xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 001c68bd94f75ce9270b69227c4199fbf34ee396)
1 /*	$NetBSD: lfs_segment.c,v 1.126 2003/07/02 13:43:03 yamt 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. All advertising materials mentioning features or use of this software
51  *    must display the following acknowledgement:
52  *	This product includes software developed by the University of
53  *	California, Berkeley and its contributors.
54  * 4. Neither the name of the University nor the names of its contributors
55  *    may be used to endorse or promote products derived from this software
56  *    without specific prior written permission.
57  *
58  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
59  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
60  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
61  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
62  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
63  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
64  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
65  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
66  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
67  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68  * SUCH DAMAGE.
69  *
70  *	@(#)lfs_segment.c	8.10 (Berkeley) 6/10/95
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: lfs_segment.c,v 1.126 2003/07/02 13:43:03 yamt Exp $");
75 
76 #define ivndebug(vp,str) printf("ino %d: %s\n",VTOI(vp)->i_number,(str))
77 
78 #if defined(_KERNEL_OPT)
79 #include "opt_ddb.h"
80 #endif
81 
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/namei.h>
85 #include <sys/kernel.h>
86 #include <sys/resourcevar.h>
87 #include <sys/file.h>
88 #include <sys/stat.h>
89 #include <sys/buf.h>
90 #include <sys/proc.h>
91 #include <sys/vnode.h>
92 #include <sys/mount.h>
93 
94 #include <miscfs/specfs/specdev.h>
95 #include <miscfs/fifofs/fifo.h>
96 
97 #include <ufs/ufs/inode.h>
98 #include <ufs/ufs/dir.h>
99 #include <ufs/ufs/ufsmount.h>
100 #include <ufs/ufs/ufs_extern.h>
101 
102 #include <ufs/lfs/lfs.h>
103 #include <ufs/lfs/lfs_extern.h>
104 
105 #include <uvm/uvm.h>
106 #include <uvm/uvm_extern.h>
107 
108 MALLOC_DEFINE(M_SEGMENT, "LFS segment", "Segment for LFS");
109 
110 extern int count_lock_queue(void);
111 extern struct simplelock vnode_free_list_slock;		/* XXX */
112 
113 static void lfs_generic_callback(struct buf *, void (*)(struct buf *));
114 static void lfs_super_aiodone(struct buf *);
115 static void lfs_cluster_aiodone(struct buf *);
116 static void lfs_cluster_callback(struct buf *);
117 
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 	((fs)->lfs_fsbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
124 	fragstofsb((fs), (fs)->lfs_frag))
125 
126 void	 lfs_callback(struct buf *);
127 int	 lfs_gather(struct lfs *, struct segment *,
128 	     struct vnode *, int (*)(struct lfs *, struct buf *));
129 int	 lfs_gatherblock(struct segment *, struct buf *, int *);
130 void	 lfs_iset(struct inode *, daddr_t, time_t);
131 int	 lfs_match_fake(struct lfs *, struct buf *);
132 int	 lfs_match_data(struct lfs *, struct buf *);
133 int	 lfs_match_dindir(struct lfs *, struct buf *);
134 int	 lfs_match_indir(struct lfs *, struct buf *);
135 int	 lfs_match_tindir(struct lfs *, struct buf *);
136 void	 lfs_newseg(struct lfs *);
137 /* XXX ondisk32 */
138 void	 lfs_shellsort(struct buf **, int32_t *, int, int);
139 void	 lfs_supercallback(struct buf *);
140 void	 lfs_updatemeta(struct segment *);
141 int	 lfs_vref(struct vnode *);
142 void	 lfs_vunref(struct vnode *);
143 void	 lfs_writefile(struct lfs *, struct segment *, struct vnode *);
144 int	 lfs_writeinode(struct lfs *, struct segment *, struct inode *);
145 int	 lfs_writeseg(struct lfs *, struct segment *);
146 void	 lfs_writesuper(struct lfs *, daddr_t);
147 int	 lfs_writevnodes(struct lfs *fs, struct mount *mp,
148 	    struct segment *sp, int dirops);
149 
150 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
151 int	lfs_writeindir = 1;		/* whether to flush indir on non-ckp */
152 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
153 int	lfs_dirvcount = 0;		/* # active dirops */
154 
155 /* Statistics Counters */
156 int lfs_dostats = 1;
157 struct lfs_stats lfs_stats;
158 
159 extern int locked_queue_count;
160 extern long locked_queue_bytes;
161 
162 /* op values to lfs_writevnodes */
163 #define	VN_REG		0
164 #define	VN_DIROP	1
165 #define	VN_EMPTY	2
166 #define VN_CLEAN	3
167 
168 #define LFS_MAX_ACTIVE		10
169 
170 /*
171  * XXX KS - Set modification time on the Ifile, so the cleaner can
172  * read the fs mod time off of it.  We don't set IN_UPDATE here,
173  * since we don't really need this to be flushed to disk (and in any
174  * case that wouldn't happen to the Ifile until we checkpoint).
175  */
176 void
177 lfs_imtime(struct lfs *fs)
178 {
179 	struct timespec ts;
180 	struct inode *ip;
181 
182 	TIMEVAL_TO_TIMESPEC(&time, &ts);
183 	ip = VTOI(fs->lfs_ivnode);
184 	ip->i_ffs1_mtime = ts.tv_sec;
185 	ip->i_ffs1_mtimensec = ts.tv_nsec;
186 }
187 
188 /*
189  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
190  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
191  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
192  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
193  */
194 
195 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
196 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
197 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
198 
199 int
200 lfs_vflush(struct vnode *vp)
201 {
202 	struct inode *ip;
203 	struct lfs *fs;
204 	struct segment *sp;
205 	struct buf *bp, *nbp, *tbp, *tnbp;
206 	int error, s;
207 	int flushed;
208 #if 0
209 	int redo;
210 #endif
211 
212 	ip = VTOI(vp);
213 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
214 
215 	if (ip->i_flag & IN_CLEANING) {
216 #ifdef DEBUG_LFS
217 		ivndebug(vp,"vflush/in_cleaning");
218 #endif
219 		LFS_CLR_UINO(ip, IN_CLEANING);
220 		LFS_SET_UINO(ip, IN_MODIFIED);
221 
222 		/*
223 		 * Toss any cleaning buffers that have real counterparts
224 		 * to avoid losing new data.
225 		 */
226 		s = splbio();
227 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
228 			nbp = LIST_NEXT(bp, b_vnbufs);
229 			if (!LFS_IS_MALLOC_BUF(bp))
230 				continue;
231 			/*
232 			 * Look for pages matching the range covered
233 			 * by cleaning blocks.  It's okay if more dirty
234 			 * pages appear, so long as none disappear out
235 			 * from under us.
236 			 */
237 			if (bp->b_lblkno > 0 && vp->v_type == VREG &&
238 			    vp != fs->lfs_ivnode) {
239 				struct vm_page *pg;
240 				voff_t off;
241 
242 				simple_lock(&vp->v_interlock);
243 				for (off = lblktosize(fs, bp->b_lblkno);
244 				     off < lblktosize(fs, bp->b_lblkno + 1);
245 				     off += PAGE_SIZE) {
246 					pg = uvm_pagelookup(&vp->v_uobj, off);
247 					if (pg == NULL)
248 						continue;
249 					if ((pg->flags & PG_CLEAN) == 0 ||
250 					    pmap_is_modified(pg)) {
251 						fs->lfs_avail += btofsb(fs,
252 							bp->b_bcount);
253 						wakeup(&fs->lfs_avail);
254 						lfs_freebuf(fs, bp);
255 						bp = NULL;
256 						goto nextbp;
257 					}
258 				}
259 				simple_unlock(&vp->v_interlock);
260 			}
261 			for (tbp = LIST_FIRST(&vp->v_dirtyblkhd); tbp;
262 			    tbp = tnbp)
263 			{
264 				tnbp = LIST_NEXT(tbp, b_vnbufs);
265 				if (tbp->b_vp == bp->b_vp
266 				   && tbp->b_lblkno == bp->b_lblkno
267 				   && tbp != bp)
268 				{
269 					fs->lfs_avail += btofsb(fs,
270 						bp->b_bcount);
271 					wakeup(&fs->lfs_avail);
272 					lfs_freebuf(fs, bp);
273 					bp = NULL;
274 					break;
275 				}
276 			}
277 		    nextbp:
278 			;
279 		}
280 		splx(s);
281 	}
282 
283 	/* If the node is being written, wait until that is done */
284 	s = splbio();
285 	if (WRITEINPROG(vp)) {
286 #ifdef DEBUG_LFS
287 		ivndebug(vp,"vflush/writeinprog");
288 #endif
289 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
290 	}
291 	splx(s);
292 
293 	/* Protect against VXLOCK deadlock in vinvalbuf() */
294 	lfs_seglock(fs, SEGM_SYNC);
295 
296 	/* If we're supposed to flush a freed inode, just toss it */
297 	/* XXX - seglock, so these buffers can't be gathered, right? */
298 	if (ip->i_mode == 0) {
299 		printf("lfs_vflush: ino %d is freed, not flushing\n",
300 			ip->i_number);
301 		s = splbio();
302 		for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
303 			nbp = LIST_NEXT(bp, b_vnbufs);
304 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
305 				fs->lfs_avail += btofsb(fs, bp->b_bcount);
306 				wakeup(&fs->lfs_avail);
307 			}
308 			/* Copied from lfs_writeseg */
309 			if (bp->b_flags & B_CALL) {
310 				biodone(bp);
311 			} else {
312 				bremfree(bp);
313 				LFS_UNLOCK_BUF(bp);
314 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
315 					 B_GATHERED);
316 				bp->b_flags |= B_DONE;
317 				reassignbuf(bp, vp);
318 				brelse(bp);
319 			}
320 		}
321 		splx(s);
322 		LFS_CLR_UINO(ip, IN_CLEANING);
323 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
324 		ip->i_flag &= ~IN_ALLMOD;
325 		printf("lfs_vflush: done not flushing ino %d\n",
326 			ip->i_number);
327 		lfs_segunlock(fs);
328 		return 0;
329 	}
330 
331 	SET_FLUSHING(fs,vp);
332 	if (fs->lfs_nactive > LFS_MAX_ACTIVE ||
333 	    (fs->lfs_sp->seg_flags & SEGM_CKP)) {
334 		error = lfs_segwrite(vp->v_mount, SEGM_CKP | SEGM_SYNC);
335 		CLR_FLUSHING(fs,vp);
336 		lfs_segunlock(fs);
337 		return error;
338 	}
339 	sp = fs->lfs_sp;
340 
341 	flushed = 0;
342 	if (VPISEMPTY(vp)) {
343 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
344 		++flushed;
345 	} else if ((ip->i_flag & IN_CLEANING) &&
346 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
347 #ifdef DEBUG_LFS
348 		ivndebug(vp,"vflush/clean");
349 #endif
350 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
351 		++flushed;
352 	} else if (lfs_dostats) {
353 		if (!VPISEMPTY(vp) || (VTOI(vp)->i_flag & IN_ALLMOD))
354 			++lfs_stats.vflush_invoked;
355 #ifdef DEBUG_LFS
356 		ivndebug(vp,"vflush");
357 #endif
358 	}
359 
360 #ifdef DIAGNOSTIC
361 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
362 	if (vp->v_flag & VDIROP) {
363 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
364 		/* panic("VDIROP being flushed...this can\'t happen"); */
365 	}
366 	if (vp->v_usecount < 0) {
367 		printf("usecount=%ld\n", (long)vp->v_usecount);
368 		panic("lfs_vflush: usecount<0");
369 	}
370 #endif
371 
372 #if 1
373 	do {
374 		do {
375 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
376 				lfs_writefile(fs, sp, vp);
377 		} while (lfs_writeinode(fs, sp, ip));
378 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
379 #else
380 	if (flushed && vp != fs->lfs_ivnode)
381 		lfs_writeseg(fs, sp);
382 	else do {
383 		fs->lfs_flags &= ~LFS_IFDIRTY;
384 		lfs_writefile(fs, sp, vp);
385 		redo = lfs_writeinode(fs, sp, ip);
386 		redo += lfs_writeseg(fs, sp);
387 		redo += (fs->lfs_flags & LFS_IFDIRTY);
388 	} while (redo && vp == fs->lfs_ivnode);
389 #endif
390 	if (lfs_dostats) {
391 		++lfs_stats.nwrites;
392 		if (sp->seg_flags & SEGM_SYNC)
393 			++lfs_stats.nsync_writes;
394 		if (sp->seg_flags & SEGM_CKP)
395 			++lfs_stats.ncheckpoints;
396 	}
397 	/*
398 	 * If we were called from somewhere that has already held the seglock
399 	 * (e.g., lfs_markv()), the lfs_segunlock will not wait for
400 	 * the write to complete because we are still locked.
401 	 * Since lfs_vflush() must return the vnode with no dirty buffers,
402 	 * we must explicitly wait, if that is the case.
403 	 *
404 	 * We compare the iocount against 1, not 0, because it is
405 	 * artificially incremented by lfs_seglock().
406 	 */
407 	simple_lock(&fs->lfs_interlock);
408 	if (fs->lfs_seglock > 1) {
409 		simple_unlock(&fs->lfs_interlock);
410 		while (fs->lfs_iocount > 1)
411 			(void)tsleep(&fs->lfs_iocount, PRIBIO + 1,
412 				     "lfs_vflush", 0);
413 	} else
414 		simple_unlock(&fs->lfs_interlock);
415 
416 	lfs_segunlock(fs);
417 
418 	/* Wait for these buffers to be recovered by aiodoned */
419 	s = splbio();
420 	simple_lock(&global_v_numoutput_slock);
421 	while (vp->v_numoutput > 0) {
422 		ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vf2", 0,
423 			&global_v_numoutput_slock);
424 	}
425 	simple_unlock(&global_v_numoutput_slock);
426 	splx(s);
427 
428 	CLR_FLUSHING(fs,vp);
429 	return (0);
430 }
431 
432 #ifdef DEBUG_LFS_VERBOSE
433 # 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)
434 #else
435 # define vndebug(vp,str)
436 #endif
437 
438 int
439 lfs_writevnodes(struct lfs *fs, struct mount *mp, struct segment *sp, int op)
440 {
441 	struct inode *ip;
442 	struct vnode *vp, *nvp;
443 	int inodes_written = 0, only_cleaning;
444 
445 #ifndef LFS_NO_BACKVP_HACK
446 	/* BEGIN HACK */
447 #define	VN_OFFSET	(((caddr_t)&LIST_NEXT(vp, v_mntvnodes)) - (caddr_t)vp)
448 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)(VP)->v_mntvnodes.le_prev) - VN_OFFSET))
449 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&(LIST_FIRST(&mp->mnt_vnodelist))) - VN_OFFSET))
450 
451 	/* Find last vnode. */
452  loop:	for (vp = LIST_FIRST(&mp->mnt_vnodelist);
453 	     vp && LIST_NEXT(vp, v_mntvnodes) != NULL;
454 	     vp = LIST_NEXT(vp, v_mntvnodes));
455 	for (; vp && vp != BEG_OF_VLIST; vp = nvp) {
456 		nvp = BACK_VP(vp);
457 #else
458 	loop:
459 	for (vp = LIST_FIRST(&mp->mnt_vnodelist); vp; vp = nvp) {
460 		nvp = LIST_NEXT(vp, v_mntvnodes);
461 #endif
462 		/*
463 		 * If the vnode that we are about to sync is no longer
464 		 * associated with this mount point, start over.
465 		 */
466 		if (vp->v_mount != mp) {
467 			printf("lfs_writevnodes: starting over\n");
468 			/*
469 			 * After this, pages might be busy
470 			 * due to our own previous putpages.
471 			 * Start actual segment write here to avoid deadlock.
472 			 */
473 			(void)lfs_writeseg(fs, sp);
474 			goto loop;
475 		}
476 
477 		if (vp->v_type == VNON) {
478 			continue;
479 		}
480 
481 		ip = VTOI(vp);
482 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
483 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
484 			vndebug(vp,"dirop");
485 			continue;
486 		}
487 
488 		if (op == VN_EMPTY && !VPISEMPTY(vp)) {
489 			vndebug(vp,"empty");
490 			continue;
491 		}
492 
493 		if (op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
494 		   && vp != fs->lfs_flushvp
495 		   && !(ip->i_flag & IN_CLEANING)) {
496 			vndebug(vp,"cleaning");
497 			continue;
498 		}
499 
500 		if (lfs_vref(vp)) {
501 			vndebug(vp,"vref");
502 			continue;
503 		}
504 
505 		only_cleaning = 0;
506 		/*
507 		 * Write the inode/file if dirty and it's not the IFILE.
508 		 */
509 		if ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp)) {
510 			only_cleaning = ((ip->i_flag & IN_ALLMOD) == IN_CLEANING);
511 
512 			if (ip->i_number != LFS_IFILE_INUM)
513 				lfs_writefile(fs, sp, vp);
514 			if (!VPISEMPTY(vp)) {
515 				if (WRITEINPROG(vp)) {
516 #ifdef DEBUG_LFS
517 					ivndebug(vp,"writevnodes/write2");
518 #endif
519 				} else if (!(ip->i_flag & IN_ALLMOD)) {
520 #ifdef DEBUG_LFS
521 					printf("<%d>",ip->i_number);
522 #endif
523 					LFS_SET_UINO(ip, IN_MODIFIED);
524 				}
525 			}
526 			(void) lfs_writeinode(fs, sp, ip);
527 			inodes_written++;
528 		}
529 
530 		if (lfs_clean_vnhead && only_cleaning)
531 			lfs_vunref_head(vp);
532 		else
533 			lfs_vunref(vp);
534 	}
535 	return inodes_written;
536 }
537 
538 /*
539  * Do a checkpoint.
540  */
541 int
542 lfs_segwrite(struct mount *mp, int flags)
543 {
544 	struct buf *bp;
545 	struct inode *ip;
546 	struct lfs *fs;
547 	struct segment *sp;
548 	struct vnode *vp;
549 	SEGUSE *segusep;
550 	int do_ckp, did_ckp, error, s;
551 	unsigned n, segleft, maxseg, sn, i, curseg;
552 	int writer_set = 0;
553 	int dirty;
554 	int redo;
555 
556 	fs = VFSTOUFS(mp)->um_lfs;
557 
558 	if (fs->lfs_ronly)
559 		return EROFS;
560 
561 	lfs_imtime(fs);
562 
563 	/*
564 	 * Allocate a segment structure and enough space to hold pointers to
565 	 * the maximum possible number of buffers which can be described in a
566 	 * single summary block.
567 	 */
568 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
569 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
570 	sp = fs->lfs_sp;
571 
572 	/*
573 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
574 	 * in which case we have to flush *all* buffers off of this vnode.
575 	 * We don't care about other nodes, but write any non-dirop nodes
576 	 * anyway in anticipation of another getnewvnode().
577 	 *
578 	 * If we're cleaning we only write cleaning and ifile blocks, and
579 	 * no dirops, since otherwise we'd risk corruption in a crash.
580 	 */
581 	if (sp->seg_flags & SEGM_CLEAN)
582 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
583 	else if (!(sp->seg_flags & SEGM_FORCE_CKP)) {
584 		lfs_writevnodes(fs, mp, sp, VN_REG);
585 		if (!fs->lfs_dirops || !fs->lfs_flushvp) {
586 			error = lfs_writer_enter(fs, "lfs writer");
587 			if (error) {
588 				printf("segwrite mysterious error\n");
589 				/* XXX why not segunlock? */
590 				pool_put(&fs->lfs_bpppool, sp->bpp);
591 				sp->bpp = NULL;
592 				pool_put(&fs->lfs_segpool, sp);
593 				sp = fs->lfs_sp = NULL;
594 				return (error);
595 			}
596 			writer_set = 1;
597 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
598 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
599 		}
600 	}
601 
602 	/*
603 	 * If we are doing a checkpoint, mark everything since the
604 	 * last checkpoint as no longer ACTIVE.
605 	 */
606 	if (do_ckp) {
607 		segleft = fs->lfs_nseg;
608 		curseg = 0;
609 		for (n = 0; n < fs->lfs_segtabsz; n++) {
610 			dirty = 0;
611 			if (bread(fs->lfs_ivnode,
612 			    fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &bp))
613 
614 				panic("lfs_segwrite: ifile read");
615 			segusep = (SEGUSE *)bp->b_data;
616 			maxseg = min(segleft, fs->lfs_sepb);
617 			for (i = 0; i < maxseg; i++) {
618 				sn = curseg + i;
619 				if (sn != fs->lfs_curseg &&
620 				    segusep->su_flags & SEGUSE_ACTIVE) {
621 					segusep->su_flags &= ~SEGUSE_ACTIVE;
622 					--fs->lfs_nactive;
623 					++dirty;
624 				}
625 				fs->lfs_suflags[fs->lfs_activesb][sn] =
626 					segusep->su_flags;
627 				if (fs->lfs_version > 1)
628 					++segusep;
629 				else
630 					segusep = (SEGUSE *)
631 						((SEGUSE_V1 *)segusep + 1);
632 			}
633 
634 			if (dirty)
635 				error = LFS_BWRITE_LOG(bp); /* Ifile */
636 			else
637 				brelse(bp);
638 			segleft -= fs->lfs_sepb;
639 			curseg += fs->lfs_sepb;
640 		}
641 	}
642 
643 	did_ckp = 0;
644 	if (do_ckp || fs->lfs_doifile) {
645 		do {
646 			vp = fs->lfs_ivnode;
647 
648 #ifdef DEBUG
649 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
650 #endif
651 			fs->lfs_flags &= ~LFS_IFDIRTY;
652 
653 			ip = VTOI(vp);
654 
655 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
656 				lfs_writefile(fs, sp, vp);
657 
658 			if (ip->i_flag & IN_ALLMOD)
659 				++did_ckp;
660 			redo = lfs_writeinode(fs, sp, ip);
661 			redo += lfs_writeseg(fs, sp);
662 			redo += (fs->lfs_flags & LFS_IFDIRTY);
663 		} while (redo && do_ckp);
664 
665 		/*
666 		 * Unless we are unmounting, the Ifile may continue to have
667 		 * dirty blocks even after a checkpoint, due to changes to
668 		 * inodes' atime.  If we're checkpointing, it's "impossible"
669 		 * for other parts of the Ifile to be dirty after the loop
670 		 * above, since we hold the segment lock.
671 		 */
672 		s = splbio();
673 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
674 			LFS_CLR_UINO(ip, IN_ALLMOD);
675 		}
676 #ifdef DIAGNOSTIC
677 		else if (do_ckp) {
678 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
679 				if (bp->b_lblkno < fs->lfs_cleansz +
680 				    fs->lfs_segtabsz &&
681 				    !(bp->b_flags & B_GATHERED)) {
682 					panic("dirty blocks");
683 				}
684 			}
685 		}
686 #endif
687 		splx(s);
688 	} else {
689 		(void) lfs_writeseg(fs, sp);
690 	}
691 
692 	/* Note Ifile no longer needs to be written */
693 	fs->lfs_doifile = 0;
694 	if (writer_set)
695 		lfs_writer_leave(fs);
696 
697 	/*
698 	 * If we didn't write the Ifile, we didn't really do anything.
699 	 * That means that (1) there is a checkpoint on disk and (2)
700 	 * nothing has changed since it was written.
701 	 *
702 	 * Take the flags off of the segment so that lfs_segunlock
703 	 * doesn't have to write the superblock either.
704 	 */
705 	if (do_ckp && !did_ckp) {
706 		sp->seg_flags &= ~SEGM_CKP;
707 	}
708 
709 	if (lfs_dostats) {
710 		++lfs_stats.nwrites;
711 		if (sp->seg_flags & SEGM_SYNC)
712 			++lfs_stats.nsync_writes;
713 		if (sp->seg_flags & SEGM_CKP)
714 			++lfs_stats.ncheckpoints;
715 	}
716 	lfs_segunlock(fs);
717 	return (0);
718 }
719 
720 /*
721  * Write the dirty blocks associated with a vnode.
722  */
723 void
724 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
725 {
726 	struct buf *bp;
727 	struct finfo *fip;
728 	struct inode *ip;
729 	IFILE *ifp;
730 	int i, frag;
731 
732 	ip = VTOI(vp);
733 
734 	if (sp->seg_bytes_left < fs->lfs_bsize ||
735 	    sp->sum_bytes_left < sizeof(struct finfo))
736 		(void) lfs_writeseg(fs, sp);
737 
738 	sp->sum_bytes_left -= FINFOSIZE;
739 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
740 
741 	if (vp->v_flag & VDIROP)
742 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
743 
744 	fip = sp->fip;
745 	fip->fi_nblocks = 0;
746 	fip->fi_ino = ip->i_number;
747 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
748 	fip->fi_version = ifp->if_version;
749 	brelse(bp);
750 
751 	if (sp->seg_flags & SEGM_CLEAN) {
752 		lfs_gather(fs, sp, vp, lfs_match_fake);
753 		/*
754 		 * For a file being flushed, we need to write *all* blocks.
755 		 * This means writing the cleaning blocks first, and then
756 		 * immediately following with any non-cleaning blocks.
757 		 * The same is true of the Ifile since checkpoints assume
758 		 * that all valid Ifile blocks are written.
759 		 */
760 		if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode) {
761 			lfs_gather(fs, sp, vp, lfs_match_data);
762 			/*
763 			 * Don't call VOP_PUTPAGES: if we're flushing,
764 			 * we've already done it, and the Ifile doesn't
765 			 * use the page cache.
766 			 */
767 		}
768 	} else {
769 		lfs_gather(fs, sp, vp, lfs_match_data);
770 		/*
771 		 * If we're flushing, we've already called VOP_PUTPAGES
772 		 * so don't do it again.  Otherwise, we want to write
773 		 * everything we've got.
774 		 */
775 		if (!IS_FLUSHING(fs, vp)) {
776 			simple_lock(&vp->v_interlock);
777 			VOP_PUTPAGES(vp, 0, 0,
778 				     PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
779 		}
780 	}
781 
782 	/*
783 	 * It may not be necessary to write the meta-data blocks at this point,
784 	 * as the roll-forward recovery code should be able to reconstruct the
785 	 * list.
786 	 *
787 	 * We have to write them anyway, though, under two conditions: (1) the
788 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
789 	 * checkpointing.
790 	 *
791 	 * BUT if we are cleaning, we might have indirect blocks that refer to
792 	 * new blocks not being written yet, in addition to fragments being
793 	 * moved out of a cleaned segment.  If that is the case, don't
794 	 * write the indirect blocks, or the finfo will have a small block
795 	 * in the middle of it!
796 	 * XXX in this case isn't the inode size wrong too?
797 	 */
798 	frag = 0;
799 	if (sp->seg_flags & SEGM_CLEAN) {
800 		for (i = 0; i < NDADDR; i++)
801 			if (ip->i_lfs_fragsize[i] > 0 &&
802 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
803 				++frag;
804 	}
805 #ifdef DIAGNOSTIC
806 	if (frag > 1)
807 		panic("lfs_writefile: more than one fragment!");
808 #endif
809 	if (IS_FLUSHING(fs, vp) ||
810 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
811 		lfs_gather(fs, sp, vp, lfs_match_indir);
812 		lfs_gather(fs, sp, vp, lfs_match_dindir);
813 		lfs_gather(fs, sp, vp, lfs_match_tindir);
814 	}
815 	fip = sp->fip;
816 	if (fip->fi_nblocks != 0) {
817 		sp->fip = (FINFO*)((caddr_t)fip + FINFOSIZE +
818 				   sizeof(int32_t) * (fip->fi_nblocks));
819 		sp->start_lbp = &sp->fip->fi_blocks[0];
820 	} else {
821 		sp->sum_bytes_left += FINFOSIZE;
822 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
823 	}
824 }
825 
826 int
827 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
828 {
829 	struct buf *bp, *ibp;
830 	struct ufs1_dinode *cdp;
831 	IFILE *ifp;
832 	SEGUSE *sup;
833 	daddr_t daddr;
834 	int32_t *daddrp;	/* XXX ondisk32 */
835 	ino_t ino;
836 	int error, i, ndx, fsb = 0;
837 	int redo_ifile = 0;
838 	struct timespec ts;
839 	int gotblk = 0;
840 
841 	if (!(ip->i_flag & IN_ALLMOD))
842 		return (0);
843 
844 	/* Allocate a new inode block if necessary. */
845 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
846 		/* Allocate a new segment if necessary. */
847 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
848 		    sp->sum_bytes_left < sizeof(int32_t))
849 			(void) lfs_writeseg(fs, sp);
850 
851 		/* Get next inode block. */
852 		daddr = fs->lfs_offset;
853 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
854 		sp->ibp = *sp->cbpp++ =
855 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
856 			       fs->lfs_ibsize, 0, 0);
857 		gotblk++;
858 
859 		/* Zero out inode numbers */
860 		for (i = 0; i < INOPB(fs); ++i)
861 			((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber = 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) * (1 - sp->ndupino)
1020 				   - 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, int (*match)(struct lfs *, struct buf *))
1101 {
1102 	struct buf *bp, *nbp;
1103 	int s, count = 0;
1104 
1105 	sp->vp = vp;
1106 	s = splbio();
1107 
1108 #ifndef LFS_NO_BACKBUF_HACK
1109 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1110 # define	BUF_OFFSET	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1111 # define	BACK_BUF(BP)	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1112 # define	BEG_OF_LIST	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1113 /* Find last buffer. */
1114 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1115 	    bp = LIST_NEXT(bp, b_vnbufs));
1116 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1117 		nbp = BACK_BUF(bp);
1118 #else /* LFS_NO_BACKBUF_HACK */
1119 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1120 		nbp = LIST_NEXT(bp, b_vnbufs);
1121 #endif /* LFS_NO_BACKBUF_HACK */
1122 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1123 #ifdef DEBUG_LFS
1124 			if (vp == fs->lfs_ivnode && (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1125 				printf("(%" PRId64 ":%lx)", bp->b_lblkno, bp->b_flags);
1126 #endif
1127 			continue;
1128 		}
1129 		if (vp->v_type == VBLK) {
1130 			/* For block devices, just write the blocks. */
1131 			/* XXX Do we really need to even do this? */
1132 #ifdef DEBUG_LFS
1133 			if (count == 0)
1134 				printf("BLK(");
1135 			printf(".");
1136 #endif
1137 			/* Get the block before bwrite, so we don't corrupt the free list */
1138 			bp->b_flags |= B_BUSY;
1139 			bremfree(bp);
1140 			bwrite(bp);
1141 		} else {
1142 #ifdef DIAGNOSTIC
1143 # ifdef LFS_USE_B_INVAL
1144 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1145 				printf("lfs_gather: lbn %" PRId64 " is "
1146 					"B_INVAL\n", bp->b_lblkno);
1147 				VOP_PRINT(bp->b_vp);
1148 			}
1149 # endif /* LFS_USE_B_INVAL */
1150 			if (!(bp->b_flags & B_DELWRI))
1151 				panic("lfs_gather: bp not B_DELWRI");
1152 			if (!(bp->b_flags & B_LOCKED)) {
1153 				printf("lfs_gather: lbn %" PRId64 " blk "
1154 					"%" PRId64 " not B_LOCKED\n",
1155 					bp->b_lblkno,
1156 					dbtofsb(fs, bp->b_blkno));
1157 				VOP_PRINT(bp->b_vp);
1158 				panic("lfs_gather: bp not B_LOCKED");
1159 			}
1160 #endif
1161 			if (lfs_gatherblock(sp, bp, &s)) {
1162 				goto loop;
1163 			}
1164 		}
1165 		count++;
1166 	}
1167 	splx(s);
1168 #ifdef DEBUG_LFS
1169 	if (vp->v_type == VBLK && count)
1170 		printf(")\n");
1171 #endif
1172 	lfs_updatemeta(sp);
1173 	sp->vp = NULL;
1174 	return count;
1175 }
1176 
1177 #if DEBUG
1178 # define DEBUG_OOFF(n) do {						\
1179 	if (ooff == 0) {						\
1180 		printf("lfs_updatemeta[%d]: warning: writing "		\
1181 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1182 			", was 0x0 (or %" PRId64 ")\n",			\
1183 			(n), ip->i_number, lbn, ndaddr, daddr);		\
1184 	}								\
1185 } while (0)
1186 #else
1187 # define DEBUG_OOFF(n)
1188 #endif
1189 
1190 /*
1191  * Change the given block's address to ndaddr, finding its previous
1192  * location using ufs_bmaparray().
1193  *
1194  * Account for this change in the segment table.
1195  */
1196 void
1197 lfs_update_single(struct lfs *fs, struct segment *sp, daddr_t lbn,
1198 		  int32_t ndaddr, int size)
1199 {
1200 	SEGUSE *sup;
1201 	struct buf *bp;
1202 	struct indir a[NIADDR + 2], *ap;
1203 	struct inode *ip;
1204 	struct vnode *vp;
1205 	daddr_t daddr, ooff;
1206 	int num, error;
1207 	int bb, osize, obb;
1208 
1209 	vp = sp->vp;
1210 	ip = VTOI(vp);
1211 
1212 	error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1213 	if (error)
1214 		panic("lfs_updatemeta: ufs_bmaparray returned %d", error);
1215 
1216 	KASSERT(daddr <= LFS_MAX_DADDR);
1217 	if (daddr > 0)
1218 		daddr = dbtofsb(fs, daddr);
1219 
1220 	bb = fragstofsb(fs, numfrags(fs, size));
1221 	switch (num) {
1222 	    case 0:
1223 		    ooff = ip->i_ffs1_db[lbn];
1224 		    DEBUG_OOFF(0);
1225 		    if (ooff == UNWRITTEN)
1226 			    ip->i_ffs1_blocks += bb;
1227 		    else {
1228 			    /* possible fragment truncation or extension */
1229 			    obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1230 			    ip->i_ffs1_blocks += (bb - obb);
1231 		    }
1232 		    ip->i_ffs1_db[lbn] = ndaddr;
1233 		    break;
1234 	    case 1:
1235 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1236 		    DEBUG_OOFF(1);
1237 		    if (ooff == UNWRITTEN)
1238 			    ip->i_ffs1_blocks += bb;
1239 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1240 		    break;
1241 	    default:
1242 		    ap = &a[num - 1];
1243 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1244 			    panic("lfs_updatemeta: bread bno %" PRId64,
1245 				  ap->in_lbn);
1246 
1247 		    /* XXX ondisk32 */
1248 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1249 		    DEBUG_OOFF(num);
1250 		    if (ooff == UNWRITTEN)
1251 			    ip->i_ffs1_blocks += bb;
1252 		    /* XXX ondisk32 */
1253 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1254 		    (void) VOP_BWRITE(bp);
1255 	}
1256 
1257 	/*
1258 	 * Though we'd rather it couldn't, this *can* happen right now
1259 	 * if cleaning blocks and regular blocks coexist.
1260 	 */
1261 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1262 
1263 	/*
1264 	 * Update segment usage information, based on old size
1265 	 * and location.
1266 	 */
1267 	if (daddr > 0) {
1268 		u_int32_t oldsn = dtosn(fs, daddr);
1269 #ifdef DIAGNOSTIC
1270 		int ndupino = (sp->seg_number == oldsn) ?
1271 			sp->ndupino : 0;
1272 #endif
1273 		KASSERT(oldsn >= 0 && oldsn < fs->lfs_nseg);
1274 		if (lbn >= 0 && lbn < NDADDR)
1275 			osize = ip->i_lfs_fragsize[lbn];
1276 		else
1277 			osize = fs->lfs_bsize;
1278 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1279 #ifdef DIAGNOSTIC
1280 		if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino
1281 		    < osize) {
1282 			printf("lfs_updatemeta: negative bytes "
1283 			       "(segment %" PRIu32 " short by %" PRId64
1284 			       ")\n", dtosn(fs, daddr),
1285 			       (int64_t)osize -
1286 			       (sizeof (struct ufs1_dinode) * sp->ndupino +
1287 				sup->su_nbytes));
1288 			printf("lfs_updatemeta: ino %d, lbn %" PRId64
1289 			       ", addr = 0x%" PRIx64 "\n",
1290 			       VTOI(sp->vp)->i_number, lbn, daddr);
1291 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1292 			panic("lfs_updatemeta: negative bytes");
1293 			sup->su_nbytes = osize -
1294 			    sizeof (struct ufs1_dinode) * sp->ndupino;
1295 		}
1296 #endif
1297 #ifdef DEBUG_SU_NBYTES
1298 		printf("seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1299 		       " db 0x%" PRIx64 "\n",
1300 		       dtosn(fs, daddr), osize,
1301 		       VTOI(sp->vp)->i_number, lbn, daddr);
1302 #endif
1303 		sup->su_nbytes -= osize;
1304 		if (!(bp->b_flags & B_GATHERED))
1305 			fs->lfs_flags |= LFS_IFDIRTY;
1306 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1307 	}
1308 	/*
1309 	 * Now that this block has a new address, and its old
1310 	 * segment no longer owns it, we can forget about its
1311 	 * old size.
1312 	 */
1313 	if (lbn >= 0 && lbn < NDADDR)
1314 		ip->i_lfs_fragsize[lbn] = size;
1315 }
1316 
1317 /*
1318  * Update the metadata that points to the blocks listed in the FINFO
1319  * array.
1320  */
1321 void
1322 lfs_updatemeta(struct segment *sp)
1323 {
1324 	struct buf *sbp;
1325 	struct lfs *fs;
1326 	struct vnode *vp;
1327 	daddr_t lbn;
1328 	int i, nblocks, num;
1329 	int bb;
1330 	int bytesleft, size;
1331 
1332 	vp = sp->vp;
1333 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1334 	KASSERT(nblocks >= 0);
1335 	if (vp == NULL || nblocks == 0)
1336 		return;
1337 
1338 	/*
1339 	 * This count may be high due to oversize blocks from lfs_gop_write.
1340 	 * Correct for this. (XXX we should be able to keep track of these.)
1341 	 */
1342 	fs = sp->fs;
1343 	for (i = 0; i < nblocks; i++) {
1344 		if (sp->start_bpp[i] == NULL) {
1345 			printf("nblocks = %d, not %d\n", i, nblocks);
1346 			nblocks = i;
1347 			break;
1348 		}
1349 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1350 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1351 		nblocks -= num - 1;
1352 	}
1353 
1354 	KASSERT(vp->v_type == VREG ||
1355 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1356 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1357 
1358 	/*
1359 	 * Sort the blocks.
1360 	 *
1361 	 * We have to sort even if the blocks come from the
1362 	 * cleaner, because there might be other pending blocks on the
1363 	 * same inode...and if we don't sort, and there are fragments
1364 	 * present, blocks may be written in the wrong place.
1365 	 */
1366 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1367 
1368 	/*
1369 	 * Record the length of the last block in case it's a fragment.
1370 	 * If there are indirect blocks present, they sort last.  An
1371 	 * indirect block will be lfs_bsize and its presence indicates
1372 	 * that you cannot have fragments.
1373 	 *
1374 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1375 	 * XXX a later indirect block.	This will continue to be
1376 	 * XXX true until lfs_markv is fixed to do everything with
1377 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1378 	 */
1379 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1380 		fs->lfs_bmask) + 1;
1381 
1382 	/*
1383 	 * Assign disk addresses, and update references to the logical
1384 	 * block and the segment usage information.
1385 	 */
1386 	for (i = nblocks; i--; ++sp->start_bpp) {
1387 		sbp = *sp->start_bpp;
1388 		lbn = *sp->start_lbp;
1389 		KASSERT(sbp->b_lblkno == lbn);
1390 
1391 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1392 
1393 		/*
1394 		 * If we write a frag in the wrong place, the cleaner won't
1395 		 * be able to correctly identify its size later, and the
1396 		 * segment will be uncleanable.	 (Even worse, it will assume
1397 		 * that the indirect block that actually ends the list
1398 		 * is of a smaller size!)
1399 		 */
1400 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1401 			panic("lfs_updatemeta: fragment is not last block");
1402 
1403 		/*
1404 		 * For each subblock in this possibly oversized block,
1405 		 * update its address on disk.
1406 		 */
1407 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1408 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1409 		     bytesleft -= fs->lfs_bsize) {
1410 			size = MIN(bytesleft, fs->lfs_bsize);
1411 			bb = fragstofsb(fs, numfrags(fs, size));
1412 			lbn = *sp->start_lbp++;
1413 			lfs_update_single(fs, sp, lbn, fs->lfs_offset, size);
1414 			fs->lfs_offset += bb;
1415 		}
1416 
1417 	}
1418 }
1419 
1420 /*
1421  * Start a new segment.
1422  */
1423 int
1424 lfs_initseg(struct lfs *fs)
1425 {
1426 	struct segment *sp;
1427 	SEGUSE *sup;
1428 	SEGSUM *ssp;
1429 	struct buf *bp, *sbp;
1430 	int repeat;
1431 
1432 	sp = fs->lfs_sp;
1433 
1434 	repeat = 0;
1435 
1436 	/* Advance to the next segment. */
1437 	if (!LFS_PARTIAL_FITS(fs)) {
1438 		/* lfs_avail eats the remaining space */
1439 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1440 						   fs->lfs_curseg);
1441 		/* Wake up any cleaning procs waiting on this file system. */
1442 		wakeup(&lfs_allclean_wakeup);
1443 		wakeup(&fs->lfs_nextseg);
1444 		lfs_newseg(fs);
1445 		repeat = 1;
1446 		fs->lfs_offset = fs->lfs_curseg;
1447 
1448 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1449 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1450 
1451 		/*
1452 		 * If the segment contains a superblock, update the offset
1453 		 * and summary address to skip over it.
1454 		 */
1455 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1456 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1457 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1458 			sp->seg_bytes_left -= LFS_SBPAD;
1459 		}
1460 		brelse(bp);
1461 		/* Segment zero could also contain the labelpad */
1462 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1463 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1464 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1465 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1466 		}
1467 	} else {
1468 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1469 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1470 				      (fs->lfs_offset - fs->lfs_curseg));
1471 	}
1472 	fs->lfs_lastpseg = fs->lfs_offset;
1473 
1474 	/* Record first address of this partial segment */
1475 	if (sp->seg_flags & SEGM_CLEAN) {
1476 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1477 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1478 			/* "1" is the artificial inc in lfs_seglock */
1479 			while (fs->lfs_iocount > 1) {
1480 				tsleep(&fs->lfs_iocount, PRIBIO + 1, "lfs_initseg", 0);
1481 			}
1482 			fs->lfs_cleanind = 0;
1483 		}
1484 	}
1485 
1486 	sp->fs = fs;
1487 	sp->ibp = NULL;
1488 	sp->idp = NULL;
1489 	sp->ninodes = 0;
1490 	sp->ndupino = 0;
1491 
1492 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1493 	sp->cbpp = sp->bpp;
1494 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1495 				     fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1496 	sp->segsum = (*sp->cbpp)->b_data;
1497 	memset(sp->segsum, 0, fs->lfs_sumsize);
1498 	sp->start_bpp = ++sp->cbpp;
1499 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1500 
1501 	/* Set point to SEGSUM, initialize it. */
1502 	ssp = sp->segsum;
1503 	ssp->ss_next = fs->lfs_nextseg;
1504 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1505 	ssp->ss_magic = SS_MAGIC;
1506 
1507 	/* Set pointer to first FINFO, initialize it. */
1508 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1509 	sp->fip->fi_nblocks = 0;
1510 	sp->start_lbp = &sp->fip->fi_blocks[0];
1511 	sp->fip->fi_lastlength = 0;
1512 
1513 	sp->seg_bytes_left -= fs->lfs_sumsize;
1514 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1515 
1516 	return (repeat);
1517 }
1518 
1519 /*
1520  * Return the next segment to write.
1521  */
1522 void
1523 lfs_newseg(struct lfs *fs)
1524 {
1525 	CLEANERINFO *cip;
1526 	SEGUSE *sup;
1527 	struct buf *bp;
1528 	int curseg, isdirty, sn;
1529 
1530 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1531 #ifdef DEBUG_SU_NBYTES
1532 	printf("lfs_newseg: seg %d := 0 in newseg\n",	/* XXXDEBUG */
1533 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1534 #endif
1535 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1536 	sup->su_nbytes = 0;
1537 	sup->su_nsums = 0;
1538 	sup->su_ninos = 0;
1539 	LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1540 
1541 	LFS_CLEANERINFO(cip, fs, bp);
1542 	--cip->clean;
1543 	++cip->dirty;
1544 	fs->lfs_nclean = cip->clean;
1545 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1546 
1547 	fs->lfs_lastseg = fs->lfs_curseg;
1548 	fs->lfs_curseg = fs->lfs_nextseg;
1549 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1550 		sn = (sn + 1) % fs->lfs_nseg;
1551 		if (sn == curseg)
1552 			panic("lfs_nextseg: no clean segments");
1553 		LFS_SEGENTRY(sup, fs, sn, bp);
1554 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1555 		/* Check SEGUSE_EMPTY as we go along */
1556 		if (isdirty && sup->su_nbytes == 0 && !(sup->su_flags & SEGUSE_EMPTY))
1557 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1558 		else
1559 			brelse(bp);
1560 
1561 		if (!isdirty)
1562 			break;
1563 	}
1564 
1565 	++fs->lfs_nactive;
1566 	fs->lfs_nextseg = sntod(fs, sn);
1567 	if (lfs_dostats) {
1568 		++lfs_stats.segsused;
1569 	}
1570 }
1571 
1572 #define BQUEUES 4 /* XXX */
1573 #define BQ_EMPTY 3 /* XXX */
1574 extern TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
1575 extern struct simplelock bqueue_slock;
1576 
1577 #define	BUFHASH(dvp, lbn)	\
1578 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
1579 extern LIST_HEAD(bufhashhdr, buf) invalhash;
1580 /*
1581  * Insq/Remq for the buffer hash lists.
1582  */
1583 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
1584 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
1585 
1586 static struct buf *
1587 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1588 {
1589 	struct lfs_cluster *cl;
1590 	struct buf **bpp, *bp;
1591 	int s;
1592 
1593 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1594 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1595 	memset(cl, 0, sizeof(*cl));
1596 	cl->fs = fs;
1597 	cl->bpp = bpp;
1598 	cl->bufcount = 0;
1599 	cl->bufsize = 0;
1600 
1601 	/* If this segment is being written synchronously, note that */
1602 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1603 		cl->flags |= LFS_CL_SYNC;
1604 		cl->seg = fs->lfs_sp;
1605 		++cl->seg->seg_iocount;
1606 		/* printf("+ %x => %d\n", cl->seg, cl->seg->seg_iocount); */
1607 	}
1608 
1609 	/* Get an empty buffer header, or maybe one with something on it */
1610 	s = splbio();
1611 	simple_lock(&bqueue_slock);
1612 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_EMPTY])) != NULL) {
1613 		simple_lock(&bp->b_interlock);
1614 		bremfree(bp);
1615 		/* clear out various other fields */
1616 		bp->b_flags = B_BUSY;
1617 		bp->b_dev = NODEV;
1618 		bp->b_blkno = bp->b_lblkno = 0;
1619 		bp->b_error = 0;
1620 		bp->b_resid = 0;
1621 		bp->b_bcount = 0;
1622 
1623 		/* nuke any credentials we were holding */
1624 		/* XXXXXX */
1625 
1626 		bremhash(bp);
1627 
1628 		/* disassociate us from our vnode, if we had one... */
1629 		if (bp->b_vp)
1630 			brelvp(bp);
1631 	}
1632 	while (!bp)
1633 		bp = getnewbuf(0, 0);
1634 	bgetvp(vp, bp);
1635 	binshash(bp,&invalhash);
1636 	simple_unlock(&bp->b_interlock);
1637 	simple_unlock(&bqueue_slock);
1638 	splx(s);
1639 	bp->b_bcount = 0;
1640 	bp->b_blkno = bp->b_lblkno = addr;
1641 
1642 	bp->b_flags |= B_CALL;
1643 	bp->b_iodone = lfs_cluster_callback;
1644 	cl->saveaddr = bp->b_saveaddr; /* XXX is this ever used? */
1645 	bp->b_saveaddr = (caddr_t)cl;
1646 
1647 	return bp;
1648 }
1649 
1650 int
1651 lfs_writeseg(struct lfs *fs, struct segment *sp)
1652 {
1653 	struct buf **bpp, *bp, *cbp, *newbp;
1654 	SEGUSE *sup;
1655 	SEGSUM *ssp;
1656 	dev_t i_dev;
1657 	char *datap, *dp;
1658 	int i, s;
1659 	int do_again, nblocks, byteoffset;
1660 	size_t el_size;
1661 	struct lfs_cluster *cl;
1662 	int (*strategy)(void *);
1663 	struct vop_strategy_args vop_strategy_a;
1664 	u_short ninos;
1665 	struct vnode *devvp;
1666 	char *p;
1667 	struct vnode *vp;
1668 	int32_t *daddrp;	/* XXX ondisk32 */
1669 	int changed;
1670 #if defined(DEBUG) && defined(LFS_PROPELLER)
1671 	static int propeller;
1672 	char propstring[4] = "-\\|/";
1673 
1674 	printf("%c\b",propstring[propeller++]);
1675 	if (propeller == 4)
1676 		propeller = 0;
1677 #endif
1678 
1679 	/*
1680 	 * If there are no buffers other than the segment summary to write
1681 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1682 	 * even if there aren't any buffers, you need to write the superblock.
1683 	 */
1684 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1685 		return (0);
1686 
1687 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1688 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1689 
1690 	/* Update the segment usage information. */
1691 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1692 
1693 	/* Loop through all blocks, except the segment summary. */
1694 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1695 		if ((*bpp)->b_vp != devvp) {
1696 			sup->su_nbytes += (*bpp)->b_bcount;
1697 #ifdef DEBUG_SU_NBYTES
1698 		printf("seg %" PRIu32 " += %ld for ino %d lbn %" PRId64
1699 		    " db 0x%" PRIx64 "\n", sp->seg_number, (*bpp)->b_bcount,
1700 		    VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
1701 		    (*bpp)->b_blkno);
1702 #endif
1703 		}
1704 	}
1705 
1706 	ssp = (SEGSUM *)sp->segsum;
1707 
1708 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1709 #ifdef DEBUG_SU_NBYTES
1710 	printf("seg %d += %d for %d inodes\n",	 /* XXXDEBUG */
1711 	       sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode),
1712 	       ssp->ss_ninos);
1713 #endif
1714 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode);
1715 	/* sup->su_nbytes += fs->lfs_sumsize; */
1716 	if (fs->lfs_version == 1)
1717 		sup->su_olastmod = time.tv_sec;
1718 	else
1719 		sup->su_lastmod = time.tv_sec;
1720 	sup->su_ninos += ninos;
1721 	++sup->su_nsums;
1722 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1723 							 fs->lfs_ibsize));
1724 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1725 
1726 	do_again = !(bp->b_flags & B_GATHERED);
1727 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
1728 
1729 	/*
1730 	 * Mark blocks B_BUSY, to prevent then from being changed between
1731 	 * the checksum computation and the actual write.
1732 	 *
1733 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1734 	 * there are any, replace them with copies that have UNASSIGNED
1735 	 * instead.
1736 	 */
1737 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1738 		++bpp;
1739 		bp = *bpp;
1740 		if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */
1741 			bp->b_flags |= B_BUSY;
1742 			continue;
1743 		}
1744 	    again:
1745 		s = splbio();
1746 		if (bp->b_flags & B_BUSY) {
1747 #ifdef DEBUG
1748 			printf("lfs_writeseg: avoiding potential data summary "
1749 			       "corruption for ino %d, lbn %" PRId64 "\n",
1750 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1751 #endif
1752 			bp->b_flags |= B_WANTED;
1753 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1754 			splx(s);
1755 			goto again;
1756 		}
1757 		bp->b_flags |= B_BUSY;
1758 		splx(s);
1759 		/*
1760 		 * Check and replace indirect block UNWRITTEN bogosity.
1761 		 * XXX See comment in lfs_writefile.
1762 		 */
1763 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1764 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
1765 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1766 #ifdef DEBUG_LFS
1767 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1768 			       VTOI(bp->b_vp)->i_number,
1769 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1770 			       VTOI(bp->b_vp)->i_ffs1_blocks);
1771 #endif
1772 			/* Make a copy we'll make changes to */
1773 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1774 					   bp->b_bcount, LFS_NB_IBLOCK);
1775 			newbp->b_blkno = bp->b_blkno;
1776 			memcpy(newbp->b_data, bp->b_data,
1777 			       newbp->b_bcount);
1778 
1779 			changed = 0;
1780 			/* XXX ondisk32 */
1781 			for (daddrp = (int32_t *)(newbp->b_data);
1782 			     daddrp < (int32_t *)(newbp->b_data +
1783 						  newbp->b_bcount); daddrp++) {
1784 				if (*daddrp == UNWRITTEN) {
1785 #ifdef DEBUG_LFS
1786 					off_t doff;
1787 					int32_t ioff;
1788 
1789 					ioff = daddrp - (int32_t *)(newbp->b_data);
1790 					doff = (-bp->b_lblkno + ioff) * fs->lfs_bsize;
1791 					printf("ino %d lbn %" PRId64 " entry %d off %" PRIx64 "\n",
1792 					       VTOI(bp->b_vp)->i_number,
1793 					       bp->b_lblkno, ioff, doff);
1794 					if (bp->b_vp->v_type == VREG) {
1795 						/*
1796 						 * What is up with this page?
1797 						 */
1798 						struct vm_page *pg;
1799 						for (; doff / fs->lfs_bsize == (-bp->b_lblkno + ioff); doff += PAGE_SIZE) {
1800 							pg = uvm_pagelookup(&bp->b_vp->v_uobj, doff);
1801 							if (pg == NULL)
1802 								printf("  page at %" PRIx64 " is NULL\n", doff);
1803 							else
1804 								printf("  page at %" PRIx64 " flags 0x%x pqflags 0x%x\n", doff, pg->flags, pg->pqflags);
1805 						}
1806 					}
1807 #endif /* DEBUG_LFS */
1808 					++changed;
1809 					*daddrp = 0;
1810 				}
1811 			}
1812 			/*
1813 			 * Get rid of the old buffer.  Don't mark it clean,
1814 			 * though, if it still has dirty data on it.
1815 			 */
1816 			if (changed) {
1817 #ifdef DEBUG_LFS
1818 				printf("lfs_writeseg: replacing UNWRITTEN(%d):"
1819 					" bp = %p newbp = %p\n", changed, bp,
1820 					newbp);
1821 #endif
1822 				*bpp = newbp;
1823 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1824 				if (bp->b_flags & B_CALL) {
1825 					printf("lfs_writeseg: indir bp should not be B_CALL\n");
1826 					s = splbio();
1827 					biodone(bp);
1828 					splx(s);
1829 					bp = NULL;
1830 				} else {
1831 					/* Still on free list, leave it there */
1832 					s = splbio();
1833 					bp->b_flags &= ~B_BUSY;
1834 					if (bp->b_flags & B_WANTED)
1835 						wakeup(bp);
1836 					splx(s);
1837 					/*
1838 					 * We have to re-decrement lfs_avail
1839 					 * since this block is going to come
1840 					 * back around to us in the next
1841 					 * segment.
1842 					 */
1843 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1844 				}
1845 			} else {
1846 				lfs_freebuf(fs, newbp);
1847 			}
1848 		}
1849 	}
1850 	/*
1851 	 * Compute checksum across data and then across summary; the first
1852 	 * block (the summary block) is skipped.  Set the create time here
1853 	 * so that it's guaranteed to be later than the inode mod times.
1854 	 *
1855 	 * XXX
1856 	 * Fix this to do it inline, instead of malloc/copy.
1857 	 */
1858 	datap = dp = pool_get(&fs->lfs_bpppool, PR_WAITOK);
1859 	if (fs->lfs_version == 1)
1860 		el_size = sizeof(u_long);
1861 	else
1862 		el_size = sizeof(u_int32_t);
1863 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
1864 		++bpp;
1865 		/* Loop through gop_write cluster blocks */
1866 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
1867 		     byteoffset += fs->lfs_bsize) {
1868 #ifdef LFS_USE_B_INVAL
1869 			if (((*bpp)->b_flags & (B_CALL | B_INVAL)) ==
1870 			    (B_CALL | B_INVAL)) {
1871 				if (copyin((caddr_t)(*bpp)->b_saveaddr +
1872 					   byteoffset, dp, el_size)) {
1873 					panic("lfs_writeseg: copyin failed [1]: "
1874 						"ino %d blk %" PRId64,
1875 						VTOI((*bpp)->b_vp)->i_number,
1876 						(*bpp)->b_lblkno);
1877 				}
1878 			} else
1879 #endif /* LFS_USE_B_INVAL */
1880 			{
1881 				memcpy(dp, (*bpp)->b_data + byteoffset,
1882 				       el_size);
1883 			}
1884 			dp += el_size;
1885 		}
1886 	}
1887 	if (fs->lfs_version == 1)
1888 		ssp->ss_ocreate = time.tv_sec;
1889 	else {
1890 		ssp->ss_create = time.tv_sec;
1891 		ssp->ss_serial = ++fs->lfs_serial;
1892 		ssp->ss_ident  = fs->lfs_ident;
1893 	}
1894 	ssp->ss_datasum = cksum(datap, dp - datap);
1895 	ssp->ss_sumsum =
1896 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1897 	pool_put(&fs->lfs_bpppool, datap);
1898 	datap = dp = NULL;
1899 #ifdef DIAGNOSTIC
1900 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1901 		panic("lfs_writeseg: No diskspace for summary");
1902 #endif
1903 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1904 			  btofsb(fs, fs->lfs_sumsize));
1905 
1906 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1907 
1908 	/*
1909 	 * When we simply write the blocks we lose a rotation for every block
1910 	 * written.  To avoid this problem, we cluster the buffers into a
1911 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
1912 	 * devices can handle, use that for the size of the chunks.
1913 	 *
1914 	 * Blocks that are already clusters (from GOP_WRITE), however, we
1915 	 * don't bother to copy into other clusters.
1916 	 */
1917 
1918 #define CHUNKSIZE MAXPHYS
1919 
1920 	if (devvp == NULL)
1921 		panic("devvp is NULL");
1922 	for (bpp = sp->bpp, i = nblocks; i;) {
1923 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1924 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1925 
1926 		cbp->b_dev = i_dev;
1927 		cbp->b_flags |= B_ASYNC | B_BUSY;
1928 		cbp->b_bcount = 0;
1929 
1930 		cl->olddata = cbp->b_data;
1931 #if defined(DEBUG) && defined(DIAGNOSTIC)
1932 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
1933 		    / sizeof(int32_t)) {
1934 			panic("lfs_writeseg: real bpp overwrite");
1935 		}
1936 		if (bpp - sp->bpp > fs->lfs_ssize / fs->lfs_fsize) {
1937 			panic("lfs_writeseg: theoretical bpp overwrite");
1938 		}
1939 #endif
1940 
1941 		/*
1942 		 * Construct the cluster.
1943 		 */
1944 		++fs->lfs_iocount;
1945 		while (i && cbp->b_bcount < CHUNKSIZE) {
1946 			bp = *bpp;
1947 
1948 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1949 				break;
1950 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
1951 				break;
1952 
1953 			/* Clusters from GOP_WRITE are expedited */
1954 			if (bp->b_bcount > fs->lfs_bsize) {
1955 				if (cbp->b_bcount > 0)
1956 					/* Put in its own buffer */
1957 					break;
1958 				else {
1959 					cbp->b_data = bp->b_data;
1960 				}
1961 			} else if (cbp->b_bcount == 0) {
1962 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
1963 							     LFS_NB_CLUSTER);
1964 				cl->flags |= LFS_CL_MALLOC;
1965 			}
1966 #ifdef DIAGNOSTIC
1967 			if (dtosn(fs, dbtofsb(fs, bp->b_blkno +
1968 					      btodb(bp->b_bcount - 1))) !=
1969 			    sp->seg_number) {
1970 				printf("blk size %ld daddr %" PRIx64 " not in seg %d\n",
1971 					bp->b_bcount, bp->b_blkno,
1972 					sp->seg_number);
1973 				panic("segment overwrite");
1974 			}
1975 #endif
1976 
1977 #ifdef LFS_USE_B_INVAL
1978 			/*
1979 			 * Fake buffers from the cleaner are marked as B_INVAL.
1980 			 * We need to copy the data from user space rather than
1981 			 * from the buffer indicated.
1982 			 * XXX == what do I do on an error?
1983 			 */
1984 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1985 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1986 					panic("lfs_writeseg: copyin failed [2]");
1987 			} else
1988 #endif /* LFS_USE_B_INVAL */
1989 			if (cl->flags & LFS_CL_MALLOC) {
1990 				bcopy(bp->b_data, p, bp->b_bcount);
1991 			}
1992 
1993 			p += bp->b_bcount;
1994 			cbp->b_bcount += bp->b_bcount;
1995 			cl->bufsize += bp->b_bcount;
1996 
1997 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
1998 			cl->bpp[cl->bufcount++] = bp;
1999 			vp = bp->b_vp;
2000 			s = splbio();
2001 			V_INCR_NUMOUTPUT(vp);
2002 			splx(s);
2003 
2004 			bpp++;
2005 			i--;
2006 		}
2007 		s = splbio();
2008 		V_INCR_NUMOUTPUT(devvp);
2009 		splx(s);
2010 		vop_strategy_a.a_desc = VDESC(vop_strategy);
2011 		vop_strategy_a.a_bp = cbp;
2012 		(strategy)(&vop_strategy_a);
2013 		curproc->p_stats->p_ru.ru_oublock++;
2014 	}
2015 
2016 	if (lfs_dostats) {
2017 		++lfs_stats.psegwrites;
2018 		lfs_stats.blocktot += nblocks - 1;
2019 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2020 			++lfs_stats.psyncwrites;
2021 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2022 			++lfs_stats.pcleanwrites;
2023 			lfs_stats.cleanblocks += nblocks - 1;
2024 		}
2025 	}
2026 	return (lfs_initseg(fs) || do_again);
2027 }
2028 
2029 void
2030 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2031 {
2032 	struct buf *bp;
2033 	dev_t i_dev;
2034 	int (*strategy)(void *);
2035 	int s;
2036 	struct vop_strategy_args vop_strategy_a;
2037 
2038 	/*
2039 	 * If we can write one superblock while another is in
2040 	 * progress, we risk not having a complete checkpoint if we crash.
2041 	 * So, block here if a superblock write is in progress.
2042 	 */
2043 	s = splbio();
2044 	while (fs->lfs_sbactive) {
2045 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
2046 	}
2047 	fs->lfs_sbactive = daddr;
2048 	splx(s);
2049 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
2050 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
2051 
2052 	/* Set timestamp of this version of the superblock */
2053 	if (fs->lfs_version == 1)
2054 		fs->lfs_otstamp = time.tv_sec;
2055 	fs->lfs_tstamp = time.tv_sec;
2056 
2057 	/* Checksum the superblock and copy it into a buffer. */
2058 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2059 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2060 	memset(bp->b_data + sizeof(struct dlfs), 0, LFS_SBPAD - sizeof(struct dlfs));
2061 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2062 
2063 	bp->b_dev = i_dev;
2064 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2065 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2066 	bp->b_iodone = lfs_supercallback;
2067 	/* XXX KS - same nasty hack as above */
2068 	bp->b_saveaddr = (caddr_t)fs;
2069 
2070 	vop_strategy_a.a_desc = VDESC(vop_strategy);
2071 	vop_strategy_a.a_bp = bp;
2072 	curproc->p_stats->p_ru.ru_oublock++;
2073 	s = splbio();
2074 	V_INCR_NUMOUTPUT(bp->b_vp);
2075 	splx(s);
2076 	++fs->lfs_iocount;
2077 	(strategy)(&vop_strategy_a);
2078 }
2079 
2080 /*
2081  * Logical block number match routines used when traversing the dirty block
2082  * chain.
2083  */
2084 int
2085 lfs_match_fake(struct lfs *fs, struct buf *bp)
2086 {
2087 	return LFS_IS_MALLOC_BUF(bp);
2088 }
2089 
2090 #if 0
2091 int
2092 lfs_match_real(struct lfs *fs, struct buf *bp)
2093 {
2094 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2095 }
2096 #endif
2097 
2098 int
2099 lfs_match_data(struct lfs *fs, struct buf *bp)
2100 {
2101 	return (bp->b_lblkno >= 0);
2102 }
2103 
2104 int
2105 lfs_match_indir(struct lfs *fs, struct buf *bp)
2106 {
2107 	daddr_t lbn;
2108 
2109 	lbn = bp->b_lblkno;
2110 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2111 }
2112 
2113 int
2114 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2115 {
2116 	daddr_t lbn;
2117 
2118 	lbn = bp->b_lblkno;
2119 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2120 }
2121 
2122 int
2123 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2124 {
2125 	daddr_t lbn;
2126 
2127 	lbn = bp->b_lblkno;
2128 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2129 }
2130 
2131 /*
2132  * XXX - The only buffers that are going to hit these functions are the
2133  * segment write blocks, or the segment summaries, or the superblocks.
2134  *
2135  * All of the above are created by lfs_newbuf, and so do not need to be
2136  * released via brelse.
2137  */
2138 void
2139 lfs_callback(struct buf *bp)
2140 {
2141 	struct lfs *fs;
2142 
2143 	fs = (struct lfs *)bp->b_saveaddr;
2144 	lfs_freebuf(fs, bp);
2145 }
2146 
2147 static void
2148 lfs_super_aiodone(struct buf *bp)
2149 {
2150 	struct lfs *fs;
2151 
2152 	fs = (struct lfs *)bp->b_saveaddr;
2153 	fs->lfs_sbactive = 0;
2154 	wakeup(&fs->lfs_sbactive);
2155 	if (--fs->lfs_iocount <= 1)
2156 		wakeup(&fs->lfs_iocount);
2157 	lfs_freebuf(fs, bp);
2158 }
2159 
2160 static void
2161 lfs_cluster_aiodone(struct buf *bp)
2162 {
2163 	struct lfs_cluster *cl;
2164 	struct lfs *fs;
2165 	struct buf *tbp, *fbp;
2166 	struct vnode *vp, *devvp;
2167 	struct inode *ip;
2168 	int s, error=0;
2169 	char *cp;
2170 	extern int locked_queue_count;
2171 	extern long locked_queue_bytes;
2172 
2173 	if (bp->b_flags & B_ERROR)
2174 		error = bp->b_error;
2175 
2176 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2177 	fs = cl->fs;
2178 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2179 	bp->b_saveaddr = cl->saveaddr;
2180 
2181 	cp = (char *)bp->b_data + cl->bufsize;
2182 	/* Put the pages back, and release the buffer */
2183 	while (cl->bufcount--) {
2184 		tbp = cl->bpp[cl->bufcount];
2185 		if (error) {
2186 			tbp->b_flags |= B_ERROR;
2187 			tbp->b_error = error;
2188 		}
2189 
2190 		/*
2191 		 * We're done with tbp.	 If it has not been re-dirtied since
2192 		 * the cluster was written, free it.  Otherwise, keep it on
2193 		 * the locked list to be written again.
2194 		 */
2195 		vp = tbp->b_vp;
2196 
2197 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2198 			LFS_UNLOCK_BUF(tbp);
2199 
2200 		tbp->b_flags &= ~B_GATHERED;
2201 
2202 		LFS_BCLEAN_LOG(fs, tbp);
2203 
2204 		if (!(tbp->b_flags & B_CALL)) {
2205 			bremfree(tbp);
2206 			s = splbio();
2207 			if (vp)
2208 				reassignbuf(tbp, vp);
2209 			splx(s);
2210 			tbp->b_flags |= B_ASYNC; /* for biodone */
2211 		}
2212 #ifdef DIAGNOSTIC
2213 		if (tbp->b_flags & B_DONE) {
2214 			printf("blk %d biodone already (flags %lx)\n",
2215 				cl->bufcount, (long)tbp->b_flags);
2216 		}
2217 #endif
2218 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2219 			if ((tbp->b_flags & B_CALL) && !LFS_IS_MALLOC_BUF(tbp)) {
2220 				/* printf("flags 0x%lx\n", tbp->b_flags); */
2221 				/*
2222 				 * A buffer from the page daemon.
2223 				 * We use the same iodone as it does,
2224 				 * so we must manually disassociate its
2225 				 * buffers from the vp.
2226 				 */
2227 				if (tbp->b_vp) {
2228 					/* This is just silly */
2229 					s = splbio();
2230 					brelvp(tbp);
2231 					tbp->b_vp = vp;
2232 					splx(s);
2233 				}
2234 				/* Put it back the way it was */
2235 				tbp->b_flags |= B_ASYNC;
2236 				/* Master buffers have B_AGE */
2237 				if (tbp->b_private == tbp)
2238 					tbp->b_flags |= B_AGE;
2239 			}
2240 			s = splbio();
2241 			biodone(tbp);
2242 
2243 			/*
2244 			 * If this is the last block for this vnode, but
2245 			 * there are other blocks on its dirty list,
2246 			 * set IN_MODIFIED/IN_CLEANING depending on what
2247 			 * sort of block.  Only do this for our mount point,
2248 			 * not for, e.g., inode blocks that are attached to
2249 			 * the devvp.
2250 			 * XXX KS - Shouldn't we set *both* if both types
2251 			 * of blocks are present (traverse the dirty list?)
2252 			 */
2253 			simple_lock(&global_v_numoutput_slock);
2254 			if (vp != devvp && vp->v_numoutput == 0 &&
2255 			    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2256 				ip = VTOI(vp);
2257 #ifdef DEBUG_LFS
2258 				printf("lfs_cluster_aiodone: marking ino %d\n",
2259 				       ip->i_number);
2260 #endif
2261 				if (LFS_IS_MALLOC_BUF(fbp))
2262 					LFS_SET_UINO(ip, IN_CLEANING);
2263 				else
2264 					LFS_SET_UINO(ip, IN_MODIFIED);
2265 			}
2266 			simple_unlock(&global_v_numoutput_slock);
2267 			splx(s);
2268 			wakeup(vp);
2269 		}
2270 	}
2271 
2272 	/* Fix up the cluster buffer, and release it */
2273 	if (cl->flags & LFS_CL_MALLOC)
2274 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2275 	bp->b_data = cl->olddata;
2276 	bp->b_bcount = 0;
2277 	bp->b_iodone = NULL;
2278 	bp->b_flags &= ~B_DELWRI;
2279 	bp->b_flags |= B_DONE;
2280 	s = splbio();
2281 	reassignbuf(bp, bp->b_vp);
2282 	splx(s);
2283 	brelse(bp);
2284 
2285 	/* Note i/o done */
2286 	if (cl->flags & LFS_CL_SYNC) {
2287 		if (--cl->seg->seg_iocount == 0)
2288 			wakeup(&cl->seg->seg_iocount);
2289 		/* printf("- %x => %d\n", cl->seg, cl->seg->seg_iocount); */
2290 	}
2291 #ifdef DIAGNOSTIC
2292 	if (fs->lfs_iocount == 0)
2293 		panic("lfs_cluster_aiodone: zero iocount");
2294 #endif
2295 	if (--fs->lfs_iocount <= 1)
2296 		wakeup(&fs->lfs_iocount);
2297 
2298 	pool_put(&fs->lfs_bpppool, cl->bpp);
2299 	cl->bpp = NULL;
2300 	pool_put(&fs->lfs_clpool, cl);
2301 }
2302 
2303 static void
2304 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2305 {
2306 	/* reset b_iodone for when this is a single-buf i/o. */
2307 	bp->b_iodone = aiodone;
2308 
2309 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
2310 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2311 	wakeup(&uvm.aiodoned);
2312 	simple_unlock(&uvm.aiodoned_lock);
2313 }
2314 
2315 static void
2316 lfs_cluster_callback(struct buf *bp)
2317 {
2318 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2319 }
2320 
2321 void
2322 lfs_supercallback(struct buf *bp)
2323 {
2324 	lfs_generic_callback(bp, lfs_super_aiodone);
2325 }
2326 
2327 /*
2328  * Shellsort (diminishing increment sort) from Data Structures and
2329  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2330  * see also Knuth Vol. 3, page 84.  The increments are selected from
2331  * formula (8), page 95.  Roughly O(N^3/2).
2332  */
2333 /*
2334  * This is our own private copy of shellsort because we want to sort
2335  * two parallel arrays (the array of buffer pointers and the array of
2336  * logical block numbers) simultaneously.  Note that we cast the array
2337  * of logical block numbers to a unsigned in this routine so that the
2338  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2339  */
2340 
2341 void
2342 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2343 {
2344 	static int __rsshell_increments[] = { 4, 1, 0 };
2345 	int incr, *incrp, t1, t2;
2346 	struct buf *bp_temp;
2347 
2348 #ifdef DEBUG
2349 	incr = 0;
2350 	for (t1 = 0; t1 < nmemb; t1++) {
2351 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2352 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2353 				/* dump before panic */
2354 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2355 				    nmemb, size);
2356 				incr = 0;
2357 				for (t1 = 0; t1 < nmemb; t1++) {
2358 					const struct buf *bp = bp_array[t1];
2359 
2360 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2361 					    PRIu64 "\n", t1,
2362 					    (uint64_t)bp->b_bcount,
2363 					    (uint64_t)bp->b_lblkno);
2364 					printf("lbns:");
2365 					for (t2 = 0; t2 * size < bp->b_bcount;
2366 					    t2++) {
2367 						printf(" %" PRId32,
2368 						    lb_array[incr++]);
2369 					}
2370 					printf("\n");
2371 				}
2372 				panic("lfs_shellsort: inconsistent input");
2373 			}
2374 		}
2375 	}
2376 #endif
2377 
2378 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2379 		for (t1 = incr; t1 < nmemb; ++t1)
2380 			for (t2 = t1 - incr; t2 >= 0;)
2381 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2382 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2383 					bp_temp = bp_array[t2];
2384 					bp_array[t2] = bp_array[t2 + incr];
2385 					bp_array[t2 + incr] = bp_temp;
2386 					t2 -= incr;
2387 				} else
2388 					break;
2389 
2390 	/* Reform the list of logical blocks */
2391 	incr = 0;
2392 	for (t1 = 0; t1 < nmemb; t1++) {
2393 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2394 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2395 		}
2396 	}
2397 }
2398 
2399 /*
2400  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2401  */
2402 int
2403 lfs_vref(struct vnode *vp)
2404 {
2405 	/*
2406 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2407 	 * being able to flush all of the pages from this vnode, which
2408 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2409 	 */
2410 	if (vp->v_flag & VXLOCK) {
2411 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2412 			return 0;
2413 		}
2414 		return (1);
2415 	}
2416 	return (vget(vp, 0));
2417 }
2418 
2419 /*
2420  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2421  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2422  */
2423 void
2424 lfs_vunref(struct vnode *vp)
2425 {
2426 	/*
2427 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2428 	 */
2429 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2430 		return;
2431 	}
2432 
2433 	simple_lock(&vp->v_interlock);
2434 #ifdef DIAGNOSTIC
2435 	if (vp->v_usecount <= 0) {
2436 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2437 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2438 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2439 		panic("lfs_vunref: v_usecount<0");
2440 	}
2441 #endif
2442 	vp->v_usecount--;
2443 	if (vp->v_usecount > 0) {
2444 		simple_unlock(&vp->v_interlock);
2445 		return;
2446 	}
2447 	/*
2448 	 * insert at tail of LRU list
2449 	 */
2450 	simple_lock(&vnode_free_list_slock);
2451 	if (vp->v_holdcnt > 0)
2452 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2453 	else
2454 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2455 	simple_unlock(&vnode_free_list_slock);
2456 	simple_unlock(&vp->v_interlock);
2457 }
2458 
2459 /*
2460  * We use this when we have vnodes that were loaded in solely for cleaning.
2461  * There is no reason to believe that these vnodes will be referenced again
2462  * soon, since the cleaning process is unrelated to normal filesystem
2463  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2464  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2465  * cleaning at the head of the list, instead.
2466  */
2467 void
2468 lfs_vunref_head(struct vnode *vp)
2469 {
2470 	simple_lock(&vp->v_interlock);
2471 #ifdef DIAGNOSTIC
2472 	if (vp->v_usecount == 0) {
2473 		panic("lfs_vunref: v_usecount<0");
2474 	}
2475 #endif
2476 	vp->v_usecount--;
2477 	if (vp->v_usecount > 0) {
2478 		simple_unlock(&vp->v_interlock);
2479 		return;
2480 	}
2481 	/*
2482 	 * insert at head of LRU list
2483 	 */
2484 	simple_lock(&vnode_free_list_slock);
2485 	if (vp->v_holdcnt > 0)
2486 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2487 	else
2488 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2489 	simple_unlock(&vnode_free_list_slock);
2490 	simple_unlock(&vp->v_interlock);
2491 }
2492 
2493