xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 8a705cc17d9661734a1beae130cdbb7daf4a7d17)
1 /*	$NetBSD: lfs_segment.c,v 1.75 2002/05/17 21:42:38 perseant Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *      This product includes software developed by the NetBSD
21  *      Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 /*
39  * Copyright (c) 1991, 1993
40  *	The Regents of the University of California.  All rights reserved.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. All advertising materials mentioning features or use of this software
51  *    must display the following acknowledgement:
52  *	This product includes software developed by the University of
53  *	California, Berkeley and its contributors.
54  * 4. Neither the name of the University nor the names of its contributors
55  *    may be used to endorse or promote products derived from this software
56  *    without specific prior written permission.
57  *
58  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
59  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
60  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
61  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
62  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
63  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
64  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
65  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
66  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
67  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68  * SUCH DAMAGE.
69  *
70  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.75 2002/05/17 21:42:38 perseant Exp $");
75 
76 #define ivndebug(vp,str) printf("ino %d: %s\n",VTOI(vp)->i_number,(str))
77 
78 #if defined(_KERNEL_OPT)
79 #include "opt_ddb.h"
80 #endif
81 
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/namei.h>
85 #include <sys/kernel.h>
86 #include <sys/resourcevar.h>
87 #include <sys/file.h>
88 #include <sys/stat.h>
89 #include <sys/buf.h>
90 #include <sys/proc.h>
91 #include <sys/conf.h>
92 #include <sys/vnode.h>
93 #include <sys/malloc.h>
94 #include <sys/mount.h>
95 
96 #include <miscfs/specfs/specdev.h>
97 #include <miscfs/fifofs/fifo.h>
98 
99 #include <ufs/ufs/inode.h>
100 #include <ufs/ufs/dir.h>
101 #include <ufs/ufs/ufsmount.h>
102 #include <ufs/ufs/ufs_extern.h>
103 
104 #include <ufs/lfs/lfs.h>
105 #include <ufs/lfs/lfs_extern.h>
106 
107 #include <uvm/uvm_extern.h>
108 
109 extern int count_lock_queue(void);
110 extern struct simplelock vnode_free_list_slock;		/* XXX */
111 
112 static void lfs_cluster_callback(struct buf *);
113 static struct buf **lookahead_pagemove(struct buf **, int, size_t *);
114 
115 /*
116  * Determine if it's OK to start a partial in this segment, or if we need
117  * to go on to a new segment.
118  */
119 #define	LFS_PARTIAL_FITS(fs) \
120 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
121 	fragstofsb((fs), (fs)->lfs_frag))
122 
123 void	 lfs_callback(struct buf *);
124 int	 lfs_gather(struct lfs *, struct segment *,
125 	     struct vnode *, int (*)(struct lfs *, struct buf *));
126 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
127 void	 lfs_iset(struct inode *, ufs_daddr_t, time_t);
128 int	 lfs_match_fake(struct lfs *, struct buf *);
129 int	 lfs_match_data(struct lfs *, struct buf *);
130 int	 lfs_match_dindir(struct lfs *, struct buf *);
131 int	 lfs_match_indir(struct lfs *, struct buf *);
132 int	 lfs_match_tindir(struct lfs *, struct buf *);
133 void	 lfs_newseg(struct lfs *);
134 void	 lfs_shellsort(struct buf **, ufs_daddr_t *, int);
135 void	 lfs_supercallback(struct buf *);
136 void	 lfs_updatemeta(struct segment *);
137 int	 lfs_vref(struct vnode *);
138 void	 lfs_vunref(struct vnode *);
139 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
140 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
141 int	 lfs_writeseg(struct lfs *, struct segment *);
142 void	 lfs_writesuper(struct lfs *, daddr_t);
143 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
144 	    struct segment *sp, int dirops);
145 
146 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
147 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
148 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
149 int	lfs_dirvcount = 0;		/* # active dirops */
150 
151 /* Statistics Counters */
152 int lfs_dostats = 1;
153 struct lfs_stats lfs_stats;
154 
155 extern int locked_queue_count;
156 extern long locked_queue_bytes;
157 
158 /* op values to lfs_writevnodes */
159 #define	VN_REG	        0
160 #define	VN_DIROP	1
161 #define	VN_EMPTY	2
162 #define VN_CLEAN        3
163 
164 #define LFS_MAX_ACTIVE          10
165 
166 /*
167  * XXX KS - Set modification time on the Ifile, so the cleaner can
168  * read the fs mod time off of it.  We don't set IN_UPDATE here,
169  * since we don't really need this to be flushed to disk (and in any
170  * case that wouldn't happen to the Ifile until we checkpoint).
171  */
172 void
173 lfs_imtime(struct lfs *fs)
174 {
175 	struct timespec ts;
176 	struct inode *ip;
177 
178 	TIMEVAL_TO_TIMESPEC(&time, &ts);
179 	ip = VTOI(fs->lfs_ivnode);
180 	ip->i_ffs_mtime = ts.tv_sec;
181 	ip->i_ffs_mtimensec = ts.tv_nsec;
182 }
183 
184 /*
185  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
186  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
187  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
188  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
189  */
190 
191 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
192 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
193 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
194 
195 int
196 lfs_vflush(struct vnode *vp)
197 {
198 	struct inode *ip;
199 	struct lfs *fs;
200 	struct segment *sp;
201 	struct buf *bp, *nbp, *tbp, *tnbp;
202 	int error, s;
203 
204 	ip = VTOI(vp);
205 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
206 
207 	if (ip->i_flag & IN_CLEANING) {
208 #ifdef DEBUG_LFS
209 		ivndebug(vp,"vflush/in_cleaning");
210 #endif
211 		LFS_CLR_UINO(ip, IN_CLEANING);
212 		LFS_SET_UINO(ip, IN_MODIFIED);
213 
214 		/*
215 		 * Toss any cleaning buffers that have real counterparts
216 		 * to avoid losing new data
217 		 */
218 		s = splbio();
219 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
220 			nbp = LIST_NEXT(bp, b_vnbufs);
221 			if (bp->b_flags & B_CALL) {
222 				for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
223 				    tbp = tnbp)
224 				{
225 					tnbp = LIST_NEXT(tbp, b_vnbufs);
226 					if (tbp->b_vp == bp->b_vp
227 					   && tbp->b_lblkno == bp->b_lblkno
228 					   && tbp != bp)
229 					{
230 						fs->lfs_avail += btofsb(fs, bp->b_bcount);
231 						wakeup(&fs->lfs_avail);
232 						lfs_freebuf(bp);
233 						bp = NULL;
234 						break;
235 					}
236 				}
237 			}
238 		}
239 		splx(s);
240 	}
241 
242 	/* If the node is being written, wait until that is done */
243 	s = splbio();
244 	if (WRITEINPROG(vp)) {
245 #ifdef DEBUG_LFS
246 		ivndebug(vp,"vflush/writeinprog");
247 #endif
248 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
249 	}
250 	splx(s);
251 
252 	/* Protect against VXLOCK deadlock in vinvalbuf() */
253 	lfs_seglock(fs, SEGM_SYNC);
254 
255 	/* If we're supposed to flush a freed inode, just toss it */
256 	/* XXX - seglock, so these buffers can't be gathered, right? */
257 	if (ip->i_ffs_mode == 0) {
258 		printf("lfs_vflush: ino %d is freed, not flushing\n",
259 			ip->i_number);
260 		s = splbio();
261 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
262 			nbp = LIST_NEXT(bp, b_vnbufs);
263 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
264 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
265 				wakeup(&fs->lfs_avail);
266 			}
267 			/* Copied from lfs_writeseg */
268 			if (bp->b_flags & B_CALL) {
269 				/* if B_CALL, it was created with newbuf */
270 				lfs_freebuf(bp);
271 				bp = NULL;
272 			} else {
273 				bremfree(bp);
274 				LFS_UNLOCK_BUF(bp);
275 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
276                                          B_GATHERED);
277 				bp->b_flags |= B_DONE;
278 				reassignbuf(bp, vp);
279 				brelse(bp);
280 			}
281 		}
282 		splx(s);
283 		LFS_CLR_UINO(ip, IN_CLEANING);
284 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
285 		ip->i_flag &= ~IN_ALLMOD;
286 		printf("lfs_vflush: done not flushing ino %d\n",
287 			ip->i_number);
288 		lfs_segunlock(fs);
289 		return 0;
290 	}
291 
292 	SET_FLUSHING(fs,vp);
293 	if (fs->lfs_nactive > LFS_MAX_ACTIVE) {
294 		error = lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP);
295 		CLR_FLUSHING(fs,vp);
296 		lfs_segunlock(fs);
297 		return error;
298 	}
299 	sp = fs->lfs_sp;
300 
301 	if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) {
302 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
303 	} else if ((ip->i_flag & IN_CLEANING) &&
304 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
305 #ifdef DEBUG_LFS
306 		ivndebug(vp,"vflush/clean");
307 #endif
308 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
309 	} else if (lfs_dostats) {
310 		if (LIST_FIRST(&vp->v_dirtyblkhd) || (VTOI(vp)->i_flag & IN_ALLMOD))
311 			++lfs_stats.vflush_invoked;
312 #ifdef DEBUG_LFS
313 		ivndebug(vp,"vflush");
314 #endif
315 	}
316 
317 #ifdef DIAGNOSTIC
318 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
319 	if (vp->v_flag & VDIROP) {
320 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
321 		/* panic("VDIROP being flushed...this can\'t happen"); */
322 	}
323 	if (vp->v_usecount < 0) {
324 		printf("usecount=%ld\n", (long)vp->v_usecount);
325 		panic("lfs_vflush: usecount<0");
326 	}
327 #endif
328 
329 	do {
330 		do {
331 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
332 				lfs_writefile(fs, sp, vp);
333 		} while (lfs_writeinode(fs, sp, ip));
334 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
335 
336 	if (lfs_dostats) {
337 		++lfs_stats.nwrites;
338 		if (sp->seg_flags & SEGM_SYNC)
339 			++lfs_stats.nsync_writes;
340 		if (sp->seg_flags & SEGM_CKP)
341 			++lfs_stats.ncheckpoints;
342 	}
343 	/*
344 	 * If we were called from somewhere that has already held the seglock
345 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
346 	 * the write to complete because we are still locked.
347 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
348 	 * we must explicitly wait, if that is the case.
349 	 *
350 	 * We compare the iocount against 1, not 0, because it is
351 	 * artificially incremented by lfs_seglock().
352 	 */
353 	if (fs->lfs_seglock > 1) {
354 		s = splbio();
355 		while (fs->lfs_iocount > 1)
356 			(void)tsleep(&fs->lfs_iocount, PRIBIO + 1,
357 				     "lfs_vflush", 0);
358 		splx(s);
359 	}
360 	lfs_segunlock(fs);
361 
362 	CLR_FLUSHING(fs,vp);
363 	return (0);
364 }
365 
366 #ifdef DEBUG_LFS_VERBOSE
367 # define vndebug(vp,str) if (VTOI(vp)->i_flag & IN_CLEANING) printf("not writing ino %d because %s (op %d)\n",VTOI(vp)->i_number,(str),op)
368 #else
369 # define vndebug(vp,str)
370 #endif
371 
372 int
373 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
374 {
375 	struct inode *ip;
376 	struct vnode *vp;
377 	int inodes_written = 0, only_cleaning;
378 	int needs_unlock;
379 
380 #ifndef LFS_NO_BACKVP_HACK
381 	/* BEGIN HACK */
382 #define	VN_OFFSET	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
383 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
384 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&(LIST_FIRST(&mp->mnt_vnodelist))) - VN_OFFSET))
385 
386 	/* Find last vnode. */
387  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
388 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
389 	     vp = LIST_NEXT(vp, v_mntvnodes));
390 	for (; vp && vp != BEG_OF_VLIST; vp = BACK_VP(vp)) {
391 #else
392 	loop:
393 	LIST_FOREACH(vp, &mp->mnt_vnodelist, v_mntvnodes) {
394 #endif
395 		/*
396 		 * If the vnode that we are about to sync is no longer
397 		 * associated with this mount point, start over.
398 		 */
399 		if (vp->v_mount != mp) {
400 			printf("lfs_writevnodes: starting over\n");
401 			goto loop;
402 		}
403 
404 		ip = VTOI(vp);
405 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
406 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
407 			vndebug(vp,"dirop");
408 			continue;
409 		}
410 
411 		if (op == VN_EMPTY && LIST_FIRST(&vp->v_dirtyblkhd)) {
412 			vndebug(vp,"empty");
413 			continue;
414 		}
415 
416 		if (vp->v_type == VNON) {
417 			continue;
418 		}
419 
420 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
421 		   && vp != fs->lfs_flushvp
422 		   && !(ip->i_flag & IN_CLEANING)) {
423 			vndebug(vp,"cleaning");
424 			continue;
425 		}
426 
427 		if (lfs_vref(vp)) {
428 			vndebug(vp,"vref");
429 			continue;
430 		}
431 
432 		needs_unlock = 0;
433 		if (VOP_ISLOCKED(vp)) {
434 			if (vp != fs->lfs_ivnode &&
435 			    vp->v_lock.lk_lockholder != curproc->p_pid) {
436 #ifdef DEBUG_LFS
437 				printf("lfs_writevnodes: not writing ino %d,"
438 				       " locked by pid %d\n",
439 				       VTOI(vp)->i_number,
440 				       vp->v_lock.lk_lockholder);
441 #endif
442 				lfs_vunref(vp);
443 				continue;
444 			}
445 		} else if (vp != fs->lfs_ivnode) {
446 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
447 			needs_unlock = 1;
448 		}
449 
450 		only_cleaning = 0;
451 		/*
452 		 * Write the inode/file if dirty and it's not the IFILE.
453 		 */
454 		if ((ip->i_flag & IN_ALLMOD) ||
455 		     (LIST_FIRST(&vp->v_dirtyblkhd) != NULL))
456 		{
457 			only_cleaning = ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
458 
459 			if (ip->i_number != LFS_IFILE_INUM
460 			   && LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
461 			{
462 				lfs_writefile(fs, sp, vp);
463 			}
464 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
465 				if (WRITEINPROG(vp)) {
466 #ifdef DEBUG_LFS
467 					ivndebug(vp,"writevnodes/write2");
468 #endif
469 				} else if (!(ip->i_flag & IN_ALLMOD)) {
470 #ifdef DEBUG_LFS
471 					printf("<%d>",ip->i_number);
472 #endif
473 					LFS_SET_UINO(ip, IN_MODIFIED);
474 				}
475 			}
476 			(void) lfs_writeinode(fs, sp, ip);
477 			inodes_written++;
478 		}
479 
480 		if (needs_unlock)
481 			VOP_UNLOCK(vp, 0);
482 
483 		if (lfs_clean_vnhead && only_cleaning)
484 			lfs_vunref_head(vp);
485 		else
486 			lfs_vunref(vp);
487 	}
488 	return inodes_written;
489 }
490 
491 /*
492  * Do a checkpoint.
493  */
494 int
495 lfs_segwrite(struct mount *mp, int flags)
496 {
497 	struct buf *bp;
498 	struct inode *ip;
499 	struct lfs *fs;
500 	struct segment *sp;
501 	struct vnode *vp;
502 	SEGUSE *segusep;
503 	ufs_daddr_t ibno;
504 	int do_ckp, did_ckp, error, i;
505 	int writer_set = 0;
506 	int dirty;
507 	int redo;
508 
509 	fs = VFSTOUFS(mp)->um_lfs;
510 
511 	if (fs->lfs_ronly)
512 		return EROFS;
513 
514 	lfs_imtime(fs);
515 
516 	/* printf("lfs_segwrite: ifile flags are 0x%lx\n",
517 	       (long)(VTOI(fs->lfs_ivnode)->i_flag)); */
518 
519 #if 0
520 	/*
521 	 * If we are not the cleaner, and there is no space available,
522 	 * wait until cleaner writes.
523 	 */
524 	if (!(flags & SEGM_CLEAN) && !(fs->lfs_seglock && fs->lfs_sp &&
525 				      (fs->lfs_sp->seg_flags & SEGM_CLEAN)))
526 	{
527 		while (fs->lfs_avail <= 0) {
528 			LFS_CLEANERINFO(cip, fs, bp);
529 			LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
530 
531 			wakeup(&lfs_allclean_wakeup);
532 			wakeup(&fs->lfs_nextseg);
533 			error = tsleep(&fs->lfs_avail, PRIBIO + 1, "lfs_av2",
534 				       0);
535 			if (error) {
536 				return (error);
537 			}
538 		}
539 	}
540 #endif
541 	/*
542 	 * Allocate a segment structure and enough space to hold pointers to
543 	 * the maximum possible number of buffers which can be described in a
544 	 * single summary block.
545 	 */
546 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
547 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
548 	sp = fs->lfs_sp;
549 
550 	/*
551 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
552 	 * in which case we have to flush *all* buffers off of this vnode.
553 	 * We don't care about other nodes, but write any non-dirop nodes
554 	 * anyway in anticipation of another getnewvnode().
555 	 *
556 	 * If we're cleaning we only write cleaning and ifile blocks, and
557 	 * no dirops, since otherwise we'd risk corruption in a crash.
558 	 */
559 	if (sp->seg_flags & SEGM_CLEAN)
560 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
561 	else {
562 		lfs_writevnodes(fs, mp, sp, VN_REG);
563 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
564 			while (fs->lfs_dirops)
565 				if ((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
566 						"lfs writer", 0)))
567 				{
568 					/* XXX why not segunlock? */
569 					free(sp->bpp, M_SEGMENT);
570 					sp->bpp = NULL;
571 					free(sp, M_SEGMENT);
572 					fs->lfs_sp = NULL;
573 					return (error);
574 				}
575 			fs->lfs_writer++;
576 			writer_set = 1;
577 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
578 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
579 		}
580 	}
581 
582 	/*
583 	 * If we are doing a checkpoint, mark everything since the
584 	 * last checkpoint as no longer ACTIVE.
585 	 */
586 	if (do_ckp) {
587 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
588 		     --ibno >= fs->lfs_cleansz; ) {
589 			dirty = 0;
590 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
591 
592 				panic("lfs_segwrite: ifile read");
593 			segusep = (SEGUSE *)bp->b_data;
594 			for (i = fs->lfs_sepb; i--;) {
595 				if (segusep->su_flags & SEGUSE_ACTIVE) {
596 					segusep->su_flags &= ~SEGUSE_ACTIVE;
597 					++dirty;
598 				}
599 				if (fs->lfs_version > 1)
600 					++segusep;
601 				else
602 					segusep = (SEGUSE *)
603 						((SEGUSE_V1 *)segusep + 1);
604 			}
605 
606 			/* But the current segment is still ACTIVE */
607 			segusep = (SEGUSE *)bp->b_data;
608 			if (dtosn(fs, fs->lfs_curseg) / fs->lfs_sepb ==
609 			    (ibno-fs->lfs_cleansz)) {
610 				if (fs->lfs_version > 1)
611 					segusep[dtosn(fs, fs->lfs_curseg) %
612 					     fs->lfs_sepb].su_flags |=
613 						     SEGUSE_ACTIVE;
614 				else
615 					((SEGUSE *)
616 					 ((SEGUSE_V1 *)(bp->b_data) +
617 					  (dtosn(fs, fs->lfs_curseg) %
618 					   fs->lfs_sepb)))->su_flags
619 						   |= SEGUSE_ACTIVE;
620 				--dirty;
621 			}
622 			if (dirty)
623 				error = LFS_BWRITE_LOG(bp); /* Ifile */
624 			else
625 				brelse(bp);
626 		}
627 	}
628 
629 	did_ckp = 0;
630 	if (do_ckp || fs->lfs_doifile) {
631 		do {
632 			vp = fs->lfs_ivnode;
633 
634 			vget(vp, LK_EXCLUSIVE | LK_CANRECURSE | LK_RETRY);
635 #ifdef DEBUG
636 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
637 #endif
638 			fs->lfs_flags &= ~LFS_IFDIRTY;
639 
640 			ip = VTOI(vp);
641 			/* if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) */
642 				lfs_writefile(fs, sp, vp);
643 			if (ip->i_flag & IN_ALLMOD)
644 				++did_ckp;
645 			redo = lfs_writeinode(fs, sp, ip);
646 
647 			vput(vp);
648 			redo += lfs_writeseg(fs, sp);
649 			redo += (fs->lfs_flags & LFS_IFDIRTY);
650 		} while (redo && do_ckp);
651 
652 		/* The ifile should now be all clear */
653 		if (do_ckp && LIST_FIRST(&vp->v_dirtyblkhd)) {
654 			struct buf *bp;
655 			int s, warned = 0, dopanic = 0;
656 			s = splbio();
657 			for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = LIST_NEXT(bp, b_vnbufs)) {
658 				if (!(bp->b_flags & B_GATHERED)) {
659 					if (!warned)
660 						printf("lfs_segwrite: ifile still has dirty blocks?!\n");
661 					++dopanic;
662 					++warned;
663 					printf("bp=%p, lbn %d, flags 0x%lx\n",
664 						bp, bp->b_lblkno, bp->b_flags);
665 				}
666 			}
667 			if (dopanic)
668 				panic("dirty blocks");
669 			splx(s);
670 		}
671 		LFS_CLR_UINO(ip, IN_ALLMOD);
672 	} else {
673 		(void) lfs_writeseg(fs, sp);
674 	}
675 
676 	/*
677 	 * If the I/O count is non-zero, sleep until it reaches zero.
678 	 * At the moment, the user's process hangs around so we can
679 	 * sleep.
680 	 */
681 	fs->lfs_doifile = 0;
682 	if (writer_set && --fs->lfs_writer == 0)
683 		wakeup(&fs->lfs_dirops);
684 
685 	/*
686 	 * If we didn't write the Ifile, we didn't really do anything.
687 	 * That means that (1) there is a checkpoint on disk and (2)
688 	 * nothing has changed since it was written.
689 	 *
690 	 * Take the flags off of the segment so that lfs_segunlock
691 	 * doesn't have to write the superblock either.
692 	 */
693 	if (did_ckp == 0) {
694 		sp->seg_flags &= ~(SEGM_SYNC|SEGM_CKP);
695 		/* if (do_ckp) printf("lfs_segwrite: no checkpoint\n"); */
696 	}
697 
698 	if (lfs_dostats) {
699 		++lfs_stats.nwrites;
700 		if (sp->seg_flags & SEGM_SYNC)
701 			++lfs_stats.nsync_writes;
702 		if (sp->seg_flags & SEGM_CKP)
703 			++lfs_stats.ncheckpoints;
704 	}
705 	lfs_segunlock(fs);
706 	return (0);
707 }
708 
709 /*
710  * Write the dirty blocks associated with a vnode.
711  */
712 void
713 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
714 {
715 	struct buf *bp;
716 	struct finfo *fip;
717 	IFILE *ifp;
718 
719 
720 	if (sp->seg_bytes_left < fs->lfs_bsize ||
721 	    sp->sum_bytes_left < sizeof(struct finfo))
722 		(void) lfs_writeseg(fs, sp);
723 
724 	sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t);
725 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
726 
727 	if (vp->v_flag & VDIROP)
728 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
729 
730 	fip = sp->fip;
731 	fip->fi_nblocks = 0;
732 	fip->fi_ino = VTOI(vp)->i_number;
733 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
734 	fip->fi_version = ifp->if_version;
735 	brelse(bp);
736 
737 	if (sp->seg_flags & SEGM_CLEAN) {
738 		lfs_gather(fs, sp, vp, lfs_match_fake);
739 		/*
740 		 * For a file being flushed, we need to write *all* blocks.
741 		 * This means writing the cleaning blocks first, and then
742 		 * immediately following with any non-cleaning blocks.
743 		 * The same is true of the Ifile since checkpoints assume
744 		 * that all valid Ifile blocks are written.
745 		 */
746 	   	if (IS_FLUSHING(fs,vp) || VTOI(vp)->i_number == LFS_IFILE_INUM)
747 			lfs_gather(fs, sp, vp, lfs_match_data);
748 	} else
749 		lfs_gather(fs, sp, vp, lfs_match_data);
750 
751 	/*
752 	 * It may not be necessary to write the meta-data blocks at this point,
753 	 * as the roll-forward recovery code should be able to reconstruct the
754 	 * list.
755 	 *
756 	 * We have to write them anyway, though, under two conditions: (1) the
757 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
758 	 * checkpointing.
759 	 */
760 	if (lfs_writeindir
761 	   || IS_FLUSHING(fs,vp)
762 	   || (sp->seg_flags & SEGM_CKP))
763 	{
764 		lfs_gather(fs, sp, vp, lfs_match_indir);
765 		lfs_gather(fs, sp, vp, lfs_match_dindir);
766 		lfs_gather(fs, sp, vp, lfs_match_tindir);
767 	}
768 	fip = sp->fip;
769 	if (fip->fi_nblocks != 0) {
770 		sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) +
771 				   sizeof(ufs_daddr_t) * (fip->fi_nblocks-1));
772 		sp->start_lbp = &sp->fip->fi_blocks[0];
773 	} else {
774 		sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t);
775 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
776 	}
777 }
778 
779 int
780 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
781 {
782 	struct buf *bp, *ibp;
783 	struct dinode *cdp;
784 	IFILE *ifp;
785 	SEGUSE *sup;
786 	ufs_daddr_t daddr;
787 	daddr_t *daddrp;
788 	ino_t ino;
789 	int error, i, ndx, fsb = 0;
790 	int redo_ifile = 0;
791 	struct timespec ts;
792 	int gotblk = 0;
793 
794 	if (!(ip->i_flag & IN_ALLMOD))
795 		return (0);
796 
797 	/* Allocate a new inode block if necessary. */
798 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
799 		/* Allocate a new segment if necessary. */
800 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
801 		    sp->sum_bytes_left < sizeof(ufs_daddr_t))
802 			(void) lfs_writeseg(fs, sp);
803 
804 		/* Get next inode block. */
805 		daddr = fs->lfs_offset;
806 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
807 		sp->ibp = *sp->cbpp++ =
808 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
809 			       fs->lfs_ibsize, 0, 0);
810 		gotblk++;
811 
812 		/* Zero out inode numbers */
813 		for (i = 0; i < INOPB(fs); ++i)
814 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
815 
816 		++sp->start_bpp;
817 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
818 		/* Set remaining space counters. */
819 		sp->seg_bytes_left -= fs->lfs_ibsize;
820 		sp->sum_bytes_left -= sizeof(ufs_daddr_t);
821 		ndx = fs->lfs_sumsize / sizeof(ufs_daddr_t) -
822 			sp->ninodes / INOPB(fs) - 1;
823 		((ufs_daddr_t *)(sp->segsum))[ndx] = daddr;
824 	}
825 
826 	/* Update the inode times and copy the inode onto the inode page. */
827 	TIMEVAL_TO_TIMESPEC(&time, &ts);
828 	/* XXX kludge --- don't redirty the ifile just to put times on it */
829 	if (ip->i_number != LFS_IFILE_INUM)
830 		LFS_ITIMES(ip, &ts, &ts, &ts);
831 
832 	/*
833 	 * If this is the Ifile, and we've already written the Ifile in this
834 	 * partial segment, just overwrite it (it's not on disk yet) and
835 	 * continue.
836 	 *
837 	 * XXX we know that the bp that we get the second time around has
838 	 * already been gathered.
839 	 */
840 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
841 		*(sp->idp) = ip->i_din.ffs_din;
842 		return 0;
843 	}
844 
845 	bp = sp->ibp;
846 	cdp = ((struct dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
847 	*cdp = ip->i_din.ffs_din;
848 #ifdef LFS_IFILE_FRAG_ADDRESSING
849 	if (fs->lfs_version > 1)
850 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
851 #endif
852 
853 	/*
854 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
855 	 * addresses to disk.
856 	 */
857 	if (ip->i_lfs_effnblks != ip->i_ffs_blocks) {
858 #ifdef DEBUG_LFS
859 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
860 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs_blocks);
861 #endif
862 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
863 		     daddrp++) {
864 			if (*daddrp == UNWRITTEN) {
865 #ifdef DEBUG_LFS
866 				printf("lfs_writeinode: wiping UNWRITTEN\n");
867 #endif
868 				*daddrp = 0;
869 			}
870 		}
871 	}
872 
873 	if (ip->i_flag & IN_CLEANING)
874 		LFS_CLR_UINO(ip, IN_CLEANING);
875 	else {
876 		/* XXX IN_ALLMOD */
877 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
878 			     IN_UPDATE);
879 		if (ip->i_lfs_effnblks == ip->i_ffs_blocks)
880 			LFS_CLR_UINO(ip, IN_MODIFIED);
881 #ifdef DEBUG_LFS
882 		else
883 			printf("lfs_writeinode: ino %d: real blks=%d, "
884 			       "eff=%d\n", ip->i_number, ip->i_ffs_blocks,
885 			       ip->i_lfs_effnblks);
886 #endif
887 	}
888 
889 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
890 		sp->idp = ((struct dinode *)bp->b_data) +
891 			(sp->ninodes % INOPB(fs));
892 	if (gotblk) {
893 		LFS_LOCK_BUF(bp);
894 		brelse(bp);
895 	}
896 
897 	/* Increment inode count in segment summary block. */
898 	++((SEGSUM *)(sp->segsum))->ss_ninos;
899 
900 	/* If this page is full, set flag to allocate a new page. */
901 	if (++sp->ninodes % INOPB(fs) == 0)
902 		sp->ibp = NULL;
903 
904 	/*
905 	 * If updating the ifile, update the super-block.  Update the disk
906 	 * address and access times for this inode in the ifile.
907 	 */
908 	ino = ip->i_number;
909 	if (ino == LFS_IFILE_INUM) {
910 		daddr = fs->lfs_idaddr;
911 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
912 	} else {
913 		LFS_IENTRY(ifp, fs, ino, ibp);
914 		daddr = ifp->if_daddr;
915 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
916 #ifdef LFS_DEBUG_NEXTFREE
917 		if (ino > 3 && ifp->if_nextfree) {
918 			vprint("lfs_writeinode",ITOV(ip));
919 			printf("lfs_writeinode: updating free ino %d\n",
920 				ip->i_number);
921 		}
922 #endif
923 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
924 	}
925 
926 	/*
927 	 * Account the inode: it no longer belongs to its former segment,
928 	 * though it will not belong to the new segment until that segment
929 	 * is actually written.
930 	 */
931 #ifdef DEBUG
932 	/*
933 	 * The inode's last address should not be in the current partial
934 	 * segment, except under exceptional circumstances (lfs_writevnodes
935 	 * had to start over, and in the meantime more blocks were written
936 	 * to a vnode).  Although the previous inode won't be accounted in
937 	 * su_nbytes until lfs_writeseg, this shouldn't be a problem as we
938 	 * have more data blocks in the current partial segment.
939 	 */
940 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno))
941 		printf("lfs_writeinode: last inode addr in current pseg "
942 		       "(ino %d daddr 0x%x)\n", ino, daddr);
943 #endif
944 	if (daddr != LFS_UNUSED_DADDR) {
945 		LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
946 #ifdef DIAGNOSTIC
947 		if (sup->su_nbytes < DINODE_SIZE) {
948 			printf("lfs_writeinode: negative bytes "
949 			       "(segment %d short by %d)\n",
950 			       dtosn(fs, daddr),
951 			       (int)DINODE_SIZE - sup->su_nbytes);
952 			panic("lfs_writeinode: negative bytes");
953 			sup->su_nbytes = DINODE_SIZE;
954 		}
955 #endif
956 #ifdef DEBUG_SU_NBYTES
957 		printf("seg %d -= %d for ino %d inode\n",
958 		       dtosn(fs, daddr), DINODE_SIZE, ino);
959 #endif
960 		sup->su_nbytes -= DINODE_SIZE;
961 		redo_ifile =
962 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
963 		if (redo_ifile)
964 			fs->lfs_flags |= LFS_IFDIRTY;
965 		error = LFS_BWRITE_LOG(bp); /* Ifile */
966 	}
967 	return (redo_ifile);
968 }
969 
970 int
971 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
972 {
973 	struct lfs *fs;
974 	int version;
975 
976 	/*
977 	 * If full, finish this segment.  We may be doing I/O, so
978 	 * release and reacquire the splbio().
979 	 */
980 #ifdef DIAGNOSTIC
981 	if (sp->vp == NULL)
982 		panic ("lfs_gatherblock: Null vp in segment");
983 #endif
984 	fs = sp->fs;
985 	if (sp->sum_bytes_left < sizeof(ufs_daddr_t) ||
986 	    sp->seg_bytes_left < bp->b_bcount) {
987 		if (sptr)
988 			splx(*sptr);
989 		lfs_updatemeta(sp);
990 
991 		version = sp->fip->fi_version;
992 		(void) lfs_writeseg(fs, sp);
993 
994 		sp->fip->fi_version = version;
995 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
996 		/* Add the current file to the segment summary. */
997 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
998 		sp->sum_bytes_left -=
999 			sizeof(struct finfo) - sizeof(ufs_daddr_t);
1000 
1001 		if (sptr)
1002 			*sptr = splbio();
1003 		return (1);
1004 	}
1005 
1006 #ifdef DEBUG
1007 	if (bp->b_flags & B_GATHERED) {
1008 		printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n",
1009 		       sp->fip->fi_ino, bp->b_lblkno);
1010 		return (0);
1011 	}
1012 #endif
1013 	/* Insert into the buffer list, update the FINFO block. */
1014 	bp->b_flags |= B_GATHERED;
1015 	bp->b_flags &= ~B_DONE;
1016 
1017 	*sp->cbpp++ = bp;
1018 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
1019 
1020 	sp->sum_bytes_left -= sizeof(ufs_daddr_t);
1021 	sp->seg_bytes_left -= bp->b_bcount;
1022 	return (0);
1023 }
1024 
1025 int
1026 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, int (*match)(struct lfs *, struct buf *))
1027 {
1028 	struct buf *bp;
1029 	int s, count = 0;
1030 
1031 	sp->vp = vp;
1032 	s = splbio();
1033 
1034 #ifndef LFS_NO_BACKBUF_HACK
1035 loop:	LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
1036 #else /* LFS_NO_BACKBUF_HACK */
1037 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1038 # define	BUF_OFFSET	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1039 # define	BACK_BUF(BP)	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1040 # define	BEG_OF_LIST	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1041 /* Find last buffer. */
1042 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1043 	    bp = LIST_NEXT(bp, b_vnbufs));
1044 	for (; bp && bp != BEG_OF_LIST; bp = BACK_BUF(bp)) {
1045 #endif /* LFS_NO_BACKBUF_HACK */
1046 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1047 #ifdef DEBUG_LFS
1048 			if (vp == fs->lfs_ivnode && (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1049 				printf("(%d:%lx)", bp->b_lblkno, bp->b_flags);
1050 #endif
1051 			continue;
1052 		}
1053 		if (vp->v_type == VBLK) {
1054 			/* For block devices, just write the blocks. */
1055 			/* XXX Do we really need to even do this? */
1056 #ifdef DEBUG_LFS
1057 			if (count == 0)
1058 				printf("BLK(");
1059 			printf(".");
1060 #endif
1061 			/* Get the block before bwrite, so we don't corrupt the free list */
1062 			bp->b_flags |= B_BUSY;
1063 			bremfree(bp);
1064 			bwrite(bp);
1065 		} else {
1066 #ifdef DIAGNOSTIC
1067 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1068 				printf("lfs_gather: lbn %d is B_INVAL\n",
1069 					bp->b_lblkno);
1070 				VOP_PRINT(bp->b_vp);
1071 			}
1072 			if (!(bp->b_flags & B_DELWRI))
1073 				panic("lfs_gather: bp not B_DELWRI");
1074 			if (!(bp->b_flags & B_LOCKED)) {
1075 				printf("lfs_gather: lbn %d blk %d"
1076 				       " not B_LOCKED\n", bp->b_lblkno,
1077 				       dbtofsb(fs, bp->b_blkno));
1078 				VOP_PRINT(bp->b_vp);
1079 				panic("lfs_gather: bp not B_LOCKED");
1080 			}
1081 #endif
1082 			if (lfs_gatherblock(sp, bp, &s)) {
1083 				goto loop;
1084 			}
1085 		}
1086 		count++;
1087 	}
1088 	splx(s);
1089 #ifdef DEBUG_LFS
1090 	if (vp->v_type == VBLK && count)
1091 		printf(")\n");
1092 #endif
1093 	lfs_updatemeta(sp);
1094 	sp->vp = NULL;
1095 	return count;
1096 }
1097 
1098 /*
1099  * Update the metadata that points to the blocks listed in the FINFO
1100  * array.
1101  */
1102 void
1103 lfs_updatemeta(struct segment *sp)
1104 {
1105 	SEGUSE *sup;
1106 	struct buf *bp;
1107 	struct lfs *fs;
1108 	struct vnode *vp;
1109 	struct indir a[NIADDR + 2], *ap;
1110 	struct inode *ip;
1111 	ufs_daddr_t daddr, lbn, off;
1112 	daddr_t ooff;
1113 	int error, i, nblocks, num;
1114 	int bb;
1115 
1116 	vp = sp->vp;
1117 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1118 	if (nblocks < 0)
1119 		panic("This is a bad thing\n");
1120 	if (vp == NULL || nblocks == 0)
1121 		return;
1122 
1123 	/* Sort the blocks. */
1124 	/*
1125 	 * XXX KS - We have to sort even if the blocks come from the
1126 	 * cleaner, because there might be other pending blocks on the
1127 	 * same inode...and if we don't sort, and there are fragments
1128 	 * present, blocks may be written in the wrong place.
1129 	 */
1130 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
1131 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
1132 
1133 	/*
1134 	 * Record the length of the last block in case it's a fragment.
1135 	 * If there are indirect blocks present, they sort last.  An
1136 	 * indirect block will be lfs_bsize and its presence indicates
1137 	 * that you cannot have fragments.
1138 	 */
1139 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1140 
1141 	/*
1142 	 * Assign disk addresses, and update references to the logical
1143 	 * block and the segment usage information.
1144 	 */
1145 	fs = sp->fs;
1146 	for (i = nblocks; i--; ++sp->start_bpp) {
1147 		lbn = *sp->start_lbp++;
1148 
1149 		(*sp->start_bpp)->b_blkno = fsbtodb(fs, fs->lfs_offset);
1150 		off = fs->lfs_offset;
1151 		if ((*sp->start_bpp)->b_blkno == (*sp->start_bpp)->b_lblkno) {
1152 			printf("lfs_updatemeta: ino %d blk %d"
1153 			       " has same lbn and daddr\n",
1154 			       VTOI(vp)->i_number, off);
1155 		}
1156 #ifdef DIAGNOSTIC
1157 		if ((*sp->start_bpp)->b_bcount < fs->lfs_bsize && i != 0)
1158 			panic("lfs_updatemeta: fragment is not last block\n");
1159 #endif
1160 		bb = fragstofsb(fs, numfrags(fs, (*sp->start_bpp)->b_bcount));
1161 		fs->lfs_offset += bb;
1162 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1163 		if (daddr > 0)
1164 			daddr = dbtofsb(fs, daddr);
1165 		if (error)
1166 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1167 		ip = VTOI(vp);
1168 		switch (num) {
1169 		case 0:
1170 			ooff = ip->i_ffs_db[lbn];
1171 #ifdef DEBUG
1172 			if (ooff == 0) {
1173 				printf("lfs_updatemeta[1]: warning: writing "
1174 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1175 				       ip->i_number, lbn, off);
1176 			}
1177 #endif
1178 			if (ooff == UNWRITTEN)
1179 				ip->i_ffs_blocks += bb;
1180 			ip->i_ffs_db[lbn] = off;
1181 			break;
1182 		case 1:
1183 			ooff = ip->i_ffs_ib[a[0].in_off];
1184 #ifdef DEBUG
1185 			if (ooff == 0) {
1186 				printf("lfs_updatemeta[2]: warning: writing "
1187 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1188 				       ip->i_number, lbn, off);
1189 			}
1190 #endif
1191 			if (ooff == UNWRITTEN)
1192 				ip->i_ffs_blocks += bb;
1193 			ip->i_ffs_ib[a[0].in_off] = off;
1194 			break;
1195 		default:
1196 			ap = &a[num - 1];
1197 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1198 				panic("lfs_updatemeta: bread bno %d",
1199 				      ap->in_lbn);
1200 
1201 			ooff = ((ufs_daddr_t *)bp->b_data)[ap->in_off];
1202 #if DEBUG
1203 			if (ooff == 0) {
1204 				printf("lfs_updatemeta[3]: warning: writing "
1205 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1206 				       ip->i_number, lbn, off);
1207 			}
1208 #endif
1209 			if (ooff == UNWRITTEN)
1210 				ip->i_ffs_blocks += bb;
1211 			((ufs_daddr_t *)bp->b_data)[ap->in_off] = off;
1212 			(void) VOP_BWRITE(bp);
1213 		}
1214 #ifdef DEBUG
1215 		if (daddr >= fs->lfs_lastpseg && daddr <= off) {
1216 			printf("lfs_updatemeta: ino %d, lbn %d, addr = %x "
1217 			       "in same pseg\n", VTOI(sp->vp)->i_number,
1218 			       (*sp->start_bpp)->b_lblkno, daddr);
1219 		}
1220 #endif
1221 		/* Update segment usage information. */
1222 		if (daddr > 0) {
1223 			LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
1224 #ifdef DIAGNOSTIC
1225 			if (sup->su_nbytes < (*sp->start_bpp)->b_bcount) {
1226 				/* XXX -- Change to a panic. */
1227 				printf("lfs_updatemeta: negative bytes "
1228 				       "(segment %d short by %ld)\n",
1229 				       dtosn(fs, daddr),
1230 				       (*sp->start_bpp)->b_bcount -
1231 				       sup->su_nbytes);
1232 				printf("lfs_updatemeta: ino %d, lbn %d, "
1233 				       "addr = 0x%x\n", VTOI(sp->vp)->i_number,
1234 				       (*sp->start_bpp)->b_lblkno, daddr);
1235 				panic("lfs_updatemeta: negative bytes");
1236 				sup->su_nbytes = (*sp->start_bpp)->b_bcount;
1237 			}
1238 #endif
1239 #ifdef DEBUG_SU_NBYTES
1240 			printf("seg %d -= %ld for ino %d lbn %d db 0x%x\n",
1241 			       dtosn(fs, daddr), (*sp->start_bpp)->b_bcount,
1242 			       VTOI(sp->vp)->i_number,
1243 			       (*sp->start_bpp)->b_lblkno, daddr);
1244 #endif
1245 			sup->su_nbytes -= (*sp->start_bpp)->b_bcount;
1246 			if (!(bp->b_flags & B_GATHERED))
1247 				fs->lfs_flags |= LFS_IFDIRTY;
1248 			error = LFS_BWRITE_LOG(bp); /* Ifile */
1249 		}
1250 	}
1251 }
1252 
1253 /*
1254  * Start a new segment.
1255  */
1256 int
1257 lfs_initseg(struct lfs *fs)
1258 {
1259 	struct segment *sp;
1260 	SEGUSE *sup;
1261 	SEGSUM *ssp;
1262 	struct buf *bp, *sbp;
1263 	int repeat;
1264 
1265 	sp = fs->lfs_sp;
1266 
1267 	repeat = 0;
1268 	/* Advance to the next segment. */
1269 	if (!LFS_PARTIAL_FITS(fs)) {
1270 		/* lfs_avail eats the remaining space */
1271 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1272 						   fs->lfs_curseg);
1273 		/* Wake up any cleaning procs waiting on this file system. */
1274 		wakeup(&lfs_allclean_wakeup);
1275 		wakeup(&fs->lfs_nextseg);
1276 		lfs_newseg(fs);
1277 		repeat = 1;
1278 		fs->lfs_offset = fs->lfs_curseg;
1279 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1280 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1281 		/*
1282 		 * If the segment contains a superblock, update the offset
1283 		 * and summary address to skip over it.
1284 		 */
1285 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1286 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1287 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1288 			sp->seg_bytes_left -= LFS_SBPAD;
1289 		}
1290 		brelse(bp);
1291 		/* Segment zero could also contain the labelpad */
1292 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1293 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1294 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1295 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1296 		}
1297 	} else {
1298 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1299 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1300 				      (fs->lfs_offset - fs->lfs_curseg));
1301 	}
1302 	fs->lfs_lastpseg = fs->lfs_offset;
1303 
1304 	sp->fs = fs;
1305 	sp->ibp = NULL;
1306 	sp->idp = NULL;
1307 	sp->ninodes = 0;
1308 
1309 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1310 	sp->cbpp = sp->bpp;
1311 #ifdef LFS_MALLOC_SUMMARY
1312 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1313 				     fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize);
1314   	sp->segsum = (*sp->cbpp)->b_data;
1315 #else
1316 	sbp = *sp->cbpp = getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1317 				 fsbtodb(fs, fs->lfs_offset), NBPG, 0, 0);
1318 	memset(sbp->b_data, 0x5a, NBPG);
1319 	sp->segsum = (*sp->cbpp)->b_data + NBPG - fs->lfs_sumsize;
1320 #endif
1321 	bzero(sp->segsum, fs->lfs_sumsize);
1322 	sp->start_bpp = ++sp->cbpp;
1323 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1324 
1325 	/* Set point to SEGSUM, initialize it. */
1326 	ssp = sp->segsum;
1327 	ssp->ss_next = fs->lfs_nextseg;
1328 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1329 	ssp->ss_magic = SS_MAGIC;
1330 
1331 	/* Set pointer to first FINFO, initialize it. */
1332 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1333 	sp->fip->fi_nblocks = 0;
1334 	sp->start_lbp = &sp->fip->fi_blocks[0];
1335 	sp->fip->fi_lastlength = 0;
1336 
1337 	sp->seg_bytes_left -= fs->lfs_sumsize;
1338 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1339 
1340 #ifndef LFS_MALLOC_SUMMARY
1341 	LFS_LOCK_BUF(sbp);
1342 	brelse(sbp);
1343 #endif
1344 	return (repeat);
1345 }
1346 
1347 /*
1348  * Return the next segment to write.
1349  */
1350 void
1351 lfs_newseg(struct lfs *fs)
1352 {
1353 	CLEANERINFO *cip;
1354 	SEGUSE *sup;
1355 	struct buf *bp;
1356 	int curseg, isdirty, sn;
1357 
1358 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1359 #ifdef DEBUG_SU_NBYTES
1360 	printf("lfs_newseg: seg %d := 0 in newseg\n",   /* XXXDEBUG */
1361 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1362 #endif
1363 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1364 	sup->su_nbytes = 0;
1365 	sup->su_nsums = 0;
1366 	sup->su_ninos = 0;
1367 	(void) LFS_BWRITE_LOG(bp); /* Ifile */
1368 
1369 	LFS_CLEANERINFO(cip, fs, bp);
1370 	--cip->clean;
1371 	++cip->dirty;
1372 	fs->lfs_nclean = cip->clean;
1373 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1374 
1375 	fs->lfs_lastseg = fs->lfs_curseg;
1376 	fs->lfs_curseg = fs->lfs_nextseg;
1377 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1378 		sn = (sn + 1) % fs->lfs_nseg;
1379 		if (sn == curseg)
1380 			panic("lfs_nextseg: no clean segments");
1381 		LFS_SEGENTRY(sup, fs, sn, bp);
1382 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1383 		brelse(bp);
1384 		if (!isdirty)
1385 			break;
1386 	}
1387 
1388 	++fs->lfs_nactive;
1389 	fs->lfs_nextseg = sntod(fs, sn);
1390 	if (lfs_dostats) {
1391 		++lfs_stats.segsused;
1392 	}
1393 }
1394 
1395 static struct buf **
1396 lookahead_pagemove(struct buf **bpp, int nblocks, size_t *size)
1397 {
1398 	size_t maxsize;
1399 #ifndef LFS_NO_PAGEMOVE
1400 	struct buf *bp;
1401 #endif
1402 
1403 	maxsize = *size;
1404 	*size = 0;
1405 #ifdef LFS_NO_PAGEMOVE
1406 	return bpp;
1407 #else
1408 	while((bp = *bpp) != NULL && *size < maxsize && nblocks--) {
1409 		if(bp->b_flags & B_CALL)
1410 			return bpp;
1411 		if(bp->b_bcount % NBPG)
1412 			return bpp;
1413 		*size += bp->b_bcount;
1414 		++bpp;
1415 	}
1416 	return NULL;
1417 #endif
1418 }
1419 
1420 #define BQUEUES 4 /* XXX */
1421 #define BQ_EMPTY 3 /* XXX */
1422 extern TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
1423 
1424 #define	BUFHASH(dvp, lbn)	\
1425 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
1426 extern LIST_HEAD(bufhashhdr, buf) invalhash;
1427 /*
1428  * Insq/Remq for the buffer hash lists.
1429  */
1430 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
1431 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
1432 
1433 static struct buf *
1434 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1435 {
1436 	struct lfs_cluster *cl;
1437 	struct buf **bpp, *bp;
1438 	int s;
1439 
1440 	cl = (struct lfs_cluster *)malloc(sizeof(*cl), M_SEGMENT, M_WAITOK);
1441 	bpp = (struct buf **)malloc(n*sizeof(*bpp), M_SEGMENT, M_WAITOK);
1442 	memset(cl,0,sizeof(*cl));
1443 	cl->fs = fs;
1444 	cl->bpp = bpp;
1445 	cl->bufcount = 0;
1446 	cl->bufsize = 0;
1447 
1448 	/* Get an empty buffer header, or maybe one with something on it */
1449 	s = splbio();
1450 	if((bp = bufqueues[BQ_EMPTY].tqh_first) != NULL) {
1451 		bremfree(bp);
1452 		/* clear out various other fields */
1453 		bp->b_flags = B_BUSY;
1454 		bp->b_dev = NODEV;
1455 		bp->b_blkno = bp->b_lblkno = 0;
1456 		bp->b_error = 0;
1457 		bp->b_resid = 0;
1458 		bp->b_bcount = 0;
1459 
1460 		/* nuke any credentials we were holding */
1461 		/* XXXXXX */
1462 
1463 		bremhash(bp);
1464 
1465 		/* disassociate us from our vnode, if we had one... */
1466 		if (bp->b_vp)
1467 			brelvp(bp);
1468 	}
1469 	splx(s);
1470 	while (!bp)
1471 		bp = getnewbuf(0, 0);
1472 	s = splbio();
1473 	bgetvp(vp, bp);
1474 	binshash(bp,&invalhash);
1475 	splx(s);
1476 	bp->b_bcount = 0;
1477 	bp->b_blkno = bp->b_lblkno = addr;
1478 
1479 	bp->b_flags |= B_CALL;
1480 	bp->b_iodone = lfs_cluster_callback;
1481 	cl->saveaddr = bp->b_saveaddr; /* XXX is this ever used? */
1482 	bp->b_saveaddr = (caddr_t)cl;
1483 
1484 	return bp;
1485 }
1486 
1487 int
1488 lfs_writeseg(struct lfs *fs, struct segment *sp)
1489 {
1490 	struct buf **bpp, *bp, *cbp, *newbp, **pmlastbpp;
1491 	SEGUSE *sup;
1492 	SEGSUM *ssp;
1493 	dev_t i_dev;
1494 	char *datap, *dp;
1495 	int do_again, i, nblocks, s;
1496 	size_t el_size;
1497  	struct lfs_cluster *cl;
1498 	int (*strategy)(void *);
1499 	struct vop_strategy_args vop_strategy_a;
1500 	u_short ninos;
1501 	struct vnode *devvp;
1502 	char *p;
1503 	struct vnode *vp;
1504 	struct inode *ip;
1505 	size_t pmsize;
1506 	int use_pagemove;
1507 	daddr_t pseg_daddr;
1508 	daddr_t *daddrp;
1509 	int changed;
1510 #if defined(DEBUG) && defined(LFS_PROPELLER)
1511 	static int propeller;
1512 	char propstring[4] = "-\\|/";
1513 
1514 	printf("%c\b",propstring[propeller++]);
1515 	if (propeller == 4)
1516 		propeller = 0;
1517 #endif
1518 	pseg_daddr = (*(sp->bpp))->b_blkno;
1519 
1520 	/*
1521 	 * If there are no buffers other than the segment summary to write
1522 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1523 	 * even if there aren't any buffers, you need to write the superblock.
1524 	 */
1525 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1526 		return (0);
1527 
1528 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1529 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1530 
1531 	/* Update the segment usage information. */
1532 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1533 
1534 	/* Loop through all blocks, except the segment summary. */
1535 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1536 		if ((*bpp)->b_vp != devvp) {
1537 			sup->su_nbytes += (*bpp)->b_bcount;
1538 #ifdef DEBUG_SU_NBYTES
1539 		printf("seg %d += %ld for ino %d lbn %d db 0x%x\n",
1540 		       sp->seg_number, (*bpp)->b_bcount,
1541 		       VTOI((*bpp)->b_vp)->i_number,
1542 		       (*bpp)->b_lblkno, (*bpp)->b_blkno);
1543 #endif
1544 		}
1545 	}
1546 
1547 	ssp = (SEGSUM *)sp->segsum;
1548 
1549 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1550 #ifdef DEBUG_SU_NBYTES
1551 	printf("seg %d += %d for %d inodes\n",   /* XXXDEBUG */
1552 	       sp->seg_number, ssp->ss_ninos * DINODE_SIZE,
1553 	       ssp->ss_ninos);
1554 #endif
1555 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1556 	/* sup->su_nbytes += fs->lfs_sumsize; */
1557 	if (fs->lfs_version == 1)
1558 		sup->su_olastmod = time.tv_sec;
1559 	else
1560 		sup->su_lastmod = time.tv_sec;
1561 	sup->su_ninos += ninos;
1562 	++sup->su_nsums;
1563 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1564 							 fs->lfs_ibsize));
1565 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1566 
1567 	do_again = !(bp->b_flags & B_GATHERED);
1568 	(void)LFS_BWRITE_LOG(bp); /* Ifile */
1569 	/*
1570 	 * Mark blocks B_BUSY, to prevent then from being changed between
1571 	 * the checksum computation and the actual write.
1572 	 *
1573 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1574 	 * there are any, replace them with copies that have UNASSIGNED
1575 	 * instead.
1576 	 */
1577 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1578 		++bpp;
1579 		if ((*bpp)->b_flags & B_CALL)
1580 			continue;
1581 		bp = *bpp;
1582 	    again:
1583 		s = splbio();
1584 		if (bp->b_flags & B_BUSY) {
1585 #ifdef DEBUG
1586 			printf("lfs_writeseg: avoiding potential data "
1587 			       "summary corruption for ino %d, lbn %d\n",
1588 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1589 #endif
1590 			bp->b_flags |= B_WANTED;
1591 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1592 			splx(s);
1593 			goto again;
1594 		}
1595 		bp->b_flags |= B_BUSY;
1596 		splx(s);
1597 		/* Check and replace indirect block UNWRITTEN bogosity */
1598 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1599 		   VTOI(bp->b_vp)->i_ffs_blocks !=
1600 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1601 #ifdef DEBUG_LFS
1602 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1603 			       VTOI(bp->b_vp)->i_number,
1604 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1605 			       VTOI(bp->b_vp)->i_ffs_blocks);
1606 #endif
1607 			/* Make a copy we'll make changes to */
1608 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1609 					   bp->b_bcount);
1610 			newbp->b_blkno = bp->b_blkno;
1611 			memcpy(newbp->b_data, bp->b_data,
1612 			       newbp->b_bcount);
1613 			*bpp = newbp;
1614 
1615 			changed = 0;
1616 			for (daddrp = (daddr_t *)(newbp->b_data);
1617 			     daddrp < (daddr_t *)(newbp->b_data +
1618 						  newbp->b_bcount); daddrp++) {
1619 				if (*daddrp == UNWRITTEN) {
1620 					++changed;
1621 #ifdef DEBUG_LFS
1622 					printf("lfs_writeseg: replacing UNWRITTEN\n");
1623 #endif
1624 					*daddrp = 0;
1625 				}
1626 			}
1627 			/*
1628 			 * Get rid of the old buffer.  Don't mark it clean,
1629 			 * though, if it still has dirty data on it.
1630 			 */
1631 			if (changed) {
1632 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1633 				if (bp->b_flags & B_CALL) {
1634 					lfs_freebuf(bp);
1635 					bp = NULL;
1636 				} else {
1637 					/* Still on free list, leave it there */
1638 					s = splbio();
1639 					bp->b_flags &= ~B_BUSY;
1640 					if (bp->b_flags & B_WANTED)
1641 						wakeup(bp);
1642 				 	splx(s);
1643 					/*
1644 					 * We have to re-decrement lfs_avail
1645 					 * since this block is going to come
1646 					 * back around to us in the next
1647 					 * segment.
1648 					 */
1649 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1650 				}
1651 			} else {
1652 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1653 						 B_GATHERED);
1654 				if (bp->b_flags & B_CALL) {
1655 					lfs_freebuf(bp);
1656 					bp = NULL;
1657 				} else {
1658 					bremfree(bp);
1659 					bp->b_flags |= B_DONE;
1660 					reassignbuf(bp, bp->b_vp);
1661 					LFS_UNLOCK_BUF(bp);
1662 					brelse(bp);
1663 				}
1664 			}
1665 
1666 		}
1667 	}
1668 	/*
1669 	 * Compute checksum across data and then across summary; the first
1670 	 * block (the summary block) is skipped.  Set the create time here
1671 	 * so that it's guaranteed to be later than the inode mod times.
1672 	 *
1673 	 * XXX
1674 	 * Fix this to do it inline, instead of malloc/copy.
1675 	 */
1676 	if (fs->lfs_version == 1)
1677 		el_size = sizeof(u_long);
1678 	else
1679 		el_size = sizeof(u_int32_t);
1680 	datap = dp = malloc(nblocks * el_size, M_SEGMENT, M_WAITOK);
1681 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1682 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1683 			if (copyin((*bpp)->b_saveaddr, dp, el_size))
1684 				panic("lfs_writeseg: copyin failed [1]: "
1685 				      "ino %d blk %d",
1686 				      VTOI((*bpp)->b_vp)->i_number,
1687 				      (*bpp)->b_lblkno);
1688 		} else
1689 			memcpy(dp, (*bpp)->b_data, el_size);
1690 		dp += el_size;
1691 	}
1692 	if (fs->lfs_version == 1)
1693 		ssp->ss_ocreate = time.tv_sec;
1694 	else {
1695 		ssp->ss_create = time.tv_sec;
1696 		ssp->ss_serial = ++fs->lfs_serial;
1697 		ssp->ss_ident  = fs->lfs_ident;
1698 	}
1699 #ifndef LFS_MALLOC_SUMMARY
1700 	/* Set the summary block busy too */
1701 	(*(sp->bpp))->b_flags |= B_BUSY;
1702 #endif
1703 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * el_size);
1704 	ssp->ss_sumsum =
1705 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1706 	free(datap, M_SEGMENT);
1707 	datap = dp = NULL;
1708 #ifdef DIAGNOSTIC
1709 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1710 		panic("lfs_writeseg: No diskspace for summary");
1711 #endif
1712 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1713 			  btofsb(fs, fs->lfs_sumsize));
1714 
1715 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1716 
1717 	/*
1718   	 * When we simply write the blocks we lose a rotation for every block
1719 	 * written.  To avoid this problem, we use pagemove to cluster
1720 	 * the buffers into a chunk and write the chunk.  CHUNKSIZE is the
1721   	 * largest size I/O devices can handle.
1722   	 *
1723 	 * XXX - right now MAXPHYS is only 64k; could it be larger?
1724 	 */
1725 
1726 #define CHUNKSIZE MAXPHYS
1727 
1728 	if (devvp == NULL)
1729 		panic("devvp is NULL");
1730 	for (bpp = sp->bpp, i = nblocks; i;) {
1731 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1732 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1733 
1734 		cbp->b_dev = i_dev;
1735 		cbp->b_flags |= B_ASYNC | B_BUSY;
1736 		cbp->b_bcount = 0;
1737 
1738 		/*
1739 		 * Find out if we can use pagemove to build the cluster,
1740 		 * or if we are stuck using malloc/copy.  If this is the
1741 		 * first cluster, set the shift flag (see below).
1742 		 */
1743 		pmsize = CHUNKSIZE;
1744 		use_pagemove = 0;
1745 		if(bpp == sp->bpp) {
1746 			/* Summary blocks have to get special treatment */
1747 			pmlastbpp = lookahead_pagemove(bpp + 1, i - 1, &pmsize);
1748 			if(pmsize >= CHUNKSIZE - fs->lfs_sumsize ||
1749 			   pmlastbpp == NULL) {
1750 				use_pagemove = 1;
1751 				cl->flags |= LFS_CL_SHIFT;
1752 			} else {
1753 				/*
1754 				 * If we're not using pagemove, we have
1755 				 * to copy the summary down to the bottom
1756 				 * end of the block.
1757 				 */
1758 #ifndef LFS_MALLOC_SUMMARY
1759 				memcpy((*bpp)->b_data, (*bpp)->b_data +
1760 				       NBPG - fs->lfs_sumsize,
1761 				       fs->lfs_sumsize);
1762 #endif /* LFS_MALLOC_SUMMARY */
1763 			}
1764 		} else {
1765 			pmlastbpp = lookahead_pagemove(bpp, i, &pmsize);
1766 			if(pmsize >= CHUNKSIZE || pmlastbpp == NULL) {
1767 				use_pagemove = 1;
1768 			}
1769 		}
1770 		if(use_pagemove == 0) {
1771 			cl->flags |= LFS_CL_MALLOC;
1772 			cl->olddata = cbp->b_data;
1773 			cbp->b_data = malloc(CHUNKSIZE, M_SEGMENT, M_WAITOK);
1774 		}
1775 #if defined(DEBUG) && defined(DIAGNOSTIC)
1776 		if(dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno + btodb((*bpp)->b_bcount - 1))) !=
1777 		   dtosn(fs, dbtofsb(fs, cbp->b_blkno))) {
1778 			printf("block at %x (%d), cbp at %x (%d)\n",
1779 				(*bpp)->b_blkno, dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno)),
1780 			       cbp->b_blkno, dtosn(fs, dbtofsb(fs, cbp->b_blkno)));
1781 			panic("lfs_writeseg: Segment overwrite");
1782 		}
1783 #endif
1784 
1785 		/*
1786 		 * Construct the cluster.
1787 		 */
1788 		s = splbio();
1789 		while (fs->lfs_iocount >= LFS_THROTTLE) {
1790 #ifdef DEBUG_LFS
1791 			printf("[%d]", fs->lfs_iocount);
1792 #endif
1793 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs_throttle", 0);
1794 		}
1795 		++fs->lfs_iocount;
1796 
1797 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1798 			bp = *bpp;
1799 
1800 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1801 				break;
1802 
1803 			/*
1804 			 * Fake buffers from the cleaner are marked as B_INVAL.
1805 			 * We need to copy the data from user space rather than
1806 			 * from the buffer indicated.
1807 			 * XXX == what do I do on an error?
1808 			 */
1809 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1810 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1811 					panic("lfs_writeseg: copyin failed [2]");
1812 			} else if (use_pagemove) {
1813 				pagemove(bp->b_data, p, bp->b_bcount);
1814 				cbp->b_bufsize += bp->b_bcount;
1815 				bp->b_bufsize -= bp->b_bcount;
1816   			} else {
1817 				bcopy(bp->b_data, p, bp->b_bcount);
1818 				/* printf("copy in %p\n", bp->b_data); */
1819   			}
1820 
1821 			/*
1822 			 * XXX If we are *not* shifting, the summary
1823 			 * block is only fs->lfs_sumsize.  Otherwise,
1824 			 * it is NBPG but shifted.
1825 			 */
1826 			if(bpp == sp->bpp && !(cl->flags & LFS_CL_SHIFT)) {
1827 				p += fs->lfs_sumsize;
1828 				cbp->b_bcount += fs->lfs_sumsize;
1829 				cl->bufsize += fs->lfs_sumsize;
1830 			} else {
1831 				p += bp->b_bcount;
1832 				cbp->b_bcount += bp->b_bcount;
1833 				cl->bufsize += bp->b_bcount;
1834 			}
1835 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
1836 			cl->bpp[cl->bufcount++] = bp;
1837 			vp = bp->b_vp;
1838 			++vp->v_numoutput;
1839 
1840 			/*
1841 			 * Although it cannot be freed for reuse before the
1842 			 * cluster is written to disk, this buffer does not
1843 			 * need to be held busy.  Therefore we unbusy it,
1844 			 * while leaving it on the locked list.  It will
1845 			 * be freed or requeued by the callback depending
1846 			 * on whether it has had B_DELWRI set again in the
1847 			 * meantime.
1848 			 *
1849 			 * If we are using pagemove, we have to hold the block
1850 			 * busy to prevent its contents from changing before
1851 			 * it hits the disk, and invalidating the checksum.
1852 			 */
1853 			bp->b_flags &= ~(B_DELWRI | B_READ | B_ERROR);
1854 #ifdef LFS_MNOBUSY
1855 			if (cl->flags & LFS_CL_MALLOC) {
1856 				if (!(bp->b_flags & B_CALL))
1857 					brelse(bp); /* Still B_LOCKED */
1858 			}
1859 #endif
1860 			bpp++;
1861 
1862 			/*
1863 			 * If this is the last block for this vnode, but
1864 			 * there are other blocks on its dirty list,
1865 			 * set IN_MODIFIED/IN_CLEANING depending on what
1866 			 * sort of block.  Only do this for our mount point,
1867 			 * not for, e.g., inode blocks that are attached to
1868 			 * the devvp.
1869 			 * XXX KS - Shouldn't we set *both* if both types
1870 			 * of blocks are present (traverse the dirty list?)
1871 			 */
1872 			if ((i == 1 ||
1873 			     (i > 1 && vp && *bpp && (*bpp)->b_vp != vp)) &&
1874 			    (bp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL &&
1875 			    vp->v_mount == fs->lfs_ivnode->v_mount)
1876   			{
1877 				ip = VTOI(vp);
1878 #ifdef DEBUG_LFS
1879 				printf("lfs_writeseg: marking ino %d\n",
1880 				       ip->i_number);
1881 #endif
1882 				if (bp->b_flags & B_CALL)
1883 					LFS_SET_UINO(ip, IN_CLEANING);
1884 				else
1885 					LFS_SET_UINO(ip, IN_MODIFIED);
1886 			}
1887 			wakeup(vp);
1888 		}
1889 		++cbp->b_vp->v_numoutput;
1890 		splx(s);
1891 		/*
1892 		 * In order to include the summary in a clustered block,
1893 		 * it may be necessary to shift the block forward (since
1894 		 * summary blocks are in generay smaller than can be
1895 		 * addressed by pagemove().  After the write, the block
1896 		 * will be corrected before disassembly.
1897 		 */
1898 		if(cl->flags & LFS_CL_SHIFT) {
1899 			cbp->b_data += (NBPG - fs->lfs_sumsize);
1900 			cbp->b_bcount -= (NBPG - fs->lfs_sumsize);
1901 		}
1902 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1903 		vop_strategy_a.a_bp = cbp;
1904 		(strategy)(&vop_strategy_a);
1905 	}
1906 
1907 	if (lfs_dostats) {
1908 		++lfs_stats.psegwrites;
1909 		lfs_stats.blocktot += nblocks - 1;
1910 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1911 			++lfs_stats.psyncwrites;
1912 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1913 			++lfs_stats.pcleanwrites;
1914 			lfs_stats.cleanblocks += nblocks - 1;
1915 		}
1916 	}
1917 	return (lfs_initseg(fs) || do_again);
1918 }
1919 
1920 void
1921 lfs_writesuper(struct lfs *fs, daddr_t daddr)
1922 {
1923 	struct buf *bp;
1924 	dev_t i_dev;
1925 	int (*strategy)(void *);
1926 	int s;
1927 	struct vop_strategy_args vop_strategy_a;
1928 
1929 	/*
1930 	 * If we can write one superblock while another is in
1931 	 * progress, we risk not having a complete checkpoint if we crash.
1932 	 * So, block here if a superblock write is in progress.
1933 	 */
1934 	s = splbio();
1935 	while (fs->lfs_sbactive) {
1936 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
1937 	}
1938 	fs->lfs_sbactive = daddr;
1939 	splx(s);
1940 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1941 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
1942 
1943 	/* Set timestamp of this version of the superblock */
1944 	if (fs->lfs_version == 1)
1945 		fs->lfs_otstamp = time.tv_sec;
1946 	fs->lfs_tstamp = time.tv_sec;
1947 
1948 	/* Checksum the superblock and copy it into a buffer. */
1949 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
1950 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD);
1951 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
1952 
1953 	bp->b_dev = i_dev;
1954 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
1955 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
1956 	bp->b_iodone = lfs_supercallback;
1957 	/* XXX KS - same nasty hack as above */
1958 	bp->b_saveaddr = (caddr_t)fs;
1959 
1960 	vop_strategy_a.a_desc = VDESC(vop_strategy);
1961 	vop_strategy_a.a_bp = bp;
1962 	s = splbio();
1963 	++bp->b_vp->v_numoutput;
1964 	++fs->lfs_iocount;
1965 	splx(s);
1966 	(strategy)(&vop_strategy_a);
1967 }
1968 
1969 /*
1970  * Logical block number match routines used when traversing the dirty block
1971  * chain.
1972  */
1973 int
1974 lfs_match_fake(struct lfs *fs, struct buf *bp)
1975 {
1976 	return (bp->b_flags & B_CALL);
1977 }
1978 
1979 int
1980 lfs_match_data(struct lfs *fs, struct buf *bp)
1981 {
1982 	return (bp->b_lblkno >= 0);
1983 }
1984 
1985 int
1986 lfs_match_indir(struct lfs *fs, struct buf *bp)
1987 {
1988 	int lbn;
1989 
1990 	lbn = bp->b_lblkno;
1991 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
1992 }
1993 
1994 int
1995 lfs_match_dindir(struct lfs *fs, struct buf *bp)
1996 {
1997 	int lbn;
1998 
1999 	lbn = bp->b_lblkno;
2000 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2001 }
2002 
2003 int
2004 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2005 {
2006 	int lbn;
2007 
2008 	lbn = bp->b_lblkno;
2009 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2010 }
2011 
2012 /*
2013  * XXX - The only buffers that are going to hit these functions are the
2014  * segment write blocks, or the segment summaries, or the superblocks.
2015  *
2016  * All of the above are created by lfs_newbuf, and so do not need to be
2017  * released via brelse.
2018  */
2019 void
2020 lfs_callback(struct buf *bp)
2021 {
2022 	/* struct lfs *fs; */
2023 	/* fs = (struct lfs *)bp->b_saveaddr; */
2024 	lfs_freebuf(bp);
2025 }
2026 
2027 void
2028 lfs_supercallback(struct buf *bp)
2029 {
2030 	struct lfs *fs;
2031 
2032 	fs = (struct lfs *)bp->b_saveaddr;
2033 	fs->lfs_sbactive = 0;
2034 	wakeup(&fs->lfs_sbactive);
2035 	if (--fs->lfs_iocount < LFS_THROTTLE)
2036 		wakeup(&fs->lfs_iocount);
2037 	lfs_freebuf(bp);
2038 }
2039 
2040 static void
2041 lfs_cluster_callback(struct buf *bp)
2042 {
2043 	struct lfs_cluster *cl;
2044 	struct lfs *fs;
2045 	struct buf *tbp;
2046 	struct vnode *vp;
2047 	int error=0;
2048 	char *cp;
2049 	extern int locked_queue_count;
2050 	extern long locked_queue_bytes;
2051 
2052 	if(bp->b_flags & B_ERROR)
2053 		error = bp->b_error;
2054 
2055 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2056 	fs = cl->fs;
2057 	bp->b_saveaddr = cl->saveaddr;
2058 
2059 	/* If shifted, shift back now */
2060 	if(cl->flags & LFS_CL_SHIFT) {
2061 		bp->b_data -= (NBPG - fs->lfs_sumsize);
2062 		bp->b_bcount += (NBPG - fs->lfs_sumsize);
2063 	}
2064 
2065 	cp = (char *)bp->b_data + cl->bufsize;
2066 	/* Put the pages back, and release the buffer */
2067 	while(cl->bufcount--) {
2068 		tbp = cl->bpp[cl->bufcount];
2069 		if(!(cl->flags & LFS_CL_MALLOC)) {
2070 			cp -= tbp->b_bcount;
2071 			printf("pm(%p,%p,%lx)",cp,tbp->b_data,tbp->b_bcount);
2072 			pagemove(cp, tbp->b_data, tbp->b_bcount);
2073 			bp->b_bufsize -= tbp->b_bcount;
2074 			tbp->b_bufsize += tbp->b_bcount;
2075 		}
2076 		if(error) {
2077 			tbp->b_flags |= B_ERROR;
2078 			tbp->b_error = error;
2079 		}
2080 
2081 		/*
2082 		 * We're done with tbp.  If it has not been re-dirtied since
2083 		 * the cluster was written, free it.  Otherwise, keep it on
2084 		 * the locked list to be written again.
2085 		 */
2086 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2087 			LFS_UNLOCK_BUF(tbp);
2088 		tbp->b_flags &= ~B_GATHERED;
2089 
2090 		LFS_BCLEAN_LOG(fs, tbp);
2091 
2092 		vp = tbp->b_vp;
2093 		/* Segment summary for a shifted cluster */
2094 		if(!cl->bufcount && (cl->flags & LFS_CL_SHIFT))
2095 			tbp->b_flags |= B_INVAL;
2096 		if(!(tbp->b_flags & B_CALL)) {
2097 			bremfree(tbp);
2098 			if(vp)
2099 				reassignbuf(tbp, vp);
2100 			tbp->b_flags |= B_ASYNC; /* for biodone */
2101 		}
2102 #ifdef DIAGNOSTIC
2103 		if (tbp->b_flags & B_DONE) {
2104 			printf("blk %d biodone already (flags %lx)\n",
2105 				cl->bufcount, (long)tbp->b_flags);
2106 		}
2107 #endif
2108 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2109 			biodone(tbp);
2110 		}
2111 	}
2112 
2113 	/* Fix up the cluster buffer, and release it */
2114 	if(!(cl->flags & LFS_CL_MALLOC) && bp->b_bufsize) {
2115 		printf("PM(%p,%p,%lx)", (char *)bp->b_data + bp->b_bcount,
2116 			 (char *)bp->b_data, bp->b_bufsize);
2117 		pagemove((char *)bp->b_data + bp->b_bcount,
2118 			 (char *)bp->b_data, bp->b_bufsize);
2119 	}
2120 	if(cl->flags & LFS_CL_MALLOC) {
2121 		free(bp->b_data, M_SEGMENT);
2122 		bp->b_data = cl->olddata;
2123 	}
2124 	bp->b_bcount = 0;
2125 	bp->b_iodone = NULL;
2126 	bp->b_flags &= ~B_DELWRI;
2127 	bp->b_flags |= B_DONE;
2128 	reassignbuf(bp, bp->b_vp);
2129 	brelse(bp);
2130 
2131 	free(cl->bpp, M_SEGMENT);
2132 	free(cl, M_SEGMENT);
2133 
2134 #ifdef DIAGNOSTIC
2135 	if (fs->lfs_iocount == 0)
2136 		panic("lfs_callback: zero iocount\n");
2137 #endif
2138 	if (--fs->lfs_iocount < LFS_THROTTLE)
2139 		wakeup(&fs->lfs_iocount);
2140 #if 0
2141 	if (fs->lfs_iocount == 0) {
2142 		/*
2143 		 * XXX - do we really want to do this in a callback?
2144 		 *
2145 		 * Vinvalbuf can move locked buffers off the locked queue
2146 		 * and we have no way of knowing about this.  So, after
2147 		 * doing a big write, we recalculate how many buffers are
2148 		 * really still left on the locked queue.
2149 		 */
2150 		lfs_countlocked(&locked_queue_count, &locked_queue_bytes, "lfs_cluster_callback");
2151 		wakeup(&locked_queue_count);
2152 	}
2153 #endif
2154 }
2155 
2156 /*
2157  * Shellsort (diminishing increment sort) from Data Structures and
2158  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2159  * see also Knuth Vol. 3, page 84.  The increments are selected from
2160  * formula (8), page 95.  Roughly O(N^3/2).
2161  */
2162 /*
2163  * This is our own private copy of shellsort because we want to sort
2164  * two parallel arrays (the array of buffer pointers and the array of
2165  * logical block numbers) simultaneously.  Note that we cast the array
2166  * of logical block numbers to a unsigned in this routine so that the
2167  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2168  */
2169 
2170 void
2171 lfs_shellsort(struct buf **bp_array, ufs_daddr_t *lb_array, int nmemb)
2172 {
2173 	static int __rsshell_increments[] = { 4, 1, 0 };
2174 	int incr, *incrp, t1, t2;
2175 	struct buf *bp_temp;
2176 	u_long lb_temp;
2177 
2178 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2179 		for (t1 = incr; t1 < nmemb; ++t1)
2180 			for (t2 = t1 - incr; t2 >= 0;)
2181 				if (lb_array[t2] > lb_array[t2 + incr]) {
2182 					lb_temp = lb_array[t2];
2183 					lb_array[t2] = lb_array[t2 + incr];
2184 					lb_array[t2 + incr] = lb_temp;
2185 					bp_temp = bp_array[t2];
2186 					bp_array[t2] = bp_array[t2 + incr];
2187 					bp_array[t2 + incr] = bp_temp;
2188 					t2 -= incr;
2189 				} else
2190 					break;
2191 }
2192 
2193 /*
2194  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2195  */
2196 int
2197 lfs_vref(struct vnode *vp)
2198 {
2199 	/*
2200 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2201 	 * being able to flush all of the pages from this vnode, which
2202 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2203 	 */
2204 	if (vp->v_flag & VXLOCK) {
2205 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2206 			return 0;
2207 		}
2208 		return (1);
2209 	}
2210 	return (vget(vp, 0));
2211 }
2212 
2213 /*
2214  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2215  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2216  */
2217 void
2218 lfs_vunref(struct vnode *vp)
2219 {
2220 	/*
2221 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2222 	 */
2223 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2224 		return;
2225 	}
2226 
2227 	simple_lock(&vp->v_interlock);
2228 #ifdef DIAGNOSTIC
2229 	if (vp->v_usecount <= 0) {
2230 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2231 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2232 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2233 		panic("lfs_vunref: v_usecount<0");
2234 	}
2235 #endif
2236 	vp->v_usecount--;
2237 	if (vp->v_usecount > 0) {
2238 		simple_unlock(&vp->v_interlock);
2239 		return;
2240 	}
2241 	/*
2242 	 * insert at tail of LRU list
2243 	 */
2244 	simple_lock(&vnode_free_list_slock);
2245 	if (vp->v_holdcnt > 0)
2246 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2247 	else
2248 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2249 	simple_unlock(&vnode_free_list_slock);
2250 	simple_unlock(&vp->v_interlock);
2251 }
2252 
2253 /*
2254  * We use this when we have vnodes that were loaded in solely for cleaning.
2255  * There is no reason to believe that these vnodes will be referenced again
2256  * soon, since the cleaning process is unrelated to normal filesystem
2257  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2258  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2259  * cleaning at the head of the list, instead.
2260  */
2261 void
2262 lfs_vunref_head(struct vnode *vp)
2263 {
2264 	simple_lock(&vp->v_interlock);
2265 #ifdef DIAGNOSTIC
2266 	if (vp->v_usecount == 0) {
2267 		panic("lfs_vunref: v_usecount<0");
2268 	}
2269 #endif
2270 	vp->v_usecount--;
2271 	if (vp->v_usecount > 0) {
2272 		simple_unlock(&vp->v_interlock);
2273 		return;
2274 	}
2275 	/*
2276 	 * insert at head of LRU list
2277 	 */
2278 	simple_lock(&vnode_free_list_slock);
2279 	TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2280 	simple_unlock(&vnode_free_list_slock);
2281 	simple_unlock(&vp->v_interlock);
2282 }
2283 
2284