xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision d710132b4b8ce7f7cccaaf660cb16aa16b4077a0)
1 /*	$NetBSD: lfs_segment.c,v 1.121 2003/05/18 12:59:05 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 1.121 2003/05/18 12:59:05 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 			while (fs->lfs_dirops)
587 				if ((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
588 						"lfs writer", 0)))
589 				{
590 					printf("segwrite mysterious error\n");
591 					/* XXX why not segunlock? */
592 					pool_put(&fs->lfs_bpppool, sp->bpp);
593 					sp->bpp = NULL;
594 					pool_put(&fs->lfs_segpool, sp);
595 					sp = fs->lfs_sp = NULL;
596 					return (error);
597 				}
598 			fs->lfs_writer++;
599 			writer_set = 1;
600 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
601 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
602 		}
603 	}
604 
605 	/*
606 	 * If we are doing a checkpoint, mark everything since the
607 	 * last checkpoint as no longer ACTIVE.
608 	 */
609 	if (do_ckp) {
610 		segleft = fs->lfs_nseg;
611 		curseg = 0;
612 		for (n = 0; n < fs->lfs_segtabsz; n++) {
613 			dirty = 0;
614 			if (bread(fs->lfs_ivnode,
615 			    fs->lfs_cleansz + n, fs->lfs_bsize, NOCRED, &bp))
616 
617 				panic("lfs_segwrite: ifile read");
618 			segusep = (SEGUSE *)bp->b_data;
619 			maxseg = min(segleft, fs->lfs_sepb);
620 			for (i = 0; i < maxseg; i++) {
621 				sn = curseg + i;
622 				if (sn != fs->lfs_curseg &&
623 				    segusep->su_flags & SEGUSE_ACTIVE) {
624 					segusep->su_flags &= ~SEGUSE_ACTIVE;
625 					--fs->lfs_nactive;
626 					++dirty;
627 				}
628 				fs->lfs_suflags[fs->lfs_activesb][sn] =
629 					segusep->su_flags;
630 				if (fs->lfs_version > 1)
631 					++segusep;
632 				else
633 					segusep = (SEGUSE *)
634 						((SEGUSE_V1 *)segusep + 1);
635 			}
636 
637 			if (dirty)
638 				error = LFS_BWRITE_LOG(bp); /* Ifile */
639 			else
640 				brelse(bp);
641 			segleft -= fs->lfs_sepb;
642 			curseg += fs->lfs_sepb;
643 		}
644 	}
645 
646 	did_ckp = 0;
647 	if (do_ckp || fs->lfs_doifile) {
648 		do {
649 			vp = fs->lfs_ivnode;
650 
651 #ifdef DEBUG
652 			LFS_ENTER_LOG("pretend", __FILE__, __LINE__, 0, 0);
653 #endif
654 			fs->lfs_flags &= ~LFS_IFDIRTY;
655 
656 			ip = VTOI(vp);
657 
658 			if (LIST_FIRST(&vp->v_dirtyblkhd) != NULL)
659 				lfs_writefile(fs, sp, vp);
660 
661 			if (ip->i_flag & IN_ALLMOD)
662 				++did_ckp;
663 			redo = lfs_writeinode(fs, sp, ip);
664 			redo += lfs_writeseg(fs, sp);
665 			redo += (fs->lfs_flags & LFS_IFDIRTY);
666 		} while (redo && do_ckp);
667 
668 		/*
669 		 * Unless we are unmounting, the Ifile may continue to have
670 		 * dirty blocks even after a checkpoint, due to changes to
671 		 * inodes' atime.  If we're checkpointing, it's "impossible"
672 		 * for other parts of the Ifile to be dirty after the loop
673 		 * above, since we hold the segment lock.
674 		 */
675 		s = splbio();
676 		if (LIST_EMPTY(&vp->v_dirtyblkhd)) {
677 			LFS_CLR_UINO(ip, IN_ALLMOD);
678 		}
679 #ifdef DIAGNOSTIC
680 		else if (do_ckp) {
681 			LIST_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
682 				if (bp->b_lblkno < fs->lfs_cleansz +
683 				    fs->lfs_segtabsz &&
684 				    !(bp->b_flags & B_GATHERED)) {
685 					panic("dirty blocks");
686 				}
687 			}
688 		}
689 #endif
690 		splx(s);
691 	} else {
692 		(void) lfs_writeseg(fs, sp);
693 	}
694 
695 	/* Note Ifile no longer needs to be written */
696 	fs->lfs_doifile = 0;
697 	if (writer_set && --fs->lfs_writer == 0)
698 		wakeup(&fs->lfs_dirops);
699 
700 	/*
701 	 * If we didn't write the Ifile, we didn't really do anything.
702 	 * That means that (1) there is a checkpoint on disk and (2)
703 	 * nothing has changed since it was written.
704 	 *
705 	 * Take the flags off of the segment so that lfs_segunlock
706 	 * doesn't have to write the superblock either.
707 	 */
708 	if (do_ckp && !did_ckp) {
709 		sp->seg_flags &= ~SEGM_CKP;
710 	}
711 
712 	if (lfs_dostats) {
713 		++lfs_stats.nwrites;
714 		if (sp->seg_flags & SEGM_SYNC)
715 			++lfs_stats.nsync_writes;
716 		if (sp->seg_flags & SEGM_CKP)
717 			++lfs_stats.ncheckpoints;
718 	}
719 	lfs_segunlock(fs);
720 	return (0);
721 }
722 
723 /*
724  * Write the dirty blocks associated with a vnode.
725  */
726 void
727 lfs_writefile(struct lfs *fs, struct segment *sp, struct vnode *vp)
728 {
729 	struct buf *bp;
730 	struct finfo *fip;
731 	struct inode *ip;
732 	IFILE *ifp;
733 	int i, frag;
734 
735 	ip = VTOI(vp);
736 
737 	if (sp->seg_bytes_left < fs->lfs_bsize ||
738 	    sp->sum_bytes_left < sizeof(struct finfo))
739 		(void) lfs_writeseg(fs, sp);
740 
741 	sp->sum_bytes_left -= FINFOSIZE;
742 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
743 
744 	if (vp->v_flag & VDIROP)
745 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
746 
747 	fip = sp->fip;
748 	fip->fi_nblocks = 0;
749 	fip->fi_ino = ip->i_number;
750 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
751 	fip->fi_version = ifp->if_version;
752 	brelse(bp);
753 
754 	if (sp->seg_flags & SEGM_CLEAN) {
755 		lfs_gather(fs, sp, vp, lfs_match_fake);
756 		/*
757 		 * For a file being flushed, we need to write *all* blocks.
758 		 * This means writing the cleaning blocks first, and then
759 		 * immediately following with any non-cleaning blocks.
760 		 * The same is true of the Ifile since checkpoints assume
761 		 * that all valid Ifile blocks are written.
762 		 */
763 		if (IS_FLUSHING(fs,vp) || vp == fs->lfs_ivnode) {
764 			lfs_gather(fs, sp, vp, lfs_match_data);
765 			/*
766 			 * Don't call VOP_PUTPAGES: if we're flushing,
767 			 * we've already done it, and the Ifile doesn't
768 			 * use the page cache.
769 			 */
770 		}
771 	} else {
772 		lfs_gather(fs, sp, vp, lfs_match_data);
773 		/*
774 		 * If we're flushing, we've already called VOP_PUTPAGES
775 		 * so don't do it again.  Otherwise, we want to write
776 		 * everything we've got.
777 		 */
778 		if (!IS_FLUSHING(fs, vp)) {
779 			simple_lock(&vp->v_interlock);
780 			VOP_PUTPAGES(vp, 0, 0,
781 				     PGO_CLEANIT | PGO_ALLPAGES | PGO_LOCKED);
782 		}
783 	}
784 
785 	/*
786 	 * It may not be necessary to write the meta-data blocks at this point,
787 	 * as the roll-forward recovery code should be able to reconstruct the
788 	 * list.
789 	 *
790 	 * We have to write them anyway, though, under two conditions: (1) the
791 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
792 	 * checkpointing.
793 	 *
794 	 * BUT if we are cleaning, we might have indirect blocks that refer to
795 	 * new blocks not being written yet, in addition to fragments being
796 	 * moved out of a cleaned segment.  If that is the case, don't
797 	 * write the indirect blocks, or the finfo will have a small block
798 	 * in the middle of it!
799 	 * XXX in this case isn't the inode size wrong too?
800 	 */
801 	frag = 0;
802 	if (sp->seg_flags & SEGM_CLEAN) {
803 		for (i = 0; i < NDADDR; i++)
804 			if (ip->i_lfs_fragsize[i] > 0 &&
805 			    ip->i_lfs_fragsize[i] < fs->lfs_bsize)
806 				++frag;
807 	}
808 #ifdef DIAGNOSTIC
809 	if (frag > 1)
810 		panic("lfs_writefile: more than one fragment!");
811 #endif
812 	if (IS_FLUSHING(fs, vp) ||
813 	    (frag == 0 && (lfs_writeindir || (sp->seg_flags & SEGM_CKP)))) {
814 		lfs_gather(fs, sp, vp, lfs_match_indir);
815 		lfs_gather(fs, sp, vp, lfs_match_dindir);
816 		lfs_gather(fs, sp, vp, lfs_match_tindir);
817 	}
818 	fip = sp->fip;
819 	if (fip->fi_nblocks != 0) {
820 		sp->fip = (FINFO*)((caddr_t)fip + FINFOSIZE +
821 				   sizeof(int32_t) * (fip->fi_nblocks));
822 		sp->start_lbp = &sp->fip->fi_blocks[0];
823 	} else {
824 		sp->sum_bytes_left += FINFOSIZE;
825 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
826 	}
827 }
828 
829 int
830 lfs_writeinode(struct lfs *fs, struct segment *sp, struct inode *ip)
831 {
832 	struct buf *bp, *ibp;
833 	struct ufs1_dinode *cdp;
834 	IFILE *ifp;
835 	SEGUSE *sup;
836 	daddr_t daddr;
837 	int32_t *daddrp;	/* XXX ondisk32 */
838 	ino_t ino;
839 	int error, i, ndx, fsb = 0;
840 	int redo_ifile = 0;
841 	struct timespec ts;
842 	int gotblk = 0;
843 
844 	if (!(ip->i_flag & IN_ALLMOD))
845 		return (0);
846 
847 	/* Allocate a new inode block if necessary. */
848 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp == NULL) && sp->ibp == NULL) {
849 		/* Allocate a new segment if necessary. */
850 		if (sp->seg_bytes_left < fs->lfs_ibsize ||
851 		    sp->sum_bytes_left < sizeof(int32_t))
852 			(void) lfs_writeseg(fs, sp);
853 
854 		/* Get next inode block. */
855 		daddr = fs->lfs_offset;
856 		fs->lfs_offset += btofsb(fs, fs->lfs_ibsize);
857 		sp->ibp = *sp->cbpp++ =
858 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr),
859 			       fs->lfs_ibsize, 0, 0);
860 		gotblk++;
861 
862 		/* Zero out inode numbers */
863 		for (i = 0; i < INOPB(fs); ++i)
864 			((struct ufs1_dinode *)sp->ibp->b_data)[i].di_inumber = 0;
865 
866 		++sp->start_bpp;
867 		fs->lfs_avail -= btofsb(fs, fs->lfs_ibsize);
868 		/* Set remaining space counters. */
869 		sp->seg_bytes_left -= fs->lfs_ibsize;
870 		sp->sum_bytes_left -= sizeof(int32_t);
871 		ndx = fs->lfs_sumsize / sizeof(int32_t) -
872 			sp->ninodes / INOPB(fs) - 1;
873 		((int32_t *)(sp->segsum))[ndx] = daddr;
874 	}
875 
876 	/* Update the inode times and copy the inode onto the inode page. */
877 	TIMEVAL_TO_TIMESPEC(&time, &ts);
878 	/* XXX kludge --- don't redirty the ifile just to put times on it */
879 	if (ip->i_number != LFS_IFILE_INUM)
880 		LFS_ITIMES(ip, &ts, &ts, &ts);
881 
882 	/*
883 	 * If this is the Ifile, and we've already written the Ifile in this
884 	 * partial segment, just overwrite it (it's not on disk yet) and
885 	 * continue.
886 	 *
887 	 * XXX we know that the bp that we get the second time around has
888 	 * already been gathered.
889 	 */
890 	if (ip->i_number == LFS_IFILE_INUM && sp->idp) {
891 		*(sp->idp) = *ip->i_din.ffs1_din;
892 		ip->i_lfs_osize = ip->i_size;
893 		return 0;
894 	}
895 
896 	bp = sp->ibp;
897 	cdp = ((struct ufs1_dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
898 	*cdp = *ip->i_din.ffs1_din;
899 #ifdef LFS_IFILE_FRAG_ADDRESSING
900 	if (fs->lfs_version > 1)
901 		fsb = (sp->ninodes % INOPB(fs)) / INOPF(fs);
902 #endif
903 
904 	/*
905 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
906 	 * addresses to disk; possibly revert the inode size.
907 	 * XXX By not writing these blocks, we are making the lfs_avail
908 	 * XXX count on disk wrong by the same amount.	We should be
909 	 * XXX able to "borrow" from lfs_avail and return it after the
910 	 * XXX Ifile is written.  See also in lfs_writeseg.
911 	 */
912 	if (ip->i_lfs_effnblks != ip->i_ffs1_blocks) {
913 		cdp->di_size = ip->i_lfs_osize;
914 #ifdef DEBUG_LFS
915 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
916 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs1_blocks);
917 #endif
918 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
919 		     daddrp++) {
920 			if (*daddrp == UNWRITTEN) {
921 #ifdef DEBUG_LFS
922 				printf("lfs_writeinode: wiping UNWRITTEN\n");
923 #endif
924 				*daddrp = 0;
925 			}
926 		}
927 	} else {
928 		/* If all blocks are goig to disk, update the "size on disk" */
929 		ip->i_lfs_osize = ip->i_size;
930 	}
931 
932 	if (ip->i_flag & IN_CLEANING)
933 		LFS_CLR_UINO(ip, IN_CLEANING);
934 	else {
935 		/* XXX IN_ALLMOD */
936 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
937 			     IN_UPDATE);
938 		if (ip->i_lfs_effnblks == ip->i_ffs1_blocks)
939 			LFS_CLR_UINO(ip, IN_MODIFIED);
940 #ifdef DEBUG_LFS
941 		else
942 			printf("lfs_writeinode: ino %d: real blks=%d, "
943 			       "eff=%d\n", ip->i_number, ip->i_ffs1_blocks,
944 			       ip->i_lfs_effnblks);
945 #endif
946 	}
947 
948 	if (ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
949 		sp->idp = ((struct ufs1_dinode *)bp->b_data) +
950 			(sp->ninodes % INOPB(fs));
951 	if (gotblk) {
952 		LFS_LOCK_BUF(bp);
953 		brelse(bp);
954 	}
955 
956 	/* Increment inode count in segment summary block. */
957 	++((SEGSUM *)(sp->segsum))->ss_ninos;
958 
959 	/* If this page is full, set flag to allocate a new page. */
960 	if (++sp->ninodes % INOPB(fs) == 0)
961 		sp->ibp = NULL;
962 
963 	/*
964 	 * If updating the ifile, update the super-block.  Update the disk
965 	 * address and access times for this inode in the ifile.
966 	 */
967 	ino = ip->i_number;
968 	if (ino == LFS_IFILE_INUM) {
969 		daddr = fs->lfs_idaddr;
970 		fs->lfs_idaddr = dbtofsb(fs, bp->b_blkno);
971 	} else {
972 		LFS_IENTRY(ifp, fs, ino, ibp);
973 		daddr = ifp->if_daddr;
974 		ifp->if_daddr = dbtofsb(fs, bp->b_blkno) + fsb;
975 #ifdef LFS_DEBUG_NEXTFREE
976 		if (ino > 3 && ifp->if_nextfree) {
977 			vprint("lfs_writeinode",ITOV(ip));
978 			printf("lfs_writeinode: updating free ino %d\n",
979 				ip->i_number);
980 		}
981 #endif
982 		error = LFS_BWRITE_LOG(ibp); /* Ifile */
983 	}
984 
985 	/*
986 	 * The inode's last address should not be in the current partial
987 	 * segment, except under exceptional circumstances (lfs_writevnodes
988 	 * had to start over, and in the meantime more blocks were written
989 	 * to a vnode).	 Both inodes will be accounted to this segment
990 	 * in lfs_writeseg so we need to subtract the earlier version
991 	 * here anyway.	 The segment count can temporarily dip below
992 	 * zero here; keep track of how many duplicates we have in
993 	 * "dupino" so we don't panic below.
994 	 */
995 	if (daddr >= fs->lfs_lastpseg && daddr <= dbtofsb(fs, bp->b_blkno)) {
996 		++sp->ndupino;
997 		printf("lfs_writeinode: last inode addr in current pseg "
998 		       "(ino %d daddr 0x%llx) ndupino=%d\n", ino,
999 			(long long)daddr, sp->ndupino);
1000 	}
1001 	/*
1002 	 * Account the inode: it no longer belongs to its former segment,
1003 	 * though it will not belong to the new segment until that segment
1004 	 * is actually written.
1005 	 */
1006 	if (daddr != LFS_UNUSED_DADDR) {
1007 		u_int32_t oldsn = dtosn(fs, daddr);
1008 #ifdef DIAGNOSTIC
1009 		int ndupino = (sp->seg_number == oldsn) ? sp->ndupino : 0;
1010 #endif
1011 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1012 #ifdef DIAGNOSTIC
1013 		if (sup->su_nbytes +
1014 		    sizeof (struct ufs1_dinode) * ndupino
1015 		      < sizeof (struct ufs1_dinode)) {
1016 			printf("lfs_writeinode: negative bytes "
1017 			       "(segment %" PRIu32 " short by %d, "
1018 			       "oldsn=%" PRIu32 ", cursn=%" PRIu32
1019 			       ", daddr=%" PRId64 ", su_nbytes=%u, "
1020 			       "ndupino=%d)\n",
1021 			       dtosn(fs, daddr),
1022 			       (int)sizeof (struct ufs1_dinode) * (1 - sp->ndupino)
1023 				   - sup->su_nbytes,
1024 			       oldsn, sp->seg_number, daddr,
1025 			       (unsigned int)sup->su_nbytes,
1026 			       sp->ndupino);
1027 			panic("lfs_writeinode: negative bytes");
1028 			sup->su_nbytes = sizeof (struct ufs1_dinode);
1029 		}
1030 #endif
1031 #ifdef DEBUG_SU_NBYTES
1032 		printf("seg %d -= %d for ino %d inode\n",
1033 		       dtosn(fs, daddr), sizeof (struct ufs1_dinode), ino);
1034 #endif
1035 		sup->su_nbytes -= sizeof (struct ufs1_dinode);
1036 		redo_ifile =
1037 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
1038 		if (redo_ifile)
1039 			fs->lfs_flags |= LFS_IFDIRTY;
1040 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp); /* Ifile */
1041 	}
1042 	return (redo_ifile);
1043 }
1044 
1045 int
1046 lfs_gatherblock(struct segment *sp, struct buf *bp, int *sptr)
1047 {
1048 	struct lfs *fs;
1049 	int version;
1050 	int j, blksinblk;
1051 
1052 	/*
1053 	 * If full, finish this segment.  We may be doing I/O, so
1054 	 * release and reacquire the splbio().
1055 	 */
1056 #ifdef DIAGNOSTIC
1057 	if (sp->vp == NULL)
1058 		panic ("lfs_gatherblock: Null vp in segment");
1059 #endif
1060 	fs = sp->fs;
1061 	blksinblk = howmany(bp->b_bcount, fs->lfs_bsize);
1062 	if (sp->sum_bytes_left < sizeof(int32_t) * blksinblk ||
1063 	    sp->seg_bytes_left < bp->b_bcount) {
1064 		if (sptr)
1065 			splx(*sptr);
1066 		lfs_updatemeta(sp);
1067 
1068 		version = sp->fip->fi_version;
1069 		(void) lfs_writeseg(fs, sp);
1070 
1071 		sp->fip->fi_version = version;
1072 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
1073 		/* Add the current file to the segment summary. */
1074 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
1075 		sp->sum_bytes_left -= FINFOSIZE;
1076 
1077 		if (sptr)
1078 			*sptr = splbio();
1079 		return (1);
1080 	}
1081 
1082 #ifdef DEBUG
1083 	if (bp->b_flags & B_GATHERED) {
1084 		printf("lfs_gatherblock: already gathered! Ino %d,"
1085 		       " lbn %" PRId64 "\n",
1086 		       sp->fip->fi_ino, bp->b_lblkno);
1087 		return (0);
1088 	}
1089 #endif
1090 	/* Insert into the buffer list, update the FINFO block. */
1091 	bp->b_flags |= B_GATHERED;
1092 
1093 	*sp->cbpp++ = bp;
1094 	for (j = 0; j < blksinblk; j++)
1095 		sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno + j;
1096 
1097 	sp->sum_bytes_left -= sizeof(int32_t) * blksinblk;
1098 	sp->seg_bytes_left -= bp->b_bcount;
1099 	return (0);
1100 }
1101 
1102 int
1103 lfs_gather(struct lfs *fs, struct segment *sp, struct vnode *vp, int (*match)(struct lfs *, struct buf *))
1104 {
1105 	struct buf *bp, *nbp;
1106 	int s, count = 0;
1107 
1108 	sp->vp = vp;
1109 	s = splbio();
1110 
1111 #ifndef LFS_NO_BACKBUF_HACK
1112 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
1113 # define	BUF_OFFSET	(((caddr_t)&LIST_NEXT(bp, b_vnbufs)) - (caddr_t)bp)
1114 # define	BACK_BUF(BP)	((struct buf *)(((caddr_t)(BP)->b_vnbufs.le_prev) - BUF_OFFSET))
1115 # define	BEG_OF_LIST	((struct buf *)(((caddr_t)&LIST_FIRST(&vp->v_dirtyblkhd)) - BUF_OFFSET))
1116 /* Find last buffer. */
1117 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp && LIST_NEXT(bp, b_vnbufs) != NULL;
1118 	    bp = LIST_NEXT(bp, b_vnbufs));
1119 	for (; bp && bp != BEG_OF_LIST; bp = nbp) {
1120 		nbp = BACK_BUF(bp);
1121 #else /* LFS_NO_BACKBUF_HACK */
1122 loop:	for (bp = LIST_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1123 		nbp = LIST_NEXT(bp, b_vnbufs);
1124 #endif /* LFS_NO_BACKBUF_HACK */
1125 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp)) {
1126 #ifdef DEBUG_LFS
1127 			if (vp == fs->lfs_ivnode && (bp->b_flags & (B_BUSY|B_GATHERED)) == B_BUSY)
1128 				printf("(%" PRId64 ":%lx)", bp->b_lblkno, bp->b_flags);
1129 #endif
1130 			continue;
1131 		}
1132 		if (vp->v_type == VBLK) {
1133 			/* For block devices, just write the blocks. */
1134 			/* XXX Do we really need to even do this? */
1135 #ifdef DEBUG_LFS
1136 			if (count == 0)
1137 				printf("BLK(");
1138 			printf(".");
1139 #endif
1140 			/* Get the block before bwrite, so we don't corrupt the free list */
1141 			bp->b_flags |= B_BUSY;
1142 			bremfree(bp);
1143 			bwrite(bp);
1144 		} else {
1145 #ifdef DIAGNOSTIC
1146 # ifdef LFS_USE_B_INVAL
1147 			if ((bp->b_flags & (B_CALL|B_INVAL)) == B_INVAL) {
1148 				printf("lfs_gather: lbn %" PRId64 " is "
1149 					"B_INVAL\n", bp->b_lblkno);
1150 				VOP_PRINT(bp->b_vp);
1151 			}
1152 # endif /* LFS_USE_B_INVAL */
1153 			if (!(bp->b_flags & B_DELWRI))
1154 				panic("lfs_gather: bp not B_DELWRI");
1155 			if (!(bp->b_flags & B_LOCKED)) {
1156 				printf("lfs_gather: lbn %" PRId64 " blk "
1157 					"%" PRId64 " not B_LOCKED\n",
1158 					bp->b_lblkno,
1159 					dbtofsb(fs, bp->b_blkno));
1160 				VOP_PRINT(bp->b_vp);
1161 				panic("lfs_gather: bp not B_LOCKED");
1162 			}
1163 #endif
1164 			if (lfs_gatherblock(sp, bp, &s)) {
1165 				goto loop;
1166 			}
1167 		}
1168 		count++;
1169 	}
1170 	splx(s);
1171 #ifdef DEBUG_LFS
1172 	if (vp->v_type == VBLK && count)
1173 		printf(")\n");
1174 #endif
1175 	lfs_updatemeta(sp);
1176 	sp->vp = NULL;
1177 	return count;
1178 }
1179 
1180 #if DEBUG
1181 # define DEBUG_OOFF(n) do {						\
1182 	if (ooff == 0) {						\
1183 		printf("lfs_updatemeta[%d]: warning: writing "		\
1184 			"ino %d lbn %" PRId64 " at 0x%" PRIx32		\
1185 			", was 0x0 (or %" PRId64 ")\n",			\
1186 			(n), ip->i_number, lbn, ndaddr, daddr);		\
1187 	}								\
1188 } while (0)
1189 #else
1190 # define DEBUG_OOFF(n)
1191 #endif
1192 
1193 /*
1194  * Change the given block's address to ndaddr, finding its previous
1195  * location using ufs_bmaparray().
1196  *
1197  * Account for this change in the segment table.
1198  */
1199 void
1200 lfs_update_single(struct lfs *fs, struct segment *sp, daddr_t lbn,
1201 		  int32_t ndaddr, int size)
1202 {
1203 	SEGUSE *sup;
1204 	struct buf *bp;
1205 	struct indir a[NIADDR + 2], *ap;
1206 	struct inode *ip;
1207 	struct vnode *vp;
1208 	daddr_t daddr, ooff;
1209 	int num, error;
1210 	int bb, osize, obb;
1211 
1212 	vp = sp->vp;
1213 	ip = VTOI(vp);
1214 
1215 	error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL, NULL);
1216 	if (error)
1217 		panic("lfs_updatemeta: ufs_bmaparray returned %d", error);
1218 
1219 	KASSERT(daddr <= LFS_MAX_DADDR);
1220 	if (daddr > 0)
1221 		daddr = dbtofsb(fs, daddr);
1222 
1223 	bb = fragstofsb(fs, numfrags(fs, size));
1224 	switch (num) {
1225 	    case 0:
1226 		    ooff = ip->i_ffs1_db[lbn];
1227 		    DEBUG_OOFF(0);
1228 		    if (ooff == UNWRITTEN)
1229 			    ip->i_ffs1_blocks += bb;
1230 		    else {
1231 			    /* possible fragment truncation or extension */
1232 			    obb = btofsb(fs, ip->i_lfs_fragsize[lbn]);
1233 			    ip->i_ffs1_blocks += (bb - obb);
1234 		    }
1235 		    ip->i_ffs1_db[lbn] = ndaddr;
1236 		    break;
1237 	    case 1:
1238 		    ooff = ip->i_ffs1_ib[a[0].in_off];
1239 		    DEBUG_OOFF(1);
1240 		    if (ooff == UNWRITTEN)
1241 			    ip->i_ffs1_blocks += bb;
1242 		    ip->i_ffs1_ib[a[0].in_off] = ndaddr;
1243 		    break;
1244 	    default:
1245 		    ap = &a[num - 1];
1246 		    if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1247 			    panic("lfs_updatemeta: bread bno %" PRId64,
1248 				  ap->in_lbn);
1249 
1250 		    /* XXX ondisk32 */
1251 		    ooff = ((int32_t *)bp->b_data)[ap->in_off];
1252 		    DEBUG_OOFF(num);
1253 		    if (ooff == UNWRITTEN)
1254 			    ip->i_ffs1_blocks += bb;
1255 		    /* XXX ondisk32 */
1256 		    ((int32_t *)bp->b_data)[ap->in_off] = ndaddr;
1257 		    (void) VOP_BWRITE(bp);
1258 	}
1259 
1260 	/*
1261 	 * Though we'd rather it couldn't, this *can* happen right now
1262 	 * if cleaning blocks and regular blocks coexist.
1263 	 */
1264 	/* KASSERT(daddr < fs->lfs_lastpseg || daddr > ndaddr); */
1265 
1266 	/*
1267 	 * Update segment usage information, based on old size
1268 	 * and location.
1269 	 */
1270 	if (daddr > 0) {
1271 		u_int32_t oldsn = dtosn(fs, daddr);
1272 #ifdef DIAGNOSTIC
1273 		int ndupino = (sp->seg_number == oldsn) ?
1274 			sp->ndupino : 0;
1275 #endif
1276 		KASSERT(oldsn >= 0 && oldsn < fs->lfs_nseg);
1277 		if (lbn >= 0 && lbn < NDADDR)
1278 			osize = ip->i_lfs_fragsize[lbn];
1279 		else
1280 			osize = fs->lfs_bsize;
1281 		LFS_SEGENTRY(sup, fs, oldsn, bp);
1282 #ifdef DIAGNOSTIC
1283 		if (sup->su_nbytes + sizeof (struct ufs1_dinode) * ndupino
1284 		    < osize) {
1285 			printf("lfs_updatemeta: negative bytes "
1286 			       "(segment %" PRIu32 " short by %" PRId64
1287 			       ")\n", dtosn(fs, daddr),
1288 			       (int64_t)osize -
1289 			       (sizeof (struct ufs1_dinode) * sp->ndupino +
1290 				sup->su_nbytes));
1291 			printf("lfs_updatemeta: ino %d, lbn %" PRId64
1292 			       ", addr = 0x%" PRIx64 "\n",
1293 			       VTOI(sp->vp)->i_number, lbn, daddr);
1294 			printf("lfs_updatemeta: ndupino=%d\n", ndupino);
1295 			panic("lfs_updatemeta: negative bytes");
1296 			sup->su_nbytes = osize -
1297 			    sizeof (struct ufs1_dinode) * sp->ndupino;
1298 		}
1299 #endif
1300 #ifdef DEBUG_SU_NBYTES
1301 		printf("seg %" PRIu32 " -= %d for ino %d lbn %" PRId64
1302 		       " db 0x%" PRIx64 "\n",
1303 		       dtosn(fs, daddr), osize,
1304 		       VTOI(sp->vp)->i_number, lbn, daddr);
1305 #endif
1306 		sup->su_nbytes -= osize;
1307 		if (!(bp->b_flags & B_GATHERED))
1308 			fs->lfs_flags |= LFS_IFDIRTY;
1309 		LFS_WRITESEGENTRY(sup, fs, oldsn, bp);
1310 	}
1311 	/*
1312 	 * Now that this block has a new address, and its old
1313 	 * segment no longer owns it, we can forget about its
1314 	 * old size.
1315 	 */
1316 	if (lbn >= 0 && lbn < NDADDR)
1317 		ip->i_lfs_fragsize[lbn] = size;
1318 }
1319 
1320 /*
1321  * Update the metadata that points to the blocks listed in the FINFO
1322  * array.
1323  */
1324 void
1325 lfs_updatemeta(struct segment *sp)
1326 {
1327 	struct buf *sbp;
1328 	struct lfs *fs;
1329 	struct vnode *vp;
1330 	daddr_t lbn;
1331 	int i, nblocks, num;
1332 	int bb;
1333 	int bytesleft, size;
1334 
1335 	vp = sp->vp;
1336 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1337 	KASSERT(nblocks >= 0);
1338 	if (vp == NULL || nblocks == 0)
1339 		return;
1340 
1341 	/*
1342 	 * This count may be high due to oversize blocks from lfs_gop_write.
1343 	 * Correct for this. (XXX we should be able to keep track of these.)
1344 	 */
1345 	fs = sp->fs;
1346 	for (i = 0; i < nblocks; i++) {
1347 		if (sp->start_bpp[i] == NULL) {
1348 			printf("nblocks = %d, not %d\n", i, nblocks);
1349 			nblocks = i;
1350 			break;
1351 		}
1352 		num = howmany(sp->start_bpp[i]->b_bcount, fs->lfs_bsize);
1353 		KASSERT(sp->start_bpp[i]->b_lblkno >= 0 || num == 1);
1354 		nblocks -= num - 1;
1355 	}
1356 
1357 	KASSERT(vp->v_type == VREG ||
1358 	   nblocks == &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp);
1359 	KASSERT(nblocks == sp->cbpp - sp->start_bpp);
1360 
1361 	/*
1362 	 * Sort the blocks.
1363 	 *
1364 	 * We have to sort even if the blocks come from the
1365 	 * cleaner, because there might be other pending blocks on the
1366 	 * same inode...and if we don't sort, and there are fragments
1367 	 * present, blocks may be written in the wrong place.
1368 	 */
1369 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks, fs->lfs_bsize);
1370 
1371 	/*
1372 	 * Record the length of the last block in case it's a fragment.
1373 	 * If there are indirect blocks present, they sort last.  An
1374 	 * indirect block will be lfs_bsize and its presence indicates
1375 	 * that you cannot have fragments.
1376 	 *
1377 	 * XXX This last is a lie.  A cleaned fragment can coexist with
1378 	 * XXX a later indirect block.	This will continue to be
1379 	 * XXX true until lfs_markv is fixed to do everything with
1380 	 * XXX fake blocks (including fake inodes and fake indirect blocks).
1381 	 */
1382 	sp->fip->fi_lastlength = ((sp->start_bpp[nblocks - 1]->b_bcount - 1) &
1383 		fs->lfs_bmask) + 1;
1384 
1385 	/*
1386 	 * Assign disk addresses, and update references to the logical
1387 	 * block and the segment usage information.
1388 	 */
1389 	for (i = nblocks; i--; ++sp->start_bpp) {
1390 		sbp = *sp->start_bpp;
1391 		lbn = *sp->start_lbp;
1392 		KASSERT(sbp->b_lblkno == lbn);
1393 
1394 		sbp->b_blkno = fsbtodb(fs, fs->lfs_offset);
1395 
1396 		/*
1397 		 * If we write a frag in the wrong place, the cleaner won't
1398 		 * be able to correctly identify its size later, and the
1399 		 * segment will be uncleanable.	 (Even worse, it will assume
1400 		 * that the indirect block that actually ends the list
1401 		 * is of a smaller size!)
1402 		 */
1403 		if ((sbp->b_bcount & fs->lfs_bmask) && i != 0)
1404 			panic("lfs_updatemeta: fragment is not last block");
1405 
1406 		/*
1407 		 * For each subblock in this possibly oversized block,
1408 		 * update its address on disk.
1409 		 */
1410 		KASSERT(lbn >= 0 || sbp->b_bcount == fs->lfs_bsize);
1411 		for (bytesleft = sbp->b_bcount; bytesleft > 0;
1412 		     bytesleft -= fs->lfs_bsize) {
1413 			size = MIN(bytesleft, fs->lfs_bsize);
1414 			bb = fragstofsb(fs, numfrags(fs, size));
1415 			lbn = *sp->start_lbp++;
1416 			lfs_update_single(fs, sp, lbn, fs->lfs_offset, size);
1417 			fs->lfs_offset += bb;
1418 		}
1419 
1420 	}
1421 }
1422 
1423 /*
1424  * Start a new segment.
1425  */
1426 int
1427 lfs_initseg(struct lfs *fs)
1428 {
1429 	struct segment *sp;
1430 	SEGUSE *sup;
1431 	SEGSUM *ssp;
1432 	struct buf *bp, *sbp;
1433 	int repeat;
1434 
1435 	sp = fs->lfs_sp;
1436 
1437 	repeat = 0;
1438 
1439 	/* Advance to the next segment. */
1440 	if (!LFS_PARTIAL_FITS(fs)) {
1441 		/* lfs_avail eats the remaining space */
1442 		fs->lfs_avail -= fs->lfs_fsbpseg - (fs->lfs_offset -
1443 						   fs->lfs_curseg);
1444 		/* Wake up any cleaning procs waiting on this file system. */
1445 		wakeup(&lfs_allclean_wakeup);
1446 		wakeup(&fs->lfs_nextseg);
1447 		lfs_newseg(fs);
1448 		repeat = 1;
1449 		fs->lfs_offset = fs->lfs_curseg;
1450 
1451 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1452 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg);
1453 
1454 		/*
1455 		 * If the segment contains a superblock, update the offset
1456 		 * and summary address to skip over it.
1457 		 */
1458 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1459 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1460 			fs->lfs_offset += btofsb(fs, LFS_SBPAD);
1461 			sp->seg_bytes_left -= LFS_SBPAD;
1462 		}
1463 		brelse(bp);
1464 		/* Segment zero could also contain the labelpad */
1465 		if (fs->lfs_version > 1 && sp->seg_number == 0 &&
1466 		    fs->lfs_start < btofsb(fs, LFS_LABELPAD)) {
1467 			fs->lfs_offset += btofsb(fs, LFS_LABELPAD) - fs->lfs_start;
1468 			sp->seg_bytes_left -= LFS_LABELPAD - fsbtob(fs, fs->lfs_start);
1469 		}
1470 	} else {
1471 		sp->seg_number = dtosn(fs, fs->lfs_curseg);
1472 		sp->seg_bytes_left = fsbtob(fs, fs->lfs_fsbpseg -
1473 				      (fs->lfs_offset - fs->lfs_curseg));
1474 	}
1475 	fs->lfs_lastpseg = fs->lfs_offset;
1476 
1477 	/* Record first address of this partial segment */
1478 	if (sp->seg_flags & SEGM_CLEAN) {
1479 		fs->lfs_cleanint[fs->lfs_cleanind] = fs->lfs_offset;
1480 		if (++fs->lfs_cleanind >= LFS_MAX_CLEANIND) {
1481 			/* "1" is the artificial inc in lfs_seglock */
1482 			while (fs->lfs_iocount > 1) {
1483 				tsleep(&fs->lfs_iocount, PRIBIO + 1, "lfs_initseg", 0);
1484 			}
1485 			fs->lfs_cleanind = 0;
1486 		}
1487 	}
1488 
1489 	sp->fs = fs;
1490 	sp->ibp = NULL;
1491 	sp->idp = NULL;
1492 	sp->ninodes = 0;
1493 	sp->ndupino = 0;
1494 
1495 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1496 	sp->cbpp = sp->bpp;
1497 	sbp = *sp->cbpp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp,
1498 				     fsbtodb(fs, fs->lfs_offset), fs->lfs_sumsize, LFS_NB_SUMMARY);
1499 	sp->segsum = (*sp->cbpp)->b_data;
1500 	memset(sp->segsum, 0, fs->lfs_sumsize);
1501 	sp->start_bpp = ++sp->cbpp;
1502 	fs->lfs_offset += btofsb(fs, fs->lfs_sumsize);
1503 
1504 	/* Set point to SEGSUM, initialize it. */
1505 	ssp = sp->segsum;
1506 	ssp->ss_next = fs->lfs_nextseg;
1507 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1508 	ssp->ss_magic = SS_MAGIC;
1509 
1510 	/* Set pointer to first FINFO, initialize it. */
1511 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + SEGSUM_SIZE(fs));
1512 	sp->fip->fi_nblocks = 0;
1513 	sp->start_lbp = &sp->fip->fi_blocks[0];
1514 	sp->fip->fi_lastlength = 0;
1515 
1516 	sp->seg_bytes_left -= fs->lfs_sumsize;
1517 	sp->sum_bytes_left = fs->lfs_sumsize - SEGSUM_SIZE(fs);
1518 
1519 	return (repeat);
1520 }
1521 
1522 /*
1523  * Return the next segment to write.
1524  */
1525 void
1526 lfs_newseg(struct lfs *fs)
1527 {
1528 	CLEANERINFO *cip;
1529 	SEGUSE *sup;
1530 	struct buf *bp;
1531 	int curseg, isdirty, sn;
1532 
1533 	LFS_SEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1534 #ifdef DEBUG_SU_NBYTES
1535 	printf("lfs_newseg: seg %d := 0 in newseg\n",	/* XXXDEBUG */
1536 	       dtosn(fs, fs->lfs_nextseg)); /* XXXDEBUG */
1537 #endif
1538 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1539 	sup->su_nbytes = 0;
1540 	sup->su_nsums = 0;
1541 	sup->su_ninos = 0;
1542 	LFS_WRITESEGENTRY(sup, fs, dtosn(fs, fs->lfs_nextseg), bp);
1543 
1544 	LFS_CLEANERINFO(cip, fs, bp);
1545 	--cip->clean;
1546 	++cip->dirty;
1547 	fs->lfs_nclean = cip->clean;
1548 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1549 
1550 	fs->lfs_lastseg = fs->lfs_curseg;
1551 	fs->lfs_curseg = fs->lfs_nextseg;
1552 	for (sn = curseg = dtosn(fs, fs->lfs_curseg) + fs->lfs_interleave;;) {
1553 		sn = (sn + 1) % fs->lfs_nseg;
1554 		if (sn == curseg)
1555 			panic("lfs_nextseg: no clean segments");
1556 		LFS_SEGENTRY(sup, fs, sn, bp);
1557 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1558 		/* Check SEGUSE_EMPTY as we go along */
1559 		if (isdirty && sup->su_nbytes == 0 && !(sup->su_flags & SEGUSE_EMPTY))
1560 			LFS_WRITESEGENTRY(sup, fs, sn, bp);
1561 		else
1562 			brelse(bp);
1563 
1564 		if (!isdirty)
1565 			break;
1566 	}
1567 
1568 	++fs->lfs_nactive;
1569 	fs->lfs_nextseg = sntod(fs, sn);
1570 	if (lfs_dostats) {
1571 		++lfs_stats.segsused;
1572 	}
1573 }
1574 
1575 #define BQUEUES 4 /* XXX */
1576 #define BQ_EMPTY 3 /* XXX */
1577 extern TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
1578 extern struct simplelock bqueue_slock;
1579 
1580 #define	BUFHASH(dvp, lbn)	\
1581 	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
1582 extern LIST_HEAD(bufhashhdr, buf) invalhash;
1583 /*
1584  * Insq/Remq for the buffer hash lists.
1585  */
1586 #define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
1587 #define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
1588 
1589 static struct buf *
1590 lfs_newclusterbuf(struct lfs *fs, struct vnode *vp, daddr_t addr, int n)
1591 {
1592 	struct lfs_cluster *cl;
1593 	struct buf **bpp, *bp;
1594 	int s;
1595 
1596 	cl = (struct lfs_cluster *)pool_get(&fs->lfs_clpool, PR_WAITOK);
1597 	bpp = (struct buf **)pool_get(&fs->lfs_bpppool, PR_WAITOK);
1598 	memset(cl, 0, sizeof(*cl));
1599 	cl->fs = fs;
1600 	cl->bpp = bpp;
1601 	cl->bufcount = 0;
1602 	cl->bufsize = 0;
1603 
1604 	/* If this segment is being written synchronously, note that */
1605 	if (fs->lfs_sp->seg_flags & SEGM_SYNC) {
1606 		cl->flags |= LFS_CL_SYNC;
1607 		cl->seg = fs->lfs_sp;
1608 		++cl->seg->seg_iocount;
1609 		/* printf("+ %x => %d\n", cl->seg, cl->seg->seg_iocount); */
1610 	}
1611 
1612 	/* Get an empty buffer header, or maybe one with something on it */
1613 	s = splbio();
1614 	simple_lock(&bqueue_slock);
1615 	if ((bp = bufqueues[BQ_EMPTY].tqh_first) != NULL) {
1616 		simple_lock(&bp->b_interlock);
1617 		bremfree(bp);
1618 		/* clear out various other fields */
1619 		bp->b_flags = B_BUSY;
1620 		bp->b_dev = NODEV;
1621 		bp->b_blkno = bp->b_lblkno = 0;
1622 		bp->b_error = 0;
1623 		bp->b_resid = 0;
1624 		bp->b_bcount = 0;
1625 
1626 		/* nuke any credentials we were holding */
1627 		/* XXXXXX */
1628 
1629 		bremhash(bp);
1630 
1631 		/* disassociate us from our vnode, if we had one... */
1632 		if (bp->b_vp)
1633 			brelvp(bp);
1634 	}
1635 	while (!bp)
1636 		bp = getnewbuf(0, 0);
1637 	bgetvp(vp, bp);
1638 	binshash(bp,&invalhash);
1639 	simple_unlock(&bp->b_interlock);
1640 	simple_unlock(&bqueue_slock);
1641 	splx(s);
1642 	bp->b_bcount = 0;
1643 	bp->b_blkno = bp->b_lblkno = addr;
1644 
1645 	bp->b_flags |= B_CALL;
1646 	bp->b_iodone = lfs_cluster_callback;
1647 	cl->saveaddr = bp->b_saveaddr; /* XXX is this ever used? */
1648 	bp->b_saveaddr = (caddr_t)cl;
1649 
1650 	return bp;
1651 }
1652 
1653 int
1654 lfs_writeseg(struct lfs *fs, struct segment *sp)
1655 {
1656 	struct buf **bpp, *bp, *cbp, *newbp;
1657 	SEGUSE *sup;
1658 	SEGSUM *ssp;
1659 	dev_t i_dev;
1660 	char *datap, *dp;
1661 	int i, s;
1662 	int do_again, nblocks, byteoffset;
1663 	size_t el_size;
1664 	struct lfs_cluster *cl;
1665 	int (*strategy)(void *);
1666 	struct vop_strategy_args vop_strategy_a;
1667 	u_short ninos;
1668 	struct vnode *devvp;
1669 	char *p;
1670 	struct vnode *vp;
1671 	int32_t *daddrp;	/* XXX ondisk32 */
1672 	int changed;
1673 #if defined(DEBUG) && defined(LFS_PROPELLER)
1674 	static int propeller;
1675 	char propstring[4] = "-\\|/";
1676 
1677 	printf("%c\b",propstring[propeller++]);
1678 	if (propeller == 4)
1679 		propeller = 0;
1680 #endif
1681 
1682 	/*
1683 	 * If there are no buffers other than the segment summary to write
1684 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1685 	 * even if there aren't any buffers, you need to write the superblock.
1686 	 */
1687 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1688 		return (0);
1689 
1690 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1691 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1692 
1693 	/* Update the segment usage information. */
1694 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1695 
1696 	/* Loop through all blocks, except the segment summary. */
1697 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1698 		if ((*bpp)->b_vp != devvp) {
1699 			sup->su_nbytes += (*bpp)->b_bcount;
1700 #ifdef DEBUG_SU_NBYTES
1701 		printf("seg %" PRIu32 " += %ld for ino %d lbn %" PRId64
1702 		    " db 0x%" PRIx64 "\n", sp->seg_number, (*bpp)->b_bcount,
1703 		    VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno,
1704 		    (*bpp)->b_blkno);
1705 #endif
1706 		}
1707 	}
1708 
1709 	ssp = (SEGSUM *)sp->segsum;
1710 
1711 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1712 #ifdef DEBUG_SU_NBYTES
1713 	printf("seg %d += %d for %d inodes\n",	 /* XXXDEBUG */
1714 	       sp->seg_number, ssp->ss_ninos * sizeof (struct ufs1_dinode),
1715 	       ssp->ss_ninos);
1716 #endif
1717 	sup->su_nbytes += ssp->ss_ninos * sizeof (struct ufs1_dinode);
1718 	/* sup->su_nbytes += fs->lfs_sumsize; */
1719 	if (fs->lfs_version == 1)
1720 		sup->su_olastmod = time.tv_sec;
1721 	else
1722 		sup->su_lastmod = time.tv_sec;
1723 	sup->su_ninos += ninos;
1724 	++sup->su_nsums;
1725 	fs->lfs_dmeta += (btofsb(fs, fs->lfs_sumsize) + btofsb(fs, ninos *
1726 							 fs->lfs_ibsize));
1727 	fs->lfs_avail -= btofsb(fs, fs->lfs_sumsize);
1728 
1729 	do_again = !(bp->b_flags & B_GATHERED);
1730 	LFS_WRITESEGENTRY(sup, fs, sp->seg_number, bp); /* Ifile */
1731 
1732 	/*
1733 	 * Mark blocks B_BUSY, to prevent then from being changed between
1734 	 * the checksum computation and the actual write.
1735 	 *
1736 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1737 	 * there are any, replace them with copies that have UNASSIGNED
1738 	 * instead.
1739 	 */
1740 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1741 		++bpp;
1742 		bp = *bpp;
1743 		if (bp->b_flags & B_CALL) { /* UBC or malloced buffer */
1744 			bp->b_flags |= B_BUSY;
1745 			continue;
1746 		}
1747 	    again:
1748 		s = splbio();
1749 		if (bp->b_flags & B_BUSY) {
1750 #ifdef DEBUG
1751 			printf("lfs_writeseg: avoiding potential data summary "
1752 			       "corruption for ino %d, lbn %" PRId64 "\n",
1753 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1754 #endif
1755 			bp->b_flags |= B_WANTED;
1756 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1757 			splx(s);
1758 			goto again;
1759 		}
1760 		bp->b_flags |= B_BUSY;
1761 		splx(s);
1762 		/*
1763 		 * Check and replace indirect block UNWRITTEN bogosity.
1764 		 * XXX See comment in lfs_writefile.
1765 		 */
1766 		if (bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1767 		   VTOI(bp->b_vp)->i_ffs1_blocks !=
1768 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1769 #ifdef DEBUG_LFS
1770 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1771 			       VTOI(bp->b_vp)->i_number,
1772 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1773 			       VTOI(bp->b_vp)->i_ffs1_blocks);
1774 #endif
1775 			/* Make a copy we'll make changes to */
1776 			newbp = lfs_newbuf(fs, bp->b_vp, bp->b_lblkno,
1777 					   bp->b_bcount, LFS_NB_IBLOCK);
1778 			newbp->b_blkno = bp->b_blkno;
1779 			memcpy(newbp->b_data, bp->b_data,
1780 			       newbp->b_bcount);
1781 
1782 			changed = 0;
1783 			/* XXX ondisk32 */
1784 			for (daddrp = (int32_t *)(newbp->b_data);
1785 			     daddrp < (int32_t *)(newbp->b_data +
1786 						  newbp->b_bcount); daddrp++) {
1787 				if (*daddrp == UNWRITTEN) {
1788 #ifdef DEBUG_LFS
1789 					off_t doff;
1790 					int32_t ioff;
1791 
1792 					ioff = daddrp - (int32_t *)(newbp->b_data);
1793 					doff = (-bp->b_lblkno + ioff) * fs->lfs_bsize;
1794 					printf("ino %d lbn %" PRId64 " entry %d off %" PRIx64 "\n",
1795 					       VTOI(bp->b_vp)->i_number,
1796 					       bp->b_lblkno, ioff, doff);
1797 					if (bp->b_vp->v_type == VREG) {
1798 						/*
1799 						 * What is up with this page?
1800 						 */
1801 						struct vm_page *pg;
1802 						for (; doff / fs->lfs_bsize == (-bp->b_lblkno + ioff); doff += PAGE_SIZE) {
1803 							pg = uvm_pagelookup(&bp->b_vp->v_uobj, doff);
1804 							if (pg == NULL)
1805 								printf("  page at %" PRIx64 " is NULL\n", doff);
1806 							else
1807 								printf("  page at %" PRIx64 " flags 0x%x pqflags 0x%x\n", doff, pg->flags, pg->pqflags);
1808 						}
1809 					}
1810 #endif /* DEBUG_LFS */
1811 					++changed;
1812 					*daddrp = 0;
1813 				}
1814 			}
1815 			/*
1816 			 * Get rid of the old buffer.  Don't mark it clean,
1817 			 * though, if it still has dirty data on it.
1818 			 */
1819 			if (changed) {
1820 #ifdef DEBUG_LFS
1821 				printf("lfs_writeseg: replacing UNWRITTEN(%d):"
1822 					" bp = %p newbp = %p\n", changed, bp,
1823 					newbp);
1824 #endif
1825 				*bpp = newbp;
1826 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1827 				if (bp->b_flags & B_CALL) {
1828 					printf("lfs_writeseg: indir bp should not be B_CALL\n");
1829 					s = splbio();
1830 					biodone(bp);
1831 					splx(s);
1832 					bp = NULL;
1833 				} else {
1834 					/* Still on free list, leave it there */
1835 					s = splbio();
1836 					bp->b_flags &= ~B_BUSY;
1837 					if (bp->b_flags & B_WANTED)
1838 						wakeup(bp);
1839 					splx(s);
1840 					/*
1841 					 * We have to re-decrement lfs_avail
1842 					 * since this block is going to come
1843 					 * back around to us in the next
1844 					 * segment.
1845 					 */
1846 					fs->lfs_avail -= btofsb(fs, bp->b_bcount);
1847 				}
1848 			} else {
1849 				lfs_freebuf(fs, newbp);
1850 			}
1851 		}
1852 	}
1853 	/*
1854 	 * Compute checksum across data and then across summary; the first
1855 	 * block (the summary block) is skipped.  Set the create time here
1856 	 * so that it's guaranteed to be later than the inode mod times.
1857 	 *
1858 	 * XXX
1859 	 * Fix this to do it inline, instead of malloc/copy.
1860 	 */
1861 	datap = dp = pool_get(&fs->lfs_bpppool, PR_WAITOK);
1862 	if (fs->lfs_version == 1)
1863 		el_size = sizeof(u_long);
1864 	else
1865 		el_size = sizeof(u_int32_t);
1866 	for (bpp = sp->bpp, i = nblocks - 1; i--; ) {
1867 		++bpp;
1868 		/* Loop through gop_write cluster blocks */
1869 		for (byteoffset = 0; byteoffset < (*bpp)->b_bcount;
1870 		     byteoffset += fs->lfs_bsize) {
1871 #ifdef LFS_USE_B_INVAL
1872 			if (((*bpp)->b_flags & (B_CALL | B_INVAL)) ==
1873 			    (B_CALL | B_INVAL)) {
1874 				if (copyin((caddr_t)(*bpp)->b_saveaddr +
1875 					   byteoffset, dp, el_size)) {
1876 					panic("lfs_writeseg: copyin failed [1]: "
1877 						"ino %d blk %" PRId64,
1878 						VTOI((*bpp)->b_vp)->i_number,
1879 						(*bpp)->b_lblkno);
1880 				}
1881 			} else
1882 #endif /* LFS_USE_B_INVAL */
1883 			{
1884 				memcpy(dp, (*bpp)->b_data + byteoffset,
1885 				       el_size);
1886 			}
1887 			dp += el_size;
1888 		}
1889 	}
1890 	if (fs->lfs_version == 1)
1891 		ssp->ss_ocreate = time.tv_sec;
1892 	else {
1893 		ssp->ss_create = time.tv_sec;
1894 		ssp->ss_serial = ++fs->lfs_serial;
1895 		ssp->ss_ident  = fs->lfs_ident;
1896 	}
1897 	ssp->ss_datasum = cksum(datap, dp - datap);
1898 	ssp->ss_sumsum =
1899 	    cksum(&ssp->ss_datasum, fs->lfs_sumsize - sizeof(ssp->ss_sumsum));
1900 	pool_put(&fs->lfs_bpppool, datap);
1901 	datap = dp = NULL;
1902 #ifdef DIAGNOSTIC
1903 	if (fs->lfs_bfree < btofsb(fs, ninos * fs->lfs_ibsize) + btofsb(fs, fs->lfs_sumsize))
1904 		panic("lfs_writeseg: No diskspace for summary");
1905 #endif
1906 	fs->lfs_bfree -= (btofsb(fs, ninos * fs->lfs_ibsize) +
1907 			  btofsb(fs, fs->lfs_sumsize));
1908 
1909 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1910 
1911 	/*
1912 	 * When we simply write the blocks we lose a rotation for every block
1913 	 * written.  To avoid this problem, we cluster the buffers into a
1914 	 * chunk and write the chunk.  MAXPHYS is the largest size I/O
1915 	 * devices can handle, use that for the size of the chunks.
1916 	 *
1917 	 * Blocks that are already clusters (from GOP_WRITE), however, we
1918 	 * don't bother to copy into other clusters.
1919 	 */
1920 
1921 #define CHUNKSIZE MAXPHYS
1922 
1923 	if (devvp == NULL)
1924 		panic("devvp is NULL");
1925 	for (bpp = sp->bpp, i = nblocks; i;) {
1926 		cbp = lfs_newclusterbuf(fs, devvp, (*bpp)->b_blkno, i);
1927 		cl = (struct lfs_cluster *)cbp->b_saveaddr;
1928 
1929 		cbp->b_dev = i_dev;
1930 		cbp->b_flags |= B_ASYNC | B_BUSY;
1931 		cbp->b_bcount = 0;
1932 
1933 		cl->olddata = cbp->b_data;
1934 #if defined(DEBUG) && defined(DIAGNOSTIC)
1935 		if (bpp - sp->bpp > (fs->lfs_sumsize - SEGSUM_SIZE(fs))
1936 		    / sizeof(int32_t)) {
1937 			panic("lfs_writeseg: real bpp overwrite");
1938 		}
1939 		if (bpp - sp->bpp > fs->lfs_ssize / fs->lfs_fsize) {
1940 			panic("lfs_writeseg: theoretical bpp overwrite");
1941 		}
1942 #endif
1943 
1944 		/*
1945 		 * Construct the cluster.
1946 		 */
1947 		++fs->lfs_iocount;
1948 		while (i && cbp->b_bcount < CHUNKSIZE) {
1949 			bp = *bpp;
1950 
1951 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1952 				break;
1953 			if (cbp->b_bcount > 0 && !(cl->flags & LFS_CL_MALLOC))
1954 				break;
1955 
1956 			/* Clusters from GOP_WRITE are expedited */
1957 			if (bp->b_bcount > fs->lfs_bsize) {
1958 				if (cbp->b_bcount > 0)
1959 					/* Put in its own buffer */
1960 					break;
1961 				else {
1962 					cbp->b_data = bp->b_data;
1963 				}
1964 			} else if (cbp->b_bcount == 0) {
1965 				p = cbp->b_data = lfs_malloc(fs, CHUNKSIZE,
1966 							     LFS_NB_CLUSTER);
1967 				cl->flags |= LFS_CL_MALLOC;
1968 			}
1969 #ifdef DIAGNOSTIC
1970 			if (dtosn(fs, dbtofsb(fs, bp->b_blkno +
1971 					      btodb(bp->b_bcount - 1))) !=
1972 			    sp->seg_number) {
1973 				printf("blk size %ld daddr %" PRIx64 " not in seg %d\n",
1974 					bp->b_bcount, bp->b_blkno,
1975 					sp->seg_number);
1976 				panic("segment overwrite");
1977 			}
1978 #endif
1979 
1980 #ifdef LFS_USE_B_INVAL
1981 			/*
1982 			 * Fake buffers from the cleaner are marked as B_INVAL.
1983 			 * We need to copy the data from user space rather than
1984 			 * from the buffer indicated.
1985 			 * XXX == what do I do on an error?
1986 			 */
1987 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1988 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1989 					panic("lfs_writeseg: copyin failed [2]");
1990 			} else
1991 #endif /* LFS_USE_B_INVAL */
1992 			if (cl->flags & LFS_CL_MALLOC) {
1993 				bcopy(bp->b_data, p, bp->b_bcount);
1994 			}
1995 
1996 			p += bp->b_bcount;
1997 			cbp->b_bcount += bp->b_bcount;
1998 			cl->bufsize += bp->b_bcount;
1999 
2000 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI | B_DONE);
2001 			cl->bpp[cl->bufcount++] = bp;
2002 			vp = bp->b_vp;
2003 			s = splbio();
2004 			V_INCR_NUMOUTPUT(vp);
2005 			splx(s);
2006 
2007 			bpp++;
2008 			i--;
2009 		}
2010 		s = splbio();
2011 		V_INCR_NUMOUTPUT(devvp);
2012 		splx(s);
2013 		vop_strategy_a.a_desc = VDESC(vop_strategy);
2014 		vop_strategy_a.a_bp = cbp;
2015 		(strategy)(&vop_strategy_a);
2016 		curproc->p_stats->p_ru.ru_oublock++;
2017 	}
2018 
2019 	if (lfs_dostats) {
2020 		++lfs_stats.psegwrites;
2021 		lfs_stats.blocktot += nblocks - 1;
2022 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
2023 			++lfs_stats.psyncwrites;
2024 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
2025 			++lfs_stats.pcleanwrites;
2026 			lfs_stats.cleanblocks += nblocks - 1;
2027 		}
2028 	}
2029 	return (lfs_initseg(fs) || do_again);
2030 }
2031 
2032 void
2033 lfs_writesuper(struct lfs *fs, daddr_t daddr)
2034 {
2035 	struct buf *bp;
2036 	dev_t i_dev;
2037 	int (*strategy)(void *);
2038 	int s;
2039 	struct vop_strategy_args vop_strategy_a;
2040 
2041 	/*
2042 	 * If we can write one superblock while another is in
2043 	 * progress, we risk not having a complete checkpoint if we crash.
2044 	 * So, block here if a superblock write is in progress.
2045 	 */
2046 	s = splbio();
2047 	while (fs->lfs_sbactive) {
2048 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
2049 	}
2050 	fs->lfs_sbactive = daddr;
2051 	splx(s);
2052 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
2053 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
2054 
2055 	/* Set timestamp of this version of the superblock */
2056 	if (fs->lfs_version == 1)
2057 		fs->lfs_otstamp = time.tv_sec;
2058 	fs->lfs_tstamp = time.tv_sec;
2059 
2060 	/* Checksum the superblock and copy it into a buffer. */
2061 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
2062 	bp = lfs_newbuf(fs, VTOI(fs->lfs_ivnode)->i_devvp, fsbtodb(fs, daddr), LFS_SBPAD, LFS_NB_SBLOCK);
2063 	memset(bp->b_data + sizeof(struct dlfs), 0, LFS_SBPAD - sizeof(struct dlfs));
2064 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
2065 
2066 	bp->b_dev = i_dev;
2067 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
2068 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
2069 	bp->b_iodone = lfs_supercallback;
2070 	/* XXX KS - same nasty hack as above */
2071 	bp->b_saveaddr = (caddr_t)fs;
2072 
2073 	vop_strategy_a.a_desc = VDESC(vop_strategy);
2074 	vop_strategy_a.a_bp = bp;
2075 	curproc->p_stats->p_ru.ru_oublock++;
2076 	s = splbio();
2077 	V_INCR_NUMOUTPUT(bp->b_vp);
2078 	splx(s);
2079 	++fs->lfs_iocount;
2080 	(strategy)(&vop_strategy_a);
2081 }
2082 
2083 /*
2084  * Logical block number match routines used when traversing the dirty block
2085  * chain.
2086  */
2087 int
2088 lfs_match_fake(struct lfs *fs, struct buf *bp)
2089 {
2090 	return LFS_IS_MALLOC_BUF(bp);
2091 }
2092 
2093 #if 0
2094 int
2095 lfs_match_real(struct lfs *fs, struct buf *bp)
2096 {
2097 	return (lfs_match_data(fs, bp) && !lfs_match_fake(fs, bp));
2098 }
2099 #endif
2100 
2101 int
2102 lfs_match_data(struct lfs *fs, struct buf *bp)
2103 {
2104 	return (bp->b_lblkno >= 0);
2105 }
2106 
2107 int
2108 lfs_match_indir(struct lfs *fs, struct buf *bp)
2109 {
2110 	daddr_t lbn;
2111 
2112 	lbn = bp->b_lblkno;
2113 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
2114 }
2115 
2116 int
2117 lfs_match_dindir(struct lfs *fs, struct buf *bp)
2118 {
2119 	daddr_t lbn;
2120 
2121 	lbn = bp->b_lblkno;
2122 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
2123 }
2124 
2125 int
2126 lfs_match_tindir(struct lfs *fs, struct buf *bp)
2127 {
2128 	daddr_t lbn;
2129 
2130 	lbn = bp->b_lblkno;
2131 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
2132 }
2133 
2134 /*
2135  * XXX - The only buffers that are going to hit these functions are the
2136  * segment write blocks, or the segment summaries, or the superblocks.
2137  *
2138  * All of the above are created by lfs_newbuf, and so do not need to be
2139  * released via brelse.
2140  */
2141 void
2142 lfs_callback(struct buf *bp)
2143 {
2144 	struct lfs *fs;
2145 
2146 	fs = (struct lfs *)bp->b_saveaddr;
2147 	lfs_freebuf(fs, bp);
2148 }
2149 
2150 static void
2151 lfs_super_aiodone(struct buf *bp)
2152 {
2153 	struct lfs *fs;
2154 
2155 	fs = (struct lfs *)bp->b_saveaddr;
2156 	fs->lfs_sbactive = 0;
2157 	wakeup(&fs->lfs_sbactive);
2158 	if (--fs->lfs_iocount <= 1)
2159 		wakeup(&fs->lfs_iocount);
2160 	lfs_freebuf(fs, bp);
2161 }
2162 
2163 static void
2164 lfs_cluster_aiodone(struct buf *bp)
2165 {
2166 	struct lfs_cluster *cl;
2167 	struct lfs *fs;
2168 	struct buf *tbp, *fbp;
2169 	struct vnode *vp, *devvp;
2170 	struct inode *ip;
2171 	int s, error=0;
2172 	char *cp;
2173 	extern int locked_queue_count;
2174 	extern long locked_queue_bytes;
2175 
2176 	if (bp->b_flags & B_ERROR)
2177 		error = bp->b_error;
2178 
2179 	cl = (struct lfs_cluster *)bp->b_saveaddr;
2180 	fs = cl->fs;
2181 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
2182 	bp->b_saveaddr = cl->saveaddr;
2183 
2184 	cp = (char *)bp->b_data + cl->bufsize;
2185 	/* Put the pages back, and release the buffer */
2186 	while (cl->bufcount--) {
2187 		tbp = cl->bpp[cl->bufcount];
2188 		if (error) {
2189 			tbp->b_flags |= B_ERROR;
2190 			tbp->b_error = error;
2191 		}
2192 
2193 		/*
2194 		 * We're done with tbp.	 If it has not been re-dirtied since
2195 		 * the cluster was written, free it.  Otherwise, keep it on
2196 		 * the locked list to be written again.
2197 		 */
2198 		vp = tbp->b_vp;
2199 
2200 		if ((tbp->b_flags & (B_LOCKED | B_DELWRI)) == B_LOCKED)
2201 			LFS_UNLOCK_BUF(tbp);
2202 
2203 		tbp->b_flags &= ~B_GATHERED;
2204 
2205 		LFS_BCLEAN_LOG(fs, tbp);
2206 
2207 		if (!(tbp->b_flags & B_CALL)) {
2208 			bremfree(tbp);
2209 			s = splbio();
2210 			if (vp)
2211 				reassignbuf(tbp, vp);
2212 			splx(s);
2213 			tbp->b_flags |= B_ASYNC; /* for biodone */
2214 		}
2215 #ifdef DIAGNOSTIC
2216 		if (tbp->b_flags & B_DONE) {
2217 			printf("blk %d biodone already (flags %lx)\n",
2218 				cl->bufcount, (long)tbp->b_flags);
2219 		}
2220 #endif
2221 		if (tbp->b_flags & (B_BUSY | B_CALL)) {
2222 			if ((tbp->b_flags & B_CALL) && !LFS_IS_MALLOC_BUF(tbp)) {
2223 				/* printf("flags 0x%lx\n", tbp->b_flags); */
2224 				/*
2225 				 * A buffer from the page daemon.
2226 				 * We use the same iodone as it does,
2227 				 * so we must manually disassociate its
2228 				 * buffers from the vp.
2229 				 */
2230 				if (tbp->b_vp) {
2231 					/* This is just silly */
2232 					s = splbio();
2233 					brelvp(tbp);
2234 					tbp->b_vp = vp;
2235 					splx(s);
2236 				}
2237 				/* Put it back the way it was */
2238 				tbp->b_flags |= B_ASYNC;
2239 				/* Master buffers have B_AGE */
2240 				if (tbp->b_private == tbp)
2241 					tbp->b_flags |= B_AGE;
2242 			}
2243 			s = splbio();
2244 			biodone(tbp);
2245 
2246 			/*
2247 			 * If this is the last block for this vnode, but
2248 			 * there are other blocks on its dirty list,
2249 			 * set IN_MODIFIED/IN_CLEANING depending on what
2250 			 * sort of block.  Only do this for our mount point,
2251 			 * not for, e.g., inode blocks that are attached to
2252 			 * the devvp.
2253 			 * XXX KS - Shouldn't we set *both* if both types
2254 			 * of blocks are present (traverse the dirty list?)
2255 			 */
2256 			simple_lock(&global_v_numoutput_slock);
2257 			if (vp != devvp && vp->v_numoutput == 0 &&
2258 			    (fbp = LIST_FIRST(&vp->v_dirtyblkhd)) != NULL) {
2259 				ip = VTOI(vp);
2260 #ifdef DEBUG_LFS
2261 				printf("lfs_cluster_aiodone: marking ino %d\n",
2262 				       ip->i_number);
2263 #endif
2264 				if (LFS_IS_MALLOC_BUF(fbp))
2265 					LFS_SET_UINO(ip, IN_CLEANING);
2266 				else
2267 					LFS_SET_UINO(ip, IN_MODIFIED);
2268 			}
2269 			simple_unlock(&global_v_numoutput_slock);
2270 			splx(s);
2271 			wakeup(vp);
2272 		}
2273 	}
2274 
2275 	/* Fix up the cluster buffer, and release it */
2276 	if (cl->flags & LFS_CL_MALLOC)
2277 		lfs_free(fs, bp->b_data, LFS_NB_CLUSTER);
2278 	bp->b_data = cl->olddata;
2279 	bp->b_bcount = 0;
2280 	bp->b_iodone = NULL;
2281 	bp->b_flags &= ~B_DELWRI;
2282 	bp->b_flags |= B_DONE;
2283 	s = splbio();
2284 	reassignbuf(bp, bp->b_vp);
2285 	splx(s);
2286 	brelse(bp);
2287 
2288 	/* Note i/o done */
2289 	if (cl->flags & LFS_CL_SYNC) {
2290 		if (--cl->seg->seg_iocount == 0)
2291 			wakeup(&cl->seg->seg_iocount);
2292 		/* printf("- %x => %d\n", cl->seg, cl->seg->seg_iocount); */
2293 	}
2294 #ifdef DIAGNOSTIC
2295 	if (fs->lfs_iocount == 0)
2296 		panic("lfs_cluster_aiodone: zero iocount");
2297 #endif
2298 	if (--fs->lfs_iocount <= 1)
2299 		wakeup(&fs->lfs_iocount);
2300 
2301 	pool_put(&fs->lfs_bpppool, cl->bpp);
2302 	cl->bpp = NULL;
2303 	pool_put(&fs->lfs_clpool, cl);
2304 }
2305 
2306 static void
2307 lfs_generic_callback(struct buf *bp, void (*aiodone)(struct buf *))
2308 {
2309 	/* reset b_iodone for when this is a single-buf i/o. */
2310 	bp->b_iodone = aiodone;
2311 
2312 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
2313 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
2314 	wakeup(&uvm.aiodoned);
2315 	simple_unlock(&uvm.aiodoned_lock);
2316 }
2317 
2318 static void
2319 lfs_cluster_callback(struct buf *bp)
2320 {
2321 	lfs_generic_callback(bp, lfs_cluster_aiodone);
2322 }
2323 
2324 void
2325 lfs_supercallback(struct buf *bp)
2326 {
2327 	lfs_generic_callback(bp, lfs_super_aiodone);
2328 }
2329 
2330 /*
2331  * Shellsort (diminishing increment sort) from Data Structures and
2332  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
2333  * see also Knuth Vol. 3, page 84.  The increments are selected from
2334  * formula (8), page 95.  Roughly O(N^3/2).
2335  */
2336 /*
2337  * This is our own private copy of shellsort because we want to sort
2338  * two parallel arrays (the array of buffer pointers and the array of
2339  * logical block numbers) simultaneously.  Note that we cast the array
2340  * of logical block numbers to a unsigned in this routine so that the
2341  * negative block numbers (meta data blocks) sort AFTER the data blocks.
2342  */
2343 
2344 void
2345 lfs_shellsort(struct buf **bp_array, int32_t *lb_array, int nmemb, int size)
2346 {
2347 	static int __rsshell_increments[] = { 4, 1, 0 };
2348 	int incr, *incrp, t1, t2;
2349 	struct buf *bp_temp;
2350 
2351 #ifdef DEBUG
2352 	incr = 0;
2353 	for (t1 = 0; t1 < nmemb; t1++) {
2354 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2355 			if (lb_array[incr++] != bp_array[t1]->b_lblkno + t2) {
2356 				/* dump before panic */
2357 				printf("lfs_shellsort: nmemb=%d, size=%d\n",
2358 				    nmemb, size);
2359 				incr = 0;
2360 				for (t1 = 0; t1 < nmemb; t1++) {
2361 					const struct buf *bp = bp_array[t1];
2362 
2363 					printf("bp[%d]: lbn=%" PRIu64 ", size=%"
2364 					    PRIu64 "\n", t1,
2365 					    (uint64_t)bp->b_bcount,
2366 					    (uint64_t)bp->b_lblkno);
2367 					printf("lbns:");
2368 					for (t2 = 0; t2 * size < bp->b_bcount;
2369 					    t2++) {
2370 						printf(" %" PRId32,
2371 						    lb_array[incr++]);
2372 					}
2373 					printf("\n");
2374 				}
2375 				panic("lfs_shellsort: inconsistent input");
2376 			}
2377 		}
2378 	}
2379 #endif
2380 
2381 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
2382 		for (t1 = incr; t1 < nmemb; ++t1)
2383 			for (t2 = t1 - incr; t2 >= 0;)
2384 				if ((u_int32_t)bp_array[t2]->b_lblkno >
2385 				    (u_int32_t)bp_array[t2 + incr]->b_lblkno) {
2386 					bp_temp = bp_array[t2];
2387 					bp_array[t2] = bp_array[t2 + incr];
2388 					bp_array[t2 + incr] = bp_temp;
2389 					t2 -= incr;
2390 				} else
2391 					break;
2392 
2393 	/* Reform the list of logical blocks */
2394 	incr = 0;
2395 	for (t1 = 0; t1 < nmemb; t1++) {
2396 		for (t2 = 0; t2 * size < bp_array[t1]->b_bcount; t2++) {
2397 			lb_array[incr++] = bp_array[t1]->b_lblkno + t2;
2398 		}
2399 	}
2400 }
2401 
2402 /*
2403  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
2404  */
2405 int
2406 lfs_vref(struct vnode *vp)
2407 {
2408 	/*
2409 	 * If we return 1 here during a flush, we risk vinvalbuf() not
2410 	 * being able to flush all of the pages from this vnode, which
2411 	 * will cause it to panic.  So, return 0 if a flush is in progress.
2412 	 */
2413 	if (vp->v_flag & VXLOCK) {
2414 		if (IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2415 			return 0;
2416 		}
2417 		return (1);
2418 	}
2419 	return (vget(vp, 0));
2420 }
2421 
2422 /*
2423  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
2424  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
2425  */
2426 void
2427 lfs_vunref(struct vnode *vp)
2428 {
2429 	/*
2430 	 * Analogous to lfs_vref, if the node is flushing, fake it.
2431 	 */
2432 	if ((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
2433 		return;
2434 	}
2435 
2436 	simple_lock(&vp->v_interlock);
2437 #ifdef DIAGNOSTIC
2438 	if (vp->v_usecount <= 0) {
2439 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
2440 		printf("lfs_vunref: flags are 0x%lx\n", (u_long)vp->v_flag);
2441 		printf("lfs_vunref: usecount = %ld\n", (long)vp->v_usecount);
2442 		panic("lfs_vunref: v_usecount<0");
2443 	}
2444 #endif
2445 	vp->v_usecount--;
2446 	if (vp->v_usecount > 0) {
2447 		simple_unlock(&vp->v_interlock);
2448 		return;
2449 	}
2450 	/*
2451 	 * insert at tail of LRU list
2452 	 */
2453 	simple_lock(&vnode_free_list_slock);
2454 	if (vp->v_holdcnt > 0)
2455 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2456 	else
2457 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2458 	simple_unlock(&vnode_free_list_slock);
2459 	simple_unlock(&vp->v_interlock);
2460 }
2461 
2462 /*
2463  * We use this when we have vnodes that were loaded in solely for cleaning.
2464  * There is no reason to believe that these vnodes will be referenced again
2465  * soon, since the cleaning process is unrelated to normal filesystem
2466  * activity.  Putting cleaned vnodes at the tail of the list has the effect
2467  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
2468  * cleaning at the head of the list, instead.
2469  */
2470 void
2471 lfs_vunref_head(struct vnode *vp)
2472 {
2473 	simple_lock(&vp->v_interlock);
2474 #ifdef DIAGNOSTIC
2475 	if (vp->v_usecount == 0) {
2476 		panic("lfs_vunref: v_usecount<0");
2477 	}
2478 #endif
2479 	vp->v_usecount--;
2480 	if (vp->v_usecount > 0) {
2481 		simple_unlock(&vp->v_interlock);
2482 		return;
2483 	}
2484 	/*
2485 	 * insert at head of LRU list
2486 	 */
2487 	simple_lock(&vnode_free_list_slock);
2488 	if (vp->v_holdcnt > 0)
2489 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
2490 	else
2491 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2492 	simple_unlock(&vnode_free_list_slock);
2493 	simple_unlock(&vp->v_interlock);
2494 }
2495 
2496