xref: /freebsd/sys/ufs/ffs/ffs_alloc.c (revision c3c73b4f0a91d2806e1a632b75f769fb4fa89576)
1 /*-
2  * SPDX-License-Identifier: (BSD-2-Clause-FreeBSD AND BSD-3-Clause)
3  *
4  * Copyright (c) 2002 Networks Associates Technology, Inc.
5  * All rights reserved.
6  *
7  * This software was developed for the FreeBSD Project by Marshall
8  * Kirk McKusick and Network Associates Laboratories, the Security
9  * Research Division of Network Associates, Inc. under DARPA/SPAWAR
10  * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS
11  * research program
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1982, 1986, 1989, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  *
37  * Redistribution and use in source and binary forms, with or without
38  * modification, are permitted provided that the following conditions
39  * are met:
40  * 1. Redistributions of source code must retain the above copyright
41  *    notice, this list of conditions and the following disclaimer.
42  * 2. Redistributions in binary form must reproduce the above copyright
43  *    notice, this list of conditions and the following disclaimer in the
44  *    documentation and/or other materials provided with the distribution.
45  * 3. Neither the name of the University nor the names of its contributors
46  *    may be used to endorse or promote products derived from this software
47  *    without specific prior written permission.
48  *
49  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
50  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
51  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
52  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
53  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
54  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
55  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
56  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
57  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
58  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
59  * SUCH DAMAGE.
60  *
61  *	@(#)ffs_alloc.c	8.18 (Berkeley) 5/26/95
62  */
63 
64 #include <sys/cdefs.h>
65 __FBSDID("$FreeBSD$");
66 
67 #include "opt_quota.h"
68 
69 #include <sys/param.h>
70 #include <sys/capsicum.h>
71 #include <sys/gsb_crc32.h>
72 #include <sys/systm.h>
73 #include <sys/bio.h>
74 #include <sys/buf.h>
75 #include <sys/conf.h>
76 #include <sys/fcntl.h>
77 #include <sys/file.h>
78 #include <sys/filedesc.h>
79 #include <sys/priv.h>
80 #include <sys/proc.h>
81 #include <sys/vnode.h>
82 #include <sys/mount.h>
83 #include <sys/kernel.h>
84 #include <sys/syscallsubr.h>
85 #include <sys/sysctl.h>
86 #include <sys/syslog.h>
87 #include <sys/taskqueue.h>
88 
89 #include <security/audit/audit.h>
90 
91 #include <geom/geom.h>
92 #include <geom/geom_vfs.h>
93 
94 #include <ufs/ufs/dir.h>
95 #include <ufs/ufs/extattr.h>
96 #include <ufs/ufs/quota.h>
97 #include <ufs/ufs/inode.h>
98 #include <ufs/ufs/ufs_extern.h>
99 #include <ufs/ufs/ufsmount.h>
100 
101 #include <ufs/ffs/fs.h>
102 #include <ufs/ffs/ffs_extern.h>
103 #include <ufs/ffs/softdep.h>
104 
105 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref,
106 				  int size, int rsize);
107 
108 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int);
109 static ufs2_daddr_t
110 	      ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int);
111 static void	ffs_blkfree_cg(struct ufsmount *, struct fs *,
112 		    struct vnode *, ufs2_daddr_t, long, ino_t,
113 		    struct workhead *);
114 #ifdef INVARIANTS
115 static int	ffs_checkblk(struct inode *, ufs2_daddr_t, long);
116 #endif
117 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int);
118 static ino_t	ffs_dirpref(struct inode *);
119 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t,
120 		    int, int);
121 static ufs2_daddr_t	ffs_hashalloc
122 		(struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *);
123 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int,
124 		    int);
125 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int);
126 static int	ffs_reallocblks_ufs1(struct vop_reallocblks_args *);
127 static int	ffs_reallocblks_ufs2(struct vop_reallocblks_args *);
128 static void	ffs_ckhash_cg(struct buf *);
129 
130 /*
131  * Allocate a block in the filesystem.
132  *
133  * The size of the requested block is given, which must be some
134  * multiple of fs_fsize and <= fs_bsize.
135  * A preference may be optionally specified. If a preference is given
136  * the following hierarchy is used to allocate a block:
137  *   1) allocate the requested block.
138  *   2) allocate a rotationally optimal block in the same cylinder.
139  *   3) allocate a block in the same cylinder group.
140  *   4) quadradically rehash into other cylinder groups, until an
141  *      available block is located.
142  * If no block preference is given the following hierarchy is used
143  * to allocate a block:
144  *   1) allocate a block in the cylinder group that contains the
145  *      inode for the file.
146  *   2) quadradically rehash into other cylinder groups, until an
147  *      available block is located.
148  */
149 int
150 ffs_alloc(ip, lbn, bpref, size, flags, cred, bnp)
151 	struct inode *ip;
152 	ufs2_daddr_t lbn, bpref;
153 	int size, flags;
154 	struct ucred *cred;
155 	ufs2_daddr_t *bnp;
156 {
157 	struct fs *fs;
158 	struct ufsmount *ump;
159 	ufs2_daddr_t bno;
160 	u_int cg, reclaimed;
161 	int64_t delta;
162 #ifdef QUOTA
163 	int error;
164 #endif
165 
166 	*bnp = 0;
167 	ump = ITOUMP(ip);
168 	fs = ump->um_fs;
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(ump->um_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) &&
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 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
206 		else
207 			UFS_INODE_SET_FLAG(ip, 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 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
226 		UFS_UNLOCK(ump);
227 		return (ENXIO);
228 	}
229 	if (reclaimed > 0 &&
230 	    ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
231 		UFS_UNLOCK(ump);
232 		ffs_fserr(fs, ip->i_number, "filesystem full");
233 		uprintf("\n%s: write failed, filesystem is full\n",
234 		    fs->fs_fsmnt);
235 	} else {
236 		UFS_UNLOCK(ump);
237 	}
238 	return (ENOSPC);
239 }
240 
241 /*
242  * Reallocate a fragment to a bigger size
243  *
244  * The number and size of the old block is given, and a preference
245  * and new size is also specified. The allocator attempts to extend
246  * the original block. Failing that, the regular block allocator is
247  * invoked to get an appropriate block.
248  */
249 int
250 ffs_realloccg(ip, lbprev, bprev, bpref, osize, nsize, flags, cred, bpp)
251 	struct inode *ip;
252 	ufs2_daddr_t lbprev;
253 	ufs2_daddr_t bprev;
254 	ufs2_daddr_t bpref;
255 	int osize, nsize, flags;
256 	struct ucred *cred;
257 	struct buf **bpp;
258 {
259 	struct vnode *vp;
260 	struct fs *fs;
261 	struct buf *bp;
262 	struct ufsmount *ump;
263 	u_int cg, request, reclaimed;
264 	int error, gbflags;
265 	ufs2_daddr_t bno;
266 	int64_t delta;
267 
268 	vp = ITOV(ip);
269 	ump = ITOUMP(ip);
270 	fs = ump->um_fs;
271 	bp = NULL;
272 	gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
273 
274 	mtx_assert(UFS_MTX(ump), MA_OWNED);
275 #ifdef INVARIANTS
276 	if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
277 		panic("ffs_realloccg: allocation on suspended filesystem");
278 	if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 ||
279 	    (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) {
280 		printf(
281 		"dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n",
282 		    devtoname(ump->um_dev), (long)fs->fs_bsize, osize,
283 		    nsize, fs->fs_fsmnt);
284 		panic("ffs_realloccg: bad size");
285 	}
286 	if (cred == NOCRED)
287 		panic("ffs_realloccg: missing credential");
288 #endif /* INVARIANTS */
289 	reclaimed = 0;
290 retry:
291 	if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) &&
292 	    freespace(fs, fs->fs_minfree) -  numfrags(fs, nsize - osize) < 0) {
293 		goto nospace;
294 	}
295 	if (bprev == 0) {
296 		printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n",
297 		    devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev,
298 		    fs->fs_fsmnt);
299 		panic("ffs_realloccg: bad bprev");
300 	}
301 	UFS_UNLOCK(ump);
302 	/*
303 	 * Allocate the extra space in the buffer.
304 	 */
305 	error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp);
306 	if (error) {
307 		return (error);
308 	}
309 
310 	if (bp->b_blkno == bp->b_lblkno) {
311 		if (lbprev >= UFS_NDADDR)
312 			panic("ffs_realloccg: lbprev out of range");
313 		bp->b_blkno = fsbtodb(fs, bprev);
314 	}
315 
316 #ifdef QUOTA
317 	error = chkdq(ip, btodb(nsize - osize), cred, 0);
318 	if (error) {
319 		brelse(bp);
320 		return (error);
321 	}
322 #endif
323 	/*
324 	 * Check for extension in the existing location.
325 	 */
326 	*bpp = NULL;
327 	cg = dtog(fs, bprev);
328 	UFS_LOCK(ump);
329 	bno = ffs_fragextend(ip, cg, bprev, osize, nsize);
330 	if (bno) {
331 		if (bp->b_blkno != fsbtodb(fs, bno))
332 			panic("ffs_realloccg: bad blockno");
333 		delta = btodb(nsize - osize);
334 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
335 		if (flags & IO_EXT)
336 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
337 		else
338 			UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
339 		allocbuf(bp, nsize);
340 		bp->b_flags |= B_DONE;
341 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
342 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
343 			vfs_bio_set_valid(bp, osize, nsize - osize);
344 		*bpp = bp;
345 		return (0);
346 	}
347 	/*
348 	 * Allocate a new disk location.
349 	 */
350 	if (bpref >= fs->fs_size)
351 		bpref = 0;
352 	switch ((int)fs->fs_optim) {
353 	case FS_OPTSPACE:
354 		/*
355 		 * Allocate an exact sized fragment. Although this makes
356 		 * best use of space, we will waste time relocating it if
357 		 * the file continues to grow. If the fragmentation is
358 		 * less than half of the minimum free reserve, we choose
359 		 * to begin optimizing for time.
360 		 */
361 		request = nsize;
362 		if (fs->fs_minfree <= 5 ||
363 		    fs->fs_cstotal.cs_nffree >
364 		    (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100))
365 			break;
366 		log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n",
367 			fs->fs_fsmnt);
368 		fs->fs_optim = FS_OPTTIME;
369 		break;
370 	case FS_OPTTIME:
371 		/*
372 		 * At this point we have discovered a file that is trying to
373 		 * grow a small fragment to a larger fragment. To save time,
374 		 * we allocate a full sized block, then free the unused portion.
375 		 * If the file continues to grow, the `ffs_fragextend' call
376 		 * above will be able to grow it in place without further
377 		 * copying. If aberrant programs cause disk fragmentation to
378 		 * grow within 2% of the free reserve, we choose to begin
379 		 * optimizing for space.
380 		 */
381 		request = fs->fs_bsize;
382 		if (fs->fs_cstotal.cs_nffree <
383 		    (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100)
384 			break;
385 		log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n",
386 			fs->fs_fsmnt);
387 		fs->fs_optim = FS_OPTSPACE;
388 		break;
389 	default:
390 		printf("dev = %s, optim = %ld, fs = %s\n",
391 		    devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt);
392 		panic("ffs_realloccg: bad optim");
393 		/* NOTREACHED */
394 	}
395 	bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg);
396 	if (bno > 0) {
397 		bp->b_blkno = fsbtodb(fs, bno);
398 		if (!DOINGSOFTDEP(vp))
399 			/*
400 			 * The usual case is that a smaller fragment that
401 			 * was just allocated has been replaced with a bigger
402 			 * fragment or a full-size block. If it is marked as
403 			 * B_DELWRI, the current contents have not been written
404 			 * to disk. It is possible that the block was written
405 			 * earlier, but very uncommon. If the block has never
406 			 * been written, there is no need to send a BIO_DELETE
407 			 * for it when it is freed. The gain from avoiding the
408 			 * TRIMs for the common case of unwritten blocks far
409 			 * exceeds the cost of the write amplification for the
410 			 * uncommon case of failing to send a TRIM for a block
411 			 * that had been written.
412 			 */
413 			ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize,
414 			    ip->i_number, vp->v_type, NULL,
415 			    (bp->b_flags & B_DELWRI) != 0 ?
416 			    NOTRIM_KEY : SINGLETON_KEY);
417 		delta = btodb(nsize - osize);
418 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
419 		if (flags & IO_EXT)
420 			UFS_INODE_SET_FLAG(ip, IN_CHANGE);
421 		else
422 			UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
423 		allocbuf(bp, nsize);
424 		bp->b_flags |= B_DONE;
425 		vfs_bio_bzero_buf(bp, osize, nsize - osize);
426 		if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
427 			vfs_bio_set_valid(bp, osize, nsize - osize);
428 		*bpp = bp;
429 		return (0);
430 	}
431 #ifdef QUOTA
432 	UFS_UNLOCK(ump);
433 	/*
434 	 * Restore user's disk quota because allocation failed.
435 	 */
436 	(void) chkdq(ip, -btodb(nsize - osize), cred, FORCE);
437 	UFS_LOCK(ump);
438 #endif
439 nospace:
440 	/*
441 	 * no space available
442 	 */
443 	if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
444 		reclaimed = 1;
445 		UFS_UNLOCK(ump);
446 		if (bp) {
447 			brelse(bp);
448 			bp = NULL;
449 		}
450 		UFS_LOCK(ump);
451 		softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT);
452 		goto retry;
453 	}
454 	if (bp)
455 		brelse(bp);
456 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
457 		UFS_UNLOCK(ump);
458 		return (ENXIO);
459 	}
460 	if (reclaimed > 0 &&
461 	    ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
462 		UFS_UNLOCK(ump);
463 		ffs_fserr(fs, ip->i_number, "filesystem full");
464 		uprintf("\n%s: write failed, filesystem is full\n",
465 		    fs->fs_fsmnt);
466 	} else {
467 		UFS_UNLOCK(ump);
468 	}
469 	return (ENOSPC);
470 }
471 
472 /*
473  * Reallocate a sequence of blocks into a contiguous sequence of blocks.
474  *
475  * The vnode and an array of buffer pointers for a range of sequential
476  * logical blocks to be made contiguous is given. The allocator attempts
477  * to find a range of sequential blocks starting as close as possible
478  * from the end of the allocation for the logical block immediately
479  * preceding the current range. If successful, the physical block numbers
480  * in the buffer pointers and in the inode are changed to reflect the new
481  * allocation. If unsuccessful, the allocation is left unchanged. The
482  * success in doing the reallocation is returned. Note that the error
483  * return is not reflected back to the user. Rather the previous block
484  * allocation will be used.
485  */
486 
487 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
488     "FFS filesystem");
489 
490 static int doasyncfree = 1;
491 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0,
492 "do not force synchronous writes when blocks are reallocated");
493 
494 static int doreallocblks = 1;
495 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0,
496 "enable block reallocation");
497 
498 static int dotrimcons = 1;
499 SYSCTL_INT(_vfs_ffs, OID_AUTO, dotrimcons, CTLFLAG_RWTUN, &dotrimcons, 0,
500 "enable BIO_DELETE / TRIM consolidation");
501 
502 static int maxclustersearch = 10;
503 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch,
504 0, "max number of cylinder group to search for contigous blocks");
505 
506 #ifdef DIAGNOSTIC
507 static int prtrealloc = 0;
508 SYSCTL_INT(_debug, OID_AUTO, ffs_prtrealloc, CTLFLAG_RW, &prtrealloc, 0,
509 	"print out FFS filesystem block reallocation operations");
510 #endif
511 
512 int
513 ffs_reallocblks(ap)
514 	struct vop_reallocblks_args /* {
515 		struct vnode *a_vp;
516 		struct cluster_save *a_buflist;
517 	} */ *ap;
518 {
519 	struct ufsmount *ump;
520 
521 	/*
522 	 * We used to skip reallocating the blocks of a file into a
523 	 * contiguous sequence if the underlying flash device requested
524 	 * BIO_DELETE notifications, because devices that benefit from
525 	 * BIO_DELETE also benefit from not moving the data. However,
526 	 * the destination for the data is usually moved before the data
527 	 * is written to the initially allocated location, so we rarely
528 	 * suffer the penalty of extra writes. With the addition of the
529 	 * consolidation of contiguous blocks into single BIO_DELETE
530 	 * operations, having fewer but larger contiguous blocks reduces
531 	 * the number of (slow and expensive) BIO_DELETE operations. So
532 	 * when doing BIO_DELETE consolidation, we do block reallocation.
533 	 *
534 	 * Skip if reallocblks has been disabled globally.
535 	 */
536 	ump = ap->a_vp->v_mount->mnt_data;
537 	if ((((ump->um_flags) & UM_CANDELETE) != 0 && dotrimcons == 0) ||
538 	    doreallocblks == 0)
539 		return (ENOSPC);
540 
541 	/*
542 	 * We can't wait in softdep prealloc as it may fsync and recurse
543 	 * here.  Instead we simply fail to reallocate blocks if this
544 	 * rare condition arises.
545 	 */
546 	if (DOINGSOFTDEP(ap->a_vp))
547 		if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0)
548 			return (ENOSPC);
549 	if (ump->um_fstype == UFS1)
550 		return (ffs_reallocblks_ufs1(ap));
551 	return (ffs_reallocblks_ufs2(ap));
552 }
553 
554 static int
555 ffs_reallocblks_ufs1(ap)
556 	struct vop_reallocblks_args /* {
557 		struct vnode *a_vp;
558 		struct cluster_save *a_buflist;
559 	} */ *ap;
560 {
561 	struct fs *fs;
562 	struct inode *ip;
563 	struct vnode *vp;
564 	struct buf *sbp, *ebp, *bp;
565 	ufs1_daddr_t *bap, *sbap, *ebap;
566 	struct cluster_save *buflist;
567 	struct ufsmount *ump;
568 	ufs_lbn_t start_lbn, end_lbn;
569 	ufs1_daddr_t soff, newblk, blkno;
570 	ufs2_daddr_t pref;
571 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
572 	int i, cg, len, start_lvl, end_lvl, ssize;
573 
574 	vp = ap->a_vp;
575 	ip = VTOI(vp);
576 	ump = ITOUMP(ip);
577 	fs = ump->um_fs;
578 	/*
579 	 * If we are not tracking block clusters or if we have less than 4%
580 	 * free blocks left, then do not attempt to cluster. Running with
581 	 * less than 5% free block reserve is not recommended and those that
582 	 * choose to do so do not expect to have good file layout.
583 	 */
584 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
585 		return (ENOSPC);
586 	buflist = ap->a_buflist;
587 	len = buflist->bs_nchildren;
588 	start_lbn = buflist->bs_children[0]->b_lblkno;
589 	end_lbn = start_lbn + len - 1;
590 #ifdef INVARIANTS
591 	for (i = 0; i < len; i++)
592 		if (!ffs_checkblk(ip,
593 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
594 			panic("ffs_reallocblks: unallocated block 1");
595 	for (i = 1; i < len; i++)
596 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
597 			panic("ffs_reallocblks: non-logical cluster");
598 	blkno = buflist->bs_children[0]->b_blkno;
599 	ssize = fsbtodb(fs, fs->fs_frag);
600 	for (i = 1; i < len - 1; i++)
601 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
602 			panic("ffs_reallocblks: non-physical cluster %d", i);
603 #endif
604 	/*
605 	 * If the cluster crosses the boundary for the first indirect
606 	 * block, leave space for the indirect block. Indirect blocks
607 	 * are initially laid out in a position after the last direct
608 	 * block. Block reallocation would usually destroy locality by
609 	 * moving the indirect block out of the way to make room for
610 	 * data blocks if we didn't compensate here. We should also do
611 	 * this for other indirect block boundaries, but it is only
612 	 * important for the first one.
613 	 */
614 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
615 		return (ENOSPC);
616 	/*
617 	 * If the latest allocation is in a new cylinder group, assume that
618 	 * the filesystem has decided to move and do not force it back to
619 	 * the previous cylinder group.
620 	 */
621 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
622 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
623 		return (ENOSPC);
624 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
625 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
626 		return (ENOSPC);
627 	/*
628 	 * Get the starting offset and block map for the first block.
629 	 */
630 	if (start_lvl == 0) {
631 		sbap = &ip->i_din1->di_db[0];
632 		soff = start_lbn;
633 	} else {
634 		idp = &start_ap[start_lvl - 1];
635 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
636 			brelse(sbp);
637 			return (ENOSPC);
638 		}
639 		sbap = (ufs1_daddr_t *)sbp->b_data;
640 		soff = idp->in_off;
641 	}
642 	/*
643 	 * If the block range spans two block maps, get the second map.
644 	 */
645 	ebap = NULL;
646 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
647 		ssize = len;
648 	} else {
649 #ifdef INVARIANTS
650 		if (start_lvl > 0 &&
651 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
652 			panic("ffs_reallocblk: start == end");
653 #endif
654 		ssize = len - (idp->in_off + 1);
655 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
656 			goto fail;
657 		ebap = (ufs1_daddr_t *)ebp->b_data;
658 	}
659 	/*
660 	 * Find the preferred location for the cluster. If we have not
661 	 * previously failed at this endeavor, then follow our standard
662 	 * preference calculation. If we have failed at it, then pick up
663 	 * where we last ended our search.
664 	 */
665 	UFS_LOCK(ump);
666 	if (ip->i_nextclustercg == -1)
667 		pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap);
668 	else
669 		pref = cgdata(fs, ip->i_nextclustercg);
670 	/*
671 	 * Search the block map looking for an allocation of the desired size.
672 	 * To avoid wasting too much time, we limit the number of cylinder
673 	 * groups that we will search.
674 	 */
675 	cg = dtog(fs, pref);
676 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
677 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
678 			break;
679 		cg += 1;
680 		if (cg >= fs->fs_ncg)
681 			cg = 0;
682 	}
683 	/*
684 	 * If we have failed in our search, record where we gave up for
685 	 * next time. Otherwise, fall back to our usual search citerion.
686 	 */
687 	if (newblk == 0) {
688 		ip->i_nextclustercg = cg;
689 		UFS_UNLOCK(ump);
690 		goto fail;
691 	}
692 	ip->i_nextclustercg = -1;
693 	/*
694 	 * We have found a new contiguous block.
695 	 *
696 	 * First we have to replace the old block pointers with the new
697 	 * block pointers in the inode and indirect blocks associated
698 	 * with the file.
699 	 */
700 #ifdef DIAGNOSTIC
701 	if (prtrealloc)
702 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:",
703 		    (uintmax_t)ip->i_number,
704 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
705 #endif
706 	blkno = newblk;
707 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
708 		if (i == ssize) {
709 			bap = ebap;
710 			soff = -i;
711 		}
712 #ifdef INVARIANTS
713 		if (!ffs_checkblk(ip,
714 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
715 			panic("ffs_reallocblks: unallocated block 2");
716 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
717 			panic("ffs_reallocblks: alloc mismatch");
718 #endif
719 #ifdef DIAGNOSTIC
720 		if (prtrealloc)
721 			printf(" %d,", *bap);
722 #endif
723 		if (DOINGSOFTDEP(vp)) {
724 			if (sbap == &ip->i_din1->di_db[0] && i < ssize)
725 				softdep_setup_allocdirect(ip, start_lbn + i,
726 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
727 				    buflist->bs_children[i]);
728 			else
729 				softdep_setup_allocindir_page(ip, start_lbn + i,
730 				    i < ssize ? sbp : ebp, soff + i, blkno,
731 				    *bap, buflist->bs_children[i]);
732 		}
733 		*bap++ = blkno;
734 	}
735 	/*
736 	 * Next we must write out the modified inode and indirect blocks.
737 	 * For strict correctness, the writes should be synchronous since
738 	 * the old block values may have been written to disk. In practise
739 	 * they are almost never written, but if we are concerned about
740 	 * strict correctness, the `doasyncfree' flag should be set to zero.
741 	 *
742 	 * The test on `doasyncfree' should be changed to test a flag
743 	 * that shows whether the associated buffers and inodes have
744 	 * been written. The flag should be set when the cluster is
745 	 * started and cleared whenever the buffer or inode is flushed.
746 	 * We can then check below to see if it is set, and do the
747 	 * synchronous write only when it has been cleared.
748 	 */
749 	if (sbap != &ip->i_din1->di_db[0]) {
750 		if (doasyncfree)
751 			bdwrite(sbp);
752 		else
753 			bwrite(sbp);
754 	} else {
755 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
756 		if (!doasyncfree)
757 			ffs_update(vp, 1);
758 	}
759 	if (ssize < len) {
760 		if (doasyncfree)
761 			bdwrite(ebp);
762 		else
763 			bwrite(ebp);
764 	}
765 	/*
766 	 * Last, free the old blocks and assign the new blocks to the buffers.
767 	 */
768 #ifdef DIAGNOSTIC
769 	if (prtrealloc)
770 		printf("\n\tnew:");
771 #endif
772 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
773 		bp = buflist->bs_children[i];
774 		if (!DOINGSOFTDEP(vp))
775 			/*
776 			 * The usual case is that a set of N-contiguous blocks
777 			 * that was just allocated has been replaced with a
778 			 * set of N+1-contiguous blocks. If they are marked as
779 			 * B_DELWRI, the current contents have not been written
780 			 * to disk. It is possible that the blocks were written
781 			 * earlier, but very uncommon. If the blocks have never
782 			 * been written, there is no need to send a BIO_DELETE
783 			 * for them when they are freed. The gain from avoiding
784 			 * the TRIMs for the common case of unwritten blocks
785 			 * far exceeds the cost of the write amplification for
786 			 * the uncommon case of failing to send a TRIM for the
787 			 * blocks that had been written.
788 			 */
789 			ffs_blkfree(ump, fs, ump->um_devvp,
790 			    dbtofsb(fs, bp->b_blkno),
791 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
792 			    (bp->b_flags & B_DELWRI) != 0 ?
793 			    NOTRIM_KEY : SINGLETON_KEY);
794 		bp->b_blkno = fsbtodb(fs, blkno);
795 #ifdef INVARIANTS
796 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
797 			panic("ffs_reallocblks: unallocated block 3");
798 #endif
799 #ifdef DIAGNOSTIC
800 		if (prtrealloc)
801 			printf(" %d,", blkno);
802 #endif
803 	}
804 #ifdef DIAGNOSTIC
805 	if (prtrealloc) {
806 		prtrealloc--;
807 		printf("\n");
808 	}
809 #endif
810 	return (0);
811 
812 fail:
813 	if (ssize < len)
814 		brelse(ebp);
815 	if (sbap != &ip->i_din1->di_db[0])
816 		brelse(sbp);
817 	return (ENOSPC);
818 }
819 
820 static int
821 ffs_reallocblks_ufs2(ap)
822 	struct vop_reallocblks_args /* {
823 		struct vnode *a_vp;
824 		struct cluster_save *a_buflist;
825 	} */ *ap;
826 {
827 	struct fs *fs;
828 	struct inode *ip;
829 	struct vnode *vp;
830 	struct buf *sbp, *ebp, *bp;
831 	ufs2_daddr_t *bap, *sbap, *ebap;
832 	struct cluster_save *buflist;
833 	struct ufsmount *ump;
834 	ufs_lbn_t start_lbn, end_lbn;
835 	ufs2_daddr_t soff, newblk, blkno, pref;
836 	struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
837 	int i, cg, len, start_lvl, end_lvl, ssize;
838 
839 	vp = ap->a_vp;
840 	ip = VTOI(vp);
841 	ump = ITOUMP(ip);
842 	fs = ump->um_fs;
843 	/*
844 	 * If we are not tracking block clusters or if we have less than 4%
845 	 * free blocks left, then do not attempt to cluster. Running with
846 	 * less than 5% free block reserve is not recommended and those that
847 	 * choose to do so do not expect to have good file layout.
848 	 */
849 	if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
850 		return (ENOSPC);
851 	buflist = ap->a_buflist;
852 	len = buflist->bs_nchildren;
853 	start_lbn = buflist->bs_children[0]->b_lblkno;
854 	end_lbn = start_lbn + len - 1;
855 #ifdef INVARIANTS
856 	for (i = 0; i < len; i++)
857 		if (!ffs_checkblk(ip,
858 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
859 			panic("ffs_reallocblks: unallocated block 1");
860 	for (i = 1; i < len; i++)
861 		if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
862 			panic("ffs_reallocblks: non-logical cluster");
863 	blkno = buflist->bs_children[0]->b_blkno;
864 	ssize = fsbtodb(fs, fs->fs_frag);
865 	for (i = 1; i < len - 1; i++)
866 		if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
867 			panic("ffs_reallocblks: non-physical cluster %d", i);
868 #endif
869 	/*
870 	 * If the cluster crosses the boundary for the first indirect
871 	 * block, do not move anything in it. Indirect blocks are
872 	 * usually initially laid out in a position between the data
873 	 * blocks. Block reallocation would usually destroy locality by
874 	 * moving the indirect block out of the way to make room for
875 	 * data blocks if we didn't compensate here. We should also do
876 	 * this for other indirect block boundaries, but it is only
877 	 * important for the first one.
878 	 */
879 	if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
880 		return (ENOSPC);
881 	/*
882 	 * If the latest allocation is in a new cylinder group, assume that
883 	 * the filesystem has decided to move and do not force it back to
884 	 * the previous cylinder group.
885 	 */
886 	if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
887 	    dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
888 		return (ENOSPC);
889 	if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
890 	    ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
891 		return (ENOSPC);
892 	/*
893 	 * Get the starting offset and block map for the first block.
894 	 */
895 	if (start_lvl == 0) {
896 		sbap = &ip->i_din2->di_db[0];
897 		soff = start_lbn;
898 	} else {
899 		idp = &start_ap[start_lvl - 1];
900 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
901 			brelse(sbp);
902 			return (ENOSPC);
903 		}
904 		sbap = (ufs2_daddr_t *)sbp->b_data;
905 		soff = idp->in_off;
906 	}
907 	/*
908 	 * If the block range spans two block maps, get the second map.
909 	 */
910 	ebap = NULL;
911 	if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
912 		ssize = len;
913 	} else {
914 #ifdef INVARIANTS
915 		if (start_lvl > 0 &&
916 		    start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
917 			panic("ffs_reallocblk: start == end");
918 #endif
919 		ssize = len - (idp->in_off + 1);
920 		if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
921 			goto fail;
922 		ebap = (ufs2_daddr_t *)ebp->b_data;
923 	}
924 	/*
925 	 * Find the preferred location for the cluster. If we have not
926 	 * previously failed at this endeavor, then follow our standard
927 	 * preference calculation. If we have failed at it, then pick up
928 	 * where we last ended our search.
929 	 */
930 	UFS_LOCK(ump);
931 	if (ip->i_nextclustercg == -1)
932 		pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap);
933 	else
934 		pref = cgdata(fs, ip->i_nextclustercg);
935 	/*
936 	 * Search the block map looking for an allocation of the desired size.
937 	 * To avoid wasting too much time, we limit the number of cylinder
938 	 * groups that we will search.
939 	 */
940 	cg = dtog(fs, pref);
941 	for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
942 		if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
943 			break;
944 		cg += 1;
945 		if (cg >= fs->fs_ncg)
946 			cg = 0;
947 	}
948 	/*
949 	 * If we have failed in our search, record where we gave up for
950 	 * next time. Otherwise, fall back to our usual search citerion.
951 	 */
952 	if (newblk == 0) {
953 		ip->i_nextclustercg = cg;
954 		UFS_UNLOCK(ump);
955 		goto fail;
956 	}
957 	ip->i_nextclustercg = -1;
958 	/*
959 	 * We have found a new contiguous block.
960 	 *
961 	 * First we have to replace the old block pointers with the new
962 	 * block pointers in the inode and indirect blocks associated
963 	 * with the file.
964 	 */
965 #ifdef DIAGNOSTIC
966 	if (prtrealloc)
967 		printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number,
968 		    (intmax_t)start_lbn, (intmax_t)end_lbn);
969 #endif
970 	blkno = newblk;
971 	for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
972 		if (i == ssize) {
973 			bap = ebap;
974 			soff = -i;
975 		}
976 #ifdef INVARIANTS
977 		if (!ffs_checkblk(ip,
978 		   dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
979 			panic("ffs_reallocblks: unallocated block 2");
980 		if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
981 			panic("ffs_reallocblks: alloc mismatch");
982 #endif
983 #ifdef DIAGNOSTIC
984 		if (prtrealloc)
985 			printf(" %jd,", (intmax_t)*bap);
986 #endif
987 		if (DOINGSOFTDEP(vp)) {
988 			if (sbap == &ip->i_din2->di_db[0] && i < ssize)
989 				softdep_setup_allocdirect(ip, start_lbn + i,
990 				    blkno, *bap, fs->fs_bsize, fs->fs_bsize,
991 				    buflist->bs_children[i]);
992 			else
993 				softdep_setup_allocindir_page(ip, start_lbn + i,
994 				    i < ssize ? sbp : ebp, soff + i, blkno,
995 				    *bap, buflist->bs_children[i]);
996 		}
997 		*bap++ = blkno;
998 	}
999 	/*
1000 	 * Next we must write out the modified inode and indirect blocks.
1001 	 * For strict correctness, the writes should be synchronous since
1002 	 * the old block values may have been written to disk. In practise
1003 	 * they are almost never written, but if we are concerned about
1004 	 * strict correctness, the `doasyncfree' flag should be set to zero.
1005 	 *
1006 	 * The test on `doasyncfree' should be changed to test a flag
1007 	 * that shows whether the associated buffers and inodes have
1008 	 * been written. The flag should be set when the cluster is
1009 	 * started and cleared whenever the buffer or inode is flushed.
1010 	 * We can then check below to see if it is set, and do the
1011 	 * synchronous write only when it has been cleared.
1012 	 */
1013 	if (sbap != &ip->i_din2->di_db[0]) {
1014 		if (doasyncfree)
1015 			bdwrite(sbp);
1016 		else
1017 			bwrite(sbp);
1018 	} else {
1019 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE);
1020 		if (!doasyncfree)
1021 			ffs_update(vp, 1);
1022 	}
1023 	if (ssize < len) {
1024 		if (doasyncfree)
1025 			bdwrite(ebp);
1026 		else
1027 			bwrite(ebp);
1028 	}
1029 	/*
1030 	 * Last, free the old blocks and assign the new blocks to the buffers.
1031 	 */
1032 #ifdef DIAGNOSTIC
1033 	if (prtrealloc)
1034 		printf("\n\tnew:");
1035 #endif
1036 	for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
1037 		bp = buflist->bs_children[i];
1038 		if (!DOINGSOFTDEP(vp))
1039 			/*
1040 			 * The usual case is that a set of N-contiguous blocks
1041 			 * that was just allocated has been replaced with a
1042 			 * set of N+1-contiguous blocks. If they are marked as
1043 			 * B_DELWRI, the current contents have not been written
1044 			 * to disk. It is possible that the blocks were written
1045 			 * earlier, but very uncommon. If the blocks have never
1046 			 * been written, there is no need to send a BIO_DELETE
1047 			 * for them when they are freed. The gain from avoiding
1048 			 * the TRIMs for the common case of unwritten blocks
1049 			 * far exceeds the cost of the write amplification for
1050 			 * the uncommon case of failing to send a TRIM for the
1051 			 * blocks that had been written.
1052 			 */
1053 			ffs_blkfree(ump, fs, ump->um_devvp,
1054 			    dbtofsb(fs, bp->b_blkno),
1055 			    fs->fs_bsize, ip->i_number, vp->v_type, NULL,
1056 			    (bp->b_flags & B_DELWRI) != 0 ?
1057 			    NOTRIM_KEY : SINGLETON_KEY);
1058 		bp->b_blkno = fsbtodb(fs, blkno);
1059 #ifdef INVARIANTS
1060 		if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
1061 			panic("ffs_reallocblks: unallocated block 3");
1062 #endif
1063 #ifdef DIAGNOSTIC
1064 		if (prtrealloc)
1065 			printf(" %jd,", (intmax_t)blkno);
1066 #endif
1067 	}
1068 #ifdef DIAGNOSTIC
1069 	if (prtrealloc) {
1070 		prtrealloc--;
1071 		printf("\n");
1072 	}
1073 #endif
1074 	return (0);
1075 
1076 fail:
1077 	if (ssize < len)
1078 		brelse(ebp);
1079 	if (sbap != &ip->i_din2->di_db[0])
1080 		brelse(sbp);
1081 	return (ENOSPC);
1082 }
1083 
1084 /*
1085  * Allocate an inode in the filesystem.
1086  *
1087  * If allocating a directory, use ffs_dirpref to select the inode.
1088  * If allocating in a directory, the following hierarchy is followed:
1089  *   1) allocate the preferred inode.
1090  *   2) allocate an inode in the same cylinder group.
1091  *   3) quadradically rehash into other cylinder groups, until an
1092  *      available inode is located.
1093  * If no inode preference is given the following hierarchy is used
1094  * to allocate an inode:
1095  *   1) allocate an inode in cylinder group 0.
1096  *   2) quadradically rehash into other cylinder groups, until an
1097  *      available inode is located.
1098  */
1099 int
1100 ffs_valloc(pvp, mode, cred, vpp)
1101 	struct vnode *pvp;
1102 	int mode;
1103 	struct ucred *cred;
1104 	struct vnode **vpp;
1105 {
1106 	struct inode *pip;
1107 	struct fs *fs;
1108 	struct inode *ip;
1109 	struct timespec ts;
1110 	struct ufsmount *ump;
1111 	ino_t ino, ipref;
1112 	u_int cg;
1113 	int error, reclaimed;
1114 
1115 	*vpp = NULL;
1116 	pip = VTOI(pvp);
1117 	ump = ITOUMP(pip);
1118 	fs = ump->um_fs;
1119 
1120 	UFS_LOCK(ump);
1121 	reclaimed = 0;
1122 retry:
1123 	if (fs->fs_cstotal.cs_nifree == 0)
1124 		goto noinodes;
1125 
1126 	if ((mode & IFMT) == IFDIR)
1127 		ipref = ffs_dirpref(pip);
1128 	else
1129 		ipref = pip->i_number;
1130 	if (ipref >= fs->fs_ncg * fs->fs_ipg)
1131 		ipref = 0;
1132 	cg = ino_to_cg(fs, ipref);
1133 	/*
1134 	 * Track number of dirs created one after another
1135 	 * in a same cg without intervening by files.
1136 	 */
1137 	if ((mode & IFMT) == IFDIR) {
1138 		if (fs->fs_contigdirs[cg] < 255)
1139 			fs->fs_contigdirs[cg]++;
1140 	} else {
1141 		if (fs->fs_contigdirs[cg] > 0)
1142 			fs->fs_contigdirs[cg]--;
1143 	}
1144 	ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0,
1145 					(allocfcn_t *)ffs_nodealloccg);
1146 	if (ino == 0)
1147 		goto noinodes;
1148 	/*
1149 	 * Get rid of the cached old vnode, force allocation of a new vnode
1150 	 * for this inode. If this fails, release the allocated ino and
1151 	 * return the error.
1152 	 */
1153 	if ((error = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp,
1154 	    FFSV_FORCEINSMQ | FFSV_REPLACE)) != 0) {
1155 		ffs_vfree(pvp, ino, mode);
1156 		return (error);
1157 	}
1158 	/*
1159 	 * We got an inode, so check mode and panic if it is already allocated.
1160 	 */
1161 	ip = VTOI(*vpp);
1162 	if (ip->i_mode) {
1163 		printf("mode = 0%o, inum = %ju, fs = %s\n",
1164 		    ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt);
1165 		panic("ffs_valloc: dup alloc");
1166 	}
1167 	if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) {  /* XXX */
1168 		printf("free inode %s/%lu had %ld blocks\n",
1169 		    fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks));
1170 		DIP_SET(ip, i_blocks, 0);
1171 	}
1172 	ip->i_flags = 0;
1173 	DIP_SET(ip, i_flags, 0);
1174 	/*
1175 	 * Set up a new generation number for this inode.
1176 	 */
1177 	while (ip->i_gen == 0 || ++ip->i_gen == 0)
1178 		ip->i_gen = arc4random();
1179 	DIP_SET(ip, i_gen, ip->i_gen);
1180 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1181 		vfs_timestamp(&ts);
1182 		ip->i_din2->di_birthtime = ts.tv_sec;
1183 		ip->i_din2->di_birthnsec = ts.tv_nsec;
1184 	}
1185 	ip->i_flag = 0;
1186 	(*vpp)->v_vflag = 0;
1187 	(*vpp)->v_type = VNON;
1188 	if (fs->fs_magic == FS_UFS2_MAGIC) {
1189 		(*vpp)->v_op = &ffs_vnodeops2;
1190 		UFS_INODE_SET_FLAG(ip, IN_UFS2);
1191 	} else {
1192 		(*vpp)->v_op = &ffs_vnodeops1;
1193 	}
1194 	return (0);
1195 noinodes:
1196 	if (reclaimed == 0) {
1197 		reclaimed = 1;
1198 		softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT);
1199 		goto retry;
1200 	}
1201 	if (ffs_fsfail_cleanup_locked(ump, 0)) {
1202 		UFS_UNLOCK(ump);
1203 		return (ENXIO);
1204 	}
1205 	if (ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) {
1206 		UFS_UNLOCK(ump);
1207 		ffs_fserr(fs, pip->i_number, "out of inodes");
1208 		uprintf("\n%s: create/symlink failed, no inodes free\n",
1209 		    fs->fs_fsmnt);
1210 	} else {
1211 		UFS_UNLOCK(ump);
1212 	}
1213 	return (ENOSPC);
1214 }
1215 
1216 /*
1217  * Find a cylinder group to place a directory.
1218  *
1219  * The policy implemented by this algorithm is to allocate a
1220  * directory inode in the same cylinder group as its parent
1221  * directory, but also to reserve space for its files inodes
1222  * and data. Restrict the number of directories which may be
1223  * allocated one after another in the same cylinder group
1224  * without intervening allocation of files.
1225  *
1226  * If we allocate a first level directory then force allocation
1227  * in another cylinder group.
1228  */
1229 static ino_t
1230 ffs_dirpref(pip)
1231 	struct inode *pip;
1232 {
1233 	struct fs *fs;
1234 	int cg, prefcg, dirsize, cgsize;
1235 	u_int avgifree, avgbfree, avgndir, curdirsize;
1236 	u_int minifree, minbfree, maxndir;
1237 	u_int mincg, minndir;
1238 	u_int maxcontigdirs;
1239 
1240 	mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED);
1241 	fs = ITOFS(pip);
1242 
1243 	avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg;
1244 	avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1245 	avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg;
1246 
1247 	/*
1248 	 * Force allocation in another cg if creating a first level dir.
1249 	 */
1250 	ASSERT_VOP_LOCKED(ITOV(pip), "ffs_dirpref");
1251 	if (ITOV(pip)->v_vflag & VV_ROOT) {
1252 		prefcg = arc4random() % fs->fs_ncg;
1253 		mincg = prefcg;
1254 		minndir = fs->fs_ipg;
1255 		for (cg = prefcg; cg < fs->fs_ncg; cg++)
1256 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1257 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1258 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1259 				mincg = cg;
1260 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1261 			}
1262 		for (cg = 0; cg < prefcg; cg++)
1263 			if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1264 			    fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1265 			    fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1266 				mincg = cg;
1267 				minndir = fs->fs_cs(fs, cg).cs_ndir;
1268 			}
1269 		return ((ino_t)(fs->fs_ipg * mincg));
1270 	}
1271 
1272 	/*
1273 	 * Count various limits which used for
1274 	 * optimal allocation of a directory inode.
1275 	 */
1276 	maxndir = min(avgndir + fs->fs_ipg / 16, fs->fs_ipg);
1277 	minifree = avgifree - avgifree / 4;
1278 	if (minifree < 1)
1279 		minifree = 1;
1280 	minbfree = avgbfree - avgbfree / 4;
1281 	if (minbfree < 1)
1282 		minbfree = 1;
1283 	cgsize = fs->fs_fsize * fs->fs_fpg;
1284 	dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir;
1285 	curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0;
1286 	if (dirsize < curdirsize)
1287 		dirsize = curdirsize;
1288 	if (dirsize <= 0)
1289 		maxcontigdirs = 0;		/* dirsize overflowed */
1290 	else
1291 		maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255);
1292 	if (fs->fs_avgfpdir > 0)
1293 		maxcontigdirs = min(maxcontigdirs,
1294 				    fs->fs_ipg / fs->fs_avgfpdir);
1295 	if (maxcontigdirs == 0)
1296 		maxcontigdirs = 1;
1297 
1298 	/*
1299 	 * Limit number of dirs in one cg and reserve space for
1300 	 * regular files, but only if we have no deficit in
1301 	 * inodes or space.
1302 	 *
1303 	 * We are trying to find a suitable cylinder group nearby
1304 	 * our preferred cylinder group to place a new directory.
1305 	 * We scan from our preferred cylinder group forward looking
1306 	 * for a cylinder group that meets our criterion. If we get
1307 	 * to the final cylinder group and do not find anything,
1308 	 * we start scanning forwards from the beginning of the
1309 	 * filesystem. While it might seem sensible to start scanning
1310 	 * backwards or even to alternate looking forward and backward,
1311 	 * this approach fails badly when the filesystem is nearly full.
1312 	 * Specifically, we first search all the areas that have no space
1313 	 * and finally try the one preceding that. We repeat this on
1314 	 * every request and in the case of the final block end up
1315 	 * searching the entire filesystem. By jumping to the front
1316 	 * of the filesystem, our future forward searches always look
1317 	 * in new cylinder groups so finds every possible block after
1318 	 * one pass over the filesystem.
1319 	 */
1320 	prefcg = ino_to_cg(fs, pip->i_number);
1321 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1322 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1323 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1324 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1325 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1326 				return ((ino_t)(fs->fs_ipg * cg));
1327 		}
1328 	for (cg = 0; cg < prefcg; cg++)
1329 		if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1330 		    fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1331 		    fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1332 			if (fs->fs_contigdirs[cg] < maxcontigdirs)
1333 				return ((ino_t)(fs->fs_ipg * cg));
1334 		}
1335 	/*
1336 	 * This is a backstop when we have deficit in space.
1337 	 */
1338 	for (cg = prefcg; cg < fs->fs_ncg; cg++)
1339 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1340 			return ((ino_t)(fs->fs_ipg * cg));
1341 	for (cg = 0; cg < prefcg; cg++)
1342 		if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1343 			break;
1344 	return ((ino_t)(fs->fs_ipg * cg));
1345 }
1346 
1347 /*
1348  * Select the desired position for the next block in a file.  The file is
1349  * logically divided into sections. The first section is composed of the
1350  * direct blocks and the next fs_maxbpg blocks. Each additional section
1351  * contains fs_maxbpg blocks.
1352  *
1353  * If no blocks have been allocated in the first section, the policy is to
1354  * request a block in the same cylinder group as the inode that describes
1355  * the file. The first indirect is allocated immediately following the last
1356  * direct block and the data blocks for the first indirect immediately
1357  * follow it.
1358  *
1359  * If no blocks have been allocated in any other section, the indirect
1360  * block(s) are allocated in the same cylinder group as its inode in an
1361  * area reserved immediately following the inode blocks. The policy for
1362  * the data blocks is to place them in a cylinder group with a greater than
1363  * average number of free blocks. An appropriate cylinder group is found
1364  * by using a rotor that sweeps the cylinder groups. When a new group of
1365  * blocks is needed, the sweep begins in the cylinder group following the
1366  * cylinder group from which the previous allocation was made. The sweep
1367  * continues until a cylinder group with greater than the average number
1368  * of free blocks is found. If the allocation is for the first block in an
1369  * indirect block or the previous block is a hole, then the information on
1370  * the previous allocation is unavailable; here a best guess is made based
1371  * on the logical block number being allocated.
1372  *
1373  * If a section is already partially allocated, the policy is to
1374  * allocate blocks contiguously within the section if possible.
1375  */
1376 ufs2_daddr_t
1377 ffs_blkpref_ufs1(ip, lbn, indx, bap)
1378 	struct inode *ip;
1379 	ufs_lbn_t lbn;
1380 	int indx;
1381 	ufs1_daddr_t *bap;
1382 {
1383 	struct fs *fs;
1384 	u_int cg, inocg;
1385 	u_int avgbfree, startcg;
1386 	ufs2_daddr_t pref, prevbn;
1387 
1388 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1389 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1390 	fs = ITOFS(ip);
1391 	/*
1392 	 * Allocation of indirect blocks is indicated by passing negative
1393 	 * values in indx: -1 for single indirect, -2 for double indirect,
1394 	 * -3 for triple indirect. As noted below, we attempt to allocate
1395 	 * the first indirect inline with the file data. For all later
1396 	 * indirect blocks, the data is often allocated in other cylinder
1397 	 * groups. However to speed random file access and to speed up
1398 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1399 	 * (typically half of fs_minfree) of the data area of each cylinder
1400 	 * group to hold these later indirect blocks.
1401 	 */
1402 	inocg = ino_to_cg(fs, ip->i_number);
1403 	if (indx < 0) {
1404 		/*
1405 		 * Our preference for indirect blocks is the zone at the
1406 		 * beginning of the inode's cylinder group data area that
1407 		 * we try to reserve for indirect blocks.
1408 		 */
1409 		pref = cgmeta(fs, inocg);
1410 		/*
1411 		 * If we are allocating the first indirect block, try to
1412 		 * place it immediately following the last direct block.
1413 		 */
1414 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1415 		    ip->i_din1->di_db[UFS_NDADDR - 1] != 0)
1416 			pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1417 		return (pref);
1418 	}
1419 	/*
1420 	 * If we are allocating the first data block in the first indirect
1421 	 * block and the indirect has been allocated in the data block area,
1422 	 * try to place it immediately following the indirect block.
1423 	 */
1424 	if (lbn == UFS_NDADDR) {
1425 		pref = ip->i_din1->di_ib[0];
1426 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1427 		    pref < cgbase(fs, inocg + 1))
1428 			return (pref + fs->fs_frag);
1429 	}
1430 	/*
1431 	 * If we are at the beginning of a file, or we have already allocated
1432 	 * the maximum number of blocks per cylinder group, or we do not
1433 	 * have a block allocated immediately preceding us, then we need
1434 	 * to decide where to start allocating new blocks.
1435 	 */
1436 	if (indx ==  0) {
1437 		prevbn = 0;
1438 	} else {
1439 		prevbn = bap[indx - 1];
1440 		if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn,
1441 		    fs->fs_bsize) != 0)
1442 			prevbn = 0;
1443 	}
1444 	if (indx % fs->fs_maxbpg == 0 || prevbn == 0) {
1445 		/*
1446 		 * If we are allocating a directory data block, we want
1447 		 * to place it in the metadata area.
1448 		 */
1449 		if ((ip->i_mode & IFMT) == IFDIR)
1450 			return (cgmeta(fs, inocg));
1451 		/*
1452 		 * Until we fill all the direct and all the first indirect's
1453 		 * blocks, we try to allocate in the data area of the inode's
1454 		 * cylinder group.
1455 		 */
1456 		if (lbn < UFS_NDADDR + NINDIR(fs))
1457 			return (cgdata(fs, inocg));
1458 		/*
1459 		 * Find a cylinder with greater than average number of
1460 		 * unused data blocks.
1461 		 */
1462 		if (indx == 0 || prevbn == 0)
1463 			startcg = inocg + lbn / fs->fs_maxbpg;
1464 		else
1465 			startcg = dtog(fs, prevbn) + 1;
1466 		startcg %= fs->fs_ncg;
1467 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1468 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1469 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1470 				fs->fs_cgrotor = cg;
1471 				return (cgdata(fs, cg));
1472 			}
1473 		for (cg = 0; cg <= startcg; cg++)
1474 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1475 				fs->fs_cgrotor = cg;
1476 				return (cgdata(fs, cg));
1477 			}
1478 		return (0);
1479 	}
1480 	/*
1481 	 * Otherwise, we just always try to lay things out contiguously.
1482 	 */
1483 	return (prevbn + fs->fs_frag);
1484 }
1485 
1486 /*
1487  * Same as above, but for UFS2
1488  */
1489 ufs2_daddr_t
1490 ffs_blkpref_ufs2(ip, lbn, indx, bap)
1491 	struct inode *ip;
1492 	ufs_lbn_t lbn;
1493 	int indx;
1494 	ufs2_daddr_t *bap;
1495 {
1496 	struct fs *fs;
1497 	u_int cg, inocg;
1498 	u_int avgbfree, startcg;
1499 	ufs2_daddr_t pref, prevbn;
1500 
1501 	KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1502 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1503 	fs = ITOFS(ip);
1504 	/*
1505 	 * Allocation of indirect blocks is indicated by passing negative
1506 	 * values in indx: -1 for single indirect, -2 for double indirect,
1507 	 * -3 for triple indirect. As noted below, we attempt to allocate
1508 	 * the first indirect inline with the file data. For all later
1509 	 * indirect blocks, the data is often allocated in other cylinder
1510 	 * groups. However to speed random file access and to speed up
1511 	 * fsck, the filesystem reserves the first fs_metaspace blocks
1512 	 * (typically half of fs_minfree) of the data area of each cylinder
1513 	 * group to hold these later indirect blocks.
1514 	 */
1515 	inocg = ino_to_cg(fs, ip->i_number);
1516 	if (indx < 0) {
1517 		/*
1518 		 * Our preference for indirect blocks is the zone at the
1519 		 * beginning of the inode's cylinder group data area that
1520 		 * we try to reserve for indirect blocks.
1521 		 */
1522 		pref = cgmeta(fs, inocg);
1523 		/*
1524 		 * If we are allocating the first indirect block, try to
1525 		 * place it immediately following the last direct block.
1526 		 */
1527 		if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1528 		    ip->i_din2->di_db[UFS_NDADDR - 1] != 0)
1529 			pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1530 		return (pref);
1531 	}
1532 	/*
1533 	 * If we are allocating the first data block in the first indirect
1534 	 * block and the indirect has been allocated in the data block area,
1535 	 * try to place it immediately following the indirect block.
1536 	 */
1537 	if (lbn == UFS_NDADDR) {
1538 		pref = ip->i_din2->di_ib[0];
1539 		if (pref != 0 && pref >= cgdata(fs, inocg) &&
1540 		    pref < cgbase(fs, inocg + 1))
1541 			return (pref + fs->fs_frag);
1542 	}
1543 	/*
1544 	 * If we are at the beginning of a file, or we have already allocated
1545 	 * the maximum number of blocks per cylinder group, or we do not
1546 	 * have a block allocated immediately preceding us, then we need
1547 	 * to decide where to start allocating new blocks.
1548 	 */
1549 	if (indx ==  0) {
1550 		prevbn = 0;
1551 	} else {
1552 		prevbn = bap[indx - 1];
1553 		if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn,
1554 		    fs->fs_bsize) != 0)
1555 			prevbn = 0;
1556 	}
1557 	if (indx % fs->fs_maxbpg == 0 || prevbn == 0) {
1558 		/*
1559 		 * If we are allocating a directory data block, we want
1560 		 * to place it in the metadata area.
1561 		 */
1562 		if ((ip->i_mode & IFMT) == IFDIR)
1563 			return (cgmeta(fs, inocg));
1564 		/*
1565 		 * Until we fill all the direct and all the first indirect's
1566 		 * blocks, we try to allocate in the data area of the inode's
1567 		 * cylinder group.
1568 		 */
1569 		if (lbn < UFS_NDADDR + NINDIR(fs))
1570 			return (cgdata(fs, inocg));
1571 		/*
1572 		 * Find a cylinder with greater than average number of
1573 		 * unused data blocks.
1574 		 */
1575 		if (indx == 0 || prevbn == 0)
1576 			startcg = inocg + lbn / fs->fs_maxbpg;
1577 		else
1578 			startcg = dtog(fs, prevbn) + 1;
1579 		startcg %= fs->fs_ncg;
1580 		avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1581 		for (cg = startcg; cg < fs->fs_ncg; cg++)
1582 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1583 				fs->fs_cgrotor = cg;
1584 				return (cgdata(fs, cg));
1585 			}
1586 		for (cg = 0; cg <= startcg; cg++)
1587 			if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1588 				fs->fs_cgrotor = cg;
1589 				return (cgdata(fs, cg));
1590 			}
1591 		return (0);
1592 	}
1593 	/*
1594 	 * Otherwise, we just always try to lay things out contiguously.
1595 	 */
1596 	return (prevbn + fs->fs_frag);
1597 }
1598 
1599 /*
1600  * Implement the cylinder overflow algorithm.
1601  *
1602  * The policy implemented by this algorithm is:
1603  *   1) allocate the block in its requested cylinder group.
1604  *   2) quadradically rehash on the cylinder group number.
1605  *   3) brute force search for a free block.
1606  *
1607  * Must be called with the UFS lock held.  Will release the lock on success
1608  * and return with it held on failure.
1609  */
1610 /*VARARGS5*/
1611 static ufs2_daddr_t
1612 ffs_hashalloc(ip, cg, pref, size, rsize, allocator)
1613 	struct inode *ip;
1614 	u_int cg;
1615 	ufs2_daddr_t pref;
1616 	int size;	/* Search size for data blocks, mode for inodes */
1617 	int rsize;	/* Real allocated size. */
1618 	allocfcn_t *allocator;
1619 {
1620 	struct fs *fs;
1621 	ufs2_daddr_t result;
1622 	u_int i, icg = cg;
1623 
1624 	mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1625 #ifdef INVARIANTS
1626 	if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
1627 		panic("ffs_hashalloc: allocation on suspended filesystem");
1628 #endif
1629 	fs = ITOFS(ip);
1630 	/*
1631 	 * 1: preferred cylinder group
1632 	 */
1633 	result = (*allocator)(ip, cg, pref, size, rsize);
1634 	if (result)
1635 		return (result);
1636 	/*
1637 	 * 2: quadratic rehash
1638 	 */
1639 	for (i = 1; i < fs->fs_ncg; i *= 2) {
1640 		cg += i;
1641 		if (cg >= fs->fs_ncg)
1642 			cg -= fs->fs_ncg;
1643 		result = (*allocator)(ip, cg, 0, size, rsize);
1644 		if (result)
1645 			return (result);
1646 	}
1647 	/*
1648 	 * 3: brute force search
1649 	 * Note that we start at i == 2, since 0 was checked initially,
1650 	 * and 1 is always checked in the quadratic rehash.
1651 	 */
1652 	cg = (icg + 2) % fs->fs_ncg;
1653 	for (i = 2; i < fs->fs_ncg; i++) {
1654 		result = (*allocator)(ip, cg, 0, size, rsize);
1655 		if (result)
1656 			return (result);
1657 		cg++;
1658 		if (cg == fs->fs_ncg)
1659 			cg = 0;
1660 	}
1661 	return (0);
1662 }
1663 
1664 /*
1665  * Determine whether a fragment can be extended.
1666  *
1667  * Check to see if the necessary fragments are available, and
1668  * if they are, allocate them.
1669  */
1670 static ufs2_daddr_t
1671 ffs_fragextend(ip, cg, bprev, osize, nsize)
1672 	struct inode *ip;
1673 	u_int cg;
1674 	ufs2_daddr_t bprev;
1675 	int osize, nsize;
1676 {
1677 	struct fs *fs;
1678 	struct cg *cgp;
1679 	struct buf *bp;
1680 	struct ufsmount *ump;
1681 	int nffree;
1682 	long bno;
1683 	int frags, bbase;
1684 	int i, error;
1685 	u_int8_t *blksfree;
1686 
1687 	ump = ITOUMP(ip);
1688 	fs = ump->um_fs;
1689 	if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize))
1690 		return (0);
1691 	frags = numfrags(fs, nsize);
1692 	bbase = fragnum(fs, bprev);
1693 	if (bbase > fragnum(fs, (bprev + frags - 1))) {
1694 		/* cannot extend across a block boundary */
1695 		return (0);
1696 	}
1697 	UFS_UNLOCK(ump);
1698 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0)
1699 		goto fail;
1700 	bno = dtogd(fs, bprev);
1701 	blksfree = cg_blksfree(cgp);
1702 	for (i = numfrags(fs, osize); i < frags; i++)
1703 		if (isclr(blksfree, bno + i))
1704 			goto fail;
1705 	/*
1706 	 * the current fragment can be extended
1707 	 * deduct the count on fragment being extended into
1708 	 * increase the count on the remaining fragment (if any)
1709 	 * allocate the extended piece
1710 	 */
1711 	for (i = frags; i < fs->fs_frag - bbase; i++)
1712 		if (isclr(blksfree, bno + i))
1713 			break;
1714 	cgp->cg_frsum[i - numfrags(fs, osize)]--;
1715 	if (i != frags)
1716 		cgp->cg_frsum[i - frags]++;
1717 	for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) {
1718 		clrbit(blksfree, bno + i);
1719 		cgp->cg_cs.cs_nffree--;
1720 		nffree++;
1721 	}
1722 	UFS_LOCK(ump);
1723 	fs->fs_cstotal.cs_nffree -= nffree;
1724 	fs->fs_cs(fs, cg).cs_nffree -= nffree;
1725 	fs->fs_fmod = 1;
1726 	ACTIVECLEAR(fs, cg);
1727 	UFS_UNLOCK(ump);
1728 	if (DOINGSOFTDEP(ITOV(ip)))
1729 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev,
1730 		    frags, numfrags(fs, osize));
1731 	bdwrite(bp);
1732 	return (bprev);
1733 
1734 fail:
1735 	brelse(bp);
1736 	UFS_LOCK(ump);
1737 	return (0);
1738 
1739 }
1740 
1741 /*
1742  * Determine whether a block can be allocated.
1743  *
1744  * Check to see if a block of the appropriate size is available,
1745  * and if it is, allocate it.
1746  */
1747 static ufs2_daddr_t
1748 ffs_alloccg(ip, cg, bpref, size, rsize)
1749 	struct inode *ip;
1750 	u_int cg;
1751 	ufs2_daddr_t bpref;
1752 	int size;
1753 	int rsize;
1754 {
1755 	struct fs *fs;
1756 	struct cg *cgp;
1757 	struct buf *bp;
1758 	struct ufsmount *ump;
1759 	ufs1_daddr_t bno;
1760 	ufs2_daddr_t blkno;
1761 	int i, allocsiz, error, frags;
1762 	u_int8_t *blksfree;
1763 
1764 	ump = ITOUMP(ip);
1765 	fs = ump->um_fs;
1766 	if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize)
1767 		return (0);
1768 	UFS_UNLOCK(ump);
1769 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0 ||
1770 	   (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize))
1771 		goto fail;
1772 	if (size == fs->fs_bsize) {
1773 		UFS_LOCK(ump);
1774 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1775 		ACTIVECLEAR(fs, cg);
1776 		UFS_UNLOCK(ump);
1777 		bdwrite(bp);
1778 		return (blkno);
1779 	}
1780 	/*
1781 	 * check to see if any fragments are already available
1782 	 * allocsiz is the size which will be allocated, hacking
1783 	 * it down to a smaller size if necessary
1784 	 */
1785 	blksfree = cg_blksfree(cgp);
1786 	frags = numfrags(fs, size);
1787 	for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++)
1788 		if (cgp->cg_frsum[allocsiz] != 0)
1789 			break;
1790 	if (allocsiz == fs->fs_frag) {
1791 		/*
1792 		 * no fragments were available, so a block will be
1793 		 * allocated, and hacked up
1794 		 */
1795 		if (cgp->cg_cs.cs_nbfree == 0)
1796 			goto fail;
1797 		UFS_LOCK(ump);
1798 		blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1799 		ACTIVECLEAR(fs, cg);
1800 		UFS_UNLOCK(ump);
1801 		bdwrite(bp);
1802 		return (blkno);
1803 	}
1804 	KASSERT(size == rsize,
1805 	    ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize));
1806 	bno = ffs_mapsearch(fs, cgp, bpref, allocsiz);
1807 	if (bno < 0)
1808 		goto fail;
1809 	for (i = 0; i < frags; i++)
1810 		clrbit(blksfree, bno + i);
1811 	cgp->cg_cs.cs_nffree -= frags;
1812 	cgp->cg_frsum[allocsiz]--;
1813 	if (frags != allocsiz)
1814 		cgp->cg_frsum[allocsiz - frags]++;
1815 	UFS_LOCK(ump);
1816 	fs->fs_cstotal.cs_nffree -= frags;
1817 	fs->fs_cs(fs, cg).cs_nffree -= frags;
1818 	fs->fs_fmod = 1;
1819 	blkno = cgbase(fs, cg) + bno;
1820 	ACTIVECLEAR(fs, cg);
1821 	UFS_UNLOCK(ump);
1822 	if (DOINGSOFTDEP(ITOV(ip)))
1823 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0);
1824 	bdwrite(bp);
1825 	return (blkno);
1826 
1827 fail:
1828 	brelse(bp);
1829 	UFS_LOCK(ump);
1830 	return (0);
1831 }
1832 
1833 /*
1834  * Allocate a block in a cylinder group.
1835  *
1836  * This algorithm implements the following policy:
1837  *   1) allocate the requested block.
1838  *   2) allocate a rotationally optimal block in the same cylinder.
1839  *   3) allocate the next available block on the block rotor for the
1840  *      specified cylinder group.
1841  * Note that this routine only allocates fs_bsize blocks; these
1842  * blocks may be fragmented by the routine that allocates them.
1843  */
1844 static ufs2_daddr_t
1845 ffs_alloccgblk(ip, bp, bpref, size)
1846 	struct inode *ip;
1847 	struct buf *bp;
1848 	ufs2_daddr_t bpref;
1849 	int size;
1850 {
1851 	struct fs *fs;
1852 	struct cg *cgp;
1853 	struct ufsmount *ump;
1854 	ufs1_daddr_t bno;
1855 	ufs2_daddr_t blkno;
1856 	u_int8_t *blksfree;
1857 	int i, cgbpref;
1858 
1859 	ump = ITOUMP(ip);
1860 	fs = ump->um_fs;
1861 	mtx_assert(UFS_MTX(ump), MA_OWNED);
1862 	cgp = (struct cg *)bp->b_data;
1863 	blksfree = cg_blksfree(cgp);
1864 	if (bpref == 0) {
1865 		bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag;
1866 	} else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) {
1867 		/* map bpref to correct zone in this cg */
1868 		if (bpref < cgdata(fs, cgbpref))
1869 			bpref = cgmeta(fs, cgp->cg_cgx);
1870 		else
1871 			bpref = cgdata(fs, cgp->cg_cgx);
1872 	}
1873 	/*
1874 	 * if the requested block is available, use it
1875 	 */
1876 	bno = dtogd(fs, blknum(fs, bpref));
1877 	if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno)))
1878 		goto gotit;
1879 	/*
1880 	 * Take the next available block in this cylinder group.
1881 	 */
1882 	bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag);
1883 	if (bno < 0)
1884 		return (0);
1885 	/* Update cg_rotor only if allocated from the data zone */
1886 	if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx)))
1887 		cgp->cg_rotor = bno;
1888 gotit:
1889 	blkno = fragstoblks(fs, bno);
1890 	ffs_clrblock(fs, blksfree, (long)blkno);
1891 	ffs_clusteracct(fs, cgp, blkno, -1);
1892 	cgp->cg_cs.cs_nbfree--;
1893 	fs->fs_cstotal.cs_nbfree--;
1894 	fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--;
1895 	fs->fs_fmod = 1;
1896 	blkno = cgbase(fs, cgp->cg_cgx) + bno;
1897 	/*
1898 	 * If the caller didn't want the whole block free the frags here.
1899 	 */
1900 	size = numfrags(fs, size);
1901 	if (size != fs->fs_frag) {
1902 		bno = dtogd(fs, blkno);
1903 		for (i = size; i < fs->fs_frag; i++)
1904 			setbit(blksfree, bno + i);
1905 		i = fs->fs_frag - size;
1906 		cgp->cg_cs.cs_nffree += i;
1907 		fs->fs_cstotal.cs_nffree += i;
1908 		fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i;
1909 		fs->fs_fmod = 1;
1910 		cgp->cg_frsum[i]++;
1911 	}
1912 	/* XXX Fixme. */
1913 	UFS_UNLOCK(ump);
1914 	if (DOINGSOFTDEP(ITOV(ip)))
1915 		softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0);
1916 	UFS_LOCK(ump);
1917 	return (blkno);
1918 }
1919 
1920 /*
1921  * Determine whether a cluster can be allocated.
1922  *
1923  * We do not currently check for optimal rotational layout if there
1924  * are multiple choices in the same cylinder group. Instead we just
1925  * take the first one that we find following bpref.
1926  */
1927 static ufs2_daddr_t
1928 ffs_clusteralloc(ip, cg, bpref, len)
1929 	struct inode *ip;
1930 	u_int cg;
1931 	ufs2_daddr_t bpref;
1932 	int len;
1933 {
1934 	struct fs *fs;
1935 	struct cg *cgp;
1936 	struct buf *bp;
1937 	struct ufsmount *ump;
1938 	int i, run, bit, map, got, error;
1939 	ufs2_daddr_t bno;
1940 	u_char *mapp;
1941 	int32_t *lp;
1942 	u_int8_t *blksfree;
1943 
1944 	ump = ITOUMP(ip);
1945 	fs = ump->um_fs;
1946 	if (fs->fs_maxcluster[cg] < len)
1947 		return (0);
1948 	UFS_UNLOCK(ump);
1949 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) {
1950 		UFS_LOCK(ump);
1951 		return (0);
1952 	}
1953 	/*
1954 	 * Check to see if a cluster of the needed size (or bigger) is
1955 	 * available in this cylinder group.
1956 	 */
1957 	lp = &cg_clustersum(cgp)[len];
1958 	for (i = len; i <= fs->fs_contigsumsize; i++)
1959 		if (*lp++ > 0)
1960 			break;
1961 	if (i > fs->fs_contigsumsize) {
1962 		/*
1963 		 * This is the first time looking for a cluster in this
1964 		 * cylinder group. Update the cluster summary information
1965 		 * to reflect the true maximum sized cluster so that
1966 		 * future cluster allocation requests can avoid reading
1967 		 * the cylinder group map only to find no clusters.
1968 		 */
1969 		lp = &cg_clustersum(cgp)[len - 1];
1970 		for (i = len - 1; i > 0; i--)
1971 			if (*lp-- > 0)
1972 				break;
1973 		UFS_LOCK(ump);
1974 		fs->fs_maxcluster[cg] = i;
1975 		brelse(bp);
1976 		return (0);
1977 	}
1978 	/*
1979 	 * Search the cluster map to find a big enough cluster.
1980 	 * We take the first one that we find, even if it is larger
1981 	 * than we need as we prefer to get one close to the previous
1982 	 * block allocation. We do not search before the current
1983 	 * preference point as we do not want to allocate a block
1984 	 * that is allocated before the previous one (as we will
1985 	 * then have to wait for another pass of the elevator
1986 	 * algorithm before it will be read). We prefer to fail and
1987 	 * be recalled to try an allocation in the next cylinder group.
1988 	 */
1989 	if (dtog(fs, bpref) != cg)
1990 		bpref = cgdata(fs, cg);
1991 	else
1992 		bpref = blknum(fs, bpref);
1993 	bpref = fragstoblks(fs, dtogd(fs, bpref));
1994 	mapp = &cg_clustersfree(cgp)[bpref / NBBY];
1995 	map = *mapp++;
1996 	bit = 1 << (bpref % NBBY);
1997 	for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) {
1998 		if ((map & bit) == 0) {
1999 			run = 0;
2000 		} else {
2001 			run++;
2002 			if (run == len)
2003 				break;
2004 		}
2005 		if ((got & (NBBY - 1)) != (NBBY - 1)) {
2006 			bit <<= 1;
2007 		} else {
2008 			map = *mapp++;
2009 			bit = 1;
2010 		}
2011 	}
2012 	if (got >= cgp->cg_nclusterblks) {
2013 		UFS_LOCK(ump);
2014 		brelse(bp);
2015 		return (0);
2016 	}
2017 	/*
2018 	 * Allocate the cluster that we have found.
2019 	 */
2020 	blksfree = cg_blksfree(cgp);
2021 	for (i = 1; i <= len; i++)
2022 		if (!ffs_isblock(fs, blksfree, got - run + i))
2023 			panic("ffs_clusteralloc: map mismatch");
2024 	bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1);
2025 	if (dtog(fs, bno) != cg)
2026 		panic("ffs_clusteralloc: allocated out of group");
2027 	len = blkstofrags(fs, len);
2028 	UFS_LOCK(ump);
2029 	for (i = 0; i < len; i += fs->fs_frag)
2030 		if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i)
2031 			panic("ffs_clusteralloc: lost block");
2032 	ACTIVECLEAR(fs, cg);
2033 	UFS_UNLOCK(ump);
2034 	bdwrite(bp);
2035 	return (bno);
2036 }
2037 
2038 static inline struct buf *
2039 getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags)
2040 {
2041 	struct fs *fs;
2042 
2043 	fs = ITOFS(ip);
2044 	return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs,
2045 	    cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
2046 	    gbflags));
2047 }
2048 
2049 /*
2050  * Synchronous inode initialization is needed only when barrier writes do not
2051  * work as advertised, and will impose a heavy cost on file creation in a newly
2052  * created filesystem.
2053  */
2054 static int doasyncinodeinit = 1;
2055 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN,
2056     &doasyncinodeinit, 0,
2057     "Perform inode block initialization using asynchronous writes");
2058 
2059 /*
2060  * Determine whether an inode can be allocated.
2061  *
2062  * Check to see if an inode is available, and if it is,
2063  * allocate it using the following policy:
2064  *   1) allocate the requested inode.
2065  *   2) allocate the next available inode after the requested
2066  *      inode in the specified cylinder group.
2067  */
2068 static ufs2_daddr_t
2069 ffs_nodealloccg(ip, cg, ipref, mode, unused)
2070 	struct inode *ip;
2071 	u_int cg;
2072 	ufs2_daddr_t ipref;
2073 	int mode;
2074 	int unused;
2075 {
2076 	struct fs *fs;
2077 	struct cg *cgp;
2078 	struct buf *bp, *ibp;
2079 	struct ufsmount *ump;
2080 	u_int8_t *inosused, *loc;
2081 	struct ufs2_dinode *dp2;
2082 	int error, start, len, i;
2083 	u_int32_t old_initediblk;
2084 
2085 	ump = ITOUMP(ip);
2086 	fs = ump->um_fs;
2087 check_nifree:
2088 	if (fs->fs_cs(fs, cg).cs_nifree == 0)
2089 		return (0);
2090 	UFS_UNLOCK(ump);
2091 	if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) {
2092 		UFS_LOCK(ump);
2093 		return (0);
2094 	}
2095 restart:
2096 	if (cgp->cg_cs.cs_nifree == 0) {
2097 		brelse(bp);
2098 		UFS_LOCK(ump);
2099 		return (0);
2100 	}
2101 	inosused = cg_inosused(cgp);
2102 	if (ipref) {
2103 		ipref %= fs->fs_ipg;
2104 		if (isclr(inosused, ipref))
2105 			goto gotit;
2106 	}
2107 	start = cgp->cg_irotor / NBBY;
2108 	len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY);
2109 	loc = memcchr(&inosused[start], 0xff, len);
2110 	if (loc == NULL) {
2111 		len = start + 1;
2112 		start = 0;
2113 		loc = memcchr(&inosused[start], 0xff, len);
2114 		if (loc == NULL) {
2115 			printf("cg = %d, irotor = %ld, fs = %s\n",
2116 			    cg, (long)cgp->cg_irotor, fs->fs_fsmnt);
2117 			panic("ffs_nodealloccg: map corrupted");
2118 			/* NOTREACHED */
2119 		}
2120 	}
2121 	ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
2122 gotit:
2123 	/*
2124 	 * Check to see if we need to initialize more inodes.
2125 	 */
2126 	if (fs->fs_magic == FS_UFS2_MAGIC &&
2127 	    ipref + INOPB(fs) > cgp->cg_initediblk &&
2128 	    cgp->cg_initediblk < cgp->cg_niblk) {
2129 		old_initediblk = cgp->cg_initediblk;
2130 
2131 		/*
2132 		 * Free the cylinder group lock before writing the
2133 		 * initialized inode block.  Entering the
2134 		 * babarrierwrite() with the cylinder group lock
2135 		 * causes lock order violation between the lock and
2136 		 * snaplk.
2137 		 *
2138 		 * Another thread can decide to initialize the same
2139 		 * inode block, but whichever thread first gets the
2140 		 * cylinder group lock after writing the newly
2141 		 * allocated inode block will update it and the other
2142 		 * will realize that it has lost and leave the
2143 		 * cylinder group unchanged.
2144 		 */
2145 		ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
2146 		brelse(bp);
2147 		if (ibp == NULL) {
2148 			/*
2149 			 * The inode block buffer is already owned by
2150 			 * another thread, which must initialize it.
2151 			 * Wait on the buffer to allow another thread
2152 			 * to finish the updates, with dropped cg
2153 			 * buffer lock, then retry.
2154 			 */
2155 			ibp = getinobuf(ip, cg, old_initediblk, 0);
2156 			brelse(ibp);
2157 			UFS_LOCK(ump);
2158 			goto check_nifree;
2159 		}
2160 		bzero(ibp->b_data, (int)fs->fs_bsize);
2161 		dp2 = (struct ufs2_dinode *)(ibp->b_data);
2162 		for (i = 0; i < INOPB(fs); i++) {
2163 			while (dp2->di_gen == 0)
2164 				dp2->di_gen = arc4random();
2165 			dp2++;
2166 		}
2167 
2168 		/*
2169 		 * Rather than adding a soft updates dependency to ensure
2170 		 * that the new inode block is written before it is claimed
2171 		 * by the cylinder group map, we just do a barrier write
2172 		 * here. The barrier write will ensure that the inode block
2173 		 * gets written before the updated cylinder group map can be
2174 		 * written. The barrier write should only slow down bulk
2175 		 * loading of newly created filesystems.
2176 		 */
2177 		if (doasyncinodeinit)
2178 			babarrierwrite(ibp);
2179 		else
2180 			bwrite(ibp);
2181 
2182 		/*
2183 		 * After the inode block is written, try to update the
2184 		 * cg initediblk pointer.  If another thread beat us
2185 		 * to it, then leave it unchanged as the other thread
2186 		 * has already set it correctly.
2187 		 */
2188 		error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp);
2189 		UFS_LOCK(ump);
2190 		ACTIVECLEAR(fs, cg);
2191 		UFS_UNLOCK(ump);
2192 		if (error != 0)
2193 			return (error);
2194 		if (cgp->cg_initediblk == old_initediblk)
2195 			cgp->cg_initediblk += INOPB(fs);
2196 		goto restart;
2197 	}
2198 	cgp->cg_irotor = ipref;
2199 	UFS_LOCK(ump);
2200 	ACTIVECLEAR(fs, cg);
2201 	setbit(inosused, ipref);
2202 	cgp->cg_cs.cs_nifree--;
2203 	fs->fs_cstotal.cs_nifree--;
2204 	fs->fs_cs(fs, cg).cs_nifree--;
2205 	fs->fs_fmod = 1;
2206 	if ((mode & IFMT) == IFDIR) {
2207 		cgp->cg_cs.cs_ndir++;
2208 		fs->fs_cstotal.cs_ndir++;
2209 		fs->fs_cs(fs, cg).cs_ndir++;
2210 	}
2211 	UFS_UNLOCK(ump);
2212 	if (DOINGSOFTDEP(ITOV(ip)))
2213 		softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode);
2214 	bdwrite(bp);
2215 	return ((ino_t)(cg * fs->fs_ipg + ipref));
2216 }
2217 
2218 /*
2219  * Free a block or fragment.
2220  *
2221  * The specified block or fragment is placed back in the
2222  * free map. If a fragment is deallocated, a possible
2223  * block reassembly is checked.
2224  */
2225 static void
2226 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd)
2227 	struct ufsmount *ump;
2228 	struct fs *fs;
2229 	struct vnode *devvp;
2230 	ufs2_daddr_t bno;
2231 	long size;
2232 	ino_t inum;
2233 	struct workhead *dephd;
2234 {
2235 	struct mount *mp;
2236 	struct cg *cgp;
2237 	struct buf *bp;
2238 	daddr_t dbn;
2239 	ufs1_daddr_t fragno, cgbno;
2240 	int i, blk, frags, bbase, error;
2241 	u_int cg;
2242 	u_int8_t *blksfree;
2243 	struct cdev *dev;
2244 
2245 	cg = dtog(fs, bno);
2246 	if (devvp->v_type == VREG) {
2247 		/* devvp is a snapshot */
2248 		MPASS(devvp->v_mount->mnt_data == ump);
2249 		dev = ump->um_devvp->v_rdev;
2250 	} else if (devvp->v_type == VCHR) {
2251 		/* devvp is a normal disk device */
2252 		dev = devvp->v_rdev;
2253 		ASSERT_VOP_LOCKED(devvp, "ffs_blkfree_cg");
2254 	} else
2255 		return;
2256 #ifdef INVARIANTS
2257 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 ||
2258 	    fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) {
2259 		printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n",
2260 		    devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize,
2261 		    size, fs->fs_fsmnt);
2262 		panic("ffs_blkfree_cg: bad size");
2263 	}
2264 #endif
2265 	if ((u_int)bno >= fs->fs_size) {
2266 		printf("bad block %jd, ino %lu\n", (intmax_t)bno,
2267 		    (u_long)inum);
2268 		ffs_fserr(fs, inum, "bad block");
2269 		return;
2270 	}
2271 	if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) {
2272 		if (!ffs_fsfail_cleanup(ump, error) ||
2273 		    !MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR)
2274 			return;
2275 		if (devvp->v_type == VREG)
2276 			dbn = fragstoblks(fs, cgtod(fs, cg));
2277 		else
2278 			dbn = fsbtodb(fs, cgtod(fs, cg));
2279 		error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp);
2280 		KASSERT(error == 0, ("getblkx failed"));
2281 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2282 		    numfrags(fs, size), dephd);
2283 		bp->b_flags |= B_RELBUF | B_NOCACHE;
2284 		bp->b_flags &= ~B_CACHE;
2285 		bawrite(bp);
2286 		return;
2287 	}
2288 	cgbno = dtogd(fs, bno);
2289 	blksfree = cg_blksfree(cgp);
2290 	UFS_LOCK(ump);
2291 	if (size == fs->fs_bsize) {
2292 		fragno = fragstoblks(fs, cgbno);
2293 		if (!ffs_isfreeblock(fs, blksfree, fragno)) {
2294 			if (devvp->v_type == VREG) {
2295 				UFS_UNLOCK(ump);
2296 				/* devvp is a snapshot */
2297 				brelse(bp);
2298 				return;
2299 			}
2300 			printf("dev = %s, block = %jd, fs = %s\n",
2301 			    devtoname(dev), (intmax_t)bno, fs->fs_fsmnt);
2302 			panic("ffs_blkfree_cg: freeing free block");
2303 		}
2304 		ffs_setblock(fs, blksfree, fragno);
2305 		ffs_clusteracct(fs, cgp, fragno, 1);
2306 		cgp->cg_cs.cs_nbfree++;
2307 		fs->fs_cstotal.cs_nbfree++;
2308 		fs->fs_cs(fs, cg).cs_nbfree++;
2309 	} else {
2310 		bbase = cgbno - fragnum(fs, cgbno);
2311 		/*
2312 		 * decrement the counts associated with the old frags
2313 		 */
2314 		blk = blkmap(fs, blksfree, bbase);
2315 		ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
2316 		/*
2317 		 * deallocate the fragment
2318 		 */
2319 		frags = numfrags(fs, size);
2320 		for (i = 0; i < frags; i++) {
2321 			if (isset(blksfree, cgbno + i)) {
2322 				printf("dev = %s, block = %jd, fs = %s\n",
2323 				    devtoname(dev), (intmax_t)(bno + i),
2324 				    fs->fs_fsmnt);
2325 				panic("ffs_blkfree_cg: freeing free frag");
2326 			}
2327 			setbit(blksfree, cgbno + i);
2328 		}
2329 		cgp->cg_cs.cs_nffree += i;
2330 		fs->fs_cstotal.cs_nffree += i;
2331 		fs->fs_cs(fs, cg).cs_nffree += i;
2332 		/*
2333 		 * add back in counts associated with the new frags
2334 		 */
2335 		blk = blkmap(fs, blksfree, bbase);
2336 		ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
2337 		/*
2338 		 * if a complete block has been reassembled, account for it
2339 		 */
2340 		fragno = fragstoblks(fs, bbase);
2341 		if (ffs_isblock(fs, blksfree, fragno)) {
2342 			cgp->cg_cs.cs_nffree -= fs->fs_frag;
2343 			fs->fs_cstotal.cs_nffree -= fs->fs_frag;
2344 			fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag;
2345 			ffs_clusteracct(fs, cgp, fragno, 1);
2346 			cgp->cg_cs.cs_nbfree++;
2347 			fs->fs_cstotal.cs_nbfree++;
2348 			fs->fs_cs(fs, cg).cs_nbfree++;
2349 		}
2350 	}
2351 	fs->fs_fmod = 1;
2352 	ACTIVECLEAR(fs, cg);
2353 	UFS_UNLOCK(ump);
2354 	mp = UFSTOVFS(ump);
2355 	if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR)
2356 		softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2357 		    numfrags(fs, size), dephd);
2358 	bdwrite(bp);
2359 }
2360 
2361 /*
2362  * Structures and routines associated with trim management.
2363  *
2364  * The following requests are passed to trim_lookup to indicate
2365  * the actions that should be taken.
2366  */
2367 #define	NEW	1	/* if found, error else allocate and hash it */
2368 #define	OLD	2	/* if not found, error, else return it */
2369 #define	REPLACE	3	/* if not found, error else unhash and reallocate it */
2370 #define	DONE	4	/* if not found, error else unhash and return it */
2371 #define	SINGLE	5	/* don't look up, just allocate it and don't hash it */
2372 
2373 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures");
2374 
2375 #define	TRIMLIST_HASH(ump, key) \
2376 	(&(ump)->um_trimhash[(key) & (ump)->um_trimlisthashsize])
2377 
2378 /*
2379  * These structures describe each of the block free requests aggregated
2380  * together to make up a trim request.
2381  */
2382 struct trim_blkreq {
2383 	TAILQ_ENTRY(trim_blkreq) blkreqlist;
2384 	ufs2_daddr_t bno;
2385 	long size;
2386 	struct workhead *pdephd;
2387 	struct workhead dephd;
2388 };
2389 
2390 /*
2391  * Description of a trim request.
2392  */
2393 struct ffs_blkfree_trim_params {
2394 	TAILQ_HEAD(, trim_blkreq) blklist;
2395 	LIST_ENTRY(ffs_blkfree_trim_params) hashlist;
2396 	struct task task;
2397 	struct ufsmount *ump;
2398 	struct vnode *devvp;
2399 	ino_t inum;
2400 	ufs2_daddr_t bno;
2401 	long size;
2402 	long key;
2403 };
2404 
2405 static void	ffs_blkfree_trim_completed(struct buf *);
2406 static void	ffs_blkfree_trim_task(void *ctx, int pending __unused);
2407 static struct	ffs_blkfree_trim_params *trim_lookup(struct ufsmount *,
2408 		    struct vnode *, ufs2_daddr_t, long, ino_t, u_long, int);
2409 static void	ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *);
2410 
2411 /*
2412  * Called on trim completion to start a task to free the associated block(s).
2413  */
2414 static void
2415 ffs_blkfree_trim_completed(bp)
2416 	struct buf *bp;
2417 {
2418 	struct ffs_blkfree_trim_params *tp;
2419 
2420 	tp = bp->b_fsprivate1;
2421 	free(bp, M_TRIM);
2422 	TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp);
2423 	taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task);
2424 }
2425 
2426 /*
2427  * Trim completion task that free associated block(s).
2428  */
2429 static void
2430 ffs_blkfree_trim_task(ctx, pending)
2431 	void *ctx;
2432 	int pending;
2433 {
2434 	struct ffs_blkfree_trim_params *tp;
2435 	struct trim_blkreq *blkelm;
2436 	struct ufsmount *ump;
2437 
2438 	tp = ctx;
2439 	ump = tp->ump;
2440 	while ((blkelm = TAILQ_FIRST(&tp->blklist)) != NULL) {
2441 		ffs_blkfree_cg(ump, ump->um_fs, tp->devvp, blkelm->bno,
2442 		    blkelm->size, tp->inum, blkelm->pdephd);
2443 		TAILQ_REMOVE(&tp->blklist, blkelm, blkreqlist);
2444 		free(blkelm, M_TRIM);
2445 	}
2446 	vn_finished_secondary_write(UFSTOVFS(ump));
2447 	UFS_LOCK(ump);
2448 	ump->um_trim_inflight -= 1;
2449 	ump->um_trim_inflight_blks -= numfrags(ump->um_fs, tp->size);
2450 	UFS_UNLOCK(ump);
2451 	free(tp, M_TRIM);
2452 }
2453 
2454 /*
2455  * Lookup a trim request by inode number.
2456  * Allocate if requested (NEW, REPLACE, SINGLE).
2457  */
2458 static struct ffs_blkfree_trim_params *
2459 trim_lookup(ump, devvp, bno, size, inum, key, alloctype)
2460 	struct ufsmount *ump;
2461 	struct vnode *devvp;
2462 	ufs2_daddr_t bno;
2463 	long size;
2464 	ino_t inum;
2465 	u_long key;
2466 	int alloctype;
2467 {
2468 	struct trimlist_hashhead *tphashhead;
2469 	struct ffs_blkfree_trim_params *tp, *ntp;
2470 
2471 	ntp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK);
2472 	if (alloctype != SINGLE) {
2473 		KASSERT(key >= FIRST_VALID_KEY, ("trim_lookup: invalid key"));
2474 		UFS_LOCK(ump);
2475 		tphashhead = TRIMLIST_HASH(ump, key);
2476 		LIST_FOREACH(tp, tphashhead, hashlist)
2477 			if (key == tp->key)
2478 				break;
2479 	}
2480 	switch (alloctype) {
2481 	case NEW:
2482 		KASSERT(tp == NULL, ("trim_lookup: found trim"));
2483 		break;
2484 	case OLD:
2485 		KASSERT(tp != NULL,
2486 		    ("trim_lookup: missing call to ffs_blkrelease_start()"));
2487 		UFS_UNLOCK(ump);
2488 		free(ntp, M_TRIM);
2489 		return (tp);
2490 	case REPLACE:
2491 		KASSERT(tp != NULL, ("trim_lookup: missing REPLACE trim"));
2492 		LIST_REMOVE(tp, hashlist);
2493 		/* tp will be freed by caller */
2494 		break;
2495 	case DONE:
2496 		KASSERT(tp != NULL, ("trim_lookup: missing DONE trim"));
2497 		LIST_REMOVE(tp, hashlist);
2498 		UFS_UNLOCK(ump);
2499 		free(ntp, M_TRIM);
2500 		return (tp);
2501 	}
2502 	TAILQ_INIT(&ntp->blklist);
2503 	ntp->ump = ump;
2504 	ntp->devvp = devvp;
2505 	ntp->bno = bno;
2506 	ntp->size = size;
2507 	ntp->inum = inum;
2508 	ntp->key = key;
2509 	if (alloctype != SINGLE) {
2510 		LIST_INSERT_HEAD(tphashhead, ntp, hashlist);
2511 		UFS_UNLOCK(ump);
2512 	}
2513 	return (ntp);
2514 }
2515 
2516 /*
2517  * Dispatch a trim request.
2518  */
2519 static void
2520 ffs_blkfree_sendtrim(tp)
2521 	struct ffs_blkfree_trim_params *tp;
2522 {
2523 	struct ufsmount *ump;
2524 	struct mount *mp;
2525 	struct buf *bp;
2526 
2527 	/*
2528 	 * Postpone the set of the free bit in the cg bitmap until the
2529 	 * BIO_DELETE is completed.  Otherwise, due to disk queue
2530 	 * reordering, TRIM might be issued after we reuse the block
2531 	 * and write some new data into it.
2532 	 */
2533 	ump = tp->ump;
2534 	bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO);
2535 	bp->b_iocmd = BIO_DELETE;
2536 	bp->b_iooffset = dbtob(fsbtodb(ump->um_fs, tp->bno));
2537 	bp->b_iodone = ffs_blkfree_trim_completed;
2538 	bp->b_bcount = tp->size;
2539 	bp->b_fsprivate1 = tp;
2540 	UFS_LOCK(ump);
2541 	ump->um_trim_total += 1;
2542 	ump->um_trim_inflight += 1;
2543 	ump->um_trim_inflight_blks += numfrags(ump->um_fs, tp->size);
2544 	ump->um_trim_total_blks += numfrags(ump->um_fs, tp->size);
2545 	UFS_UNLOCK(ump);
2546 
2547 	mp = UFSTOVFS(ump);
2548 	vn_start_secondary_write(NULL, &mp, 0);
2549 	g_vfs_strategy(ump->um_bo, bp);
2550 }
2551 
2552 /*
2553  * Allocate a new key to use to identify a range of blocks.
2554  */
2555 u_long
2556 ffs_blkrelease_start(ump, devvp, inum)
2557 	struct ufsmount *ump;
2558 	struct vnode *devvp;
2559 	ino_t inum;
2560 {
2561 	static u_long masterkey;
2562 	u_long key;
2563 
2564 	if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2565 		return (SINGLETON_KEY);
2566 	do {
2567 		key = atomic_fetchadd_long(&masterkey, 1);
2568 	} while (key < FIRST_VALID_KEY);
2569 	(void) trim_lookup(ump, devvp, 0, 0, inum, key, NEW);
2570 	return (key);
2571 }
2572 
2573 /*
2574  * Deallocate a key that has been used to identify a range of blocks.
2575  */
2576 void
2577 ffs_blkrelease_finish(ump, key)
2578 	struct ufsmount *ump;
2579 	u_long key;
2580 {
2581 	struct ffs_blkfree_trim_params *tp;
2582 
2583 	if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2584 		return;
2585 	/*
2586 	 * If the vfs.ffs.dotrimcons sysctl option is enabled while
2587 	 * a file deletion is active, specifically after a call
2588 	 * to ffs_blkrelease_start() but before the call to
2589 	 * ffs_blkrelease_finish(), ffs_blkrelease_start() will
2590 	 * have handed out SINGLETON_KEY rather than starting a
2591 	 * collection sequence. Thus if we get a SINGLETON_KEY
2592 	 * passed to ffs_blkrelease_finish(), we just return rather
2593 	 * than trying to finish the nonexistent sequence.
2594 	 */
2595 	if (key == SINGLETON_KEY) {
2596 #ifdef INVARIANTS
2597 		printf("%s: vfs.ffs.dotrimcons enabled on active filesystem\n",
2598 		    ump->um_mountp->mnt_stat.f_mntonname);
2599 #endif
2600 		return;
2601 	}
2602 	/*
2603 	 * We are done with sending blocks using this key. Look up the key
2604 	 * using the DONE alloctype (in tp) to request that it be unhashed
2605 	 * as we will not be adding to it. If the key has never been used,
2606 	 * tp->size will be zero, so we can just free tp. Otherwise the call
2607 	 * to ffs_blkfree_sendtrim(tp) causes the block range described by
2608 	 * tp to be issued (and then tp to be freed).
2609 	 */
2610 	tp = trim_lookup(ump, NULL, 0, 0, 0, key, DONE);
2611 	if (tp->size == 0)
2612 		free(tp, M_TRIM);
2613 	else
2614 		ffs_blkfree_sendtrim(tp);
2615 }
2616 
2617 /*
2618  * Setup to free a block or fragment.
2619  *
2620  * Check for snapshots that might want to claim the block.
2621  * If trims are requested, prepare a trim request. Attempt to
2622  * aggregate consecutive blocks into a single trim request.
2623  */
2624 void
2625 ffs_blkfree(ump, fs, devvp, bno, size, inum, vtype, dephd, key)
2626 	struct ufsmount *ump;
2627 	struct fs *fs;
2628 	struct vnode *devvp;
2629 	ufs2_daddr_t bno;
2630 	long size;
2631 	ino_t inum;
2632 	enum vtype vtype;
2633 	struct workhead *dephd;
2634 	u_long key;
2635 {
2636 	struct ffs_blkfree_trim_params *tp, *ntp;
2637 	struct trim_blkreq *blkelm;
2638 
2639 	/*
2640 	 * Check to see if a snapshot wants to claim the block.
2641 	 * Check that devvp is a normal disk device, not a snapshot,
2642 	 * it has a snapshot(s) associated with it, and one of the
2643 	 * snapshots wants to claim the block.
2644 	 */
2645 	if (devvp->v_type == VCHR &&
2646 	    (devvp->v_vflag & VV_COPYONWRITE) &&
2647 	    ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) {
2648 		return;
2649 	}
2650 	/*
2651 	 * Nothing to delay if TRIM is not required for this block or TRIM
2652 	 * is disabled or the operation is performed on a snapshot.
2653 	 */
2654 	if (key == NOTRIM_KEY || ((ump->um_flags & UM_CANDELETE) == 0) ||
2655 	    devvp->v_type == VREG) {
2656 		ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd);
2657 		return;
2658 	}
2659 	blkelm = malloc(sizeof(struct trim_blkreq), M_TRIM, M_WAITOK);
2660 	blkelm->bno = bno;
2661 	blkelm->size = size;
2662 	if (dephd == NULL) {
2663 		blkelm->pdephd = NULL;
2664 	} else {
2665 		LIST_INIT(&blkelm->dephd);
2666 		LIST_SWAP(dephd, &blkelm->dephd, worklist, wk_list);
2667 		blkelm->pdephd = &blkelm->dephd;
2668 	}
2669 	if (key == SINGLETON_KEY) {
2670 		/*
2671 		 * Just a single non-contiguous piece. Use the SINGLE
2672 		 * alloctype to return a trim request that will not be
2673 		 * hashed for future lookup.
2674 		 */
2675 		tp = trim_lookup(ump, devvp, bno, size, inum, key, SINGLE);
2676 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2677 		ffs_blkfree_sendtrim(tp);
2678 		return;
2679 	}
2680 	/*
2681 	 * The callers of this function are not tracking whether or not
2682 	 * the blocks are contiguous. They are just saying that they
2683 	 * are freeing a set of blocks. It is this code that determines
2684 	 * the pieces of that range that are actually contiguous.
2685 	 *
2686 	 * Calling ffs_blkrelease_start() will have created an entry
2687 	 * that we will use.
2688 	 */
2689 	tp = trim_lookup(ump, devvp, bno, size, inum, key, OLD);
2690 	if (tp->size == 0) {
2691 		/*
2692 		 * First block of a potential range, set block and size
2693 		 * for the trim block.
2694 		 */
2695 		tp->bno = bno;
2696 		tp->size = size;
2697 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2698 		return;
2699 	}
2700 	/*
2701 	 * If this block is a continuation of the range (either
2702 	 * follows at the end or preceeds in the front) then we
2703 	 * add it to the front or back of the list and return.
2704 	 *
2705 	 * If it is not a continuation of the trim that we were
2706 	 * building, using the REPLACE alloctype, we request that
2707 	 * the old trim request (still in tp) be unhashed and a
2708 	 * new range started (in ntp). The ffs_blkfree_sendtrim(tp)
2709 	 * call causes the block range described by tp to be issued
2710 	 * (and then tp to be freed).
2711 	 */
2712 	if (bno + numfrags(fs, size) == tp->bno) {
2713 		TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2714 		tp->bno = bno;
2715 		tp->size += size;
2716 		return;
2717 	} else if (bno == tp->bno + numfrags(fs, tp->size)) {
2718 		TAILQ_INSERT_TAIL(&tp->blklist, blkelm, blkreqlist);
2719 		tp->size += size;
2720 		return;
2721 	}
2722 	ntp = trim_lookup(ump, devvp, bno, size, inum, key, REPLACE);
2723 	TAILQ_INSERT_HEAD(&ntp->blklist, blkelm, blkreqlist);
2724 	ffs_blkfree_sendtrim(tp);
2725 }
2726 
2727 #ifdef INVARIANTS
2728 /*
2729  * Verify allocation of a block or fragment. Returns true if block or
2730  * fragment is allocated, false if it is free.
2731  */
2732 static int
2733 ffs_checkblk(ip, bno, size)
2734 	struct inode *ip;
2735 	ufs2_daddr_t bno;
2736 	long size;
2737 {
2738 	struct fs *fs;
2739 	struct cg *cgp;
2740 	struct buf *bp;
2741 	ufs1_daddr_t cgbno;
2742 	int i, error, frags, free;
2743 	u_int8_t *blksfree;
2744 
2745 	fs = ITOFS(ip);
2746 	if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
2747 		printf("bsize = %ld, size = %ld, fs = %s\n",
2748 		    (long)fs->fs_bsize, size, fs->fs_fsmnt);
2749 		panic("ffs_checkblk: bad size");
2750 	}
2751 	if ((u_int)bno >= fs->fs_size)
2752 		panic("ffs_checkblk: bad block %jd", (intmax_t)bno);
2753 	error = ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), 0, &bp, &cgp);
2754 	if (error)
2755 		panic("ffs_checkblk: cylinder group read failed");
2756 	blksfree = cg_blksfree(cgp);
2757 	cgbno = dtogd(fs, bno);
2758 	if (size == fs->fs_bsize) {
2759 		free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno));
2760 	} else {
2761 		frags = numfrags(fs, size);
2762 		for (free = 0, i = 0; i < frags; i++)
2763 			if (isset(blksfree, cgbno + i))
2764 				free++;
2765 		if (free != 0 && free != frags)
2766 			panic("ffs_checkblk: partially free fragment");
2767 	}
2768 	brelse(bp);
2769 	return (!free);
2770 }
2771 #endif /* INVARIANTS */
2772 
2773 /*
2774  * Free an inode.
2775  */
2776 int
2777 ffs_vfree(pvp, ino, mode)
2778 	struct vnode *pvp;
2779 	ino_t ino;
2780 	int mode;
2781 {
2782 	struct ufsmount *ump;
2783 
2784 	if (DOINGSOFTDEP(pvp)) {
2785 		softdep_freefile(pvp, ino, mode);
2786 		return (0);
2787 	}
2788 	ump = VFSTOUFS(pvp->v_mount);
2789 	return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL));
2790 }
2791 
2792 /*
2793  * Do the actual free operation.
2794  * The specified inode is placed back in the free map.
2795  */
2796 int
2797 ffs_freefile(ump, fs, devvp, ino, mode, wkhd)
2798 	struct ufsmount *ump;
2799 	struct fs *fs;
2800 	struct vnode *devvp;
2801 	ino_t ino;
2802 	int mode;
2803 	struct workhead *wkhd;
2804 {
2805 	struct cg *cgp;
2806 	struct buf *bp;
2807 	daddr_t dbn;
2808 	int error;
2809 	u_int cg;
2810 	u_int8_t *inosused;
2811 	struct cdev *dev;
2812 	ino_t cgino;
2813 
2814 	cg = ino_to_cg(fs, ino);
2815 	if (devvp->v_type == VREG) {
2816 		/* devvp is a snapshot */
2817 		MPASS(devvp->v_mount->mnt_data == ump);
2818 		dev = ump->um_devvp->v_rdev;
2819 	} else if (devvp->v_type == VCHR) {
2820 		/* devvp is a normal disk device */
2821 		dev = devvp->v_rdev;
2822 	} else {
2823 		bp = NULL;
2824 		return (0);
2825 	}
2826 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2827 		panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s",
2828 		    devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt);
2829 	if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) {
2830 		if (!ffs_fsfail_cleanup(ump, error) ||
2831 		    !MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR)
2832 			return (error);
2833 		if (devvp->v_type == VREG)
2834 			dbn = fragstoblks(fs, cgtod(fs, cg));
2835 		else
2836 			dbn = fsbtodb(fs, cgtod(fs, cg));
2837 		error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp);
2838 		KASSERT(error == 0, ("getblkx failed"));
2839 		softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd);
2840 		bp->b_flags |= B_RELBUF | B_NOCACHE;
2841 		bp->b_flags &= ~B_CACHE;
2842 		bawrite(bp);
2843 		return (error);
2844 	}
2845 	inosused = cg_inosused(cgp);
2846 	cgino = ino % fs->fs_ipg;
2847 	if (isclr(inosused, cgino)) {
2848 		printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev),
2849 		    (uintmax_t)ino, fs->fs_fsmnt);
2850 		if (fs->fs_ronly == 0)
2851 			panic("ffs_freefile: freeing free inode");
2852 	}
2853 	clrbit(inosused, cgino);
2854 	if (cgino < cgp->cg_irotor)
2855 		cgp->cg_irotor = cgino;
2856 	cgp->cg_cs.cs_nifree++;
2857 	UFS_LOCK(ump);
2858 	fs->fs_cstotal.cs_nifree++;
2859 	fs->fs_cs(fs, cg).cs_nifree++;
2860 	if ((mode & IFMT) == IFDIR) {
2861 		cgp->cg_cs.cs_ndir--;
2862 		fs->fs_cstotal.cs_ndir--;
2863 		fs->fs_cs(fs, cg).cs_ndir--;
2864 	}
2865 	fs->fs_fmod = 1;
2866 	ACTIVECLEAR(fs, cg);
2867 	UFS_UNLOCK(ump);
2868 	if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR)
2869 		softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd);
2870 	bdwrite(bp);
2871 	return (0);
2872 }
2873 
2874 /*
2875  * Check to see if a file is free.
2876  * Used to check for allocated files in snapshots.
2877  */
2878 int
2879 ffs_checkfreefile(fs, devvp, ino)
2880 	struct fs *fs;
2881 	struct vnode *devvp;
2882 	ino_t ino;
2883 {
2884 	struct cg *cgp;
2885 	struct buf *bp;
2886 	int ret, error;
2887 	u_int cg;
2888 	u_int8_t *inosused;
2889 
2890 	cg = ino_to_cg(fs, ino);
2891 	if ((devvp->v_type != VREG) && (devvp->v_type != VCHR))
2892 		return (1);
2893 	if (ino >= fs->fs_ipg * fs->fs_ncg)
2894 		return (1);
2895 	if ((error = ffs_getcg(fs, devvp, cg, 0, &bp, &cgp)) != 0)
2896 		return (1);
2897 	inosused = cg_inosused(cgp);
2898 	ino %= fs->fs_ipg;
2899 	ret = isclr(inosused, ino);
2900 	brelse(bp);
2901 	return (ret);
2902 }
2903 
2904 /*
2905  * Find a block of the specified size in the specified cylinder group.
2906  *
2907  * It is a panic if a request is made to find a block if none are
2908  * available.
2909  */
2910 static ufs1_daddr_t
2911 ffs_mapsearch(fs, cgp, bpref, allocsiz)
2912 	struct fs *fs;
2913 	struct cg *cgp;
2914 	ufs2_daddr_t bpref;
2915 	int allocsiz;
2916 {
2917 	ufs1_daddr_t bno;
2918 	int start, len, loc, i;
2919 	int blk, field, subfield, pos;
2920 	u_int8_t *blksfree;
2921 
2922 	/*
2923 	 * find the fragment by searching through the free block
2924 	 * map for an appropriate bit pattern
2925 	 */
2926 	if (bpref)
2927 		start = dtogd(fs, bpref) / NBBY;
2928 	else
2929 		start = cgp->cg_frotor / NBBY;
2930 	blksfree = cg_blksfree(cgp);
2931 	len = howmany(fs->fs_fpg, NBBY) - start;
2932 	loc = scanc((u_int)len, (u_char *)&blksfree[start],
2933 		fragtbl[fs->fs_frag],
2934 		(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2935 	if (loc == 0) {
2936 		len = start + 1;
2937 		start = 0;
2938 		loc = scanc((u_int)len, (u_char *)&blksfree[0],
2939 			fragtbl[fs->fs_frag],
2940 			(u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2941 		if (loc == 0) {
2942 			printf("start = %d, len = %d, fs = %s\n",
2943 			    start, len, fs->fs_fsmnt);
2944 			panic("ffs_alloccg: map corrupted");
2945 			/* NOTREACHED */
2946 		}
2947 	}
2948 	bno = (start + len - loc) * NBBY;
2949 	cgp->cg_frotor = bno;
2950 	/*
2951 	 * found the byte in the map
2952 	 * sift through the bits to find the selected frag
2953 	 */
2954 	for (i = bno + NBBY; bno < i; bno += fs->fs_frag) {
2955 		blk = blkmap(fs, blksfree, bno);
2956 		blk <<= 1;
2957 		field = around[allocsiz];
2958 		subfield = inside[allocsiz];
2959 		for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) {
2960 			if ((blk & field) == subfield)
2961 				return (bno + pos);
2962 			field <<= 1;
2963 			subfield <<= 1;
2964 		}
2965 	}
2966 	printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt);
2967 	panic("ffs_alloccg: block not in map");
2968 	return (-1);
2969 }
2970 
2971 static const struct statfs *
2972 ffs_getmntstat(struct vnode *devvp)
2973 {
2974 
2975 	if (devvp->v_type == VCHR)
2976 		return (&devvp->v_rdev->si_mountpt->mnt_stat);
2977 	return (ffs_getmntstat(VFSTOUFS(devvp->v_mount)->um_devvp));
2978 }
2979 
2980 /*
2981  * Fetch and verify a cylinder group.
2982  */
2983 int
2984 ffs_getcg(fs, devvp, cg, flags, bpp, cgpp)
2985 	struct fs *fs;
2986 	struct vnode *devvp;
2987 	u_int cg;
2988 	int flags;
2989 	struct buf **bpp;
2990 	struct cg **cgpp;
2991 {
2992 	struct buf *bp;
2993 	struct cg *cgp;
2994 	const struct statfs *sfs;
2995 	daddr_t blkno;
2996 	int error;
2997 
2998 	*bpp = NULL;
2999 	*cgpp = NULL;
3000 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
3001 		flags |= GB_CKHASH;
3002 	if (devvp->v_type == VREG)
3003 		blkno = fragstoblks(fs, cgtod(fs, cg));
3004 	else
3005 		blkno = fsbtodb(fs, cgtod(fs, cg));
3006 	error = breadn_flags(devvp, blkno, blkno, (int)fs->fs_cgsize, NULL,
3007 	    NULL, 0, NOCRED, flags, ffs_ckhash_cg, &bp);
3008 	if (error != 0)
3009 		return (error);
3010 	cgp = (struct cg *)bp->b_data;
3011 	if ((fs->fs_metackhash & CK_CYLGRP) != 0 &&
3012 	    (bp->b_flags & B_CKHASH) != 0 &&
3013 	    cgp->cg_ckhash != bp->b_ckhash) {
3014 		sfs = ffs_getmntstat(devvp);
3015 		printf("UFS %s%s (%s) cylinder checksum failed: cg %u, cgp: "
3016 		    "0x%x != bp: 0x%jx\n",
3017 		    devvp->v_type == VCHR ? "" : "snapshot of ",
3018 		    sfs->f_mntfromname, sfs->f_mntonname,
3019 		    cg, cgp->cg_ckhash, (uintmax_t)bp->b_ckhash);
3020 		bp->b_flags &= ~B_CKHASH;
3021 		bp->b_flags |= B_INVAL | B_NOCACHE;
3022 		brelse(bp);
3023 		return (EIO);
3024 	}
3025 	if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) {
3026 		sfs = ffs_getmntstat(devvp);
3027 		printf("UFS %s%s (%s)",
3028 		    devvp->v_type == VCHR ? "" : "snapshot of ",
3029 		    sfs->f_mntfromname, sfs->f_mntonname);
3030 		if (!cg_chkmagic(cgp))
3031 			printf(" cg %u: bad magic number 0x%x should be 0x%x\n",
3032 			    cg, cgp->cg_magic, CG_MAGIC);
3033 		else
3034 			printf(": wrong cylinder group cg %u != cgx %u\n", cg,
3035 			    cgp->cg_cgx);
3036 		bp->b_flags &= ~B_CKHASH;
3037 		bp->b_flags |= B_INVAL | B_NOCACHE;
3038 		brelse(bp);
3039 		return (EIO);
3040 	}
3041 	bp->b_flags &= ~B_CKHASH;
3042 	bp->b_xflags |= BX_BKGRDWRITE;
3043 	/*
3044 	 * If we are using check hashes on the cylinder group then we want
3045 	 * to limit changing the cylinder group time to when we are actually
3046 	 * going to write it to disk so that its check hash remains correct
3047 	 * in memory. If the CK_CYLGRP flag is set the time is updated in
3048 	 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we
3049 	 * update the time here as we have done historically.
3050 	 */
3051 	if ((fs->fs_metackhash & CK_CYLGRP) != 0)
3052 		bp->b_xflags |= BX_CYLGRP;
3053 	else
3054 		cgp->cg_old_time = cgp->cg_time = time_second;
3055 	*bpp = bp;
3056 	*cgpp = cgp;
3057 	return (0);
3058 }
3059 
3060 static void
3061 ffs_ckhash_cg(bp)
3062 	struct buf *bp;
3063 {
3064 	uint32_t ckhash;
3065 	struct cg *cgp;
3066 
3067 	cgp = (struct cg *)bp->b_data;
3068 	ckhash = cgp->cg_ckhash;
3069 	cgp->cg_ckhash = 0;
3070 	bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount);
3071 	cgp->cg_ckhash = ckhash;
3072 }
3073 
3074 /*
3075  * Fserr prints the name of a filesystem with an error diagnostic.
3076  *
3077  * The form of the error message is:
3078  *	fs: error message
3079  */
3080 void
3081 ffs_fserr(fs, inum, cp)
3082 	struct fs *fs;
3083 	ino_t inum;
3084 	char *cp;
3085 {
3086 	struct thread *td = curthread;	/* XXX */
3087 	struct proc *p = td->td_proc;
3088 
3089 	log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n",
3090 	    p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum,
3091 	    fs->fs_fsmnt, cp);
3092 }
3093 
3094 /*
3095  * This function provides the capability for the fsck program to
3096  * update an active filesystem. Fourteen operations are provided:
3097  *
3098  * adjrefcnt(inode, amt) - adjusts the reference count on the
3099  *	specified inode by the specified amount. Under normal
3100  *	operation the count should always go down. Decrementing
3101  *	the count to zero will cause the inode to be freed.
3102  * adjblkcnt(inode, amt) - adjust the number of blocks used by the
3103  *	inode by the specified amount.
3104  * setsize(inode, size) - set the size of the inode to the
3105  *	specified size.
3106  * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) -
3107  *	adjust the superblock summary.
3108  * freedirs(inode, count) - directory inodes [inode..inode + count - 1]
3109  *	are marked as free. Inodes should never have to be marked
3110  *	as in use.
3111  * freefiles(inode, count) - file inodes [inode..inode + count - 1]
3112  *	are marked as free. Inodes should never have to be marked
3113  *	as in use.
3114  * freeblks(blockno, size) - blocks [blockno..blockno + size - 1]
3115  *	are marked as free. Blocks should never have to be marked
3116  *	as in use.
3117  * setflags(flags, set/clear) - the fs_flags field has the specified
3118  *	flags set (second parameter +1) or cleared (second parameter -1).
3119  * setcwd(dirinode) - set the current directory to dirinode in the
3120  *	filesystem associated with the snapshot.
3121  * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".."
3122  *	in the current directory is oldvalue then change it to newvalue.
3123  * unlink(nameptr, oldvalue) - Verify that the inode number associated
3124  *	with nameptr in the current directory is oldvalue then unlink it.
3125  */
3126 
3127 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS);
3128 
3129 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt,
3130     CTLFLAG_WR | CTLTYPE_STRUCT | CTLFLAG_NEEDGIANT,
3131     0, 0, sysctl_ffs_fsck, "S,fsck",
3132     "Adjust Inode Reference Count");
3133 
3134 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt,
3135     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3136     "Adjust Inode Used Blocks Count");
3137 
3138 static SYSCTL_NODE(_vfs_ffs, FFS_SET_SIZE, setsize,
3139     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3140     "Set the inode size");
3141 
3142 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir,
3143     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3144     "Adjust number of directories");
3145 
3146 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree,
3147     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3148     "Adjust number of free blocks");
3149 
3150 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree,
3151     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3152     "Adjust number of free inodes");
3153 
3154 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree,
3155     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3156     "Adjust number of free frags");
3157 
3158 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters,
3159     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3160     "Adjust number of free clusters");
3161 
3162 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs,
3163     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3164     "Free Range of Directory Inodes");
3165 
3166 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles,
3167     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3168     "Free Range of File Inodes");
3169 
3170 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks,
3171     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3172     "Free Range of Blocks");
3173 
3174 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags,
3175     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3176     "Change Filesystem Flags");
3177 
3178 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd,
3179     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3180     "Set Current Working Directory");
3181 
3182 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot,
3183     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3184     "Change Value of .. Entry");
3185 
3186 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink,
3187     CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck,
3188     "Unlink a Duplicate Name");
3189 
3190 #ifdef DIAGNOSTIC
3191 static int fsckcmds = 0;
3192 SYSCTL_INT(_debug, OID_AUTO, ffs_fsckcmds, CTLFLAG_RW, &fsckcmds, 0,
3193 	"print out fsck_ffs-based filesystem update commands");
3194 #endif /* DIAGNOSTIC */
3195 
3196 static int
3197 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS)
3198 {
3199 	struct thread *td = curthread;
3200 	struct fsck_cmd cmd;
3201 	struct ufsmount *ump;
3202 	struct vnode *vp, *dvp, *fdvp;
3203 	struct inode *ip, *dp;
3204 	struct mount *mp;
3205 	struct fs *fs;
3206 	struct pwd *pwd;
3207 	ufs2_daddr_t blkno;
3208 	long blkcnt, blksize;
3209 	u_long key;
3210 	struct file *fp;
3211 	cap_rights_t rights;
3212 	int filetype, error;
3213 
3214 	if (req->newlen > sizeof cmd)
3215 		return (EBADRPC);
3216 	if ((error = SYSCTL_IN(req, &cmd, sizeof cmd)) != 0)
3217 		return (error);
3218 	if (cmd.version != FFS_CMD_VERSION)
3219 		return (ERPCMISMATCH);
3220 	if ((error = getvnode(td, cmd.handle,
3221 	    cap_rights_init(&rights, CAP_FSCK), &fp)) != 0)
3222 		return (error);
3223 	vp = fp->f_vnode;
3224 	if (vp->v_type != VREG && vp->v_type != VDIR) {
3225 		fdrop(fp, td);
3226 		return (EINVAL);
3227 	}
3228 	vn_start_write(vp, &mp, V_WAIT);
3229 	if (mp == NULL ||
3230 	    strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) {
3231 		vn_finished_write(mp);
3232 		fdrop(fp, td);
3233 		return (EINVAL);
3234 	}
3235 	ump = VFSTOUFS(mp);
3236 	if ((mp->mnt_flag & MNT_RDONLY) &&
3237 	    ump->um_fsckpid != td->td_proc->p_pid) {
3238 		vn_finished_write(mp);
3239 		fdrop(fp, td);
3240 		return (EROFS);
3241 	}
3242 	fs = ump->um_fs;
3243 	filetype = IFREG;
3244 
3245 	switch (oidp->oid_number) {
3246 	case FFS_SET_FLAGS:
3247 #ifdef DIAGNOSTIC
3248 		if (fsckcmds)
3249 			printf("%s: %s flags\n", mp->mnt_stat.f_mntonname,
3250 			    cmd.size > 0 ? "set" : "clear");
3251 #endif /* DIAGNOSTIC */
3252 		if (cmd.size > 0)
3253 			fs->fs_flags |= (long)cmd.value;
3254 		else
3255 			fs->fs_flags &= ~(long)cmd.value;
3256 		break;
3257 
3258 	case FFS_ADJ_REFCNT:
3259 #ifdef DIAGNOSTIC
3260 		if (fsckcmds) {
3261 			printf("%s: adjust inode %jd link count by %jd\n",
3262 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3263 			    (intmax_t)cmd.size);
3264 		}
3265 #endif /* DIAGNOSTIC */
3266 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3267 			break;
3268 		ip = VTOI(vp);
3269 		ip->i_nlink += cmd.size;
3270 		DIP_SET(ip, i_nlink, ip->i_nlink);
3271 		ip->i_effnlink += cmd.size;
3272 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED);
3273 		error = ffs_update(vp, 1);
3274 		if (DOINGSOFTDEP(vp))
3275 			softdep_change_linkcnt(ip);
3276 		vput(vp);
3277 		break;
3278 
3279 	case FFS_ADJ_BLKCNT:
3280 #ifdef DIAGNOSTIC
3281 		if (fsckcmds) {
3282 			printf("%s: adjust inode %jd block count by %jd\n",
3283 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3284 			    (intmax_t)cmd.size);
3285 		}
3286 #endif /* DIAGNOSTIC */
3287 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3288 			break;
3289 		ip = VTOI(vp);
3290 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size);
3291 		UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED);
3292 		error = ffs_update(vp, 1);
3293 		vput(vp);
3294 		break;
3295 
3296 	case FFS_SET_SIZE:
3297 #ifdef DIAGNOSTIC
3298 		if (fsckcmds) {
3299 			printf("%s: set inode %jd size to %jd\n",
3300 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3301 			    (intmax_t)cmd.size);
3302 		}
3303 #endif /* DIAGNOSTIC */
3304 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3305 			break;
3306 		ip = VTOI(vp);
3307 		DIP_SET(ip, i_size, cmd.size);
3308 		UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE | IN_MODIFIED);
3309 		error = ffs_update(vp, 1);
3310 		vput(vp);
3311 		break;
3312 
3313 	case FFS_DIR_FREE:
3314 		filetype = IFDIR;
3315 		/* fall through */
3316 
3317 	case FFS_FILE_FREE:
3318 #ifdef DIAGNOSTIC
3319 		if (fsckcmds) {
3320 			if (cmd.size == 1)
3321 				printf("%s: free %s inode %ju\n",
3322 				    mp->mnt_stat.f_mntonname,
3323 				    filetype == IFDIR ? "directory" : "file",
3324 				    (uintmax_t)cmd.value);
3325 			else
3326 				printf("%s: free %s inodes %ju-%ju\n",
3327 				    mp->mnt_stat.f_mntonname,
3328 				    filetype == IFDIR ? "directory" : "file",
3329 				    (uintmax_t)cmd.value,
3330 				    (uintmax_t)(cmd.value + cmd.size - 1));
3331 		}
3332 #endif /* DIAGNOSTIC */
3333 		while (cmd.size > 0) {
3334 			if ((error = ffs_freefile(ump, fs, ump->um_devvp,
3335 			    cmd.value, filetype, NULL)))
3336 				break;
3337 			cmd.size -= 1;
3338 			cmd.value += 1;
3339 		}
3340 		break;
3341 
3342 	case FFS_BLK_FREE:
3343 #ifdef DIAGNOSTIC
3344 		if (fsckcmds) {
3345 			if (cmd.size == 1)
3346 				printf("%s: free block %jd\n",
3347 				    mp->mnt_stat.f_mntonname,
3348 				    (intmax_t)cmd.value);
3349 			else
3350 				printf("%s: free blocks %jd-%jd\n",
3351 				    mp->mnt_stat.f_mntonname,
3352 				    (intmax_t)cmd.value,
3353 				    (intmax_t)cmd.value + cmd.size - 1);
3354 		}
3355 #endif /* DIAGNOSTIC */
3356 		blkno = cmd.value;
3357 		blkcnt = cmd.size;
3358 		blksize = fs->fs_frag - (blkno % fs->fs_frag);
3359 		key = ffs_blkrelease_start(ump, ump->um_devvp, UFS_ROOTINO);
3360 		while (blkcnt > 0) {
3361 			if (blkcnt < blksize)
3362 				blksize = blkcnt;
3363 			ffs_blkfree(ump, fs, ump->um_devvp, blkno,
3364 			    blksize * fs->fs_fsize, UFS_ROOTINO,
3365 			    VDIR, NULL, key);
3366 			blkno += blksize;
3367 			blkcnt -= blksize;
3368 			blksize = fs->fs_frag;
3369 		}
3370 		ffs_blkrelease_finish(ump, key);
3371 		break;
3372 
3373 	/*
3374 	 * Adjust superblock summaries.  fsck(8) is expected to
3375 	 * submit deltas when necessary.
3376 	 */
3377 	case FFS_ADJ_NDIR:
3378 #ifdef DIAGNOSTIC
3379 		if (fsckcmds) {
3380 			printf("%s: adjust number of directories by %jd\n",
3381 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3382 		}
3383 #endif /* DIAGNOSTIC */
3384 		fs->fs_cstotal.cs_ndir += cmd.value;
3385 		break;
3386 
3387 	case FFS_ADJ_NBFREE:
3388 #ifdef DIAGNOSTIC
3389 		if (fsckcmds) {
3390 			printf("%s: adjust number of free blocks by %+jd\n",
3391 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3392 		}
3393 #endif /* DIAGNOSTIC */
3394 		fs->fs_cstotal.cs_nbfree += cmd.value;
3395 		break;
3396 
3397 	case FFS_ADJ_NIFREE:
3398 #ifdef DIAGNOSTIC
3399 		if (fsckcmds) {
3400 			printf("%s: adjust number of free inodes by %+jd\n",
3401 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3402 		}
3403 #endif /* DIAGNOSTIC */
3404 		fs->fs_cstotal.cs_nifree += cmd.value;
3405 		break;
3406 
3407 	case FFS_ADJ_NFFREE:
3408 #ifdef DIAGNOSTIC
3409 		if (fsckcmds) {
3410 			printf("%s: adjust number of free frags by %+jd\n",
3411 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3412 		}
3413 #endif /* DIAGNOSTIC */
3414 		fs->fs_cstotal.cs_nffree += cmd.value;
3415 		break;
3416 
3417 	case FFS_ADJ_NUMCLUSTERS:
3418 #ifdef DIAGNOSTIC
3419 		if (fsckcmds) {
3420 			printf("%s: adjust number of free clusters by %+jd\n",
3421 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3422 		}
3423 #endif /* DIAGNOSTIC */
3424 		fs->fs_cstotal.cs_numclusters += cmd.value;
3425 		break;
3426 
3427 	case FFS_SET_CWD:
3428 #ifdef DIAGNOSTIC
3429 		if (fsckcmds) {
3430 			printf("%s: set current directory to inode %jd\n",
3431 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3432 		}
3433 #endif /* DIAGNOSTIC */
3434 		if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp)))
3435 			break;
3436 		AUDIT_ARG_VNODE1(vp);
3437 		if ((error = change_dir(vp, td)) != 0) {
3438 			vput(vp);
3439 			break;
3440 		}
3441 		VOP_UNLOCK(vp);
3442 		pwd_chdir(td, vp);
3443 		break;
3444 
3445 	case FFS_SET_DOTDOT:
3446 #ifdef DIAGNOSTIC
3447 		if (fsckcmds) {
3448 			printf("%s: change .. in cwd from %jd to %jd\n",
3449 			    mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3450 			    (intmax_t)cmd.size);
3451 		}
3452 #endif /* DIAGNOSTIC */
3453 		/*
3454 		 * First we have to get and lock the parent directory
3455 		 * to which ".." points.
3456 		 */
3457 		error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp);
3458 		if (error)
3459 			break;
3460 		/*
3461 		 * Now we get and lock the child directory containing "..".
3462 		 */
3463 		pwd = pwd_hold(td);
3464 		dvp = pwd->pwd_cdir;
3465 		if ((error = vget(dvp, LK_EXCLUSIVE)) != 0) {
3466 			vput(fdvp);
3467 			pwd_drop(pwd);
3468 			break;
3469 		}
3470 		dp = VTOI(dvp);
3471 		SET_I_OFFSET(dp, 12);	/* XXX mastertemplate.dot_reclen */
3472 		error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size,
3473 		    DT_DIR, 0);
3474 		cache_purge(fdvp);
3475 		cache_purge(dvp);
3476 		vput(dvp);
3477 		vput(fdvp);
3478 		pwd_drop(pwd);
3479 		break;
3480 
3481 	case FFS_UNLINK:
3482 #ifdef DIAGNOSTIC
3483 		if (fsckcmds) {
3484 			char buf[32];
3485 
3486 			if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL))
3487 				strncpy(buf, "Name_too_long", 32);
3488 			printf("%s: unlink %s (inode %jd)\n",
3489 			    mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size);
3490 		}
3491 #endif /* DIAGNOSTIC */
3492 		/*
3493 		 * kern_funlinkat will do its own start/finish writes and
3494 		 * they do not nest, so drop ours here. Setting mp == NULL
3495 		 * indicates that vn_finished_write is not needed down below.
3496 		 */
3497 		vn_finished_write(mp);
3498 		mp = NULL;
3499 		error = kern_funlinkat(td, AT_FDCWD,
3500 		    (char *)(intptr_t)cmd.value, FD_NONE, UIO_USERSPACE,
3501 		    0, (ino_t)cmd.size);
3502 		break;
3503 
3504 	default:
3505 #ifdef DIAGNOSTIC
3506 		if (fsckcmds) {
3507 			printf("Invalid request %d from fsck\n",
3508 			    oidp->oid_number);
3509 		}
3510 #endif /* DIAGNOSTIC */
3511 		error = EINVAL;
3512 		break;
3513 	}
3514 	fdrop(fp, td);
3515 	vn_finished_write(mp);
3516 	return (error);
3517 }
3518