xref: /freebsd/sys/ufs/ffs/ffs_alloc.c (revision d1d015864103b253b3fcb2f72a0da5b0cfeb31b6)
1 /*-
2  * Copyright (c) 2002 Networks Associates Technology, Inc.
3  * All rights reserved.
4  *
5  * This software was developed for the FreeBSD Project by Marshall
6  * Kirk McKusick and Network Associates Laboratories, the Security
7  * Research Division of Network Associates, Inc. under DARPA/SPAWAR
8  * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS
9  * research program
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  * Copyright (c) 1982, 1986, 1989, 1993
33  *	The Regents of the University of California.  All rights reserved.
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 4. Neither the name of the University nor the names of its contributors
44  *    may be used to endorse or promote products derived from this software
45  *    without specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  *	@(#)ffs_alloc.c	8.18 (Berkeley) 5/26/95
60  */
61 
62 #include <sys/cdefs.h>
63 __FBSDID("$FreeBSD$");
64 
65 #include "opt_quota.h"
66 
67 #include <sys/param.h>
68 #include <sys/capability.h>
69 #include <sys/systm.h>
70 #include <sys/bio.h>
71 #include <sys/buf.h>
72 #include <sys/conf.h>
73 #include <sys/fcntl.h>
74 #include <sys/file.h>
75 #include <sys/filedesc.h>
76 #include <sys/priv.h>
77 #include <sys/proc.h>
78 #include <sys/vnode.h>
79 #include <sys/mount.h>
80 #include <sys/kernel.h>
81 #include <sys/syscallsubr.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/taskqueue.h>
85 
86 #include <security/audit/audit.h>
87 
88 #include <geom/geom.h>
89 
90 #include <ufs/ufs/dir.h>
91 #include <ufs/ufs/extattr.h>
92 #include <ufs/ufs/quota.h>
93 #include <ufs/ufs/inode.h>
94 #include <ufs/ufs/ufs_extern.h>
95 #include <ufs/ufs/ufsmount.h>
96 
97 #include <ufs/ffs/fs.h>
98 #include <ufs/ffs/ffs_extern.h>
99 #include <ufs/ffs/softdep.h>
100 
101 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref,
102 				  int size, int rsize);
103 
104 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int);
105 static ufs2_daddr_t
106 	      ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int);
107 static void	ffs_blkfree_cg(struct ufsmount *, struct fs *,
108 		    struct vnode *, ufs2_daddr_t, long, ino_t,
109 		    struct workhead *);
110 static void	ffs_blkfree_trim_completed(struct bio *);
111 static void	ffs_blkfree_trim_task(void *ctx, int pending __unused);
112 #ifdef INVARIANTS
113 static int	ffs_checkblk(struct inode *, ufs2_daddr_t, long);
114 #endif
115 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int,
116 		    int);
117 static ino_t	ffs_dirpref(struct inode *);
118 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t,
119 		    int, int);
120 static ufs2_daddr_t	ffs_hashalloc
121 		(struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *);
122 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int,
123 		    int);
124 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int);
125 static int	ffs_reallocblks_ufs1(struct vop_reallocblks_args *);
126 static int	ffs_reallocblks_ufs2(struct vop_reallocblks_args *);
127 
128 /*
129  * Allocate a block in the filesystem.
130  *
131  * The size of the requested block is given, which must be some
132  * multiple of fs_fsize and <= fs_bsize.
133  * A preference may be optionally specified. If a preference is given
134  * the following hierarchy is used to allocate a block:
135  *   1) allocate the requested block.
136  *   2) allocate a rotationally optimal block in the same cylinder.
137  *   3) allocate a block in the same cylinder group.
138  *   4) quadradically rehash into other cylinder groups, until an
139  *      available block is located.
140  * If no block preference is given the following hierarchy is used
141  * to allocate a block:
142  *   1) allocate a block in the cylinder group that contains the
143  *      inode for the file.
144  *   2) quadradically rehash into other cylinder groups, until an
145  *      available block is located.
146  */
147 int
148 ffs_alloc(ip, lbn, bpref, size, flags, cred, bnp)
149 	struct inode *ip;
150 	ufs2_daddr_t lbn, bpref;
151 	int size, flags;
152 	struct ucred *cred;
153 	ufs2_daddr_t *bnp;
154 {
155 	struct fs *fs;
156 	struct ufsmount *ump;
157 	ufs2_daddr_t bno;
158 	u_int cg, reclaimed;
159 	static struct timeval lastfail;
160 	static int curfail;
161 	int64_t delta;
162 #ifdef QUOTA
163 	int error;
164 #endif
165 
166 	*bnp = 0;
167 	fs = ip->i_fs;
168 	ump = ip->i_ump;
169 	mtx_assert(UFS_MTX(ump), MA_OWNED);
170 #ifdef INVARIANTS
171 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
172 		printf("dev = %s, bsize = %ld, size = %d, fs = %s\n",
173 		    devtoname(ip->i_dev), (long)fs->fs_bsize, size,
174 		    fs->fs_fsmnt);
175 		panic("ffs_alloc: bad size");
176 	}
177 	if (cred == NOCRED)
178 		panic("ffs_alloc: missing credential");
179 #endif /* INVARIANTS */
180 	reclaimed = 0;
181 retry:
182 #ifdef QUOTA
183 	UFS_UNLOCK(ump);
184 	error = chkdq(ip, btodb(size), cred, 0);
185 	if (error)
186 		return (error);
187 	UFS_LOCK(ump);
188 #endif
189 	if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0)
190 		goto nospace;
191 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0) &&
192 	    freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0)
193 		goto nospace;
194 	if (bpref >= fs->fs_size)
195 		bpref = 0;
196 	if (bpref == 0)
197 		cg = ino_to_cg(fs, ip->i_number);
198 	else
199 		cg = dtog(fs, bpref);
200 	bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg);
201 	if (bno > 0) {
202 		delta = btodb(size);
203 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
204 		if (flags & IO_EXT)
205 			ip->i_flag |= IN_CHANGE;
206 		else
207 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
208 		*bnp = bno;
209 		return (0);
210 	}
211 nospace:
212 #ifdef QUOTA
213 	UFS_UNLOCK(ump);
214 	/*
215 	 * Restore user's disk quota because allocation failed.
216 	 */
217 	(void) chkdq(ip, -btodb(size), cred, FORCE);
218 	UFS_LOCK(ump);
219 #endif
220 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
221 		reclaimed = 1;
222 		softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT);
223 		goto retry;
224 	}
225 	UFS_UNLOCK(ump);
226 	if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
227 		ffs_fserr(fs, ip->i_number, "filesystem full");
228 		uprintf("\n%s: write failed, filesystem is full\n",
229 		    fs->fs_fsmnt);
230 	}
231 	return (ENOSPC);
232 }
233 
234 /*
235  * Reallocate a fragment to a bigger size
236  *
237  * The number and size of the old block is given, and a preference
238  * and new size is also specified. The allocator attempts to extend
239  * the original block. Failing that, the regular block allocator is
240  * invoked to get an appropriate block.
241  */
242 int
243 ffs_realloccg(ip, lbprev, bprev, bpref, osize, nsize, flags, cred, bpp)
244 	struct inode *ip;
245 	ufs2_daddr_t lbprev;
246 	ufs2_daddr_t bprev;
247 	ufs2_daddr_t bpref;
248 	int osize, nsize, flags;
249 	struct ucred *cred;
250 	struct buf **bpp;
251 {
252 	struct vnode *vp;
253 	struct fs *fs;
254 	struct buf *bp;
255 	struct ufsmount *ump;
256 	u_int cg, request, reclaimed;
257 	int error, gbflags;
258 	ufs2_daddr_t bno;
259 	static struct timeval lastfail;
260 	static int curfail;
261 	int64_t delta;
262 
263 	*bpp = 0;
264 	vp = ITOV(ip);
265 	fs = ip->i_fs;
266 	bp = NULL;
267 	ump = ip->i_ump;
268 	gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
269 
270 	mtx_assert(UFS_MTX(ump), MA_OWNED);
271 #ifdef INVARIANTS
272 	if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
273 		panic("ffs_realloccg: allocation on suspended filesystem");
274 	if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 ||
275 	    (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) {
276 		printf(
277 		"dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n",
278 		    devtoname(ip->i_dev), (long)fs->fs_bsize, osize,
279 		    nsize, fs->fs_fsmnt);
280 		panic("ffs_realloccg: bad size");
281 	}
282 	if (cred == NOCRED)
283 		panic("ffs_realloccg: missing credential");
284 #endif /* INVARIANTS */
285 	reclaimed = 0;
286 retry:
287 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0) &&
288 	    freespace(fs, fs->fs_minfree) -  numfrags(fs, nsize - osize) < 0) {
289 		goto nospace;
290 	}
291 	if (bprev == 0) {
292 		printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n",
293 		    devtoname(ip->i_dev), (long)fs->fs_bsize, (intmax_t)bprev,
294 		    fs->fs_fsmnt);
295 		panic("ffs_realloccg: bad bprev");
296 	}
297 	UFS_UNLOCK(ump);
298 	/*
299 	 * Allocate the extra space in the buffer.
300 	 */
301 	error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp);
302 	if (error) {
303 		brelse(bp);
304 		return (error);
305 	}
306 
307 	if (bp->b_blkno == bp->b_lblkno) {
308 		if (lbprev >= NDADDR)
309 			panic("ffs_realloccg: lbprev out of range");
310 		bp->b_blkno = fsbtodb(fs, bprev);
311 	}
312 
313 #ifdef QUOTA
314 	error = chkdq(ip, btodb(nsize - osize), cred, 0);
315 	if (error) {
316 		brelse(bp);
317 		return (error);
318 	}
319 #endif
320 	/*
321 	 * Check for extension in the existing location.
322 	 */
323 	cg = dtog(fs, bprev);
324 	UFS_LOCK(ump);
325 	bno = ffs_fragextend(ip, cg, bprev, osize, nsize);
326 	if (bno) {
327 		if (bp->b_blkno != fsbtodb(fs, bno))
328 			panic("ffs_realloccg: bad blockno");
329 		delta = btodb(nsize - osize);
330 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
331 		if (flags & IO_EXT)
332 			ip->i_flag |= IN_CHANGE;
333 		else
334 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
335 		allocbuf(bp, nsize);
336 		bp->b_flags |= B_DONE;
337 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
338 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
339 			vfs_bio_set_valid(bp, osize, nsize - osize);
340 		*bpp = bp;
341 		return (0);
342 	}
343 	/*
344 	 * Allocate a new disk location.
345 	 */
346 	if (bpref >= fs->fs_size)
347 		bpref = 0;
348 	switch ((int)fs->fs_optim) {
349 	case FS_OPTSPACE:
350 		/*
351 		 * Allocate an exact sized fragment. Although this makes
352 		 * best use of space, we will waste time relocating it if
353 		 * the file continues to grow. If the fragmentation is
354 		 * less than half of the minimum free reserve, we choose
355 		 * to begin optimizing for time.
356 		 */
357 		request = nsize;
358 		if (fs->fs_minfree <= 5 ||
359 		    fs->fs_cstotal.cs_nffree >
360 		    (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100))
361 			break;
362 		log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n",
363 			fs->fs_fsmnt);
364 		fs->fs_optim = FS_OPTTIME;
365 		break;
366 	case FS_OPTTIME:
367 		/*
368 		 * At this point we have discovered a file that is trying to
369 		 * grow a small fragment to a larger fragment. To save time,
370 		 * we allocate a full sized block, then free the unused portion.
371 		 * If the file continues to grow, the `ffs_fragextend' call
372 		 * above will be able to grow it in place without further
373 		 * copying. If aberrant programs cause disk fragmentation to
374 		 * grow within 2% of the free reserve, we choose to begin
375 		 * optimizing for space.
376 		 */
377 		request = fs->fs_bsize;
378 		if (fs->fs_cstotal.cs_nffree <
379 		    (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100)
380 			break;
381 		log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n",
382 			fs->fs_fsmnt);
383 		fs->fs_optim = FS_OPTSPACE;
384 		break;
385 	default:
386 		printf("dev = %s, optim = %ld, fs = %s\n",
387 		    devtoname(ip->i_dev), (long)fs->fs_optim, fs->fs_fsmnt);
388 		panic("ffs_realloccg: bad optim");
389 		/* NOTREACHED */
390 	}
391 	bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg);
392 	if (bno > 0) {
393 		bp->b_blkno = fsbtodb(fs, bno);
394 		if (!DOINGSOFTDEP(vp))
395 			ffs_blkfree(ump, fs, ip->i_devvp, bprev, (long)osize,
396 			    ip->i_number, vp->v_type, NULL);
397 		delta = btodb(nsize - osize);
398 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
399 		if (flags & IO_EXT)
400 			ip->i_flag |= IN_CHANGE;
401 		else
402 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
403 		allocbuf(bp, nsize);
404 		bp->b_flags |= B_DONE;
405 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
406 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
407 			vfs_bio_set_valid(bp, osize, nsize - osize);
408 		*bpp = bp;
409 		return (0);
410 	}
411 #ifdef QUOTA
412 	UFS_UNLOCK(ump);
413 	/*
414 	 * Restore user's disk quota because allocation failed.
415 	 */
416 	(void) chkdq(ip, -btodb(nsize - osize), cred, FORCE);
417 	UFS_LOCK(ump);
418 #endif
419 nospace:
420 	/*
421 	 * no space available
422 	 */
423 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
424 		reclaimed = 1;
425 		UFS_UNLOCK(ump);
426 		if (bp) {
427 			brelse(bp);
428 			bp = NULL;
429 		}
430 		UFS_LOCK(ump);
431 		softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT);
432 		goto retry;
433 	}
434 	UFS_UNLOCK(ump);
435 	if (bp)
436 		brelse(bp);
437 	if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
438 		ffs_fserr(fs, ip->i_number, "filesystem full");
439 		uprintf("\n%s: write failed, filesystem is full\n",
440 		    fs->fs_fsmnt);
441 	}
442 	return (ENOSPC);
443 }
444 
445 /*
446  * Reallocate a sequence of blocks into a contiguous sequence of blocks.
447  *
448  * The vnode and an array of buffer pointers for a range of sequential
449  * logical blocks to be made contiguous is given. The allocator attempts
450  * to find a range of sequential blocks starting as close as possible
451  * from the end of the allocation for the logical block immediately
452  * preceding the current range. If successful, the physical block numbers
453  * in the buffer pointers and in the inode are changed to reflect the new
454  * allocation. If unsuccessful, the allocation is left unchanged. The
455  * success in doing the reallocation is returned. Note that the error
456  * return is not reflected back to the user. Rather the previous block
457  * allocation will be used.
458  */
459 
460 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW, 0, "FFS filesystem");
461 
462 static int doasyncfree = 1;
463 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0, "");
464 
465 static int doreallocblks = 1;
466 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0, "");
467 
468 #ifdef DEBUG
469 static volatile int prtrealloc = 0;
470 #endif
471 
472 int
473 ffs_reallocblks(ap)
474 	struct vop_reallocblks_args /* {
475 		struct vnode *a_vp;
476 		struct cluster_save *a_buflist;
477 	} */ *ap;
478 {
479 
480 	if (doreallocblks == 0)
481 		return (ENOSPC);
482 	/*
483 	 * We can't wait in softdep prealloc as it may fsync and recurse
484 	 * here.  Instead we simply fail to reallocate blocks if this
485 	 * rare condition arises.
486 	 */
487 	if (DOINGSOFTDEP(ap->a_vp))
488 		if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0)
489 			return (ENOSPC);
490 	if (VTOI(ap->a_vp)->i_ump->um_fstype == UFS1)
491 		return (ffs_reallocblks_ufs1(ap));
492 	return (ffs_reallocblks_ufs2(ap));
493 }
494 
495 static int
496 ffs_reallocblks_ufs1(ap)
497 	struct vop_reallocblks_args /* {
498 		struct vnode *a_vp;
499 		struct cluster_save *a_buflist;
500 	} */ *ap;
501 {
502 	struct fs *fs;
503 	struct inode *ip;
504 	struct vnode *vp;
505 	struct buf *sbp, *ebp;
506 	ufs1_daddr_t *bap, *sbap, *ebap = 0;
507 	struct cluster_save *buflist;
508 	struct ufsmount *ump;
509 	ufs_lbn_t start_lbn, end_lbn;
510 	ufs1_daddr_t soff, newblk, blkno;
511 	ufs2_daddr_t pref;
512 	struct indir start_ap[NIADDR + 1], end_ap[NIADDR + 1], *idp;
513 	int i, len, start_lvl, end_lvl, ssize;
514 
515 	vp = ap->a_vp;
516 	ip = VTOI(vp);
517 	fs = ip->i_fs;
518 	ump = ip->i_ump;
519 	/*
520 	 * If we are not tracking block clusters or if we have less than 2%
521 	 * free blocks left, then do not attempt to cluster. Running with
522 	 * less than 5% free block reserve is not recommended and those that
523 	 * choose to do so do not expect to have good file layout.
524 	 */
525 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 2) < 0)
526 		return (ENOSPC);
527 	buflist = ap->a_buflist;
528 	len = buflist->bs_nchildren;
529 	start_lbn = buflist->bs_children[0]->b_lblkno;
530 	end_lbn = start_lbn + len - 1;
531 #ifdef INVARIANTS
532 	for (i = 0; i < len; i++)
533 		if (!ffs_checkblk(ip,
534 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
535 			panic("ffs_reallocblks: unallocated block 1");
536 	for (i = 1; i < len; i++)
537 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
538 			panic("ffs_reallocblks: non-logical cluster");
539 	blkno = buflist->bs_children[0]->b_blkno;
540 	ssize = fsbtodb(fs, fs->fs_frag);
541 	for (i = 1; i < len - 1; i++)
542 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
543 			panic("ffs_reallocblks: non-physical cluster %d", i);
544 #endif
545 	/*
546 	 * If the cluster crosses the boundary for the first indirect
547 	 * block, leave space for the indirect block. Indirect blocks
548 	 * are initially laid out in a position after the last direct
549 	 * block. Block reallocation would usually destroy locality by
550 	 * moving the indirect block out of the way to make room for
551 	 * data blocks if we didn't compensate here. We should also do
552 	 * this for other indirect block boundaries, but it is only
553 	 * important for the first one.
554 	 */
555 	if (start_lbn < NDADDR && end_lbn >= NDADDR)
556 		return (ENOSPC);
557 	/*
558 	 * If the latest allocation is in a new cylinder group, assume that
559 	 * the filesystem has decided to move and do not force it back to
560 	 * the previous cylinder group.
561 	 */
562 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
563 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
564 		return (ENOSPC);
565 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
566 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
567 		return (ENOSPC);
568 	/*
569 	 * Get the starting offset and block map for the first block.
570 	 */
571 	if (start_lvl == 0) {
572 		sbap = &ip->i_din1->di_db[0];
573 		soff = start_lbn;
574 	} else {
575 		idp = &start_ap[start_lvl - 1];
576 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
577 			brelse(sbp);
578 			return (ENOSPC);
579 		}
580 		sbap = (ufs1_daddr_t *)sbp->b_data;
581 		soff = idp->in_off;
582 	}
583 	/*
584 	 * If the block range spans two block maps, get the second map.
585 	 */
586 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
587 		ssize = len;
588 	} else {
589 #ifdef INVARIANTS
590 		if (start_lvl > 0 &&
591 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
592 			panic("ffs_reallocblk: start == end");
593 #endif
594 		ssize = len - (idp->in_off + 1);
595 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
596 			goto fail;
597 		ebap = (ufs1_daddr_t *)ebp->b_data;
598 	}
599 	/*
600 	 * Find the preferred location for the cluster.
601 	 */
602 	UFS_LOCK(ump);
603 	pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap);
604 	/*
605 	 * Search the block map looking for an allocation of the desired size.
606 	 */
607 	if ((newblk = ffs_hashalloc(ip, dtog(fs, pref), pref,
608 	    len, len, ffs_clusteralloc)) == 0) {
609 		UFS_UNLOCK(ump);
610 		goto fail;
611 	}
612 	/*
613 	 * We have found a new contiguous block.
614 	 *
615 	 * First we have to replace the old block pointers with the new
616 	 * block pointers in the inode and indirect blocks associated
617 	 * with the file.
618 	 */
619 #ifdef DEBUG
620 	if (prtrealloc)
621 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:",
622 		    (uintmax_t)ip->i_number,
623 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
624 #endif
625 	blkno = newblk;
626 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
627 		if (i == ssize) {
628 			bap = ebap;
629 			soff = -i;
630 		}
631 #ifdef INVARIANTS
632 		if (!ffs_checkblk(ip,
633 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
634 			panic("ffs_reallocblks: unallocated block 2");
635 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
636 			panic("ffs_reallocblks: alloc mismatch");
637 #endif
638 #ifdef DEBUG
639 		if (prtrealloc)
640 			printf(" %d,", *bap);
641 #endif
642 		if (DOINGSOFTDEP(vp)) {
643 			if (sbap == &ip->i_din1->di_db[0] && i < ssize)
644 				softdep_setup_allocdirect(ip, start_lbn + i,
645 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
646 				    buflist->bs_children[i]);
647 			else
648 				softdep_setup_allocindir_page(ip, start_lbn + i,
649 				    i < ssize ? sbp : ebp, soff + i, blkno,
650 				    *bap, buflist->bs_children[i]);
651 		}
652 		*bap++ = blkno;
653 	}
654 	/*
655 	 * Next we must write out the modified inode and indirect blocks.
656 	 * For strict correctness, the writes should be synchronous since
657 	 * the old block values may have been written to disk. In practise
658 	 * they are almost never written, but if we are concerned about
659 	 * strict correctness, the `doasyncfree' flag should be set to zero.
660 	 *
661 	 * The test on `doasyncfree' should be changed to test a flag
662 	 * that shows whether the associated buffers and inodes have
663 	 * been written. The flag should be set when the cluster is
664 	 * started and cleared whenever the buffer or inode is flushed.
665 	 * We can then check below to see if it is set, and do the
666 	 * synchronous write only when it has been cleared.
667 	 */
668 	if (sbap != &ip->i_din1->di_db[0]) {
669 		if (doasyncfree)
670 			bdwrite(sbp);
671 		else
672 			bwrite(sbp);
673 	} else {
674 		ip->i_flag |= IN_CHANGE | IN_UPDATE;
675 		if (!doasyncfree)
676 			ffs_update(vp, 1);
677 	}
678 	if (ssize < len) {
679 		if (doasyncfree)
680 			bdwrite(ebp);
681 		else
682 			bwrite(ebp);
683 	}
684 	/*
685 	 * Last, free the old blocks and assign the new blocks to the buffers.
686 	 */
687 #ifdef DEBUG
688 	if (prtrealloc)
689 		printf("\n\tnew:");
690 #endif
691 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
692 		if (!DOINGSOFTDEP(vp))
693 			ffs_blkfree(ump, fs, ip->i_devvp,
694 			    dbtofsb(fs, buflist->bs_children[i]->b_blkno),
695 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL);
696 		buflist->bs_children[i]->b_blkno = fsbtodb(fs, blkno);
697 #ifdef INVARIANTS
698 		if (!ffs_checkblk(ip,
699 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
700 			panic("ffs_reallocblks: unallocated block 3");
701 #endif
702 #ifdef DEBUG
703 		if (prtrealloc)
704 			printf(" %d,", blkno);
705 #endif
706 	}
707 #ifdef DEBUG
708 	if (prtrealloc) {
709 		prtrealloc--;
710 		printf("\n");
711 	}
712 #endif
713 	return (0);
714 
715 fail:
716 	if (ssize < len)
717 		brelse(ebp);
718 	if (sbap != &ip->i_din1->di_db[0])
719 		brelse(sbp);
720 	return (ENOSPC);
721 }
722 
723 static int
724 ffs_reallocblks_ufs2(ap)
725 	struct vop_reallocblks_args /* {
726 		struct vnode *a_vp;
727 		struct cluster_save *a_buflist;
728 	} */ *ap;
729 {
730 	struct fs *fs;
731 	struct inode *ip;
732 	struct vnode *vp;
733 	struct buf *sbp, *ebp;
734 	ufs2_daddr_t *bap, *sbap, *ebap = 0;
735 	struct cluster_save *buflist;
736 	struct ufsmount *ump;
737 	ufs_lbn_t start_lbn, end_lbn;
738 	ufs2_daddr_t soff, newblk, blkno, pref;
739 	struct indir start_ap[NIADDR + 1], end_ap[NIADDR + 1], *idp;
740 	int i, len, start_lvl, end_lvl, ssize;
741 
742 	vp = ap->a_vp;
743 	ip = VTOI(vp);
744 	fs = ip->i_fs;
745 	ump = ip->i_ump;
746 	/*
747 	 * If we are not tracking block clusters or if we have less than 2%
748 	 * free blocks left, then do not attempt to cluster. Running with
749 	 * less than 5% free block reserve is not recommended and those that
750 	 * choose to do so do not expect to have good file layout.
751 	 */
752 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 2) < 0)
753 		return (ENOSPC);
754 	buflist = ap->a_buflist;
755 	len = buflist->bs_nchildren;
756 	start_lbn = buflist->bs_children[0]->b_lblkno;
757 	end_lbn = start_lbn + len - 1;
758 #ifdef INVARIANTS
759 	for (i = 0; i < len; i++)
760 		if (!ffs_checkblk(ip,
761 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
762 			panic("ffs_reallocblks: unallocated block 1");
763 	for (i = 1; i < len; i++)
764 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
765 			panic("ffs_reallocblks: non-logical cluster");
766 	blkno = buflist->bs_children[0]->b_blkno;
767 	ssize = fsbtodb(fs, fs->fs_frag);
768 	for (i = 1; i < len - 1; i++)
769 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
770 			panic("ffs_reallocblks: non-physical cluster %d", i);
771 #endif
772 	/*
773 	 * If the cluster crosses the boundary for the first indirect
774 	 * block, do not move anything in it. Indirect blocks are
775 	 * usually initially laid out in a position between the data
776 	 * blocks. Block reallocation would usually destroy locality by
777 	 * moving the indirect block out of the way to make room for
778 	 * data blocks if we didn't compensate here. We should also do
779 	 * this for other indirect block boundaries, but it is only
780 	 * important for the first one.
781 	 */
782 	if (start_lbn < NDADDR && end_lbn >= NDADDR)
783 		return (ENOSPC);
784 	/*
785 	 * If the latest allocation is in a new cylinder group, assume that
786 	 * the filesystem has decided to move and do not force it back to
787 	 * the previous cylinder group.
788 	 */
789 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
790 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
791 		return (ENOSPC);
792 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
793 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
794 		return (ENOSPC);
795 	/*
796 	 * Get the starting offset and block map for the first block.
797 	 */
798 	if (start_lvl == 0) {
799 		sbap = &ip->i_din2->di_db[0];
800 		soff = start_lbn;
801 	} else {
802 		idp = &start_ap[start_lvl - 1];
803 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
804 			brelse(sbp);
805 			return (ENOSPC);
806 		}
807 		sbap = (ufs2_daddr_t *)sbp->b_data;
808 		soff = idp->in_off;
809 	}
810 	/*
811 	 * If the block range spans two block maps, get the second map.
812 	 */
813 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
814 		ssize = len;
815 	} else {
816 #ifdef INVARIANTS
817 		if (start_lvl > 0 &&
818 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
819 			panic("ffs_reallocblk: start == end");
820 #endif
821 		ssize = len - (idp->in_off + 1);
822 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
823 			goto fail;
824 		ebap = (ufs2_daddr_t *)ebp->b_data;
825 	}
826 	/*
827 	 * Find the preferred location for the cluster.
828 	 */
829 	UFS_LOCK(ump);
830 	pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap);
831 	/*
832 	 * Search the block map looking for an allocation of the desired size.
833 	 */
834 	if ((newblk = ffs_hashalloc(ip, dtog(fs, pref), pref,
835 	    len, len, ffs_clusteralloc)) == 0) {
836 		UFS_UNLOCK(ump);
837 		goto fail;
838 	}
839 	/*
840 	 * We have found a new contiguous block.
841 	 *
842 	 * First we have to replace the old block pointers with the new
843 	 * block pointers in the inode and indirect blocks associated
844 	 * with the file.
845 	 */
846 #ifdef DEBUG
847 	if (prtrealloc)
848 		printf("realloc: ino %d, lbns %jd-%jd\n\told:", ip->i_number,
849 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
850 #endif
851 	blkno = newblk;
852 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
853 		if (i == ssize) {
854 			bap = ebap;
855 			soff = -i;
856 		}
857 #ifdef INVARIANTS
858 		if (!ffs_checkblk(ip,
859 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
860 			panic("ffs_reallocblks: unallocated block 2");
861 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
862 			panic("ffs_reallocblks: alloc mismatch");
863 #endif
864 #ifdef DEBUG
865 		if (prtrealloc)
866 			printf(" %jd,", (intmax_t)*bap);
867 #endif
868 		if (DOINGSOFTDEP(vp)) {
869 			if (sbap == &ip->i_din2->di_db[0] && i < ssize)
870 				softdep_setup_allocdirect(ip, start_lbn + i,
871 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
872 				    buflist->bs_children[i]);
873 			else
874 				softdep_setup_allocindir_page(ip, start_lbn + i,
875 				    i < ssize ? sbp : ebp, soff + i, blkno,
876 				    *bap, buflist->bs_children[i]);
877 		}
878 		*bap++ = blkno;
879 	}
880 	/*
881 	 * Next we must write out the modified inode and indirect blocks.
882 	 * For strict correctness, the writes should be synchronous since
883 	 * the old block values may have been written to disk. In practise
884 	 * they are almost never written, but if we are concerned about
885 	 * strict correctness, the `doasyncfree' flag should be set to zero.
886 	 *
887 	 * The test on `doasyncfree' should be changed to test a flag
888 	 * that shows whether the associated buffers and inodes have
889 	 * been written. The flag should be set when the cluster is
890 	 * started and cleared whenever the buffer or inode is flushed.
891 	 * We can then check below to see if it is set, and do the
892 	 * synchronous write only when it has been cleared.
893 	 */
894 	if (sbap != &ip->i_din2->di_db[0]) {
895 		if (doasyncfree)
896 			bdwrite(sbp);
897 		else
898 			bwrite(sbp);
899 	} else {
900 		ip->i_flag |= IN_CHANGE | IN_UPDATE;
901 		if (!doasyncfree)
902 			ffs_update(vp, 1);
903 	}
904 	if (ssize < len) {
905 		if (doasyncfree)
906 			bdwrite(ebp);
907 		else
908 			bwrite(ebp);
909 	}
910 	/*
911 	 * Last, free the old blocks and assign the new blocks to the buffers.
912 	 */
913 #ifdef DEBUG
914 	if (prtrealloc)
915 		printf("\n\tnew:");
916 #endif
917 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
918 		if (!DOINGSOFTDEP(vp))
919 			ffs_blkfree(ump, fs, ip->i_devvp,
920 			    dbtofsb(fs, buflist->bs_children[i]->b_blkno),
921 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL);
922 		buflist->bs_children[i]->b_blkno = fsbtodb(fs, blkno);
923 #ifdef INVARIANTS
924 		if (!ffs_checkblk(ip,
925 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
926 			panic("ffs_reallocblks: unallocated block 3");
927 #endif
928 #ifdef DEBUG
929 		if (prtrealloc)
930 			printf(" %jd,", (intmax_t)blkno);
931 #endif
932 	}
933 #ifdef DEBUG
934 	if (prtrealloc) {
935 		prtrealloc--;
936 		printf("\n");
937 	}
938 #endif
939 	return (0);
940 
941 fail:
942 	if (ssize < len)
943 		brelse(ebp);
944 	if (sbap != &ip->i_din2->di_db[0])
945 		brelse(sbp);
946 	return (ENOSPC);
947 }
948 
949 /*
950  * Allocate an inode in the filesystem.
951  *
952  * If allocating a directory, use ffs_dirpref to select the inode.
953  * If allocating in a directory, the following hierarchy is followed:
954  *   1) allocate the preferred inode.
955  *   2) allocate an inode in the same cylinder group.
956  *   3) quadradically rehash into other cylinder groups, until an
957  *      available inode is located.
958  * If no inode preference is given the following hierarchy is used
959  * to allocate an inode:
960  *   1) allocate an inode in cylinder group 0.
961  *   2) quadradically rehash into other cylinder groups, until an
962  *      available inode is located.
963  */
964 int
965 ffs_valloc(pvp, mode, cred, vpp)
966 	struct vnode *pvp;
967 	int mode;
968 	struct ucred *cred;
969 	struct vnode **vpp;
970 {
971 	struct inode *pip;
972 	struct fs *fs;
973 	struct inode *ip;
974 	struct timespec ts;
975 	struct ufsmount *ump;
976 	ino_t ino, ipref;
977 	u_int cg;
978 	int error, error1, reclaimed;
979 	static struct timeval lastfail;
980 	static int curfail;
981 
982 	*vpp = NULL;
983 	pip = VTOI(pvp);
984 	fs = pip->i_fs;
985 	ump = pip->i_ump;
986 
987 	UFS_LOCK(ump);
988 	reclaimed = 0;
989 retry:
990 	if (fs->fs_cstotal.cs_nifree == 0)
991 		goto noinodes;
992 
993 	if ((mode & IFMT) == IFDIR)
994 		ipref = ffs_dirpref(pip);
995 	else
996 		ipref = pip->i_number;
997 	if (ipref >= fs->fs_ncg * fs->fs_ipg)
998 		ipref = 0;
999 	cg = ino_to_cg(fs, ipref);
1000 	/*
1001 	 * Track number of dirs created one after another
1002 	 * in a same cg without intervening by files.
1003 	 */
1004 	if ((mode & IFMT) == IFDIR) {
1005 		if (fs->fs_contigdirs[cg] < 255)
1006 			fs->fs_contigdirs[cg]++;
1007 	} else {
1008 		if (fs->fs_contigdirs[cg] > 0)
1009 			fs->fs_contigdirs[cg]--;
1010 	}
1011 	ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0,
1012 					(allocfcn_t *)ffs_nodealloccg);
1013 	if (ino == 0)
1014 		goto noinodes;
1015 	error = ffs_vget(pvp->v_mount, ino, LK_EXCLUSIVE, vpp);
1016 	if (error) {
1017 		error1 = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp,
1018 		    FFSV_FORCEINSMQ);
1019 		ffs_vfree(pvp, ino, mode);
1020 		if (error1 == 0) {
1021 			ip = VTOI(*vpp);
1022 			if (ip->i_mode)
1023 				goto dup_alloc;
1024 			ip->i_flag |= IN_MODIFIED;
1025 			vput(*vpp);
1026 		}
1027 		return (error);
1028 	}
1029 	ip = VTOI(*vpp);
1030 	if (ip->i_mode) {
1031 dup_alloc:
1032 		printf("mode = 0%o, inum = %lu, fs = %s\n",
1033 		    ip->i_mode, (u_long)ip->i_number, fs->fs_fsmnt);
1034 		panic("ffs_valloc: dup alloc");
1035 	}
1036 	if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) {  /* XXX */
1037 		printf("free inode %s/%lu had %ld blocks\n",
1038 		    fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks));
1039 		DIP_SET(ip, i_blocks, 0);
1040 	}
1041 	ip->i_flags = 0;
1042 	DIP_SET(ip, i_flags, 0);
1043 	/*
1044 	 * Set up a new generation number for this inode.
1045 	 */
1046 	if (ip->i_gen == 0 || ++ip->i_gen == 0)
1047 		ip->i_gen = arc4random() / 2 + 1;
1048 	DIP_SET(ip, i_gen, ip->i_gen);
1049 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1050 		vfs_timestamp(&ts);
1051 		ip->i_din2->di_birthtime = ts.tv_sec;
1052 		ip->i_din2->di_birthnsec = ts.tv_nsec;
1053 	}
1054 	ufs_prepare_reclaim(*vpp);
1055 	ip->i_flag = 0;
1056 	(*vpp)->v_vflag = 0;
1057 	(*vpp)->v_type = VNON;
1058 	if (fs->fs_magic == FS_UFS2_MAGIC)
1059 		(*vpp)->v_op = &ffs_vnodeops2;
1060 	else
1061 		(*vpp)->v_op = &ffs_vnodeops1;
1062 	return (0);
1063 noinodes:
1064 	if (reclaimed == 0) {
1065 		reclaimed = 1;
1066 		softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT);
1067 		goto retry;
1068 	}
1069 	UFS_UNLOCK(ump);
1070 	if (ppsratecheck(&lastfail, &curfail, 1)) {
1071 		ffs_fserr(fs, pip->i_number, "out of inodes");
1072 		uprintf("\n%s: create/symlink failed, no inodes free\n",
1073 		    fs->fs_fsmnt);
1074 	}
1075 	return (ENOSPC);
1076 }
1077 
1078 /*
1079  * Find a cylinder group to place a directory.
1080  *
1081  * The policy implemented by this algorithm is to allocate a
1082  * directory inode in the same cylinder group as its parent
1083  * directory, but also to reserve space for its files inodes
1084  * and data. Restrict the number of directories which may be
1085  * allocated one after another in the same cylinder group
1086  * without intervening allocation of files.
1087  *
1088  * If we allocate a first level directory then force allocation
1089  * in another cylinder group.
1090  */
1091 static ino_t
1092 ffs_dirpref(pip)
1093 	struct inode *pip;
1094 {
1095 	struct fs *fs;
1096 	int cg, prefcg, dirsize, cgsize;
1097 	u_int avgifree, avgbfree, avgndir, curdirsize;
1098 	u_int minifree, minbfree, maxndir;
1099 	u_int mincg, minndir;
1100 	u_int maxcontigdirs;
1101 
1102 	mtx_assert(UFS_MTX(pip->i_ump), MA_OWNED);
1103 	fs = pip->i_fs;
1104 
1105 	avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg;
1106 	avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1107 	avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg;
1108 
1109 	/*
1110 	 * Force allocation in another cg if creating a first level dir.
1111 	 */
1112 	ASSERT_VOP_LOCKED(ITOV(pip), "ffs_dirpref");
1113 	if (ITOV(pip)->v_vflag & VV_ROOT) {
1114 		prefcg = arc4random() % fs->fs_ncg;
1115 		mincg = prefcg;
1116 		minndir = fs->fs_ipg;
1117 		for (cg = prefcg; cg < fs->fs_ncg; cg++)
1118 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1119 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1120 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1121 				mincg = cg;
1122 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1123 			}
1124 		for (cg = 0; cg < prefcg; cg++)
1125 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1126 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1127 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1128 				mincg = cg;
1129 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1130 			}
1131 		return ((ino_t)(fs->fs_ipg * mincg));
1132 	}
1133 
1134 	/*
1135 	 * Count various limits which used for
1136 	 * optimal allocation of a directory inode.
1137 	 */
1138 	maxndir = min(avgndir + fs->fs_ipg / 16, fs->fs_ipg);
1139 	minifree = avgifree - avgifree / 4;
1140 	if (minifree < 1)
1141 		minifree = 1;
1142 	minbfree = avgbfree - avgbfree / 4;
1143 	if (minbfree < 1)
1144 		minbfree = 1;
1145 	cgsize = fs->fs_fsize * fs->fs_fpg;
1146 	dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir;
1147 	curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0;
1148 	if (dirsize < curdirsize)
1149 		dirsize = curdirsize;
1150 	if (dirsize <= 0)
1151 		maxcontigdirs = 0;		/* dirsize overflowed */
1152 	else
1153 		maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255);
1154 	if (fs->fs_avgfpdir > 0)
1155 		maxcontigdirs = min(maxcontigdirs,
1156 				    fs->fs_ipg / fs->fs_avgfpdir);
1157 	if (maxcontigdirs == 0)
1158 		maxcontigdirs = 1;
1159 
1160 	/*
1161 	 * Limit number of dirs in one cg and reserve space for
1162 	 * regular files, but only if we have no deficit in
1163 	 * inodes or space.
1164 	 *
1165 	 * We are trying to find a suitable cylinder group nearby
1166 	 * our preferred cylinder group to place a new directory.
1167 	 * We scan from our preferred cylinder group forward looking
1168 	 * for a cylinder group that meets our criterion. If we get
1169 	 * to the final cylinder group and do not find anything,
1170 	 * we start scanning backwards from our preferred cylinder
1171 	 * group. The ideal would be to alternate looking forward
1172 	 * and backward, but that is just too complex to code for
1173 	 * the gain it would get. The most likely place where the
1174 	 * backward scan would take effect is when we start near
1175 	 * the end of the filesystem and do not find anything from
1176 	 * where we are to the end. In that case, scanning backward
1177 	 * will likely find us a suitable cylinder group much closer
1178 	 * to our desired location than if we were to start scanning
1179 	 * forward from the beginning of the filesystem.
1180 	 */
1181 	prefcg = ino_to_cg(fs, pip->i_number);
1182 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1183 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1184 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1185 	    	    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1186 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1187 				return ((ino_t)(fs->fs_ipg * cg));
1188 		}
1189 	for (cg = 0; cg < prefcg; cg++)
1190 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1191 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1192 	    	    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1193 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1194 				return ((ino_t)(fs->fs_ipg * cg));
1195 		}
1196 	/*
1197 	 * This is a backstop when we have deficit in space.
1198 	 */
1199 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1200 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1201 			return ((ino_t)(fs->fs_ipg * cg));
1202 	for (cg = 0; cg < prefcg; cg++)
1203 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1204 			break;
1205 	return ((ino_t)(fs->fs_ipg * cg));
1206 }
1207 
1208 /*
1209  * Select the desired position for the next block in a file.  The file is
1210  * logically divided into sections. The first section is composed of the
1211  * direct blocks and the next fs_maxbpg blocks. Each additional section
1212  * contains fs_maxbpg blocks.
1213  *
1214  * If no blocks have been allocated in the first section, the policy is to
1215  * request a block in the same cylinder group as the inode that describes
1216  * the file. The first indirect is allocated immediately following the last
1217  * direct block and the data blocks for the first indirect immediately
1218  * follow it.
1219  *
1220  * If no blocks have been allocated in any other section, the indirect
1221  * block(s) are allocated in the same cylinder group as its inode in an
1222  * area reserved immediately following the inode blocks. The policy for
1223  * the data blocks is to place them in a cylinder group with a greater than
1224  * average number of free blocks. An appropriate cylinder group is found
1225  * by using a rotor that sweeps the cylinder groups. When a new group of
1226  * blocks is needed, the sweep begins in the cylinder group following the
1227  * cylinder group from which the previous allocation was made. The sweep
1228  * continues until a cylinder group with greater than the average number
1229  * of free blocks is found. If the allocation is for the first block in an
1230  * indirect block or the previous block is a hole, then the information on
1231  * the previous allocation is unavailable; here a best guess is made based
1232  * on the logical block number being allocated.
1233  *
1234  * If a section is already partially allocated, the policy is to
1235  * allocate blocks contiguously within the section if possible.
1236  */
1237 ufs2_daddr_t
1238 ffs_blkpref_ufs1(ip, lbn, indx, bap)
1239 	struct inode *ip;
1240 	ufs_lbn_t lbn;
1241 	int indx;
1242 	ufs1_daddr_t *bap;
1243 {
1244 	struct fs *fs;
1245 	u_int cg, inocg;
1246 	u_int avgbfree, startcg;
1247 	ufs2_daddr_t pref;
1248 
1249 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1250 	mtx_assert(UFS_MTX(ip->i_ump), MA_OWNED);
1251 	fs = ip->i_fs;
1252 	/*
1253 	 * Allocation of indirect blocks is indicated by passing negative
1254 	 * values in indx: -1 for single indirect, -2 for double indirect,
1255 	 * -3 for triple indirect. As noted below, we attempt to allocate
1256 	 * the first indirect inline with the file data. For all later
1257 	 * indirect blocks, the data is often allocated in other cylinder
1258 	 * groups. However to speed random file access and to speed up
1259 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1260 	 * (typically half of fs_minfree) of the data area of each cylinder
1261 	 * group to hold these later indirect blocks.
1262 	 */
1263 	inocg = ino_to_cg(fs, ip->i_number);
1264 	if (indx < 0) {
1265 		/*
1266 		 * Our preference for indirect blocks is the zone at the
1267 		 * beginning of the inode's cylinder group data area that
1268 		 * we try to reserve for indirect blocks.
1269 		 */
1270 		pref = cgmeta(fs, inocg);
1271 		/*
1272 		 * If we are allocating the first indirect block, try to
1273 		 * place it immediately following the last direct block.
1274 		 */
1275 		if (indx == -1 && lbn < NDADDR + NINDIR(fs) &&
1276 		    ip->i_din1->di_db[NDADDR - 1] != 0)
1277 			pref = ip->i_din1->di_db[NDADDR - 1] + fs->fs_frag;
1278 		return (pref);
1279 	}
1280 	/*
1281 	 * If we are allocating the first data block in the first indirect
1282 	 * block and the indirect has been allocated in the data block area,
1283 	 * try to place it immediately following the indirect block.
1284 	 */
1285 	if (lbn == NDADDR) {
1286 		pref = ip->i_din1->di_ib[0];
1287 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1288 		    pref < cgbase(fs, inocg + 1))
1289 			return (pref + fs->fs_frag);
1290 	}
1291 	/*
1292 	 * If we are at the beginning of a file, or we have already allocated
1293 	 * the maximum number of blocks per cylinder group, or we do not
1294 	 * have a block allocated immediately preceeding us, then we need
1295 	 * to decide where to start allocating new blocks.
1296 	 */
1297 	if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1298 		/*
1299 		 * If we are allocating a directory data block, we want
1300 		 * to place it in the metadata area.
1301 		 */
1302 		if ((ip->i_mode & IFMT) == IFDIR)
1303 			return (cgmeta(fs, inocg));
1304 		/*
1305 		 * Until we fill all the direct and all the first indirect's
1306 		 * blocks, we try to allocate in the data area of the inode's
1307 		 * cylinder group.
1308 		 */
1309 		if (lbn < NDADDR + NINDIR(fs))
1310 			return (cgdata(fs, inocg));
1311 		/*
1312 		 * Find a cylinder with greater than average number of
1313 		 * unused data blocks.
1314 		 */
1315 		if (indx == 0 || bap[indx - 1] == 0)
1316 			startcg = inocg + lbn / fs->fs_maxbpg;
1317 		else
1318 			startcg = dtog(fs, bap[indx - 1]) + 1;
1319 		startcg %= fs->fs_ncg;
1320 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1321 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1322 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1323 				fs->fs_cgrotor = cg;
1324 				return (cgdata(fs, cg));
1325 			}
1326 		for (cg = 0; cg <= startcg; cg++)
1327 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1328 				fs->fs_cgrotor = cg;
1329 				return (cgdata(fs, cg));
1330 			}
1331 		return (0);
1332 	}
1333 	/*
1334 	 * Otherwise, we just always try to lay things out contiguously.
1335 	 */
1336 	return (bap[indx - 1] + fs->fs_frag);
1337 }
1338 
1339 /*
1340  * Same as above, but for UFS2
1341  */
1342 ufs2_daddr_t
1343 ffs_blkpref_ufs2(ip, lbn, indx, bap)
1344 	struct inode *ip;
1345 	ufs_lbn_t lbn;
1346 	int indx;
1347 	ufs2_daddr_t *bap;
1348 {
1349 	struct fs *fs;
1350 	u_int cg, inocg;
1351 	u_int avgbfree, startcg;
1352 	ufs2_daddr_t pref;
1353 
1354 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1355 	mtx_assert(UFS_MTX(ip->i_ump), MA_OWNED);
1356 	fs = ip->i_fs;
1357 	/*
1358 	 * Allocation of indirect blocks is indicated by passing negative
1359 	 * values in indx: -1 for single indirect, -2 for double indirect,
1360 	 * -3 for triple indirect. As noted below, we attempt to allocate
1361 	 * the first indirect inline with the file data. For all later
1362 	 * indirect blocks, the data is often allocated in other cylinder
1363 	 * groups. However to speed random file access and to speed up
1364 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1365 	 * (typically half of fs_minfree) of the data area of each cylinder
1366 	 * group to hold these later indirect blocks.
1367 	 */
1368 	inocg = ino_to_cg(fs, ip->i_number);
1369 	if (indx < 0) {
1370 		/*
1371 		 * Our preference for indirect blocks is the zone at the
1372 		 * beginning of the inode's cylinder group data area that
1373 		 * we try to reserve for indirect blocks.
1374 		 */
1375 		pref = cgmeta(fs, inocg);
1376 		/*
1377 		 * If we are allocating the first indirect block, try to
1378 		 * place it immediately following the last direct block.
1379 		 */
1380 		if (indx == -1 && lbn < NDADDR + NINDIR(fs) &&
1381 		    ip->i_din2->di_db[NDADDR - 1] != 0)
1382 			pref = ip->i_din2->di_db[NDADDR - 1] + fs->fs_frag;
1383 		return (pref);
1384 	}
1385 	/*
1386 	 * If we are allocating the first data block in the first indirect
1387 	 * block and the indirect has been allocated in the data block area,
1388 	 * try to place it immediately following the indirect block.
1389 	 */
1390 	if (lbn == NDADDR) {
1391 		pref = ip->i_din2->di_ib[0];
1392 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1393 		    pref < cgbase(fs, inocg + 1))
1394 			return (pref + fs->fs_frag);
1395 	}
1396 	/*
1397 	 * If we are at the beginning of a file, or we have already allocated
1398 	 * the maximum number of blocks per cylinder group, or we do not
1399 	 * have a block allocated immediately preceeding us, then we need
1400 	 * to decide where to start allocating new blocks.
1401 	 */
1402 	if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1403 		/*
1404 		 * If we are allocating a directory data block, we want
1405 		 * to place it in the metadata area.
1406 		 */
1407 		if ((ip->i_mode & IFMT) == IFDIR)
1408 			return (cgmeta(fs, inocg));
1409 		/*
1410 		 * Until we fill all the direct and all the first indirect's
1411 		 * blocks, we try to allocate in the data area of the inode's
1412 		 * cylinder group.
1413 		 */
1414 		if (lbn < NDADDR + NINDIR(fs))
1415 			return (cgdata(fs, inocg));
1416 		/*
1417 		 * Find a cylinder with greater than average number of
1418 		 * unused data blocks.
1419 		 */
1420 		if (indx == 0 || bap[indx - 1] == 0)
1421 			startcg = inocg + lbn / fs->fs_maxbpg;
1422 		else
1423 			startcg = dtog(fs, bap[indx - 1]) + 1;
1424 		startcg %= fs->fs_ncg;
1425 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1426 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1427 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1428 				fs->fs_cgrotor = cg;
1429 				return (cgdata(fs, cg));
1430 			}
1431 		for (cg = 0; cg <= startcg; cg++)
1432 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1433 				fs->fs_cgrotor = cg;
1434 				return (cgdata(fs, cg));
1435 			}
1436 		return (0);
1437 	}
1438 	/*
1439 	 * Otherwise, we just always try to lay things out contiguously.
1440 	 */
1441 	return (bap[indx - 1] + fs->fs_frag);
1442 }
1443 
1444 /*
1445  * Implement the cylinder overflow algorithm.
1446  *
1447  * The policy implemented by this algorithm is:
1448  *   1) allocate the block in its requested cylinder group.
1449  *   2) quadradically rehash on the cylinder group number.
1450  *   3) brute force search for a free block.
1451  *
1452  * Must be called with the UFS lock held.  Will release the lock on success
1453  * and return with it held on failure.
1454  */
1455 /*VARARGS5*/
1456 static ufs2_daddr_t
1457 ffs_hashalloc(ip, cg, pref, size, rsize, allocator)
1458 	struct inode *ip;
1459 	u_int cg;
1460 	ufs2_daddr_t pref;
1461 	int size;	/* Search size for data blocks, mode for inodes */
1462 	int rsize;	/* Real allocated size. */
1463 	allocfcn_t *allocator;
1464 {
1465 	struct fs *fs;
1466 	ufs2_daddr_t result;
1467 	u_int i, icg = cg;
1468 
1469 	mtx_assert(UFS_MTX(ip->i_ump), MA_OWNED);
1470 #ifdef INVARIANTS
1471 	if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
1472 		panic("ffs_hashalloc: allocation on suspended filesystem");
1473 #endif
1474 	fs = ip->i_fs;
1475 	/*
1476 	 * 1: preferred cylinder group
1477 	 */
1478 	result = (*allocator)(ip, cg, pref, size, rsize);
1479 	if (result)
1480 		return (result);
1481 	/*
1482 	 * 2: quadratic rehash
1483 	 */
1484 	for (i = 1; i < fs->fs_ncg; i *= 2) {
1485 		cg += i;
1486 		if (cg >= fs->fs_ncg)
1487 			cg -= fs->fs_ncg;
1488 		result = (*allocator)(ip, cg, 0, size, rsize);
1489 		if (result)
1490 			return (result);
1491 	}
1492 	/*
1493 	 * 3: brute force search
1494 	 * Note that we start at i == 2, since 0 was checked initially,
1495 	 * and 1 is always checked in the quadratic rehash.
1496 	 */
1497 	cg = (icg + 2) % fs->fs_ncg;
1498 	for (i = 2; i < fs->fs_ncg; i++) {
1499 		result = (*allocator)(ip, cg, 0, size, rsize);
1500 		if (result)
1501 			return (result);
1502 		cg++;
1503 		if (cg == fs->fs_ncg)
1504 			cg = 0;
1505 	}
1506 	return (0);
1507 }
1508 
1509 /*
1510  * Determine whether a fragment can be extended.
1511  *
1512  * Check to see if the necessary fragments are available, and
1513  * if they are, allocate them.
1514  */
1515 static ufs2_daddr_t
1516 ffs_fragextend(ip, cg, bprev, osize, nsize)
1517 	struct inode *ip;
1518 	u_int cg;
1519 	ufs2_daddr_t bprev;
1520 	int osize, nsize;
1521 {
1522 	struct fs *fs;
1523 	struct cg *cgp;
1524 	struct buf *bp;
1525 	struct ufsmount *ump;
1526 	int nffree;
1527 	long bno;
1528 	int frags, bbase;
1529 	int i, error;
1530 	u_int8_t *blksfree;
1531 
1532 	ump = ip->i_ump;
1533 	fs = ip->i_fs;
1534 	if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize))
1535 		return (0);
1536 	frags = numfrags(fs, nsize);
1537 	bbase = fragnum(fs, bprev);
1538 	if (bbase > fragnum(fs, (bprev + frags - 1))) {
1539 		/* cannot extend across a block boundary */
1540 		return (0);
1541 	}
1542 	UFS_UNLOCK(ump);
1543 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
1544 		(int)fs->fs_cgsize, NOCRED, &bp);
1545 	if (error)
1546 		goto fail;
1547 	cgp = (struct cg *)bp->b_data;
1548 	if (!cg_chkmagic(cgp))
1549 		goto fail;
1550 	bp->b_xflags |= BX_BKGRDWRITE;
1551 	cgp->cg_old_time = cgp->cg_time = time_second;
1552 	bno = dtogd(fs, bprev);
1553 	blksfree = cg_blksfree(cgp);
1554 	for (i = numfrags(fs, osize); i < frags; i++)
1555 		if (isclr(blksfree, bno + i))
1556 			goto fail;
1557 	/*
1558 	 * the current fragment can be extended
1559 	 * deduct the count on fragment being extended into
1560 	 * increase the count on the remaining fragment (if any)
1561 	 * allocate the extended piece
1562 	 */
1563 	for (i = frags; i < fs->fs_frag - bbase; i++)
1564 		if (isclr(blksfree, bno + i))
1565 			break;
1566 	cgp->cg_frsum[i - numfrags(fs, osize)]--;
1567 	if (i != frags)
1568 		cgp->cg_frsum[i - frags]++;
1569 	for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) {
1570 		clrbit(blksfree, bno + i);
1571 		cgp->cg_cs.cs_nffree--;
1572 		nffree++;
1573 	}
1574 	UFS_LOCK(ump);
1575 	fs->fs_cstotal.cs_nffree -= nffree;
1576 	fs->fs_cs(fs, cg).cs_nffree -= nffree;
1577 	fs->fs_fmod = 1;
1578 	ACTIVECLEAR(fs, cg);
1579 	UFS_UNLOCK(ump);
1580 	if (DOINGSOFTDEP(ITOV(ip)))
1581 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev,
1582 		    frags, numfrags(fs, osize));
1583 	bdwrite(bp);
1584 	return (bprev);
1585 
1586 fail:
1587 	brelse(bp);
1588 	UFS_LOCK(ump);
1589 	return (0);
1590 
1591 }
1592 
1593 /*
1594  * Determine whether a block can be allocated.
1595  *
1596  * Check to see if a block of the appropriate size is available,
1597  * and if it is, allocate it.
1598  */
1599 static ufs2_daddr_t
1600 ffs_alloccg(ip, cg, bpref, size, rsize)
1601 	struct inode *ip;
1602 	u_int cg;
1603 	ufs2_daddr_t bpref;
1604 	int size;
1605 	int rsize;
1606 {
1607 	struct fs *fs;
1608 	struct cg *cgp;
1609 	struct buf *bp;
1610 	struct ufsmount *ump;
1611 	ufs1_daddr_t bno;
1612 	ufs2_daddr_t blkno;
1613 	int i, allocsiz, error, frags;
1614 	u_int8_t *blksfree;
1615 
1616 	ump = ip->i_ump;
1617 	fs = ip->i_fs;
1618 	if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize)
1619 		return (0);
1620 	UFS_UNLOCK(ump);
1621 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
1622 		(int)fs->fs_cgsize, NOCRED, &bp);
1623 	if (error)
1624 		goto fail;
1625 	cgp = (struct cg *)bp->b_data;
1626 	if (!cg_chkmagic(cgp) ||
1627 	    (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize))
1628 		goto fail;
1629 	bp->b_xflags |= BX_BKGRDWRITE;
1630 	cgp->cg_old_time = cgp->cg_time = time_second;
1631 	if (size == fs->fs_bsize) {
1632 		UFS_LOCK(ump);
1633 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1634 		ACTIVECLEAR(fs, cg);
1635 		UFS_UNLOCK(ump);
1636 		bdwrite(bp);
1637 		return (blkno);
1638 	}
1639 	/*
1640 	 * check to see if any fragments are already available
1641 	 * allocsiz is the size which will be allocated, hacking
1642 	 * it down to a smaller size if necessary
1643 	 */
1644 	blksfree = cg_blksfree(cgp);
1645 	frags = numfrags(fs, size);
1646 	for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++)
1647 		if (cgp->cg_frsum[allocsiz] != 0)
1648 			break;
1649 	if (allocsiz == fs->fs_frag) {
1650 		/*
1651 		 * no fragments were available, so a block will be
1652 		 * allocated, and hacked up
1653 		 */
1654 		if (cgp->cg_cs.cs_nbfree == 0)
1655 			goto fail;
1656 		UFS_LOCK(ump);
1657 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1658 		ACTIVECLEAR(fs, cg);
1659 		UFS_UNLOCK(ump);
1660 		bdwrite(bp);
1661 		return (blkno);
1662 	}
1663 	KASSERT(size == rsize,
1664 	    ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize));
1665 	bno = ffs_mapsearch(fs, cgp, bpref, allocsiz);
1666 	if (bno < 0)
1667 		goto fail;
1668 	for (i = 0; i < frags; i++)
1669 		clrbit(blksfree, bno + i);
1670 	cgp->cg_cs.cs_nffree -= frags;
1671 	cgp->cg_frsum[allocsiz]--;
1672 	if (frags != allocsiz)
1673 		cgp->cg_frsum[allocsiz - frags]++;
1674 	UFS_LOCK(ump);
1675 	fs->fs_cstotal.cs_nffree -= frags;
1676 	fs->fs_cs(fs, cg).cs_nffree -= frags;
1677 	fs->fs_fmod = 1;
1678 	blkno = cgbase(fs, cg) + bno;
1679 	ACTIVECLEAR(fs, cg);
1680 	UFS_UNLOCK(ump);
1681 	if (DOINGSOFTDEP(ITOV(ip)))
1682 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0);
1683 	bdwrite(bp);
1684 	return (blkno);
1685 
1686 fail:
1687 	brelse(bp);
1688 	UFS_LOCK(ump);
1689 	return (0);
1690 }
1691 
1692 /*
1693  * Allocate a block in a cylinder group.
1694  *
1695  * This algorithm implements the following policy:
1696  *   1) allocate the requested block.
1697  *   2) allocate a rotationally optimal block in the same cylinder.
1698  *   3) allocate the next available block on the block rotor for the
1699  *      specified cylinder group.
1700  * Note that this routine only allocates fs_bsize blocks; these
1701  * blocks may be fragmented by the routine that allocates them.
1702  */
1703 static ufs2_daddr_t
1704 ffs_alloccgblk(ip, bp, bpref, size)
1705 	struct inode *ip;
1706 	struct buf *bp;
1707 	ufs2_daddr_t bpref;
1708 	int size;
1709 {
1710 	struct fs *fs;
1711 	struct cg *cgp;
1712 	struct ufsmount *ump;
1713 	ufs1_daddr_t bno;
1714 	ufs2_daddr_t blkno;
1715 	u_int8_t *blksfree;
1716 	int i, cgbpref;
1717 
1718 	fs = ip->i_fs;
1719 	ump = ip->i_ump;
1720 	mtx_assert(UFS_MTX(ump), MA_OWNED);
1721 	cgp = (struct cg *)bp->b_data;
1722 	blksfree = cg_blksfree(cgp);
1723 	if (bpref == 0) {
1724 		bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag;
1725 	} else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) {
1726 		/* map bpref to correct zone in this cg */
1727 		if (bpref < cgdata(fs, cgbpref))
1728 			bpref = cgmeta(fs, cgp->cg_cgx);
1729 		else
1730 			bpref = cgdata(fs, cgp->cg_cgx);
1731 	}
1732 	/*
1733 	 * if the requested block is available, use it
1734 	 */
1735 	bno = dtogd(fs, blknum(fs, bpref));
1736 	if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno)))
1737 		goto gotit;
1738 	/*
1739 	 * Take the next available block in this cylinder group.
1740 	 */
1741 	bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag);
1742 	if (bno < 0)
1743 		return (0);
1744 	/* Update cg_rotor only if allocated from the data zone */
1745 	if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx)))
1746 		cgp->cg_rotor = bno;
1747 gotit:
1748 	blkno = fragstoblks(fs, bno);
1749 	ffs_clrblock(fs, blksfree, (long)blkno);
1750 	ffs_clusteracct(fs, cgp, blkno, -1);
1751 	cgp->cg_cs.cs_nbfree--;
1752 	fs->fs_cstotal.cs_nbfree--;
1753 	fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--;
1754 	fs->fs_fmod = 1;
1755 	blkno = cgbase(fs, cgp->cg_cgx) + bno;
1756 	/*
1757 	 * If the caller didn't want the whole block free the frags here.
1758 	 */
1759 	size = numfrags(fs, size);
1760 	if (size != fs->fs_frag) {
1761 		bno = dtogd(fs, blkno);
1762 		for (i = size; i < fs->fs_frag; i++)
1763 			setbit(blksfree, bno + i);
1764 		i = fs->fs_frag - size;
1765 		cgp->cg_cs.cs_nffree += i;
1766 		fs->fs_cstotal.cs_nffree += i;
1767 		fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i;
1768 		fs->fs_fmod = 1;
1769 		cgp->cg_frsum[i]++;
1770 	}
1771 	/* XXX Fixme. */
1772 	UFS_UNLOCK(ump);
1773 	if (DOINGSOFTDEP(ITOV(ip)))
1774 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno,
1775 		    size, 0);
1776 	UFS_LOCK(ump);
1777 	return (blkno);
1778 }
1779 
1780 /*
1781  * Determine whether a cluster can be allocated.
1782  *
1783  * We do not currently check for optimal rotational layout if there
1784  * are multiple choices in the same cylinder group. Instead we just
1785  * take the first one that we find following bpref.
1786  */
1787 static ufs2_daddr_t
1788 ffs_clusteralloc(ip, cg, bpref, len, unused)
1789 	struct inode *ip;
1790 	u_int cg;
1791 	ufs2_daddr_t bpref;
1792 	int len;
1793 	int unused;
1794 {
1795 	struct fs *fs;
1796 	struct cg *cgp;
1797 	struct buf *bp;
1798 	struct ufsmount *ump;
1799 	int i, run, bit, map, got;
1800 	ufs2_daddr_t bno;
1801 	u_char *mapp;
1802 	int32_t *lp;
1803 	u_int8_t *blksfree;
1804 
1805 	fs = ip->i_fs;
1806 	ump = ip->i_ump;
1807 	if (fs->fs_maxcluster[cg] < len)
1808 		return (0);
1809 	UFS_UNLOCK(ump);
1810 	if (bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)), (int)fs->fs_cgsize,
1811 	    NOCRED, &bp))
1812 		goto fail_lock;
1813 	cgp = (struct cg *)bp->b_data;
1814 	if (!cg_chkmagic(cgp))
1815 		goto fail_lock;
1816 	bp->b_xflags |= BX_BKGRDWRITE;
1817 	/*
1818 	 * Check to see if a cluster of the needed size (or bigger) is
1819 	 * available in this cylinder group.
1820 	 */
1821 	lp = &cg_clustersum(cgp)[len];
1822 	for (i = len; i <= fs->fs_contigsumsize; i++)
1823 		if (*lp++ > 0)
1824 			break;
1825 	if (i > fs->fs_contigsumsize) {
1826 		/*
1827 		 * This is the first time looking for a cluster in this
1828 		 * cylinder group. Update the cluster summary information
1829 		 * to reflect the true maximum sized cluster so that
1830 		 * future cluster allocation requests can avoid reading
1831 		 * the cylinder group map only to find no clusters.
1832 		 */
1833 		lp = &cg_clustersum(cgp)[len - 1];
1834 		for (i = len - 1; i > 0; i--)
1835 			if (*lp-- > 0)
1836 				break;
1837 		UFS_LOCK(ump);
1838 		fs->fs_maxcluster[cg] = i;
1839 		goto fail;
1840 	}
1841 	/*
1842 	 * Search the cluster map to find a big enough cluster.
1843 	 * We take the first one that we find, even if it is larger
1844 	 * than we need as we prefer to get one close to the previous
1845 	 * block allocation. We do not search before the current
1846 	 * preference point as we do not want to allocate a block
1847 	 * that is allocated before the previous one (as we will
1848 	 * then have to wait for another pass of the elevator
1849 	 * algorithm before it will be read). We prefer to fail and
1850 	 * be recalled to try an allocation in the next cylinder group.
1851 	 */
1852 	if (dtog(fs, bpref) != cg)
1853 		bpref = cgdata(fs, cg);
1854 	else
1855 		bpref = blknum(fs, bpref);
1856 	bpref = fragstoblks(fs, dtogd(fs, bpref));
1857 	mapp = &cg_clustersfree(cgp)[bpref / NBBY];
1858 	map = *mapp++;
1859 	bit = 1 << (bpref % NBBY);
1860 	for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) {
1861 		if ((map & bit) == 0) {
1862 			run = 0;
1863 		} else {
1864 			run++;
1865 			if (run == len)
1866 				break;
1867 		}
1868 		if ((got & (NBBY - 1)) != (NBBY - 1)) {
1869 			bit <<= 1;
1870 		} else {
1871 			map = *mapp++;
1872 			bit = 1;
1873 		}
1874 	}
1875 	if (got >= cgp->cg_nclusterblks)
1876 		goto fail_lock;
1877 	/*
1878 	 * Allocate the cluster that we have found.
1879 	 */
1880 	blksfree = cg_blksfree(cgp);
1881 	for (i = 1; i <= len; i++)
1882 		if (!ffs_isblock(fs, blksfree, got - run + i))
1883 			panic("ffs_clusteralloc: map mismatch");
1884 	bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1);
1885 	if (dtog(fs, bno) != cg)
1886 		panic("ffs_clusteralloc: allocated out of group");
1887 	len = blkstofrags(fs, len);
1888 	UFS_LOCK(ump);
1889 	for (i = 0; i < len; i += fs->fs_frag)
1890 		if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i)
1891 			panic("ffs_clusteralloc: lost block");
1892 	ACTIVECLEAR(fs, cg);
1893 	UFS_UNLOCK(ump);
1894 	bdwrite(bp);
1895 	return (bno);
1896 
1897 fail_lock:
1898 	UFS_LOCK(ump);
1899 fail:
1900 	brelse(bp);
1901 	return (0);
1902 }
1903 
1904 static inline struct buf *
1905 getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags)
1906 {
1907 	struct fs *fs;
1908 
1909 	fs = ip->i_fs;
1910 	return (getblk(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs,
1911 	    cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
1912 	    gbflags));
1913 }
1914 
1915 /*
1916  * Determine whether an inode can be allocated.
1917  *
1918  * Check to see if an inode is available, and if it is,
1919  * allocate it using the following policy:
1920  *   1) allocate the requested inode.
1921  *   2) allocate the next available inode after the requested
1922  *      inode in the specified cylinder group.
1923  */
1924 static ufs2_daddr_t
1925 ffs_nodealloccg(ip, cg, ipref, mode, unused)
1926 	struct inode *ip;
1927 	u_int cg;
1928 	ufs2_daddr_t ipref;
1929 	int mode;
1930 	int unused;
1931 {
1932 	struct fs *fs;
1933 	struct cg *cgp;
1934 	struct buf *bp, *ibp;
1935 	struct ufsmount *ump;
1936 	u_int8_t *inosused, *loc;
1937 	struct ufs2_dinode *dp2;
1938 	int error, start, len, i;
1939 	u_int32_t old_initediblk;
1940 
1941 	fs = ip->i_fs;
1942 	ump = ip->i_ump;
1943 check_nifree:
1944 	if (fs->fs_cs(fs, cg).cs_nifree == 0)
1945 		return (0);
1946 	UFS_UNLOCK(ump);
1947 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
1948 		(int)fs->fs_cgsize, NOCRED, &bp);
1949 	if (error) {
1950 		brelse(bp);
1951 		UFS_LOCK(ump);
1952 		return (0);
1953 	}
1954 	cgp = (struct cg *)bp->b_data;
1955 restart:
1956 	if (!cg_chkmagic(cgp) || cgp->cg_cs.cs_nifree == 0) {
1957 		brelse(bp);
1958 		UFS_LOCK(ump);
1959 		return (0);
1960 	}
1961 	bp->b_xflags |= BX_BKGRDWRITE;
1962 	inosused = cg_inosused(cgp);
1963 	if (ipref) {
1964 		ipref %= fs->fs_ipg;
1965 		if (isclr(inosused, ipref))
1966 			goto gotit;
1967 	}
1968 	start = cgp->cg_irotor / NBBY;
1969 	len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY);
1970 	loc = memcchr(&inosused[start], 0xff, len);
1971 	if (loc == NULL) {
1972 		len = start + 1;
1973 		start = 0;
1974 		loc = memcchr(&inosused[start], 0xff, len);
1975 		if (loc == NULL) {
1976 			printf("cg = %d, irotor = %ld, fs = %s\n",
1977 			    cg, (long)cgp->cg_irotor, fs->fs_fsmnt);
1978 			panic("ffs_nodealloccg: map corrupted");
1979 			/* NOTREACHED */
1980 		}
1981 	}
1982 	ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
1983 gotit:
1984 	/*
1985 	 * Check to see if we need to initialize more inodes.
1986 	 */
1987 	if (fs->fs_magic == FS_UFS2_MAGIC &&
1988 	    ipref + INOPB(fs) > cgp->cg_initediblk &&
1989 	    cgp->cg_initediblk < cgp->cg_niblk) {
1990 		old_initediblk = cgp->cg_initediblk;
1991 
1992 		/*
1993 		 * Free the cylinder group lock before writing the
1994 		 * initialized inode block.  Entering the
1995 		 * babarrierwrite() with the cylinder group lock
1996 		 * causes lock order violation between the lock and
1997 		 * snaplk.
1998 		 *
1999 		 * Another thread can decide to initialize the same
2000 		 * inode block, but whichever thread first gets the
2001 		 * cylinder group lock after writing the newly
2002 		 * allocated inode block will update it and the other
2003 		 * will realize that it has lost and leave the
2004 		 * cylinder group unchanged.
2005 		 */
2006 		ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
2007 		brelse(bp);
2008 		if (ibp == NULL) {
2009 			/*
2010 			 * The inode block buffer is already owned by
2011 			 * another thread, which must initialize it.
2012 			 * Wait on the buffer to allow another thread
2013 			 * to finish the updates, with dropped cg
2014 			 * buffer lock, then retry.
2015 			 */
2016 			ibp = getinobuf(ip, cg, old_initediblk, 0);
2017 			brelse(ibp);
2018 			UFS_LOCK(ump);
2019 			goto check_nifree;
2020 		}
2021 		bzero(ibp->b_data, (int)fs->fs_bsize);
2022 		dp2 = (struct ufs2_dinode *)(ibp->b_data);
2023 		for (i = 0; i < INOPB(fs); i++) {
2024 			dp2->di_gen = arc4random() / 2 + 1;
2025 			dp2++;
2026 		}
2027 		/*
2028 		 * Rather than adding a soft updates dependency to ensure
2029 		 * that the new inode block is written before it is claimed
2030 		 * by the cylinder group map, we just do a barrier write
2031 		 * here. The barrier write will ensure that the inode block
2032 		 * gets written before the updated cylinder group map can be
2033 		 * written. The barrier write should only slow down bulk
2034 		 * loading of newly created filesystems.
2035 		 */
2036 		babarrierwrite(ibp);
2037 
2038 		/*
2039 		 * After the inode block is written, try to update the
2040 		 * cg initediblk pointer.  If another thread beat us
2041 		 * to it, then leave it unchanged as the other thread
2042 		 * has already set it correctly.
2043 		 */
2044 		error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
2045 		    (int)fs->fs_cgsize, NOCRED, &bp);
2046 		UFS_LOCK(ump);
2047 		ACTIVECLEAR(fs, cg);
2048 		UFS_UNLOCK(ump);
2049 		if (error != 0) {
2050 			brelse(bp);
2051 			return (error);
2052 		}
2053 		cgp = (struct cg *)bp->b_data;
2054 		if (cgp->cg_initediblk == old_initediblk)
2055 			cgp->cg_initediblk += INOPB(fs);
2056 		goto restart;
2057 	}
2058 	cgp->cg_old_time = cgp->cg_time = time_second;
2059 	cgp->cg_irotor = ipref;
2060 	UFS_LOCK(ump);
2061 	ACTIVECLEAR(fs, cg);
2062 	setbit(inosused, ipref);
2063 	cgp->cg_cs.cs_nifree--;
2064 	fs->fs_cstotal.cs_nifree--;
2065 	fs->fs_cs(fs, cg).cs_nifree--;
2066 	fs->fs_fmod = 1;
2067 	if ((mode & IFMT) == IFDIR) {
2068 		cgp->cg_cs.cs_ndir++;
2069 		fs->fs_cstotal.cs_ndir++;
2070 		fs->fs_cs(fs, cg).cs_ndir++;
2071 	}
2072 	UFS_UNLOCK(ump);
2073 	if (DOINGSOFTDEP(ITOV(ip)))
2074 		softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode);
2075 	bdwrite(bp);
2076 	return ((ino_t)(cg * fs->fs_ipg + ipref));
2077 }
2078 
2079 /*
2080  * Free a block or fragment.
2081  *
2082  * The specified block or fragment is placed back in the
2083  * free map. If a fragment is deallocated, a possible
2084  * block reassembly is checked.
2085  */
2086 static void
2087 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd)
2088 	struct ufsmount *ump;
2089 	struct fs *fs;
2090 	struct vnode *devvp;
2091 	ufs2_daddr_t bno;
2092 	long size;
2093 	ino_t inum;
2094 	struct workhead *dephd;
2095 {
2096 	struct mount *mp;
2097 	struct cg *cgp;
2098 	struct buf *bp;
2099 	ufs1_daddr_t fragno, cgbno;
2100 	ufs2_daddr_t cgblkno;
2101 	int i, blk, frags, bbase;
2102 	u_int cg;
2103 	u_int8_t *blksfree;
2104 	struct cdev *dev;
2105 
2106 	cg = dtog(fs, bno);
2107 	if (devvp->v_type == VREG) {
2108 		/* devvp is a snapshot */
2109 		dev = VTOI(devvp)->i_devvp->v_rdev;
2110 		cgblkno = fragstoblks(fs, cgtod(fs, cg));
2111 	} else {
2112 		/* devvp is a normal disk device */
2113 		dev = devvp->v_rdev;
2114 		cgblkno = fsbtodb(fs, cgtod(fs, cg));
2115 		ASSERT_VOP_LOCKED(devvp, "ffs_blkfree_cg");
2116 	}
2117 #ifdef INVARIANTS
2118 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 ||
2119 	    fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) {
2120 		printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n",
2121 		    devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize,
2122 		    size, fs->fs_fsmnt);
2123 		panic("ffs_blkfree_cg: bad size");
2124 	}
2125 #endif
2126 	if ((u_int)bno >= fs->fs_size) {
2127 		printf("bad block %jd, ino %lu\n", (intmax_t)bno,
2128 		    (u_long)inum);
2129 		ffs_fserr(fs, inum, "bad block");
2130 		return;
2131 	}
2132 	if (bread(devvp, cgblkno, (int)fs->fs_cgsize, NOCRED, &bp)) {
2133 		brelse(bp);
2134 		return;
2135 	}
2136 	cgp = (struct cg *)bp->b_data;
2137 	if (!cg_chkmagic(cgp)) {
2138 		brelse(bp);
2139 		return;
2140 	}
2141 	bp->b_xflags |= BX_BKGRDWRITE;
2142 	cgp->cg_old_time = cgp->cg_time = time_second;
2143 	cgbno = dtogd(fs, bno);
2144 	blksfree = cg_blksfree(cgp);
2145 	UFS_LOCK(ump);
2146 	if (size == fs->fs_bsize) {
2147 		fragno = fragstoblks(fs, cgbno);
2148 		if (!ffs_isfreeblock(fs, blksfree, fragno)) {
2149 			if (devvp->v_type == VREG) {
2150 				UFS_UNLOCK(ump);
2151 				/* devvp is a snapshot */
2152 				brelse(bp);
2153 				return;
2154 			}
2155 			printf("dev = %s, block = %jd, fs = %s\n",
2156 			    devtoname(dev), (intmax_t)bno, fs->fs_fsmnt);
2157 			panic("ffs_blkfree_cg: freeing free block");
2158 		}
2159 		ffs_setblock(fs, blksfree, fragno);
2160 		ffs_clusteracct(fs, cgp, fragno, 1);
2161 		cgp->cg_cs.cs_nbfree++;
2162 		fs->fs_cstotal.cs_nbfree++;
2163 		fs->fs_cs(fs, cg).cs_nbfree++;
2164 	} else {
2165 		bbase = cgbno - fragnum(fs, cgbno);
2166 		/*
2167 		 * decrement the counts associated with the old frags
2168 		 */
2169 		blk = blkmap(fs, blksfree, bbase);
2170 		ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
2171 		/*
2172 		 * deallocate the fragment
2173 		 */
2174 		frags = numfrags(fs, size);
2175 		for (i = 0; i < frags; i++) {
2176 			if (isset(blksfree, cgbno + i)) {
2177 				printf("dev = %s, block = %jd, fs = %s\n",
2178 				    devtoname(dev), (intmax_t)(bno + i),
2179 				    fs->fs_fsmnt);
2180 				panic("ffs_blkfree_cg: freeing free frag");
2181 			}
2182 			setbit(blksfree, cgbno + i);
2183 		}
2184 		cgp->cg_cs.cs_nffree += i;
2185 		fs->fs_cstotal.cs_nffree += i;
2186 		fs->fs_cs(fs, cg).cs_nffree += i;
2187 		/*
2188 		 * add back in counts associated with the new frags
2189 		 */
2190 		blk = blkmap(fs, blksfree, bbase);
2191 		ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
2192 		/*
2193 		 * if a complete block has been reassembled, account for it
2194 		 */
2195 		fragno = fragstoblks(fs, bbase);
2196 		if (ffs_isblock(fs, blksfree, fragno)) {
2197 			cgp->cg_cs.cs_nffree -= fs->fs_frag;
2198 			fs->fs_cstotal.cs_nffree -= fs->fs_frag;
2199 			fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag;
2200 			ffs_clusteracct(fs, cgp, fragno, 1);
2201 			cgp->cg_cs.cs_nbfree++;
2202 			fs->fs_cstotal.cs_nbfree++;
2203 			fs->fs_cs(fs, cg).cs_nbfree++;
2204 		}
2205 	}
2206 	fs->fs_fmod = 1;
2207 	ACTIVECLEAR(fs, cg);
2208 	UFS_UNLOCK(ump);
2209 	mp = UFSTOVFS(ump);
2210 	if (MOUNTEDSOFTDEP(mp) && devvp->v_type != VREG)
2211 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2212 		    numfrags(fs, size), dephd);
2213 	bdwrite(bp);
2214 }
2215 
2216 TASKQUEUE_DEFINE_THREAD(ffs_trim);
2217 
2218 struct ffs_blkfree_trim_params {
2219 	struct task task;
2220 	struct ufsmount *ump;
2221 	struct vnode *devvp;
2222 	ufs2_daddr_t bno;
2223 	long size;
2224 	ino_t inum;
2225 	struct workhead *pdephd;
2226 	struct workhead dephd;
2227 };
2228 
2229 static void
2230 ffs_blkfree_trim_task(ctx, pending)
2231 	void *ctx;
2232 	int pending;
2233 {
2234 	struct ffs_blkfree_trim_params *tp;
2235 
2236 	tp = ctx;
2237 	ffs_blkfree_cg(tp->ump, tp->ump->um_fs, tp->devvp, tp->bno, tp->size,
2238 	    tp->inum, tp->pdephd);
2239 	vn_finished_secondary_write(UFSTOVFS(tp->ump));
2240 	free(tp, M_TEMP);
2241 }
2242 
2243 static void
2244 ffs_blkfree_trim_completed(bip)
2245 	struct bio *bip;
2246 {
2247 	struct ffs_blkfree_trim_params *tp;
2248 
2249 	tp = bip->bio_caller2;
2250 	g_destroy_bio(bip);
2251 	TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp);
2252 	taskqueue_enqueue(taskqueue_ffs_trim, &tp->task);
2253 }
2254 
2255 void
2256 ffs_blkfree(ump, fs, devvp, bno, size, inum, vtype, dephd)
2257 	struct ufsmount *ump;
2258 	struct fs *fs;
2259 	struct vnode *devvp;
2260 	ufs2_daddr_t bno;
2261 	long size;
2262 	ino_t inum;
2263 	enum vtype vtype;
2264 	struct workhead *dephd;
2265 {
2266 	struct mount *mp;
2267 	struct bio *bip;
2268 	struct ffs_blkfree_trim_params *tp;
2269 
2270 	/*
2271 	 * Check to see if a snapshot wants to claim the block.
2272 	 * Check that devvp is a normal disk device, not a snapshot,
2273 	 * it has a snapshot(s) associated with it, and one of the
2274 	 * snapshots wants to claim the block.
2275 	 */
2276 	if (devvp->v_type != VREG &&
2277 	    (devvp->v_vflag & VV_COPYONWRITE) &&
2278 	    ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) {
2279 		return;
2280 	}
2281 	/*
2282 	 * Nothing to delay if TRIM is disabled, or the operation is
2283 	 * performed on the snapshot.
2284 	 */
2285 	if (!ump->um_candelete || devvp->v_type == VREG) {
2286 		ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd);
2287 		return;
2288 	}
2289 
2290 	/*
2291 	 * Postpone the set of the free bit in the cg bitmap until the
2292 	 * BIO_DELETE is completed.  Otherwise, due to disk queue
2293 	 * reordering, TRIM might be issued after we reuse the block
2294 	 * and write some new data into it.
2295 	 */
2296 	tp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TEMP, M_WAITOK);
2297 	tp->ump = ump;
2298 	tp->devvp = devvp;
2299 	tp->bno = bno;
2300 	tp->size = size;
2301 	tp->inum = inum;
2302 	if (dephd != NULL) {
2303 		LIST_INIT(&tp->dephd);
2304 		LIST_SWAP(dephd, &tp->dephd, worklist, wk_list);
2305 		tp->pdephd = &tp->dephd;
2306 	} else
2307 		tp->pdephd = NULL;
2308 
2309 	bip = g_alloc_bio();
2310 	bip->bio_cmd = BIO_DELETE;
2311 	bip->bio_offset = dbtob(fsbtodb(fs, bno));
2312 	bip->bio_done = ffs_blkfree_trim_completed;
2313 	bip->bio_length = size;
2314 	bip->bio_caller2 = tp;
2315 
2316 	mp = UFSTOVFS(ump);
2317 	vn_start_secondary_write(NULL, &mp, 0);
2318 	g_io_request(bip, (struct g_consumer *)devvp->v_bufobj.bo_private);
2319 }
2320 
2321 #ifdef INVARIANTS
2322 /*
2323  * Verify allocation of a block or fragment. Returns true if block or
2324  * fragment is allocated, false if it is free.
2325  */
2326 static int
2327 ffs_checkblk(ip, bno, size)
2328 	struct inode *ip;
2329 	ufs2_daddr_t bno;
2330 	long size;
2331 {
2332 	struct fs *fs;
2333 	struct cg *cgp;
2334 	struct buf *bp;
2335 	ufs1_daddr_t cgbno;
2336 	int i, error, frags, free;
2337 	u_int8_t *blksfree;
2338 
2339 	fs = ip->i_fs;
2340 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
2341 		printf("bsize = %ld, size = %ld, fs = %s\n",
2342 		    (long)fs->fs_bsize, size, fs->fs_fsmnt);
2343 		panic("ffs_checkblk: bad size");
2344 	}
2345 	if ((u_int)bno >= fs->fs_size)
2346 		panic("ffs_checkblk: bad block %jd", (intmax_t)bno);
2347 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, dtog(fs, bno))),
2348 		(int)fs->fs_cgsize, NOCRED, &bp);
2349 	if (error)
2350 		panic("ffs_checkblk: cg bread failed");
2351 	cgp = (struct cg *)bp->b_data;
2352 	if (!cg_chkmagic(cgp))
2353 		panic("ffs_checkblk: cg magic mismatch");
2354 	bp->b_xflags |= BX_BKGRDWRITE;
2355 	blksfree = cg_blksfree(cgp);
2356 	cgbno = dtogd(fs, bno);
2357 	if (size == fs->fs_bsize) {
2358 		free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno));
2359 	} else {
2360 		frags = numfrags(fs, size);
2361 		for (free = 0, i = 0; i < frags; i++)
2362 			if (isset(blksfree, cgbno + i))
2363 				free++;
2364 		if (free != 0 && free != frags)
2365 			panic("ffs_checkblk: partially free fragment");
2366 	}
2367 	brelse(bp);
2368 	return (!free);
2369 }
2370 #endif /* INVARIANTS */
2371 
2372 /*
2373  * Free an inode.
2374  */
2375 int
2376 ffs_vfree(pvp, ino, mode)
2377 	struct vnode *pvp;
2378 	ino_t ino;
2379 	int mode;
2380 {
2381 	struct inode *ip;
2382 
2383 	if (DOINGSOFTDEP(pvp)) {
2384 		softdep_freefile(pvp, ino, mode);
2385 		return (0);
2386 	}
2387 	ip = VTOI(pvp);
2388 	return (ffs_freefile(ip->i_ump, ip->i_fs, ip->i_devvp, ino, mode,
2389 	    NULL));
2390 }
2391 
2392 /*
2393  * Do the actual free operation.
2394  * The specified inode is placed back in the free map.
2395  */
2396 int
2397 ffs_freefile(ump, fs, devvp, ino, mode, wkhd)
2398 	struct ufsmount *ump;
2399 	struct fs *fs;
2400 	struct vnode *devvp;
2401 	ino_t ino;
2402 	int mode;
2403 	struct workhead *wkhd;
2404 {
2405 	struct cg *cgp;
2406 	struct buf *bp;
2407 	ufs2_daddr_t cgbno;
2408 	int error;
2409 	u_int cg;
2410 	u_int8_t *inosused;
2411 	struct cdev *dev;
2412 
2413 	cg = ino_to_cg(fs, ino);
2414 	if (devvp->v_type == VREG) {
2415 		/* devvp is a snapshot */
2416 		dev = VTOI(devvp)->i_devvp->v_rdev;
2417 		cgbno = fragstoblks(fs, cgtod(fs, cg));
2418 	} else {
2419 		/* devvp is a normal disk device */
2420 		dev = devvp->v_rdev;
2421 		cgbno = fsbtodb(fs, cgtod(fs, cg));
2422 	}
2423 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2424 		panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s",
2425 		    devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt);
2426 	if ((error = bread(devvp, cgbno, (int)fs->fs_cgsize, NOCRED, &bp))) {
2427 		brelse(bp);
2428 		return (error);
2429 	}
2430 	cgp = (struct cg *)bp->b_data;
2431 	if (!cg_chkmagic(cgp)) {
2432 		brelse(bp);
2433 		return (0);
2434 	}
2435 	bp->b_xflags |= BX_BKGRDWRITE;
2436 	cgp->cg_old_time = cgp->cg_time = time_second;
2437 	inosused = cg_inosused(cgp);
2438 	ino %= fs->fs_ipg;
2439 	if (isclr(inosused, ino)) {
2440 		printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev),
2441 		    (uintmax_t)(ino + cg * fs->fs_ipg), fs->fs_fsmnt);
2442 		if (fs->fs_ronly == 0)
2443 			panic("ffs_freefile: freeing free inode");
2444 	}
2445 	clrbit(inosused, ino);
2446 	if (ino < cgp->cg_irotor)
2447 		cgp->cg_irotor = ino;
2448 	cgp->cg_cs.cs_nifree++;
2449 	UFS_LOCK(ump);
2450 	fs->fs_cstotal.cs_nifree++;
2451 	fs->fs_cs(fs, cg).cs_nifree++;
2452 	if ((mode & IFMT) == IFDIR) {
2453 		cgp->cg_cs.cs_ndir--;
2454 		fs->fs_cstotal.cs_ndir--;
2455 		fs->fs_cs(fs, cg).cs_ndir--;
2456 	}
2457 	fs->fs_fmod = 1;
2458 	ACTIVECLEAR(fs, cg);
2459 	UFS_UNLOCK(ump);
2460 	if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type != VREG)
2461 		softdep_setup_inofree(UFSTOVFS(ump), bp,
2462 		    ino + cg * fs->fs_ipg, wkhd);
2463 	bdwrite(bp);
2464 	return (0);
2465 }
2466 
2467 /*
2468  * Check to see if a file is free.
2469  */
2470 int
2471 ffs_checkfreefile(fs, devvp, ino)
2472 	struct fs *fs;
2473 	struct vnode *devvp;
2474 	ino_t ino;
2475 {
2476 	struct cg *cgp;
2477 	struct buf *bp;
2478 	ufs2_daddr_t cgbno;
2479 	int ret;
2480 	u_int cg;
2481 	u_int8_t *inosused;
2482 
2483 	cg = ino_to_cg(fs, ino);
2484 	if (devvp->v_type == VREG) {
2485 		/* devvp is a snapshot */
2486 		cgbno = fragstoblks(fs, cgtod(fs, cg));
2487 	} else {
2488 		/* devvp is a normal disk device */
2489 		cgbno = fsbtodb(fs, cgtod(fs, cg));
2490 	}
2491 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2492 		return (1);
2493 	if (bread(devvp, cgbno, (int)fs->fs_cgsize, NOCRED, &bp)) {
2494 		brelse(bp);
2495 		return (1);
2496 	}
2497 	cgp = (struct cg *)bp->b_data;
2498 	if (!cg_chkmagic(cgp)) {
2499 		brelse(bp);
2500 		return (1);
2501 	}
2502 	inosused = cg_inosused(cgp);
2503 	ino %= fs->fs_ipg;
2504 	ret = isclr(inosused, ino);
2505 	brelse(bp);
2506 	return (ret);
2507 }
2508 
2509 /*
2510  * Find a block of the specified size in the specified cylinder group.
2511  *
2512  * It is a panic if a request is made to find a block if none are
2513  * available.
2514  */
2515 static ufs1_daddr_t
2516 ffs_mapsearch(fs, cgp, bpref, allocsiz)
2517 	struct fs *fs;
2518 	struct cg *cgp;
2519 	ufs2_daddr_t bpref;
2520 	int allocsiz;
2521 {
2522 	ufs1_daddr_t bno;
2523 	int start, len, loc, i;
2524 	int blk, field, subfield, pos;
2525 	u_int8_t *blksfree;
2526 
2527 	/*
2528 	 * find the fragment by searching through the free block
2529 	 * map for an appropriate bit pattern
2530 	 */
2531 	if (bpref)
2532 		start = dtogd(fs, bpref) / NBBY;
2533 	else
2534 		start = cgp->cg_frotor / NBBY;
2535 	blksfree = cg_blksfree(cgp);
2536 	len = howmany(fs->fs_fpg, NBBY) - start;
2537 	loc = scanc((u_int)len, (u_char *)&blksfree[start],
2538 		fragtbl[fs->fs_frag],
2539 		(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2540 	if (loc == 0) {
2541 		len = start + 1;
2542 		start = 0;
2543 		loc = scanc((u_int)len, (u_char *)&blksfree[0],
2544 			fragtbl[fs->fs_frag],
2545 			(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2546 		if (loc == 0) {
2547 			printf("start = %d, len = %d, fs = %s\n",
2548 			    start, len, fs->fs_fsmnt);
2549 			panic("ffs_alloccg: map corrupted");
2550 			/* NOTREACHED */
2551 		}
2552 	}
2553 	bno = (start + len - loc) * NBBY;
2554 	cgp->cg_frotor = bno;
2555 	/*
2556 	 * found the byte in the map
2557 	 * sift through the bits to find the selected frag
2558 	 */
2559 	for (i = bno + NBBY; bno < i; bno += fs->fs_frag) {
2560 		blk = blkmap(fs, blksfree, bno);
2561 		blk <<= 1;
2562 		field = around[allocsiz];
2563 		subfield = inside[allocsiz];
2564 		for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) {
2565 			if ((blk & field) == subfield)
2566 				return (bno + pos);
2567 			field <<= 1;
2568 			subfield <<= 1;
2569 		}
2570 	}
2571 	printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt);
2572 	panic("ffs_alloccg: block not in map");
2573 	return (-1);
2574 }
2575 
2576 /*
2577  * Fserr prints the name of a filesystem with an error diagnostic.
2578  *
2579  * The form of the error message is:
2580  *	fs: error message
2581  */
2582 void
2583 ffs_fserr(fs, inum, cp)
2584 	struct fs *fs;
2585 	ino_t inum;
2586 	char *cp;
2587 {
2588 	struct thread *td = curthread;	/* XXX */
2589 	struct proc *p = td->td_proc;
2590 
2591 	log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n",
2592 	    p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum,
2593 	    fs->fs_fsmnt, cp);
2594 }
2595 
2596 /*
2597  * This function provides the capability for the fsck program to
2598  * update an active filesystem. Fourteen operations are provided:
2599  *
2600  * adjrefcnt(inode, amt) - adjusts the reference count on the
2601  *	specified inode by the specified amount. Under normal
2602  *	operation the count should always go down. Decrementing
2603  *	the count to zero will cause the inode to be freed.
2604  * adjblkcnt(inode, amt) - adjust the number of blocks used by the
2605  *	inode by the specified amount.
2606  * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) -
2607  *	adjust the superblock summary.
2608  * freedirs(inode, count) - directory inodes [inode..inode + count - 1]
2609  *	are marked as free. Inodes should never have to be marked
2610  *	as in use.
2611  * freefiles(inode, count) - file inodes [inode..inode + count - 1]
2612  *	are marked as free. Inodes should never have to be marked
2613  *	as in use.
2614  * freeblks(blockno, size) - blocks [blockno..blockno + size - 1]
2615  *	are marked as free. Blocks should never have to be marked
2616  *	as in use.
2617  * setflags(flags, set/clear) - the fs_flags field has the specified
2618  *	flags set (second parameter +1) or cleared (second parameter -1).
2619  * setcwd(dirinode) - set the current directory to dirinode in the
2620  *	filesystem associated with the snapshot.
2621  * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".."
2622  *	in the current directory is oldvalue then change it to newvalue.
2623  * unlink(nameptr, oldvalue) - Verify that the inode number associated
2624  *	with nameptr in the current directory is oldvalue then unlink it.
2625  *
2626  * The following functions may only be used on a quiescent filesystem
2627  * by the soft updates journal. They are not safe to be run on an active
2628  * filesystem.
2629  *
2630  * setinode(inode, dip) - the specified disk inode is replaced with the
2631  *	contents pointed to by dip.
2632  * setbufoutput(fd, flags) - output associated with the specified file
2633  *	descriptor (which must reference the character device supporting
2634  *	the filesystem) switches from using physio to running through the
2635  *	buffer cache when flags is set to 1. The descriptor reverts to
2636  *	physio for output when flags is set to zero.
2637  */
2638 
2639 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS);
2640 
2641 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt, CTLFLAG_WR|CTLTYPE_STRUCT,
2642 	0, 0, sysctl_ffs_fsck, "S,fsck", "Adjust Inode Reference Count");
2643 
2644 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt, CTLFLAG_WR,
2645 	sysctl_ffs_fsck, "Adjust Inode Used Blocks Count");
2646 
2647 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir, CTLFLAG_WR,
2648 	sysctl_ffs_fsck, "Adjust number of directories");
2649 
2650 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree, CTLFLAG_WR,
2651 	sysctl_ffs_fsck, "Adjust number of free blocks");
2652 
2653 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree, CTLFLAG_WR,
2654 	sysctl_ffs_fsck, "Adjust number of free inodes");
2655 
2656 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree, CTLFLAG_WR,
2657 	sysctl_ffs_fsck, "Adjust number of free frags");
2658 
2659 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters, CTLFLAG_WR,
2660 	sysctl_ffs_fsck, "Adjust number of free clusters");
2661 
2662 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs, CTLFLAG_WR,
2663 	sysctl_ffs_fsck, "Free Range of Directory Inodes");
2664 
2665 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles, CTLFLAG_WR,
2666 	sysctl_ffs_fsck, "Free Range of File Inodes");
2667 
2668 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks, CTLFLAG_WR,
2669 	sysctl_ffs_fsck, "Free Range of Blocks");
2670 
2671 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags, CTLFLAG_WR,
2672 	sysctl_ffs_fsck, "Change Filesystem Flags");
2673 
2674 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd, CTLFLAG_WR,
2675 	sysctl_ffs_fsck, "Set Current Working Directory");
2676 
2677 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot, CTLFLAG_WR,
2678 	sysctl_ffs_fsck, "Change Value of .. Entry");
2679 
2680 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink, CTLFLAG_WR,
2681 	sysctl_ffs_fsck, "Unlink a Duplicate Name");
2682 
2683 static SYSCTL_NODE(_vfs_ffs, FFS_SET_INODE, setinode, CTLFLAG_WR,
2684 	sysctl_ffs_fsck, "Update an On-Disk Inode");
2685 
2686 static SYSCTL_NODE(_vfs_ffs, FFS_SET_BUFOUTPUT, setbufoutput, CTLFLAG_WR,
2687 	sysctl_ffs_fsck, "Set Buffered Writing for Descriptor");
2688 
2689 #define DEBUG 1
2690 #ifdef DEBUG
2691 static int fsckcmds = 0;
2692 SYSCTL_INT(_debug, OID_AUTO, fsckcmds, CTLFLAG_RW, &fsckcmds, 0, "");
2693 #endif /* DEBUG */
2694 
2695 static int buffered_write(struct file *, struct uio *, struct ucred *,
2696 	int, struct thread *);
2697 
2698 static int
2699 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS)
2700 {
2701 	struct thread *td = curthread;
2702 	struct fsck_cmd cmd;
2703 	struct ufsmount *ump;
2704 	struct vnode *vp, *vpold, *dvp, *fdvp;
2705 	struct inode *ip, *dp;
2706 	struct mount *mp;
2707 	struct fs *fs;
2708 	ufs2_daddr_t blkno;
2709 	long blkcnt, blksize;
2710 	struct filedesc *fdp;
2711 	struct file *fp, *vfp;
2712 	cap_rights_t rights;
2713 	int filetype, error;
2714 	static struct fileops *origops, bufferedops;
2715 
2716 	if (req->newlen > sizeof cmd)
2717 		return (EBADRPC);
2718 	if ((error = SYSCTL_IN(req, &cmd, sizeof cmd)) != 0)
2719 		return (error);
2720 	if (cmd.version != FFS_CMD_VERSION)
2721 		return (ERPCMISMATCH);
2722 	if ((error = getvnode(td->td_proc->p_fd, cmd.handle,
2723 	    cap_rights_init(&rights, CAP_FSCK), &fp)) != 0)
2724 		return (error);
2725 	vp = fp->f_data;
2726 	if (vp->v_type != VREG && vp->v_type != VDIR) {
2727 		fdrop(fp, td);
2728 		return (EINVAL);
2729 	}
2730 	vn_start_write(vp, &mp, V_WAIT);
2731 	if (mp == 0 || strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) {
2732 		vn_finished_write(mp);
2733 		fdrop(fp, td);
2734 		return (EINVAL);
2735 	}
2736 	ump = VFSTOUFS(mp);
2737 	if ((mp->mnt_flag & MNT_RDONLY) &&
2738 	    ump->um_fsckpid != td->td_proc->p_pid) {
2739 		vn_finished_write(mp);
2740 		fdrop(fp, td);
2741 		return (EROFS);
2742 	}
2743 	fs = ump->um_fs;
2744 	filetype = IFREG;
2745 
2746 	switch (oidp->oid_number) {
2747 
2748 	case FFS_SET_FLAGS:
2749 #ifdef DEBUG
2750 		if (fsckcmds)
2751 			printf("%s: %s flags\n", mp->mnt_stat.f_mntonname,
2752 			    cmd.size > 0 ? "set" : "clear");
2753 #endif /* DEBUG */
2754 		if (cmd.size > 0)
2755 			fs->fs_flags |= (long)cmd.value;
2756 		else
2757 			fs->fs_flags &= ~(long)cmd.value;
2758 		break;
2759 
2760 	case FFS_ADJ_REFCNT:
2761 #ifdef DEBUG
2762 		if (fsckcmds) {
2763 			printf("%s: adjust inode %jd link count by %jd\n",
2764 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
2765 			    (intmax_t)cmd.size);
2766 		}
2767 #endif /* DEBUG */
2768 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
2769 			break;
2770 		ip = VTOI(vp);
2771 		ip->i_nlink += cmd.size;
2772 		DIP_SET(ip, i_nlink, ip->i_nlink);
2773 		ip->i_effnlink += cmd.size;
2774 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
2775 		error = ffs_update(vp, 1);
2776 		if (DOINGSOFTDEP(vp))
2777 			softdep_change_linkcnt(ip);
2778 		vput(vp);
2779 		break;
2780 
2781 	case FFS_ADJ_BLKCNT:
2782 #ifdef DEBUG
2783 		if (fsckcmds) {
2784 			printf("%s: adjust inode %jd block count by %jd\n",
2785 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
2786 			    (intmax_t)cmd.size);
2787 		}
2788 #endif /* DEBUG */
2789 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
2790 			break;
2791 		ip = VTOI(vp);
2792 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size);
2793 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
2794 		error = ffs_update(vp, 1);
2795 		vput(vp);
2796 		break;
2797 
2798 	case FFS_DIR_FREE:
2799 		filetype = IFDIR;
2800 		/* fall through */
2801 
2802 	case FFS_FILE_FREE:
2803 #ifdef DEBUG
2804 		if (fsckcmds) {
2805 			if (cmd.size == 1)
2806 				printf("%s: free %s inode %ju\n",
2807 				    mp->mnt_stat.f_mntonname,
2808 				    filetype == IFDIR ? "directory" : "file",
2809 				    (uintmax_t)cmd.value);
2810 			else
2811 				printf("%s: free %s inodes %ju-%ju\n",
2812 				    mp->mnt_stat.f_mntonname,
2813 				    filetype == IFDIR ? "directory" : "file",
2814 				    (uintmax_t)cmd.value,
2815 				    (uintmax_t)(cmd.value + cmd.size - 1));
2816 		}
2817 #endif /* DEBUG */
2818 		while (cmd.size > 0) {
2819 			if ((error = ffs_freefile(ump, fs, ump->um_devvp,
2820 			    cmd.value, filetype, NULL)))
2821 				break;
2822 			cmd.size -= 1;
2823 			cmd.value += 1;
2824 		}
2825 		break;
2826 
2827 	case FFS_BLK_FREE:
2828 #ifdef DEBUG
2829 		if (fsckcmds) {
2830 			if (cmd.size == 1)
2831 				printf("%s: free block %jd\n",
2832 				    mp->mnt_stat.f_mntonname,
2833 				    (intmax_t)cmd.value);
2834 			else
2835 				printf("%s: free blocks %jd-%jd\n",
2836 				    mp->mnt_stat.f_mntonname,
2837 				    (intmax_t)cmd.value,
2838 				    (intmax_t)cmd.value + cmd.size - 1);
2839 		}
2840 #endif /* DEBUG */
2841 		blkno = cmd.value;
2842 		blkcnt = cmd.size;
2843 		blksize = fs->fs_frag - (blkno % fs->fs_frag);
2844 		while (blkcnt > 0) {
2845 			if (blksize > blkcnt)
2846 				blksize = blkcnt;
2847 			ffs_blkfree(ump, fs, ump->um_devvp, blkno,
2848 			    blksize * fs->fs_fsize, ROOTINO, VDIR, NULL);
2849 			blkno += blksize;
2850 			blkcnt -= blksize;
2851 			blksize = fs->fs_frag;
2852 		}
2853 		break;
2854 
2855 	/*
2856 	 * Adjust superblock summaries.  fsck(8) is expected to
2857 	 * submit deltas when necessary.
2858 	 */
2859 	case FFS_ADJ_NDIR:
2860 #ifdef DEBUG
2861 		if (fsckcmds) {
2862 			printf("%s: adjust number of directories by %jd\n",
2863 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2864 		}
2865 #endif /* DEBUG */
2866 		fs->fs_cstotal.cs_ndir += cmd.value;
2867 		break;
2868 
2869 	case FFS_ADJ_NBFREE:
2870 #ifdef DEBUG
2871 		if (fsckcmds) {
2872 			printf("%s: adjust number of free blocks by %+jd\n",
2873 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2874 		}
2875 #endif /* DEBUG */
2876 		fs->fs_cstotal.cs_nbfree += cmd.value;
2877 		break;
2878 
2879 	case FFS_ADJ_NIFREE:
2880 #ifdef DEBUG
2881 		if (fsckcmds) {
2882 			printf("%s: adjust number of free inodes by %+jd\n",
2883 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2884 		}
2885 #endif /* DEBUG */
2886 		fs->fs_cstotal.cs_nifree += cmd.value;
2887 		break;
2888 
2889 	case FFS_ADJ_NFFREE:
2890 #ifdef DEBUG
2891 		if (fsckcmds) {
2892 			printf("%s: adjust number of free frags by %+jd\n",
2893 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2894 		}
2895 #endif /* DEBUG */
2896 		fs->fs_cstotal.cs_nffree += cmd.value;
2897 		break;
2898 
2899 	case FFS_ADJ_NUMCLUSTERS:
2900 #ifdef DEBUG
2901 		if (fsckcmds) {
2902 			printf("%s: adjust number of free clusters by %+jd\n",
2903 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2904 		}
2905 #endif /* DEBUG */
2906 		fs->fs_cstotal.cs_numclusters += cmd.value;
2907 		break;
2908 
2909 	case FFS_SET_CWD:
2910 #ifdef DEBUG
2911 		if (fsckcmds) {
2912 			printf("%s: set current directory to inode %jd\n",
2913 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2914 		}
2915 #endif /* DEBUG */
2916 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp)))
2917 			break;
2918 		AUDIT_ARG_VNODE1(vp);
2919 		if ((error = change_dir(vp, td)) != 0) {
2920 			vput(vp);
2921 			break;
2922 		}
2923 		VOP_UNLOCK(vp, 0);
2924 		fdp = td->td_proc->p_fd;
2925 		FILEDESC_XLOCK(fdp);
2926 		vpold = fdp->fd_cdir;
2927 		fdp->fd_cdir = vp;
2928 		FILEDESC_XUNLOCK(fdp);
2929 		vrele(vpold);
2930 		break;
2931 
2932 	case FFS_SET_DOTDOT:
2933 #ifdef DEBUG
2934 		if (fsckcmds) {
2935 			printf("%s: change .. in cwd from %jd to %jd\n",
2936 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
2937 			    (intmax_t)cmd.size);
2938 		}
2939 #endif /* DEBUG */
2940 		/*
2941 		 * First we have to get and lock the parent directory
2942 		 * to which ".." points.
2943 		 */
2944 		error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp);
2945 		if (error)
2946 			break;
2947 		/*
2948 		 * Now we get and lock the child directory containing "..".
2949 		 */
2950 		FILEDESC_SLOCK(td->td_proc->p_fd);
2951 		dvp = td->td_proc->p_fd->fd_cdir;
2952 		FILEDESC_SUNLOCK(td->td_proc->p_fd);
2953 		if ((error = vget(dvp, LK_EXCLUSIVE, td)) != 0) {
2954 			vput(fdvp);
2955 			break;
2956 		}
2957 		dp = VTOI(dvp);
2958 		dp->i_offset = 12;	/* XXX mastertemplate.dot_reclen */
2959 		error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size,
2960 		    DT_DIR, 0);
2961 		cache_purge(fdvp);
2962 		cache_purge(dvp);
2963 		vput(dvp);
2964 		vput(fdvp);
2965 		break;
2966 
2967 	case FFS_UNLINK:
2968 #ifdef DEBUG
2969 		if (fsckcmds) {
2970 			char buf[32];
2971 
2972 			if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL))
2973 				strncpy(buf, "Name_too_long", 32);
2974 			printf("%s: unlink %s (inode %jd)\n",
2975 			    mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size);
2976 		}
2977 #endif /* DEBUG */
2978 		/*
2979 		 * kern_unlinkat will do its own start/finish writes and
2980 		 * they do not nest, so drop ours here. Setting mp == NULL
2981 		 * indicates that vn_finished_write is not needed down below.
2982 		 */
2983 		vn_finished_write(mp);
2984 		mp = NULL;
2985 		error = kern_unlinkat(td, AT_FDCWD, (char *)(intptr_t)cmd.value,
2986 		    UIO_USERSPACE, (ino_t)cmd.size);
2987 		break;
2988 
2989 	case FFS_SET_INODE:
2990 		if (ump->um_fsckpid != td->td_proc->p_pid) {
2991 			error = EPERM;
2992 			break;
2993 		}
2994 #ifdef DEBUG
2995 		if (fsckcmds) {
2996 			printf("%s: update inode %jd\n",
2997 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
2998 		}
2999 #endif /* DEBUG */
3000 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3001 			break;
3002 		AUDIT_ARG_VNODE1(vp);
3003 		ip = VTOI(vp);
3004 		if (ip->i_ump->um_fstype == UFS1)
3005 			error = copyin((void *)(intptr_t)cmd.size, ip->i_din1,
3006 			    sizeof(struct ufs1_dinode));
3007 		else
3008 			error = copyin((void *)(intptr_t)cmd.size, ip->i_din2,
3009 			    sizeof(struct ufs2_dinode));
3010 		if (error) {
3011 			vput(vp);
3012 			break;
3013 		}
3014 		ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3015 		error = ffs_update(vp, 1);
3016 		vput(vp);
3017 		break;
3018 
3019 	case FFS_SET_BUFOUTPUT:
3020 		if (ump->um_fsckpid != td->td_proc->p_pid) {
3021 			error = EPERM;
3022 			break;
3023 		}
3024 		if (VTOI(vp)->i_ump != ump) {
3025 			error = EINVAL;
3026 			break;
3027 		}
3028 #ifdef DEBUG
3029 		if (fsckcmds) {
3030 			printf("%s: %s buffered output for descriptor %jd\n",
3031 			    mp->mnt_stat.f_mntonname,
3032 			    cmd.size == 1 ? "enable" : "disable",
3033 			    (intmax_t)cmd.value);
3034 		}
3035 #endif /* DEBUG */
3036 		if ((error = getvnode(td->td_proc->p_fd, cmd.value,
3037 		    cap_rights_init(&rights, CAP_FSCK), &vfp)) != 0)
3038 			break;
3039 		if (vfp->f_vnode->v_type != VCHR) {
3040 			fdrop(vfp, td);
3041 			error = EINVAL;
3042 			break;
3043 		}
3044 		if (origops == NULL) {
3045 			origops = vfp->f_ops;
3046 			bcopy((void *)origops, (void *)&bufferedops,
3047 			    sizeof(bufferedops));
3048 			bufferedops.fo_write = buffered_write;
3049 		}
3050 		if (cmd.size == 1)
3051 			atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3052 			    (uintptr_t)&bufferedops);
3053 		else
3054 			atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3055 			    (uintptr_t)origops);
3056 		fdrop(vfp, td);
3057 		break;
3058 
3059 	default:
3060 #ifdef DEBUG
3061 		if (fsckcmds) {
3062 			printf("Invalid request %d from fsck\n",
3063 			    oidp->oid_number);
3064 		}
3065 #endif /* DEBUG */
3066 		error = EINVAL;
3067 		break;
3068 
3069 	}
3070 	fdrop(fp, td);
3071 	vn_finished_write(mp);
3072 	return (error);
3073 }
3074 
3075 /*
3076  * Function to switch a descriptor to use the buffer cache to stage
3077  * its I/O. This is needed so that writes to the filesystem device
3078  * will give snapshots a chance to copy modified blocks for which it
3079  * needs to retain copies.
3080  */
3081 static int
3082 buffered_write(fp, uio, active_cred, flags, td)
3083 	struct file *fp;
3084 	struct uio *uio;
3085 	struct ucred *active_cred;
3086 	int flags;
3087 	struct thread *td;
3088 {
3089 	struct vnode *devvp, *vp;
3090 	struct inode *ip;
3091 	struct buf *bp;
3092 	struct fs *fs;
3093 	struct filedesc *fdp;
3094 	int error;
3095 	daddr_t lbn;
3096 
3097 	/*
3098 	 * The devvp is associated with the /dev filesystem. To discover
3099 	 * the filesystem with which the device is associated, we depend
3100 	 * on the application setting the current directory to a location
3101 	 * within the filesystem being written. Yes, this is an ugly hack.
3102 	 */
3103 	devvp = fp->f_vnode;
3104 	if (!vn_isdisk(devvp, NULL))
3105 		return (EINVAL);
3106 	fdp = td->td_proc->p_fd;
3107 	FILEDESC_SLOCK(fdp);
3108 	vp = fdp->fd_cdir;
3109 	vref(vp);
3110 	FILEDESC_SUNLOCK(fdp);
3111 	vn_lock(vp, LK_SHARED | LK_RETRY);
3112 	/*
3113 	 * Check that the current directory vnode indeed belongs to
3114 	 * UFS before trying to dereference UFS-specific v_data fields.
3115 	 */
3116 	if (vp->v_op != &ffs_vnodeops1 && vp->v_op != &ffs_vnodeops2) {
3117 		vput(vp);
3118 		return (EINVAL);
3119 	}
3120 	ip = VTOI(vp);
3121 	if (ip->i_devvp != devvp) {
3122 		vput(vp);
3123 		return (EINVAL);
3124 	}
3125 	fs = ip->i_fs;
3126 	vput(vp);
3127 	foffset_lock_uio(fp, uio, flags);
3128 	vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
3129 #ifdef DEBUG
3130 	if (fsckcmds) {
3131 		printf("%s: buffered write for block %jd\n",
3132 		    fs->fs_fsmnt, (intmax_t)btodb(uio->uio_offset));
3133 	}
3134 #endif /* DEBUG */
3135 	/*
3136 	 * All I/O must be contained within a filesystem block, start on
3137 	 * a fragment boundary, and be a multiple of fragments in length.
3138 	 */
3139 	if (uio->uio_resid > fs->fs_bsize - (uio->uio_offset % fs->fs_bsize) ||
3140 	    fragoff(fs, uio->uio_offset) != 0 ||
3141 	    fragoff(fs, uio->uio_resid) != 0) {
3142 		error = EINVAL;
3143 		goto out;
3144 	}
3145 	lbn = numfrags(fs, uio->uio_offset);
3146 	bp = getblk(devvp, lbn, uio->uio_resid, 0, 0, 0);
3147 	bp->b_flags |= B_RELBUF;
3148 	if ((error = uiomove((char *)bp->b_data, uio->uio_resid, uio)) != 0) {
3149 		brelse(bp);
3150 		goto out;
3151 	}
3152 	error = bwrite(bp);
3153 out:
3154 	VOP_UNLOCK(devvp, 0);
3155 	foffset_unlock_uio(fp, uio, flags | FOF_NEXTOFF);
3156 	return (error);
3157 }
3158