xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 5aefcfdc06931dd97e76246d2fe0302f7b3fe094)
1 /*	$NetBSD: lfs_segment.c,v 1.66 2000/12/03 05:56:27 perseant Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999, 2000 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 #define ivndebug(vp,str) printf("ino %d: %s\n",VTOI(vp)->i_number,(str))
74 
75 #if defined(_KERNEL) && !defined(_LKM)
76 #include "opt_ddb.h"
77 #endif
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/namei.h>
82 #include <sys/kernel.h>
83 #include <sys/resourcevar.h>
84 #include <sys/file.h>
85 #include <sys/stat.h>
86 #include <sys/buf.h>
87 #include <sys/proc.h>
88 #include <sys/conf.h>
89 #include <sys/vnode.h>
90 #include <sys/malloc.h>
91 #include <sys/mount.h>
92 
93 #include <miscfs/specfs/specdev.h>
94 #include <miscfs/fifofs/fifo.h>
95 
96 #include <ufs/ufs/quota.h>
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 extern int count_lock_queue __P((void));
106 extern struct simplelock vnode_free_list_slock;		/* XXX */
107 
108 /*
109  * Determine if it's OK to start a partial in this segment, or if we need
110  * to go on to a new segment.
111  */
112 #define	LFS_PARTIAL_FITS(fs) \
113 	((fs)->lfs_dbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
114 	1 << (fs)->lfs_fsbtodb)
115 
116 void	 lfs_callback __P((struct buf *));
117 int	 lfs_gather __P((struct lfs *, struct segment *,
118 	     struct vnode *, int (*) __P((struct lfs *, struct buf *))));
119 int	 lfs_gatherblock __P((struct segment *, struct buf *, int *));
120 void	 lfs_iset __P((struct inode *, ufs_daddr_t, time_t));
121 int	 lfs_match_fake __P((struct lfs *, struct buf *));
122 int	 lfs_match_data __P((struct lfs *, struct buf *));
123 int	 lfs_match_dindir __P((struct lfs *, struct buf *));
124 int	 lfs_match_indir __P((struct lfs *, struct buf *));
125 int	 lfs_match_tindir __P((struct lfs *, struct buf *));
126 void	 lfs_newseg __P((struct lfs *));
127 void	 lfs_shellsort __P((struct buf **, ufs_daddr_t *, int));
128 void	 lfs_supercallback __P((struct buf *));
129 void	 lfs_updatemeta __P((struct segment *));
130 int	 lfs_vref __P((struct vnode *));
131 void	 lfs_vunref __P((struct vnode *));
132 void	 lfs_writefile __P((struct lfs *, struct segment *, struct vnode *));
133 int	 lfs_writeinode __P((struct lfs *, struct segment *, struct inode *));
134 int	 lfs_writeseg __P((struct lfs *, struct segment *));
135 void	 lfs_writesuper __P((struct lfs *, daddr_t));
136 int	 lfs_writevnodes __P((struct lfs *fs, struct mount *mp,
137 	    struct segment *sp, int dirops));
138 
139 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
140 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
141 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
142 int	lfs_dirvcount = 0;		/* # active dirops */
143 
144 /* Statistics Counters */
145 int lfs_dostats = 1;
146 struct lfs_stats lfs_stats;
147 
148 extern int locked_queue_count;
149 extern long locked_queue_bytes;
150 
151 /* op values to lfs_writevnodes */
152 #define	VN_REG	        0
153 #define	VN_DIROP	1
154 #define	VN_EMPTY	2
155 #define VN_CLEAN        3
156 
157 #define LFS_MAX_ACTIVE          10
158 
159 /*
160  * XXX KS - Set modification time on the Ifile, so the cleaner can
161  * read the fs mod time off of it.  We don't set IN_UPDATE here,
162  * since we don't really need this to be flushed to disk (and in any
163  * case that wouldn't happen to the Ifile until we checkpoint).
164  */
165 void
166 lfs_imtime(fs)
167 	struct lfs *fs;
168 {
169 	struct timespec ts;
170 	struct inode *ip;
171 
172 	TIMEVAL_TO_TIMESPEC(&time, &ts);
173 	ip = VTOI(fs->lfs_ivnode);
174 	ip->i_ffs_mtime = ts.tv_sec;
175 	ip->i_ffs_mtimensec = ts.tv_nsec;
176 }
177 
178 /*
179  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
180  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
181  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
182  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
183  */
184 
185 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
186 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
187 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
188 
189 int
190 lfs_vflush(vp)
191 	struct vnode *vp;
192 {
193 	struct inode *ip;
194 	struct lfs *fs;
195 	struct segment *sp;
196 	struct buf *bp, *nbp, *tbp, *tnbp;
197 	int error, s;
198 
199 	ip = VTOI(vp);
200 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
201 
202 	if(ip->i_flag & IN_CLEANING) {
203 #ifdef DEBUG_LFS
204 		ivndebug(vp,"vflush/in_cleaning");
205 #endif
206 		LFS_CLR_UINO(ip, IN_CLEANING);
207 		LFS_SET_UINO(ip, IN_MODIFIED);
208 
209 		/*
210 		 * Toss any cleaning buffers that have real counterparts
211 		 * to avoid losing new data
212 		 */
213 		s = splbio();
214 		for(bp=vp->v_dirtyblkhd.lh_first; bp; bp=nbp) {
215 			nbp = bp->b_vnbufs.le_next;
216 			if(bp->b_flags & B_CALL) {
217 				for(tbp=vp->v_dirtyblkhd.lh_first; tbp;
218 				    tbp=tnbp)
219 				{
220 					tnbp = tbp->b_vnbufs.le_next;
221 					if(tbp->b_vp == bp->b_vp
222 					   && tbp->b_lblkno == bp->b_lblkno
223 					   && tbp != bp)
224 					{
225 						fs->lfs_avail += btodb(bp->b_bcount);
226 						wakeup(&fs->lfs_avail);
227 						lfs_freebuf(bp);
228 					}
229 				}
230 			}
231 		}
232 		splx(s);
233 	}
234 
235 	/* If the node is being written, wait until that is done */
236 	if(WRITEINPROG(vp)) {
237 #ifdef DEBUG_LFS
238 		ivndebug(vp,"vflush/writeinprog");
239 #endif
240 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
241 	}
242 
243 	/* Protect against VXLOCK deadlock in vinvalbuf() */
244 	lfs_seglock(fs, SEGM_SYNC);
245 
246 	/* If we're supposed to flush a freed inode, just toss it */
247 	/* XXX - seglock, so these buffers can't be gathered, right? */
248 	if(ip->i_ffs_mode == 0) {
249 		printf("lfs_vflush: ino %d is freed, not flushing\n",
250 			ip->i_number);
251 		s = splbio();
252 		for(bp=vp->v_dirtyblkhd.lh_first; bp; bp=nbp) {
253 			nbp = bp->b_vnbufs.le_next;
254 			if (bp->b_flags & B_DELWRI) { /* XXX always true? */
255 				fs->lfs_avail += btodb(bp->b_bcount);
256 				wakeup(&fs->lfs_avail);
257 			}
258 			/* Copied from lfs_writeseg */
259 			if (bp->b_flags & B_CALL) {
260 				/* if B_CALL, it was created with newbuf */
261 				lfs_freebuf(bp);
262 			} else {
263 				bremfree(bp);
264 				LFS_UNLOCK_BUF(bp);
265 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
266                                          B_GATHERED);
267 				bp->b_flags |= B_DONE;
268 				reassignbuf(bp, vp);
269 				brelse(bp);
270 			}
271 		}
272 		splx(s);
273 		LFS_CLR_UINO(ip, IN_CLEANING);
274 		LFS_CLR_UINO(ip, IN_MODIFIED | IN_ACCESSED);
275 		ip->i_flag &= ~IN_ALLMOD;
276 		printf("lfs_vflush: done not flushing ino %d\n",
277 			ip->i_number);
278 		lfs_segunlock(fs);
279 		return 0;
280 	}
281 
282 	SET_FLUSHING(fs,vp);
283 	if (fs->lfs_nactive > LFS_MAX_ACTIVE) {
284 		error = lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP);
285 		CLR_FLUSHING(fs,vp);
286 		lfs_segunlock(fs);
287 		return error;
288 	}
289 	sp = fs->lfs_sp;
290 
291 	if (vp->v_dirtyblkhd.lh_first == NULL) {
292 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
293 	} else if((ip->i_flag & IN_CLEANING) &&
294 		  (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
295 #ifdef DEBUG_LFS
296 		ivndebug(vp,"vflush/clean");
297 #endif
298 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
299 	}
300 	else if(lfs_dostats) {
301 		if(vp->v_dirtyblkhd.lh_first || (VTOI(vp)->i_flag & IN_ALLMOD))
302 			++lfs_stats.vflush_invoked;
303 #ifdef DEBUG_LFS
304 		ivndebug(vp,"vflush");
305 #endif
306 	}
307 
308 #ifdef DIAGNOSTIC
309 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
310 	if(vp->v_flag & VDIROP) {
311 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
312 		/* panic("VDIROP being flushed...this can\'t happen"); */
313 	}
314 	if(vp->v_usecount<0) {
315 		printf("usecount=%d\n",vp->v_usecount);
316 		panic("lfs_vflush: usecount<0");
317 	}
318 #endif
319 
320 	do {
321 		do {
322 			if (vp->v_dirtyblkhd.lh_first != NULL)
323 				lfs_writefile(fs, sp, vp);
324 		} while (lfs_writeinode(fs, sp, ip));
325 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
326 
327 	if(lfs_dostats) {
328 		++lfs_stats.nwrites;
329 		if (sp->seg_flags & SEGM_SYNC)
330 			++lfs_stats.nsync_writes;
331 		if (sp->seg_flags & SEGM_CKP)
332 			++lfs_stats.ncheckpoints;
333 	}
334 	lfs_segunlock(fs);
335 
336 	CLR_FLUSHING(fs,vp);
337 	return (0);
338 }
339 
340 #ifdef DEBUG_LFS_VERBOSE
341 # 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)
342 #else
343 # define vndebug(vp,str)
344 #endif
345 
346 int
347 lfs_writevnodes(fs, mp, sp, op)
348 	struct lfs *fs;
349 	struct mount *mp;
350 	struct segment *sp;
351 	int op;
352 {
353 	struct inode *ip;
354 	struct vnode *vp;
355 	int inodes_written=0, only_cleaning;
356 	int needs_unlock;
357 
358 #ifndef LFS_NO_BACKVP_HACK
359 	/* BEGIN HACK */
360 #define	VN_OFFSET	(((caddr_t)&vp->v_mntvnodes.le_next) - (caddr_t)vp)
361 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)VP->v_mntvnodes.le_prev) - VN_OFFSET))
362 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&mp->mnt_vnodelist.lh_first) - VN_OFFSET))
363 
364 	/* Find last vnode. */
365  loop:	for (vp = mp->mnt_vnodelist.lh_first;
366 	     vp && vp->v_mntvnodes.le_next != NULL;
367 	     vp = vp->v_mntvnodes.le_next);
368 	for (; vp && vp != BEG_OF_VLIST; vp = BACK_VP(vp)) {
369 #else
370 	loop:
371 	for (vp = mp->mnt_vnodelist.lh_first;
372 	     vp != NULL;
373 	     vp = vp->v_mntvnodes.le_next) {
374 #endif
375 		/*
376 		 * If the vnode that we are about to sync is no longer
377 		 * associated with this mount point, start over.
378 		 */
379 		if (vp->v_mount != mp) {
380 			printf("lfs_writevnodes: starting over\n");
381 			goto loop;
382 		}
383 
384 		ip = VTOI(vp);
385 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
386 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
387 			vndebug(vp,"dirop");
388 			continue;
389 		}
390 
391 		if (op == VN_EMPTY && vp->v_dirtyblkhd.lh_first) {
392 			vndebug(vp,"empty");
393 			continue;
394 		}
395 
396 		if (vp->v_type == VNON) {
397 			continue;
398 		}
399 
400 		if(op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
401 		   && vp != fs->lfs_flushvp
402 		   && !(ip->i_flag & IN_CLEANING)) {
403 			vndebug(vp,"cleaning");
404 			continue;
405 		}
406 
407 		if (lfs_vref(vp)) {
408 			vndebug(vp,"vref");
409 			continue;
410 		}
411 
412 		needs_unlock = 0;
413 		if (VOP_ISLOCKED(vp)) {
414 			if (vp != fs->lfs_ivnode &&
415 			    vp->v_lock.lk_lockholder != curproc->p_pid) {
416 #ifdef DEBUG_LFS
417 				printf("lfs_writevnodes: not writing ino %d,"
418 				       " locked by pid %d\n",
419 				       VTOI(vp)->i_number,
420 				       vp->v_lock.lk_lockholder);
421 #endif
422 				lfs_vunref(vp);
423 				continue;
424 			}
425 		} else if (vp != fs->lfs_ivnode) {
426 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
427 			needs_unlock = 1;
428 		}
429 
430 		only_cleaning = 0;
431 		/*
432 		 * Write the inode/file if dirty and it's not the IFILE.
433 		 */
434 		if ((ip->i_flag & IN_ALLMOD) ||
435 		     (vp->v_dirtyblkhd.lh_first != NULL))
436 		{
437 			only_cleaning = ((ip->i_flag & IN_ALLMOD)==IN_CLEANING);
438 
439 			if(ip->i_number != LFS_IFILE_INUM
440 			   && vp->v_dirtyblkhd.lh_first != NULL)
441 			{
442 				lfs_writefile(fs, sp, vp);
443 			}
444 			if(vp->v_dirtyblkhd.lh_first != NULL) {
445 				if(WRITEINPROG(vp)) {
446 #ifdef DEBUG_LFS
447 					ivndebug(vp,"writevnodes/write2");
448 #endif
449 				} else if(!(ip->i_flag & IN_ALLMOD)) {
450 #ifdef DEBUG_LFS
451 					printf("<%d>",ip->i_number);
452 #endif
453 					LFS_SET_UINO(ip, IN_MODIFIED);
454 				}
455 			}
456 			(void) lfs_writeinode(fs, sp, ip);
457 			inodes_written++;
458 		}
459 
460 		if (needs_unlock)
461 			VOP_UNLOCK(vp, 0);
462 
463 		if (lfs_clean_vnhead && only_cleaning)
464 			lfs_vunref_head(vp);
465 		else
466 			lfs_vunref(vp);
467 	}
468 	return inodes_written;
469 }
470 
471 int
472 lfs_segwrite(mp, flags)
473 	struct mount *mp;
474 	int flags;			/* Do a checkpoint. */
475 {
476 	struct buf *bp;
477 	struct inode *ip;
478 	struct lfs *fs;
479 	struct segment *sp;
480 	struct vnode *vp;
481 	SEGUSE *segusep;
482 	ufs_daddr_t ibno;
483 	int do_ckp, did_ckp, error, i;
484 	int writer_set = 0;
485 	int dirty;
486 
487 	fs = VFSTOUFS(mp)->um_lfs;
488 
489 	if (fs->lfs_ronly)
490 		return EROFS;
491 
492 	lfs_imtime(fs);
493 
494 	/* printf("lfs_segwrite: ifile flags are 0x%lx\n",
495 	       (long)(VTOI(fs->lfs_ivnode)->i_flag)); */
496 
497 #if 0
498 	/*
499 	 * If we are not the cleaner, and there is no space available,
500 	 * wait until cleaner writes.
501 	 */
502 	if(!(flags & SEGM_CLEAN) && !(fs->lfs_seglock && fs->lfs_sp &&
503 				      (fs->lfs_sp->seg_flags & SEGM_CLEAN)))
504 	{
505 		while (fs->lfs_avail <= 0) {
506 			LFS_CLEANERINFO(cip, fs, bp);
507 			LFS_SYNC_CLEANERINFO(cip, fs, bp, 0);
508 
509 			wakeup(&lfs_allclean_wakeup);
510 			wakeup(&fs->lfs_nextseg);
511 			error = tsleep(&fs->lfs_avail, PRIBIO + 1, "lfs_av2",
512 				       0);
513 			if (error) {
514 				return (error);
515 			}
516 		}
517 	}
518 #endif
519 	/*
520 	 * Allocate a segment structure and enough space to hold pointers to
521 	 * the maximum possible number of buffers which can be described in a
522 	 * single summary block.
523 	 */
524 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
525 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
526 	sp = fs->lfs_sp;
527 
528 	/*
529 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
530 	 * in which case we have to flush *all* buffers off of this vnode.
531 	 * We don't care about other nodes, but write any non-dirop nodes
532 	 * anyway in anticipation of another getnewvnode().
533 	 *
534 	 * If we're cleaning we only write cleaning and ifile blocks, and
535 	 * no dirops, since otherwise we'd risk corruption in a crash.
536 	 */
537 	if(sp->seg_flags & SEGM_CLEAN)
538 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
539 	else {
540 		lfs_writevnodes(fs, mp, sp, VN_REG);
541 		if(!fs->lfs_dirops || !fs->lfs_flushvp) {
542 			while(fs->lfs_dirops)
543 				if((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
544 						"lfs writer", 0)))
545 				{
546 					free(sp->bpp, M_SEGMENT);
547 					free(sp, M_SEGMENT);
548 					return (error);
549 				}
550 			fs->lfs_writer++;
551 			writer_set=1;
552 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
553 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
554 		}
555 	}
556 
557 	/*
558 	 * If we are doing a checkpoint, mark everything since the
559 	 * last checkpoint as no longer ACTIVE.
560 	 */
561 	if (do_ckp) {
562 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
563 		     --ibno >= fs->lfs_cleansz; ) {
564 			dirty = 0;
565 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
566 
567 				panic("lfs_segwrite: ifile read");
568 			segusep = (SEGUSE *)bp->b_data;
569 			for (i = fs->lfs_sepb; i--; segusep++) {
570 				if (segusep->su_flags & SEGUSE_ACTIVE) {
571 					segusep->su_flags &= ~SEGUSE_ACTIVE;
572 					++dirty;
573 				}
574 			}
575 
576 			/* But the current segment is still ACTIVE */
577 			segusep = (SEGUSE *)bp->b_data;
578 			if (datosn(fs, fs->lfs_curseg) / fs->lfs_sepb ==
579 			    (ibno-fs->lfs_cleansz)) {
580 				segusep[datosn(fs, fs->lfs_curseg) %
581 					fs->lfs_sepb].su_flags |= SEGUSE_ACTIVE;
582 				--dirty;
583 			}
584 			if (dirty)
585 				error = VOP_BWRITE(bp); /* Ifile */
586 			else
587 				brelse(bp);
588 		}
589 	}
590 
591 	did_ckp = 0;
592 	if (do_ckp || fs->lfs_doifile) {
593 		do {
594 			vp = fs->lfs_ivnode;
595 
596 			vget(vp, LK_EXCLUSIVE | LK_CANRECURSE | LK_RETRY);
597 
598 			ip = VTOI(vp);
599 			if (vp->v_dirtyblkhd.lh_first != NULL)
600 				lfs_writefile(fs, sp, vp);
601 			if (ip->i_flag & IN_ALLMOD)
602 				++did_ckp;
603 			(void) lfs_writeinode(fs, sp, ip);
604 
605 			vput(vp);
606 		} while (lfs_writeseg(fs, sp) && do_ckp);
607 
608 		/* The ifile should now be all clear */
609 		LFS_CLR_UINO(ip, IN_ALLMOD);
610 	} else {
611 		(void) lfs_writeseg(fs, sp);
612 	}
613 
614 	/*
615 	 * If the I/O count is non-zero, sleep until it reaches zero.
616 	 * At the moment, the user's process hangs around so we can
617 	 * sleep.
618 	 */
619 	fs->lfs_doifile = 0;
620 	if(writer_set && --fs->lfs_writer==0)
621 		wakeup(&fs->lfs_dirops);
622 
623 	/*
624 	 * If we didn't write the Ifile, we didn't really do anything.
625 	 * That means that (1) there is a checkpoint on disk and (2)
626 	 * nothing has changed since it was written.
627 	 *
628 	 * Take the flags off of the segment so that lfs_segunlock
629 	 * doesn't have to write the superblock either.
630 	 */
631 	if (did_ckp == 0) {
632 		sp->seg_flags &= ~(SEGM_SYNC|SEGM_CKP);
633 		/* if(do_ckp) printf("lfs_segwrite: no checkpoint\n"); */
634 	}
635 
636 	if(lfs_dostats) {
637 		++lfs_stats.nwrites;
638 		if (sp->seg_flags & SEGM_SYNC)
639 			++lfs_stats.nsync_writes;
640 		if (sp->seg_flags & SEGM_CKP)
641 			++lfs_stats.ncheckpoints;
642 	}
643 	lfs_segunlock(fs);
644 	return (0);
645 }
646 
647 /*
648  * Write the dirty blocks associated with a vnode.
649  */
650 void
651 lfs_writefile(fs, sp, vp)
652 	struct lfs *fs;
653 	struct segment *sp;
654 	struct vnode *vp;
655 {
656 	struct buf *bp;
657 	struct finfo *fip;
658 	IFILE *ifp;
659 
660 
661 	if (sp->seg_bytes_left < fs->lfs_bsize ||
662 	    sp->sum_bytes_left < sizeof(struct finfo))
663 		(void) lfs_writeseg(fs, sp);
664 
665 	sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t);
666 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
667 
668 	if(vp->v_flag & VDIROP)
669 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
670 
671 	fip = sp->fip;
672 	fip->fi_nblocks = 0;
673 	fip->fi_ino = VTOI(vp)->i_number;
674 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
675 	fip->fi_version = ifp->if_version;
676 	brelse(bp);
677 
678 	if(sp->seg_flags & SEGM_CLEAN)
679 	{
680 		lfs_gather(fs, sp, vp, lfs_match_fake);
681 		/*
682 		 * For a file being flushed, we need to write *all* blocks.
683 		 * This means writing the cleaning blocks first, and then
684 		 * immediately following with any non-cleaning blocks.
685 		 * The same is true of the Ifile since checkpoints assume
686 		 * that all valid Ifile blocks are written.
687 		 */
688 	   	if(IS_FLUSHING(fs,vp) || VTOI(vp)->i_number == LFS_IFILE_INUM)
689 			lfs_gather(fs, sp, vp, lfs_match_data);
690 	} else
691 		lfs_gather(fs, sp, vp, lfs_match_data);
692 
693 	/*
694 	 * It may not be necessary to write the meta-data blocks at this point,
695 	 * as the roll-forward recovery code should be able to reconstruct the
696 	 * list.
697 	 *
698 	 * We have to write them anyway, though, under two conditions: (1) the
699 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
700 	 * checkpointing.
701 	 */
702 	if(lfs_writeindir
703 	   || IS_FLUSHING(fs,vp)
704 	   || (sp->seg_flags & SEGM_CKP))
705 	{
706 		lfs_gather(fs, sp, vp, lfs_match_indir);
707 		lfs_gather(fs, sp, vp, lfs_match_dindir);
708 		lfs_gather(fs, sp, vp, lfs_match_tindir);
709 	}
710 	fip = sp->fip;
711 	if (fip->fi_nblocks != 0) {
712 		sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) +
713 				   sizeof(ufs_daddr_t) * (fip->fi_nblocks-1));
714 		sp->start_lbp = &sp->fip->fi_blocks[0];
715 	} else {
716 		sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t);
717 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
718 	}
719 }
720 
721 int
722 lfs_writeinode(fs, sp, ip)
723 	struct lfs *fs;
724 	struct segment *sp;
725 	struct inode *ip;
726 {
727 	struct buf *bp, *ibp;
728 	struct dinode *cdp;
729 	IFILE *ifp;
730 	SEGUSE *sup;
731 	ufs_daddr_t daddr;
732 	daddr_t *daddrp;
733 	ino_t ino;
734 	int error, i, ndx;
735 	int redo_ifile = 0;
736 	struct timespec ts;
737 	int gotblk=0;
738 
739 	if (!(ip->i_flag & IN_ALLMOD))
740 		return(0);
741 
742 	/* Allocate a new inode block if necessary. */
743 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp==NULL) && sp->ibp == NULL) {
744 		/* Allocate a new segment if necessary. */
745 		if (sp->seg_bytes_left < fs->lfs_bsize ||
746 		    sp->sum_bytes_left < sizeof(ufs_daddr_t))
747 			(void) lfs_writeseg(fs, sp);
748 
749 		/* Get next inode block. */
750 		daddr = fs->lfs_offset;
751 		fs->lfs_offset += fsbtodb(fs, 1);
752 		sp->ibp = *sp->cbpp++ =
753 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, daddr, fs->lfs_bsize, 0, 0);
754 		gotblk++;
755 
756 		/* Zero out inode numbers */
757 		for (i = 0; i < INOPB(fs); ++i)
758 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
759 
760 		++sp->start_bpp;
761 		fs->lfs_avail -= fsbtodb(fs, 1);
762 		/* Set remaining space counters. */
763 		sp->seg_bytes_left -= fs->lfs_bsize;
764 		sp->sum_bytes_left -= sizeof(ufs_daddr_t);
765 		ndx = LFS_SUMMARY_SIZE / sizeof(ufs_daddr_t) -
766 			sp->ninodes / INOPB(fs) - 1;
767 		((ufs_daddr_t *)(sp->segsum))[ndx] = daddr;
768 	}
769 
770 	/* Update the inode times and copy the inode onto the inode page. */
771 	TIMEVAL_TO_TIMESPEC(&time, &ts);
772 	LFS_ITIMES(ip, &ts, &ts, &ts);
773 
774 	/*
775 	 * If this is the Ifile, and we've already written the Ifile in this
776 	 * partial segment, just overwrite it (it's not on disk yet) and
777 	 * continue.
778 	 *
779 	 * XXX we know that the bp that we get the second time around has
780 	 * already been gathered.
781 	 */
782 	if(ip->i_number == LFS_IFILE_INUM && sp->idp) {
783 		*(sp->idp) = ip->i_din.ffs_din;
784 		return 0;
785 	}
786 
787 	bp = sp->ibp;
788 	cdp = ((struct dinode *)bp->b_data) + (sp->ninodes % INOPB(fs));
789 	*cdp = ip->i_din.ffs_din;
790 
791 	/*
792 	 * If we are cleaning, ensure that we don't write UNWRITTEN disk
793 	 * addresses to disk.
794 	 */
795 	if (ip->i_lfs_effnblks != ip->i_ffs_blocks) {
796 #ifdef DEBUG_LFS
797 		printf("lfs_writeinode: cleansing ino %d (%d != %d)\n",
798 		       ip->i_number, ip->i_lfs_effnblks, ip->i_ffs_blocks);
799 #endif
800 		for (daddrp = cdp->di_db; daddrp < cdp->di_ib + NIADDR;
801 		     daddrp++) {
802 			if (*daddrp == UNWRITTEN) {
803 #ifdef DEBUG_LFS
804 				printf("lfs_writeinode: wiping UNWRITTEN\n");
805 #endif
806 				*daddrp = 0;
807 			}
808 		}
809 	}
810 
811 	if(ip->i_flag & IN_CLEANING)
812 		LFS_CLR_UINO(ip, IN_CLEANING);
813 	else {
814 		/* XXX IN_ALLMOD */
815 		LFS_CLR_UINO(ip, IN_ACCESSED | IN_ACCESS | IN_CHANGE |
816 			     IN_UPDATE);
817 		if (ip->i_lfs_effnblks == ip->i_ffs_blocks)
818 			LFS_CLR_UINO(ip, IN_MODIFIED);
819 #ifdef DEBUG_LFS
820 		else
821 			printf("lfs_writeinode: ino %d: real blks=%d, "
822 			       "eff=%d\n", ip->i_number, ip->i_ffs_blocks,
823 			       ip->i_lfs_effnblks);
824 #endif
825 	}
826 
827 	if(ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
828 		sp->idp = ((struct dinode *)bp->b_data) +
829 			(sp->ninodes % INOPB(fs));
830 	if(gotblk) {
831 		LFS_LOCK_BUF(bp);
832 		brelse(bp);
833 	}
834 
835 	/* Increment inode count in segment summary block. */
836 	++((SEGSUM *)(sp->segsum))->ss_ninos;
837 
838 	/* If this page is full, set flag to allocate a new page. */
839 	if (++sp->ninodes % INOPB(fs) == 0)
840 		sp->ibp = NULL;
841 
842 	/*
843 	 * If updating the ifile, update the super-block.  Update the disk
844 	 * address and access times for this inode in the ifile.
845 	 */
846 	ino = ip->i_number;
847 	if (ino == LFS_IFILE_INUM) {
848 		daddr = fs->lfs_idaddr;
849 		fs->lfs_idaddr = bp->b_blkno;
850 	} else {
851 		LFS_IENTRY(ifp, fs, ino, ibp);
852 		daddr = ifp->if_daddr;
853 		ifp->if_daddr = bp->b_blkno;
854 #ifdef LFS_DEBUG_NEXTFREE
855 		if(ino > 3 && ifp->if_nextfree) {
856 			vprint("lfs_writeinode",ITOV(ip));
857 			printf("lfs_writeinode: updating free ino %d\n",
858 				ip->i_number);
859 		}
860 #endif
861 		error = VOP_BWRITE(ibp); /* Ifile */
862 	}
863 
864 	/*
865 	 * Account the inode: it no longer belongs to its former segment,
866 	 * though it will not belong to the new segment until that segment
867 	 * is actually written.
868 	 */
869 #ifdef DEBUG
870 	/*
871 	 * The inode's last address should not be in the current partial
872 	 * segment, except under exceptional circumstances (lfs_writevnodes
873 	 * had to start over, and in the meantime more blocks were written
874 	 * to a vnode).  Although the previous inode won't be accounted in
875 	 * su_nbytes until lfs_writeseg, this shouldn't be a problem as we
876 	 * have more data blocks in the current partial segment.
877 	 */
878 	if (daddr >= fs->lfs_lastpseg && daddr <= bp->b_blkno)
879 		printf("lfs_writeinode: last inode addr in current pseg "
880 		       "(ino %d daddr 0x%x)\n", ino, daddr);
881 #endif
882 	if (daddr != LFS_UNUSED_DADDR) {
883 		LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp);
884 #ifdef DIAGNOSTIC
885 		if (sup->su_nbytes < DINODE_SIZE) {
886 			printf("lfs_writeinode: negative bytes "
887 			       "(segment %d short by %d)\n",
888 			       datosn(fs, daddr),
889 			       (int)DINODE_SIZE - sup->su_nbytes);
890 			panic("lfs_writeinode: negative bytes");
891 			sup->su_nbytes = DINODE_SIZE;
892 		}
893 #endif
894 		sup->su_nbytes -= DINODE_SIZE;
895 		redo_ifile =
896 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
897 		error = VOP_BWRITE(bp); /* Ifile */
898 	}
899 	return (redo_ifile);
900 }
901 
902 int
903 lfs_gatherblock(sp, bp, sptr)
904 	struct segment *sp;
905 	struct buf *bp;
906 	int *sptr;
907 {
908 	struct lfs *fs;
909 	int version;
910 
911 	/*
912 	 * If full, finish this segment.  We may be doing I/O, so
913 	 * release and reacquire the splbio().
914 	 */
915 #ifdef DIAGNOSTIC
916 	if (sp->vp == NULL)
917 		panic ("lfs_gatherblock: Null vp in segment");
918 #endif
919 	fs = sp->fs;
920 	if (sp->sum_bytes_left < sizeof(ufs_daddr_t) ||
921 	    sp->seg_bytes_left < bp->b_bcount) {
922 		if (sptr)
923 			splx(*sptr);
924 		lfs_updatemeta(sp);
925 
926 		version = sp->fip->fi_version;
927 		(void) lfs_writeseg(fs, sp);
928 
929 		sp->fip->fi_version = version;
930 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
931 		/* Add the current file to the segment summary. */
932 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
933 		sp->sum_bytes_left -=
934 			sizeof(struct finfo) - sizeof(ufs_daddr_t);
935 
936 		if (sptr)
937 			*sptr = splbio();
938 		return(1);
939 	}
940 
941 #ifdef DEBUG
942 	if(bp->b_flags & B_GATHERED) {
943 		printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n",
944 		       sp->fip->fi_ino, bp->b_lblkno);
945 		return(0);
946 	}
947 #endif
948 	/* Insert into the buffer list, update the FINFO block. */
949 	bp->b_flags |= B_GATHERED;
950 	*sp->cbpp++ = bp;
951 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
952 
953 	sp->sum_bytes_left -= sizeof(ufs_daddr_t);
954 	sp->seg_bytes_left -= bp->b_bcount;
955 	return(0);
956 }
957 
958 int
959 lfs_gather(fs, sp, vp, match)
960 	struct lfs *fs;
961 	struct segment *sp;
962 	struct vnode *vp;
963 	int (*match) __P((struct lfs *, struct buf *));
964 {
965 	struct buf *bp;
966 	int s, count=0;
967 
968 	sp->vp = vp;
969 	s = splbio();
970 
971 #ifndef LFS_NO_BACKBUF_HACK
972 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = bp->b_vnbufs.le_next) {
973 #else /* LFS_NO_BACKBUF_HACK */
974 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
975 # define	BUF_OFFSET	(((void *)&bp->b_vnbufs.le_next) - (void *)bp)
976 # define	BACK_BUF(BP)	((struct buf *)(((void *)BP->b_vnbufs.le_prev) - BUF_OFFSET))
977 # define	BEG_OF_LIST	((struct buf *)(((void *)&vp->v_dirtyblkhd.lh_first) - BUF_OFFSET))
978 /* Find last buffer. */
979 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp && bp->b_vnbufs.le_next != NULL;
980 	    bp = bp->b_vnbufs.le_next);
981 	for (; bp && bp != BEG_OF_LIST; bp = BACK_BUF(bp)) {
982 #endif /* LFS_NO_BACKBUF_HACK */
983 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp))
984 			continue;
985 		if(vp->v_type == VBLK) {
986 			/* For block devices, just write the blocks. */
987 			/* XXX Do we really need to even do this? */
988 #ifdef DEBUG_LFS
989 			if(count==0)
990 				printf("BLK(");
991 			printf(".");
992 #endif
993 			/* Get the block before bwrite, so we don't corrupt the free list */
994 			bp->b_flags |= B_BUSY;
995 			bremfree(bp);
996 			bwrite(bp);
997 		} else {
998 #ifdef DIAGNOSTIC
999 			if ((bp->b_flags & (B_CALL|B_INVAL))==B_INVAL) {
1000 				printf("lfs_gather: lbn %d is B_INVAL\n",
1001 					bp->b_lblkno);
1002 				VOP_PRINT(bp->b_vp);
1003 			}
1004 			if (!(bp->b_flags & B_DELWRI))
1005 				panic("lfs_gather: bp not B_DELWRI");
1006 			if (!(bp->b_flags & B_LOCKED)) {
1007 				printf("lfs_gather: lbn %d blk %d"
1008 				       " not B_LOCKED\n", bp->b_lblkno,
1009 				       bp->b_blkno);
1010 				VOP_PRINT(bp->b_vp);
1011 				panic("lfs_gather: bp not B_LOCKED");
1012 			}
1013 #endif
1014 			if (lfs_gatherblock(sp, bp, &s)) {
1015 				goto loop;
1016 			}
1017 		}
1018 		count++;
1019 	}
1020 	splx(s);
1021 #ifdef DEBUG_LFS
1022 	if(vp->v_type == VBLK && count)
1023 		printf(")\n");
1024 #endif
1025 	lfs_updatemeta(sp);
1026 	sp->vp = NULL;
1027 	return count;
1028 }
1029 
1030 /*
1031  * Update the metadata that points to the blocks listed in the FINFO
1032  * array.
1033  */
1034 void
1035 lfs_updatemeta(sp)
1036 	struct segment *sp;
1037 {
1038 	SEGUSE *sup;
1039 	struct buf *bp;
1040 	struct lfs *fs;
1041 	struct vnode *vp;
1042 	struct indir a[NIADDR + 2], *ap;
1043 	struct inode *ip;
1044 	ufs_daddr_t daddr, lbn, off;
1045 	daddr_t ooff;
1046 	int error, i, nblocks, num;
1047 	int bb;
1048 
1049 	vp = sp->vp;
1050 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
1051 	if (nblocks < 0)
1052 		panic("This is a bad thing\n");
1053 	if (vp == NULL || nblocks == 0)
1054 		return;
1055 
1056 	/* Sort the blocks. */
1057 	/*
1058 	 * XXX KS - We have to sort even if the blocks come from the
1059 	 * cleaner, because there might be other pending blocks on the
1060 	 * same inode...and if we don't sort, and there are fragments
1061 	 * present, blocks may be written in the wrong place.
1062 	 */
1063 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
1064 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
1065 
1066 	/*
1067 	 * Record the length of the last block in case it's a fragment.
1068 	 * If there are indirect blocks present, they sort last.  An
1069 	 * indirect block will be lfs_bsize and its presence indicates
1070 	 * that you cannot have fragments.
1071 	 */
1072 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1073 
1074 	/*
1075 	 * Assign disk addresses, and update references to the logical
1076 	 * block and the segment usage information.
1077 	 */
1078 	fs = sp->fs;
1079 	for (i = nblocks; i--; ++sp->start_bpp) {
1080 		lbn = *sp->start_lbp++;
1081 
1082 		(*sp->start_bpp)->b_blkno = off = fs->lfs_offset;
1083 		if((*sp->start_bpp)->b_blkno == (*sp->start_bpp)->b_lblkno) {
1084 			printf("lfs_updatemeta: ino %d blk %d"
1085 			       " has same lbn and daddr\n",
1086 			       VTOI(vp)->i_number, off);
1087 		}
1088 		bb = fragstodb(fs, numfrags(fs, (*sp->start_bpp)->b_bcount));
1089 		fs->lfs_offset += bb;
1090 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1091 		if (error)
1092 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1093 		ip = VTOI(vp);
1094 		switch (num) {
1095 		case 0:
1096 			ooff = ip->i_ffs_db[lbn];
1097 #ifdef DEBUG
1098 			if (ooff == 0) {
1099 				printf("lfs_updatemeta[1]: warning: writing "
1100 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1101 				       ip->i_number, lbn, off);
1102 			}
1103 #endif
1104 			if (ooff == UNWRITTEN)
1105 				ip->i_ffs_blocks += bb;
1106 			ip->i_ffs_db[lbn] = off;
1107 			break;
1108 		case 1:
1109 			ooff = ip->i_ffs_ib[a[0].in_off];
1110 #ifdef DEBUG
1111 			if (ooff == 0) {
1112 				printf("lfs_updatemeta[2]: warning: writing "
1113 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1114 				       ip->i_number, lbn, off);
1115 			}
1116 #endif
1117 			if (ooff == UNWRITTEN)
1118 				ip->i_ffs_blocks += bb;
1119 			ip->i_ffs_ib[a[0].in_off] = off;
1120 			break;
1121 		default:
1122 			ap = &a[num - 1];
1123 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1124 				panic("lfs_updatemeta: bread bno %d",
1125 				      ap->in_lbn);
1126 
1127 			ooff = ((ufs_daddr_t *)bp->b_data)[ap->in_off];
1128 #if DEBUG
1129 			if (ooff == 0) {
1130 				printf("lfs_updatemeta[3]: warning: writing "
1131 				       "ino %d lbn %d at 0x%x, was 0x0\n",
1132 				       ip->i_number, lbn, off);
1133 			}
1134 #endif
1135 			if (ooff == UNWRITTEN)
1136 				ip->i_ffs_blocks += bb;
1137 			((ufs_daddr_t *)bp->b_data)[ap->in_off] = off;
1138 			(void) VOP_BWRITE(bp);
1139 		}
1140 #ifdef DEBUG
1141 		if (daddr >= fs->lfs_lastpseg && daddr <= off) {
1142 			printf("lfs_updatemeta: ino %d, lbn %d, addr = %x "
1143 			       "in same pseg\n", VTOI(sp->vp)->i_number,
1144 			       (*sp->start_bpp)->b_lblkno, daddr);
1145 		}
1146 #endif
1147 		/* Update segment usage information. */
1148 		if (daddr > 0) {
1149 			LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp);
1150 #ifdef DIAGNOSTIC
1151 			if (sup->su_nbytes < (*sp->start_bpp)->b_bcount) {
1152 				/* XXX -- Change to a panic. */
1153 				printf("lfs_updatemeta: negative bytes "
1154 				       "(segment %d short by %ld)\n",
1155 				       datosn(fs, daddr),
1156 				       (*sp->start_bpp)->b_bcount -
1157 				       sup->su_nbytes);
1158 				printf("lfs_updatemeta: ino %d, lbn %d, "
1159 				       "addr = %x\n", VTOI(sp->vp)->i_number,
1160 				       (*sp->start_bpp)->b_lblkno, daddr);
1161 				panic("lfs_updatemeta: negative bytes");
1162 				sup->su_nbytes = (*sp->start_bpp)->b_bcount;
1163 			}
1164 #endif
1165 			sup->su_nbytes -= (*sp->start_bpp)->b_bcount;
1166 			error = VOP_BWRITE(bp); /* Ifile */
1167 		}
1168 	}
1169 }
1170 
1171 /*
1172  * Start a new segment.
1173  */
1174 int
1175 lfs_initseg(fs)
1176 	struct lfs *fs;
1177 {
1178 	struct segment *sp;
1179 	SEGUSE *sup;
1180 	SEGSUM *ssp;
1181 	struct buf *bp;
1182 	int repeat;
1183 
1184 	sp = fs->lfs_sp;
1185 
1186 	repeat = 0;
1187 	/* Advance to the next segment. */
1188 	if (!LFS_PARTIAL_FITS(fs)) {
1189 		/* lfs_avail eats the remaining space */
1190 		fs->lfs_avail -= fs->lfs_dbpseg - (fs->lfs_offset -
1191 						   fs->lfs_curseg);
1192 		/* Wake up any cleaning procs waiting on this file system. */
1193 		wakeup(&lfs_allclean_wakeup);
1194 		wakeup(&fs->lfs_nextseg);
1195 		lfs_newseg(fs);
1196 		repeat = 1;
1197 		fs->lfs_offset = fs->lfs_curseg;
1198 		sp->seg_number = datosn(fs, fs->lfs_curseg);
1199 		sp->seg_bytes_left = dbtob(fs->lfs_dbpseg);
1200 		/*
1201 		 * If the segment contains a superblock, update the offset
1202 		 * and summary address to skip over it.
1203 		 */
1204 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1205 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1206 			fs->lfs_offset += btodb(LFS_SBPAD);
1207 			sp->seg_bytes_left -= LFS_SBPAD;
1208 		}
1209 		brelse(bp);
1210 	} else {
1211 		sp->seg_number = datosn(fs, fs->lfs_curseg);
1212 		sp->seg_bytes_left = dbtob(fs->lfs_dbpseg -
1213 				      (fs->lfs_offset - fs->lfs_curseg));
1214 	}
1215 	fs->lfs_lastpseg = fs->lfs_offset;
1216 
1217 	sp->fs = fs;
1218 	sp->ibp = NULL;
1219 	sp->idp = NULL;
1220 	sp->ninodes = 0;
1221 
1222 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1223 	sp->cbpp = sp->bpp;
1224 	*sp->cbpp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp,
1225 			       fs->lfs_offset, LFS_SUMMARY_SIZE);
1226 	sp->segsum = (*sp->cbpp)->b_data;
1227 	bzero(sp->segsum, LFS_SUMMARY_SIZE);
1228 	sp->start_bpp = ++sp->cbpp;
1229 	fs->lfs_offset += btodb(LFS_SUMMARY_SIZE);
1230 
1231 	/* Set point to SEGSUM, initialize it. */
1232 	ssp = sp->segsum;
1233 	ssp->ss_next = fs->lfs_nextseg;
1234 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1235 	ssp->ss_magic = SS_MAGIC;
1236 
1237 	/* Set pointer to first FINFO, initialize it. */
1238 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + sizeof(SEGSUM));
1239 	sp->fip->fi_nblocks = 0;
1240 	sp->start_lbp = &sp->fip->fi_blocks[0];
1241 	sp->fip->fi_lastlength = 0;
1242 
1243 	sp->seg_bytes_left -= LFS_SUMMARY_SIZE;
1244 	sp->sum_bytes_left = LFS_SUMMARY_SIZE - sizeof(SEGSUM);
1245 
1246 	return(repeat);
1247 }
1248 
1249 /*
1250  * Return the next segment to write.
1251  */
1252 void
1253 lfs_newseg(fs)
1254 	struct lfs *fs;
1255 {
1256 	CLEANERINFO *cip;
1257 	SEGUSE *sup;
1258 	struct buf *bp;
1259 	int curseg, isdirty, sn;
1260 
1261 	LFS_SEGENTRY(sup, fs, datosn(fs, fs->lfs_nextseg), bp);
1262 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1263 	sup->su_nbytes = 0;
1264 	sup->su_nsums = 0;
1265 	sup->su_ninos = 0;
1266 	(void) VOP_BWRITE(bp); /* Ifile */
1267 
1268 	LFS_CLEANERINFO(cip, fs, bp);
1269 	--cip->clean;
1270 	++cip->dirty;
1271 	fs->lfs_nclean = cip->clean;
1272 	LFS_SYNC_CLEANERINFO(cip, fs, bp, 1);
1273 
1274 	fs->lfs_lastseg = fs->lfs_curseg;
1275 	fs->lfs_curseg = fs->lfs_nextseg;
1276 	for (sn = curseg = datosn(fs, fs->lfs_curseg);;) {
1277 		sn = (sn + 1) % fs->lfs_nseg;
1278 		if (sn == curseg)
1279 			panic("lfs_nextseg: no clean segments");
1280 		LFS_SEGENTRY(sup, fs, sn, bp);
1281 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1282 		brelse(bp);
1283 		if (!isdirty)
1284 			break;
1285 	}
1286 
1287 	++fs->lfs_nactive;
1288 	fs->lfs_nextseg = sntoda(fs, sn);
1289 	if(lfs_dostats) {
1290 		++lfs_stats.segsused;
1291 	}
1292 }
1293 
1294 int
1295 lfs_writeseg(fs, sp)
1296 	struct lfs *fs;
1297 	struct segment *sp;
1298 {
1299 	struct buf **bpp, *bp, *cbp, *newbp;
1300 	SEGUSE *sup;
1301 	SEGSUM *ssp;
1302 	dev_t i_dev;
1303 	u_long *datap, *dp;
1304 	int do_again, i, nblocks, s;
1305 #ifdef LFS_TRACK_IOS
1306 	int j;
1307 #endif
1308 	int (*strategy)__P((void *));
1309 	struct vop_strategy_args vop_strategy_a;
1310 	u_short ninos;
1311 	struct vnode *devvp;
1312 	char *p;
1313 	struct vnode *vn;
1314 	struct inode *ip;
1315 	daddr_t *daddrp;
1316 	int changed;
1317 #if defined(DEBUG) && defined(LFS_PROPELLER)
1318 	static int propeller;
1319 	char propstring[4] = "-\\|/";
1320 
1321 	printf("%c\b",propstring[propeller++]);
1322 	if(propeller==4)
1323 		propeller = 0;
1324 #endif
1325 
1326 	/*
1327 	 * If there are no buffers other than the segment summary to write
1328 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1329 	 * even if there aren't any buffers, you need to write the superblock.
1330 	 */
1331 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1332 		return (0);
1333 
1334 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1335 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1336 
1337 	/* Update the segment usage information. */
1338 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1339 
1340 	/* Loop through all blocks, except the segment summary. */
1341 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1342 		if((*bpp)->b_vp != devvp)
1343 			sup->su_nbytes += (*bpp)->b_bcount;
1344 	}
1345 
1346 	ssp = (SEGSUM *)sp->segsum;
1347 
1348 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1349 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1350 	/* sup->su_nbytes += LFS_SUMMARY_SIZE; */
1351 	sup->su_lastmod = time.tv_sec;
1352 	sup->su_ninos += ninos;
1353 	++sup->su_nsums;
1354 	fs->lfs_dmeta += (btodb(LFS_SUMMARY_SIZE) + fsbtodb(fs, ninos));
1355 	fs->lfs_avail -= btodb(LFS_SUMMARY_SIZE);
1356 
1357 	do_again = !(bp->b_flags & B_GATHERED);
1358 	(void)VOP_BWRITE(bp); /* Ifile */
1359 	/*
1360 	 * Mark blocks B_BUSY, to prevent then from being changed between
1361 	 * the checksum computation and the actual write.
1362 	 *
1363 	 * If we are cleaning, check indirect blocks for UNWRITTEN, and if
1364 	 * there are any, replace them with copies that have UNASSIGNED
1365 	 * instead.
1366 	 */
1367 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1368 		++bpp;
1369 		if((*bpp)->b_flags & B_CALL)
1370 			continue;
1371 		bp = *bpp;
1372 	    again:
1373 		s = splbio();
1374 		if(bp->b_flags & B_BUSY) {
1375 #ifdef DEBUG
1376 			printf("lfs_writeseg: avoiding potential data "
1377 			       "summary corruption for ino %d, lbn %d\n",
1378 			       VTOI(bp->b_vp)->i_number, bp->b_lblkno);
1379 #endif
1380 			bp->b_flags |= B_WANTED;
1381 			tsleep(bp, (PRIBIO + 1), "lfs_writeseg", 0);
1382 			splx(s);
1383 			goto again;
1384 		}
1385 		bp->b_flags |= B_BUSY;
1386 		splx(s);
1387 		/* Check and replace indirect block UNWRITTEN bogosity */
1388 		if(bp->b_lblkno < 0 && bp->b_vp != devvp && bp->b_vp &&
1389 		   VTOI(bp->b_vp)->i_ffs_blocks !=
1390 		   VTOI(bp->b_vp)->i_lfs_effnblks) {
1391 #ifdef DEBUG_LFS
1392 			printf("lfs_writeseg: cleansing ino %d (%d != %d)\n",
1393 			       VTOI(bp->b_vp)->i_number,
1394 			       VTOI(bp->b_vp)->i_lfs_effnblks,
1395 			       VTOI(bp->b_vp)->i_ffs_blocks);
1396 #endif
1397 			/* Make a copy we'll make changes to */
1398 			newbp = lfs_newbuf(bp->b_vp, bp->b_lblkno,
1399 					   bp->b_bcount);
1400 			newbp->b_blkno = bp->b_blkno;
1401 			memcpy(newbp->b_data, bp->b_data,
1402 			       newbp->b_bcount);
1403 			*bpp = newbp;
1404 
1405 			changed = 0;
1406 			for (daddrp = (daddr_t *)(newbp->b_data);
1407 			     daddrp < (daddr_t *)(newbp->b_data +
1408 						  newbp->b_bcount); daddrp++) {
1409 				if (*daddrp == UNWRITTEN) {
1410 					++changed;
1411 #ifdef DEBUG_LFS
1412 					printf("lfs_writeseg: replacing UNWRITTEN\n");
1413 #endif
1414 					*daddrp = 0;
1415 				}
1416 			}
1417 			/*
1418 			 * Get rid of the old buffer.  Don't mark it clean,
1419 			 * though, if it still has dirty data on it.
1420 			 */
1421 			if (changed) {
1422 				bp->b_flags &= ~(B_ERROR | B_GATHERED);
1423 				if (bp->b_flags & B_CALL)
1424 					lfs_freebuf(bp);
1425 				else {
1426 					/* Still on free list, leave it there */
1427 					s = splbio();
1428 					bp->b_flags &= ~B_BUSY;
1429 					if (bp->b_flags & B_WANTED)
1430 						wakeup(bp);
1431 				 	splx(s);
1432 					/*
1433 					 * We have to re-decrement lfs_avail
1434 					 * since this block is going to come
1435 					 * back around to us in the next
1436 					 * segment.
1437 					 */
1438 					fs->lfs_avail -= btodb(bp->b_bcount);
1439 				}
1440 			} else {
1441 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1442 						 B_GATHERED);
1443 				LFS_UNLOCK_BUF(bp);
1444 				if (bp->b_flags & B_CALL)
1445 					lfs_freebuf(bp);
1446 				else {
1447 					bremfree(bp);
1448 					bp->b_flags |= B_DONE;
1449 					reassignbuf(bp, bp->b_vp);
1450 					brelse(bp);
1451 				}
1452 			}
1453 
1454 		}
1455 	}
1456 	/*
1457 	 * Compute checksum across data and then across summary; the first
1458 	 * block (the summary block) is skipped.  Set the create time here
1459 	 * so that it's guaranteed to be later than the inode mod times.
1460 	 *
1461 	 * XXX
1462 	 * Fix this to do it inline, instead of malloc/copy.
1463 	 */
1464 	datap = dp = malloc(nblocks * sizeof(u_long), M_SEGMENT, M_WAITOK);
1465 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1466 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1467 			if (copyin((*bpp)->b_saveaddr, dp++, sizeof(u_long)))
1468 				panic("lfs_writeseg: copyin failed [1]: "
1469 				      "ino %d blk %d",
1470 				      VTOI((*bpp)->b_vp)->i_number,
1471 				      (*bpp)->b_lblkno);
1472 		} else
1473 			*dp++ = ((u_long *)(*bpp)->b_data)[0];
1474 	}
1475 	ssp->ss_create = time.tv_sec;
1476 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * sizeof(u_long));
1477 	ssp->ss_sumsum =
1478 	    cksum(&ssp->ss_datasum, LFS_SUMMARY_SIZE - sizeof(ssp->ss_sumsum));
1479 	free(datap, M_SEGMENT);
1480 
1481 	fs->lfs_bfree -= (fsbtodb(fs, ninos) + btodb(LFS_SUMMARY_SIZE));
1482 
1483 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1484 
1485 	/*
1486 	 * When we simply write the blocks we lose a rotation for every block
1487 	 * written.  To avoid this problem, we allocate memory in chunks, copy
1488 	 * the buffers into the chunk and write the chunk.  CHUNKSIZE is the
1489 	 * largest size I/O devices can handle.
1490 	 * When the data is copied to the chunk, turn off the B_LOCKED bit
1491 	 * and brelse the buffer (which will move them to the LRU list).  Add
1492 	 * the B_CALL flag to the buffer header so we can count I/O's for the
1493 	 * checkpoints and so we can release the allocated memory.
1494 	 *
1495 	 * XXX
1496 	 * This should be removed if the new virtual memory system allows us to
1497 	 * easily make the buffers contiguous in kernel memory and if that's
1498 	 * fast enough.
1499 	 */
1500 
1501 #define CHUNKSIZE MAXPHYS
1502 
1503 	if(devvp==NULL)
1504 		panic("devvp is NULL");
1505 	for (bpp = sp->bpp,i = nblocks; i;) {
1506 		cbp = lfs_newbuf(devvp, (*bpp)->b_blkno, CHUNKSIZE);
1507 		cbp->b_dev = i_dev;
1508 		cbp->b_flags |= B_ASYNC | B_BUSY;
1509 		cbp->b_bcount = 0;
1510 
1511 #ifdef DIAGNOSTIC
1512 		if(datosn(fs, (*bpp)->b_blkno + btodb((*bpp)->b_bcount) - 1) !=
1513 		   datosn(fs, cbp->b_blkno)) {
1514 			panic("lfs_writeseg: Segment overwrite");
1515 		}
1516 #endif
1517 
1518 		s = splbio();
1519 		if(fs->lfs_iocount >= LFS_THROTTLE) {
1520 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs throttle", 0);
1521 		}
1522 		++fs->lfs_iocount;
1523 #ifdef LFS_TRACK_IOS
1524 		for(j=0;j<LFS_THROTTLE;j++) {
1525 			if(fs->lfs_pending[j]==LFS_UNUSED_DADDR) {
1526 				fs->lfs_pending[j] = cbp->b_blkno;
1527 				break;
1528 			}
1529 		}
1530 #endif /* LFS_TRACK_IOS */
1531 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1532 			bp = *bpp;
1533 
1534 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1535 				break;
1536 
1537 			/*
1538 			 * Fake buffers from the cleaner are marked as B_INVAL.
1539 			 * We need to copy the data from user space rather than
1540 			 * from the buffer indicated.
1541 			 * XXX == what do I do on an error?
1542 			 */
1543 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1544 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1545 					panic("lfs_writeseg: copyin failed [2]");
1546 			} else
1547 				bcopy(bp->b_data, p, bp->b_bcount);
1548 			p += bp->b_bcount;
1549 			cbp->b_bcount += bp->b_bcount;
1550 			LFS_UNLOCK_BUF(bp);
1551 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1552 					 B_GATHERED);
1553 			vn = bp->b_vp;
1554 			if (bp->b_flags & B_CALL) {
1555 				/* if B_CALL, it was created with newbuf */
1556 				lfs_freebuf(bp);
1557 			} else {
1558 				bremfree(bp);
1559 				bp->b_flags |= B_DONE;
1560 				if(vn)
1561 					reassignbuf(bp, vn);
1562 				brelse(bp);
1563 			}
1564 
1565 			bpp++;
1566 
1567 			/*
1568 			 * If this is the last block for this vnode, but
1569 			 * there are other blocks on its dirty list,
1570 			 * set IN_MODIFIED/IN_CLEANING depending on what
1571 			 * sort of block.  Only do this for our mount point,
1572 			 * not for, e.g., inode blocks that are attached to
1573 			 * the devvp.
1574 			 */
1575 			if(i>1 && vn && *bpp && (*bpp)->b_vp != vn
1576 			   && (*bpp)->b_vp && (bp=vn->v_dirtyblkhd.lh_first)!=NULL &&
1577 			   vn->v_mount == fs->lfs_ivnode->v_mount)
1578 			{
1579 				ip = VTOI(vn);
1580 #ifdef DEBUG_LFS
1581 				printf("lfs_writeseg: marking ino %d\n",ip->i_number);
1582 #endif
1583 				if(bp->b_flags & B_CALL)
1584 					LFS_SET_UINO(ip, IN_CLEANING);
1585 				else
1586 					LFS_SET_UINO(ip, IN_MODIFIED);
1587 			}
1588 			/* if(vn->v_dirtyblkhd.lh_first == NULL) */
1589 				wakeup(vn);
1590 		}
1591 		++cbp->b_vp->v_numoutput;
1592 		splx(s);
1593 		/*
1594 		 * XXXX This is a gross and disgusting hack.  Since these
1595 		 * buffers are physically addressed, they hang off the
1596 		 * device vnode (devvp).  As a result, they have no way
1597 		 * of getting to the LFS superblock or lfs structure to
1598 		 * keep track of the number of I/O's pending.  So, I am
1599 		 * going to stuff the fs into the saveaddr field of
1600 		 * the buffer (yuk).
1601 		 */
1602 		cbp->b_saveaddr = (caddr_t)fs;
1603 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1604 		vop_strategy_a.a_bp = cbp;
1605 		(strategy)(&vop_strategy_a);
1606 	}
1607 #if 1 || defined(DEBUG)
1608 	/*
1609 	 * After doing a big write, we recalculate how many buffers are
1610 	 * really still left on the locked queue.
1611 	 */
1612 	s = splbio();
1613 	lfs_countlocked(&locked_queue_count, &locked_queue_bytes);
1614 	splx(s);
1615 	wakeup(&locked_queue_count);
1616 #endif /* 1 || DEBUG */
1617 	if(lfs_dostats) {
1618 		++lfs_stats.psegwrites;
1619 		lfs_stats.blocktot += nblocks - 1;
1620 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1621 			++lfs_stats.psyncwrites;
1622 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1623 			++lfs_stats.pcleanwrites;
1624 			lfs_stats.cleanblocks += nblocks - 1;
1625 		}
1626 	}
1627 	return (lfs_initseg(fs) || do_again);
1628 }
1629 
1630 void
1631 lfs_writesuper(fs, daddr)
1632 	struct lfs *fs;
1633 	daddr_t daddr;
1634 {
1635 	struct buf *bp;
1636 	dev_t i_dev;
1637 	int (*strategy) __P((void *));
1638 	int s;
1639 	struct vop_strategy_args vop_strategy_a;
1640 
1641 #ifdef LFS_CANNOT_ROLLFW
1642 	/*
1643 	 * If we can write one superblock while another is in
1644 	 * progress, we risk not having a complete checkpoint if we crash.
1645 	 * So, block here if a superblock write is in progress.
1646 	 */
1647 	s = splbio();
1648 	while(fs->lfs_sbactive) {
1649 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
1650 	}
1651 	fs->lfs_sbactive = daddr;
1652 	splx(s);
1653 #endif
1654 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1655 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
1656 
1657 	/* Set timestamp of this version of the superblock */
1658 	fs->lfs_tstamp = time.tv_sec;
1659 
1660 	/* Checksum the superblock and copy it into a buffer. */
1661 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
1662 	bp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, daddr, LFS_SBPAD);
1663 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
1664 
1665 	bp->b_dev = i_dev;
1666 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
1667 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
1668 	bp->b_iodone = lfs_supercallback;
1669 	/* XXX KS - same nasty hack as above */
1670 	bp->b_saveaddr = (caddr_t)fs;
1671 
1672 	vop_strategy_a.a_desc = VDESC(vop_strategy);
1673 	vop_strategy_a.a_bp = bp;
1674 	s = splbio();
1675 	++bp->b_vp->v_numoutput;
1676 	++fs->lfs_iocount;
1677 	splx(s);
1678 	(strategy)(&vop_strategy_a);
1679 }
1680 
1681 /*
1682  * Logical block number match routines used when traversing the dirty block
1683  * chain.
1684  */
1685 int
1686 lfs_match_fake(fs, bp)
1687 	struct lfs *fs;
1688 	struct buf *bp;
1689 {
1690 	return (bp->b_flags & B_CALL);
1691 }
1692 
1693 int
1694 lfs_match_data(fs, bp)
1695 	struct lfs *fs;
1696 	struct buf *bp;
1697 {
1698 	return (bp->b_lblkno >= 0);
1699 }
1700 
1701 int
1702 lfs_match_indir(fs, bp)
1703 	struct lfs *fs;
1704 	struct buf *bp;
1705 {
1706 	int lbn;
1707 
1708 	lbn = bp->b_lblkno;
1709 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
1710 }
1711 
1712 int
1713 lfs_match_dindir(fs, bp)
1714 	struct lfs *fs;
1715 	struct buf *bp;
1716 {
1717 	int lbn;
1718 
1719 	lbn = bp->b_lblkno;
1720 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
1721 }
1722 
1723 int
1724 lfs_match_tindir(fs, bp)
1725 	struct lfs *fs;
1726 	struct buf *bp;
1727 {
1728 	int lbn;
1729 
1730 	lbn = bp->b_lblkno;
1731 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
1732 }
1733 
1734 /*
1735  * XXX - The only buffers that are going to hit these functions are the
1736  * segment write blocks, or the segment summaries, or the superblocks.
1737  *
1738  * All of the above are created by lfs_newbuf, and so do not need to be
1739  * released via brelse.
1740  */
1741 void
1742 lfs_callback(bp)
1743 	struct buf *bp;
1744 {
1745 	struct lfs *fs;
1746 #ifdef LFS_TRACK_IOS
1747 	int j;
1748 #endif
1749 
1750 	fs = (struct lfs *)bp->b_saveaddr;
1751 #ifdef DIAGNOSTIC
1752 	if (fs->lfs_iocount == 0)
1753 		panic("lfs_callback: zero iocount\n");
1754 #endif
1755 	if (--fs->lfs_iocount < LFS_THROTTLE)
1756 		wakeup(&fs->lfs_iocount);
1757 #ifdef LFS_TRACK_IOS
1758 	for(j=0;j<LFS_THROTTLE;j++) {
1759 		if(fs->lfs_pending[j]==bp->b_blkno) {
1760 			fs->lfs_pending[j] = LFS_UNUSED_DADDR;
1761 			wakeup(&(fs->lfs_pending[j]));
1762 			break;
1763 		}
1764 	}
1765 #endif /* LFS_TRACK_IOS */
1766 
1767 	lfs_freebuf(bp);
1768 }
1769 
1770 void
1771 lfs_supercallback(bp)
1772 	struct buf *bp;
1773 {
1774 	struct lfs *fs;
1775 
1776 	fs = (struct lfs *)bp->b_saveaddr;
1777 #ifdef LFS_CANNOT_ROLLFW
1778 	fs->lfs_sbactive = 0;
1779 	wakeup(&fs->lfs_sbactive);
1780 #endif
1781 	if (--fs->lfs_iocount < LFS_THROTTLE)
1782 		wakeup(&fs->lfs_iocount);
1783 	lfs_freebuf(bp);
1784 }
1785 
1786 /*
1787  * Shellsort (diminishing increment sort) from Data Structures and
1788  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
1789  * see also Knuth Vol. 3, page 84.  The increments are selected from
1790  * formula (8), page 95.  Roughly O(N^3/2).
1791  */
1792 /*
1793  * This is our own private copy of shellsort because we want to sort
1794  * two parallel arrays (the array of buffer pointers and the array of
1795  * logical block numbers) simultaneously.  Note that we cast the array
1796  * of logical block numbers to a unsigned in this routine so that the
1797  * negative block numbers (meta data blocks) sort AFTER the data blocks.
1798  */
1799 
1800 void
1801 lfs_shellsort(bp_array, lb_array, nmemb)
1802 	struct buf **bp_array;
1803 	ufs_daddr_t *lb_array;
1804 	int nmemb;
1805 {
1806 	static int __rsshell_increments[] = { 4, 1, 0 };
1807 	int incr, *incrp, t1, t2;
1808 	struct buf *bp_temp;
1809 	u_long lb_temp;
1810 
1811 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
1812 		for (t1 = incr; t1 < nmemb; ++t1)
1813 			for (t2 = t1 - incr; t2 >= 0;)
1814 				if (lb_array[t2] > lb_array[t2 + incr]) {
1815 					lb_temp = lb_array[t2];
1816 					lb_array[t2] = lb_array[t2 + incr];
1817 					lb_array[t2 + incr] = lb_temp;
1818 					bp_temp = bp_array[t2];
1819 					bp_array[t2] = bp_array[t2 + incr];
1820 					bp_array[t2 + incr] = bp_temp;
1821 					t2 -= incr;
1822 				} else
1823 					break;
1824 }
1825 
1826 /*
1827  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
1828  */
1829 int
1830 lfs_vref(vp)
1831 	struct vnode *vp;
1832 {
1833 	/*
1834 	 * If we return 1 here during a flush, we risk vinvalbuf() not
1835 	 * being able to flush all of the pages from this vnode, which
1836 	 * will cause it to panic.  So, return 0 if a flush is in progress.
1837 	 */
1838 	if (vp->v_flag & VXLOCK) {
1839 		if(IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1840 			return 0;
1841 		}
1842 		return(1);
1843 	}
1844 	return (vget(vp, 0));
1845 }
1846 
1847 /*
1848  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
1849  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
1850  */
1851 void
1852 lfs_vunref(vp)
1853 	struct vnode *vp;
1854 {
1855 	/*
1856 	 * Analogous to lfs_vref, if the node is flushing, fake it.
1857 	 */
1858 	if((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1859 		return;
1860 	}
1861 
1862 	simple_lock(&vp->v_interlock);
1863 #ifdef DIAGNOSTIC
1864 	if(vp->v_usecount<=0) {
1865 		printf("lfs_vunref: inum is %d\n", VTOI(vp)->i_number);
1866 		printf("lfs_vunref: flags are 0x%x\n", vp->v_flag);
1867 		printf("lfs_vunref: usecount = %d\n", vp->v_usecount);
1868 		panic("lfs_vunref: v_usecount<0");
1869 	}
1870 #endif
1871 	vp->v_usecount--;
1872 	if (vp->v_usecount > 0) {
1873 		simple_unlock(&vp->v_interlock);
1874 		return;
1875 	}
1876 	/*
1877 	 * insert at tail of LRU list
1878 	 */
1879 	simple_lock(&vnode_free_list_slock);
1880 	if (vp->v_holdcnt > 0)
1881 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
1882 	else
1883 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
1884 	simple_unlock(&vnode_free_list_slock);
1885 	simple_unlock(&vp->v_interlock);
1886 }
1887 
1888 /*
1889  * We use this when we have vnodes that were loaded in solely for cleaning.
1890  * There is no reason to believe that these vnodes will be referenced again
1891  * soon, since the cleaning process is unrelated to normal filesystem
1892  * activity.  Putting cleaned vnodes at the tail of the list has the effect
1893  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
1894  * cleaning at the head of the list, instead.
1895  */
1896 void
1897 lfs_vunref_head(vp)
1898 	struct vnode *vp;
1899 {
1900 	simple_lock(&vp->v_interlock);
1901 #ifdef DIAGNOSTIC
1902 	if(vp->v_usecount==0) {
1903 		panic("lfs_vunref: v_usecount<0");
1904 	}
1905 #endif
1906 	vp->v_usecount--;
1907 	if (vp->v_usecount > 0) {
1908 		simple_unlock(&vp->v_interlock);
1909 		return;
1910 	}
1911 	/*
1912 	 * insert at head of LRU list
1913 	 */
1914 	simple_lock(&vnode_free_list_slock);
1915 	TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
1916 	simple_unlock(&vnode_free_list_slock);
1917 	simple_unlock(&vp->v_interlock);
1918 }
1919 
1920