xref: /freebsd/sys/ufs/ffs/ffs_alloc.c (revision ccfd87fe2ac0e2e6aeb1911a7d7cce6712a8564f)
1 /*-
2  * SPDX-License-Identifier: (BSD-2-Clause 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/systm.h>
71 #include <sys/bio.h>
72 #include <sys/buf.h>
73 #include <sys/capsicum.h>
74 #include <sys/conf.h>
75 #include <sys/fcntl.h>
76 #include <sys/file.h>
77 #include <sys/filedesc.h>
78 #include <sys/gsb_crc32.h>
79 #include <sys/kernel.h>
80 #include <sys/mount.h>
81 #include <sys/priv.h>
82 #include <sys/proc.h>
83 #include <sys/stat.h>
84 #include <sys/syscallsubr.h>
85 #include <sys/sysctl.h>
86 #include <sys/syslog.h>
87 #include <sys/taskqueue.h>
88 #include <sys/vnode.h>
89 
90 #include <security/audit/audit.h>
91 
92 #include <geom/geom.h>
93 #include <geom/geom_vfs.h>
94 
95 #include <ufs/ufs/dir.h>
96 #include <ufs/ufs/extattr.h>
97 #include <ufs/ufs/quota.h>
98 #include <ufs/ufs/inode.h>
99 #include <ufs/ufs/ufs_extern.h>
100 #include <ufs/ufs/ufsmount.h>
101 
102 #include <ufs/ffs/fs.h>
103 #include <ufs/ffs/ffs_extern.h>
104 #include <ufs/ffs/softdep.h>
105 
106 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref,
107 				  int size, int rsize);
108 
109 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int);
110 static ufs2_daddr_t
111 	      ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int);
112 static void	ffs_blkfree_cg(struct ufsmount *, struct fs *,
113 		    struct vnode *, ufs2_daddr_t, long, ino_t,
114 		    struct workhead *);
115 #ifdef INVARIANTS
116 static int	ffs_checkblk(struct inode *, ufs2_daddr_t, long);
117 #endif
118 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int);
119 static ino_t	ffs_dirpref(struct inode *);
120 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t,
121 		    int, int);
122 static ufs2_daddr_t	ffs_hashalloc
123 		(struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *);
124 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int,
125 		    int);
126 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int);
127 static int	ffs_reallocblks_ufs1(struct vop_reallocblks_args *);
128 static int	ffs_reallocblks_ufs2(struct vop_reallocblks_args *);
129 static void	ffs_ckhash_cg(struct buf *);
130 
131 /*
132  * Allocate a block in the filesystem.
133  *
134  * The size of the requested block is given, which must be some
135  * multiple of fs_fsize and <= fs_bsize.
136  * A preference may be optionally specified. If a preference is given
137  * the following hierarchy is used to allocate a block:
138  *   1) allocate the requested block.
139  *   2) allocate a rotationally optimal block in the same cylinder.
140  *   3) allocate a block in the same cylinder group.
141  *   4) quadratically rehash into other cylinder groups, until an
142  *      available block is located.
143  * If no block preference is given the following hierarchy is used
144  * to allocate a block:
145  *   1) allocate a block in the cylinder group that contains the
146  *      inode for the file.
147  *   2) quadratically rehash into other cylinder groups, until an
148  *      available block is located.
149  */
150 int
151 ffs_alloc(struct inode *ip,
152 	ufs2_daddr_t lbn,
153 	ufs2_daddr_t bpref,
154 	int size,
155 	int flags,
156 	struct ucred *cred,
157 	ufs2_daddr_t *bnp)
158 {
159 	struct fs *fs;
160 	struct ufsmount *ump;
161 	ufs2_daddr_t bno;
162 	u_int cg, reclaimed;
163 	int64_t delta;
164 #ifdef QUOTA
165 	int error;
166 #endif
167 
168 	*bnp = 0;
169 	ump = ITOUMP(ip);
170 	fs = ump->um_fs;
171 	mtx_assert(UFS_MTX(ump), MA_OWNED);
172 #ifdef INVARIANTS
173 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
174 		printf("dev = %s, bsize = %ld, size = %d, fs = %s\n",
175 		    devtoname(ump->um_dev), (long)fs->fs_bsize, size,
176 		    fs->fs_fsmnt);
177 		panic("ffs_alloc: bad size");
178 	}
179 	if (cred == NOCRED)
180 		panic("ffs_alloc: missing credential");
181 #endif /* INVARIANTS */
182 	reclaimed = 0;
183 retry:
184 #ifdef QUOTA
185 	UFS_UNLOCK(ump);
186 	error = chkdq(ip, btodb(size), cred, 0);
187 	if (error)
188 		return (error);
189 	UFS_LOCK(ump);
190 #endif
191 	if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0)
192 		goto nospace;
193 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) &&
194 	    freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0)
195 		goto nospace;
196 	if (bpref >= fs->fs_size)
197 		bpref = 0;
198 	if (bpref == 0)
199 		cg = ino_to_cg(fs, ip->i_number);
200 	else
201 		cg = dtog(fs, bpref);
202 	bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg);
203 	if (bno > 0) {
204 		delta = btodb(size);
205 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
206 		if (flags & IO_EXT)
207 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
208 		else
209 			UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
210 		*bnp = bno;
211 		return (0);
212 	}
213 nospace:
214 #ifdef QUOTA
215 	UFS_UNLOCK(ump);
216 	/*
217 	 * Restore user's disk quota because allocation failed.
218 	 */
219 	(void) chkdq(ip, -btodb(size), cred, FORCE);
220 	UFS_LOCK(ump);
221 #endif
222 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
223 		reclaimed = 1;
224 		softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT);
225 		goto retry;
226 	}
227 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
228 		UFS_UNLOCK(ump);
229 		return (ENXIO);
230 	}
231 	if (reclaimed > 0 &&
232 	    ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
233 		UFS_UNLOCK(ump);
234 		ffs_fserr(fs, ip->i_number, "filesystem full");
235 		uprintf("\n%s: write failed, filesystem is full\n",
236 		    fs->fs_fsmnt);
237 	} else {
238 		UFS_UNLOCK(ump);
239 	}
240 	return (ENOSPC);
241 }
242 
243 /*
244  * Reallocate a fragment to a bigger size
245  *
246  * The number and size of the old block is given, and a preference
247  * and new size is also specified. The allocator attempts to extend
248  * the original block. Failing that, the regular block allocator is
249  * invoked to get an appropriate block.
250  */
251 int
252 ffs_realloccg(struct inode *ip,
253 	ufs2_daddr_t lbprev,
254 	ufs2_daddr_t bprev,
255 	ufs2_daddr_t bpref,
256 	int osize,
257 	int nsize,
258 	int flags,
259 	struct ucred *cred,
260 	struct buf **bpp)
261 {
262 	struct vnode *vp;
263 	struct fs *fs;
264 	struct buf *bp;
265 	struct ufsmount *ump;
266 	u_int cg, request, reclaimed;
267 	int error, gbflags;
268 	ufs2_daddr_t bno;
269 	int64_t delta;
270 
271 	vp = ITOV(ip);
272 	ump = ITOUMP(ip);
273 	fs = ump->um_fs;
274 	bp = NULL;
275 	gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
276 #ifdef WITNESS
277 	gbflags |= IS_SNAPSHOT(ip) ? GB_NOWITNESS : 0;
278 #endif
279 
280 	mtx_assert(UFS_MTX(ump), MA_OWNED);
281 #ifdef INVARIANTS
282 	if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
283 		panic("ffs_realloccg: allocation on suspended filesystem");
284 	if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 ||
285 	    (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) {
286 		printf(
287 		"dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n",
288 		    devtoname(ump->um_dev), (long)fs->fs_bsize, osize,
289 		    nsize, fs->fs_fsmnt);
290 		panic("ffs_realloccg: bad size");
291 	}
292 	if (cred == NOCRED)
293 		panic("ffs_realloccg: missing credential");
294 #endif /* INVARIANTS */
295 	reclaimed = 0;
296 retry:
297 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) &&
298 	    freespace(fs, fs->fs_minfree) -  numfrags(fs, nsize - osize) < 0) {
299 		goto nospace;
300 	}
301 	if (bprev == 0) {
302 		printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n",
303 		    devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev,
304 		    fs->fs_fsmnt);
305 		panic("ffs_realloccg: bad bprev");
306 	}
307 	UFS_UNLOCK(ump);
308 	/*
309 	 * Allocate the extra space in the buffer.
310 	 */
311 	error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp);
312 	if (error) {
313 		return (error);
314 	}
315 
316 	if (bp->b_blkno == bp->b_lblkno) {
317 		if (lbprev >= UFS_NDADDR)
318 			panic("ffs_realloccg: lbprev out of range");
319 		bp->b_blkno = fsbtodb(fs, bprev);
320 	}
321 
322 #ifdef QUOTA
323 	error = chkdq(ip, btodb(nsize - osize), cred, 0);
324 	if (error) {
325 		brelse(bp);
326 		return (error);
327 	}
328 #endif
329 	/*
330 	 * Check for extension in the existing location.
331 	 */
332 	*bpp = NULL;
333 	cg = dtog(fs, bprev);
334 	UFS_LOCK(ump);
335 	bno = ffs_fragextend(ip, cg, bprev, osize, nsize);
336 	if (bno) {
337 		if (bp->b_blkno != fsbtodb(fs, bno))
338 			panic("ffs_realloccg: bad blockno");
339 		delta = btodb(nsize - osize);
340 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
341 		if (flags & IO_EXT)
342 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
343 		else
344 			UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
345 		allocbuf(bp, nsize);
346 		bp->b_flags |= B_DONE;
347 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
348 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
349 			vfs_bio_set_valid(bp, osize, nsize - osize);
350 		*bpp = bp;
351 		return (0);
352 	}
353 	/*
354 	 * Allocate a new disk location.
355 	 */
356 	if (bpref >= fs->fs_size)
357 		bpref = 0;
358 	switch ((int)fs->fs_optim) {
359 	case FS_OPTSPACE:
360 		/*
361 		 * Allocate an exact sized fragment. Although this makes
362 		 * best use of space, we will waste time relocating it if
363 		 * the file continues to grow. If the fragmentation is
364 		 * less than half of the minimum free reserve, we choose
365 		 * to begin optimizing for time.
366 		 */
367 		request = nsize;
368 		if (fs->fs_minfree <= 5 ||
369 		    fs->fs_cstotal.cs_nffree >
370 		    (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100))
371 			break;
372 		log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n",
373 			fs->fs_fsmnt);
374 		fs->fs_optim = FS_OPTTIME;
375 		break;
376 	case FS_OPTTIME:
377 		/*
378 		 * At this point we have discovered a file that is trying to
379 		 * grow a small fragment to a larger fragment. To save time,
380 		 * we allocate a full sized block, then free the unused portion.
381 		 * If the file continues to grow, the `ffs_fragextend' call
382 		 * above will be able to grow it in place without further
383 		 * copying. If aberrant programs cause disk fragmentation to
384 		 * grow within 2% of the free reserve, we choose to begin
385 		 * optimizing for space.
386 		 */
387 		request = fs->fs_bsize;
388 		if (fs->fs_cstotal.cs_nffree <
389 		    (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100)
390 			break;
391 		log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n",
392 			fs->fs_fsmnt);
393 		fs->fs_optim = FS_OPTSPACE;
394 		break;
395 	default:
396 		printf("dev = %s, optim = %ld, fs = %s\n",
397 		    devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt);
398 		panic("ffs_realloccg: bad optim");
399 		/* NOTREACHED */
400 	}
401 	bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg);
402 	if (bno > 0) {
403 		bp->b_blkno = fsbtodb(fs, bno);
404 		if (!DOINGSOFTDEP(vp))
405 			/*
406 			 * The usual case is that a smaller fragment that
407 			 * was just allocated has been replaced with a bigger
408 			 * fragment or a full-size block. If it is marked as
409 			 * B_DELWRI, the current contents have not been written
410 			 * to disk. It is possible that the block was written
411 			 * earlier, but very uncommon. If the block has never
412 			 * been written, there is no need to send a BIO_DELETE
413 			 * for it when it is freed. The gain from avoiding the
414 			 * TRIMs for the common case of unwritten blocks far
415 			 * exceeds the cost of the write amplification for the
416 			 * uncommon case of failing to send a TRIM for a block
417 			 * that had been written.
418 			 */
419 			ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize,
420 			    ip->i_number, vp->v_type, NULL,
421 			    (bp->b_flags & B_DELWRI) != 0 ?
422 			    NOTRIM_KEY : SINGLETON_KEY);
423 		delta = btodb(nsize - osize);
424 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
425 		if (flags & IO_EXT)
426 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
427 		else
428 			UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
429 		allocbuf(bp, nsize);
430 		bp->b_flags |= B_DONE;
431 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
432 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
433 			vfs_bio_set_valid(bp, osize, nsize - osize);
434 		*bpp = bp;
435 		return (0);
436 	}
437 #ifdef QUOTA
438 	UFS_UNLOCK(ump);
439 	/*
440 	 * Restore user's disk quota because allocation failed.
441 	 */
442 	(void) chkdq(ip, -btodb(nsize - osize), cred, FORCE);
443 	UFS_LOCK(ump);
444 #endif
445 nospace:
446 	/*
447 	 * no space available
448 	 */
449 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
450 		reclaimed = 1;
451 		UFS_UNLOCK(ump);
452 		if (bp) {
453 			brelse(bp);
454 			bp = NULL;
455 		}
456 		UFS_LOCK(ump);
457 		softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT);
458 		goto retry;
459 	}
460 	if (bp)
461 		brelse(bp);
462 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
463 		UFS_UNLOCK(ump);
464 		return (ENXIO);
465 	}
466 	if (reclaimed > 0 &&
467 	    ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
468 		UFS_UNLOCK(ump);
469 		ffs_fserr(fs, ip->i_number, "filesystem full");
470 		uprintf("\n%s: write failed, filesystem is full\n",
471 		    fs->fs_fsmnt);
472 	} else {
473 		UFS_UNLOCK(ump);
474 	}
475 	return (ENOSPC);
476 }
477 
478 /*
479  * Reallocate a sequence of blocks into a contiguous sequence of blocks.
480  *
481  * The vnode and an array of buffer pointers for a range of sequential
482  * logical blocks to be made contiguous is given. The allocator attempts
483  * to find a range of sequential blocks starting as close as possible
484  * from the end of the allocation for the logical block immediately
485  * preceding the current range. If successful, the physical block numbers
486  * in the buffer pointers and in the inode are changed to reflect the new
487  * allocation. If unsuccessful, the allocation is left unchanged. The
488  * success in doing the reallocation is returned. Note that the error
489  * return is not reflected back to the user. Rather the previous block
490  * allocation will be used.
491  */
492 
493 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
494     "FFS filesystem");
495 
496 static int doasyncfree = 1;
497 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0,
498 "do not force synchronous writes when blocks are reallocated");
499 
500 static int doreallocblks = 1;
501 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0,
502 "enable block reallocation");
503 
504 static int dotrimcons = 1;
505 SYSCTL_INT(_vfs_ffs, OID_AUTO, dotrimcons, CTLFLAG_RWTUN, &dotrimcons, 0,
506 "enable BIO_DELETE / TRIM consolidation");
507 
508 static int maxclustersearch = 10;
509 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch,
510 0, "max number of cylinder group to search for contigous blocks");
511 
512 #ifdef DIAGNOSTIC
513 static int prtrealloc = 0;
514 SYSCTL_INT(_debug, OID_AUTO, ffs_prtrealloc, CTLFLAG_RW, &prtrealloc, 0,
515 	"print out FFS filesystem block reallocation operations");
516 #endif
517 
518 int
519 ffs_reallocblks(
520 	struct vop_reallocblks_args /* {
521 		struct vnode *a_vp;
522 		struct cluster_save *a_buflist;
523 	} */ *ap)
524 {
525 	struct ufsmount *ump;
526 	int error;
527 
528 	/*
529 	 * We used to skip reallocating the blocks of a file into a
530 	 * contiguous sequence if the underlying flash device requested
531 	 * BIO_DELETE notifications, because devices that benefit from
532 	 * BIO_DELETE also benefit from not moving the data. However,
533 	 * the destination for the data is usually moved before the data
534 	 * is written to the initially allocated location, so we rarely
535 	 * suffer the penalty of extra writes. With the addition of the
536 	 * consolidation of contiguous blocks into single BIO_DELETE
537 	 * operations, having fewer but larger contiguous blocks reduces
538 	 * the number of (slow and expensive) BIO_DELETE operations. So
539 	 * when doing BIO_DELETE consolidation, we do block reallocation.
540 	 *
541 	 * Skip if reallocblks has been disabled globally.
542 	 */
543 	ump = ap->a_vp->v_mount->mnt_data;
544 	if ((((ump->um_flags) & UM_CANDELETE) != 0 && dotrimcons == 0) ||
545 	    doreallocblks == 0)
546 		return (ENOSPC);
547 
548 	/*
549 	 * We can't wait in softdep prealloc as it may fsync and recurse
550 	 * here.  Instead we simply fail to reallocate blocks if this
551 	 * rare condition arises.
552 	 */
553 	if (DOINGSUJ(ap->a_vp))
554 		if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0)
555 			return (ENOSPC);
556 	vn_seqc_write_begin(ap->a_vp);
557 	error = ump->um_fstype == UFS1 ? ffs_reallocblks_ufs1(ap) :
558 	    ffs_reallocblks_ufs2(ap);
559 	vn_seqc_write_end(ap->a_vp);
560 	return (error);
561 }
562 
563 static int
564 ffs_reallocblks_ufs1(
565 	struct vop_reallocblks_args /* {
566 		struct vnode *a_vp;
567 		struct cluster_save *a_buflist;
568 	} */ *ap)
569 {
570 	struct fs *fs;
571 	struct inode *ip;
572 	struct vnode *vp;
573 	struct buf *sbp, *ebp, *bp;
574 	ufs1_daddr_t *bap, *sbap, *ebap;
575 	struct cluster_save *buflist;
576 	struct ufsmount *ump;
577 	ufs_lbn_t start_lbn, end_lbn;
578 	ufs1_daddr_t soff, newblk, blkno;
579 	ufs2_daddr_t pref;
580 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
581 	int i, cg, len, start_lvl, end_lvl, ssize;
582 
583 	vp = ap->a_vp;
584 	ip = VTOI(vp);
585 	ump = ITOUMP(ip);
586 	fs = ump->um_fs;
587 	/*
588 	 * If we are not tracking block clusters or if we have less than 4%
589 	 * free blocks left, then do not attempt to cluster. Running with
590 	 * less than 5% free block reserve is not recommended and those that
591 	 * choose to do so do not expect to have good file layout.
592 	 */
593 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
594 		return (ENOSPC);
595 	buflist = ap->a_buflist;
596 	len = buflist->bs_nchildren;
597 	start_lbn = buflist->bs_children[0]->b_lblkno;
598 	end_lbn = start_lbn + len - 1;
599 #ifdef INVARIANTS
600 	for (i = 0; i < len; i++)
601 		if (!ffs_checkblk(ip,
602 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
603 			panic("ffs_reallocblks: unallocated block 1");
604 	for (i = 1; i < len; i++)
605 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
606 			panic("ffs_reallocblks: non-logical cluster");
607 	blkno = buflist->bs_children[0]->b_blkno;
608 	ssize = fsbtodb(fs, fs->fs_frag);
609 	for (i = 1; i < len - 1; i++)
610 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
611 			panic("ffs_reallocblks: non-physical cluster %d", i);
612 #endif
613 	/*
614 	 * If the cluster crosses the boundary for the first indirect
615 	 * block, leave space for the indirect block. Indirect blocks
616 	 * are initially laid out in a position after the last direct
617 	 * block. Block reallocation would usually destroy locality by
618 	 * moving the indirect block out of the way to make room for
619 	 * data blocks if we didn't compensate here. We should also do
620 	 * this for other indirect block boundaries, but it is only
621 	 * important for the first one.
622 	 */
623 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
624 		return (ENOSPC);
625 	/*
626 	 * If the latest allocation is in a new cylinder group, assume that
627 	 * the filesystem has decided to move and do not force it back to
628 	 * the previous cylinder group.
629 	 */
630 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
631 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
632 		return (ENOSPC);
633 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
634 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
635 		return (ENOSPC);
636 	/*
637 	 * Get the starting offset and block map for the first block.
638 	 */
639 	if (start_lvl == 0) {
640 		sbap = &ip->i_din1->di_db[0];
641 		soff = start_lbn;
642 	} else {
643 		idp = &start_ap[start_lvl - 1];
644 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
645 			brelse(sbp);
646 			return (ENOSPC);
647 		}
648 		sbap = (ufs1_daddr_t *)sbp->b_data;
649 		soff = idp->in_off;
650 	}
651 	/*
652 	 * If the block range spans two block maps, get the second map.
653 	 */
654 	ebap = NULL;
655 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
656 		ssize = len;
657 	} else {
658 #ifdef INVARIANTS
659 		if (start_lvl > 0 &&
660 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
661 			panic("ffs_reallocblk: start == end");
662 #endif
663 		ssize = len - (idp->in_off + 1);
664 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
665 			goto fail;
666 		ebap = (ufs1_daddr_t *)ebp->b_data;
667 	}
668 	/*
669 	 * Find the preferred location for the cluster. If we have not
670 	 * previously failed at this endeavor, then follow our standard
671 	 * preference calculation. If we have failed at it, then pick up
672 	 * where we last ended our search.
673 	 */
674 	UFS_LOCK(ump);
675 	if (ip->i_nextclustercg == -1)
676 		pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap);
677 	else
678 		pref = cgdata(fs, ip->i_nextclustercg);
679 	/*
680 	 * Search the block map looking for an allocation of the desired size.
681 	 * To avoid wasting too much time, we limit the number of cylinder
682 	 * groups that we will search.
683 	 */
684 	cg = dtog(fs, pref);
685 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
686 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
687 			break;
688 		cg += 1;
689 		if (cg >= fs->fs_ncg)
690 			cg = 0;
691 	}
692 	/*
693 	 * If we have failed in our search, record where we gave up for
694 	 * next time. Otherwise, fall back to our usual search citerion.
695 	 */
696 	if (newblk == 0) {
697 		ip->i_nextclustercg = cg;
698 		UFS_UNLOCK(ump);
699 		goto fail;
700 	}
701 	ip->i_nextclustercg = -1;
702 	/*
703 	 * We have found a new contiguous block.
704 	 *
705 	 * First we have to replace the old block pointers with the new
706 	 * block pointers in the inode and indirect blocks associated
707 	 * with the file.
708 	 */
709 #ifdef DIAGNOSTIC
710 	if (prtrealloc)
711 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:",
712 		    (uintmax_t)ip->i_number,
713 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
714 #endif
715 	blkno = newblk;
716 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
717 		if (i == ssize) {
718 			bap = ebap;
719 			soff = -i;
720 		}
721 #ifdef INVARIANTS
722 		if (!ffs_checkblk(ip,
723 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
724 			panic("ffs_reallocblks: unallocated block 2");
725 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
726 			panic("ffs_reallocblks: alloc mismatch");
727 #endif
728 #ifdef DIAGNOSTIC
729 		if (prtrealloc)
730 			printf(" %d,", *bap);
731 #endif
732 		if (DOINGSOFTDEP(vp)) {
733 			if (sbap == &ip->i_din1->di_db[0] && i < ssize)
734 				softdep_setup_allocdirect(ip, start_lbn + i,
735 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
736 				    buflist->bs_children[i]);
737 			else
738 				softdep_setup_allocindir_page(ip, start_lbn + i,
739 				    i < ssize ? sbp : ebp, soff + i, blkno,
740 				    *bap, buflist->bs_children[i]);
741 		}
742 		*bap++ = blkno;
743 	}
744 	/*
745 	 * Next we must write out the modified inode and indirect blocks.
746 	 * For strict correctness, the writes should be synchronous since
747 	 * the old block values may have been written to disk. In practise
748 	 * they are almost never written, but if we are concerned about
749 	 * strict correctness, the `doasyncfree' flag should be set to zero.
750 	 *
751 	 * The test on `doasyncfree' should be changed to test a flag
752 	 * that shows whether the associated buffers and inodes have
753 	 * been written. The flag should be set when the cluster is
754 	 * started and cleared whenever the buffer or inode is flushed.
755 	 * We can then check below to see if it is set, and do the
756 	 * synchronous write only when it has been cleared.
757 	 */
758 	if (sbap != &ip->i_din1->di_db[0]) {
759 		if (doasyncfree)
760 			bdwrite(sbp);
761 		else
762 			bwrite(sbp);
763 	} else {
764 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
765 		if (!doasyncfree)
766 			ffs_update(vp, 1);
767 	}
768 	if (ssize < len) {
769 		if (doasyncfree)
770 			bdwrite(ebp);
771 		else
772 			bwrite(ebp);
773 	}
774 	/*
775 	 * Last, free the old blocks and assign the new blocks to the buffers.
776 	 */
777 #ifdef DIAGNOSTIC
778 	if (prtrealloc)
779 		printf("\n\tnew:");
780 #endif
781 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
782 		bp = buflist->bs_children[i];
783 		if (!DOINGSOFTDEP(vp))
784 			/*
785 			 * The usual case is that a set of N-contiguous blocks
786 			 * that was just allocated has been replaced with a
787 			 * set of N+1-contiguous blocks. If they are marked as
788 			 * B_DELWRI, the current contents have not been written
789 			 * to disk. It is possible that the blocks were written
790 			 * earlier, but very uncommon. If the blocks have never
791 			 * been written, there is no need to send a BIO_DELETE
792 			 * for them when they are freed. The gain from avoiding
793 			 * the TRIMs for the common case of unwritten blocks
794 			 * far exceeds the cost of the write amplification for
795 			 * the uncommon case of failing to send a TRIM for the
796 			 * blocks that had been written.
797 			 */
798 			ffs_blkfree(ump, fs, ump->um_devvp,
799 			    dbtofsb(fs, bp->b_blkno),
800 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
801 			    (bp->b_flags & B_DELWRI) != 0 ?
802 			    NOTRIM_KEY : SINGLETON_KEY);
803 		bp->b_blkno = fsbtodb(fs, blkno);
804 #ifdef INVARIANTS
805 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
806 			panic("ffs_reallocblks: unallocated block 3");
807 #endif
808 #ifdef DIAGNOSTIC
809 		if (prtrealloc)
810 			printf(" %d,", blkno);
811 #endif
812 	}
813 #ifdef DIAGNOSTIC
814 	if (prtrealloc) {
815 		prtrealloc--;
816 		printf("\n");
817 	}
818 #endif
819 	return (0);
820 
821 fail:
822 	if (ssize < len)
823 		brelse(ebp);
824 	if (sbap != &ip->i_din1->di_db[0])
825 		brelse(sbp);
826 	return (ENOSPC);
827 }
828 
829 static int
830 ffs_reallocblks_ufs2(
831 	struct vop_reallocblks_args /* {
832 		struct vnode *a_vp;
833 		struct cluster_save *a_buflist;
834 	} */ *ap)
835 {
836 	struct fs *fs;
837 	struct inode *ip;
838 	struct vnode *vp;
839 	struct buf *sbp, *ebp, *bp;
840 	ufs2_daddr_t *bap, *sbap, *ebap;
841 	struct cluster_save *buflist;
842 	struct ufsmount *ump;
843 	ufs_lbn_t start_lbn, end_lbn;
844 	ufs2_daddr_t soff, newblk, blkno, pref;
845 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
846 	int i, cg, len, start_lvl, end_lvl, ssize;
847 
848 	vp = ap->a_vp;
849 	ip = VTOI(vp);
850 	ump = ITOUMP(ip);
851 	fs = ump->um_fs;
852 	/*
853 	 * If we are not tracking block clusters or if we have less than 4%
854 	 * free blocks left, then do not attempt to cluster. Running with
855 	 * less than 5% free block reserve is not recommended and those that
856 	 * choose to do so do not expect to have good file layout.
857 	 */
858 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
859 		return (ENOSPC);
860 	buflist = ap->a_buflist;
861 	len = buflist->bs_nchildren;
862 	start_lbn = buflist->bs_children[0]->b_lblkno;
863 	end_lbn = start_lbn + len - 1;
864 #ifdef INVARIANTS
865 	for (i = 0; i < len; i++)
866 		if (!ffs_checkblk(ip,
867 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
868 			panic("ffs_reallocblks: unallocated block 1");
869 	for (i = 1; i < len; i++)
870 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
871 			panic("ffs_reallocblks: non-logical cluster");
872 	blkno = buflist->bs_children[0]->b_blkno;
873 	ssize = fsbtodb(fs, fs->fs_frag);
874 	for (i = 1; i < len - 1; i++)
875 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
876 			panic("ffs_reallocblks: non-physical cluster %d", i);
877 #endif
878 	/*
879 	 * If the cluster crosses the boundary for the first indirect
880 	 * block, do not move anything in it. Indirect blocks are
881 	 * usually initially laid out in a position between the data
882 	 * blocks. Block reallocation would usually destroy locality by
883 	 * moving the indirect block out of the way to make room for
884 	 * data blocks if we didn't compensate here. We should also do
885 	 * this for other indirect block boundaries, but it is only
886 	 * important for the first one.
887 	 */
888 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
889 		return (ENOSPC);
890 	/*
891 	 * If the latest allocation is in a new cylinder group, assume that
892 	 * the filesystem has decided to move and do not force it back to
893 	 * the previous cylinder group.
894 	 */
895 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
896 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
897 		return (ENOSPC);
898 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
899 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
900 		return (ENOSPC);
901 	/*
902 	 * Get the starting offset and block map for the first block.
903 	 */
904 	if (start_lvl == 0) {
905 		sbap = &ip->i_din2->di_db[0];
906 		soff = start_lbn;
907 	} else {
908 		idp = &start_ap[start_lvl - 1];
909 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
910 			brelse(sbp);
911 			return (ENOSPC);
912 		}
913 		sbap = (ufs2_daddr_t *)sbp->b_data;
914 		soff = idp->in_off;
915 	}
916 	/*
917 	 * If the block range spans two block maps, get the second map.
918 	 */
919 	ebap = NULL;
920 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
921 		ssize = len;
922 	} else {
923 #ifdef INVARIANTS
924 		if (start_lvl > 0 &&
925 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
926 			panic("ffs_reallocblk: start == end");
927 #endif
928 		ssize = len - (idp->in_off + 1);
929 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
930 			goto fail;
931 		ebap = (ufs2_daddr_t *)ebp->b_data;
932 	}
933 	/*
934 	 * Find the preferred location for the cluster. If we have not
935 	 * previously failed at this endeavor, then follow our standard
936 	 * preference calculation. If we have failed at it, then pick up
937 	 * where we last ended our search.
938 	 */
939 	UFS_LOCK(ump);
940 	if (ip->i_nextclustercg == -1)
941 		pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap);
942 	else
943 		pref = cgdata(fs, ip->i_nextclustercg);
944 	/*
945 	 * Search the block map looking for an allocation of the desired size.
946 	 * To avoid wasting too much time, we limit the number of cylinder
947 	 * groups that we will search.
948 	 */
949 	cg = dtog(fs, pref);
950 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
951 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
952 			break;
953 		cg += 1;
954 		if (cg >= fs->fs_ncg)
955 			cg = 0;
956 	}
957 	/*
958 	 * If we have failed in our search, record where we gave up for
959 	 * next time. Otherwise, fall back to our usual search citerion.
960 	 */
961 	if (newblk == 0) {
962 		ip->i_nextclustercg = cg;
963 		UFS_UNLOCK(ump);
964 		goto fail;
965 	}
966 	ip->i_nextclustercg = -1;
967 	/*
968 	 * We have found a new contiguous block.
969 	 *
970 	 * First we have to replace the old block pointers with the new
971 	 * block pointers in the inode and indirect blocks associated
972 	 * with the file.
973 	 */
974 #ifdef DIAGNOSTIC
975 	if (prtrealloc)
976 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number,
977 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
978 #endif
979 	blkno = newblk;
980 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
981 		if (i == ssize) {
982 			bap = ebap;
983 			soff = -i;
984 		}
985 #ifdef INVARIANTS
986 		if (!ffs_checkblk(ip,
987 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
988 			panic("ffs_reallocblks: unallocated block 2");
989 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
990 			panic("ffs_reallocblks: alloc mismatch");
991 #endif
992 #ifdef DIAGNOSTIC
993 		if (prtrealloc)
994 			printf(" %jd,", (intmax_t)*bap);
995 #endif
996 		if (DOINGSOFTDEP(vp)) {
997 			if (sbap == &ip->i_din2->di_db[0] && i < ssize)
998 				softdep_setup_allocdirect(ip, start_lbn + i,
999 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
1000 				    buflist->bs_children[i]);
1001 			else
1002 				softdep_setup_allocindir_page(ip, start_lbn + i,
1003 				    i < ssize ? sbp : ebp, soff + i, blkno,
1004 				    *bap, buflist->bs_children[i]);
1005 		}
1006 		*bap++ = blkno;
1007 	}
1008 	/*
1009 	 * Next we must write out the modified inode and indirect blocks.
1010 	 * For strict correctness, the writes should be synchronous since
1011 	 * the old block values may have been written to disk. In practise
1012 	 * they are almost never written, but if we are concerned about
1013 	 * strict correctness, the `doasyncfree' flag should be set to zero.
1014 	 *
1015 	 * The test on `doasyncfree' should be changed to test a flag
1016 	 * that shows whether the associated buffers and inodes have
1017 	 * been written. The flag should be set when the cluster is
1018 	 * started and cleared whenever the buffer or inode is flushed.
1019 	 * We can then check below to see if it is set, and do the
1020 	 * synchronous write only when it has been cleared.
1021 	 */
1022 	if (sbap != &ip->i_din2->di_db[0]) {
1023 		if (doasyncfree)
1024 			bdwrite(sbp);
1025 		else
1026 			bwrite(sbp);
1027 	} else {
1028 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
1029 		if (!doasyncfree)
1030 			ffs_update(vp, 1);
1031 	}
1032 	if (ssize < len) {
1033 		if (doasyncfree)
1034 			bdwrite(ebp);
1035 		else
1036 			bwrite(ebp);
1037 	}
1038 	/*
1039 	 * Last, free the old blocks and assign the new blocks to the buffers.
1040 	 */
1041 #ifdef DIAGNOSTIC
1042 	if (prtrealloc)
1043 		printf("\n\tnew:");
1044 #endif
1045 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
1046 		bp = buflist->bs_children[i];
1047 		if (!DOINGSOFTDEP(vp))
1048 			/*
1049 			 * The usual case is that a set of N-contiguous blocks
1050 			 * that was just allocated has been replaced with a
1051 			 * set of N+1-contiguous blocks. If they are marked as
1052 			 * B_DELWRI, the current contents have not been written
1053 			 * to disk. It is possible that the blocks were written
1054 			 * earlier, but very uncommon. If the blocks have never
1055 			 * been written, there is no need to send a BIO_DELETE
1056 			 * for them when they are freed. The gain from avoiding
1057 			 * the TRIMs for the common case of unwritten blocks
1058 			 * far exceeds the cost of the write amplification for
1059 			 * the uncommon case of failing to send a TRIM for the
1060 			 * blocks that had been written.
1061 			 */
1062 			ffs_blkfree(ump, fs, ump->um_devvp,
1063 			    dbtofsb(fs, bp->b_blkno),
1064 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
1065 			    (bp->b_flags & B_DELWRI) != 0 ?
1066 			    NOTRIM_KEY : SINGLETON_KEY);
1067 		bp->b_blkno = fsbtodb(fs, blkno);
1068 #ifdef INVARIANTS
1069 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
1070 			panic("ffs_reallocblks: unallocated block 3");
1071 #endif
1072 #ifdef DIAGNOSTIC
1073 		if (prtrealloc)
1074 			printf(" %jd,", (intmax_t)blkno);
1075 #endif
1076 	}
1077 #ifdef DIAGNOSTIC
1078 	if (prtrealloc) {
1079 		prtrealloc--;
1080 		printf("\n");
1081 	}
1082 #endif
1083 	return (0);
1084 
1085 fail:
1086 	if (ssize < len)
1087 		brelse(ebp);
1088 	if (sbap != &ip->i_din2->di_db[0])
1089 		brelse(sbp);
1090 	return (ENOSPC);
1091 }
1092 
1093 /*
1094  * Allocate an inode in the filesystem.
1095  *
1096  * If allocating a directory, use ffs_dirpref to select the inode.
1097  * If allocating in a directory, the following hierarchy is followed:
1098  *   1) allocate the preferred inode.
1099  *   2) allocate an inode in the same cylinder group.
1100  *   3) quadratically rehash into other cylinder groups, until an
1101  *      available inode is located.
1102  * If no inode preference is given the following hierarchy is used
1103  * to allocate an inode:
1104  *   1) allocate an inode in cylinder group 0.
1105  *   2) quadratically rehash into other cylinder groups, until an
1106  *      available inode is located.
1107  */
1108 int
1109 ffs_valloc(struct vnode *pvp,
1110 	int mode,
1111 	struct ucred *cred,
1112 	struct vnode **vpp)
1113 {
1114 	struct inode *pip;
1115 	struct fs *fs;
1116 	struct inode *ip;
1117 	struct timespec ts;
1118 	struct ufsmount *ump;
1119 	ino_t ino, ipref;
1120 	u_int cg;
1121 	int error, reclaimed;
1122 
1123 	*vpp = NULL;
1124 	pip = VTOI(pvp);
1125 	ump = ITOUMP(pip);
1126 	fs = ump->um_fs;
1127 
1128 	UFS_LOCK(ump);
1129 	reclaimed = 0;
1130 retry:
1131 	if (fs->fs_cstotal.cs_nifree == 0)
1132 		goto noinodes;
1133 
1134 	if ((mode & IFMT) == IFDIR)
1135 		ipref = ffs_dirpref(pip);
1136 	else
1137 		ipref = pip->i_number;
1138 	if (ipref >= fs->fs_ncg * fs->fs_ipg)
1139 		ipref = 0;
1140 	cg = ino_to_cg(fs, ipref);
1141 	/*
1142 	 * Track number of dirs created one after another
1143 	 * in a same cg without intervening by files.
1144 	 */
1145 	if ((mode & IFMT) == IFDIR) {
1146 		if (fs->fs_contigdirs[cg] < 255)
1147 			fs->fs_contigdirs[cg]++;
1148 	} else {
1149 		if (fs->fs_contigdirs[cg] > 0)
1150 			fs->fs_contigdirs[cg]--;
1151 	}
1152 	ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0,
1153 					(allocfcn_t *)ffs_nodealloccg);
1154 	if (ino == 0)
1155 		goto noinodes;
1156 	/*
1157 	 * Get rid of the cached old vnode, force allocation of a new vnode
1158 	 * for this inode. If this fails, release the allocated ino and
1159 	 * return the error.
1160 	 */
1161 	if ((error = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp,
1162 	    FFSV_FORCEINSMQ | FFSV_REPLACE | FFSV_NEWINODE)) != 0) {
1163 		ffs_vfree(pvp, ino, mode);
1164 		return (error);
1165 	}
1166 	/*
1167 	 * We got an inode, so check mode and panic if it is already allocated.
1168 	 */
1169 	ip = VTOI(*vpp);
1170 	if (ip->i_mode) {
1171 		printf("mode = 0%o, inum = %ju, fs = %s\n",
1172 		    ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt);
1173 		panic("ffs_valloc: dup alloc");
1174 	}
1175 	if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) {  /* XXX */
1176 		printf("free inode %s/%lu had %ld blocks\n",
1177 		    fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks));
1178 		DIP_SET(ip, i_blocks, 0);
1179 	}
1180 	ip->i_flags = 0;
1181 	DIP_SET(ip, i_flags, 0);
1182 	if ((mode & IFMT) == IFDIR)
1183 		DIP_SET(ip, i_dirdepth, DIP(pip, i_dirdepth) + 1);
1184 	/*
1185 	 * Set up a new generation number for this inode.
1186 	 */
1187 	while (ip->i_gen == 0 || ++ip->i_gen == 0)
1188 		ip->i_gen = arc4random();
1189 	DIP_SET(ip, i_gen, ip->i_gen);
1190 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1191 		vfs_timestamp(&ts);
1192 		ip->i_din2->di_birthtime = ts.tv_sec;
1193 		ip->i_din2->di_birthnsec = ts.tv_nsec;
1194 	}
1195 	ip->i_flag = 0;
1196 	(*vpp)->v_vflag = 0;
1197 	(*vpp)->v_type = VNON;
1198 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1199 		(*vpp)->v_op = &ffs_vnodeops2;
1200 		UFS_INODE_SET_FLAG(ip, IN_UFS2);
1201 	} else {
1202 		(*vpp)->v_op = &ffs_vnodeops1;
1203 	}
1204 	return (0);
1205 noinodes:
1206 	if (reclaimed == 0) {
1207 		reclaimed = 1;
1208 		softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT);
1209 		goto retry;
1210 	}
1211 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
1212 		UFS_UNLOCK(ump);
1213 		return (ENXIO);
1214 	}
1215 	if (ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
1216 		UFS_UNLOCK(ump);
1217 		ffs_fserr(fs, pip->i_number, "out of inodes");
1218 		uprintf("\n%s: create/symlink failed, no inodes free\n",
1219 		    fs->fs_fsmnt);
1220 	} else {
1221 		UFS_UNLOCK(ump);
1222 	}
1223 	return (ENOSPC);
1224 }
1225 
1226 /*
1227  * Find a cylinder group to place a directory.
1228  *
1229  * The policy implemented by this algorithm is to allocate a
1230  * directory inode in the same cylinder group as its parent
1231  * directory, but also to reserve space for its files inodes
1232  * and data. Restrict the number of directories which may be
1233  * allocated one after another in the same cylinder group
1234  * without intervening allocation of files.
1235  *
1236  * If we allocate a first level directory then force allocation
1237  * in another cylinder group.
1238  */
1239 static ino_t
1240 ffs_dirpref(struct inode *pip)
1241 {
1242 	struct fs *fs;
1243 	int cg, prefcg, curcg, dirsize, cgsize;
1244 	int depth, range, start, end, numdirs, power, numerator, denominator;
1245 	u_int avgifree, avgbfree, avgndir, curdirsize;
1246 	u_int minifree, minbfree, maxndir;
1247 	u_int maxcontigdirs;
1248 
1249 	mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED);
1250 	fs = ITOFS(pip);
1251 
1252 	avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg;
1253 	avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1254 	avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg;
1255 
1256 	/*
1257 	 * Select a preferred cylinder group to place a new directory.
1258 	 * If we are near the root of the filesystem we aim to spread
1259 	 * them out as much as possible. As we descend deeper from the
1260 	 * root we cluster them closer together around their parent as
1261 	 * we expect them to be more closely interactive. Higher-level
1262 	 * directories like usr/src/sys and usr/src/bin should be
1263 	 * separated while the directories in these areas are more
1264 	 * likely to be accessed together so should be closer.
1265 	 *
1266 	 * We pick a range of cylinder groups around the cylinder group
1267 	 * of the directory in which we are being created. The size of
1268 	 * the range for our search is based on our depth from the root
1269 	 * of our filesystem. We then probe that range based on how many
1270 	 * directories are already present. The first new directory is at
1271 	 * 1/2 (middle) of the range; the second is in the first 1/4 of the
1272 	 * range, then at 3/4, 1/8, 3/8, 5/8, 7/8, 1/16, 3/16, 5/16, etc.
1273 	 */
1274 	depth = DIP(pip, i_dirdepth);
1275 	range = fs->fs_ncg / (1 << depth);
1276 	curcg = ino_to_cg(fs, pip->i_number);
1277 	start = curcg - (range / 2);
1278 	if (start < 0)
1279 		start += fs->fs_ncg;
1280 	end = curcg + (range / 2);
1281 	if (end >= fs->fs_ncg)
1282 		end -= fs->fs_ncg;
1283 	numdirs = pip->i_effnlink - 1;
1284 	power = fls(numdirs);
1285 	numerator = (numdirs & ~(1 << (power - 1))) * 2 + 1;
1286 	denominator = 1 << power;
1287 	prefcg = (curcg - (range / 2) + (range * numerator / denominator));
1288 	if (prefcg < 0)
1289 		prefcg += fs->fs_ncg;
1290 	if (prefcg >= fs->fs_ncg)
1291 		prefcg -= fs->fs_ncg;
1292 	/*
1293 	 * If this filesystem is not tracking directory depths,
1294 	 * revert to the old algorithm.
1295 	 */
1296 	if (depth == 0 && pip->i_number != UFS_ROOTINO)
1297 		prefcg = curcg;
1298 
1299 	/*
1300 	 * Count various limits which used for
1301 	 * optimal allocation of a directory inode.
1302 	 */
1303 	maxndir = min(avgndir + (1 << depth), fs->fs_ipg);
1304 	minifree = avgifree - avgifree / 4;
1305 	if (minifree < 1)
1306 		minifree = 1;
1307 	minbfree = avgbfree - avgbfree / 4;
1308 	if (minbfree < 1)
1309 		minbfree = 1;
1310 	cgsize = fs->fs_fsize * fs->fs_fpg;
1311 	dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir;
1312 	curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0;
1313 	if (dirsize < curdirsize)
1314 		dirsize = curdirsize;
1315 	if (dirsize <= 0)
1316 		maxcontigdirs = 0;		/* dirsize overflowed */
1317 	else
1318 		maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255);
1319 	if (fs->fs_avgfpdir > 0)
1320 		maxcontigdirs = min(maxcontigdirs,
1321 				    fs->fs_ipg / fs->fs_avgfpdir);
1322 	if (maxcontigdirs == 0)
1323 		maxcontigdirs = 1;
1324 
1325 	/*
1326 	 * Limit number of dirs in one cg and reserve space for
1327 	 * regular files, but only if we have no deficit in
1328 	 * inodes or space.
1329 	 *
1330 	 * We are trying to find a suitable cylinder group nearby
1331 	 * our preferred cylinder group to place a new directory.
1332 	 * We scan from our preferred cylinder group forward looking
1333 	 * for a cylinder group that meets our criterion. If we get
1334 	 * to the final cylinder group and do not find anything,
1335 	 * we start scanning forwards from the beginning of the
1336 	 * filesystem. While it might seem sensible to start scanning
1337 	 * backwards or even to alternate looking forward and backward,
1338 	 * this approach fails badly when the filesystem is nearly full.
1339 	 * Specifically, we first search all the areas that have no space
1340 	 * and finally try the one preceding that. We repeat this on
1341 	 * every request and in the case of the final block end up
1342 	 * searching the entire filesystem. By jumping to the front
1343 	 * of the filesystem, our future forward searches always look
1344 	 * in new cylinder groups so finds every possible block after
1345 	 * one pass over the filesystem.
1346 	 */
1347 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1348 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1349 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1350 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1351 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1352 				return ((ino_t)(fs->fs_ipg * cg));
1353 		}
1354 	for (cg = 0; cg < prefcg; cg++)
1355 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1356 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1357 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1358 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1359 				return ((ino_t)(fs->fs_ipg * cg));
1360 		}
1361 	/*
1362 	 * This is a backstop when we have deficit in space.
1363 	 */
1364 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1365 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1366 			return ((ino_t)(fs->fs_ipg * cg));
1367 	for (cg = 0; cg < prefcg; cg++)
1368 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1369 			break;
1370 	return ((ino_t)(fs->fs_ipg * cg));
1371 }
1372 
1373 /*
1374  * Select the desired position for the next block in a file.  The file is
1375  * logically divided into sections. The first section is composed of the
1376  * direct blocks and the next fs_maxbpg blocks. Each additional section
1377  * contains fs_maxbpg blocks.
1378  *
1379  * If no blocks have been allocated in the first section, the policy is to
1380  * request a block in the same cylinder group as the inode that describes
1381  * the file. The first indirect is allocated immediately following the last
1382  * direct block and the data blocks for the first indirect immediately
1383  * follow it.
1384  *
1385  * If no blocks have been allocated in any other section, the indirect
1386  * block(s) are allocated in the same cylinder group as its inode in an
1387  * area reserved immediately following the inode blocks. The policy for
1388  * the data blocks is to place them in a cylinder group with a greater than
1389  * average number of free blocks. An appropriate cylinder group is found
1390  * by using a rotor that sweeps the cylinder groups. When a new group of
1391  * blocks is needed, the sweep begins in the cylinder group following the
1392  * cylinder group from which the previous allocation was made. The sweep
1393  * continues until a cylinder group with greater than the average number
1394  * of free blocks is found. If the allocation is for the first block in an
1395  * indirect block or the previous block is a hole, then the information on
1396  * the previous allocation is unavailable; here a best guess is made based
1397  * on the logical block number being allocated.
1398  *
1399  * If a section is already partially allocated, the policy is to
1400  * allocate blocks contiguously within the section if possible.
1401  */
1402 ufs2_daddr_t
1403 ffs_blkpref_ufs1(struct inode *ip,
1404 	ufs_lbn_t lbn,
1405 	int indx,
1406 	ufs1_daddr_t *bap)
1407 {
1408 	struct fs *fs;
1409 	u_int cg, inocg;
1410 	u_int avgbfree, startcg;
1411 	ufs2_daddr_t pref, prevbn;
1412 
1413 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1414 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1415 	fs = ITOFS(ip);
1416 	/*
1417 	 * Allocation of indirect blocks is indicated by passing negative
1418 	 * values in indx: -1 for single indirect, -2 for double indirect,
1419 	 * -3 for triple indirect. As noted below, we attempt to allocate
1420 	 * the first indirect inline with the file data. For all later
1421 	 * indirect blocks, the data is often allocated in other cylinder
1422 	 * groups. However to speed random file access and to speed up
1423 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1424 	 * (typically half of fs_minfree) of the data area of each cylinder
1425 	 * group to hold these later indirect blocks.
1426 	 */
1427 	inocg = ino_to_cg(fs, ip->i_number);
1428 	if (indx < 0) {
1429 		/*
1430 		 * Our preference for indirect blocks is the zone at the
1431 		 * beginning of the inode's cylinder group data area that
1432 		 * we try to reserve for indirect blocks.
1433 		 */
1434 		pref = cgmeta(fs, inocg);
1435 		/*
1436 		 * If we are allocating the first indirect block, try to
1437 		 * place it immediately following the last direct block.
1438 		 */
1439 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1440 		    ip->i_din1->di_db[UFS_NDADDR - 1] != 0)
1441 			pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1442 		return (pref);
1443 	}
1444 	/*
1445 	 * If we are allocating the first data block in the first indirect
1446 	 * block and the indirect has been allocated in the data block area,
1447 	 * try to place it immediately following the indirect block.
1448 	 */
1449 	if (lbn == UFS_NDADDR) {
1450 		pref = ip->i_din1->di_ib[0];
1451 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1452 		    pref < cgbase(fs, inocg + 1))
1453 			return (pref + fs->fs_frag);
1454 	}
1455 	/*
1456 	 * If we are at the beginning of a file, or we have already allocated
1457 	 * the maximum number of blocks per cylinder group, or we do not
1458 	 * have a block allocated immediately preceding us, then we need
1459 	 * to decide where to start allocating new blocks.
1460 	 */
1461 	if (indx ==  0) {
1462 		prevbn = 0;
1463 	} else {
1464 		prevbn = bap[indx - 1];
1465 		if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn,
1466 		    fs->fs_bsize) != 0)
1467 			prevbn = 0;
1468 	}
1469 	if (indx % fs->fs_maxbpg == 0 || prevbn == 0) {
1470 		/*
1471 		 * If we are allocating a directory data block, we want
1472 		 * to place it in the metadata area.
1473 		 */
1474 		if ((ip->i_mode & IFMT) == IFDIR)
1475 			return (cgmeta(fs, inocg));
1476 		/*
1477 		 * Until we fill all the direct and all the first indirect's
1478 		 * blocks, we try to allocate in the data area of the inode's
1479 		 * cylinder group.
1480 		 */
1481 		if (lbn < UFS_NDADDR + NINDIR(fs))
1482 			return (cgdata(fs, inocg));
1483 		/*
1484 		 * Find a cylinder with greater than average number of
1485 		 * unused data blocks.
1486 		 */
1487 		if (indx == 0 || prevbn == 0)
1488 			startcg = inocg + lbn / fs->fs_maxbpg;
1489 		else
1490 			startcg = dtog(fs, prevbn) + 1;
1491 		startcg %= fs->fs_ncg;
1492 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1493 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1494 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1495 				fs->fs_cgrotor = cg;
1496 				return (cgdata(fs, cg));
1497 			}
1498 		for (cg = 0; cg <= startcg; cg++)
1499 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1500 				fs->fs_cgrotor = cg;
1501 				return (cgdata(fs, cg));
1502 			}
1503 		return (0);
1504 	}
1505 	/*
1506 	 * Otherwise, we just always try to lay things out contiguously.
1507 	 */
1508 	return (prevbn + fs->fs_frag);
1509 }
1510 
1511 /*
1512  * Same as above, but for UFS2
1513  */
1514 ufs2_daddr_t
1515 ffs_blkpref_ufs2(struct inode *ip,
1516 	ufs_lbn_t lbn,
1517 	int indx,
1518 	ufs2_daddr_t *bap)
1519 {
1520 	struct fs *fs;
1521 	u_int cg, inocg;
1522 	u_int avgbfree, startcg;
1523 	ufs2_daddr_t pref, prevbn;
1524 
1525 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1526 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1527 	fs = ITOFS(ip);
1528 	/*
1529 	 * Allocation of indirect blocks is indicated by passing negative
1530 	 * values in indx: -1 for single indirect, -2 for double indirect,
1531 	 * -3 for triple indirect. As noted below, we attempt to allocate
1532 	 * the first indirect inline with the file data. For all later
1533 	 * indirect blocks, the data is often allocated in other cylinder
1534 	 * groups. However to speed random file access and to speed up
1535 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1536 	 * (typically half of fs_minfree) of the data area of each cylinder
1537 	 * group to hold these later indirect blocks.
1538 	 */
1539 	inocg = ino_to_cg(fs, ip->i_number);
1540 	if (indx < 0) {
1541 		/*
1542 		 * Our preference for indirect blocks is the zone at the
1543 		 * beginning of the inode's cylinder group data area that
1544 		 * we try to reserve for indirect blocks.
1545 		 */
1546 		pref = cgmeta(fs, inocg);
1547 		/*
1548 		 * If we are allocating the first indirect block, try to
1549 		 * place it immediately following the last direct block.
1550 		 */
1551 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1552 		    ip->i_din2->di_db[UFS_NDADDR - 1] != 0)
1553 			pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1554 		return (pref);
1555 	}
1556 	/*
1557 	 * If we are allocating the first data block in the first indirect
1558 	 * block and the indirect has been allocated in the data block area,
1559 	 * try to place it immediately following the indirect block.
1560 	 */
1561 	if (lbn == UFS_NDADDR) {
1562 		pref = ip->i_din2->di_ib[0];
1563 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1564 		    pref < cgbase(fs, inocg + 1))
1565 			return (pref + fs->fs_frag);
1566 	}
1567 	/*
1568 	 * If we are at the beginning of a file, or we have already allocated
1569 	 * the maximum number of blocks per cylinder group, or we do not
1570 	 * have a block allocated immediately preceding us, then we need
1571 	 * to decide where to start allocating new blocks.
1572 	 */
1573 	if (indx ==  0) {
1574 		prevbn = 0;
1575 	} else {
1576 		prevbn = bap[indx - 1];
1577 		if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn,
1578 		    fs->fs_bsize) != 0)
1579 			prevbn = 0;
1580 	}
1581 	if (indx % fs->fs_maxbpg == 0 || prevbn == 0) {
1582 		/*
1583 		 * If we are allocating a directory data block, we want
1584 		 * to place it in the metadata area.
1585 		 */
1586 		if ((ip->i_mode & IFMT) == IFDIR)
1587 			return (cgmeta(fs, inocg));
1588 		/*
1589 		 * Until we fill all the direct and all the first indirect's
1590 		 * blocks, we try to allocate in the data area of the inode's
1591 		 * cylinder group.
1592 		 */
1593 		if (lbn < UFS_NDADDR + NINDIR(fs))
1594 			return (cgdata(fs, inocg));
1595 		/*
1596 		 * Find a cylinder with greater than average number of
1597 		 * unused data blocks.
1598 		 */
1599 		if (indx == 0 || prevbn == 0)
1600 			startcg = inocg + lbn / fs->fs_maxbpg;
1601 		else
1602 			startcg = dtog(fs, prevbn) + 1;
1603 		startcg %= fs->fs_ncg;
1604 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1605 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1606 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1607 				fs->fs_cgrotor = cg;
1608 				return (cgdata(fs, cg));
1609 			}
1610 		for (cg = 0; cg <= startcg; cg++)
1611 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1612 				fs->fs_cgrotor = cg;
1613 				return (cgdata(fs, cg));
1614 			}
1615 		return (0);
1616 	}
1617 	/*
1618 	 * Otherwise, we just always try to lay things out contiguously.
1619 	 */
1620 	return (prevbn + fs->fs_frag);
1621 }
1622 
1623 /*
1624  * Implement the cylinder overflow algorithm.
1625  *
1626  * The policy implemented by this algorithm is:
1627  *   1) allocate the block in its requested cylinder group.
1628  *   2) quadratically rehash on the cylinder group number.
1629  *   3) brute force search for a free block.
1630  *
1631  * Must be called with the UFS lock held.  Will release the lock on success
1632  * and return with it held on failure.
1633  */
1634 /*VARARGS5*/
1635 static ufs2_daddr_t
1636 ffs_hashalloc(struct inode *ip,
1637 	u_int cg,
1638 	ufs2_daddr_t pref,
1639 	int size,	/* Search size for data blocks, mode for inodes */
1640 	int rsize,	/* Real allocated size. */
1641 	allocfcn_t *allocator)
1642 {
1643 	struct fs *fs;
1644 	ufs2_daddr_t result;
1645 	u_int i, icg = cg;
1646 
1647 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1648 #ifdef INVARIANTS
1649 	if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
1650 		panic("ffs_hashalloc: allocation on suspended filesystem");
1651 #endif
1652 	fs = ITOFS(ip);
1653 	/*
1654 	 * 1: preferred cylinder group
1655 	 */
1656 	result = (*allocator)(ip, cg, pref, size, rsize);
1657 	if (result)
1658 		return (result);
1659 	/*
1660 	 * 2: quadratic rehash
1661 	 */
1662 	for (i = 1; i < fs->fs_ncg; i *= 2) {
1663 		cg += i;
1664 		if (cg >= fs->fs_ncg)
1665 			cg -= fs->fs_ncg;
1666 		result = (*allocator)(ip, cg, 0, size, rsize);
1667 		if (result)
1668 			return (result);
1669 	}
1670 	/*
1671 	 * 3: brute force search
1672 	 * Note that we start at i == 2, since 0 was checked initially,
1673 	 * and 1 is always checked in the quadratic rehash.
1674 	 */
1675 	cg = (icg + 2) % fs->fs_ncg;
1676 	for (i = 2; i < fs->fs_ncg; i++) {
1677 		result = (*allocator)(ip, cg, 0, size, rsize);
1678 		if (result)
1679 			return (result);
1680 		cg++;
1681 		if (cg == fs->fs_ncg)
1682 			cg = 0;
1683 	}
1684 	return (0);
1685 }
1686 
1687 /*
1688  * Determine whether a fragment can be extended.
1689  *
1690  * Check to see if the necessary fragments are available, and
1691  * if they are, allocate them.
1692  */
1693 static ufs2_daddr_t
1694 ffs_fragextend(struct inode *ip,
1695 	u_int cg,
1696 	ufs2_daddr_t bprev,
1697 	int osize,
1698 	int nsize)
1699 {
1700 	struct fs *fs;
1701 	struct cg *cgp;
1702 	struct buf *bp;
1703 	struct ufsmount *ump;
1704 	int nffree;
1705 	long bno;
1706 	int frags, bbase;
1707 	int i, error;
1708 	u_int8_t *blksfree;
1709 
1710 	ump = ITOUMP(ip);
1711 	fs = ump->um_fs;
1712 	if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize))
1713 		return (0);
1714 	frags = numfrags(fs, nsize);
1715 	bbase = fragnum(fs, bprev);
1716 	if (bbase > fragnum(fs, (bprev + frags - 1))) {
1717 		/* cannot extend across a block boundary */
1718 		return (0);
1719 	}
1720 	UFS_UNLOCK(ump);
1721 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0)
1722 		goto fail;
1723 	bno = dtogd(fs, bprev);
1724 	blksfree = cg_blksfree(cgp);
1725 	for (i = numfrags(fs, osize); i < frags; i++)
1726 		if (isclr(blksfree, bno + i))
1727 			goto fail;
1728 	/*
1729 	 * the current fragment can be extended
1730 	 * deduct the count on fragment being extended into
1731 	 * increase the count on the remaining fragment (if any)
1732 	 * allocate the extended piece
1733 	 */
1734 	for (i = frags; i < fs->fs_frag - bbase; i++)
1735 		if (isclr(blksfree, bno + i))
1736 			break;
1737 	cgp->cg_frsum[i - numfrags(fs, osize)]--;
1738 	if (i != frags)
1739 		cgp->cg_frsum[i - frags]++;
1740 	for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) {
1741 		clrbit(blksfree, bno + i);
1742 		cgp->cg_cs.cs_nffree--;
1743 		nffree++;
1744 	}
1745 	UFS_LOCK(ump);
1746 	fs->fs_cstotal.cs_nffree -= nffree;
1747 	fs->fs_cs(fs, cg).cs_nffree -= nffree;
1748 	fs->fs_fmod = 1;
1749 	ACTIVECLEAR(fs, cg);
1750 	UFS_UNLOCK(ump);
1751 	if (DOINGSOFTDEP(ITOV(ip)))
1752 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev,
1753 		    frags, numfrags(fs, osize));
1754 	bdwrite(bp);
1755 	return (bprev);
1756 
1757 fail:
1758 	brelse(bp);
1759 	UFS_LOCK(ump);
1760 	return (0);
1761 
1762 }
1763 
1764 /*
1765  * Determine whether a block can be allocated.
1766  *
1767  * Check to see if a block of the appropriate size is available,
1768  * and if it is, allocate it.
1769  */
1770 static ufs2_daddr_t
1771 ffs_alloccg(struct inode *ip,
1772 	u_int cg,
1773 	ufs2_daddr_t bpref,
1774 	int size,
1775 	int rsize)
1776 {
1777 	struct fs *fs;
1778 	struct cg *cgp;
1779 	struct buf *bp;
1780 	struct ufsmount *ump;
1781 	ufs1_daddr_t bno;
1782 	ufs2_daddr_t blkno;
1783 	int i, allocsiz, error, frags;
1784 	u_int8_t *blksfree;
1785 
1786 	ump = ITOUMP(ip);
1787 	fs = ump->um_fs;
1788 	if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize)
1789 		return (0);
1790 	UFS_UNLOCK(ump);
1791 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0 ||
1792 	   (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize))
1793 		goto fail;
1794 	if (size == fs->fs_bsize) {
1795 		UFS_LOCK(ump);
1796 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1797 		ACTIVECLEAR(fs, cg);
1798 		UFS_UNLOCK(ump);
1799 		bdwrite(bp);
1800 		return (blkno);
1801 	}
1802 	/*
1803 	 * check to see if any fragments are already available
1804 	 * allocsiz is the size which will be allocated, hacking
1805 	 * it down to a smaller size if necessary
1806 	 */
1807 	blksfree = cg_blksfree(cgp);
1808 	frags = numfrags(fs, size);
1809 	for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++)
1810 		if (cgp->cg_frsum[allocsiz] != 0)
1811 			break;
1812 	if (allocsiz == fs->fs_frag) {
1813 		/*
1814 		 * no fragments were available, so a block will be
1815 		 * allocated, and hacked up
1816 		 */
1817 		if (cgp->cg_cs.cs_nbfree == 0)
1818 			goto fail;
1819 		UFS_LOCK(ump);
1820 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1821 		ACTIVECLEAR(fs, cg);
1822 		UFS_UNLOCK(ump);
1823 		bdwrite(bp);
1824 		return (blkno);
1825 	}
1826 	KASSERT(size == rsize,
1827 	    ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize));
1828 	bno = ffs_mapsearch(fs, cgp, bpref, allocsiz);
1829 	if (bno < 0)
1830 		goto fail;
1831 	for (i = 0; i < frags; i++)
1832 		clrbit(blksfree, bno + i);
1833 	cgp->cg_cs.cs_nffree -= frags;
1834 	cgp->cg_frsum[allocsiz]--;
1835 	if (frags != allocsiz)
1836 		cgp->cg_frsum[allocsiz - frags]++;
1837 	UFS_LOCK(ump);
1838 	fs->fs_cstotal.cs_nffree -= frags;
1839 	fs->fs_cs(fs, cg).cs_nffree -= frags;
1840 	fs->fs_fmod = 1;
1841 	blkno = cgbase(fs, cg) + bno;
1842 	ACTIVECLEAR(fs, cg);
1843 	UFS_UNLOCK(ump);
1844 	if (DOINGSOFTDEP(ITOV(ip)))
1845 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0);
1846 	bdwrite(bp);
1847 	return (blkno);
1848 
1849 fail:
1850 	brelse(bp);
1851 	UFS_LOCK(ump);
1852 	return (0);
1853 }
1854 
1855 /*
1856  * Allocate a block in a cylinder group.
1857  *
1858  * This algorithm implements the following policy:
1859  *   1) allocate the requested block.
1860  *   2) allocate a rotationally optimal block in the same cylinder.
1861  *   3) allocate the next available block on the block rotor for the
1862  *      specified cylinder group.
1863  * Note that this routine only allocates fs_bsize blocks; these
1864  * blocks may be fragmented by the routine that allocates them.
1865  */
1866 static ufs2_daddr_t
1867 ffs_alloccgblk(struct inode *ip,
1868 	struct buf *bp,
1869 	ufs2_daddr_t bpref,
1870 	int size)
1871 {
1872 	struct fs *fs;
1873 	struct cg *cgp;
1874 	struct ufsmount *ump;
1875 	ufs1_daddr_t bno;
1876 	ufs2_daddr_t blkno;
1877 	u_int8_t *blksfree;
1878 	int i, cgbpref;
1879 
1880 	ump = ITOUMP(ip);
1881 	fs = ump->um_fs;
1882 	mtx_assert(UFS_MTX(ump), MA_OWNED);
1883 	cgp = (struct cg *)bp->b_data;
1884 	blksfree = cg_blksfree(cgp);
1885 	if (bpref == 0) {
1886 		bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag;
1887 	} else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) {
1888 		/* map bpref to correct zone in this cg */
1889 		if (bpref < cgdata(fs, cgbpref))
1890 			bpref = cgmeta(fs, cgp->cg_cgx);
1891 		else
1892 			bpref = cgdata(fs, cgp->cg_cgx);
1893 	}
1894 	/*
1895 	 * if the requested block is available, use it
1896 	 */
1897 	bno = dtogd(fs, blknum(fs, bpref));
1898 	if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno)))
1899 		goto gotit;
1900 	/*
1901 	 * Take the next available block in this cylinder group.
1902 	 */
1903 	bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag);
1904 	if (bno < 0)
1905 		return (0);
1906 	/* Update cg_rotor only if allocated from the data zone */
1907 	if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx)))
1908 		cgp->cg_rotor = bno;
1909 gotit:
1910 	blkno = fragstoblks(fs, bno);
1911 	ffs_clrblock(fs, blksfree, (long)blkno);
1912 	ffs_clusteracct(fs, cgp, blkno, -1);
1913 	cgp->cg_cs.cs_nbfree--;
1914 	fs->fs_cstotal.cs_nbfree--;
1915 	fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--;
1916 	fs->fs_fmod = 1;
1917 	blkno = cgbase(fs, cgp->cg_cgx) + bno;
1918 	/*
1919 	 * If the caller didn't want the whole block free the frags here.
1920 	 */
1921 	size = numfrags(fs, size);
1922 	if (size != fs->fs_frag) {
1923 		bno = dtogd(fs, blkno);
1924 		for (i = size; i < fs->fs_frag; i++)
1925 			setbit(blksfree, bno + i);
1926 		i = fs->fs_frag - size;
1927 		cgp->cg_cs.cs_nffree += i;
1928 		fs->fs_cstotal.cs_nffree += i;
1929 		fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i;
1930 		fs->fs_fmod = 1;
1931 		cgp->cg_frsum[i]++;
1932 	}
1933 	/* XXX Fixme. */
1934 	UFS_UNLOCK(ump);
1935 	if (DOINGSOFTDEP(ITOV(ip)))
1936 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0);
1937 	UFS_LOCK(ump);
1938 	return (blkno);
1939 }
1940 
1941 /*
1942  * Determine whether a cluster can be allocated.
1943  *
1944  * We do not currently check for optimal rotational layout if there
1945  * are multiple choices in the same cylinder group. Instead we just
1946  * take the first one that we find following bpref.
1947  */
1948 static ufs2_daddr_t
1949 ffs_clusteralloc(struct inode *ip,
1950 	u_int cg,
1951 	ufs2_daddr_t bpref,
1952 	int len)
1953 {
1954 	struct fs *fs;
1955 	struct cg *cgp;
1956 	struct buf *bp;
1957 	struct ufsmount *ump;
1958 	int i, run, bit, map, got, error;
1959 	ufs2_daddr_t bno;
1960 	u_char *mapp;
1961 	int32_t *lp;
1962 	u_int8_t *blksfree;
1963 
1964 	ump = ITOUMP(ip);
1965 	fs = ump->um_fs;
1966 	if (fs->fs_maxcluster[cg] < len)
1967 		return (0);
1968 	UFS_UNLOCK(ump);
1969 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) {
1970 		UFS_LOCK(ump);
1971 		return (0);
1972 	}
1973 	/*
1974 	 * Check to see if a cluster of the needed size (or bigger) is
1975 	 * available in this cylinder group.
1976 	 */
1977 	lp = &cg_clustersum(cgp)[len];
1978 	for (i = len; i <= fs->fs_contigsumsize; i++)
1979 		if (*lp++ > 0)
1980 			break;
1981 	if (i > fs->fs_contigsumsize) {
1982 		/*
1983 		 * This is the first time looking for a cluster in this
1984 		 * cylinder group. Update the cluster summary information
1985 		 * to reflect the true maximum sized cluster so that
1986 		 * future cluster allocation requests can avoid reading
1987 		 * the cylinder group map only to find no clusters.
1988 		 */
1989 		lp = &cg_clustersum(cgp)[len - 1];
1990 		for (i = len - 1; i > 0; i--)
1991 			if (*lp-- > 0)
1992 				break;
1993 		UFS_LOCK(ump);
1994 		fs->fs_maxcluster[cg] = i;
1995 		brelse(bp);
1996 		return (0);
1997 	}
1998 	/*
1999 	 * Search the cluster map to find a big enough cluster.
2000 	 * We take the first one that we find, even if it is larger
2001 	 * than we need as we prefer to get one close to the previous
2002 	 * block allocation. We do not search before the current
2003 	 * preference point as we do not want to allocate a block
2004 	 * that is allocated before the previous one (as we will
2005 	 * then have to wait for another pass of the elevator
2006 	 * algorithm before it will be read). We prefer to fail and
2007 	 * be recalled to try an allocation in the next cylinder group.
2008 	 */
2009 	if (dtog(fs, bpref) != cg)
2010 		bpref = cgdata(fs, cg);
2011 	else
2012 		bpref = blknum(fs, bpref);
2013 	bpref = fragstoblks(fs, dtogd(fs, bpref));
2014 	mapp = &cg_clustersfree(cgp)[bpref / NBBY];
2015 	map = *mapp++;
2016 	bit = 1 << (bpref % NBBY);
2017 	for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) {
2018 		if ((map & bit) == 0) {
2019 			run = 0;
2020 		} else {
2021 			run++;
2022 			if (run == len)
2023 				break;
2024 		}
2025 		if ((got & (NBBY - 1)) != (NBBY - 1)) {
2026 			bit <<= 1;
2027 		} else {
2028 			map = *mapp++;
2029 			bit = 1;
2030 		}
2031 	}
2032 	if (got >= cgp->cg_nclusterblks) {
2033 		UFS_LOCK(ump);
2034 		brelse(bp);
2035 		return (0);
2036 	}
2037 	/*
2038 	 * Allocate the cluster that we have found.
2039 	 */
2040 	blksfree = cg_blksfree(cgp);
2041 	for (i = 1; i <= len; i++)
2042 		if (!ffs_isblock(fs, blksfree, got - run + i))
2043 			panic("ffs_clusteralloc: map mismatch");
2044 	bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1);
2045 	if (dtog(fs, bno) != cg)
2046 		panic("ffs_clusteralloc: allocated out of group");
2047 	len = blkstofrags(fs, len);
2048 	UFS_LOCK(ump);
2049 	for (i = 0; i < len; i += fs->fs_frag)
2050 		if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i)
2051 			panic("ffs_clusteralloc: lost block");
2052 	ACTIVECLEAR(fs, cg);
2053 	UFS_UNLOCK(ump);
2054 	bdwrite(bp);
2055 	return (bno);
2056 }
2057 
2058 static inline struct buf *
2059 getinobuf(struct inode *ip,
2060 	u_int cg,
2061 	u_int32_t cginoblk,
2062 	int gbflags)
2063 {
2064 	struct fs *fs;
2065 
2066 	fs = ITOFS(ip);
2067 	return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs,
2068 	    cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
2069 	    gbflags));
2070 }
2071 
2072 /*
2073  * Synchronous inode initialization is needed only when barrier writes do not
2074  * work as advertised, and will impose a heavy cost on file creation in a newly
2075  * created filesystem.
2076  */
2077 static int doasyncinodeinit = 1;
2078 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN,
2079     &doasyncinodeinit, 0,
2080     "Perform inode block initialization using asynchronous writes");
2081 
2082 /*
2083  * Determine whether an inode can be allocated.
2084  *
2085  * Check to see if an inode is available, and if it is,
2086  * allocate it using the following policy:
2087  *   1) allocate the requested inode.
2088  *   2) allocate the next available inode after the requested
2089  *      inode in the specified cylinder group.
2090  */
2091 static ufs2_daddr_t
2092 ffs_nodealloccg(struct inode *ip,
2093 	u_int cg,
2094 	ufs2_daddr_t ipref,
2095 	int mode,
2096 	int unused)
2097 {
2098 	struct fs *fs;
2099 	struct cg *cgp;
2100 	struct buf *bp, *ibp;
2101 	struct ufsmount *ump;
2102 	u_int8_t *inosused, *loc;
2103 	struct ufs2_dinode *dp2;
2104 	int error, start, len, i;
2105 	u_int32_t old_initediblk;
2106 
2107 	ump = ITOUMP(ip);
2108 	fs = ump->um_fs;
2109 check_nifree:
2110 	if (fs->fs_cs(fs, cg).cs_nifree == 0)
2111 		return (0);
2112 	UFS_UNLOCK(ump);
2113 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) {
2114 		UFS_LOCK(ump);
2115 		return (0);
2116 	}
2117 restart:
2118 	if (cgp->cg_cs.cs_nifree == 0) {
2119 		brelse(bp);
2120 		UFS_LOCK(ump);
2121 		return (0);
2122 	}
2123 	inosused = cg_inosused(cgp);
2124 	if (ipref) {
2125 		ipref %= fs->fs_ipg;
2126 		if (isclr(inosused, ipref))
2127 			goto gotit;
2128 	}
2129 	start = cgp->cg_irotor / NBBY;
2130 	len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY);
2131 	loc = memcchr(&inosused[start], 0xff, len);
2132 	if (loc == NULL) {
2133 		len = start + 1;
2134 		start = 0;
2135 		loc = memcchr(&inosused[start], 0xff, len);
2136 		if (loc == NULL) {
2137 			printf("cg = %d, irotor = %ld, fs = %s\n",
2138 			    cg, (long)cgp->cg_irotor, fs->fs_fsmnt);
2139 			panic("ffs_nodealloccg: map corrupted");
2140 			/* NOTREACHED */
2141 		}
2142 	}
2143 	ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
2144 gotit:
2145 	/*
2146 	 * Check to see if we need to initialize more inodes.
2147 	 */
2148 	if (fs->fs_magic == FS_UFS2_MAGIC &&
2149 	    ipref + INOPB(fs) > cgp->cg_initediblk &&
2150 	    cgp->cg_initediblk < cgp->cg_niblk) {
2151 		old_initediblk = cgp->cg_initediblk;
2152 
2153 		/*
2154 		 * Free the cylinder group lock before writing the
2155 		 * initialized inode block.  Entering the
2156 		 * babarrierwrite() with the cylinder group lock
2157 		 * causes lock order violation between the lock and
2158 		 * snaplk.
2159 		 *
2160 		 * Another thread can decide to initialize the same
2161 		 * inode block, but whichever thread first gets the
2162 		 * cylinder group lock after writing the newly
2163 		 * allocated inode block will update it and the other
2164 		 * will realize that it has lost and leave the
2165 		 * cylinder group unchanged.
2166 		 */
2167 		ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
2168 		brelse(bp);
2169 		if (ibp == NULL) {
2170 			/*
2171 			 * The inode block buffer is already owned by
2172 			 * another thread, which must initialize it.
2173 			 * Wait on the buffer to allow another thread
2174 			 * to finish the updates, with dropped cg
2175 			 * buffer lock, then retry.
2176 			 */
2177 			ibp = getinobuf(ip, cg, old_initediblk, 0);
2178 			brelse(ibp);
2179 			UFS_LOCK(ump);
2180 			goto check_nifree;
2181 		}
2182 		bzero(ibp->b_data, (int)fs->fs_bsize);
2183 		dp2 = (struct ufs2_dinode *)(ibp->b_data);
2184 		for (i = 0; i < INOPB(fs); i++) {
2185 			while (dp2->di_gen == 0)
2186 				dp2->di_gen = arc4random();
2187 			dp2++;
2188 		}
2189 
2190 		/*
2191 		 * Rather than adding a soft updates dependency to ensure
2192 		 * that the new inode block is written before it is claimed
2193 		 * by the cylinder group map, we just do a barrier write
2194 		 * here. The barrier write will ensure that the inode block
2195 		 * gets written before the updated cylinder group map can be
2196 		 * written. The barrier write should only slow down bulk
2197 		 * loading of newly created filesystems.
2198 		 */
2199 		if (doasyncinodeinit)
2200 			babarrierwrite(ibp);
2201 		else
2202 			bwrite(ibp);
2203 
2204 		/*
2205 		 * After the inode block is written, try to update the
2206 		 * cg initediblk pointer.  If another thread beat us
2207 		 * to it, then leave it unchanged as the other thread
2208 		 * has already set it correctly.
2209 		 */
2210 		error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp);
2211 		UFS_LOCK(ump);
2212 		ACTIVECLEAR(fs, cg);
2213 		UFS_UNLOCK(ump);
2214 		if (error != 0)
2215 			return (error);
2216 		if (cgp->cg_initediblk == old_initediblk)
2217 			cgp->cg_initediblk += INOPB(fs);
2218 		goto restart;
2219 	}
2220 	cgp->cg_irotor = ipref;
2221 	UFS_LOCK(ump);
2222 	ACTIVECLEAR(fs, cg);
2223 	setbit(inosused, ipref);
2224 	cgp->cg_cs.cs_nifree--;
2225 	fs->fs_cstotal.cs_nifree--;
2226 	fs->fs_cs(fs, cg).cs_nifree--;
2227 	fs->fs_fmod = 1;
2228 	if ((mode & IFMT) == IFDIR) {
2229 		cgp->cg_cs.cs_ndir++;
2230 		fs->fs_cstotal.cs_ndir++;
2231 		fs->fs_cs(fs, cg).cs_ndir++;
2232 	}
2233 	UFS_UNLOCK(ump);
2234 	if (DOINGSOFTDEP(ITOV(ip)))
2235 		softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode);
2236 	bdwrite(bp);
2237 	return ((ino_t)(cg * fs->fs_ipg + ipref));
2238 }
2239 
2240 /*
2241  * Free a block or fragment.
2242  *
2243  * The specified block or fragment is placed back in the
2244  * free map. If a fragment is deallocated, a possible
2245  * block reassembly is checked.
2246  */
2247 static void
2248 ffs_blkfree_cg(struct ufsmount *ump,
2249 	struct fs *fs,
2250 	struct vnode *devvp,
2251 	ufs2_daddr_t bno,
2252 	long size,
2253 	ino_t inum,
2254 	struct workhead *dephd)
2255 {
2256 	struct mount *mp;
2257 	struct cg *cgp;
2258 	struct buf *bp;
2259 	daddr_t dbn;
2260 	ufs1_daddr_t fragno, cgbno;
2261 	int i, blk, frags, bbase, error;
2262 	u_int cg;
2263 	u_int8_t *blksfree;
2264 	struct cdev *dev;
2265 
2266 	cg = dtog(fs, bno);
2267 	if (devvp->v_type == VREG) {
2268 		/* devvp is a snapshot */
2269 		MPASS(devvp->v_mount->mnt_data == ump);
2270 		dev = ump->um_devvp->v_rdev;
2271 	} else if (devvp->v_type == VCHR) {
2272 		/*
2273 		 * devvp is a normal disk device
2274 		 * XXXKIB: devvp is not locked there, v_rdev access depends on
2275 		 * busy mount, which prevents mntfs devvp from reclamation.
2276 		 */
2277 		dev = devvp->v_rdev;
2278 	} else
2279 		return;
2280 #ifdef INVARIANTS
2281 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 ||
2282 	    fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) {
2283 		printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n",
2284 		    devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize,
2285 		    size, fs->fs_fsmnt);
2286 		panic("ffs_blkfree_cg: bad size");
2287 	}
2288 #endif
2289 	if ((u_int)bno >= fs->fs_size) {
2290 		printf("bad block %jd, ino %lu\n", (intmax_t)bno,
2291 		    (u_long)inum);
2292 		ffs_fserr(fs, inum, "bad block");
2293 		return;
2294 	}
2295 	if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) {
2296 		if (!ffs_fsfail_cleanup(ump, error) ||
2297 		    !MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR)
2298 			return;
2299 		if (devvp->v_type == VREG)
2300 			dbn = fragstoblks(fs, cgtod(fs, cg));
2301 		else
2302 			dbn = fsbtodb(fs, cgtod(fs, cg));
2303 		error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp);
2304 		KASSERT(error == 0, ("getblkx failed"));
2305 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2306 		    numfrags(fs, size), dephd);
2307 		bp->b_flags |= B_RELBUF | B_NOCACHE;
2308 		bp->b_flags &= ~B_CACHE;
2309 		bawrite(bp);
2310 		return;
2311 	}
2312 	cgbno = dtogd(fs, bno);
2313 	blksfree = cg_blksfree(cgp);
2314 	UFS_LOCK(ump);
2315 	if (size == fs->fs_bsize) {
2316 		fragno = fragstoblks(fs, cgbno);
2317 		if (!ffs_isfreeblock(fs, blksfree, fragno)) {
2318 			if (devvp->v_type == VREG) {
2319 				UFS_UNLOCK(ump);
2320 				/* devvp is a snapshot */
2321 				brelse(bp);
2322 				return;
2323 			}
2324 			printf("dev = %s, block = %jd, fs = %s\n",
2325 			    devtoname(dev), (intmax_t)bno, fs->fs_fsmnt);
2326 			panic("ffs_blkfree_cg: freeing free block");
2327 		}
2328 		ffs_setblock(fs, blksfree, fragno);
2329 		ffs_clusteracct(fs, cgp, fragno, 1);
2330 		cgp->cg_cs.cs_nbfree++;
2331 		fs->fs_cstotal.cs_nbfree++;
2332 		fs->fs_cs(fs, cg).cs_nbfree++;
2333 	} else {
2334 		bbase = cgbno - fragnum(fs, cgbno);
2335 		/*
2336 		 * decrement the counts associated with the old frags
2337 		 */
2338 		blk = blkmap(fs, blksfree, bbase);
2339 		ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
2340 		/*
2341 		 * deallocate the fragment
2342 		 */
2343 		frags = numfrags(fs, size);
2344 		for (i = 0; i < frags; i++) {
2345 			if (isset(blksfree, cgbno + i)) {
2346 				printf("dev = %s, block = %jd, fs = %s\n",
2347 				    devtoname(dev), (intmax_t)(bno + i),
2348 				    fs->fs_fsmnt);
2349 				panic("ffs_blkfree_cg: freeing free frag");
2350 			}
2351 			setbit(blksfree, cgbno + i);
2352 		}
2353 		cgp->cg_cs.cs_nffree += i;
2354 		fs->fs_cstotal.cs_nffree += i;
2355 		fs->fs_cs(fs, cg).cs_nffree += i;
2356 		/*
2357 		 * add back in counts associated with the new frags
2358 		 */
2359 		blk = blkmap(fs, blksfree, bbase);
2360 		ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
2361 		/*
2362 		 * if a complete block has been reassembled, account for it
2363 		 */
2364 		fragno = fragstoblks(fs, bbase);
2365 		if (ffs_isblock(fs, blksfree, fragno)) {
2366 			cgp->cg_cs.cs_nffree -= fs->fs_frag;
2367 			fs->fs_cstotal.cs_nffree -= fs->fs_frag;
2368 			fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag;
2369 			ffs_clusteracct(fs, cgp, fragno, 1);
2370 			cgp->cg_cs.cs_nbfree++;
2371 			fs->fs_cstotal.cs_nbfree++;
2372 			fs->fs_cs(fs, cg).cs_nbfree++;
2373 		}
2374 	}
2375 	fs->fs_fmod = 1;
2376 	ACTIVECLEAR(fs, cg);
2377 	UFS_UNLOCK(ump);
2378 	mp = UFSTOVFS(ump);
2379 	if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR)
2380 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2381 		    numfrags(fs, size), dephd);
2382 	bdwrite(bp);
2383 }
2384 
2385 /*
2386  * Structures and routines associated with trim management.
2387  *
2388  * The following requests are passed to trim_lookup to indicate
2389  * the actions that should be taken.
2390  */
2391 #define	NEW	1	/* if found, error else allocate and hash it */
2392 #define	OLD	2	/* if not found, error, else return it */
2393 #define	REPLACE	3	/* if not found, error else unhash and reallocate it */
2394 #define	DONE	4	/* if not found, error else unhash and return it */
2395 #define	SINGLE	5	/* don't look up, just allocate it and don't hash it */
2396 
2397 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures");
2398 
2399 #define	TRIMLIST_HASH(ump, key) \
2400 	(&(ump)->um_trimhash[(key) & (ump)->um_trimlisthashsize])
2401 
2402 /*
2403  * These structures describe each of the block free requests aggregated
2404  * together to make up a trim request.
2405  */
2406 struct trim_blkreq {
2407 	TAILQ_ENTRY(trim_blkreq) blkreqlist;
2408 	ufs2_daddr_t bno;
2409 	long size;
2410 	struct workhead *pdephd;
2411 	struct workhead dephd;
2412 };
2413 
2414 /*
2415  * Description of a trim request.
2416  */
2417 struct ffs_blkfree_trim_params {
2418 	TAILQ_HEAD(, trim_blkreq) blklist;
2419 	LIST_ENTRY(ffs_blkfree_trim_params) hashlist;
2420 	struct task task;
2421 	struct ufsmount *ump;
2422 	struct vnode *devvp;
2423 	ino_t inum;
2424 	ufs2_daddr_t bno;
2425 	long size;
2426 	long key;
2427 };
2428 
2429 static void	ffs_blkfree_trim_completed(struct buf *);
2430 static void	ffs_blkfree_trim_task(void *ctx, int pending __unused);
2431 static struct	ffs_blkfree_trim_params *trim_lookup(struct ufsmount *,
2432 		    struct vnode *, ufs2_daddr_t, long, ino_t, u_long, int);
2433 static void	ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *);
2434 
2435 /*
2436  * Called on trim completion to start a task to free the associated block(s).
2437  */
2438 static void
2439 ffs_blkfree_trim_completed(struct buf *bp)
2440 {
2441 	struct ffs_blkfree_trim_params *tp;
2442 
2443 	tp = bp->b_fsprivate1;
2444 	free(bp, M_TRIM);
2445 	TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp);
2446 	taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task);
2447 }
2448 
2449 /*
2450  * Trim completion task that free associated block(s).
2451  */
2452 static void
2453 ffs_blkfree_trim_task(void *ctx, int pending)
2454 {
2455 	struct ffs_blkfree_trim_params *tp;
2456 	struct trim_blkreq *blkelm;
2457 	struct ufsmount *ump;
2458 
2459 	tp = ctx;
2460 	ump = tp->ump;
2461 	while ((blkelm = TAILQ_FIRST(&tp->blklist)) != NULL) {
2462 		ffs_blkfree_cg(ump, ump->um_fs, tp->devvp, blkelm->bno,
2463 		    blkelm->size, tp->inum, blkelm->pdephd);
2464 		TAILQ_REMOVE(&tp->blklist, blkelm, blkreqlist);
2465 		free(blkelm, M_TRIM);
2466 	}
2467 	vn_finished_secondary_write(UFSTOVFS(ump));
2468 	UFS_LOCK(ump);
2469 	ump->um_trim_inflight -= 1;
2470 	ump->um_trim_inflight_blks -= numfrags(ump->um_fs, tp->size);
2471 	UFS_UNLOCK(ump);
2472 	free(tp, M_TRIM);
2473 }
2474 
2475 /*
2476  * Lookup a trim request by inode number.
2477  * Allocate if requested (NEW, REPLACE, SINGLE).
2478  */
2479 static struct ffs_blkfree_trim_params *
2480 trim_lookup(struct ufsmount *ump,
2481 	struct vnode *devvp,
2482 	ufs2_daddr_t bno,
2483 	long size,
2484 	ino_t inum,
2485 	u_long key,
2486 	int alloctype)
2487 {
2488 	struct trimlist_hashhead *tphashhead;
2489 	struct ffs_blkfree_trim_params *tp, *ntp;
2490 
2491 	ntp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK);
2492 	if (alloctype != SINGLE) {
2493 		KASSERT(key >= FIRST_VALID_KEY, ("trim_lookup: invalid key"));
2494 		UFS_LOCK(ump);
2495 		tphashhead = TRIMLIST_HASH(ump, key);
2496 		LIST_FOREACH(tp, tphashhead, hashlist)
2497 			if (key == tp->key)
2498 				break;
2499 	}
2500 	switch (alloctype) {
2501 	case NEW:
2502 		KASSERT(tp == NULL, ("trim_lookup: found trim"));
2503 		break;
2504 	case OLD:
2505 		KASSERT(tp != NULL,
2506 		    ("trim_lookup: missing call to ffs_blkrelease_start()"));
2507 		UFS_UNLOCK(ump);
2508 		free(ntp, M_TRIM);
2509 		return (tp);
2510 	case REPLACE:
2511 		KASSERT(tp != NULL, ("trim_lookup: missing REPLACE trim"));
2512 		LIST_REMOVE(tp, hashlist);
2513 		/* tp will be freed by caller */
2514 		break;
2515 	case DONE:
2516 		KASSERT(tp != NULL, ("trim_lookup: missing DONE trim"));
2517 		LIST_REMOVE(tp, hashlist);
2518 		UFS_UNLOCK(ump);
2519 		free(ntp, M_TRIM);
2520 		return (tp);
2521 	}
2522 	TAILQ_INIT(&ntp->blklist);
2523 	ntp->ump = ump;
2524 	ntp->devvp = devvp;
2525 	ntp->bno = bno;
2526 	ntp->size = size;
2527 	ntp->inum = inum;
2528 	ntp->key = key;
2529 	if (alloctype != SINGLE) {
2530 		LIST_INSERT_HEAD(tphashhead, ntp, hashlist);
2531 		UFS_UNLOCK(ump);
2532 	}
2533 	return (ntp);
2534 }
2535 
2536 /*
2537  * Dispatch a trim request.
2538  */
2539 static void
2540 ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *tp)
2541 {
2542 	struct ufsmount *ump;
2543 	struct mount *mp;
2544 	struct buf *bp;
2545 
2546 	/*
2547 	 * Postpone the set of the free bit in the cg bitmap until the
2548 	 * BIO_DELETE is completed.  Otherwise, due to disk queue
2549 	 * reordering, TRIM might be issued after we reuse the block
2550 	 * and write some new data into it.
2551 	 */
2552 	ump = tp->ump;
2553 	bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO);
2554 	bp->b_iocmd = BIO_DELETE;
2555 	bp->b_iooffset = dbtob(fsbtodb(ump->um_fs, tp->bno));
2556 	bp->b_iodone = ffs_blkfree_trim_completed;
2557 	bp->b_bcount = tp->size;
2558 	bp->b_fsprivate1 = tp;
2559 	UFS_LOCK(ump);
2560 	ump->um_trim_total += 1;
2561 	ump->um_trim_inflight += 1;
2562 	ump->um_trim_inflight_blks += numfrags(ump->um_fs, tp->size);
2563 	ump->um_trim_total_blks += numfrags(ump->um_fs, tp->size);
2564 	UFS_UNLOCK(ump);
2565 
2566 	mp = UFSTOVFS(ump);
2567 	vn_start_secondary_write(NULL, &mp, 0);
2568 	g_vfs_strategy(ump->um_bo, bp);
2569 }
2570 
2571 /*
2572  * Allocate a new key to use to identify a range of blocks.
2573  */
2574 u_long
2575 ffs_blkrelease_start(struct ufsmount *ump,
2576 	struct vnode *devvp,
2577 	ino_t inum)
2578 {
2579 	static u_long masterkey;
2580 	u_long key;
2581 
2582 	if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2583 		return (SINGLETON_KEY);
2584 	do {
2585 		key = atomic_fetchadd_long(&masterkey, 1);
2586 	} while (key < FIRST_VALID_KEY);
2587 	(void) trim_lookup(ump, devvp, 0, 0, inum, key, NEW);
2588 	return (key);
2589 }
2590 
2591 /*
2592  * Deallocate a key that has been used to identify a range of blocks.
2593  */
2594 void
2595 ffs_blkrelease_finish(struct ufsmount *ump, u_long key)
2596 {
2597 	struct ffs_blkfree_trim_params *tp;
2598 
2599 	if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2600 		return;
2601 	/*
2602 	 * If the vfs.ffs.dotrimcons sysctl option is enabled while
2603 	 * a file deletion is active, specifically after a call
2604 	 * to ffs_blkrelease_start() but before the call to
2605 	 * ffs_blkrelease_finish(), ffs_blkrelease_start() will
2606 	 * have handed out SINGLETON_KEY rather than starting a
2607 	 * collection sequence. Thus if we get a SINGLETON_KEY
2608 	 * passed to ffs_blkrelease_finish(), we just return rather
2609 	 * than trying to finish the nonexistent sequence.
2610 	 */
2611 	if (key == SINGLETON_KEY) {
2612 #ifdef INVARIANTS
2613 		printf("%s: vfs.ffs.dotrimcons enabled on active filesystem\n",
2614 		    ump->um_mountp->mnt_stat.f_mntonname);
2615 #endif
2616 		return;
2617 	}
2618 	/*
2619 	 * We are done with sending blocks using this key. Look up the key
2620 	 * using the DONE alloctype (in tp) to request that it be unhashed
2621 	 * as we will not be adding to it. If the key has never been used,
2622 	 * tp->size will be zero, so we can just free tp. Otherwise the call
2623 	 * to ffs_blkfree_sendtrim(tp) causes the block range described by
2624 	 * tp to be issued (and then tp to be freed).
2625 	 */
2626 	tp = trim_lookup(ump, NULL, 0, 0, 0, key, DONE);
2627 	if (tp->size == 0)
2628 		free(tp, M_TRIM);
2629 	else
2630 		ffs_blkfree_sendtrim(tp);
2631 }
2632 
2633 /*
2634  * Setup to free a block or fragment.
2635  *
2636  * Check for snapshots that might want to claim the block.
2637  * If trims are requested, prepare a trim request. Attempt to
2638  * aggregate consecutive blocks into a single trim request.
2639  */
2640 void
2641 ffs_blkfree(struct ufsmount *ump,
2642 	struct fs *fs,
2643 	struct vnode *devvp,
2644 	ufs2_daddr_t bno,
2645 	long size,
2646 	ino_t inum,
2647 	enum vtype vtype,
2648 	struct workhead *dephd,
2649 	u_long key)
2650 {
2651 	struct ffs_blkfree_trim_params *tp, *ntp;
2652 	struct trim_blkreq *blkelm;
2653 
2654 	/*
2655 	 * Check to see if a snapshot wants to claim the block.
2656 	 * Check that devvp is a normal disk device, not a snapshot,
2657 	 * it has a snapshot(s) associated with it, and one of the
2658 	 * snapshots wants to claim the block.
2659 	 */
2660 	if (devvp->v_type == VCHR &&
2661 	    (devvp->v_vflag & VV_COPYONWRITE) &&
2662 	    ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) {
2663 		return;
2664 	}
2665 	/*
2666 	 * Nothing to delay if TRIM is not required for this block or TRIM
2667 	 * is disabled or the operation is performed on a snapshot.
2668 	 */
2669 	if (key == NOTRIM_KEY || ((ump->um_flags & UM_CANDELETE) == 0) ||
2670 	    devvp->v_type == VREG) {
2671 		ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd);
2672 		return;
2673 	}
2674 	blkelm = malloc(sizeof(struct trim_blkreq), M_TRIM, M_WAITOK);
2675 	blkelm->bno = bno;
2676 	blkelm->size = size;
2677 	if (dephd == NULL) {
2678 		blkelm->pdephd = NULL;
2679 	} else {
2680 		LIST_INIT(&blkelm->dephd);
2681 		LIST_SWAP(dephd, &blkelm->dephd, worklist, wk_list);
2682 		blkelm->pdephd = &blkelm->dephd;
2683 	}
2684 	if (key == SINGLETON_KEY) {
2685 		/*
2686 		 * Just a single non-contiguous piece. Use the SINGLE
2687 		 * alloctype to return a trim request that will not be
2688 		 * hashed for future lookup.
2689 		 */
2690 		tp = trim_lookup(ump, devvp, bno, size, inum, key, SINGLE);
2691 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2692 		ffs_blkfree_sendtrim(tp);
2693 		return;
2694 	}
2695 	/*
2696 	 * The callers of this function are not tracking whether or not
2697 	 * the blocks are contiguous. They are just saying that they
2698 	 * are freeing a set of blocks. It is this code that determines
2699 	 * the pieces of that range that are actually contiguous.
2700 	 *
2701 	 * Calling ffs_blkrelease_start() will have created an entry
2702 	 * that we will use.
2703 	 */
2704 	tp = trim_lookup(ump, devvp, bno, size, inum, key, OLD);
2705 	if (tp->size == 0) {
2706 		/*
2707 		 * First block of a potential range, set block and size
2708 		 * for the trim block.
2709 		 */
2710 		tp->bno = bno;
2711 		tp->size = size;
2712 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2713 		return;
2714 	}
2715 	/*
2716 	 * If this block is a continuation of the range (either
2717 	 * follows at the end or preceeds in the front) then we
2718 	 * add it to the front or back of the list and return.
2719 	 *
2720 	 * If it is not a continuation of the trim that we were
2721 	 * building, using the REPLACE alloctype, we request that
2722 	 * the old trim request (still in tp) be unhashed and a
2723 	 * new range started (in ntp). The ffs_blkfree_sendtrim(tp)
2724 	 * call causes the block range described by tp to be issued
2725 	 * (and then tp to be freed).
2726 	 */
2727 	if (bno + numfrags(fs, size) == tp->bno) {
2728 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2729 		tp->bno = bno;
2730 		tp->size += size;
2731 		return;
2732 	} else if (bno == tp->bno + numfrags(fs, tp->size)) {
2733 		TAILQ_INSERT_TAIL(&tp->blklist, blkelm, blkreqlist);
2734 		tp->size += size;
2735 		return;
2736 	}
2737 	ntp = trim_lookup(ump, devvp, bno, size, inum, key, REPLACE);
2738 	TAILQ_INSERT_HEAD(&ntp->blklist, blkelm, blkreqlist);
2739 	ffs_blkfree_sendtrim(tp);
2740 }
2741 
2742 #ifdef INVARIANTS
2743 /*
2744  * Verify allocation of a block or fragment. Returns true if block or
2745  * fragment is allocated, false if it is free.
2746  */
2747 static int
2748 ffs_checkblk(struct inode *ip,
2749 	ufs2_daddr_t bno,
2750 	long size)
2751 {
2752 	struct fs *fs;
2753 	struct cg *cgp;
2754 	struct buf *bp;
2755 	ufs1_daddr_t cgbno;
2756 	int i, error, frags, free;
2757 	u_int8_t *blksfree;
2758 
2759 	fs = ITOFS(ip);
2760 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
2761 		printf("bsize = %ld, size = %ld, fs = %s\n",
2762 		    (long)fs->fs_bsize, size, fs->fs_fsmnt);
2763 		panic("ffs_checkblk: bad size");
2764 	}
2765 	if ((u_int)bno >= fs->fs_size)
2766 		panic("ffs_checkblk: bad block %jd", (intmax_t)bno);
2767 	error = ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), 0, &bp, &cgp);
2768 	if (error)
2769 		panic("ffs_checkblk: cylinder group read failed");
2770 	blksfree = cg_blksfree(cgp);
2771 	cgbno = dtogd(fs, bno);
2772 	if (size == fs->fs_bsize) {
2773 		free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno));
2774 	} else {
2775 		frags = numfrags(fs, size);
2776 		for (free = 0, i = 0; i < frags; i++)
2777 			if (isset(blksfree, cgbno + i))
2778 				free++;
2779 		if (free != 0 && free != frags)
2780 			panic("ffs_checkblk: partially free fragment");
2781 	}
2782 	brelse(bp);
2783 	return (!free);
2784 }
2785 #endif /* INVARIANTS */
2786 
2787 /*
2788  * Free an inode.
2789  */
2790 int
2791 ffs_vfree(struct vnode *pvp,
2792 	ino_t ino,
2793 	int mode)
2794 {
2795 	struct ufsmount *ump;
2796 
2797 	if (DOINGSOFTDEP(pvp)) {
2798 		softdep_freefile(pvp, ino, mode);
2799 		return (0);
2800 	}
2801 	ump = VFSTOUFS(pvp->v_mount);
2802 	return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL));
2803 }
2804 
2805 /*
2806  * Do the actual free operation.
2807  * The specified inode is placed back in the free map.
2808  */
2809 int
2810 ffs_freefile(struct ufsmount *ump,
2811 	struct fs *fs,
2812 	struct vnode *devvp,
2813 	ino_t ino,
2814 	int mode,
2815 	struct workhead *wkhd)
2816 {
2817 	struct cg *cgp;
2818 	struct buf *bp;
2819 	daddr_t dbn;
2820 	int error;
2821 	u_int cg;
2822 	u_int8_t *inosused;
2823 	struct cdev *dev;
2824 	ino_t cgino;
2825 
2826 	cg = ino_to_cg(fs, ino);
2827 	if (devvp->v_type == VREG) {
2828 		/* devvp is a snapshot */
2829 		MPASS(devvp->v_mount->mnt_data == ump);
2830 		dev = ump->um_devvp->v_rdev;
2831 	} else if (devvp->v_type == VCHR) {
2832 		/* devvp is a normal disk device */
2833 		dev = devvp->v_rdev;
2834 	} else {
2835 		bp = NULL;
2836 		return (0);
2837 	}
2838 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2839 		panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s",
2840 		    devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt);
2841 	if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) {
2842 		if (!ffs_fsfail_cleanup(ump, error) ||
2843 		    !MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR)
2844 			return (error);
2845 		if (devvp->v_type == VREG)
2846 			dbn = fragstoblks(fs, cgtod(fs, cg));
2847 		else
2848 			dbn = fsbtodb(fs, cgtod(fs, cg));
2849 		error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp);
2850 		KASSERT(error == 0, ("getblkx failed"));
2851 		softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd);
2852 		bp->b_flags |= B_RELBUF | B_NOCACHE;
2853 		bp->b_flags &= ~B_CACHE;
2854 		bawrite(bp);
2855 		return (error);
2856 	}
2857 	inosused = cg_inosused(cgp);
2858 	cgino = ino % fs->fs_ipg;
2859 	if (isclr(inosused, cgino)) {
2860 		printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev),
2861 		    (uintmax_t)ino, fs->fs_fsmnt);
2862 		if (fs->fs_ronly == 0)
2863 			panic("ffs_freefile: freeing free inode");
2864 	}
2865 	clrbit(inosused, cgino);
2866 	if (cgino < cgp->cg_irotor)
2867 		cgp->cg_irotor = cgino;
2868 	cgp->cg_cs.cs_nifree++;
2869 	UFS_LOCK(ump);
2870 	fs->fs_cstotal.cs_nifree++;
2871 	fs->fs_cs(fs, cg).cs_nifree++;
2872 	if ((mode & IFMT) == IFDIR) {
2873 		cgp->cg_cs.cs_ndir--;
2874 		fs->fs_cstotal.cs_ndir--;
2875 		fs->fs_cs(fs, cg).cs_ndir--;
2876 	}
2877 	fs->fs_fmod = 1;
2878 	ACTIVECLEAR(fs, cg);
2879 	UFS_UNLOCK(ump);
2880 	if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR)
2881 		softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd);
2882 	bdwrite(bp);
2883 	return (0);
2884 }
2885 
2886 /*
2887  * Check to see if a file is free.
2888  * Used to check for allocated files in snapshots.
2889  */
2890 int
2891 ffs_checkfreefile(struct fs *fs,
2892 	struct vnode *devvp,
2893 	ino_t ino)
2894 {
2895 	struct cg *cgp;
2896 	struct buf *bp;
2897 	int ret, error;
2898 	u_int cg;
2899 	u_int8_t *inosused;
2900 
2901 	cg = ino_to_cg(fs, ino);
2902 	if ((devvp->v_type != VREG) && (devvp->v_type != VCHR))
2903 		return (1);
2904 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2905 		return (1);
2906 	if ((error = ffs_getcg(fs, devvp, cg, 0, &bp, &cgp)) != 0)
2907 		return (1);
2908 	inosused = cg_inosused(cgp);
2909 	ino %= fs->fs_ipg;
2910 	ret = isclr(inosused, ino);
2911 	brelse(bp);
2912 	return (ret);
2913 }
2914 
2915 /*
2916  * Find a block of the specified size in the specified cylinder group.
2917  *
2918  * It is a panic if a request is made to find a block if none are
2919  * available.
2920  */
2921 static ufs1_daddr_t
2922 ffs_mapsearch(struct fs *fs,
2923 	struct cg *cgp,
2924 	ufs2_daddr_t bpref,
2925 	int allocsiz)
2926 {
2927 	ufs1_daddr_t bno;
2928 	int start, len, loc, i;
2929 	int blk, field, subfield, pos;
2930 	u_int8_t *blksfree;
2931 
2932 	/*
2933 	 * find the fragment by searching through the free block
2934 	 * map for an appropriate bit pattern
2935 	 */
2936 	if (bpref)
2937 		start = dtogd(fs, bpref) / NBBY;
2938 	else
2939 		start = cgp->cg_frotor / NBBY;
2940 	blksfree = cg_blksfree(cgp);
2941 	len = howmany(fs->fs_fpg, NBBY) - start;
2942 	loc = scanc((u_int)len, (u_char *)&blksfree[start],
2943 		fragtbl[fs->fs_frag],
2944 		(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2945 	if (loc == 0) {
2946 		len = start + 1;
2947 		start = 0;
2948 		loc = scanc((u_int)len, (u_char *)&blksfree[0],
2949 			fragtbl[fs->fs_frag],
2950 			(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2951 		if (loc == 0) {
2952 			printf("start = %d, len = %d, fs = %s\n",
2953 			    start, len, fs->fs_fsmnt);
2954 			panic("ffs_alloccg: map corrupted");
2955 			/* NOTREACHED */
2956 		}
2957 	}
2958 	bno = (start + len - loc) * NBBY;
2959 	cgp->cg_frotor = bno;
2960 	/*
2961 	 * found the byte in the map
2962 	 * sift through the bits to find the selected frag
2963 	 */
2964 	for (i = bno + NBBY; bno < i; bno += fs->fs_frag) {
2965 		blk = blkmap(fs, blksfree, bno);
2966 		blk <<= 1;
2967 		field = around[allocsiz];
2968 		subfield = inside[allocsiz];
2969 		for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) {
2970 			if ((blk & field) == subfield)
2971 				return (bno + pos);
2972 			field <<= 1;
2973 			subfield <<= 1;
2974 		}
2975 	}
2976 	printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt);
2977 	panic("ffs_alloccg: block not in map");
2978 	return (-1);
2979 }
2980 
2981 static const struct statfs *
2982 ffs_getmntstat(struct vnode *devvp)
2983 {
2984 
2985 	if (devvp->v_type == VCHR)
2986 		return (&devvp->v_rdev->si_mountpt->mnt_stat);
2987 	return (ffs_getmntstat(VFSTOUFS(devvp->v_mount)->um_devvp));
2988 }
2989 
2990 /*
2991  * Fetch and verify a cylinder group.
2992  */
2993 int
2994 ffs_getcg(struct fs *fs,
2995 	struct vnode *devvp,
2996 	u_int cg,
2997 	int flags,
2998 	struct buf **bpp,
2999 	struct cg **cgpp)
3000 {
3001 	struct buf *bp;
3002 	struct cg *cgp;
3003 	const struct statfs *sfs;
3004 	daddr_t blkno;
3005 	int error;
3006 
3007 	*bpp = NULL;
3008 	*cgpp = NULL;
3009 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
3010 		flags |= GB_CKHASH;
3011 	if (devvp->v_type == VREG)
3012 		blkno = fragstoblks(fs, cgtod(fs, cg));
3013 	else
3014 		blkno = fsbtodb(fs, cgtod(fs, cg));
3015 	error = breadn_flags(devvp, blkno, blkno, (int)fs->fs_cgsize, NULL,
3016 	    NULL, 0, NOCRED, flags, ffs_ckhash_cg, &bp);
3017 	if (error != 0)
3018 		return (error);
3019 	cgp = (struct cg *)bp->b_data;
3020 	if ((fs->fs_metackhash & CK_CYLGRP) != 0 &&
3021 	    (bp->b_flags & B_CKHASH) != 0 &&
3022 	    cgp->cg_ckhash != bp->b_ckhash) {
3023 		sfs = ffs_getmntstat(devvp);
3024 		printf("UFS %s%s (%s) cylinder checksum failed: cg %u, cgp: "
3025 		    "0x%x != bp: 0x%jx\n",
3026 		    devvp->v_type == VCHR ? "" : "snapshot of ",
3027 		    sfs->f_mntfromname, sfs->f_mntonname,
3028 		    cg, cgp->cg_ckhash, (uintmax_t)bp->b_ckhash);
3029 		bp->b_flags &= ~B_CKHASH;
3030 		bp->b_flags |= B_INVAL | B_NOCACHE;
3031 		brelse(bp);
3032 		return (EIO);
3033 	}
3034 	if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) {
3035 		sfs = ffs_getmntstat(devvp);
3036 		printf("UFS %s%s (%s)",
3037 		    devvp->v_type == VCHR ? "" : "snapshot of ",
3038 		    sfs->f_mntfromname, sfs->f_mntonname);
3039 		if (!cg_chkmagic(cgp))
3040 			printf(" cg %u: bad magic number 0x%x should be 0x%x\n",
3041 			    cg, cgp->cg_magic, CG_MAGIC);
3042 		else
3043 			printf(": wrong cylinder group cg %u != cgx %u\n", cg,
3044 			    cgp->cg_cgx);
3045 		bp->b_flags &= ~B_CKHASH;
3046 		bp->b_flags |= B_INVAL | B_NOCACHE;
3047 		brelse(bp);
3048 		return (EIO);
3049 	}
3050 	bp->b_flags &= ~B_CKHASH;
3051 	bp->b_xflags |= BX_BKGRDWRITE;
3052 	/*
3053 	 * If we are using check hashes on the cylinder group then we want
3054 	 * to limit changing the cylinder group time to when we are actually
3055 	 * going to write it to disk so that its check hash remains correct
3056 	 * in memory. If the CK_CYLGRP flag is set the time is updated in
3057 	 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we
3058 	 * update the time here as we have done historically.
3059 	 */
3060 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
3061 		bp->b_xflags |= BX_CYLGRP;
3062 	else
3063 		cgp->cg_old_time = cgp->cg_time = time_second;
3064 	*bpp = bp;
3065 	*cgpp = cgp;
3066 	return (0);
3067 }
3068 
3069 static void
3070 ffs_ckhash_cg(struct buf *bp)
3071 {
3072 	uint32_t ckhash;
3073 	struct cg *cgp;
3074 
3075 	cgp = (struct cg *)bp->b_data;
3076 	ckhash = cgp->cg_ckhash;
3077 	cgp->cg_ckhash = 0;
3078 	bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount);
3079 	cgp->cg_ckhash = ckhash;
3080 }
3081 
3082 /*
3083  * Fserr prints the name of a filesystem with an error diagnostic.
3084  *
3085  * The form of the error message is:
3086  *	fs: error message
3087  */
3088 void
3089 ffs_fserr(struct fs *fs,
3090 	ino_t inum,
3091 	char *cp)
3092 {
3093 	struct thread *td = curthread;	/* XXX */
3094 	struct proc *p = td->td_proc;
3095 
3096 	log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n",
3097 	    p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum,
3098 	    fs->fs_fsmnt, cp);
3099 }
3100 
3101 /*
3102  * This function provides the capability for the fsck program to
3103  * update an active filesystem. Fourteen operations are provided:
3104  *
3105  * adjrefcnt(inode, amt) - adjusts the reference count on the
3106  *	specified inode by the specified amount. Under normal
3107  *	operation the count should always go down. Decrementing
3108  *	the count to zero will cause the inode to be freed.
3109  * adjblkcnt(inode, amt) - adjust the number of blocks used by the
3110  *	inode by the specified amount.
3111  * adjdepth(inode, amt) - adjust the depth of the specified directory
3112  *	inode by the specified amount.
3113  * setsize(inode, size) - set the size of the inode to the
3114  *	specified size.
3115  * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) -
3116  *	adjust the superblock summary.
3117  * freedirs(inode, count) - directory inodes [inode..inode + count - 1]
3118  *	are marked as free. Inodes should never have to be marked
3119  *	as in use.
3120  * freefiles(inode, count) - file inodes [inode..inode + count - 1]
3121  *	are marked as free. Inodes should never have to be marked
3122  *	as in use.
3123  * freeblks(blockno, size) - blocks [blockno..blockno + size - 1]
3124  *	are marked as free. Blocks should never have to be marked
3125  *	as in use.
3126  * setflags(flags, set/clear) - the fs_flags field has the specified
3127  *	flags set (second parameter +1) or cleared (second parameter -1).
3128  * setcwd(dirinode) - set the current directory to dirinode in the
3129  *	filesystem associated with the snapshot.
3130  * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".."
3131  *	in the current directory is oldvalue then change it to newvalue.
3132  * unlink(nameptr, oldvalue) - Verify that the inode number associated
3133  *	with nameptr in the current directory is oldvalue then unlink it.
3134  */
3135 
3136 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS);
3137 
3138 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt,
3139     CTLFLAG_WR | CTLTYPE_STRUCT | CTLFLAG_NEEDGIANT,
3140     0, 0, sysctl_ffs_fsck, "S,fsck",
3141     "Adjust Inode Reference Count");
3142 
3143 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt,
3144     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3145     "Adjust Inode Used Blocks Count");
3146 
3147 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_DEPTH, adjdepth,
3148     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3149     "Adjust Directory Inode Depth");
3150 
3151 static SYSCTL_NODE(_vfs_ffs, FFS_SET_SIZE, setsize,
3152     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3153     "Set the inode size");
3154 
3155 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir,
3156     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3157     "Adjust number of directories");
3158 
3159 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree,
3160     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3161     "Adjust number of free blocks");
3162 
3163 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree,
3164     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3165     "Adjust number of free inodes");
3166 
3167 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree,
3168     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3169     "Adjust number of free frags");
3170 
3171 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters,
3172     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3173     "Adjust number of free clusters");
3174 
3175 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs,
3176     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3177     "Free Range of Directory Inodes");
3178 
3179 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles,
3180     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3181     "Free Range of File Inodes");
3182 
3183 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks,
3184     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3185     "Free Range of Blocks");
3186 
3187 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags,
3188     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3189     "Change Filesystem Flags");
3190 
3191 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd,
3192     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3193     "Set Current Working Directory");
3194 
3195 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot,
3196     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3197     "Change Value of .. Entry");
3198 
3199 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink,
3200     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3201     "Unlink a Duplicate Name");
3202 
3203 #ifdef DIAGNOSTIC
3204 static int fsckcmds = 0;
3205 SYSCTL_INT(_debug, OID_AUTO, ffs_fsckcmds, CTLFLAG_RW, &fsckcmds, 0,
3206 	"print out fsck_ffs-based filesystem update commands");
3207 #endif /* DIAGNOSTIC */
3208 
3209 static int
3210 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS)
3211 {
3212 	struct thread *td = curthread;
3213 	struct fsck_cmd cmd;
3214 	struct ufsmount *ump;
3215 	struct vnode *vp, *dvp, *fdvp;
3216 	struct inode *ip, *dp;
3217 	struct mount *mp;
3218 	struct fs *fs;
3219 	struct pwd *pwd;
3220 	ufs2_daddr_t blkno;
3221 	long blkcnt, blksize;
3222 	u_long key;
3223 	struct file *fp;
3224 	cap_rights_t rights;
3225 	int filetype, error;
3226 
3227 	if (req->newptr == NULL || req->newlen > sizeof(cmd))
3228 		return (EBADRPC);
3229 	if ((error = SYSCTL_IN(req, &cmd, sizeof(cmd))) != 0)
3230 		return (error);
3231 	if (cmd.version != FFS_CMD_VERSION)
3232 		return (ERPCMISMATCH);
3233 	if ((error = getvnode(td, cmd.handle,
3234 	    cap_rights_init_one(&rights, CAP_FSCK), &fp)) != 0)
3235 		return (error);
3236 	vp = fp->f_vnode;
3237 	if (vp->v_type != VREG && vp->v_type != VDIR) {
3238 		fdrop(fp, td);
3239 		return (EINVAL);
3240 	}
3241 	vn_start_write(vp, &mp, V_WAIT);
3242 	if (mp == NULL ||
3243 	    strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) {
3244 		vn_finished_write(mp);
3245 		fdrop(fp, td);
3246 		return (EINVAL);
3247 	}
3248 	ump = VFSTOUFS(mp);
3249 	if (mp->mnt_flag & MNT_RDONLY) {
3250 		vn_finished_write(mp);
3251 		fdrop(fp, td);
3252 		return (EROFS);
3253 	}
3254 	fs = ump->um_fs;
3255 	filetype = IFREG;
3256 
3257 	switch (oidp->oid_number) {
3258 	case FFS_SET_FLAGS:
3259 #ifdef DIAGNOSTIC
3260 		if (fsckcmds)
3261 			printf("%s: %s flags\n", mp->mnt_stat.f_mntonname,
3262 			    cmd.size > 0 ? "set" : "clear");
3263 #endif /* DIAGNOSTIC */
3264 		if (cmd.size > 0)
3265 			fs->fs_flags |= (long)cmd.value;
3266 		else
3267 			fs->fs_flags &= ~(long)cmd.value;
3268 		break;
3269 
3270 	case FFS_ADJ_REFCNT:
3271 #ifdef DIAGNOSTIC
3272 		if (fsckcmds) {
3273 			printf("%s: adjust inode %jd link count by %jd\n",
3274 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3275 			    (intmax_t)cmd.size);
3276 		}
3277 #endif /* DIAGNOSTIC */
3278 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3279 			break;
3280 		ip = VTOI(vp);
3281 		ip->i_nlink += cmd.size;
3282 		DIP_SET(ip, i_nlink, ip->i_nlink);
3283 		ip->i_effnlink += cmd.size;
3284 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED);
3285 		error = ffs_update(vp, 1);
3286 		if (DOINGSOFTDEP(vp))
3287 			softdep_change_linkcnt(ip);
3288 		vput(vp);
3289 		break;
3290 
3291 	case FFS_ADJ_BLKCNT:
3292 #ifdef DIAGNOSTIC
3293 		if (fsckcmds) {
3294 			printf("%s: adjust inode %jd block count by %jd\n",
3295 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3296 			    (intmax_t)cmd.size);
3297 		}
3298 #endif /* DIAGNOSTIC */
3299 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3300 			break;
3301 		ip = VTOI(vp);
3302 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size);
3303 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED);
3304 		error = ffs_update(vp, 1);
3305 		vput(vp);
3306 		break;
3307 
3308 	case FFS_ADJ_DEPTH:
3309 #ifdef DIAGNOSTIC
3310 		if (fsckcmds) {
3311 			printf("%s: adjust directory inode %jd depth by %jd\n",
3312 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3313 			    (intmax_t)cmd.size);
3314 		}
3315 #endif /* DIAGNOSTIC */
3316 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3317 			break;
3318 		if (vp->v_type != VDIR) {
3319 			vput(vp);
3320 			error = ENOTDIR;
3321 			break;
3322 		}
3323 		ip = VTOI(vp);
3324 		DIP_SET(ip, i_dirdepth, DIP(ip, i_dirdepth) + cmd.size);
3325 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED);
3326 		error = ffs_update(vp, 1);
3327 		vput(vp);
3328 		break;
3329 
3330 	case FFS_SET_SIZE:
3331 #ifdef DIAGNOSTIC
3332 		if (fsckcmds) {
3333 			printf("%s: set inode %jd size to %jd\n",
3334 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3335 			    (intmax_t)cmd.size);
3336 		}
3337 #endif /* DIAGNOSTIC */
3338 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3339 			break;
3340 		ip = VTOI(vp);
3341 		DIP_SET(ip, i_size, cmd.size);
3342 		UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE | IN_MODIFIED);
3343 		error = ffs_update(vp, 1);
3344 		vput(vp);
3345 		break;
3346 
3347 	case FFS_DIR_FREE:
3348 		filetype = IFDIR;
3349 		/* fall through */
3350 
3351 	case FFS_FILE_FREE:
3352 #ifdef DIAGNOSTIC
3353 		if (fsckcmds) {
3354 			if (cmd.size == 1)
3355 				printf("%s: free %s inode %ju\n",
3356 				    mp->mnt_stat.f_mntonname,
3357 				    filetype == IFDIR ? "directory" : "file",
3358 				    (uintmax_t)cmd.value);
3359 			else
3360 				printf("%s: free %s inodes %ju-%ju\n",
3361 				    mp->mnt_stat.f_mntonname,
3362 				    filetype == IFDIR ? "directory" : "file",
3363 				    (uintmax_t)cmd.value,
3364 				    (uintmax_t)(cmd.value + cmd.size - 1));
3365 		}
3366 #endif /* DIAGNOSTIC */
3367 		while (cmd.size > 0) {
3368 			if ((error = ffs_freefile(ump, fs, ump->um_devvp,
3369 			    cmd.value, filetype, NULL)))
3370 				break;
3371 			cmd.size -= 1;
3372 			cmd.value += 1;
3373 		}
3374 		break;
3375 
3376 	case FFS_BLK_FREE:
3377 #ifdef DIAGNOSTIC
3378 		if (fsckcmds) {
3379 			if (cmd.size == 1)
3380 				printf("%s: free block %jd\n",
3381 				    mp->mnt_stat.f_mntonname,
3382 				    (intmax_t)cmd.value);
3383 			else
3384 				printf("%s: free blocks %jd-%jd\n",
3385 				    mp->mnt_stat.f_mntonname,
3386 				    (intmax_t)cmd.value,
3387 				    (intmax_t)cmd.value + cmd.size - 1);
3388 		}
3389 #endif /* DIAGNOSTIC */
3390 		blkno = cmd.value;
3391 		blkcnt = cmd.size;
3392 		blksize = fs->fs_frag - (blkno % fs->fs_frag);
3393 		key = ffs_blkrelease_start(ump, ump->um_devvp, UFS_ROOTINO);
3394 		while (blkcnt > 0) {
3395 			if (blkcnt < blksize)
3396 				blksize = blkcnt;
3397 			ffs_blkfree(ump, fs, ump->um_devvp, blkno,
3398 			    blksize * fs->fs_fsize, UFS_ROOTINO,
3399 			    VDIR, NULL, key);
3400 			blkno += blksize;
3401 			blkcnt -= blksize;
3402 			blksize = fs->fs_frag;
3403 		}
3404 		ffs_blkrelease_finish(ump, key);
3405 		break;
3406 
3407 	/*
3408 	 * Adjust superblock summaries.  fsck(8) is expected to
3409 	 * submit deltas when necessary.
3410 	 */
3411 	case FFS_ADJ_NDIR:
3412 #ifdef DIAGNOSTIC
3413 		if (fsckcmds) {
3414 			printf("%s: adjust number of directories by %jd\n",
3415 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3416 		}
3417 #endif /* DIAGNOSTIC */
3418 		fs->fs_cstotal.cs_ndir += cmd.value;
3419 		break;
3420 
3421 	case FFS_ADJ_NBFREE:
3422 #ifdef DIAGNOSTIC
3423 		if (fsckcmds) {
3424 			printf("%s: adjust number of free blocks by %+jd\n",
3425 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3426 		}
3427 #endif /* DIAGNOSTIC */
3428 		fs->fs_cstotal.cs_nbfree += cmd.value;
3429 		break;
3430 
3431 	case FFS_ADJ_NIFREE:
3432 #ifdef DIAGNOSTIC
3433 		if (fsckcmds) {
3434 			printf("%s: adjust number of free inodes by %+jd\n",
3435 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3436 		}
3437 #endif /* DIAGNOSTIC */
3438 		fs->fs_cstotal.cs_nifree += cmd.value;
3439 		break;
3440 
3441 	case FFS_ADJ_NFFREE:
3442 #ifdef DIAGNOSTIC
3443 		if (fsckcmds) {
3444 			printf("%s: adjust number of free frags by %+jd\n",
3445 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3446 		}
3447 #endif /* DIAGNOSTIC */
3448 		fs->fs_cstotal.cs_nffree += cmd.value;
3449 		break;
3450 
3451 	case FFS_ADJ_NUMCLUSTERS:
3452 #ifdef DIAGNOSTIC
3453 		if (fsckcmds) {
3454 			printf("%s: adjust number of free clusters by %+jd\n",
3455 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3456 		}
3457 #endif /* DIAGNOSTIC */
3458 		fs->fs_cstotal.cs_numclusters += cmd.value;
3459 		break;
3460 
3461 	case FFS_SET_CWD:
3462 #ifdef DIAGNOSTIC
3463 		if (fsckcmds) {
3464 			printf("%s: set current directory to inode %jd\n",
3465 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3466 		}
3467 #endif /* DIAGNOSTIC */
3468 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp)))
3469 			break;
3470 		AUDIT_ARG_VNODE1(vp);
3471 		if ((error = change_dir(vp, td)) != 0) {
3472 			vput(vp);
3473 			break;
3474 		}
3475 		VOP_UNLOCK(vp);
3476 		pwd_chdir(td, vp);
3477 		break;
3478 
3479 	case FFS_SET_DOTDOT:
3480 #ifdef DIAGNOSTIC
3481 		if (fsckcmds) {
3482 			printf("%s: change .. in cwd from %jd to %jd\n",
3483 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3484 			    (intmax_t)cmd.size);
3485 		}
3486 #endif /* DIAGNOSTIC */
3487 		/*
3488 		 * First we have to get and lock the parent directory
3489 		 * to which ".." points.
3490 		 */
3491 		error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp);
3492 		if (error)
3493 			break;
3494 		/*
3495 		 * Now we get and lock the child directory containing "..".
3496 		 */
3497 		pwd = pwd_hold(td);
3498 		dvp = pwd->pwd_cdir;
3499 		if ((error = vget(dvp, LK_EXCLUSIVE)) != 0) {
3500 			vput(fdvp);
3501 			pwd_drop(pwd);
3502 			break;
3503 		}
3504 		dp = VTOI(dvp);
3505 		SET_I_OFFSET(dp, 12);	/* XXX mastertemplate.dot_reclen */
3506 		error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size,
3507 		    DT_DIR, 0);
3508 		cache_purge(fdvp);
3509 		cache_purge(dvp);
3510 		vput(dvp);
3511 		vput(fdvp);
3512 		pwd_drop(pwd);
3513 		break;
3514 
3515 	case FFS_UNLINK:
3516 #ifdef DIAGNOSTIC
3517 		if (fsckcmds) {
3518 			char buf[32];
3519 
3520 			if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL))
3521 				strncpy(buf, "Name_too_long", 32);
3522 			printf("%s: unlink %s (inode %jd)\n",
3523 			    mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size);
3524 		}
3525 #endif /* DIAGNOSTIC */
3526 		/*
3527 		 * kern_funlinkat will do its own start/finish writes and
3528 		 * they do not nest, so drop ours here. Setting mp == NULL
3529 		 * indicates that vn_finished_write is not needed down below.
3530 		 */
3531 		vn_finished_write(mp);
3532 		mp = NULL;
3533 		error = kern_funlinkat(td, AT_FDCWD,
3534 		    (char *)(intptr_t)cmd.value, FD_NONE, UIO_USERSPACE,
3535 		    0, (ino_t)cmd.size);
3536 		break;
3537 
3538 	default:
3539 #ifdef DIAGNOSTIC
3540 		if (fsckcmds) {
3541 			printf("Invalid request %d from fsck\n",
3542 			    oidp->oid_number);
3543 		}
3544 #endif /* DIAGNOSTIC */
3545 		error = EINVAL;
3546 		break;
3547 	}
3548 	fdrop(fp, td);
3549 	vn_finished_write(mp);
3550 	return (error);
3551 }
3552