xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 220b5c059a84c51ea44107ea8951a57ffaecdc8c)
1 /*	$NetBSD: lfs_segment.c,v 1.73 2001/11/23 21:44:27 chs 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 1.73 2001/11/23 21:44:27 chs 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 extern int count_lock_queue(void);
108 extern struct simplelock vnode_free_list_slock;		/* XXX */
109 
110 /*
111  * Determine if it's OK to start a partial in this segment, or if we need
112  * to go on to a new segment.
113  */
114 #define	LFS_PARTIAL_FITS(fs) \
115 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
116 	fragstofsb((fs), (fs)->lfs_frag))
117 
118 void	 lfs_callback(struct buf *);
119 int	 lfs_gather(struct lfs *, struct segment *,
120 	     struct vnode *, int (*)(struct lfs *, struct buf *));
121 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
122 void	 lfs_iset(struct inode *, ufs_daddr_t, time_t);
123 int	 lfs_match_fake(struct lfs *, struct buf *);
124 int	 lfs_match_data(struct lfs *, struct buf *);
125 int	 lfs_match_dindir(struct lfs *, struct buf *);
126 int	 lfs_match_indir(struct lfs *, struct buf *);
127 int	 lfs_match_tindir(struct lfs *, struct buf *);
128 void	 lfs_newseg(struct lfs *);
129 void	 lfs_shellsort(struct buf **, ufs_daddr_t *, int);
130 void	 lfs_supercallback(struct buf *);
131 void	 lfs_updatemeta(struct segment *);
132 int	 lfs_vref(struct vnode *);
133 void	 lfs_vunref(struct vnode *);
134 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
135 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
136 int	 lfs_writeseg(struct lfs *, struct segment *);
137 void	 lfs_writesuper(struct lfs *, daddr_t);
138 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
139 	    struct segment *sp, int dirops);
140 
141 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
142 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
143 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
144 int	lfs_dirvcount = 0;		/* # active dirops */
145 
146 /* Statistics Counters */
147 int lfs_dostats = 1;
148 struct lfs_stats lfs_stats;
149 
150 extern int locked_queue_count;
151 extern long locked_queue_bytes;
152 
153 /* op values to lfs_writevnodes */
154 #define	VN_REG	        0
155 #define	VN_DIROP	1
156 #define	VN_EMPTY	2
157 #define VN_CLEAN        3
158 
159 #define LFS_MAX_ACTIVE          10
160 
161 /*
162  * XXX KS - Set modification time on the Ifile, so the cleaner can
163  * read the fs mod time off of it.  We don't set IN_UPDATE here,
164  * since we don't really need this to be flushed to disk (and in any
165  * case that wouldn't happen to the Ifile until we checkpoint).
166  */
167 void
168 lfs_imtime(struct lfs *fs)
169 {
170 	struct timespec ts;
171 	struct inode *ip;
172 
173 	TIMEVAL_TO_TIMESPEC(&time, &ts);
174 	ip = VTOI(fs->lfs_ivnode);
175 	ip->i_ffs_mtime = ts.tv_sec;
176 	ip->i_ffs_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 SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
187 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
188 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
189 
190 int
191 lfs_vflush(struct vnode *vp)
192 {
193 	struct inode *ip;
194 	struct lfs *fs;
195 	struct segment *sp;
196 	struct buf *bp, *nbp, *tbp, *tnbp;
197 	int error, s;
198 
199 	ip = VTOI(vp);
200 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
201 
202 	if (ip->i_flag & IN_CLEANING) {
203 #ifdef DEBUG_LFS
204 		ivndebug(vp,"vflush/in_cleaning");
205 #endif
206 		LFS_CLR_UINO(ip, IN_CLEANING);
207 		LFS_SET_UINO(ip, IN_MODIFIED);
208 
209 		/*
210 		 * Toss any cleaning buffers that have real counterparts
211 		 * to avoid losing new data
212 		 */
213 		s = splbio();
214 		for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = nbp) {
215 			nbp = bp->b_vnbufs.le_next;
216 			if (bp->b_flags & B_CALL) {
217 				for (tbp = vp->v_dirtyblkhd.lh_first; tbp;
218 				    tbp = tnbp)
219 				{
220 					tnbp = tbp->b_vnbufs.le_next;
221 					if (tbp->b_vp == bp->b_vp
222 					   && tbp->b_lblkno == bp->b_lblkno
223 					   && tbp != bp)
224 					{
225 						fs->lfs_avail += btofsb(fs, bp->b_bcount);
226 						wakeup(&fs->lfs_avail);
227 						lfs_freebuf(bp);
228 						bp = NULL;
229 						break;
230 					}
231 				}
232 			}
233 		}
234 		splx(s);
235 	}
236 
237 	/* If the node is being written, wait until that is done */
238 	if (WRITEINPROG(vp)) {
239 #ifdef DEBUG_LFS
240 		ivndebug(vp,"vflush/writeinprog");
241 #endif
242 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
243 	}
244 
245 	/* Protect against VXLOCK deadlock in vinvalbuf() */
246 	lfs_seglock(fs, SEGM_SYNC);
247 
248 	/* If we're supposed to flush a freed inode, just toss it */
249 	/* XXX - seglock, so these buffers can't be gathered, right? */
250 	if (ip->i_ffs_mode == 0) {
251 		printf("lfs_vflush: ino %d is freed, not flushing\n",
252 			ip->i_number);
253 		s = splbio();
254 		for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = nbp) {
255 			nbp = bp->b_vnbufs.le_next;
256 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
257 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
258 				wakeup(&fs->lfs_avail);
259 			}
260 			/* Copied from lfs_writeseg */
261 			if (bp->b_flags & B_CALL) {
262 				/* if B_CALL, it was created with newbuf */
263 				lfs_freebuf(bp);
264 				bp = NULL;
265 			} else {
266 				bremfree(bp);
267 				LFS_UNLOCK_BUF(bp);
268 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
269                                          B_GATHERED);
270 				bp->b_flags |= B_DONE;
271 				reassignbuf(bp, vp);
272 				brelse(bp);
273 			}
274 		}
275 		splx(s);
276 		LFS_CLR_UINO(ip, IN_CLEANING);
277 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
278 		ip->i_flag &= ~IN_ALLMOD;
279 		printf("lfs_vflush: done not flushing ino %d\n",
280 			ip->i_number);
281 		lfs_segunlock(fs);
282 		return 0;
283 	}
284 
285 	SET_FLUSHING(fs,vp);
286 	if (fs->lfs_nactive > LFS_MAX_ACTIVE) {
287 		error = lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP);
288 		CLR_FLUSHING(fs,vp);
289 		lfs_segunlock(fs);
290 		return error;
291 	}
292 	sp = fs->lfs_sp;
293 
294 	if (vp->v_dirtyblkhd.lh_first == NULL) {
295 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
296 	} else if ((ip->i_flag & IN_CLEANING) &&
297 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
298 #ifdef DEBUG_LFS
299 		ivndebug(vp,"vflush/clean");
300 #endif
301 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
302 	}
303 	else if (lfs_dostats) {
304 		if (vp->v_dirtyblkhd.lh_first || (VTOI(vp)->i_flag & IN_ALLMOD))
305 			++lfs_stats.vflush_invoked;
306 #ifdef DEBUG_LFS
307 		ivndebug(vp,"vflush");
308 #endif
309 	}
310 
311 #ifdef DIAGNOSTIC
312 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
313 	if (vp->v_flag & VDIROP) {
314 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
315 		/* panic("VDIROP being flushed...this can\'t happen"); */
316 	}
317 	if (vp->v_usecount < 0) {
318 		printf("usecount=%ld\n", (long)vp->v_usecount);
319 		panic("lfs_vflush: usecount<0");
320 	}
321 #endif
322 
323 	do {
324 		do {
325 			if (vp->v_dirtyblkhd.lh_first != NULL)
326 				lfs_writefile(fs, sp, vp);
327 		} while (lfs_writeinode(fs, sp, ip));
328 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
329 
330 	if (lfs_dostats) {
331 		++lfs_stats.nwrites;
332 		if (sp->seg_flags & SEGM_SYNC)
333 			++lfs_stats.nsync_writes;
334 		if (sp->seg_flags & SEGM_CKP)
335 			++lfs_stats.ncheckpoints;
336 	}
337 	lfs_segunlock(fs);
338 
339 	CLR_FLUSHING(fs,vp);
340 	return (0);
341 }
342 
343 #ifdef DEBUG_LFS_VERBOSE
344 # 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)
345 #else
346 # define vndebug(vp,str)
347 #endif
348 
349 int
350 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
351 {
352 	struct inode *ip;
353 	struct vnode *vp;
354 	int inodes_written = 0, only_cleaning;
355 	int needs_unlock;
356 
357 #ifndef LFS_NO_BACKVP_HACK
358 	/* BEGIN HACK */
359 #define	VN_OFFSET	(((caddr_t)&vp->v_mntvnodes.le_next) - (caddr_t)vp)
360 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)VP->v_mntvnodes.le_prev) - VN_OFFSET))
361 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&mp->mnt_vnodelist.lh_first) - VN_OFFSET))
362 
363 	/* Find last vnode. */
364  loop:	for (vp = mp->mnt_vnodelist.lh_first;
365 	     vp && vp->v_mntvnodes.le_next != NULL;
366 	     vp = vp->v_mntvnodes.le_next);
367 	for (; vp && vp != BEG_OF_VLIST; vp = BACK_VP(vp)) {
368 #else
369 	loop:
370 	for (vp = mp->mnt_vnodelist.lh_first;
371 	     vp != NULL;
372 	     vp = vp->v_mntvnodes.le_next) {
373 #endif
374 		/*
375 		 * If the vnode that we are about to sync is no longer
376 		 * associated with this mount point, start over.
377 		 */
378 		if (vp->v_mount != mp) {
379 			printf("lfs_writevnodes: starting over\n");
380 			goto loop;
381 		}
382 
383 		ip = VTOI(vp);
384 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
385 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
386 			vndebug(vp,"dirop");
387 			continue;
388 		}
389 
390 		if (op == VN_EMPTY && vp->v_dirtyblkhd.lh_first) {
391 			vndebug(vp,"empty");
392 			continue;
393 		}
394 
395 		if (vp->v_type == VNON) {
396 			continue;
397 		}
398 
399 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
400 		   && vp != fs->lfs_flushvp
401 		   && !(ip->i_flag & IN_CLEANING)) {
402 			vndebug(vp,"cleaning");
403 			continue;
404 		}
405 
406 		if (lfs_vref(vp)) {
407 			vndebug(vp,"vref");
408 			continue;
409 		}
410 
411 		needs_unlock = 0;
412 		if (VOP_ISLOCKED(vp)) {
413 			if (vp != fs->lfs_ivnode &&
414 			    vp->v_lock.lk_lockholder != curproc->p_pid) {
415 #ifdef DEBUG_LFS
416 				printf("lfs_writevnodes: not writing ino %d,"
417 				       " locked by pid %d\n",
418 				       VTOI(vp)->i_number,
419 				       vp->v_lock.lk_lockholder);
420 #endif
421 				lfs_vunref(vp);
422 				continue;
423 			}
424 		} else if (vp != fs->lfs_ivnode) {
425 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
426 			needs_unlock = 1;
427 		}
428 
429 		only_cleaning = 0;
430 		/*
431 		 * Write the inode/file if dirty and it's not the IFILE.
432 		 */
433 		if ((ip->i_flag & IN_ALLMOD) ||
434 		     (vp->v_dirtyblkhd.lh_first != NULL))
435 		{
436 			only_cleaning = ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
437 
438 			if (ip->i_number != LFS_IFILE_INUM
439 			   && vp->v_dirtyblkhd.lh_first != NULL)
440 			{
441 				lfs_writefile(fs, sp, vp);
442 			}
443 			if (vp->v_dirtyblkhd.lh_first != NULL) {
444 				if (WRITEINPROG(vp)) {
445 #ifdef DEBUG_LFS
446 					ivndebug(vp,"writevnodes/write2");
447 #endif
448 				} else if (!(ip->i_flag & IN_ALLMOD)) {
449 #ifdef DEBUG_LFS
450 					printf("<%d>",ip->i_number);
451 #endif
452 					LFS_SET_UINO(ip, IN_MODIFIED);
453 				}
454 			}
455 			(void) lfs_writeinode(fs, sp, ip);
456 			inodes_written++;
457 		}
458 
459 		if (needs_unlock)
460 			VOP_UNLOCK(vp, 0);
461 
462 		if (lfs_clean_vnhead && only_cleaning)
463 			lfs_vunref_head(vp);
464 		else
465 			lfs_vunref(vp);
466 	}
467 	return inodes_written;
468 }
469 
470 /*
471  * Do a checkpoint.
472  */
473 int
474 lfs_segwrite(struct mount *mp, int flags)
475 {
476 	struct buf *bp;
477 	struct inode *ip;
478 	struct lfs *fs;
479 	struct segment *sp;
480 	struct vnode *vp;
481 	SEGUSE *segusep;
482 	ufs_daddr_t ibno;
483 	int do_ckp, did_ckp, error, i;
484 	int writer_set = 0;
485 	int dirty;
486 
487 	fs = VFSTOUFS(mp)->um_lfs;
488 
489 	if (fs->lfs_ronly)
490 		return EROFS;
491 
492 	lfs_imtime(fs);
493 
494 	/* printf("lfs_segwrite: ifile flags are 0x%lx\n",
495 	       (long)(VTOI(fs->lfs_ivnode)->i_flag)); */
496 
497 #if 0
498 	/*
499 	 * If we are not the cleaner, and there is no space available,
500 	 * wait until cleaner writes.
501 	 */
502 	if (!(flags & SEGM_CLEAN) && !(fs->lfs_seglock && fs->lfs_sp &&
503 				      (fs->lfs_sp->seg_flags & SEGM_CLEAN)))
504 	{
505 		while (fs->lfs_avail <= 0) {
506 			LFS_CLEANERINFO(cip, fs, bp);
507 			LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
508 
509 			wakeup(&lfs_allclean_wakeup);
510 			wakeup(&fs->lfs_nextseg);
511 			error = tsleep(&fs->lfs_avail, PRIBIO + 1, "lfs_av2",
512 				       0);
513 			if (error) {
514 				return (error);
515 			}
516 		}
517 	}
518 #endif
519 	/*
520 	 * Allocate a segment structure and enough space to hold pointers to
521 	 * the maximum possible number of buffers which can be described in a
522 	 * single summary block.
523 	 */
524 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
525 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
526 	sp = fs->lfs_sp;
527 
528 	/*
529 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
530 	 * in which case we have to flush *all* buffers off of this vnode.
531 	 * We don't care about other nodes, but write any non-dirop nodes
532 	 * anyway in anticipation of another getnewvnode().
533 	 *
534 	 * If we're cleaning we only write cleaning and ifile blocks, and
535 	 * no dirops, since otherwise we'd risk corruption in a crash.
536 	 */
537 	if (sp->seg_flags & SEGM_CLEAN)
538 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
539 	else {
540 		lfs_writevnodes(fs, mp, sp, VN_REG);
541 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
542 			while (fs->lfs_dirops)
543 				if ((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
544 						"lfs writer", 0)))
545 				{
546 					/* XXX why not segunlock? */
547 					free(sp->bpp, M_SEGMENT);
548 					sp->bpp = NULL;
549 					free(sp, M_SEGMENT);
550 					fs->lfs_sp = NULL;
551 					return (error);
552 				}
553 			fs->lfs_writer++;
554 			writer_set = 1;
555 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
556 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
557 		}
558 	}
559 
560 	/*
561 	 * If we are doing a checkpoint, mark everything since the
562 	 * last checkpoint as no longer ACTIVE.
563 	 */
564 	if (do_ckp) {
565 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
566 		     --ibno >= fs->lfs_cleansz; ) {
567 			dirty = 0;
568 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
569 
570 				panic("lfs_segwrite: ifile read");
571 			segusep = (SEGUSE *)bp->b_data;
572 			for (i = fs->lfs_sepb; i--;) {
573 				if (segusep->su_flags & SEGUSE_ACTIVE) {
574 					segusep->su_flags &= ~SEGUSE_ACTIVE;
575 					++dirty;
576 				}
577 				if (fs->lfs_version > 1)
578 					++segusep;
579 				else
580 					segusep = (SEGUSE *)
581 						((SEGUSE_V1 *)segusep + 1);
582 			}
583 
584 			/* But the current segment is still ACTIVE */
585 			segusep = (SEGUSE *)bp->b_data;
586 			if (dtosn(fs, fs->lfs_curseg) / fs->lfs_sepb ==
587 			    (ibno-fs->lfs_cleansz)) {
588 				if (fs->lfs_version > 1)
589 					segusep[dtosn(fs, fs->lfs_curseg) %
590 					     fs->lfs_sepb].su_flags |=
591 						     SEGUSE_ACTIVE;
592 				else
593 					((SEGUSE *)
594 					 ((SEGUSE_V1 *)(bp->b_data) +
595 					  (dtosn(fs, fs->lfs_curseg) %
596 					   fs->lfs_sepb)))->su_flags
597 						   |= SEGUSE_ACTIVE;
598 				--dirty;
599 			}
600 			if (dirty)
601 				error = VOP_BWRITE(bp); /* Ifile */
602 			else
603 				brelse(bp);
604 		}
605 	}
606 
607 	did_ckp = 0;
608 	if (do_ckp || fs->lfs_doifile) {
609 		do {
610 			vp = fs->lfs_ivnode;
611 
612 			vget(vp, LK_EXCLUSIVE | LK_CANRECURSE | LK_RETRY);
613 
614 			ip = VTOI(vp);
615 			if (vp->v_dirtyblkhd.lh_first != NULL)
616 				lfs_writefile(fs, sp, vp);
617 			if (ip->i_flag & IN_ALLMOD)
618 				++did_ckp;
619 			(void) lfs_writeinode(fs, sp, ip);
620 
621 			vput(vp);
622 		} while (lfs_writeseg(fs, sp) && do_ckp);
623 
624 		/* The ifile should now be all clear */
625 		LFS_CLR_UINO(ip, IN_ALLMOD);
626 	} else {
627 		(void) lfs_writeseg(fs, sp);
628 	}
629 
630 	/*
631 	 * If the I/O count is non-zero, sleep until it reaches zero.
632 	 * At the moment, the user's process hangs around so we can
633 	 * sleep.
634 	 */
635 	fs->lfs_doifile = 0;
636 	if (writer_set && --fs->lfs_writer == 0)
637 		wakeup(&fs->lfs_dirops);
638 
639 	/*
640 	 * If we didn't write the Ifile, we didn't really do anything.
641 	 * That means that (1) there is a checkpoint on disk and (2)
642 	 * nothing has changed since it was written.
643 	 *
644 	 * Take the flags off of the segment so that lfs_segunlock
645 	 * doesn't have to write the superblock either.
646 	 */
647 	if (did_ckp == 0) {
648 		sp->seg_flags &= ~(SEGM_SYNC|SEGM_CKP);
649 		/* if (do_ckp) printf("lfs_segwrite: no checkpoint\n"); */
650 	}
651 
652 	if (lfs_dostats) {
653 		++lfs_stats.nwrites;
654 		if (sp->seg_flags & SEGM_SYNC)
655 			++lfs_stats.nsync_writes;
656 		if (sp->seg_flags & SEGM_CKP)
657 			++lfs_stats.ncheckpoints;
658 	}
659 	lfs_segunlock(fs);
660 	return (0);
661 }
662 
663 /*
664  * Write the dirty blocks associated with a vnode.
665  */
666 void
667 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
668 {
669 	struct buf *bp;
670 	struct finfo *fip;
671 	IFILE *ifp;
672 
673 
674 	if (sp->seg_bytes_left < fs->lfs_bsize ||
675 	    sp->sum_bytes_left < sizeof(struct finfo))
676 		(void) lfs_writeseg(fs, sp);
677 
678 	sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t);
679 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
680 
681 	if (vp->v_flag & VDIROP)
682 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
683 
684 	fip = sp->fip;
685 	fip->fi_nblocks = 0;
686 	fip->fi_ino = VTOI(vp)->i_number;
687 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
688 	fip->fi_version = ifp->if_version;
689 	brelse(bp);
690 
691 	if (sp->seg_flags & SEGM_CLEAN)
692 	{
693 		lfs_gather(fs, sp, vp, lfs_match_fake);
694 		/*
695 		 * For a file being flushed, we need to write *all* blocks.
696 		 * This means writing the cleaning blocks first, and then
697 		 * immediately following with any non-cleaning blocks.
698 		 * The same is true of the Ifile since checkpoints assume
699 		 * that all valid Ifile blocks are written.
700 		 */
701 	   	if (IS_FLUSHING(fs,vp) || VTOI(vp)->i_number == LFS_IFILE_INUM)
702 			lfs_gather(fs, sp, vp, lfs_match_data);
703 	} else
704 		lfs_gather(fs, sp, vp, lfs_match_data);
705 
706 	/*
707 	 * It may not be necessary to write the meta-data blocks at this point,
708 	 * as the roll-forward recovery code should be able to reconstruct the
709 	 * list.
710 	 *
711 	 * We have to write them anyway, though, under two conditions: (1) the
712 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
713 	 * checkpointing.
714 	 */
715 	if (lfs_writeindir
716 	   || IS_FLUSHING(fs,vp)
717 	   || (sp->seg_flags & SEGM_CKP))
718 	{
719 		lfs_gather(fs, sp, vp, lfs_match_indir);
720 		lfs_gather(fs, sp, vp, lfs_match_dindir);
721 		lfs_gather(fs, sp, vp, lfs_match_tindir);
722 	}
723 	fip = sp->fip;
724 	if (fip->fi_nblocks != 0) {
725 		sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) +
726 				   sizeof(ufs_daddr_t) * (fip->fi_nblocks-1));
727 		sp->start_lbp = &sp->fip->fi_blocks[0];
728 	} else {
729 		sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t);
730 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
731 	}
732 }
733 
734 int
735 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
736 {
737 	struct buf *bp, *ibp;
738 	struct dinode *cdp;
739 	IFILE *ifp;
740 	SEGUSE *sup;
741 	ufs_daddr_t daddr;
742 	daddr_t *daddrp;
743 	ino_t ino;
744 	int error, i, ndx, fsb = 0;
745 	int redo_ifile = 0;
746 	struct timespec ts;
747 	int gotblk = 0;
748 
749 	if (!(ip->i_flag & IN_ALLMOD))
750 		return (0);
751 
752 	/* Allocate a new inode block if necessary. */
753 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
754 		/* Allocate a new segment if necessary. */
755 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
756 		    sp->sum_bytes_left < sizeof(ufs_daddr_t))
757 			(void) lfs_writeseg(fs, sp);
758 
759 		/* Get next inode block. */
760 		daddr = fs->lfs_offset;
761 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
762 		sp->ibp = *sp->cbpp++ =
763 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
764 			       fs->lfs_ibsize, 0, 0);
765 		gotblk++;
766 
767 		/* Zero out inode numbers */
768 		for (i = 0; i < INOPB(fs); ++i)
769 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
770 
771 		++sp->start_bpp;
772 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
773 		/* Set remaining space counters. */
774 		sp->seg_bytes_left -= fs->lfs_ibsize;
775 		sp->sum_bytes_left -= sizeof(ufs_daddr_t);
776 		ndx = fs->lfs_sumsize / sizeof(ufs_daddr_t) -
777 			sp->ninodes / INOPB(fs) - 1;
778 		((ufs_daddr_t *)(sp->segsum))[ndx] = daddr;
779 	}
780 
781 	/* Update the inode times and copy the inode onto the inode page. */
782 	TIMEVAL_TO_TIMESPEC(&time, &ts);
783 	LFS_ITIMES(ip, &ts, &ts, &ts);
784 
785 	/*
786 	 * If this is the Ifile, and we've already written the Ifile in this
787 	 * partial segment, just overwrite it (it's not on disk yet) and
788 	 * continue.
789 	 *
790 	 * XXX we know that the bp that we get the second time around has
791 	 * already been gathered.
792 	 */
793 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
794 		*(sp->idp) = ip->i_din.ffs_din;
795 		return 0;
796 	}
797 
798 	bp = sp->ibp;
799 	cdp = ((struct dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
800 	*cdp = ip->i_din.ffs_din;
801 #ifdef LFS_IFILE_FRAG_ADDRESSING
802 	if (fs->lfs_version > 1)
803 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
804 #endif
805 
806 	/*
807 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
808 	 * addresses to disk.
809 	 */
810 	if (ip->i_lfs_effnblks != ip->i_ffs_blocks) {
811 #ifdef DEBUG_LFS
812 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
813 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs_blocks);
814 #endif
815 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
816 		     daddrp++) {
817 			if (*daddrp == UNWRITTEN) {
818 #ifdef DEBUG_LFS
819 				printf("lfs_writeinode: wiping UNWRITTEN\n");
820 #endif
821 				*daddrp = 0;
822 			}
823 		}
824 	}
825 
826 	if (ip->i_flag & IN_CLEANING)
827 		LFS_CLR_UINO(ip, IN_CLEANING);
828 	else {
829 		/* XXX IN_ALLMOD */
830 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
831 			     IN_UPDATE);
832 		if (ip->i_lfs_effnblks == ip->i_ffs_blocks)
833 			LFS_CLR_UINO(ip, IN_MODIFIED);
834 #ifdef DEBUG_LFS
835 		else
836 			printf("lfs_writeinode: ino %d: real blks=%d, "
837 			       "eff=%d\n", ip->i_number, ip->i_ffs_blocks,
838 			       ip->i_lfs_effnblks);
839 #endif
840 	}
841 
842 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
843 		sp->idp = ((struct dinode *)bp->b_data) +
844 			(sp->ninodes % INOPB(fs));
845 	if (gotblk) {
846 		LFS_LOCK_BUF(bp);
847 		brelse(bp);
848 	}
849 
850 	/* Increment inode count in segment summary block. */
851 	++((SEGSUM *)(sp->segsum))->ss_ninos;
852 
853 	/* If this page is full, set flag to allocate a new page. */
854 	if (++sp->ninodes % INOPB(fs) == 0)
855 		sp->ibp = NULL;
856 
857 	/*
858 	 * If updating the ifile, update the super-block.  Update the disk
859 	 * address and access times for this inode in the ifile.
860 	 */
861 	ino = ip->i_number;
862 	if (ino == LFS_IFILE_INUM) {
863 		daddr = fs->lfs_idaddr;
864 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
865 	} else {
866 		LFS_IENTRY(ifp, fs, ino, ibp);
867 		daddr = ifp->if_daddr;
868 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
869 #ifdef LFS_DEBUG_NEXTFREE
870 		if (ino > 3 && ifp->if_nextfree) {
871 			vprint("lfs_writeinode",ITOV(ip));
872 			printf("lfs_writeinode: updating free ino %d\n",
873 				ip->i_number);
874 		}
875 #endif
876 		error = VOP_BWRITE(ibp); /* Ifile */
877 	}
878 
879 	/*
880 	 * Account the inode: it no longer belongs to its former segment,
881 	 * though it will not belong to the new segment until that segment
882 	 * is actually written.
883 	 */
884 #ifdef DEBUG
885 	/*
886 	 * The inode's last address should not be in the current partial
887 	 * segment, except under exceptional circumstances (lfs_writevnodes
888 	 * had to start over, and in the meantime more blocks were written
889 	 * to a vnode).  Although the previous inode won't be accounted in
890 	 * su_nbytes until lfs_writeseg, this shouldn't be a problem as we
891 	 * have more data blocks in the current partial segment.
892 	 */
893 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno))
894 		printf("lfs_writeinode: last inode addr in current pseg "
895 		       "(ino %d daddr 0x%x)\n", ino, daddr);
896 #endif
897 	if (daddr != LFS_UNUSED_DADDR) {
898 		LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
899 #ifdef DIAGNOSTIC
900 		if (sup->su_nbytes < DINODE_SIZE) {
901 			printf("lfs_writeinode: negative bytes "
902 			       "(segment %d short by %d)\n",
903 			       dtosn(fs, daddr),
904 			       (int)DINODE_SIZE - sup->su_nbytes);
905 			panic("lfs_writeinode: negative bytes");
906 			sup->su_nbytes = DINODE_SIZE;
907 		}
908 #endif
909 #ifdef DEBUG_SU_NBYTES
910 		printf("seg %d -= %d for ino %d inode\n",
911 		       dtosn(fs, daddr), DINODE_SIZE, ino);
912 #endif
913 		sup->su_nbytes -= DINODE_SIZE;
914 		redo_ifile =
915 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
916 		error = VOP_BWRITE(bp); /* Ifile */
917 	}
918 	return (redo_ifile);
919 }
920 
921 int
922 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
923 {
924 	struct lfs *fs;
925 	int version;
926 
927 	/*
928 	 * If full, finish this segment.  We may be doing I/O, so
929 	 * release and reacquire the splbio().
930 	 */
931 #ifdef DIAGNOSTIC
932 	if (sp->vp == NULL)
933 		panic ("lfs_gatherblock: Null vp in segment");
934 #endif
935 	fs = sp->fs;
936 	if (sp->sum_bytes_left < sizeof(ufs_daddr_t) ||
937 	    sp->seg_bytes_left < bp->b_bcount) {
938 		if (sptr)
939 			splx(*sptr);
940 		lfs_updatemeta(sp);
941 
942 		version = sp->fip->fi_version;
943 		(void) lfs_writeseg(fs, sp);
944 
945 		sp->fip->fi_version = version;
946 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
947 		/* Add the current file to the segment summary. */
948 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
949 		sp->sum_bytes_left -=
950 			sizeof(struct finfo) - sizeof(ufs_daddr_t);
951 
952 		if (sptr)
953 			*sptr = splbio();
954 		return (1);
955 	}
956 
957 #ifdef DEBUG
958 	if (bp->b_flags & B_GATHERED) {
959 		printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n",
960 		       sp->fip->fi_ino, bp->b_lblkno);
961 		return (0);
962 	}
963 #endif
964 	/* Insert into the buffer list, update the FINFO block. */
965 	bp->b_flags |= B_GATHERED;
966 	*sp->cbpp++ = bp;
967 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
968 
969 	sp->sum_bytes_left -= sizeof(ufs_daddr_t);
970 	sp->seg_bytes_left -= bp->b_bcount;
971 	return (0);
972 }
973 
974 int
975 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, int (*match)(struct lfs *, struct buf *))
976 {
977 	struct buf *bp;
978 	int s, count = 0;
979 
980 	sp->vp = vp;
981 	s = splbio();
982 
983 #ifndef LFS_NO_BACKBUF_HACK
984 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = bp->b_vnbufs.le_next) {
985 #else /* LFS_NO_BACKBUF_HACK */
986 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
987 # define	BUF_OFFSET	(((void *)&bp->b_vnbufs.le_next) - (void *)bp)
988 # define	BACK_BUF(BP)	((struct buf *)(((void *)BP->b_vnbufs.le_prev) - BUF_OFFSET))
989 # define	BEG_OF_LIST	((struct buf *)(((void *)&vp->v_dirtyblkhd.lh_first) - BUF_OFFSET))
990 /* Find last buffer. */
991 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp && bp->b_vnbufs.le_next != NULL;
992 	    bp = bp->b_vnbufs.le_next);
993 	for (; bp && bp != BEG_OF_LIST; bp = BACK_BUF(bp)) {
994 #endif /* LFS_NO_BACKBUF_HACK */
995 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp))
996 			continue;
997 		if (vp->v_type == VBLK) {
998 			/* For block devices, just write the blocks. */
999 			/* XXX Do we really need to even do this? */
1000 #ifdef DEBUG_LFS
1001 			if (count == 0)
1002 				printf("BLK(");
1003 			printf(".");
1004 #endif
1005 			/* Get the block before bwrite, so we don't corrupt the free list */
1006 			bp->b_flags |= B_BUSY;
1007 			bremfree(bp);
1008 			bwrite(bp);
1009 		} else {
1010 #ifdef DIAGNOSTIC
1011 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1012 				printf("lfs_gather: lbn %d is B_INVAL\n",
1013 					bp->b_lblkno);
1014 				VOP_PRINT(bp->b_vp);
1015 			}
1016 			if (!(bp->b_flags & B_DELWRI))
1017 				panic("lfs_gather: bp not B_DELWRI");
1018 			if (!(bp->b_flags & B_LOCKED)) {
1019 				printf("lfs_gather: lbn %d blk %d"
1020 				       " not B_LOCKED\n", bp->b_lblkno,
1021 				       dbtofsb(fs, bp->b_blkno));
1022 				VOP_PRINT(bp->b_vp);
1023 				panic("lfs_gather: bp not B_LOCKED");
1024 			}
1025 #endif
1026 			if (lfs_gatherblock(sp, bp, &s)) {
1027 				goto loop;
1028 			}
1029 		}
1030 		count++;
1031 	}
1032 	splx(s);
1033 #ifdef DEBUG_LFS
1034 	if (vp->v_type == VBLK && count)
1035 		printf(")\n");
1036 #endif
1037 	lfs_updatemeta(sp);
1038 	sp->vp = NULL;
1039 	return count;
1040 }
1041 
1042 /*
1043  * Update the metadata that points to the blocks listed in the FINFO
1044  * array.
1045  */
1046 void
1047 lfs_updatemeta(struct segment *sp)
1048 {
1049 	SEGUSE *sup;
1050 	struct buf *bp;
1051 	struct lfs *fs;
1052 	struct vnode *vp;
1053 	struct indir a[NIADDR + 2], *ap;
1054 	struct inode *ip;
1055 	ufs_daddr_t daddr, lbn, off;
1056 	daddr_t ooff;
1057 	int error, i, nblocks, num;
1058 	int bb;
1059 
1060 	vp = sp->vp;
1061 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1062 	if (nblocks < 0)
1063 		panic("This is a bad thing\n");
1064 	if (vp == NULL || nblocks == 0)
1065 		return;
1066 
1067 	/* Sort the blocks. */
1068 	/*
1069 	 * XXX KS - We have to sort even if the blocks come from the
1070 	 * cleaner, because there might be other pending blocks on the
1071 	 * same inode...and if we don't sort, and there are fragments
1072 	 * present, blocks may be written in the wrong place.
1073 	 */
1074 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
1075 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
1076 
1077 	/*
1078 	 * Record the length of the last block in case it's a fragment.
1079 	 * If there are indirect blocks present, they sort last.  An
1080 	 * indirect block will be lfs_bsize and its presence indicates
1081 	 * that you cannot have fragments.
1082 	 */
1083 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1084 
1085 	/*
1086 	 * Assign disk addresses, and update references to the logical
1087 	 * block and the segment usage information.
1088 	 */
1089 	fs = sp->fs;
1090 	for (i = nblocks; i--; ++sp->start_bpp) {
1091 		lbn = *sp->start_lbp++;
1092 
1093 		(*sp->start_bpp)->b_blkno = fsbtodb(fs, fs->lfs_offset);
1094 		off = fs->lfs_offset;
1095 		if ((*sp->start_bpp)->b_blkno == (*sp->start_bpp)->b_lblkno) {
1096 			printf("lfs_updatemeta: ino %d blk %d"
1097 			       " has same lbn and daddr\n",
1098 			       VTOI(vp)->i_number, off);
1099 		}
1100 #ifdef DIAGNOSTIC
1101 		if ((*sp->start_bpp)->b_bcount < fs->lfs_bsize && i != 0)
1102 			panic("lfs_updatemeta: fragment is not last block\n");
1103 #endif
1104 		bb = fragstofsb(fs, numfrags(fs, (*sp->start_bpp)->b_bcount));
1105 		fs->lfs_offset += bb;
1106 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1107 		if (daddr > 0)
1108 			daddr = dbtofsb(fs, daddr);
1109 		if (error)
1110 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1111 		ip = VTOI(vp);
1112 		switch (num) {
1113 		case 0:
1114 			ooff = ip->i_ffs_db[lbn];
1115 #ifdef DEBUG
1116 			if (ooff == 0) {
1117 				printf("lfs_updatemeta[1]: warning: writing "
1118 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1119 				       ip->i_number, lbn, off);
1120 			}
1121 #endif
1122 			if (ooff == UNWRITTEN)
1123 				ip->i_ffs_blocks += bb;
1124 			ip->i_ffs_db[lbn] = off;
1125 			break;
1126 		case 1:
1127 			ooff = ip->i_ffs_ib[a[0].in_off];
1128 #ifdef DEBUG
1129 			if (ooff == 0) {
1130 				printf("lfs_updatemeta[2]: warning: writing "
1131 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1132 				       ip->i_number, lbn, off);
1133 			}
1134 #endif
1135 			if (ooff == UNWRITTEN)
1136 				ip->i_ffs_blocks += bb;
1137 			ip->i_ffs_ib[a[0].in_off] = off;
1138 			break;
1139 		default:
1140 			ap = &a[num - 1];
1141 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1142 				panic("lfs_updatemeta: bread bno %d",
1143 				      ap->in_lbn);
1144 
1145 			ooff = ((ufs_daddr_t *)bp->b_data)[ap->in_off];
1146 #if DEBUG
1147 			if (ooff == 0) {
1148 				printf("lfs_updatemeta[3]: warning: writing "
1149 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1150 				       ip->i_number, lbn, off);
1151 			}
1152 #endif
1153 			if (ooff == UNWRITTEN)
1154 				ip->i_ffs_blocks += bb;
1155 			((ufs_daddr_t *)bp->b_data)[ap->in_off] = off;
1156 			(void) VOP_BWRITE(bp);
1157 		}
1158 #ifdef DEBUG
1159 		if (daddr >= fs->lfs_lastpseg && daddr <= off) {
1160 			printf("lfs_updatemeta: ino %d, lbn %d, addr = %x "
1161 			       "in same pseg\n", VTOI(sp->vp)->i_number,
1162 			       (*sp->start_bpp)->b_lblkno, daddr);
1163 		}
1164 #endif
1165 		/* Update segment usage information. */
1166 		if (daddr > 0) {
1167 			LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
1168 #ifdef DIAGNOSTIC
1169 			if (sup->su_nbytes < (*sp->start_bpp)->b_bcount) {
1170 				/* XXX -- Change to a panic. */
1171 				printf("lfs_updatemeta: negative bytes "
1172 				       "(segment %d short by %ld)\n",
1173 				       dtosn(fs, daddr),
1174 				       (*sp->start_bpp)->b_bcount -
1175 				       sup->su_nbytes);
1176 				printf("lfs_updatemeta: ino %d, lbn %d, "
1177 				       "addr = 0x%x\n", VTOI(sp->vp)->i_number,
1178 				       (*sp->start_bpp)->b_lblkno, daddr);
1179 				panic("lfs_updatemeta: negative bytes");
1180 				sup->su_nbytes = (*sp->start_bpp)->b_bcount;
1181 			}
1182 #endif
1183 #ifdef DEBUG_SU_NBYTES
1184 			printf("seg %d -= %ld for ino %d lbn %d db 0x%x\n",
1185 			       dtosn(fs, daddr), (*sp->start_bpp)->b_bcount,
1186 			       VTOI(sp->vp)->i_number,
1187 			       (*sp->start_bpp)->b_lblkno, daddr);
1188 #endif
1189 			sup->su_nbytes -= (*sp->start_bpp)->b_bcount;
1190 			error = VOP_BWRITE(bp); /* Ifile */
1191 		}
1192 	}
1193 }
1194 
1195 /*
1196  * Start a new segment.
1197  */
1198 int
1199 lfs_initseg(struct lfs *fs)
1200 {
1201 	struct segment *sp;
1202 	SEGUSE *sup;
1203 	SEGSUM *ssp;
1204 	struct buf *bp;
1205 	int repeat;
1206 
1207 	sp = fs->lfs_sp;
1208 
1209 	repeat = 0;
1210 	/* Advance to the next segment. */
1211 	if (!LFS_PARTIAL_FITS(fs)) {
1212 		/* lfs_avail eats the remaining space */
1213 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1214 						   fs->lfs_curseg);
1215 		/* Wake up any cleaning procs waiting on this file system. */
1216 		wakeup(&lfs_allclean_wakeup);
1217 		wakeup(&fs->lfs_nextseg);
1218 		lfs_newseg(fs);
1219 		repeat = 1;
1220 		fs->lfs_offset = fs->lfs_curseg;
1221 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1222 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1223 		/*
1224 		 * If the segment contains a superblock, update the offset
1225 		 * and summary address to skip over it.
1226 		 */
1227 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1228 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1229 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1230 			sp->seg_bytes_left -= LFS_SBPAD;
1231 		}
1232 		brelse(bp);
1233 		/* Segment zero could also contain the labelpad */
1234 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1235 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1236 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1237 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1238 		}
1239 	} else {
1240 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1241 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1242 				      (fs->lfs_offset - fs->lfs_curseg));
1243 	}
1244 	fs->lfs_lastpseg = fs->lfs_offset;
1245 
1246 	sp->fs = fs;
1247 	sp->ibp = NULL;
1248 	sp->idp = NULL;
1249 	sp->ninodes = 0;
1250 
1251 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1252 	sp->cbpp = sp->bpp;
1253 	*sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1254 			       fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize);
1255 	sp->segsum = (*sp->cbpp)->b_data;
1256 	bzero(sp->segsum, fs->lfs_sumsize);
1257 	sp->start_bpp = ++sp->cbpp;
1258 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1259 
1260 	/* Set point to SEGSUM, initialize it. */
1261 	ssp = sp->segsum;
1262 	ssp->ss_next = fs->lfs_nextseg;
1263 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1264 	ssp->ss_magic = SS_MAGIC;
1265 
1266 	/* Set pointer to first FINFO, initialize it. */
1267 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1268 	sp->fip->fi_nblocks = 0;
1269 	sp->start_lbp = &sp->fip->fi_blocks[0];
1270 	sp->fip->fi_lastlength = 0;
1271 
1272 	sp->seg_bytes_left -= fs->lfs_sumsize;
1273 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1274 
1275 	return (repeat);
1276 }
1277 
1278 /*
1279  * Return the next segment to write.
1280  */
1281 void
1282 lfs_newseg(struct lfs *fs)
1283 {
1284 	CLEANERINFO *cip;
1285 	SEGUSE *sup;
1286 	struct buf *bp;
1287 	int curseg, isdirty, sn;
1288 
1289 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1290 #ifdef DEBUG_SU_NBYTES
1291 	printf("lfs_newseg: seg %d := 0 in newseg\n",   /* XXXDEBUG */
1292 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1293 #endif
1294 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1295 	sup->su_nbytes = 0;
1296 	sup->su_nsums = 0;
1297 	sup->su_ninos = 0;
1298 	(void) VOP_BWRITE(bp); /* Ifile */
1299 
1300 	LFS_CLEANERINFO(cip, fs, bp);
1301 	--cip->clean;
1302 	++cip->dirty;
1303 	fs->lfs_nclean = cip->clean;
1304 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1305 
1306 	fs->lfs_lastseg = fs->lfs_curseg;
1307 	fs->lfs_curseg = fs->lfs_nextseg;
1308 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1309 		sn = (sn + 1) % fs->lfs_nseg;
1310 		if (sn == curseg)
1311 			panic("lfs_nextseg: no clean segments");
1312 		LFS_SEGENTRY(sup, fs, sn, bp);
1313 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1314 		brelse(bp);
1315 		if (!isdirty)
1316 			break;
1317 	}
1318 
1319 	++fs->lfs_nactive;
1320 	fs->lfs_nextseg = sntod(fs, sn);
1321 	if (lfs_dostats) {
1322 		++lfs_stats.segsused;
1323 	}
1324 }
1325 
1326 int
1327 lfs_writeseg(struct lfs *fs, struct segment *sp)
1328 {
1329 	struct buf **bpp, *bp, *cbp, *newbp;
1330 	SEGUSE *sup;
1331 	SEGSUM *ssp;
1332 	dev_t i_dev;
1333 	char *datap, *dp;
1334 	int do_again, i, nblocks, s;
1335 	size_t el_size;
1336 #ifdef LFS_TRACK_IOS
1337 	int j;
1338 #endif
1339 	int (*strategy)(void *);
1340 	struct vop_strategy_args vop_strategy_a;
1341 	u_short ninos;
1342 	struct vnode *devvp;
1343 	char *p;
1344 	struct vnode *vp;
1345 	struct inode *ip;
1346 	daddr_t *daddrp;
1347 	int changed;
1348 #if defined(DEBUG) && defined(LFS_PROPELLER)
1349 	static int propeller;
1350 	char propstring[4] = "-\\|/";
1351 
1352 	printf("%c\b",propstring[propeller++]);
1353 	if (propeller == 4)
1354 		propeller = 0;
1355 #endif
1356 
1357 	/*
1358 	 * If there are no buffers other than the segment summary to write
1359 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1360 	 * even if there aren't any buffers, you need to write the superblock.
1361 	 */
1362 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1363 		return (0);
1364 
1365 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1366 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1367 
1368 	/* Update the segment usage information. */
1369 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1370 
1371 	/* Loop through all blocks, except the segment summary. */
1372 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1373 		if ((*bpp)->b_vp != devvp) {
1374 			sup->su_nbytes += (*bpp)->b_bcount;
1375 #ifdef DEBUG_SU_NBYTES
1376 		printf("seg %d += %ld for ino %d lbn %d db 0x%x\n",
1377 		       sp->seg_number, (*bpp)->b_bcount,
1378 		       VTOI((*bpp)->b_vp)->i_number,
1379 		       (*bpp)->b_lblkno, (*bpp)->b_blkno);
1380 #endif
1381 		}
1382 	}
1383 
1384 	ssp = (SEGSUM *)sp->segsum;
1385 
1386 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1387 #ifdef DEBUG_SU_NBYTES
1388 	printf("seg %d += %d for %d inodes\n",   /* XXXDEBUG */
1389 	       sp->seg_number, ssp->ss_ninos * DINODE_SIZE,
1390 	       ssp->ss_ninos);
1391 #endif
1392 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1393 	/* sup->su_nbytes += fs->lfs_sumsize; */
1394 	if (fs->lfs_version == 1)
1395 		sup->su_olastmod = time.tv_sec;
1396 	else
1397 		sup->su_lastmod = time.tv_sec;
1398 	sup->su_ninos += ninos;
1399 	++sup->su_nsums;
1400 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1401 							 fs->lfs_ibsize));
1402 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1403 
1404 	do_again = !(bp->b_flags & B_GATHERED);
1405 	(void)VOP_BWRITE(bp); /* Ifile */
1406 	/*
1407 	 * Mark blocks B_BUSY, to prevent then from being changed between
1408 	 * the checksum computation and the actual write.
1409 	 *
1410 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1411 	 * there are any, replace them with copies that have UNASSIGNED
1412 	 * instead.
1413 	 */
1414 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1415 		++bpp;
1416 		if ((*bpp)->b_flags & B_CALL)
1417 			continue;
1418 		bp = *bpp;
1419 	    again:
1420 		s = splbio();
1421 		if (bp->b_flags & B_BUSY) {
1422 #ifdef DEBUG
1423 			printf("lfs_writeseg: avoiding potential data "
1424 			       "summary corruption for ino %d, lbn %d\n",
1425 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1426 #endif
1427 			bp->b_flags |= B_WANTED;
1428 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1429 			splx(s);
1430 			goto again;
1431 		}
1432 		bp->b_flags |= B_BUSY;
1433 		splx(s);
1434 		/* Check and replace indirect block UNWRITTEN bogosity */
1435 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1436 		   VTOI(bp->b_vp)->i_ffs_blocks !=
1437 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1438 #ifdef DEBUG_LFS
1439 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1440 			       VTOI(bp->b_vp)->i_number,
1441 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1442 			       VTOI(bp->b_vp)->i_ffs_blocks);
1443 #endif
1444 			/* Make a copy we'll make changes to */
1445 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1446 					   bp->b_bcount);
1447 			newbp->b_blkno = bp->b_blkno;
1448 			memcpy(newbp->b_data, bp->b_data,
1449 			       newbp->b_bcount);
1450 			*bpp = newbp;
1451 
1452 			changed = 0;
1453 			for (daddrp = (daddr_t *)(newbp->b_data);
1454 			     daddrp < (daddr_t *)(newbp->b_data +
1455 						  newbp->b_bcount); daddrp++) {
1456 				if (*daddrp == UNWRITTEN) {
1457 					++changed;
1458 #ifdef DEBUG_LFS
1459 					printf("lfs_writeseg: replacing UNWRITTEN\n");
1460 #endif
1461 					*daddrp = 0;
1462 				}
1463 			}
1464 			/*
1465 			 * Get rid of the old buffer.  Don't mark it clean,
1466 			 * though, if it still has dirty data on it.
1467 			 */
1468 			if (changed) {
1469 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1470 				if (bp->b_flags & B_CALL) {
1471 					lfs_freebuf(bp);
1472 					bp = NULL;
1473 				} else {
1474 					/* Still on free list, leave it there */
1475 					s = splbio();
1476 					bp->b_flags &= ~B_BUSY;
1477 					if (bp->b_flags & B_WANTED)
1478 						wakeup(bp);
1479 				 	splx(s);
1480 					/*
1481 					 * We have to re-decrement lfs_avail
1482 					 * since this block is going to come
1483 					 * back around to us in the next
1484 					 * segment.
1485 					 */
1486 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1487 				}
1488 			} else {
1489 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1490 						 B_GATHERED);
1491 				LFS_UNLOCK_BUF(bp);
1492 				if (bp->b_flags & B_CALL) {
1493 					lfs_freebuf(bp);
1494 					bp = NULL;
1495 				} else {
1496 					bremfree(bp);
1497 					bp->b_flags |= B_DONE;
1498 					reassignbuf(bp, bp->b_vp);
1499 					brelse(bp);
1500 				}
1501 			}
1502 
1503 		}
1504 	}
1505 	/*
1506 	 * Compute checksum across data and then across summary; the first
1507 	 * block (the summary block) is skipped.  Set the create time here
1508 	 * so that it's guaranteed to be later than the inode mod times.
1509 	 *
1510 	 * XXX
1511 	 * Fix this to do it inline, instead of malloc/copy.
1512 	 */
1513 	if (fs->lfs_version == 1)
1514 		el_size = sizeof(u_long);
1515 	else
1516 		el_size = sizeof(u_int32_t);
1517 	datap = dp = malloc(nblocks * el_size, M_SEGMENT, M_WAITOK);
1518 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1519 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1520 			if (copyin((*bpp)->b_saveaddr, dp, el_size))
1521 				panic("lfs_writeseg: copyin failed [1]: "
1522 				      "ino %d blk %d",
1523 				      VTOI((*bpp)->b_vp)->i_number,
1524 				      (*bpp)->b_lblkno);
1525 		} else
1526 			memcpy(dp, (*bpp)->b_data, el_size);
1527 		dp += el_size;
1528 	}
1529 	if (fs->lfs_version == 1)
1530 		ssp->ss_ocreate = time.tv_sec;
1531 	else {
1532 		ssp->ss_create = time.tv_sec;
1533 		ssp->ss_serial = ++fs->lfs_serial;
1534 		ssp->ss_ident  = fs->lfs_ident;
1535 	}
1536 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * el_size);
1537 	ssp->ss_sumsum =
1538 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1539 	free(datap, M_SEGMENT);
1540 	datap = dp = NULL;
1541 #ifdef DIAGNOSTIC
1542 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1543 		panic("lfs_writeseg: No diskspace for summary");
1544 #endif
1545 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1546 			  btofsb(fs, fs->lfs_sumsize));
1547 
1548 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1549 
1550 	/*
1551 	 * When we simply write the blocks we lose a rotation for every block
1552 	 * written.  To avoid this problem, we allocate memory in chunks, copy
1553 	 * the buffers into the chunk and write the chunk.  CHUNKSIZE is the
1554 	 * largest size I/O devices can handle.
1555 	 * When the data is copied to the chunk, turn off the B_LOCKED bit
1556 	 * and brelse the buffer (which will move them to the LRU list).  Add
1557 	 * the B_CALL flag to the buffer header so we can count I/O's for the
1558 	 * checkpoints and so we can release the allocated memory.
1559 	 *
1560 	 * XXX
1561 	 * This should be removed if the new virtual memory system allows us to
1562 	 * easily make the buffers contiguous in kernel memory and if that's
1563 	 * fast enough.
1564 	 */
1565 
1566 #define CHUNKSIZE MAXPHYS
1567 
1568 	if (devvp == NULL)
1569 		panic("devvp is NULL");
1570 	for (bpp = sp->bpp,i = nblocks; i;) {
1571 		cbp = lfs_newbuf(fs, devvp, (*bpp)->b_blkno, CHUNKSIZE);
1572 		cbp->b_dev = i_dev;
1573 		cbp->b_flags |= B_ASYNC | B_BUSY;
1574 		cbp->b_bcount = 0;
1575 
1576 #ifdef DIAGNOSTIC
1577 		if (dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno) + btofsb(fs, (*bpp)->b_bcount) - 1) !=
1578 		   dtosn(fs, dbtofsb(fs, cbp->b_blkno))) {
1579 			panic("lfs_writeseg: Segment overwrite");
1580 		}
1581 #endif
1582 
1583 		s = splbio();
1584 		if (fs->lfs_iocount >= LFS_THROTTLE) {
1585 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs throttle", 0);
1586 		}
1587 		++fs->lfs_iocount;
1588 #ifdef LFS_TRACK_IOS
1589 		for (j = 0; j < LFS_THROTTLE; j++) {
1590 			if (fs->lfs_pending[j] == LFS_UNUSED_DADDR) {
1591 				fs->lfs_pending[j] = dbtofsb(fs, cbp->b_blkno);
1592 				break;
1593 			}
1594 		}
1595 #endif /* LFS_TRACK_IOS */
1596 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1597 			bp = *bpp;
1598 
1599 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1600 				break;
1601 
1602 			/*
1603 			 * Fake buffers from the cleaner are marked as B_INVAL.
1604 			 * We need to copy the data from user space rather than
1605 			 * from the buffer indicated.
1606 			 * XXX == what do I do on an error?
1607 			 */
1608 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1609 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1610 					panic("lfs_writeseg: copyin failed [2]");
1611 			} else
1612 				bcopy(bp->b_data, p, bp->b_bcount);
1613 			p += bp->b_bcount;
1614 			cbp->b_bcount += bp->b_bcount;
1615 			LFS_UNLOCK_BUF(bp);
1616 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1617 					 B_GATHERED);
1618 			vp = bp->b_vp;
1619 			if (bp->b_flags & B_CALL) {
1620 				/* if B_CALL, it was created with newbuf */
1621 				lfs_freebuf(bp);
1622 				bp = NULL;
1623 			} else {
1624 				bremfree(bp);
1625 				bp->b_flags |= B_DONE;
1626 				if (vp)
1627 					reassignbuf(bp, vp);
1628 				brelse(bp);
1629 			}
1630 
1631 			bpp++;
1632 
1633 			/*
1634 			 * If this is the last block for this vnode, but
1635 			 * there are other blocks on its dirty list,
1636 			 * set IN_MODIFIED/IN_CLEANING depending on what
1637 			 * sort of block.  Only do this for our mount point,
1638 			 * not for, e.g., inode blocks that are attached to
1639 			 * the devvp.
1640 			 * XXX KS - Shouldn't we set *both* if both types
1641 			 * of blocks are present (traverse the dirty list?)
1642 			 */
1643 			if ((i == 1 ||
1644 			    (i > 1 && vp && *bpp && (*bpp)->b_vp != vp)) &&
1645 			   (bp = vp->v_dirtyblkhd.lh_first) != NULL &&
1646 			   vp->v_mount == fs->lfs_ivnode->v_mount)
1647 			{
1648 				ip = VTOI(vp);
1649 #ifdef DEBUG_LFS
1650 				printf("lfs_writeseg: marking ino %d\n",
1651 				       ip->i_number);
1652 #endif
1653 				if (bp->b_flags & B_CALL)
1654 					LFS_SET_UINO(ip, IN_CLEANING);
1655 				else
1656 					LFS_SET_UINO(ip, IN_MODIFIED);
1657 			}
1658 			wakeup(vp);
1659 		}
1660 		++cbp->b_vp->v_numoutput;
1661 		splx(s);
1662 		/*
1663 		 * XXXX This is a gross and disgusting hack.  Since these
1664 		 * buffers are physically addressed, they hang off the
1665 		 * device vnode (devvp).  As a result, they have no way
1666 		 * of getting to the LFS superblock or lfs structure to
1667 		 * keep track of the number of I/O's pending.  So, I am
1668 		 * going to stuff the fs into the saveaddr field of
1669 		 * the buffer (yuk).
1670 		 */
1671 		cbp->b_saveaddr = (caddr_t)fs;
1672 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1673 		vop_strategy_a.a_bp = cbp;
1674 		(strategy)(&vop_strategy_a);
1675 	}
1676 #if 1 || defined(DEBUG)
1677 	/*
1678 	 * After doing a big write, we recalculate how many buffers are
1679 	 * really still left on the locked queue.
1680 	 */
1681 	s = splbio();
1682 	lfs_countlocked(&locked_queue_count, &locked_queue_bytes);
1683 	splx(s);
1684 	wakeup(&locked_queue_count);
1685 #endif /* 1 || DEBUG */
1686 	if (lfs_dostats) {
1687 		++lfs_stats.psegwrites;
1688 		lfs_stats.blocktot += nblocks - 1;
1689 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1690 			++lfs_stats.psyncwrites;
1691 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1692 			++lfs_stats.pcleanwrites;
1693 			lfs_stats.cleanblocks += nblocks - 1;
1694 		}
1695 	}
1696 	return (lfs_initseg(fs) || do_again);
1697 }
1698 
1699 void
1700 lfs_writesuper(struct lfs *fs, daddr_t daddr)
1701 {
1702 	struct buf *bp;
1703 	dev_t i_dev;
1704 	int (*strategy)(void *);
1705 	int s;
1706 	struct vop_strategy_args vop_strategy_a;
1707 
1708 	/*
1709 	 * If we can write one superblock while another is in
1710 	 * progress, we risk not having a complete checkpoint if we crash.
1711 	 * So, block here if a superblock write is in progress.
1712 	 */
1713 	s = splbio();
1714 	while (fs->lfs_sbactive) {
1715 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
1716 	}
1717 	fs->lfs_sbactive = daddr;
1718 	splx(s);
1719 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1720 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
1721 
1722 	/* Set timestamp of this version of the superblock */
1723 	if (fs->lfs_version == 1)
1724 		fs->lfs_otstamp = time.tv_sec;
1725 	fs->lfs_tstamp = time.tv_sec;
1726 
1727 	/* Checksum the superblock and copy it into a buffer. */
1728 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
1729 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD);
1730 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
1731 
1732 	bp->b_dev = i_dev;
1733 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
1734 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
1735 	bp->b_iodone = lfs_supercallback;
1736 	/* XXX KS - same nasty hack as above */
1737 	bp->b_saveaddr = (caddr_t)fs;
1738 
1739 	vop_strategy_a.a_desc = VDESC(vop_strategy);
1740 	vop_strategy_a.a_bp = bp;
1741 	s = splbio();
1742 	++bp->b_vp->v_numoutput;
1743 	++fs->lfs_iocount;
1744 	splx(s);
1745 	(strategy)(&vop_strategy_a);
1746 }
1747 
1748 /*
1749  * Logical block number match routines used when traversing the dirty block
1750  * chain.
1751  */
1752 int
1753 lfs_match_fake(struct lfs *fs, struct buf *bp)
1754 {
1755 	return (bp->b_flags & B_CALL);
1756 }
1757 
1758 int
1759 lfs_match_data(struct lfs *fs, struct buf *bp)
1760 {
1761 	return (bp->b_lblkno >= 0);
1762 }
1763 
1764 int
1765 lfs_match_indir(struct lfs *fs, struct buf *bp)
1766 {
1767 	int lbn;
1768 
1769 	lbn = bp->b_lblkno;
1770 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
1771 }
1772 
1773 int
1774 lfs_match_dindir(struct lfs *fs, struct buf *bp)
1775 {
1776 	int lbn;
1777 
1778 	lbn = bp->b_lblkno;
1779 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
1780 }
1781 
1782 int
1783 lfs_match_tindir(struct lfs *fs, struct buf *bp)
1784 {
1785 	int lbn;
1786 
1787 	lbn = bp->b_lblkno;
1788 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
1789 }
1790 
1791 /*
1792  * XXX - The only buffers that are going to hit these functions are the
1793  * segment write blocks, or the segment summaries, or the superblocks.
1794  *
1795  * All of the above are created by lfs_newbuf, and so do not need to be
1796  * released via brelse.
1797  */
1798 void
1799 lfs_callback(struct buf *bp)
1800 {
1801 	struct lfs *fs;
1802 #ifdef LFS_TRACK_IOS
1803 	int j;
1804 #endif
1805 
1806 	fs = (struct lfs *)bp->b_saveaddr;
1807 #ifdef DIAGNOSTIC
1808 	if (fs->lfs_iocount == 0)
1809 		panic("lfs_callback: zero iocount\n");
1810 #endif
1811 	if (--fs->lfs_iocount < LFS_THROTTLE)
1812 		wakeup(&fs->lfs_iocount);
1813 #ifdef LFS_TRACK_IOS
1814 	for (j = 0; j < LFS_THROTTLE; j++) {
1815 		if (fs->lfs_pending[j] == dbtofsb(fs, bp->b_blkno)) {
1816 			fs->lfs_pending[j] = LFS_UNUSED_DADDR;
1817 			wakeup(&(fs->lfs_pending[j]));
1818 			break;
1819 		}
1820 	}
1821 #endif /* LFS_TRACK_IOS */
1822 
1823 	lfs_freebuf(bp);
1824 }
1825 
1826 void
1827 lfs_supercallback(struct buf *bp)
1828 {
1829 	struct lfs *fs;
1830 
1831 	fs = (struct lfs *)bp->b_saveaddr;
1832 	fs->lfs_sbactive = 0;
1833 	wakeup(&fs->lfs_sbactive);
1834 	if (--fs->lfs_iocount < LFS_THROTTLE)
1835 		wakeup(&fs->lfs_iocount);
1836 	lfs_freebuf(bp);
1837 }
1838 
1839 /*
1840  * Shellsort (diminishing increment sort) from Data Structures and
1841  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
1842  * see also Knuth Vol. 3, page 84.  The increments are selected from
1843  * formula (8), page 95.  Roughly O(N^3/2).
1844  */
1845 /*
1846  * This is our own private copy of shellsort because we want to sort
1847  * two parallel arrays (the array of buffer pointers and the array of
1848  * logical block numbers) simultaneously.  Note that we cast the array
1849  * of logical block numbers to a unsigned in this routine so that the
1850  * negative block numbers (meta data blocks) sort AFTER the data blocks.
1851  */
1852 
1853 void
1854 lfs_shellsort(struct buf **bp_array, ufs_daddr_t *lb_array, int nmemb)
1855 {
1856 	static int __rsshell_increments[] = { 4, 1, 0 };
1857 	int incr, *incrp, t1, t2;
1858 	struct buf *bp_temp;
1859 	u_long lb_temp;
1860 
1861 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
1862 		for (t1 = incr; t1 < nmemb; ++t1)
1863 			for (t2 = t1 - incr; t2 >= 0;)
1864 				if (lb_array[t2] > lb_array[t2 + incr]) {
1865 					lb_temp = lb_array[t2];
1866 					lb_array[t2] = lb_array[t2 + incr];
1867 					lb_array[t2 + incr] = lb_temp;
1868 					bp_temp = bp_array[t2];
1869 					bp_array[t2] = bp_array[t2 + incr];
1870 					bp_array[t2 + incr] = bp_temp;
1871 					t2 -= incr;
1872 				} else
1873 					break;
1874 }
1875 
1876 /*
1877  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
1878  */
1879 int
1880 lfs_vref(struct vnode *vp)
1881 {
1882 	/*
1883 	 * If we return 1 here during a flush, we risk vinvalbuf() not
1884 	 * being able to flush all of the pages from this vnode, which
1885 	 * will cause it to panic.  So, return 0 if a flush is in progress.
1886 	 */
1887 	if (vp->v_flag & VXLOCK) {
1888 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1889 			return 0;
1890 		}
1891 		return (1);
1892 	}
1893 	return (vget(vp, 0));
1894 }
1895 
1896 /*
1897  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
1898  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
1899  */
1900 void
1901 lfs_vunref(struct vnode *vp)
1902 {
1903 	/*
1904 	 * Analogous to lfs_vref, if the node is flushing, fake it.
1905 	 */
1906 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1907 		return;
1908 	}
1909 
1910 	simple_lock(&vp->v_interlock);
1911 #ifdef DIAGNOSTIC
1912 	if (vp->v_usecount <= 0) {
1913 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
1914 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
1915 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
1916 		panic("lfs_vunref: v_usecount<0");
1917 	}
1918 #endif
1919 	vp->v_usecount--;
1920 	if (vp->v_usecount > 0) {
1921 		simple_unlock(&vp->v_interlock);
1922 		return;
1923 	}
1924 	/*
1925 	 * insert at tail of LRU list
1926 	 */
1927 	simple_lock(&vnode_free_list_slock);
1928 	if (vp->v_holdcnt > 0)
1929 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
1930 	else
1931 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
1932 	simple_unlock(&vnode_free_list_slock);
1933 	simple_unlock(&vp->v_interlock);
1934 }
1935 
1936 /*
1937  * We use this when we have vnodes that were loaded in solely for cleaning.
1938  * There is no reason to believe that these vnodes will be referenced again
1939  * soon, since the cleaning process is unrelated to normal filesystem
1940  * activity.  Putting cleaned vnodes at the tail of the list has the effect
1941  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
1942  * cleaning at the head of the list, instead.
1943  */
1944 void
1945 lfs_vunref_head(struct vnode *vp)
1946 {
1947 	simple_lock(&vp->v_interlock);
1948 #ifdef DIAGNOSTIC
1949 	if (vp->v_usecount == 0) {
1950 		panic("lfs_vunref: v_usecount<0");
1951 	}
1952 #endif
1953 	vp->v_usecount--;
1954 	if (vp->v_usecount > 0) {
1955 		simple_unlock(&vp->v_interlock);
1956 		return;
1957 	}
1958 	/*
1959 	 * insert at head of LRU list
1960 	 */
1961 	simple_lock(&vnode_free_list_slock);
1962 	TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
1963 	simple_unlock(&vnode_free_list_slock);
1964 	simple_unlock(&vp->v_interlock);
1965 }
1966 
1967