xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision b8c616269f5ebf18ab2e35cb8099d683130a177c)
1 /*	$NetBSD: lfs_segment.c,v 1.99 2003/02/01 06:23:53 thorpej 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 1.99 2003/02/01 06:23:53 thorpej 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/vnode.h>
92 #include <sys/malloc.h>
93 #include <sys/mount.h>
94 
95 #include <miscfs/specfs/specdev.h>
96 #include <miscfs/fifofs/fifo.h>
97 
98 #include <ufs/ufs/inode.h>
99 #include <ufs/ufs/dir.h>
100 #include <ufs/ufs/ufsmount.h>
101 #include <ufs/ufs/ufs_extern.h>
102 
103 #include <ufs/lfs/lfs.h>
104 #include <ufs/lfs/lfs_extern.h>
105 
106 #include <uvm/uvm.h>
107 #include <uvm/uvm_extern.h>
108 
109 MALLOC_DEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
110 
111 extern int count_lock_queue(void);
112 extern struct simplelock vnode_free_list_slock;		/* XXX */
113 
114 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
115 static void lfs_super_aiodone(struct buf *);
116 static void lfs_cluster_aiodone(struct buf *);
117 static void lfs_cluster_callback(struct buf *);
118 static struct buf **lookahead_pagemove(struct buf **, int, size_t *);
119 
120 /*
121  * Determine if it's OK to start a partial in this segment, or if we need
122  * to go on to a new segment.
123  */
124 #define	LFS_PARTIAL_FITS(fs) \
125 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
126 	fragstofsb((fs), (fs)->lfs_frag))
127 
128 void	 lfs_callback(struct buf *);
129 int	 lfs_gather(struct lfs *, struct segment *,
130 	     struct vnode *, int (*)(struct lfs *, struct buf *));
131 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
132 void	 lfs_iset(struct inode *, daddr_t, time_t);
133 int	 lfs_match_fake(struct lfs *, struct buf *);
134 int	 lfs_match_data(struct lfs *, struct buf *);
135 int	 lfs_match_dindir(struct lfs *, struct buf *);
136 int	 lfs_match_indir(struct lfs *, struct buf *);
137 int	 lfs_match_tindir(struct lfs *, struct buf *);
138 void	 lfs_newseg(struct lfs *);
139 /* XXX ondisk32 */
140 void	 lfs_shellsort(struct buf **, int32_t *, int);
141 void	 lfs_supercallback(struct buf *);
142 void	 lfs_updatemeta(struct segment *);
143 int	 lfs_vref(struct vnode *);
144 void	 lfs_vunref(struct vnode *);
145 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
146 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
147 int	 lfs_writeseg(struct lfs *, struct segment *);
148 void	 lfs_writesuper(struct lfs *, daddr_t);
149 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
150 	    struct segment *sp, int dirops);
151 
152 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
153 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
154 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
155 int	lfs_dirvcount = 0;		/* # active dirops */
156 
157 /* Statistics Counters */
158 int lfs_dostats = 1;
159 struct lfs_stats lfs_stats;
160 
161 extern int locked_queue_count;
162 extern long locked_queue_bytes;
163 
164 /* op values to lfs_writevnodes */
165 #define	VN_REG	        0
166 #define	VN_DIROP	1
167 #define	VN_EMPTY	2
168 #define VN_CLEAN        3
169 
170 #define LFS_MAX_ACTIVE          10
171 
172 /*
173  * XXX KS - Set modification time on the Ifile, so the cleaner can
174  * read the fs mod time off of it.  We don't set IN_UPDATE here,
175  * since we don't really need this to be flushed to disk (and in any
176  * case that wouldn't happen to the Ifile until we checkpoint).
177  */
178 void
179 lfs_imtime(struct lfs *fs)
180 {
181 	struct timespec ts;
182 	struct inode *ip;
183 
184 	TIMEVAL_TO_TIMESPEC(&time, &ts);
185 	ip = VTOI(fs->lfs_ivnode);
186 	ip->i_ffs_mtime = ts.tv_sec;
187 	ip->i_ffs_mtimensec = ts.tv_nsec;
188 }
189 
190 /*
191  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
192  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
193  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
194  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
195  */
196 
197 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
198 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
199 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
200 
201 int
202 lfs_vflush(struct vnode *vp)
203 {
204 	struct inode *ip;
205 	struct lfs *fs;
206 	struct segment *sp;
207 	struct buf *bp, *nbp, *tbp, *tnbp;
208 	int error, s;
209 
210 	ip = VTOI(vp);
211 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
212 
213 	if (ip->i_flag & IN_CLEANING) {
214 #ifdef DEBUG_LFS
215 		ivndebug(vp,"vflush/in_cleaning");
216 #endif
217 		LFS_CLR_UINO(ip, IN_CLEANING);
218 		LFS_SET_UINO(ip, IN_MODIFIED);
219 
220 		/*
221 		 * Toss any cleaning buffers that have real counterparts
222 		 * to avoid losing new data
223 		 */
224 		s = splbio();
225 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
226 			nbp = LIST_NEXT(bp, b_vnbufs);
227 			if (bp->b_flags & B_CALL) {
228 				for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
229 				    tbp = tnbp)
230 				{
231 					tnbp = LIST_NEXT(tbp, b_vnbufs);
232 					if (tbp->b_vp == bp->b_vp
233 					   && tbp->b_lblkno == bp->b_lblkno
234 					   && tbp != bp)
235 					{
236 						fs->lfs_avail += btofsb(fs, bp->b_bcount);
237 						wakeup(&fs->lfs_avail);
238 						lfs_freebuf(bp);
239 						bp = NULL;
240 						break;
241 					}
242 				}
243 			}
244 		}
245 		splx(s);
246 	}
247 
248 	/* If the node is being written, wait until that is done */
249 	s = splbio();
250 	if (WRITEINPROG(vp)) {
251 #ifdef DEBUG_LFS
252 		ivndebug(vp,"vflush/writeinprog");
253 #endif
254 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
255 	}
256 	splx(s);
257 
258 	/* Protect against VXLOCK deadlock in vinvalbuf() */
259 	lfs_seglock(fs, SEGM_SYNC);
260 
261 	/* If we're supposed to flush a freed inode, just toss it */
262 	/* XXX - seglock, so these buffers can't be gathered, right? */
263 	if (ip->i_ffs_mode == 0) {
264 		printf("lfs_vflush: ino %d is freed, not flushing\n",
265 			ip->i_number);
266 		s = splbio();
267 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
268 			nbp = LIST_NEXT(bp, b_vnbufs);
269 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
270 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
271 				wakeup(&fs->lfs_avail);
272 			}
273 			/* Copied from lfs_writeseg */
274 			if (bp->b_flags & B_CALL) {
275 				/* if B_CALL, it was created with newbuf */
276 				lfs_freebuf(bp);
277 				bp = NULL;
278 			} else {
279 				bremfree(bp);
280 				LFS_UNLOCK_BUF(bp);
281 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
282                                          B_GATHERED);
283 				bp->b_flags |= B_DONE;
284 				reassignbuf(bp, vp);
285 				brelse(bp);
286 			}
287 		}
288 		splx(s);
289 		LFS_CLR_UINO(ip, IN_CLEANING);
290 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
291 		ip->i_flag &= ~IN_ALLMOD;
292 		printf("lfs_vflush: done not flushing ino %d\n",
293 			ip->i_number);
294 		lfs_segunlock(fs);
295 		return 0;
296 	}
297 
298 	SET_FLUSHING(fs,vp);
299 	if (fs->lfs_nactive > LFS_MAX_ACTIVE ||
300 	    (fs->lfs_sp->seg_flags & SEGM_CKP)) {
301 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
302 		CLR_FLUSHING(fs,vp);
303 		lfs_segunlock(fs);
304 		return error;
305 	}
306 	sp = fs->lfs_sp;
307 
308 	if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) {
309 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
310 	} else if ((ip->i_flag & IN_CLEANING) &&
311 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
312 #ifdef DEBUG_LFS
313 		ivndebug(vp,"vflush/clean");
314 #endif
315 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
316 	} else if (lfs_dostats) {
317 		if (LIST_FIRST(&vp->v_dirtyblkhd) || (VTOI(vp)->i_flag & IN_ALLMOD))
318 			++lfs_stats.vflush_invoked;
319 #ifdef DEBUG_LFS
320 		ivndebug(vp,"vflush");
321 #endif
322 	}
323 
324 #ifdef DIAGNOSTIC
325 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
326 	if (vp->v_flag & VDIROP) {
327 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
328 		/* panic("VDIROP being flushed...this can\'t happen"); */
329 	}
330 	if (vp->v_usecount < 0) {
331 		printf("usecount=%ld\n", (long)vp->v_usecount);
332 		panic("lfs_vflush: usecount<0");
333 	}
334 #endif
335 
336 	do {
337 		do {
338 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
339 				lfs_writefile(fs, sp, vp);
340 		} while (lfs_writeinode(fs, sp, ip));
341 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
342 
343 	if (lfs_dostats) {
344 		++lfs_stats.nwrites;
345 		if (sp->seg_flags & SEGM_SYNC)
346 			++lfs_stats.nsync_writes;
347 		if (sp->seg_flags & SEGM_CKP)
348 			++lfs_stats.ncheckpoints;
349 	}
350 	/*
351 	 * If we were called from somewhere that has already held the seglock
352 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
353 	 * the write to complete because we are still locked.
354 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
355 	 * we must explicitly wait, if that is the case.
356 	 *
357 	 * We compare the iocount against 1, not 0, because it is
358 	 * artificially incremented by lfs_seglock().
359 	 */
360 	if (fs->lfs_seglock > 1) {
361 		while (fs->lfs_iocount > 1)
362 			(void)tsleep(&fs->lfs_iocount, PRIBIO + 1,
363 				     "lfs_vflush", 0);
364 	}
365 	lfs_segunlock(fs);
366 
367 	CLR_FLUSHING(fs,vp);
368 	return (0);
369 }
370 
371 #ifdef DEBUG_LFS_VERBOSE
372 # 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)
373 #else
374 # define vndebug(vp,str)
375 #endif
376 
377 int
378 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
379 {
380 	struct inode *ip;
381 	struct vnode *vp, *nvp;
382 	int inodes_written = 0, only_cleaning;
383 
384 #ifndef LFS_NO_BACKVP_HACK
385 	/* BEGIN HACK */
386 #define	VN_OFFSET	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
387 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
388 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&(LIST_FIRST(&mp->mnt_vnodelist))) - VN_OFFSET))
389 
390 	/* Find last vnode. */
391  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
392 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
393 	     vp = LIST_NEXT(vp, v_mntvnodes));
394 	for (; vp && vp != BEG_OF_VLIST; vp = nvp) {
395 		nvp = BACK_VP(vp);
396 #else
397 	loop:
398 	for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) {
399 		nvp = LIST_NEXT(vp, v_mntvnodes);
400 #endif
401 		/*
402 		 * If the vnode that we are about to sync is no longer
403 		 * associated with this mount point, start over.
404 		 */
405 		if (vp->v_mount != mp) {
406 			printf("lfs_writevnodes: starting over\n");
407 			goto loop;
408 		}
409 
410 		if (vp->v_type == VNON) {
411 			continue;
412 		}
413 
414 		ip = VTOI(vp);
415 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
416 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
417 			vndebug(vp,"dirop");
418 			continue;
419 		}
420 
421 		if (op == VN_EMPTY && LIST_FIRST(&vp->v_dirtyblkhd)) {
422 			vndebug(vp,"empty");
423 			continue;
424 		}
425 
426 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
427 		   && vp != fs->lfs_flushvp
428 		   && !(ip->i_flag & IN_CLEANING)) {
429 			vndebug(vp,"cleaning");
430 			continue;
431 		}
432 
433 		if (lfs_vref(vp)) {
434 			vndebug(vp,"vref");
435 			continue;
436 		}
437 
438 		only_cleaning = 0;
439 		/*
440 		 * Write the inode/file if dirty and it's not the IFILE.
441 		 */
442 		if ((ip->i_flag & IN_ALLMOD) ||
443 		     (LIST_FIRST(&vp->v_dirtyblkhd) != NULL))
444 		{
445 			only_cleaning = ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
446 
447 			if (ip->i_number != LFS_IFILE_INUM
448 			   && LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
449 			{
450 				lfs_writefile(fs, sp, vp);
451 			}
452 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
453 				if (WRITEINPROG(vp)) {
454 #ifdef DEBUG_LFS
455 					ivndebug(vp,"writevnodes/write2");
456 #endif
457 				} else if (!(ip->i_flag & IN_ALLMOD)) {
458 #ifdef DEBUG_LFS
459 					printf("<%d>",ip->i_number);
460 #endif
461 					LFS_SET_UINO(ip, IN_MODIFIED);
462 				}
463 			}
464 			(void) lfs_writeinode(fs, sp, ip);
465 			inodes_written++;
466 		}
467 
468 		if (lfs_clean_vnhead && only_cleaning)
469 			lfs_vunref_head(vp);
470 		else
471 			lfs_vunref(vp);
472 	}
473 	return inodes_written;
474 }
475 
476 /*
477  * Do a checkpoint.
478  */
479 int
480 lfs_segwrite(struct mount *mp, int flags)
481 {
482 	struct buf *bp;
483 	struct inode *ip;
484 	struct lfs *fs;
485 	struct segment *sp;
486 	struct vnode *vp;
487 	SEGUSE *segusep;
488 	daddr_t ibno;
489 	int do_ckp, did_ckp, error, i;
490 	int writer_set = 0;
491 	int dirty;
492 	int redo;
493 
494 	fs = VFSTOUFS(mp)->um_lfs;
495 
496 	if (fs->lfs_ronly)
497 		return EROFS;
498 
499 	lfs_imtime(fs);
500 
501 	/* printf("lfs_segwrite: ifile flags are 0x%lx\n",
502 	       (long)(VTOI(fs->lfs_ivnode)->i_flag)); */
503 
504 #if 0
505 	/*
506 	 * If we are not the cleaner, and there is no space available,
507 	 * wait until cleaner writes.
508 	 */
509 	if (!(flags & SEGM_CLEAN) && !(fs->lfs_seglock && fs->lfs_sp &&
510 				      (fs->lfs_sp->seg_flags & SEGM_CLEAN)))
511 	{
512 		while (fs->lfs_avail <= 0) {
513 			LFS_CLEANERINFO(cip, fs, bp);
514 			LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
515 
516 			wakeup(&lfs_allclean_wakeup);
517 			wakeup(&fs->lfs_nextseg);
518 			error = tsleep(&fs->lfs_avail, PRIBIO + 1, "lfs_av2",
519 				       0);
520 			if (error) {
521 				return (error);
522 			}
523 		}
524 	}
525 #endif
526 	/*
527 	 * Allocate a segment structure and enough space to hold pointers to
528 	 * the maximum possible number of buffers which can be described in a
529 	 * single summary block.
530 	 */
531 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
532 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
533 	sp = fs->lfs_sp;
534 
535 	/*
536 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
537 	 * in which case we have to flush *all* buffers off of this vnode.
538 	 * We don't care about other nodes, but write any non-dirop nodes
539 	 * anyway in anticipation of another getnewvnode().
540 	 *
541 	 * If we're cleaning we only write cleaning and ifile blocks, and
542 	 * no dirops, since otherwise we'd risk corruption in a crash.
543 	 */
544 	if (sp->seg_flags & SEGM_CLEAN)
545 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
546 	else {
547 		lfs_writevnodes(fs, mp, sp, VN_REG);
548 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
549 			while (fs->lfs_dirops)
550 				if ((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
551 						"lfs writer", 0)))
552 				{
553 					/* XXX why not segunlock? */
554 					free(sp->bpp, M_SEGMENT);
555 					sp->bpp = NULL;
556 					free(sp, M_SEGMENT);
557 					fs->lfs_sp = NULL;
558 					return (error);
559 				}
560 			fs->lfs_writer++;
561 			writer_set = 1;
562 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
563 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
564 		}
565 	}
566 
567 	/*
568 	 * If we are doing a checkpoint, mark everything since the
569 	 * last checkpoint as no longer ACTIVE.
570 	 */
571 	if (do_ckp) {
572 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
573 		     --ibno >= fs->lfs_cleansz; ) {
574 			dirty = 0;
575 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
576 
577 				panic("lfs_segwrite: ifile read");
578 			segusep = (SEGUSE *)bp->b_data;
579 			for (i = fs->lfs_sepb; i--;) {
580 				if (segusep->su_flags & SEGUSE_ACTIVE) {
581 					segusep->su_flags &= ~SEGUSE_ACTIVE;
582 					++dirty;
583 				}
584 				if (fs->lfs_version > 1)
585 					++segusep;
586 				else
587 					segusep = (SEGUSE *)
588 						((SEGUSE_V1 *)segusep + 1);
589 			}
590 
591 			/* But the current segment is still ACTIVE */
592 			segusep = (SEGUSE *)bp->b_data;
593 			if (dtosn(fs, fs->lfs_curseg) / fs->lfs_sepb ==
594 			    (ibno-fs->lfs_cleansz)) {
595 				if (fs->lfs_version > 1)
596 					segusep[dtosn(fs, fs->lfs_curseg) %
597 					     fs->lfs_sepb].su_flags |=
598 						     SEGUSE_ACTIVE;
599 				else
600 					((SEGUSE *)
601 					 ((SEGUSE_V1 *)(bp->b_data) +
602 					  (dtosn(fs, fs->lfs_curseg) %
603 					   fs->lfs_sepb)))->su_flags
604 						   |= SEGUSE_ACTIVE;
605 				--dirty;
606 			}
607 			if (dirty)
608 				error = LFS_BWRITE_LOG(bp); /* Ifile */
609 			else
610 				brelse(bp);
611 		}
612 	}
613 
614 	did_ckp = 0;
615 	if (do_ckp || fs->lfs_doifile) {
616 		do {
617 			vp = fs->lfs_ivnode;
618 
619 			vget(vp, LK_EXCLUSIVE | LK_CANRECURSE | LK_RETRY);
620 #ifdef DEBUG
621 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
622 #endif
623 			fs->lfs_flags &= ~LFS_IFDIRTY;
624 
625 			ip = VTOI(vp);
626 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
627 				lfs_writefile(fs, sp, vp);
628 			if (ip->i_flag & IN_ALLMOD)
629 				++did_ckp;
630 			redo = lfs_writeinode(fs, sp, ip);
631 
632 			vput(vp);
633 			/*
634 			 * if we know we'll redo, no need to writeseg here.
635 			 */
636 			if (!(redo && do_ckp)) {
637 				redo += lfs_writeseg(fs, sp);
638 			}
639 			redo += (fs->lfs_flags & LFS_IFDIRTY);
640 		} while (redo && do_ckp);
641 
642 		/* The ifile should now be all clear */
643 		if (do_ckp && LIST_FIRST(&vp->v_dirtyblkhd)) {
644 			struct buf *bp;
645 			int s, warned = 0, dopanic = 0;
646 			s = splbio();
647 			for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = LIST_NEXT(bp, b_vnbufs)) {
648 				if (!(bp->b_flags & B_GATHERED)) {
649 					if (!warned)
650 						printf("lfs_segwrite: ifile still has dirty blocks?!\n");
651 					++dopanic;
652 					++warned;
653 					printf("bp=%p, lbn %" PRId64 ", "
654 						"flags 0x%lx\n",
655 						bp, bp->b_lblkno,
656 						bp->b_flags);
657 				}
658 			}
659 			if (dopanic)
660 				panic("dirty blocks");
661 			splx(s);
662 		}
663 		LFS_CLR_UINO(ip, IN_ALLMOD);
664 	} else {
665 		(void) lfs_writeseg(fs, sp);
666 	}
667 
668 	/*
669 	 * If the I/O count is non-zero, sleep until it reaches zero.
670 	 * At the moment, the user's process hangs around so we can
671 	 * sleep.
672 	 */
673 	fs->lfs_doifile = 0;
674 	if (writer_set && --fs->lfs_writer == 0)
675 		wakeup(&fs->lfs_dirops);
676 
677 	/*
678 	 * If we didn't write the Ifile, we didn't really do anything.
679 	 * That means that (1) there is a checkpoint on disk and (2)
680 	 * nothing has changed since it was written.
681 	 *
682 	 * Take the flags off of the segment so that lfs_segunlock
683 	 * doesn't have to write the superblock either.
684 	 */
685 	if (do_ckp && !did_ckp) {
686 		sp->seg_flags &= ~SEGM_CKP;
687 		/* if (do_ckp) printf("lfs_segwrite: no checkpoint\n"); */
688 	}
689 
690 	if (lfs_dostats) {
691 		++lfs_stats.nwrites;
692 		if (sp->seg_flags & SEGM_SYNC)
693 			++lfs_stats.nsync_writes;
694 		if (sp->seg_flags & SEGM_CKP)
695 			++lfs_stats.ncheckpoints;
696 	}
697 	lfs_segunlock(fs);
698 	return (0);
699 }
700 
701 /*
702  * Write the dirty blocks associated with a vnode.
703  */
704 void
705 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
706 {
707 	struct buf *bp;
708 	struct finfo *fip;
709 	struct inode *ip;
710 	IFILE *ifp;
711 	int i, frag;
712 
713 	ip = VTOI(vp);
714 
715 	if (sp->seg_bytes_left < fs->lfs_bsize ||
716 	    sp->sum_bytes_left < sizeof(struct finfo))
717 		(void) lfs_writeseg(fs, sp);
718 
719 	sp->sum_bytes_left -= FINFOSIZE;
720 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
721 
722 	if (vp->v_flag & VDIROP)
723 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
724 
725 	fip = sp->fip;
726 	fip->fi_nblocks = 0;
727 	fip->fi_ino = ip->i_number;
728 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
729 	fip->fi_version = ifp->if_version;
730 	brelse(bp);
731 
732 	if (sp->seg_flags & SEGM_CLEAN) {
733 		lfs_gather(fs, sp, vp, lfs_match_fake);
734 		/*
735 		 * For a file being flushed, we need to write *all* blocks.
736 		 * This means writing the cleaning blocks first, and then
737 		 * immediately following with any non-cleaning blocks.
738 		 * The same is true of the Ifile since checkpoints assume
739 		 * that all valid Ifile blocks are written.
740 		 */
741 	   	if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode)
742 			lfs_gather(fs, sp, vp, lfs_match_data);
743 	} else
744 		lfs_gather(fs, sp, vp, lfs_match_data);
745 
746 	/*
747 	 * It may not be necessary to write the meta-data blocks at this point,
748 	 * as the roll-forward recovery code should be able to reconstruct the
749 	 * list.
750 	 *
751 	 * We have to write them anyway, though, under two conditions: (1) the
752 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
753 	 * checkpointing.
754 	 *
755 	 * BUT if we are cleaning, we might have indirect blocks that refer to
756 	 * new blocks not being written yet, in addition to fragments being
757 	 * moved out of a cleaned segment.  If that is the case, don't
758 	 * write the indirect blocks, or the finfo will have a small block
759 	 * in the middle of it!
760 	 * XXX in this case isn't the inode size wrong too?
761 	 */
762 	frag = 0;
763 	if (sp->seg_flags & SEGM_CLEAN) {
764 		for (i = 0; i < NDADDR; i++)
765 			if (ip->i_lfs_fragsize[i] > 0 &&
766 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
767 				++frag;
768 	}
769 #ifdef DIAGNOSTIC
770 	if (frag > 1)
771 		panic("lfs_writefile: more than one fragment!");
772 #endif
773 	if (IS_FLUSHING(fs, vp) ||
774 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
775 		lfs_gather(fs, sp, vp, lfs_match_indir);
776 		lfs_gather(fs, sp, vp, lfs_match_dindir);
777 		lfs_gather(fs, sp, vp, lfs_match_tindir);
778 	}
779 	fip = sp->fip;
780 	if (fip->fi_nblocks != 0) {
781 		sp->fip = (FINFO*)((caddr_t)fip + FINFOSIZE +
782 				   sizeof(int32_t) * (fip->fi_nblocks));
783 		sp->start_lbp = &sp->fip->fi_blocks[0];
784 	} else {
785 		sp->sum_bytes_left += FINFOSIZE;
786 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
787 	}
788 }
789 
790 int
791 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
792 {
793 	struct buf *bp, *ibp;
794 	struct dinode *cdp;
795 	IFILE *ifp;
796 	SEGUSE *sup;
797 	daddr_t daddr;
798 	int32_t *daddrp;	/* XXX ondisk32 */
799 	ino_t ino;
800 	int error, i, ndx, fsb = 0;
801 	int redo_ifile = 0;
802 	struct timespec ts;
803 	int gotblk = 0;
804 
805 	if (!(ip->i_flag & IN_ALLMOD))
806 		return (0);
807 
808 	/* Allocate a new inode block if necessary. */
809 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
810 		/* Allocate a new segment if necessary. */
811 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
812 		    sp->sum_bytes_left < sizeof(int32_t))
813 			(void) lfs_writeseg(fs, sp);
814 
815 		/* Get next inode block. */
816 		daddr = fs->lfs_offset;
817 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
818 		sp->ibp = *sp->cbpp++ =
819 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
820 			       fs->lfs_ibsize, 0, 0);
821 		gotblk++;
822 
823 		/* Zero out inode numbers */
824 		for (i = 0; i < INOPB(fs); ++i)
825 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
826 
827 		++sp->start_bpp;
828 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
829 		/* Set remaining space counters. */
830 		sp->seg_bytes_left -= fs->lfs_ibsize;
831 		sp->sum_bytes_left -= sizeof(int32_t);
832 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
833 			sp->ninodes / INOPB(fs) - 1;
834 		((int32_t *)(sp->segsum))[ndx] = daddr;
835 	}
836 
837 	/* Update the inode times and copy the inode onto the inode page. */
838 	TIMEVAL_TO_TIMESPEC(&time, &ts);
839 	/* XXX kludge --- don't redirty the ifile just to put times on it */
840 	if (ip->i_number != LFS_IFILE_INUM)
841 		LFS_ITIMES(ip, &ts, &ts, &ts);
842 
843 	/*
844 	 * If this is the Ifile, and we've already written the Ifile in this
845 	 * partial segment, just overwrite it (it's not on disk yet) and
846 	 * continue.
847 	 *
848 	 * XXX we know that the bp that we get the second time around has
849 	 * already been gathered.
850 	 */
851 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
852 		*(sp->idp) = ip->i_din.ffs_din;
853 		ip->i_lfs_osize = ip->i_ffs_size;
854 		return 0;
855 	}
856 
857 	bp = sp->ibp;
858 	cdp = ((struct dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
859 	*cdp = ip->i_din.ffs_din;
860 #ifdef LFS_IFILE_FRAG_ADDRESSING
861 	if (fs->lfs_version > 1)
862 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
863 #endif
864 
865 	/*
866 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
867 	 * addresses to disk; possibly revert the inode size.
868 	 */
869 	if (ip->i_lfs_effnblks != ip->i_ffs_blocks) {
870 		cdp->di_size = ip->i_lfs_osize;
871 #ifdef DEBUG_LFS
872 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
873 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs_blocks);
874 #endif
875 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
876 		     daddrp++) {
877 			if (*daddrp == UNWRITTEN) {
878 #ifdef DEBUG_LFS
879 				printf("lfs_writeinode: wiping UNWRITTEN\n");
880 #endif
881 				*daddrp = 0;
882 			}
883 		}
884 	} else {
885 		/* If all blocks are goig to disk, update the "size on disk" */
886 		ip->i_lfs_osize = ip->i_ffs_size;
887 	}
888 
889 	if (ip->i_flag & IN_CLEANING)
890 		LFS_CLR_UINO(ip, IN_CLEANING);
891 	else {
892 		/* XXX IN_ALLMOD */
893 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
894 			     IN_UPDATE);
895 		if (ip->i_lfs_effnblks == ip->i_ffs_blocks)
896 			LFS_CLR_UINO(ip, IN_MODIFIED);
897 #ifdef DEBUG_LFS
898 		else
899 			printf("lfs_writeinode: ino %d: real blks=%d, "
900 			       "eff=%d\n", ip->i_number, ip->i_ffs_blocks,
901 			       ip->i_lfs_effnblks);
902 #endif
903 	}
904 
905 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
906 		sp->idp = ((struct dinode *)bp->b_data) +
907 			(sp->ninodes % INOPB(fs));
908 	if (gotblk) {
909 		LFS_LOCK_BUF(bp);
910 		brelse(bp);
911 	}
912 
913 	/* Increment inode count in segment summary block. */
914 	++((SEGSUM *)(sp->segsum))->ss_ninos;
915 
916 	/* If this page is full, set flag to allocate a new page. */
917 	if (++sp->ninodes % INOPB(fs) == 0)
918 		sp->ibp = NULL;
919 
920 	/*
921 	 * If updating the ifile, update the super-block.  Update the disk
922 	 * address and access times for this inode in the ifile.
923 	 */
924 	ino = ip->i_number;
925 	if (ino == LFS_IFILE_INUM) {
926 		daddr = fs->lfs_idaddr;
927 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
928 	} else {
929 		LFS_IENTRY(ifp, fs, ino, ibp);
930 		daddr = ifp->if_daddr;
931 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
932 #ifdef LFS_DEBUG_NEXTFREE
933 		if (ino > 3 && ifp->if_nextfree) {
934 			vprint("lfs_writeinode",ITOV(ip));
935 			printf("lfs_writeinode: updating free ino %d\n",
936 				ip->i_number);
937 		}
938 #endif
939 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
940 	}
941 
942 	/*
943 	 * The inode's last address should not be in the current partial
944 	 * segment, except under exceptional circumstances (lfs_writevnodes
945 	 * had to start over, and in the meantime more blocks were written
946 	 * to a vnode).  Both inodes will be accounted to this segment
947 	 * in lfs_writeseg so we need to subtract the earlier version
948 	 * here anyway.  The segment count can temporarily dip below
949 	 * zero here; keep track of how many duplicates we have in
950 	 * "dupino" so we don't panic below.
951 	 */
952 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno)) {
953 		++sp->ndupino;
954 		printf("lfs_writeinode: last inode addr in current pseg "
955 		       "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
956 			(long long)daddr, sp->ndupino);
957 	}
958 	/*
959 	 * Account the inode: it no longer belongs to its former segment,
960 	 * though it will not belong to the new segment until that segment
961 	 * is actually written.
962 	 */
963 	if (daddr != LFS_UNUSED_DADDR) {
964 		u_int32_t oldsn = dtosn(fs, daddr);
965 #ifdef DIAGNOSTIC
966 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
967 #endif
968 		LFS_SEGENTRY(sup, fs, oldsn, bp);
969 #ifdef DIAGNOSTIC
970 		if (sup->su_nbytes + DINODE_SIZE * ndupino < DINODE_SIZE) {
971 			printf("lfs_writeinode: negative bytes "
972 			       "(segment %" PRIu32 " short by %d, "
973 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
974 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
975 			       "ndupino=%d)\n",
976 			       dtosn(fs, daddr),
977 			       (int)DINODE_SIZE * (1 - sp->ndupino)
978 				   - sup->su_nbytes,
979 			       oldsn, sp->seg_number, daddr,
980 			       (unsigned int)sup->su_nbytes,
981 			       sp->ndupino);
982 			panic("lfs_writeinode: negative bytes");
983 			sup->su_nbytes = DINODE_SIZE;
984 		}
985 #endif
986 #ifdef DEBUG_SU_NBYTES
987 		printf("seg %d -= %d for ino %d inode\n",
988 		       dtosn(fs, daddr), DINODE_SIZE, ino);
989 #endif
990 		sup->su_nbytes -= DINODE_SIZE;
991 		redo_ifile =
992 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
993 		if (redo_ifile)
994 			fs->lfs_flags |= LFS_IFDIRTY;
995 		error = LFS_BWRITE_LOG(bp); /* Ifile */
996 	}
997 	return (redo_ifile);
998 }
999 
1000 int
1001 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1002 {
1003 	struct lfs *fs;
1004 	int version;
1005 
1006 	/*
1007 	 * If full, finish this segment.  We may be doing I/O, so
1008 	 * release and reacquire the splbio().
1009 	 */
1010 #ifdef DIAGNOSTIC
1011 	if (sp->vp == NULL)
1012 		panic ("lfs_gatherblock: Null vp in segment");
1013 #endif
1014 	fs = sp->fs;
1015 	if (sp->sum_bytes_left < sizeof(int32_t) ||
1016 	    sp->seg_bytes_left < bp->b_bcount) {
1017 		if (sptr)
1018 			splx(*sptr);
1019 		lfs_updatemeta(sp);
1020 
1021 		version = sp->fip->fi_version;
1022 		(void) lfs_writeseg(fs, sp);
1023 
1024 		sp->fip->fi_version = version;
1025 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
1026 		/* Add the current file to the segment summary. */
1027 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
1028 		sp->sum_bytes_left -= FINFOSIZE;
1029 
1030 		if (sptr)
1031 			*sptr = splbio();
1032 		return (1);
1033 	}
1034 
1035 #ifdef DEBUG
1036 	if (bp->b_flags & B_GATHERED) {
1037 		printf("lfs_gatherblock: already gathered! Ino %d,"
1038 		       " lbn %" PRId64 "\n",
1039 		       sp->fip->fi_ino, bp->b_lblkno);
1040 		return (0);
1041 	}
1042 #endif
1043 	/* Insert into the buffer list, update the FINFO block. */
1044 	bp->b_flags |= B_GATHERED;
1045 	bp->b_flags &= ~B_DONE;
1046 
1047 	*sp->cbpp++ = bp;
1048 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
1049 
1050 	sp->sum_bytes_left -= sizeof(int32_t);
1051 	sp->seg_bytes_left -= bp->b_bcount;
1052 	return (0);
1053 }
1054 
1055 int
1056 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, int (*match)(struct lfs *, struct buf *))
1057 {
1058 	struct buf *bp, *nbp;
1059 	int s, count = 0;
1060 
1061 	sp->vp = vp;
1062 	s = splbio();
1063 
1064 #ifndef LFS_NO_BACKBUF_HACK
1065 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1066 # define	BUF_OFFSET	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1067 # define	BACK_BUF(BP)	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1068 # define	BEG_OF_LIST	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1069 /* Find last buffer. */
1070 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1071 	    bp = LIST_NEXT(bp, b_vnbufs));
1072 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1073 		nbp = BACK_BUF(bp);
1074 #else /* LFS_NO_BACKBUF_HACK */
1075 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1076 		nbp = LIST_NEXT(bp, b_vnbufs);
1077 #endif /* LFS_NO_BACKBUF_HACK */
1078 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1079 #ifdef DEBUG_LFS
1080 			if (vp == fs->lfs_ivnode && (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1081 				printf("(%" PRId64 ":%lx)", bp->b_lblkno, bp->b_flags);
1082 #endif
1083 			continue;
1084 		}
1085 		if (vp->v_type == VBLK) {
1086 			/* For block devices, just write the blocks. */
1087 			/* XXX Do we really need to even do this? */
1088 #ifdef DEBUG_LFS
1089 			if (count == 0)
1090 				printf("BLK(");
1091 			printf(".");
1092 #endif
1093 			/* Get the block before bwrite, so we don't corrupt the free list */
1094 			bp->b_flags |= B_BUSY;
1095 			bremfree(bp);
1096 			bwrite(bp);
1097 		} else {
1098 #ifdef DIAGNOSTIC
1099 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1100 				printf("lfs_gather: lbn %" PRId64 " is "
1101 					"B_INVAL\n", bp->b_lblkno);
1102 				VOP_PRINT(bp->b_vp);
1103 			}
1104 			if (!(bp->b_flags & B_DELWRI))
1105 				panic("lfs_gather: bp not B_DELWRI");
1106 			if (!(bp->b_flags & B_LOCKED)) {
1107 				printf("lfs_gather: lbn %" PRId64 " blk "
1108 					"%" PRId64 " not B_LOCKED\n",
1109 					bp->b_lblkno,
1110 					dbtofsb(fs, bp->b_blkno));
1111 				VOP_PRINT(bp->b_vp);
1112 				panic("lfs_gather: bp not B_LOCKED");
1113 			}
1114 #endif
1115 			if (lfs_gatherblock(sp, bp, &s)) {
1116 				goto loop;
1117 			}
1118 		}
1119 		count++;
1120 	}
1121 	splx(s);
1122 #ifdef DEBUG_LFS
1123 	if (vp->v_type == VBLK && count)
1124 		printf(")\n");
1125 #endif
1126 	lfs_updatemeta(sp);
1127 	sp->vp = NULL;
1128 	return count;
1129 }
1130 
1131 /*
1132  * Update the metadata that points to the blocks listed in the FINFO
1133  * array.
1134  */
1135 void
1136 lfs_updatemeta(struct segment *sp)
1137 {
1138 	SEGUSE *sup;
1139 	struct buf *bp, *sbp;
1140 	struct lfs *fs;
1141 	struct vnode *vp;
1142 	struct indir a[NIADDR + 2], *ap;
1143 	struct inode *ip;
1144 	daddr_t daddr, lbn, off;
1145 	daddr_t ooff;
1146 	int error, i, nblocks, num;
1147 	int bb, osize, obb;
1148 
1149 	vp = sp->vp;
1150 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1151 	if (nblocks < 0)
1152 		panic("This is a bad thing");
1153 	if (vp == NULL || nblocks == 0)
1154 		return;
1155 
1156 	/* Sort the blocks. */
1157 	/*
1158 	 * XXX KS - We have to sort even if the blocks come from the
1159 	 * cleaner, because there might be other pending blocks on the
1160 	 * same inode...and if we don't sort, and there are fragments
1161 	 * present, blocks may be written in the wrong place.
1162 	 */
1163 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
1164 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
1165 
1166 	/*
1167 	 * Record the length of the last block in case it's a fragment.
1168 	 * If there are indirect blocks present, they sort last.  An
1169 	 * indirect block will be lfs_bsize and its presence indicates
1170 	 * that you cannot have fragments.
1171 	 *
1172 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1173 	 * XXX a later indirect block.  This will continue to be
1174 	 * XXX true until lfs_markv is fixed to do everything with
1175 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1176 	 */
1177 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1178 
1179 	/*
1180 	 * Assign disk addresses, and update references to the logical
1181 	 * block and the segment usage information.
1182 	 */
1183 	fs = sp->fs;
1184 	for (i = nblocks; i--; ++sp->start_bpp) {
1185 		lbn = *sp->start_lbp++;
1186 		sbp = *sp->start_bpp;
1187 
1188 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1189 		off = fs->lfs_offset;
1190 		if (sbp->b_blkno == sbp->b_lblkno) {
1191 			printf("lfs_updatemeta: ino %d blk %" PRId64
1192 			       " has same lbn and daddr\n",
1193 			       VTOI(vp)->i_number, off);
1194 		}
1195 
1196 		/*
1197 		 * If we write a frag in the wrong place, the cleaner won't
1198 		 * be able to correctly identify its size later, and the
1199 		 * segment will be uncleanable.  (Even worse, it will assume
1200 		 * that the indirect block that actually ends the list
1201 		 * is of a smaller size!)
1202 		 */
1203 		if (sbp->b_bcount < fs->lfs_bsize && i != 0)
1204 			panic("lfs_updatemeta: fragment is not last block");
1205 
1206 		bb = fragstofsb(fs, numfrags(fs, sbp->b_bcount));
1207 		fs->lfs_offset += bb;
1208 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1209 		if (daddr > 0)
1210 			daddr = dbtofsb(fs, daddr);
1211 		if (error)
1212 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1213 		ip = VTOI(vp);
1214 		switch (num) {
1215 		case 0:
1216 			ooff = ip->i_ffs_db[lbn];
1217 #ifdef DEBUG
1218 			if (ooff == 0) {
1219 				printf("lfs_updatemeta[1]: warning: writing "
1220 					"ino %d lbn %" PRId64 " at 0x%" PRIx64
1221 					", was 0x0\n", ip->i_number, lbn, off);
1222 			}
1223 #endif
1224 			if (ooff == UNWRITTEN)
1225 				ip->i_ffs_blocks += bb;
1226 			else {
1227 				/* possible fragment truncation or extension */
1228 				obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1229 				ip->i_ffs_blocks += (bb - obb);
1230 			}
1231 			ip->i_ffs_db[lbn] = off;
1232 			break;
1233 		case 1:
1234 			ooff = ip->i_ffs_ib[a[0].in_off];
1235 #ifdef DEBUG
1236 			if (ooff == 0) {
1237 				printf("lfs_updatemeta[2]: warning: writing "
1238 					"ino %d lbn %" PRId64 " at 0x%" PRIx64
1239 					", was 0x0\n", ip->i_number, lbn, off);
1240 			}
1241 #endif
1242 			if (ooff == UNWRITTEN)
1243 				ip->i_ffs_blocks += bb;
1244 			ip->i_ffs_ib[a[0].in_off] = off;
1245 			break;
1246 		default:
1247 			ap = &a[num - 1];
1248 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1249 				panic("lfs_updatemeta: bread bno %" PRId64,
1250 				      ap->in_lbn);
1251 
1252 			/* XXX ondisk32 */
1253 			ooff = ((int32_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 %" PRId64 " at 0x%" PRIx64
1258 					", was 0x0\n", ip->i_number, lbn, off);
1259 			}
1260 #endif
1261 			if (ooff == UNWRITTEN)
1262 				ip->i_ffs_blocks += bb;
1263 			/* XXX ondisk32 */
1264 			((int32_t *)bp->b_data)[ap->in_off] = off;
1265 			(void) VOP_BWRITE(bp);
1266 		}
1267 #ifdef DEBUG
1268 		if (daddr >= fs->lfs_lastpseg && daddr <= off) {
1269 			printf("lfs_updatemeta: ino %d, lbn %" PRId64 ", "
1270 				"addr = %" PRIx64 " in same pseg\n",
1271 				VTOI(sp->vp)->i_number, sbp->b_lblkno, daddr);
1272 		}
1273 #endif
1274 		/*
1275 		 * Update segment usage information, based on old size
1276 		 * and location.
1277 		 */
1278 		if (daddr > 0) {
1279 			u_int32_t oldsn = dtosn(fs, daddr);
1280 #ifdef DIAGNOSTIC
1281 			int ndupino = (sp->seg_number == oldsn) ?
1282 			    sp->ndupino : 0;
1283 #endif
1284 			if (lbn >= 0 && lbn < NDADDR)
1285 				osize = ip->i_lfs_fragsize[lbn];
1286 			else
1287 				osize = fs->lfs_bsize;
1288 			LFS_SEGENTRY(sup, fs, oldsn, bp);
1289 #ifdef DIAGNOSTIC
1290 			if (sup->su_nbytes + DINODE_SIZE * ndupino < osize) {
1291 				printf("lfs_updatemeta: negative bytes "
1292 				       "(segment %" PRIu32 " short by %d)\n",
1293 				       dtosn(fs, daddr),
1294 				       osize - sup->su_nbytes);
1295 				printf("lfs_updatemeta: ino %d, lbn %" PRId64
1296 				       ", addr = 0x%" PRIx64 "\n",
1297 				       VTOI(sp->vp)->i_number, lbn, daddr);
1298 				printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1299 				panic("lfs_updatemeta: negative bytes");
1300 				sup->su_nbytes = osize;
1301 			}
1302 #endif
1303 #ifdef DEBUG_SU_NBYTES
1304 			printf("seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1305 				" db 0x%" PRIx64 "\n",
1306 				dtosn(fs, daddr), osize,
1307 				VTOI(sp->vp)->i_number, lbn, daddr);
1308 #endif
1309 			sup->su_nbytes -= osize;
1310 			if (!(bp->b_flags & B_GATHERED))
1311 				fs->lfs_flags |= LFS_IFDIRTY;
1312 			error = LFS_BWRITE_LOG(bp); /* Ifile */
1313 		}
1314 		/*
1315 		 * Now that this block has a new address, and its old
1316 		 * segment no longer owns it, we can forget about its
1317 		 * old size.
1318 		 */
1319 		if (lbn >= 0 && lbn < NDADDR)
1320 			ip->i_lfs_fragsize[lbn] = sbp->b_bcount;
1321 	}
1322 }
1323 
1324 /*
1325  * Start a new segment.
1326  */
1327 int
1328 lfs_initseg(struct lfs *fs)
1329 {
1330 	struct segment *sp;
1331 	SEGUSE *sup;
1332 	SEGSUM *ssp;
1333 	struct buf *bp, *sbp;
1334 	int repeat;
1335 
1336 	sp = fs->lfs_sp;
1337 
1338 	repeat = 0;
1339 	/* Advance to the next segment. */
1340 	if (!LFS_PARTIAL_FITS(fs)) {
1341 		/* lfs_avail eats the remaining space */
1342 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1343 						   fs->lfs_curseg);
1344 		/* Wake up any cleaning procs waiting on this file system. */
1345 		wakeup(&lfs_allclean_wakeup);
1346 		wakeup(&fs->lfs_nextseg);
1347 		lfs_newseg(fs);
1348 		repeat = 1;
1349 		fs->lfs_offset = fs->lfs_curseg;
1350 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1351 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1352 		/*
1353 		 * If the segment contains a superblock, update the offset
1354 		 * and summary address to skip over it.
1355 		 */
1356 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1357 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1358 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1359 			sp->seg_bytes_left -= LFS_SBPAD;
1360 		}
1361 		brelse(bp);
1362 		/* Segment zero could also contain the labelpad */
1363 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1364 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1365 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1366 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1367 		}
1368 	} else {
1369 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1370 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1371 				      (fs->lfs_offset - fs->lfs_curseg));
1372 	}
1373 	fs->lfs_lastpseg = fs->lfs_offset;
1374 
1375 	sp->fs = fs;
1376 	sp->ibp = NULL;
1377 	sp->idp = NULL;
1378 	sp->ninodes = 0;
1379 	sp->ndupino = 0;
1380 
1381 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1382 	sp->cbpp = sp->bpp;
1383 #ifdef LFS_MALLOC_SUMMARY
1384 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1385 				     fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize);
1386   	sp->segsum = (*sp->cbpp)->b_data;
1387 #else
1388 	sbp = *sp->cbpp = getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1389 				 fsbtodb(fs, fs->lfs_offset), NBPG, 0, 0);
1390 	memset(sbp->b_data, 0x5a, NBPG);
1391 	sp->segsum = (*sp->cbpp)->b_data + NBPG - fs->lfs_sumsize;
1392 #endif
1393 	bzero(sp->segsum, fs->lfs_sumsize);
1394 	sp->start_bpp = ++sp->cbpp;
1395 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1396 
1397 	/* Set point to SEGSUM, initialize it. */
1398 	ssp = sp->segsum;
1399 	ssp->ss_next = fs->lfs_nextseg;
1400 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1401 	ssp->ss_magic = SS_MAGIC;
1402 
1403 	/* Set pointer to first FINFO, initialize it. */
1404 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1405 	sp->fip->fi_nblocks = 0;
1406 	sp->start_lbp = &sp->fip->fi_blocks[0];
1407 	sp->fip->fi_lastlength = 0;
1408 
1409 	sp->seg_bytes_left -= fs->lfs_sumsize;
1410 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1411 
1412 #ifndef LFS_MALLOC_SUMMARY
1413 	LFS_LOCK_BUF(sbp);
1414 	brelse(sbp);
1415 #endif
1416 	return (repeat);
1417 }
1418 
1419 /*
1420  * Return the next segment to write.
1421  */
1422 void
1423 lfs_newseg(struct lfs *fs)
1424 {
1425 	CLEANERINFO *cip;
1426 	SEGUSE *sup;
1427 	struct buf *bp;
1428 	int curseg, isdirty, sn;
1429 
1430 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1431 #ifdef DEBUG_SU_NBYTES
1432 	printf("lfs_newseg: seg %d := 0 in newseg\n",   /* XXXDEBUG */
1433 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1434 #endif
1435 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1436 	sup->su_nbytes = 0;
1437 	sup->su_nsums = 0;
1438 	sup->su_ninos = 0;
1439 	(void) LFS_BWRITE_LOG(bp); /* Ifile */
1440 
1441 	LFS_CLEANERINFO(cip, fs, bp);
1442 	--cip->clean;
1443 	++cip->dirty;
1444 	fs->lfs_nclean = cip->clean;
1445 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1446 
1447 	fs->lfs_lastseg = fs->lfs_curseg;
1448 	fs->lfs_curseg = fs->lfs_nextseg;
1449 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1450 		sn = (sn + 1) % fs->lfs_nseg;
1451 		if (sn == curseg)
1452 			panic("lfs_nextseg: no clean segments");
1453 		LFS_SEGENTRY(sup, fs, sn, bp);
1454 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1455 		brelse(bp);
1456 		if (!isdirty)
1457 			break;
1458 	}
1459 
1460 	++fs->lfs_nactive;
1461 	fs->lfs_nextseg = sntod(fs, sn);
1462 	if (lfs_dostats) {
1463 		++lfs_stats.segsused;
1464 	}
1465 }
1466 
1467 static struct buf **
1468 lookahead_pagemove(struct buf **bpp, int nblocks, size_t *size)
1469 {
1470 	size_t maxsize;
1471 #ifndef LFS_NO_PAGEMOVE
1472 	struct buf *bp;
1473 #endif
1474 
1475 	maxsize = *size;
1476 	*size = 0;
1477 #ifdef LFS_NO_PAGEMOVE
1478 	return bpp;
1479 #else
1480 	while((bp = *bpp) != NULL && *size < maxsize && nblocks--) {
1481 		if(bp->b_flags & B_CALL)
1482 			return bpp;
1483 		if(bp->b_bcount % NBPG)
1484 			return bpp;
1485 		*size += bp->b_bcount;
1486 		++bpp;
1487 	}
1488 	return NULL;
1489 #endif
1490 }
1491 
1492 #define BQUEUES 4 /* XXX */
1493 #define BQ_EMPTY 3 /* XXX */
1494 extern TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
1495 
1496 #define	BUFHASH(dvp, lbn)	\
1497 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
1498 extern LIST_HEAD(bufhashhdr, buf) invalhash;
1499 /*
1500  * Insq/Remq for the buffer hash lists.
1501  */
1502 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
1503 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
1504 
1505 static struct buf *
1506 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1507 {
1508 	struct lfs_cluster *cl;
1509 	struct buf **bpp, *bp;
1510 	int s;
1511 
1512 	cl = (struct lfs_cluster *)malloc(sizeof(*cl), M_SEGMENT, M_WAITOK);
1513 	bpp = (struct buf **)malloc(n*sizeof(*bpp), M_SEGMENT, M_WAITOK);
1514 	memset(cl, 0, sizeof(*cl));
1515 	cl->fs = fs;
1516 	cl->bpp = bpp;
1517 	cl->bufcount = 0;
1518 	cl->bufsize = 0;
1519 
1520 	/* If this segment is being written synchronously, note that */
1521 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1522 		cl->flags |= LFS_CL_SYNC;
1523 		cl->seg = fs->lfs_sp;
1524 		++cl->seg->seg_iocount;
1525 		/* printf("+ %x => %d\n", cl->seg, cl->seg->seg_iocount); */
1526 	}
1527 
1528 	/* Get an empty buffer header, or maybe one with something on it */
1529 	s = splbio();
1530 	if((bp = bufqueues[BQ_EMPTY].tqh_first) != NULL) {
1531 		bremfree(bp);
1532 		/* clear out various other fields */
1533 		bp->b_flags = B_BUSY;
1534 		bp->b_dev = NODEV;
1535 		bp->b_blkno = bp->b_lblkno = 0;
1536 		bp->b_error = 0;
1537 		bp->b_resid = 0;
1538 		bp->b_bcount = 0;
1539 
1540 		/* nuke any credentials we were holding */
1541 		/* XXXXXX */
1542 
1543 		bremhash(bp);
1544 
1545 		/* disassociate us from our vnode, if we had one... */
1546 		if (bp->b_vp)
1547 			brelvp(bp);
1548 	}
1549 	splx(s);
1550 	while (!bp)
1551 		bp = getnewbuf(0, 0);
1552 	s = splbio();
1553 	bgetvp(vp, bp);
1554 	binshash(bp,&invalhash);
1555 	splx(s);
1556 	bp->b_bcount = 0;
1557 	bp->b_blkno = bp->b_lblkno = addr;
1558 
1559 	bp->b_flags |= B_CALL;
1560 	bp->b_iodone = lfs_cluster_callback;
1561 	cl->saveaddr = bp->b_saveaddr; /* XXX is this ever used? */
1562 	bp->b_saveaddr = (caddr_t)cl;
1563 
1564 	return bp;
1565 }
1566 
1567 int
1568 lfs_writeseg(struct lfs *fs, struct segment *sp)
1569 {
1570 	struct buf **bpp, *bp, *cbp, *newbp, **pmlastbpp;
1571 	SEGUSE *sup;
1572 	SEGSUM *ssp;
1573 	dev_t i_dev;
1574 	char *datap, *dp;
1575 	int do_again, i, nblocks, s;
1576 	size_t el_size;
1577  	struct lfs_cluster *cl;
1578 	int (*strategy)(void *);
1579 	struct vop_strategy_args vop_strategy_a;
1580 	u_short ninos;
1581 	struct vnode *devvp;
1582 	char *p;
1583 	struct vnode *vp;
1584 	struct inode *ip;
1585 	size_t pmsize;
1586 	int use_pagemove;
1587 	int32_t *daddrp;	/* XXX ondisk32 */
1588 	int changed;
1589 #if defined(DEBUG) && defined(LFS_PROPELLER)
1590 	static int propeller;
1591 	char propstring[4] = "-\\|/";
1592 
1593 	printf("%c\b",propstring[propeller++]);
1594 	if (propeller == 4)
1595 		propeller = 0;
1596 #endif
1597 
1598 	/*
1599 	 * If there are no buffers other than the segment summary to write
1600 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1601 	 * even if there aren't any buffers, you need to write the superblock.
1602 	 */
1603 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1604 		return (0);
1605 
1606 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1607 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1608 
1609 	/* Update the segment usage information. */
1610 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1611 
1612 	/* Loop through all blocks, except the segment summary. */
1613 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1614 		if ((*bpp)->b_vp != devvp) {
1615 			sup->su_nbytes += (*bpp)->b_bcount;
1616 #ifdef DEBUG_SU_NBYTES
1617 		printf("seg %" PRIu32 " += %ld for ino %d lbn %" PRId64
1618 		    " db 0x%" PRIx64 "\n", sp->seg_number, (*bpp)->b_bcount,
1619 		    VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
1620 		    (*bpp)->b_blkno);
1621 #endif
1622 		}
1623 	}
1624 
1625 	ssp = (SEGSUM *)sp->segsum;
1626 
1627 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1628 #ifdef DEBUG_SU_NBYTES
1629 	printf("seg %d += %d for %d inodes\n",   /* XXXDEBUG */
1630 	       sp->seg_number, ssp->ss_ninos * DINODE_SIZE,
1631 	       ssp->ss_ninos);
1632 #endif
1633 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1634 	/* sup->su_nbytes += fs->lfs_sumsize; */
1635 	if (fs->lfs_version == 1)
1636 		sup->su_olastmod = time.tv_sec;
1637 	else
1638 		sup->su_lastmod = time.tv_sec;
1639 	sup->su_ninos += ninos;
1640 	++sup->su_nsums;
1641 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1642 							 fs->lfs_ibsize));
1643 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1644 
1645 	do_again = !(bp->b_flags & B_GATHERED);
1646 	(void)LFS_BWRITE_LOG(bp); /* Ifile */
1647 	/*
1648 	 * Mark blocks B_BUSY, to prevent then from being changed between
1649 	 * the checksum computation and the actual write.
1650 	 *
1651 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1652 	 * there are any, replace them with copies that have UNASSIGNED
1653 	 * instead.
1654 	 */
1655 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1656 		++bpp;
1657 		if ((*bpp)->b_flags & B_CALL)
1658 			continue;
1659 		bp = *bpp;
1660 	    again:
1661 		s = splbio();
1662 		if (bp->b_flags & B_BUSY) {
1663 #ifdef DEBUG
1664 			printf("lfs_writeseg: avoiding potential data summary "
1665 			       "corruption for ino %d, lbn %" PRId64 "\n",
1666 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1667 #endif
1668 			bp->b_flags |= B_WANTED;
1669 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1670 			splx(s);
1671 			goto again;
1672 		}
1673 		bp->b_flags |= B_BUSY;
1674 		splx(s);
1675 		/* Check and replace indirect block UNWRITTEN bogosity */
1676 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1677 		   VTOI(bp->b_vp)->i_ffs_blocks !=
1678 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1679 #ifdef DEBUG_LFS
1680 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1681 			       VTOI(bp->b_vp)->i_number,
1682 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1683 			       VTOI(bp->b_vp)->i_ffs_blocks);
1684 #endif
1685 			/* Make a copy we'll make changes to */
1686 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1687 					   bp->b_bcount);
1688 			newbp->b_blkno = bp->b_blkno;
1689 			memcpy(newbp->b_data, bp->b_data,
1690 			       newbp->b_bcount);
1691 			*bpp = newbp;
1692 
1693 			changed = 0;
1694 			/* XXX ondisk32 */
1695 			for (daddrp = (int32_t *)(newbp->b_data);
1696 			     daddrp < (int32_t *)(newbp->b_data +
1697 						  newbp->b_bcount); daddrp++) {
1698 				if (*daddrp == UNWRITTEN) {
1699 					++changed;
1700 #ifdef DEBUG_LFS
1701 					printf("lfs_writeseg: replacing UNWRITTEN\n");
1702 #endif
1703 					*daddrp = 0;
1704 				}
1705 			}
1706 			/*
1707 			 * Get rid of the old buffer.  Don't mark it clean,
1708 			 * though, if it still has dirty data on it.
1709 			 */
1710 			if (changed) {
1711 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1712 				if (bp->b_flags & B_CALL) {
1713 					lfs_freebuf(bp);
1714 					bp = NULL;
1715 				} else {
1716 					/* Still on free list, leave it there */
1717 					s = splbio();
1718 					bp->b_flags &= ~B_BUSY;
1719 					if (bp->b_flags & B_WANTED)
1720 						wakeup(bp);
1721 				 	splx(s);
1722 					/*
1723 					 * We have to re-decrement lfs_avail
1724 					 * since this block is going to come
1725 					 * back around to us in the next
1726 					 * segment.
1727 					 */
1728 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1729 				}
1730 			} else {
1731 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1732 						 B_GATHERED);
1733 				if (bp->b_flags & B_CALL) {
1734 					lfs_freebuf(bp);
1735 					bp = NULL;
1736 				} else {
1737 					bremfree(bp);
1738 					bp->b_flags |= B_DONE;
1739 					s = splbio();
1740 					reassignbuf(bp, bp->b_vp);
1741 					splx(s);
1742 					LFS_UNLOCK_BUF(bp);
1743 					brelse(bp);
1744 				}
1745 			}
1746 
1747 		}
1748 	}
1749 	/*
1750 	 * Compute checksum across data and then across summary; the first
1751 	 * block (the summary block) is skipped.  Set the create time here
1752 	 * so that it's guaranteed to be later than the inode mod times.
1753 	 *
1754 	 * XXX
1755 	 * Fix this to do it inline, instead of malloc/copy.
1756 	 */
1757 	if (fs->lfs_version == 1)
1758 		el_size = sizeof(u_long);
1759 	else
1760 		el_size = sizeof(u_int32_t);
1761 	datap = dp = malloc(nblocks * el_size, M_SEGMENT, M_WAITOK);
1762 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1763 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1764 			if (copyin((*bpp)->b_saveaddr, dp, el_size))
1765 				panic("lfs_writeseg: copyin failed [1]: "
1766 				      "ino %d blk %" PRId64,
1767 				      VTOI((*bpp)->b_vp)->i_number,
1768 				      (*bpp)->b_lblkno);
1769 		} else
1770 			memcpy(dp, (*bpp)->b_data, el_size);
1771 		dp += el_size;
1772 	}
1773 	if (fs->lfs_version == 1)
1774 		ssp->ss_ocreate = time.tv_sec;
1775 	else {
1776 		ssp->ss_create = time.tv_sec;
1777 		ssp->ss_serial = ++fs->lfs_serial;
1778 		ssp->ss_ident  = fs->lfs_ident;
1779 	}
1780 #ifndef LFS_MALLOC_SUMMARY
1781 	/* Set the summary block busy too */
1782 	(*(sp->bpp))->b_flags |= B_BUSY;
1783 #endif
1784 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * el_size);
1785 	ssp->ss_sumsum =
1786 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1787 	free(datap, M_SEGMENT);
1788 	datap = dp = NULL;
1789 #ifdef DIAGNOSTIC
1790 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1791 		panic("lfs_writeseg: No diskspace for summary");
1792 #endif
1793 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1794 			  btofsb(fs, fs->lfs_sumsize));
1795 
1796 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1797 
1798 	/*
1799   	 * When we simply write the blocks we lose a rotation for every block
1800 	 * written.  To avoid this problem, we use pagemove to cluster
1801 	 * the buffers into a chunk and write the chunk.  CHUNKSIZE is the
1802   	 * largest size I/O devices can handle.
1803   	 *
1804 	 * XXX - right now MAXPHYS is only 64k; could it be larger?
1805 	 */
1806 
1807 #define CHUNKSIZE MAXPHYS
1808 
1809 	if (devvp == NULL)
1810 		panic("devvp is NULL");
1811 	for (bpp = sp->bpp, i = nblocks; i;) {
1812 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1813 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1814 
1815 		cbp->b_dev = i_dev;
1816 		cbp->b_flags |= B_ASYNC | B_BUSY;
1817 		cbp->b_bcount = 0;
1818 
1819 		/*
1820 		 * Find out if we can use pagemove to build the cluster,
1821 		 * or if we are stuck using malloc/copy.  If this is the
1822 		 * first cluster, set the shift flag (see below).
1823 		 */
1824 		pmsize = CHUNKSIZE;
1825 		use_pagemove = 0;
1826 		if(bpp == sp->bpp) {
1827 			/* Summary blocks have to get special treatment */
1828 			pmlastbpp = lookahead_pagemove(bpp + 1, i - 1, &pmsize);
1829 			if(pmsize >= CHUNKSIZE - fs->lfs_sumsize ||
1830 			   pmlastbpp == NULL) {
1831 				use_pagemove = 1;
1832 				cl->flags |= LFS_CL_SHIFT;
1833 			} else {
1834 				/*
1835 				 * If we're not using pagemove, we have
1836 				 * to copy the summary down to the bottom
1837 				 * end of the block.
1838 				 */
1839 #ifndef LFS_MALLOC_SUMMARY
1840 				memcpy((*bpp)->b_data, (*bpp)->b_data +
1841 				       NBPG - fs->lfs_sumsize,
1842 				       fs->lfs_sumsize);
1843 #endif /* LFS_MALLOC_SUMMARY */
1844 			}
1845 		} else {
1846 			pmlastbpp = lookahead_pagemove(bpp, i, &pmsize);
1847 			if(pmsize >= CHUNKSIZE || pmlastbpp == NULL) {
1848 				use_pagemove = 1;
1849 			}
1850 		}
1851 		if(use_pagemove == 0) {
1852 			cl->flags |= LFS_CL_MALLOC;
1853 			cl->olddata = cbp->b_data;
1854 			cbp->b_data = malloc(CHUNKSIZE, M_SEGMENT, M_WAITOK);
1855 		}
1856 #if defined(DEBUG) && defined(DIAGNOSTIC)
1857 		if(dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno + btodb((*bpp)->b_bcount - 1))) !=
1858 		   dtosn(fs, dbtofsb(fs, cbp->b_blkno))) {
1859 			printf("block at %" PRId64 " (%" PRIu32 "), "
1860 			       "cbp at %" PRId64 " (%" PRIu32 ")\n",
1861 				(*bpp)->b_blkno, dtosn(fs, dbtofsb(fs, (*bpp)->b_blkno)),
1862 			       cbp->b_blkno, dtosn(fs, dbtofsb(fs, cbp->b_blkno)));
1863 			panic("lfs_writeseg: Segment overwrite");
1864 		}
1865 #endif
1866 
1867 		/*
1868 		 * Construct the cluster.
1869 		 */
1870 		while (fs->lfs_iocount >= LFS_THROTTLE) {
1871 #ifdef DEBUG_LFS
1872 			printf("[%d]", fs->lfs_iocount);
1873 #endif
1874 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs_throttle", 0);
1875 		}
1876 		++fs->lfs_iocount;
1877 
1878 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1879 			bp = *bpp;
1880 
1881 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1882 				break;
1883 
1884 			/*
1885 			 * Fake buffers from the cleaner are marked as B_INVAL.
1886 			 * We need to copy the data from user space rather than
1887 			 * from the buffer indicated.
1888 			 * XXX == what do I do on an error?
1889 			 */
1890 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1891 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1892 					panic("lfs_writeseg: copyin failed [2]");
1893 			} else if (use_pagemove) {
1894 				pagemove(bp->b_data, p, bp->b_bcount);
1895 				cbp->b_bufsize += bp->b_bcount;
1896 				bp->b_bufsize -= bp->b_bcount;
1897   			} else {
1898 				bcopy(bp->b_data, p, bp->b_bcount);
1899 				/* printf("copy in %p\n", bp->b_data); */
1900   			}
1901 
1902 			/*
1903 			 * XXX If we are *not* shifting, the summary
1904 			 * block is only fs->lfs_sumsize.  Otherwise,
1905 			 * it is NBPG but shifted.
1906 			 */
1907 			if(bpp == sp->bpp && !(cl->flags & LFS_CL_SHIFT)) {
1908 				p += fs->lfs_sumsize;
1909 				cbp->b_bcount += fs->lfs_sumsize;
1910 				cl->bufsize += fs->lfs_sumsize;
1911 			} else {
1912 				p += bp->b_bcount;
1913 				cbp->b_bcount += bp->b_bcount;
1914 				cl->bufsize += bp->b_bcount;
1915 			}
1916 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
1917 			cl->bpp[cl->bufcount++] = bp;
1918 			vp = bp->b_vp;
1919 			s = splbio();
1920 			++vp->v_numoutput;
1921 			splx(s);
1922 
1923 			/*
1924 			 * Although it cannot be freed for reuse before the
1925 			 * cluster is written to disk, this buffer does not
1926 			 * need to be held busy.  Therefore we unbusy it,
1927 			 * while leaving it on the locked list.  It will
1928 			 * be freed or requeued by the callback depending
1929 			 * on whether it has had B_DELWRI set again in the
1930 			 * meantime.
1931 			 *
1932 			 * If we are using pagemove, we have to hold the block
1933 			 * busy to prevent its contents from changing before
1934 			 * it hits the disk, and invalidating the checksum.
1935 			 */
1936 			bp->b_flags &= ~(B_DELWRI | B_READ | B_ERROR);
1937 #ifdef LFS_MNOBUSY
1938 			if (cl->flags & LFS_CL_MALLOC) {
1939 				if (!(bp->b_flags & B_CALL))
1940 					brelse(bp); /* Still B_LOCKED */
1941 			}
1942 #endif
1943 			bpp++;
1944 
1945 			/*
1946 			 * If this is the last block for this vnode, but
1947 			 * there are other blocks on its dirty list,
1948 			 * set IN_MODIFIED/IN_CLEANING depending on what
1949 			 * sort of block.  Only do this for our mount point,
1950 			 * not for, e.g., inode blocks that are attached to
1951 			 * the devvp.
1952 			 * XXX KS - Shouldn't we set *both* if both types
1953 			 * of blocks are present (traverse the dirty list?)
1954 			 */
1955 			s = splbio();
1956 			if ((i == 1 ||
1957 			     (i > 1 && vp && *bpp && (*bpp)->b_vp != vp)) &&
1958 			    (bp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL &&
1959 			    vp->v_mount == fs->lfs_ivnode->v_mount)
1960   			{
1961 				ip = VTOI(vp);
1962 #ifdef DEBUG_LFS
1963 				printf("lfs_writeseg: marking ino %d\n",
1964 				       ip->i_number);
1965 #endif
1966 				if (bp->b_flags & B_CALL)
1967 					LFS_SET_UINO(ip, IN_CLEANING);
1968 				else
1969 					LFS_SET_UINO(ip, IN_MODIFIED);
1970 			}
1971 			splx(s);
1972 			wakeup(vp);
1973 		}
1974 		s = splbio();
1975 		++cbp->b_vp->v_numoutput;
1976 		splx(s);
1977 		/*
1978 		 * In order to include the summary in a clustered block,
1979 		 * it may be necessary to shift the block forward (since
1980 		 * summary blocks are in generay smaller than can be
1981 		 * addressed by pagemove().  After the write, the block
1982 		 * will be corrected before disassembly.
1983 		 */
1984 		if(cl->flags & LFS_CL_SHIFT) {
1985 			cbp->b_data += (NBPG - fs->lfs_sumsize);
1986 			cbp->b_bcount -= (NBPG - fs->lfs_sumsize);
1987 		}
1988 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1989 		vop_strategy_a.a_bp = cbp;
1990 		(strategy)(&vop_strategy_a);
1991 	}
1992 
1993 	if (lfs_dostats) {
1994 		++lfs_stats.psegwrites;
1995 		lfs_stats.blocktot += nblocks - 1;
1996 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1997 			++lfs_stats.psyncwrites;
1998 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1999 			++lfs_stats.pcleanwrites;
2000 			lfs_stats.cleanblocks += nblocks - 1;
2001 		}
2002 	}
2003 	return (lfs_initseg(fs) || do_again);
2004 }
2005 
2006 void
2007 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2008 {
2009 	struct buf *bp;
2010 	dev_t i_dev;
2011 	int (*strategy)(void *);
2012 	int s;
2013 	struct vop_strategy_args vop_strategy_a;
2014 
2015 	/*
2016 	 * If we can write one superblock while another is in
2017 	 * progress, we risk not having a complete checkpoint if we crash.
2018 	 * So, block here if a superblock write is in progress.
2019 	 */
2020 	s = splbio();
2021 	while (fs->lfs_sbactive) {
2022 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
2023 	}
2024 	fs->lfs_sbactive = daddr;
2025 	splx(s);
2026 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
2027 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
2028 
2029 	/* Set timestamp of this version of the superblock */
2030 	if (fs->lfs_version == 1)
2031 		fs->lfs_otstamp = time.tv_sec;
2032 	fs->lfs_tstamp = time.tv_sec;
2033 
2034 	/* Checksum the superblock and copy it into a buffer. */
2035 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2036 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD);
2037 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2038 
2039 	bp->b_dev = i_dev;
2040 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2041 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2042 	bp->b_iodone = lfs_supercallback;
2043 	/* XXX KS - same nasty hack as above */
2044 	bp->b_saveaddr = (caddr_t)fs;
2045 
2046 	vop_strategy_a.a_desc = VDESC(vop_strategy);
2047 	vop_strategy_a.a_bp = bp;
2048 	s = splbio();
2049 	++bp->b_vp->v_numoutput;
2050 	splx(s);
2051 	++fs->lfs_iocount;
2052 	(strategy)(&vop_strategy_a);
2053 }
2054 
2055 /*
2056  * Logical block number match routines used when traversing the dirty block
2057  * chain.
2058  */
2059 int
2060 lfs_match_fake(struct lfs *fs, struct buf *bp)
2061 {
2062 	return (bp->b_flags & B_CALL);
2063 }
2064 
2065 int
2066 lfs_match_data(struct lfs *fs, struct buf *bp)
2067 {
2068 	return (bp->b_lblkno >= 0);
2069 }
2070 
2071 int
2072 lfs_match_indir(struct lfs *fs, struct buf *bp)
2073 {
2074 	daddr_t lbn;
2075 
2076 	lbn = bp->b_lblkno;
2077 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2078 }
2079 
2080 int
2081 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2082 {
2083 	daddr_t lbn;
2084 
2085 	lbn = bp->b_lblkno;
2086 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2087 }
2088 
2089 int
2090 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2091 {
2092 	daddr_t lbn;
2093 
2094 	lbn = bp->b_lblkno;
2095 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2096 }
2097 
2098 /*
2099  * XXX - The only buffers that are going to hit these functions are the
2100  * segment write blocks, or the segment summaries, or the superblocks.
2101  *
2102  * All of the above are created by lfs_newbuf, and so do not need to be
2103  * released via brelse.
2104  */
2105 void
2106 lfs_callback(struct buf *bp)
2107 {
2108 	/* struct lfs *fs; */
2109 	/* fs = (struct lfs *)bp->b_saveaddr; */
2110 	lfs_freebuf(bp);
2111 }
2112 
2113 static void
2114 lfs_super_aiodone(struct buf *bp)
2115 {
2116 	struct lfs *fs;
2117 
2118 	fs = (struct lfs *)bp->b_saveaddr;
2119 	fs->lfs_sbactive = 0;
2120 	wakeup(&fs->lfs_sbactive);
2121 	if (--fs->lfs_iocount < LFS_THROTTLE)
2122 		wakeup(&fs->lfs_iocount);
2123 	lfs_freebuf(bp);
2124 }
2125 
2126 static void
2127 lfs_cluster_aiodone(struct buf *bp)
2128 {
2129 	struct lfs_cluster *cl;
2130 	struct lfs *fs;
2131 	struct buf *tbp;
2132 	struct vnode *vp;
2133 	int s, error=0;
2134 	char *cp;
2135 	extern int locked_queue_count;
2136 	extern long locked_queue_bytes;
2137 
2138 	if(bp->b_flags & B_ERROR)
2139 		error = bp->b_error;
2140 
2141 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2142 	fs = cl->fs;
2143 	bp->b_saveaddr = cl->saveaddr;
2144 
2145 	/* If shifted, shift back now */
2146 	if(cl->flags & LFS_CL_SHIFT) {
2147 		bp->b_data -= (NBPG - fs->lfs_sumsize);
2148 		bp->b_bcount += (NBPG - fs->lfs_sumsize);
2149 	}
2150 
2151 	cp = (char *)bp->b_data + cl->bufsize;
2152 	/* Put the pages back, and release the buffer */
2153 	while(cl->bufcount--) {
2154 		tbp = cl->bpp[cl->bufcount];
2155 		if(!(cl->flags & LFS_CL_MALLOC)) {
2156 			cp -= tbp->b_bcount;
2157 			printf("pm(%p,%p,%lx)",cp,tbp->b_data,tbp->b_bcount);
2158 			pagemove(cp, tbp->b_data, tbp->b_bcount);
2159 			bp->b_bufsize -= tbp->b_bcount;
2160 			tbp->b_bufsize += tbp->b_bcount;
2161 		}
2162 		if(error) {
2163 			tbp->b_flags |= B_ERROR;
2164 			tbp->b_error = error;
2165 		}
2166 
2167 		/*
2168 		 * We're done with tbp.  If it has not been re-dirtied since
2169 		 * the cluster was written, free it.  Otherwise, keep it on
2170 		 * the locked list to be written again.
2171 		 */
2172 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2173 			LFS_UNLOCK_BUF(tbp);
2174 		tbp->b_flags &= ~B_GATHERED;
2175 
2176 		LFS_BCLEAN_LOG(fs, tbp);
2177 
2178 		vp = tbp->b_vp;
2179 		/* Segment summary for a shifted cluster */
2180 		if(!cl->bufcount && (cl->flags & LFS_CL_SHIFT))
2181 			tbp->b_flags |= B_INVAL;
2182 		if(!(tbp->b_flags & B_CALL)) {
2183 			bremfree(tbp);
2184 			s = splbio();
2185 			if(vp)
2186 				reassignbuf(tbp, vp);
2187 			splx(s);
2188 			tbp->b_flags |= B_ASYNC; /* for biodone */
2189 		}
2190 #ifdef DIAGNOSTIC
2191 		if (tbp->b_flags & B_DONE) {
2192 			printf("blk %d biodone already (flags %lx)\n",
2193 				cl->bufcount, (long)tbp->b_flags);
2194 		}
2195 #endif
2196 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2197 			biodone(tbp);
2198 		}
2199 	}
2200 
2201 	/* Fix up the cluster buffer, and release it */
2202 	if(!(cl->flags & LFS_CL_MALLOC) && bp->b_bufsize) {
2203 		printf("PM(%p,%p,%lx)", (char *)bp->b_data + bp->b_bcount,
2204 			 (char *)bp->b_data, bp->b_bufsize);
2205 		pagemove((char *)bp->b_data + bp->b_bcount,
2206 			 (char *)bp->b_data, bp->b_bufsize);
2207 	}
2208 	if(cl->flags & LFS_CL_MALLOC) {
2209 		free(bp->b_data, M_SEGMENT);
2210 		bp->b_data = cl->olddata;
2211 	}
2212 	bp->b_bcount = 0;
2213 	bp->b_iodone = NULL;
2214 	bp->b_flags &= ~B_DELWRI;
2215 	bp->b_flags |= B_DONE;
2216 	s = splbio();
2217 	reassignbuf(bp, bp->b_vp);
2218 	splx(s);
2219 	brelse(bp);
2220 
2221 	/* Note i/o done */
2222 	if (cl->flags & LFS_CL_SYNC) {
2223 		if (--cl->seg->seg_iocount == 0)
2224 			wakeup(&cl->seg->seg_iocount);
2225 		/* printf("- %x => %d\n", cl->seg, cl->seg->seg_iocount); */
2226 	}
2227 #ifdef DIAGNOSTIC
2228 	if (fs->lfs_iocount == 0)
2229 		panic("lfs_cluster_aiodone: zero iocount");
2230 #endif
2231 	if (--fs->lfs_iocount < LFS_THROTTLE)
2232 		wakeup(&fs->lfs_iocount);
2233 #if 0
2234 	if (fs->lfs_iocount == 0) {
2235 		/*
2236 		 * Vinvalbuf can move locked buffers off the locked queue
2237 		 * and we have no way of knowing about this.  So, after
2238 		 * doing a big write, we recalculate how many buffers are
2239 		 * really still left on the locked queue.
2240 		 */
2241 		lfs_countlocked(&locked_queue_count, &locked_queue_bytes, "lfs_cluster_callback");
2242 		wakeup(&locked_queue_count);
2243 	}
2244 #endif
2245 
2246 	free(cl->bpp, M_SEGMENT);
2247 	free(cl, M_SEGMENT);
2248 }
2249 
2250 static void
2251 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2252 {
2253 	/* reset b_iodone for when this is a single-buf i/o. */
2254 	bp->b_iodone = aiodone;
2255 
2256 	simple_lock(&uvm.aiodoned_lock);        /* locks uvm.aio_done */
2257 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2258 	wakeup(&uvm.aiodoned);
2259 	simple_unlock(&uvm.aiodoned_lock);
2260 }
2261 
2262 static void
2263 lfs_cluster_callback(struct buf *bp)
2264 {
2265 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2266 }
2267 
2268 void
2269 lfs_supercallback(struct buf *bp)
2270 {
2271 	lfs_generic_callback(bp, lfs_super_aiodone);
2272 }
2273 
2274 /*
2275  * Shellsort (diminishing increment sort) from Data Structures and
2276  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2277  * see also Knuth Vol. 3, page 84.  The increments are selected from
2278  * formula (8), page 95.  Roughly O(N^3/2).
2279  */
2280 /*
2281  * This is our own private copy of shellsort because we want to sort
2282  * two parallel arrays (the array of buffer pointers and the array of
2283  * logical block numbers) simultaneously.  Note that we cast the array
2284  * of logical block numbers to a unsigned in this routine so that the
2285  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2286  */
2287 
2288 void
2289 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb)
2290 {
2291 	static int __rsshell_increments[] = { 4, 1, 0 };
2292 	int incr, *incrp, t1, t2;
2293 	struct buf *bp_temp;
2294 	u_long lb_temp;
2295 
2296 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2297 		for (t1 = incr; t1 < nmemb; ++t1)
2298 			for (t2 = t1 - incr; t2 >= 0;)
2299 				if (lb_array[t2] > lb_array[t2 + incr]) {
2300 					lb_temp = lb_array[t2];
2301 					lb_array[t2] = lb_array[t2 + incr];
2302 					lb_array[t2 + incr] = lb_temp;
2303 					bp_temp = bp_array[t2];
2304 					bp_array[t2] = bp_array[t2 + incr];
2305 					bp_array[t2 + incr] = bp_temp;
2306 					t2 -= incr;
2307 				} else
2308 					break;
2309 }
2310 
2311 /*
2312  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2313  */
2314 int
2315 lfs_vref(struct vnode *vp)
2316 {
2317 	/*
2318 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2319 	 * being able to flush all of the pages from this vnode, which
2320 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2321 	 */
2322 	if (vp->v_flag & VXLOCK) {
2323 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2324 			return 0;
2325 		}
2326 		return (1);
2327 	}
2328 	return (vget(vp, 0));
2329 }
2330 
2331 /*
2332  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2333  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2334  */
2335 void
2336 lfs_vunref(struct vnode *vp)
2337 {
2338 	/*
2339 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2340 	 */
2341 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2342 		return;
2343 	}
2344 
2345 	simple_lock(&vp->v_interlock);
2346 #ifdef DIAGNOSTIC
2347 	if (vp->v_usecount <= 0) {
2348 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2349 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2350 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2351 		panic("lfs_vunref: v_usecount<0");
2352 	}
2353 #endif
2354 	vp->v_usecount--;
2355 	if (vp->v_usecount > 0) {
2356 		simple_unlock(&vp->v_interlock);
2357 		return;
2358 	}
2359 	/*
2360 	 * insert at tail of LRU list
2361 	 */
2362 	simple_lock(&vnode_free_list_slock);
2363 	if (vp->v_holdcnt > 0)
2364 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2365 	else
2366 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2367 	simple_unlock(&vnode_free_list_slock);
2368 	simple_unlock(&vp->v_interlock);
2369 }
2370 
2371 /*
2372  * We use this when we have vnodes that were loaded in solely for cleaning.
2373  * There is no reason to believe that these vnodes will be referenced again
2374  * soon, since the cleaning process is unrelated to normal filesystem
2375  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2376  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2377  * cleaning at the head of the list, instead.
2378  */
2379 void
2380 lfs_vunref_head(struct vnode *vp)
2381 {
2382 	simple_lock(&vp->v_interlock);
2383 #ifdef DIAGNOSTIC
2384 	if (vp->v_usecount == 0) {
2385 		panic("lfs_vunref: v_usecount<0");
2386 	}
2387 #endif
2388 	vp->v_usecount--;
2389 	if (vp->v_usecount > 0) {
2390 		simple_unlock(&vp->v_interlock);
2391 		return;
2392 	}
2393 	/*
2394 	 * insert at head of LRU list
2395 	 */
2396 	simple_lock(&vnode_free_list_slock);
2397 	if (vp->v_holdcnt > 0)
2398 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2399 	else
2400 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2401 	simple_unlock(&vnode_free_list_slock);
2402 	simple_unlock(&vp->v_interlock);
2403 }
2404 
2405