xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision df0caa2637da0538ecdf6b878c4d08e684b43d8f)
1 /*	$NetBSD: lfs_segment.c,v 1.164 2005/05/29 21:25:24 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 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. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
67  */
68 
69 #include <sys/cdefs.h>
70 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.164 2005/05/29 21:25:24 christos Exp $");
71 
72 #ifdef DEBUG
73 # define vndebug(vp, str) do {						\
74 	if (VTOI(vp)->i_flag & IN_CLEANING)				\
75 		DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \
76 		     VTOI(vp)->i_number, (str), op));			\
77 } while(0)
78 #else
79 # define vndebug(vp, str)
80 #endif
81 #define ivndebug(vp, str) \
82 	DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str)))
83 
84 #if defined(_KERNEL_OPT)
85 #include "opt_ddb.h"
86 #endif
87 
88 #include <sys/param.h>
89 #include <sys/systm.h>
90 #include <sys/namei.h>
91 #include <sys/kernel.h>
92 #include <sys/resourcevar.h>
93 #include <sys/file.h>
94 #include <sys/stat.h>
95 #include <sys/buf.h>
96 #include <sys/proc.h>
97 #include <sys/vnode.h>
98 #include <sys/mount.h>
99 
100 #include <miscfs/specfs/specdev.h>
101 #include <miscfs/fifofs/fifo.h>
102 
103 #include <ufs/ufs/inode.h>
104 #include <ufs/ufs/dir.h>
105 #include <ufs/ufs/ufsmount.h>
106 #include <ufs/ufs/ufs_extern.h>
107 
108 #include <ufs/lfs/lfs.h>
109 #include <ufs/lfs/lfs_extern.h>
110 
111 #include <uvm/uvm.h>
112 #include <uvm/uvm_extern.h>
113 
114 MALLOC_DEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
115 
116 extern int count_lock_queue(void);
117 extern struct simplelock vnode_free_list_slock;		/* XXX */
118 extern struct simplelock bqueue_slock;			/* XXX */
119 
120 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
121 static void lfs_super_aiodone(struct buf *);
122 static void lfs_cluster_aiodone(struct buf *);
123 static void lfs_cluster_callback(struct buf *);
124 
125 /*
126  * Determine if it's OK to start a partial in this segment, or if we need
127  * to go on to a new segment.
128  */
129 #define	LFS_PARTIAL_FITS(fs) \
130 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
131 	fragstofsb((fs), (fs)->lfs_frag))
132 
133 int	 lfs_match_fake(struct lfs *, struct buf *);
134 void	 lfs_newseg(struct lfs *);
135 /* XXX ondisk32 */
136 void	 lfs_shellsort(struct buf **, int32_t *, int, int);
137 void	 lfs_supercallback(struct buf *);
138 void	 lfs_updatemeta(struct segment *);
139 void	 lfs_writesuper(struct lfs *, daddr_t);
140 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
141 	    struct segment *sp, int dirops);
142 
143 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
144 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
145 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
146 int	lfs_dirvcount = 0;		/* # active dirops */
147 
148 /* Statistics Counters */
149 int lfs_dostats = 1;
150 struct lfs_stats lfs_stats;
151 
152 /* op values to lfs_writevnodes */
153 #define	VN_REG		0
154 #define	VN_DIROP	1
155 #define	VN_EMPTY	2
156 #define VN_CLEAN	3
157 
158 /*
159  * XXX KS - Set modification time on the Ifile, so the cleaner can
160  * read the fs mod time off of it.  We don't set IN_UPDATE here,
161  * since we don't really need this to be flushed to disk (and in any
162  * case that wouldn't happen to the Ifile until we checkpoint).
163  */
164 void
165 lfs_imtime(struct lfs *fs)
166 {
167 	struct timespec ts;
168 	struct inode *ip;
169 
170 	ASSERT_MAYBE_SEGLOCK(fs);
171 	TIMEVAL_TO_TIMESPEC(&time, &ts);
172 	ip = VTOI(fs->lfs_ivnode);
173 	ip->i_ffs1_mtime = ts.tv_sec;
174 	ip->i_ffs1_mtimensec = ts.tv_nsec;
175 }
176 
177 /*
178  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
179  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
180  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
181  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
182  */
183 
184 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
185 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
186 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
187 
188 int
189 lfs_vflush(struct vnode *vp)
190 {
191 	struct inode *ip;
192 	struct lfs *fs;
193 	struct segment *sp;
194 	struct buf *bp, *nbp, *tbp, *tnbp;
195 	int error, s;
196 	int flushed;
197 #if 0
198 	int redo;
199 #endif
200 
201 	ip = VTOI(vp);
202 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
203 
204 	ASSERT_NO_SEGLOCK(fs);
205 	if (ip->i_flag & IN_CLEANING) {
206 		ivndebug(vp,"vflush/in_cleaning");
207 		LFS_CLR_UINO(ip, IN_CLEANING);
208 		LFS_SET_UINO(ip, IN_MODIFIED);
209 
210 		/*
211 		 * Toss any cleaning buffers that have real counterparts
212 		 * to avoid losing new data.
213 		 */
214 		s = splbio();
215 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
216 			nbp = LIST_NEXT(bp, b_vnbufs);
217 			if (!LFS_IS_MALLOC_BUF(bp))
218 				continue;
219 			/*
220 			 * Look for pages matching the range covered
221 			 * by cleaning blocks.  It's okay if more dirty
222 			 * pages appear, so long as none disappear out
223 			 * from under us.
224 			 */
225 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
226 			    vp != fs->lfs_ivnode) {
227 				struct vm_page *pg;
228 				voff_t off;
229 
230 				simple_lock(&vp->v_interlock);
231 				for (off = lblktosize(fs, bp->b_lblkno);
232 				     off < lblktosize(fs, bp->b_lblkno + 1);
233 				     off += PAGE_SIZE) {
234 					pg = uvm_pagelookup(&vp->v_uobj, off);
235 					if (pg == NULL)
236 						continue;
237 					if ((pg->flags & PG_CLEAN) == 0 ||
238 					    pmap_is_modified(pg)) {
239 						fs->lfs_avail += btofsb(fs,
240 							bp->b_bcount);
241 						wakeup(&fs->lfs_avail);
242 						lfs_freebuf(fs, bp);
243 						bp = NULL;
244 						goto nextbp;
245 					}
246 				}
247 				simple_unlock(&vp->v_interlock);
248 			}
249 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
250 			    tbp = tnbp)
251 			{
252 				tnbp = LIST_NEXT(tbp, b_vnbufs);
253 				if (tbp->b_vp == bp->b_vp
254 				   && tbp->b_lblkno == bp->b_lblkno
255 				   && tbp != bp)
256 				{
257 					fs->lfs_avail += btofsb(fs,
258 						bp->b_bcount);
259 					wakeup(&fs->lfs_avail);
260 					lfs_freebuf(fs, bp);
261 					bp = NULL;
262 					break;
263 				}
264 			}
265 		    nextbp:
266 			;
267 		}
268 		splx(s);
269 	}
270 
271 	/* If the node is being written, wait until that is done */
272 	simple_lock(&vp->v_interlock);
273 	s = splbio();
274 	if (WRITEINPROG(vp)) {
275 		ivndebug(vp,"vflush/writeinprog");
276 		ltsleep(vp, (PRIBIO+1), "lfs_vw", 0, &vp->v_interlock);
277 	}
278 	splx(s);
279 	simple_unlock(&vp->v_interlock);
280 
281 	/* Protect against VXLOCK deadlock in vinvalbuf() */
282 	lfs_seglock(fs, SEGM_SYNC);
283 
284 	/* If we're supposed to flush a freed inode, just toss it */
285 	/* XXX - seglock, so these buffers can't be gathered, right? */
286 	if (ip->i_mode == 0) {
287 		DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n",
288 		      ip->i_number));
289 		s = splbio();
290 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
291 			nbp = LIST_NEXT(bp, b_vnbufs);
292 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
293 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
294 				wakeup(&fs->lfs_avail);
295 			}
296 			/* Copied from lfs_writeseg */
297 			if (bp->b_flags & B_CALL) {
298 				biodone(bp);
299 			} else {
300 				bremfree(bp);
301 				LFS_UNLOCK_BUF(bp);
302 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
303 					 B_GATHERED);
304 				bp->b_flags |= B_DONE;
305 				reassignbuf(bp, vp);
306 				brelse(bp);
307 			}
308 		}
309 		splx(s);
310 		LFS_CLR_UINO(ip, IN_CLEANING);
311 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
312 		ip->i_flag &= ~IN_ALLMOD;
313 		DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n",
314 		      ip->i_number));
315 		lfs_segunlock(fs);
316 		return 0;
317 	}
318 
319 	SET_FLUSHING(fs,vp);
320 	if (fs->lfs_nactive > LFS_MAX_ACTIVE ||
321 	    (fs->lfs_sp->seg_flags & SEGM_CKP)) {
322 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
323 		CLR_FLUSHING(fs,vp);
324 		lfs_segunlock(fs);
325 		return error;
326 	}
327 	sp = fs->lfs_sp;
328 
329 	flushed = 0;
330 	if (VPISEMPTY(vp)) {
331 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
332 		++flushed;
333 	} else if ((ip->i_flag & IN_CLEANING) &&
334 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
335 		ivndebug(vp,"vflush/clean");
336 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
337 		++flushed;
338 	} else if (lfs_dostats) {
339 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
340 			++lfs_stats.vflush_invoked;
341 		ivndebug(vp,"vflush");
342 	}
343 
344 #ifdef DIAGNOSTIC
345 	if (vp->v_flag & VDIROP) {
346 		DLOG((DLOG_VNODE, "lfs_vflush: flushing VDIROP\n"));
347 		/* panic("lfs_vflush: VDIROP being flushed...this can\'t happen"); */
348 	}
349 	if (vp->v_usecount < 0) {
350 		printf("usecount=%ld\n", (long)vp->v_usecount);
351 		panic("lfs_vflush: usecount<0");
352 	}
353 #endif
354 
355 #if 1
356 	do {
357 		do {
358 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
359 				lfs_writefile(fs, sp, vp);
360 		} while (lfs_writeinode(fs, sp, ip));
361 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
362 #else
363 	if (flushed && vp != fs->lfs_ivnode)
364 		lfs_writeseg(fs, sp);
365 	else do {
366 		simple_lock(&fs->lfs_interlock);
367 		fs->lfs_flags &= ~LFS_IFDIRTY;
368 		simple_unlock(&fs->lfs_interlock);
369 		lfs_writefile(fs, sp, vp);
370 		redo = lfs_writeinode(fs, sp, ip);
371 		redo += lfs_writeseg(fs, sp);
372 		simple_lock(&fs->lfs_interlock);
373 		redo += (fs->lfs_flags & LFS_IFDIRTY);
374 		simple_unlock(&fs->lfs_interlock);
375 	} while (redo && vp == fs->lfs_ivnode);
376 #endif
377 	if (lfs_dostats) {
378 		++lfs_stats.nwrites;
379 		if (sp->seg_flags & SEGM_SYNC)
380 			++lfs_stats.nsync_writes;
381 		if (sp->seg_flags & SEGM_CKP)
382 			++lfs_stats.ncheckpoints;
383 	}
384 	/*
385 	 * If we were called from somewhere that has already held the seglock
386 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
387 	 * the write to complete because we are still locked.
388 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
389 	 * we must explicitly wait, if that is the case.
390 	 *
391 	 * We compare the iocount against 1, not 0, because it is
392 	 * artificially incremented by lfs_seglock().
393 	 */
394 	simple_lock(&fs->lfs_interlock);
395 	if (fs->lfs_seglock > 1) {
396 		while (fs->lfs_iocount > 1)
397 			(void)ltsleep(&fs->lfs_iocount, PRIBIO + 1,
398 				     "lfs_vflush", 0, &fs->lfs_interlock);
399 	}
400 	simple_unlock(&fs->lfs_interlock);
401 
402 	lfs_segunlock(fs);
403 
404 	/* Wait for these buffers to be recovered by aiodoned */
405 	s = splbio();
406 	simple_lock(&global_v_numoutput_slock);
407 	while (vp->v_numoutput > 0) {
408 		vp->v_flag |= VBWAIT;
409 		ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf2", 0,
410 			&global_v_numoutput_slock);
411 	}
412 	simple_unlock(&global_v_numoutput_slock);
413 	splx(s);
414 
415 	CLR_FLUSHING(fs,vp);
416 	return (0);
417 }
418 
419 int
420 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
421 {
422 	struct inode *ip;
423 	struct vnode *vp, *nvp;
424 	int inodes_written = 0, only_cleaning;
425 
426 	ASSERT_SEGLOCK(fs);
427 #ifndef LFS_NO_BACKVP_HACK
428 	/* BEGIN HACK */
429 #define	VN_OFFSET	\
430 	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
431 #define	BACK_VP(VP)	\
432 	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
433 #define	BEG_OF_VLIST	\
434 	((struct vnode *)(((caddr_t)&LIST_FIRST(&mp->mnt_vnodelist)) \
435 	- VN_OFFSET))
436 
437 	/* Find last vnode. */
438  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
439 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
440 	     vp = LIST_NEXT(vp, v_mntvnodes));
441 	for (; vp && vp != BEG_OF_VLIST; vp = nvp) {
442 		nvp = BACK_VP(vp);
443 #else
444 	loop:
445 	for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) {
446 		nvp = LIST_NEXT(vp, v_mntvnodes);
447 #endif
448 		/*
449 		 * If the vnode that we are about to sync is no longer
450 		 * associated with this mount point, start over.
451 		 */
452 		if (vp->v_mount != mp) {
453 			DLOG((DLOG_VNODE, "lfs_writevnodes: starting over\n"));
454 			/*
455 			 * After this, pages might be busy
456 			 * due to our own previous putpages.
457 			 * Start actual segment write here to avoid deadlock.
458 			 */
459 			(void)lfs_writeseg(fs, sp);
460 			goto loop;
461 		}
462 
463 		if (vp->v_type == VNON) {
464 			continue;
465 		}
466 
467 		ip = VTOI(vp);
468 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
469 		    (op != VN_DIROP && op != VN_CLEAN &&
470 		    (vp->v_flag & VDIROP))) {
471 			vndebug(vp,"dirop");
472 			continue;
473 		}
474 
475 		if (op == VN_EMPTY && !VPISEMPTY(vp)) {
476 			vndebug(vp,"empty");
477 			continue;
478 		}
479 
480 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
481 		   && vp != fs->lfs_flushvp
482 		   && !(ip->i_flag & IN_CLEANING)) {
483 			vndebug(vp,"cleaning");
484 			continue;
485 		}
486 
487 		if (lfs_vref(vp)) {
488 			vndebug(vp,"vref");
489 			continue;
490 		}
491 
492 		only_cleaning = 0;
493 		/*
494 		 * Write the inode/file if dirty and it's not the IFILE.
495 		 */
496 		if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) {
497 			only_cleaning =
498 			    ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
499 
500 			if (ip->i_number != LFS_IFILE_INUM) {
501 				lfs_writefile(fs, sp, vp);
502 				if (!VPISEMPTY(vp)) {
503 					if (WRITEINPROG(vp)) {
504 						ivndebug(vp,"writevnodes/write2");
505 					} else if (!(ip->i_flag & IN_ALLMOD)) {
506 						LFS_SET_UINO(ip, IN_MODIFIED);
507 					}
508 				}
509 				(void) lfs_writeinode(fs, sp, ip);
510 				inodes_written++;
511 			}
512 		}
513 
514 		if (lfs_clean_vnhead && only_cleaning)
515 			lfs_vunref_head(vp);
516 		else
517 			lfs_vunref(vp);
518 	}
519 	return inodes_written;
520 }
521 
522 /*
523  * Do a checkpoint.
524  */
525 int
526 lfs_segwrite(struct mount *mp, int flags)
527 {
528 	struct buf *bp;
529 	struct inode *ip;
530 	struct lfs *fs;
531 	struct segment *sp;
532 	struct vnode *vp;
533 	SEGUSE *segusep;
534 	int do_ckp, did_ckp, error, s;
535 	unsigned n, segleft, maxseg, sn, i, curseg;
536 	int writer_set = 0;
537 	int dirty;
538 	int redo;
539 
540 	fs = VFSTOUFS(mp)->um_lfs;
541 	ASSERT_MAYBE_SEGLOCK(fs);
542 
543 	if (fs->lfs_ronly)
544 		return EROFS;
545 
546 	lfs_imtime(fs);
547 
548 	/*
549 	 * Allocate a segment structure and enough space to hold pointers to
550 	 * the maximum possible number of buffers which can be described in a
551 	 * single summary block.
552 	 */
553 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
554 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
555 	sp = fs->lfs_sp;
556 
557 	/*
558 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
559 	 * in which case we have to flush *all* buffers off of this vnode.
560 	 * We don't care about other nodes, but write any non-dirop nodes
561 	 * anyway in anticipation of another getnewvnode().
562 	 *
563 	 * If we're cleaning we only write cleaning and ifile blocks, and
564 	 * no dirops, since otherwise we'd risk corruption in a crash.
565 	 */
566 	if (sp->seg_flags & SEGM_CLEAN)
567 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
568 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
569 		lfs_writevnodes(fs, mp, sp, VN_REG);
570 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
571 			error = lfs_writer_enter(fs, "lfs writer");
572 			if (error) {
573 				DLOG((DLOG_SEG, "segwrite mysterious error\n"));
574 				/* XXX why not segunlock? */
575 				pool_put(&fs->lfs_bpppool, sp->bpp);
576 				sp->bpp = NULL;
577 				pool_put(&fs->lfs_segpool, sp);
578 				sp = fs->lfs_sp = NULL;
579 				return (error);
580 			}
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 		segleft = fs->lfs_nseg;
593 		curseg = 0;
594 		for (n = 0; n < fs->lfs_segtabsz; n++) {
595 			dirty = 0;
596 			if (bread(fs->lfs_ivnode,
597 			    fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &bp))
598 				panic("lfs_segwrite: ifile read");
599 			segusep = (SEGUSE *)bp->b_data;
600 			maxseg = min(segleft, fs->lfs_sepb);
601 			for (i = 0; i < maxseg; i++) {
602 				sn = curseg + i;
603 				if (sn != dtosn(fs, fs->lfs_curseg) &&
604 				    segusep->su_flags & SEGUSE_ACTIVE) {
605 					segusep->su_flags &= ~SEGUSE_ACTIVE;
606 					--fs->lfs_nactive;
607 					++dirty;
608 				}
609 				fs->lfs_suflags[fs->lfs_activesb][sn] =
610 					segusep->su_flags;
611 				if (fs->lfs_version > 1)
612 					++segusep;
613 				else
614 					segusep = (SEGUSE *)
615 						((SEGUSE_V1 *)segusep + 1);
616 			}
617 
618 			if (dirty)
619 				error = LFS_BWRITE_LOG(bp); /* Ifile */
620 			else
621 				brelse(bp);
622 			segleft -= fs->lfs_sepb;
623 			curseg += fs->lfs_sepb;
624 		}
625 	}
626 
627 	LOCK_ASSERT(LFS_SEGLOCK_HELD(fs));
628 
629 	did_ckp = 0;
630 	if (do_ckp || fs->lfs_doifile) {
631 		vp = fs->lfs_ivnode;
632 		vn_lock(vp, LK_EXCLUSIVE);
633 		do {
634 #ifdef DEBUG
635 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid);
636 #endif
637 			simple_lock(&fs->lfs_interlock);
638 			fs->lfs_flags &= ~LFS_IFDIRTY;
639 			simple_unlock(&fs->lfs_interlock);
640 
641 			ip = VTOI(vp);
642 
643 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
644 				lfs_writefile(fs, sp, vp);
645 
646 			if (ip->i_flag & IN_ALLMOD)
647 				++did_ckp;
648 			redo = lfs_writeinode(fs, sp, ip);
649 			redo += lfs_writeseg(fs, sp);
650 			simple_lock(&fs->lfs_interlock);
651 			redo += (fs->lfs_flags & LFS_IFDIRTY);
652 			simple_unlock(&fs->lfs_interlock);
653 		} while (redo && do_ckp);
654 
655 		/*
656 		 * Unless we are unmounting, the Ifile may continue to have
657 		 * dirty blocks even after a checkpoint, due to changes to
658 		 * inodes' atime.  If we're checkpointing, it's "impossible"
659 		 * for other parts of the Ifile to be dirty after the loop
660 		 * above, since we hold the segment lock.
661 		 */
662 		s = splbio();
663 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
664 			LFS_CLR_UINO(ip, IN_ALLMOD);
665 		}
666 #ifdef DIAGNOSTIC
667 		else if (do_ckp) {
668 			int do_panic = 0;
669 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
670 				if (bp->b_lblkno < fs->lfs_cleansz +
671 				    fs->lfs_segtabsz &&
672 				    !(bp->b_flags & B_GATHERED)) {
673 					printf("ifile lbn %ld still dirty (flags %lx)\n",
674 						(long)bp->b_lblkno,
675 						(long)bp->b_flags);
676 					++do_panic;
677 				}
678 			}
679 			if (do_panic)
680 				panic("dirty blocks");
681 		}
682 #endif
683 		splx(s);
684 		VOP_UNLOCK(vp, 0);
685 	} else {
686 		(void) lfs_writeseg(fs, sp);
687 	}
688 
689 	/* Note Ifile no longer needs to be written */
690 	fs->lfs_doifile = 0;
691 	if (writer_set)
692 		lfs_writer_leave(fs);
693 
694 	/*
695 	 * If we didn't write the Ifile, we didn't really do anything.
696 	 * That means that (1) there is a checkpoint on disk and (2)
697 	 * nothing has changed since it was written.
698 	 *
699 	 * Take the flags off of the segment so that lfs_segunlock
700 	 * doesn't have to write the superblock either.
701 	 */
702 	if (do_ckp && !did_ckp) {
703 		sp->seg_flags &= ~SEGM_CKP;
704 	}
705 
706 	if (lfs_dostats) {
707 		++lfs_stats.nwrites;
708 		if (sp->seg_flags & SEGM_SYNC)
709 			++lfs_stats.nsync_writes;
710 		if (sp->seg_flags & SEGM_CKP)
711 			++lfs_stats.ncheckpoints;
712 	}
713 	lfs_segunlock(fs);
714 	return (0);
715 }
716 
717 /*
718  * Write the dirty blocks associated with a vnode.
719  */
720 void
721 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
722 {
723 	struct buf *bp;
724 	struct finfo *fip;
725 	struct inode *ip;
726 	IFILE *ifp;
727 	int i, frag;
728 
729 	ASSERT_SEGLOCK(fs);
730 	ip = VTOI(vp);
731 
732 	if (sp->seg_bytes_left < fs->lfs_bsize ||
733 	    sp->sum_bytes_left < sizeof(struct finfo))
734 		(void) lfs_writeseg(fs, sp);
735 
736 	sp->sum_bytes_left -= FINFOSIZE;
737 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
738 
739 	if (vp->v_flag & VDIROP)
740 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
741 
742 	fip = sp->fip;
743 	fip->fi_nblocks = 0;
744 	fip->fi_ino = ip->i_number;
745 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
746 	fip->fi_version = ifp->if_version;
747 	brelse(bp);
748 
749 	if (sp->seg_flags & SEGM_CLEAN) {
750 		lfs_gather(fs, sp, vp, lfs_match_fake);
751 		/*
752 		 * For a file being flushed, we need to write *all* blocks.
753 		 * This means writing the cleaning blocks first, and then
754 		 * immediately following with any non-cleaning blocks.
755 		 * The same is true of the Ifile since checkpoints assume
756 		 * that all valid Ifile blocks are written.
757 		 */
758 		if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode) {
759 			lfs_gather(fs, sp, vp, lfs_match_data);
760 			/*
761 			 * Don't call VOP_PUTPAGES: if we're flushing,
762 			 * we've already done it, and the Ifile doesn't
763 			 * use the page cache.
764 			 */
765 		}
766 	} else {
767 		lfs_gather(fs, sp, vp, lfs_match_data);
768 		/*
769 		 * If we're flushing, we've already called VOP_PUTPAGES
770 		 * so don't do it again.  Otherwise, we want to write
771 		 * everything we've got.
772 		 */
773 		if (!IS_FLUSHING(fs, vp)) {
774 			simple_lock(&vp->v_interlock);
775 			VOP_PUTPAGES(vp, 0, 0,
776 				     PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
777 		}
778 	}
779 
780 	/*
781 	 * It may not be necessary to write the meta-data blocks at this point,
782 	 * as the roll-forward recovery code should be able to reconstruct the
783 	 * list.
784 	 *
785 	 * We have to write them anyway, though, under two conditions: (1) the
786 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
787 	 * checkpointing.
788 	 *
789 	 * BUT if we are cleaning, we might have indirect blocks that refer to
790 	 * new blocks not being written yet, in addition to fragments being
791 	 * moved out of a cleaned segment.  If that is the case, don't
792 	 * write the indirect blocks, or the finfo will have a small block
793 	 * in the middle of it!
794 	 * XXX in this case isn't the inode size wrong too?
795 	 */
796 	frag = 0;
797 	if (sp->seg_flags & SEGM_CLEAN) {
798 		for (i = 0; i < NDADDR; i++)
799 			if (ip->i_lfs_fragsize[i] > 0 &&
800 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
801 				++frag;
802 	}
803 #ifdef DIAGNOSTIC
804 	if (frag > 1)
805 		panic("lfs_writefile: more than one fragment!");
806 #endif
807 	if (IS_FLUSHING(fs, vp) ||
808 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
809 		lfs_gather(fs, sp, vp, lfs_match_indir);
810 		lfs_gather(fs, sp, vp, lfs_match_dindir);
811 		lfs_gather(fs, sp, vp, lfs_match_tindir);
812 	}
813 	fip = sp->fip;
814 	if (fip->fi_nblocks != 0) {
815 		sp->fip = (FINFO*)((caddr_t)fip + FINFOSIZE +
816 				   sizeof(int32_t) * (fip->fi_nblocks));
817 		sp->start_lbp = &sp->fip->fi_blocks[0];
818 	} else {
819 		sp->sum_bytes_left += FINFOSIZE;
820 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
821 	}
822 }
823 
824 int
825 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
826 {
827 	struct buf *bp, *ibp;
828 	struct ufs1_dinode *cdp;
829 	IFILE *ifp;
830 	SEGUSE *sup;
831 	daddr_t daddr;
832 	int32_t *daddrp;	/* XXX ondisk32 */
833 	ino_t ino;
834 	int error, i, ndx, fsb = 0;
835 	int redo_ifile = 0;
836 	struct timespec ts;
837 	int gotblk = 0;
838 
839 	ASSERT_SEGLOCK(fs);
840 	if (!(ip->i_flag & IN_ALLMOD))
841 		return (0);
842 
843 	/* Allocate a new inode block if necessary. */
844 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
845 	    sp->ibp == NULL) {
846 		/* Allocate a new segment if necessary. */
847 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
848 		    sp->sum_bytes_left < sizeof(int32_t))
849 			(void) lfs_writeseg(fs, sp);
850 
851 		/* Get next inode block. */
852 		daddr = fs->lfs_offset;
853 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
854 		sp->ibp = *sp->cbpp++ =
855 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
856 			    fsbtodb(fs, daddr), fs->lfs_ibsize, 0, 0);
857 		gotblk++;
858 
859 		/* Zero out inode numbers */
860 		for (i = 0; i < INOPB(fs); ++i)
861 			((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber =
862 			    0;
863 
864 		++sp->start_bpp;
865 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
866 		/* Set remaining space counters. */
867 		sp->seg_bytes_left -= fs->lfs_ibsize;
868 		sp->sum_bytes_left -= sizeof(int32_t);
869 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
870 			sp->ninodes / INOPB(fs) - 1;
871 		((int32_t *)(sp->segsum))[ndx] = daddr;
872 	}
873 
874 	/* Update the inode times and copy the inode onto the inode page. */
875 	TIMEVAL_TO_TIMESPEC(&time, &ts);
876 	/* XXX kludge --- don't redirty the ifile just to put times on it */
877 	if (ip->i_number != LFS_IFILE_INUM)
878 		LFS_ITIMES(ip, &ts, &ts, &ts);
879 
880 	/*
881 	 * If this is the Ifile, and we've already written the Ifile in this
882 	 * partial segment, just overwrite it (it's not on disk yet) and
883 	 * continue.
884 	 *
885 	 * XXX we know that the bp that we get the second time around has
886 	 * already been gathered.
887 	 */
888 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
889 		*(sp->idp) = *ip->i_din.ffs1_din;
890 		ip->i_lfs_osize = ip->i_size;
891 		return 0;
892 	}
893 
894 	bp = sp->ibp;
895 	cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
896 	*cdp = *ip->i_din.ffs1_din;
897 
898 	/*
899 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
900 	 * addresses to disk; possibly change the on-disk record of
901 	 * the inode size, either by reverting to the previous size
902 	 * (in the case of cleaning) or by verifying the inode's block
903 	 * holdings (in the case of files being allocated as they are being
904 	 * written).
905 	 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail
906 	 * XXX count on disk wrong by the same amount.	We should be
907 	 * XXX able to "borrow" from lfs_avail and return it after the
908 	 * XXX Ifile is written.  See also in lfs_writeseg.
909 	 */
910 
911 	/* Check file size based on highest allocated block */
912 	if (((ip->i_ffs1_mode & IFMT) == IFREG ||
913 	     (ip->i_ffs1_mode & IFMT) == IFDIR) &&
914 	    ip->i_size > ((ip->i_lfs_hiblk + 1) << fs->lfs_bshift)) {
915 		cdp->di_size = (ip->i_lfs_hiblk + 1) << fs->lfs_bshift;
916 		DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %"
917 		      PRId64 "\n", (int)ip->i_number, ip->i_size, cdp->di_size));
918 	}
919 	if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) {
920 		if (ip->i_flags & IN_CLEANING)
921 			cdp->di_size = ip->i_lfs_osize;
922 		DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %d != nblk %d)"
923 		      " at %x\n", ip->i_number, ip->i_lfs_effnblks,
924 		      ip->i_ffs1_blocks, fs->lfs_offset));
925 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
926 		     daddrp++) {
927 			if (*daddrp == UNWRITTEN) {
928 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
929 				*daddrp = 0;
930 			}
931 		}
932 	} else {
933 		/* If all blocks are going to disk, update "size on disk" */
934 		ip->i_lfs_osize = ip->i_size;
935 	}
936 
937 #ifdef DIAGNOSTIC
938 	/*
939 	 * Check dinode held blocks against dinode size.
940 	 * This should be identical to the check in lfs_vget().
941 	 */
942 	for (i = (cdp->di_size + fs->lfs_bsize - 1) >> fs->lfs_bshift;
943 	     i < NDADDR; i++) {
944 		KASSERT(i >= 0);
945 		if ((cdp->di_mode & IFMT) == IFLNK)
946 			continue;
947 		if (((cdp->di_mode & IFMT) == IFBLK ||
948 		     (cdp->di_mode & IFMT) == IFCHR) && i == 0)
949 			continue;
950 		if (cdp->di_db[i] != 0) {
951 # ifdef DEBUG
952 			lfs_dump_dinode(cdp);
953 # endif
954 			panic("writing inconsistent inode");
955 		}
956 	}
957 #endif /* DIAGNOSTIC */
958 
959 	if (ip->i_flag & IN_CLEANING)
960 		LFS_CLR_UINO(ip, IN_CLEANING);
961 	else {
962 		/* XXX IN_ALLMOD */
963 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
964 			     IN_UPDATE | IN_MODIFY);
965 		if (ip->i_lfs_effnblks == ip->i_ffs1_blocks)
966 			LFS_CLR_UINO(ip, IN_MODIFIED);
967 		else
968 			DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real blks=%d, "
969 			      "eff=%d\n", ip->i_number, ip->i_ffs1_blocks,
970 			      ip->i_lfs_effnblks));
971 	}
972 
973 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
974 		sp->idp = ((struct ufs1_dinode *)bp->b_data) +
975 			(sp->ninodes % INOPB(fs));
976 	if (gotblk) {
977 		LFS_LOCK_BUF(bp);
978 		brelse(bp);
979 	}
980 
981 	/* Increment inode count in segment summary block. */
982 	++((SEGSUM *)(sp->segsum))->ss_ninos;
983 
984 	/* If this page is full, set flag to allocate a new page. */
985 	if (++sp->ninodes % INOPB(fs) == 0)
986 		sp->ibp = NULL;
987 
988 	/*
989 	 * If updating the ifile, update the super-block.  Update the disk
990 	 * address and access times for this inode in the ifile.
991 	 */
992 	ino = ip->i_number;
993 	if (ino == LFS_IFILE_INUM) {
994 		daddr = fs->lfs_idaddr;
995 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
996 	} else {
997 		LFS_IENTRY(ifp, fs, ino, ibp);
998 		daddr = ifp->if_daddr;
999 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
1000 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
1001 	}
1002 
1003 	/*
1004 	 * The inode's last address should not be in the current partial
1005 	 * segment, except under exceptional circumstances (lfs_writevnodes
1006 	 * had to start over, and in the meantime more blocks were written
1007 	 * to a vnode).	 Both inodes will be accounted to this segment
1008 	 * in lfs_writeseg so we need to subtract the earlier version
1009 	 * here anyway.	 The segment count can temporarily dip below
1010 	 * zero here; keep track of how many duplicates we have in
1011 	 * "dupino" so we don't panic below.
1012 	 */
1013 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno)) {
1014 		++sp->ndupino;
1015 		DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg "
1016 		      "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
1017 		      (long long)daddr, sp->ndupino));
1018 	}
1019 	/*
1020 	 * Account the inode: it no longer belongs to its former segment,
1021 	 * though it will not belong to the new segment until that segment
1022 	 * is actually written.
1023 	 */
1024 	if (daddr != LFS_UNUSED_DADDR) {
1025 		u_int32_t oldsn = dtosn(fs, daddr);
1026 #ifdef DIAGNOSTIC
1027 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
1028 #endif
1029 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1030 #ifdef DIAGNOSTIC
1031 		if (sup->su_nbytes +
1032 		    sizeof (struct ufs1_dinode) * ndupino
1033 		      < sizeof (struct ufs1_dinode)) {
1034 			printf("lfs_writeinode: negative bytes "
1035 			       "(segment %" PRIu32 " short by %d, "
1036 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
1037 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
1038 			       "ndupino=%d)\n",
1039 			       dtosn(fs, daddr),
1040 			       (int)sizeof (struct ufs1_dinode) *
1041 				   (1 - sp->ndupino) - sup->su_nbytes,
1042 			       oldsn, sp->seg_number, daddr,
1043 			       (unsigned int)sup->su_nbytes,
1044 			       sp->ndupino);
1045 			panic("lfs_writeinode: negative bytes");
1046 			sup->su_nbytes = sizeof (struct ufs1_dinode);
1047 		}
1048 #endif
1049 		DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n",
1050 		      dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino));
1051 		sup->su_nbytes -= sizeof (struct ufs1_dinode);
1052 		redo_ifile =
1053 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1054 		if (redo_ifile) {
1055 			simple_lock(&fs->lfs_interlock);
1056 			fs->lfs_flags |= LFS_IFDIRTY;
1057 			simple_unlock(&fs->lfs_interlock);
1058 		}
1059 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1060 	}
1061 	return (redo_ifile);
1062 }
1063 
1064 int
1065 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1066 {
1067 	struct lfs *fs;
1068 	int vers;
1069 	int j, blksinblk;
1070 
1071 	ASSERT_SEGLOCK(sp->fs);
1072 	/*
1073 	 * If full, finish this segment.  We may be doing I/O, so
1074 	 * release and reacquire the splbio().
1075 	 */
1076 #ifdef DIAGNOSTIC
1077 	if (sp->vp == NULL)
1078 		panic ("lfs_gatherblock: Null vp in segment");
1079 #endif
1080 	fs = sp->fs;
1081 	blksinblk = howmany(bp->b_bcount, fs->lfs_bsize);
1082 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1083 	    sp->seg_bytes_left < bp->b_bcount) {
1084 		if (sptr)
1085 			splx(*sptr);
1086 		lfs_updatemeta(sp);
1087 
1088 		vers = sp->fip->fi_version;
1089 		(void) lfs_writeseg(fs, sp);
1090 
1091 		sp->fip->fi_version = vers;
1092 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
1093 		/* Add the current file to the segment summary. */
1094 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
1095 		sp->sum_bytes_left -= FINFOSIZE;
1096 
1097 		if (sptr)
1098 			*sptr = splbio();
1099 		return (1);
1100 	}
1101 
1102 	if (bp->b_flags & B_GATHERED) {
1103 		DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %d,"
1104 		      " lbn %" PRId64 "\n",
1105 		      sp->fip->fi_ino, bp->b_lblkno));
1106 		return (0);
1107 	}
1108 
1109 	/* Insert into the buffer list, update the FINFO block. */
1110 	bp->b_flags |= B_GATHERED;
1111 
1112 	*sp->cbpp++ = bp;
1113 	for (j = 0; j < blksinblk; j++) {
1114 		sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j;
1115 		/* This block's accounting moves from lfs_favail to lfs_avail */
1116 		lfs_deregister_block(sp->vp, bp->b_lblkno + j);
1117 	}
1118 
1119 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1120 	sp->seg_bytes_left -= bp->b_bcount;
1121 	return (0);
1122 }
1123 
1124 int
1125 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1126     int (*match)(struct lfs *, struct buf *))
1127 {
1128 	struct buf *bp, *nbp;
1129 	int s, count = 0;
1130 
1131 	ASSERT_SEGLOCK(fs);
1132 	KASSERT(sp->vp == NULL);
1133 	sp->vp = vp;
1134 	s = splbio();
1135 
1136 #ifndef LFS_NO_BACKBUF_HACK
1137 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1138 # define	BUF_OFFSET	\
1139 	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1140 # define	BACK_BUF(BP)	\
1141 	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1142 # define	BEG_OF_LIST	\
1143 	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1144 
1145 loop:
1146 	/* Find last buffer. */
1147 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1148 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1149 	     bp = LIST_NEXT(bp, b_vnbufs))
1150 		/* nothing */;
1151 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1152 		nbp = BACK_BUF(bp);
1153 #else /* LFS_NO_BACKBUF_HACK */
1154 loop:
1155 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1156 		nbp = LIST_NEXT(bp, b_vnbufs);
1157 #endif /* LFS_NO_BACKBUF_HACK */
1158 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1159 #ifdef DEBUG
1160 			if (vp == fs->lfs_ivnode &&
1161 			    (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1162 				DLOG((DLOG_SEG, "lfs_gather: ifile lbn %"
1163 				      PRId64 " busy (%x)",
1164 				      bp->b_lblkno, bp->b_flags));
1165 #endif
1166 			continue;
1167 		}
1168 		if (vp->v_type == VBLK) {
1169 			/* For block devices, just write the blocks. */
1170 			/* XXX Do we even need to do this? */
1171 			/*
1172 			 * Get the block before bwrite,
1173 			 * so we don't corrupt the free list
1174 			 */
1175 			bp->b_flags |= B_BUSY;
1176 			bremfree(bp);
1177 			bwrite(bp);
1178 		} else {
1179 #ifdef DIAGNOSTIC
1180 # ifdef LFS_USE_B_INVAL
1181 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1182 				DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1183 				      " is B_INVAL\n", bp->b_lblkno));
1184 				VOP_PRINT(bp->b_vp);
1185 			}
1186 # endif /* LFS_USE_B_INVAL */
1187 			if (!(bp->b_flags & B_DELWRI))
1188 				panic("lfs_gather: bp not B_DELWRI");
1189 			if (!(bp->b_flags & B_LOCKED)) {
1190 				DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1191 				      " blk %" PRId64 " not B_LOCKED\n",
1192 				      bp->b_lblkno,
1193 				      dbtofsb(fs, bp->b_blkno)));
1194 				VOP_PRINT(bp->b_vp);
1195 				panic("lfs_gather: bp not B_LOCKED");
1196 			}
1197 #endif
1198 			if (lfs_gatherblock(sp, bp, &s)) {
1199 				goto loop;
1200 			}
1201 		}
1202 		count++;
1203 	}
1204 	splx(s);
1205 	lfs_updatemeta(sp);
1206 	KASSERT(sp->vp == vp);
1207 	sp->vp = NULL;
1208 	return count;
1209 }
1210 
1211 #if DEBUG
1212 # define DEBUG_OOFF(n) do {						\
1213 	if (ooff == 0) {						\
1214 		DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \
1215 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1216 			", was 0x0 (or %" PRId64 ")\n",			\
1217 			(n), ip->i_number, lbn, ndaddr, daddr));	\
1218 	}								\
1219 } while (0)
1220 #else
1221 # define DEBUG_OOFF(n)
1222 #endif
1223 
1224 /*
1225  * Change the given block's address to ndaddr, finding its previous
1226  * location using ufs_bmaparray().
1227  *
1228  * Account for this change in the segment table.
1229  *
1230  * called with sp == NULL by roll-forwarding code.
1231  */
1232 void
1233 lfs_update_single(struct lfs *fs, struct segment *sp, struct vnode *vp,
1234     daddr_t lbn, int32_t ndaddr, int size)
1235 {
1236 	SEGUSE *sup;
1237 	struct buf *bp;
1238 	struct indir a[NIADDR + 2], *ap;
1239 	struct inode *ip;
1240 	daddr_t daddr, ooff;
1241 	int num, error;
1242 	int bb, osize, obb;
1243 
1244 	ASSERT_SEGLOCK(fs);
1245 	KASSERT(sp == NULL || sp->vp == vp);
1246 	ip = VTOI(vp);
1247 
1248 	error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1249 	if (error)
1250 		panic("lfs_updatemeta: ufs_bmaparray returned %d", error);
1251 
1252 	daddr = (daddr_t)((int32_t)daddr); /* XXX ondisk32 */
1253 	KASSERT(daddr <= LFS_MAX_DADDR);
1254 	if (daddr > 0)
1255 		daddr = dbtofsb(fs, daddr);
1256 
1257 	bb = fragstofsb(fs, numfrags(fs, size));
1258 	switch (num) {
1259 	    case 0:
1260 		    ooff = ip->i_ffs1_db[lbn];
1261 		    DEBUG_OOFF(0);
1262 		    if (ooff == UNWRITTEN)
1263 			    ip->i_ffs1_blocks += bb;
1264 		    else {
1265 			    /* possible fragment truncation or extension */
1266 			    obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1267 			    ip->i_ffs1_blocks += (bb - obb);
1268 		    }
1269 		    ip->i_ffs1_db[lbn] = ndaddr;
1270 		    break;
1271 	    case 1:
1272 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1273 		    DEBUG_OOFF(1);
1274 		    if (ooff == UNWRITTEN)
1275 			    ip->i_ffs1_blocks += bb;
1276 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1277 		    break;
1278 	    default:
1279 		    ap = &a[num - 1];
1280 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1281 			    panic("lfs_updatemeta: bread bno %" PRId64,
1282 				  ap->in_lbn);
1283 
1284 		    /* XXX ondisk32 */
1285 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1286 		    DEBUG_OOFF(num);
1287 		    if (ooff == UNWRITTEN)
1288 			    ip->i_ffs1_blocks += bb;
1289 		    /* XXX ondisk32 */
1290 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1291 		    (void) VOP_BWRITE(bp);
1292 	}
1293 
1294 	KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr);
1295 
1296 	/* Update hiblk when extending the file */
1297 	if (lbn > ip->i_lfs_hiblk)
1298 		ip->i_lfs_hiblk = lbn;
1299 
1300 	/*
1301 	 * Though we'd rather it couldn't, this *can* happen right now
1302 	 * if cleaning blocks and regular blocks coexist.
1303 	 */
1304 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1305 
1306 	/*
1307 	 * Update segment usage information, based on old size
1308 	 * and location.
1309 	 */
1310 	if (daddr > 0) {
1311 		u_int32_t oldsn = dtosn(fs, daddr);
1312 #ifdef DIAGNOSTIC
1313 		int ndupino;
1314 
1315 		if (sp && sp->seg_number == oldsn) {
1316 			ndupino = sp->ndupino;
1317 		} else {
1318 			ndupino = 0;
1319 		}
1320 #endif
1321 		KASSERT(oldsn >= 0 && oldsn < fs->lfs_nseg);
1322 		if (lbn >= 0 && lbn < NDADDR)
1323 			osize = ip->i_lfs_fragsize[lbn];
1324 		else
1325 			osize = fs->lfs_bsize;
1326 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1327 #ifdef DIAGNOSTIC
1328 		if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino
1329 		    < osize) {
1330 			printf("lfs_updatemeta: negative bytes "
1331 			       "(segment %" PRIu32 " short by %" PRId64
1332 			       ")\n", dtosn(fs, daddr),
1333 			       (int64_t)osize -
1334 			       (sizeof (struct ufs1_dinode) * ndupino +
1335 				sup->su_nbytes));
1336 			printf("lfs_updatemeta: ino %d, lbn %" PRId64
1337 			       ", addr = 0x%" PRIx64 "\n",
1338 			       ip->i_number, lbn, daddr);
1339 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1340 			panic("lfs_updatemeta: negative bytes");
1341 			sup->su_nbytes = osize -
1342 			    sizeof (struct ufs1_dinode) * ndupino;
1343 		}
1344 #endif
1345 		DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1346 		      " db 0x%" PRIx64 "\n",
1347 		      dtosn(fs, daddr), osize,
1348 		      ip->i_number, lbn, daddr));
1349 		sup->su_nbytes -= osize;
1350 		if (!(bp->b_flags & B_GATHERED)) {
1351 			simple_lock(&fs->lfs_interlock);
1352 			fs->lfs_flags |= LFS_IFDIRTY;
1353 			simple_unlock(&fs->lfs_interlock);
1354 		}
1355 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1356 	}
1357 	/*
1358 	 * Now that this block has a new address, and its old
1359 	 * segment no longer owns it, we can forget about its
1360 	 * old size.
1361 	 */
1362 	if (lbn >= 0 && lbn < NDADDR)
1363 		ip->i_lfs_fragsize[lbn] = size;
1364 }
1365 
1366 /*
1367  * Update the metadata that points to the blocks listed in the FINFO
1368  * array.
1369  */
1370 void
1371 lfs_updatemeta(struct segment *sp)
1372 {
1373 	struct buf *sbp;
1374 	struct lfs *fs;
1375 	struct vnode *vp;
1376 	daddr_t lbn;
1377 	int i, nblocks, num;
1378 	int bb;
1379 	int bytesleft, size;
1380 
1381 	ASSERT_SEGLOCK(sp->fs);
1382 	vp = sp->vp;
1383 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1384 	KASSERT(nblocks >= 0);
1385 	KASSERT(vp != NULL);
1386 	if (nblocks == 0)
1387 		return;
1388 
1389 	/*
1390 	 * This count may be high due to oversize blocks from lfs_gop_write.
1391 	 * Correct for this. (XXX we should be able to keep track of these.)
1392 	 */
1393 	fs = sp->fs;
1394 	for (i = 0; i < nblocks; i++) {
1395 		if (sp->start_bpp[i] == NULL) {
1396 			DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks));
1397 			nblocks = i;
1398 			break;
1399 		}
1400 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1401 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1402 		nblocks -= num - 1;
1403 	}
1404 
1405 	KASSERT(vp->v_type == VREG ||
1406 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1407 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1408 
1409 	/*
1410 	 * Sort the blocks.
1411 	 *
1412 	 * We have to sort even if the blocks come from the
1413 	 * cleaner, because there might be other pending blocks on the
1414 	 * same inode...and if we don't sort, and there are fragments
1415 	 * present, blocks may be written in the wrong place.
1416 	 */
1417 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1418 
1419 	/*
1420 	 * Record the length of the last block in case it's a fragment.
1421 	 * If there are indirect blocks present, they sort last.  An
1422 	 * indirect block will be lfs_bsize and its presence indicates
1423 	 * that you cannot have fragments.
1424 	 *
1425 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1426 	 * XXX a later indirect block.	This will continue to be
1427 	 * XXX true until lfs_markv is fixed to do everything with
1428 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1429 	 */
1430 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1431 		fs->lfs_bmask) + 1;
1432 
1433 	/*
1434 	 * Assign disk addresses, and update references to the logical
1435 	 * block and the segment usage information.
1436 	 */
1437 	for (i = nblocks; i--; ++sp->start_bpp) {
1438 		sbp = *sp->start_bpp;
1439 		lbn = *sp->start_lbp;
1440 		KASSERT(sbp->b_lblkno == lbn);
1441 
1442 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1443 
1444 		/*
1445 		 * If we write a frag in the wrong place, the cleaner won't
1446 		 * be able to correctly identify its size later, and the
1447 		 * segment will be uncleanable.	 (Even worse, it will assume
1448 		 * that the indirect block that actually ends the list
1449 		 * is of a smaller size!)
1450 		 */
1451 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1452 			panic("lfs_updatemeta: fragment is not last block");
1453 
1454 		/*
1455 		 * For each subblock in this possibly oversized block,
1456 		 * update its address on disk.
1457 		 */
1458 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1459 		KASSERT(vp == sbp->b_vp);
1460 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1461 		     bytesleft -= fs->lfs_bsize) {
1462 			size = MIN(bytesleft, fs->lfs_bsize);
1463 			bb = fragstofsb(fs, numfrags(fs, size));
1464 			lbn = *sp->start_lbp++;
1465 			lfs_update_single(fs, sp, sp->vp, lbn, fs->lfs_offset,
1466 			    size);
1467 			fs->lfs_offset += bb;
1468 		}
1469 
1470 	}
1471 }
1472 
1473 /*
1474  * Move lfs_offset to a segment earlier than sn.
1475  */
1476 int
1477 lfs_rewind(struct lfs *fs, int newsn)
1478 {
1479 	int sn, osn, isdirty;
1480 	struct buf *bp;
1481 	SEGUSE *sup;
1482 
1483 	ASSERT_SEGLOCK(fs);
1484 
1485 	osn = dtosn(fs, fs->lfs_offset);
1486 	if (osn < newsn)
1487 		return 0;
1488 
1489 	/* lfs_avail eats the remaining space in this segment */
1490 	fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset - fs->lfs_curseg);
1491 
1492 	/* Find a low-numbered segment */
1493 	for (sn = 0; sn < fs->lfs_nseg; ++sn) {
1494 		LFS_SEGENTRY(sup, fs, sn, bp);
1495 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1496 		brelse(bp);
1497 
1498 		if (!isdirty)
1499 			break;
1500 	}
1501 	if (sn == fs->lfs_nseg)
1502 		panic("lfs_rewind: no clean segments");
1503 	if (sn >= newsn)
1504 		return ENOENT;
1505 	fs->lfs_nextseg = sn;
1506 	lfs_newseg(fs);
1507 	fs->lfs_offset = fs->lfs_curseg;
1508 
1509 	return 0;
1510 }
1511 
1512 /*
1513  * Start a new partial segment.
1514  *
1515  * Return 1 when we entered to a new segment.
1516  * Otherwise, return 0.
1517  */
1518 int
1519 lfs_initseg(struct lfs *fs)
1520 {
1521 	struct segment *sp = fs->lfs_sp;
1522 	SEGSUM *ssp;
1523 	struct buf *sbp;	/* buffer for SEGSUM */
1524 	int repeat = 0;		/* return value */
1525 
1526 	ASSERT_SEGLOCK(fs);
1527 	/* Advance to the next segment. */
1528 	if (!LFS_PARTIAL_FITS(fs)) {
1529 		SEGUSE *sup;
1530 		struct buf *bp;
1531 
1532 		/* lfs_avail eats the remaining space */
1533 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1534 						   fs->lfs_curseg);
1535 		/* Wake up any cleaning procs waiting on this file system. */
1536 		wakeup(&lfs_allclean_wakeup);
1537 		wakeup(&fs->lfs_nextseg);
1538 		lfs_newseg(fs);
1539 		repeat = 1;
1540 		fs->lfs_offset = fs->lfs_curseg;
1541 
1542 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1543 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1544 
1545 		/*
1546 		 * If the segment contains a superblock, update the offset
1547 		 * and summary address to skip over it.
1548 		 */
1549 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1550 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1551 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1552 			sp->seg_bytes_left -= LFS_SBPAD;
1553 		}
1554 		brelse(bp);
1555 		/* Segment zero could also contain the labelpad */
1556 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1557 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1558 			fs->lfs_offset +=
1559 			    btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1560 			sp->seg_bytes_left -=
1561 			    LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1562 		}
1563 	} else {
1564 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1565 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1566 				      (fs->lfs_offset - fs->lfs_curseg));
1567 	}
1568 	fs->lfs_lastpseg = fs->lfs_offset;
1569 
1570 	/* Record first address of this partial segment */
1571 	if (sp->seg_flags & SEGM_CLEAN) {
1572 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1573 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1574 			/* "1" is the artificial inc in lfs_seglock */
1575 			simple_lock(&fs->lfs_interlock);
1576 			while (fs->lfs_iocount > 1) {
1577 				ltsleep(&fs->lfs_iocount, PRIBIO + 1,
1578 				    "lfs_initseg", 0, &fs->lfs_interlock);
1579 			}
1580 			simple_unlock(&fs->lfs_interlock);
1581 			fs->lfs_cleanind = 0;
1582 		}
1583 	}
1584 
1585 	sp->fs = fs;
1586 	sp->ibp = NULL;
1587 	sp->idp = NULL;
1588 	sp->ninodes = 0;
1589 	sp->ndupino = 0;
1590 
1591 	sp->cbpp = sp->bpp;
1592 
1593 	/* Get a new buffer for SEGSUM */
1594 	sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1595 	    fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1596 
1597 	/* ... and enter it into the buffer list. */
1598 	*sp->cbpp = sbp;
1599 	sp->cbpp++;
1600 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1601 
1602 	sp->start_bpp = sp->cbpp;
1603 
1604 	/* Set point to SEGSUM, initialize it. */
1605 	ssp = sp->segsum = sbp->b_data;
1606 	memset(ssp, 0, fs->lfs_sumsize);
1607 	ssp->ss_next = fs->lfs_nextseg;
1608 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1609 	ssp->ss_magic = SS_MAGIC;
1610 
1611 	/* Set pointer to first FINFO, initialize it. */
1612 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1613 	sp->fip->fi_nblocks = 0;
1614 	sp->start_lbp = &sp->fip->fi_blocks[0];
1615 	sp->fip->fi_lastlength = 0;
1616 
1617 	sp->seg_bytes_left -= fs->lfs_sumsize;
1618 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1619 
1620 	return (repeat);
1621 }
1622 
1623 /*
1624  * Remove SEGUSE_INVAL from all segments.
1625  */
1626 void
1627 lfs_unset_inval_all(struct lfs *fs)
1628 {
1629 	SEGUSE *sup;
1630 	struct buf *bp;
1631 	int i;
1632 
1633 	for (i = 0; i < fs->lfs_nseg; i++) {
1634 		LFS_SEGENTRY(sup, fs, i, bp);
1635 		if (sup->su_flags & SEGUSE_INVAL) {
1636 			sup->su_flags &= ~SEGUSE_INVAL;
1637 			VOP_BWRITE(bp);
1638 		} else
1639 			brelse(bp);
1640 	}
1641 }
1642 
1643 /*
1644  * Return the next segment to write.
1645  */
1646 void
1647 lfs_newseg(struct lfs *fs)
1648 {
1649 	CLEANERINFO *cip;
1650 	SEGUSE *sup;
1651 	struct buf *bp;
1652 	int curseg, isdirty, sn, skip_inval;
1653 
1654 	ASSERT_SEGLOCK(fs);
1655 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1656 	DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n",
1657 	      dtosn(fs, fs->lfs_nextseg)));
1658 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1659 	sup->su_nbytes = 0;
1660 	sup->su_nsums = 0;
1661 	sup->su_ninos = 0;
1662 	LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1663 
1664 	LFS_CLEANERINFO(cip, fs, bp);
1665 	--cip->clean;
1666 	++cip->dirty;
1667 	fs->lfs_nclean = cip->clean;
1668 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1669 
1670 	fs->lfs_lastseg = fs->lfs_curseg;
1671 	fs->lfs_curseg = fs->lfs_nextseg;
1672 	skip_inval = 1;
1673 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1674 		sn = (sn + 1) % fs->lfs_nseg;
1675 		if (sn == curseg) {
1676 			if (skip_inval)
1677 				skip_inval = 0;
1678 			else
1679 				panic("lfs_nextseg: no clean segments");
1680 		}
1681 		LFS_SEGENTRY(sup, fs, sn, bp);
1682 		isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0));
1683 		/* Check SEGUSE_EMPTY as we go along */
1684 		if (isdirty && sup->su_nbytes == 0 &&
1685 		    !(sup->su_flags & SEGUSE_EMPTY))
1686 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1687 		else
1688 			brelse(bp);
1689 
1690 		if (!isdirty)
1691 			break;
1692 	}
1693 	if (skip_inval == 0)
1694 		lfs_unset_inval_all(fs);
1695 
1696 	++fs->lfs_nactive;
1697 	fs->lfs_nextseg = sntod(fs, sn);
1698 	if (lfs_dostats) {
1699 		++lfs_stats.segsused;
1700 	}
1701 }
1702 
1703 static struct buf *
1704 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1705 {
1706 	struct lfs_cluster *cl;
1707 	struct buf **bpp, *bp;
1708 	int s;
1709 
1710 	ASSERT_SEGLOCK(fs);
1711 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1712 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1713 	memset(cl, 0, sizeof(*cl));
1714 	cl->fs = fs;
1715 	cl->bpp = bpp;
1716 	cl->bufcount = 0;
1717 	cl->bufsize = 0;
1718 
1719 	/* If this segment is being written synchronously, note that */
1720 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1721 		cl->flags |= LFS_CL_SYNC;
1722 		cl->seg = fs->lfs_sp;
1723 		++cl->seg->seg_iocount;
1724 	}
1725 
1726 	/* Get an empty buffer header, or maybe one with something on it */
1727 	s = splbio();
1728 	bp = pool_get(&bufpool, PR_WAITOK); /* XXX should use lfs_malloc? */
1729 	splx(s);
1730 	memset(bp, 0, sizeof(*bp));
1731 	BUF_INIT(bp);
1732 
1733 	bp->b_flags = B_BUSY | B_CALL;
1734 	bp->b_dev = NODEV;
1735 	bp->b_blkno = bp->b_lblkno = addr;
1736 	bp->b_iodone = lfs_cluster_callback;
1737 	bp->b_private = cl;
1738 	bp->b_vp = vp;
1739 
1740 	return bp;
1741 }
1742 
1743 int
1744 lfs_writeseg(struct lfs *fs, struct segment *sp)
1745 {
1746 	struct buf **bpp, *bp, *cbp, *newbp;
1747 	SEGUSE *sup;
1748 	SEGSUM *ssp;
1749 	int i, s;
1750 	int do_again, nblocks, byteoffset;
1751 	size_t el_size;
1752 	struct lfs_cluster *cl;
1753 	u_short ninos;
1754 	struct vnode *devvp;
1755 	char *p = NULL;
1756 	struct vnode *vp;
1757 	int32_t *daddrp;	/* XXX ondisk32 */
1758 	int changed;
1759 	u_int32_t sum;
1760 
1761 	ASSERT_SEGLOCK(fs);
1762 	/*
1763 	 * If there are no buffers other than the segment summary to write
1764 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1765 	 * even if there aren't any buffers, you need to write the superblock.
1766 	 */
1767 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1768 		return (0);
1769 
1770 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1771 
1772 	/* Update the segment usage information. */
1773 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1774 
1775 	/* Loop through all blocks, except the segment summary. */
1776 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1777 		if ((*bpp)->b_vp != devvp) {
1778 			sup->su_nbytes += (*bpp)->b_bcount;
1779 			DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d"
1780 			      " lbn %" PRId64 " db 0x%" PRIx64 "\n",
1781 			      sp->seg_number, (*bpp)->b_bcount,
1782 			      VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
1783 			      (*bpp)->b_blkno));
1784 		}
1785 	}
1786 
1787 	ssp = (SEGSUM *)sp->segsum;
1788 
1789 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1790 	DLOG((DLOG_SU, "seg %d += %d for %d inodes\n",
1791 	      sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode),
1792 	      ssp->ss_ninos));
1793 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode);
1794 	/* sup->su_nbytes += fs->lfs_sumsize; */
1795 	if (fs->lfs_version == 1)
1796 		sup->su_olastmod = time.tv_sec;
1797 	else
1798 		sup->su_lastmod = time.tv_sec;
1799 	sup->su_ninos += ninos;
1800 	++sup->su_nsums;
1801 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1802 							 fs->lfs_ibsize));
1803 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1804 
1805 	do_again = !(bp->b_flags & B_GATHERED);
1806 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
1807 
1808 	/*
1809 	 * Mark blocks B_BUSY, to prevent then from being changed between
1810 	 * the checksum computation and the actual write.
1811 	 *
1812 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1813 	 * there are any, replace them with copies that have UNASSIGNED
1814 	 * instead.
1815 	 */
1816 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1817 		++bpp;
1818 		bp = *bpp;
1819 		if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */
1820 			bp->b_flags |= B_BUSY;
1821 			continue;
1822 		}
1823 
1824 		simple_lock(&bp->b_interlock);
1825 		s = splbio();
1826 		while (bp->b_flags & B_BUSY) {
1827 			DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential"
1828 			      " data summary corruption for ino %d, lbn %"
1829 			      PRId64 "\n",
1830 			      VTOI(bp->b_vp)->i_number, bp->b_lblkno));
1831 			bp->b_flags |= B_WANTED;
1832 			ltsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0,
1833 				&bp->b_interlock);
1834 			splx(s);
1835 			s = splbio();
1836 		}
1837 		bp->b_flags |= B_BUSY;
1838 		splx(s);
1839 		simple_unlock(&bp->b_interlock);
1840 
1841 		/*
1842 		 * Check and replace indirect block UNWRITTEN bogosity.
1843 		 * XXX See comment in lfs_writefile.
1844 		 */
1845 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1846 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
1847 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1848 			DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%d != %d)\n",
1849 			      VTOI(bp->b_vp)->i_number,
1850 			      VTOI(bp->b_vp)->i_lfs_effnblks,
1851 			      VTOI(bp->b_vp)->i_ffs1_blocks));
1852 			/* Make a copy we'll make changes to */
1853 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1854 					   bp->b_bcount, LFS_NB_IBLOCK);
1855 			newbp->b_blkno = bp->b_blkno;
1856 			memcpy(newbp->b_data, bp->b_data,
1857 			       newbp->b_bcount);
1858 
1859 			changed = 0;
1860 			/* XXX ondisk32 */
1861 			for (daddrp = (int32_t *)(newbp->b_data);
1862 			     daddrp < (int32_t *)(newbp->b_data +
1863 						  newbp->b_bcount); daddrp++) {
1864 				if (*daddrp == UNWRITTEN) {
1865 					++changed;
1866 					*daddrp = 0;
1867 				}
1868 			}
1869 			/*
1870 			 * Get rid of the old buffer.  Don't mark it clean,
1871 			 * though, if it still has dirty data on it.
1872 			 */
1873 			if (changed) {
1874 				DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):"
1875 				      " bp = %p newbp = %p\n", changed, bp,
1876 				      newbp));
1877 				*bpp = newbp;
1878 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1879 				if (bp->b_flags & B_CALL) {
1880 					DLOG((DLOG_SEG, "lfs_writeseg: "
1881 					      "indir bp should not be B_CALL\n"));
1882 					s = splbio();
1883 					biodone(bp);
1884 					splx(s);
1885 					bp = NULL;
1886 				} else {
1887 					/* Still on free list, leave it there */
1888 					s = splbio();
1889 					bp->b_flags &= ~B_BUSY;
1890 					if (bp->b_flags & B_WANTED)
1891 						wakeup(bp);
1892 					splx(s);
1893 					/*
1894 					 * We have to re-decrement lfs_avail
1895 					 * since this block is going to come
1896 					 * back around to us in the next
1897 					 * segment.
1898 					 */
1899 					fs->lfs_avail -=
1900 					    btofsb(fs, bp->b_bcount);
1901 				}
1902 			} else {
1903 				lfs_freebuf(fs, newbp);
1904 			}
1905 		}
1906 	}
1907 	/*
1908 	 * Compute checksum across data and then across summary; the first
1909 	 * block (the summary block) is skipped.  Set the create time here
1910 	 * so that it's guaranteed to be later than the inode mod times.
1911 	 */
1912 	sum = 0;
1913 	if (fs->lfs_version == 1)
1914 		el_size = sizeof(u_long);
1915 	else
1916 		el_size = sizeof(u_int32_t);
1917 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
1918 		++bpp;
1919 		/* Loop through gop_write cluster blocks */
1920 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
1921 		     byteoffset += fs->lfs_bsize) {
1922 #ifdef LFS_USE_B_INVAL
1923 			if (((*bpp)->b_flags & (B_CALL | B_INVAL)) ==
1924 			    (B_CALL | B_INVAL)) {
1925 				if (copyin((caddr_t)(*bpp)->b_saveaddr +
1926 					   byteoffset, dp, el_size)) {
1927 					panic("lfs_writeseg: copyin failed [1]:"
1928 						" ino %d blk %" PRId64,
1929 						VTOI((*bpp)->b_vp)->i_number,
1930 						(*bpp)->b_lblkno);
1931 				}
1932 			} else
1933 #endif /* LFS_USE_B_INVAL */
1934 			{
1935 				sum = lfs_cksum_part(
1936 				    (*bpp)->b_data + byteoffset, el_size, sum);
1937 			}
1938 		}
1939 	}
1940 	if (fs->lfs_version == 1)
1941 		ssp->ss_ocreate = time.tv_sec;
1942 	else {
1943 		ssp->ss_create = time.tv_sec;
1944 		ssp->ss_serial = ++fs->lfs_serial;
1945 		ssp->ss_ident  = fs->lfs_ident;
1946 	}
1947 	ssp->ss_datasum = lfs_cksum_fold(sum);
1948 	ssp->ss_sumsum = cksum(&ssp->ss_datasum,
1949 	    fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1950 
1951 	simple_lock(&fs->lfs_interlock);
1952 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1953 			  btofsb(fs, fs->lfs_sumsize));
1954 	simple_unlock(&fs->lfs_interlock);
1955 
1956 	/*
1957 	 * When we simply write the blocks we lose a rotation for every block
1958 	 * written.  To avoid this problem, we cluster the buffers into a
1959 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
1960 	 * devices can handle, use that for the size of the chunks.
1961 	 *
1962 	 * Blocks that are already clusters (from GOP_WRITE), however, we
1963 	 * don't bother to copy into other clusters.
1964 	 */
1965 
1966 #define CHUNKSIZE MAXPHYS
1967 
1968 	if (devvp == NULL)
1969 		panic("devvp is NULL");
1970 	for (bpp = sp->bpp, i = nblocks; i;) {
1971 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1972 		cl = cbp->b_private;
1973 
1974 		cbp->b_flags |= B_ASYNC | B_BUSY;
1975 		cbp->b_bcount = 0;
1976 
1977 #if defined(DEBUG) && defined(DIAGNOSTIC)
1978 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
1979 		    / sizeof(int32_t)) {
1980 			panic("lfs_writeseg: real bpp overwrite");
1981 		}
1982 		if (bpp - sp->bpp > segsize(fs) / fs->lfs_fsize) {
1983 			panic("lfs_writeseg: theoretical bpp overwrite");
1984 		}
1985 #endif
1986 
1987 		/*
1988 		 * Construct the cluster.
1989 		 */
1990 		simple_lock(&fs->lfs_interlock);
1991 		++fs->lfs_iocount;
1992 		simple_unlock(&fs->lfs_interlock);
1993 		while (i && cbp->b_bcount < CHUNKSIZE) {
1994 			bp = *bpp;
1995 
1996 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1997 				break;
1998 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
1999 				break;
2000 
2001 			/* Clusters from GOP_WRITE are expedited */
2002 			if (bp->b_bcount > fs->lfs_bsize) {
2003 				if (cbp->b_bcount > 0)
2004 					/* Put in its own buffer */
2005 					break;
2006 				else {
2007 					cbp->b_data = bp->b_data;
2008 				}
2009 			} else if (cbp->b_bcount == 0) {
2010 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
2011 							     LFS_NB_CLUSTER);
2012 				cl->flags |= LFS_CL_MALLOC;
2013 			}
2014 #ifdef DIAGNOSTIC
2015 			if (dtosn(fs, dbtofsb(fs, bp->b_blkno +
2016 					      btodb(bp->b_bcount - 1))) !=
2017 			    sp->seg_number) {
2018 				printf("blk size %d daddr %" PRIx64
2019 				    " not in seg %d\n",
2020 				    bp->b_bcount, bp->b_blkno,
2021 				    sp->seg_number);
2022 				panic("segment overwrite");
2023 			}
2024 #endif
2025 
2026 #ifdef LFS_USE_B_INVAL
2027 			/*
2028 			 * Fake buffers from the cleaner are marked as B_INVAL.
2029 			 * We need to copy the data from user space rather than
2030 			 * from the buffer indicated.
2031 			 * XXX == what do I do on an error?
2032 			 */
2033 			if ((bp->b_flags & (B_CALL|B_INVAL)) ==
2034 			    (B_CALL|B_INVAL)) {
2035 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
2036 					panic("lfs_writeseg: "
2037 					    "copyin failed [2]");
2038 			} else
2039 #endif /* LFS_USE_B_INVAL */
2040 			if (cl->flags & LFS_CL_MALLOC) {
2041 				/* copy data into our cluster. */
2042 				memcpy(p, bp->b_data, bp->b_bcount);
2043 				p += bp->b_bcount;
2044 			}
2045 
2046 			cbp->b_bcount += bp->b_bcount;
2047 			cl->bufsize += bp->b_bcount;
2048 
2049 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
2050 			cl->bpp[cl->bufcount++] = bp;
2051 			vp = bp->b_vp;
2052 			s = splbio();
2053 			reassignbuf(bp, vp);
2054 			V_INCR_NUMOUTPUT(vp);
2055 			splx(s);
2056 
2057 			bpp++;
2058 			i--;
2059 		}
2060 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2061 			BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL);
2062 		else
2063 			BIO_SETPRIO(cbp, BPRIO_TIMELIMITED);
2064 		s = splbio();
2065 		V_INCR_NUMOUTPUT(devvp);
2066 		splx(s);
2067 		VOP_STRATEGY(devvp, cbp);
2068 		curproc->p_stats->p_ru.ru_oublock++;
2069 	}
2070 
2071 	if (lfs_dostats) {
2072 		++lfs_stats.psegwrites;
2073 		lfs_stats.blocktot += nblocks - 1;
2074 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2075 			++lfs_stats.psyncwrites;
2076 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2077 			++lfs_stats.pcleanwrites;
2078 			lfs_stats.cleanblocks += nblocks - 1;
2079 		}
2080 	}
2081 	return (lfs_initseg(fs) || do_again);
2082 }
2083 
2084 void
2085 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2086 {
2087 	struct buf *bp;
2088 	int s;
2089 	struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2090 
2091 	ASSERT_MAYBE_SEGLOCK(fs);
2092 #ifdef DIAGNOSTIC
2093 	KASSERT(fs->lfs_magic == LFS_MAGIC);
2094 #endif
2095 	/*
2096 	 * If we can write one superblock while another is in
2097 	 * progress, we risk not having a complete checkpoint if we crash.
2098 	 * So, block here if a superblock write is in progress.
2099 	 */
2100 	simple_lock(&fs->lfs_interlock);
2101 	s = splbio();
2102 	while (fs->lfs_sbactive) {
2103 		ltsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0,
2104 			&fs->lfs_interlock);
2105 	}
2106 	fs->lfs_sbactive = daddr;
2107 	splx(s);
2108 	simple_unlock(&fs->lfs_interlock);
2109 
2110 	/* Set timestamp of this version of the superblock */
2111 	if (fs->lfs_version == 1)
2112 		fs->lfs_otstamp = time.tv_sec;
2113 	fs->lfs_tstamp = time.tv_sec;
2114 
2115 	/* Checksum the superblock and copy it into a buffer. */
2116 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2117 	bp = lfs_newbuf(fs, devvp,
2118 	    fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2119 	memset(bp->b_data + sizeof(struct dlfs), 0,
2120 	    LFS_SBPAD - sizeof(struct dlfs));
2121 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2122 
2123 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2124 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2125 	bp->b_iodone = lfs_supercallback;
2126 
2127 	if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC)
2128 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
2129 	else
2130 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
2131 	curproc->p_stats->p_ru.ru_oublock++;
2132 	s = splbio();
2133 	V_INCR_NUMOUTPUT(bp->b_vp);
2134 	splx(s);
2135 	simple_lock(&fs->lfs_interlock);
2136 	++fs->lfs_iocount;
2137 	simple_unlock(&fs->lfs_interlock);
2138 	VOP_STRATEGY(devvp, bp);
2139 }
2140 
2141 /*
2142  * Logical block number match routines used when traversing the dirty block
2143  * chain.
2144  */
2145 int
2146 lfs_match_fake(struct lfs *fs, struct buf *bp)
2147 {
2148 
2149 	ASSERT_SEGLOCK(fs);
2150 	return LFS_IS_MALLOC_BUF(bp);
2151 }
2152 
2153 #if 0
2154 int
2155 lfs_match_real(struct lfs *fs, struct buf *bp)
2156 {
2157 
2158 	ASSERT_SEGLOCK(fs);
2159 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2160 }
2161 #endif
2162 
2163 int
2164 lfs_match_data(struct lfs *fs, struct buf *bp)
2165 {
2166 
2167 	ASSERT_SEGLOCK(fs);
2168 	return (bp->b_lblkno >= 0);
2169 }
2170 
2171 int
2172 lfs_match_indir(struct lfs *fs, struct buf *bp)
2173 {
2174 	daddr_t lbn;
2175 
2176 	ASSERT_SEGLOCK(fs);
2177 	lbn = bp->b_lblkno;
2178 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2179 }
2180 
2181 int
2182 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2183 {
2184 	daddr_t lbn;
2185 
2186 	ASSERT_SEGLOCK(fs);
2187 	lbn = bp->b_lblkno;
2188 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2189 }
2190 
2191 int
2192 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2193 {
2194 	daddr_t lbn;
2195 
2196 	ASSERT_SEGLOCK(fs);
2197 	lbn = bp->b_lblkno;
2198 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2199 }
2200 
2201 /*
2202  * XXX - The only buffers that are going to hit these functions are the
2203  * segment write blocks, or the segment summaries, or the superblocks.
2204  *
2205  * All of the above are created by lfs_newbuf, and so do not need to be
2206  * released via brelse.
2207  */
2208 void
2209 lfs_callback(struct buf *bp)
2210 {
2211 	struct lfs *fs;
2212 
2213 	fs = bp->b_private;
2214 	ASSERT_NO_SEGLOCK(fs);
2215 	lfs_freebuf(fs, bp);
2216 }
2217 
2218 static void
2219 lfs_super_aiodone(struct buf *bp)
2220 {
2221 	struct lfs *fs;
2222 
2223 	fs = bp->b_private;
2224 	ASSERT_NO_SEGLOCK(fs);
2225 	simple_lock(&fs->lfs_interlock);
2226 	fs->lfs_sbactive = 0;
2227 	if (--fs->lfs_iocount <= 1)
2228 		wakeup(&fs->lfs_iocount);
2229 	simple_unlock(&fs->lfs_interlock);
2230 	wakeup(&fs->lfs_sbactive);
2231 	lfs_freebuf(fs, bp);
2232 }
2233 
2234 static void
2235 lfs_cluster_aiodone(struct buf *bp)
2236 {
2237 	struct lfs_cluster *cl;
2238 	struct lfs *fs;
2239 	struct buf *tbp, *fbp;
2240 	struct vnode *vp, *devvp;
2241 	struct inode *ip;
2242 	int s, error=0;
2243 
2244 	if (bp->b_flags & B_ERROR)
2245 		error = bp->b_error;
2246 
2247 	cl = bp->b_private;
2248 	fs = cl->fs;
2249 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2250 	ASSERT_NO_SEGLOCK(fs);
2251 
2252 	/* Put the pages back, and release the buffer */
2253 	while (cl->bufcount--) {
2254 		tbp = cl->bpp[cl->bufcount];
2255 		KASSERT(tbp->b_flags & B_BUSY);
2256 		if (error) {
2257 			tbp->b_flags |= B_ERROR;
2258 			tbp->b_error = error;
2259 		}
2260 
2261 		/*
2262 		 * We're done with tbp.	 If it has not been re-dirtied since
2263 		 * the cluster was written, free it.  Otherwise, keep it on
2264 		 * the locked list to be written again.
2265 		 */
2266 		vp = tbp->b_vp;
2267 
2268 		tbp->b_flags &= ~B_GATHERED;
2269 
2270 		LFS_BCLEAN_LOG(fs, tbp);
2271 
2272 		if (!(tbp->b_flags & B_CALL)) {
2273 			KASSERT(tbp->b_flags & B_LOCKED);
2274 			s = splbio();
2275 			simple_lock(&bqueue_slock);
2276 			bremfree(tbp);
2277 			simple_unlock(&bqueue_slock);
2278 			if (vp)
2279 				reassignbuf(tbp, vp);
2280 			splx(s);
2281 			tbp->b_flags |= B_ASYNC; /* for biodone */
2282 		}
2283 
2284 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2285 			LFS_UNLOCK_BUF(tbp);
2286 
2287 		if (tbp->b_flags & B_DONE) {
2288 			DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n",
2289 				cl->bufcount, (long)tbp->b_flags));
2290 		}
2291 
2292 		if ((tbp->b_flags & B_CALL) && !LFS_IS_MALLOC_BUF(tbp)) {
2293 			/*
2294 			 * A buffer from the page daemon.
2295 			 * We use the same iodone as it does,
2296 			 * so we must manually disassociate its
2297 			 * buffers from the vp.
2298 			 */
2299 			if (tbp->b_vp) {
2300 				/* This is just silly */
2301 				s = splbio();
2302 				brelvp(tbp);
2303 				tbp->b_vp = vp;
2304 				splx(s);
2305 			}
2306 			/* Put it back the way it was */
2307 			tbp->b_flags |= B_ASYNC;
2308 			/* Master buffers have B_AGE */
2309 			if (tbp->b_private == tbp)
2310 				tbp->b_flags |= B_AGE;
2311 		}
2312 		s = splbio();
2313 		biodone(tbp);
2314 
2315 		/*
2316 		 * If this is the last block for this vnode, but
2317 		 * there are other blocks on its dirty list,
2318 		 * set IN_MODIFIED/IN_CLEANING depending on what
2319 		 * sort of block.  Only do this for our mount point,
2320 		 * not for, e.g., inode blocks that are attached to
2321 		 * the devvp.
2322 		 * XXX KS - Shouldn't we set *both* if both types
2323 		 * of blocks are present (traverse the dirty list?)
2324 		 */
2325 		simple_lock(&global_v_numoutput_slock);
2326 		if (vp != devvp && vp->v_numoutput == 0 &&
2327 		    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2328 			ip = VTOI(vp);
2329 			DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n",
2330 			       ip->i_number));
2331 			if (LFS_IS_MALLOC_BUF(fbp))
2332 				LFS_SET_UINO(ip, IN_CLEANING);
2333 			else
2334 				LFS_SET_UINO(ip, IN_MODIFIED);
2335 		}
2336 		simple_unlock(&global_v_numoutput_slock);
2337 		splx(s);
2338 		wakeup(vp);
2339 	}
2340 
2341 	/* Fix up the cluster buffer, and release it */
2342 	if (cl->flags & LFS_CL_MALLOC)
2343 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2344 	s = splbio();
2345 	pool_put(&bufpool, bp); /* XXX should use lfs_free? */
2346 	splx(s);
2347 
2348 	/* Note i/o done */
2349 	if (cl->flags & LFS_CL_SYNC) {
2350 		if (--cl->seg->seg_iocount == 0)
2351 			wakeup(&cl->seg->seg_iocount);
2352 	}
2353 	simple_lock(&fs->lfs_interlock);
2354 #ifdef DIAGNOSTIC
2355 	if (fs->lfs_iocount == 0)
2356 		panic("lfs_cluster_aiodone: zero iocount");
2357 #endif
2358 	if (--fs->lfs_iocount <= 1)
2359 		wakeup(&fs->lfs_iocount);
2360 	simple_unlock(&fs->lfs_interlock);
2361 
2362 	pool_put(&fs->lfs_bpppool, cl->bpp);
2363 	cl->bpp = NULL;
2364 	pool_put(&fs->lfs_clpool, cl);
2365 }
2366 
2367 static void
2368 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2369 {
2370 	/* reset b_iodone for when this is a single-buf i/o. */
2371 	bp->b_iodone = aiodone;
2372 
2373 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
2374 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2375 	wakeup(&uvm.aiodoned);
2376 	simple_unlock(&uvm.aiodoned_lock);
2377 }
2378 
2379 static void
2380 lfs_cluster_callback(struct buf *bp)
2381 {
2382 
2383 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2384 }
2385 
2386 void
2387 lfs_supercallback(struct buf *bp)
2388 {
2389 
2390 	lfs_generic_callback(bp, lfs_super_aiodone);
2391 }
2392 
2393 /*
2394  * Shellsort (diminishing increment sort) from Data Structures and
2395  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2396  * see also Knuth Vol. 3, page 84.  The increments are selected from
2397  * formula (8), page 95.  Roughly O(N^3/2).
2398  */
2399 /*
2400  * This is our own private copy of shellsort because we want to sort
2401  * two parallel arrays (the array of buffer pointers and the array of
2402  * logical block numbers) simultaneously.  Note that we cast the array
2403  * of logical block numbers to a unsigned in this routine so that the
2404  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2405  */
2406 
2407 void
2408 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2409 {
2410 	static int __rsshell_increments[] = { 4, 1, 0 };
2411 	int incr, *incrp, t1, t2;
2412 	struct buf *bp_temp;
2413 
2414 #ifdef DEBUG
2415 	incr = 0;
2416 	for (t1 = 0; t1 < nmemb; t1++) {
2417 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2418 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2419 				/* dump before panic */
2420 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2421 				    nmemb, size);
2422 				incr = 0;
2423 				for (t1 = 0; t1 < nmemb; t1++) {
2424 					const struct buf *bp = bp_array[t1];
2425 
2426 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2427 					    PRIu64 "\n", t1,
2428 					    (uint64_t)bp->b_bcount,
2429 					    (uint64_t)bp->b_lblkno);
2430 					printf("lbns:");
2431 					for (t2 = 0; t2 * size < bp->b_bcount;
2432 					    t2++) {
2433 						printf(" %" PRId32,
2434 						    lb_array[incr++]);
2435 					}
2436 					printf("\n");
2437 				}
2438 				panic("lfs_shellsort: inconsistent input");
2439 			}
2440 		}
2441 	}
2442 #endif
2443 
2444 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2445 		for (t1 = incr; t1 < nmemb; ++t1)
2446 			for (t2 = t1 - incr; t2 >= 0;)
2447 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2448 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2449 					bp_temp = bp_array[t2];
2450 					bp_array[t2] = bp_array[t2 + incr];
2451 					bp_array[t2 + incr] = bp_temp;
2452 					t2 -= incr;
2453 				} else
2454 					break;
2455 
2456 	/* Reform the list of logical blocks */
2457 	incr = 0;
2458 	for (t1 = 0; t1 < nmemb; t1++) {
2459 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2460 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2461 		}
2462 	}
2463 }
2464 
2465 /*
2466  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2467  */
2468 int
2469 lfs_vref(struct vnode *vp)
2470 {
2471 	ASSERT_MAYBE_SEGLOCK(VTOI(vp)->i_lfs);
2472 	/*
2473 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2474 	 * being able to flush all of the pages from this vnode, which
2475 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2476 	 */
2477 	if (vp->v_flag & VXLOCK) {
2478 		if (IS_FLUSHING(VTOI(vp)->i_lfs, vp)) {
2479 			return 0;
2480 		}
2481 		return (1);
2482 	}
2483 	return (vget(vp, 0));
2484 }
2485 
2486 /*
2487  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2488  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2489  */
2490 void
2491 lfs_vunref(struct vnode *vp)
2492 {
2493 	ASSERT_MAYBE_SEGLOCK(VTOI(vp)->i_lfs);
2494 	/*
2495 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2496 	 */
2497 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs, vp)) {
2498 		return;
2499 	}
2500 
2501 	simple_lock(&vp->v_interlock);
2502 #ifdef DIAGNOSTIC
2503 	if (vp->v_usecount <= 0) {
2504 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2505 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2506 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2507 		panic("lfs_vunref: v_usecount < 0");
2508 	}
2509 #endif
2510 	vp->v_usecount--;
2511 	if (vp->v_usecount > 0) {
2512 		simple_unlock(&vp->v_interlock);
2513 		return;
2514 	}
2515 	/*
2516 	 * insert at tail of LRU list
2517 	 */
2518 	simple_lock(&vnode_free_list_slock);
2519 	if (vp->v_holdcnt > 0)
2520 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2521 	else
2522 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2523 	simple_unlock(&vnode_free_list_slock);
2524 	simple_unlock(&vp->v_interlock);
2525 }
2526 
2527 /*
2528  * We use this when we have vnodes that were loaded in solely for cleaning.
2529  * There is no reason to believe that these vnodes will be referenced again
2530  * soon, since the cleaning process is unrelated to normal filesystem
2531  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2532  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2533  * cleaning at the head of the list, instead.
2534  */
2535 void
2536 lfs_vunref_head(struct vnode *vp)
2537 {
2538 
2539 	ASSERT_SEGLOCK(VTOI(vp)->i_lfs);
2540 	simple_lock(&vp->v_interlock);
2541 #ifdef DIAGNOSTIC
2542 	if (vp->v_usecount == 0) {
2543 		panic("lfs_vunref: v_usecount<0");
2544 	}
2545 #endif
2546 	vp->v_usecount--;
2547 	if (vp->v_usecount > 0) {
2548 		simple_unlock(&vp->v_interlock);
2549 		return;
2550 	}
2551 	/*
2552 	 * insert at head of LRU list
2553 	 */
2554 	simple_lock(&vnode_free_list_slock);
2555 	if (vp->v_holdcnt > 0)
2556 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2557 	else
2558 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2559 	simple_unlock(&vnode_free_list_slock);
2560 	simple_unlock(&vp->v_interlock);
2561 }
2562 
2563