xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision eb7c1594f145c931049e1fd9eb056a5987e87e59)
1 /*	$NetBSD: lfs_segment.c,v 1.131 2003/08/07 16:34:37 agc Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Konrad E. Schroder <perseant@hhhh.org>.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the NetBSD
21  *	Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 /*
39  * Copyright (c) 1991, 1993
40  *	The Regents of the University of California.  All rights reserved.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
67  */
68 
69 #include <sys/cdefs.h>
70 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.131 2003/08/07 16:34:37 agc Exp $");
71 
72 #define ivndebug(vp,str) printf("ino %d: %s\n",VTOI(vp)->i_number,(str))
73 
74 #if defined(_KERNEL_OPT)
75 #include "opt_ddb.h"
76 #endif
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/namei.h>
81 #include <sys/kernel.h>
82 #include <sys/resourcevar.h>
83 #include <sys/file.h>
84 #include <sys/stat.h>
85 #include <sys/buf.h>
86 #include <sys/proc.h>
87 #include <sys/vnode.h>
88 #include <sys/mount.h>
89 
90 #include <miscfs/specfs/specdev.h>
91 #include <miscfs/fifofs/fifo.h>
92 
93 #include <ufs/ufs/inode.h>
94 #include <ufs/ufs/dir.h>
95 #include <ufs/ufs/ufsmount.h>
96 #include <ufs/ufs/ufs_extern.h>
97 
98 #include <ufs/lfs/lfs.h>
99 #include <ufs/lfs/lfs_extern.h>
100 
101 #include <uvm/uvm.h>
102 #include <uvm/uvm_extern.h>
103 
104 MALLOC_DEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
105 
106 extern int count_lock_queue(void);
107 extern struct simplelock vnode_free_list_slock;		/* XXX */
108 
109 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
110 static void lfs_super_aiodone(struct buf *);
111 static void lfs_cluster_aiodone(struct buf *);
112 static void lfs_cluster_callback(struct buf *);
113 
114 /*
115  * Determine if it's OK to start a partial in this segment, or if we need
116  * to go on to a new segment.
117  */
118 #define	LFS_PARTIAL_FITS(fs) \
119 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
120 	fragstofsb((fs), (fs)->lfs_frag))
121 
122 void	 lfs_callback(struct buf *);
123 int	 lfs_gather(struct lfs *, struct segment *,
124 	     struct vnode *, int (*)(struct lfs *, struct buf *));
125 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
126 void	 lfs_iset(struct inode *, daddr_t, time_t);
127 int	 lfs_match_fake(struct lfs *, struct buf *);
128 int	 lfs_match_data(struct lfs *, struct buf *);
129 int	 lfs_match_dindir(struct lfs *, struct buf *);
130 int	 lfs_match_indir(struct lfs *, struct buf *);
131 int	 lfs_match_tindir(struct lfs *, struct buf *);
132 void	 lfs_newseg(struct lfs *);
133 /* XXX ondisk32 */
134 void	 lfs_shellsort(struct buf **, int32_t *, int, int);
135 void	 lfs_supercallback(struct buf *);
136 void	 lfs_updatemeta(struct segment *);
137 int	 lfs_vref(struct vnode *);
138 void	 lfs_vunref(struct vnode *);
139 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
140 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
141 int	 lfs_writeseg(struct lfs *, struct segment *);
142 void	 lfs_writesuper(struct lfs *, daddr_t);
143 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
144 	    struct segment *sp, int dirops);
145 
146 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
147 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
148 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
149 int	lfs_dirvcount = 0;		/* # active dirops */
150 
151 /* Statistics Counters */
152 int lfs_dostats = 1;
153 struct lfs_stats lfs_stats;
154 
155 /* op values to lfs_writevnodes */
156 #define	VN_REG		0
157 #define	VN_DIROP	1
158 #define	VN_EMPTY	2
159 #define VN_CLEAN	3
160 
161 #define LFS_MAX_ACTIVE		10
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 	TIMEVAL_TO_TIMESPEC(&time, &ts);
176 	ip = VTOI(fs->lfs_ivnode);
177 	ip->i_ffs1_mtime = ts.tv_sec;
178 	ip->i_ffs1_mtimensec = ts.tv_nsec;
179 }
180 
181 /*
182  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
183  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
184  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
185  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
186  */
187 
188 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
189 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
190 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
191 
192 int
193 lfs_vflush(struct vnode *vp)
194 {
195 	struct inode *ip;
196 	struct lfs *fs;
197 	struct segment *sp;
198 	struct buf *bp, *nbp, *tbp, *tnbp;
199 	int error, s;
200 	int flushed;
201 #if 0
202 	int redo;
203 #endif
204 
205 	ip = VTOI(vp);
206 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
207 
208 	if (ip->i_flag & IN_CLEANING) {
209 #ifdef DEBUG_LFS
210 		ivndebug(vp,"vflush/in_cleaning");
211 #endif
212 		LFS_CLR_UINO(ip, IN_CLEANING);
213 		LFS_SET_UINO(ip, IN_MODIFIED);
214 
215 		/*
216 		 * Toss any cleaning buffers that have real counterparts
217 		 * to avoid losing new data.
218 		 */
219 		s = splbio();
220 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
221 			nbp = LIST_NEXT(bp, b_vnbufs);
222 			if (!LFS_IS_MALLOC_BUF(bp))
223 				continue;
224 			/*
225 			 * Look for pages matching the range covered
226 			 * by cleaning blocks.  It's okay if more dirty
227 			 * pages appear, so long as none disappear out
228 			 * from under us.
229 			 */
230 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
231 			    vp != fs->lfs_ivnode) {
232 				struct vm_page *pg;
233 				voff_t off;
234 
235 				simple_lock(&vp->v_interlock);
236 				for (off = lblktosize(fs, bp->b_lblkno);
237 				     off < lblktosize(fs, bp->b_lblkno + 1);
238 				     off += PAGE_SIZE) {
239 					pg = uvm_pagelookup(&vp->v_uobj, off);
240 					if (pg == NULL)
241 						continue;
242 					if ((pg->flags & PG_CLEAN) == 0 ||
243 					    pmap_is_modified(pg)) {
244 						fs->lfs_avail += btofsb(fs,
245 							bp->b_bcount);
246 						wakeup(&fs->lfs_avail);
247 						lfs_freebuf(fs, bp);
248 						bp = NULL;
249 						goto nextbp;
250 					}
251 				}
252 				simple_unlock(&vp->v_interlock);
253 			}
254 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
255 			    tbp = tnbp)
256 			{
257 				tnbp = LIST_NEXT(tbp, b_vnbufs);
258 				if (tbp->b_vp == bp->b_vp
259 				   && tbp->b_lblkno == bp->b_lblkno
260 				   && tbp != bp)
261 				{
262 					fs->lfs_avail += btofsb(fs,
263 						bp->b_bcount);
264 					wakeup(&fs->lfs_avail);
265 					lfs_freebuf(fs, bp);
266 					bp = NULL;
267 					break;
268 				}
269 			}
270 		    nextbp:
271 			;
272 		}
273 		splx(s);
274 	}
275 
276 	/* If the node is being written, wait until that is done */
277 	s = splbio();
278 	if (WRITEINPROG(vp)) {
279 #ifdef DEBUG_LFS
280 		ivndebug(vp,"vflush/writeinprog");
281 #endif
282 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
283 	}
284 	splx(s);
285 
286 	/* Protect against VXLOCK deadlock in vinvalbuf() */
287 	lfs_seglock(fs, SEGM_SYNC);
288 
289 	/* If we're supposed to flush a freed inode, just toss it */
290 	/* XXX - seglock, so these buffers can't be gathered, right? */
291 	if (ip->i_mode == 0) {
292 		printf("lfs_vflush: ino %d is freed, not flushing\n",
293 			ip->i_number);
294 		s = splbio();
295 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
296 			nbp = LIST_NEXT(bp, b_vnbufs);
297 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
298 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
299 				wakeup(&fs->lfs_avail);
300 			}
301 			/* Copied from lfs_writeseg */
302 			if (bp->b_flags & B_CALL) {
303 				biodone(bp);
304 			} else {
305 				bremfree(bp);
306 				LFS_UNLOCK_BUF(bp);
307 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
308 					 B_GATHERED);
309 				bp->b_flags |= B_DONE;
310 				reassignbuf(bp, vp);
311 				brelse(bp);
312 			}
313 		}
314 		splx(s);
315 		LFS_CLR_UINO(ip, IN_CLEANING);
316 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
317 		ip->i_flag &= ~IN_ALLMOD;
318 		printf("lfs_vflush: done not flushing ino %d\n",
319 			ip->i_number);
320 		lfs_segunlock(fs);
321 		return 0;
322 	}
323 
324 	SET_FLUSHING(fs,vp);
325 	if (fs->lfs_nactive > LFS_MAX_ACTIVE ||
326 	    (fs->lfs_sp->seg_flags & SEGM_CKP)) {
327 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
328 		CLR_FLUSHING(fs,vp);
329 		lfs_segunlock(fs);
330 		return error;
331 	}
332 	sp = fs->lfs_sp;
333 
334 	flushed = 0;
335 	if (VPISEMPTY(vp)) {
336 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
337 		++flushed;
338 	} else if ((ip->i_flag & IN_CLEANING) &&
339 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
340 #ifdef DEBUG_LFS
341 		ivndebug(vp,"vflush/clean");
342 #endif
343 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
344 		++flushed;
345 	} else if (lfs_dostats) {
346 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
347 			++lfs_stats.vflush_invoked;
348 #ifdef DEBUG_LFS
349 		ivndebug(vp,"vflush");
350 #endif
351 	}
352 
353 #ifdef DIAGNOSTIC
354 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
355 	if (vp->v_flag & VDIROP) {
356 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
357 		/* panic("VDIROP being flushed...this can\'t happen"); */
358 	}
359 	if (vp->v_usecount < 0) {
360 		printf("usecount=%ld\n", (long)vp->v_usecount);
361 		panic("lfs_vflush: usecount<0");
362 	}
363 #endif
364 
365 #if 1
366 	do {
367 		do {
368 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
369 				lfs_writefile(fs, sp, vp);
370 		} while (lfs_writeinode(fs, sp, ip));
371 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
372 #else
373 	if (flushed && vp != fs->lfs_ivnode)
374 		lfs_writeseg(fs, sp);
375 	else do {
376 		fs->lfs_flags &= ~LFS_IFDIRTY;
377 		lfs_writefile(fs, sp, vp);
378 		redo = lfs_writeinode(fs, sp, ip);
379 		redo += lfs_writeseg(fs, sp);
380 		redo += (fs->lfs_flags & LFS_IFDIRTY);
381 	} while (redo && vp == fs->lfs_ivnode);
382 #endif
383 	if (lfs_dostats) {
384 		++lfs_stats.nwrites;
385 		if (sp->seg_flags & SEGM_SYNC)
386 			++lfs_stats.nsync_writes;
387 		if (sp->seg_flags & SEGM_CKP)
388 			++lfs_stats.ncheckpoints;
389 	}
390 	/*
391 	 * If we were called from somewhere that has already held the seglock
392 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
393 	 * the write to complete because we are still locked.
394 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
395 	 * we must explicitly wait, if that is the case.
396 	 *
397 	 * We compare the iocount against 1, not 0, because it is
398 	 * artificially incremented by lfs_seglock().
399 	 */
400 	simple_lock(&fs->lfs_interlock);
401 	if (fs->lfs_seglock > 1) {
402 		simple_unlock(&fs->lfs_interlock);
403 		while (fs->lfs_iocount > 1)
404 			(void)tsleep(&fs->lfs_iocount, PRIBIO + 1,
405 				     "lfs_vflush", 0);
406 	} else
407 		simple_unlock(&fs->lfs_interlock);
408 
409 	lfs_segunlock(fs);
410 
411 	/* Wait for these buffers to be recovered by aiodoned */
412 	s = splbio();
413 	simple_lock(&global_v_numoutput_slock);
414 	while (vp->v_numoutput > 0) {
415 		ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf2", 0,
416 			&global_v_numoutput_slock);
417 	}
418 	simple_unlock(&global_v_numoutput_slock);
419 	splx(s);
420 
421 	CLR_FLUSHING(fs,vp);
422 	return (0);
423 }
424 
425 #ifdef DEBUG_LFS_VERBOSE
426 # define vndebug(vp,str) if (VTOI(vp)->i_flag & IN_CLEANING) printf("not writing ino %d because %s (op %d)\n",VTOI(vp)->i_number,(str),op)
427 #else
428 # define vndebug(vp,str)
429 #endif
430 
431 int
432 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
433 {
434 	struct inode *ip;
435 	struct vnode *vp, *nvp;
436 	int inodes_written = 0, only_cleaning;
437 
438 #ifndef LFS_NO_BACKVP_HACK
439 	/* BEGIN HACK */
440 #define	VN_OFFSET	\
441 	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
442 #define	BACK_VP(VP)	\
443 	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
444 #define	BEG_OF_VLIST	\
445 	((struct vnode *)(((caddr_t)&LIST_FIRST(&mp->mnt_vnodelist)) \
446 	- VN_OFFSET))
447 
448 	/* Find last vnode. */
449  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
450 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
451 	     vp = LIST_NEXT(vp, v_mntvnodes));
452 	for (; vp && vp != BEG_OF_VLIST; vp = nvp) {
453 		nvp = BACK_VP(vp);
454 #else
455 	loop:
456 	for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) {
457 		nvp = LIST_NEXT(vp, v_mntvnodes);
458 #endif
459 		/*
460 		 * If the vnode that we are about to sync is no longer
461 		 * associated with this mount point, start over.
462 		 */
463 		if (vp->v_mount != mp) {
464 			printf("lfs_writevnodes: starting over\n");
465 			/*
466 			 * After this, pages might be busy
467 			 * due to our own previous putpages.
468 			 * Start actual segment write here to avoid deadlock.
469 			 */
470 			(void)lfs_writeseg(fs, sp);
471 			goto loop;
472 		}
473 
474 		if (vp->v_type == VNON) {
475 			continue;
476 		}
477 
478 		ip = VTOI(vp);
479 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
480 		    (op != VN_DIROP && op != VN_CLEAN &&
481 		    (vp->v_flag & VDIROP))) {
482 			vndebug(vp,"dirop");
483 			continue;
484 		}
485 
486 		if (op == VN_EMPTY && !VPISEMPTY(vp)) {
487 			vndebug(vp,"empty");
488 			continue;
489 		}
490 
491 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
492 		   && vp != fs->lfs_flushvp
493 		   && !(ip->i_flag & IN_CLEANING)) {
494 			vndebug(vp,"cleaning");
495 			continue;
496 		}
497 
498 		if (lfs_vref(vp)) {
499 			vndebug(vp,"vref");
500 			continue;
501 		}
502 
503 		only_cleaning = 0;
504 		/*
505 		 * Write the inode/file if dirty and it's not the IFILE.
506 		 */
507 		if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) {
508 			only_cleaning =
509 			    ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
510 
511 			if (ip->i_number != LFS_IFILE_INUM)
512 				lfs_writefile(fs, sp, vp);
513 			if (!VPISEMPTY(vp)) {
514 				if (WRITEINPROG(vp)) {
515 #ifdef DEBUG_LFS
516 					ivndebug(vp,"writevnodes/write2");
517 #endif
518 				} else if (!(ip->i_flag & IN_ALLMOD)) {
519 #ifdef DEBUG_LFS
520 					printf("<%d>",ip->i_number);
521 #endif
522 					LFS_SET_UINO(ip, IN_MODIFIED);
523 				}
524 			}
525 			(void) lfs_writeinode(fs, sp, ip);
526 			inodes_written++;
527 		}
528 
529 		if (lfs_clean_vnhead && only_cleaning)
530 			lfs_vunref_head(vp);
531 		else
532 			lfs_vunref(vp);
533 	}
534 	return inodes_written;
535 }
536 
537 /*
538  * Do a checkpoint.
539  */
540 int
541 lfs_segwrite(struct mount *mp, int flags)
542 {
543 	struct buf *bp;
544 	struct inode *ip;
545 	struct lfs *fs;
546 	struct segment *sp;
547 	struct vnode *vp;
548 	SEGUSE *segusep;
549 	int do_ckp, did_ckp, error, s;
550 	unsigned n, segleft, maxseg, sn, i, curseg;
551 	int writer_set = 0;
552 	int dirty;
553 	int redo;
554 
555 	fs = VFSTOUFS(mp)->um_lfs;
556 
557 	if (fs->lfs_ronly)
558 		return EROFS;
559 
560 	lfs_imtime(fs);
561 
562 	/*
563 	 * Allocate a segment structure and enough space to hold pointers to
564 	 * the maximum possible number of buffers which can be described in a
565 	 * single summary block.
566 	 */
567 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
568 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
569 	sp = fs->lfs_sp;
570 
571 	/*
572 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
573 	 * in which case we have to flush *all* buffers off of this vnode.
574 	 * We don't care about other nodes, but write any non-dirop nodes
575 	 * anyway in anticipation of another getnewvnode().
576 	 *
577 	 * If we're cleaning we only write cleaning and ifile blocks, and
578 	 * no dirops, since otherwise we'd risk corruption in a crash.
579 	 */
580 	if (sp->seg_flags & SEGM_CLEAN)
581 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
582 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
583 		lfs_writevnodes(fs, mp, sp, VN_REG);
584 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
585 			error = lfs_writer_enter(fs, "lfs writer");
586 			if (error) {
587 				printf("segwrite mysterious error\n");
588 				/* XXX why not segunlock? */
589 				pool_put(&fs->lfs_bpppool, sp->bpp);
590 				sp->bpp = NULL;
591 				pool_put(&fs->lfs_segpool, sp);
592 				sp = fs->lfs_sp = NULL;
593 				return (error);
594 			}
595 			writer_set = 1;
596 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
597 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
598 		}
599 	}
600 
601 	/*
602 	 * If we are doing a checkpoint, mark everything since the
603 	 * last checkpoint as no longer ACTIVE.
604 	 */
605 	if (do_ckp) {
606 		segleft = fs->lfs_nseg;
607 		curseg = 0;
608 		for (n = 0; n < fs->lfs_segtabsz; n++) {
609 			dirty = 0;
610 			if (bread(fs->lfs_ivnode,
611 			    fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &bp))
612 				panic("lfs_segwrite: ifile read");
613 			segusep = (SEGUSE *)bp->b_data;
614 			maxseg = min(segleft, fs->lfs_sepb);
615 			for (i = 0; i < maxseg; i++) {
616 				sn = curseg + i;
617 				if (sn != fs->lfs_curseg &&
618 				    segusep->su_flags & SEGUSE_ACTIVE) {
619 					segusep->su_flags &= ~SEGUSE_ACTIVE;
620 					--fs->lfs_nactive;
621 					++dirty;
622 				}
623 				fs->lfs_suflags[fs->lfs_activesb][sn] =
624 					segusep->su_flags;
625 				if (fs->lfs_version > 1)
626 					++segusep;
627 				else
628 					segusep = (SEGUSE *)
629 						((SEGUSE_V1 *)segusep + 1);
630 			}
631 
632 			if (dirty)
633 				error = LFS_BWRITE_LOG(bp); /* Ifile */
634 			else
635 				brelse(bp);
636 			segleft -= fs->lfs_sepb;
637 			curseg += fs->lfs_sepb;
638 		}
639 	}
640 
641 	did_ckp = 0;
642 	if (do_ckp || fs->lfs_doifile) {
643 		do {
644 			vp = fs->lfs_ivnode;
645 
646 #ifdef DEBUG
647 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
648 #endif
649 			fs->lfs_flags &= ~LFS_IFDIRTY;
650 
651 			ip = VTOI(vp);
652 
653 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
654 				lfs_writefile(fs, sp, vp);
655 
656 			if (ip->i_flag & IN_ALLMOD)
657 				++did_ckp;
658 			redo = lfs_writeinode(fs, sp, ip);
659 			redo += lfs_writeseg(fs, sp);
660 			redo += (fs->lfs_flags & LFS_IFDIRTY);
661 		} while (redo && do_ckp);
662 
663 		/*
664 		 * Unless we are unmounting, the Ifile may continue to have
665 		 * dirty blocks even after a checkpoint, due to changes to
666 		 * inodes' atime.  If we're checkpointing, it's "impossible"
667 		 * for other parts of the Ifile to be dirty after the loop
668 		 * above, since we hold the segment lock.
669 		 */
670 		s = splbio();
671 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
672 			LFS_CLR_UINO(ip, IN_ALLMOD);
673 		}
674 #ifdef DIAGNOSTIC
675 		else if (do_ckp) {
676 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
677 				if (bp->b_lblkno < fs->lfs_cleansz +
678 				    fs->lfs_segtabsz &&
679 				    !(bp->b_flags & B_GATHERED)) {
680 					panic("dirty blocks");
681 				}
682 			}
683 		}
684 #endif
685 		splx(s);
686 	} else {
687 		(void) lfs_writeseg(fs, sp);
688 	}
689 
690 	/* Note Ifile no longer needs to be written */
691 	fs->lfs_doifile = 0;
692 	if (writer_set)
693 		lfs_writer_leave(fs);
694 
695 	/*
696 	 * If we didn't write the Ifile, we didn't really do anything.
697 	 * That means that (1) there is a checkpoint on disk and (2)
698 	 * nothing has changed since it was written.
699 	 *
700 	 * Take the flags off of the segment so that lfs_segunlock
701 	 * doesn't have to write the superblock either.
702 	 */
703 	if (do_ckp && !did_ckp) {
704 		sp->seg_flags &= ~SEGM_CKP;
705 	}
706 
707 	if (lfs_dostats) {
708 		++lfs_stats.nwrites;
709 		if (sp->seg_flags & SEGM_SYNC)
710 			++lfs_stats.nsync_writes;
711 		if (sp->seg_flags & SEGM_CKP)
712 			++lfs_stats.ncheckpoints;
713 	}
714 	lfs_segunlock(fs);
715 	return (0);
716 }
717 
718 /*
719  * Write the dirty blocks associated with a vnode.
720  */
721 void
722 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
723 {
724 	struct buf *bp;
725 	struct finfo *fip;
726 	struct inode *ip;
727 	IFILE *ifp;
728 	int i, frag;
729 
730 	ip = VTOI(vp);
731 
732 	if (sp->seg_bytes_left < fs->lfs_bsize ||
733 	    sp->sum_bytes_left < sizeof(struct finfo))
734 		(void) lfs_writeseg(fs, sp);
735 
736 	sp->sum_bytes_left -= FINFOSIZE;
737 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
738 
739 	if (vp->v_flag & VDIROP)
740 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
741 
742 	fip = sp->fip;
743 	fip->fi_nblocks = 0;
744 	fip->fi_ino = ip->i_number;
745 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
746 	fip->fi_version = ifp->if_version;
747 	brelse(bp);
748 
749 	if (sp->seg_flags & SEGM_CLEAN) {
750 		lfs_gather(fs, sp, vp, lfs_match_fake);
751 		/*
752 		 * For a file being flushed, we need to write *all* blocks.
753 		 * This means writing the cleaning blocks first, and then
754 		 * immediately following with any non-cleaning blocks.
755 		 * The same is true of the Ifile since checkpoints assume
756 		 * that all valid Ifile blocks are written.
757 		 */
758 		if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode) {
759 			lfs_gather(fs, sp, vp, lfs_match_data);
760 			/*
761 			 * Don't call VOP_PUTPAGES: if we're flushing,
762 			 * we've already done it, and the Ifile doesn't
763 			 * use the page cache.
764 			 */
765 		}
766 	} else {
767 		lfs_gather(fs, sp, vp, lfs_match_data);
768 		/*
769 		 * If we're flushing, we've already called VOP_PUTPAGES
770 		 * so don't do it again.  Otherwise, we want to write
771 		 * everything we've got.
772 		 */
773 		if (!IS_FLUSHING(fs, vp)) {
774 			simple_lock(&vp->v_interlock);
775 			VOP_PUTPAGES(vp, 0, 0,
776 				     PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
777 		}
778 	}
779 
780 	/*
781 	 * It may not be necessary to write the meta-data blocks at this point,
782 	 * as the roll-forward recovery code should be able to reconstruct the
783 	 * list.
784 	 *
785 	 * We have to write them anyway, though, under two conditions: (1) the
786 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
787 	 * checkpointing.
788 	 *
789 	 * BUT if we are cleaning, we might have indirect blocks that refer to
790 	 * new blocks not being written yet, in addition to fragments being
791 	 * moved out of a cleaned segment.  If that is the case, don't
792 	 * write the indirect blocks, or the finfo will have a small block
793 	 * in the middle of it!
794 	 * XXX in this case isn't the inode size wrong too?
795 	 */
796 	frag = 0;
797 	if (sp->seg_flags & SEGM_CLEAN) {
798 		for (i = 0; i < NDADDR; i++)
799 			if (ip->i_lfs_fragsize[i] > 0 &&
800 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
801 				++frag;
802 	}
803 #ifdef DIAGNOSTIC
804 	if (frag > 1)
805 		panic("lfs_writefile: more than one fragment!");
806 #endif
807 	if (IS_FLUSHING(fs, vp) ||
808 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
809 		lfs_gather(fs, sp, vp, lfs_match_indir);
810 		lfs_gather(fs, sp, vp, lfs_match_dindir);
811 		lfs_gather(fs, sp, vp, lfs_match_tindir);
812 	}
813 	fip = sp->fip;
814 	if (fip->fi_nblocks != 0) {
815 		sp->fip = (FINFO*)((caddr_t)fip + FINFOSIZE +
816 				   sizeof(int32_t) * (fip->fi_nblocks));
817 		sp->start_lbp = &sp->fip->fi_blocks[0];
818 	} else {
819 		sp->sum_bytes_left += FINFOSIZE;
820 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
821 	}
822 }
823 
824 int
825 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
826 {
827 	struct buf *bp, *ibp;
828 	struct ufs1_dinode *cdp;
829 	IFILE *ifp;
830 	SEGUSE *sup;
831 	daddr_t daddr;
832 	int32_t *daddrp;	/* XXX ondisk32 */
833 	ino_t ino;
834 	int error, i, ndx, fsb = 0;
835 	int redo_ifile = 0;
836 	struct timespec ts;
837 	int gotblk = 0;
838 
839 	if (!(ip->i_flag & IN_ALLMOD))
840 		return (0);
841 
842 	/* Allocate a new inode block if necessary. */
843 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) &&
844 	    sp->ibp == NULL) {
845 		/* Allocate a new segment if necessary. */
846 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
847 		    sp->sum_bytes_left < sizeof(int32_t))
848 			(void) lfs_writeseg(fs, sp);
849 
850 		/* Get next inode block. */
851 		daddr = fs->lfs_offset;
852 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
853 		sp->ibp = *sp->cbpp++ =
854 			getblk(VTOI(fs->lfs_ivnode)->i_devvp,
855 			    fsbtodb(fs, daddr), fs->lfs_ibsize, 0, 0);
856 		gotblk++;
857 
858 		/* Zero out inode numbers */
859 		for (i = 0; i < INOPB(fs); ++i)
860 			((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber =
861 			    0;
862 
863 		++sp->start_bpp;
864 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
865 		/* Set remaining space counters. */
866 		sp->seg_bytes_left -= fs->lfs_ibsize;
867 		sp->sum_bytes_left -= sizeof(int32_t);
868 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
869 			sp->ninodes / INOPB(fs) - 1;
870 		((int32_t *)(sp->segsum))[ndx] = daddr;
871 	}
872 
873 	/* Update the inode times and copy the inode onto the inode page. */
874 	TIMEVAL_TO_TIMESPEC(&time, &ts);
875 	/* XXX kludge --- don't redirty the ifile just to put times on it */
876 	if (ip->i_number != LFS_IFILE_INUM)
877 		LFS_ITIMES(ip, &ts, &ts, &ts);
878 
879 	/*
880 	 * If this is the Ifile, and we've already written the Ifile in this
881 	 * partial segment, just overwrite it (it's not on disk yet) and
882 	 * continue.
883 	 *
884 	 * XXX we know that the bp that we get the second time around has
885 	 * already been gathered.
886 	 */
887 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
888 		*(sp->idp) = *ip->i_din.ffs1_din;
889 		ip->i_lfs_osize = ip->i_size;
890 		return 0;
891 	}
892 
893 	bp = sp->ibp;
894 	cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
895 	*cdp = *ip->i_din.ffs1_din;
896 #ifdef LFS_IFILE_FRAG_ADDRESSING
897 	if (fs->lfs_version > 1)
898 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
899 #endif
900 
901 	/*
902 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
903 	 * addresses to disk; possibly revert the inode size.
904 	 * XXX By not writing these blocks, we are making the lfs_avail
905 	 * XXX count on disk wrong by the same amount.	We should be
906 	 * XXX able to "borrow" from lfs_avail and return it after the
907 	 * XXX Ifile is written.  See also in lfs_writeseg.
908 	 */
909 	if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) {
910 		cdp->di_size = ip->i_lfs_osize;
911 #ifdef DEBUG_LFS
912 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
913 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs1_blocks);
914 #endif
915 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
916 		     daddrp++) {
917 			if (*daddrp == UNWRITTEN) {
918 #ifdef DEBUG_LFS
919 				printf("lfs_writeinode: wiping UNWRITTEN\n");
920 #endif
921 				*daddrp = 0;
922 			}
923 		}
924 	} else {
925 		/* If all blocks are goig to disk, update the "size on disk" */
926 		ip->i_lfs_osize = ip->i_size;
927 	}
928 
929 	if (ip->i_flag & IN_CLEANING)
930 		LFS_CLR_UINO(ip, IN_CLEANING);
931 	else {
932 		/* XXX IN_ALLMOD */
933 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
934 			     IN_UPDATE);
935 		if (ip->i_lfs_effnblks == ip->i_ffs1_blocks)
936 			LFS_CLR_UINO(ip, IN_MODIFIED);
937 #ifdef DEBUG_LFS
938 		else
939 			printf("lfs_writeinode: ino %d: real blks=%d, "
940 			       "eff=%d\n", ip->i_number, ip->i_ffs1_blocks,
941 			       ip->i_lfs_effnblks);
942 #endif
943 	}
944 
945 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
946 		sp->idp = ((struct ufs1_dinode *)bp->b_data) +
947 			(sp->ninodes % INOPB(fs));
948 	if (gotblk) {
949 		LFS_LOCK_BUF(bp);
950 		brelse(bp);
951 	}
952 
953 	/* Increment inode count in segment summary block. */
954 	++((SEGSUM *)(sp->segsum))->ss_ninos;
955 
956 	/* If this page is full, set flag to allocate a new page. */
957 	if (++sp->ninodes % INOPB(fs) == 0)
958 		sp->ibp = NULL;
959 
960 	/*
961 	 * If updating the ifile, update the super-block.  Update the disk
962 	 * address and access times for this inode in the ifile.
963 	 */
964 	ino = ip->i_number;
965 	if (ino == LFS_IFILE_INUM) {
966 		daddr = fs->lfs_idaddr;
967 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
968 	} else {
969 		LFS_IENTRY(ifp, fs, ino, ibp);
970 		daddr = ifp->if_daddr;
971 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
972 #ifdef LFS_DEBUG_NEXTFREE
973 		if (ino > 3 && ifp->if_nextfree) {
974 			vprint("lfs_writeinode",ITOV(ip));
975 			printf("lfs_writeinode: updating free ino %d\n",
976 				ip->i_number);
977 		}
978 #endif
979 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
980 	}
981 
982 	/*
983 	 * The inode's last address should not be in the current partial
984 	 * segment, except under exceptional circumstances (lfs_writevnodes
985 	 * had to start over, and in the meantime more blocks were written
986 	 * to a vnode).	 Both inodes will be accounted to this segment
987 	 * in lfs_writeseg so we need to subtract the earlier version
988 	 * here anyway.	 The segment count can temporarily dip below
989 	 * zero here; keep track of how many duplicates we have in
990 	 * "dupino" so we don't panic below.
991 	 */
992 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno)) {
993 		++sp->ndupino;
994 		printf("lfs_writeinode: last inode addr in current pseg "
995 		       "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
996 			(long long)daddr, sp->ndupino);
997 	}
998 	/*
999 	 * Account the inode: it no longer belongs to its former segment,
1000 	 * though it will not belong to the new segment until that segment
1001 	 * is actually written.
1002 	 */
1003 	if (daddr != LFS_UNUSED_DADDR) {
1004 		u_int32_t oldsn = dtosn(fs, daddr);
1005 #ifdef DIAGNOSTIC
1006 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
1007 #endif
1008 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1009 #ifdef DIAGNOSTIC
1010 		if (sup->su_nbytes +
1011 		    sizeof (struct ufs1_dinode) * ndupino
1012 		      < sizeof (struct ufs1_dinode)) {
1013 			printf("lfs_writeinode: negative bytes "
1014 			       "(segment %" PRIu32 " short by %d, "
1015 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
1016 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
1017 			       "ndupino=%d)\n",
1018 			       dtosn(fs, daddr),
1019 			       (int)sizeof (struct ufs1_dinode) *
1020 				   (1 - sp->ndupino) - sup->su_nbytes,
1021 			       oldsn, sp->seg_number, daddr,
1022 			       (unsigned int)sup->su_nbytes,
1023 			       sp->ndupino);
1024 			panic("lfs_writeinode: negative bytes");
1025 			sup->su_nbytes = sizeof (struct ufs1_dinode);
1026 		}
1027 #endif
1028 #ifdef DEBUG_SU_NBYTES
1029 		printf("seg %d -= %d for ino %d inode\n",
1030 		       dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino);
1031 #endif
1032 		sup->su_nbytes -= sizeof (struct ufs1_dinode);
1033 		redo_ifile =
1034 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1035 		if (redo_ifile)
1036 			fs->lfs_flags |= LFS_IFDIRTY;
1037 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1038 	}
1039 	return (redo_ifile);
1040 }
1041 
1042 int
1043 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1044 {
1045 	struct lfs *fs;
1046 	int version;
1047 	int j, blksinblk;
1048 
1049 	/*
1050 	 * If full, finish this segment.  We may be doing I/O, so
1051 	 * release and reacquire the splbio().
1052 	 */
1053 #ifdef DIAGNOSTIC
1054 	if (sp->vp == NULL)
1055 		panic ("lfs_gatherblock: Null vp in segment");
1056 #endif
1057 	fs = sp->fs;
1058 	blksinblk = howmany(bp->b_bcount, fs->lfs_bsize);
1059 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1060 	    sp->seg_bytes_left < bp->b_bcount) {
1061 		if (sptr)
1062 			splx(*sptr);
1063 		lfs_updatemeta(sp);
1064 
1065 		version = sp->fip->fi_version;
1066 		(void) lfs_writeseg(fs, sp);
1067 
1068 		sp->fip->fi_version = version;
1069 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
1070 		/* Add the current file to the segment summary. */
1071 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
1072 		sp->sum_bytes_left -= FINFOSIZE;
1073 
1074 		if (sptr)
1075 			*sptr = splbio();
1076 		return (1);
1077 	}
1078 
1079 #ifdef DEBUG
1080 	if (bp->b_flags & B_GATHERED) {
1081 		printf("lfs_gatherblock: already gathered! Ino %d,"
1082 		       " lbn %" PRId64 "\n",
1083 		       sp->fip->fi_ino, bp->b_lblkno);
1084 		return (0);
1085 	}
1086 #endif
1087 	/* Insert into the buffer list, update the FINFO block. */
1088 	bp->b_flags |= B_GATHERED;
1089 
1090 	*sp->cbpp++ = bp;
1091 	for (j = 0; j < blksinblk; j++)
1092 		sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j;
1093 
1094 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1095 	sp->seg_bytes_left -= bp->b_bcount;
1096 	return (0);
1097 }
1098 
1099 int
1100 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp,
1101     int (*match)(struct lfs *, struct buf *))
1102 {
1103 	struct buf *bp, *nbp;
1104 	int s, count = 0;
1105 
1106 	sp->vp = vp;
1107 	s = splbio();
1108 
1109 #ifndef LFS_NO_BACKBUF_HACK
1110 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1111 # define	BUF_OFFSET	\
1112 	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1113 # define	BACK_BUF(BP)	\
1114 	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1115 # define	BEG_OF_LIST	\
1116 	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1117 
1118 loop:
1119 	/* Find last buffer. */
1120 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd);
1121 	     bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1122 	     bp = LIST_NEXT(bp, b_vnbufs))
1123 		/* nothing */;
1124 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1125 		nbp = BACK_BUF(bp);
1126 #else /* LFS_NO_BACKBUF_HACK */
1127 loop:
1128 	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1129 		nbp = LIST_NEXT(bp, b_vnbufs);
1130 #endif /* LFS_NO_BACKBUF_HACK */
1131 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1132 #ifdef DEBUG_LFS
1133 			if (vp == fs->lfs_ivnode &&
1134 			    (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1135 				printf("(%" PRId64 ":%lx)",
1136 				    bp->b_lblkno, bp->b_flags);
1137 #endif
1138 			continue;
1139 		}
1140 		if (vp->v_type == VBLK) {
1141 			/* For block devices, just write the blocks. */
1142 			/* XXX Do we really need to even do this? */
1143 #ifdef DEBUG_LFS
1144 			if (count == 0)
1145 				printf("BLK(");
1146 			printf(".");
1147 #endif
1148 			/*
1149 			 * Get the block before bwrite,
1150 			 * so we don't corrupt the free list
1151 			 */
1152 			bp->b_flags |= B_BUSY;
1153 			bremfree(bp);
1154 			bwrite(bp);
1155 		} else {
1156 #ifdef DIAGNOSTIC
1157 # ifdef LFS_USE_B_INVAL
1158 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1159 				printf("lfs_gather: lbn %" PRId64 " is "
1160 					"B_INVAL\n", bp->b_lblkno);
1161 				VOP_PRINT(bp->b_vp);
1162 			}
1163 # endif /* LFS_USE_B_INVAL */
1164 			if (!(bp->b_flags & B_DELWRI))
1165 				panic("lfs_gather: bp not B_DELWRI");
1166 			if (!(bp->b_flags & B_LOCKED)) {
1167 				printf("lfs_gather: lbn %" PRId64 " blk "
1168 					"%" PRId64 " not B_LOCKED\n",
1169 					bp->b_lblkno,
1170 					dbtofsb(fs, bp->b_blkno));
1171 				VOP_PRINT(bp->b_vp);
1172 				panic("lfs_gather: bp not B_LOCKED");
1173 			}
1174 #endif
1175 			if (lfs_gatherblock(sp, bp, &s)) {
1176 				goto loop;
1177 			}
1178 		}
1179 		count++;
1180 	}
1181 	splx(s);
1182 #ifdef DEBUG_LFS
1183 	if (vp->v_type == VBLK && count)
1184 		printf(")\n");
1185 #endif
1186 	lfs_updatemeta(sp);
1187 	sp->vp = NULL;
1188 	return count;
1189 }
1190 
1191 #if DEBUG
1192 # define DEBUG_OOFF(n) do {						\
1193 	if (ooff == 0) {						\
1194 		printf("lfs_updatemeta[%d]: warning: writing "		\
1195 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1196 			", was 0x0 (or %" PRId64 ")\n",			\
1197 			(n), ip->i_number, lbn, ndaddr, daddr);		\
1198 	}								\
1199 } while (0)
1200 #else
1201 # define DEBUG_OOFF(n)
1202 #endif
1203 
1204 /*
1205  * Change the given block's address to ndaddr, finding its previous
1206  * location using ufs_bmaparray().
1207  *
1208  * Account for this change in the segment table.
1209  */
1210 void
1211 lfs_update_single(struct lfs *fs, struct segment *sp, daddr_t lbn,
1212 		  int32_t ndaddr, int size)
1213 {
1214 	SEGUSE *sup;
1215 	struct buf *bp;
1216 	struct indir a[NIADDR + 2], *ap;
1217 	struct inode *ip;
1218 	struct vnode *vp;
1219 	daddr_t daddr, ooff;
1220 	int num, error;
1221 	int bb, osize, obb;
1222 
1223 	vp = sp->vp;
1224 	ip = VTOI(vp);
1225 
1226 	error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1227 	if (error)
1228 		panic("lfs_updatemeta: ufs_bmaparray returned %d", error);
1229 
1230 	KASSERT(daddr <= LFS_MAX_DADDR);
1231 	if (daddr > 0)
1232 		daddr = dbtofsb(fs, daddr);
1233 
1234 	bb = fragstofsb(fs, numfrags(fs, size));
1235 	switch (num) {
1236 	    case 0:
1237 		    ooff = ip->i_ffs1_db[lbn];
1238 		    DEBUG_OOFF(0);
1239 		    if (ooff == UNWRITTEN)
1240 			    ip->i_ffs1_blocks += bb;
1241 		    else {
1242 			    /* possible fragment truncation or extension */
1243 			    obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1244 			    ip->i_ffs1_blocks += (bb - obb);
1245 		    }
1246 		    ip->i_ffs1_db[lbn] = ndaddr;
1247 		    break;
1248 	    case 1:
1249 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1250 		    DEBUG_OOFF(1);
1251 		    if (ooff == UNWRITTEN)
1252 			    ip->i_ffs1_blocks += bb;
1253 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1254 		    break;
1255 	    default:
1256 		    ap = &a[num - 1];
1257 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1258 			    panic("lfs_updatemeta: bread bno %" PRId64,
1259 				  ap->in_lbn);
1260 
1261 		    /* XXX ondisk32 */
1262 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1263 		    DEBUG_OOFF(num);
1264 		    if (ooff == UNWRITTEN)
1265 			    ip->i_ffs1_blocks += bb;
1266 		    /* XXX ondisk32 */
1267 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1268 		    (void) VOP_BWRITE(bp);
1269 	}
1270 
1271 	/*
1272 	 * Though we'd rather it couldn't, this *can* happen right now
1273 	 * if cleaning blocks and regular blocks coexist.
1274 	 */
1275 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1276 
1277 	/*
1278 	 * Update segment usage information, based on old size
1279 	 * and location.
1280 	 */
1281 	if (daddr > 0) {
1282 		u_int32_t oldsn = dtosn(fs, daddr);
1283 #ifdef DIAGNOSTIC
1284 		int ndupino = (sp->seg_number == oldsn) ?
1285 			sp->ndupino : 0;
1286 #endif
1287 		KASSERT(oldsn >= 0 && oldsn < fs->lfs_nseg);
1288 		if (lbn >= 0 && lbn < NDADDR)
1289 			osize = ip->i_lfs_fragsize[lbn];
1290 		else
1291 			osize = fs->lfs_bsize;
1292 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1293 #ifdef DIAGNOSTIC
1294 		if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino
1295 		    < osize) {
1296 			printf("lfs_updatemeta: negative bytes "
1297 			       "(segment %" PRIu32 " short by %" PRId64
1298 			       ")\n", dtosn(fs, daddr),
1299 			       (int64_t)osize -
1300 			       (sizeof (struct ufs1_dinode) * sp->ndupino +
1301 				sup->su_nbytes));
1302 			printf("lfs_updatemeta: ino %d, lbn %" PRId64
1303 			       ", addr = 0x%" PRIx64 "\n",
1304 			       VTOI(sp->vp)->i_number, lbn, daddr);
1305 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1306 			panic("lfs_updatemeta: negative bytes");
1307 			sup->su_nbytes = osize -
1308 			    sizeof (struct ufs1_dinode) * sp->ndupino;
1309 		}
1310 #endif
1311 #ifdef DEBUG_SU_NBYTES
1312 		printf("seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1313 		       " db 0x%" PRIx64 "\n",
1314 		       dtosn(fs, daddr), osize,
1315 		       VTOI(sp->vp)->i_number, lbn, daddr);
1316 #endif
1317 		sup->su_nbytes -= osize;
1318 		if (!(bp->b_flags & B_GATHERED))
1319 			fs->lfs_flags |= LFS_IFDIRTY;
1320 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1321 	}
1322 	/*
1323 	 * Now that this block has a new address, and its old
1324 	 * segment no longer owns it, we can forget about its
1325 	 * old size.
1326 	 */
1327 	if (lbn >= 0 && lbn < NDADDR)
1328 		ip->i_lfs_fragsize[lbn] = size;
1329 }
1330 
1331 /*
1332  * Update the metadata that points to the blocks listed in the FINFO
1333  * array.
1334  */
1335 void
1336 lfs_updatemeta(struct segment *sp)
1337 {
1338 	struct buf *sbp;
1339 	struct lfs *fs;
1340 	struct vnode *vp;
1341 	daddr_t lbn;
1342 	int i, nblocks, num;
1343 	int bb;
1344 	int bytesleft, size;
1345 
1346 	vp = sp->vp;
1347 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1348 	KASSERT(nblocks >= 0);
1349 	if (vp == NULL || nblocks == 0)
1350 		return;
1351 
1352 	/*
1353 	 * This count may be high due to oversize blocks from lfs_gop_write.
1354 	 * Correct for this. (XXX we should be able to keep track of these.)
1355 	 */
1356 	fs = sp->fs;
1357 	for (i = 0; i < nblocks; i++) {
1358 		if (sp->start_bpp[i] == NULL) {
1359 			printf("nblocks = %d, not %d\n", i, nblocks);
1360 			nblocks = i;
1361 			break;
1362 		}
1363 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1364 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1365 		nblocks -= num - 1;
1366 	}
1367 
1368 	KASSERT(vp->v_type == VREG ||
1369 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1370 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1371 
1372 	/*
1373 	 * Sort the blocks.
1374 	 *
1375 	 * We have to sort even if the blocks come from the
1376 	 * cleaner, because there might be other pending blocks on the
1377 	 * same inode...and if we don't sort, and there are fragments
1378 	 * present, blocks may be written in the wrong place.
1379 	 */
1380 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1381 
1382 	/*
1383 	 * Record the length of the last block in case it's a fragment.
1384 	 * If there are indirect blocks present, they sort last.  An
1385 	 * indirect block will be lfs_bsize and its presence indicates
1386 	 * that you cannot have fragments.
1387 	 *
1388 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1389 	 * XXX a later indirect block.	This will continue to be
1390 	 * XXX true until lfs_markv is fixed to do everything with
1391 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1392 	 */
1393 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1394 		fs->lfs_bmask) + 1;
1395 
1396 	/*
1397 	 * Assign disk addresses, and update references to the logical
1398 	 * block and the segment usage information.
1399 	 */
1400 	for (i = nblocks; i--; ++sp->start_bpp) {
1401 		sbp = *sp->start_bpp;
1402 		lbn = *sp->start_lbp;
1403 		KASSERT(sbp->b_lblkno == lbn);
1404 
1405 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1406 
1407 		/*
1408 		 * If we write a frag in the wrong place, the cleaner won't
1409 		 * be able to correctly identify its size later, and the
1410 		 * segment will be uncleanable.	 (Even worse, it will assume
1411 		 * that the indirect block that actually ends the list
1412 		 * is of a smaller size!)
1413 		 */
1414 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1415 			panic("lfs_updatemeta: fragment is not last block");
1416 
1417 		/*
1418 		 * For each subblock in this possibly oversized block,
1419 		 * update its address on disk.
1420 		 */
1421 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1422 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1423 		     bytesleft -= fs->lfs_bsize) {
1424 			size = MIN(bytesleft, fs->lfs_bsize);
1425 			bb = fragstofsb(fs, numfrags(fs, size));
1426 			lbn = *sp->start_lbp++;
1427 			lfs_update_single(fs, sp, lbn, fs->lfs_offset, size);
1428 			fs->lfs_offset += bb;
1429 		}
1430 
1431 	}
1432 }
1433 
1434 /*
1435  * Start a new segment.
1436  */
1437 int
1438 lfs_initseg(struct lfs *fs)
1439 {
1440 	struct segment *sp;
1441 	SEGUSE *sup;
1442 	SEGSUM *ssp;
1443 	struct buf *bp, *sbp;
1444 	int repeat;
1445 
1446 	sp = fs->lfs_sp;
1447 
1448 	repeat = 0;
1449 
1450 	/* Advance to the next segment. */
1451 	if (!LFS_PARTIAL_FITS(fs)) {
1452 		/* lfs_avail eats the remaining space */
1453 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1454 						   fs->lfs_curseg);
1455 		/* Wake up any cleaning procs waiting on this file system. */
1456 		wakeup(&lfs_allclean_wakeup);
1457 		wakeup(&fs->lfs_nextseg);
1458 		lfs_newseg(fs);
1459 		repeat = 1;
1460 		fs->lfs_offset = fs->lfs_curseg;
1461 
1462 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1463 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1464 
1465 		/*
1466 		 * If the segment contains a superblock, update the offset
1467 		 * and summary address to skip over it.
1468 		 */
1469 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1470 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1471 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1472 			sp->seg_bytes_left -= LFS_SBPAD;
1473 		}
1474 		brelse(bp);
1475 		/* Segment zero could also contain the labelpad */
1476 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1477 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1478 			fs->lfs_offset +=
1479 			    btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1480 			sp->seg_bytes_left -=
1481 			    LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1482 		}
1483 	} else {
1484 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1485 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1486 				      (fs->lfs_offset - fs->lfs_curseg));
1487 	}
1488 	fs->lfs_lastpseg = fs->lfs_offset;
1489 
1490 	/* Record first address of this partial segment */
1491 	if (sp->seg_flags & SEGM_CLEAN) {
1492 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1493 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1494 			/* "1" is the artificial inc in lfs_seglock */
1495 			while (fs->lfs_iocount > 1) {
1496 				tsleep(&fs->lfs_iocount, PRIBIO + 1,
1497 				    "lfs_initseg", 0);
1498 			}
1499 			fs->lfs_cleanind = 0;
1500 		}
1501 	}
1502 
1503 	sp->fs = fs;
1504 	sp->ibp = NULL;
1505 	sp->idp = NULL;
1506 	sp->ninodes = 0;
1507 	sp->ndupino = 0;
1508 
1509 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1510 	sp->cbpp = sp->bpp;
1511 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1512 	    fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1513 	sp->segsum = (*sp->cbpp)->b_data;
1514 	memset(sp->segsum, 0, fs->lfs_sumsize);
1515 	sp->start_bpp = ++sp->cbpp;
1516 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1517 
1518 	/* Set point to SEGSUM, initialize it. */
1519 	ssp = sp->segsum;
1520 	ssp->ss_next = fs->lfs_nextseg;
1521 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1522 	ssp->ss_magic = SS_MAGIC;
1523 
1524 	/* Set pointer to first FINFO, initialize it. */
1525 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1526 	sp->fip->fi_nblocks = 0;
1527 	sp->start_lbp = &sp->fip->fi_blocks[0];
1528 	sp->fip->fi_lastlength = 0;
1529 
1530 	sp->seg_bytes_left -= fs->lfs_sumsize;
1531 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1532 
1533 	return (repeat);
1534 }
1535 
1536 /*
1537  * Return the next segment to write.
1538  */
1539 void
1540 lfs_newseg(struct lfs *fs)
1541 {
1542 	CLEANERINFO *cip;
1543 	SEGUSE *sup;
1544 	struct buf *bp;
1545 	int curseg, isdirty, sn;
1546 
1547 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1548 #ifdef DEBUG_SU_NBYTES
1549 	printf("lfs_newseg: seg %d := 0 in newseg\n",	/* XXXDEBUG */
1550 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1551 #endif
1552 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1553 	sup->su_nbytes = 0;
1554 	sup->su_nsums = 0;
1555 	sup->su_ninos = 0;
1556 	LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1557 
1558 	LFS_CLEANERINFO(cip, fs, bp);
1559 	--cip->clean;
1560 	++cip->dirty;
1561 	fs->lfs_nclean = cip->clean;
1562 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1563 
1564 	fs->lfs_lastseg = fs->lfs_curseg;
1565 	fs->lfs_curseg = fs->lfs_nextseg;
1566 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1567 		sn = (sn + 1) % fs->lfs_nseg;
1568 		if (sn == curseg)
1569 			panic("lfs_nextseg: no clean segments");
1570 		LFS_SEGENTRY(sup, fs, sn, bp);
1571 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1572 		/* Check SEGUSE_EMPTY as we go along */
1573 		if (isdirty && sup->su_nbytes == 0 &&
1574 		    !(sup->su_flags & SEGUSE_EMPTY))
1575 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1576 		else
1577 			brelse(bp);
1578 
1579 		if (!isdirty)
1580 			break;
1581 	}
1582 
1583 	++fs->lfs_nactive;
1584 	fs->lfs_nextseg = sntod(fs, sn);
1585 	if (lfs_dostats) {
1586 		++lfs_stats.segsused;
1587 	}
1588 }
1589 
1590 static struct buf *
1591 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1592 {
1593 	struct lfs_cluster *cl;
1594 	struct buf **bpp, *bp;
1595 	int s;
1596 
1597 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1598 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1599 	memset(cl, 0, sizeof(*cl));
1600 	cl->fs = fs;
1601 	cl->bpp = bpp;
1602 	cl->bufcount = 0;
1603 	cl->bufsize = 0;
1604 
1605 	/* If this segment is being written synchronously, note that */
1606 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1607 		cl->flags |= LFS_CL_SYNC;
1608 		cl->seg = fs->lfs_sp;
1609 		++cl->seg->seg_iocount;
1610 		/* printf("+ %x => %d\n", cl->seg, cl->seg->seg_iocount); */
1611 	}
1612 
1613 	/* Get an empty buffer header, or maybe one with something on it */
1614 	s = splbio();
1615 	bp = pool_get(&bufpool, PR_WAITOK); /* XXX should use lfs_malloc? */
1616 	splx(s);
1617 	memset(bp, 0, sizeof(*bp));
1618 	BUF_INIT(bp);
1619 
1620 	bp->b_flags = B_BUSY | B_CALL;
1621 	bp->b_dev = NODEV;
1622 	bp->b_blkno = bp->b_lblkno = addr;
1623 	bp->b_iodone = lfs_cluster_callback;
1624 	bp->b_saveaddr = (caddr_t)cl;
1625 	bp->b_vp = vp;
1626 
1627 	return bp;
1628 }
1629 
1630 int
1631 lfs_writeseg(struct lfs *fs, struct segment *sp)
1632 {
1633 	struct buf **bpp, *bp, *cbp, *newbp;
1634 	SEGUSE *sup;
1635 	SEGSUM *ssp;
1636 	dev_t i_dev;
1637 	char *datap, *dp;
1638 	int i, s;
1639 	int do_again, nblocks, byteoffset;
1640 	size_t el_size;
1641 	struct lfs_cluster *cl;
1642 	int (*strategy)(void *);
1643 	struct vop_strategy_args vop_strategy_a;
1644 	u_short ninos;
1645 	struct vnode *devvp;
1646 	char *p;
1647 	struct vnode *vp;
1648 	int32_t *daddrp;	/* XXX ondisk32 */
1649 	int changed;
1650 #if defined(DEBUG) && defined(LFS_PROPELLER)
1651 	static int propeller;
1652 	char propstring[4] = "-\\|/";
1653 
1654 	printf("%c\b",propstring[propeller++]);
1655 	if (propeller == 4)
1656 		propeller = 0;
1657 #endif
1658 
1659 	/*
1660 	 * If there are no buffers other than the segment summary to write
1661 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1662 	 * even if there aren't any buffers, you need to write the superblock.
1663 	 */
1664 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1665 		return (0);
1666 
1667 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1668 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1669 
1670 	/* Update the segment usage information. */
1671 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1672 
1673 	/* Loop through all blocks, except the segment summary. */
1674 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1675 		if ((*bpp)->b_vp != devvp) {
1676 			sup->su_nbytes += (*bpp)->b_bcount;
1677 #ifdef DEBUG_SU_NBYTES
1678 		printf("seg %" PRIu32 " += %ld for ino %d lbn %" PRId64
1679 		    " db 0x%" PRIx64 "\n", sp->seg_number, (*bpp)->b_bcount,
1680 		    VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
1681 		    (*bpp)->b_blkno);
1682 #endif
1683 		}
1684 	}
1685 
1686 	ssp = (SEGSUM *)sp->segsum;
1687 
1688 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1689 #ifdef DEBUG_SU_NBYTES
1690 	printf("seg %d += %d for %d inodes\n",	 /* XXXDEBUG */
1691 	       sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode),
1692 	       ssp->ss_ninos);
1693 #endif
1694 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode);
1695 	/* sup->su_nbytes += fs->lfs_sumsize; */
1696 	if (fs->lfs_version == 1)
1697 		sup->su_olastmod = time.tv_sec;
1698 	else
1699 		sup->su_lastmod = time.tv_sec;
1700 	sup->su_ninos += ninos;
1701 	++sup->su_nsums;
1702 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1703 							 fs->lfs_ibsize));
1704 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1705 
1706 	do_again = !(bp->b_flags & B_GATHERED);
1707 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
1708 
1709 	/*
1710 	 * Mark blocks B_BUSY, to prevent then from being changed between
1711 	 * the checksum computation and the actual write.
1712 	 *
1713 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1714 	 * there are any, replace them with copies that have UNASSIGNED
1715 	 * instead.
1716 	 */
1717 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1718 		++bpp;
1719 		bp = *bpp;
1720 		if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */
1721 			bp->b_flags |= B_BUSY;
1722 			continue;
1723 		}
1724 	    again:
1725 		s = splbio();
1726 		if (bp->b_flags & B_BUSY) {
1727 #ifdef DEBUG
1728 			printf("lfs_writeseg: avoiding potential data summary "
1729 			       "corruption for ino %d, lbn %" PRId64 "\n",
1730 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1731 #endif
1732 			bp->b_flags |= B_WANTED;
1733 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1734 			splx(s);
1735 			goto again;
1736 		}
1737 		bp->b_flags |= B_BUSY;
1738 		splx(s);
1739 		/*
1740 		 * Check and replace indirect block UNWRITTEN bogosity.
1741 		 * XXX See comment in lfs_writefile.
1742 		 */
1743 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1744 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
1745 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1746 #ifdef DEBUG_LFS
1747 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1748 			       VTOI(bp->b_vp)->i_number,
1749 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1750 			       VTOI(bp->b_vp)->i_ffs1_blocks);
1751 #endif
1752 			/* Make a copy we'll make changes to */
1753 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1754 					   bp->b_bcount, LFS_NB_IBLOCK);
1755 			newbp->b_blkno = bp->b_blkno;
1756 			memcpy(newbp->b_data, bp->b_data,
1757 			       newbp->b_bcount);
1758 
1759 			changed = 0;
1760 			/* XXX ondisk32 */
1761 			for (daddrp = (int32_t *)(newbp->b_data);
1762 			     daddrp < (int32_t *)(newbp->b_data +
1763 						  newbp->b_bcount); daddrp++) {
1764 				if (*daddrp == UNWRITTEN) {
1765 #ifdef DEBUG_LFS
1766 					off_t doff;
1767 					int32_t ioff;
1768 
1769 					ioff =
1770 					    daddrp - (int32_t *)(newbp->b_data);
1771 					doff =
1772 					 (-bp->b_lblkno + ioff) * fs->lfs_bsize;
1773 					printf("ino %d lbn %" PRId64
1774 					    " entry %d off %" PRIx64 "\n",
1775 					    VTOI(bp->b_vp)->i_number,
1776 					    bp->b_lblkno, ioff, doff);
1777 					if (bp->b_vp->v_type == VREG) {
1778 						/*
1779 						 * What is up with this page?
1780 						 */
1781 		struct vm_page *pg;
1782 		for (; doff / fs->lfs_bsize == (-bp->b_lblkno + ioff);
1783 		    doff += PAGE_SIZE) {
1784 			pg = uvm_pagelookup(&bp->b_vp->v_uobj, doff);
1785 			if (pg == NULL)
1786 				printf("  page at %" PRIx64 " is NULL\n", doff);
1787 			else
1788 				printf("  page at %" PRIx64
1789 				    " flags 0x%x pqflags 0x%x\n",
1790 				    doff, pg->flags, pg->pqflags);
1791 		}
1792 					}
1793 #endif /* DEBUG_LFS */
1794 					++changed;
1795 					*daddrp = 0;
1796 				}
1797 			}
1798 			/*
1799 			 * Get rid of the old buffer.  Don't mark it clean,
1800 			 * though, if it still has dirty data on it.
1801 			 */
1802 			if (changed) {
1803 #ifdef DEBUG_LFS
1804 				printf("lfs_writeseg: replacing UNWRITTEN(%d):"
1805 					" bp = %p newbp = %p\n", changed, bp,
1806 					newbp);
1807 #endif
1808 				*bpp = newbp;
1809 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1810 				if (bp->b_flags & B_CALL) {
1811 					printf("lfs_writeseg: "
1812 					    "indir bp should not be B_CALL\n");
1813 					s = splbio();
1814 					biodone(bp);
1815 					splx(s);
1816 					bp = NULL;
1817 				} else {
1818 					/* Still on free list, leave it there */
1819 					s = splbio();
1820 					bp->b_flags &= ~B_BUSY;
1821 					if (bp->b_flags & B_WANTED)
1822 						wakeup(bp);
1823 					splx(s);
1824 					/*
1825 					 * We have to re-decrement lfs_avail
1826 					 * since this block is going to come
1827 					 * back around to us in the next
1828 					 * segment.
1829 					 */
1830 					fs->lfs_avail -=
1831 					    btofsb(fs, bp->b_bcount);
1832 				}
1833 			} else {
1834 				lfs_freebuf(fs, newbp);
1835 			}
1836 		}
1837 	}
1838 	/*
1839 	 * Compute checksum across data and then across summary; the first
1840 	 * block (the summary block) is skipped.  Set the create time here
1841 	 * so that it's guaranteed to be later than the inode mod times.
1842 	 *
1843 	 * XXX
1844 	 * Fix this to do it inline, instead of malloc/copy.
1845 	 */
1846 	datap = dp = pool_get(&fs->lfs_bpppool, PR_WAITOK);
1847 	if (fs->lfs_version == 1)
1848 		el_size = sizeof(u_long);
1849 	else
1850 		el_size = sizeof(u_int32_t);
1851 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
1852 		++bpp;
1853 		/* Loop through gop_write cluster blocks */
1854 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
1855 		     byteoffset += fs->lfs_bsize) {
1856 #ifdef LFS_USE_B_INVAL
1857 			if (((*bpp)->b_flags & (B_CALL | B_INVAL)) ==
1858 			    (B_CALL | B_INVAL)) {
1859 				if (copyin((caddr_t)(*bpp)->b_saveaddr +
1860 					   byteoffset, dp, el_size)) {
1861 					panic("lfs_writeseg: copyin failed [1]:"
1862 						" ino %d blk %" PRId64,
1863 						VTOI((*bpp)->b_vp)->i_number,
1864 						(*bpp)->b_lblkno);
1865 				}
1866 			} else
1867 #endif /* LFS_USE_B_INVAL */
1868 			{
1869 				memcpy(dp, (*bpp)->b_data + byteoffset,
1870 				       el_size);
1871 			}
1872 			dp += el_size;
1873 		}
1874 	}
1875 	if (fs->lfs_version == 1)
1876 		ssp->ss_ocreate = time.tv_sec;
1877 	else {
1878 		ssp->ss_create = time.tv_sec;
1879 		ssp->ss_serial = ++fs->lfs_serial;
1880 		ssp->ss_ident  = fs->lfs_ident;
1881 	}
1882 	ssp->ss_datasum = cksum(datap, dp - datap);
1883 	ssp->ss_sumsum =
1884 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1885 	pool_put(&fs->lfs_bpppool, datap);
1886 	datap = dp = NULL;
1887 #ifdef DIAGNOSTIC
1888 	if (fs->lfs_bfree <
1889 	    btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1890 		panic("lfs_writeseg: No diskspace for summary");
1891 #endif
1892 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1893 			  btofsb(fs, fs->lfs_sumsize));
1894 
1895 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1896 
1897 	/*
1898 	 * When we simply write the blocks we lose a rotation for every block
1899 	 * written.  To avoid this problem, we cluster the buffers into a
1900 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
1901 	 * devices can handle, use that for the size of the chunks.
1902 	 *
1903 	 * Blocks that are already clusters (from GOP_WRITE), however, we
1904 	 * don't bother to copy into other clusters.
1905 	 */
1906 
1907 #define CHUNKSIZE MAXPHYS
1908 
1909 	if (devvp == NULL)
1910 		panic("devvp is NULL");
1911 	for (bpp = sp->bpp, i = nblocks; i;) {
1912 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1913 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1914 
1915 		cbp->b_dev = i_dev;
1916 		cbp->b_flags |= B_ASYNC | B_BUSY;
1917 		cbp->b_bcount = 0;
1918 
1919 #if defined(DEBUG) && defined(DIAGNOSTIC)
1920 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
1921 		    / sizeof(int32_t)) {
1922 			panic("lfs_writeseg: real bpp overwrite");
1923 		}
1924 		if (bpp - sp->bpp > fs->lfs_ssize / fs->lfs_fsize) {
1925 			panic("lfs_writeseg: theoretical bpp overwrite");
1926 		}
1927 #endif
1928 
1929 		/*
1930 		 * Construct the cluster.
1931 		 */
1932 		++fs->lfs_iocount;
1933 		while (i && cbp->b_bcount < CHUNKSIZE) {
1934 			bp = *bpp;
1935 
1936 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1937 				break;
1938 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
1939 				break;
1940 
1941 			/* Clusters from GOP_WRITE are expedited */
1942 			if (bp->b_bcount > fs->lfs_bsize) {
1943 				if (cbp->b_bcount > 0)
1944 					/* Put in its own buffer */
1945 					break;
1946 				else {
1947 					cbp->b_data = bp->b_data;
1948 				}
1949 			} else if (cbp->b_bcount == 0) {
1950 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
1951 							     LFS_NB_CLUSTER);
1952 				cl->flags |= LFS_CL_MALLOC;
1953 			}
1954 #ifdef DIAGNOSTIC
1955 			if (dtosn(fs, dbtofsb(fs, bp->b_blkno +
1956 					      btodb(bp->b_bcount - 1))) !=
1957 			    sp->seg_number) {
1958 				printf("blk size %ld daddr %" PRIx64
1959 				    " not in seg %d\n",
1960 				    bp->b_bcount, bp->b_blkno,
1961 				    sp->seg_number);
1962 				panic("segment overwrite");
1963 			}
1964 #endif
1965 
1966 #ifdef LFS_USE_B_INVAL
1967 			/*
1968 			 * Fake buffers from the cleaner are marked as B_INVAL.
1969 			 * We need to copy the data from user space rather than
1970 			 * from the buffer indicated.
1971 			 * XXX == what do I do on an error?
1972 			 */
1973 			if ((bp->b_flags & (B_CALL|B_INVAL)) ==
1974 			    (B_CALL|B_INVAL)) {
1975 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1976 					panic("lfs_writeseg: "
1977 					    "copyin failed [2]");
1978 			} else
1979 #endif /* LFS_USE_B_INVAL */
1980 			if (cl->flags & LFS_CL_MALLOC) {
1981 				bcopy(bp->b_data, p, bp->b_bcount);
1982 			}
1983 
1984 			p += bp->b_bcount;
1985 			cbp->b_bcount += bp->b_bcount;
1986 			cl->bufsize += bp->b_bcount;
1987 
1988 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
1989 			cl->bpp[cl->bufcount++] = bp;
1990 			vp = bp->b_vp;
1991 			s = splbio();
1992 			V_INCR_NUMOUTPUT(vp);
1993 			splx(s);
1994 
1995 			bpp++;
1996 			i--;
1997 		}
1998 		s = splbio();
1999 		V_INCR_NUMOUTPUT(devvp);
2000 		splx(s);
2001 		vop_strategy_a.a_desc = VDESC(vop_strategy);
2002 		vop_strategy_a.a_bp = cbp;
2003 		(strategy)(&vop_strategy_a);
2004 		curproc->p_stats->p_ru.ru_oublock++;
2005 	}
2006 
2007 	if (lfs_dostats) {
2008 		++lfs_stats.psegwrites;
2009 		lfs_stats.blocktot += nblocks - 1;
2010 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2011 			++lfs_stats.psyncwrites;
2012 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2013 			++lfs_stats.pcleanwrites;
2014 			lfs_stats.cleanblocks += nblocks - 1;
2015 		}
2016 	}
2017 	return (lfs_initseg(fs) || do_again);
2018 }
2019 
2020 void
2021 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2022 {
2023 	struct buf *bp;
2024 	dev_t i_dev;
2025 	int (*strategy)(void *);
2026 	int s;
2027 	struct vop_strategy_args vop_strategy_a;
2028 
2029 	/*
2030 	 * If we can write one superblock while another is in
2031 	 * progress, we risk not having a complete checkpoint if we crash.
2032 	 * So, block here if a superblock write is in progress.
2033 	 */
2034 	s = splbio();
2035 	while (fs->lfs_sbactive) {
2036 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
2037 	}
2038 	fs->lfs_sbactive = daddr;
2039 	splx(s);
2040 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
2041 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
2042 
2043 	/* Set timestamp of this version of the superblock */
2044 	if (fs->lfs_version == 1)
2045 		fs->lfs_otstamp = time.tv_sec;
2046 	fs->lfs_tstamp = time.tv_sec;
2047 
2048 	/* Checksum the superblock and copy it into a buffer. */
2049 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2050 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
2051 	    fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2052 	memset(bp->b_data + sizeof(struct dlfs), 0,
2053 	    LFS_SBPAD - sizeof(struct dlfs));
2054 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2055 
2056 	bp->b_dev = i_dev;
2057 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2058 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2059 	bp->b_iodone = lfs_supercallback;
2060 	/* XXX KS - same nasty hack as above */
2061 	bp->b_saveaddr = (caddr_t)fs;
2062 
2063 	vop_strategy_a.a_desc = VDESC(vop_strategy);
2064 	vop_strategy_a.a_bp = bp;
2065 	curproc->p_stats->p_ru.ru_oublock++;
2066 	s = splbio();
2067 	V_INCR_NUMOUTPUT(bp->b_vp);
2068 	splx(s);
2069 	++fs->lfs_iocount;
2070 	(strategy)(&vop_strategy_a);
2071 }
2072 
2073 /*
2074  * Logical block number match routines used when traversing the dirty block
2075  * chain.
2076  */
2077 int
2078 lfs_match_fake(struct lfs *fs, struct buf *bp)
2079 {
2080 
2081 	return LFS_IS_MALLOC_BUF(bp);
2082 }
2083 
2084 #if 0
2085 int
2086 lfs_match_real(struct lfs *fs, struct buf *bp)
2087 {
2088 
2089 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2090 }
2091 #endif
2092 
2093 int
2094 lfs_match_data(struct lfs *fs, struct buf *bp)
2095 {
2096 
2097 	return (bp->b_lblkno >= 0);
2098 }
2099 
2100 int
2101 lfs_match_indir(struct lfs *fs, struct buf *bp)
2102 {
2103 	daddr_t lbn;
2104 
2105 	lbn = bp->b_lblkno;
2106 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2107 }
2108 
2109 int
2110 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2111 {
2112 	daddr_t lbn;
2113 
2114 	lbn = bp->b_lblkno;
2115 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2116 }
2117 
2118 int
2119 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2120 {
2121 	daddr_t lbn;
2122 
2123 	lbn = bp->b_lblkno;
2124 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2125 }
2126 
2127 /*
2128  * XXX - The only buffers that are going to hit these functions are the
2129  * segment write blocks, or the segment summaries, or the superblocks.
2130  *
2131  * All of the above are created by lfs_newbuf, and so do not need to be
2132  * released via brelse.
2133  */
2134 void
2135 lfs_callback(struct buf *bp)
2136 {
2137 	struct lfs *fs;
2138 
2139 	fs = (struct lfs *)bp->b_saveaddr;
2140 	lfs_freebuf(fs, bp);
2141 }
2142 
2143 static void
2144 lfs_super_aiodone(struct buf *bp)
2145 {
2146 	struct lfs *fs;
2147 
2148 	fs = (struct lfs *)bp->b_saveaddr;
2149 	fs->lfs_sbactive = 0;
2150 	wakeup(&fs->lfs_sbactive);
2151 	if (--fs->lfs_iocount <= 1)
2152 		wakeup(&fs->lfs_iocount);
2153 	lfs_freebuf(fs, bp);
2154 }
2155 
2156 static void
2157 lfs_cluster_aiodone(struct buf *bp)
2158 {
2159 	struct lfs_cluster *cl;
2160 	struct lfs *fs;
2161 	struct buf *tbp, *fbp;
2162 	struct vnode *vp, *devvp;
2163 	struct inode *ip;
2164 	int s, error=0;
2165 	char *cp;
2166 
2167 	if (bp->b_flags & B_ERROR)
2168 		error = bp->b_error;
2169 
2170 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2171 	fs = cl->fs;
2172 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2173 
2174 	cp = (char *)bp->b_data + cl->bufsize;
2175 	/* Put the pages back, and release the buffer */
2176 	while (cl->bufcount--) {
2177 		tbp = cl->bpp[cl->bufcount];
2178 		if (error) {
2179 			tbp->b_flags |= B_ERROR;
2180 			tbp->b_error = error;
2181 		}
2182 
2183 		/*
2184 		 * We're done with tbp.	 If it has not been re-dirtied since
2185 		 * the cluster was written, free it.  Otherwise, keep it on
2186 		 * the locked list to be written again.
2187 		 */
2188 		vp = tbp->b_vp;
2189 
2190 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2191 			LFS_UNLOCK_BUF(tbp);
2192 
2193 		tbp->b_flags &= ~B_GATHERED;
2194 
2195 		LFS_BCLEAN_LOG(fs, tbp);
2196 
2197 		if (!(tbp->b_flags & B_CALL)) {
2198 			bremfree(tbp);
2199 			s = splbio();
2200 			if (vp)
2201 				reassignbuf(tbp, vp);
2202 			splx(s);
2203 			tbp->b_flags |= B_ASYNC; /* for biodone */
2204 		}
2205 #ifdef DIAGNOSTIC
2206 		if (tbp->b_flags & B_DONE) {
2207 			printf("blk %d biodone already (flags %lx)\n",
2208 				cl->bufcount, (long)tbp->b_flags);
2209 		}
2210 #endif
2211 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2212 			if ((tbp->b_flags & B_CALL) &&
2213 			    !LFS_IS_MALLOC_BUF(tbp)) {
2214 				/* printf("flags 0x%lx\n", tbp->b_flags); */
2215 				/*
2216 				 * A buffer from the page daemon.
2217 				 * We use the same iodone as it does,
2218 				 * so we must manually disassociate its
2219 				 * buffers from the vp.
2220 				 */
2221 				if (tbp->b_vp) {
2222 					/* This is just silly */
2223 					s = splbio();
2224 					brelvp(tbp);
2225 					tbp->b_vp = vp;
2226 					splx(s);
2227 				}
2228 				/* Put it back the way it was */
2229 				tbp->b_flags |= B_ASYNC;
2230 				/* Master buffers have B_AGE */
2231 				if (tbp->b_private == tbp)
2232 					tbp->b_flags |= B_AGE;
2233 			}
2234 			s = splbio();
2235 			biodone(tbp);
2236 
2237 			/*
2238 			 * If this is the last block for this vnode, but
2239 			 * there are other blocks on its dirty list,
2240 			 * set IN_MODIFIED/IN_CLEANING depending on what
2241 			 * sort of block.  Only do this for our mount point,
2242 			 * not for, e.g., inode blocks that are attached to
2243 			 * the devvp.
2244 			 * XXX KS - Shouldn't we set *both* if both types
2245 			 * of blocks are present (traverse the dirty list?)
2246 			 */
2247 			simple_lock(&global_v_numoutput_slock);
2248 			if (vp != devvp && vp->v_numoutput == 0 &&
2249 			    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2250 				ip = VTOI(vp);
2251 #ifdef DEBUG_LFS
2252 				printf("lfs_cluster_aiodone: marking ino %d\n",
2253 				       ip->i_number);
2254 #endif
2255 				if (LFS_IS_MALLOC_BUF(fbp))
2256 					LFS_SET_UINO(ip, IN_CLEANING);
2257 				else
2258 					LFS_SET_UINO(ip, IN_MODIFIED);
2259 			}
2260 			simple_unlock(&global_v_numoutput_slock);
2261 			splx(s);
2262 			wakeup(vp);
2263 		}
2264 	}
2265 
2266 	/* Fix up the cluster buffer, and release it */
2267 	if (cl->flags & LFS_CL_MALLOC)
2268 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2269 	s = splbio();
2270 	pool_put(&bufpool, bp); /* XXX should use lfs_free? */
2271 	splx(s);
2272 
2273 	/* Note i/o done */
2274 	if (cl->flags & LFS_CL_SYNC) {
2275 		if (--cl->seg->seg_iocount == 0)
2276 			wakeup(&cl->seg->seg_iocount);
2277 		/* printf("- %x => %d\n", cl->seg, cl->seg->seg_iocount); */
2278 	}
2279 #ifdef DIAGNOSTIC
2280 	if (fs->lfs_iocount == 0)
2281 		panic("lfs_cluster_aiodone: zero iocount");
2282 #endif
2283 	if (--fs->lfs_iocount <= 1)
2284 		wakeup(&fs->lfs_iocount);
2285 
2286 	pool_put(&fs->lfs_bpppool, cl->bpp);
2287 	cl->bpp = NULL;
2288 	pool_put(&fs->lfs_clpool, cl);
2289 }
2290 
2291 static void
2292 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2293 {
2294 	/* reset b_iodone for when this is a single-buf i/o. */
2295 	bp->b_iodone = aiodone;
2296 
2297 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
2298 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2299 	wakeup(&uvm.aiodoned);
2300 	simple_unlock(&uvm.aiodoned_lock);
2301 }
2302 
2303 static void
2304 lfs_cluster_callback(struct buf *bp)
2305 {
2306 
2307 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2308 }
2309 
2310 void
2311 lfs_supercallback(struct buf *bp)
2312 {
2313 
2314 	lfs_generic_callback(bp, lfs_super_aiodone);
2315 }
2316 
2317 /*
2318  * Shellsort (diminishing increment sort) from Data Structures and
2319  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2320  * see also Knuth Vol. 3, page 84.  The increments are selected from
2321  * formula (8), page 95.  Roughly O(N^3/2).
2322  */
2323 /*
2324  * This is our own private copy of shellsort because we want to sort
2325  * two parallel arrays (the array of buffer pointers and the array of
2326  * logical block numbers) simultaneously.  Note that we cast the array
2327  * of logical block numbers to a unsigned in this routine so that the
2328  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2329  */
2330 
2331 void
2332 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2333 {
2334 	static int __rsshell_increments[] = { 4, 1, 0 };
2335 	int incr, *incrp, t1, t2;
2336 	struct buf *bp_temp;
2337 
2338 #ifdef DEBUG
2339 	incr = 0;
2340 	for (t1 = 0; t1 < nmemb; t1++) {
2341 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2342 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2343 				/* dump before panic */
2344 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2345 				    nmemb, size);
2346 				incr = 0;
2347 				for (t1 = 0; t1 < nmemb; t1++) {
2348 					const struct buf *bp = bp_array[t1];
2349 
2350 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2351 					    PRIu64 "\n", t1,
2352 					    (uint64_t)bp->b_bcount,
2353 					    (uint64_t)bp->b_lblkno);
2354 					printf("lbns:");
2355 					for (t2 = 0; t2 * size < bp->b_bcount;
2356 					    t2++) {
2357 						printf(" %" PRId32,
2358 						    lb_array[incr++]);
2359 					}
2360 					printf("\n");
2361 				}
2362 				panic("lfs_shellsort: inconsistent input");
2363 			}
2364 		}
2365 	}
2366 #endif
2367 
2368 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2369 		for (t1 = incr; t1 < nmemb; ++t1)
2370 			for (t2 = t1 - incr; t2 >= 0;)
2371 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2372 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2373 					bp_temp = bp_array[t2];
2374 					bp_array[t2] = bp_array[t2 + incr];
2375 					bp_array[t2 + incr] = bp_temp;
2376 					t2 -= incr;
2377 				} else
2378 					break;
2379 
2380 	/* Reform the list of logical blocks */
2381 	incr = 0;
2382 	for (t1 = 0; t1 < nmemb; t1++) {
2383 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2384 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2385 		}
2386 	}
2387 }
2388 
2389 /*
2390  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2391  */
2392 int
2393 lfs_vref(struct vnode *vp)
2394 {
2395 	/*
2396 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2397 	 * being able to flush all of the pages from this vnode, which
2398 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2399 	 */
2400 	if (vp->v_flag & VXLOCK) {
2401 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2402 			return 0;
2403 		}
2404 		return (1);
2405 	}
2406 	return (vget(vp, 0));
2407 }
2408 
2409 /*
2410  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2411  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2412  */
2413 void
2414 lfs_vunref(struct vnode *vp)
2415 {
2416 	/*
2417 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2418 	 */
2419 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2420 		return;
2421 	}
2422 
2423 	simple_lock(&vp->v_interlock);
2424 #ifdef DIAGNOSTIC
2425 	if (vp->v_usecount <= 0) {
2426 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2427 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2428 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2429 		panic("lfs_vunref: v_usecount<0");
2430 	}
2431 #endif
2432 	vp->v_usecount--;
2433 	if (vp->v_usecount > 0) {
2434 		simple_unlock(&vp->v_interlock);
2435 		return;
2436 	}
2437 	/*
2438 	 * insert at tail of LRU list
2439 	 */
2440 	simple_lock(&vnode_free_list_slock);
2441 	if (vp->v_holdcnt > 0)
2442 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2443 	else
2444 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2445 	simple_unlock(&vnode_free_list_slock);
2446 	simple_unlock(&vp->v_interlock);
2447 }
2448 
2449 /*
2450  * We use this when we have vnodes that were loaded in solely for cleaning.
2451  * There is no reason to believe that these vnodes will be referenced again
2452  * soon, since the cleaning process is unrelated to normal filesystem
2453  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2454  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2455  * cleaning at the head of the list, instead.
2456  */
2457 void
2458 lfs_vunref_head(struct vnode *vp)
2459 {
2460 
2461 	simple_lock(&vp->v_interlock);
2462 #ifdef DIAGNOSTIC
2463 	if (vp->v_usecount == 0) {
2464 		panic("lfs_vunref: v_usecount<0");
2465 	}
2466 #endif
2467 	vp->v_usecount--;
2468 	if (vp->v_usecount > 0) {
2469 		simple_unlock(&vp->v_interlock);
2470 		return;
2471 	}
2472 	/*
2473 	 * insert at head of LRU list
2474 	 */
2475 	simple_lock(&vnode_free_list_slock);
2476 	if (vp->v_holdcnt > 0)
2477 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2478 	else
2479 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2480 	simple_unlock(&vnode_free_list_slock);
2481 	simple_unlock(&vp->v_interlock);
2482 }
2483 
2484