xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /*	$NetBSD: lfs_segment.c,v 1.277 2018/06/09 18:48:31 zafer 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  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*
32  * Copyright (c) 1991, 1993
33  *	The Regents of the University of California.  All rights reserved.
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. Neither the name of the University nor the names of its contributors
44  *    may be used to endorse or promote products derived from this software
45  *    without specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
60  */
61 
62 #include <sys/cdefs.h>
63 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.277 2018/06/09 18:48:31 zafer Exp $");
64 
65 #ifdef DEBUG
66 # define vndebug(vp, str) do {						\
67 	if (VTOI(vp)->i_state & IN_CLEANING)				\
68 		DLOG((DLOG_WVNODE, "not writing ino %d because %s (op %d)\n", \
69 		     VTOI(vp)->i_number, (str), op));			\
70 } while(0)
71 #else
72 # define vndebug(vp, str)
73 #endif
74 #define ivndebug(vp, str) \
75 	DLOG((DLOG_WVNODE, "ino %d: %s\n", VTOI(vp)->i_number, (str)))
76 
77 #if defined(_KERNEL_OPT)
78 #include "opt_ddb.h"
79 #endif
80 
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/namei.h>
84 #include <sys/kernel.h>
85 #include <sys/resourcevar.h>
86 #include <sys/file.h>
87 #include <sys/stat.h>
88 #include <sys/buf.h>
89 #include <sys/proc.h>
90 #include <sys/vnode.h>
91 #include <sys/mount.h>
92 #include <sys/kauth.h>
93 #include <sys/syslog.h>
94 
95 #include <miscfs/specfs/specdev.h>
96 #include <miscfs/fifofs/fifo.h>
97 
98 #include <ufs/lfs/ulfs_inode.h>
99 #include <ufs/lfs/ulfsmount.h>
100 #include <ufs/lfs/ulfs_extern.h>
101 
102 #include <ufs/lfs/lfs.h>
103 #include <ufs/lfs/lfs_accessors.h>
104 #include <ufs/lfs/lfs_kernel.h>
105 #include <ufs/lfs/lfs_extern.h>
106 
107 #include <uvm/uvm.h>
108 #include <uvm/uvm_extern.h>
109 
110 MALLOC_JUSTDEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
111 
112 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
113 static void lfs_free_aiodone(struct buf *);
114 static void lfs_super_aiodone(struct buf *);
115 static void lfs_cluster_aiodone(struct buf *);
116 static void lfs_cluster_callback(struct buf *);
117 
118 /*
119  * Determine if it's OK to start a partial in this segment, or if we need
120  * to go on to a new segment.
121  */
122 #define	LFS_PARTIAL_FITS(fs) \
123 	(lfs_sb_getfsbpseg(fs) - \
124 	    (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)) > \
125 	lfs_sb_getfrag(fs))
126 
127 /*
128  * Figure out whether we should do a checkpoint write or go ahead with
129  * an ordinary write.
130  */
131 #define LFS_SHOULD_CHECKPOINT(fs, flags) \
132         ((flags & SEGM_CLEAN) == 0 &&					\
133 	  ((fs->lfs_nactive > LFS_MAX_ACTIVE ||				\
134 	    (flags & SEGM_CKP) ||					\
135 	    lfs_sb_getnclean(fs) < LFS_MAX_ACTIVE)))
136 
137 int	 lfs_match_fake(struct lfs *, struct buf *);
138 void	 lfs_newseg(struct lfs *);
139 void	 lfs_supercallback(struct buf *);
140 void	 lfs_updatemeta(struct segment *);
141 void	 lfs_writesuper(struct lfs *, daddr_t);
142 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
143 	    struct segment *sp, int dirops);
144 
145 static void lfs_shellsort(struct lfs *, struct buf **, union lfs_blocks *,
146 			  int, int);
147 
148 kcondvar_t	lfs_allclean_wakeup;	/* Cleaner wakeup address. */
149 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
150 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
151 int	lfs_dirvcount = 0;		/* # active dirops */
152 
153 /* Statistics Counters */
154 int lfs_dostats = 1;
155 struct lfs_stats lfs_stats;
156 
157 /* op values to lfs_writevnodes */
158 #define	VN_REG		0
159 #define	VN_DIROP	1
160 #define	VN_EMPTY	2
161 #define VN_CLEAN	3
162 
163 /*
164  * XXX KS - Set modification time on the Ifile, so the cleaner can
165  * read the fs mod time off of it.  We don't set IN_UPDATE here,
166  * since we don't really need this to be flushed to disk (and in any
167  * case that wouldn't happen to the Ifile until we checkpoint).
168  */
169 void
170 lfs_imtime(struct lfs *fs)
171 {
172 	struct timespec ts;
173 	struct inode *ip;
174 
175 	ASSERT_MAYBE_SEGLOCK(fs);
176 	vfs_timestamp(&ts);
177 	ip = VTOI(fs->lfs_ivnode);
178 	lfs_dino_setmtime(fs, ip->i_din, ts.tv_sec);
179 	lfs_dino_setmtimensec(fs, ip->i_din, ts.tv_nsec);
180 }
181 
182 /*
183  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
184  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
185  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
186  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
187  */
188 
189 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
190 
191 int
192 lfs_vflush(struct vnode *vp)
193 {
194 	struct inode *ip;
195 	struct lfs *fs;
196 	struct segment *sp;
197 	struct buf *bp, *nbp, *tbp, *tnbp;
198 	int error;
199 	int flushed;
200 	int relock;
201 
202 	ip = VTOI(vp);
203 	fs = VFSTOULFS(vp->v_mount)->um_lfs;
204 	relock = 0;
205 
206     top:
207 	KASSERT(mutex_owned(vp->v_interlock) == false);
208 	KASSERT(mutex_owned(&lfs_lock) == false);
209 	KASSERT(mutex_owned(&bufcache_lock) == false);
210 	ASSERT_NO_SEGLOCK(fs);
211 	if (ip->i_state & IN_CLEANING) {
212 		ivndebug(vp,"vflush/in_cleaning");
213 		mutex_enter(&lfs_lock);
214 		LFS_CLR_UINO(ip, IN_CLEANING);
215 		LFS_SET_UINO(ip, IN_MODIFIED);
216 		mutex_exit(&lfs_lock);
217 
218 		/*
219 		 * Toss any cleaning buffers that have real counterparts
220 		 * to avoid losing new data.
221 		 */
222 		mutex_enter(vp->v_interlock);
223 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
224 			nbp = LIST_NEXT(bp, b_vnbufs);
225 			if (!LFS_IS_MALLOC_BUF(bp))
226 				continue;
227 			/*
228 			 * Look for pages matching the range covered
229 			 * by cleaning blocks.  It's okay if more dirty
230 			 * pages appear, so long as none disappear out
231 			 * from under us.
232 			 */
233 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
234 			    vp != fs->lfs_ivnode) {
235 				struct vm_page *pg;
236 				voff_t off;
237 
238 				for (off = lfs_lblktosize(fs, bp->b_lblkno);
239 				     off < lfs_lblktosize(fs, bp->b_lblkno + 1);
240 				     off += PAGE_SIZE) {
241 					pg = uvm_pagelookup(&vp->v_uobj, off);
242 					if (pg == NULL)
243 						continue;
244 					if ((pg->flags & PG_CLEAN) == 0 ||
245 					    pmap_is_modified(pg)) {
246 						lfs_sb_addavail(fs,
247 							lfs_btofsb(fs,
248 								bp->b_bcount));
249 						wakeup(&fs->lfs_availsleep);
250 						mutex_exit(vp->v_interlock);
251 						lfs_freebuf(fs, bp);
252 						mutex_enter(vp->v_interlock);
253 						bp = NULL;
254 						break;
255 					}
256 				}
257 			}
258 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
259 			    tbp = tnbp)
260 			{
261 				tnbp = LIST_NEXT(tbp, b_vnbufs);
262 				if (tbp->b_vp == bp->b_vp
263 				   && tbp->b_lblkno == bp->b_lblkno
264 				   && tbp != bp)
265 				{
266 					lfs_sb_addavail(fs, lfs_btofsb(fs,
267 						bp->b_bcount));
268 					wakeup(&fs->lfs_availsleep);
269 					mutex_exit(vp->v_interlock);
270 					lfs_freebuf(fs, bp);
271 					mutex_enter(vp->v_interlock);
272 					bp = NULL;
273 					break;
274 				}
275 			}
276 		}
277 	} else {
278 		mutex_enter(vp->v_interlock);
279 	}
280 
281 	/* If the node is being written, wait until that is done */
282 	while (WRITEINPROG(vp)) {
283 		ivndebug(vp,"vflush/writeinprog");
284 		cv_wait(&vp->v_cv, vp->v_interlock);
285 	}
286 	error = vdead_check(vp, VDEAD_NOWAIT);
287 	mutex_exit(vp->v_interlock);
288 
289 	/* Protect against deadlock in vinvalbuf() */
290 	lfs_seglock(fs, SEGM_SYNC | ((error != 0) ? SEGM_RECLAIM : 0));
291 	if (error != 0) {
292 		fs->lfs_reclino = ip->i_number;
293 	}
294 
295 	/* If we're supposed to flush a freed inode, just toss it */
296 	if (ip->i_lfs_iflags & LFSI_DELETED) {
297 		DLOG((DLOG_VNODE, "lfs_vflush: ino %d freed, not flushing\n",
298 		      ip->i_number));
299 		/* Drain v_numoutput */
300 		mutex_enter(vp->v_interlock);
301 		while (vp->v_numoutput > 0) {
302 			cv_wait(&vp->v_cv, vp->v_interlock);
303 		}
304 		KASSERT(vp->v_numoutput == 0);
305 		mutex_exit(vp->v_interlock);
306 
307 		mutex_enter(&bufcache_lock);
308 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
309 			nbp = LIST_NEXT(bp, b_vnbufs);
310 
311 			KASSERT((bp->b_flags & B_GATHERED) == 0);
312 			if (bp->b_oflags & BO_DELWRI) { /* XXX always true? */
313 				lfs_sb_addavail(fs, lfs_btofsb(fs, bp->b_bcount));
314 				wakeup(&fs->lfs_availsleep);
315 			}
316 			/* Copied from lfs_writeseg */
317 			if (bp->b_iodone != NULL) {
318 				mutex_exit(&bufcache_lock);
319 				biodone(bp);
320 				mutex_enter(&bufcache_lock);
321 			} else {
322 				bremfree(bp);
323 				LFS_UNLOCK_BUF(bp);
324 				mutex_enter(vp->v_interlock);
325 				bp->b_flags &= ~(B_READ | B_GATHERED);
326 				bp->b_oflags = (bp->b_oflags & ~BO_DELWRI) | BO_DONE;
327 				bp->b_error = 0;
328 				reassignbuf(bp, vp);
329 				mutex_exit(vp->v_interlock);
330 				brelse(bp, 0);
331 			}
332 		}
333 		mutex_exit(&bufcache_lock);
334 		LFS_CLR_UINO(ip, IN_CLEANING);
335 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
336 		ip->i_state &= ~IN_ALLMOD;
337 		DLOG((DLOG_VNODE, "lfs_vflush: done not flushing ino %d\n",
338 		      ip->i_number));
339 		lfs_segunlock(fs);
340 
341 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
342 
343 		return 0;
344 	}
345 
346 	fs->lfs_flushvp = vp;
347 	if (LFS_SHOULD_CHECKPOINT(fs, fs->lfs_sp->seg_flags)) {
348 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
349 		fs->lfs_flushvp = NULL;
350 		KASSERT(fs->lfs_flushvp_fakevref == 0);
351 		lfs_segunlock(fs);
352 
353 		/* Make sure that any pending buffers get written */
354 		mutex_enter(vp->v_interlock);
355 		while (vp->v_numoutput > 0) {
356 			cv_wait(&vp->v_cv, vp->v_interlock);
357 		}
358 		KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
359 		KASSERT(vp->v_numoutput == 0);
360 		mutex_exit(vp->v_interlock);
361 
362 		return error;
363 	}
364 	sp = fs->lfs_sp;
365 
366 	flushed = 0;
367 	if (VPISEMPTY(vp)) {
368 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
369 		++flushed;
370 	} else if ((ip->i_state & IN_CLEANING) &&
371 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
372 		ivndebug(vp,"vflush/clean");
373 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
374 		++flushed;
375 	} else if (lfs_dostats) {
376 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_state & IN_ALLMOD))
377 			++lfs_stats.vflush_invoked;
378 		ivndebug(vp,"vflush");
379 	}
380 
381 #ifdef DIAGNOSTIC
382 	if (vp->v_uflag & VU_DIROP) {
383 		DLOG((DLOG_VNODE, "lfs_vflush: flushing VU_DIROP\n"));
384 		/* panic("lfs_vflush: VU_DIROP being flushed...this can\'t happen"); */
385 	}
386 #endif
387 
388 	do {
389 #ifdef DEBUG
390 		int loopcount = 0;
391 #endif
392 		do {
393 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
394 				relock = lfs_writefile(fs, sp, vp);
395 				if (relock && vp != fs->lfs_ivnode) {
396 					/*
397 					 * Might have to wait for the
398 					 * cleaner to run; but we're
399 					 * still not done with this vnode.
400 					 * XXX we can do better than this.
401 					 */
402 					KDASSERT(ip->i_number != LFS_IFILE_INUM);
403 					lfs_writeinode(fs, sp, ip);
404 					mutex_enter(&lfs_lock);
405 					LFS_SET_UINO(ip, IN_MODIFIED);
406 					mutex_exit(&lfs_lock);
407 					lfs_writeseg(fs, sp);
408 					lfs_segunlock(fs);
409 					lfs_segunlock_relock(fs);
410 					goto top;
411 				}
412 			}
413 			/*
414 			 * If we begin a new segment in the middle of writing
415 			 * the Ifile, it creates an inconsistent checkpoint,
416 			 * since the Ifile information for the new segment
417 			 * is not up-to-date.  Take care of this here by
418 			 * sending the Ifile through again in case there
419 			 * are newly dirtied blocks.  But wait, there's more!
420 			 * This second Ifile write could *also* cross a segment
421 			 * boundary, if the first one was large.  The second
422 			 * one is guaranteed to be no more than 8 blocks,
423 			 * though (two segment blocks and supporting indirects)
424 			 * so the third write *will not* cross the boundary.
425 			 */
426 			if (vp == fs->lfs_ivnode) {
427 				lfs_writefile(fs, sp, vp);
428 				lfs_writefile(fs, sp, vp);
429 			}
430 #ifdef DEBUG
431 			if (++loopcount > 2)
432 				log(LOG_NOTICE, "lfs_vflush: looping count=%d\n", loopcount);
433 #endif
434 		} while (lfs_writeinode(fs, sp, ip));
435 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
436 
437 	if (lfs_dostats) {
438 		++lfs_stats.nwrites;
439 		if (sp->seg_flags & SEGM_SYNC)
440 			++lfs_stats.nsync_writes;
441 		if (sp->seg_flags & SEGM_CKP)
442 			++lfs_stats.ncheckpoints;
443 	}
444 	/*
445 	 * If we were called from somewhere that has already held the seglock
446 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
447 	 * the write to complete because we are still locked.
448 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
449 	 * we must explicitly wait, if that is the case.
450 	 *
451 	 * We compare the iocount against 1, not 0, because it is
452 	 * artificially incremented by lfs_seglock().
453 	 */
454 	mutex_enter(&lfs_lock);
455 	if (fs->lfs_seglock > 1) {
456 		while (fs->lfs_iocount > 1)
457 			(void)mtsleep(&fs->lfs_iocount, PRIBIO + 1,
458 				     "lfs_vflush", 0, &lfs_lock);
459 	}
460 	mutex_exit(&lfs_lock);
461 
462 	lfs_segunlock(fs);
463 
464 	/* Wait for these buffers to be recovered by aiodoned */
465 	mutex_enter(vp->v_interlock);
466 	while (vp->v_numoutput > 0) {
467 		cv_wait(&vp->v_cv, vp->v_interlock);
468 	}
469 	KASSERT(LIST_FIRST(&vp->v_dirtyblkhd) == NULL);
470 	KASSERT(vp->v_numoutput == 0);
471 	mutex_exit(vp->v_interlock);
472 
473 	fs->lfs_flushvp = NULL;
474 	KASSERT(fs->lfs_flushvp_fakevref == 0);
475 
476 	return (0);
477 }
478 
479 struct lfs_writevnodes_ctx {
480 	int op;
481 	struct lfs *fs;
482 };
483 static bool
484 lfs_writevnodes_selector(void *cl, struct vnode *vp)
485 {
486 	struct lfs_writevnodes_ctx *c = cl;
487 	struct inode *ip;
488 	int op = c->op;
489 
490 	KASSERT(mutex_owned(vp->v_interlock));
491 
492 	ip = VTOI(vp);
493 	if (ip == NULL || vp->v_type == VNON)
494 		return false;
495 	if ((op == VN_DIROP && !(vp->v_uflag & VU_DIROP)) ||
496 	    (op != VN_DIROP && op != VN_CLEAN && (vp->v_uflag & VU_DIROP))) {
497 		vndebug(vp, "dirop");
498 		return false;
499 	}
500 	if (op == VN_EMPTY && !VPISEMPTY(vp)) {
501 		vndebug(vp,"empty");
502 		return false;
503 	}
504 	if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM &&
505 	    vp != c->fs->lfs_flushvp && !(ip->i_state & IN_CLEANING)) {
506 		vndebug(vp,"cleaning");
507 		return false;
508 	}
509 	mutex_enter(&lfs_lock);
510 	if (vp == c->fs->lfs_unlockvp) {
511 		mutex_exit(&lfs_lock);
512 		return false;
513 	}
514 	mutex_exit(&lfs_lock);
515 
516 	return true;
517 }
518 
519 int
520 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
521 {
522 	struct inode *ip;
523 	struct vnode *vp;
524 	struct vnode_iterator *marker;
525 	struct lfs_writevnodes_ctx ctx;
526 	int inodes_written = 0;
527 	int error = 0;
528 
529 	/*
530 	 * XXX This was TAILQ_FOREACH_REVERSE on &mp->mnt_vnodelist.
531 	 * XXX The rationale is unclear, the initial commit had no information.
532 	 * XXX If the order really matters we have to sort the vnodes first.
533 	*/
534 
535 	ASSERT_SEGLOCK(fs);
536 	vfs_vnode_iterator_init(mp, &marker);
537 	ctx.op = op;
538 	ctx.fs = fs;
539 	while ((vp = vfs_vnode_iterator_next(marker,
540 	    lfs_writevnodes_selector, &ctx)) != NULL) {
541 		ip = VTOI(vp);
542 
543 		/*
544 		 * Write the inode/file if dirty and it's not the IFILE.
545 		 */
546 		if (((ip->i_state & IN_ALLMOD) || !VPISEMPTY(vp)) &&
547 		    ip->i_number != LFS_IFILE_INUM) {
548 			error = lfs_writefile(fs, sp, vp);
549 			if (error) {
550 				vrele(vp);
551 				if (error == EAGAIN) {
552 					/*
553 					 * This error from lfs_putpages
554 					 * indicates we need to drop
555 					 * the segment lock and start
556 					 * over after the cleaner has
557 					 * had a chance to run.
558 					 */
559 					lfs_writeinode(fs, sp, ip);
560 					lfs_writeseg(fs, sp);
561 					if (!VPISEMPTY(vp) &&
562 					    !WRITEINPROG(vp) &&
563 					    !(ip->i_state & IN_ALLMOD)) {
564 						mutex_enter(&lfs_lock);
565 						LFS_SET_UINO(ip, IN_MODIFIED);
566 						mutex_exit(&lfs_lock);
567 					}
568 					break;
569 				}
570 				error = 0; /* XXX not quite right */
571 				continue;
572 			}
573 
574 			if (!VPISEMPTY(vp)) {
575 				if (WRITEINPROG(vp)) {
576 					ivndebug(vp,"writevnodes/write2");
577 				} else if (!(ip->i_state & IN_ALLMOD)) {
578 					mutex_enter(&lfs_lock);
579 					LFS_SET_UINO(ip, IN_MODIFIED);
580 					mutex_exit(&lfs_lock);
581 				}
582 			}
583 			(void) lfs_writeinode(fs, sp, ip);
584 			inodes_written++;
585 		}
586 		vrele(vp);
587 	}
588 	vfs_vnode_iterator_destroy(marker);
589 	return error;
590 }
591 
592 /*
593  * Do a checkpoint.
594  */
595 int
596 lfs_segwrite(struct mount *mp, int flags)
597 {
598 	struct buf *bp;
599 	struct inode *ip;
600 	struct lfs *fs;
601 	struct segment *sp;
602 	struct vnode *vp;
603 	SEGUSE *segusep;
604 	int do_ckp, did_ckp, error;
605 	unsigned n, segleft, maxseg, sn, i, curseg;
606 	int writer_set = 0;
607 	int dirty;
608 	int redo;
609 	SEGSUM *ssp;
610 	int um_error;
611 
612 	fs = VFSTOULFS(mp)->um_lfs;
613 	ASSERT_MAYBE_SEGLOCK(fs);
614 
615 	if (fs->lfs_ronly)
616 		return EROFS;
617 
618 	lfs_imtime(fs);
619 
620 	/*
621 	 * Allocate a segment structure and enough space to hold pointers to
622 	 * the maximum possible number of buffers which can be described in a
623 	 * single summary block.
624 	 */
625 	do_ckp = LFS_SHOULD_CHECKPOINT(fs, flags);
626 
627 	/* We can't do a partial write and checkpoint at the same time. */
628 	if (do_ckp)
629 		flags &= ~SEGM_SINGLE;
630 
631 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
632 	sp = fs->lfs_sp;
633 	if (sp->seg_flags & (SEGM_CLEAN | SEGM_CKP))
634 		do_ckp = 1;
635 
636 	/*
637 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
638 	 * in which case we have to flush *all* buffers off of this vnode.
639 	 * We don't care about other nodes, but write any non-dirop nodes
640 	 * anyway in anticipation of another getnewvnode().
641 	 *
642 	 * If we're cleaning we only write cleaning and ifile blocks, and
643 	 * no dirops, since otherwise we'd risk corruption in a crash.
644 	 */
645 	if (sp->seg_flags & SEGM_CLEAN)
646 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
647 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
648 		do {
649 			um_error = lfs_writevnodes(fs, mp, sp, VN_REG);
650 			if ((sp->seg_flags & SEGM_SINGLE) &&
651 			    lfs_sb_getcurseg(fs) != fs->lfs_startseg) {
652 				DLOG((DLOG_SEG, "lfs_segwrite: breaking out of segment write at daddr 0x%jx\n", (uintmax_t)lfs_sb_getoffset(fs)));
653 				break;
654 			}
655 
656 			if (do_ckp || fs->lfs_dirops == 0) {
657 				if (!writer_set) {
658 					lfs_writer_enter(fs, "lfs writer");
659 					writer_set = 1;
660 				}
661 				error = lfs_writevnodes(fs, mp, sp, VN_DIROP);
662 				if (um_error == 0)
663 					um_error = error;
664 				/*
665 				 * In case writevnodes errored out
666 				 * XXX why are we always doing this and not
667 				 * just on error?
668 				 */
669 				lfs_flush_dirops(fs);
670 				ssp = (SEGSUM *)(sp->segsum);
671 				lfs_ss_setflags(fs, ssp,
672 						lfs_ss_getflags(fs, ssp) & ~(SS_CONT));
673 				lfs_finalize_fs_seguse(fs);
674 			}
675 			if (do_ckp && um_error) {
676 				lfs_segunlock_relock(fs);
677 				sp = fs->lfs_sp;
678 			}
679 		} while (do_ckp && um_error != 0);
680 	}
681 
682 	/*
683 	 * If we are doing a checkpoint, mark everything since the
684 	 * last checkpoint as no longer ACTIVE.
685 	 */
686 	if (do_ckp || fs->lfs_doifile) {
687 		segleft = lfs_sb_getnseg(fs);
688 		curseg = 0;
689 		for (n = 0; n < lfs_sb_getsegtabsz(fs); n++) {
690 			int bread_error;
691 
692 			dirty = 0;
693 			bread_error = bread(fs->lfs_ivnode,
694 			    lfs_sb_getcleansz(fs) + n,
695 			    lfs_sb_getbsize(fs), B_MODIFY, &bp);
696 			if (bread_error)
697 				panic("lfs_segwrite: ifile read: "
698 				      "seguse %u: error %d\n",
699 				      n, bread_error);
700 			segusep = (SEGUSE *)bp->b_data;
701 			maxseg = min(segleft, lfs_sb_getsepb(fs));
702 			for (i = 0; i < maxseg; i++) {
703 				sn = curseg + i;
704 				if (sn != lfs_dtosn(fs, lfs_sb_getcurseg(fs)) &&
705 				    segusep->su_flags & SEGUSE_ACTIVE) {
706 					segusep->su_flags &= ~SEGUSE_ACTIVE;
707 					--fs->lfs_nactive;
708 					++dirty;
709 				}
710 				fs->lfs_suflags[fs->lfs_activesb][sn] =
711 					segusep->su_flags;
712 				if (lfs_sb_getversion(fs) > 1)
713 					++segusep;
714 				else
715 					segusep = (SEGUSE *)
716 						((SEGUSE_V1 *)segusep + 1);
717 			}
718 
719 			if (dirty)
720 				error = LFS_BWRITE_LOG(bp); /* Ifile */
721 			else
722 				brelse(bp, 0);
723 			segleft -= lfs_sb_getsepb(fs);
724 			curseg += lfs_sb_getsepb(fs);
725 		}
726 	}
727 
728 	KASSERT(LFS_SEGLOCK_HELD(fs));
729 
730 	did_ckp = 0;
731 	if (do_ckp || fs->lfs_doifile) {
732 		vp = fs->lfs_ivnode;
733 #ifdef DEBUG
734 		int loopcount = 0;
735 #endif
736 		do {
737 
738 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0, curproc->p_pid);
739 
740 			mutex_enter(&lfs_lock);
741 			fs->lfs_flags &= ~LFS_IFDIRTY;
742 			mutex_exit(&lfs_lock);
743 
744 			ip = VTOI(vp);
745 
746 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL) {
747 				/*
748 				 * Ifile has no pages, so we don't need
749 				 * to check error return here.
750 				 */
751 				lfs_writefile(fs, sp, vp);
752 				/*
753 				 * Ensure the Ifile takes the current segment
754 				 * into account.  See comment in lfs_vflush.
755 				 */
756 				lfs_writefile(fs, sp, vp);
757 				lfs_writefile(fs, sp, vp);
758 			}
759 
760 			if (ip->i_state & IN_ALLMOD)
761 				++did_ckp;
762 #if 0
763 			redo = (do_ckp ? lfs_writeinode(fs, sp, ip) : 0);
764 #else
765 			redo = lfs_writeinode(fs, sp, ip);
766 #endif
767 			redo += lfs_writeseg(fs, sp);
768 			mutex_enter(&lfs_lock);
769 			redo += (fs->lfs_flags & LFS_IFDIRTY);
770 			mutex_exit(&lfs_lock);
771 #ifdef DEBUG
772 			if (++loopcount > 2)
773 				log(LOG_NOTICE, "lfs_segwrite: looping count=%d\n",
774 					loopcount);
775 #endif
776 		} while (redo && do_ckp);
777 
778 		/*
779 		 * Unless we are unmounting, the Ifile may continue to have
780 		 * dirty blocks even after a checkpoint, due to changes to
781 		 * inodes' atime.  If we're checkpointing, it's "impossible"
782 		 * for other parts of the Ifile to be dirty after the loop
783 		 * above, since we hold the segment lock.
784 		 */
785 		mutex_enter(vp->v_interlock);
786 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
787 			LFS_CLR_UINO(ip, IN_ALLMOD);
788 		}
789 #ifdef DIAGNOSTIC
790 		else if (do_ckp) {
791 			int do_panic = 0;
792 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
793 				if (bp->b_lblkno < lfs_sb_getcleansz(fs) +
794 				    lfs_sb_getsegtabsz(fs) &&
795 				    !(bp->b_flags & B_GATHERED)) {
796 					printf("ifile lbn %ld still dirty (flags %lx)\n",
797 						(long)bp->b_lblkno,
798 						(long)bp->b_flags);
799 					++do_panic;
800 				}
801 			}
802 			if (do_panic)
803 				panic("dirty blocks");
804 		}
805 #endif
806 		mutex_exit(vp->v_interlock);
807 	} else {
808 		(void) lfs_writeseg(fs, sp);
809 	}
810 
811 	/* Note Ifile no longer needs to be written */
812 	fs->lfs_doifile = 0;
813 	if (writer_set)
814 		lfs_writer_leave(fs);
815 
816 	/*
817 	 * If we didn't write the Ifile, we didn't really do anything.
818 	 * That means that (1) there is a checkpoint on disk and (2)
819 	 * nothing has changed since it was written.
820 	 *
821 	 * Take the flags off of the segment so that lfs_segunlock
822 	 * doesn't have to write the superblock either.
823 	 */
824 	if (do_ckp && !did_ckp) {
825 		sp->seg_flags &= ~SEGM_CKP;
826 	}
827 
828 	if (lfs_dostats) {
829 		++lfs_stats.nwrites;
830 		if (sp->seg_flags & SEGM_SYNC)
831 			++lfs_stats.nsync_writes;
832 		if (sp->seg_flags & SEGM_CKP)
833 			++lfs_stats.ncheckpoints;
834 	}
835 	lfs_segunlock(fs);
836 	return (0);
837 }
838 
839 /*
840  * Write the dirty blocks associated with a vnode.
841  */
842 int
843 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
844 {
845 	struct inode *ip;
846 	int i, frag;
847 	SEGSUM *ssp;
848 	int error;
849 
850 	ASSERT_SEGLOCK(fs);
851 	error = 0;
852 	ip = VTOI(vp);
853 
854 	lfs_acquire_finfo(fs, ip->i_number, ip->i_gen);
855 
856 	if (vp->v_uflag & VU_DIROP) {
857 		ssp = (SEGSUM *)sp->segsum;
858 		lfs_ss_setflags(fs, ssp,
859 				lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT));
860 	}
861 
862 	if (sp->seg_flags & SEGM_CLEAN) {
863 		lfs_gather(fs, sp, vp, lfs_match_fake);
864 		/*
865 		 * For a file being flushed, we need to write *all* blocks.
866 		 * This means writing the cleaning blocks first, and then
867 		 * immediately following with any non-cleaning blocks.
868 		 * The same is true of the Ifile since checkpoints assume
869 		 * that all valid Ifile blocks are written.
870 		 */
871 		if (IS_FLUSHING(fs, vp) || vp == fs->lfs_ivnode) {
872 			lfs_gather(fs, sp, vp, lfs_match_data);
873 			/*
874 			 * Don't call VOP_PUTPAGES: if we're flushing,
875 			 * we've already done it, and the Ifile doesn't
876 			 * use the page cache.
877 			 */
878 		}
879 	} else {
880 		lfs_gather(fs, sp, vp, lfs_match_data);
881 		/*
882 		 * If we're flushing, we've already called VOP_PUTPAGES
883 		 * so don't do it again.  Otherwise, we want to write
884 		 * everything we've got.
885 		 */
886 		if (!IS_FLUSHING(fs, vp)) {
887 			mutex_enter(vp->v_interlock);
888 			error = VOP_PUTPAGES(vp, 0, 0,
889 				PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
890 		}
891 	}
892 
893 	/*
894 	 * It may not be necessary to write the meta-data blocks at this point,
895 	 * as the roll-forward recovery code should be able to reconstruct the
896 	 * list.
897 	 *
898 	 * We have to write them anyway, though, under two conditions: (1) the
899 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
900 	 * checkpointing.
901 	 *
902 	 * BUT if we are cleaning, we might have indirect blocks that refer to
903 	 * new blocks not being written yet, in addition to fragments being
904 	 * moved out of a cleaned segment.  If that is the case, don't
905 	 * write the indirect blocks, or the finfo will have a small block
906 	 * in the middle of it!
907 	 * XXX in this case isn't the inode size wrong too?
908 	 */
909 	frag = 0;
910 	if (sp->seg_flags & SEGM_CLEAN) {
911 		for (i = 0; i < ULFS_NDADDR; i++)
912 			if (ip->i_lfs_fragsize[i] > 0 &&
913 			    ip->i_lfs_fragsize[i] < lfs_sb_getbsize(fs))
914 				++frag;
915 	}
916 	KASSERTMSG((frag <= 1),
917 	    "lfs_writefile: more than one fragment! frag=%d", frag);
918 	if (IS_FLUSHING(fs, vp) ||
919 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
920 		lfs_gather(fs, sp, vp, lfs_match_indir);
921 		lfs_gather(fs, sp, vp, lfs_match_dindir);
922 		lfs_gather(fs, sp, vp, lfs_match_tindir);
923 	}
924 	lfs_release_finfo(fs);
925 
926 	return error;
927 }
928 
929 /*
930  * Update segment accounting to reflect this inode's change of address.
931  */
932 static int
933 lfs_update_iaddr(struct lfs *fs, struct segment *sp, struct inode *ip, daddr_t ndaddr)
934 {
935 	struct buf *bp;
936 	daddr_t daddr;
937 	IFILE *ifp;
938 	SEGUSE *sup;
939 	ino_t ino;
940 	int redo_ifile;
941 	u_int32_t sn;
942 
943 	redo_ifile = 0;
944 
945 	/*
946 	 * If updating the ifile, update the super-block.  Update the disk
947 	 * address and access times for this inode in the ifile.
948 	 */
949 	ino = ip->i_number;
950 	if (ino == LFS_IFILE_INUM) {
951 		daddr = lfs_sb_getidaddr(fs);
952 		lfs_sb_setidaddr(fs, LFS_DBTOFSB(fs, ndaddr));
953 	} else {
954 		LFS_IENTRY(ifp, fs, ino, bp);
955 		daddr = lfs_if_getdaddr(fs, ifp);
956 		lfs_if_setdaddr(fs, ifp, LFS_DBTOFSB(fs, ndaddr));
957 		(void)LFS_BWRITE_LOG(bp); /* Ifile */
958 	}
959 
960 	/*
961 	 * If this is the Ifile and lfs_offset is set to the first block
962 	 * in the segment, dirty the new segment's accounting block
963 	 * (XXX should already be dirty?) and tell the caller to do it again.
964 	 */
965 	if (ip->i_number == LFS_IFILE_INUM) {
966 		sn = lfs_dtosn(fs, lfs_sb_getoffset(fs));
967 		if (lfs_sntod(fs, sn) + lfs_btofsb(fs, lfs_sb_getsumsize(fs)) ==
968 		    lfs_sb_getoffset(fs)) {
969 			LFS_SEGENTRY(sup, fs, sn, bp);
970 			KASSERT(bp->b_oflags & BO_DELWRI);
971 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
972 			/* fs->lfs_flags |= LFS_IFDIRTY; */
973 			redo_ifile |= 1;
974 		}
975 	}
976 
977 	/*
978 	 * The inode's last address should not be in the current partial
979 	 * segment, except under exceptional circumstances (lfs_writevnodes
980 	 * had to start over, and in the meantime more blocks were written
981 	 * to a vnode).	 Both inodes will be accounted to this segment
982 	 * in lfs_writeseg so we need to subtract the earlier version
983 	 * here anyway.	 The segment count can temporarily dip below
984 	 * zero here; keep track of how many duplicates we have in
985 	 * "dupino" so we don't panic below.
986 	 */
987 	if (daddr >= lfs_sb_getlastpseg(fs) && daddr <= lfs_sb_getoffset(fs)) {
988 		++sp->ndupino;
989 		DLOG((DLOG_SEG, "lfs_writeinode: last inode addr in current pseg "
990 		      "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
991 		      (long long)daddr, sp->ndupino));
992 	}
993 	/*
994 	 * Account the inode: it no longer belongs to its former segment,
995 	 * though it will not belong to the new segment until that segment
996 	 * is actually written.
997 	 */
998 	if (daddr != LFS_UNUSED_DADDR) {
999 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
1000 		int ndupino __diagused =
1001 		    (sp->seg_number == oldsn) ? sp->ndupino : 0;
1002 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1003 		KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino)
1004 			>= DINOSIZE(fs)),
1005 		    "lfs_writeinode: negative bytes "
1006 		    "(segment %" PRIu32 " short by %d, "
1007 		    "oldsn=%" PRIu32 ", cursn=%" PRIu32
1008 		    ", daddr=%" PRId64 ", su_nbytes=%u, "
1009 		    "ndupino=%d)\n",
1010 		    lfs_dtosn(fs, daddr),
1011 		    (int)DINOSIZE(fs) * (1 - sp->ndupino) - sup->su_nbytes,
1012 		    oldsn, sp->seg_number, daddr,
1013 		    (unsigned int)sup->su_nbytes,
1014 		    sp->ndupino);
1015 		DLOG((DLOG_SU, "seg %d -= %d for ino %d inode\n",
1016 		      lfs_dtosn(fs, daddr), DINOSIZE(fs), ino));
1017 		sup->su_nbytes -= DINOSIZE(fs);
1018 		redo_ifile |=
1019 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1020 		if (redo_ifile) {
1021 			mutex_enter(&lfs_lock);
1022 			fs->lfs_flags |= LFS_IFDIRTY;
1023 			mutex_exit(&lfs_lock);
1024 			/* Don't double-account */
1025 			lfs_sb_setidaddr(fs, 0x0);
1026 		}
1027 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1028 	}
1029 
1030 	return redo_ifile;
1031 }
1032 
1033 int
1034 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
1035 {
1036 	struct buf *bp;
1037 	union lfs_dinode *cdp;
1038 	struct vnode *vp = ITOV(ip);
1039 	daddr_t daddr;
1040 	IINFO *iip;
1041 	int i;
1042 	int redo_ifile = 0;
1043 	int gotblk = 0;
1044 	int count;
1045 	SEGSUM *ssp;
1046 
1047 	ASSERT_SEGLOCK(fs);
1048 	if (!(ip->i_state & IN_ALLMOD) && !(vp->v_uflag & VU_DIROP))
1049 		return (0);
1050 
1051 	/* Can't write ifile when writer is not set */
1052 	KASSERT(ip->i_number != LFS_IFILE_INUM || fs->lfs_writer > 0 ||
1053 		(sp->seg_flags & SEGM_CLEAN));
1054 
1055 	/*
1056 	 * If this is the Ifile, see if writing it here will generate a
1057 	 * temporary misaccounting.  If it will, do the accounting and write
1058 	 * the blocks, postponing the inode write until the accounting is
1059 	 * solid.
1060 	 */
1061 	count = 0;
1062 	while (vp == fs->lfs_ivnode) {
1063 		int redo = 0;
1064 
1065 		if (sp->idp == NULL && sp->ibp == NULL &&
1066 		    (sp->seg_bytes_left < lfs_sb_getibsize(fs) ||
1067 		     sp->sum_bytes_left < sizeof(int32_t))) {
1068 			(void) lfs_writeseg(fs, sp);
1069 			continue;
1070 		}
1071 
1072 		/* Look for dirty Ifile blocks */
1073 		LIST_FOREACH(bp, &fs->lfs_ivnode->v_dirtyblkhd, b_vnbufs) {
1074 			if (!(bp->b_flags & B_GATHERED)) {
1075 				redo = 1;
1076 				break;
1077 			}
1078 		}
1079 
1080 		if (redo == 0)
1081 			redo = lfs_update_iaddr(fs, sp, ip, 0x0);
1082 		if (redo == 0)
1083 			break;
1084 
1085 		if (sp->idp) {
1086 			lfs_dino_setinumber(fs, sp->idp, 0);
1087 			sp->idp = NULL;
1088 		}
1089 		++count;
1090 		if (count > 2)
1091 			log(LOG_NOTICE, "lfs_writeinode: looping count=%d\n", count);
1092 		lfs_writefile(fs, sp, fs->lfs_ivnode);
1093 	}
1094 
1095 	/* Allocate a new inode block if necessary. */
1096 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
1097 	    sp->ibp == NULL) {
1098 		/* Allocate a new segment if necessary. */
1099 		if (sp->seg_bytes_left < lfs_sb_getibsize(fs) ||
1100 		    sp->sum_bytes_left < sizeof(int32_t))
1101 			(void) lfs_writeseg(fs, sp);
1102 
1103 		/* Get next inode block. */
1104 		daddr = lfs_sb_getoffset(fs);
1105 		lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs)));
1106 		sp->ibp = *sp->cbpp++ =
1107 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
1108 			    LFS_FSBTODB(fs, daddr), lfs_sb_getibsize(fs), 0, 0);
1109 		gotblk++;
1110 
1111 		/* Zero out inode numbers */
1112 		for (i = 0; i < LFS_INOPB(fs); ++i) {
1113 			union lfs_dinode *tmpdi;
1114 
1115 			tmpdi = (union lfs_dinode *)((char *)sp->ibp->b_data +
1116 						     DINOSIZE(fs) * i);
1117 			lfs_dino_setinumber(fs, tmpdi, 0);
1118 		}
1119 
1120 		++sp->start_bpp;
1121 		lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getibsize(fs)));
1122 		/* Set remaining space counters. */
1123 		sp->seg_bytes_left -= lfs_sb_getibsize(fs);
1124 		sp->sum_bytes_left -= sizeof(int32_t);
1125 
1126 		/* Store the address in the segment summary. */
1127 		iip = NTH_IINFO(fs, sp->segsum, sp->ninodes / LFS_INOPB(fs));
1128 		lfs_ii_setblock(fs, iip, daddr);
1129 	}
1130 
1131 	/* Check VU_DIROP in case there is a new file with no data blocks */
1132 	if (vp->v_uflag & VU_DIROP) {
1133 		ssp = (SEGSUM *)sp->segsum;
1134 		lfs_ss_setflags(fs, ssp,
1135 				lfs_ss_getflags(fs, ssp) | (SS_DIROP|SS_CONT));
1136 	}
1137 
1138 	/* Update the inode times and copy the inode onto the inode page. */
1139 	/* XXX kludge --- don't redirty the ifile just to put times on it */
1140 	if (ip->i_number != LFS_IFILE_INUM)
1141 		LFS_ITIMES(ip, NULL, NULL, NULL);
1142 
1143 	/*
1144 	 * If this is the Ifile, and we've already written the Ifile in this
1145 	 * partial segment, just overwrite it (it's not on disk yet) and
1146 	 * continue.
1147 	 *
1148 	 * XXX we know that the bp that we get the second time around has
1149 	 * already been gathered.
1150 	 */
1151 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
1152 		lfs_copy_dinode(fs, sp->idp, ip->i_din);
1153 		ip->i_lfs_osize = ip->i_size;
1154 		return 0;
1155 	}
1156 
1157 	bp = sp->ibp;
1158 	cdp = DINO_IN_BLOCK(fs, bp->b_data, sp->ninodes % LFS_INOPB(fs));
1159 	lfs_copy_dinode(fs, cdp, ip->i_din);
1160 
1161 	/*
1162 	 * This inode is on its way to disk; clear its VU_DIROP status when
1163 	 * the write is complete.
1164 	 */
1165 	if (vp->v_uflag & VU_DIROP) {
1166 		if (!(sp->seg_flags & SEGM_CLEAN))
1167 			ip->i_state |= IN_CDIROP;
1168 		else {
1169 			DLOG((DLOG_DIROP, "lfs_writeinode: not clearing dirop for cleaned ino %d\n", (int)ip->i_number));
1170 		}
1171 	}
1172 
1173 	/*
1174 	 * If cleaning, link counts and directory file sizes cannot change,
1175 	 * since those would be directory operations---even if the file
1176 	 * we are writing is marked VU_DIROP we should write the old values.
1177 	 * If we're not cleaning, of course, update the values so we get
1178 	 * current values the next time we clean.
1179 	 */
1180 	if (sp->seg_flags & SEGM_CLEAN) {
1181 		if (vp->v_uflag & VU_DIROP) {
1182 			lfs_dino_setnlink(fs, cdp, ip->i_lfs_odnlink);
1183 			/* if (vp->v_type == VDIR) */
1184 			lfs_dino_setsize(fs, cdp, ip->i_lfs_osize);
1185 		}
1186 	} else {
1187 		ip->i_lfs_odnlink = lfs_dino_getnlink(fs, cdp);
1188 		ip->i_lfs_osize = ip->i_size;
1189 	}
1190 
1191 
1192 	/* We can finish the segment accounting for truncations now */
1193 	lfs_finalize_ino_seguse(fs, ip);
1194 
1195 	/*
1196 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
1197 	 * addresses to disk; possibly change the on-disk record of
1198 	 * the inode size, either by reverting to the previous size
1199 	 * (in the case of cleaning) or by verifying the inode's block
1200 	 * holdings (in the case of files being allocated as they are being
1201 	 * written).
1202 	 * XXX By not writing UNWRITTEN blocks, we are making the lfs_avail
1203 	 * XXX count on disk wrong by the same amount.	We should be
1204 	 * XXX able to "borrow" from lfs_avail and return it after the
1205 	 * XXX Ifile is written.  See also in lfs_writeseg.
1206 	 */
1207 
1208 	/* Check file size based on highest allocated block */
1209 	if (((lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFREG ||
1210 	     (lfs_dino_getmode(fs, ip->i_din) & LFS_IFMT) == LFS_IFDIR) &&
1211 	    ip->i_size > ((ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs))) {
1212 		lfs_dino_setsize(fs, cdp, (ip->i_lfs_hiblk + 1) << lfs_sb_getbshift(fs));
1213 		DLOG((DLOG_SEG, "lfs_writeinode: ino %d size %" PRId64 " -> %"
1214 		      PRId64 "\n", (int)ip->i_number, ip->i_size, lfs_dino_getsize(fs, cdp)));
1215 	}
1216 	if (ip->i_lfs_effnblks != lfs_dino_getblocks(fs, ip->i_din)) {
1217 		DLOG((DLOG_SEG, "lfs_writeinode: cleansing ino %d eff %jd != nblk %d)"
1218 		      " at %jx\n", ip->i_number, (intmax_t)ip->i_lfs_effnblks,
1219 		      lfs_dino_getblocks(fs, ip->i_din), (uintmax_t)lfs_sb_getoffset(fs)));
1220 		for (i=0; i<ULFS_NDADDR; i++) {
1221 			if (lfs_dino_getdb(fs, cdp, i) == UNWRITTEN) {
1222 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1223 				lfs_dino_setdb(fs, cdp, i, 0);
1224 			}
1225 		}
1226 		for (i=0; i<ULFS_NIADDR; i++) {
1227 			if (lfs_dino_getib(fs, cdp, i) == UNWRITTEN) {
1228 				DLOG((DLOG_SEG, "lfs_writeinode: wiping UNWRITTEN\n"));
1229 				lfs_dino_setib(fs, cdp, i, 0);
1230 			}
1231 		}
1232 	}
1233 
1234 #ifdef DIAGNOSTIC
1235 	/*
1236 	 * Check dinode held blocks against dinode size.
1237 	 * This should be identical to the check in lfs_vget().
1238 	 */
1239 	for (i = (lfs_dino_getsize(fs, cdp) + lfs_sb_getbsize(fs) - 1) >> lfs_sb_getbshift(fs);
1240 	     i < ULFS_NDADDR; i++) {
1241 		KASSERT(i >= 0);
1242 		if ((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFLNK)
1243 			continue;
1244 		if (((lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFBLK ||
1245 		     (lfs_dino_getmode(fs, cdp) & LFS_IFMT) == LFS_IFCHR) && i == 0)
1246 			continue;
1247 		if (lfs_dino_getdb(fs, cdp, i) != 0) {
1248 # ifdef DEBUG
1249 			lfs_dump_dinode(fs, cdp);
1250 # endif
1251 			panic("writing inconsistent inode");
1252 		}
1253 	}
1254 #endif /* DIAGNOSTIC */
1255 
1256 	if (ip->i_state & IN_CLEANING)
1257 		LFS_CLR_UINO(ip, IN_CLEANING);
1258 	else {
1259 		/* XXX IN_ALLMOD */
1260 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
1261 			     IN_UPDATE | IN_MODIFY);
1262 		if (ip->i_lfs_effnblks == lfs_dino_getblocks(fs, ip->i_din))
1263 			LFS_CLR_UINO(ip, IN_MODIFIED);
1264 		else {
1265 			DLOG((DLOG_VNODE, "lfs_writeinode: ino %d: real "
1266 			    "blks=%d, eff=%jd\n", ip->i_number,
1267 			    lfs_dino_getblocks(fs, ip->i_din), (intmax_t)ip->i_lfs_effnblks));
1268 		}
1269 	}
1270 
1271 	if (ip->i_number == LFS_IFILE_INUM) {
1272 		/* We know sp->idp == NULL */
1273 		sp->idp = DINO_IN_BLOCK(fs, bp, sp->ninodes % LFS_INOPB(fs));
1274 
1275 		/* Not dirty any more */
1276 		mutex_enter(&lfs_lock);
1277 		fs->lfs_flags &= ~LFS_IFDIRTY;
1278 		mutex_exit(&lfs_lock);
1279 	}
1280 
1281 	if (gotblk) {
1282 		mutex_enter(&bufcache_lock);
1283 		LFS_LOCK_BUF(bp);
1284 		brelsel(bp, 0);
1285 		mutex_exit(&bufcache_lock);
1286 	}
1287 
1288 	/* Increment inode count in segment summary block. */
1289 
1290 	ssp = (SEGSUM *)sp->segsum;
1291 	lfs_ss_setninos(fs, ssp, lfs_ss_getninos(fs, ssp) + 1);
1292 
1293 	/* If this page is full, set flag to allocate a new page. */
1294 	if (++sp->ninodes % LFS_INOPB(fs) == 0)
1295 		sp->ibp = NULL;
1296 
1297 	redo_ifile = lfs_update_iaddr(fs, sp, ip, bp->b_blkno);
1298 
1299 	KASSERT(redo_ifile == 0);
1300 	return (redo_ifile);
1301 }
1302 
1303 int
1304 lfs_gatherblock(struct segment *sp, struct buf *bp, kmutex_t *mptr)
1305 {
1306 	struct lfs *fs;
1307 	int vers;
1308 	int j, blksinblk;
1309 
1310 	ASSERT_SEGLOCK(sp->fs);
1311 	KASSERTMSG((sp->vp != NULL),
1312 	    "lfs_gatherblock: Null vp in segment");
1313 
1314 	/* If full, finish this segment. */
1315 	fs = sp->fs;
1316 	blksinblk = howmany(bp->b_bcount, lfs_sb_getbsize(fs));
1317 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1318 	    sp->seg_bytes_left < bp->b_bcount) {
1319 		if (mptr)
1320 			mutex_exit(mptr);
1321 		lfs_updatemeta(sp);
1322 
1323 		vers = lfs_fi_getversion(fs, sp->fip);
1324 		(void) lfs_writeseg(fs, sp);
1325 
1326 		/* Add the current file to the segment summary. */
1327 		lfs_acquire_finfo(fs, VTOI(sp->vp)->i_number, vers);
1328 
1329 		if (mptr)
1330 			mutex_enter(mptr);
1331 		return (1);
1332 	}
1333 
1334 	if (bp->b_flags & B_GATHERED) {
1335 		DLOG((DLOG_SEG, "lfs_gatherblock: already gathered! Ino %ju,"
1336 		      " lbn %" PRId64 "\n",
1337 		      (uintmax_t)lfs_fi_getino(fs, sp->fip), bp->b_lblkno));
1338 		return (0);
1339 	}
1340 
1341 	/* Insert into the buffer list, update the FINFO block. */
1342 	bp->b_flags |= B_GATHERED;
1343 
1344 	*sp->cbpp++ = bp;
1345 	for (j = 0; j < blksinblk; j++) {
1346 		unsigned bn;
1347 
1348 		bn = lfs_fi_getnblocks(fs, sp->fip);
1349 		lfs_fi_setnblocks(fs, sp->fip, bn+1);
1350 		lfs_fi_setblock(fs, sp->fip, bn, bp->b_lblkno + j);
1351 		/* This block's accounting moves from lfs_favail to lfs_avail */
1352 		lfs_deregister_block(sp->vp, bp->b_lblkno + j);
1353 	}
1354 
1355 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1356 	sp->seg_bytes_left -= bp->b_bcount;
1357 	return (0);
1358 }
1359 
1360 int
1361 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1362     int (*match)(struct lfs *, struct buf *))
1363 {
1364 	struct buf *bp, *nbp;
1365 	int count = 0;
1366 
1367 	ASSERT_SEGLOCK(fs);
1368 	if (vp->v_type == VBLK)
1369 		return 0;
1370 	KASSERT(sp->vp == NULL);
1371 	sp->vp = vp;
1372 	mutex_enter(&bufcache_lock);
1373 
1374 #ifndef LFS_NO_BACKBUF_HACK
1375 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1376 # define	BUF_OFFSET	\
1377 	(((char *)&LIST_NEXT(bp, b_vnbufs)) - (char *)bp)
1378 # define	BACK_BUF(BP)	\
1379 	((struct buf *)(((char *)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1380 # define	BEG_OF_LIST	\
1381 	((struct buf *)(((char *)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1382 
1383 loop:
1384 	/* Find last buffer. */
1385 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1386 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1387 	     bp = LIST_NEXT(bp, b_vnbufs))
1388 		continue;
1389 
1390 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1391 		nbp = BACK_BUF(bp);
1392 #else /* LFS_NO_BACKBUF_HACK */
1393 loop:
1394 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1395 		nbp = LIST_NEXT(bp, b_vnbufs);
1396 #endif /* LFS_NO_BACKBUF_HACK */
1397 		if ((bp->b_cflags & BC_BUSY) != 0 ||
1398 		    (bp->b_flags & B_GATHERED) != 0 || !match(fs, bp)) {
1399 #ifdef DEBUG
1400 			if (vp == fs->lfs_ivnode &&
1401 			    (bp->b_cflags & BC_BUSY) != 0 &&
1402 			    (bp->b_flags & B_GATHERED) == 0)
1403 				log(LOG_NOTICE, "lfs_gather: ifile lbn %"
1404 				      PRId64 " busy (%x) at 0x%jx",
1405 				      bp->b_lblkno, bp->b_flags,
1406 				      (uintmax_t)lfs_sb_getoffset(fs));
1407 #endif
1408 			continue;
1409 		}
1410 #ifdef DIAGNOSTIC
1411 # ifdef LFS_USE_BC_INVAL
1412 		if ((bp->b_cflags & BC_INVAL) != 0 && bp->b_iodone == NULL) {
1413 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1414 			      " is BC_INVAL\n", bp->b_lblkno));
1415 			VOP_PRINT(bp->b_vp);
1416 		}
1417 # endif /* LFS_USE_BC_INVAL */
1418 		if (!(bp->b_oflags & BO_DELWRI))
1419 			panic("lfs_gather: bp not BO_DELWRI");
1420 		if (!(bp->b_flags & B_LOCKED)) {
1421 			DLOG((DLOG_SEG, "lfs_gather: lbn %" PRId64
1422 			      " blk %" PRId64 " not B_LOCKED\n",
1423 			      bp->b_lblkno,
1424 			      LFS_DBTOFSB(fs, bp->b_blkno)));
1425 			VOP_PRINT(bp->b_vp);
1426 			panic("lfs_gather: bp not B_LOCKED");
1427 		}
1428 #endif
1429 		if (lfs_gatherblock(sp, bp, &bufcache_lock)) {
1430 			goto loop;
1431 		}
1432 		count++;
1433 	}
1434 	mutex_exit(&bufcache_lock);
1435 	lfs_updatemeta(sp);
1436 	KASSERT(sp->vp == vp);
1437 	sp->vp = NULL;
1438 	return count;
1439 }
1440 
1441 #if DEBUG
1442 # define DEBUG_OOFF(n) do {						\
1443 	if (ooff == 0) {						\
1444 		DLOG((DLOG_SEG, "lfs_updatemeta[%d]: warning: writing " \
1445 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1446 			", was 0x0 (or %" PRId64 ")\n",			\
1447 			(n), ip->i_number, lbn, ndaddr, daddr));	\
1448 	}								\
1449 } while (0)
1450 #else
1451 # define DEBUG_OOFF(n)
1452 #endif
1453 
1454 /*
1455  * Change the given block's address to ndaddr, finding its previous
1456  * location using ulfs_bmaparray().
1457  *
1458  * Account for this change in the segment table.
1459  *
1460  * called with sp == NULL by roll-forwarding code.
1461  */
1462 void
1463 lfs_update_single(struct lfs *fs, struct segment *sp,
1464     struct vnode *vp, daddr_t lbn, daddr_t ndaddr, int size)
1465 {
1466 	SEGUSE *sup;
1467 	struct buf *bp;
1468 	struct indir a[ULFS_NIADDR + 2], *ap;
1469 	struct inode *ip;
1470 	daddr_t daddr, ooff;
1471 	int num, error;
1472 	int bb, osize, obb;
1473 
1474 	ASSERT_SEGLOCK(fs);
1475 	KASSERT(sp == NULL || sp->vp == vp);
1476 	ip = VTOI(vp);
1477 
1478 	error = ulfs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1479 	if (error)
1480 		panic("lfs_updatemeta: ulfs_bmaparray returned %d", error);
1481 
1482 	KASSERT(daddr <= LFS_MAX_DADDR(fs));
1483 	if (daddr > 0)
1484 		daddr = LFS_DBTOFSB(fs, daddr);
1485 
1486 	bb = lfs_numfrags(fs, size);
1487 	switch (num) {
1488 	    case 0:
1489 		    ooff = lfs_dino_getdb(fs, ip->i_din, lbn);
1490 		    DEBUG_OOFF(0);
1491 		    if (ooff == UNWRITTEN)
1492 			    lfs_dino_setblocks(fs, ip->i_din,
1493 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1494 		    else {
1495 			    /* possible fragment truncation or extension */
1496 			    obb = lfs_btofsb(fs, ip->i_lfs_fragsize[lbn]);
1497 			    lfs_dino_setblocks(fs, ip->i_din,
1498 				lfs_dino_getblocks(fs, ip->i_din) + (bb-obb));
1499 		    }
1500 		    lfs_dino_setdb(fs, ip->i_din, lbn, ndaddr);
1501 		    break;
1502 	    case 1:
1503 		    ooff = lfs_dino_getib(fs, ip->i_din, a[0].in_off);
1504 		    DEBUG_OOFF(1);
1505 		    if (ooff == UNWRITTEN)
1506 			    lfs_dino_setblocks(fs, ip->i_din,
1507 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1508 		    lfs_dino_setib(fs, ip->i_din, a[0].in_off, ndaddr);
1509 		    break;
1510 	    default:
1511 		    ap = &a[num - 1];
1512 		    if (bread(vp, ap->in_lbn, lfs_sb_getbsize(fs),
1513 			B_MODIFY, &bp))
1514 			    panic("lfs_updatemeta: bread bno %" PRId64,
1515 				  ap->in_lbn);
1516 
1517 		    ooff = lfs_iblock_get(fs, bp->b_data, ap->in_off);
1518 		    DEBUG_OOFF(num);
1519 		    if (ooff == UNWRITTEN)
1520 			    lfs_dino_setblocks(fs, ip->i_din,
1521 				lfs_dino_getblocks(fs, ip->i_din) + bb);
1522 		    lfs_iblock_set(fs, bp->b_data, ap->in_off, ndaddr);
1523 		    (void) VOP_BWRITE(bp->b_vp, bp);
1524 	}
1525 
1526 	KASSERT(ooff == 0 || ooff == UNWRITTEN || ooff == daddr);
1527 
1528 	/* Update hiblk when extending the file */
1529 	if (lbn > ip->i_lfs_hiblk)
1530 		ip->i_lfs_hiblk = lbn;
1531 
1532 	/*
1533 	 * Though we'd rather it couldn't, this *can* happen right now
1534 	 * if cleaning blocks and regular blocks coexist.
1535 	 */
1536 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1537 
1538 	/*
1539 	 * Update segment usage information, based on old size
1540 	 * and location.
1541 	 */
1542 	if (daddr > 0) {
1543 		u_int32_t oldsn = lfs_dtosn(fs, daddr);
1544 		int ndupino __diagused = (sp && sp->seg_number == oldsn ?
1545 		    sp->ndupino : 0);
1546 
1547 		KASSERT(oldsn < lfs_sb_getnseg(fs));
1548 		if (lbn >= 0 && lbn < ULFS_NDADDR)
1549 			osize = ip->i_lfs_fragsize[lbn];
1550 		else
1551 			osize = lfs_sb_getbsize(fs);
1552 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1553 		KASSERTMSG(((sup->su_nbytes + DINOSIZE(fs)*ndupino) >= osize),
1554 		    "lfs_updatemeta: negative bytes "
1555 		    "(segment %" PRIu32 " short by %" PRId64
1556 		    ")\n"
1557 		    "lfs_updatemeta: ino %llu, lbn %" PRId64
1558 		    ", addr = 0x%" PRIx64 "\n"
1559 		    "lfs_updatemeta: ndupino=%d",
1560 		    lfs_dtosn(fs, daddr),
1561 		    (int64_t)osize - (DINOSIZE(fs) * ndupino + sup->su_nbytes),
1562 		    (unsigned long long)ip->i_number, lbn, daddr,
1563 		    ndupino);
1564 		DLOG((DLOG_SU, "seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1565 		      " db 0x%" PRIx64 "\n",
1566 		      lfs_dtosn(fs, daddr), osize,
1567 		      ip->i_number, lbn, daddr));
1568 		sup->su_nbytes -= osize;
1569 		if (!(bp->b_flags & B_GATHERED)) {
1570 			mutex_enter(&lfs_lock);
1571 			fs->lfs_flags |= LFS_IFDIRTY;
1572 			mutex_exit(&lfs_lock);
1573 		}
1574 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1575 	}
1576 	/*
1577 	 * Now that this block has a new address, and its old
1578 	 * segment no longer owns it, we can forget about its
1579 	 * old size.
1580 	 */
1581 	if (lbn >= 0 && lbn < ULFS_NDADDR)
1582 		ip->i_lfs_fragsize[lbn] = size;
1583 }
1584 
1585 /*
1586  * Update the metadata that points to the blocks listed in the FINFO
1587  * array.
1588  */
1589 void
1590 lfs_updatemeta(struct segment *sp)
1591 {
1592 	struct buf *sbp;
1593 	struct lfs *fs;
1594 	struct vnode *vp;
1595 	daddr_t lbn;
1596 	int i, nblocks, num;
1597 	int __diagused nblocks_orig;
1598 	int bb;
1599 	int bytesleft, size;
1600 	unsigned lastlength;
1601 	union lfs_blocks tmpptr;
1602 
1603 	fs = sp->fs;
1604 	vp = sp->vp;
1605 	ASSERT_SEGLOCK(fs);
1606 
1607 	/*
1608 	 * This used to be:
1609 	 *
1610 	 *  nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1611 	 *
1612 	 * that is, it allowed for the possibility that start_lbp did
1613 	 * not point to the beginning of the finfo block pointer area.
1614 	 * This particular formulation is six kinds of painful in the
1615 	 * lfs64 world where we have two sizes of block pointer, so
1616 	 * unless/until everything can be cleaned up to not move
1617 	 * start_lbp around but instead use an offset, we do the
1618 	 * following:
1619 	 *    1. Get NEXT_FINFO(sp->fip). This is the same pointer as
1620 	 * &sp->fip->fi_blocks[sp->fip->fi_nblocks], just the wrong
1621 	 * type. (Ugh.)
1622 	 *    2. Cast it to void *, then assign it to a temporary
1623 	 * union lfs_blocks.
1624 	 *    3. Subtract start_lbp from that.
1625 	 *    4. Save the value of nblocks in blocks_orig so we can
1626 	 * assert below that it hasn't changed without repeating this
1627 	 * rubbish.
1628 	 *
1629 	 * XXX.
1630 	 */
1631 	lfs_blocks_fromvoid(fs, &tmpptr, (void *)NEXT_FINFO(fs, sp->fip));
1632 	nblocks = lfs_blocks_sub(fs, &tmpptr, &sp->start_lbp);
1633 	nblocks_orig = nblocks;
1634 
1635 	KASSERT(nblocks >= 0);
1636 	KASSERT(vp != NULL);
1637 	if (nblocks == 0)
1638 		return;
1639 
1640 	/*
1641 	 * This count may be high due to oversize blocks from lfs_gop_write.
1642 	 * Correct for this. (XXX we should be able to keep track of these.)
1643 	 */
1644 	for (i = 0; i < nblocks; i++) {
1645 		if (sp->start_bpp[i] == NULL) {
1646 			DLOG((DLOG_SEG, "lfs_updatemeta: nblocks = %d, not %d\n", i, nblocks));
1647 			nblocks = i;
1648 			break;
1649 		}
1650 		num = howmany(sp->start_bpp[i]->b_bcount, lfs_sb_getbsize(fs));
1651 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1652 		nblocks -= num - 1;
1653 	}
1654 
1655 #if 0
1656 	/* pre-lfs64 assertion */
1657 	KASSERT(vp->v_type == VREG ||
1658 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1659 #else
1660 	KASSERT(vp->v_type == VREG || nblocks == nblocks_orig);
1661 #endif
1662 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1663 
1664 	/*
1665 	 * Sort the blocks.
1666 	 *
1667 	 * We have to sort even if the blocks come from the
1668 	 * cleaner, because there might be other pending blocks on the
1669 	 * same inode...and if we don't sort, and there are fragments
1670 	 * present, blocks may be written in the wrong place.
1671 	 */
1672 	lfs_shellsort(fs, sp->start_bpp, &sp->start_lbp, nblocks, lfs_sb_getbsize(fs));
1673 
1674 	/*
1675 	 * Record the length of the last block in case it's a fragment.
1676 	 * If there are indirect blocks present, they sort last.  An
1677 	 * indirect block will be lfs_bsize and its presence indicates
1678 	 * that you cannot have fragments.
1679 	 *
1680 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1681 	 * XXX a later indirect block.	This will continue to be
1682 	 * XXX true until lfs_markv is fixed to do everything with
1683 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1684 	 */
1685 	lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1686 		lfs_sb_getbmask(fs)) + 1;
1687 	lfs_fi_setlastlength(fs, sp->fip, lastlength);
1688 
1689 	/*
1690 	 * Assign disk addresses, and update references to the logical
1691 	 * block and the segment usage information.
1692 	 */
1693 	for (i = nblocks; i--; ++sp->start_bpp) {
1694 		sbp = *sp->start_bpp;
1695 		lbn = lfs_blocks_get(fs, &sp->start_lbp, 0);
1696 		KASSERT(sbp->b_lblkno == lbn);
1697 
1698 		sbp->b_blkno = LFS_FSBTODB(fs, lfs_sb_getoffset(fs));
1699 
1700 		/*
1701 		 * If we write a frag in the wrong place, the cleaner won't
1702 		 * be able to correctly identify its size later, and the
1703 		 * segment will be uncleanable.	 (Even worse, it will assume
1704 		 * that the indirect block that actually ends the list
1705 		 * is of a smaller size!)
1706 		 */
1707 		if ((sbp->b_bcount & lfs_sb_getbmask(fs)) && i != 0)
1708 			panic("lfs_updatemeta: fragment is not last block");
1709 
1710 		/*
1711 		 * For each subblock in this possibly oversized block,
1712 		 * update its address on disk.
1713 		 */
1714 		KASSERT(lbn >= 0 || sbp->b_bcount == lfs_sb_getbsize(fs));
1715 		KASSERT(vp == sbp->b_vp);
1716 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1717 		     bytesleft -= lfs_sb_getbsize(fs)) {
1718 			size = MIN(bytesleft, lfs_sb_getbsize(fs));
1719 			bb = lfs_numfrags(fs, size);
1720 			lbn = lfs_blocks_get(fs, &sp->start_lbp, 0);
1721 			lfs_blocks_inc(fs, &sp->start_lbp);
1722 			lfs_update_single(fs, sp, sp->vp, lbn, lfs_sb_getoffset(fs),
1723 			    size);
1724 			lfs_sb_addoffset(fs, bb);
1725 		}
1726 
1727 	}
1728 
1729 	/* This inode has been modified */
1730 	LFS_SET_UINO(VTOI(vp), IN_MODIFIED);
1731 }
1732 
1733 /*
1734  * Move lfs_offset to a segment earlier than newsn.
1735  */
1736 int
1737 lfs_rewind(struct lfs *fs, int newsn)
1738 {
1739 	int sn, osn, isdirty;
1740 	struct buf *bp;
1741 	SEGUSE *sup;
1742 
1743 	ASSERT_SEGLOCK(fs);
1744 
1745 	osn = lfs_dtosn(fs, lfs_sb_getoffset(fs));
1746 	if (osn < newsn)
1747 		return 0;
1748 
1749 	/* lfs_avail eats the remaining space in this segment */
1750 	lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)));
1751 
1752 	/* Find a low-numbered segment */
1753 	for (sn = 0; sn < lfs_sb_getnseg(fs); ++sn) {
1754 		LFS_SEGENTRY(sup, fs, sn, bp);
1755 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1756 		brelse(bp, 0);
1757 
1758 		if (!isdirty)
1759 			break;
1760 	}
1761 	if (sn == lfs_sb_getnseg(fs))
1762 		panic("lfs_rewind: no clean segments");
1763 	if (newsn >= 0 && sn >= newsn)
1764 		return ENOENT;
1765 	lfs_sb_setnextseg(fs, lfs_sntod(fs, sn));
1766 	lfs_newseg(fs);
1767 	lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs));
1768 
1769 	return 0;
1770 }
1771 
1772 /*
1773  * Start a new partial segment.
1774  *
1775  * Return 1 when we entered to a new segment.
1776  * Otherwise, return 0.
1777  */
1778 int
1779 lfs_initseg(struct lfs *fs)
1780 {
1781 	struct segment *sp = fs->lfs_sp;
1782 	SEGSUM *ssp;
1783 	struct buf *sbp;	/* buffer for SEGSUM */
1784 	int repeat = 0;		/* return value */
1785 
1786 	ASSERT_SEGLOCK(fs);
1787 	/* Advance to the next segment. */
1788 	if (!LFS_PARTIAL_FITS(fs)) {
1789 		SEGUSE *sup;
1790 		struct buf *bp;
1791 
1792 		/* lfs_avail eats the remaining space */
1793 		lfs_sb_subavail(fs, lfs_sb_getfsbpseg(fs) - (lfs_sb_getoffset(fs) -
1794 						   lfs_sb_getcurseg(fs)));
1795 		/* Wake up any cleaning procs waiting on this file system. */
1796 		lfs_wakeup_cleaner(fs);
1797 		lfs_newseg(fs);
1798 		repeat = 1;
1799 		lfs_sb_setoffset(fs, lfs_sb_getcurseg(fs));
1800 
1801 		sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs));
1802 		sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs));
1803 
1804 		/*
1805 		 * If the segment contains a superblock, update the offset
1806 		 * and summary address to skip over it.
1807 		 */
1808 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1809 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1810 			lfs_sb_addoffset(fs, lfs_btofsb(fs, LFS_SBPAD));
1811 			sp->seg_bytes_left -= LFS_SBPAD;
1812 		}
1813 		brelse(bp, 0);
1814 		/* Segment zero could also contain the labelpad */
1815 		if (lfs_sb_getversion(fs) > 1 && sp->seg_number == 0 &&
1816 		    lfs_sb_gets0addr(fs) < lfs_btofsb(fs, LFS_LABELPAD)) {
1817 			lfs_sb_addoffset(fs,
1818 			    lfs_btofsb(fs, LFS_LABELPAD) - lfs_sb_gets0addr(fs));
1819 			sp->seg_bytes_left -=
1820 			    LFS_LABELPAD - lfs_fsbtob(fs, lfs_sb_gets0addr(fs));
1821 		}
1822 	} else {
1823 		sp->seg_number = lfs_dtosn(fs, lfs_sb_getcurseg(fs));
1824 		sp->seg_bytes_left = lfs_fsbtob(fs, lfs_sb_getfsbpseg(fs) -
1825 				      (lfs_sb_getoffset(fs) - lfs_sb_getcurseg(fs)));
1826 	}
1827 	lfs_sb_setlastpseg(fs, lfs_sb_getoffset(fs));
1828 
1829 	/* Record first address of this partial segment */
1830 	if (sp->seg_flags & SEGM_CLEAN) {
1831 		fs->lfs_cleanint[fs->lfs_cleanind] = lfs_sb_getoffset(fs);
1832 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1833 			/* "1" is the artificial inc in lfs_seglock */
1834 			mutex_enter(&lfs_lock);
1835 			while (fs->lfs_iocount > 1) {
1836 				mtsleep(&fs->lfs_iocount, PRIBIO + 1,
1837 				    "lfs_initseg", 0, &lfs_lock);
1838 			}
1839 			mutex_exit(&lfs_lock);
1840 			fs->lfs_cleanind = 0;
1841 		}
1842 	}
1843 
1844 	sp->fs = fs;
1845 	sp->ibp = NULL;
1846 	sp->idp = NULL;
1847 	sp->ninodes = 0;
1848 	sp->ndupino = 0;
1849 
1850 	sp->cbpp = sp->bpp;
1851 
1852 	/* Get a new buffer for SEGSUM */
1853 	sbp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1854 	    LFS_FSBTODB(fs, lfs_sb_getoffset(fs)), lfs_sb_getsumsize(fs), LFS_NB_SUMMARY);
1855 
1856 	/* ... and enter it into the buffer list. */
1857 	*sp->cbpp = sbp;
1858 	sp->cbpp++;
1859 	lfs_sb_addoffset(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs)));
1860 
1861 	sp->start_bpp = sp->cbpp;
1862 
1863 	/* Set point to SEGSUM, initialize it. */
1864 	ssp = sp->segsum = sbp->b_data;
1865 	memset(ssp, 0, lfs_sb_getsumsize(fs));
1866 	lfs_ss_setnext(fs, ssp, lfs_sb_getnextseg(fs));
1867 	lfs_ss_setnfinfo(fs, ssp, 0);
1868 	lfs_ss_setninos(fs, ssp, 0);
1869 	lfs_ss_setmagic(fs, ssp, SS_MAGIC);
1870 
1871 	/* Set pointer to first FINFO, initialize it. */
1872 	sp->fip = SEGSUM_FINFOBASE(fs, sp->segsum);
1873 	lfs_fi_setnblocks(fs, sp->fip, 0);
1874 	lfs_fi_setlastlength(fs, sp->fip, 0);
1875 	lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip);
1876 
1877 	sp->seg_bytes_left -= lfs_sb_getsumsize(fs);
1878 	sp->sum_bytes_left = lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs);
1879 
1880 	return (repeat);
1881 }
1882 
1883 /*
1884  * Remove SEGUSE_INVAL from all segments.
1885  */
1886 void
1887 lfs_unset_inval_all(struct lfs *fs)
1888 {
1889 	SEGUSE *sup;
1890 	struct buf *bp;
1891 	int i;
1892 
1893 	for (i = 0; i < lfs_sb_getnseg(fs); i++) {
1894 		LFS_SEGENTRY(sup, fs, i, bp);
1895 		if (sup->su_flags & SEGUSE_INVAL) {
1896 			sup->su_flags &= ~SEGUSE_INVAL;
1897 			LFS_WRITESEGENTRY(sup, fs, i, bp);
1898 		} else
1899 			brelse(bp, 0);
1900 	}
1901 }
1902 
1903 /*
1904  * Return the next segment to write.
1905  */
1906 void
1907 lfs_newseg(struct lfs *fs)
1908 {
1909 	CLEANERINFO *cip;
1910 	SEGUSE *sup;
1911 	struct buf *bp;
1912 	int curseg, isdirty, sn, skip_inval;
1913 
1914 	ASSERT_SEGLOCK(fs);
1915 
1916 	/* Honor LFCNWRAPSTOP */
1917 	mutex_enter(&lfs_lock);
1918 	while (lfs_sb_getnextseg(fs) < lfs_sb_getcurseg(fs) && fs->lfs_nowrap) {
1919 		if (fs->lfs_wrappass) {
1920 			log(LOG_NOTICE, "%s: wrappass=%d\n",
1921 				lfs_sb_getfsmnt(fs), fs->lfs_wrappass);
1922 			fs->lfs_wrappass = 0;
1923 			break;
1924 		}
1925 		fs->lfs_wrapstatus = LFS_WRAP_WAITING;
1926 		wakeup(&fs->lfs_nowrap);
1927 		log(LOG_NOTICE, "%s: waiting at log wrap\n", lfs_sb_getfsmnt(fs));
1928 		mtsleep(&fs->lfs_wrappass, PVFS, "newseg", 10 * hz,
1929 			&lfs_lock);
1930 	}
1931 	fs->lfs_wrapstatus = LFS_WRAP_GOING;
1932 	mutex_exit(&lfs_lock);
1933 
1934 	LFS_SEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp);
1935 	DLOG((DLOG_SU, "lfs_newseg: seg %d := 0 in newseg\n",
1936 	      lfs_dtosn(fs, lfs_sb_getnextseg(fs))));
1937 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1938 	sup->su_nbytes = 0;
1939 	sup->su_nsums = 0;
1940 	sup->su_ninos = 0;
1941 	LFS_WRITESEGENTRY(sup, fs, lfs_dtosn(fs, lfs_sb_getnextseg(fs)), bp);
1942 
1943 	LFS_CLEANERINFO(cip, fs, bp);
1944 	lfs_ci_shiftcleantodirty(fs, cip, 1);
1945 	lfs_sb_setnclean(fs, lfs_ci_getclean(fs, cip));
1946 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1947 
1948 	lfs_sb_setlastseg(fs, lfs_sb_getcurseg(fs));
1949 	lfs_sb_setcurseg(fs, lfs_sb_getnextseg(fs));
1950 	skip_inval = 1;
1951 	for (sn = curseg = lfs_dtosn(fs, lfs_sb_getcurseg(fs)) + lfs_sb_getinterleave(fs);;) {
1952 		sn = (sn + 1) % lfs_sb_getnseg(fs);
1953 
1954 		if (sn == curseg) {
1955 			if (skip_inval)
1956 				skip_inval = 0;
1957 			else
1958 				panic("lfs_nextseg: no clean segments");
1959 		}
1960 		LFS_SEGENTRY(sup, fs, sn, bp);
1961 		isdirty = sup->su_flags & (SEGUSE_DIRTY | (skip_inval ? SEGUSE_INVAL : 0));
1962 		/* Check SEGUSE_EMPTY as we go along */
1963 		if (isdirty && sup->su_nbytes == 0 &&
1964 		    !(sup->su_flags & SEGUSE_EMPTY))
1965 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1966 		else
1967 			brelse(bp, 0);
1968 
1969 		if (!isdirty)
1970 			break;
1971 	}
1972 	if (skip_inval == 0)
1973 		lfs_unset_inval_all(fs);
1974 
1975 	++fs->lfs_nactive;
1976 	lfs_sb_setnextseg(fs, lfs_sntod(fs, sn));
1977 	if (lfs_dostats) {
1978 		++lfs_stats.segsused;
1979 	}
1980 }
1981 
1982 static struct buf *
1983 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr,
1984     int n)
1985 {
1986 	struct lfs_cluster *cl;
1987 	struct buf **bpp, *bp;
1988 
1989 	ASSERT_SEGLOCK(fs);
1990 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1991 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1992 	memset(cl, 0, sizeof(*cl));
1993 	cl->fs = fs;
1994 	cl->bpp = bpp;
1995 	cl->bufcount = 0;
1996 	cl->bufsize = 0;
1997 
1998 	/* If this segment is being written synchronously, note that */
1999 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
2000 		cl->flags |= LFS_CL_SYNC;
2001 		cl->seg = fs->lfs_sp;
2002 		++cl->seg->seg_iocount;
2003 	}
2004 
2005 	/* Get an empty buffer header, or maybe one with something on it */
2006 	bp = getiobuf(vp, true);
2007 	bp->b_dev = NODEV;
2008 	bp->b_blkno = bp->b_lblkno = addr;
2009 	bp->b_iodone = lfs_cluster_callback;
2010 	bp->b_private = cl;
2011 
2012 	return bp;
2013 }
2014 
2015 int
2016 lfs_writeseg(struct lfs *fs, struct segment *sp)
2017 {
2018 	struct buf **bpp, *bp, *cbp, *newbp, *unbusybp;
2019 	SEGUSE *sup;
2020 	SEGSUM *ssp;
2021 	int i;
2022 	int do_again, nblocks, byteoffset;
2023 	size_t el_size;
2024 	struct lfs_cluster *cl;
2025 	u_short ninos;
2026 	struct vnode *devvp;
2027 	char *p = NULL;
2028 	struct vnode *vp;
2029 	unsigned ibindex, iblimit;
2030 	int changed;
2031 	u_int32_t sum;
2032 	size_t sumstart;
2033 #ifdef DEBUG
2034 	FINFO *fip;
2035 	int findex;
2036 #endif
2037 
2038 	ASSERT_SEGLOCK(fs);
2039 
2040 	ssp = (SEGSUM *)sp->segsum;
2041 
2042 	/*
2043 	 * If there are no buffers other than the segment summary to write,
2044 	 * don't do anything.  If we are the end of a dirop sequence, however,
2045 	 * write the empty segment summary anyway, to help out the
2046 	 * roll-forward agent.
2047 	 */
2048 	if ((nblocks = sp->cbpp - sp->bpp) == 1) {
2049 		if ((lfs_ss_getflags(fs, ssp) & (SS_DIROP | SS_CONT)) != SS_DIROP)
2050 			return 0;
2051 	}
2052 
2053 	/* Note if partial segment is being written by the cleaner */
2054 	if (sp->seg_flags & SEGM_CLEAN)
2055 		lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_CLEAN);
2056 
2057 	/* Note if we are writing to reclaim */
2058 	if (sp->seg_flags & SEGM_RECLAIM) {
2059 		lfs_ss_setflags(fs, ssp, lfs_ss_getflags(fs, ssp) | SS_RECLAIM);
2060 		lfs_ss_setreclino(fs, ssp, fs->lfs_reclino);
2061 	}
2062 
2063 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2064 
2065 	/* Update the segment usage information. */
2066 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
2067 
2068 	/* Loop through all blocks, except the segment summary. */
2069 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
2070 		if ((*bpp)->b_vp != devvp) {
2071 			sup->su_nbytes += (*bpp)->b_bcount;
2072 			DLOG((DLOG_SU, "seg %" PRIu32 " += %ld for ino %d"
2073 			      " lbn %" PRId64 " db 0x%" PRIx64 "\n",
2074 			      sp->seg_number, (*bpp)->b_bcount,
2075 			      VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
2076 			      (*bpp)->b_blkno));
2077 		}
2078 	}
2079 
2080 #ifdef DEBUG
2081 	/* Check for zero-length and zero-version FINFO entries. */
2082 	fip = SEGSUM_FINFOBASE(fs, ssp);
2083 	for (findex = 0; findex < lfs_ss_getnfinfo(fs, ssp); findex++) {
2084 		KDASSERT(lfs_fi_getnblocks(fs, fip) > 0);
2085 		KDASSERT(lfs_fi_getversion(fs, fip) > 0);
2086 		fip = NEXT_FINFO(fs, fip);
2087 	}
2088 #endif /* DEBUG */
2089 
2090 	ninos = (lfs_ss_getninos(fs, ssp) + LFS_INOPB(fs) - 1) / LFS_INOPB(fs);
2091 	DLOG((DLOG_SU, "seg %d += %d for %d inodes\n",
2092 	      sp->seg_number,
2093 	      lfs_ss_getninos(fs, ssp) * DINOSIZE(fs),
2094 	      lfs_ss_getninos(fs, ssp)));
2095 	sup->su_nbytes += lfs_ss_getninos(fs, ssp) * DINOSIZE(fs);
2096 	/* sup->su_nbytes += lfs_sb_getsumsize(fs); */
2097 	if (lfs_sb_getversion(fs) == 1)
2098 		sup->su_olastmod = time_second;
2099 	else
2100 		sup->su_lastmod = time_second;
2101 	sup->su_ninos += ninos;
2102 	++sup->su_nsums;
2103 	lfs_sb_subavail(fs, lfs_btofsb(fs, lfs_sb_getsumsize(fs)));
2104 
2105 	do_again = !(bp->b_flags & B_GATHERED);
2106 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
2107 
2108 	/*
2109 	 * Mark blocks BC_BUSY, to prevent then from being changed between
2110 	 * the checksum computation and the actual write.
2111 	 *
2112 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
2113 	 * there are any, replace them with copies that have UNASSIGNED
2114 	 * instead.
2115 	 */
2116 	mutex_enter(&bufcache_lock);
2117 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
2118 		++bpp;
2119 		bp = *bpp;
2120 		if (bp->b_iodone != NULL) {	 /* UBC or malloced buffer */
2121 			bp->b_cflags |= BC_BUSY;
2122 			continue;
2123 		}
2124 
2125 		while (bp->b_cflags & BC_BUSY) {
2126 			DLOG((DLOG_SEG, "lfs_writeseg: avoiding potential"
2127 			      " data summary corruption for ino %d, lbn %"
2128 			      PRId64 "\n",
2129 			      VTOI(bp->b_vp)->i_number, bp->b_lblkno));
2130 			bp->b_cflags |= BC_WANTED;
2131 			cv_wait(&bp->b_busy, &bufcache_lock);
2132 		}
2133 		bp->b_cflags |= BC_BUSY;
2134 		mutex_exit(&bufcache_lock);
2135 		unbusybp = NULL;
2136 
2137 		/*
2138 		 * Check and replace indirect block UNWRITTEN bogosity.
2139 		 * XXX See comment in lfs_writefile.
2140 		 */
2141 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
2142 		   lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din) !=
2143 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
2144 			DLOG((DLOG_VNODE, "lfs_writeseg: cleansing ino %d (%jd != %d)\n",
2145 			      VTOI(bp->b_vp)->i_number,
2146 			      (intmax_t)VTOI(bp->b_vp)->i_lfs_effnblks,
2147 			      lfs_dino_getblocks(fs, VTOI(bp->b_vp)->i_din)));
2148 			/* Make a copy we'll make changes to */
2149 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
2150 					   bp->b_bcount, LFS_NB_IBLOCK);
2151 			newbp->b_blkno = bp->b_blkno;
2152 			memcpy(newbp->b_data, bp->b_data,
2153 			       newbp->b_bcount);
2154 
2155 			changed = 0;
2156 			iblimit = newbp->b_bcount / LFS_BLKPTRSIZE(fs);
2157 			for (ibindex = 0; ibindex < iblimit; ibindex++) {
2158 				if (lfs_iblock_get(fs, newbp->b_data, ibindex) == UNWRITTEN) {
2159 					++changed;
2160 					lfs_iblock_set(fs, newbp->b_data,
2161 						       ibindex, 0);
2162 				}
2163 			}
2164 			/*
2165 			 * Get rid of the old buffer.  Don't mark it clean,
2166 			 * though, if it still has dirty data on it.
2167 			 */
2168 			if (changed) {
2169 				DLOG((DLOG_SEG, "lfs_writeseg: replacing UNWRITTEN(%d):"
2170 				      " bp = %p newbp = %p\n", changed, bp,
2171 				      newbp));
2172 				*bpp = newbp;
2173 				bp->b_flags &= ~B_GATHERED;
2174 				bp->b_error = 0;
2175 				if (bp->b_iodone != NULL) {
2176 					DLOG((DLOG_SEG, "lfs_writeseg: "
2177 					      "indir bp should not be B_CALL\n"));
2178 					biodone(bp);
2179 					bp = NULL;
2180 				} else {
2181 					/* Still on free list, leave it there */
2182 					unbusybp = bp;
2183 					/*
2184 					 * We have to re-decrement lfs_avail
2185 					 * since this block is going to come
2186 					 * back around to us in the next
2187 					 * segment.
2188 					 */
2189 					lfs_sb_subavail(fs,
2190 					    lfs_btofsb(fs, bp->b_bcount));
2191 				}
2192 			} else {
2193 				lfs_freebuf(fs, newbp);
2194 			}
2195 		}
2196 		mutex_enter(&bufcache_lock);
2197 		if (unbusybp != NULL) {
2198 			unbusybp->b_cflags &= ~BC_BUSY;
2199 			if (unbusybp->b_cflags & BC_WANTED)
2200 				cv_broadcast(&bp->b_busy);
2201 		}
2202 	}
2203 	mutex_exit(&bufcache_lock);
2204 
2205 	/*
2206 	 * Compute checksum across data and then across summary; the first
2207 	 * block (the summary block) is skipped.  Set the create time here
2208 	 * so that it's guaranteed to be later than the inode mod times.
2209 	 */
2210 	sum = 0;
2211 	if (lfs_sb_getversion(fs) == 1)
2212 		el_size = sizeof(u_long);
2213 	else
2214 		el_size = sizeof(u_int32_t);
2215 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
2216 		++bpp;
2217 		/* Loop through gop_write cluster blocks */
2218 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
2219 		     byteoffset += lfs_sb_getbsize(fs)) {
2220 #ifdef LFS_USE_BC_INVAL
2221 			if (((*bpp)->b_cflags & BC_INVAL) != 0 &&
2222 			    (*bpp)->b_iodone != NULL) {
2223 				if (copyin((void *)(*bpp)->b_saveaddr +
2224 					   byteoffset, dp, el_size)) {
2225 					panic("lfs_writeseg: copyin failed [1]:"
2226 						" ino %" PRIu64 " blk %" PRId64,
2227 						VTOI((*bpp)->b_vp)->i_number,
2228 						(*bpp)->b_lblkno);
2229 				}
2230 			} else
2231 #endif /* LFS_USE_BC_INVAL */
2232 			{
2233 				sum = lfs_cksum_part((char *)
2234 				    (*bpp)->b_data + byteoffset, el_size, sum);
2235 			}
2236 		}
2237 	}
2238 	if (lfs_sb_getversion(fs) == 1)
2239 		lfs_ss_setocreate(fs, ssp, time_second);
2240 	else {
2241 		lfs_ss_setcreate(fs, ssp, time_second);
2242 		lfs_sb_addserial(fs, 1);
2243 		lfs_ss_setserial(fs, ssp, lfs_sb_getserial(fs));
2244 		lfs_ss_setident(fs, ssp, lfs_sb_getident(fs));
2245 	}
2246 	lfs_ss_setdatasum(fs, ssp, lfs_cksum_fold(sum));
2247 	sumstart = lfs_ss_getsumstart(fs);
2248 	lfs_ss_setsumsum(fs, ssp, cksum((char *)ssp + sumstart,
2249 	    lfs_sb_getsumsize(fs) - sumstart));
2250 
2251 	mutex_enter(&lfs_lock);
2252 	lfs_sb_subbfree(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) +
2253 			  lfs_btofsb(fs, lfs_sb_getsumsize(fs))));
2254 	lfs_sb_adddmeta(fs, (lfs_btofsb(fs, ninos * lfs_sb_getibsize(fs)) +
2255 			  lfs_btofsb(fs, lfs_sb_getsumsize(fs))));
2256 	mutex_exit(&lfs_lock);
2257 
2258 	/*
2259 	 * When we simply write the blocks we lose a rotation for every block
2260 	 * written.  To avoid this problem, we cluster the buffers into a
2261 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
2262 	 * devices can handle, use that for the size of the chunks.
2263 	 *
2264 	 * Blocks that are already clusters (from GOP_WRITE), however, we
2265 	 * don't bother to copy into other clusters.
2266 	 */
2267 
2268 #define CHUNKSIZE MAXPHYS
2269 
2270 	if (devvp == NULL)
2271 		panic("devvp is NULL");
2272 	for (bpp = sp->bpp, i = nblocks; i;) {
2273 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
2274 		cl = cbp->b_private;
2275 
2276 		cbp->b_flags |= B_ASYNC;
2277 		cbp->b_cflags |= BC_BUSY;
2278 		cbp->b_bcount = 0;
2279 
2280 		KASSERTMSG((bpp - sp->bpp <=
2281 			(lfs_sb_getsumsize(fs) - SEGSUM_SIZE(fs))
2282 			/ sizeof(int32_t)),
2283 		    "lfs_writeseg: real bpp overwrite");
2284 		KASSERTMSG((bpp - sp->bpp <=
2285 			lfs_segsize(fs) / lfs_sb_getfsize(fs)),
2286 		    "lfs_writeseg: theoretical bpp overwrite");
2287 
2288 		/*
2289 		 * Construct the cluster.
2290 		 */
2291 		mutex_enter(&lfs_lock);
2292 		++fs->lfs_iocount;
2293 		mutex_exit(&lfs_lock);
2294 		while (i && cbp->b_bcount < CHUNKSIZE) {
2295 			bp = *bpp;
2296 
2297 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
2298 				break;
2299 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
2300 				break;
2301 
2302 			/* Clusters from GOP_WRITE are expedited */
2303 			if (bp->b_bcount > lfs_sb_getbsize(fs)) {
2304 				if (cbp->b_bcount > 0)
2305 					/* Put in its own buffer */
2306 					break;
2307 				else {
2308 					cbp->b_data = bp->b_data;
2309 				}
2310 			} else if (cbp->b_bcount == 0) {
2311 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
2312 							     LFS_NB_CLUSTER);
2313 				cl->flags |= LFS_CL_MALLOC;
2314 			}
2315 			KASSERTMSG((lfs_dtosn(fs, LFS_DBTOFSB(fs, bp->b_blkno +
2316 					btodb(bp->b_bcount - 1))) ==
2317 				sp->seg_number),
2318 			    "segment overwrite: blk size %d daddr %" PRIx64
2319 			    " not in seg %d\n",
2320 			    bp->b_bcount, bp->b_blkno,
2321 			    sp->seg_number);
2322 
2323 #ifdef LFS_USE_BC_INVAL
2324 			/*
2325 			 * Fake buffers from the cleaner are marked as BC_INVAL.
2326 			 * We need to copy the data from user space rather than
2327 			 * from the buffer indicated.
2328 			 * XXX == what do I do on an error?
2329 			 */
2330 			if ((bp->b_cflags & BC_INVAL) != 0 &&
2331 			    bp->b_iodone != NULL) {
2332 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
2333 					panic("lfs_writeseg: "
2334 					    "copyin failed [2]");
2335 			} else
2336 #endif /* LFS_USE_BC_INVAL */
2337 			if (cl->flags & LFS_CL_MALLOC) {
2338 				/* copy data into our cluster. */
2339 				memcpy(p, bp->b_data, bp->b_bcount);
2340 				p += bp->b_bcount;
2341 			}
2342 
2343 			cbp->b_bcount += bp->b_bcount;
2344 			cl->bufsize += bp->b_bcount;
2345 
2346 			bp->b_flags &= ~B_READ;
2347 			bp->b_error = 0;
2348 			cl->bpp[cl->bufcount++] = bp;
2349 
2350 			vp = bp->b_vp;
2351 			mutex_enter(&bufcache_lock);
2352 			mutex_enter(vp->v_interlock);
2353 			bp->b_oflags &= ~(BO_DELWRI | BO_DONE);
2354 			reassignbuf(bp, vp);
2355 			vp->v_numoutput++;
2356 			mutex_exit(vp->v_interlock);
2357 			mutex_exit(&bufcache_lock);
2358 
2359 			bpp++;
2360 			i--;
2361 		}
2362 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2363 			BIO_SETPRIO(cbp, BPRIO_TIMECRITICAL);
2364 		else
2365 			BIO_SETPRIO(cbp, BPRIO_TIMELIMITED);
2366 		mutex_enter(devvp->v_interlock);
2367 		devvp->v_numoutput++;
2368 		mutex_exit(devvp->v_interlock);
2369 		VOP_STRATEGY(devvp, cbp);
2370 		curlwp->l_ru.ru_oublock++;
2371 	}
2372 
2373 	if (lfs_dostats) {
2374 		++lfs_stats.psegwrites;
2375 		lfs_stats.blocktot += nblocks - 1;
2376 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2377 			++lfs_stats.psyncwrites;
2378 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2379 			++lfs_stats.pcleanwrites;
2380 			lfs_stats.cleanblocks += nblocks - 1;
2381 		}
2382 	}
2383 
2384 	return (lfs_initseg(fs) || do_again);
2385 }
2386 
2387 void
2388 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2389 {
2390 	struct buf *bp;
2391 	struct vnode *devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2392 
2393 	ASSERT_MAYBE_SEGLOCK(fs);
2394 	if (fs->lfs_is64) {
2395 		KASSERT(fs->lfs_dlfs_u.u_64.dlfs_magic == LFS64_MAGIC);
2396 	} else {
2397 		KASSERT(fs->lfs_dlfs_u.u_32.dlfs_magic == LFS_MAGIC);
2398 	}
2399 	/*
2400 	 * If we can write one superblock while another is in
2401 	 * progress, we risk not having a complete checkpoint if we crash.
2402 	 * So, block here if a superblock write is in progress.
2403 	 */
2404 	mutex_enter(&lfs_lock);
2405 	while (fs->lfs_sbactive) {
2406 		mtsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0,
2407 			&lfs_lock);
2408 	}
2409 	fs->lfs_sbactive = daddr;
2410 	mutex_exit(&lfs_lock);
2411 
2412 	/* Set timestamp of this version of the superblock */
2413 	if (lfs_sb_getversion(fs) == 1)
2414 		lfs_sb_setotstamp(fs, time_second);
2415 	lfs_sb_settstamp(fs, time_second);
2416 
2417 	/* The next chunk of code relies on this assumption */
2418 	CTASSERT(sizeof(struct dlfs) == sizeof(struct dlfs64));
2419 
2420 	/* Checksum the superblock and copy it into a buffer. */
2421 	lfs_sb_setcksum(fs, lfs_sb_cksum(fs));
2422 	bp = lfs_newbuf(fs, devvp,
2423 	    LFS_FSBTODB(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2424 	memcpy(bp->b_data, &fs->lfs_dlfs_u, sizeof(struct dlfs));
2425 	memset((char *)bp->b_data + sizeof(struct dlfs), 0,
2426 	    LFS_SBPAD - sizeof(struct dlfs));
2427 
2428 	bp->b_cflags |= BC_BUSY;
2429 	bp->b_flags = (bp->b_flags & ~B_READ) | B_ASYNC;
2430 	bp->b_oflags &= ~(BO_DONE | BO_DELWRI);
2431 	bp->b_error = 0;
2432 	bp->b_iodone = lfs_supercallback;
2433 
2434 	if (fs->lfs_sp != NULL && fs->lfs_sp->seg_flags & SEGM_SYNC)
2435 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
2436 	else
2437 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
2438 	curlwp->l_ru.ru_oublock++;
2439 
2440 	mutex_enter(devvp->v_interlock);
2441 	devvp->v_numoutput++;
2442 	mutex_exit(devvp->v_interlock);
2443 
2444 	mutex_enter(&lfs_lock);
2445 	++fs->lfs_iocount;
2446 	mutex_exit(&lfs_lock);
2447 	VOP_STRATEGY(devvp, bp);
2448 }
2449 
2450 /*
2451  * Logical block number match routines used when traversing the dirty block
2452  * chain.
2453  */
2454 int
2455 lfs_match_fake(struct lfs *fs, struct buf *bp)
2456 {
2457 
2458 	ASSERT_SEGLOCK(fs);
2459 	return LFS_IS_MALLOC_BUF(bp);
2460 }
2461 
2462 #if 0
2463 int
2464 lfs_match_real(struct lfs *fs, struct buf *bp)
2465 {
2466 
2467 	ASSERT_SEGLOCK(fs);
2468 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2469 }
2470 #endif
2471 
2472 int
2473 lfs_match_data(struct lfs *fs, struct buf *bp)
2474 {
2475 
2476 	ASSERT_SEGLOCK(fs);
2477 	return (bp->b_lblkno >= 0);
2478 }
2479 
2480 int
2481 lfs_match_indir(struct lfs *fs, struct buf *bp)
2482 {
2483 	daddr_t lbn;
2484 
2485 	ASSERT_SEGLOCK(fs);
2486 	lbn = bp->b_lblkno;
2487 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 0);
2488 }
2489 
2490 int
2491 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2492 {
2493 	daddr_t lbn;
2494 
2495 	ASSERT_SEGLOCK(fs);
2496 	lbn = bp->b_lblkno;
2497 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 1);
2498 }
2499 
2500 int
2501 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2502 {
2503 	daddr_t lbn;
2504 
2505 	ASSERT_SEGLOCK(fs);
2506 	lbn = bp->b_lblkno;
2507 	return (lbn < 0 && (-lbn - ULFS_NDADDR) % LFS_NINDIR(fs) == 2);
2508 }
2509 
2510 static void
2511 lfs_free_aiodone(struct buf *bp)
2512 {
2513 	struct lfs *fs;
2514 
2515 	KERNEL_LOCK(1, curlwp);
2516 	fs = bp->b_private;
2517 	ASSERT_NO_SEGLOCK(fs);
2518 	lfs_freebuf(fs, bp);
2519 	KERNEL_UNLOCK_LAST(curlwp);
2520 }
2521 
2522 static void
2523 lfs_super_aiodone(struct buf *bp)
2524 {
2525 	struct lfs *fs;
2526 
2527 	KERNEL_LOCK(1, curlwp);
2528 	fs = bp->b_private;
2529 	ASSERT_NO_SEGLOCK(fs);
2530 	mutex_enter(&lfs_lock);
2531 	fs->lfs_sbactive = 0;
2532 	if (--fs->lfs_iocount <= 1)
2533 		wakeup(&fs->lfs_iocount);
2534 	wakeup(&fs->lfs_sbactive);
2535 	mutex_exit(&lfs_lock);
2536 	lfs_freebuf(fs, bp);
2537 	KERNEL_UNLOCK_LAST(curlwp);
2538 }
2539 
2540 static void
2541 lfs_cluster_aiodone(struct buf *bp)
2542 {
2543 	struct lfs_cluster *cl;
2544 	struct lfs *fs;
2545 	struct buf *tbp, *fbp;
2546 	struct vnode *vp, *devvp, *ovp;
2547 	struct inode *ip;
2548 	int error;
2549 
2550 	KERNEL_LOCK(1, curlwp);
2551 
2552 	error = bp->b_error;
2553 	cl = bp->b_private;
2554 	fs = cl->fs;
2555 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2556 	ASSERT_NO_SEGLOCK(fs);
2557 
2558 	/* Put the pages back, and release the buffer */
2559 	while (cl->bufcount--) {
2560 		tbp = cl->bpp[cl->bufcount];
2561 		KASSERT(tbp->b_cflags & BC_BUSY);
2562 		if (error) {
2563 			tbp->b_error = error;
2564 		}
2565 
2566 		/*
2567 		 * We're done with tbp.	 If it has not been re-dirtied since
2568 		 * the cluster was written, free it.  Otherwise, keep it on
2569 		 * the locked list to be written again.
2570 		 */
2571 		vp = tbp->b_vp;
2572 
2573 		tbp->b_flags &= ~B_GATHERED;
2574 
2575 #ifdef DEBUG
2576 		if ((tbp)->b_vp == (fs)->lfs_ivnode)
2577 			LFS_ENTER_LOG("clear", __FILE__, __LINE__,
2578 			    tbp->b_lblkno, tbp->b_flags, curproc->p_pid);
2579 #endif
2580 
2581 		mutex_enter(&bufcache_lock);
2582 		if (tbp->b_iodone == NULL) {
2583 			KASSERT(tbp->b_flags & B_LOCKED);
2584 			bremfree(tbp);
2585 			if (vp) {
2586 				mutex_enter(vp->v_interlock);
2587 				reassignbuf(tbp, vp);
2588 				mutex_exit(vp->v_interlock);
2589 			}
2590 			tbp->b_flags |= B_ASYNC; /* for biodone */
2591 		}
2592 
2593 		if ((tbp->b_flags & B_LOCKED) && !(tbp->b_oflags & BO_DELWRI))
2594 			LFS_UNLOCK_BUF(tbp);
2595 
2596 		if (tbp->b_oflags & BO_DONE) {
2597 			DLOG((DLOG_SEG, "blk %d biodone already (flags %lx)\n",
2598 				cl->bufcount, (long)tbp->b_flags));
2599 		}
2600 
2601 		if (tbp->b_iodone != NULL && !LFS_IS_MALLOC_BUF(tbp)) {
2602 			/*
2603 			 * A buffer from the page daemon.
2604 			 * We use the same iodone as it does,
2605 			 * so we must manually disassociate its
2606 			 * buffers from the vp.
2607 			 */
2608 			if ((ovp = tbp->b_vp) != NULL) {
2609 				/* This is just silly */
2610 				mutex_enter(ovp->v_interlock);
2611 				brelvp(tbp);
2612 				mutex_exit(ovp->v_interlock);
2613 				tbp->b_vp = vp;
2614 				tbp->b_objlock = vp->v_interlock;
2615 			}
2616 			/* Put it back the way it was */
2617 			tbp->b_flags |= B_ASYNC;
2618 			/* Master buffers have BC_AGE */
2619 			if (tbp->b_private == tbp)
2620 				tbp->b_cflags |= BC_AGE;
2621 		}
2622 		mutex_exit(&bufcache_lock);
2623 
2624 		biodone(tbp);
2625 
2626 		/*
2627 		 * If this is the last block for this vnode, but
2628 		 * there are other blocks on its dirty list,
2629 		 * set IN_MODIFIED/IN_CLEANING depending on what
2630 		 * sort of block.  Only do this for our mount point,
2631 		 * not for, e.g., inode blocks that are attached to
2632 		 * the devvp.
2633 		 * XXX KS - Shouldn't we set *both* if both types
2634 		 * of blocks are present (traverse the dirty list?)
2635 		 */
2636 		mutex_enter(vp->v_interlock);
2637 		mutex_enter(&lfs_lock);
2638 		if (vp != devvp && vp->v_numoutput == 0 &&
2639 		    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2640 			ip = VTOI(vp);
2641 			DLOG((DLOG_SEG, "lfs_cluster_aiodone: mark ino %d\n",
2642 			       ip->i_number));
2643 			if (LFS_IS_MALLOC_BUF(fbp))
2644 				LFS_SET_UINO(ip, IN_CLEANING);
2645 			else
2646 				LFS_SET_UINO(ip, IN_MODIFIED);
2647 		}
2648 		cv_broadcast(&vp->v_cv);
2649 		mutex_exit(&lfs_lock);
2650 		mutex_exit(vp->v_interlock);
2651 	}
2652 
2653 	/* Fix up the cluster buffer, and release it */
2654 	if (cl->flags & LFS_CL_MALLOC)
2655 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2656 	putiobuf(bp);
2657 
2658 	/* Note i/o done */
2659 	if (cl->flags & LFS_CL_SYNC) {
2660 		if (--cl->seg->seg_iocount == 0)
2661 			wakeup(&cl->seg->seg_iocount);
2662 	}
2663 	mutex_enter(&lfs_lock);
2664 	KASSERTMSG((fs->lfs_iocount != 0),
2665 	    "lfs_cluster_aiodone: zero iocount");
2666 	if (--fs->lfs_iocount <= 1)
2667 		wakeup(&fs->lfs_iocount);
2668 	mutex_exit(&lfs_lock);
2669 
2670 	KERNEL_UNLOCK_LAST(curlwp);
2671 
2672 	pool_put(&fs->lfs_bpppool, cl->bpp);
2673 	cl->bpp = NULL;
2674 	pool_put(&fs->lfs_clpool, cl);
2675 }
2676 
2677 static void
2678 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2679 {
2680 	/* reset b_iodone for when this is a single-buf i/o. */
2681 	bp->b_iodone = aiodone;
2682 
2683 	workqueue_enqueue(uvm.aiodone_queue, &bp->b_work, NULL);
2684 }
2685 
2686 static void
2687 lfs_cluster_callback(struct buf *bp)
2688 {
2689 
2690 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2691 }
2692 
2693 void
2694 lfs_supercallback(struct buf *bp)
2695 {
2696 
2697 	lfs_generic_callback(bp, lfs_super_aiodone);
2698 }
2699 
2700 /*
2701  * The only buffers that are going to hit these functions are the
2702  * segment write blocks, or the segment summaries, or the superblocks.
2703  *
2704  * All of the above are created by lfs_newbuf, and so do not need to be
2705  * released via brelse.
2706  */
2707 void
2708 lfs_callback(struct buf *bp)
2709 {
2710 
2711 	lfs_generic_callback(bp, lfs_free_aiodone);
2712 }
2713 
2714 /*
2715  * Shellsort (diminishing increment sort) from Data Structures and
2716  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2717  * see also Knuth Vol. 3, page 84.  The increments are selected from
2718  * formula (8), page 95.  Roughly O(N^3/2).
2719  */
2720 /*
2721  * This is our own private copy of shellsort because we want to sort
2722  * two parallel arrays (the array of buffer pointers and the array of
2723  * logical block numbers) simultaneously.  Note that we cast the array
2724  * of logical block numbers to a unsigned in this routine so that the
2725  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2726  */
2727 
2728 static void
2729 lfs_shellsort(struct lfs *fs,
2730 	      struct buf **bp_array, union lfs_blocks *lb_array,
2731 	      int nmemb, int size)
2732 {
2733 	static int __rsshell_increments[] = { 4, 1, 0 };
2734 	int incr, *incrp, t1, t2;
2735 	struct buf *bp_temp;
2736 
2737 #ifdef DEBUG
2738 	incr = 0;
2739 	for (t1 = 0; t1 < nmemb; t1++) {
2740 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2741 			if (lfs_blocks_get(fs, lb_array, incr++) != bp_array[t1]->b_lblkno + t2) {
2742 				/* dump before panic */
2743 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2744 				    nmemb, size);
2745 				incr = 0;
2746 				for (t1 = 0; t1 < nmemb; t1++) {
2747 					const struct buf *bp = bp_array[t1];
2748 
2749 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2750 					    PRIu64 "\n", t1,
2751 					    (uint64_t)bp->b_bcount,
2752 					    (uint64_t)bp->b_lblkno);
2753 					printf("lbns:");
2754 					for (t2 = 0; t2 * size < bp->b_bcount;
2755 					    t2++) {
2756 						printf(" %jd",
2757 						    (intmax_t)lfs_blocks_get(fs, lb_array, incr++));
2758 					}
2759 					printf("\n");
2760 				}
2761 				panic("lfs_shellsort: inconsistent input");
2762 			}
2763 		}
2764 	}
2765 #endif
2766 
2767 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2768 		for (t1 = incr; t1 < nmemb; ++t1)
2769 			for (t2 = t1 - incr; t2 >= 0;)
2770 				if ((u_int64_t)bp_array[t2]->b_lblkno >
2771 				    (u_int64_t)bp_array[t2 + incr]->b_lblkno) {
2772 					bp_temp = bp_array[t2];
2773 					bp_array[t2] = bp_array[t2 + incr];
2774 					bp_array[t2 + incr] = bp_temp;
2775 					t2 -= incr;
2776 				} else
2777 					break;
2778 
2779 	/* Reform the list of logical blocks */
2780 	incr = 0;
2781 	for (t1 = 0; t1 < nmemb; t1++) {
2782 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2783 			lfs_blocks_set(fs, lb_array, incr++,
2784 				       bp_array[t1]->b_lblkno + t2);
2785 		}
2786 	}
2787 }
2788 
2789 /*
2790  * Set up an FINFO entry for a new file.  The fip pointer is assumed to
2791  * point at uninitialized space.
2792  */
2793 void
2794 lfs_acquire_finfo(struct lfs *fs, ino_t ino, int vers)
2795 {
2796 	struct segment *sp = fs->lfs_sp;
2797 	SEGSUM *ssp;
2798 
2799 	KASSERT(vers > 0);
2800 
2801 	if (sp->seg_bytes_left < lfs_sb_getbsize(fs) ||
2802 	    sp->sum_bytes_left < FINFOSIZE(fs) + LFS_BLKPTRSIZE(fs))
2803 		(void) lfs_writeseg(fs, fs->lfs_sp);
2804 
2805 	sp->sum_bytes_left -= FINFOSIZE(fs);
2806 	ssp = (SEGSUM *)sp->segsum;
2807 	lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) + 1);
2808 	lfs_fi_setnblocks(fs, sp->fip, 0);
2809 	lfs_fi_setino(fs, sp->fip, ino);
2810 	lfs_fi_setversion(fs, sp->fip, vers);
2811 }
2812 
2813 /*
2814  * Release the FINFO entry, either clearing out an unused entry or
2815  * advancing us to the next available entry.
2816  */
2817 void
2818 lfs_release_finfo(struct lfs *fs)
2819 {
2820 	struct segment *sp = fs->lfs_sp;
2821 	SEGSUM *ssp;
2822 
2823 	if (lfs_fi_getnblocks(fs, sp->fip) != 0) {
2824 		sp->fip = NEXT_FINFO(fs, sp->fip);
2825 		lfs_blocks_fromfinfo(fs, &sp->start_lbp, sp->fip);
2826 	} else {
2827 		/* XXX shouldn't this update sp->fip? */
2828 		sp->sum_bytes_left += FINFOSIZE(fs);
2829 		ssp = (SEGSUM *)sp->segsum;
2830 		lfs_ss_setnfinfo(fs, ssp, lfs_ss_getnfinfo(fs, ssp) - 1);
2831 	}
2832 }
2833