xref: /freebsd/sys/ufs/ffs/ffs_alloc.c (revision 78b9f0095b4af3aca6c931b2c7b009ddb8a05125)
1 /*-
2  * SPDX-License-Identifier: (BSD-2-Clause-FreeBSD AND BSD-3-Clause)
3  *
4  * Copyright (c) 2002 Networks Associates Technology, Inc.
5  * All rights reserved.
6  *
7  * This software was developed for the FreeBSD Project by Marshall
8  * Kirk McKusick and Network Associates Laboratories, the Security
9  * Research Division of Network Associates, Inc. under DARPA/SPAWAR
10  * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS
11  * research program
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1982, 1986, 1989, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  *
37  * Redistribution and use in source and binary forms, with or without
38  * modification, are permitted provided that the following conditions
39  * are met:
40  * 1. Redistributions of source code must retain the above copyright
41  *    notice, this list of conditions and the following disclaimer.
42  * 2. Redistributions in binary form must reproduce the above copyright
43  *    notice, this list of conditions and the following disclaimer in the
44  *    documentation and/or other materials provided with the distribution.
45  * 3. Neither the name of the University nor the names of its contributors
46  *    may be used to endorse or promote products derived from this software
47  *    without specific prior written permission.
48  *
49  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
50  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
51  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
52  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
53  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
54  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
55  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
56  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
57  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
58  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
59  * SUCH DAMAGE.
60  *
61  *	@(#)ffs_alloc.c	8.18 (Berkeley) 5/26/95
62  */
63 
64 #include <sys/cdefs.h>
65 __FBSDID("$FreeBSD$");
66 
67 #include "opt_quota.h"
68 
69 #include <sys/param.h>
70 #include <sys/capsicum.h>
71 #include <sys/systm.h>
72 #include <sys/bio.h>
73 #include <sys/buf.h>
74 #include <sys/conf.h>
75 #include <sys/fcntl.h>
76 #include <sys/file.h>
77 #include <sys/filedesc.h>
78 #include <sys/priv.h>
79 #include <sys/proc.h>
80 #include <sys/vnode.h>
81 #include <sys/mount.h>
82 #include <sys/kernel.h>
83 #include <sys/syscallsubr.h>
84 #include <sys/sysctl.h>
85 #include <sys/syslog.h>
86 #include <sys/taskqueue.h>
87 
88 #include <security/audit/audit.h>
89 
90 #include <geom/geom.h>
91 #include <geom/geom_vfs.h>
92 
93 #include <ufs/ufs/dir.h>
94 #include <ufs/ufs/extattr.h>
95 #include <ufs/ufs/quota.h>
96 #include <ufs/ufs/inode.h>
97 #include <ufs/ufs/ufs_extern.h>
98 #include <ufs/ufs/ufsmount.h>
99 
100 #include <ufs/ffs/fs.h>
101 #include <ufs/ffs/ffs_extern.h>
102 #include <ufs/ffs/softdep.h>
103 
104 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref,
105 				  int size, int rsize);
106 
107 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int);
108 static ufs2_daddr_t
109 	      ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int);
110 static void	ffs_blkfree_cg(struct ufsmount *, struct fs *,
111 		    struct vnode *, ufs2_daddr_t, long, ino_t,
112 		    struct workhead *);
113 #ifdef INVARIANTS
114 static int	ffs_checkblk(struct inode *, ufs2_daddr_t, long);
115 #endif
116 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int);
117 static ino_t	ffs_dirpref(struct inode *);
118 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t,
119 		    int, int);
120 static ufs2_daddr_t	ffs_hashalloc
121 		(struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *);
122 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int,
123 		    int);
124 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int);
125 static int	ffs_reallocblks_ufs1(struct vop_reallocblks_args *);
126 static int	ffs_reallocblks_ufs2(struct vop_reallocblks_args *);
127 static void	ffs_ckhash_cg(struct buf *);
128 
129 /*
130  * Allocate a block in the filesystem.
131  *
132  * The size of the requested block is given, which must be some
133  * multiple of fs_fsize and <= fs_bsize.
134  * A preference may be optionally specified. If a preference is given
135  * the following hierarchy is used to allocate a block:
136  *   1) allocate the requested block.
137  *   2) allocate a rotationally optimal block in the same cylinder.
138  *   3) allocate a block in the same cylinder group.
139  *   4) quadradically rehash into other cylinder groups, until an
140  *      available block is located.
141  * If no block preference is given the following hierarchy is used
142  * to allocate a block:
143  *   1) allocate a block in the cylinder group that contains the
144  *      inode for the file.
145  *   2) quadradically rehash into other cylinder groups, until an
146  *      available block is located.
147  */
148 int
149 ffs_alloc(ip, lbn, bpref, size, flags, cred, bnp)
150 	struct inode *ip;
151 	ufs2_daddr_t lbn, bpref;
152 	int size, flags;
153 	struct ucred *cred;
154 	ufs2_daddr_t *bnp;
155 {
156 	struct fs *fs;
157 	struct ufsmount *ump;
158 	ufs2_daddr_t bno;
159 	u_int cg, reclaimed;
160 	static struct timeval lastfail;
161 	static int curfail;
162 	int64_t delta;
163 #ifdef QUOTA
164 	int error;
165 #endif
166 
167 	*bnp = 0;
168 	ump = ITOUMP(ip);
169 	fs = ump->um_fs;
170 	mtx_assert(UFS_MTX(ump), MA_OWNED);
171 #ifdef INVARIANTS
172 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
173 		printf("dev = %s, bsize = %ld, size = %d, fs = %s\n",
174 		    devtoname(ump->um_dev), (long)fs->fs_bsize, size,
175 		    fs->fs_fsmnt);
176 		panic("ffs_alloc: bad size");
177 	}
178 	if (cred == NOCRED)
179 		panic("ffs_alloc: missing credential");
180 #endif /* INVARIANTS */
181 	reclaimed = 0;
182 retry:
183 #ifdef QUOTA
184 	UFS_UNLOCK(ump);
185 	error = chkdq(ip, btodb(size), cred, 0);
186 	if (error)
187 		return (error);
188 	UFS_LOCK(ump);
189 #endif
190 	if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0)
191 		goto nospace;
192 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0) &&
193 	    freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0)
194 		goto nospace;
195 	if (bpref >= fs->fs_size)
196 		bpref = 0;
197 	if (bpref == 0)
198 		cg = ino_to_cg(fs, ip->i_number);
199 	else
200 		cg = dtog(fs, bpref);
201 	bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg);
202 	if (bno > 0) {
203 		delta = btodb(size);
204 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
205 		if (flags & IO_EXT)
206 			ip->i_flag |= IN_CHANGE;
207 		else
208 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
209 		*bnp = bno;
210 		return (0);
211 	}
212 nospace:
213 #ifdef QUOTA
214 	UFS_UNLOCK(ump);
215 	/*
216 	 * Restore user's disk quota because allocation failed.
217 	 */
218 	(void) chkdq(ip, -btodb(size), cred, FORCE);
219 	UFS_LOCK(ump);
220 #endif
221 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
222 		reclaimed = 1;
223 		softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT);
224 		goto retry;
225 	}
226 	UFS_UNLOCK(ump);
227 	if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
228 		ffs_fserr(fs, ip->i_number, "filesystem full");
229 		uprintf("\n%s: write failed, filesystem is full\n",
230 		    fs->fs_fsmnt);
231 	}
232 	return (ENOSPC);
233 }
234 
235 /*
236  * Reallocate a fragment to a bigger size
237  *
238  * The number and size of the old block is given, and a preference
239  * and new size is also specified. The allocator attempts to extend
240  * the original block. Failing that, the regular block allocator is
241  * invoked to get an appropriate block.
242  */
243 int
244 ffs_realloccg(ip, lbprev, bprev, bpref, osize, nsize, flags, cred, bpp)
245 	struct inode *ip;
246 	ufs2_daddr_t lbprev;
247 	ufs2_daddr_t bprev;
248 	ufs2_daddr_t bpref;
249 	int osize, nsize, flags;
250 	struct ucred *cred;
251 	struct buf **bpp;
252 {
253 	struct vnode *vp;
254 	struct fs *fs;
255 	struct buf *bp;
256 	struct ufsmount *ump;
257 	u_int cg, request, reclaimed;
258 	int error, gbflags;
259 	ufs2_daddr_t bno;
260 	static struct timeval lastfail;
261 	static int curfail;
262 	int64_t delta;
263 
264 	vp = ITOV(ip);
265 	ump = ITOUMP(ip);
266 	fs = ump->um_fs;
267 	bp = NULL;
268 	gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
269 
270 	mtx_assert(UFS_MTX(ump), MA_OWNED);
271 #ifdef INVARIANTS
272 	if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
273 		panic("ffs_realloccg: allocation on suspended filesystem");
274 	if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 ||
275 	    (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) {
276 		printf(
277 		"dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n",
278 		    devtoname(ump->um_dev), (long)fs->fs_bsize, osize,
279 		    nsize, fs->fs_fsmnt);
280 		panic("ffs_realloccg: bad size");
281 	}
282 	if (cred == NOCRED)
283 		panic("ffs_realloccg: missing credential");
284 #endif /* INVARIANTS */
285 	reclaimed = 0;
286 retry:
287 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0) &&
288 	    freespace(fs, fs->fs_minfree) -  numfrags(fs, nsize - osize) < 0) {
289 		goto nospace;
290 	}
291 	if (bprev == 0) {
292 		printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n",
293 		    devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev,
294 		    fs->fs_fsmnt);
295 		panic("ffs_realloccg: bad bprev");
296 	}
297 	UFS_UNLOCK(ump);
298 	/*
299 	 * Allocate the extra space in the buffer.
300 	 */
301 	error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp);
302 	if (error) {
303 		brelse(bp);
304 		return (error);
305 	}
306 
307 	if (bp->b_blkno == bp->b_lblkno) {
308 		if (lbprev >= UFS_NDADDR)
309 			panic("ffs_realloccg: lbprev out of range");
310 		bp->b_blkno = fsbtodb(fs, bprev);
311 	}
312 
313 #ifdef QUOTA
314 	error = chkdq(ip, btodb(nsize - osize), cred, 0);
315 	if (error) {
316 		brelse(bp);
317 		return (error);
318 	}
319 #endif
320 	/*
321 	 * Check for extension in the existing location.
322 	 */
323 	*bpp = NULL;
324 	cg = dtog(fs, bprev);
325 	UFS_LOCK(ump);
326 	bno = ffs_fragextend(ip, cg, bprev, osize, nsize);
327 	if (bno) {
328 		if (bp->b_blkno != fsbtodb(fs, bno))
329 			panic("ffs_realloccg: bad blockno");
330 		delta = btodb(nsize - osize);
331 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
332 		if (flags & IO_EXT)
333 			ip->i_flag |= IN_CHANGE;
334 		else
335 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
336 		allocbuf(bp, nsize);
337 		bp->b_flags |= B_DONE;
338 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
339 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
340 			vfs_bio_set_valid(bp, osize, nsize - osize);
341 		*bpp = bp;
342 		return (0);
343 	}
344 	/*
345 	 * Allocate a new disk location.
346 	 */
347 	if (bpref >= fs->fs_size)
348 		bpref = 0;
349 	switch ((int)fs->fs_optim) {
350 	case FS_OPTSPACE:
351 		/*
352 		 * Allocate an exact sized fragment. Although this makes
353 		 * best use of space, we will waste time relocating it if
354 		 * the file continues to grow. If the fragmentation is
355 		 * less than half of the minimum free reserve, we choose
356 		 * to begin optimizing for time.
357 		 */
358 		request = nsize;
359 		if (fs->fs_minfree <= 5 ||
360 		    fs->fs_cstotal.cs_nffree >
361 		    (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100))
362 			break;
363 		log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n",
364 			fs->fs_fsmnt);
365 		fs->fs_optim = FS_OPTTIME;
366 		break;
367 	case FS_OPTTIME:
368 		/*
369 		 * At this point we have discovered a file that is trying to
370 		 * grow a small fragment to a larger fragment. To save time,
371 		 * we allocate a full sized block, then free the unused portion.
372 		 * If the file continues to grow, the `ffs_fragextend' call
373 		 * above will be able to grow it in place without further
374 		 * copying. If aberrant programs cause disk fragmentation to
375 		 * grow within 2% of the free reserve, we choose to begin
376 		 * optimizing for space.
377 		 */
378 		request = fs->fs_bsize;
379 		if (fs->fs_cstotal.cs_nffree <
380 		    (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100)
381 			break;
382 		log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n",
383 			fs->fs_fsmnt);
384 		fs->fs_optim = FS_OPTSPACE;
385 		break;
386 	default:
387 		printf("dev = %s, optim = %ld, fs = %s\n",
388 		    devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt);
389 		panic("ffs_realloccg: bad optim");
390 		/* NOTREACHED */
391 	}
392 	bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg);
393 	if (bno > 0) {
394 		bp->b_blkno = fsbtodb(fs, bno);
395 		if (!DOINGSOFTDEP(vp))
396 			/*
397 			 * The usual case is that a smaller fragment that
398 			 * was just allocated has been replaced with a bigger
399 			 * fragment or a full-size block. If it is marked as
400 			 * B_DELWRI, the current contents have not been written
401 			 * to disk. It is possible that the block was written
402 			 * earlier, but very uncommon. If the block has never
403 			 * been written, there is no need to send a BIO_DELETE
404 			 * for it when it is freed. The gain from avoiding the
405 			 * TRIMs for the common case of unwritten blocks far
406 			 * exceeds the cost of the write amplification for the
407 			 * uncommon case of failing to send a TRIM for a block
408 			 * that had been written.
409 			 */
410 			ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize,
411 			    ip->i_number, vp->v_type, NULL,
412 			    (bp->b_flags & B_DELWRI) != 0 ? NOTRIM : SINGLETON);
413 		delta = btodb(nsize - osize);
414 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
415 		if (flags & IO_EXT)
416 			ip->i_flag |= IN_CHANGE;
417 		else
418 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
419 		allocbuf(bp, nsize);
420 		bp->b_flags |= B_DONE;
421 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
422 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
423 			vfs_bio_set_valid(bp, osize, nsize - osize);
424 		*bpp = bp;
425 		return (0);
426 	}
427 #ifdef QUOTA
428 	UFS_UNLOCK(ump);
429 	/*
430 	 * Restore user's disk quota because allocation failed.
431 	 */
432 	(void) chkdq(ip, -btodb(nsize - osize), cred, FORCE);
433 	UFS_LOCK(ump);
434 #endif
435 nospace:
436 	/*
437 	 * no space available
438 	 */
439 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
440 		reclaimed = 1;
441 		UFS_UNLOCK(ump);
442 		if (bp) {
443 			brelse(bp);
444 			bp = NULL;
445 		}
446 		UFS_LOCK(ump);
447 		softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT);
448 		goto retry;
449 	}
450 	UFS_UNLOCK(ump);
451 	if (bp)
452 		brelse(bp);
453 	if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
454 		ffs_fserr(fs, ip->i_number, "filesystem full");
455 		uprintf("\n%s: write failed, filesystem is full\n",
456 		    fs->fs_fsmnt);
457 	}
458 	return (ENOSPC);
459 }
460 
461 /*
462  * Reallocate a sequence of blocks into a contiguous sequence of blocks.
463  *
464  * The vnode and an array of buffer pointers for a range of sequential
465  * logical blocks to be made contiguous is given. The allocator attempts
466  * to find a range of sequential blocks starting as close as possible
467  * from the end of the allocation for the logical block immediately
468  * preceding the current range. If successful, the physical block numbers
469  * in the buffer pointers and in the inode are changed to reflect the new
470  * allocation. If unsuccessful, the allocation is left unchanged. The
471  * success in doing the reallocation is returned. Note that the error
472  * return is not reflected back to the user. Rather the previous block
473  * allocation will be used.
474  */
475 
476 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW, 0, "FFS filesystem");
477 
478 static int doasyncfree = 1;
479 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0,
480 "do not force synchronous writes when blocks are reallocated");
481 
482 static int doreallocblks = 1;
483 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0,
484 "enable block reallocation");
485 
486 static int maxclustersearch = 10;
487 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch,
488 0, "max number of cylinder group to search for contigous blocks");
489 
490 #ifdef DEBUG
491 static volatile int prtrealloc = 0;
492 #endif
493 
494 int
495 ffs_reallocblks(ap)
496 	struct vop_reallocblks_args /* {
497 		struct vnode *a_vp;
498 		struct cluster_save *a_buflist;
499 	} */ *ap;
500 {
501 	struct ufsmount *ump;
502 
503 	/*
504 	 * If the underlying device can do deletes, then skip reallocating
505 	 * the blocks of this file into contiguous sequences. Devices that
506 	 * benefit from BIO_DELETE also benefit from not moving the data.
507 	 * These devices are flash and therefore work less well with this
508 	 * optimization. Also skip if reallocblks has been disabled globally.
509 	 */
510 	ump = ap->a_vp->v_mount->mnt_data;
511 	if (((ump->um_flags) & UM_CANDELETE) != 0 || doreallocblks == 0)
512 		return (ENOSPC);
513 
514 	/*
515 	 * We can't wait in softdep prealloc as it may fsync and recurse
516 	 * here.  Instead we simply fail to reallocate blocks if this
517 	 * rare condition arises.
518 	 */
519 	if (DOINGSOFTDEP(ap->a_vp))
520 		if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0)
521 			return (ENOSPC);
522 	if (ump->um_fstype == UFS1)
523 		return (ffs_reallocblks_ufs1(ap));
524 	return (ffs_reallocblks_ufs2(ap));
525 }
526 
527 static int
528 ffs_reallocblks_ufs1(ap)
529 	struct vop_reallocblks_args /* {
530 		struct vnode *a_vp;
531 		struct cluster_save *a_buflist;
532 	} */ *ap;
533 {
534 	struct fs *fs;
535 	struct inode *ip;
536 	struct vnode *vp;
537 	struct buf *sbp, *ebp, *bp;
538 	ufs1_daddr_t *bap, *sbap, *ebap;
539 	struct cluster_save *buflist;
540 	struct ufsmount *ump;
541 	ufs_lbn_t start_lbn, end_lbn;
542 	ufs1_daddr_t soff, newblk, blkno;
543 	ufs2_daddr_t pref;
544 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
545 	int i, cg, len, start_lvl, end_lvl, ssize;
546 
547 	vp = ap->a_vp;
548 	ip = VTOI(vp);
549 	ump = ITOUMP(ip);
550 	fs = ump->um_fs;
551 	/*
552 	 * If we are not tracking block clusters or if we have less than 4%
553 	 * free blocks left, then do not attempt to cluster. Running with
554 	 * less than 5% free block reserve is not recommended and those that
555 	 * choose to do so do not expect to have good file layout.
556 	 */
557 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
558 		return (ENOSPC);
559 	buflist = ap->a_buflist;
560 	len = buflist->bs_nchildren;
561 	start_lbn = buflist->bs_children[0]->b_lblkno;
562 	end_lbn = start_lbn + len - 1;
563 #ifdef INVARIANTS
564 	for (i = 0; i < len; i++)
565 		if (!ffs_checkblk(ip,
566 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
567 			panic("ffs_reallocblks: unallocated block 1");
568 	for (i = 1; i < len; i++)
569 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
570 			panic("ffs_reallocblks: non-logical cluster");
571 	blkno = buflist->bs_children[0]->b_blkno;
572 	ssize = fsbtodb(fs, fs->fs_frag);
573 	for (i = 1; i < len - 1; i++)
574 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
575 			panic("ffs_reallocblks: non-physical cluster %d", i);
576 #endif
577 	/*
578 	 * If the cluster crosses the boundary for the first indirect
579 	 * block, leave space for the indirect block. Indirect blocks
580 	 * are initially laid out in a position after the last direct
581 	 * block. Block reallocation would usually destroy locality by
582 	 * moving the indirect block out of the way to make room for
583 	 * data blocks if we didn't compensate here. We should also do
584 	 * this for other indirect block boundaries, but it is only
585 	 * important for the first one.
586 	 */
587 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
588 		return (ENOSPC);
589 	/*
590 	 * If the latest allocation is in a new cylinder group, assume that
591 	 * the filesystem has decided to move and do not force it back to
592 	 * the previous cylinder group.
593 	 */
594 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
595 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
596 		return (ENOSPC);
597 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
598 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
599 		return (ENOSPC);
600 	/*
601 	 * Get the starting offset and block map for the first block.
602 	 */
603 	if (start_lvl == 0) {
604 		sbap = &ip->i_din1->di_db[0];
605 		soff = start_lbn;
606 	} else {
607 		idp = &start_ap[start_lvl - 1];
608 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
609 			brelse(sbp);
610 			return (ENOSPC);
611 		}
612 		sbap = (ufs1_daddr_t *)sbp->b_data;
613 		soff = idp->in_off;
614 	}
615 	/*
616 	 * If the block range spans two block maps, get the second map.
617 	 */
618 	ebap = NULL;
619 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
620 		ssize = len;
621 	} else {
622 #ifdef INVARIANTS
623 		if (start_lvl > 0 &&
624 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
625 			panic("ffs_reallocblk: start == end");
626 #endif
627 		ssize = len - (idp->in_off + 1);
628 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
629 			goto fail;
630 		ebap = (ufs1_daddr_t *)ebp->b_data;
631 	}
632 	/*
633 	 * Find the preferred location for the cluster. If we have not
634 	 * previously failed at this endeavor, then follow our standard
635 	 * preference calculation. If we have failed at it, then pick up
636 	 * where we last ended our search.
637 	 */
638 	UFS_LOCK(ump);
639 	if (ip->i_nextclustercg == -1)
640 		pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap);
641 	else
642 		pref = cgdata(fs, ip->i_nextclustercg);
643 	/*
644 	 * Search the block map looking for an allocation of the desired size.
645 	 * To avoid wasting too much time, we limit the number of cylinder
646 	 * groups that we will search.
647 	 */
648 	cg = dtog(fs, pref);
649 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
650 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
651 			break;
652 		cg += 1;
653 		if (cg >= fs->fs_ncg)
654 			cg = 0;
655 	}
656 	/*
657 	 * If we have failed in our search, record where we gave up for
658 	 * next time. Otherwise, fall back to our usual search citerion.
659 	 */
660 	if (newblk == 0) {
661 		ip->i_nextclustercg = cg;
662 		UFS_UNLOCK(ump);
663 		goto fail;
664 	}
665 	ip->i_nextclustercg = -1;
666 	/*
667 	 * We have found a new contiguous block.
668 	 *
669 	 * First we have to replace the old block pointers with the new
670 	 * block pointers in the inode and indirect blocks associated
671 	 * with the file.
672 	 */
673 #ifdef DEBUG
674 	if (prtrealloc)
675 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:",
676 		    (uintmax_t)ip->i_number,
677 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
678 #endif
679 	blkno = newblk;
680 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
681 		if (i == ssize) {
682 			bap = ebap;
683 			soff = -i;
684 		}
685 #ifdef INVARIANTS
686 		if (!ffs_checkblk(ip,
687 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
688 			panic("ffs_reallocblks: unallocated block 2");
689 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
690 			panic("ffs_reallocblks: alloc mismatch");
691 #endif
692 #ifdef DEBUG
693 		if (prtrealloc)
694 			printf(" %d,", *bap);
695 #endif
696 		if (DOINGSOFTDEP(vp)) {
697 			if (sbap == &ip->i_din1->di_db[0] && i < ssize)
698 				softdep_setup_allocdirect(ip, start_lbn + i,
699 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
700 				    buflist->bs_children[i]);
701 			else
702 				softdep_setup_allocindir_page(ip, start_lbn + i,
703 				    i < ssize ? sbp : ebp, soff + i, blkno,
704 				    *bap, buflist->bs_children[i]);
705 		}
706 		*bap++ = blkno;
707 	}
708 	/*
709 	 * Next we must write out the modified inode and indirect blocks.
710 	 * For strict correctness, the writes should be synchronous since
711 	 * the old block values may have been written to disk. In practise
712 	 * they are almost never written, but if we are concerned about
713 	 * strict correctness, the `doasyncfree' flag should be set to zero.
714 	 *
715 	 * The test on `doasyncfree' should be changed to test a flag
716 	 * that shows whether the associated buffers and inodes have
717 	 * been written. The flag should be set when the cluster is
718 	 * started and cleared whenever the buffer or inode is flushed.
719 	 * We can then check below to see if it is set, and do the
720 	 * synchronous write only when it has been cleared.
721 	 */
722 	if (sbap != &ip->i_din1->di_db[0]) {
723 		if (doasyncfree)
724 			bdwrite(sbp);
725 		else
726 			bwrite(sbp);
727 	} else {
728 		ip->i_flag |= IN_CHANGE | IN_UPDATE;
729 		if (!doasyncfree)
730 			ffs_update(vp, 1);
731 	}
732 	if (ssize < len) {
733 		if (doasyncfree)
734 			bdwrite(ebp);
735 		else
736 			bwrite(ebp);
737 	}
738 	/*
739 	 * Last, free the old blocks and assign the new blocks to the buffers.
740 	 */
741 #ifdef DEBUG
742 	if (prtrealloc)
743 		printf("\n\tnew:");
744 #endif
745 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
746 		bp = buflist->bs_children[i];
747 		if (!DOINGSOFTDEP(vp))
748 			/*
749 			 * The usual case is that a set of N-contiguous blocks
750 			 * that was just allocated has been replaced with a
751 			 * set of N+1-contiguous blocks. If they are marked as
752 			 * B_DELWRI, the current contents have not been written
753 			 * to disk. It is possible that the blocks were written
754 			 * earlier, but very uncommon. If the blocks have never
755 			 * been written, there is no need to send a BIO_DELETE
756 			 * for them when they are freed. The gain from avoiding
757 			 * the TRIMs for the common case of unwritten blocks
758 			 * far exceeds the cost of the write amplification for
759 			 * the uncommon case of failing to send a TRIM for the
760 			 * blocks that had been written.
761 			 */
762 			ffs_blkfree(ump, fs, ump->um_devvp,
763 			    dbtofsb(fs, bp->b_blkno),
764 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
765 			    (bp->b_flags & B_DELWRI) != 0 ? NOTRIM : SINGLETON);
766 		bp->b_blkno = fsbtodb(fs, blkno);
767 #ifdef INVARIANTS
768 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
769 			panic("ffs_reallocblks: unallocated block 3");
770 #endif
771 #ifdef DEBUG
772 		if (prtrealloc)
773 			printf(" %d,", blkno);
774 #endif
775 	}
776 #ifdef DEBUG
777 	if (prtrealloc) {
778 		prtrealloc--;
779 		printf("\n");
780 	}
781 #endif
782 	return (0);
783 
784 fail:
785 	if (ssize < len)
786 		brelse(ebp);
787 	if (sbap != &ip->i_din1->di_db[0])
788 		brelse(sbp);
789 	return (ENOSPC);
790 }
791 
792 static int
793 ffs_reallocblks_ufs2(ap)
794 	struct vop_reallocblks_args /* {
795 		struct vnode *a_vp;
796 		struct cluster_save *a_buflist;
797 	} */ *ap;
798 {
799 	struct fs *fs;
800 	struct inode *ip;
801 	struct vnode *vp;
802 	struct buf *sbp, *ebp, *bp;
803 	ufs2_daddr_t *bap, *sbap, *ebap;
804 	struct cluster_save *buflist;
805 	struct ufsmount *ump;
806 	ufs_lbn_t start_lbn, end_lbn;
807 	ufs2_daddr_t soff, newblk, blkno, pref;
808 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
809 	int i, cg, len, start_lvl, end_lvl, ssize;
810 
811 	vp = ap->a_vp;
812 	ip = VTOI(vp);
813 	ump = ITOUMP(ip);
814 	fs = ump->um_fs;
815 	/*
816 	 * If we are not tracking block clusters or if we have less than 4%
817 	 * free blocks left, then do not attempt to cluster. Running with
818 	 * less than 5% free block reserve is not recommended and those that
819 	 * choose to do so do not expect to have good file layout.
820 	 */
821 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
822 		return (ENOSPC);
823 	buflist = ap->a_buflist;
824 	len = buflist->bs_nchildren;
825 	start_lbn = buflist->bs_children[0]->b_lblkno;
826 	end_lbn = start_lbn + len - 1;
827 #ifdef INVARIANTS
828 	for (i = 0; i < len; i++)
829 		if (!ffs_checkblk(ip,
830 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
831 			panic("ffs_reallocblks: unallocated block 1");
832 	for (i = 1; i < len; i++)
833 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
834 			panic("ffs_reallocblks: non-logical cluster");
835 	blkno = buflist->bs_children[0]->b_blkno;
836 	ssize = fsbtodb(fs, fs->fs_frag);
837 	for (i = 1; i < len - 1; i++)
838 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
839 			panic("ffs_reallocblks: non-physical cluster %d", i);
840 #endif
841 	/*
842 	 * If the cluster crosses the boundary for the first indirect
843 	 * block, do not move anything in it. Indirect blocks are
844 	 * usually initially laid out in a position between the data
845 	 * blocks. Block reallocation would usually destroy locality by
846 	 * moving the indirect block out of the way to make room for
847 	 * data blocks if we didn't compensate here. We should also do
848 	 * this for other indirect block boundaries, but it is only
849 	 * important for the first one.
850 	 */
851 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
852 		return (ENOSPC);
853 	/*
854 	 * If the latest allocation is in a new cylinder group, assume that
855 	 * the filesystem has decided to move and do not force it back to
856 	 * the previous cylinder group.
857 	 */
858 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
859 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
860 		return (ENOSPC);
861 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
862 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
863 		return (ENOSPC);
864 	/*
865 	 * Get the starting offset and block map for the first block.
866 	 */
867 	if (start_lvl == 0) {
868 		sbap = &ip->i_din2->di_db[0];
869 		soff = start_lbn;
870 	} else {
871 		idp = &start_ap[start_lvl - 1];
872 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
873 			brelse(sbp);
874 			return (ENOSPC);
875 		}
876 		sbap = (ufs2_daddr_t *)sbp->b_data;
877 		soff = idp->in_off;
878 	}
879 	/*
880 	 * If the block range spans two block maps, get the second map.
881 	 */
882 	ebap = NULL;
883 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
884 		ssize = len;
885 	} else {
886 #ifdef INVARIANTS
887 		if (start_lvl > 0 &&
888 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
889 			panic("ffs_reallocblk: start == end");
890 #endif
891 		ssize = len - (idp->in_off + 1);
892 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
893 			goto fail;
894 		ebap = (ufs2_daddr_t *)ebp->b_data;
895 	}
896 	/*
897 	 * Find the preferred location for the cluster. If we have not
898 	 * previously failed at this endeavor, then follow our standard
899 	 * preference calculation. If we have failed at it, then pick up
900 	 * where we last ended our search.
901 	 */
902 	UFS_LOCK(ump);
903 	if (ip->i_nextclustercg == -1)
904 		pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap);
905 	else
906 		pref = cgdata(fs, ip->i_nextclustercg);
907 	/*
908 	 * Search the block map looking for an allocation of the desired size.
909 	 * To avoid wasting too much time, we limit the number of cylinder
910 	 * groups that we will search.
911 	 */
912 	cg = dtog(fs, pref);
913 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
914 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
915 			break;
916 		cg += 1;
917 		if (cg >= fs->fs_ncg)
918 			cg = 0;
919 	}
920 	/*
921 	 * If we have failed in our search, record where we gave up for
922 	 * next time. Otherwise, fall back to our usual search citerion.
923 	 */
924 	if (newblk == 0) {
925 		ip->i_nextclustercg = cg;
926 		UFS_UNLOCK(ump);
927 		goto fail;
928 	}
929 	ip->i_nextclustercg = -1;
930 	/*
931 	 * We have found a new contiguous block.
932 	 *
933 	 * First we have to replace the old block pointers with the new
934 	 * block pointers in the inode and indirect blocks associated
935 	 * with the file.
936 	 */
937 #ifdef DEBUG
938 	if (prtrealloc)
939 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number,
940 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
941 #endif
942 	blkno = newblk;
943 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
944 		if (i == ssize) {
945 			bap = ebap;
946 			soff = -i;
947 		}
948 #ifdef INVARIANTS
949 		if (!ffs_checkblk(ip,
950 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
951 			panic("ffs_reallocblks: unallocated block 2");
952 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
953 			panic("ffs_reallocblks: alloc mismatch");
954 #endif
955 #ifdef DEBUG
956 		if (prtrealloc)
957 			printf(" %jd,", (intmax_t)*bap);
958 #endif
959 		if (DOINGSOFTDEP(vp)) {
960 			if (sbap == &ip->i_din2->di_db[0] && i < ssize)
961 				softdep_setup_allocdirect(ip, start_lbn + i,
962 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
963 				    buflist->bs_children[i]);
964 			else
965 				softdep_setup_allocindir_page(ip, start_lbn + i,
966 				    i < ssize ? sbp : ebp, soff + i, blkno,
967 				    *bap, buflist->bs_children[i]);
968 		}
969 		*bap++ = blkno;
970 	}
971 	/*
972 	 * Next we must write out the modified inode and indirect blocks.
973 	 * For strict correctness, the writes should be synchronous since
974 	 * the old block values may have been written to disk. In practise
975 	 * they are almost never written, but if we are concerned about
976 	 * strict correctness, the `doasyncfree' flag should be set to zero.
977 	 *
978 	 * The test on `doasyncfree' should be changed to test a flag
979 	 * that shows whether the associated buffers and inodes have
980 	 * been written. The flag should be set when the cluster is
981 	 * started and cleared whenever the buffer or inode is flushed.
982 	 * We can then check below to see if it is set, and do the
983 	 * synchronous write only when it has been cleared.
984 	 */
985 	if (sbap != &ip->i_din2->di_db[0]) {
986 		if (doasyncfree)
987 			bdwrite(sbp);
988 		else
989 			bwrite(sbp);
990 	} else {
991 		ip->i_flag |= IN_CHANGE | IN_UPDATE;
992 		if (!doasyncfree)
993 			ffs_update(vp, 1);
994 	}
995 	if (ssize < len) {
996 		if (doasyncfree)
997 			bdwrite(ebp);
998 		else
999 			bwrite(ebp);
1000 	}
1001 	/*
1002 	 * Last, free the old blocks and assign the new blocks to the buffers.
1003 	 */
1004 #ifdef DEBUG
1005 	if (prtrealloc)
1006 		printf("\n\tnew:");
1007 #endif
1008 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
1009 		bp = buflist->bs_children[i];
1010 		if (!DOINGSOFTDEP(vp))
1011 			/*
1012 			 * The usual case is that a set of N-contiguous blocks
1013 			 * that was just allocated has been replaced with a
1014 			 * set of N+1-contiguous blocks. If they are marked as
1015 			 * B_DELWRI, the current contents have not been written
1016 			 * to disk. It is possible that the blocks were written
1017 			 * earlier, but very uncommon. If the blocks have never
1018 			 * been written, there is no need to send a BIO_DELETE
1019 			 * for them when they are freed. The gain from avoiding
1020 			 * the TRIMs for the common case of unwritten blocks
1021 			 * far exceeds the cost of the write amplification for
1022 			 * the uncommon case of failing to send a TRIM for the
1023 			 * blocks that had been written.
1024 			 */
1025 			ffs_blkfree(ump, fs, ump->um_devvp,
1026 			    dbtofsb(fs, bp->b_blkno),
1027 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
1028 			    (bp->b_flags & B_DELWRI) != 0 ? NOTRIM : SINGLETON);
1029 		bp->b_blkno = fsbtodb(fs, blkno);
1030 #ifdef INVARIANTS
1031 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
1032 			panic("ffs_reallocblks: unallocated block 3");
1033 #endif
1034 #ifdef DEBUG
1035 		if (prtrealloc)
1036 			printf(" %jd,", (intmax_t)blkno);
1037 #endif
1038 	}
1039 #ifdef DEBUG
1040 	if (prtrealloc) {
1041 		prtrealloc--;
1042 		printf("\n");
1043 	}
1044 #endif
1045 	return (0);
1046 
1047 fail:
1048 	if (ssize < len)
1049 		brelse(ebp);
1050 	if (sbap != &ip->i_din2->di_db[0])
1051 		brelse(sbp);
1052 	return (ENOSPC);
1053 }
1054 
1055 /*
1056  * Allocate an inode in the filesystem.
1057  *
1058  * If allocating a directory, use ffs_dirpref to select the inode.
1059  * If allocating in a directory, the following hierarchy is followed:
1060  *   1) allocate the preferred inode.
1061  *   2) allocate an inode in the same cylinder group.
1062  *   3) quadradically rehash into other cylinder groups, until an
1063  *      available inode is located.
1064  * If no inode preference is given the following hierarchy is used
1065  * to allocate an inode:
1066  *   1) allocate an inode in cylinder group 0.
1067  *   2) quadradically rehash into other cylinder groups, until an
1068  *      available inode is located.
1069  */
1070 int
1071 ffs_valloc(pvp, mode, cred, vpp)
1072 	struct vnode *pvp;
1073 	int mode;
1074 	struct ucred *cred;
1075 	struct vnode **vpp;
1076 {
1077 	struct inode *pip;
1078 	struct fs *fs;
1079 	struct inode *ip;
1080 	struct timespec ts;
1081 	struct ufsmount *ump;
1082 	ino_t ino, ipref;
1083 	u_int cg;
1084 	int error, error1, reclaimed;
1085 	static struct timeval lastfail;
1086 	static int curfail;
1087 
1088 	*vpp = NULL;
1089 	pip = VTOI(pvp);
1090 	ump = ITOUMP(pip);
1091 	fs = ump->um_fs;
1092 
1093 	UFS_LOCK(ump);
1094 	reclaimed = 0;
1095 retry:
1096 	if (fs->fs_cstotal.cs_nifree == 0)
1097 		goto noinodes;
1098 
1099 	if ((mode & IFMT) == IFDIR)
1100 		ipref = ffs_dirpref(pip);
1101 	else
1102 		ipref = pip->i_number;
1103 	if (ipref >= fs->fs_ncg * fs->fs_ipg)
1104 		ipref = 0;
1105 	cg = ino_to_cg(fs, ipref);
1106 	/*
1107 	 * Track number of dirs created one after another
1108 	 * in a same cg without intervening by files.
1109 	 */
1110 	if ((mode & IFMT) == IFDIR) {
1111 		if (fs->fs_contigdirs[cg] < 255)
1112 			fs->fs_contigdirs[cg]++;
1113 	} else {
1114 		if (fs->fs_contigdirs[cg] > 0)
1115 			fs->fs_contigdirs[cg]--;
1116 	}
1117 	ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0,
1118 					(allocfcn_t *)ffs_nodealloccg);
1119 	if (ino == 0)
1120 		goto noinodes;
1121 	error = ffs_vget(pvp->v_mount, ino, LK_EXCLUSIVE, vpp);
1122 	if (error) {
1123 		error1 = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp,
1124 		    FFSV_FORCEINSMQ);
1125 		ffs_vfree(pvp, ino, mode);
1126 		if (error1 == 0) {
1127 			ip = VTOI(*vpp);
1128 			if (ip->i_mode)
1129 				goto dup_alloc;
1130 			ip->i_flag |= IN_MODIFIED;
1131 			vput(*vpp);
1132 		}
1133 		return (error);
1134 	}
1135 	ip = VTOI(*vpp);
1136 	if (ip->i_mode) {
1137 dup_alloc:
1138 		printf("mode = 0%o, inum = %ju, fs = %s\n",
1139 		    ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt);
1140 		panic("ffs_valloc: dup alloc");
1141 	}
1142 	if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) {  /* XXX */
1143 		printf("free inode %s/%lu had %ld blocks\n",
1144 		    fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks));
1145 		DIP_SET(ip, i_blocks, 0);
1146 	}
1147 	ip->i_flags = 0;
1148 	DIP_SET(ip, i_flags, 0);
1149 	/*
1150 	 * Set up a new generation number for this inode.
1151 	 */
1152 	while (ip->i_gen == 0 || ++ip->i_gen == 0)
1153 		ip->i_gen = arc4random();
1154 	DIP_SET(ip, i_gen, ip->i_gen);
1155 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1156 		vfs_timestamp(&ts);
1157 		ip->i_din2->di_birthtime = ts.tv_sec;
1158 		ip->i_din2->di_birthnsec = ts.tv_nsec;
1159 	}
1160 	ufs_prepare_reclaim(*vpp);
1161 	ip->i_flag = 0;
1162 	(*vpp)->v_vflag = 0;
1163 	(*vpp)->v_type = VNON;
1164 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1165 		(*vpp)->v_op = &ffs_vnodeops2;
1166 		ip->i_flag |= IN_UFS2;
1167 	} else {
1168 		(*vpp)->v_op = &ffs_vnodeops1;
1169 	}
1170 	return (0);
1171 noinodes:
1172 	if (reclaimed == 0) {
1173 		reclaimed = 1;
1174 		softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT);
1175 		goto retry;
1176 	}
1177 	UFS_UNLOCK(ump);
1178 	if (ppsratecheck(&lastfail, &curfail, 1)) {
1179 		ffs_fserr(fs, pip->i_number, "out of inodes");
1180 		uprintf("\n%s: create/symlink failed, no inodes free\n",
1181 		    fs->fs_fsmnt);
1182 	}
1183 	return (ENOSPC);
1184 }
1185 
1186 /*
1187  * Find a cylinder group to place a directory.
1188  *
1189  * The policy implemented by this algorithm is to allocate a
1190  * directory inode in the same cylinder group as its parent
1191  * directory, but also to reserve space for its files inodes
1192  * and data. Restrict the number of directories which may be
1193  * allocated one after another in the same cylinder group
1194  * without intervening allocation of files.
1195  *
1196  * If we allocate a first level directory then force allocation
1197  * in another cylinder group.
1198  */
1199 static ino_t
1200 ffs_dirpref(pip)
1201 	struct inode *pip;
1202 {
1203 	struct fs *fs;
1204 	int cg, prefcg, dirsize, cgsize;
1205 	u_int avgifree, avgbfree, avgndir, curdirsize;
1206 	u_int minifree, minbfree, maxndir;
1207 	u_int mincg, minndir;
1208 	u_int maxcontigdirs;
1209 
1210 	mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED);
1211 	fs = ITOFS(pip);
1212 
1213 	avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg;
1214 	avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1215 	avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg;
1216 
1217 	/*
1218 	 * Force allocation in another cg if creating a first level dir.
1219 	 */
1220 	ASSERT_VOP_LOCKED(ITOV(pip), "ffs_dirpref");
1221 	if (ITOV(pip)->v_vflag & VV_ROOT) {
1222 		prefcg = arc4random() % fs->fs_ncg;
1223 		mincg = prefcg;
1224 		minndir = fs->fs_ipg;
1225 		for (cg = prefcg; cg < fs->fs_ncg; cg++)
1226 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1227 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1228 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1229 				mincg = cg;
1230 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1231 			}
1232 		for (cg = 0; cg < prefcg; cg++)
1233 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1234 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1235 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1236 				mincg = cg;
1237 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1238 			}
1239 		return ((ino_t)(fs->fs_ipg * mincg));
1240 	}
1241 
1242 	/*
1243 	 * Count various limits which used for
1244 	 * optimal allocation of a directory inode.
1245 	 */
1246 	maxndir = min(avgndir + fs->fs_ipg / 16, fs->fs_ipg);
1247 	minifree = avgifree - avgifree / 4;
1248 	if (minifree < 1)
1249 		minifree = 1;
1250 	minbfree = avgbfree - avgbfree / 4;
1251 	if (minbfree < 1)
1252 		minbfree = 1;
1253 	cgsize = fs->fs_fsize * fs->fs_fpg;
1254 	dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir;
1255 	curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0;
1256 	if (dirsize < curdirsize)
1257 		dirsize = curdirsize;
1258 	if (dirsize <= 0)
1259 		maxcontigdirs = 0;		/* dirsize overflowed */
1260 	else
1261 		maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255);
1262 	if (fs->fs_avgfpdir > 0)
1263 		maxcontigdirs = min(maxcontigdirs,
1264 				    fs->fs_ipg / fs->fs_avgfpdir);
1265 	if (maxcontigdirs == 0)
1266 		maxcontigdirs = 1;
1267 
1268 	/*
1269 	 * Limit number of dirs in one cg and reserve space for
1270 	 * regular files, but only if we have no deficit in
1271 	 * inodes or space.
1272 	 *
1273 	 * We are trying to find a suitable cylinder group nearby
1274 	 * our preferred cylinder group to place a new directory.
1275 	 * We scan from our preferred cylinder group forward looking
1276 	 * for a cylinder group that meets our criterion. If we get
1277 	 * to the final cylinder group and do not find anything,
1278 	 * we start scanning forwards from the beginning of the
1279 	 * filesystem. While it might seem sensible to start scanning
1280 	 * backwards or even to alternate looking forward and backward,
1281 	 * this approach fails badly when the filesystem is nearly full.
1282 	 * Specifically, we first search all the areas that have no space
1283 	 * and finally try the one preceding that. We repeat this on
1284 	 * every request and in the case of the final block end up
1285 	 * searching the entire filesystem. By jumping to the front
1286 	 * of the filesystem, our future forward searches always look
1287 	 * in new cylinder groups so finds every possible block after
1288 	 * one pass over the filesystem.
1289 	 */
1290 	prefcg = ino_to_cg(fs, pip->i_number);
1291 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1292 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1293 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1294 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1295 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1296 				return ((ino_t)(fs->fs_ipg * cg));
1297 		}
1298 	for (cg = 0; cg < prefcg; cg++)
1299 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1300 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1301 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1302 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1303 				return ((ino_t)(fs->fs_ipg * cg));
1304 		}
1305 	/*
1306 	 * This is a backstop when we have deficit in space.
1307 	 */
1308 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1309 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1310 			return ((ino_t)(fs->fs_ipg * cg));
1311 	for (cg = 0; cg < prefcg; cg++)
1312 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1313 			break;
1314 	return ((ino_t)(fs->fs_ipg * cg));
1315 }
1316 
1317 /*
1318  * Select the desired position for the next block in a file.  The file is
1319  * logically divided into sections. The first section is composed of the
1320  * direct blocks and the next fs_maxbpg blocks. Each additional section
1321  * contains fs_maxbpg blocks.
1322  *
1323  * If no blocks have been allocated in the first section, the policy is to
1324  * request a block in the same cylinder group as the inode that describes
1325  * the file. The first indirect is allocated immediately following the last
1326  * direct block and the data blocks for the first indirect immediately
1327  * follow it.
1328  *
1329  * If no blocks have been allocated in any other section, the indirect
1330  * block(s) are allocated in the same cylinder group as its inode in an
1331  * area reserved immediately following the inode blocks. The policy for
1332  * the data blocks is to place them in a cylinder group with a greater than
1333  * average number of free blocks. An appropriate cylinder group is found
1334  * by using a rotor that sweeps the cylinder groups. When a new group of
1335  * blocks is needed, the sweep begins in the cylinder group following the
1336  * cylinder group from which the previous allocation was made. The sweep
1337  * continues until a cylinder group with greater than the average number
1338  * of free blocks is found. If the allocation is for the first block in an
1339  * indirect block or the previous block is a hole, then the information on
1340  * the previous allocation is unavailable; here a best guess is made based
1341  * on the logical block number being allocated.
1342  *
1343  * If a section is already partially allocated, the policy is to
1344  * allocate blocks contiguously within the section if possible.
1345  */
1346 ufs2_daddr_t
1347 ffs_blkpref_ufs1(ip, lbn, indx, bap)
1348 	struct inode *ip;
1349 	ufs_lbn_t lbn;
1350 	int indx;
1351 	ufs1_daddr_t *bap;
1352 {
1353 	struct fs *fs;
1354 	u_int cg, inocg;
1355 	u_int avgbfree, startcg;
1356 	ufs2_daddr_t pref;
1357 
1358 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1359 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1360 	fs = ITOFS(ip);
1361 	/*
1362 	 * Allocation of indirect blocks is indicated by passing negative
1363 	 * values in indx: -1 for single indirect, -2 for double indirect,
1364 	 * -3 for triple indirect. As noted below, we attempt to allocate
1365 	 * the first indirect inline with the file data. For all later
1366 	 * indirect blocks, the data is often allocated in other cylinder
1367 	 * groups. However to speed random file access and to speed up
1368 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1369 	 * (typically half of fs_minfree) of the data area of each cylinder
1370 	 * group to hold these later indirect blocks.
1371 	 */
1372 	inocg = ino_to_cg(fs, ip->i_number);
1373 	if (indx < 0) {
1374 		/*
1375 		 * Our preference for indirect blocks is the zone at the
1376 		 * beginning of the inode's cylinder group data area that
1377 		 * we try to reserve for indirect blocks.
1378 		 */
1379 		pref = cgmeta(fs, inocg);
1380 		/*
1381 		 * If we are allocating the first indirect block, try to
1382 		 * place it immediately following the last direct block.
1383 		 */
1384 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1385 		    ip->i_din1->di_db[UFS_NDADDR - 1] != 0)
1386 			pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1387 		return (pref);
1388 	}
1389 	/*
1390 	 * If we are allocating the first data block in the first indirect
1391 	 * block and the indirect has been allocated in the data block area,
1392 	 * try to place it immediately following the indirect block.
1393 	 */
1394 	if (lbn == UFS_NDADDR) {
1395 		pref = ip->i_din1->di_ib[0];
1396 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1397 		    pref < cgbase(fs, inocg + 1))
1398 			return (pref + fs->fs_frag);
1399 	}
1400 	/*
1401 	 * If we are at the beginning of a file, or we have already allocated
1402 	 * the maximum number of blocks per cylinder group, or we do not
1403 	 * have a block allocated immediately preceding us, then we need
1404 	 * to decide where to start allocating new blocks.
1405 	 */
1406 	if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1407 		/*
1408 		 * If we are allocating a directory data block, we want
1409 		 * to place it in the metadata area.
1410 		 */
1411 		if ((ip->i_mode & IFMT) == IFDIR)
1412 			return (cgmeta(fs, inocg));
1413 		/*
1414 		 * Until we fill all the direct and all the first indirect's
1415 		 * blocks, we try to allocate in the data area of the inode's
1416 		 * cylinder group.
1417 		 */
1418 		if (lbn < UFS_NDADDR + NINDIR(fs))
1419 			return (cgdata(fs, inocg));
1420 		/*
1421 		 * Find a cylinder with greater than average number of
1422 		 * unused data blocks.
1423 		 */
1424 		if (indx == 0 || bap[indx - 1] == 0)
1425 			startcg = inocg + lbn / fs->fs_maxbpg;
1426 		else
1427 			startcg = dtog(fs, bap[indx - 1]) + 1;
1428 		startcg %= fs->fs_ncg;
1429 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1430 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1431 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1432 				fs->fs_cgrotor = cg;
1433 				return (cgdata(fs, cg));
1434 			}
1435 		for (cg = 0; cg <= startcg; cg++)
1436 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1437 				fs->fs_cgrotor = cg;
1438 				return (cgdata(fs, cg));
1439 			}
1440 		return (0);
1441 	}
1442 	/*
1443 	 * Otherwise, we just always try to lay things out contiguously.
1444 	 */
1445 	return (bap[indx - 1] + fs->fs_frag);
1446 }
1447 
1448 /*
1449  * Same as above, but for UFS2
1450  */
1451 ufs2_daddr_t
1452 ffs_blkpref_ufs2(ip, lbn, indx, bap)
1453 	struct inode *ip;
1454 	ufs_lbn_t lbn;
1455 	int indx;
1456 	ufs2_daddr_t *bap;
1457 {
1458 	struct fs *fs;
1459 	u_int cg, inocg;
1460 	u_int avgbfree, startcg;
1461 	ufs2_daddr_t pref;
1462 
1463 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1464 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1465 	fs = ITOFS(ip);
1466 	/*
1467 	 * Allocation of indirect blocks is indicated by passing negative
1468 	 * values in indx: -1 for single indirect, -2 for double indirect,
1469 	 * -3 for triple indirect. As noted below, we attempt to allocate
1470 	 * the first indirect inline with the file data. For all later
1471 	 * indirect blocks, the data is often allocated in other cylinder
1472 	 * groups. However to speed random file access and to speed up
1473 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1474 	 * (typically half of fs_minfree) of the data area of each cylinder
1475 	 * group to hold these later indirect blocks.
1476 	 */
1477 	inocg = ino_to_cg(fs, ip->i_number);
1478 	if (indx < 0) {
1479 		/*
1480 		 * Our preference for indirect blocks is the zone at the
1481 		 * beginning of the inode's cylinder group data area that
1482 		 * we try to reserve for indirect blocks.
1483 		 */
1484 		pref = cgmeta(fs, inocg);
1485 		/*
1486 		 * If we are allocating the first indirect block, try to
1487 		 * place it immediately following the last direct block.
1488 		 */
1489 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1490 		    ip->i_din2->di_db[UFS_NDADDR - 1] != 0)
1491 			pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1492 		return (pref);
1493 	}
1494 	/*
1495 	 * If we are allocating the first data block in the first indirect
1496 	 * block and the indirect has been allocated in the data block area,
1497 	 * try to place it immediately following the indirect block.
1498 	 */
1499 	if (lbn == UFS_NDADDR) {
1500 		pref = ip->i_din2->di_ib[0];
1501 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1502 		    pref < cgbase(fs, inocg + 1))
1503 			return (pref + fs->fs_frag);
1504 	}
1505 	/*
1506 	 * If we are at the beginning of a file, or we have already allocated
1507 	 * the maximum number of blocks per cylinder group, or we do not
1508 	 * have a block allocated immediately preceding us, then we need
1509 	 * to decide where to start allocating new blocks.
1510 	 */
1511 	if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1512 		/*
1513 		 * If we are allocating a directory data block, we want
1514 		 * to place it in the metadata area.
1515 		 */
1516 		if ((ip->i_mode & IFMT) == IFDIR)
1517 			return (cgmeta(fs, inocg));
1518 		/*
1519 		 * Until we fill all the direct and all the first indirect's
1520 		 * blocks, we try to allocate in the data area of the inode's
1521 		 * cylinder group.
1522 		 */
1523 		if (lbn < UFS_NDADDR + NINDIR(fs))
1524 			return (cgdata(fs, inocg));
1525 		/*
1526 		 * Find a cylinder with greater than average number of
1527 		 * unused data blocks.
1528 		 */
1529 		if (indx == 0 || bap[indx - 1] == 0)
1530 			startcg = inocg + lbn / fs->fs_maxbpg;
1531 		else
1532 			startcg = dtog(fs, bap[indx - 1]) + 1;
1533 		startcg %= fs->fs_ncg;
1534 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1535 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1536 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1537 				fs->fs_cgrotor = cg;
1538 				return (cgdata(fs, cg));
1539 			}
1540 		for (cg = 0; cg <= startcg; cg++)
1541 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1542 				fs->fs_cgrotor = cg;
1543 				return (cgdata(fs, cg));
1544 			}
1545 		return (0);
1546 	}
1547 	/*
1548 	 * Otherwise, we just always try to lay things out contiguously.
1549 	 */
1550 	return (bap[indx - 1] + fs->fs_frag);
1551 }
1552 
1553 /*
1554  * Implement the cylinder overflow algorithm.
1555  *
1556  * The policy implemented by this algorithm is:
1557  *   1) allocate the block in its requested cylinder group.
1558  *   2) quadradically rehash on the cylinder group number.
1559  *   3) brute force search for a free block.
1560  *
1561  * Must be called with the UFS lock held.  Will release the lock on success
1562  * and return with it held on failure.
1563  */
1564 /*VARARGS5*/
1565 static ufs2_daddr_t
1566 ffs_hashalloc(ip, cg, pref, size, rsize, allocator)
1567 	struct inode *ip;
1568 	u_int cg;
1569 	ufs2_daddr_t pref;
1570 	int size;	/* Search size for data blocks, mode for inodes */
1571 	int rsize;	/* Real allocated size. */
1572 	allocfcn_t *allocator;
1573 {
1574 	struct fs *fs;
1575 	ufs2_daddr_t result;
1576 	u_int i, icg = cg;
1577 
1578 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1579 #ifdef INVARIANTS
1580 	if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
1581 		panic("ffs_hashalloc: allocation on suspended filesystem");
1582 #endif
1583 	fs = ITOFS(ip);
1584 	/*
1585 	 * 1: preferred cylinder group
1586 	 */
1587 	result = (*allocator)(ip, cg, pref, size, rsize);
1588 	if (result)
1589 		return (result);
1590 	/*
1591 	 * 2: quadratic rehash
1592 	 */
1593 	for (i = 1; i < fs->fs_ncg; i *= 2) {
1594 		cg += i;
1595 		if (cg >= fs->fs_ncg)
1596 			cg -= fs->fs_ncg;
1597 		result = (*allocator)(ip, cg, 0, size, rsize);
1598 		if (result)
1599 			return (result);
1600 	}
1601 	/*
1602 	 * 3: brute force search
1603 	 * Note that we start at i == 2, since 0 was checked initially,
1604 	 * and 1 is always checked in the quadratic rehash.
1605 	 */
1606 	cg = (icg + 2) % fs->fs_ncg;
1607 	for (i = 2; i < fs->fs_ncg; i++) {
1608 		result = (*allocator)(ip, cg, 0, size, rsize);
1609 		if (result)
1610 			return (result);
1611 		cg++;
1612 		if (cg == fs->fs_ncg)
1613 			cg = 0;
1614 	}
1615 	return (0);
1616 }
1617 
1618 /*
1619  * Determine whether a fragment can be extended.
1620  *
1621  * Check to see if the necessary fragments are available, and
1622  * if they are, allocate them.
1623  */
1624 static ufs2_daddr_t
1625 ffs_fragextend(ip, cg, bprev, osize, nsize)
1626 	struct inode *ip;
1627 	u_int cg;
1628 	ufs2_daddr_t bprev;
1629 	int osize, nsize;
1630 {
1631 	struct fs *fs;
1632 	struct cg *cgp;
1633 	struct buf *bp;
1634 	struct ufsmount *ump;
1635 	int nffree;
1636 	long bno;
1637 	int frags, bbase;
1638 	int i, error;
1639 	u_int8_t *blksfree;
1640 
1641 	ump = ITOUMP(ip);
1642 	fs = ump->um_fs;
1643 	if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize))
1644 		return (0);
1645 	frags = numfrags(fs, nsize);
1646 	bbase = fragnum(fs, bprev);
1647 	if (bbase > fragnum(fs, (bprev + frags - 1))) {
1648 		/* cannot extend across a block boundary */
1649 		return (0);
1650 	}
1651 	UFS_UNLOCK(ump);
1652 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0)
1653 		goto fail;
1654 	bno = dtogd(fs, bprev);
1655 	blksfree = cg_blksfree(cgp);
1656 	for (i = numfrags(fs, osize); i < frags; i++)
1657 		if (isclr(blksfree, bno + i))
1658 			goto fail;
1659 	/*
1660 	 * the current fragment can be extended
1661 	 * deduct the count on fragment being extended into
1662 	 * increase the count on the remaining fragment (if any)
1663 	 * allocate the extended piece
1664 	 */
1665 	for (i = frags; i < fs->fs_frag - bbase; i++)
1666 		if (isclr(blksfree, bno + i))
1667 			break;
1668 	cgp->cg_frsum[i - numfrags(fs, osize)]--;
1669 	if (i != frags)
1670 		cgp->cg_frsum[i - frags]++;
1671 	for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) {
1672 		clrbit(blksfree, bno + i);
1673 		cgp->cg_cs.cs_nffree--;
1674 		nffree++;
1675 	}
1676 	UFS_LOCK(ump);
1677 	fs->fs_cstotal.cs_nffree -= nffree;
1678 	fs->fs_cs(fs, cg).cs_nffree -= nffree;
1679 	fs->fs_fmod = 1;
1680 	ACTIVECLEAR(fs, cg);
1681 	UFS_UNLOCK(ump);
1682 	if (DOINGSOFTDEP(ITOV(ip)))
1683 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev,
1684 		    frags, numfrags(fs, osize));
1685 	bdwrite(bp);
1686 	return (bprev);
1687 
1688 fail:
1689 	brelse(bp);
1690 	UFS_LOCK(ump);
1691 	return (0);
1692 
1693 }
1694 
1695 /*
1696  * Determine whether a block can be allocated.
1697  *
1698  * Check to see if a block of the appropriate size is available,
1699  * and if it is, allocate it.
1700  */
1701 static ufs2_daddr_t
1702 ffs_alloccg(ip, cg, bpref, size, rsize)
1703 	struct inode *ip;
1704 	u_int cg;
1705 	ufs2_daddr_t bpref;
1706 	int size;
1707 	int rsize;
1708 {
1709 	struct fs *fs;
1710 	struct cg *cgp;
1711 	struct buf *bp;
1712 	struct ufsmount *ump;
1713 	ufs1_daddr_t bno;
1714 	ufs2_daddr_t blkno;
1715 	int i, allocsiz, error, frags;
1716 	u_int8_t *blksfree;
1717 
1718 	ump = ITOUMP(ip);
1719 	fs = ump->um_fs;
1720 	if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize)
1721 		return (0);
1722 	UFS_UNLOCK(ump);
1723 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0 ||
1724 	   (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize))
1725 		goto fail;
1726 	if (size == fs->fs_bsize) {
1727 		UFS_LOCK(ump);
1728 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1729 		ACTIVECLEAR(fs, cg);
1730 		UFS_UNLOCK(ump);
1731 		bdwrite(bp);
1732 		return (blkno);
1733 	}
1734 	/*
1735 	 * check to see if any fragments are already available
1736 	 * allocsiz is the size which will be allocated, hacking
1737 	 * it down to a smaller size if necessary
1738 	 */
1739 	blksfree = cg_blksfree(cgp);
1740 	frags = numfrags(fs, size);
1741 	for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++)
1742 		if (cgp->cg_frsum[allocsiz] != 0)
1743 			break;
1744 	if (allocsiz == fs->fs_frag) {
1745 		/*
1746 		 * no fragments were available, so a block will be
1747 		 * allocated, and hacked up
1748 		 */
1749 		if (cgp->cg_cs.cs_nbfree == 0)
1750 			goto fail;
1751 		UFS_LOCK(ump);
1752 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1753 		ACTIVECLEAR(fs, cg);
1754 		UFS_UNLOCK(ump);
1755 		bdwrite(bp);
1756 		return (blkno);
1757 	}
1758 	KASSERT(size == rsize,
1759 	    ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize));
1760 	bno = ffs_mapsearch(fs, cgp, bpref, allocsiz);
1761 	if (bno < 0)
1762 		goto fail;
1763 	for (i = 0; i < frags; i++)
1764 		clrbit(blksfree, bno + i);
1765 	cgp->cg_cs.cs_nffree -= frags;
1766 	cgp->cg_frsum[allocsiz]--;
1767 	if (frags != allocsiz)
1768 		cgp->cg_frsum[allocsiz - frags]++;
1769 	UFS_LOCK(ump);
1770 	fs->fs_cstotal.cs_nffree -= frags;
1771 	fs->fs_cs(fs, cg).cs_nffree -= frags;
1772 	fs->fs_fmod = 1;
1773 	blkno = cgbase(fs, cg) + bno;
1774 	ACTIVECLEAR(fs, cg);
1775 	UFS_UNLOCK(ump);
1776 	if (DOINGSOFTDEP(ITOV(ip)))
1777 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0);
1778 	bdwrite(bp);
1779 	return (blkno);
1780 
1781 fail:
1782 	brelse(bp);
1783 	UFS_LOCK(ump);
1784 	return (0);
1785 }
1786 
1787 /*
1788  * Allocate a block in a cylinder group.
1789  *
1790  * This algorithm implements the following policy:
1791  *   1) allocate the requested block.
1792  *   2) allocate a rotationally optimal block in the same cylinder.
1793  *   3) allocate the next available block on the block rotor for the
1794  *      specified cylinder group.
1795  * Note that this routine only allocates fs_bsize blocks; these
1796  * blocks may be fragmented by the routine that allocates them.
1797  */
1798 static ufs2_daddr_t
1799 ffs_alloccgblk(ip, bp, bpref, size)
1800 	struct inode *ip;
1801 	struct buf *bp;
1802 	ufs2_daddr_t bpref;
1803 	int size;
1804 {
1805 	struct fs *fs;
1806 	struct cg *cgp;
1807 	struct ufsmount *ump;
1808 	ufs1_daddr_t bno;
1809 	ufs2_daddr_t blkno;
1810 	u_int8_t *blksfree;
1811 	int i, cgbpref;
1812 
1813 	ump = ITOUMP(ip);
1814 	fs = ump->um_fs;
1815 	mtx_assert(UFS_MTX(ump), MA_OWNED);
1816 	cgp = (struct cg *)bp->b_data;
1817 	blksfree = cg_blksfree(cgp);
1818 	if (bpref == 0) {
1819 		bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag;
1820 	} else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) {
1821 		/* map bpref to correct zone in this cg */
1822 		if (bpref < cgdata(fs, cgbpref))
1823 			bpref = cgmeta(fs, cgp->cg_cgx);
1824 		else
1825 			bpref = cgdata(fs, cgp->cg_cgx);
1826 	}
1827 	/*
1828 	 * if the requested block is available, use it
1829 	 */
1830 	bno = dtogd(fs, blknum(fs, bpref));
1831 	if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno)))
1832 		goto gotit;
1833 	/*
1834 	 * Take the next available block in this cylinder group.
1835 	 */
1836 	bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag);
1837 	if (bno < 0)
1838 		return (0);
1839 	/* Update cg_rotor only if allocated from the data zone */
1840 	if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx)))
1841 		cgp->cg_rotor = bno;
1842 gotit:
1843 	blkno = fragstoblks(fs, bno);
1844 	ffs_clrblock(fs, blksfree, (long)blkno);
1845 	ffs_clusteracct(fs, cgp, blkno, -1);
1846 	cgp->cg_cs.cs_nbfree--;
1847 	fs->fs_cstotal.cs_nbfree--;
1848 	fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--;
1849 	fs->fs_fmod = 1;
1850 	blkno = cgbase(fs, cgp->cg_cgx) + bno;
1851 	/*
1852 	 * If the caller didn't want the whole block free the frags here.
1853 	 */
1854 	size = numfrags(fs, size);
1855 	if (size != fs->fs_frag) {
1856 		bno = dtogd(fs, blkno);
1857 		for (i = size; i < fs->fs_frag; i++)
1858 			setbit(blksfree, bno + i);
1859 		i = fs->fs_frag - size;
1860 		cgp->cg_cs.cs_nffree += i;
1861 		fs->fs_cstotal.cs_nffree += i;
1862 		fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i;
1863 		fs->fs_fmod = 1;
1864 		cgp->cg_frsum[i]++;
1865 	}
1866 	/* XXX Fixme. */
1867 	UFS_UNLOCK(ump);
1868 	if (DOINGSOFTDEP(ITOV(ip)))
1869 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0);
1870 	UFS_LOCK(ump);
1871 	return (blkno);
1872 }
1873 
1874 /*
1875  * Determine whether a cluster can be allocated.
1876  *
1877  * We do not currently check for optimal rotational layout if there
1878  * are multiple choices in the same cylinder group. Instead we just
1879  * take the first one that we find following bpref.
1880  */
1881 static ufs2_daddr_t
1882 ffs_clusteralloc(ip, cg, bpref, len)
1883 	struct inode *ip;
1884 	u_int cg;
1885 	ufs2_daddr_t bpref;
1886 	int len;
1887 {
1888 	struct fs *fs;
1889 	struct cg *cgp;
1890 	struct buf *bp;
1891 	struct ufsmount *ump;
1892 	int i, run, bit, map, got, error;
1893 	ufs2_daddr_t bno;
1894 	u_char *mapp;
1895 	int32_t *lp;
1896 	u_int8_t *blksfree;
1897 
1898 	ump = ITOUMP(ip);
1899 	fs = ump->um_fs;
1900 	if (fs->fs_maxcluster[cg] < len)
1901 		return (0);
1902 	UFS_UNLOCK(ump);
1903 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) {
1904 		UFS_LOCK(ump);
1905 		return (0);
1906 	}
1907 	/*
1908 	 * Check to see if a cluster of the needed size (or bigger) is
1909 	 * available in this cylinder group.
1910 	 */
1911 	lp = &cg_clustersum(cgp)[len];
1912 	for (i = len; i <= fs->fs_contigsumsize; i++)
1913 		if (*lp++ > 0)
1914 			break;
1915 	if (i > fs->fs_contigsumsize) {
1916 		/*
1917 		 * This is the first time looking for a cluster in this
1918 		 * cylinder group. Update the cluster summary information
1919 		 * to reflect the true maximum sized cluster so that
1920 		 * future cluster allocation requests can avoid reading
1921 		 * the cylinder group map only to find no clusters.
1922 		 */
1923 		lp = &cg_clustersum(cgp)[len - 1];
1924 		for (i = len - 1; i > 0; i--)
1925 			if (*lp-- > 0)
1926 				break;
1927 		UFS_LOCK(ump);
1928 		fs->fs_maxcluster[cg] = i;
1929 		brelse(bp);
1930 		return (0);
1931 	}
1932 	/*
1933 	 * Search the cluster map to find a big enough cluster.
1934 	 * We take the first one that we find, even if it is larger
1935 	 * than we need as we prefer to get one close to the previous
1936 	 * block allocation. We do not search before the current
1937 	 * preference point as we do not want to allocate a block
1938 	 * that is allocated before the previous one (as we will
1939 	 * then have to wait for another pass of the elevator
1940 	 * algorithm before it will be read). We prefer to fail and
1941 	 * be recalled to try an allocation in the next cylinder group.
1942 	 */
1943 	if (dtog(fs, bpref) != cg)
1944 		bpref = cgdata(fs, cg);
1945 	else
1946 		bpref = blknum(fs, bpref);
1947 	bpref = fragstoblks(fs, dtogd(fs, bpref));
1948 	mapp = &cg_clustersfree(cgp)[bpref / NBBY];
1949 	map = *mapp++;
1950 	bit = 1 << (bpref % NBBY);
1951 	for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) {
1952 		if ((map & bit) == 0) {
1953 			run = 0;
1954 		} else {
1955 			run++;
1956 			if (run == len)
1957 				break;
1958 		}
1959 		if ((got & (NBBY - 1)) != (NBBY - 1)) {
1960 			bit <<= 1;
1961 		} else {
1962 			map = *mapp++;
1963 			bit = 1;
1964 		}
1965 	}
1966 	if (got >= cgp->cg_nclusterblks) {
1967 		UFS_LOCK(ump);
1968 		brelse(bp);
1969 		return (0);
1970 	}
1971 	/*
1972 	 * Allocate the cluster that we have found.
1973 	 */
1974 	blksfree = cg_blksfree(cgp);
1975 	for (i = 1; i <= len; i++)
1976 		if (!ffs_isblock(fs, blksfree, got - run + i))
1977 			panic("ffs_clusteralloc: map mismatch");
1978 	bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1);
1979 	if (dtog(fs, bno) != cg)
1980 		panic("ffs_clusteralloc: allocated out of group");
1981 	len = blkstofrags(fs, len);
1982 	UFS_LOCK(ump);
1983 	for (i = 0; i < len; i += fs->fs_frag)
1984 		if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i)
1985 			panic("ffs_clusteralloc: lost block");
1986 	ACTIVECLEAR(fs, cg);
1987 	UFS_UNLOCK(ump);
1988 	bdwrite(bp);
1989 	return (bno);
1990 }
1991 
1992 static inline struct buf *
1993 getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags)
1994 {
1995 	struct fs *fs;
1996 
1997 	fs = ITOFS(ip);
1998 	return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs,
1999 	    cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
2000 	    gbflags));
2001 }
2002 
2003 /*
2004  * Synchronous inode initialization is needed only when barrier writes do not
2005  * work as advertised, and will impose a heavy cost on file creation in a newly
2006  * created filesystem.
2007  */
2008 static int doasyncinodeinit = 1;
2009 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN,
2010     &doasyncinodeinit, 0,
2011     "Perform inode block initialization using asynchronous writes");
2012 
2013 /*
2014  * Determine whether an inode can be allocated.
2015  *
2016  * Check to see if an inode is available, and if it is,
2017  * allocate it using the following policy:
2018  *   1) allocate the requested inode.
2019  *   2) allocate the next available inode after the requested
2020  *      inode in the specified cylinder group.
2021  */
2022 static ufs2_daddr_t
2023 ffs_nodealloccg(ip, cg, ipref, mode, unused)
2024 	struct inode *ip;
2025 	u_int cg;
2026 	ufs2_daddr_t ipref;
2027 	int mode;
2028 	int unused;
2029 {
2030 	struct fs *fs;
2031 	struct cg *cgp;
2032 	struct buf *bp, *ibp;
2033 	struct ufsmount *ump;
2034 	u_int8_t *inosused, *loc;
2035 	struct ufs2_dinode *dp2;
2036 	int error, start, len, i;
2037 	u_int32_t old_initediblk;
2038 
2039 	ump = ITOUMP(ip);
2040 	fs = ump->um_fs;
2041 check_nifree:
2042 	if (fs->fs_cs(fs, cg).cs_nifree == 0)
2043 		return (0);
2044 	UFS_UNLOCK(ump);
2045 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) {
2046 		UFS_LOCK(ump);
2047 		return (0);
2048 	}
2049 restart:
2050 	if (cgp->cg_cs.cs_nifree == 0) {
2051 		brelse(bp);
2052 		UFS_LOCK(ump);
2053 		return (0);
2054 	}
2055 	inosused = cg_inosused(cgp);
2056 	if (ipref) {
2057 		ipref %= fs->fs_ipg;
2058 		if (isclr(inosused, ipref))
2059 			goto gotit;
2060 	}
2061 	start = cgp->cg_irotor / NBBY;
2062 	len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY);
2063 	loc = memcchr(&inosused[start], 0xff, len);
2064 	if (loc == NULL) {
2065 		len = start + 1;
2066 		start = 0;
2067 		loc = memcchr(&inosused[start], 0xff, len);
2068 		if (loc == NULL) {
2069 			printf("cg = %d, irotor = %ld, fs = %s\n",
2070 			    cg, (long)cgp->cg_irotor, fs->fs_fsmnt);
2071 			panic("ffs_nodealloccg: map corrupted");
2072 			/* NOTREACHED */
2073 		}
2074 	}
2075 	ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
2076 gotit:
2077 	/*
2078 	 * Check to see if we need to initialize more inodes.
2079 	 */
2080 	if (fs->fs_magic == FS_UFS2_MAGIC &&
2081 	    ipref + INOPB(fs) > cgp->cg_initediblk &&
2082 	    cgp->cg_initediblk < cgp->cg_niblk) {
2083 		old_initediblk = cgp->cg_initediblk;
2084 
2085 		/*
2086 		 * Free the cylinder group lock before writing the
2087 		 * initialized inode block.  Entering the
2088 		 * babarrierwrite() with the cylinder group lock
2089 		 * causes lock order violation between the lock and
2090 		 * snaplk.
2091 		 *
2092 		 * Another thread can decide to initialize the same
2093 		 * inode block, but whichever thread first gets the
2094 		 * cylinder group lock after writing the newly
2095 		 * allocated inode block will update it and the other
2096 		 * will realize that it has lost and leave the
2097 		 * cylinder group unchanged.
2098 		 */
2099 		ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
2100 		brelse(bp);
2101 		if (ibp == NULL) {
2102 			/*
2103 			 * The inode block buffer is already owned by
2104 			 * another thread, which must initialize it.
2105 			 * Wait on the buffer to allow another thread
2106 			 * to finish the updates, with dropped cg
2107 			 * buffer lock, then retry.
2108 			 */
2109 			ibp = getinobuf(ip, cg, old_initediblk, 0);
2110 			brelse(ibp);
2111 			UFS_LOCK(ump);
2112 			goto check_nifree;
2113 		}
2114 		bzero(ibp->b_data, (int)fs->fs_bsize);
2115 		dp2 = (struct ufs2_dinode *)(ibp->b_data);
2116 		for (i = 0; i < INOPB(fs); i++) {
2117 			while (dp2->di_gen == 0)
2118 				dp2->di_gen = arc4random();
2119 			dp2++;
2120 		}
2121 
2122 		/*
2123 		 * Rather than adding a soft updates dependency to ensure
2124 		 * that the new inode block is written before it is claimed
2125 		 * by the cylinder group map, we just do a barrier write
2126 		 * here. The barrier write will ensure that the inode block
2127 		 * gets written before the updated cylinder group map can be
2128 		 * written. The barrier write should only slow down bulk
2129 		 * loading of newly created filesystems.
2130 		 */
2131 		if (doasyncinodeinit)
2132 			babarrierwrite(ibp);
2133 		else
2134 			bwrite(ibp);
2135 
2136 		/*
2137 		 * After the inode block is written, try to update the
2138 		 * cg initediblk pointer.  If another thread beat us
2139 		 * to it, then leave it unchanged as the other thread
2140 		 * has already set it correctly.
2141 		 */
2142 		error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp);
2143 		UFS_LOCK(ump);
2144 		ACTIVECLEAR(fs, cg);
2145 		UFS_UNLOCK(ump);
2146 		if (error != 0)
2147 			return (error);
2148 		if (cgp->cg_initediblk == old_initediblk)
2149 			cgp->cg_initediblk += INOPB(fs);
2150 		goto restart;
2151 	}
2152 	cgp->cg_irotor = ipref;
2153 	UFS_LOCK(ump);
2154 	ACTIVECLEAR(fs, cg);
2155 	setbit(inosused, ipref);
2156 	cgp->cg_cs.cs_nifree--;
2157 	fs->fs_cstotal.cs_nifree--;
2158 	fs->fs_cs(fs, cg).cs_nifree--;
2159 	fs->fs_fmod = 1;
2160 	if ((mode & IFMT) == IFDIR) {
2161 		cgp->cg_cs.cs_ndir++;
2162 		fs->fs_cstotal.cs_ndir++;
2163 		fs->fs_cs(fs, cg).cs_ndir++;
2164 	}
2165 	UFS_UNLOCK(ump);
2166 	if (DOINGSOFTDEP(ITOV(ip)))
2167 		softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode);
2168 	bdwrite(bp);
2169 	return ((ino_t)(cg * fs->fs_ipg + ipref));
2170 }
2171 
2172 /*
2173  * Free a block or fragment.
2174  *
2175  * The specified block or fragment is placed back in the
2176  * free map. If a fragment is deallocated, a possible
2177  * block reassembly is checked.
2178  */
2179 static void
2180 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd)
2181 	struct ufsmount *ump;
2182 	struct fs *fs;
2183 	struct vnode *devvp;
2184 	ufs2_daddr_t bno;
2185 	long size;
2186 	ino_t inum;
2187 	struct workhead *dephd;
2188 {
2189 	struct mount *mp;
2190 	struct cg *cgp;
2191 	struct buf *bp;
2192 	ufs1_daddr_t fragno, cgbno;
2193 	int i, blk, frags, bbase, error;
2194 	u_int cg;
2195 	u_int8_t *blksfree;
2196 	struct cdev *dev;
2197 
2198 	cg = dtog(fs, bno);
2199 	if (devvp->v_type == VREG) {
2200 		/* devvp is a snapshot */
2201 		MPASS(devvp->v_mount->mnt_data == ump);
2202 		dev = ump->um_devvp->v_rdev;
2203 	} else if (devvp->v_type == VCHR) {
2204 		/* devvp is a normal disk device */
2205 		dev = devvp->v_rdev;
2206 		ASSERT_VOP_LOCKED(devvp, "ffs_blkfree_cg");
2207 	} else
2208 		return;
2209 #ifdef INVARIANTS
2210 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 ||
2211 	    fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) {
2212 		printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n",
2213 		    devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize,
2214 		    size, fs->fs_fsmnt);
2215 		panic("ffs_blkfree_cg: bad size");
2216 	}
2217 #endif
2218 	if ((u_int)bno >= fs->fs_size) {
2219 		printf("bad block %jd, ino %lu\n", (intmax_t)bno,
2220 		    (u_long)inum);
2221 		ffs_fserr(fs, inum, "bad block");
2222 		return;
2223 	}
2224 	if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2225 		return;
2226 	cgbno = dtogd(fs, bno);
2227 	blksfree = cg_blksfree(cgp);
2228 	UFS_LOCK(ump);
2229 	if (size == fs->fs_bsize) {
2230 		fragno = fragstoblks(fs, cgbno);
2231 		if (!ffs_isfreeblock(fs, blksfree, fragno)) {
2232 			if (devvp->v_type == VREG) {
2233 				UFS_UNLOCK(ump);
2234 				/* devvp is a snapshot */
2235 				brelse(bp);
2236 				return;
2237 			}
2238 			printf("dev = %s, block = %jd, fs = %s\n",
2239 			    devtoname(dev), (intmax_t)bno, fs->fs_fsmnt);
2240 			panic("ffs_blkfree_cg: freeing free block");
2241 		}
2242 		ffs_setblock(fs, blksfree, fragno);
2243 		ffs_clusteracct(fs, cgp, fragno, 1);
2244 		cgp->cg_cs.cs_nbfree++;
2245 		fs->fs_cstotal.cs_nbfree++;
2246 		fs->fs_cs(fs, cg).cs_nbfree++;
2247 	} else {
2248 		bbase = cgbno - fragnum(fs, cgbno);
2249 		/*
2250 		 * decrement the counts associated with the old frags
2251 		 */
2252 		blk = blkmap(fs, blksfree, bbase);
2253 		ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
2254 		/*
2255 		 * deallocate the fragment
2256 		 */
2257 		frags = numfrags(fs, size);
2258 		for (i = 0; i < frags; i++) {
2259 			if (isset(blksfree, cgbno + i)) {
2260 				printf("dev = %s, block = %jd, fs = %s\n",
2261 				    devtoname(dev), (intmax_t)(bno + i),
2262 				    fs->fs_fsmnt);
2263 				panic("ffs_blkfree_cg: freeing free frag");
2264 			}
2265 			setbit(blksfree, cgbno + i);
2266 		}
2267 		cgp->cg_cs.cs_nffree += i;
2268 		fs->fs_cstotal.cs_nffree += i;
2269 		fs->fs_cs(fs, cg).cs_nffree += i;
2270 		/*
2271 		 * add back in counts associated with the new frags
2272 		 */
2273 		blk = blkmap(fs, blksfree, bbase);
2274 		ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
2275 		/*
2276 		 * if a complete block has been reassembled, account for it
2277 		 */
2278 		fragno = fragstoblks(fs, bbase);
2279 		if (ffs_isblock(fs, blksfree, fragno)) {
2280 			cgp->cg_cs.cs_nffree -= fs->fs_frag;
2281 			fs->fs_cstotal.cs_nffree -= fs->fs_frag;
2282 			fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag;
2283 			ffs_clusteracct(fs, cgp, fragno, 1);
2284 			cgp->cg_cs.cs_nbfree++;
2285 			fs->fs_cstotal.cs_nbfree++;
2286 			fs->fs_cs(fs, cg).cs_nbfree++;
2287 		}
2288 	}
2289 	fs->fs_fmod = 1;
2290 	ACTIVECLEAR(fs, cg);
2291 	UFS_UNLOCK(ump);
2292 	mp = UFSTOVFS(ump);
2293 	if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR)
2294 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2295 		    numfrags(fs, size), dephd);
2296 	bdwrite(bp);
2297 }
2298 
2299 /*
2300  * Structures and routines associated with trim management.
2301  */
2302 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures");
2303 
2304 #define	TRIMLIST_HASH(ump, inum) \
2305 	(&(ump)->um_trimhash[(inum) & (ump)->um_trimlisthashsize])
2306 
2307 static void	ffs_blkfree_trim_completed(struct buf *);
2308 static void	ffs_blkfree_trim_task(void *ctx, int pending __unused);
2309 
2310 struct ffs_blkfree_trim_params {
2311 	struct task task;
2312 	struct ufsmount *ump;
2313 	struct vnode *devvp;
2314 	ufs2_daddr_t bno;
2315 	long size;
2316 	ino_t inum;
2317 	struct workhead *pdephd;
2318 	struct workhead dephd;
2319 };
2320 
2321 static void
2322 ffs_blkfree_trim_task(ctx, pending)
2323 	void *ctx;
2324 	int pending;
2325 {
2326 	struct ffs_blkfree_trim_params *tp;
2327 
2328 	tp = ctx;
2329 	ffs_blkfree_cg(tp->ump, tp->ump->um_fs, tp->devvp, tp->bno, tp->size,
2330 	    tp->inum, tp->pdephd);
2331 	vn_finished_secondary_write(UFSTOVFS(tp->ump));
2332 	atomic_add_int(&tp->ump->um_trim_inflight, -1);
2333 	free(tp, M_TRIM);
2334 }
2335 
2336 static void
2337 ffs_blkfree_trim_completed(bp)
2338 	struct buf *bp;
2339 {
2340 	struct ffs_blkfree_trim_params *tp;
2341 
2342 	tp = bp->b_fsprivate1;
2343 	free(bp, M_TRIM);
2344 	TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp);
2345 	taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task);
2346 }
2347 
2348 void
2349 ffs_blkfree(ump, fs, devvp, bno, size, inum, vtype, dephd, trimtype)
2350 	struct ufsmount *ump;
2351 	struct fs *fs;
2352 	struct vnode *devvp;
2353 	ufs2_daddr_t bno;
2354 	long size;
2355 	ino_t inum;
2356 	enum vtype vtype;
2357 	struct workhead *dephd;
2358 	int trimtype;
2359 {
2360 	struct mount *mp;
2361 	struct buf *bp;
2362 	struct ffs_blkfree_trim_params *tp;
2363 
2364 	/*
2365 	 * Check to see if a snapshot wants to claim the block.
2366 	 * Check that devvp is a normal disk device, not a snapshot,
2367 	 * it has a snapshot(s) associated with it, and one of the
2368 	 * snapshots wants to claim the block.
2369 	 */
2370 	if (devvp->v_type == VCHR &&
2371 	    (devvp->v_vflag & VV_COPYONWRITE) &&
2372 	    ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) {
2373 		return;
2374 	}
2375 	/*
2376 	 * Nothing to delay if TRIM is not required for this block or TRIM
2377 	 * is disabled or the operation is performed on a snapshot.
2378 	 */
2379 	if (trimtype == NOTRIM || ((ump->um_flags & UM_CANDELETE) == 0) ||
2380 	    devvp->v_type == VREG) {
2381 		ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd);
2382 		return;
2383 	}
2384 
2385 	/*
2386 	 * Postpone the set of the free bit in the cg bitmap until the
2387 	 * BIO_DELETE is completed.  Otherwise, due to disk queue
2388 	 * reordering, TRIM might be issued after we reuse the block
2389 	 * and write some new data into it.
2390 	 */
2391 	atomic_add_int(&ump->um_trim_inflight, 1);
2392 	tp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK);
2393 	tp->ump = ump;
2394 	tp->devvp = devvp;
2395 	tp->bno = bno;
2396 	tp->size = size;
2397 	tp->inum = inum;
2398 	if (dephd != NULL) {
2399 		LIST_INIT(&tp->dephd);
2400 		LIST_SWAP(dephd, &tp->dephd, worklist, wk_list);
2401 		tp->pdephd = &tp->dephd;
2402 	} else
2403 		tp->pdephd = NULL;
2404 
2405 	bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO);
2406 	bp->b_iocmd = BIO_DELETE;
2407 	bp->b_iooffset = dbtob(fsbtodb(fs, bno));
2408 	bp->b_iodone = ffs_blkfree_trim_completed;
2409 	bp->b_bcount = size;
2410 	bp->b_fsprivate1 = tp;
2411 
2412 	mp = UFSTOVFS(ump);
2413 	vn_start_secondary_write(NULL, &mp, 0);
2414 	g_vfs_strategy(ump->um_bo, bp);
2415 }
2416 
2417 #ifdef INVARIANTS
2418 /*
2419  * Verify allocation of a block or fragment. Returns true if block or
2420  * fragment is allocated, false if it is free.
2421  */
2422 static int
2423 ffs_checkblk(ip, bno, size)
2424 	struct inode *ip;
2425 	ufs2_daddr_t bno;
2426 	long size;
2427 {
2428 	struct fs *fs;
2429 	struct cg *cgp;
2430 	struct buf *bp;
2431 	ufs1_daddr_t cgbno;
2432 	int i, error, frags, free;
2433 	u_int8_t *blksfree;
2434 
2435 	fs = ITOFS(ip);
2436 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
2437 		printf("bsize = %ld, size = %ld, fs = %s\n",
2438 		    (long)fs->fs_bsize, size, fs->fs_fsmnt);
2439 		panic("ffs_checkblk: bad size");
2440 	}
2441 	if ((u_int)bno >= fs->fs_size)
2442 		panic("ffs_checkblk: bad block %jd", (intmax_t)bno);
2443 	error = ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), &bp, &cgp);
2444 	if (error)
2445 		panic("ffs_checkblk: cylinder group read failed");
2446 	blksfree = cg_blksfree(cgp);
2447 	cgbno = dtogd(fs, bno);
2448 	if (size == fs->fs_bsize) {
2449 		free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno));
2450 	} else {
2451 		frags = numfrags(fs, size);
2452 		for (free = 0, i = 0; i < frags; i++)
2453 			if (isset(blksfree, cgbno + i))
2454 				free++;
2455 		if (free != 0 && free != frags)
2456 			panic("ffs_checkblk: partially free fragment");
2457 	}
2458 	brelse(bp);
2459 	return (!free);
2460 }
2461 #endif /* INVARIANTS */
2462 
2463 /*
2464  * Free an inode.
2465  */
2466 int
2467 ffs_vfree(pvp, ino, mode)
2468 	struct vnode *pvp;
2469 	ino_t ino;
2470 	int mode;
2471 {
2472 	struct ufsmount *ump;
2473 
2474 	if (DOINGSOFTDEP(pvp)) {
2475 		softdep_freefile(pvp, ino, mode);
2476 		return (0);
2477 	}
2478 	ump = VFSTOUFS(pvp->v_mount);
2479 	return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL));
2480 }
2481 
2482 /*
2483  * Do the actual free operation.
2484  * The specified inode is placed back in the free map.
2485  */
2486 int
2487 ffs_freefile(ump, fs, devvp, ino, mode, wkhd)
2488 	struct ufsmount *ump;
2489 	struct fs *fs;
2490 	struct vnode *devvp;
2491 	ino_t ino;
2492 	int mode;
2493 	struct workhead *wkhd;
2494 {
2495 	struct cg *cgp;
2496 	struct buf *bp;
2497 	int error;
2498 	u_int cg;
2499 	u_int8_t *inosused;
2500 	struct cdev *dev;
2501 
2502 	cg = ino_to_cg(fs, ino);
2503 	if (devvp->v_type == VREG) {
2504 		/* devvp is a snapshot */
2505 		MPASS(devvp->v_mount->mnt_data == ump);
2506 		dev = ump->um_devvp->v_rdev;
2507 	} else if (devvp->v_type == VCHR) {
2508 		/* devvp is a normal disk device */
2509 		dev = devvp->v_rdev;
2510 	} else {
2511 		bp = NULL;
2512 		return (0);
2513 	}
2514 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2515 		panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s",
2516 		    devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt);
2517 	if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2518 		return (error);
2519 	inosused = cg_inosused(cgp);
2520 	ino %= fs->fs_ipg;
2521 	if (isclr(inosused, ino)) {
2522 		printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev),
2523 		    (uintmax_t)(ino + cg * fs->fs_ipg), fs->fs_fsmnt);
2524 		if (fs->fs_ronly == 0)
2525 			panic("ffs_freefile: freeing free inode");
2526 	}
2527 	clrbit(inosused, ino);
2528 	if (ino < cgp->cg_irotor)
2529 		cgp->cg_irotor = ino;
2530 	cgp->cg_cs.cs_nifree++;
2531 	UFS_LOCK(ump);
2532 	fs->fs_cstotal.cs_nifree++;
2533 	fs->fs_cs(fs, cg).cs_nifree++;
2534 	if ((mode & IFMT) == IFDIR) {
2535 		cgp->cg_cs.cs_ndir--;
2536 		fs->fs_cstotal.cs_ndir--;
2537 		fs->fs_cs(fs, cg).cs_ndir--;
2538 	}
2539 	fs->fs_fmod = 1;
2540 	ACTIVECLEAR(fs, cg);
2541 	UFS_UNLOCK(ump);
2542 	if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR)
2543 		softdep_setup_inofree(UFSTOVFS(ump), bp,
2544 		    ino + cg * fs->fs_ipg, wkhd);
2545 	bdwrite(bp);
2546 	return (0);
2547 }
2548 
2549 /*
2550  * Check to see if a file is free.
2551  * Used to check for allocated files in snapshots.
2552  */
2553 int
2554 ffs_checkfreefile(fs, devvp, ino)
2555 	struct fs *fs;
2556 	struct vnode *devvp;
2557 	ino_t ino;
2558 {
2559 	struct cg *cgp;
2560 	struct buf *bp;
2561 	int ret, error;
2562 	u_int cg;
2563 	u_int8_t *inosused;
2564 
2565 	cg = ino_to_cg(fs, ino);
2566 	if ((devvp->v_type != VREG) && (devvp->v_type != VCHR))
2567 		return (1);
2568 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2569 		return (1);
2570 	if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2571 		return (1);
2572 	inosused = cg_inosused(cgp);
2573 	ino %= fs->fs_ipg;
2574 	ret = isclr(inosused, ino);
2575 	brelse(bp);
2576 	return (ret);
2577 }
2578 
2579 /*
2580  * Find a block of the specified size in the specified cylinder group.
2581  *
2582  * It is a panic if a request is made to find a block if none are
2583  * available.
2584  */
2585 static ufs1_daddr_t
2586 ffs_mapsearch(fs, cgp, bpref, allocsiz)
2587 	struct fs *fs;
2588 	struct cg *cgp;
2589 	ufs2_daddr_t bpref;
2590 	int allocsiz;
2591 {
2592 	ufs1_daddr_t bno;
2593 	int start, len, loc, i;
2594 	int blk, field, subfield, pos;
2595 	u_int8_t *blksfree;
2596 
2597 	/*
2598 	 * find the fragment by searching through the free block
2599 	 * map for an appropriate bit pattern
2600 	 */
2601 	if (bpref)
2602 		start = dtogd(fs, bpref) / NBBY;
2603 	else
2604 		start = cgp->cg_frotor / NBBY;
2605 	blksfree = cg_blksfree(cgp);
2606 	len = howmany(fs->fs_fpg, NBBY) - start;
2607 	loc = scanc((u_int)len, (u_char *)&blksfree[start],
2608 		fragtbl[fs->fs_frag],
2609 		(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2610 	if (loc == 0) {
2611 		len = start + 1;
2612 		start = 0;
2613 		loc = scanc((u_int)len, (u_char *)&blksfree[0],
2614 			fragtbl[fs->fs_frag],
2615 			(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2616 		if (loc == 0) {
2617 			printf("start = %d, len = %d, fs = %s\n",
2618 			    start, len, fs->fs_fsmnt);
2619 			panic("ffs_alloccg: map corrupted");
2620 			/* NOTREACHED */
2621 		}
2622 	}
2623 	bno = (start + len - loc) * NBBY;
2624 	cgp->cg_frotor = bno;
2625 	/*
2626 	 * found the byte in the map
2627 	 * sift through the bits to find the selected frag
2628 	 */
2629 	for (i = bno + NBBY; bno < i; bno += fs->fs_frag) {
2630 		blk = blkmap(fs, blksfree, bno);
2631 		blk <<= 1;
2632 		field = around[allocsiz];
2633 		subfield = inside[allocsiz];
2634 		for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) {
2635 			if ((blk & field) == subfield)
2636 				return (bno + pos);
2637 			field <<= 1;
2638 			subfield <<= 1;
2639 		}
2640 	}
2641 	printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt);
2642 	panic("ffs_alloccg: block not in map");
2643 	return (-1);
2644 }
2645 
2646 static const struct statfs *
2647 ffs_getmntstat(struct vnode *devvp)
2648 {
2649 
2650 	if (devvp->v_type == VCHR)
2651 		return (&devvp->v_rdev->si_mountpt->mnt_stat);
2652 	return (ffs_getmntstat(VFSTOUFS(devvp->v_mount)->um_devvp));
2653 }
2654 
2655 /*
2656  * Fetch and verify a cylinder group.
2657  */
2658 int
2659 ffs_getcg(fs, devvp, cg, bpp, cgpp)
2660 	struct fs *fs;
2661 	struct vnode *devvp;
2662 	u_int cg;
2663 	struct buf **bpp;
2664 	struct cg **cgpp;
2665 {
2666 	struct buf *bp;
2667 	struct cg *cgp;
2668 	const struct statfs *sfs;
2669 	int flags, error;
2670 
2671 	*bpp = NULL;
2672 	*cgpp = NULL;
2673 	flags = 0;
2674 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
2675 		flags |= GB_CKHASH;
2676 	error = breadn_flags(devvp, devvp->v_type == VREG ?
2677 	    fragstoblks(fs, cgtod(fs, cg)) : fsbtodb(fs, cgtod(fs, cg)),
2678 	    (int)fs->fs_cgsize, NULL, NULL, 0, NOCRED, flags,
2679 	    ffs_ckhash_cg, &bp);
2680 	if (error != 0)
2681 		return (error);
2682 	cgp = (struct cg *)bp->b_data;
2683 	if ((fs->fs_metackhash & CK_CYLGRP) != 0 &&
2684 	    (bp->b_flags & B_CKHASH) != 0 &&
2685 	    cgp->cg_ckhash != bp->b_ckhash) {
2686 		sfs = ffs_getmntstat(devvp);
2687 		printf("UFS %s%s (%s) cylinder checksum failed: cg %u, cgp: "
2688 		    "0x%x != bp: 0x%jx\n",
2689 		    devvp->v_type == VCHR ? "" : "snapshot of ",
2690 		    sfs->f_mntfromname, sfs->f_mntonname,
2691 		    cg, cgp->cg_ckhash, (uintmax_t)bp->b_ckhash);
2692 		bp->b_flags &= ~B_CKHASH;
2693 		bp->b_flags |= B_INVAL | B_NOCACHE;
2694 		brelse(bp);
2695 		return (EIO);
2696 	}
2697 	if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) {
2698 		sfs = ffs_getmntstat(devvp);
2699 		printf("UFS %s%s (%s)",
2700 		    devvp->v_type == VCHR ? "" : "snapshot of ",
2701 		    sfs->f_mntfromname, sfs->f_mntonname);
2702 		if (!cg_chkmagic(cgp))
2703 			printf(" cg %u: bad magic number 0x%x should be 0x%x\n",
2704 			    cg, cgp->cg_magic, CG_MAGIC);
2705 		else
2706 			printf(": wrong cylinder group cg %u != cgx %u\n", cg,
2707 			    cgp->cg_cgx);
2708 		bp->b_flags &= ~B_CKHASH;
2709 		bp->b_flags |= B_INVAL | B_NOCACHE;
2710 		brelse(bp);
2711 		return (EIO);
2712 	}
2713 	bp->b_flags &= ~B_CKHASH;
2714 	bp->b_xflags |= BX_BKGRDWRITE;
2715 	/*
2716 	 * If we are using check hashes on the cylinder group then we want
2717 	 * to limit changing the cylinder group time to when we are actually
2718 	 * going to write it to disk so that its check hash remains correct
2719 	 * in memory. If the CK_CYLGRP flag is set the time is updated in
2720 	 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we
2721 	 * update the time here as we have done historically.
2722 	 */
2723 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
2724 		bp->b_xflags |= BX_CYLGRP;
2725 	else
2726 		cgp->cg_old_time = cgp->cg_time = time_second;
2727 	*bpp = bp;
2728 	*cgpp = cgp;
2729 	return (0);
2730 }
2731 
2732 static void
2733 ffs_ckhash_cg(bp)
2734 	struct buf *bp;
2735 {
2736 	uint32_t ckhash;
2737 	struct cg *cgp;
2738 
2739 	cgp = (struct cg *)bp->b_data;
2740 	ckhash = cgp->cg_ckhash;
2741 	cgp->cg_ckhash = 0;
2742 	bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount);
2743 	cgp->cg_ckhash = ckhash;
2744 }
2745 
2746 /*
2747  * Fserr prints the name of a filesystem with an error diagnostic.
2748  *
2749  * The form of the error message is:
2750  *	fs: error message
2751  */
2752 void
2753 ffs_fserr(fs, inum, cp)
2754 	struct fs *fs;
2755 	ino_t inum;
2756 	char *cp;
2757 {
2758 	struct thread *td = curthread;	/* XXX */
2759 	struct proc *p = td->td_proc;
2760 
2761 	log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n",
2762 	    p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum,
2763 	    fs->fs_fsmnt, cp);
2764 }
2765 
2766 /*
2767  * This function provides the capability for the fsck program to
2768  * update an active filesystem. Fourteen operations are provided:
2769  *
2770  * adjrefcnt(inode, amt) - adjusts the reference count on the
2771  *	specified inode by the specified amount. Under normal
2772  *	operation the count should always go down. Decrementing
2773  *	the count to zero will cause the inode to be freed.
2774  * adjblkcnt(inode, amt) - adjust the number of blocks used by the
2775  *	inode by the specified amount.
2776  * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) -
2777  *	adjust the superblock summary.
2778  * freedirs(inode, count) - directory inodes [inode..inode + count - 1]
2779  *	are marked as free. Inodes should never have to be marked
2780  *	as in use.
2781  * freefiles(inode, count) - file inodes [inode..inode + count - 1]
2782  *	are marked as free. Inodes should never have to be marked
2783  *	as in use.
2784  * freeblks(blockno, size) - blocks [blockno..blockno + size - 1]
2785  *	are marked as free. Blocks should never have to be marked
2786  *	as in use.
2787  * setflags(flags, set/clear) - the fs_flags field has the specified
2788  *	flags set (second parameter +1) or cleared (second parameter -1).
2789  * setcwd(dirinode) - set the current directory to dirinode in the
2790  *	filesystem associated with the snapshot.
2791  * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".."
2792  *	in the current directory is oldvalue then change it to newvalue.
2793  * unlink(nameptr, oldvalue) - Verify that the inode number associated
2794  *	with nameptr in the current directory is oldvalue then unlink it.
2795  *
2796  * The following functions may only be used on a quiescent filesystem
2797  * by the soft updates journal. They are not safe to be run on an active
2798  * filesystem.
2799  *
2800  * setinode(inode, dip) - the specified disk inode is replaced with the
2801  *	contents pointed to by dip.
2802  * setbufoutput(fd, flags) - output associated with the specified file
2803  *	descriptor (which must reference the character device supporting
2804  *	the filesystem) switches from using physio to running through the
2805  *	buffer cache when flags is set to 1. The descriptor reverts to
2806  *	physio for output when flags is set to zero.
2807  */
2808 
2809 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS);
2810 
2811 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt, CTLFLAG_WR|CTLTYPE_STRUCT,
2812 	0, 0, sysctl_ffs_fsck, "S,fsck", "Adjust Inode Reference Count");
2813 
2814 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt, CTLFLAG_WR,
2815 	sysctl_ffs_fsck, "Adjust Inode Used Blocks Count");
2816 
2817 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir, CTLFLAG_WR,
2818 	sysctl_ffs_fsck, "Adjust number of directories");
2819 
2820 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree, CTLFLAG_WR,
2821 	sysctl_ffs_fsck, "Adjust number of free blocks");
2822 
2823 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree, CTLFLAG_WR,
2824 	sysctl_ffs_fsck, "Adjust number of free inodes");
2825 
2826 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree, CTLFLAG_WR,
2827 	sysctl_ffs_fsck, "Adjust number of free frags");
2828 
2829 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters, CTLFLAG_WR,
2830 	sysctl_ffs_fsck, "Adjust number of free clusters");
2831 
2832 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs, CTLFLAG_WR,
2833 	sysctl_ffs_fsck, "Free Range of Directory Inodes");
2834 
2835 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles, CTLFLAG_WR,
2836 	sysctl_ffs_fsck, "Free Range of File Inodes");
2837 
2838 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks, CTLFLAG_WR,
2839 	sysctl_ffs_fsck, "Free Range of Blocks");
2840 
2841 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags, CTLFLAG_WR,
2842 	sysctl_ffs_fsck, "Change Filesystem Flags");
2843 
2844 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd, CTLFLAG_WR,
2845 	sysctl_ffs_fsck, "Set Current Working Directory");
2846 
2847 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot, CTLFLAG_WR,
2848 	sysctl_ffs_fsck, "Change Value of .. Entry");
2849 
2850 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink, CTLFLAG_WR,
2851 	sysctl_ffs_fsck, "Unlink a Duplicate Name");
2852 
2853 static SYSCTL_NODE(_vfs_ffs, FFS_SET_INODE, setinode, CTLFLAG_WR,
2854 	sysctl_ffs_fsck, "Update an On-Disk Inode");
2855 
2856 static SYSCTL_NODE(_vfs_ffs, FFS_SET_BUFOUTPUT, setbufoutput, CTLFLAG_WR,
2857 	sysctl_ffs_fsck, "Set Buffered Writing for Descriptor");
2858 
2859 #define DEBUG 1
2860 #ifdef DEBUG
2861 static int fsckcmds = 0;
2862 SYSCTL_INT(_debug, OID_AUTO, fsckcmds, CTLFLAG_RW, &fsckcmds, 0, "");
2863 #endif /* DEBUG */
2864 
2865 static int buffered_write(struct file *, struct uio *, struct ucred *,
2866 	int, struct thread *);
2867 
2868 static int
2869 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS)
2870 {
2871 	struct thread *td = curthread;
2872 	struct fsck_cmd cmd;
2873 	struct ufsmount *ump;
2874 	struct vnode *vp, *dvp, *fdvp;
2875 	struct inode *ip, *dp;
2876 	struct mount *mp;
2877 	struct fs *fs;
2878 	ufs2_daddr_t blkno;
2879 	long blkcnt, blksize;
2880 	struct file *fp, *vfp;
2881 	cap_rights_t rights;
2882 	int filetype, trimtype, error;
2883 	static struct fileops *origops, bufferedops;
2884 
2885 	if (req->newlen > sizeof cmd)
2886 		return (EBADRPC);
2887 	if ((error = SYSCTL_IN(req, &cmd, sizeof cmd)) != 0)
2888 		return (error);
2889 	if (cmd.version != FFS_CMD_VERSION)
2890 		return (ERPCMISMATCH);
2891 	if ((error = getvnode(td, cmd.handle,
2892 	    cap_rights_init(&rights, CAP_FSCK), &fp)) != 0)
2893 		return (error);
2894 	vp = fp->f_data;
2895 	if (vp->v_type != VREG && vp->v_type != VDIR) {
2896 		fdrop(fp, td);
2897 		return (EINVAL);
2898 	}
2899 	vn_start_write(vp, &mp, V_WAIT);
2900 	if (mp == NULL ||
2901 	    strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) {
2902 		vn_finished_write(mp);
2903 		fdrop(fp, td);
2904 		return (EINVAL);
2905 	}
2906 	ump = VFSTOUFS(mp);
2907 	if ((mp->mnt_flag & MNT_RDONLY) &&
2908 	    ump->um_fsckpid != td->td_proc->p_pid) {
2909 		vn_finished_write(mp);
2910 		fdrop(fp, td);
2911 		return (EROFS);
2912 	}
2913 	fs = ump->um_fs;
2914 	filetype = IFREG;
2915 
2916 	switch (oidp->oid_number) {
2917 
2918 	case FFS_SET_FLAGS:
2919 #ifdef DEBUG
2920 		if (fsckcmds)
2921 			printf("%s: %s flags\n", mp->mnt_stat.f_mntonname,
2922 			    cmd.size > 0 ? "set" : "clear");
2923 #endif /* DEBUG */
2924 		if (cmd.size > 0)
2925 			fs->fs_flags |= (long)cmd.value;
2926 		else
2927 			fs->fs_flags &= ~(long)cmd.value;
2928 		break;
2929 
2930 	case FFS_ADJ_REFCNT:
2931 #ifdef DEBUG
2932 		if (fsckcmds) {
2933 			printf("%s: adjust inode %jd link count by %jd\n",
2934 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
2935 			    (intmax_t)cmd.size);
2936 		}
2937 #endif /* DEBUG */
2938 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
2939 			break;
2940 		ip = VTOI(vp);
2941 		ip->i_nlink += cmd.size;
2942 		DIP_SET(ip, i_nlink, ip->i_nlink);
2943 		ip->i_effnlink += cmd.size;
2944 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
2945 		error = ffs_update(vp, 1);
2946 		if (DOINGSOFTDEP(vp))
2947 			softdep_change_linkcnt(ip);
2948 		vput(vp);
2949 		break;
2950 
2951 	case FFS_ADJ_BLKCNT:
2952 #ifdef DEBUG
2953 		if (fsckcmds) {
2954 			printf("%s: adjust inode %jd block count by %jd\n",
2955 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
2956 			    (intmax_t)cmd.size);
2957 		}
2958 #endif /* DEBUG */
2959 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
2960 			break;
2961 		ip = VTOI(vp);
2962 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size);
2963 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
2964 		error = ffs_update(vp, 1);
2965 		vput(vp);
2966 		break;
2967 
2968 	case FFS_DIR_FREE:
2969 		filetype = IFDIR;
2970 		/* fall through */
2971 
2972 	case FFS_FILE_FREE:
2973 #ifdef DEBUG
2974 		if (fsckcmds) {
2975 			if (cmd.size == 1)
2976 				printf("%s: free %s inode %ju\n",
2977 				    mp->mnt_stat.f_mntonname,
2978 				    filetype == IFDIR ? "directory" : "file",
2979 				    (uintmax_t)cmd.value);
2980 			else
2981 				printf("%s: free %s inodes %ju-%ju\n",
2982 				    mp->mnt_stat.f_mntonname,
2983 				    filetype == IFDIR ? "directory" : "file",
2984 				    (uintmax_t)cmd.value,
2985 				    (uintmax_t)(cmd.value + cmd.size - 1));
2986 		}
2987 #endif /* DEBUG */
2988 		while (cmd.size > 0) {
2989 			if ((error = ffs_freefile(ump, fs, ump->um_devvp,
2990 			    cmd.value, filetype, NULL)))
2991 				break;
2992 			cmd.size -= 1;
2993 			cmd.value += 1;
2994 		}
2995 		break;
2996 
2997 	case FFS_BLK_FREE:
2998 #ifdef DEBUG
2999 		if (fsckcmds) {
3000 			if (cmd.size == 1)
3001 				printf("%s: free block %jd\n",
3002 				    mp->mnt_stat.f_mntonname,
3003 				    (intmax_t)cmd.value);
3004 			else
3005 				printf("%s: free blocks %jd-%jd\n",
3006 				    mp->mnt_stat.f_mntonname,
3007 				    (intmax_t)cmd.value,
3008 				    (intmax_t)cmd.value + cmd.size - 1);
3009 		}
3010 #endif /* DEBUG */
3011 		blkno = cmd.value;
3012 		blkcnt = cmd.size;
3013 		blksize = fs->fs_frag - (blkno % fs->fs_frag);
3014 		trimtype = (blksize < blkcnt) ? STARTFREE : SINGLETON;
3015 		while (blkcnt > 0) {
3016 			if (blksize > blkcnt)
3017 				blksize = blkcnt;
3018 			ffs_blkfree(ump, fs, ump->um_devvp, blkno,
3019 			    blksize * fs->fs_fsize, UFS_ROOTINO,
3020 			    VDIR, NULL, trimtype);
3021 			blkno += blksize;
3022 			blkcnt -= blksize;
3023 			blksize = fs->fs_frag;
3024 			trimtype = (blksize < blkcnt) ? CONTINUEFREE : ENDFREE;
3025 		}
3026 		break;
3027 
3028 	/*
3029 	 * Adjust superblock summaries.  fsck(8) is expected to
3030 	 * submit deltas when necessary.
3031 	 */
3032 	case FFS_ADJ_NDIR:
3033 #ifdef DEBUG
3034 		if (fsckcmds) {
3035 			printf("%s: adjust number of directories by %jd\n",
3036 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3037 		}
3038 #endif /* DEBUG */
3039 		fs->fs_cstotal.cs_ndir += cmd.value;
3040 		break;
3041 
3042 	case FFS_ADJ_NBFREE:
3043 #ifdef DEBUG
3044 		if (fsckcmds) {
3045 			printf("%s: adjust number of free blocks by %+jd\n",
3046 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3047 		}
3048 #endif /* DEBUG */
3049 		fs->fs_cstotal.cs_nbfree += cmd.value;
3050 		break;
3051 
3052 	case FFS_ADJ_NIFREE:
3053 #ifdef DEBUG
3054 		if (fsckcmds) {
3055 			printf("%s: adjust number of free inodes by %+jd\n",
3056 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3057 		}
3058 #endif /* DEBUG */
3059 		fs->fs_cstotal.cs_nifree += cmd.value;
3060 		break;
3061 
3062 	case FFS_ADJ_NFFREE:
3063 #ifdef DEBUG
3064 		if (fsckcmds) {
3065 			printf("%s: adjust number of free frags by %+jd\n",
3066 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3067 		}
3068 #endif /* DEBUG */
3069 		fs->fs_cstotal.cs_nffree += cmd.value;
3070 		break;
3071 
3072 	case FFS_ADJ_NUMCLUSTERS:
3073 #ifdef DEBUG
3074 		if (fsckcmds) {
3075 			printf("%s: adjust number of free clusters by %+jd\n",
3076 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3077 		}
3078 #endif /* DEBUG */
3079 		fs->fs_cstotal.cs_numclusters += cmd.value;
3080 		break;
3081 
3082 	case FFS_SET_CWD:
3083 #ifdef DEBUG
3084 		if (fsckcmds) {
3085 			printf("%s: set current directory to inode %jd\n",
3086 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3087 		}
3088 #endif /* DEBUG */
3089 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp)))
3090 			break;
3091 		AUDIT_ARG_VNODE1(vp);
3092 		if ((error = change_dir(vp, td)) != 0) {
3093 			vput(vp);
3094 			break;
3095 		}
3096 		VOP_UNLOCK(vp, 0);
3097 		pwd_chdir(td, vp);
3098 		break;
3099 
3100 	case FFS_SET_DOTDOT:
3101 #ifdef DEBUG
3102 		if (fsckcmds) {
3103 			printf("%s: change .. in cwd from %jd to %jd\n",
3104 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3105 			    (intmax_t)cmd.size);
3106 		}
3107 #endif /* DEBUG */
3108 		/*
3109 		 * First we have to get and lock the parent directory
3110 		 * to which ".." points.
3111 		 */
3112 		error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp);
3113 		if (error)
3114 			break;
3115 		/*
3116 		 * Now we get and lock the child directory containing "..".
3117 		 */
3118 		FILEDESC_SLOCK(td->td_proc->p_fd);
3119 		dvp = td->td_proc->p_fd->fd_cdir;
3120 		FILEDESC_SUNLOCK(td->td_proc->p_fd);
3121 		if ((error = vget(dvp, LK_EXCLUSIVE, td)) != 0) {
3122 			vput(fdvp);
3123 			break;
3124 		}
3125 		dp = VTOI(dvp);
3126 		dp->i_offset = 12;	/* XXX mastertemplate.dot_reclen */
3127 		error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size,
3128 		    DT_DIR, 0);
3129 		cache_purge(fdvp);
3130 		cache_purge(dvp);
3131 		vput(dvp);
3132 		vput(fdvp);
3133 		break;
3134 
3135 	case FFS_UNLINK:
3136 #ifdef DEBUG
3137 		if (fsckcmds) {
3138 			char buf[32];
3139 
3140 			if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL))
3141 				strncpy(buf, "Name_too_long", 32);
3142 			printf("%s: unlink %s (inode %jd)\n",
3143 			    mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size);
3144 		}
3145 #endif /* DEBUG */
3146 		/*
3147 		 * kern_unlinkat will do its own start/finish writes and
3148 		 * they do not nest, so drop ours here. Setting mp == NULL
3149 		 * indicates that vn_finished_write is not needed down below.
3150 		 */
3151 		vn_finished_write(mp);
3152 		mp = NULL;
3153 		error = kern_unlinkat(td, AT_FDCWD, (char *)(intptr_t)cmd.value,
3154 		    UIO_USERSPACE, (ino_t)cmd.size);
3155 		break;
3156 
3157 	case FFS_SET_INODE:
3158 		if (ump->um_fsckpid != td->td_proc->p_pid) {
3159 			error = EPERM;
3160 			break;
3161 		}
3162 #ifdef DEBUG
3163 		if (fsckcmds) {
3164 			printf("%s: update inode %jd\n",
3165 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3166 		}
3167 #endif /* DEBUG */
3168 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3169 			break;
3170 		AUDIT_ARG_VNODE1(vp);
3171 		ip = VTOI(vp);
3172 		if (I_IS_UFS1(ip))
3173 			error = copyin((void *)(intptr_t)cmd.size, ip->i_din1,
3174 			    sizeof(struct ufs1_dinode));
3175 		else
3176 			error = copyin((void *)(intptr_t)cmd.size, ip->i_din2,
3177 			    sizeof(struct ufs2_dinode));
3178 		if (error) {
3179 			vput(vp);
3180 			break;
3181 		}
3182 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3183 		error = ffs_update(vp, 1);
3184 		vput(vp);
3185 		break;
3186 
3187 	case FFS_SET_BUFOUTPUT:
3188 		if (ump->um_fsckpid != td->td_proc->p_pid) {
3189 			error = EPERM;
3190 			break;
3191 		}
3192 		if (ITOUMP(VTOI(vp)) != ump) {
3193 			error = EINVAL;
3194 			break;
3195 		}
3196 #ifdef DEBUG
3197 		if (fsckcmds) {
3198 			printf("%s: %s buffered output for descriptor %jd\n",
3199 			    mp->mnt_stat.f_mntonname,
3200 			    cmd.size == 1 ? "enable" : "disable",
3201 			    (intmax_t)cmd.value);
3202 		}
3203 #endif /* DEBUG */
3204 		if ((error = getvnode(td, cmd.value,
3205 		    cap_rights_init(&rights, CAP_FSCK), &vfp)) != 0)
3206 			break;
3207 		if (vfp->f_vnode->v_type != VCHR) {
3208 			fdrop(vfp, td);
3209 			error = EINVAL;
3210 			break;
3211 		}
3212 		if (origops == NULL) {
3213 			origops = vfp->f_ops;
3214 			bcopy((void *)origops, (void *)&bufferedops,
3215 			    sizeof(bufferedops));
3216 			bufferedops.fo_write = buffered_write;
3217 		}
3218 		if (cmd.size == 1)
3219 			atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3220 			    (uintptr_t)&bufferedops);
3221 		else
3222 			atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3223 			    (uintptr_t)origops);
3224 		fdrop(vfp, td);
3225 		break;
3226 
3227 	default:
3228 #ifdef DEBUG
3229 		if (fsckcmds) {
3230 			printf("Invalid request %d from fsck\n",
3231 			    oidp->oid_number);
3232 		}
3233 #endif /* DEBUG */
3234 		error = EINVAL;
3235 		break;
3236 
3237 	}
3238 	fdrop(fp, td);
3239 	vn_finished_write(mp);
3240 	return (error);
3241 }
3242 
3243 /*
3244  * Function to switch a descriptor to use the buffer cache to stage
3245  * its I/O. This is needed so that writes to the filesystem device
3246  * will give snapshots a chance to copy modified blocks for which it
3247  * needs to retain copies.
3248  */
3249 static int
3250 buffered_write(fp, uio, active_cred, flags, td)
3251 	struct file *fp;
3252 	struct uio *uio;
3253 	struct ucred *active_cred;
3254 	int flags;
3255 	struct thread *td;
3256 {
3257 	struct vnode *devvp, *vp;
3258 	struct inode *ip;
3259 	struct buf *bp;
3260 	struct fs *fs;
3261 	struct filedesc *fdp;
3262 	int error;
3263 	daddr_t lbn;
3264 
3265 	/*
3266 	 * The devvp is associated with the /dev filesystem. To discover
3267 	 * the filesystem with which the device is associated, we depend
3268 	 * on the application setting the current directory to a location
3269 	 * within the filesystem being written. Yes, this is an ugly hack.
3270 	 */
3271 	devvp = fp->f_vnode;
3272 	if (!vn_isdisk(devvp, NULL))
3273 		return (EINVAL);
3274 	fdp = td->td_proc->p_fd;
3275 	FILEDESC_SLOCK(fdp);
3276 	vp = fdp->fd_cdir;
3277 	vref(vp);
3278 	FILEDESC_SUNLOCK(fdp);
3279 	vn_lock(vp, LK_SHARED | LK_RETRY);
3280 	/*
3281 	 * Check that the current directory vnode indeed belongs to
3282 	 * UFS before trying to dereference UFS-specific v_data fields.
3283 	 */
3284 	if (vp->v_op != &ffs_vnodeops1 && vp->v_op != &ffs_vnodeops2) {
3285 		vput(vp);
3286 		return (EINVAL);
3287 	}
3288 	ip = VTOI(vp);
3289 	if (ITODEVVP(ip) != devvp) {
3290 		vput(vp);
3291 		return (EINVAL);
3292 	}
3293 	fs = ITOFS(ip);
3294 	vput(vp);
3295 	foffset_lock_uio(fp, uio, flags);
3296 	vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
3297 #ifdef DEBUG
3298 	if (fsckcmds) {
3299 		printf("%s: buffered write for block %jd\n",
3300 		    fs->fs_fsmnt, (intmax_t)btodb(uio->uio_offset));
3301 	}
3302 #endif /* DEBUG */
3303 	/*
3304 	 * All I/O must be contained within a filesystem block, start on
3305 	 * a fragment boundary, and be a multiple of fragments in length.
3306 	 */
3307 	if (uio->uio_resid > fs->fs_bsize - (uio->uio_offset % fs->fs_bsize) ||
3308 	    fragoff(fs, uio->uio_offset) != 0 ||
3309 	    fragoff(fs, uio->uio_resid) != 0) {
3310 		error = EINVAL;
3311 		goto out;
3312 	}
3313 	lbn = numfrags(fs, uio->uio_offset);
3314 	bp = getblk(devvp, lbn, uio->uio_resid, 0, 0, 0);
3315 	bp->b_flags |= B_RELBUF;
3316 	if ((error = uiomove((char *)bp->b_data, uio->uio_resid, uio)) != 0) {
3317 		brelse(bp);
3318 		goto out;
3319 	}
3320 	error = bwrite(bp);
3321 out:
3322 	VOP_UNLOCK(devvp, 0);
3323 	foffset_unlock_uio(fp, uio, flags | FOF_NEXTOFF);
3324 	return (error);
3325 }
3326