824 	 *
825 	 * Take the flags off of the segment so that lfs_segunlock
826 	 * doesn't have to write the superblock either.
827 	 */
828 	if (do_ckp && !did_ckp) {
829 		sp->seg_flags &= ~SEGM_CKP;
830 	}
831 
832 	if (lfs_dostats) {
833 		++lfs_stats.nwrites;
834 		if (sp->seg_flags & SEGM_SYNC)
835 			++lfs_stats.nsync_writes;
836 		if (sp->seg_flags & SEGM_CKP)
837 			++lfs_stats.ncheckpoints;
838 	}
839 	lfs_segunlock(fs);
840 	return (0);
841 }
842 
843 /*
844  * Write the dirty blocks associated with a vnode.
845  */
846 int
847 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
848 {
849 	struct inode *ip;
850 	int i, frag;
851 	int error;
852 
853 	ASSERT_SEGLOCK(fs);
854 	error = 0;
855 	ip = VTOI(vp);
856 
857 	lfs_acquire_finfo(fs, ip->i_number, ip->i_gen);
858 
859 	if (vp->v_uflag & VU_DIROP)
860 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
861 
862 	if (sp->seg_flags & SEGM_CLEAN) {
863 		lfs_gather(fs, sp, vp, lfs_match_fake);
864 		/*
865 		 * For a file being flushed, we need to write *all* blocks.
866 		 * This means writing the cleaning blocks first, and then
867 		 * immediately following with any non-cleaning blocks.
868 		 * The same is true of the Ifile since checkpoints assume
869 		 * that all valid Ifile blocks are written.
870 		 */
871 		if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) {
872 			lfs_gather(fs, sp, vp, lfs_match_data);
873 			/*
874 			 * Don't call VOP_PUTPAGES: if we're flushing,
875 			 * we've already done it, and the Ifile doesn't
876 			 * use the page cache.
877 			 */
878 		}
879 	} else {
880 		lfs_gather(fs, sp, vp, lfs_match_data);
881 		/*
882 		 * If we're flushing, we've already called VOP_PUTPAGES
883 		 * so don't do it again.  Otherwise, we want to write
884 		 * everything we've got.
885 		 */
886 		if (!IS_FLUSHING(fs, vp)) {
887 			mutex_enter(vp->v_interlock);
888 			error = VOP_PUTPAGES(vp, 0, 0,
889 				PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
890 		}
891 	}
892 
893 	/*
894 	 * It may not be necessary to write the meta-data blocks at this point,
895 	 * as the roll-forward recovery code should be able to reconstruct the
896 	 * list.
897 	 *
898 	 * We have to write them anyway, though, under two conditions: (1) the
899 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
900 	 * checkpointing.
901 	 *
902 	 * BUT if we are cleaning, we might have indirect blocks that refer to
903 	 * new blocks not being written yet, in addition to fragments being
904 	 * moved out of a cleaned segment.  If that is the case, don't
905 	 * write the indirect blocks, or the finfo will have a small block
906 	 * in the middle of it!
907 	 * XXX in this case isn't the inode size wrong too?
908 	 */
909 	frag = 0;
910 	if (sp->seg_flags & SEGM_CLEAN) {
911 		for (i = 0; i < ULFS_NDADDR; i++)
912 			if (ip->i_lfs_fragsize[i] > 0 &&
913 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
914 				++frag;
915 	}
916 #ifdef DIAGNOSTIC
917 	if (frag > 1)
918 		panic("lfs_writefile: more than one fragment!");
919 #endif
920 	if (IS_FLUSHING(fs, vp) ||
921 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
922 		lfs_gather(fs, sp, vp, lfs_match_indir);
923 		lfs_gather(fs, sp, vp, lfs_match_dindir);
924 		lfs_gather(fs, sp, vp, lfs_match_tindir);
925 	}
926 	lfs_release_finfo(fs);
927 
928 	return error;
929 }
930 
931 /*
932  * Update segment accounting to reflect this inode's change of address.
933  */
934 static int
935 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr)
936 {
937 	struct buf *bp;
938 	daddr_t daddr;
939 	IFILE *ifp;
940 	SEGUSE *sup;
941 	ino_t ino;
942 	int redo_ifile;
943 	u_int32_t sn;
944 
945 	redo_ifile = 0;
946 
947 	/*
948 	 * If updating the ifile, update the super-block.  Update the disk
949 	 * address and access times for this inode in the ifile.
950 	 */
951 	ino = ip->i_number;
952 	if (ino == LFS_IFILE_INUM) {
953 		daddr = fs->lfs_idaddr;
954 		fs->lfs_idaddr = LFS_DBTOFSB(fs, ndaddr);
955 	} else {
956 		LFS_IENTRY(ifp, fs, ino, bp);
957 		daddr = ifp->if_daddr;
958 		ifp->if_daddr = LFS_DBTOFSB(fs, ndaddr);
959 		(void)LFS_BWRITE_LOG(bp); /* Ifile */
960 	}
961 
962 	/*
963 	 * If this is the Ifile and lfs_offset is set to the first block
964 	 * in the segment, dirty the new segment's accounting block
965 	 * (XXX should already be dirty?) and tell the caller to do it again.
966 	 */
967 	if (ip->i_number == LFS_IFILE_INUM) {
968 		sn = lfs_dtosn(fs, fs->lfs_offset);
969 		if (lfs_sntod(fs, sn) + lfs_btofsb(fs, fs->lfs_sumsize) ==
970 		    fs->lfs_offset) {
971 			LFS_SEGENTRY(sup, fs, sn, bp);
972 			KASSERT(bp->b_oflags & BO_DELWRI);
973 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
974 			/* fs->lfs_flags |= LFS_IFDIRTY; */
975 			redo_ifile |= 1;
976 		}
977 	}
978 
979 	/*
980 	 * The inode's last address should not be in the current partial
981 	 * segment, except under exceptional circumstances (lfs_writevnodes
982 	 * had to start over, and in the meantime more blocks were written
983 	 * to a vnode).	 Both inodes will be accounted to this segment
984 	 * in lfs_writeseg so we need to subtract the earlier version
985 	 * here anyway.	 The segment count can temporarily dip below
986 	 * zero here; keep track of how many duplicates we have in
987 	 * "dupino" so we don't panic below.
988 	 */
989 	if (daddr >= fs->lfs_lastpseg && daddr <= fs->lfs_offset) {
990 		++sp->ndupino;
991 		DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg "
992 		      "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
993 		      (long long)daddr, sp->ndupino));
994 	}
995 	/*
996 	 * Account the inode: it no longer belongs to its former segment,
997 	 * though it will not belong to the new segment until that segment
998 	 * is actually written.
999 	 */
1000 	if (daddr != LFS_UNUSED_DADDR) {
1001 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
1002 #ifdef DIAGNOSTIC
1003 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
1004 #endif
1005 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1006 #ifdef DIAGNOSTIC
1007 		if (sup->su_nbytes +
1008 		    sizeof (struct ulfs1_dinode) * ndupino
1009 		      < sizeof (struct ulfs1_dinode)) {
1010 			printf("lfs_writeinode: negative bytes "
1011 			       "(segment %" PRIu32 " short by %d, "
1012 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
1013 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
1014 			       "ndupino=%d)\n",
1015 			       lfs_dtosn(fs, daddr),
1016 			       (int)sizeof (struct ulfs1_dinode) *
1017 				   (1 - sp->ndupino) - sup->su_nbytes,
1018 			       oldsn, sp->seg_number, daddr,
1019 			       (unsigned int)sup->su_nbytes,
1020 			       sp->ndupino);
1021 			panic("lfs_writeinode: negative bytes");
1022 			sup->su_nbytes = sizeof (struct ulfs1_dinode);
1023 		}
1024 #endif
1025 		DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n",
1026 		      lfs_dtosn(fs, daddr), sizeof (struct ulfs1_dinode), ino));
1027 		sup->su_nbytes -= sizeof (struct ulfs1_dinode);
1028 		redo_ifile |=
1029 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1030 		if (redo_ifile) {
1031 			mutex_enter(&lfs_lock);
1032 			fs->lfs_flags |= LFS_IFDIRTY;
1033 			mutex_exit(&lfs_lock);
1034 			/* Don't double-account */
1035 			fs->lfs_idaddr = 0x0;
1036 		}
1037 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1038 	}
1039 
1040 	return redo_ifile;
1041 }
1042 
1043 int
1044 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
1045 {
1046 	struct buf *bp;
1047 	struct ulfs1_dinode *cdp;
1048 	struct vnode *vp = ITOV(ip);
1049 	daddr_t daddr;
1050 	int32_t *daddrp;	/* XXX ondisk32 */
1051 	int i, ndx;
1052 	int redo_ifile = 0;
1053 	int gotblk = 0;
1054 	int count;
1055 
1056 	ASSERT_SEGLOCK(fs);
1057 	if (!(ip->i_flag & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP))
1058 		return (0);
1059 
1060 	/* Can't write ifile when writer is not set */
1061 	KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 ||
1062 		(sp->seg_flags & SEGM_CLEAN));
1063 
1064 	/*
1065 	 * If this is the Ifile, see if writing it here will generate a
1066 	 * temporary misaccounting.  If it will, do the accounting and write
1067 	 * the blocks, postponing the inode write until the accounting is
1068 	 * solid.
1069 	 */
1070 	count = 0;
1071 	while (vp == fs->lfs_ivnode) {
1072 		int redo = 0;
1073 
1074 		if (sp->idp == NULL && sp->ibp == NULL &&
1075 		    (sp->seg_bytes_left < fs->lfs_ibsize ||
1076 		     sp->sum_bytes_left < sizeof(int32_t))) {
1077 			(void) lfs_writeseg(fs, sp);
1078 			continue;
1079 		}
1080 
1081 		/* Look for dirty Ifile blocks */
1082 		LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) {
1083 			if (!(bp->b_flags & B_GATHERED)) {
1084 				redo = 1;
1085 				break;
1086 			}
1087 		}
1088 
1089 		if (redo == 0)
1090 			redo = lfs_update_iaddr(fs, sp, ip, 0x0);
1091 		if (redo == 0)
1092 			break;
1093 
1094 		if (sp->idp) {
1095 			sp->idp->di_inumber = 0;
1096 			sp->idp = NULL;
1097 		}
1098 		++count;
1099 		if (count > 2)
1100 			log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count);
1101 		lfs_writefile(fs, sp, fs->lfs_ivnode);
1102 	}
1103 
1104 	/* Allocate a new inode block if necessary. */
1105 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
1106 	    sp->ibp == NULL) {
1107 		/* Allocate a new segment if necessary. */
1108 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
1109 		    sp->sum_bytes_left < sizeof(int32_t))
1110 			(void) lfs_writeseg(fs, sp);
1111 
1112 		/* Get next inode block. */
1113 		daddr = fs->lfs_offset;
1114 		fs->lfs_offset += lfs_btofsb(fs, fs->lfs_ibsize);
1115 		sp->ibp = *sp->cbpp++ =
1116 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1117 			    LFS_FSBTODB(fs, daddr), fs->lfs_ibsize, 0, 0);
1118 		gotblk++;
1119 
1120 		/* Zero out inode numbers */
1121 		for (i = 0; i < LFS_INOPB(fs); ++i)
1122 			((struct ulfs1_dinode *)sp->ibp->b_data)[i].di_inumber =
1123 			    0;
1124 
1125 		++sp->start_bpp;
1126 		fs->lfs_avail -= lfs_btofsb(fs, fs->lfs_ibsize);
1127 		/* Set remaining space counters. */
1128 		sp->seg_bytes_left -= fs->lfs_ibsize;
1129 		sp->sum_bytes_left -= sizeof(int32_t);
1130 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
1131 			sp->ninodes / LFS_INOPB(fs) - 1;
1132 		((int32_t *)(sp->segsum))[ndx] = daddr;
1133 	}
1134 
1135 	/* Check VU_DIROP in case there is a new file with no data blocks */
1136 	if (vp->v_uflag & VU_DIROP)
1137 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
1138 
1139 	/* Update the inode times and copy the inode onto the inode page. */
1140 	/* XXX kludge --- don't redirty the ifile just to put times on it */
1141 	if (ip->i_number != LFS_IFILE_INUM)
1142 		LFS_ITIMES(ip, NULL, NULL, NULL);
1143 
1144 	/*
1145 	 * If this is the Ifile, and we've already written the Ifile in this
1146 	 * partial segment, just overwrite it (it's not on disk yet) and
1147 	 * continue.
1148 	 *
1149 	 * XXX we know that the bp that we get the second time around has
1150 	 * already been gathered.
1151 	 */
1152 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
1153 		*(sp->idp) = *ip->i_din.ffs1_din;
1154 		ip->i_lfs_osize = ip->i_size;
1155 		return 0;
1156 	}
1157 
1158 	bp = sp->ibp;
1159 	cdp = ((struct ulfs1_dinode *)bp->b_data) + (sp->ninodes % LFS_INOPB(fs));
1160 	*cdp = *ip->i_din.ffs1_din;
1161 
1162 	/*
1163 	 * This inode is on its way to disk; clear its VU_DIROP status when
1164 	 * the write is complete.
1165 	 */
1166 	if (vp->v_uflag & VU_DIROP) {
1167 		if (!(sp->seg_flags & SEGM_CLEAN))
1168 			ip->i_flag |= IN_CDIROP;
1169 		else {
1170 			DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number));
1171 		}
1172 	}
1173 
1174 	/*
1175 	 * If cleaning, link counts and directory file sizes cannot change,
1176 	 * since those would be directory operations---even if the file
1177 	 * we are writing is marked VU_DIROP we should write the old values.
1178 	 * If we're not cleaning, of course, update the values so we get
1179 	 * current values the next time we clean.
1180 	 */
1181 	if (sp->seg_flags & SEGM_CLEAN) {
1182 		if (vp->v_uflag & VU_DIROP) {
1183 			cdp->di_nlink = ip->i_lfs_odnlink;
1184 			/* if (vp->v_type == VDIR) */
1185 			cdp->di_size = ip->i_lfs_osize;
1186 		}
1187 	} else {
1188 		ip->i_lfs_odnlink = cdp->di_nlink;
1189 		ip->i_lfs_osize = ip->i_size;
1190 	}
1191 
1192 
1193 	/* We can finish the segment accounting for truncations now */
1194 	lfs_finalize_ino_seguse(fs, ip);
1195 
1196 	/*
1197 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
1198 	 * addresses to disk; possibly change the on-disk record of
1199 	 * the inode size, either by reverting to the previous size
1200 	 * (in the case of cleaning) or by verifying the inode's block
1201 	 * holdings (in the case of files being allocated as they are being
1202 	 * written).
1203 	 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail
1204 	 * XXX count on disk wrong by the same amount.	We should be
1205 	 * XXX able to "borrow" from lfs_avail and return it after the
1206 	 * XXX Ifile is written.  See also in lfs_writeseg.
1207 	 */
1208 
1209 	/* Check file size based on highest allocated block */
1210 	if (((ip->i_ffs1_mode & LFS_IFMT) == LFS_IFREG ||
1211 	     (ip->i_ffs1_mode & LFS_IFMT) == LFS_IFDIR) &&
1212 	    ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) {
1213 		cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift;
1214 		DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %"
1215 		      PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size));
1216 	}
1217 	if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) {
1218 		DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)"
1219 		      " at %x\n", ip->i_number, ip->i_lfs_effnblks,
1220 		      ip->i_ffs1_blocks, fs->lfs_offset));
1221 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + ULFS_NIADDR;
1222 		     daddrp++) {
1223 			if (*daddrp == UNWRITTEN) {
1224 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1225 				*daddrp = 0;
1226 			}
1227 		}
1228 	}
1229 
1230 #ifdef DIAGNOSTIC
1231 	/*
1232 	 * Check dinode held blocks against dinode size.
1233 	 * This should be identical to the check in lfs_vget().
1234 	 */
1235 	for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift;
1236 	     i < ULFS_NDADDR; i++) {
1237 		KASSERT(i >= 0);
1238 		if ((cdp->di_mode & LFS_IFMT) == LFS_IFLNK)
1239 			continue;
1240 		if (((cdp->di_mode & LFS_IFMT) == LFS_IFBLK ||
1241 		     (cdp->di_mode & LFS_IFMT) == LFS_IFCHR) && i == 0)
1242 			continue;
1243 		if (cdp->di_db[i] != 0) {
1244 # ifdef DEBUG
1245 			lfs_dump_dinode(cdp);
1246 # endif
1247 			panic("writing inconsistent inode");
1248 		}
1249 	}
1250 #endif /* DIAGNOSTIC */
1251 
1252 	if (ip->i_flag & IN_CLEANING)
1253 		LFS_CLR_UINO(ip, IN_CLEANING);
1254 	else {
1255 		/* XXX IN_ALLMOD */
1256 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
1257 			     IN_UPDATE | IN_MODIFY);
1258 		if (ip->i_lfs_effnblks == ip->i_ffs1_blocks)
1259 			LFS_CLR_UINO(ip, IN_MODIFIED);
1260 		else {
1261 			DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real "
1262 			    "blks=%d, eff=%d\n", ip->i_number,
1263 			    ip->i_ffs1_blocks, ip->i_lfs_effnblks));
1264 		}
1265 	}
1266 
1267 	if (ip->i_number == LFS_IFILE_INUM) {
1268 		/* We know sp->idp == NULL */
1269 		sp->idp = ((struct ulfs1_dinode *)bp->b_data) +
1270 			(sp->ninodes % LFS_INOPB(fs));
1271 
1272 		/* Not dirty any more */
1273 		mutex_enter(&lfs_lock);
1274 		fs->lfs_flags &= ~LFS_IFDIRTY;
1275 		mutex_exit(&lfs_lock);
1276 	}
1277 
1278 	if (gotblk) {
1279 		mutex_enter(&bufcache_lock);
1280 		LFS_LOCK_BUF(bp);
1281 		brelsel(bp, 0);
1282 		mutex_exit(&bufcache_lock);
1283 	}
1284 
1285 	/* Increment inode count in segment summary block. */
1286 	++((SEGSUM *)(sp->segsum))->ss_ninos;
1287 
1288 	/* If this page is full, set flag to allocate a new page. */
1289 	if (++sp->ninodes % LFS_INOPB(fs) == 0)
1290 		sp->ibp = NULL;
1291 
1292 	redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno);
1293 
1294 	KASSERT(redo_ifile == 0);
1295 	return (redo_ifile);
1296 }
1297 
1298 int
1299 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr)
1300 {
1301 	struct lfs *fs;
1302 	int vers;
1303 	int j, blksinblk;
1304 
1305 	ASSERT_SEGLOCK(sp->fs);
1306 	/*
1307 	 * If full, finish this segment.  We may be doing I/O, so
1308 	 * release and reacquire the splbio().
1309 	 */
1310 #ifdef DIAGNOSTIC
1311 	if (sp->vp == NULL)
1312 		panic ("lfs_gatherblock: Null vp in segment");
1313 #endif
1314 	fs = sp->fs;
1315 	blksinblk = howmany(bp->b_bcount, fs->lfs_bsize);
1316 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1317 	    sp->seg_bytes_left < bp->b_bcount) {
1318 		if (mptr)
1319 			mutex_exit(mptr);
1320 		lfs_updatemeta(sp);
1321 
1322 		vers = sp->fip->fi_version;
1323 		(void) lfs_writeseg(fs, sp);
1324 
1325 		/* Add the current file to the segment summary. */
1326 		lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers);
1327 
1328 		if (mptr)
1329 			mutex_enter(mptr);
1330 		return (1);
1331 	}
1332 
1333 	if (bp->b_flags & B_GATHERED) {
1334 		DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d,"
1335 		      " lbn %" PRId64 "\n",
1336 		      sp->fip->fi_ino, bp->b_lblkno));
1337 		return (0);
1338 	}
1339 
1340 	/* Insert into the buffer list, update the FINFO block. */
1341 	bp->b_flags |= B_GATHERED;
1342 
1343 	*sp->cbpp++ = bp;
1344 	for (j = 0; j < blksinblk; j++) {
1345 		sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j;
1346 		/* This block's accounting moves from lfs_favail to lfs_avail */
1347 		lfs_deregister_block(sp->vp, bp->b_lblkno + j);
1348 	}
1349 
1350 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1351 	sp->seg_bytes_left -= bp->b_bcount;
1352 	return (0);
1353 }
1354 
1355 int
1356 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1357     int (*match)(struct lfs *, struct buf *))
1358 {
1359 	struct buf *bp, *nbp;
1360 	int count = 0;
1361 
1362 	ASSERT_SEGLOCK(fs);
1363 	if (vp->v_type == VBLK)
1364 		return 0;
1365 	KASSERT(sp->vp == NULL);
1366 	sp->vp = vp;
1367 	mutex_enter(&bufcache_lock);
1368 
1369 #ifndef LFS_NO_BACKBUF_HACK
1370 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1371 # define	BUF_OFFSET	\
1372 	(((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp)
1373 # define	BACK_BUF(BP)	\
1374 	((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1375 # define	BEG_OF_LIST	\
1376 	((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1377 
1378 loop:
1379 	/* Find last buffer. */
1380 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1381 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1382 	     bp = LIST_NEXT(bp, b_vnbufs))
1383 		/* nothing */;
1384 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1385 		nbp = BACK_BUF(bp);
1386 #else /* LFS_NO_BACKBUF_HACK */
1387 loop:
1388 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1389 		nbp = LIST_NEXT(bp, b_vnbufs);
1390 #endif /* LFS_NO_BACKBUF_HACK */
1391 		if ((bp->b_cflags & BC_BUSY) != 0 ||
1392 		    (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) {
1393 #ifdef DEBUG
1394 			if (vp == fs->lfs_ivnode &&
1395 			    (bp->b_cflags & BC_BUSY) != 0 &&
1396 			    (bp->b_flags & B_GATHERED) == 0)
1397 				log(LOG_NOTICE, "lfs_gather: ifile lbn %"
1398 				      PRId64 " busy (%x) at 0x%x",
1399 				      bp->b_lblkno, bp->b_flags,
1400 				      (unsigned)fs->lfs_offset);
1401 #endif
1402 			continue;
1403 		}
1404 #ifdef DIAGNOSTIC
1405 # ifdef LFS_USE_B_INVAL
1406 		if ((bp->b_flags & BC_INVAL) != 0 && bp->b_iodone == NULL) {
1407 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1408 			      " is BC_INVAL\n", bp->b_lblkno));
1409 			VOP_PRINT(bp->b_vp);
1410 		}
1411 # endif /* LFS_USE_B_INVAL */
1412 		if (!(bp->b_oflags & BO_DELWRI))
1413 			panic("lfs_gather: bp not BO_DELWRI");
1414 		if (!(bp->b_flags & B_LOCKED)) {
1415 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1416 			      " blk %" PRId64 " not B_LOCKED\n",
1417 			      bp->b_lblkno,
1418 			      LFS_DBTOFSB(fs, bp->b_blkno)));
1419 			VOP_PRINT(bp->b_vp);
1420 			panic("lfs_gather: bp not B_LOCKED");
1421 		}
1422 #endif
1423 		if (lfs_gatherblock(sp, bp, &bufcache_lock)) {
1424 			goto loop;
1425 		}
1426 		count++;
1427 	}
1428 	mutex_exit(&bufcache_lock);
1429 	lfs_updatemeta(sp);
1430 	KASSERT(sp->vp == vp);
1431 	sp->vp = NULL;
1432 	return count;
1433 }
1434 
1435 #if DEBUG
1436 # define DEBUG_OOFF(n) do {						\
1437 	if (ooff == 0) {						\
1438 		DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \
1439 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1440 			", was 0x0 (or %" PRId64 ")\n",			\
1441 			(n), ip->i_number, lbn, ndaddr, daddr));	\
1442 	}								\
1443 } while (0)
1444 #else
1445 # define DEBUG_OOFF(n)
1446 #endif
1447 
1448 /*
1449  * Change the given block's address to ndaddr, finding its previous
1450  * location using ulfs_bmaparray().
1451  *
1452  * Account for this change in the segment table.
1453  *
1454  * called with sp == NULL by roll-forwarding code.
1455  */
1456 void
1457 lfs_update_single(struct lfs *fs, struct segment *sp,
1458     struct vnode *vp, daddr_t lbn, int32_t ndaddr, int size)
1459 {
1460 	SEGUSE *sup;
1461 	struct buf *bp;
1462 	struct indir a[ULFS_NIADDR + 2], *ap;
1463 	struct inode *ip;
1464 	daddr_t daddr, ooff;
1465 	int num, error;
1466 	int bb, osize, obb;
1467 
1468 	ASSERT_SEGLOCK(fs);
1469 	KASSERT(sp == NULL || sp->vp == vp);
1470 	ip = VTOI(vp);
1471 
1472 	error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1473 	if (error)
1474 		panic("lfs_updatemeta: ulfs_bmaparray returned %d", error);
1475 
1476 	daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */
1477 	KASSERT(daddr <= LFS_MAX_DADDR);
1478 	if (daddr > 0)
1479 		daddr = LFS_DBTOFSB(fs, daddr);
1480 
1481 	bb = lfs_numfrags(fs, size);
1482 	switch (num) {
1483 	    case 0:
1484 		    ooff = ip->i_ffs1_db[lbn];
1485 		    DEBUG_OOFF(0);
1486 		    if (ooff == UNWRITTEN)
1487 			    ip->i_ffs1_blocks += bb;
1488 		    else {
1489 			    /* possible fragment truncation or extension */
1490 			    obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]);
1491 			    ip->i_ffs1_blocks += (bb - obb);
1492 		    }
1493 		    ip->i_ffs1_db[lbn] = ndaddr;
1494 		    break;
1495 	    case 1:
1496 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1497 		    DEBUG_OOFF(1);
1498 		    if (ooff == UNWRITTEN)
1499 			    ip->i_ffs1_blocks += bb;
1500 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1501 		    break;
1502 	    default:
1503 		    ap = &a[num - 1];
1504 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED,
1505 			B_MODIFY, &bp))
1506 			    panic("lfs_updatemeta: bread bno %" PRId64,
1507 				  ap->in_lbn);
1508 
1509 		    /* XXX ondisk32 */
1510 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1511 		    DEBUG_OOFF(num);
1512 		    if (ooff == UNWRITTEN)
1513 			    ip->i_ffs1_blocks += bb;
1514 		    /* XXX ondisk32 */
1515 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1516 		    (void) VOP_BWRITE(bp->b_vp, bp);
1517 	}
1518 
1519 	KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr);
1520 
1521 	/* Update hiblk when extending the file */
1522 	if (lbn > ip->i_lfs_hiblk)
1523 		ip->i_lfs_hiblk = lbn;
1524 
1525 	/*
1526 	 * Though we'd rather it couldn't, this *can* happen right now
1527 	 * if cleaning blocks and regular blocks coexist.
1528 	 */
1529 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1530 
1531 	/*
1532 	 * Update segment usage information, based on old size
1533 	 * and location.
1534 	 */
1535 	if (daddr > 0) {
1536 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
1537 #ifdef DIAGNOSTIC
1538 		int ndupino;
1539 
1540 		if (sp && sp->seg_number == oldsn) {
1541 			ndupino = sp->ndupino;
1542 		} else {
1543 			ndupino = 0;
1544 		}
1545 #endif
1546 		KASSERT(oldsn < fs->lfs_nseg);
1547 		if (lbn >= 0 && lbn < ULFS_NDADDR)
1548 			osize = ip->i_lfs_fragsize[lbn];
1549 		else
1550 			osize = fs->lfs_bsize;
1551 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1552 #ifdef DIAGNOSTIC
1553 		if (sup->su_nbytes + sizeof (struct ulfs1_dinode) * ndupino
1554 		    < osize) {
1555 			printf("lfs_updatemeta: negative bytes "
1556 			       "(segment %" PRIu32 " short by %" PRId64
1557 			       ")\n", lfs_dtosn(fs, daddr),
1558 			       (int64_t)osize -
1559 			       (sizeof (struct ulfs1_dinode) * ndupino +
1560 				sup->su_nbytes));
1561 			printf("lfs_updatemeta: ino %llu, lbn %" PRId64
1562 			       ", addr = 0x%" PRIx64 "\n",
1563 			       (unsigned long long)ip->i_number, lbn, daddr);
1564 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1565 			panic("lfs_updatemeta: negative bytes");
1566 			sup->su_nbytes = osize -
1567 			    sizeof (struct ulfs1_dinode) * ndupino;
1568 		}
1569 #endif
1570 		DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1571 		      " db 0x%" PRIx64 "\n",
1572 		      lfs_dtosn(fs, daddr), osize,
1573 		      ip->i_number, lbn, daddr));
1574 		sup->su_nbytes -= osize;
1575 		if (!(bp->b_flags & B_GATHERED)) {
1576 			mutex_enter(&lfs_lock);
1577 			fs->lfs_flags |= LFS_IFDIRTY;
1578 			mutex_exit(&lfs_lock);
1579 		}
1580 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1581 	}
1582 	/*
1583 	 * Now that this block has a new address, and its old
1584 	 * segment no longer owns it, we can forget about its
1585 	 * old size.
1586 	 */
1587 	if (lbn >= 0 && lbn < ULFS_NDADDR)
1588 		ip->i_lfs_fragsize[lbn] = size;
1589 }
1590 
1591 /*
1592  * Update the metadata that points to the blocks listed in the FINFO
1593  * array.
1594  */
1595 void
1596 lfs_updatemeta(struct segment *sp)
1597 {
1598 	struct buf *sbp;
1599 	struct lfs *fs;
1600 	struct vnode *vp;
1601 	daddr_t lbn;
1602 	int i, nblocks, num;
1603 	int bb;
1604 	int bytesleft, size;
1605 
1606 	ASSERT_SEGLOCK(sp->fs);
1607 	vp = sp->vp;
1608 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1609 	KASSERT(nblocks >= 0);
1610 	KASSERT(vp != NULL);
1611 	if (nblocks == 0)
1612 		return;
1613 
1614 	/*
1615 	 * This count may be high due to oversize blocks from lfs_gop_write.
1616 	 * Correct for this. (XXX we should be able to keep track of these.)
1617 	 */
1618 	fs = sp->fs;
1619 	for (i = 0; i < nblocks; i++) {
1620 		if (sp->start_bpp[i] == NULL) {
1621 			DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks));
1622 			nblocks = i;
1623 			break;
1624 		}
1625 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1626 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1627 		nblocks -= num - 1;
1628 	}
1629 
1630 	KASSERT(vp->v_type == VREG ||
1631 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1632 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1633 
1634 	/*
1635 	 * Sort the blocks.
1636 	 *
1637 	 * We have to sort even if the blocks come from the
1638 	 * cleaner, because there might be other pending blocks on the
1639 	 * same inode...and if we don't sort, and there are fragments
1640 	 * present, blocks may be written in the wrong place.
1641 	 */
1642 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1643 
1644 	/*
1645 	 * Record the length of the last block in case it's a fragment.
1646 	 * If there are indirect blocks present, they sort last.  An
1647 	 * indirect block will be lfs_bsize and its presence indicates
1648 	 * that you cannot have fragments.
1649 	 *
1650 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1651 	 * XXX a later indirect block.	This will continue to be
1652 	 * XXX true until lfs_markv is fixed to do everything with
1653 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1654 	 */
1655 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1656 		fs->lfs_bmask) + 1;
1657 
1658 	/*
1659 	 * Assign disk addresses, and update references to the logical
1660 	 * block and the segment usage information.
1661 	 */
1662 	for (i = nblocks; i--; ++sp->start_bpp) {
1663 		sbp = *sp->start_bpp;
1664 		lbn = *sp->start_lbp;
1665 		KASSERT(sbp->b_lblkno == lbn);
1666 
1667 		sbp->b_blkno = LFS_FSBTODB(fs, fs->lfs_offset);
1668 
1669 		/*
1670 		 * If we write a frag in the wrong place, the cleaner won't
1671 		 * be able to correctly identify its size later, and the
1672 		 * segment will be uncleanable.	 (Even worse, it will assume
1673 		 * that the indirect block that actually ends the list
1674 		 * is of a smaller size!)
1675 		 */
1676 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1677 			panic("lfs_updatemeta: fragment is not last block");
1678 
1679 		/*
1680 		 * For each subblock in this possibly oversized block,
1681 		 * update its address on disk.
1682 		 */
1683 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1684 		KASSERT(vp == sbp->b_vp);
1685 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1686 		     bytesleft -= fs->lfs_bsize) {
1687 			size = MIN(bytesleft, fs->lfs_bsize);
1688 			bb = lfs_numfrags(fs, size);
1689 			lbn = *sp->start_lbp++;
1690 			lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset,
1691 			    size);
1692 			fs->lfs_offset += bb;
1693 		}
1694 
1695 	}
1696 
1697 	/* This inode has been modified */
1698 	LFS_SET_UINO(VTOI(vp), IN_MODIFIED);
1699 }
1700 
1701 /*
1702  * Move lfs_offset to a segment earlier than sn.
1703  */
1704 int
1705 lfs_rewind(struct lfs *fs, int newsn)
1706 {
1707 	int sn, osn, isdirty;
1708 	struct buf *bp;
1709 	SEGUSE *sup;
1710 
1711 	ASSERT_SEGLOCK(fs);
1712 
1713 	osn = lfs_dtosn(fs, fs->lfs_offset);
1714 	if (osn < newsn)
1715 		return 0;
1716 
1717 	/* lfs_avail eats the remaining space in this segment */
1718 	fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg);
1719 
1720 	/* Find a low-numbered segment */
1721 	for (sn = 0; sn < fs->lfs_nseg; ++sn) {
1722 		LFS_SEGENTRY(sup, fs, sn, bp);
1723 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1724 		brelse(bp, 0);
1725 
1726 		if (!isdirty)
1727 			break;
1728 	}
1729 	if (sn == fs->lfs_nseg)
1730 		panic("lfs_rewind: no clean segments");
1731 	if (newsn >= 0 && sn >= newsn)
1732 		return ENOENT;
1733 	fs->lfs_nextseg = sn;
1734 	lfs_newseg(fs);
1735 	fs->lfs_offset = fs->lfs_curseg;
1736 
1737 	return 0;
1738 }
1739 
1740 /*
1741  * Start a new partial segment.
1742  *
1743  * Return 1 when we entered to a new segment.
1744  * Otherwise, return 0.
1745  */
1746 int
1747 lfs_initseg(struct lfs *fs)
1748 {
1749 	struct segment *sp = fs->lfs_sp;
1750 	SEGSUM *ssp;
1751 	struct buf *sbp;	/* buffer for SEGSUM */
1752 	int repeat = 0;		/* return value */
1753 
1754 	ASSERT_SEGLOCK(fs);
1755 	/* Advance to the next segment. */
1756 	if (!LFS_PARTIAL_FITS(fs)) {
1757 		SEGUSE *sup;
1758 		struct buf *bp;
1759 
1760 		/* lfs_avail eats the remaining space */
1761 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1762 						   fs->lfs_curseg);
1763 		/* Wake up any cleaning procs waiting on this file system. */
1764 		lfs_wakeup_cleaner(fs);
1765 		lfs_newseg(fs);
1766 		repeat = 1;
1767 		fs->lfs_offset = fs->lfs_curseg;
1768 
1769 		sp->seg_number = lfs_dtosn(fs, fs->lfs_curseg);
1770 		sp->seg_bytes_left = lfs_fsbtob(fs, fs->lfs_fsbpseg);
1771 
1772 		/*
1773 		 * If the segment contains a superblock, update the offset
1774 		 * and summary address to skip over it.
1775 		 */
1776 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1777 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1778 			fs->lfs_offset += lfs_btofsb(fs, LFS_SBPAD);
1779 			sp->seg_bytes_left -= LFS_SBPAD;
1780 		}
1781 		brelse(bp, 0);
1782 		/* Segment zero could also contain the labelpad */
1783 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1784 		    fs->lfs_start < lfs_btofsb(fs, LFS_LABELPAD)) {
1785 			fs->lfs_offset +=
1786 			    lfs_btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1787 			sp->seg_bytes_left -=
1788 			    LFS_LABELPAD - lfs_fsbtob(fs, fs->lfs_start);
1789 		}
1790 	} else {
1791 		sp->seg_number = lfs_dtosn(fs, fs->lfs_curseg);
1792 		sp->seg_bytes_left = lfs_fsbtob(fs, fs->lfs_fsbpseg -
1793 				      (fs->lfs_offset - fs->lfs_curseg));
1794 	}
1795 	fs->lfs_lastpseg = fs->lfs_offset;
1796 
1797 	/* Record first address of this partial segment */
1798 	if (sp->seg_flags & SEGM_CLEAN) {
1799 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1800 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1801 			/* "1" is the artificial inc in lfs_seglock */
1802 			mutex_enter(&lfs_lock);
1803 			while (fs->lfs_iocount > 1) {
1804 				mtsleep(&fs->lfs_iocount, PRIBIO + 1,
1805 				    "lfs_initseg", 0, &lfs_lock);
1806 			}
1807 			mutex_exit(&lfs_lock);
1808 			fs->lfs_cleanind = 0;
1809 		}
1810 	}
1811 
1812 	sp->fs = fs;
1813 	sp->ibp = NULL;
1814 	sp->idp = NULL;
1815 	sp->ninodes = 0;
1816 	sp->ndupino = 0;
1817 
1818 	sp->cbpp = sp->bpp;
1819 
1820 	/* Get a new buffer for SEGSUM */
1821 	sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1822 	    LFS_FSBTODB(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1823 
1824 	/* ... and enter it into the buffer list. */
1825 	*sp->cbpp = sbp;
1826 	sp->cbpp++;
1827 	fs->lfs_offset += lfs_btofsb(fs, fs->lfs_sumsize);
1828 
1829 	sp->start_bpp = sp->cbpp;
1830 
1831 	/* Set point to SEGSUM, initialize it. */
1832 	ssp = sp->segsum = sbp->b_data;
1833 	memset(ssp, 0, fs->lfs_sumsize);
1834 	ssp->ss_next = fs->lfs_nextseg;
1835 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1836 	ssp->ss_magic = SS_MAGIC;
1837 
1838 	/* Set pointer to first FINFO, initialize it. */
1839 	sp->fip = (struct finfo *)((char *)sp->segsum + SEGSUM_SIZE(fs));
1840 	sp->fip->fi_nblocks = 0;
1841 	sp->start_lbp = &sp->fip->fi_blocks[0];
1842 	sp->fip->fi_lastlength = 0;
1843 
1844 	sp->seg_bytes_left -= fs->lfs_sumsize;
1845 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1846 
1847 	return (repeat);
1848 }
1849 
1850 /*
1851  * Remove SEGUSE_INVAL from all segments.
1852  */
1853 void
1854 lfs_unset_inval_all(struct lfs *fs)
1855 {
1856 	SEGUSE *sup;
1857 	struct buf *bp;
1858 	int i;
1859 
1860 	for (i = 0; i < fs->lfs_nseg; i++) {
1861 		LFS_SEGENTRY(sup, fs, i, bp);
1862 		if (sup->su_flags & SEGUSE_INVAL) {
1863 			sup->su_flags &= ~SEGUSE_INVAL;
1864 			LFS_WRITESEGENTRY(sup, fs, i, bp);
1865 		} else
1866 			brelse(bp, 0);
1867 	}
1868 }
1869 
1870 /*
1871  * Return the next segment to write.
1872  */
1873 void
1874 lfs_newseg(struct lfs *fs)
1875 {
1876 	CLEANERINFO *cip;
1877 	SEGUSE *sup;
1878 	struct buf *bp;
1879 	int curseg, isdirty, sn, skip_inval;
1880 
1881 	ASSERT_SEGLOCK(fs);
1882 
1883 	/* Honor LFCNWRAPSTOP */
1884 	mutex_enter(&lfs_lock);
1885 	while (fs->lfs_nextseg < fs->lfs_curseg && fs->lfs_nowrap) {
1886 		if (fs->lfs_wrappass) {
1887 			log(LOG_NOTICE, "%s: wrappass=%d\n",
1888 				fs->lfs_fsmnt, fs->lfs_wrappass);
1889 			fs->lfs_wrappass = 0;
1890 			break;
1891 		}
1892 		fs->lfs_wrapstatus = LFS_WRAP_WAITING;
1893 		wakeup(&fs->lfs_nowrap);
1894 		log(LOG_NOTICE, "%s: waiting at log wrap\n", fs->lfs_fsmnt);
1895 		mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz,
1896 			&lfs_lock);
1897 	}
1898 	fs->lfs_wrapstatus = LFS_WRAP_GOING;
1899 	mutex_exit(&lfs_lock);
1900 
1901 	LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, fs->lfs_nextseg), bp);
1902 	DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n",
1903 	      lfs_dtosn(fs, fs->lfs_nextseg)));
1904 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1905 	sup->su_nbytes = 0;
1906 	sup->su_nsums = 0;
1907 	sup->su_ninos = 0;
1908 	LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, fs->lfs_nextseg), bp);
1909 
1910 	LFS_CLEANERINFO(cip, fs, bp);
1911 	--cip->clean;
1912 	++cip->dirty;
1913 	fs->lfs_nclean = cip->clean;
1914 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1915 
1916 	fs->lfs_lastseg = fs->lfs_curseg;
1917 	fs->lfs_curseg = fs->lfs_nextseg;
1918 	skip_inval = 1;
1919 	for (sn = curseg = lfs_dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1920 		sn = (sn + 1) % fs->lfs_nseg;
1921 
1922 		if (sn == curseg) {
1923 			if (skip_inval)
1924 				skip_inval = 0;
1925 			else
1926 				panic("lfs_nextseg: no clean segments");
1927 		}
1928 		LFS_SEGENTRY(sup, fs, sn, bp);
1929 		isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0));
1930 		/* Check SEGUSE_EMPTY as we go along */
1931 		if (isdirty && sup->su_nbytes == 0 &&
1932 		    !(sup->su_flags & SEGUSE_EMPTY))
1933 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1934 		else
1935 			brelse(bp, 0);
1936 
1937 		if (!isdirty)
1938 			break;
1939 	}
1940 	if (skip_inval == 0)
1941 		lfs_unset_inval_all(fs);
1942 
1943 	++fs->lfs_nactive;
1944 	fs->lfs_nextseg = lfs_sntod(fs, sn);
1945 	if (lfs_dostats) {
1946 		++lfs_stats.segsused;
1947 	}
1948 }
1949 
1950 static struct buf *
1951 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr,
1952     int n)
1953 {
1954 	struct lfs_cluster *cl;
1955 	struct buf **bpp, *bp;
1956 
1957 	ASSERT_SEGLOCK(fs);
1958 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1959 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1960 	memset(cl, 0, sizeof(*cl));
1961 	cl->fs = fs;
1962 	cl->bpp = bpp;
1963 	cl->bufcount = 0;
1964 	cl->bufsize = 0;
1965 
1966 	/* If this segment is being written synchronously, note that */
1967 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1968 		cl->flags |= LFS_CL_SYNC;
1969 		cl->seg = fs->lfs_sp;
1970 		++cl->seg->seg_iocount;
1971 	}
1972 
1973 	/* Get an empty buffer header, or maybe one with something on it */
1974 	bp = getiobuf(vp, true);
1975 	bp->b_dev = NODEV;
1976 	bp->b_blkno = bp->b_lblkno = addr;
1977 	bp->b_iodone = lfs_cluster_callback;
1978 	bp->b_private = cl;
1979 
1980 	return bp;
1981 }
1982 
1983 int
1984 lfs_writeseg(struct lfs *fs, struct segment *sp)
1985 {
1986 	struct buf **bpp, *bp, *cbp, *newbp, *unbusybp;
1987 	SEGUSE *sup;
1988 	SEGSUM *ssp;
1989 	int i;
1990 	int do_again, nblocks, byteoffset;
1991 	size_t el_size;
1992 	struct lfs_cluster *cl;
1993 	u_short ninos;
1994 	struct vnode *devvp;
1995 	char *p = NULL;
1996 	struct vnode *vp;
1997 	int32_t *daddrp;	/* XXX ondisk32 */
1998 	int changed;
1999 	u_int32_t sum;
2000 #ifdef DEBUG
2001 	FINFO *fip;
2002 	int findex;
2003 #endif
2004 
2005 	ASSERT_SEGLOCK(fs);
2006 
2007 	ssp = (SEGSUM *)sp->segsum;
2008 
2009 	/*
2010 	 * If there are no buffers other than the segment summary to write,
2011 	 * don't do anything.  If we are the end of a dirop sequence, however,
2012 	 * write the empty segment summary anyway, to help out the
2013 	 * roll-forward agent.
2014 	 */
2015 	if ((nblocks = sp->cbpp - sp->bpp) == 1) {
2016 		if ((ssp->ss_flags & (SS_DIROP | SS_CONT)) != SS_DIROP)
2017 			return 0;
2018 	}
2019 
2020 	/* Note if partial segment is being written by the cleaner */
2021 	if (sp->seg_flags & SEGM_CLEAN)
2022 		ssp->ss_flags |= SS_CLEAN;
2023 
2024 	/* Note if we are writing to reclaim */
2025 	if (sp->seg_flags & SEGM_RECLAIM) {
2026 		ssp->ss_flags |= SS_RECLAIM;
2027 		ssp->ss_reclino = fs->lfs_reclino;
2028 	}
2029 
2030 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2031 
2032 	/* Update the segment usage information. */
2033 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
2034 
2035 	/* Loop through all blocks, except the segment summary. */
2036 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
2037 		if ((*bpp)->b_vp != devvp) {
2038 			sup->su_nbytes += (*bpp)->b_bcount;
2039 			DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d"
2040 			      " lbn %" PRId64 " db 0x%" PRIx64 "\n",
2041 			      sp->seg_number, (*bpp)->b_bcount,
2042 			      VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
2043 			      (*bpp)->b_blkno));
2044 		}
2045 	}
2046 
2047 #ifdef DEBUG
2048 	/* Check for zero-length and zero-version FINFO entries. */
2049 	fip = (struct finfo *)((char *)ssp + SEGSUM_SIZE(fs));
2050 	for (findex = 0; findex < ssp->ss_nfinfo; findex++) {
2051 		KDASSERT(fip->fi_nblocks > 0);
2052 		KDASSERT(fip->fi_version > 0);
2053 		fip = (FINFO *)((char *)fip + FINFOSIZE +
2054 			sizeof(int32_t) * fip->fi_nblocks);
2055 	}
2056 #endif /* DEBUG */
2057 
2058 	ninos = (ssp->ss_ninos + LFS_INOPB(fs) - 1) / LFS_INOPB(fs);
2059 	DLOG((DLOG_SU, "seg %d += %d for %d inodes\n",
2060 	      sp->seg_number, ssp->ss_ninos * sizeof (struct ulfs1_dinode),
2061 	      ssp->ss_ninos));
2062 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ulfs1_dinode);
2063 	/* sup->su_nbytes += fs->lfs_sumsize; */
2064 	if (fs->lfs_version == 1)
2065 		sup->su_olastmod = time_second;
2066 	else
2067 		sup->su_lastmod = time_second;
2068 	sup->su_ninos += ninos;
2069 	++sup->su_nsums;
2070 	fs->lfs_avail -= lfs_btofsb(fs, fs->lfs_sumsize);
2071 
2072 	do_again = !(bp->b_flags & B_GATHERED);
2073 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
2074 
2075 	/*
2076 	 * Mark blocks B_BUSY, to prevent then from being changed between
2077 	 * the checksum computation and the actual write.
2078 	 *
2079 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
2080 	 * there are any, replace them with copies that have UNASSIGNED
2081 	 * instead.
2082 	 */
2083 	mutex_enter(&bufcache_lock);
2084 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
2085 		++bpp;
2086 		bp = *bpp;
2087 		if (bp->b_iodone != NULL) {	 /* UBC or malloced buffer */
2088 			bp->b_cflags |= BC_BUSY;
2089 			continue;
2090 		}
2091 
2092 		while (bp->b_cflags & BC_BUSY) {
2093 			DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential"
2094 			      " data summary corruption for ino %d, lbn %"
2095 			      PRId64 "\n",
2096 			      VTOI(bp->b_vp)->i_number, bp->b_lblkno));
2097 			bp->b_cflags |= BC_WANTED;
2098 			cv_wait(&bp->b_busy, &bufcache_lock);
2099 		}
2100 		bp->b_cflags |= BC_BUSY;
2101 		mutex_exit(&bufcache_lock);
2102 		unbusybp = NULL;
2103 
2104 		/*
2105 		 * Check and replace indirect block UNWRITTEN bogosity.
2106 		 * XXX See comment in lfs_writefile.
2107 		 */
2108 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
2109 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
2110 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
2111 			DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n",
2112 			      VTOI(bp->b_vp)->i_number,
2113 			      VTOI(bp->b_vp)->i_lfs_effnblks,
2114 			      VTOI(bp->b_vp)->i_ffs1_blocks));
2115 			/* Make a copy we'll make changes to */
2116 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
2117 					   bp->b_bcount, LFS_NB_IBLOCK);
2118 			newbp->b_blkno = bp->b_blkno;
2119 			memcpy(newbp->b_data, bp->b_data,
2120 			       newbp->b_bcount);
2121 
2122 			changed = 0;
2123 			/* XXX ondisk32 */
2124 			for (daddrp = (int32_t *)(newbp->b_data);
2125 			     daddrp < (int32_t *)((char *)newbp->b_data +
2126 						  newbp->b_bcount); daddrp++) {
2127 				if (*daddrp == UNWRITTEN) {
2128 					++changed;
2129 					*daddrp = 0;
2130 				}
2131 			}
2132 			/*
2133 			 * Get rid of the old buffer.  Don't mark it clean,
2134 			 * though, if it still has dirty data on it.
2135 			 */
2136 			if (changed) {
2137 				DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):"
2138 				      " bp = %p newbp = %p\n", changed, bp,
2139 				      newbp));
2140 				*bpp = newbp;
2141 				bp->b_flags &= ~B_GATHERED;
2142 				bp->b_error = 0;
2143 				if (bp->b_iodone != NULL) {
2144 					DLOG((DLOG_SEG, "lfs_writeseg: "
2145 					      "indir bp should not be B_CALL\n"));
2146 					biodone(bp);
2147 					bp = NULL;
2148 				} else {
2149 					/* Still on free list, leave it there */
2150 					unbusybp = bp;
2151 					/*
2152 					 * We have to re-decrement lfs_avail
2153 					 * since this block is going to come
2154 					 * back around to us in the next
2155 					 * segment.
2156 					 */
2157 					fs->lfs_avail -=
2158 					    lfs_btofsb(fs, bp->b_bcount);
2159 				}
2160 			} else {
2161 				lfs_freebuf(fs, newbp);
2162 			}
2163 		}
2164 		mutex_enter(&bufcache_lock);
2165 		if (unbusybp != NULL) {
2166 			unbusybp->b_cflags &= ~BC_BUSY;
2167 			if (unbusybp->b_cflags & BC_WANTED)
2168 				cv_broadcast(&bp->b_busy);
2169 		}
2170 	}
2171 	mutex_exit(&bufcache_lock);
2172 
2173 	/*
2174 	 * Compute checksum across data and then across summary; the first
2175 	 * block (the summary block) is skipped.  Set the create time here
2176 	 * so that it's guaranteed to be later than the inode mod times.
2177 	 */
2178 	sum = 0;
2179 	if (fs->lfs_version == 1)
2180 		el_size = sizeof(u_long);
2181 	else
2182 		el_size = sizeof(u_int32_t);
2183 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
2184 		++bpp;
2185 		/* Loop through gop_write cluster blocks */
2186 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
2187 		     byteoffset += fs->lfs_bsize) {
2188 #ifdef LFS_USE_B_INVAL
2189 			if (((*bpp)->b_cflags & BC_INVAL) != 0 &&
2190 			    (*bpp)->b_iodone != NULL) {
2191 				if (copyin((void *)(*bpp)->b_saveaddr +
2192 					   byteoffset, dp, el_size)) {
2193 					panic("lfs_writeseg: copyin failed [1]:"
2194 						" ino %d blk %" PRId64,
2195 						VTOI((*bpp)->b_vp)->i_number,
2196 						(*bpp)->b_lblkno);
2197 				}
2198 			} else
2199 #endif /* LFS_USE_B_INVAL */
2200 			{
2201 				sum = lfs_cksum_part((char *)
2202 				    (*bpp)->b_data + byteoffset, el_size, sum);
2203 			}
2204 		}
2205 	}
2206 	if (fs->lfs_version == 1)
2207 		ssp->ss_ocreate = time_second;
2208 	else {
2209 		ssp->ss_create = time_second;
2210 		ssp->ss_serial = ++fs->lfs_serial;
2211 		ssp->ss_ident  = fs->lfs_ident;
2212 	}
2213 	ssp->ss_datasum = lfs_cksum_fold(sum);
2214 	ssp->ss_sumsum = cksum(&ssp->ss_datasum,
2215 	    fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
2216 
2217 	mutex_enter(&lfs_lock);
2218 	fs->lfs_bfree -= (lfs_btofsb(fs, ninos * fs->lfs_ibsize) +
2219 			  lfs_btofsb(fs, fs->lfs_sumsize));
2220 	fs->lfs_dmeta += (lfs_btofsb(fs, ninos * fs->lfs_ibsize) +
2221 			  lfs_btofsb(fs, fs->lfs_sumsize));
2222 	mutex_exit(&lfs_lock);
2223 
2224 	/*
2225 	 * When we simply write the blocks we lose a rotation for every block
2226 	 * written.  To avoid this problem, we cluster the buffers into a
2227 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
2228 	 * devices can handle, use that for the size of the chunks.
2229 	 *
2230 	 * Blocks that are already clusters (from GOP_WRITE), however, we
2231 	 * don't bother to copy into other clusters.
2232 	 */
2233 
2234 #define CHUNKSIZE MAXPHYS
2235 
2236 	if (devvp == NULL)
2237 		panic("devvp is NULL");
2238 	for (bpp = sp->bpp, i = nblocks; i;) {
2239 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
2240 		cl = cbp->b_private;
2241 
2242 		cbp->b_flags |= B_ASYNC;
2243 		cbp->b_cflags |= BC_BUSY;
2244 		cbp->b_bcount = 0;
2245 
2246 #if defined(DEBUG) && defined(DIAGNOSTIC)
2247 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
2248 		    / sizeof(int32_t)) {
2249 			panic("lfs_writeseg: real bpp overwrite");
2250 		}
2251 		if (bpp - sp->bpp > lfs_segsize(fs) / fs->lfs_fsize) {
2252 			panic("lfs_writeseg: theoretical bpp overwrite");
2253 		}
2254 #endif
2255 
2256 		/*
2257 		 * Construct the cluster.
2258 		 */
2259 		mutex_enter(&lfs_lock);
2260 		++fs->lfs_iocount;
2261 		mutex_exit(&lfs_lock);
2262 		while (i && cbp->b_bcount < CHUNKSIZE) {
2263 			bp = *bpp;
2264 
2265 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
2266 				break;
2267 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
2268 				break;
2269 
2270 			/* Clusters from GOP_WRITE are expedited */
2271 			if (bp->b_bcount > fs->lfs_bsize) {
2272 				if (cbp->b_bcount > 0)
2273 					/* Put in its own buffer */
2274 					break;
2275 				else {
2276 					cbp->b_data = bp->b_data;
2277 				}
2278 			} else if (cbp->b_bcount == 0) {
2279 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
2280 							     LFS_NB_CLUSTER);
2281 				cl->flags |= LFS_CL_MALLOC;
2282 			}
2283 #ifdef DIAGNOSTIC
2284 			if (lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno +
2285 					      btodb(bp->b_bcount - 1))) !=
2286 			    sp->seg_number) {
2287 				printf("blk size %d daddr %" PRIx64
2288 				    " not in seg %d\n",
2289 				    bp->b_bcount, bp->b_blkno,
2290 				    sp->seg_number);
2291 				panic("segment overwrite");
2292 			}
2293 #endif
2294 
2295 #ifdef LFS_USE_B_INVAL
2296 			/*
2297 			 * Fake buffers from the cleaner are marked as B_INVAL.
2298 			 * We need to copy the data from user space rather than
2299 			 * from the buffer indicated.
2300 			 * XXX == what do I do on an error?
2301 			 */
2302 			if ((bp->b_cflags & BC_INVAL) != 0 &&
2303 			    bp->b_iodone != NULL) {
2304 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
2305 					panic("lfs_writeseg: "
2306 					    "copyin failed [2]");
2307 			} else
2308 #endif /* LFS_USE_B_INVAL */
2309 			if (cl->flags & LFS_CL_MALLOC) {
2310 				/* copy data into our cluster. */
2311 				memcpy(p, bp->b_data, bp->b_bcount);
2312 				p += bp->b_bcount;
2313 			}
2314 
2315 			cbp->b_bcount += bp->b_bcount;
2316 			cl->bufsize += bp->b_bcount;
2317 
2318 			bp->b_flags &= ~B_READ;
2319 			bp->b_error = 0;
2320 			cl->bpp[cl->bufcount++] = bp;
2321 
2322 			vp = bp->b_vp;
2323 			mutex_enter(&bufcache_lock);
2324 			mutex_enter(vp->v_interlock);
2325 			bp->b_oflags &= ~(BO_DELWRI | BO_DONE);
2326 			reassignbuf(bp, vp);
2327 			vp->v_numoutput++;
2328 			mutex_exit(vp->v_interlock);
2329 			mutex_exit(&bufcache_lock);
2330 
2331 			bpp++;
2332 			i--;
2333 		}
2334 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2335 			BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL);
2336 		else
2337 			BIO_SETPRIO(cbp, BPRIO_TIMELIMITED);
2338 		mutex_enter(devvp->v_interlock);
2339 		devvp->v_numoutput++;
2340 		mutex_exit(devvp->v_interlock);
2341 		VOP_STRATEGY(devvp, cbp);
2342 		curlwp->l_ru.ru_oublock++;
2343 	}
2344 
2345 	if (lfs_dostats) {
2346 		++lfs_stats.psegwrites;
2347 		lfs_stats.blocktot += nblocks - 1;
2348 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2349 			++lfs_stats.psyncwrites;
2350 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2351 			++lfs_stats.pcleanwrites;
2352 			lfs_stats.cleanblocks += nblocks - 1;
2353 		}
2354 	}
2355 
2356 	return (lfs_initseg(fs) || do_again);
2357 }
2358 
2359 void
2360 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2361 {
2362 	struct buf *bp;
2363 	struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2364 	int s;
2365 
2366 	ASSERT_MAYBE_SEGLOCK(fs);
2367 #ifdef DIAGNOSTIC
2368 	KASSERT(fs->lfs_magic == LFS_MAGIC);
2369 #endif
2370 	/*
2371 	 * If we can write one superblock while another is in
2372 	 * progress, we risk not having a complete checkpoint if we crash.
2373 	 * So, block here if a superblock write is in progress.
2374 	 */
2375 	mutex_enter(&lfs_lock);
2376 	s = splbio();
2377 	while (fs->lfs_sbactive) {
2378 		mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0,
2379 			&lfs_lock);
2380 	}
2381 	fs->lfs_sbactive = daddr;
2382 	splx(s);
2383 	mutex_exit(&lfs_lock);
2384 
2385 	/* Set timestamp of this version of the superblock */
2386 	if (fs->lfs_version == 1)
2387 		fs->lfs_otstamp = time_second;
2388 	fs->lfs_tstamp = time_second;
2389 
2390 	/* Checksum the superblock and copy it into a buffer. */
2391 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2392 	bp = lfs_newbuf(fs, devvp,
2393 	    LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2394 	memset((char *)bp->b_data + sizeof(struct dlfs), 0,
2395 	    LFS_SBPAD - sizeof(struct dlfs));
2396 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2397 
2398 	bp->b_cflags |= BC_BUSY;
2399 	bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC;
2400 	bp->b_oflags &= ~(BO_DONE | BO_DELWRI);
2401 	bp->b_error = 0;
2402 	bp->b_iodone = lfs_supercallback;
2403 
2404 	if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC)
2405 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
2406 	else
2407 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
2408 	curlwp->l_ru.ru_oublock++;
2409 
2410 	mutex_enter(devvp->v_interlock);
2411 	devvp->v_numoutput++;
2412 	mutex_exit(devvp->v_interlock);
2413 
2414 	mutex_enter(&lfs_lock);
2415 	++fs->lfs_iocount;
2416 	mutex_exit(&lfs_lock);
2417 	VOP_STRATEGY(devvp, bp);
2418 }
2419 
2420 /*
2421  * Logical block number match routines used when traversing the dirty block
2422  * chain.
2423  */
2424 int
2425 lfs_match_fake(struct lfs *fs, struct buf *bp)
2426 {
2427 
2428 	ASSERT_SEGLOCK(fs);
2429 	return LFS_IS_MALLOC_BUF(bp);
2430 }
2431 
2432 #if 0
2433 int
2434 lfs_match_real(struct lfs *fs, struct buf *bp)
2435 {
2436 
2437 	ASSERT_SEGLOCK(fs);
2438 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2439 }
2440 #endif
2441 
2442 int
2443 lfs_match_data(struct lfs *fs, struct buf *bp)
2444 {
2445 
2446 	ASSERT_SEGLOCK(fs);
2447 	return (bp->b_lblkno >= 0);
2448 }
2449 
2450 int
2451 lfs_match_indir(struct lfs *fs, struct buf *bp)
2452 {
2453 	daddr_t lbn;
2454 
2455 	ASSERT_SEGLOCK(fs);
2456 	lbn = bp->b_lblkno;
2457 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0);
2458 }
2459 
2460 int
2461 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2462 {
2463 	daddr_t lbn;
2464 
2465 	ASSERT_SEGLOCK(fs);
2466 	lbn = bp->b_lblkno;
2467 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1);
2468 }
2469 
2470 int
2471 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2472 {
2473 	daddr_t lbn;
2474 
2475 	ASSERT_SEGLOCK(fs);
2476 	lbn = bp->b_lblkno;
2477 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2);
2478 }
2479 
2480 static void
2481 lfs_free_aiodone(struct buf *bp)
2482 {
2483 	struct lfs *fs;
2484 
2485 	KERNEL_LOCK(1, curlwp);
2486 	fs = bp->b_private;
2487 	ASSERT_NO_SEGLOCK(fs);
2488 	lfs_freebuf(fs, bp);
2489 	KERNEL_UNLOCK_LAST(curlwp);
2490 }
2491 
2492 static void
2493 lfs_super_aiodone(struct buf *bp)
2494 {
2495 	struct lfs *fs;
2496 
2497 	KERNEL_LOCK(1, curlwp);
2498 	fs = bp->b_private;
2499 	ASSERT_NO_SEGLOCK(fs);
2500 	mutex_enter(&lfs_lock);
2501 	fs->lfs_sbactive = 0;
2502 	if (--fs->lfs_iocount <= 1)
2503 		wakeup(&fs->lfs_iocount);
2504 	wakeup(&fs->lfs_sbactive);
2505 	mutex_exit(&lfs_lock);
2506 	lfs_freebuf(fs, bp);
2507 	KERNEL_UNLOCK_LAST(curlwp);
2508 }
2509 
2510 static void
2511 lfs_cluster_aiodone(struct buf *bp)
2512 {
2513 	struct lfs_cluster *cl;
2514 	struct lfs *fs;
2515 	struct buf *tbp, *fbp;
2516 	struct vnode *vp, *devvp, *ovp;
2517 	struct inode *ip;
2518 	int error;
2519 
2520 	KERNEL_LOCK(1, curlwp);
2521 
2522 	error = bp->b_error;
2523 	cl = bp->b_private;
2524 	fs = cl->fs;
2525 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2526 	ASSERT_NO_SEGLOCK(fs);
2527 
2528 	/* Put the pages back, and release the buffer */
2529 	while (cl->bufcount--) {
2530 		tbp = cl->bpp[cl->bufcount];
2531 		KASSERT(tbp->b_cflags & BC_BUSY);
2532 		if (error) {
2533 			tbp->b_error = error;
2534 		}
2535 
2536 		/*
2537 		 * We're done with tbp.	 If it has not been re-dirtied since
2538 		 * the cluster was written, free it.  Otherwise, keep it on
2539 		 * the locked list to be written again.
2540 		 */
2541 		vp = tbp->b_vp;
2542 
2543 		tbp->b_flags &= ~B_GATHERED;
2544 
2545 		LFS_BCLEAN_LOG(fs, tbp);
2546 
2547 		mutex_enter(&bufcache_lock);
2548 		if (tbp->b_iodone == NULL) {
2549 			KASSERT(tbp->b_flags & B_LOCKED);
2550 			bremfree(tbp);
2551 			if (vp) {
2552 				mutex_enter(vp->v_interlock);
2553 				reassignbuf(tbp, vp);
2554 				mutex_exit(vp->v_interlock);
2555 			}
2556 			tbp->b_flags |= B_ASYNC; /* for biodone */
2557 		}
2558 
2559 		if (((tbp->b_flags | tbp->b_oflags) &
2560 		    (B_LOCKED | BO_DELWRI)) == B_LOCKED)
2561 			LFS_UNLOCK_BUF(tbp);
2562 
2563 		if (tbp->b_oflags & BO_DONE) {
2564 			DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n",
2565 				cl->bufcount, (long)tbp->b_flags));
2566 		}
2567 
2568 		if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) {
2569 			/*
2570 			 * A buffer from the page daemon.
2571 			 * We use the same iodone as it does,
2572 			 * so we must manually disassociate its
2573 			 * buffers from the vp.
2574 			 */
2575 			if ((ovp = tbp->b_vp) != NULL) {
2576 				/* This is just silly */
2577 				mutex_enter(ovp->v_interlock);
2578 				brelvp(tbp);
2579 				mutex_exit(ovp->v_interlock);
2580 				tbp->b_vp = vp;
2581 				tbp->b_objlock = vp->v_interlock;
2582 			}
2583 			/* Put it back the way it was */
2584 			tbp->b_flags |= B_ASYNC;
2585 			/* Master buffers have BC_AGE */
2586 			if (tbp->b_private == tbp)
2587 				tbp->b_cflags |= BC_AGE;
2588 		}
2589 		mutex_exit(&bufcache_lock);
2590 
2591 		biodone(tbp);
2592 
2593 		/*
2594 		 * If this is the last block for this vnode, but
2595 		 * there are other blocks on its dirty list,
2596 		 * set IN_MODIFIED/IN_CLEANING depending on what
2597 		 * sort of block.  Only do this for our mount point,
2598 		 * not for, e.g., inode blocks that are attached to
2599 		 * the devvp.
2600 		 * XXX KS - Shouldn't we set *both* if both types
2601 		 * of blocks are present (traverse the dirty list?)
2602 		 */
2603 		mutex_enter(vp->v_interlock);
2604 		mutex_enter(&lfs_lock);
2605 		if (vp != devvp && vp->v_numoutput == 0 &&
2606 		    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2607 			ip = VTOI(vp);
2608 			DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n",
2609 			       ip->i_number));
2610 			if (LFS_IS_MALLOC_BUF(fbp))
2611 				LFS_SET_UINO(ip, IN_CLEANING);
2612 			else
2613 				LFS_SET_UINO(ip, IN_MODIFIED);
2614 		}
2615 		cv_broadcast(&vp->v_cv);
2616 		mutex_exit(&lfs_lock);
2617 		mutex_exit(vp->v_interlock);
2618 	}
2619 
2620 	/* Fix up the cluster buffer, and release it */
2621 	if (cl->flags & LFS_CL_MALLOC)
2622 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2623 	putiobuf(bp);
2624 
2625 	/* Note i/o done */
2626 	if (cl->flags & LFS_CL_SYNC) {
2627 		if (--cl->seg->seg_iocount == 0)
2628 			wakeup(&cl->seg->seg_iocount);
2629 	}
2630 	mutex_enter(&lfs_lock);
2631 #ifdef DIAGNOSTIC
2632 	if (fs->lfs_iocount == 0)
2633 		panic("lfs_cluster_aiodone: zero iocount");
2634 #endif
2635 	if (--fs->lfs_iocount <= 1)
2636 		wakeup(&fs->lfs_iocount);
2637 	mutex_exit(&lfs_lock);
2638 
2639 	KERNEL_UNLOCK_LAST(curlwp);
2640 
2641 	pool_put(&fs->lfs_bpppool, cl->bpp);
2642 	cl->bpp = NULL;
2643 	pool_put(&fs->lfs_clpool, cl);
2644 }
2645 
2646 static void
2647 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2648 {
2649 	/* reset b_iodone for when this is a single-buf i/o. */
2650 	bp->b_iodone = aiodone;
2651 
2652 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
2653 }
2654 
2655 static void
2656 lfs_cluster_callback(struct buf *bp)
2657 {
2658 
2659 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2660 }
2661 
2662 void
2663 lfs_supercallback(struct buf *bp)
2664 {
2665 
2666 	lfs_generic_callback(bp, lfs_super_aiodone);
2667 }
2668 
2669 /*
2670  * The only buffers that are going to hit these functions are the
2671  * segment write blocks, or the segment summaries, or the superblocks.
2672  *
2673  * All of the above are created by lfs_newbuf, and so do not need to be
2674  * released via brelse.
2675  */
2676 void
2677 lfs_callback(struct buf *bp)
2678 {
2679 
2680 	lfs_generic_callback(bp, lfs_free_aiodone);
2681 }
2682 
2683 /*
2684  * Shellsort (diminishing increment sort) from Data Structures and
2685  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2686  * see also Knuth Vol. 3, page 84.  The increments are selected from
2687  * formula (8), page 95.  Roughly O(N^3/2).
2688  */
2689 /*
2690  * This is our own private copy of shellsort because we want to sort
2691  * two parallel arrays (the array of buffer pointers and the array of
2692  * logical block numbers) simultaneously.  Note that we cast the array
2693  * of logical block numbers to a unsigned in this routine so that the
2694  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2695  */
2696 
2697 void
2698 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2699 {
2700 	static int __rsshell_increments[] = { 4, 1, 0 };
2701 	int incr, *incrp, t1, t2;
2702 	struct buf *bp_temp;
2703 
2704 #ifdef DEBUG
2705 	incr = 0;
2706 	for (t1 = 0; t1 < nmemb; t1++) {
2707 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2708 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2709 				/* dump before panic */
2710 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2711 				    nmemb, size);
2712 				incr = 0;
2713 				for (t1 = 0; t1 < nmemb; t1++) {
2714 					const struct buf *bp = bp_array[t1];
2715 
2716 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2717 					    PRIu64 "\n", t1,
2718 					    (uint64_t)bp->b_bcount,
2719 					    (uint64_t)bp->b_lblkno);
2720 					printf("lbns:");
2721 					for (t2 = 0; t2 * size < bp->b_bcount;
2722 					    t2++) {
2723 						printf(" %" PRId32,
2724 						    lb_array[incr++]);
2725 					}
2726 					printf("\n");
2727 				}
2728 				panic("lfs_shellsort: inconsistent input");
2729 			}
2730 		}
2731 	}
2732 #endif
2733 
2734 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2735 		for (t1 = incr; t1 < nmemb; ++t1)
2736 			for (t2 = t1 - incr; t2 >= 0;)
2737 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2738 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2739 					bp_temp = bp_array[t2];
2740 					bp_array[t2] = bp_array[t2 + incr];
2741 					bp_array[t2 + incr] = bp_temp;
2742 					t2 -= incr;
2743 				} else
2744 					break;
2745 
2746 	/* Reform the list of logical blocks */
2747 	incr = 0;
2748 	for (t1 = 0; t1 < nmemb; t1++) {
2749 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2750 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2751 		}
2752 	}
2753 }
2754 
2755 /*
2756  * Call vget with LK_NOWAIT.  If we are the one who holds VI_XLOCK,
2757  * however, we must press on.  Just fake success in that case.
2758  */
2759 int
2760 lfs_vref(struct vnode *vp)
2761 {
2762 	struct lfs *fs;
2763 
2764 	KASSERT(mutex_owned(vp->v_interlock));
2765 
2766 	fs = VTOI(vp)->i_lfs;
2767 
2768 	ASSERT_MAYBE_SEGLOCK(fs);
2769 
2770 	/*
2771 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2772 	 * being able to flush all of the pages from this vnode, which
2773 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2774 	 */
2775 	if (IS_FLUSHING(VTOI(vp)->i_lfs, vp)) {
2776  		++fs->lfs_flushvp_fakevref;
2777 		mutex_exit(vp->v_interlock);
2778  		return 0;
2779  	}
2780 
2781 	return vget(vp, LK_NOWAIT);
2782 }
2783 
2784 /*
2785  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2786  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2787  */
2788 void
2789 lfs_vunref(struct vnode *vp)
2790 {
2791 	struct lfs *fs;
2792 
2793 	fs = VTOI(vp)->i_lfs;
2794 	ASSERT_MAYBE_SEGLOCK(fs);
2795 
2796 	/*
2797 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2798 	 */
2799 	if (IS_FLUSHING(fs, vp) && fs->lfs_flushvp_fakevref) {
2800 		--fs->lfs_flushvp_fakevref;
2801 		return;
2802 	}
2803 
2804 	/* does not call inactive XXX sure it does XXX */
2805 	vrele(vp);
2806 }
2807 
2808 /*
2809  * We use this when we have vnodes that were loaded in solely for cleaning.
2810  * There is no reason to believe that these vnodes will be referenced again
2811  * soon, since the cleaning process is unrelated to normal filesystem
2812  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2813  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2814  * cleaning at the head of the list, instead.
2815  */
2816 void
2817 lfs_vunref_head(struct vnode *vp)
2818 {
2819 
2820 	ASSERT_SEGLOCK(VTOI(vp)->i_lfs);
2821 
2822 	/* does not call inactive XXX sure it does XXX,
2823 	   inserts non-held vnode at head of freelist */
2824 	vrele(vp);
2825 }
2826 
2827 
2828 /*
2829  * Set up an FINFO entry for a new file.  The fip pointer is assumed to
2830  * point at uninitialized space.
2831  */
2832 void
2833 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers)
2834 {
2835 	struct segment *sp = fs->lfs_sp;
2836 
2837 	KASSERT(vers > 0);
2838 
2839 	if (sp->seg_bytes_left < fs->lfs_bsize ||
2840 	    sp->sum_bytes_left < sizeof(struct finfo))
2841 		(void) lfs_writeseg(fs, fs->lfs_sp);
2842 
2843 	sp->sum_bytes_left -= FINFOSIZE;
2844 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
2845 	sp->fip->fi_nblocks = 0;
2846 	sp->fip->fi_ino = ino;
2847 	sp->fip->fi_version = vers;
2848 }
2849 
2850 /*
2851  * Release the FINFO entry, either clearing out an unused entry or
2852  * advancing us to the next available entry.
2853  */
2854 void
2855 lfs_release_finfo(struct lfs *fs)
2856 {
2857 	struct segment *sp = fs->lfs_sp;
2858 
2859 	if (sp->fip->fi_nblocks != 0) {
2860 		sp->fip = (FINFO*)((char *)sp->fip + FINFOSIZE +
2861 			sizeof(int32_t) * sp->fip->fi_nblocks);
2862 		sp->start_lbp = &sp->fip->fi_blocks[0];
2863 	} else {
2864 		sp->sum_bytes_left += FINFOSIZE;
2865 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
2866 	}
2867 }
2868