xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 08c81a9c2dc8c7300e893321eb65c0925d60871c)
1 /*	$NetBSD: lfs_segment.c,v 1.80 2002/07/06 01:30:13 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 1.80 2002/07/06 01:30:13 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.h>
108 #include <uvm/uvm_extern.h>
109 
110 extern int count_lock_queue(void);
111 extern struct simplelock vnode_free_list_slock;		/* XXX */
112 
113 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
114 static void lfs_super_aiodone(struct buf *);
115 static void lfs_cluster_aiodone(struct buf *);
116 static void lfs_cluster_callback(struct buf *);
117 static struct buf **lookahead_pagemove(struct buf **, int, size_t *);
118 
119 /*
120  * Determine if it's OK to start a partial in this segment, or if we need
121  * to go on to a new segment.
122  */
123 #define	LFS_PARTIAL_FITS(fs) \
124 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
125 	fragstofsb((fs), (fs)->lfs_frag))
126 
127 void	 lfs_callback(struct buf *);
128 int	 lfs_gather(struct lfs *, struct segment *,
129 	     struct vnode *, int (*)(struct lfs *, struct buf *));
130 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
131 void	 lfs_iset(struct inode *, ufs_daddr_t, time_t);
132 int	 lfs_match_fake(struct lfs *, struct buf *);
133 int	 lfs_match_data(struct lfs *, struct buf *);
134 int	 lfs_match_dindir(struct lfs *, struct buf *);
135 int	 lfs_match_indir(struct lfs *, struct buf *);
136 int	 lfs_match_tindir(struct lfs *, struct buf *);
137 void	 lfs_newseg(struct lfs *);
138 void	 lfs_shellsort(struct buf **, ufs_daddr_t *, int);
139 void	 lfs_supercallback(struct buf *);
140 void	 lfs_updatemeta(struct segment *);
141 int	 lfs_vref(struct vnode *);
142 void	 lfs_vunref(struct vnode *);
143 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
144 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
145 int	 lfs_writeseg(struct lfs *, struct segment *);
146 void	 lfs_writesuper(struct lfs *, daddr_t);
147 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
148 	    struct segment *sp, int dirops);
149 
150 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
151 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
152 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
153 int	lfs_dirvcount = 0;		/* # active dirops */
154 
155 /* Statistics Counters */
156 int lfs_dostats = 1;
157 struct lfs_stats lfs_stats;
158 
159 extern int locked_queue_count;
160 extern long locked_queue_bytes;
161 
162 /* op values to lfs_writevnodes */
163 #define	VN_REG	        0
164 #define	VN_DIROP	1
165 #define	VN_EMPTY	2
166 #define VN_CLEAN        3
167 
168 #define LFS_MAX_ACTIVE          10
169 
170 /*
171  * XXX KS - Set modification time on the Ifile, so the cleaner can
172  * read the fs mod time off of it.  We don't set IN_UPDATE here,
173  * since we don't really need this to be flushed to disk (and in any
174  * case that wouldn't happen to the Ifile until we checkpoint).
175  */
176 void
177 lfs_imtime(struct lfs *fs)
178 {
179 	struct timespec ts;
180 	struct inode *ip;
181 
182 	TIMEVAL_TO_TIMESPEC(&time, &ts);
183 	ip = VTOI(fs->lfs_ivnode);
184 	ip->i_ffs_mtime = ts.tv_sec;
185 	ip->i_ffs_mtimensec = ts.tv_nsec;
186 }
187 
188 /*
189  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
190  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
191  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
192  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
193  */
194 
195 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
196 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
197 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
198 
199 int
200 lfs_vflush(struct vnode *vp)
201 {
202 	struct inode *ip;
203 	struct lfs *fs;
204 	struct segment *sp;
205 	struct buf *bp, *nbp, *tbp, *tnbp;
206 	int error, s;
207 
208 	ip = VTOI(vp);
209 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
210 
211 	if (ip->i_flag & IN_CLEANING) {
212 #ifdef DEBUG_LFS
213 		ivndebug(vp,"vflush/in_cleaning");
214 #endif
215 		LFS_CLR_UINO(ip, IN_CLEANING);
216 		LFS_SET_UINO(ip, IN_MODIFIED);
217 
218 		/*
219 		 * Toss any cleaning buffers that have real counterparts
220 		 * to avoid losing new data
221 		 */
222 		s = splbio();
223 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
224 			nbp = LIST_NEXT(bp, b_vnbufs);
225 			if (bp->b_flags & B_CALL) {
226 				for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
227 				    tbp = tnbp)
228 				{
229 					tnbp = LIST_NEXT(tbp, b_vnbufs);
230 					if (tbp->b_vp == bp->b_vp
231 					   && tbp->b_lblkno == bp->b_lblkno
232 					   && tbp != bp)
233 					{
234 						fs->lfs_avail += btofsb(fs, bp->b_bcount);
235 						wakeup(&fs->lfs_avail);
236 						lfs_freebuf(bp);
237 						bp = NULL;
238 						break;
239 					}
240 				}
241 			}
242 		}
243 		splx(s);
244 	}
245 
246 	/* If the node is being written, wait until that is done */
247 	s = splbio();
248 	if (WRITEINPROG(vp)) {
249 #ifdef DEBUG_LFS
250 		ivndebug(vp,"vflush/writeinprog");
251 #endif
252 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
253 	}
254 	splx(s);
255 
256 	/* Protect against VXLOCK deadlock in vinvalbuf() */
257 	lfs_seglock(fs, SEGM_SYNC);
258 
259 	/* If we're supposed to flush a freed inode, just toss it */
260 	/* XXX - seglock, so these buffers can't be gathered, right? */
261 	if (ip->i_ffs_mode == 0) {
262 		printf("lfs_vflush: ino %d is freed, not flushing\n",
263 			ip->i_number);
264 		s = splbio();
265 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
266 			nbp = LIST_NEXT(bp, b_vnbufs);
267 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
268 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
269 				wakeup(&fs->lfs_avail);
270 			}
271 			/* Copied from lfs_writeseg */
272 			if (bp->b_flags & B_CALL) {
273 				/* if B_CALL, it was created with newbuf */
274 				lfs_freebuf(bp);
275 				bp = NULL;
276 			} else {
277 				bremfree(bp);
278 				LFS_UNLOCK_BUF(bp);
279 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
280                                          B_GATHERED);
281 				bp->b_flags |= B_DONE;
282 				reassignbuf(bp, vp);
283 				brelse(bp);
284 			}
285 		}
286 		splx(s);
287 		LFS_CLR_UINO(ip, IN_CLEANING);
288 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
289 		ip->i_flag &= ~IN_ALLMOD;
290 		printf("lfs_vflush: done not flushing ino %d\n",
291 			ip->i_number);
292 		lfs_segunlock(fs);
293 		return 0;
294 	}
295 
296 	SET_FLUSHING(fs,vp);
297 	if (fs->lfs_nactive > LFS_MAX_ACTIVE ||
298 	    (fs->lfs_sp->seg_flags & SEGM_CKP)) {
299 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
300 		CLR_FLUSHING(fs,vp);
301 		lfs_segunlock(fs);
302 		return error;
303 	}
304 	sp = fs->lfs_sp;
305 
306 	if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) {
307 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
308 	} else if ((ip->i_flag & IN_CLEANING) &&
309 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
310 #ifdef DEBUG_LFS
311 		ivndebug(vp,"vflush/clean");
312 #endif
313 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
314 	} else if (lfs_dostats) {
315 		if (LIST_FIRST(&vp->v_dirtyblkhd) || (VTOI(vp)->i_flag & IN_ALLMOD))
316 			++lfs_stats.vflush_invoked;
317 #ifdef DEBUG_LFS
318 		ivndebug(vp,"vflush");
319 #endif
320 	}
321 
322 #ifdef DIAGNOSTIC
323 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
324 	if (vp->v_flag & VDIROP) {
325 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
326 		/* panic("VDIROP being flushed...this can\'t happen"); */
327 	}
328 	if (vp->v_usecount < 0) {
329 		printf("usecount=%ld\n", (long)vp->v_usecount);
330 		panic("lfs_vflush: usecount<0");
331 	}
332 #endif
333 
334 	do {
335 		do {
336 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
337 				lfs_writefile(fs, sp, vp);
338 		} while (lfs_writeinode(fs, sp, ip));
339 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
340 
341 	if (lfs_dostats) {
342 		++lfs_stats.nwrites;
343 		if (sp->seg_flags & SEGM_SYNC)
344 			++lfs_stats.nsync_writes;
345 		if (sp->seg_flags & SEGM_CKP)
346 			++lfs_stats.ncheckpoints;
347 	}
348 	/*
349 	 * If we were called from somewhere that has already held the seglock
350 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
351 	 * the write to complete because we are still locked.
352 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
353 	 * we must explicitly wait, if that is the case.
354 	 *
355 	 * We compare the iocount against 1, not 0, because it is
356 	 * artificially incremented by lfs_seglock().
357 	 */
358 	if (fs->lfs_seglock > 1) {
359 		while (fs->lfs_iocount > 1)
360 			(void)tsleep(&fs->lfs_iocount, PRIBIO + 1,
361 				     "lfs_vflush", 0);
362 	}
363 	lfs_segunlock(fs);
364 
365 	CLR_FLUSHING(fs,vp);
366 	return (0);
367 }
368 
369 #ifdef DEBUG_LFS_VERBOSE
370 # 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)
371 #else
372 # define vndebug(vp,str)
373 #endif
374 
375 int
376 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
377 {
378 	struct inode *ip;
379 	struct vnode *vp, *nvp;
380 	int inodes_written = 0, only_cleaning;
381 	int needs_unlock;
382 
383 #ifndef LFS_NO_BACKVP_HACK
384 	/* BEGIN HACK */
385 #define	VN_OFFSET	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
386 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
387 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&(LIST_FIRST(&mp->mnt_vnodelist))) - VN_OFFSET))
388 
389 	/* Find last vnode. */
390  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
391 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
392 	     vp = LIST_NEXT(vp, v_mntvnodes));
393 	for (; vp && vp != BEG_OF_VLIST; vp = nvp) {
394 		nvp = BACK_VP(vp);
395 #else
396 	loop:
397 	for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) {
398 		nvp = LIST_NEXT(vp, v_mntvnodes);
399 #endif
400 		/*
401 		 * If the vnode that we are about to sync is no longer
402 		 * associated with this mount point, start over.
403 		 */
404 		if (vp->v_mount != mp) {
405 			printf("lfs_writevnodes: starting over\n");
406 			goto loop;
407 		}
408 
409 		ip = VTOI(vp);
410 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
411 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
412 			vndebug(vp,"dirop");
413 			continue;
414 		}
415 
416 		if (op == VN_EMPTY && LIST_FIRST(&vp->v_dirtyblkhd)) {
417 			vndebug(vp,"empty");
418 			continue;
419 		}
420 
421 		if (vp->v_type == VNON) {
422 			continue;
423 		}
424 
425 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
426 		   && vp != fs->lfs_flushvp
427 		   && !(ip->i_flag & IN_CLEANING)) {
428 			vndebug(vp,"cleaning");
429 			continue;
430 		}
431 
432 		if (lfs_vref(vp)) {
433 			vndebug(vp,"vref");
434 			continue;
435 		}
436 
437 		needs_unlock = 0;
438 		if (VOP_ISLOCKED(vp)) {
439 			if (vp != fs->lfs_ivnode &&
440 			    vp->v_lock.lk_lockholder != curproc->p_pid) {
441 #ifdef DEBUG_LFS
442 				printf("lfs_writevnodes: not writing ino %d,"
443 				       " locked by pid %d\n",
444 				       VTOI(vp)->i_number,
445 				       vp->v_lock.lk_lockholder);
446 #endif
447 				lfs_vunref(vp);
448 				continue;
449 			}
450 		} else if (vp != fs->lfs_ivnode) {
451 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
452 			needs_unlock = 1;
453 		}
454 
455 		only_cleaning = 0;
456 		/*
457 		 * Write the inode/file if dirty and it's not the IFILE.
458 		 */
459 		if ((ip->i_flag & IN_ALLMOD) ||
460 		     (LIST_FIRST(&vp->v_dirtyblkhd) != NULL))
461 		{
462 			only_cleaning = ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
463 
464 			if (ip->i_number != LFS_IFILE_INUM
465 			   && LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
466 			{
467 				lfs_writefile(fs, sp, vp);
468 			}
469 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
470 				if (WRITEINPROG(vp)) {
471 #ifdef DEBUG_LFS
472 					ivndebug(vp,"writevnodes/write2");
473 #endif
474 				} else if (!(ip->i_flag & IN_ALLMOD)) {
475 #ifdef DEBUG_LFS
476 					printf("<%d>",ip->i_number);
477 #endif
478 					LFS_SET_UINO(ip, IN_MODIFIED);
479 				}
480 			}
481 			(void) lfs_writeinode(fs, sp, ip);
482 			inodes_written++;
483 		}
484 
485 		if (needs_unlock)
486 			VOP_UNLOCK(vp, 0);
487 
488 		if (lfs_clean_vnhead && only_cleaning)
489 			lfs_vunref_head(vp);
490 		else
491 			lfs_vunref(vp);
492 	}
493 	return inodes_written;
494 }
495 
496 /*
497  * Do a checkpoint.
498  */
499 int
500 lfs_segwrite(struct mount *mp, int flags)
501 {
502 	struct buf *bp;
503 	struct inode *ip;
504 	struct lfs *fs;
505 	struct segment *sp;
506 	struct vnode *vp;
507 	SEGUSE *segusep;
508 	ufs_daddr_t ibno;
509 	int do_ckp, did_ckp, error, i;
510 	int writer_set = 0;
511 	int dirty;
512 	int redo;
513 
514 	fs = VFSTOUFS(mp)->um_lfs;
515 
516 	if (fs->lfs_ronly)
517 		return EROFS;
518 
519 	lfs_imtime(fs);
520 
521 	/* printf("lfs_segwrite: ifile flags are 0x%lx\n",
522 	       (long)(VTOI(fs->lfs_ivnode)->i_flag)); */
523 
524 #if 0
525 	/*
526 	 * If we are not the cleaner, and there is no space available,
527 	 * wait until cleaner writes.
528 	 */
529 	if (!(flags & SEGM_CLEAN) && !(fs->lfs_seglock && fs->lfs_sp &&
530 				      (fs->lfs_sp->seg_flags & SEGM_CLEAN)))
531 	{
532 		while (fs->lfs_avail <= 0) {
533 			LFS_CLEANERINFO(cip, fs, bp);
534 			LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
535 
536 			wakeup(&lfs_allclean_wakeup);
537 			wakeup(&fs->lfs_nextseg);
538 			error = tsleep(&fs->lfs_avail, PRIBIO + 1, "lfs_av2",
539 				       0);
540 			if (error) {
541 				return (error);
542 			}
543 		}
544 	}
545 #endif
546 	/*
547 	 * Allocate a segment structure and enough space to hold pointers to
548 	 * the maximum possible number of buffers which can be described in a
549 	 * single summary block.
550 	 */
551 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
552 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
553 	sp = fs->lfs_sp;
554 
555 	/*
556 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
557 	 * in which case we have to flush *all* buffers off of this vnode.
558 	 * We don't care about other nodes, but write any non-dirop nodes
559 	 * anyway in anticipation of another getnewvnode().
560 	 *
561 	 * If we're cleaning we only write cleaning and ifile blocks, and
562 	 * no dirops, since otherwise we'd risk corruption in a crash.
563 	 */
564 	if (sp->seg_flags & SEGM_CLEAN)
565 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
566 	else {
567 		lfs_writevnodes(fs, mp, sp, VN_REG);
568 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
569 			while (fs->lfs_dirops)
570 				if ((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
571 						"lfs writer", 0)))
572 				{
573 					/* XXX why not segunlock? */
574 					free(sp->bpp, M_SEGMENT);
575 					sp->bpp = NULL;
576 					free(sp, M_SEGMENT);
577 					fs->lfs_sp = NULL;
578 					return (error);
579 				}
580 			fs->lfs_writer++;
581 			writer_set = 1;
582 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
583 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
584 		}
585 	}
586 
587 	/*
588 	 * If we are doing a checkpoint, mark everything since the
589 	 * last checkpoint as no longer ACTIVE.
590 	 */
591 	if (do_ckp) {
592 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
593 		     --ibno >= fs->lfs_cleansz; ) {
594 			dirty = 0;
595 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
596 
597 				panic("lfs_segwrite: ifile read");
598 			segusep = (SEGUSE *)bp->b_data;
599 			for (i = fs->lfs_sepb; i--;) {
600 				if (segusep->su_flags & SEGUSE_ACTIVE) {
601 					segusep->su_flags &= ~SEGUSE_ACTIVE;
602 					++dirty;
603 				}
604 				if (fs->lfs_version > 1)
605 					++segusep;
606 				else
607 					segusep = (SEGUSE *)
608 						((SEGUSE_V1 *)segusep + 1);
609 			}
610 
611 			/* But the current segment is still ACTIVE */
612 			segusep = (SEGUSE *)bp->b_data;
613 			if (dtosn(fs, fs->lfs_curseg) / fs->lfs_sepb ==
614 			    (ibno-fs->lfs_cleansz)) {
615 				if (fs->lfs_version > 1)
616 					segusep[dtosn(fs, fs->lfs_curseg) %
617 					     fs->lfs_sepb].su_flags |=
618 						     SEGUSE_ACTIVE;
619 				else
620 					((SEGUSE *)
621 					 ((SEGUSE_V1 *)(bp->b_data) +
622 					  (dtosn(fs, fs->lfs_curseg) %
623 					   fs->lfs_sepb)))->su_flags
624 						   |= SEGUSE_ACTIVE;
625 				--dirty;
626 			}
627 			if (dirty)
628 				error = LFS_BWRITE_LOG(bp); /* Ifile */
629 			else
630 				brelse(bp);
631 		}
632 	}
633 
634 	did_ckp = 0;
635 	if (do_ckp || fs->lfs_doifile) {
636 		do {
637 			vp = fs->lfs_ivnode;
638 
639 			vget(vp, LK_EXCLUSIVE | LK_CANRECURSE | LK_RETRY);
640 #ifdef DEBUG
641 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
642 #endif
643 			fs->lfs_flags &= ~LFS_IFDIRTY;
644 
645 			ip = VTOI(vp);
646 			/* if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) */
647 				lfs_writefile(fs, sp, vp);
648 			if (ip->i_flag & IN_ALLMOD)
649 				++did_ckp;
650 			redo = lfs_writeinode(fs, sp, ip);
651 
652 			vput(vp);
653 			redo += lfs_writeseg(fs, sp);
654 			redo += (fs->lfs_flags & LFS_IFDIRTY);
655 		} while (redo && do_ckp);
656 
657 		/* The ifile should now be all clear */
658 		if (do_ckp && LIST_FIRST(&vp->v_dirtyblkhd)) {
659 			struct buf *bp;
660 			int s, warned = 0, dopanic = 0;
661 			s = splbio();
662 			for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = LIST_NEXT(bp, b_vnbufs)) {
663 				if (!(bp->b_flags & B_GATHERED)) {
664 					if (!warned)
665 						printf("lfs_segwrite: ifile still has dirty blocks?!\n");
666 					++dopanic;
667 					++warned;
668 					printf("bp=%p, lbn %d, flags 0x%lx\n",
669 						bp, bp->b_lblkno, bp->b_flags);
670 				}
671 			}
672 			if (dopanic)
673 				panic("dirty blocks");
674 			splx(s);
675 		}
676 		LFS_CLR_UINO(ip, IN_ALLMOD);
677 	} else {
678 		(void) lfs_writeseg(fs, sp);
679 	}
680 
681 	/*
682 	 * If the I/O count is non-zero, sleep until it reaches zero.
683 	 * At the moment, the user's process hangs around so we can
684 	 * sleep.
685 	 */
686 	fs->lfs_doifile = 0;
687 	if (writer_set && --fs->lfs_writer == 0)
688 		wakeup(&fs->lfs_dirops);
689 
690 	/*
691 	 * If we didn't write the Ifile, we didn't really do anything.
692 	 * That means that (1) there is a checkpoint on disk and (2)
693 	 * nothing has changed since it was written.
694 	 *
695 	 * Take the flags off of the segment so that lfs_segunlock
696 	 * doesn't have to write the superblock either.
697 	 */
698 	if (do_ckp && !did_ckp) {
699 		sp->seg_flags &= ~SEGM_CKP;
700 		/* if (do_ckp) printf("lfs_segwrite: no checkpoint\n"); */
701 	}
702 
703 	if (lfs_dostats) {
704 		++lfs_stats.nwrites;
705 		if (sp->seg_flags & SEGM_SYNC)
706 			++lfs_stats.nsync_writes;
707 		if (sp->seg_flags & SEGM_CKP)
708 			++lfs_stats.ncheckpoints;
709 	}
710 	lfs_segunlock(fs);
711 	return (0);
712 }
713 
714 /*
715  * Write the dirty blocks associated with a vnode.
716  */
717 void
718 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
719 {
720 	struct buf *bp;
721 	struct finfo *fip;
722 	struct inode *ip;
723 	IFILE *ifp;
724 	int i, frag;
725 
726 	ip = VTOI(vp);
727 
728 	if (sp->seg_bytes_left < fs->lfs_bsize ||
729 	    sp->sum_bytes_left < sizeof(struct finfo))
730 		(void) lfs_writeseg(fs, sp);
731 
732 	sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t);
733 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
734 
735 	if (vp->v_flag & VDIROP)
736 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
737 
738 	fip = sp->fip;
739 	fip->fi_nblocks = 0;
740 	fip->fi_ino = ip->i_number;
741 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
742 	fip->fi_version = ifp->if_version;
743 	brelse(bp);
744 
745 	if (sp->seg_flags & SEGM_CLEAN) {
746 		lfs_gather(fs, sp, vp, lfs_match_fake);
747 		/*
748 		 * For a file being flushed, we need to write *all* blocks.
749 		 * This means writing the cleaning blocks first, and then
750 		 * immediately following with any non-cleaning blocks.
751 		 * The same is true of the Ifile since checkpoints assume
752 		 * that all valid Ifile blocks are written.
753 		 */
754 	   	if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode)
755 			lfs_gather(fs, sp, vp, lfs_match_data);
756 	} else
757 		lfs_gather(fs, sp, vp, lfs_match_data);
758 
759 	/*
760 	 * It may not be necessary to write the meta-data blocks at this point,
761 	 * as the roll-forward recovery code should be able to reconstruct the
762 	 * list.
763 	 *
764 	 * We have to write them anyway, though, under two conditions: (1) the
765 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
766 	 * checkpointing.
767 	 *
768 	 * BUT if we are cleaning, we might have indirect blocks that refer to
769 	 * new blocks not being written yet, in addition to fragments being
770 	 * moved out of a cleaned segment.  If that is the case, don't
771 	 * write the indirect blocks, or the finfo will have a small block
772 	 * in the middle of it!
773 	 * XXX in this case isn't the inode size wrong too?
774 	 */
775 	frag = 0;
776 	if (sp->seg_flags & SEGM_CLEAN) {
777 		for (i = 0; i < NDADDR; i++)
778 			if (ip->i_lfs_fragsize[i] > 0 &&
779 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
780 				++frag;
781 	}
782 #ifdef DIAGNOSTIC
783 	if (frag > 1)
784 		panic("lfs_writefile: more than one fragment!");
785 #endif
786 	if (IS_FLUSHING(fs, vp) ||
787 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
788 		lfs_gather(fs, sp, vp, lfs_match_indir);
789 		lfs_gather(fs, sp, vp, lfs_match_dindir);
790 		lfs_gather(fs, sp, vp, lfs_match_tindir);
791 	}
792 	fip = sp->fip;
793 	if (fip->fi_nblocks != 0) {
794 		sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) +
795 				   sizeof(ufs_daddr_t) * (fip->fi_nblocks-1));
796 		sp->start_lbp = &sp->fip->fi_blocks[0];
797 	} else {
798 		sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t);
799 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
800 	}
801 }
802 
803 int
804 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
805 {
806 	struct buf *bp, *ibp;
807 	struct dinode *cdp;
808 	IFILE *ifp;
809 	SEGUSE *sup;
810 	ufs_daddr_t daddr;
811 	daddr_t *daddrp;
812 	ino_t ino;
813 	int error, i, ndx, fsb = 0;
814 	int redo_ifile = 0;
815 	struct timespec ts;
816 	int gotblk = 0;
817 
818 	if (!(ip->i_flag & IN_ALLMOD))
819 		return (0);
820 
821 	/* Allocate a new inode block if necessary. */
822 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
823 		/* Allocate a new segment if necessary. */
824 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
825 		    sp->sum_bytes_left < sizeof(ufs_daddr_t))
826 			(void) lfs_writeseg(fs, sp);
827 
828 		/* Get next inode block. */
829 		daddr = fs->lfs_offset;
830 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
831 		sp->ibp = *sp->cbpp++ =
832 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
833 			       fs->lfs_ibsize, 0, 0);
834 		gotblk++;
835 
836 		/* Zero out inode numbers */
837 		for (i = 0; i < INOPB(fs); ++i)
838 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
839 
840 		++sp->start_bpp;
841 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
842 		/* Set remaining space counters. */
843 		sp->seg_bytes_left -= fs->lfs_ibsize;
844 		sp->sum_bytes_left -= sizeof(ufs_daddr_t);
845 		ndx = fs->lfs_sumsize / sizeof(ufs_daddr_t) -
846 			sp->ninodes / INOPB(fs) - 1;
847 		((ufs_daddr_t *)(sp->segsum))[ndx] = daddr;
848 	}
849 
850 	/* Update the inode times and copy the inode onto the inode page. */
851 	TIMEVAL_TO_TIMESPEC(&time, &ts);
852 	/* XXX kludge --- don't redirty the ifile just to put times on it */
853 	if (ip->i_number != LFS_IFILE_INUM)
854 		LFS_ITIMES(ip, &ts, &ts, &ts);
855 
856 	/*
857 	 * If this is the Ifile, and we've already written the Ifile in this
858 	 * partial segment, just overwrite it (it's not on disk yet) and
859 	 * continue.
860 	 *
861 	 * XXX we know that the bp that we get the second time around has
862 	 * already been gathered.
863 	 */
864 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
865 		*(sp->idp) = ip->i_din.ffs_din;
866 		ip->i_lfs_osize = ip->i_ffs_size;
867 		return 0;
868 	}
869 
870 	bp = sp->ibp;
871 	cdp = ((struct dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
872 	*cdp = ip->i_din.ffs_din;
873 #ifdef LFS_IFILE_FRAG_ADDRESSING
874 	if (fs->lfs_version > 1)
875 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
876 #endif
877 
878 	/*
879 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
880 	 * addresses to disk; possibly revert the inode size.
881 	 */
882 	if (ip->i_lfs_effnblks != ip->i_ffs_blocks) {
883 		cdp->di_size = ip->i_lfs_osize;
884 #ifdef DEBUG_LFS
885 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
886 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs_blocks);
887 #endif
888 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
889 		     daddrp++) {
890 			if (*daddrp == UNWRITTEN) {
891 #ifdef DEBUG_LFS
892 				printf("lfs_writeinode: wiping UNWRITTEN\n");
893 #endif
894 				*daddrp = 0;
895 			}
896 		}
897 	} else {
898 		/* If all blocks are goig to disk, update the "size on disk" */
899 		ip->i_lfs_osize = ip->i_ffs_size;
900 	}
901 
902 	if (ip->i_flag & IN_CLEANING)
903 		LFS_CLR_UINO(ip, IN_CLEANING);
904 	else {
905 		/* XXX IN_ALLMOD */
906 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
907 			     IN_UPDATE);
908 		if (ip->i_lfs_effnblks == ip->i_ffs_blocks)
909 			LFS_CLR_UINO(ip, IN_MODIFIED);
910 #ifdef DEBUG_LFS
911 		else
912 			printf("lfs_writeinode: ino %d: real blks=%d, "
913 			       "eff=%d\n", ip->i_number, ip->i_ffs_blocks,
914 			       ip->i_lfs_effnblks);
915 #endif
916 	}
917 
918 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
919 		sp->idp = ((struct dinode *)bp->b_data) +
920 			(sp->ninodes % INOPB(fs));
921 	if (gotblk) {
922 		LFS_LOCK_BUF(bp);
923 		brelse(bp);
924 	}
925 
926 	/* Increment inode count in segment summary block. */
927 	++((SEGSUM *)(sp->segsum))->ss_ninos;
928 
929 	/* If this page is full, set flag to allocate a new page. */
930 	if (++sp->ninodes % INOPB(fs) == 0)
931 		sp->ibp = NULL;
932 
933 	/*
934 	 * If updating the ifile, update the super-block.  Update the disk
935 	 * address and access times for this inode in the ifile.
936 	 */
937 	ino = ip->i_number;
938 	if (ino == LFS_IFILE_INUM) {
939 		daddr = fs->lfs_idaddr;
940 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
941 	} else {
942 		LFS_IENTRY(ifp, fs, ino, ibp);
943 		daddr = ifp->if_daddr;
944 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
945 #ifdef LFS_DEBUG_NEXTFREE
946 		if (ino > 3 && ifp->if_nextfree) {
947 			vprint("lfs_writeinode",ITOV(ip));
948 			printf("lfs_writeinode: updating free ino %d\n",
949 				ip->i_number);
950 		}
951 #endif
952 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
953 	}
954 
955 	/*
956 	 * The inode's last address should not be in the current partial
957 	 * segment, except under exceptional circumstances (lfs_writevnodes
958 	 * had to start over, and in the meantime more blocks were written
959 	 * to a vnode).  Both inodes will be accounted to this segment
960 	 * in lfs_writeseg so we need to subtract the earlier version
961 	 * here anyway.  The segment count can temporarily dip below
962 	 * zero here; keep track of how many duplicates we have in
963 	 * "dupino" so we don't panic below.
964 	 */
965 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno)) {
966 		++sp->ndupino;
967 		printf("lfs_writeinode: last inode addr in current pseg "
968 		       "(ino %d daddr 0x%x) ndupino=%d\n", ino, daddr,
969 			sp->ndupino);
970 	}
971 	/*
972 	 * Account the inode: it no longer belongs to its former segment,
973 	 * though it will not belong to the new segment until that segment
974 	 * is actually written.
975 	 */
976 	if (daddr != LFS_UNUSED_DADDR) {
977 		LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
978 #ifdef DIAGNOSTIC
979 		if (sup->su_nbytes < DINODE_SIZE * (1 + sp->ndupino)) {
980 			printf("lfs_writeinode: negative bytes "
981 			       "(segment %d short by %d)\n",
982 			       dtosn(fs, daddr),
983 			       (int)DINODE_SIZE - sup->su_nbytes);
984 			panic("lfs_writeinode: negative bytes");
985 			sup->su_nbytes = DINODE_SIZE;
986 		}
987 #endif
988 #ifdef DEBUG_SU_NBYTES
989 		printf("seg %d -= %d for ino %d inode\n",
990 		       dtosn(fs, daddr), DINODE_SIZE, ino);
991 #endif
992 		sup->su_nbytes -= DINODE_SIZE;
993 		redo_ifile =
994 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
995 		if (redo_ifile)
996 			fs->lfs_flags |= LFS_IFDIRTY;
997 		error = LFS_BWRITE_LOG(bp); /* Ifile */
998 	}
999 	return (redo_ifile);
1000 }
1001 
1002 int
1003 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1004 {
1005 	struct lfs *fs;
1006 	int version;
1007 
1008 	/*
1009 	 * If full, finish this segment.  We may be doing I/O, so
1010 	 * release and reacquire the splbio().
1011 	 */
1012 #ifdef DIAGNOSTIC
1013 	if (sp->vp == NULL)
1014 		panic ("lfs_gatherblock: Null vp in segment");
1015 #endif
1016 	fs = sp->fs;
1017 	if (sp->sum_bytes_left < sizeof(ufs_daddr_t) ||
1018 	    sp->seg_bytes_left < bp->b_bcount) {
1019 		if (sptr)
1020 			splx(*sptr);
1021 		lfs_updatemeta(sp);
1022 
1023 		version = sp->fip->fi_version;
1024 		(void) lfs_writeseg(fs, sp);
1025 
1026 		sp->fip->fi_version = version;
1027 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
1028 		/* Add the current file to the segment summary. */
1029 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
1030 		sp->sum_bytes_left -=
1031 			sizeof(struct finfo) - sizeof(ufs_daddr_t);
1032 
1033 		if (sptr)
1034 			*sptr = splbio();
1035 		return (1);
1036 	}
1037 
1038 #ifdef DEBUG
1039 	if (bp->b_flags & B_GATHERED) {
1040 		printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n",
1041 		       sp->fip->fi_ino, bp->b_lblkno);
1042 		return (0);
1043 	}
1044 #endif
1045 	/* Insert into the buffer list, update the FINFO block. */
1046 	bp->b_flags |= B_GATHERED;
1047 	bp->b_flags &= ~B_DONE;
1048 
1049 	*sp->cbpp++ = bp;
1050 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
1051 
1052 	sp->sum_bytes_left -= sizeof(ufs_daddr_t);
1053 	sp->seg_bytes_left -= bp->b_bcount;
1054 	return (0);
1055 }
1056 
1057 int
1058 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, int (*match)(struct lfs *, struct buf *))
1059 {
1060 	struct buf *bp, *nbp;
1061 	int s, count = 0;
1062 
1063 	sp->vp = vp;
1064 	s = splbio();
1065 
1066 #ifndef LFS_NO_BACKBUF_HACK
1067 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1068 # define	BUF_OFFSET	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1069 # define	BACK_BUF(BP)	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1070 # define	BEG_OF_LIST	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1071 /* Find last buffer. */
1072 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1073 	    bp = LIST_NEXT(bp, b_vnbufs));
1074 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1075 		nbp = BACK_BUF(bp);
1076 #else /* LFS_NO_BACKBUF_HACK */
1077 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1078 		nbp = LIST_NEXT(bp, b_vnbufs);
1079 #endif /* LFS_NO_BACKBUF_HACK */
1080 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1081 #ifdef DEBUG_LFS
1082 			if (vp == fs->lfs_ivnode && (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1083 				printf("(%d:%lx)", bp->b_lblkno, bp->b_flags);
1084 #endif
1085 			continue;
1086 		}
1087 		if (vp->v_type == VBLK) {
1088 			/* For block devices, just write the blocks. */
1089 			/* XXX Do we really need to even do this? */
1090 #ifdef DEBUG_LFS
1091 			if (count == 0)
1092 				printf("BLK(");
1093 			printf(".");
1094 #endif
1095 			/* Get the block before bwrite, so we don't corrupt the free list */
1096 			bp->b_flags |= B_BUSY;
1097 			bremfree(bp);
1098 			bwrite(bp);
1099 		} else {
1100 #ifdef DIAGNOSTIC
1101 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1102 				printf("lfs_gather: lbn %d is B_INVAL\n",
1103 					bp->b_lblkno);
1104 				VOP_PRINT(bp->b_vp);
1105 			}
1106 			if (!(bp->b_flags & B_DELWRI))
1107 				panic("lfs_gather: bp not B_DELWRI");
1108 			if (!(bp->b_flags & B_LOCKED)) {
1109 				printf("lfs_gather: lbn %d blk %d"
1110 				       " not B_LOCKED\n", bp->b_lblkno,
1111 				       dbtofsb(fs, bp->b_blkno));
1112 				VOP_PRINT(bp->b_vp);
1113 				panic("lfs_gather: bp not B_LOCKED");
1114 			}
1115 #endif
1116 			if (lfs_gatherblock(sp, bp, &s)) {
1117 				goto loop;
1118 			}
1119 		}
1120 		count++;
1121 	}
1122 	splx(s);
1123 #ifdef DEBUG_LFS
1124 	if (vp->v_type == VBLK && count)
1125 		printf(")\n");
1126 #endif
1127 	lfs_updatemeta(sp);
1128 	sp->vp = NULL;
1129 	return count;
1130 }
1131 
1132 /*
1133  * Update the metadata that points to the blocks listed in the FINFO
1134  * array.
1135  */
1136 void
1137 lfs_updatemeta(struct segment *sp)
1138 {
1139 	SEGUSE *sup;
1140 	struct buf *bp, *sbp;
1141 	struct lfs *fs;
1142 	struct vnode *vp;
1143 	struct indir a[NIADDR + 2], *ap;
1144 	struct inode *ip;
1145 	ufs_daddr_t daddr, lbn, off;
1146 	daddr_t ooff;
1147 	int error, i, nblocks, num;
1148 	int bb, osize, obb;
1149 
1150 	vp = sp->vp;
1151 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1152 	if (nblocks < 0)
1153 		panic("This is a bad thing\n");
1154 	if (vp == NULL || nblocks == 0)
1155 		return;
1156 
1157 	/* Sort the blocks. */
1158 	/*
1159 	 * XXX KS - We have to sort even if the blocks come from the
1160 	 * cleaner, because there might be other pending blocks on the
1161 	 * same inode...and if we don't sort, and there are fragments
1162 	 * present, blocks may be written in the wrong place.
1163 	 */
1164 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
1165 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
1166 
1167 	/*
1168 	 * Record the length of the last block in case it's a fragment.
1169 	 * If there are indirect blocks present, they sort last.  An
1170 	 * indirect block will be lfs_bsize and its presence indicates
1171 	 * that you cannot have fragments.
1172 	 *
1173 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1174 	 * XXX a later indirect block.  This will continue to be
1175 	 * XXX true until lfs_markv is fixed to do everything with
1176 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1177 	 */
1178 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1179 
1180 	/*
1181 	 * Assign disk addresses, and update references to the logical
1182 	 * block and the segment usage information.
1183 	 */
1184 	fs = sp->fs;
1185 	for (i = nblocks; i--; ++sp->start_bpp) {
1186 		lbn = *sp->start_lbp++;
1187 		sbp = *sp->start_bpp;
1188 
1189 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1190 		off = fs->lfs_offset;
1191 		if (sbp->b_blkno == sbp->b_lblkno) {
1192 			printf("lfs_updatemeta: ino %d blk %d"
1193 			       " has same lbn and daddr\n",
1194 			       VTOI(vp)->i_number, off);
1195 		}
1196 
1197 		/*
1198 		 * If we write a frag in the wrong place, the cleaner won't
1199 		 * be able to correctly identify its size later, and the
1200 		 * segment will be uncleanable.  (Even worse, it will assume
1201 		 * that the indirect block that actually ends the list
1202 		 * is of a smaller size!)
1203 		 */
1204 		if (sbp->b_bcount < fs->lfs_bsize && i != 0)
1205 			panic("lfs_updatemeta: fragment is not last block\n");
1206 
1207 		bb = fragstofsb(fs, numfrags(fs, sbp->b_bcount));
1208 		fs->lfs_offset += bb;
1209 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1210 		if (daddr > 0)
1211 			daddr = dbtofsb(fs, daddr);
1212 		if (error)
1213 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1214 		ip = VTOI(vp);
1215 		switch (num) {
1216 		case 0:
1217 			ooff = ip->i_ffs_db[lbn];
1218 #ifdef DEBUG
1219 			if (ooff == 0) {
1220 				printf("lfs_updatemeta[1]: warning: writing "
1221 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1222 				       ip->i_number, lbn, off);
1223 			}
1224 #endif
1225 			if (ooff == UNWRITTEN)
1226 				ip->i_ffs_blocks += bb;
1227 			else {
1228 				/* possible fragment truncation or extension */
1229 				obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1230 				ip->i_ffs_blocks += (bb - obb);
1231 			}
1232 			ip->i_ffs_db[lbn] = off;
1233 			break;
1234 		case 1:
1235 			ooff = ip->i_ffs_ib[a[0].in_off];
1236 #ifdef DEBUG
1237 			if (ooff == 0) {
1238 				printf("lfs_updatemeta[2]: warning: writing "
1239 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1240 				       ip->i_number, lbn, off);
1241 			}
1242 #endif
1243 			if (ooff == UNWRITTEN)
1244 				ip->i_ffs_blocks += bb;
1245 			ip->i_ffs_ib[a[0].in_off] = off;
1246 			break;
1247 		default:
1248 			ap = &a[num - 1];
1249 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1250 				panic("lfs_updatemeta: bread bno %d",
1251 				      ap->in_lbn);
1252 
1253 			ooff = ((ufs_daddr_t *)bp->b_data)[ap->in_off];
1254 #if DEBUG
1255 			if (ooff == 0) {
1256 				printf("lfs_updatemeta[3]: warning: writing "
1257 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1258 				       ip->i_number, lbn, off);
1259 			}
1260 #endif
1261 			if (ooff == UNWRITTEN)
1262 				ip->i_ffs_blocks += bb;
1263 			((ufs_daddr_t *)bp->b_data)[ap->in_off] = off;
1264 			(void) VOP_BWRITE(bp);
1265 		}
1266 #ifdef DEBUG
1267 		if (daddr >= fs->lfs_lastpseg && daddr <= off) {
1268 			printf("lfs_updatemeta: ino %d, lbn %d, addr = %x "
1269 			       "in same pseg\n", VTOI(sp->vp)->i_number,
1270 			       sbp->b_lblkno, daddr);
1271 		}
1272 #endif
1273 		/*
1274 		 * Update segment usage information, based on old size
1275 		 * and location.
1276 		 */
1277 		if (daddr > 0) {
1278 			if (lbn >= 0 && lbn < NDADDR)
1279 				osize = ip->i_lfs_fragsize[lbn];
1280 			else
1281 				osize = fs->lfs_bsize;
1282 			LFS_SEGENTRY(sup, fs, dtosn(fs, daddr), bp);
1283 #ifdef DIAGNOSTIC
1284 			if (sup->su_nbytes < osize + DINODE_SIZE * sp->ndupino) {
1285 				printf("lfs_updatemeta: negative bytes "
1286 				       "(segment %d short by %d)\n",
1287 				       dtosn(fs, daddr),
1288 				       osize - sup->su_nbytes);
1289 				printf("lfs_updatemeta: ino %d, lbn %d, "
1290 				       "addr = 0x%x\n", VTOI(sp->vp)->i_number,
1291 				       lbn, daddr);
1292 				panic("lfs_updatemeta: negative bytes");
1293 				sup->su_nbytes = osize + DINODE_SIZE * sp->ndupino;
1294 			}
1295 #endif
1296 #ifdef DEBUG_SU_NBYTES
1297 			printf("seg %d -= %ld for ino %d lbn %d db 0x%x\n",
1298 			       dtosn(fs, daddr), osize, VTOI(sp->vp)->i_number,
1299 			       lbn, daddr);
1300 #endif
1301 			sup->su_nbytes -= osize;
1302 			if (!(bp->b_flags & B_GATHERED))
1303 				fs->lfs_flags |= LFS_IFDIRTY;
1304 			error = LFS_BWRITE_LOG(bp); /* Ifile */
1305 		}
1306 		/*
1307 		 * Now that this block has a new address, and its old
1308 		 * segment no longer owns it, we can forget about its
1309 		 * old size.
1310 		 */
1311 		if (lbn >= 0 && lbn < NDADDR)
1312 			ip->i_lfs_fragsize[lbn] = sbp->b_bcount;
1313 	}
1314 }
1315 
1316 /*
1317  * Start a new segment.
1318  */
1319 int
1320 lfs_initseg(struct lfs *fs)
1321 {
1322 	struct segment *sp;
1323 	SEGUSE *sup;
1324 	SEGSUM *ssp;
1325 	struct buf *bp, *sbp;
1326 	int repeat;
1327 
1328 	sp = fs->lfs_sp;
1329 
1330 	repeat = 0;
1331 	/* Advance to the next segment. */
1332 	if (!LFS_PARTIAL_FITS(fs)) {
1333 		/* lfs_avail eats the remaining space */
1334 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1335 						   fs->lfs_curseg);
1336 		/* Wake up any cleaning procs waiting on this file system. */
1337 		wakeup(&lfs_allclean_wakeup);
1338 		wakeup(&fs->lfs_nextseg);
1339 		lfs_newseg(fs);
1340 		repeat = 1;
1341 		fs->lfs_offset = fs->lfs_curseg;
1342 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1343 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1344 		/*
1345 		 * If the segment contains a superblock, update the offset
1346 		 * and summary address to skip over it.
1347 		 */
1348 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1349 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1350 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1351 			sp->seg_bytes_left -= LFS_SBPAD;
1352 		}
1353 		brelse(bp);
1354 		/* Segment zero could also contain the labelpad */
1355 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1356 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1357 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1358 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1359 		}
1360 	} else {
1361 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1362 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1363 				      (fs->lfs_offset - fs->lfs_curseg));
1364 	}
1365 	fs->lfs_lastpseg = fs->lfs_offset;
1366 
1367 	sp->fs = fs;
1368 	sp->ibp = NULL;
1369 	sp->idp = NULL;
1370 	sp->ninodes = 0;
1371 	sp->ndupino = 0;
1372 
1373 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1374 	sp->cbpp = sp->bpp;
1375 #ifdef LFS_MALLOC_SUMMARY
1376 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1377 				     fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize);
1378   	sp->segsum = (*sp->cbpp)->b_data;
1379 #else
1380 	sbp = *sp->cbpp = getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1381 				 fsbtodb(fs, fs->lfs_offset), NBPG, 0, 0);
1382 	memset(sbp->b_data, 0x5a, NBPG);
1383 	sp->segsum = (*sp->cbpp)->b_data + NBPG - fs->lfs_sumsize;
1384 #endif
1385 	bzero(sp->segsum, fs->lfs_sumsize);
1386 	sp->start_bpp = ++sp->cbpp;
1387 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1388 
1389 	/* Set point to SEGSUM, initialize it. */
1390 	ssp = sp->segsum;
1391 	ssp->ss_next = fs->lfs_nextseg;
1392 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1393 	ssp->ss_magic = SS_MAGIC;
1394 
1395 	/* Set pointer to first FINFO, initialize it. */
1396 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1397 	sp->fip->fi_nblocks = 0;
1398 	sp->start_lbp = &sp->fip->fi_blocks[0];
1399 	sp->fip->fi_lastlength = 0;
1400 
1401 	sp->seg_bytes_left -= fs->lfs_sumsize;
1402 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1403 
1404 #ifndef LFS_MALLOC_SUMMARY
1405 	LFS_LOCK_BUF(sbp);
1406 	brelse(sbp);
1407 #endif
1408 	return (repeat);
1409 }
1410 
1411 /*
1412  * Return the next segment to write.
1413  */
1414 void
1415 lfs_newseg(struct lfs *fs)
1416 {
1417 	CLEANERINFO *cip;
1418 	SEGUSE *sup;
1419 	struct buf *bp;
1420 	int curseg, isdirty, sn;
1421 
1422 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1423 #ifdef DEBUG_SU_NBYTES
1424 	printf("lfs_newseg: seg %d := 0 in newseg\n",   /* XXXDEBUG */
1425 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1426 #endif
1427 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1428 	sup->su_nbytes = 0;
1429 	sup->su_nsums = 0;
1430 	sup->su_ninos = 0;
1431 	(void) LFS_BWRITE_LOG(bp); /* Ifile */
1432 
1433 	LFS_CLEANERINFO(cip, fs, bp);
1434 	--cip->clean;
1435 	++cip->dirty;
1436 	fs->lfs_nclean = cip->clean;
1437 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1438 
1439 	fs->lfs_lastseg = fs->lfs_curseg;
1440 	fs->lfs_curseg = fs->lfs_nextseg;
1441 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1442 		sn = (sn + 1) % fs->lfs_nseg;
1443 		if (sn == curseg)
1444 			panic("lfs_nextseg: no clean segments");
1445 		LFS_SEGENTRY(sup, fs, sn, bp);
1446 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1447 		brelse(bp);
1448 		if (!isdirty)
1449 			break;
1450 	}
1451 
1452 	++fs->lfs_nactive;
1453 	fs->lfs_nextseg = sntod(fs, sn);
1454 	if (lfs_dostats) {
1455 		++lfs_stats.segsused;
1456 	}
1457 }
1458 
1459 static struct buf **
1460 lookahead_pagemove(struct buf **bpp, int nblocks, size_t *size)
1461 {
1462 	size_t maxsize;
1463 #ifndef LFS_NO_PAGEMOVE
1464 	struct buf *bp;
1465 #endif
1466 
1467 	maxsize = *size;
1468 	*size = 0;
1469 #ifdef LFS_NO_PAGEMOVE
1470 	return bpp;
1471 #else
1472 	while((bp = *bpp) != NULL && *size < maxsize && nblocks--) {
1473 		if(bp->b_flags & B_CALL)
1474 			return bpp;
1475 		if(bp->b_bcount % NBPG)
1476 			return bpp;
1477 		*size += bp->b_bcount;
1478 		++bpp;
1479 	}
1480 	return NULL;
1481 #endif
1482 }
1483 
1484 #define BQUEUES 4 /* XXX */
1485 #define BQ_EMPTY 3 /* XXX */
1486 extern TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
1487 
1488 #define	BUFHASH(dvp, lbn)	\
1489 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
1490 extern LIST_HEAD(bufhashhdr, buf) invalhash;
1491 /*
1492  * Insq/Remq for the buffer hash lists.
1493  */
1494 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
1495 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
1496 
1497 static struct buf *
1498 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1499 {
1500 	struct lfs_cluster *cl;
1501 	struct buf **bpp, *bp;
1502 	int s;
1503 
1504 	cl = (struct lfs_cluster *)malloc(sizeof(*cl), M_SEGMENT, M_WAITOK);
1505 	bpp = (struct buf **)malloc(n*sizeof(*bpp), M_SEGMENT, M_WAITOK);
1506 	memset(cl, 0, sizeof(*cl));
1507 	cl->fs = fs;
1508 	cl->bpp = bpp;
1509 	cl->bufcount = 0;
1510 	cl->bufsize = 0;
1511 
1512 	/* If this segment is being written synchronously, note that */
1513 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1514 		cl->flags |= LFS_CL_SYNC;
1515 		cl->seg = fs->lfs_sp;
1516 		++cl->seg->seg_iocount;
1517 		/* printf("+ %x => %d\n", cl->seg, cl->seg->seg_iocount); */
1518 	}
1519 
1520 	/* Get an empty buffer header, or maybe one with something on it */
1521 	s = splbio();
1522 	if((bp = bufqueues[BQ_EMPTY].tqh_first) != NULL) {
1523 		bremfree(bp);
1524 		/* clear out various other fields */
1525 		bp->b_flags = B_BUSY;
1526 		bp->b_dev = NODEV;
1527 		bp->b_blkno = bp->b_lblkno = 0;
1528 		bp->b_error = 0;
1529 		bp->b_resid = 0;
1530 		bp->b_bcount = 0;
1531 
1532 		/* nuke any credentials we were holding */
1533 		/* XXXXXX */
1534 
1535 		bremhash(bp);
1536 
1537 		/* disassociate us from our vnode, if we had one... */
1538 		if (bp->b_vp)
1539 			brelvp(bp);
1540 	}
1541 	splx(s);
1542 	while (!bp)
1543 		bp = getnewbuf(0, 0);
1544 	s = splbio();
1545 	bgetvp(vp, bp);
1546 	binshash(bp,&invalhash);
1547 	splx(s);
1548 	bp->b_bcount = 0;
1549 	bp->b_blkno = bp->b_lblkno = addr;
1550 
1551 	bp->b_flags |= B_CALL;
1552 	bp->b_iodone = lfs_cluster_callback;
1553 	cl->saveaddr = bp->b_saveaddr; /* XXX is this ever used? */
1554 	bp->b_saveaddr = (caddr_t)cl;
1555 
1556 	return bp;
1557 }
1558 
1559 int
1560 lfs_writeseg(struct lfs *fs, struct segment *sp)
1561 {
1562 	struct buf **bpp, *bp, *cbp, *newbp, **pmlastbpp;
1563 	SEGUSE *sup;
1564 	SEGSUM *ssp;
1565 	dev_t i_dev;
1566 	char *datap, *dp;
1567 	int do_again, i, nblocks, s;
1568 	size_t el_size;
1569  	struct lfs_cluster *cl;
1570 	int (*strategy)(void *);
1571 	struct vop_strategy_args vop_strategy_a;
1572 	u_short ninos;
1573 	struct vnode *devvp;
1574 	char *p;
1575 	struct vnode *vp;
1576 	struct inode *ip;
1577 	size_t pmsize;
1578 	int use_pagemove;
1579 	daddr_t pseg_daddr;
1580 	daddr_t *daddrp;
1581 	int changed;
1582 #if defined(DEBUG) && defined(LFS_PROPELLER)
1583 	static int propeller;
1584 	char propstring[4] = "-\\|/";
1585 
1586 	printf("%c\b",propstring[propeller++]);
1587 	if (propeller == 4)
1588 		propeller = 0;
1589 #endif
1590 	pseg_daddr = (*(sp->bpp))->b_blkno;
1591 
1592 	/*
1593 	 * If there are no buffers other than the segment summary to write
1594 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1595 	 * even if there aren't any buffers, you need to write the superblock.
1596 	 */
1597 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1598 		return (0);
1599 
1600 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1601 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1602 
1603 	/* Update the segment usage information. */
1604 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1605 
1606 	/* Loop through all blocks, except the segment summary. */
1607 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1608 		if ((*bpp)->b_vp != devvp) {
1609 			sup->su_nbytes += (*bpp)->b_bcount;
1610 #ifdef DEBUG_SU_NBYTES
1611 		printf("seg %d += %ld for ino %d lbn %d db 0x%x\n",
1612 		       sp->seg_number, (*bpp)->b_bcount,
1613 		       VTOI((*bpp)->b_vp)->i_number,
1614 		       (*bpp)->b_lblkno, (*bpp)->b_blkno);
1615 #endif
1616 		}
1617 	}
1618 
1619 	ssp = (SEGSUM *)sp->segsum;
1620 
1621 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1622 #ifdef DEBUG_SU_NBYTES
1623 	printf("seg %d += %d for %d inodes\n",   /* XXXDEBUG */
1624 	       sp->seg_number, ssp->ss_ninos * DINODE_SIZE,
1625 	       ssp->ss_ninos);
1626 #endif
1627 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1628 	/* sup->su_nbytes += fs->lfs_sumsize; */
1629 	if (fs->lfs_version == 1)
1630 		sup->su_olastmod = time.tv_sec;
1631 	else
1632 		sup->su_lastmod = time.tv_sec;
1633 	sup->su_ninos += ninos;
1634 	++sup->su_nsums;
1635 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1636 							 fs->lfs_ibsize));
1637 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1638 
1639 	do_again = !(bp->b_flags & B_GATHERED);
1640 	(void)LFS_BWRITE_LOG(bp); /* Ifile */
1641 	/*
1642 	 * Mark blocks B_BUSY, to prevent then from being changed between
1643 	 * the checksum computation and the actual write.
1644 	 *
1645 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1646 	 * there are any, replace them with copies that have UNASSIGNED
1647 	 * instead.
1648 	 */
1649 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1650 		++bpp;
1651 		if ((*bpp)->b_flags & B_CALL)
1652 			continue;
1653 		bp = *bpp;
1654 	    again:
1655 		s = splbio();
1656 		if (bp->b_flags & B_BUSY) {
1657 #ifdef DEBUG
1658 			printf("lfs_writeseg: avoiding potential data "
1659 			       "summary corruption for ino %d, lbn %d\n",
1660 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1661 #endif
1662 			bp->b_flags |= B_WANTED;
1663 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1664 			splx(s);
1665 			goto again;
1666 		}
1667 		bp->b_flags |= B_BUSY;
1668 		splx(s);
1669 		/* Check and replace indirect block UNWRITTEN bogosity */
1670 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1671 		   VTOI(bp->b_vp)->i_ffs_blocks !=
1672 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1673 #ifdef DEBUG_LFS
1674 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1675 			       VTOI(bp->b_vp)->i_number,
1676 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1677 			       VTOI(bp->b_vp)->i_ffs_blocks);
1678 #endif
1679 			/* Make a copy we'll make changes to */
1680 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1681 					   bp->b_bcount);
1682 			newbp->b_blkno = bp->b_blkno;
1683 			memcpy(newbp->b_data, bp->b_data,
1684 			       newbp->b_bcount);
1685 			*bpp = newbp;
1686 
1687 			changed = 0;
1688 			for (daddrp = (daddr_t *)(newbp->b_data);
1689 			     daddrp < (daddr_t *)(newbp->b_data +
1690 						  newbp->b_bcount); daddrp++) {
1691 				if (*daddrp == UNWRITTEN) {
1692 					++changed;
1693 #ifdef DEBUG_LFS
1694 					printf("lfs_writeseg: replacing UNWRITTEN\n");
1695 #endif
1696 					*daddrp = 0;
1697 				}
1698 			}
1699 			/*
1700 			 * Get rid of the old buffer.  Don't mark it clean,
1701 			 * though, if it still has dirty data on it.
1702 			 */
1703 			if (changed) {
1704 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1705 				if (bp->b_flags & B_CALL) {
1706 					lfs_freebuf(bp);
1707 					bp = NULL;
1708 				} else {
1709 					/* Still on free list, leave it there */
1710 					s = splbio();
1711 					bp->b_flags &= ~B_BUSY;
1712 					if (bp->b_flags & B_WANTED)
1713 						wakeup(bp);
1714 				 	splx(s);
1715 					/*
1716 					 * We have to re-decrement lfs_avail
1717 					 * since this block is going to come
1718 					 * back around to us in the next
1719 					 * segment.
1720 					 */
1721 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1722 				}
1723 			} else {
1724 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1725 						 B_GATHERED);
1726 				if (bp->b_flags & B_CALL) {
1727 					lfs_freebuf(bp);
1728 					bp = NULL;
1729 				} else {
1730 					bremfree(bp);
1731 					bp->b_flags |= B_DONE;
1732 					s = splbio();
1733 					reassignbuf(bp, bp->b_vp);
1734 					splx(s);
1735 					LFS_UNLOCK_BUF(bp);
1736 					brelse(bp);
1737 				}
1738 			}
1739 
1740 		}
1741 	}
1742 	/*
1743 	 * Compute checksum across data and then across summary; the first
1744 	 * block (the summary block) is skipped.  Set the create time here
1745 	 * so that it's guaranteed to be later than the inode mod times.
1746 	 *
1747 	 * XXX
1748 	 * Fix this to do it inline, instead of malloc/copy.
1749 	 */
1750 	if (fs->lfs_version == 1)
1751 		el_size = sizeof(u_long);
1752 	else
1753 		el_size = sizeof(u_int32_t);
1754 	datap = dp = malloc(nblocks * el_size, M_SEGMENT, M_WAITOK);
1755 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1756 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1757 			if (copyin((*bpp)->b_saveaddr, dp, el_size))
1758 				panic("lfs_writeseg: copyin failed [1]: "
1759 				      "ino %d blk %d",
1760 				      VTOI((*bpp)->b_vp)->i_number,
1761 				      (*bpp)->b_lblkno);
1762 		} else
1763 			memcpy(dp, (*bpp)->b_data, el_size);
1764 		dp += el_size;
1765 	}
1766 	if (fs->lfs_version == 1)
1767 		ssp->ss_ocreate = time.tv_sec;
1768 	else {
1769 		ssp->ss_create = time.tv_sec;
1770 		ssp->ss_serial = ++fs->lfs_serial;
1771 		ssp->ss_ident  = fs->lfs_ident;
1772 	}
1773 #ifndef LFS_MALLOC_SUMMARY
1774 	/* Set the summary block busy too */
1775 	(*(sp->bpp))->b_flags |= B_BUSY;
1776 #endif
1777 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * el_size);
1778 	ssp->ss_sumsum =
1779 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1780 	free(datap, M_SEGMENT);
1781 	datap = dp = NULL;
1782 #ifdef DIAGNOSTIC
1783 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1784 		panic("lfs_writeseg: No diskspace for summary");
1785 #endif
1786 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1787 			  btofsb(fs, fs->lfs_sumsize));
1788 
1789 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1790 
1791 	/*
1792   	 * When we simply write the blocks we lose a rotation for every block
1793 	 * written.  To avoid this problem, we use pagemove to cluster
1794 	 * the buffers into a chunk and write the chunk.  CHUNKSIZE is the
1795   	 * largest size I/O devices can handle.
1796   	 *
1797 	 * XXX - right now MAXPHYS is only 64k; could it be larger?
1798 	 */
1799 
1800 #define CHUNKSIZE MAXPHYS
1801 
1802 	if (devvp == NULL)
1803 		panic("devvp is NULL");
1804 	for (bpp = sp->bpp, i = nblocks; i;) {
1805 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1806 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1807 
1808 		cbp->b_dev = i_dev;
1809 		cbp->b_flags |= B_ASYNC | B_BUSY;
1810 		cbp->b_bcount = 0;
1811 
1812 		/*
1813 		 * Find out if we can use pagemove to build the cluster,
1814 		 * or if we are stuck using malloc/copy.  If this is the
1815 		 * first cluster, set the shift flag (see below).
1816 		 */
1817 		pmsize = CHUNKSIZE;
1818 		use_pagemove = 0;
1819 		if(bpp == sp->bpp) {
1820 			/* Summary blocks have to get special treatment */
1821 			pmlastbpp = lookahead_pagemove(bpp + 1, i - 1, &pmsize);
1822 			if(pmsize >= CHUNKSIZE - fs->lfs_sumsize ||
1823 			   pmlastbpp == NULL) {
1824 				use_pagemove = 1;
1825 				cl->flags |= LFS_CL_SHIFT;
1826 			} else {
1827 				/*
1828 				 * If we're not using pagemove, we have
1829 				 * to copy the summary down to the bottom
1830 				 * end of the block.
1831 				 */
1832 #ifndef LFS_MALLOC_SUMMARY
1833 				memcpy((*bpp)->b_data, (*bpp)->b_data +
1834 				       NBPG - fs->lfs_sumsize,
1835 				       fs->lfs_sumsize);
1836 #endif /* LFS_MALLOC_SUMMARY */
1837 			}
1838 		} else {
1839 			pmlastbpp = lookahead_pagemove(bpp, i, &pmsize);
1840 			if(pmsize >= CHUNKSIZE || pmlastbpp == NULL) {
1841 				use_pagemove = 1;
1842 			}
1843 		}
1844 		if(use_pagemove == 0) {
1845 			cl->flags |= LFS_CL_MALLOC;
1846 			cl->olddata = cbp->b_data;
1847 			cbp->b_data = malloc(CHUNKSIZE, M_SEGMENT, M_WAITOK);
1848 		}
1849 #if defined(DEBUG) && defined(DIAGNOSTIC)
1850 		if(dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno + btodb((*bpp)->b_bcount - 1))) !=
1851 		   dtosn(fs, dbtofsb(fs, cbp->b_blkno))) {
1852 			printf("block at %x (%d), cbp at %x (%d)\n",
1853 				(*bpp)->b_blkno, dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno)),
1854 			       cbp->b_blkno, dtosn(fs, dbtofsb(fs, cbp->b_blkno)));
1855 			panic("lfs_writeseg: Segment overwrite");
1856 		}
1857 #endif
1858 
1859 		/*
1860 		 * Construct the cluster.
1861 		 */
1862 		while (fs->lfs_iocount >= LFS_THROTTLE) {
1863 #ifdef DEBUG_LFS
1864 			printf("[%d]", fs->lfs_iocount);
1865 #endif
1866 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs_throttle", 0);
1867 		}
1868 		++fs->lfs_iocount;
1869 
1870 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1871 			bp = *bpp;
1872 
1873 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1874 				break;
1875 
1876 			/*
1877 			 * Fake buffers from the cleaner are marked as B_INVAL.
1878 			 * We need to copy the data from user space rather than
1879 			 * from the buffer indicated.
1880 			 * XXX == what do I do on an error?
1881 			 */
1882 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1883 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1884 					panic("lfs_writeseg: copyin failed [2]");
1885 			} else if (use_pagemove) {
1886 				pagemove(bp->b_data, p, bp->b_bcount);
1887 				cbp->b_bufsize += bp->b_bcount;
1888 				bp->b_bufsize -= bp->b_bcount;
1889   			} else {
1890 				bcopy(bp->b_data, p, bp->b_bcount);
1891 				/* printf("copy in %p\n", bp->b_data); */
1892   			}
1893 
1894 			/*
1895 			 * XXX If we are *not* shifting, the summary
1896 			 * block is only fs->lfs_sumsize.  Otherwise,
1897 			 * it is NBPG but shifted.
1898 			 */
1899 			if(bpp == sp->bpp && !(cl->flags & LFS_CL_SHIFT)) {
1900 				p += fs->lfs_sumsize;
1901 				cbp->b_bcount += fs->lfs_sumsize;
1902 				cl->bufsize += fs->lfs_sumsize;
1903 			} else {
1904 				p += bp->b_bcount;
1905 				cbp->b_bcount += bp->b_bcount;
1906 				cl->bufsize += bp->b_bcount;
1907 			}
1908 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
1909 			cl->bpp[cl->bufcount++] = bp;
1910 			vp = bp->b_vp;
1911 			s = splbio();
1912 			++vp->v_numoutput;
1913 			splx(s);
1914 
1915 			/*
1916 			 * Although it cannot be freed for reuse before the
1917 			 * cluster is written to disk, this buffer does not
1918 			 * need to be held busy.  Therefore we unbusy it,
1919 			 * while leaving it on the locked list.  It will
1920 			 * be freed or requeued by the callback depending
1921 			 * on whether it has had B_DELWRI set again in the
1922 			 * meantime.
1923 			 *
1924 			 * If we are using pagemove, we have to hold the block
1925 			 * busy to prevent its contents from changing before
1926 			 * it hits the disk, and invalidating the checksum.
1927 			 */
1928 			bp->b_flags &= ~(B_DELWRI | B_READ | B_ERROR);
1929 #ifdef LFS_MNOBUSY
1930 			if (cl->flags & LFS_CL_MALLOC) {
1931 				if (!(bp->b_flags & B_CALL))
1932 					brelse(bp); /* Still B_LOCKED */
1933 			}
1934 #endif
1935 			bpp++;
1936 
1937 			/*
1938 			 * If this is the last block for this vnode, but
1939 			 * there are other blocks on its dirty list,
1940 			 * set IN_MODIFIED/IN_CLEANING depending on what
1941 			 * sort of block.  Only do this for our mount point,
1942 			 * not for, e.g., inode blocks that are attached to
1943 			 * the devvp.
1944 			 * XXX KS - Shouldn't we set *both* if both types
1945 			 * of blocks are present (traverse the dirty list?)
1946 			 */
1947 			s = splbio();
1948 			if ((i == 1 ||
1949 			     (i > 1 && vp && *bpp && (*bpp)->b_vp != vp)) &&
1950 			    (bp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL &&
1951 			    vp->v_mount == fs->lfs_ivnode->v_mount)
1952   			{
1953 				ip = VTOI(vp);
1954 #ifdef DEBUG_LFS
1955 				printf("lfs_writeseg: marking ino %d\n",
1956 				       ip->i_number);
1957 #endif
1958 				if (bp->b_flags & B_CALL)
1959 					LFS_SET_UINO(ip, IN_CLEANING);
1960 				else
1961 					LFS_SET_UINO(ip, IN_MODIFIED);
1962 			}
1963 			splx(s);
1964 			wakeup(vp);
1965 		}
1966 		s = splbio();
1967 		++cbp->b_vp->v_numoutput;
1968 		splx(s);
1969 		/*
1970 		 * In order to include the summary in a clustered block,
1971 		 * it may be necessary to shift the block forward (since
1972 		 * summary blocks are in generay smaller than can be
1973 		 * addressed by pagemove().  After the write, the block
1974 		 * will be corrected before disassembly.
1975 		 */
1976 		if(cl->flags & LFS_CL_SHIFT) {
1977 			cbp->b_data += (NBPG - fs->lfs_sumsize);
1978 			cbp->b_bcount -= (NBPG - fs->lfs_sumsize);
1979 		}
1980 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1981 		vop_strategy_a.a_bp = cbp;
1982 		(strategy)(&vop_strategy_a);
1983 	}
1984 
1985 	if (lfs_dostats) {
1986 		++lfs_stats.psegwrites;
1987 		lfs_stats.blocktot += nblocks - 1;
1988 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1989 			++lfs_stats.psyncwrites;
1990 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1991 			++lfs_stats.pcleanwrites;
1992 			lfs_stats.cleanblocks += nblocks - 1;
1993 		}
1994 	}
1995 	return (lfs_initseg(fs) || do_again);
1996 }
1997 
1998 void
1999 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2000 {
2001 	struct buf *bp;
2002 	dev_t i_dev;
2003 	int (*strategy)(void *);
2004 	int s;
2005 	struct vop_strategy_args vop_strategy_a;
2006 
2007 	/*
2008 	 * If we can write one superblock while another is in
2009 	 * progress, we risk not having a complete checkpoint if we crash.
2010 	 * So, block here if a superblock write is in progress.
2011 	 */
2012 	s = splbio();
2013 	while (fs->lfs_sbactive) {
2014 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
2015 	}
2016 	fs->lfs_sbactive = daddr;
2017 	splx(s);
2018 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
2019 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
2020 
2021 	/* Set timestamp of this version of the superblock */
2022 	if (fs->lfs_version == 1)
2023 		fs->lfs_otstamp = time.tv_sec;
2024 	fs->lfs_tstamp = time.tv_sec;
2025 
2026 	/* Checksum the superblock and copy it into a buffer. */
2027 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2028 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD);
2029 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2030 
2031 	bp->b_dev = i_dev;
2032 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2033 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2034 	bp->b_iodone = lfs_supercallback;
2035 	/* XXX KS - same nasty hack as above */
2036 	bp->b_saveaddr = (caddr_t)fs;
2037 
2038 	vop_strategy_a.a_desc = VDESC(vop_strategy);
2039 	vop_strategy_a.a_bp = bp;
2040 	s = splbio();
2041 	++bp->b_vp->v_numoutput;
2042 	splx(s);
2043 	++fs->lfs_iocount;
2044 	(strategy)(&vop_strategy_a);
2045 }
2046 
2047 /*
2048  * Logical block number match routines used when traversing the dirty block
2049  * chain.
2050  */
2051 int
2052 lfs_match_fake(struct lfs *fs, struct buf *bp)
2053 {
2054 	return (bp->b_flags & B_CALL);
2055 }
2056 
2057 int
2058 lfs_match_data(struct lfs *fs, struct buf *bp)
2059 {
2060 	return (bp->b_lblkno >= 0);
2061 }
2062 
2063 int
2064 lfs_match_indir(struct lfs *fs, struct buf *bp)
2065 {
2066 	int lbn;
2067 
2068 	lbn = bp->b_lblkno;
2069 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2070 }
2071 
2072 int
2073 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2074 {
2075 	int lbn;
2076 
2077 	lbn = bp->b_lblkno;
2078 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2079 }
2080 
2081 int
2082 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2083 {
2084 	int lbn;
2085 
2086 	lbn = bp->b_lblkno;
2087 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2088 }
2089 
2090 /*
2091  * XXX - The only buffers that are going to hit these functions are the
2092  * segment write blocks, or the segment summaries, or the superblocks.
2093  *
2094  * All of the above are created by lfs_newbuf, and so do not need to be
2095  * released via brelse.
2096  */
2097 void
2098 lfs_callback(struct buf *bp)
2099 {
2100 	/* struct lfs *fs; */
2101 	/* fs = (struct lfs *)bp->b_saveaddr; */
2102 	lfs_freebuf(bp);
2103 }
2104 
2105 static void
2106 lfs_super_aiodone(struct buf *bp)
2107 {
2108 	struct lfs *fs;
2109 
2110 	fs = (struct lfs *)bp->b_saveaddr;
2111 	fs->lfs_sbactive = 0;
2112 	wakeup(&fs->lfs_sbactive);
2113 	if (--fs->lfs_iocount < LFS_THROTTLE)
2114 		wakeup(&fs->lfs_iocount);
2115 	lfs_freebuf(bp);
2116 }
2117 
2118 static void
2119 lfs_cluster_aiodone(struct buf *bp)
2120 {
2121 	struct lfs_cluster *cl;
2122 	struct lfs *fs;
2123 	struct buf *tbp;
2124 	struct vnode *vp;
2125 	int s, error=0;
2126 	char *cp;
2127 	extern int locked_queue_count;
2128 	extern long locked_queue_bytes;
2129 
2130 	if(bp->b_flags & B_ERROR)
2131 		error = bp->b_error;
2132 
2133 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2134 	fs = cl->fs;
2135 	bp->b_saveaddr = cl->saveaddr;
2136 
2137 	/* If shifted, shift back now */
2138 	if(cl->flags & LFS_CL_SHIFT) {
2139 		bp->b_data -= (NBPG - fs->lfs_sumsize);
2140 		bp->b_bcount += (NBPG - fs->lfs_sumsize);
2141 	}
2142 
2143 	cp = (char *)bp->b_data + cl->bufsize;
2144 	/* Put the pages back, and release the buffer */
2145 	while(cl->bufcount--) {
2146 		tbp = cl->bpp[cl->bufcount];
2147 		if(!(cl->flags & LFS_CL_MALLOC)) {
2148 			cp -= tbp->b_bcount;
2149 			printf("pm(%p,%p,%lx)",cp,tbp->b_data,tbp->b_bcount);
2150 			pagemove(cp, tbp->b_data, tbp->b_bcount);
2151 			bp->b_bufsize -= tbp->b_bcount;
2152 			tbp->b_bufsize += tbp->b_bcount;
2153 		}
2154 		if(error) {
2155 			tbp->b_flags |= B_ERROR;
2156 			tbp->b_error = error;
2157 		}
2158 
2159 		/*
2160 		 * We're done with tbp.  If it has not been re-dirtied since
2161 		 * the cluster was written, free it.  Otherwise, keep it on
2162 		 * the locked list to be written again.
2163 		 */
2164 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2165 			LFS_UNLOCK_BUF(tbp);
2166 		tbp->b_flags &= ~B_GATHERED;
2167 
2168 		LFS_BCLEAN_LOG(fs, tbp);
2169 
2170 		vp = tbp->b_vp;
2171 		/* Segment summary for a shifted cluster */
2172 		if(!cl->bufcount && (cl->flags & LFS_CL_SHIFT))
2173 			tbp->b_flags |= B_INVAL;
2174 		if(!(tbp->b_flags & B_CALL)) {
2175 			bremfree(tbp);
2176 			s = splbio();
2177 			if(vp)
2178 				reassignbuf(tbp, vp);
2179 			splx(s);
2180 			tbp->b_flags |= B_ASYNC; /* for biodone */
2181 		}
2182 #ifdef DIAGNOSTIC
2183 		if (tbp->b_flags & B_DONE) {
2184 			printf("blk %d biodone already (flags %lx)\n",
2185 				cl->bufcount, (long)tbp->b_flags);
2186 		}
2187 #endif
2188 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2189 			biodone(tbp);
2190 		}
2191 	}
2192 
2193 	/* Fix up the cluster buffer, and release it */
2194 	if(!(cl->flags & LFS_CL_MALLOC) && bp->b_bufsize) {
2195 		printf("PM(%p,%p,%lx)", (char *)bp->b_data + bp->b_bcount,
2196 			 (char *)bp->b_data, bp->b_bufsize);
2197 		pagemove((char *)bp->b_data + bp->b_bcount,
2198 			 (char *)bp->b_data, bp->b_bufsize);
2199 	}
2200 	if(cl->flags & LFS_CL_MALLOC) {
2201 		free(bp->b_data, M_SEGMENT);
2202 		bp->b_data = cl->olddata;
2203 	}
2204 	bp->b_bcount = 0;
2205 	bp->b_iodone = NULL;
2206 	bp->b_flags &= ~B_DELWRI;
2207 	bp->b_flags |= B_DONE;
2208 	s = splbio();
2209 	reassignbuf(bp, bp->b_vp);
2210 	splx(s);
2211 	brelse(bp);
2212 
2213 	/* Note i/o done */
2214 	if (cl->flags & LFS_CL_SYNC) {
2215 		if (--cl->seg->seg_iocount == 0)
2216 			wakeup(&cl->seg->seg_iocount);
2217 		/* printf("- %x => %d\n", cl->seg, cl->seg->seg_iocount); */
2218 	}
2219 #ifdef DIAGNOSTIC
2220 	if (fs->lfs_iocount == 0)
2221 		panic("lfs_cluster_aiodone: zero iocount\n");
2222 #endif
2223 	if (--fs->lfs_iocount < LFS_THROTTLE)
2224 		wakeup(&fs->lfs_iocount);
2225 #if 0
2226 	if (fs->lfs_iocount == 0) {
2227 		/*
2228 		 * Vinvalbuf can move locked buffers off the locked queue
2229 		 * and we have no way of knowing about this.  So, after
2230 		 * doing a big write, we recalculate how many buffers are
2231 		 * really still left on the locked queue.
2232 		 */
2233 		lfs_countlocked(&locked_queue_count, &locked_queue_bytes, "lfs_cluster_callback");
2234 		wakeup(&locked_queue_count);
2235 	}
2236 #endif
2237 
2238 	free(cl->bpp, M_SEGMENT);
2239 	free(cl, M_SEGMENT);
2240 }
2241 
2242 static void
2243 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2244 {
2245 	/* reset b_iodone for when this is a single-buf i/o. */
2246 	bp->b_iodone = aiodone;
2247 
2248 	simple_lock(&uvm.aiodoned_lock);        /* locks uvm.aio_done */
2249 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2250 	wakeup(&uvm.aiodoned);
2251 	simple_unlock(&uvm.aiodoned_lock);
2252 }
2253 
2254 static void
2255 lfs_cluster_callback(struct buf *bp)
2256 {
2257 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2258 }
2259 
2260 void
2261 lfs_supercallback(struct buf *bp)
2262 {
2263 	lfs_generic_callback(bp, lfs_super_aiodone);
2264 }
2265 
2266 /*
2267  * Shellsort (diminishing increment sort) from Data Structures and
2268  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2269  * see also Knuth Vol. 3, page 84.  The increments are selected from
2270  * formula (8), page 95.  Roughly O(N^3/2).
2271  */
2272 /*
2273  * This is our own private copy of shellsort because we want to sort
2274  * two parallel arrays (the array of buffer pointers and the array of
2275  * logical block numbers) simultaneously.  Note that we cast the array
2276  * of logical block numbers to a unsigned in this routine so that the
2277  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2278  */
2279 
2280 void
2281 lfs_shellsort(struct buf **bp_array, ufs_daddr_t *lb_array, int nmemb)
2282 {
2283 	static int __rsshell_increments[] = { 4, 1, 0 };
2284 	int incr, *incrp, t1, t2;
2285 	struct buf *bp_temp;
2286 	u_long lb_temp;
2287 
2288 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2289 		for (t1 = incr; t1 < nmemb; ++t1)
2290 			for (t2 = t1 - incr; t2 >= 0;)
2291 				if (lb_array[t2] > lb_array[t2 + incr]) {
2292 					lb_temp = lb_array[t2];
2293 					lb_array[t2] = lb_array[t2 + incr];
2294 					lb_array[t2 + incr] = lb_temp;
2295 					bp_temp = bp_array[t2];
2296 					bp_array[t2] = bp_array[t2 + incr];
2297 					bp_array[t2 + incr] = bp_temp;
2298 					t2 -= incr;
2299 				} else
2300 					break;
2301 }
2302 
2303 /*
2304  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2305  */
2306 int
2307 lfs_vref(struct vnode *vp)
2308 {
2309 	/*
2310 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2311 	 * being able to flush all of the pages from this vnode, which
2312 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2313 	 */
2314 	if (vp->v_flag & VXLOCK) {
2315 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2316 			return 0;
2317 		}
2318 		return (1);
2319 	}
2320 	return (vget(vp, 0));
2321 }
2322 
2323 /*
2324  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2325  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2326  */
2327 void
2328 lfs_vunref(struct vnode *vp)
2329 {
2330 	/*
2331 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2332 	 */
2333 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2334 		return;
2335 	}
2336 
2337 	simple_lock(&vp->v_interlock);
2338 #ifdef DIAGNOSTIC
2339 	if (vp->v_usecount <= 0) {
2340 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2341 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2342 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2343 		panic("lfs_vunref: v_usecount<0");
2344 	}
2345 #endif
2346 	vp->v_usecount--;
2347 	if (vp->v_usecount > 0) {
2348 		simple_unlock(&vp->v_interlock);
2349 		return;
2350 	}
2351 	/*
2352 	 * insert at tail of LRU list
2353 	 */
2354 	simple_lock(&vnode_free_list_slock);
2355 	if (vp->v_holdcnt > 0)
2356 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2357 	else
2358 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2359 	simple_unlock(&vnode_free_list_slock);
2360 	simple_unlock(&vp->v_interlock);
2361 }
2362 
2363 /*
2364  * We use this when we have vnodes that were loaded in solely for cleaning.
2365  * There is no reason to believe that these vnodes will be referenced again
2366  * soon, since the cleaning process is unrelated to normal filesystem
2367  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2368  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2369  * cleaning at the head of the list, instead.
2370  */
2371 void
2372 lfs_vunref_head(struct vnode *vp)
2373 {
2374 	simple_lock(&vp->v_interlock);
2375 #ifdef DIAGNOSTIC
2376 	if (vp->v_usecount == 0) {
2377 		panic("lfs_vunref: v_usecount<0");
2378 	}
2379 #endif
2380 	vp->v_usecount--;
2381 	if (vp->v_usecount > 0) {
2382 		simple_unlock(&vp->v_interlock);
2383 		return;
2384 	}
2385 	/*
2386 	 * insert at head of LRU list
2387 	 */
2388 	simple_lock(&vnode_free_list_slock);
2389 	if (vp->v_holdcnt > 0)
2390 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2391 	else
2392 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2393 	simple_unlock(&vnode_free_list_slock);
2394 	simple_unlock(&vp->v_interlock);
2395 }
2396 
2397