xref: /netbsd-src/sys/ufs/lfs/lfs_segment.c (revision 4472dbe5e3bd91ef2540bada7a7ca7384627ff9b)
1 /*	$NetBSD: lfs_segment.c,v 1.48 2000/05/31 03:37:35 fredb Exp $	*/
2 
3 /*-
4  * Copyright (c) 1999 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 #include "opt_ddb.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/namei.h>
79 #include <sys/kernel.h>
80 #include <sys/resourcevar.h>
81 #include <sys/file.h>
82 #include <sys/stat.h>
83 #include <sys/buf.h>
84 #include <sys/proc.h>
85 #include <sys/conf.h>
86 #include <sys/vnode.h>
87 #include <sys/malloc.h>
88 #include <sys/mount.h>
89 
90 #include <miscfs/specfs/specdev.h>
91 #include <miscfs/fifofs/fifo.h>
92 
93 #include <ufs/ufs/quota.h>
94 #include <ufs/ufs/inode.h>
95 #include <ufs/ufs/dir.h>
96 #include <ufs/ufs/ufsmount.h>
97 #include <ufs/ufs/ufs_extern.h>
98 
99 #include <ufs/lfs/lfs.h>
100 #include <ufs/lfs/lfs_extern.h>
101 
102 extern int count_lock_queue __P((void));
103 extern struct simplelock vnode_free_list_slock;		/* XXX */
104 
105 /*
106  * Determine if it's OK to start a partial in this segment, or if we need
107  * to go on to a new segment.
108  */
109 #define	LFS_PARTIAL_FITS(fs) \
110 	((fs)->lfs_dbpseg - ((fs)->lfs_offset - (fs)->lfs_curseg) > \
111 	1 << (fs)->lfs_fsbtodb)
112 
113 void	 lfs_callback __P((struct buf *));
114 int	 lfs_gather __P((struct lfs *, struct segment *,
115 	     struct vnode *, int (*) __P((struct lfs *, struct buf *))));
116 int	 lfs_gatherblock __P((struct segment *, struct buf *, int *));
117 void	 lfs_iset __P((struct inode *, ufs_daddr_t, time_t));
118 int	 lfs_match_fake __P((struct lfs *, struct buf *));
119 int	 lfs_match_data __P((struct lfs *, struct buf *));
120 int	 lfs_match_dindir __P((struct lfs *, struct buf *));
121 int	 lfs_match_indir __P((struct lfs *, struct buf *));
122 int	 lfs_match_tindir __P((struct lfs *, struct buf *));
123 void	 lfs_newseg __P((struct lfs *));
124 void	 lfs_shellsort __P((struct buf **, ufs_daddr_t *, int));
125 void	 lfs_supercallback __P((struct buf *));
126 void	 lfs_updatemeta __P((struct segment *));
127 int	 lfs_vref __P((struct vnode *));
128 void	 lfs_vunref __P((struct vnode *));
129 void	 lfs_writefile __P((struct lfs *, struct segment *, struct vnode *));
130 int	 lfs_writeinode __P((struct lfs *, struct segment *, struct inode *));
131 int	 lfs_writeseg __P((struct lfs *, struct segment *));
132 void	 lfs_writesuper __P((struct lfs *, daddr_t));
133 int	 lfs_writevnodes __P((struct lfs *fs, struct mount *mp,
134 	    struct segment *sp, int dirops));
135 
136 int	lfs_allclean_wakeup;		/* Cleaner wakeup address. */
137 int	lfs_writeindir = 1;             /* whether to flush indir on non-ckp */
138 int	lfs_clean_vnhead = 0;		/* Allow freeing to head of vn list */
139 int	lfs_dirvcount = 0;		/* # active dirops */
140 
141 /* Statistics Counters */
142 int lfs_dostats = 1;
143 struct lfs_stats lfs_stats;
144 
145 /* op values to lfs_writevnodes */
146 #define	VN_REG	        0
147 #define	VN_DIROP	1
148 #define	VN_EMPTY	2
149 #define VN_CLEAN        3
150 
151 #define LFS_MAX_ACTIVE          10
152 
153 /*
154  * XXX KS - Set modification time on the Ifile, so the cleaner can
155  * read the fs mod time off of it.  We don't set IN_UPDATE here,
156  * since we don't really need this to be flushed to disk (and in any
157  * case that wouldn't happen to the Ifile until we checkpoint).
158  */
159 void
160 lfs_imtime(fs)
161 	struct lfs *fs;
162 {
163 	struct timespec ts;
164 	struct inode *ip;
165 
166 	TIMEVAL_TO_TIMESPEC(&time, &ts);
167 	ip = VTOI(fs->lfs_ivnode);
168 	ip->i_ffs_mtime = ts.tv_sec;
169 	ip->i_ffs_mtimensec = ts.tv_nsec;
170 }
171 
172 /*
173  * Ifile and meta data blocks are not marked busy, so segment writes MUST be
174  * single threaded.  Currently, there are two paths into lfs_segwrite, sync()
175  * and getnewbuf().  They both mark the file system busy.  Lfs_vflush()
176  * explicitly marks the file system busy.  So lfs_segwrite is safe.  I think.
177  */
178 
179 #define SET_FLUSHING(fs,vp) (fs)->lfs_flushvp = (vp)
180 #define IS_FLUSHING(fs,vp)  ((fs)->lfs_flushvp == (vp))
181 #define CLR_FLUSHING(fs,vp) (fs)->lfs_flushvp = NULL
182 
183 int
184 lfs_vflush(vp)
185 	struct vnode *vp;
186 {
187 	struct inode *ip;
188 	struct lfs *fs;
189 	struct segment *sp;
190 	struct buf *bp, *nbp, *tbp, *tnbp;
191 	int error, s;
192 
193 	ip = VTOI(vp);
194 	fs = VFSTOUFS(vp->v_mount)->um_lfs;
195 
196 	if(ip->i_flag & IN_CLEANING) {
197 #ifdef DEBUG_LFS
198 		ivndebug(vp,"vflush/in_cleaning");
199 #endif
200 		ip->i_flag &= ~IN_CLEANING;
201 		if(ip->i_flag & (IN_MODIFIED | IN_ACCESSED)) {
202 			fs->lfs_uinodes--;
203 		} else
204 			ip->i_flag |= IN_MODIFIED;
205 		/*
206 		 * Toss any cleaning buffers that have real counterparts
207 		 * to avoid losing new data
208 		 */
209 		s = splbio();
210 		for(bp=vp->v_dirtyblkhd.lh_first; bp; bp=nbp) {
211 			nbp = bp->b_vnbufs.le_next;
212 			if(bp->b_flags & B_CALL) {
213 				for(tbp=vp->v_dirtyblkhd.lh_first; tbp;
214 				    tbp=tnbp)
215 				{
216 					tnbp = tbp->b_vnbufs.le_next;
217 					if(tbp->b_vp == bp->b_vp
218 					   && tbp->b_lblkno == bp->b_lblkno
219 					   && tbp != bp)
220 					{
221 						lfs_freebuf(bp);
222 					}
223 				}
224 			}
225 		}
226 		splx(s);
227 	}
228 
229 	/* If the node is being written, wait until that is done */
230 	if(WRITEINPROG(vp)) {
231 #ifdef DEBUG_LFS
232 		ivndebug(vp,"vflush/writeinprog");
233 #endif
234 		tsleep(vp, PRIBIO+1, "lfs_vw", 0);
235 	}
236 
237 	/* Protect against VXLOCK deadlock in vinvalbuf() */
238 	lfs_seglock(fs, SEGM_SYNC);
239 
240 	/* If we're supposed to flush a freed inode, just toss it */
241 	/* XXX - seglock, so these buffers can't be gathered, right? */
242 	if(ip->i_ffs_mode == 0) {
243 		printf("lfs_vflush: ino %d is freed, not flushing\n",
244 			ip->i_number);
245 		s = splbio();
246 		for(bp=vp->v_dirtyblkhd.lh_first; bp; bp=nbp) {
247 			nbp = bp->b_vnbufs.le_next;
248 			/* Copied from lfs_writeseg */
249 			if (bp->b_flags & B_CALL) {
250 				/* if B_CALL, it was created with newbuf */
251 				lfs_freebuf(bp);
252 			} else {
253 				bremfree(bp);
254 				bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
255                                          B_LOCKED | B_GATHERED);
256 				bp->b_flags |= B_DONE;
257 				reassignbuf(bp, vp);
258 				brelse(bp);
259 			}
260 		}
261 		splx(s);
262 		if(ip->i_flag & IN_CLEANING)
263 			fs->lfs_uinodes--;
264 		if(ip->i_flag & (IN_MODIFIED | IN_ACCESSED))
265 			fs->lfs_uinodes--;
266 		ip->i_flag &= ~IN_ALLMOD;
267 		printf("lfs_vflush: done not flushing ino %d\n",
268 			ip->i_number);
269 		lfs_segunlock(fs);
270 		return 0;
271 	}
272 
273 	SET_FLUSHING(fs,vp);
274 	if (fs->lfs_nactive > LFS_MAX_ACTIVE) {
275 		error = lfs_segwrite(vp->v_mount, SEGM_SYNC|SEGM_CKP);
276 		CLR_FLUSHING(fs,vp);
277 		lfs_segunlock(fs);
278 		return error;
279 	}
280 	sp = fs->lfs_sp;
281 
282 	if (vp->v_dirtyblkhd.lh_first == NULL) {
283 		lfs_writevnodes(fs, vp->v_mount, sp, VN_EMPTY);
284 	} else if((ip->i_flag & IN_CLEANING) && (fs->lfs_sp->seg_flags & SEGM_CLEAN)) {
285 #ifdef DEBUG_LFS
286 		ivndebug(vp,"vflush/clean");
287 #endif
288 		lfs_writevnodes(fs, vp->v_mount, sp, VN_CLEAN);
289 	}
290 	else if(lfs_dostats) {
291 		if(vp->v_dirtyblkhd.lh_first || (VTOI(vp)->i_flag & IN_ALLMOD))
292 			++lfs_stats.vflush_invoked;
293 #ifdef DEBUG_LFS
294 		ivndebug(vp,"vflush");
295 #endif
296 	}
297 
298 #ifdef DIAGNOSTIC
299 	/* XXX KS This actually can happen right now, though it shouldn't(?) */
300 	if(vp->v_flag & VDIROP) {
301 		printf("lfs_vflush: flushing VDIROP, this shouldn\'t be\n");
302 		/* panic("VDIROP being flushed...this can\'t happen"); */
303 	}
304 	if(vp->v_usecount<0) {
305 		printf("usecount=%ld\n",vp->v_usecount);
306 		panic("lfs_vflush: usecount<0");
307 	}
308 #endif
309 
310 	do {
311 		do {
312 			if (vp->v_dirtyblkhd.lh_first != NULL)
313 				lfs_writefile(fs, sp, vp);
314 		} while (lfs_writeinode(fs, sp, ip));
315 	} while (lfs_writeseg(fs, sp) && ip->i_number == LFS_IFILE_INUM);
316 
317 	if(lfs_dostats) {
318 		++lfs_stats.nwrites;
319 		if (sp->seg_flags & SEGM_SYNC)
320 			++lfs_stats.nsync_writes;
321 		if (sp->seg_flags & SEGM_CKP)
322 			++lfs_stats.ncheckpoints;
323 	}
324 	lfs_segunlock(fs);
325 
326 	CLR_FLUSHING(fs,vp);
327 	return (0);
328 }
329 
330 #ifdef DEBUG_LFS_VERBOSE
331 # 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)
332 #else
333 # define vndebug(vp,str)
334 #endif
335 
336 int
337 lfs_writevnodes(fs, mp, sp, op)
338 	struct lfs *fs;
339 	struct mount *mp;
340 	struct segment *sp;
341 	int op;
342 {
343 	struct inode *ip;
344 	struct vnode *vp;
345 	int inodes_written=0, only_cleaning;
346 	int needs_unlock;
347 
348 #ifndef LFS_NO_BACKVP_HACK
349 	/* BEGIN HACK */
350 #define	VN_OFFSET	(((caddr_t)&vp->v_mntvnodes.le_next) - (caddr_t)vp)
351 #define	BACK_VP(VP)	((struct vnode *)(((caddr_t)VP->v_mntvnodes.le_prev) - VN_OFFSET))
352 #define	BEG_OF_VLIST	((struct vnode *)(((caddr_t)&mp->mnt_vnodelist.lh_first) - VN_OFFSET))
353 
354 	/* Find last vnode. */
355  loop:	for (vp = mp->mnt_vnodelist.lh_first;
356 	     vp && vp->v_mntvnodes.le_next != NULL;
357 	     vp = vp->v_mntvnodes.le_next);
358 	for (; vp && vp != BEG_OF_VLIST; vp = BACK_VP(vp)) {
359 #else
360 	loop:
361 	for (vp = mp->mnt_vnodelist.lh_first;
362 	     vp != NULL;
363 	     vp = vp->v_mntvnodes.le_next) {
364 #endif
365 		/*
366 		 * If the vnode that we are about to sync is no longer
367 		 * associated with this mount point, start over.
368 		 */
369 		if (vp->v_mount != mp)
370 			goto loop;
371 
372 		ip = VTOI(vp);
373 		if ((op == VN_DIROP && !(vp->v_flag & VDIROP)) ||
374 		    (op != VN_DIROP && op != VN_CLEAN && (vp->v_flag & VDIROP))) {
375 			vndebug(vp,"dirop");
376 			continue;
377 		}
378 
379 		if (op == VN_EMPTY && vp->v_dirtyblkhd.lh_first) {
380 			vndebug(vp,"empty");
381 			continue;
382 		}
383 
384 		if (vp->v_type == VNON) {
385 			continue;
386 		}
387 
388 		if(op == VN_CLEAN && ip->i_number != LFS_IFILE_INUM
389 		   && vp != fs->lfs_flushvp
390 		   && !(ip->i_flag & IN_CLEANING)) {
391 			vndebug(vp,"cleaning");
392 			continue;
393 		}
394 
395 		if (lfs_vref(vp)) {
396 			vndebug(vp,"vref");
397 			continue;
398 		}
399 
400 #if 0 /* XXX KS - if we skip the ifile, things could go badly for us. */
401 		if(WRITEINPROG(vp)) {
402 			lfs_vunref(vp);
403 #ifdef DEBUG_LFS
404 			ivndebug(vp,"writevnodes/writeinprog");
405 #endif
406 			continue;
407 		}
408 #endif
409 
410 		needs_unlock = 0;
411 		if(VOP_ISLOCKED(vp)) {
412 			if (vp != fs->lfs_ivnode &&
413 			    vp->v_lock.lk_lockholder != curproc->p_pid) {
414 #ifdef DEBUG_LFS
415 				printf("lfs_writevnodes: not writing ino %d, locked by pid %d\n",
416 					VTOI(vp)->i_number,
417 					vp->v_lock.lk_lockholder);
418 #endif
419 				lfs_vunref(vp);
420 				continue;
421 			}
422 		} else if (vp != fs->lfs_ivnode) {
423 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
424 			needs_unlock = 1;
425 		}
426 
427 		only_cleaning = 0;
428 		/*
429 		 * Write the inode/file if dirty and it's not the
430 		 * the IFILE.
431 		 */
432 		if ((ip->i_flag & IN_ALLMOD) ||
433 		     (vp->v_dirtyblkhd.lh_first != NULL))
434 		{
435 			only_cleaning = ((ip->i_flag & IN_ALLMOD)==IN_CLEANING);
436 
437 			if(ip->i_number != LFS_IFILE_INUM
438 			   && vp->v_dirtyblkhd.lh_first != NULL)
439 			{
440 				lfs_writefile(fs, sp, vp);
441 			}
442 			if(vp->v_dirtyblkhd.lh_first != NULL) {
443 				if(WRITEINPROG(vp)) {
444 #ifdef DEBUG_LFS
445 					ivndebug(vp,"writevnodes/write2");
446 #endif
447 				} else if(!(ip->i_flag & IN_ALLMOD)) {
448 #ifdef DEBUG_LFS
449 					printf("<%d>",ip->i_number);
450 #endif
451 					ip->i_flag |= IN_MODIFIED;
452 					++fs->lfs_uinodes;
453 				}
454 			}
455 			(void) lfs_writeinode(fs, sp, ip);
456 			inodes_written++;
457 		}
458 
459 		if(needs_unlock)
460 			VOP_UNLOCK(vp,0);
461 
462 		if(lfs_clean_vnhead && only_cleaning)
463 			lfs_vunref_head(vp);
464 		else
465 			lfs_vunref(vp);
466 	}
467 	return inodes_written;
468 }
469 
470 int
471 lfs_segwrite(mp, flags)
472 	struct mount *mp;
473 	int flags;			/* Do a checkpoint. */
474 {
475 	struct buf *bp;
476 	struct inode *ip;
477 	struct lfs *fs;
478 	struct segment *sp;
479 	struct vnode *vp;
480 	SEGUSE *segusep;
481 	ufs_daddr_t ibno;
482 	int do_ckp, error, i;
483 	int writer_set = 0;
484 	int need_unlock = 0;
485 
486 	fs = VFSTOUFS(mp)->um_lfs;
487 
488 	lfs_imtime(fs);
489 
490 	/*
491 	 * If we are not the cleaner, and we have fewer than MIN_FREE_SEGS
492 	 * clean segments, wait until cleaner writes.
493 	 */
494 	if(!(flags & SEGM_CLEAN)
495 	   && (!fs->lfs_seglock || !(fs->lfs_sp->seg_flags & SEGM_CLEAN)))
496 	{
497 		do {
498 			if (fs->lfs_nclean <= MIN_FREE_SEGS
499 			    || fs->lfs_avail <= 0)
500 			{
501 				wakeup(&lfs_allclean_wakeup);
502 				wakeup(&fs->lfs_nextseg);
503 				error = tsleep(&fs->lfs_avail, PRIBIO + 1,
504 					       "lfs_avail", 0);
505 				if (error) {
506 					return (error);
507 				}
508 			}
509 		} while (fs->lfs_nclean <= MIN_FREE_SEGS || fs->lfs_avail <= 0);
510 	}
511 
512 	/*
513 	 * Allocate a segment structure and enough space to hold pointers to
514 	 * the maximum possible number of buffers which can be described in a
515 	 * single summary block.
516 	 */
517 	do_ckp = (flags & SEGM_CKP) || fs->lfs_nactive > LFS_MAX_ACTIVE;
518 	lfs_seglock(fs, flags | (do_ckp ? SEGM_CKP : 0));
519 	sp = fs->lfs_sp;
520 
521 	/*
522 	 * If lfs_flushvp is non-NULL, we are called from lfs_vflush,
523 	 * in which case we have to flush *all* buffers off of this vnode.
524 	 * We don't care about other nodes, but write any non-dirop nodes
525 	 * anyway in anticipation of another getnewvnode().
526 	 *
527 	 * If we're cleaning we only write cleaning and ifile blocks, and
528 	 * no dirops, since otherwise we'd risk corruption in a crash.
529 	 */
530 	if(sp->seg_flags & SEGM_CLEAN)
531 		lfs_writevnodes(fs, mp, sp, VN_CLEAN);
532 	else {
533 		lfs_writevnodes(fs, mp, sp, VN_REG);
534 		if(!fs->lfs_dirops || !fs->lfs_flushvp) {
535 			while(fs->lfs_dirops)
536 				if((error = tsleep(&fs->lfs_writer, PRIBIO + 1,
537 						"lfs writer", 0)))
538 				{
539 					free(sp->bpp, M_SEGMENT);
540 					free(sp, M_SEGMENT);
541 					return (error);
542 				}
543 			fs->lfs_writer++;
544 			writer_set=1;
545 			lfs_writevnodes(fs, mp, sp, VN_DIROP);
546 			((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT);
547 		}
548 	}
549 
550 	/*
551 	 * If we are doing a checkpoint, mark everything since the
552 	 * last checkpoint as no longer ACTIVE.
553 	 */
554 	if (do_ckp) {
555 		for (ibno = fs->lfs_cleansz + fs->lfs_segtabsz;
556 		     --ibno >= fs->lfs_cleansz; ) {
557 			if (bread(fs->lfs_ivnode, ibno, fs->lfs_bsize, NOCRED, &bp))
558 
559 				panic("lfs_segwrite: ifile read");
560 			segusep = (SEGUSE *)bp->b_data;
561 			for (i = fs->lfs_sepb; i--; segusep++)
562 				segusep->su_flags &= ~SEGUSE_ACTIVE;
563 
564 			/* But the current segment is still ACTIVE */
565 			if (fs->lfs_curseg/fs->lfs_sepb==(ibno-fs->lfs_cleansz))
566 				((SEGUSE *)(bp->b_data))[fs->lfs_curseg%fs->lfs_sepb].su_flags |= SEGUSE_ACTIVE;
567 			error = VOP_BWRITE(bp);
568 		}
569 	}
570 
571 	if (do_ckp || fs->lfs_doifile) {
572 	redo:
573 		vp = fs->lfs_ivnode;
574 		/*
575 		 * Depending on the circumstances of our calling, the ifile
576 		 * inode might be locked.  If it is, and if it is locked by
577 		 * us, we should VREF instead of vget here.
578 		 */
579 		need_unlock = 0;
580 		if(VOP_ISLOCKED(vp)
581 		   && vp->v_lock.lk_lockholder == curproc->p_pid) {
582 			VREF(vp);
583 		} else {
584 			while (vget(vp, LK_EXCLUSIVE))
585 				continue;
586 			need_unlock = 1;
587 		}
588 		ip = VTOI(vp);
589 		if (vp->v_dirtyblkhd.lh_first != NULL)
590 			lfs_writefile(fs, sp, vp);
591 		(void)lfs_writeinode(fs, sp, ip);
592 
593 		/* Only vput if we used vget() above. */
594 		if(need_unlock)
595 			vput(vp);
596 		else
597 			vrele(vp);
598 
599 		if (lfs_writeseg(fs, sp) && do_ckp)
600 			goto redo;
601 	} else {
602 		(void) lfs_writeseg(fs, sp);
603 	}
604 
605 	/*
606 	 * If the I/O count is non-zero, sleep until it reaches zero.
607 	 * At the moment, the user's process hangs around so we can
608 	 * sleep.
609 	 */
610 	fs->lfs_doifile = 0;
611 	if(writer_set && --fs->lfs_writer==0)
612 		wakeup(&fs->lfs_dirops);
613 
614 	if(lfs_dostats) {
615 		++lfs_stats.nwrites;
616 		if (sp->seg_flags & SEGM_SYNC)
617 			++lfs_stats.nsync_writes;
618 		if (sp->seg_flags & SEGM_CKP)
619 			++lfs_stats.ncheckpoints;
620 	}
621 	lfs_segunlock(fs);
622 	return (0);
623 }
624 
625 /*
626  * Write the dirty blocks associated with a vnode.
627  */
628 void
629 lfs_writefile(fs, sp, vp)
630 	struct lfs *fs;
631 	struct segment *sp;
632 	struct vnode *vp;
633 {
634 	struct buf *bp;
635 	struct finfo *fip;
636 	IFILE *ifp;
637 
638 
639 	if (sp->seg_bytes_left < fs->lfs_bsize ||
640 	    sp->sum_bytes_left < sizeof(struct finfo))
641 		(void) lfs_writeseg(fs, sp);
642 
643 	sp->sum_bytes_left -= sizeof(struct finfo) - sizeof(ufs_daddr_t);
644 	++((SEGSUM *)(sp->segsum))->ss_nfinfo;
645 
646 	if(vp->v_flag & VDIROP)
647 		((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT);
648 
649 	fip = sp->fip;
650 	fip->fi_nblocks = 0;
651 	fip->fi_ino = VTOI(vp)->i_number;
652 	LFS_IENTRY(ifp, fs, fip->fi_ino, bp);
653 	fip->fi_version = ifp->if_version;
654 	brelse(bp);
655 
656 	if(sp->seg_flags & SEGM_CLEAN)
657 	{
658 		lfs_gather(fs, sp, vp, lfs_match_fake);
659 		/*
660 		 * For a file being flushed, we need to write *all* blocks.
661 		 * This means writing the cleaning blocks first, and then
662 		 * immediately following with any non-cleaning blocks.
663 		 * The same is true of the Ifile since checkpoints assume
664 		 * that all valid Ifile blocks are written.
665 		 */
666 	   	if(IS_FLUSHING(fs,vp) || VTOI(vp)->i_number == LFS_IFILE_INUM)
667 			lfs_gather(fs, sp, vp, lfs_match_data);
668 	} else
669 		lfs_gather(fs, sp, vp, lfs_match_data);
670 
671 	/*
672 	 * It may not be necessary to write the meta-data blocks at this point,
673 	 * as the roll-forward recovery code should be able to reconstruct the
674 	 * list.
675 	 *
676 	 * We have to write them anyway, though, under two conditions: (1) the
677 	 * vnode is being flushed (for reuse by vinvalbuf); or (2) we are
678 	 * checkpointing.
679 	 */
680 	if(lfs_writeindir
681 	   || IS_FLUSHING(fs,vp)
682 	   || (sp->seg_flags & SEGM_CKP))
683 	{
684 		lfs_gather(fs, sp, vp, lfs_match_indir);
685 		lfs_gather(fs, sp, vp, lfs_match_dindir);
686 		lfs_gather(fs, sp, vp, lfs_match_tindir);
687 	}
688 	fip = sp->fip;
689 	if (fip->fi_nblocks != 0) {
690 		sp->fip = (FINFO*)((caddr_t)fip + sizeof(struct finfo) +
691 				   sizeof(ufs_daddr_t) * (fip->fi_nblocks-1));
692 		sp->start_lbp = &sp->fip->fi_blocks[0];
693 	} else {
694 		sp->sum_bytes_left += sizeof(FINFO) - sizeof(ufs_daddr_t);
695 		--((SEGSUM *)(sp->segsum))->ss_nfinfo;
696 	}
697 }
698 
699 int
700 lfs_writeinode(fs, sp, ip)
701 	struct lfs *fs;
702 	struct segment *sp;
703 	struct inode *ip;
704 {
705 	struct buf *bp, *ibp;
706 	IFILE *ifp;
707 	SEGUSE *sup;
708 	ufs_daddr_t daddr;
709 	ino_t ino;
710 	int error, i, ndx;
711 	int redo_ifile = 0;
712 	struct timespec ts;
713 	int gotblk=0;
714 
715 	if (!(ip->i_flag & IN_ALLMOD))
716 		return(0);
717 
718 	/* Allocate a new inode block if necessary. */
719 	if ((ip->i_number != LFS_IFILE_INUM || sp->idp==NULL) && sp->ibp == NULL) {
720 		/* Allocate a new segment if necessary. */
721 		if (sp->seg_bytes_left < fs->lfs_bsize ||
722 		    sp->sum_bytes_left < sizeof(ufs_daddr_t))
723 			(void) lfs_writeseg(fs, sp);
724 
725 		/* Get next inode block. */
726 		daddr = fs->lfs_offset;
727 		fs->lfs_offset += fsbtodb(fs, 1);
728 		sp->ibp = *sp->cbpp++ =
729 			getblk(VTOI(fs->lfs_ivnode)->i_devvp, daddr, fs->lfs_bsize, 0, 0);
730 		gotblk++;
731 
732 		/* Zero out inode numbers */
733 		for (i = 0; i < INOPB(fs); ++i)
734 			((struct dinode *)sp->ibp->b_data)[i].di_inumber = 0;
735 
736 		++sp->start_bpp;
737 		fs->lfs_avail -= fsbtodb(fs, 1);
738 		/* Set remaining space counters. */
739 		sp->seg_bytes_left -= fs->lfs_bsize;
740 		sp->sum_bytes_left -= sizeof(ufs_daddr_t);
741 		ndx = LFS_SUMMARY_SIZE / sizeof(ufs_daddr_t) -
742 			sp->ninodes / INOPB(fs) - 1;
743 		((ufs_daddr_t *)(sp->segsum))[ndx] = daddr;
744 	}
745 
746 	/* Update the inode times and copy the inode onto the inode page. */
747 	if (ip->i_flag & (IN_CLEANING | IN_MODIFIED | IN_ACCESSED))
748 		--fs->lfs_uinodes;
749 	TIMEVAL_TO_TIMESPEC(&time, &ts);
750 	LFS_ITIMES(ip, &ts, &ts, &ts);
751 
752 	/* XXX IN_ALLMOD */
753 	if(ip->i_flag & IN_CLEANING)
754 		ip->i_flag &= ~IN_CLEANING;
755 	else
756 		ip->i_flag &= ~(IN_ACCESS | IN_CHANGE | IN_MODIFIED |
757 				IN_UPDATE | IN_ACCESSED);
758 
759 	/*
760 	 * If this is the Ifile, and we've already written the Ifile in this
761 	 * partial segment, just overwrite it (it's not on disk yet) and
762 	 * continue.
763 	 *
764 	 * XXX we know that the bp that we get the second time around has
765 	 * already been gathered.
766 	 */
767 	if(ip->i_number == LFS_IFILE_INUM && sp->idp) {
768 		*(sp->idp) = ip->i_din.ffs_din;
769 		return 0;
770 	}
771 
772 	bp = sp->ibp;
773 	((struct dinode *)bp->b_data)[sp->ninodes % INOPB(fs)] =
774 		ip->i_din.ffs_din;
775 
776 	if(ip->i_number == LFS_IFILE_INUM) /* We know sp->idp == NULL */
777 		sp->idp = ((struct dinode *)bp->b_data)+(sp->ninodes % INOPB(fs));
778 	if(gotblk) {
779 		bp->b_flags |= B_LOCKED;
780 		brelse(bp);
781 	}
782 
783 	/* Increment inode count in segment summary block. */
784 	++((SEGSUM *)(sp->segsum))->ss_ninos;
785 
786 	/* If this page is full, set flag to allocate a new page. */
787 	if (++sp->ninodes % INOPB(fs) == 0)
788 		sp->ibp = NULL;
789 
790 	/*
791 	 * If updating the ifile, update the super-block.  Update the disk
792 	 * address and access times for this inode in the ifile.
793 	 */
794 	ino = ip->i_number;
795 	if (ino == LFS_IFILE_INUM) {
796 		daddr = fs->lfs_idaddr;
797 		fs->lfs_idaddr = bp->b_blkno;
798 	} else {
799 		LFS_IENTRY(ifp, fs, ino, ibp);
800 		daddr = ifp->if_daddr;
801 		ifp->if_daddr = bp->b_blkno;
802 #ifdef LFS_DEBUG_NEXTFREE
803 		if(ino > 3 && ifp->if_nextfree) {
804 			vprint("lfs_writeinode",ITOV(ip));
805 			printf("lfs_writeinode: updating free ino %d\n",
806 				ip->i_number);
807 		}
808 #endif
809 		error = VOP_BWRITE(ibp);
810 	}
811 
812 	/*
813 	 * No need to update segment usage if there was no former inode address
814 	 * or if the last inode address is in the current partial segment.
815 	 */
816 	if (daddr != LFS_UNUSED_DADDR &&
817 	    !(daddr >= fs->lfs_lastpseg && daddr <= bp->b_blkno)) {
818 		LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp);
819 #ifdef DIAGNOSTIC
820 		if (sup->su_nbytes < DINODE_SIZE) {
821 			/* XXX -- Change to a panic. */
822 			printf("lfs_writeinode: negative bytes (segment %d short by %d)\n",
823 			       datosn(fs, daddr), (int)DINODE_SIZE - sup->su_nbytes);
824 			panic("lfs_writeinode: negative bytes");
825 			sup->su_nbytes = DINODE_SIZE;
826 		}
827 #endif
828 		sup->su_nbytes -= DINODE_SIZE;
829 		redo_ifile =
830 			(ino == LFS_IFILE_INUM && !(bp->b_flags & B_GATHERED));
831 		error = VOP_BWRITE(bp);
832 	}
833 	return (redo_ifile);
834 }
835 
836 int
837 lfs_gatherblock(sp, bp, sptr)
838 	struct segment *sp;
839 	struct buf *bp;
840 	int *sptr;
841 {
842 	struct lfs *fs;
843 	int version;
844 
845 	/*
846 	 * If full, finish this segment.  We may be doing I/O, so
847 	 * release and reacquire the splbio().
848 	 */
849 #ifdef DIAGNOSTIC
850 	if (sp->vp == NULL)
851 		panic ("lfs_gatherblock: Null vp in segment");
852 #endif
853 	fs = sp->fs;
854 	if (sp->sum_bytes_left < sizeof(ufs_daddr_t) ||
855 	    sp->seg_bytes_left < bp->b_bcount) {
856 		if (sptr)
857 			splx(*sptr);
858 		lfs_updatemeta(sp);
859 
860 		version = sp->fip->fi_version;
861 		(void) lfs_writeseg(fs, sp);
862 
863 		sp->fip->fi_version = version;
864 		sp->fip->fi_ino = VTOI(sp->vp)->i_number;
865 		/* Add the current file to the segment summary. */
866 		++((SEGSUM *)(sp->segsum))->ss_nfinfo;
867 		sp->sum_bytes_left -=
868 			sizeof(struct finfo) - sizeof(ufs_daddr_t);
869 
870 		if (sptr)
871 			*sptr = splbio();
872 		return(1);
873 	}
874 
875 #ifdef DEBUG
876 	if(bp->b_flags & B_GATHERED) {
877 		printf("lfs_gatherblock: already gathered! Ino %d, lbn %d\n",
878 		       sp->fip->fi_ino, bp->b_lblkno);
879 		return(0);
880 	}
881 #endif
882 	/* Insert into the buffer list, update the FINFO block. */
883 	bp->b_flags |= B_GATHERED;
884 	*sp->cbpp++ = bp;
885 	sp->fip->fi_blocks[sp->fip->fi_nblocks++] = bp->b_lblkno;
886 
887 	sp->sum_bytes_left -= sizeof(ufs_daddr_t);
888 	sp->seg_bytes_left -= bp->b_bcount;
889 	return(0);
890 }
891 
892 int
893 lfs_gather(fs, sp, vp, match)
894 	struct lfs *fs;
895 	struct segment *sp;
896 	struct vnode *vp;
897 	int (*match) __P((struct lfs *, struct buf *));
898 {
899 	struct buf *bp;
900 	int s, count=0;
901 
902 	sp->vp = vp;
903 	s = splbio();
904 
905 #ifndef LFS_NO_BACKBUF_HACK
906 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp; bp = bp->b_vnbufs.le_next) {
907 #else /* LFS_NO_BACKBUF_HACK */
908 /* This is a hack to see if ordering the blocks in LFS makes a difference. */
909 # define	BUF_OFFSET	(((void *)&bp->b_vnbufs.le_next) - (void *)bp)
910 # define	BACK_BUF(BP)	((struct buf *)(((void *)BP->b_vnbufs.le_prev) - BUF_OFFSET))
911 # define	BEG_OF_LIST	((struct buf *)(((void *)&vp->v_dirtyblkhd.lh_first) - BUF_OFFSET))
912 /* Find last buffer. */
913 loop:	for (bp = vp->v_dirtyblkhd.lh_first; bp && bp->b_vnbufs.le_next != NULL;
914 	    bp = bp->b_vnbufs.le_next);
915 	for (; bp && bp != BEG_OF_LIST; bp = BACK_BUF(bp)) {
916 #endif /* LFS_NO_BACKBUF_HACK */
917 		if ((bp->b_flags & (B_BUSY|B_GATHERED)) || !match(fs, bp))
918 			continue;
919 		if(vp->v_type == VBLK) {
920 			/* For block devices, just write the blocks. */
921 			/* XXX Do we really need to even do this? */
922 #ifdef DEBUG_LFS
923 			if(count==0)
924 				printf("BLK(");
925 			printf(".");
926 #endif
927 			/* Get the block before bwrite, so we don't corrupt the free list */
928 			bp->b_flags |= B_BUSY;
929 			bremfree(bp);
930 			bwrite(bp);
931 		} else {
932 #ifdef DIAGNOSTIC
933 			if ((bp->b_flags & (B_CALL|B_INVAL))==B_INVAL) {
934 				printf("lfs_gather: lbn %d is B_INVAL\n",
935 					bp->b_lblkno);
936 				VOP_PRINT(bp->b_vp);
937 			}
938 			if (!(bp->b_flags & B_DELWRI))
939 				panic("lfs_gather: bp not B_DELWRI");
940 			if (!(bp->b_flags & B_LOCKED)) {
941 				printf("lfs_gather: lbn %d blk %d not B_LOCKED\n", bp->b_lblkno, bp->b_blkno);
942 				VOP_PRINT(bp->b_vp);
943 				panic("lfs_gather: bp not B_LOCKED");
944 			}
945 #endif
946 			if (lfs_gatherblock(sp, bp, &s)) {
947 				goto loop;
948 			}
949 		}
950 		count++;
951 	}
952 	splx(s);
953 #ifdef DEBUG_LFS
954 	if(vp->v_type == VBLK && count)
955 		printf(")\n");
956 #endif
957 	lfs_updatemeta(sp);
958 	sp->vp = NULL;
959 	return count;
960 }
961 
962 /*
963  * Update the metadata that points to the blocks listed in the FINFO
964  * array.
965  */
966 void
967 lfs_updatemeta(sp)
968 	struct segment *sp;
969 {
970 	SEGUSE *sup;
971 	struct buf *bp, *ibp;
972 	struct lfs *fs;
973 	struct vnode *vp;
974 	struct indir a[NIADDR + 2], *ap;
975 	struct inode *ip;
976 	ufs_daddr_t daddr, lbn, off;
977 	daddr_t ooff;
978 	int error, i, nblocks, num;
979 
980 	vp = sp->vp;
981 	nblocks = &sp->fip->fi_blocks[sp->fip->fi_nblocks] - sp->start_lbp;
982 	if (nblocks < 0)
983 		panic("This is a bad thing\n");
984 	if (vp == NULL || nblocks == 0)
985 		return;
986 
987 	/* Sort the blocks. */
988 	/*
989 	 * XXX KS - We have to sort even if the blocks come from the
990 	 * cleaner, because there might be other pending blocks on the
991 	 * same inode...and if we don't sort, and there are fragments
992 	 * present, blocks may be written in the wrong place.
993 	 */
994 	/* if (!(sp->seg_flags & SEGM_CLEAN)) */
995 	lfs_shellsort(sp->start_bpp, sp->start_lbp, nblocks);
996 
997 	/*
998 	 * Record the length of the last block in case it's a fragment.
999 	 * If there are indirect blocks present, they sort last.  An
1000 	 * indirect block will be lfs_bsize and its presence indicates
1001 	 * that you cannot have fragments.
1002 	 */
1003 	sp->fip->fi_lastlength = sp->start_bpp[nblocks - 1]->b_bcount;
1004 
1005 	/*
1006 	 * Assign disk addresses, and update references to the logical
1007 	 * block and the segment usage information.
1008 	 */
1009 	fs = sp->fs;
1010 	for (i = nblocks; i--; ++sp->start_bpp) {
1011 		lbn = *sp->start_lbp++;
1012 
1013 		(*sp->start_bpp)->b_blkno = off = fs->lfs_offset;
1014 		if((*sp->start_bpp)->b_blkno == (*sp->start_bpp)->b_lblkno) {
1015 			printf("lfs_updatemeta: ino %d blk %d has same lbn and daddr\n", VTOI(vp)->i_number, off);
1016 		}
1017 		fs->lfs_offset +=
1018 			fragstodb(fs, numfrags(fs, (*sp->start_bpp)->b_bcount));
1019 		error = ufs_bmaparray(vp, lbn, &daddr, a, &num, NULL);
1020 		if (error)
1021 			panic("lfs_updatemeta: ufs_bmaparray %d", error);
1022 		ip = VTOI(vp);
1023 		switch (num) {
1024 		case 0:
1025 			ooff = ip->i_ffs_db[lbn];
1026 			if(vp != fs->lfs_ivnode && (ooff == 0 || ooff == UNASSIGNED)) {
1027 #ifdef DEBUG_LFS
1028 				printf("lfs_updatemeta[1]: warning: writing ino %d lbn %d at 0x%x, was 0x%x\n", ip->i_number, lbn, off, ooff);
1029 #endif
1030 			} else
1031 				ip->i_ffs_db[lbn] = off;
1032 			break;
1033 		case 1:
1034 			ooff = ip->i_ffs_ib[a[0].in_off];
1035 			if(vp != fs->lfs_ivnode && (ooff == 0 || ooff == UNASSIGNED)) {
1036 #ifdef DEBUG_LFS
1037 				printf("lfs_updatemeta[2]: warning: writing ino %d lbn %d at 0x%x, was 0x%x\n", ip->i_number, lbn, off, ooff);
1038 #endif
1039 			} else
1040 				ip->i_ffs_ib[a[0].in_off] = off;
1041 			break;
1042 		default:
1043 			ap = &a[num - 1];
1044 			if (bread(vp, ap->in_lbn, fs->lfs_bsize, NOCRED, &bp))
1045 				panic("lfs_updatemeta: bread bno %d",
1046 				      ap->in_lbn);
1047 			/*
1048 			 * Bread may create a new (indirect) block which needs
1049 			 * to get counted for the inode.
1050 			 *
1051 			 * XXX - why would it ever do this (except possibly
1052 			 * for the Ifile)?  lfs_balloc is supposed to take
1053 			 * care of this.
1054 			 */
1055 			if (bp->b_blkno == UNASSIGNED) {
1056 				ip->i_ffs_blocks += fsbtodb(fs, 1);
1057 				fs->lfs_bfree -= fragstodb(fs, fs->lfs_frag);
1058 
1059 				/* Note the new address */
1060 				bp->b_blkno = UNWRITTEN;
1061 
1062 				if(num == 2) {
1063 					ip->i_ffs_ib[a[0].in_off] = UNWRITTEN;
1064 				} else {
1065 					ap = &a[num - 2];
1066 					if (bread(vp, ap->in_lbn,
1067 						  fs->lfs_bsize, NOCRED, &ibp))
1068 						panic("lfs_updatemeta: bread bno %d",
1069 						      ap->in_lbn);
1070 					((ufs_daddr_t *)ibp->b_data)[ap->in_off] = UNWRITTEN;
1071 					VOP_BWRITE(ibp);
1072 				}
1073 			}
1074 #ifdef DEBUG
1075 			else if(!(bp->b_flags & (B_DONE|B_DELWRI)))
1076                         	printf("lfs_updatemeta: unaccounted indirect block ino %d block %d\n", ip->i_number, ap->in_lbn);
1077 #endif
1078 			ooff = ((ufs_daddr_t *)bp->b_data)[ap->in_off];
1079 			if(vp != fs->lfs_ivnode && (ooff == 0 || ooff == UNASSIGNED)) {
1080 #ifdef DEBUG_LFS
1081 				printf("lfs_updatemeta[3]: warning: writing ino %d lbn %d at 0x%x, was 0x%x\n", ip->i_number, lbn, off, ooff);
1082 #endif
1083 				brelse(bp);
1084 			} else {
1085 				((ufs_daddr_t *)bp->b_data)[ap->in_off] = off;
1086 				VOP_BWRITE(bp);
1087 			}
1088 		}
1089 		/* Update segment usage information. */
1090 		if (daddr > 0 && !(daddr >= fs->lfs_lastpseg && daddr <= off)) {
1091 			LFS_SEGENTRY(sup, fs, datosn(fs, daddr), bp);
1092 #ifdef DIAGNOSTIC
1093 			if (sup->su_nbytes < (*sp->start_bpp)->b_bcount) {
1094 				/* XXX -- Change to a panic. */
1095 				printf("lfs_updatemeta: negative bytes (segment %d short by %ld)\n",
1096 				       datosn(fs, daddr), (*sp->start_bpp)->b_bcount - sup->su_nbytes);
1097 				printf("lfs_updatemeta: ino %d, lbn %d, addr = %x\n",
1098 				       VTOI(sp->vp)->i_number, (*sp->start_bpp)->b_lblkno, daddr);
1099 				panic("lfs_updatemeta: negative bytes");
1100 				sup->su_nbytes = (*sp->start_bpp)->b_bcount;
1101 			}
1102 #endif
1103 			sup->su_nbytes -= (*sp->start_bpp)->b_bcount;
1104 			error = VOP_BWRITE(bp);
1105 		}
1106 	}
1107 }
1108 
1109 /*
1110  * Start a new segment.
1111  */
1112 int
1113 lfs_initseg(fs)
1114 	struct lfs *fs;
1115 {
1116 	struct segment *sp;
1117 	SEGUSE *sup;
1118 	SEGSUM *ssp;
1119 	struct buf *bp;
1120 	int repeat;
1121 
1122 	sp = fs->lfs_sp;
1123 
1124 	repeat = 0;
1125 	/* Advance to the next segment. */
1126 	if (!LFS_PARTIAL_FITS(fs)) {
1127 		/* Wake up any cleaning procs waiting on this file system. */
1128 		wakeup(&lfs_allclean_wakeup);
1129 		wakeup(&fs->lfs_nextseg);
1130 		lfs_newseg(fs);
1131 		repeat = 1;
1132 		fs->lfs_offset = fs->lfs_curseg;
1133 		sp->seg_number = datosn(fs, fs->lfs_curseg);
1134 		sp->seg_bytes_left = fs->lfs_dbpseg * DEV_BSIZE;
1135 		/*
1136 		 * If the segment contains a superblock, update the offset
1137 		 * and summary address to skip over it.
1138 		 */
1139 		LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1140 		if (sup->su_flags & SEGUSE_SUPERBLOCK) {
1141 			fs->lfs_offset += LFS_SBPAD / DEV_BSIZE;
1142 			sp->seg_bytes_left -= LFS_SBPAD;
1143 		}
1144 		brelse(bp);
1145 	} else {
1146 		sp->seg_number = datosn(fs, fs->lfs_curseg);
1147 		sp->seg_bytes_left = (fs->lfs_dbpseg -
1148 				      (fs->lfs_offset - fs->lfs_curseg)) * DEV_BSIZE;
1149 	}
1150 	fs->lfs_lastpseg = fs->lfs_offset;
1151 
1152 	sp->fs = fs;
1153 	sp->ibp = NULL;
1154 	sp->idp = NULL;
1155 	sp->ninodes = 0;
1156 
1157 	/* Get a new buffer for SEGSUM and enter it into the buffer list. */
1158 	sp->cbpp = sp->bpp;
1159 	*sp->cbpp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp,
1160 			       fs->lfs_offset, LFS_SUMMARY_SIZE);
1161 	sp->segsum = (*sp->cbpp)->b_data;
1162 	bzero(sp->segsum, LFS_SUMMARY_SIZE);
1163 	sp->start_bpp = ++sp->cbpp;
1164 	fs->lfs_offset += LFS_SUMMARY_SIZE / DEV_BSIZE;
1165 
1166 	/* Set point to SEGSUM, initialize it. */
1167 	ssp = sp->segsum;
1168 	ssp->ss_next = fs->lfs_nextseg;
1169 	ssp->ss_nfinfo = ssp->ss_ninos = 0;
1170 	ssp->ss_magic = SS_MAGIC;
1171 
1172 	/* Set pointer to first FINFO, initialize it. */
1173 	sp->fip = (struct finfo *)((caddr_t)sp->segsum + sizeof(SEGSUM));
1174 	sp->fip->fi_nblocks = 0;
1175 	sp->start_lbp = &sp->fip->fi_blocks[0];
1176 	sp->fip->fi_lastlength = 0;
1177 
1178 	sp->seg_bytes_left -= LFS_SUMMARY_SIZE;
1179 	sp->sum_bytes_left = LFS_SUMMARY_SIZE - sizeof(SEGSUM);
1180 
1181 	return(repeat);
1182 }
1183 
1184 /*
1185  * Return the next segment to write.
1186  */
1187 void
1188 lfs_newseg(fs)
1189 	struct lfs *fs;
1190 {
1191 	CLEANERINFO *cip;
1192 	SEGUSE *sup;
1193 	struct buf *bp;
1194 	int curseg, isdirty, sn;
1195 
1196 	LFS_SEGENTRY(sup, fs, datosn(fs, fs->lfs_nextseg), bp);
1197 	sup->su_flags |= SEGUSE_DIRTY | SEGUSE_ACTIVE;
1198 	sup->su_nbytes = 0;
1199 	sup->su_nsums = 0;
1200 	sup->su_ninos = 0;
1201 	(void) VOP_BWRITE(bp);
1202 
1203 	LFS_CLEANERINFO(cip, fs, bp);
1204 	--cip->clean;
1205 	++cip->dirty;
1206 	fs->lfs_nclean = cip->clean;
1207 	(void) VOP_BWRITE(bp);
1208 
1209 	fs->lfs_lastseg = fs->lfs_curseg;
1210 	fs->lfs_curseg = fs->lfs_nextseg;
1211 	for (sn = curseg = datosn(fs, fs->lfs_curseg);;) {
1212 		sn = (sn + 1) % fs->lfs_nseg;
1213 		if (sn == curseg)
1214 			panic("lfs_nextseg: no clean segments");
1215 		LFS_SEGENTRY(sup, fs, sn, bp);
1216 		isdirty = sup->su_flags & SEGUSE_DIRTY;
1217 		brelse(bp);
1218 		if (!isdirty)
1219 			break;
1220 	}
1221 
1222 	++fs->lfs_nactive;
1223 	fs->lfs_nextseg = sntoda(fs, sn);
1224 	if(lfs_dostats) {
1225 		++lfs_stats.segsused;
1226 	}
1227 }
1228 
1229 int
1230 lfs_writeseg(fs, sp)
1231 	struct lfs *fs;
1232 	struct segment *sp;
1233 {
1234 	extern int locked_queue_count;
1235 	extern long locked_queue_bytes;
1236 	struct buf **bpp, *bp, *cbp;
1237 	SEGUSE *sup;
1238 	SEGSUM *ssp;
1239 	dev_t i_dev;
1240 	u_long *datap, *dp;
1241 	int do_again, i, nblocks, s;
1242 #ifdef LFS_TRACK_IOS
1243 	int j;
1244 #endif
1245 	int (*strategy)__P((void *));
1246 	struct vop_strategy_args vop_strategy_a;
1247 	u_short ninos;
1248 	struct vnode *devvp;
1249 	char *p;
1250 	struct vnode *vn;
1251 	struct inode *ip;
1252 #if defined(DEBUG) && defined(LFS_PROPELLER)
1253 	static int propeller;
1254 	char propstring[4] = "-\\|/";
1255 
1256 	printf("%c\b",propstring[propeller++]);
1257 	if(propeller==4)
1258 		propeller = 0;
1259 #endif
1260 
1261 	/*
1262 	 * If there are no buffers other than the segment summary to write
1263 	 * and it is not a checkpoint, don't do anything.  On a checkpoint,
1264 	 * even if there aren't any buffers, you need to write the superblock.
1265 	 */
1266 	if ((nblocks = sp->cbpp - sp->bpp) == 1)
1267 		return (0);
1268 
1269 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1270 	devvp = VTOI(fs->lfs_ivnode)->i_devvp;
1271 
1272 	/* Update the segment usage information. */
1273 	LFS_SEGENTRY(sup, fs, sp->seg_number, bp);
1274 
1275 	/* Loop through all blocks, except the segment summary. */
1276 	for (bpp = sp->bpp; ++bpp < sp->cbpp; ) {
1277 		if((*bpp)->b_vp != devvp)
1278 			sup->su_nbytes += (*bpp)->b_bcount;
1279 	}
1280 
1281 	ssp = (SEGSUM *)sp->segsum;
1282 
1283 	ninos = (ssp->ss_ninos + INOPB(fs) - 1) / INOPB(fs);
1284 	sup->su_nbytes += ssp->ss_ninos * DINODE_SIZE;
1285 	/* sup->su_nbytes += LFS_SUMMARY_SIZE; */
1286 	sup->su_lastmod = time.tv_sec;
1287 	sup->su_ninos += ninos;
1288 	++sup->su_nsums;
1289 
1290 	do_again = !(bp->b_flags & B_GATHERED);
1291 	(void)VOP_BWRITE(bp);
1292 	/*
1293 	 * Compute checksum across data and then across summary; the first
1294 	 * block (the summary block) is skipped.  Set the create time here
1295 	 * so that it's guaranteed to be later than the inode mod times.
1296 	 *
1297 	 * XXX
1298 	 * Fix this to do it inline, instead of malloc/copy.
1299 	 */
1300 	datap = dp = malloc(nblocks * sizeof(u_long), M_SEGMENT, M_WAITOK);
1301 	for (bpp = sp->bpp, i = nblocks - 1; i--;) {
1302 		if (((*++bpp)->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1303 			if (copyin((*bpp)->b_saveaddr, dp++, sizeof(u_long)))
1304 				panic("lfs_writeseg: copyin failed [1]: ino %d blk %d", VTOI((*bpp)->b_vp)->i_number, (*bpp)->b_lblkno);
1305 		} else {
1306 			if( !((*bpp)->b_flags & B_CALL) ) {
1307 				/*
1308 				 * Before we record data for a checksm,
1309 				 * make sure the data won't change in between
1310 				 * the checksum calculation and the write,
1311 				 * by marking the buffer B_BUSY.  It will
1312 				 * be freed later by brelse().
1313 				 */
1314 			again:
1315 				s = splbio();
1316 				if((*bpp)->b_flags & B_BUSY) {
1317 #ifdef DEBUG
1318 					printf("lfs_writeseg: avoiding potential data summary corruption for ino %d, lbn %d\n",
1319 					       VTOI((*bpp)->b_vp)->i_number,
1320 					       bp->b_lblkno);
1321 #endif
1322 					(*bpp)->b_flags |= B_WANTED;
1323 					tsleep((*bpp), (PRIBIO + 1),
1324 					       "lfs_writeseg", 0);
1325 					splx(s);
1326 					goto again;
1327 				}
1328 				(*bpp)->b_flags |= B_BUSY;
1329 				splx(s);
1330 			}
1331 			*dp++ = ((u_long *)(*bpp)->b_data)[0];
1332 		}
1333 	}
1334 	ssp->ss_create = time.tv_sec;
1335 	ssp->ss_datasum = cksum(datap, (nblocks - 1) * sizeof(u_long));
1336 	ssp->ss_sumsum =
1337 	    cksum(&ssp->ss_datasum, LFS_SUMMARY_SIZE - sizeof(ssp->ss_sumsum));
1338 	free(datap, M_SEGMENT);
1339 #ifdef DIAGNOSTIC
1340 	if (fs->lfs_bfree < fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE)
1341 		panic("lfs_writeseg: No diskspace for summary");
1342 #endif
1343 	fs->lfs_bfree -= (fsbtodb(fs, ninos) + LFS_SUMMARY_SIZE / DEV_BSIZE);
1344 
1345 	strategy = devvp->v_op[VOFFSET(vop_strategy)];
1346 
1347 	/*
1348 	 * When we simply write the blocks we lose a rotation for every block
1349 	 * written.  To avoid this problem, we allocate memory in chunks, copy
1350 	 * the buffers into the chunk and write the chunk.  CHUNKSIZE is the
1351 	 * largest size I/O devices can handle.
1352 	 * When the data is copied to the chunk, turn off the B_LOCKED bit
1353 	 * and brelse the buffer (which will move them to the LRU list).  Add
1354 	 * the B_CALL flag to the buffer header so we can count I/O's for the
1355 	 * checkpoints and so we can release the allocated memory.
1356 	 *
1357 	 * XXX
1358 	 * This should be removed if the new virtual memory system allows us to
1359 	 * easily make the buffers contiguous in kernel memory and if that's
1360 	 * fast enough.
1361 	 */
1362 
1363 #define CHUNKSIZE MAXPHYS
1364 
1365 	if(devvp==NULL)
1366 		panic("devvp is NULL");
1367 	for (bpp = sp->bpp,i = nblocks; i;) {
1368 		cbp = lfs_newbuf(devvp, (*bpp)->b_blkno, CHUNKSIZE);
1369 		cbp->b_dev = i_dev;
1370 		cbp->b_flags |= B_ASYNC | B_BUSY;
1371 		cbp->b_bcount = 0;
1372 
1373 #ifdef DIAGNOSTIC
1374 		if(datosn(fs,(*bpp)->b_blkno + ((*bpp)->b_bcount - 1)/DEV_BSIZE) != datosn(fs,cbp->b_blkno)) {
1375 			panic("lfs_writeseg: Segment overwrite");
1376 		}
1377 #endif
1378 
1379 		s = splbio();
1380 		if(fs->lfs_iocount >= LFS_THROTTLE) {
1381 			tsleep(&fs->lfs_iocount, PRIBIO+1, "lfs throttle", 0);
1382 		}
1383 		++fs->lfs_iocount;
1384 #ifdef LFS_TRACK_IOS
1385 		for(j=0;j<LFS_THROTTLE;j++) {
1386 			if(fs->lfs_pending[j]==LFS_UNUSED_DADDR) {
1387 				fs->lfs_pending[j] = cbp->b_blkno;
1388 				break;
1389 			}
1390 		}
1391 #endif /* LFS_TRACK_IOS */
1392 		for (p = cbp->b_data; i && cbp->b_bcount < CHUNKSIZE; i--) {
1393 			bp = *bpp;
1394 
1395 			if (bp->b_bcount > (CHUNKSIZE - cbp->b_bcount))
1396 				break;
1397 
1398 			/*
1399 			 * Fake buffers from the cleaner are marked as B_INVAL.
1400 			 * We need to copy the data from user space rather than
1401 			 * from the buffer indicated.
1402 			 * XXX == what do I do on an error?
1403 			 */
1404 			if ((bp->b_flags & (B_CALL|B_INVAL)) == (B_CALL|B_INVAL)) {
1405 				if (copyin(bp->b_saveaddr, p, bp->b_bcount))
1406 					panic("lfs_writeseg: copyin failed [2]");
1407 			} else
1408 				bcopy(bp->b_data, p, bp->b_bcount);
1409 			p += bp->b_bcount;
1410 			cbp->b_bcount += bp->b_bcount;
1411 			if (bp->b_flags & B_LOCKED) {
1412 				--locked_queue_count;
1413 				locked_queue_bytes -= bp->b_bufsize;
1414 			}
1415 			bp->b_flags &= ~(B_ERROR | B_READ | B_DELWRI |
1416 					 B_LOCKED | B_GATHERED);
1417 			vn = bp->b_vp;
1418 			if (bp->b_flags & B_CALL) {
1419 				/* if B_CALL, it was created with newbuf */
1420 				lfs_freebuf(bp);
1421 			} else {
1422 				bremfree(bp);
1423 				bp->b_flags |= B_DONE;
1424 				if(vn)
1425 					reassignbuf(bp, vn);
1426 				brelse(bp);
1427 			}
1428 			if(bp->b_flags & B_NEEDCOMMIT) { /* XXX */
1429 				bp->b_flags &= ~B_NEEDCOMMIT;
1430 				wakeup(bp);
1431 			}
1432 
1433 			bpp++;
1434 
1435 			/*
1436 			 * If this is the last block for this vnode, but
1437 			 * there are other blocks on its dirty list,
1438 			 * set IN_MODIFIED/IN_CLEANING depending on what
1439 			 * sort of block.  Only do this for our mount point,
1440 			 * not for, e.g., inode blocks that are attached to
1441 			 * the devvp.
1442 			 */
1443 			if(i>1 && vn && *bpp && (*bpp)->b_vp != vn
1444 			   && (*bpp)->b_vp && (bp=vn->v_dirtyblkhd.lh_first)!=NULL &&
1445 			   vn->v_mount == fs->lfs_ivnode->v_mount)
1446 			{
1447 				ip = VTOI(vn);
1448 #ifdef DEBUG_LFS
1449 				printf("lfs_writeseg: marking ino %d\n",ip->i_number);
1450 #endif
1451 		       		if(!(ip->i_flag & (IN_CLEANING | IN_MODIFIED |
1452 				                   IN_ACCESSED))) {
1453 					fs->lfs_uinodes++;
1454 					if(bp->b_flags & B_CALL)
1455 						ip->i_flag |= IN_CLEANING;
1456 					else
1457 						ip->i_flag |= IN_MODIFIED;
1458 				}
1459 			}
1460 			/* if(vn->v_dirtyblkhd.lh_first == NULL) */
1461 				wakeup(vn);
1462 		}
1463 		++cbp->b_vp->v_numoutput;
1464 		splx(s);
1465 		/*
1466 		 * XXXX This is a gross and disgusting hack.  Since these
1467 		 * buffers are physically addressed, they hang off the
1468 		 * device vnode (devvp).  As a result, they have no way
1469 		 * of getting to the LFS superblock or lfs structure to
1470 		 * keep track of the number of I/O's pending.  So, I am
1471 		 * going to stuff the fs into the saveaddr field of
1472 		 * the buffer (yuk).
1473 		 */
1474 		cbp->b_saveaddr = (caddr_t)fs;
1475 		vop_strategy_a.a_desc = VDESC(vop_strategy);
1476 		vop_strategy_a.a_bp = cbp;
1477 		(strategy)(&vop_strategy_a);
1478 	}
1479 	/*
1480 	 * XXX
1481 	 * Vinvalbuf can move locked buffers off the locked queue
1482 	 * and we have no way of knowing about this.  So, after
1483 	 * doing a big write, we recalculate how many buffers are
1484 	 * really still left on the locked queue.
1485 	 */
1486 	lfs_countlocked(&locked_queue_count,&locked_queue_bytes);
1487 	wakeup(&locked_queue_count);
1488 	if(lfs_dostats) {
1489 		++lfs_stats.psegwrites;
1490 		lfs_stats.blocktot += nblocks - 1;
1491 		if (fs->lfs_sp->seg_flags & SEGM_SYNC)
1492 			++lfs_stats.psyncwrites;
1493 		if (fs->lfs_sp->seg_flags & SEGM_CLEAN) {
1494 			++lfs_stats.pcleanwrites;
1495 			lfs_stats.cleanblocks += nblocks - 1;
1496 		}
1497 	}
1498 	return (lfs_initseg(fs) || do_again);
1499 }
1500 
1501 void
1502 lfs_writesuper(fs, daddr)
1503 	struct lfs *fs;
1504 	daddr_t daddr;
1505 {
1506 	struct buf *bp;
1507 	dev_t i_dev;
1508 	int (*strategy) __P((void *));
1509 	int s;
1510 	struct vop_strategy_args vop_strategy_a;
1511 
1512 #ifdef LFS_CANNOT_ROLLFW
1513 	/*
1514 	 * If we can write one superblock while another is in
1515 	 * progress, we risk not having a complete checkpoint if we crash.
1516 	 * So, block here if a superblock write is in progress.
1517 	 */
1518 	s = splbio();
1519 	while(fs->lfs_sbactive) {
1520 		tsleep(&fs->lfs_sbactive, PRIBIO+1, "lfs sb", 0);
1521 	}
1522 	fs->lfs_sbactive = daddr;
1523 	splx(s);
1524 #endif
1525 	i_dev = VTOI(fs->lfs_ivnode)->i_dev;
1526 	strategy = VTOI(fs->lfs_ivnode)->i_devvp->v_op[VOFFSET(vop_strategy)];
1527 
1528 	/* Set timestamp of this version of the superblock */
1529 	fs->lfs_tstamp = time.tv_sec;
1530 
1531 	/* Checksum the superblock and copy it into a buffer. */
1532 	fs->lfs_cksum = lfs_sb_cksum(&(fs->lfs_dlfs));
1533 	bp = lfs_newbuf(VTOI(fs->lfs_ivnode)->i_devvp, daddr, LFS_SBPAD);
1534 	*(struct dlfs *)bp->b_data = fs->lfs_dlfs;
1535 
1536 	bp->b_dev = i_dev;
1537 	bp->b_flags |= B_BUSY | B_CALL | B_ASYNC;
1538 	bp->b_flags &= ~(B_DONE | B_ERROR | B_READ | B_DELWRI);
1539 	bp->b_iodone = lfs_supercallback;
1540 	/* XXX KS - same nasty hack as above */
1541 	bp->b_saveaddr = (caddr_t)fs;
1542 
1543 	vop_strategy_a.a_desc = VDESC(vop_strategy);
1544 	vop_strategy_a.a_bp = bp;
1545 	s = splbio();
1546 	++bp->b_vp->v_numoutput;
1547 	splx(s);
1548 	(strategy)(&vop_strategy_a);
1549 }
1550 
1551 /*
1552  * Logical block number match routines used when traversing the dirty block
1553  * chain.
1554  */
1555 int
1556 lfs_match_fake(fs, bp)
1557 	struct lfs *fs;
1558 	struct buf *bp;
1559 {
1560 	return (bp->b_flags & B_CALL);
1561 }
1562 
1563 int
1564 lfs_match_data(fs, bp)
1565 	struct lfs *fs;
1566 	struct buf *bp;
1567 {
1568 	return (bp->b_lblkno >= 0);
1569 }
1570 
1571 int
1572 lfs_match_indir(fs, bp)
1573 	struct lfs *fs;
1574 	struct buf *bp;
1575 {
1576 	int lbn;
1577 
1578 	lbn = bp->b_lblkno;
1579 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 0);
1580 }
1581 
1582 int
1583 lfs_match_dindir(fs, bp)
1584 	struct lfs *fs;
1585 	struct buf *bp;
1586 {
1587 	int lbn;
1588 
1589 	lbn = bp->b_lblkno;
1590 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 1);
1591 }
1592 
1593 int
1594 lfs_match_tindir(fs, bp)
1595 	struct lfs *fs;
1596 	struct buf *bp;
1597 {
1598 	int lbn;
1599 
1600 	lbn = bp->b_lblkno;
1601 	return (lbn < 0 && (-lbn - NDADDR) % NINDIR(fs) == 2);
1602 }
1603 
1604 /*
1605  * XXX - The only buffers that are going to hit these functions are the
1606  * segment write blocks, or the segment summaries, or the superblocks.
1607  *
1608  * All of the above are created by lfs_newbuf, and so do not need to be
1609  * released via brelse.
1610  */
1611 void
1612 lfs_callback(bp)
1613 	struct buf *bp;
1614 {
1615 	struct lfs *fs;
1616 #ifdef LFS_TRACK_IOS
1617 	int j;
1618 #endif
1619 
1620 	fs = (struct lfs *)bp->b_saveaddr;
1621 #ifdef DIAGNOSTIC
1622 	if (fs->lfs_iocount == 0)
1623 		panic("lfs_callback: zero iocount\n");
1624 #endif
1625 	if (--fs->lfs_iocount < LFS_THROTTLE)
1626 		wakeup(&fs->lfs_iocount);
1627 #ifdef LFS_TRACK_IOS
1628 	for(j=0;j<LFS_THROTTLE;j++) {
1629 		if(fs->lfs_pending[j]==bp->b_blkno) {
1630 			fs->lfs_pending[j] = LFS_UNUSED_DADDR;
1631 			wakeup(&(fs->lfs_pending[j]));
1632 			break;
1633 		}
1634 	}
1635 #endif /* LFS_TRACK_IOS */
1636 
1637 	lfs_freebuf(bp);
1638 }
1639 
1640 void
1641 lfs_supercallback(bp)
1642 	struct buf *bp;
1643 {
1644 #ifdef LFS_CANNOT_ROLLFW
1645 	struct lfs *fs;
1646 
1647 	fs = (struct lfs *)bp->b_saveaddr;
1648 	fs->lfs_sbactive = 0;
1649 	wakeup(&fs->lfs_sbactive);
1650 #endif
1651 	lfs_freebuf(bp);
1652 }
1653 
1654 /*
1655  * Shellsort (diminishing increment sort) from Data Structures and
1656  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
1657  * see also Knuth Vol. 3, page 84.  The increments are selected from
1658  * formula (8), page 95.  Roughly O(N^3/2).
1659  */
1660 /*
1661  * This is our own private copy of shellsort because we want to sort
1662  * two parallel arrays (the array of buffer pointers and the array of
1663  * logical block numbers) simultaneously.  Note that we cast the array
1664  * of logical block numbers to a unsigned in this routine so that the
1665  * negative block numbers (meta data blocks) sort AFTER the data blocks.
1666  */
1667 
1668 void
1669 lfs_shellsort(bp_array, lb_array, nmemb)
1670 	struct buf **bp_array;
1671 	ufs_daddr_t *lb_array;
1672 	int nmemb;
1673 {
1674 	static int __rsshell_increments[] = { 4, 1, 0 };
1675 	int incr, *incrp, t1, t2;
1676 	struct buf *bp_temp;
1677 	u_long lb_temp;
1678 
1679 	for (incrp = __rsshell_increments; (incr = *incrp++) != 0;)
1680 		for (t1 = incr; t1 < nmemb; ++t1)
1681 			for (t2 = t1 - incr; t2 >= 0;)
1682 				if (lb_array[t2] > lb_array[t2 + incr]) {
1683 					lb_temp = lb_array[t2];
1684 					lb_array[t2] = lb_array[t2 + incr];
1685 					lb_array[t2 + incr] = lb_temp;
1686 					bp_temp = bp_array[t2];
1687 					bp_array[t2] = bp_array[t2 + incr];
1688 					bp_array[t2 + incr] = bp_temp;
1689 					t2 -= incr;
1690 				} else
1691 					break;
1692 }
1693 
1694 /*
1695  * Check VXLOCK.  Return 1 if the vnode is locked.  Otherwise, vget it.
1696  */
1697 int
1698 lfs_vref(vp)
1699 	struct vnode *vp;
1700 {
1701 	/*
1702 	 * If we return 1 here during a flush, we risk vinvalbuf() not
1703 	 * being able to flush all of the pages from this vnode, which
1704 	 * will cause it to panic.  So, return 0 if a flush is in progress.
1705 	 */
1706 	if (vp->v_flag & VXLOCK) {
1707 		if(IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1708 			return 0;
1709 		}
1710 		return(1);
1711 	}
1712 	return (vget(vp, 0));
1713 }
1714 
1715 /*
1716  * This is vrele except that we do not want to VOP_INACTIVE this vnode. We
1717  * inline vrele here to avoid the vn_lock and VOP_INACTIVE call at the end.
1718  */
1719 void
1720 lfs_vunref(vp)
1721 	struct vnode *vp;
1722 {
1723 	/*
1724 	 * Analogous to lfs_vref, if the node is flushing, fake it.
1725 	 */
1726 	if((vp->v_flag & VXLOCK) && IS_FLUSHING(VTOI(vp)->i_lfs,vp)) {
1727 		return;
1728 	}
1729 
1730 	simple_lock(&vp->v_interlock);
1731 #ifdef DIAGNOSTIC
1732 	if(vp->v_usecount<=0) {
1733 		printf("lfs_vunref: flags are 0x%lx\n", vp->v_flag);
1734 		printf("lfs_vunref: usecount = %ld\n", vp->v_usecount);
1735 		panic("lfs_vunref: v_usecount<0");
1736 	}
1737 #endif
1738 	vp->v_usecount--;
1739 	if (vp->v_usecount > 0) {
1740 		simple_unlock(&vp->v_interlock);
1741 		return;
1742 	}
1743 #ifdef DIAGNOSTIC
1744 	if(VOP_ISLOCKED(vp))
1745 		panic("lfs_vunref: vnode locked");
1746 #endif
1747 	/*
1748 	 * insert at tail of LRU list
1749 	 */
1750 	simple_lock(&vnode_free_list_slock);
1751 	if (vp->v_holdcnt > 0)
1752 		TAILQ_INSERT_TAIL(&vnode_hold_list, vp, v_freelist);
1753 	else
1754 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
1755 	simple_unlock(&vnode_free_list_slock);
1756 	simple_unlock(&vp->v_interlock);
1757 }
1758 
1759 /*
1760  * We use this when we have vnodes that were loaded in solely for cleaning.
1761  * There is no reason to believe that these vnodes will be referenced again
1762  * soon, since the cleaning process is unrelated to normal filesystem
1763  * activity.  Putting cleaned vnodes at the tail of the list has the effect
1764  * of flushing the vnode LRU.  So, put vnodes that were loaded only for
1765  * cleaning at the head of the list, instead.
1766  */
1767 void
1768 lfs_vunref_head(vp)
1769 	struct vnode *vp;
1770 {
1771 	simple_lock(&vp->v_interlock);
1772 #ifdef DIAGNOSTIC
1773 	if(vp->v_usecount==0) {
1774 		panic("lfs_vunref: v_usecount<0");
1775 	}
1776 #endif
1777 	vp->v_usecount--;
1778 	if (vp->v_usecount > 0) {
1779 		simple_unlock(&vp->v_interlock);
1780 		return;
1781 	}
1782 #ifdef DIAGNOSTIC
1783 	if(VOP_ISLOCKED(vp))
1784 		panic("lfs_vunref_head: vnode locked");
1785 #endif
1786 	/*
1787 	 * insert at head of LRU list
1788 	 */
1789 	simple_lock(&vnode_free_list_slock);
1790 	TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
1791 	simple_unlock(&vnode_free_list_slock);
1792 	simple_unlock(&vp->v_interlock);
1793 }
1794 
1795