2 * SPDX-License-Identifier: (BSD-2-Clause-FreeBSD AND BSD-3-Clause)
4 * Copyright (c) 2002 Networks Associates Technology, Inc.
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
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
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.
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
34 * Copyright (c) 1982, 1986, 1989, 1993
35 * The Regents of the University of California. All rights reserved.
37 * Redistribution and use in source and binary forms, with or without
38 * modification, are permitted provided that the following conditions
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.
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
61 * @(#)ffs_alloc.c 8.18 (Berkeley) 5/26/95
64 #include <sys/cdefs.h>
65 __FBSDID("$FreeBSD$");
67 #include "opt_quota.h"
69 #include <sys/param.h>
70 #include <sys/capsicum.h>
71 #include <sys/systm.h>
75 #include <sys/fcntl.h>
77 #include <sys/filedesc.h>
80 #include <sys/vnode.h>
81 #include <sys/mount.h>
82 #include <sys/kernel.h>
83 #include <sys/syscallsubr.h>
84 #include <sys/sysctl.h>
85 #include <sys/syslog.h>
86 #include <sys/taskqueue.h>
88 #include <security/audit/audit.h>
90 #include <geom/geom.h>
91 #include <geom/geom_vfs.h>
93 #include <ufs/ufs/dir.h>
94 #include <ufs/ufs/extattr.h>
95 #include <ufs/ufs/quota.h>
96 #include <ufs/ufs/inode.h>
97 #include <ufs/ufs/ufs_extern.h>
98 #include <ufs/ufs/ufsmount.h>
100 #include <ufs/ffs/fs.h>
101 #include <ufs/ffs/ffs_extern.h>
102 #include <ufs/ffs/softdep.h>
104 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref,
105 int size, int rsize);
107 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int);
109 ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int);
110 static void ffs_blkfree_cg(struct ufsmount *, struct fs *,
111 struct vnode *, ufs2_daddr_t, long, ino_t,
114 static int ffs_checkblk(struct inode *, ufs2_daddr_t, long);
116 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int);
117 static ino_t ffs_dirpref(struct inode *);
118 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t,
120 static ufs2_daddr_t ffs_hashalloc
121 (struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *);
122 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int,
124 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int);
125 static int ffs_reallocblks_ufs1(struct vop_reallocblks_args *);
126 static int ffs_reallocblks_ufs2(struct vop_reallocblks_args *);
127 static void ffs_ckhash_cg(struct buf *);
130 * Allocate a block in the filesystem.
132 * The size of the requested block is given, which must be some
133 * multiple of fs_fsize and <= fs_bsize.
134 * A preference may be optionally specified. If a preference is given
135 * the following hierarchy is used to allocate a block:
136 * 1) allocate the requested block.
137 * 2) allocate a rotationally optimal block in the same cylinder.
138 * 3) allocate a block in the same cylinder group.
139 * 4) quadradically rehash into other cylinder groups, until an
140 * available block is located.
141 * If no block preference is given the following hierarchy is used
142 * to allocate a block:
143 * 1) allocate a block in the cylinder group that contains the
144 * inode for the file.
145 * 2) quadradically rehash into other cylinder groups, until an
146 * available block is located.
149 ffs_alloc(ip, lbn, bpref, size, flags, cred, bnp)
151 ufs2_daddr_t lbn, bpref;
157 struct ufsmount *ump;
160 static struct timeval lastfail;
170 mtx_assert(UFS_MTX(ump), MA_OWNED);
172 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
173 printf("dev = %s, bsize = %ld, size = %d, fs = %s\n",
174 devtoname(ump->um_dev), (long)fs->fs_bsize, size,
176 panic("ffs_alloc: bad size");
179 panic("ffs_alloc: missing credential");
180 #endif /* INVARIANTS */
185 error = chkdq(ip, btodb(size), cred, 0);
190 if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0)
192 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) &&
193 freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0)
195 if (bpref >= fs->fs_size)
198 cg = ino_to_cg(fs, ip->i_number);
200 cg = dtog(fs, bpref);
201 bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg);
204 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
206 ip->i_flag |= IN_CHANGE;
208 ip->i_flag |= IN_CHANGE | IN_UPDATE;
216 * Restore user's disk quota because allocation failed.
218 (void) chkdq(ip, -btodb(size), cred, FORCE);
221 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
223 softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT);
227 if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
228 ffs_fserr(fs, ip->i_number, "filesystem full");
229 uprintf("\n%s: write failed, filesystem is full\n",
236 * Reallocate a fragment to a bigger size
238 * The number and size of the old block is given, and a preference
239 * and new size is also specified. The allocator attempts to extend
240 * the original block. Failing that, the regular block allocator is
241 * invoked to get an appropriate block.
244 ffs_realloccg(ip, lbprev, bprev, bpref, osize, nsize, flags, cred, bpp)
249 int osize, nsize, flags;
256 struct ufsmount *ump;
257 u_int cg, request, reclaimed;
260 static struct timeval lastfail;
268 gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
270 mtx_assert(UFS_MTX(ump), MA_OWNED);
272 if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
273 panic("ffs_realloccg: allocation on suspended filesystem");
274 if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 ||
275 (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) {
277 "dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n",
278 devtoname(ump->um_dev), (long)fs->fs_bsize, osize,
279 nsize, fs->fs_fsmnt);
280 panic("ffs_realloccg: bad size");
283 panic("ffs_realloccg: missing credential");
284 #endif /* INVARIANTS */
287 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) &&
288 freespace(fs, fs->fs_minfree) - numfrags(fs, nsize - osize) < 0) {
292 printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n",
293 devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev,
295 panic("ffs_realloccg: bad bprev");
299 * Allocate the extra space in the buffer.
301 error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp);
307 if (bp->b_blkno == bp->b_lblkno) {
308 if (lbprev >= UFS_NDADDR)
309 panic("ffs_realloccg: lbprev out of range");
310 bp->b_blkno = fsbtodb(fs, bprev);
314 error = chkdq(ip, btodb(nsize - osize), cred, 0);
321 * Check for extension in the existing location.
324 cg = dtog(fs, bprev);
326 bno = ffs_fragextend(ip, cg, bprev, osize, nsize);
328 if (bp->b_blkno != fsbtodb(fs, bno))
329 panic("ffs_realloccg: bad blockno");
330 delta = btodb(nsize - osize);
331 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
333 ip->i_flag |= IN_CHANGE;
335 ip->i_flag |= IN_CHANGE | IN_UPDATE;
337 bp->b_flags |= B_DONE;
338 vfs_bio_bzero_buf(bp, osize, nsize - osize);
339 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
340 vfs_bio_set_valid(bp, osize, nsize - osize);
345 * Allocate a new disk location.
347 if (bpref >= fs->fs_size)
349 switch ((int)fs->fs_optim) {
352 * Allocate an exact sized fragment. Although this makes
353 * best use of space, we will waste time relocating it if
354 * the file continues to grow. If the fragmentation is
355 * less than half of the minimum free reserve, we choose
356 * to begin optimizing for time.
359 if (fs->fs_minfree <= 5 ||
360 fs->fs_cstotal.cs_nffree >
361 (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100))
363 log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n",
365 fs->fs_optim = FS_OPTTIME;
369 * At this point we have discovered a file that is trying to
370 * grow a small fragment to a larger fragment. To save time,
371 * we allocate a full sized block, then free the unused portion.
372 * If the file continues to grow, the `ffs_fragextend' call
373 * above will be able to grow it in place without further
374 * copying. If aberrant programs cause disk fragmentation to
375 * grow within 2% of the free reserve, we choose to begin
376 * optimizing for space.
378 request = fs->fs_bsize;
379 if (fs->fs_cstotal.cs_nffree <
380 (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100)
382 log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n",
384 fs->fs_optim = FS_OPTSPACE;
387 printf("dev = %s, optim = %ld, fs = %s\n",
388 devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt);
389 panic("ffs_realloccg: bad optim");
392 bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg);
394 bp->b_blkno = fsbtodb(fs, bno);
395 if (!DOINGSOFTDEP(vp))
397 * The usual case is that a smaller fragment that
398 * was just allocated has been replaced with a bigger
399 * fragment or a full-size block. If it is marked as
400 * B_DELWRI, the current contents have not been written
401 * to disk. It is possible that the block was written
402 * earlier, but very uncommon. If the block has never
403 * been written, there is no need to send a BIO_DELETE
404 * for it when it is freed. The gain from avoiding the
405 * TRIMs for the common case of unwritten blocks far
406 * exceeds the cost of the write amplification for the
407 * uncommon case of failing to send a TRIM for a block
408 * that had been written.
410 ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize,
411 ip->i_number, vp->v_type, NULL,
412 (bp->b_flags & B_DELWRI) != 0 ?
413 NOTRIM_KEY : SINGLETON_KEY);
414 delta = btodb(nsize - osize);
415 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta);
417 ip->i_flag |= IN_CHANGE;
419 ip->i_flag |= IN_CHANGE | IN_UPDATE;
421 bp->b_flags |= B_DONE;
422 vfs_bio_bzero_buf(bp, osize, nsize - osize);
423 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO)
424 vfs_bio_set_valid(bp, osize, nsize - osize);
431 * Restore user's disk quota because allocation failed.
433 (void) chkdq(ip, -btodb(nsize - osize), cred, FORCE);
440 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) {
448 softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT);
454 if (reclaimed > 0 && ppsratecheck(&lastfail, &curfail, 1)) {
455 ffs_fserr(fs, ip->i_number, "filesystem full");
456 uprintf("\n%s: write failed, filesystem is full\n",
463 * Reallocate a sequence of blocks into a contiguous sequence of blocks.
465 * The vnode and an array of buffer pointers for a range of sequential
466 * logical blocks to be made contiguous is given. The allocator attempts
467 * to find a range of sequential blocks starting as close as possible
468 * from the end of the allocation for the logical block immediately
469 * preceding the current range. If successful, the physical block numbers
470 * in the buffer pointers and in the inode are changed to reflect the new
471 * allocation. If unsuccessful, the allocation is left unchanged. The
472 * success in doing the reallocation is returned. Note that the error
473 * return is not reflected back to the user. Rather the previous block
474 * allocation will be used.
477 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW, 0, "FFS filesystem");
479 static int doasyncfree = 1;
480 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0,
481 "do not force synchronous writes when blocks are reallocated");
483 static int doreallocblks = 1;
484 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0,
485 "enable block reallocation");
487 static int dotrimcons = 1;
488 SYSCTL_INT(_vfs_ffs, OID_AUTO, dotrimcons, CTLFLAG_RWTUN, &dotrimcons, 0,
489 "enable BIO_DELETE / TRIM consolidation");
491 static int maxclustersearch = 10;
492 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch,
493 0, "max number of cylinder group to search for contigous blocks");
496 static volatile int prtrealloc = 0;
501 struct vop_reallocblks_args /* {
503 struct cluster_save *a_buflist;
506 struct ufsmount *ump;
509 * We used to skip reallocating the blocks of a file into a
510 * contiguous sequence if the underlying flash device requested
511 * BIO_DELETE notifications, because devices that benefit from
512 * BIO_DELETE also benefit from not moving the data. However,
513 * the destination for the data is usually moved before the data
514 * is written to the initially allocated location, so we rarely
515 * suffer the penalty of extra writes. With the addition of the
516 * consolidation of contiguous blocks into single BIO_DELETE
517 * operations, having fewer but larger contiguous blocks reduces
518 * the number of (slow and expensive) BIO_DELETE operations. So
519 * when doing BIO_DELETE consolidation, we do block reallocation.
521 * Skip if reallocblks has been disabled globally.
523 ump = ap->a_vp->v_mount->mnt_data;
524 if ((((ump->um_flags) & UM_CANDELETE) != 0 && dotrimcons == 0) ||
529 * We can't wait in softdep prealloc as it may fsync and recurse
530 * here. Instead we simply fail to reallocate blocks if this
531 * rare condition arises.
533 if (DOINGSOFTDEP(ap->a_vp))
534 if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0)
536 if (ump->um_fstype == UFS1)
537 return (ffs_reallocblks_ufs1(ap));
538 return (ffs_reallocblks_ufs2(ap));
542 ffs_reallocblks_ufs1(ap)
543 struct vop_reallocblks_args /* {
545 struct cluster_save *a_buflist;
551 struct buf *sbp, *ebp, *bp;
552 ufs1_daddr_t *bap, *sbap, *ebap;
553 struct cluster_save *buflist;
554 struct ufsmount *ump;
555 ufs_lbn_t start_lbn, end_lbn;
556 ufs1_daddr_t soff, newblk, blkno;
558 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
559 int i, cg, len, start_lvl, end_lvl, ssize;
566 * If we are not tracking block clusters or if we have less than 4%
567 * free blocks left, then do not attempt to cluster. Running with
568 * less than 5% free block reserve is not recommended and those that
569 * choose to do so do not expect to have good file layout.
571 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
573 buflist = ap->a_buflist;
574 len = buflist->bs_nchildren;
575 start_lbn = buflist->bs_children[0]->b_lblkno;
576 end_lbn = start_lbn + len - 1;
578 for (i = 0; i < len; i++)
579 if (!ffs_checkblk(ip,
580 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
581 panic("ffs_reallocblks: unallocated block 1");
582 for (i = 1; i < len; i++)
583 if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
584 panic("ffs_reallocblks: non-logical cluster");
585 blkno = buflist->bs_children[0]->b_blkno;
586 ssize = fsbtodb(fs, fs->fs_frag);
587 for (i = 1; i < len - 1; i++)
588 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
589 panic("ffs_reallocblks: non-physical cluster %d", i);
592 * If the cluster crosses the boundary for the first indirect
593 * block, leave space for the indirect block. Indirect blocks
594 * are initially laid out in a position after the last direct
595 * block. Block reallocation would usually destroy locality by
596 * moving the indirect block out of the way to make room for
597 * data blocks if we didn't compensate here. We should also do
598 * this for other indirect block boundaries, but it is only
599 * important for the first one.
601 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
604 * If the latest allocation is in a new cylinder group, assume that
605 * the filesystem has decided to move and do not force it back to
606 * the previous cylinder group.
608 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
609 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
611 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
612 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
615 * Get the starting offset and block map for the first block.
617 if (start_lvl == 0) {
618 sbap = &ip->i_din1->di_db[0];
621 idp = &start_ap[start_lvl - 1];
622 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
626 sbap = (ufs1_daddr_t *)sbp->b_data;
630 * If the block range spans two block maps, get the second map.
633 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
638 start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
639 panic("ffs_reallocblk: start == end");
641 ssize = len - (idp->in_off + 1);
642 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
644 ebap = (ufs1_daddr_t *)ebp->b_data;
647 * Find the preferred location for the cluster. If we have not
648 * previously failed at this endeavor, then follow our standard
649 * preference calculation. If we have failed at it, then pick up
650 * where we last ended our search.
653 if (ip->i_nextclustercg == -1)
654 pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap);
656 pref = cgdata(fs, ip->i_nextclustercg);
658 * Search the block map looking for an allocation of the desired size.
659 * To avoid wasting too much time, we limit the number of cylinder
660 * groups that we will search.
663 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
664 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
667 if (cg >= fs->fs_ncg)
671 * If we have failed in our search, record where we gave up for
672 * next time. Otherwise, fall back to our usual search citerion.
675 ip->i_nextclustercg = cg;
679 ip->i_nextclustercg = -1;
681 * We have found a new contiguous block.
683 * First we have to replace the old block pointers with the new
684 * block pointers in the inode and indirect blocks associated
689 printf("realloc: ino %ju, lbns %jd-%jd\n\told:",
690 (uintmax_t)ip->i_number,
691 (intmax_t)start_lbn, (intmax_t)end_lbn);
694 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
700 if (!ffs_checkblk(ip,
701 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
702 panic("ffs_reallocblks: unallocated block 2");
703 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
704 panic("ffs_reallocblks: alloc mismatch");
708 printf(" %d,", *bap);
710 if (DOINGSOFTDEP(vp)) {
711 if (sbap == &ip->i_din1->di_db[0] && i < ssize)
712 softdep_setup_allocdirect(ip, start_lbn + i,
713 blkno, *bap, fs->fs_bsize, fs->fs_bsize,
714 buflist->bs_children[i]);
716 softdep_setup_allocindir_page(ip, start_lbn + i,
717 i < ssize ? sbp : ebp, soff + i, blkno,
718 *bap, buflist->bs_children[i]);
723 * Next we must write out the modified inode and indirect blocks.
724 * For strict correctness, the writes should be synchronous since
725 * the old block values may have been written to disk. In practise
726 * they are almost never written, but if we are concerned about
727 * strict correctness, the `doasyncfree' flag should be set to zero.
729 * The test on `doasyncfree' should be changed to test a flag
730 * that shows whether the associated buffers and inodes have
731 * been written. The flag should be set when the cluster is
732 * started and cleared whenever the buffer or inode is flushed.
733 * We can then check below to see if it is set, and do the
734 * synchronous write only when it has been cleared.
736 if (sbap != &ip->i_din1->di_db[0]) {
742 ip->i_flag |= IN_CHANGE | IN_UPDATE;
753 * Last, free the old blocks and assign the new blocks to the buffers.
759 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
760 bp = buflist->bs_children[i];
761 if (!DOINGSOFTDEP(vp))
763 * The usual case is that a set of N-contiguous blocks
764 * that was just allocated has been replaced with a
765 * set of N+1-contiguous blocks. If they are marked as
766 * B_DELWRI, the current contents have not been written
767 * to disk. It is possible that the blocks were written
768 * earlier, but very uncommon. If the blocks have never
769 * been written, there is no need to send a BIO_DELETE
770 * for them when they are freed. The gain from avoiding
771 * the TRIMs for the common case of unwritten blocks
772 * far exceeds the cost of the write amplification for
773 * the uncommon case of failing to send a TRIM for the
774 * blocks that had been written.
776 ffs_blkfree(ump, fs, ump->um_devvp,
777 dbtofsb(fs, bp->b_blkno),
778 fs->fs_bsize, ip->i_number, vp->v_type, NULL,
779 (bp->b_flags & B_DELWRI) != 0 ?
780 NOTRIM_KEY : SINGLETON_KEY);
781 bp->b_blkno = fsbtodb(fs, blkno);
783 if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
784 panic("ffs_reallocblks: unallocated block 3");
788 printf(" %d,", blkno);
802 if (sbap != &ip->i_din1->di_db[0])
808 ffs_reallocblks_ufs2(ap)
809 struct vop_reallocblks_args /* {
811 struct cluster_save *a_buflist;
817 struct buf *sbp, *ebp, *bp;
818 ufs2_daddr_t *bap, *sbap, *ebap;
819 struct cluster_save *buflist;
820 struct ufsmount *ump;
821 ufs_lbn_t start_lbn, end_lbn;
822 ufs2_daddr_t soff, newblk, blkno, pref;
823 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp;
824 int i, cg, len, start_lvl, end_lvl, ssize;
831 * If we are not tracking block clusters or if we have less than 4%
832 * free blocks left, then do not attempt to cluster. Running with
833 * less than 5% free block reserve is not recommended and those that
834 * choose to do so do not expect to have good file layout.
836 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0)
838 buflist = ap->a_buflist;
839 len = buflist->bs_nchildren;
840 start_lbn = buflist->bs_children[0]->b_lblkno;
841 end_lbn = start_lbn + len - 1;
843 for (i = 0; i < len; i++)
844 if (!ffs_checkblk(ip,
845 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
846 panic("ffs_reallocblks: unallocated block 1");
847 for (i = 1; i < len; i++)
848 if (buflist->bs_children[i]->b_lblkno != start_lbn + i)
849 panic("ffs_reallocblks: non-logical cluster");
850 blkno = buflist->bs_children[0]->b_blkno;
851 ssize = fsbtodb(fs, fs->fs_frag);
852 for (i = 1; i < len - 1; i++)
853 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize))
854 panic("ffs_reallocblks: non-physical cluster %d", i);
857 * If the cluster crosses the boundary for the first indirect
858 * block, do not move anything in it. Indirect blocks are
859 * usually initially laid out in a position between the data
860 * blocks. Block reallocation would usually destroy locality by
861 * moving the indirect block out of the way to make room for
862 * data blocks if we didn't compensate here. We should also do
863 * this for other indirect block boundaries, but it is only
864 * important for the first one.
866 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR)
869 * If the latest allocation is in a new cylinder group, assume that
870 * the filesystem has decided to move and do not force it back to
871 * the previous cylinder group.
873 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) !=
874 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno)))
876 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) ||
877 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl))
880 * Get the starting offset and block map for the first block.
882 if (start_lvl == 0) {
883 sbap = &ip->i_din2->di_db[0];
886 idp = &start_ap[start_lvl - 1];
887 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) {
891 sbap = (ufs2_daddr_t *)sbp->b_data;
895 * If the block range spans two block maps, get the second map.
898 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) {
903 start_ap[start_lvl - 1].in_lbn == idp->in_lbn)
904 panic("ffs_reallocblk: start == end");
906 ssize = len - (idp->in_off + 1);
907 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp))
909 ebap = (ufs2_daddr_t *)ebp->b_data;
912 * Find the preferred location for the cluster. If we have not
913 * previously failed at this endeavor, then follow our standard
914 * preference calculation. If we have failed at it, then pick up
915 * where we last ended our search.
918 if (ip->i_nextclustercg == -1)
919 pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap);
921 pref = cgdata(fs, ip->i_nextclustercg);
923 * Search the block map looking for an allocation of the desired size.
924 * To avoid wasting too much time, we limit the number of cylinder
925 * groups that we will search.
928 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) {
929 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0)
932 if (cg >= fs->fs_ncg)
936 * If we have failed in our search, record where we gave up for
937 * next time. Otherwise, fall back to our usual search citerion.
940 ip->i_nextclustercg = cg;
944 ip->i_nextclustercg = -1;
946 * We have found a new contiguous block.
948 * First we have to replace the old block pointers with the new
949 * block pointers in the inode and indirect blocks associated
954 printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number,
955 (intmax_t)start_lbn, (intmax_t)end_lbn);
958 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) {
964 if (!ffs_checkblk(ip,
965 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize))
966 panic("ffs_reallocblks: unallocated block 2");
967 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap)
968 panic("ffs_reallocblks: alloc mismatch");
972 printf(" %jd,", (intmax_t)*bap);
974 if (DOINGSOFTDEP(vp)) {
975 if (sbap == &ip->i_din2->di_db[0] && i < ssize)
976 softdep_setup_allocdirect(ip, start_lbn + i,
977 blkno, *bap, fs->fs_bsize, fs->fs_bsize,
978 buflist->bs_children[i]);
980 softdep_setup_allocindir_page(ip, start_lbn + i,
981 i < ssize ? sbp : ebp, soff + i, blkno,
982 *bap, buflist->bs_children[i]);
987 * Next we must write out the modified inode and indirect blocks.
988 * For strict correctness, the writes should be synchronous since
989 * the old block values may have been written to disk. In practise
990 * they are almost never written, but if we are concerned about
991 * strict correctness, the `doasyncfree' flag should be set to zero.
993 * The test on `doasyncfree' should be changed to test a flag
994 * that shows whether the associated buffers and inodes have
995 * been written. The flag should be set when the cluster is
996 * started and cleared whenever the buffer or inode is flushed.
997 * We can then check below to see if it is set, and do the
998 * synchronous write only when it has been cleared.
1000 if (sbap != &ip->i_din2->di_db[0]) {
1006 ip->i_flag |= IN_CHANGE | IN_UPDATE;
1017 * Last, free the old blocks and assign the new blocks to the buffers.
1023 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) {
1024 bp = buflist->bs_children[i];
1025 if (!DOINGSOFTDEP(vp))
1027 * The usual case is that a set of N-contiguous blocks
1028 * that was just allocated has been replaced with a
1029 * set of N+1-contiguous blocks. If they are marked as
1030 * B_DELWRI, the current contents have not been written
1031 * to disk. It is possible that the blocks were written
1032 * earlier, but very uncommon. If the blocks have never
1033 * been written, there is no need to send a BIO_DELETE
1034 * for them when they are freed. The gain from avoiding
1035 * the TRIMs for the common case of unwritten blocks
1036 * far exceeds the cost of the write amplification for
1037 * the uncommon case of failing to send a TRIM for the
1038 * blocks that had been written.
1040 ffs_blkfree(ump, fs, ump->um_devvp,
1041 dbtofsb(fs, bp->b_blkno),
1042 fs->fs_bsize, ip->i_number, vp->v_type, NULL,
1043 (bp->b_flags & B_DELWRI) != 0 ?
1044 NOTRIM_KEY : SINGLETON_KEY);
1045 bp->b_blkno = fsbtodb(fs, blkno);
1047 if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize))
1048 panic("ffs_reallocblks: unallocated block 3");
1052 printf(" %jd,", (intmax_t)blkno);
1066 if (sbap != &ip->i_din2->di_db[0])
1072 * Allocate an inode in the filesystem.
1074 * If allocating a directory, use ffs_dirpref to select the inode.
1075 * If allocating in a directory, the following hierarchy is followed:
1076 * 1) allocate the preferred inode.
1077 * 2) allocate an inode in the same cylinder group.
1078 * 3) quadradically rehash into other cylinder groups, until an
1079 * available inode is located.
1080 * If no inode preference is given the following hierarchy is used
1081 * to allocate an inode:
1082 * 1) allocate an inode in cylinder group 0.
1083 * 2) quadradically rehash into other cylinder groups, until an
1084 * available inode is located.
1087 ffs_valloc(pvp, mode, cred, vpp)
1097 struct ufsmount *ump;
1100 int error, error1, reclaimed;
1101 static struct timeval lastfail;
1112 if (fs->fs_cstotal.cs_nifree == 0)
1115 if ((mode & IFMT) == IFDIR)
1116 ipref = ffs_dirpref(pip);
1118 ipref = pip->i_number;
1119 if (ipref >= fs->fs_ncg * fs->fs_ipg)
1121 cg = ino_to_cg(fs, ipref);
1123 * Track number of dirs created one after another
1124 * in a same cg without intervening by files.
1126 if ((mode & IFMT) == IFDIR) {
1127 if (fs->fs_contigdirs[cg] < 255)
1128 fs->fs_contigdirs[cg]++;
1130 if (fs->fs_contigdirs[cg] > 0)
1131 fs->fs_contigdirs[cg]--;
1133 ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0,
1134 (allocfcn_t *)ffs_nodealloccg);
1137 error = ffs_vget(pvp->v_mount, ino, LK_EXCLUSIVE, vpp);
1139 error1 = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp,
1141 ffs_vfree(pvp, ino, mode);
1146 ip->i_flag |= IN_MODIFIED;
1154 printf("mode = 0%o, inum = %ju, fs = %s\n",
1155 ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt);
1156 panic("ffs_valloc: dup alloc");
1158 if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) { /* XXX */
1159 printf("free inode %s/%lu had %ld blocks\n",
1160 fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks));
1161 DIP_SET(ip, i_blocks, 0);
1164 DIP_SET(ip, i_flags, 0);
1166 * Set up a new generation number for this inode.
1168 while (ip->i_gen == 0 || ++ip->i_gen == 0)
1169 ip->i_gen = arc4random();
1170 DIP_SET(ip, i_gen, ip->i_gen);
1171 if (fs->fs_magic == FS_UFS2_MAGIC) {
1173 ip->i_din2->di_birthtime = ts.tv_sec;
1174 ip->i_din2->di_birthnsec = ts.tv_nsec;
1176 ufs_prepare_reclaim(*vpp);
1178 (*vpp)->v_vflag = 0;
1179 (*vpp)->v_type = VNON;
1180 if (fs->fs_magic == FS_UFS2_MAGIC) {
1181 (*vpp)->v_op = &ffs_vnodeops2;
1182 ip->i_flag |= IN_UFS2;
1184 (*vpp)->v_op = &ffs_vnodeops1;
1188 if (reclaimed == 0) {
1190 softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT);
1194 if (ppsratecheck(&lastfail, &curfail, 1)) {
1195 ffs_fserr(fs, pip->i_number, "out of inodes");
1196 uprintf("\n%s: create/symlink failed, no inodes free\n",
1203 * Find a cylinder group to place a directory.
1205 * The policy implemented by this algorithm is to allocate a
1206 * directory inode in the same cylinder group as its parent
1207 * directory, but also to reserve space for its files inodes
1208 * and data. Restrict the number of directories which may be
1209 * allocated one after another in the same cylinder group
1210 * without intervening allocation of files.
1212 * If we allocate a first level directory then force allocation
1213 * in another cylinder group.
1220 int cg, prefcg, dirsize, cgsize;
1221 u_int avgifree, avgbfree, avgndir, curdirsize;
1222 u_int minifree, minbfree, maxndir;
1223 u_int mincg, minndir;
1224 u_int maxcontigdirs;
1226 mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED);
1229 avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg;
1230 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1231 avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg;
1234 * Force allocation in another cg if creating a first level dir.
1236 ASSERT_VOP_LOCKED(ITOV(pip), "ffs_dirpref");
1237 if (ITOV(pip)->v_vflag & VV_ROOT) {
1238 prefcg = arc4random() % fs->fs_ncg;
1240 minndir = fs->fs_ipg;
1241 for (cg = prefcg; cg < fs->fs_ncg; cg++)
1242 if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1243 fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1244 fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1246 minndir = fs->fs_cs(fs, cg).cs_ndir;
1248 for (cg = 0; cg < prefcg; cg++)
1249 if (fs->fs_cs(fs, cg).cs_ndir < minndir &&
1250 fs->fs_cs(fs, cg).cs_nifree >= avgifree &&
1251 fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1253 minndir = fs->fs_cs(fs, cg).cs_ndir;
1255 return ((ino_t)(fs->fs_ipg * mincg));
1259 * Count various limits which used for
1260 * optimal allocation of a directory inode.
1262 maxndir = min(avgndir + fs->fs_ipg / 16, fs->fs_ipg);
1263 minifree = avgifree - avgifree / 4;
1266 minbfree = avgbfree - avgbfree / 4;
1269 cgsize = fs->fs_fsize * fs->fs_fpg;
1270 dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir;
1271 curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0;
1272 if (dirsize < curdirsize)
1273 dirsize = curdirsize;
1275 maxcontigdirs = 0; /* dirsize overflowed */
1277 maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255);
1278 if (fs->fs_avgfpdir > 0)
1279 maxcontigdirs = min(maxcontigdirs,
1280 fs->fs_ipg / fs->fs_avgfpdir);
1281 if (maxcontigdirs == 0)
1285 * Limit number of dirs in one cg and reserve space for
1286 * regular files, but only if we have no deficit in
1289 * We are trying to find a suitable cylinder group nearby
1290 * our preferred cylinder group to place a new directory.
1291 * We scan from our preferred cylinder group forward looking
1292 * for a cylinder group that meets our criterion. If we get
1293 * to the final cylinder group and do not find anything,
1294 * we start scanning forwards from the beginning of the
1295 * filesystem. While it might seem sensible to start scanning
1296 * backwards or even to alternate looking forward and backward,
1297 * this approach fails badly when the filesystem is nearly full.
1298 * Specifically, we first search all the areas that have no space
1299 * and finally try the one preceding that. We repeat this on
1300 * every request and in the case of the final block end up
1301 * searching the entire filesystem. By jumping to the front
1302 * of the filesystem, our future forward searches always look
1303 * in new cylinder groups so finds every possible block after
1304 * one pass over the filesystem.
1306 prefcg = ino_to_cg(fs, pip->i_number);
1307 for (cg = prefcg; cg < fs->fs_ncg; cg++)
1308 if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1309 fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1310 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1311 if (fs->fs_contigdirs[cg] < maxcontigdirs)
1312 return ((ino_t)(fs->fs_ipg * cg));
1314 for (cg = 0; cg < prefcg; cg++)
1315 if (fs->fs_cs(fs, cg).cs_ndir < maxndir &&
1316 fs->fs_cs(fs, cg).cs_nifree >= minifree &&
1317 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) {
1318 if (fs->fs_contigdirs[cg] < maxcontigdirs)
1319 return ((ino_t)(fs->fs_ipg * cg));
1322 * This is a backstop when we have deficit in space.
1324 for (cg = prefcg; cg < fs->fs_ncg; cg++)
1325 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1326 return ((ino_t)(fs->fs_ipg * cg));
1327 for (cg = 0; cg < prefcg; cg++)
1328 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree)
1330 return ((ino_t)(fs->fs_ipg * cg));
1334 * Select the desired position for the next block in a file. The file is
1335 * logically divided into sections. The first section is composed of the
1336 * direct blocks and the next fs_maxbpg blocks. Each additional section
1337 * contains fs_maxbpg blocks.
1339 * If no blocks have been allocated in the first section, the policy is to
1340 * request a block in the same cylinder group as the inode that describes
1341 * the file. The first indirect is allocated immediately following the last
1342 * direct block and the data blocks for the first indirect immediately
1345 * If no blocks have been allocated in any other section, the indirect
1346 * block(s) are allocated in the same cylinder group as its inode in an
1347 * area reserved immediately following the inode blocks. The policy for
1348 * the data blocks is to place them in a cylinder group with a greater than
1349 * average number of free blocks. An appropriate cylinder group is found
1350 * by using a rotor that sweeps the cylinder groups. When a new group of
1351 * blocks is needed, the sweep begins in the cylinder group following the
1352 * cylinder group from which the previous allocation was made. The sweep
1353 * continues until a cylinder group with greater than the average number
1354 * of free blocks is found. If the allocation is for the first block in an
1355 * indirect block or the previous block is a hole, then the information on
1356 * the previous allocation is unavailable; here a best guess is made based
1357 * on the logical block number being allocated.
1359 * If a section is already partially allocated, the policy is to
1360 * allocate blocks contiguously within the section if possible.
1363 ffs_blkpref_ufs1(ip, lbn, indx, bap)
1371 u_int avgbfree, startcg;
1374 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1375 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1378 * Allocation of indirect blocks is indicated by passing negative
1379 * values in indx: -1 for single indirect, -2 for double indirect,
1380 * -3 for triple indirect. As noted below, we attempt to allocate
1381 * the first indirect inline with the file data. For all later
1382 * indirect blocks, the data is often allocated in other cylinder
1383 * groups. However to speed random file access and to speed up
1384 * fsck, the filesystem reserves the first fs_metaspace blocks
1385 * (typically half of fs_minfree) of the data area of each cylinder
1386 * group to hold these later indirect blocks.
1388 inocg = ino_to_cg(fs, ip->i_number);
1391 * Our preference for indirect blocks is the zone at the
1392 * beginning of the inode's cylinder group data area that
1393 * we try to reserve for indirect blocks.
1395 pref = cgmeta(fs, inocg);
1397 * If we are allocating the first indirect block, try to
1398 * place it immediately following the last direct block.
1400 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1401 ip->i_din1->di_db[UFS_NDADDR - 1] != 0)
1402 pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1406 * If we are allocating the first data block in the first indirect
1407 * block and the indirect has been allocated in the data block area,
1408 * try to place it immediately following the indirect block.
1410 if (lbn == UFS_NDADDR) {
1411 pref = ip->i_din1->di_ib[0];
1412 if (pref != 0 && pref >= cgdata(fs, inocg) &&
1413 pref < cgbase(fs, inocg + 1))
1414 return (pref + fs->fs_frag);
1417 * If we are at the beginning of a file, or we have already allocated
1418 * the maximum number of blocks per cylinder group, or we do not
1419 * have a block allocated immediately preceding us, then we need
1420 * to decide where to start allocating new blocks.
1422 if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1424 * If we are allocating a directory data block, we want
1425 * to place it in the metadata area.
1427 if ((ip->i_mode & IFMT) == IFDIR)
1428 return (cgmeta(fs, inocg));
1430 * Until we fill all the direct and all the first indirect's
1431 * blocks, we try to allocate in the data area of the inode's
1434 if (lbn < UFS_NDADDR + NINDIR(fs))
1435 return (cgdata(fs, inocg));
1437 * Find a cylinder with greater than average number of
1438 * unused data blocks.
1440 if (indx == 0 || bap[indx - 1] == 0)
1441 startcg = inocg + lbn / fs->fs_maxbpg;
1443 startcg = dtog(fs, bap[indx - 1]) + 1;
1444 startcg %= fs->fs_ncg;
1445 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1446 for (cg = startcg; cg < fs->fs_ncg; cg++)
1447 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1448 fs->fs_cgrotor = cg;
1449 return (cgdata(fs, cg));
1451 for (cg = 0; cg <= startcg; cg++)
1452 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1453 fs->fs_cgrotor = cg;
1454 return (cgdata(fs, cg));
1459 * Otherwise, we just always try to lay things out contiguously.
1461 return (bap[indx - 1] + fs->fs_frag);
1465 * Same as above, but for UFS2
1468 ffs_blkpref_ufs2(ip, lbn, indx, bap)
1476 u_int avgbfree, startcg;
1479 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap"));
1480 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1483 * Allocation of indirect blocks is indicated by passing negative
1484 * values in indx: -1 for single indirect, -2 for double indirect,
1485 * -3 for triple indirect. As noted below, we attempt to allocate
1486 * the first indirect inline with the file data. For all later
1487 * indirect blocks, the data is often allocated in other cylinder
1488 * groups. However to speed random file access and to speed up
1489 * fsck, the filesystem reserves the first fs_metaspace blocks
1490 * (typically half of fs_minfree) of the data area of each cylinder
1491 * group to hold these later indirect blocks.
1493 inocg = ino_to_cg(fs, ip->i_number);
1496 * Our preference for indirect blocks is the zone at the
1497 * beginning of the inode's cylinder group data area that
1498 * we try to reserve for indirect blocks.
1500 pref = cgmeta(fs, inocg);
1502 * If we are allocating the first indirect block, try to
1503 * place it immediately following the last direct block.
1505 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) &&
1506 ip->i_din2->di_db[UFS_NDADDR - 1] != 0)
1507 pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag;
1511 * If we are allocating the first data block in the first indirect
1512 * block and the indirect has been allocated in the data block area,
1513 * try to place it immediately following the indirect block.
1515 if (lbn == UFS_NDADDR) {
1516 pref = ip->i_din2->di_ib[0];
1517 if (pref != 0 && pref >= cgdata(fs, inocg) &&
1518 pref < cgbase(fs, inocg + 1))
1519 return (pref + fs->fs_frag);
1522 * If we are at the beginning of a file, or we have already allocated
1523 * the maximum number of blocks per cylinder group, or we do not
1524 * have a block allocated immediately preceding us, then we need
1525 * to decide where to start allocating new blocks.
1527 if (indx % fs->fs_maxbpg == 0 || bap[indx - 1] == 0) {
1529 * If we are allocating a directory data block, we want
1530 * to place it in the metadata area.
1532 if ((ip->i_mode & IFMT) == IFDIR)
1533 return (cgmeta(fs, inocg));
1535 * Until we fill all the direct and all the first indirect's
1536 * blocks, we try to allocate in the data area of the inode's
1539 if (lbn < UFS_NDADDR + NINDIR(fs))
1540 return (cgdata(fs, inocg));
1542 * Find a cylinder with greater than average number of
1543 * unused data blocks.
1545 if (indx == 0 || bap[indx - 1] == 0)
1546 startcg = inocg + lbn / fs->fs_maxbpg;
1548 startcg = dtog(fs, bap[indx - 1]) + 1;
1549 startcg %= fs->fs_ncg;
1550 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg;
1551 for (cg = startcg; cg < fs->fs_ncg; cg++)
1552 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1553 fs->fs_cgrotor = cg;
1554 return (cgdata(fs, cg));
1556 for (cg = 0; cg <= startcg; cg++)
1557 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) {
1558 fs->fs_cgrotor = cg;
1559 return (cgdata(fs, cg));
1564 * Otherwise, we just always try to lay things out contiguously.
1566 return (bap[indx - 1] + fs->fs_frag);
1570 * Implement the cylinder overflow algorithm.
1572 * The policy implemented by this algorithm is:
1573 * 1) allocate the block in its requested cylinder group.
1574 * 2) quadradically rehash on the cylinder group number.
1575 * 3) brute force search for a free block.
1577 * Must be called with the UFS lock held. Will release the lock on success
1578 * and return with it held on failure.
1582 ffs_hashalloc(ip, cg, pref, size, rsize, allocator)
1586 int size; /* Search size for data blocks, mode for inodes */
1587 int rsize; /* Real allocated size. */
1588 allocfcn_t *allocator;
1591 ufs2_daddr_t result;
1594 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED);
1596 if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED)
1597 panic("ffs_hashalloc: allocation on suspended filesystem");
1601 * 1: preferred cylinder group
1603 result = (*allocator)(ip, cg, pref, size, rsize);
1607 * 2: quadratic rehash
1609 for (i = 1; i < fs->fs_ncg; i *= 2) {
1611 if (cg >= fs->fs_ncg)
1613 result = (*allocator)(ip, cg, 0, size, rsize);
1618 * 3: brute force search
1619 * Note that we start at i == 2, since 0 was checked initially,
1620 * and 1 is always checked in the quadratic rehash.
1622 cg = (icg + 2) % fs->fs_ncg;
1623 for (i = 2; i < fs->fs_ncg; i++) {
1624 result = (*allocator)(ip, cg, 0, size, rsize);
1628 if (cg == fs->fs_ncg)
1635 * Determine whether a fragment can be extended.
1637 * Check to see if the necessary fragments are available, and
1638 * if they are, allocate them.
1641 ffs_fragextend(ip, cg, bprev, osize, nsize)
1650 struct ufsmount *ump;
1659 if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize))
1661 frags = numfrags(fs, nsize);
1662 bbase = fragnum(fs, bprev);
1663 if (bbase > fragnum(fs, (bprev + frags - 1))) {
1664 /* cannot extend across a block boundary */
1668 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0)
1670 bno = dtogd(fs, bprev);
1671 blksfree = cg_blksfree(cgp);
1672 for (i = numfrags(fs, osize); i < frags; i++)
1673 if (isclr(blksfree, bno + i))
1676 * the current fragment can be extended
1677 * deduct the count on fragment being extended into
1678 * increase the count on the remaining fragment (if any)
1679 * allocate the extended piece
1681 for (i = frags; i < fs->fs_frag - bbase; i++)
1682 if (isclr(blksfree, bno + i))
1684 cgp->cg_frsum[i - numfrags(fs, osize)]--;
1686 cgp->cg_frsum[i - frags]++;
1687 for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) {
1688 clrbit(blksfree, bno + i);
1689 cgp->cg_cs.cs_nffree--;
1693 fs->fs_cstotal.cs_nffree -= nffree;
1694 fs->fs_cs(fs, cg).cs_nffree -= nffree;
1696 ACTIVECLEAR(fs, cg);
1698 if (DOINGSOFTDEP(ITOV(ip)))
1699 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev,
1700 frags, numfrags(fs, osize));
1712 * Determine whether a block can be allocated.
1714 * Check to see if a block of the appropriate size is available,
1715 * and if it is, allocate it.
1718 ffs_alloccg(ip, cg, bpref, size, rsize)
1728 struct ufsmount *ump;
1731 int i, allocsiz, error, frags;
1736 if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize)
1739 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0 ||
1740 (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize))
1742 if (size == fs->fs_bsize) {
1744 blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1745 ACTIVECLEAR(fs, cg);
1751 * check to see if any fragments are already available
1752 * allocsiz is the size which will be allocated, hacking
1753 * it down to a smaller size if necessary
1755 blksfree = cg_blksfree(cgp);
1756 frags = numfrags(fs, size);
1757 for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++)
1758 if (cgp->cg_frsum[allocsiz] != 0)
1760 if (allocsiz == fs->fs_frag) {
1762 * no fragments were available, so a block will be
1763 * allocated, and hacked up
1765 if (cgp->cg_cs.cs_nbfree == 0)
1768 blkno = ffs_alloccgblk(ip, bp, bpref, rsize);
1769 ACTIVECLEAR(fs, cg);
1774 KASSERT(size == rsize,
1775 ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize));
1776 bno = ffs_mapsearch(fs, cgp, bpref, allocsiz);
1779 for (i = 0; i < frags; i++)
1780 clrbit(blksfree, bno + i);
1781 cgp->cg_cs.cs_nffree -= frags;
1782 cgp->cg_frsum[allocsiz]--;
1783 if (frags != allocsiz)
1784 cgp->cg_frsum[allocsiz - frags]++;
1786 fs->fs_cstotal.cs_nffree -= frags;
1787 fs->fs_cs(fs, cg).cs_nffree -= frags;
1789 blkno = cgbase(fs, cg) + bno;
1790 ACTIVECLEAR(fs, cg);
1792 if (DOINGSOFTDEP(ITOV(ip)))
1793 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0);
1804 * Allocate a block in a cylinder group.
1806 * This algorithm implements the following policy:
1807 * 1) allocate the requested block.
1808 * 2) allocate a rotationally optimal block in the same cylinder.
1809 * 3) allocate the next available block on the block rotor for the
1810 * specified cylinder group.
1811 * Note that this routine only allocates fs_bsize blocks; these
1812 * blocks may be fragmented by the routine that allocates them.
1815 ffs_alloccgblk(ip, bp, bpref, size)
1823 struct ufsmount *ump;
1831 mtx_assert(UFS_MTX(ump), MA_OWNED);
1832 cgp = (struct cg *)bp->b_data;
1833 blksfree = cg_blksfree(cgp);
1835 bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag;
1836 } else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) {
1837 /* map bpref to correct zone in this cg */
1838 if (bpref < cgdata(fs, cgbpref))
1839 bpref = cgmeta(fs, cgp->cg_cgx);
1841 bpref = cgdata(fs, cgp->cg_cgx);
1844 * if the requested block is available, use it
1846 bno = dtogd(fs, blknum(fs, bpref));
1847 if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno)))
1850 * Take the next available block in this cylinder group.
1852 bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag);
1855 /* Update cg_rotor only if allocated from the data zone */
1856 if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx)))
1857 cgp->cg_rotor = bno;
1859 blkno = fragstoblks(fs, bno);
1860 ffs_clrblock(fs, blksfree, (long)blkno);
1861 ffs_clusteracct(fs, cgp, blkno, -1);
1862 cgp->cg_cs.cs_nbfree--;
1863 fs->fs_cstotal.cs_nbfree--;
1864 fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--;
1866 blkno = cgbase(fs, cgp->cg_cgx) + bno;
1868 * If the caller didn't want the whole block free the frags here.
1870 size = numfrags(fs, size);
1871 if (size != fs->fs_frag) {
1872 bno = dtogd(fs, blkno);
1873 for (i = size; i < fs->fs_frag; i++)
1874 setbit(blksfree, bno + i);
1875 i = fs->fs_frag - size;
1876 cgp->cg_cs.cs_nffree += i;
1877 fs->fs_cstotal.cs_nffree += i;
1878 fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i;
1884 if (DOINGSOFTDEP(ITOV(ip)))
1885 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0);
1891 * Determine whether a cluster can be allocated.
1893 * We do not currently check for optimal rotational layout if there
1894 * are multiple choices in the same cylinder group. Instead we just
1895 * take the first one that we find following bpref.
1898 ffs_clusteralloc(ip, cg, bpref, len)
1907 struct ufsmount *ump;
1908 int i, run, bit, map, got, error;
1916 if (fs->fs_maxcluster[cg] < len)
1919 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) {
1924 * Check to see if a cluster of the needed size (or bigger) is
1925 * available in this cylinder group.
1927 lp = &cg_clustersum(cgp)[len];
1928 for (i = len; i <= fs->fs_contigsumsize; i++)
1931 if (i > fs->fs_contigsumsize) {
1933 * This is the first time looking for a cluster in this
1934 * cylinder group. Update the cluster summary information
1935 * to reflect the true maximum sized cluster so that
1936 * future cluster allocation requests can avoid reading
1937 * the cylinder group map only to find no clusters.
1939 lp = &cg_clustersum(cgp)[len - 1];
1940 for (i = len - 1; i > 0; i--)
1944 fs->fs_maxcluster[cg] = i;
1949 * Search the cluster map to find a big enough cluster.
1950 * We take the first one that we find, even if it is larger
1951 * than we need as we prefer to get one close to the previous
1952 * block allocation. We do not search before the current
1953 * preference point as we do not want to allocate a block
1954 * that is allocated before the previous one (as we will
1955 * then have to wait for another pass of the elevator
1956 * algorithm before it will be read). We prefer to fail and
1957 * be recalled to try an allocation in the next cylinder group.
1959 if (dtog(fs, bpref) != cg)
1960 bpref = cgdata(fs, cg);
1962 bpref = blknum(fs, bpref);
1963 bpref = fragstoblks(fs, dtogd(fs, bpref));
1964 mapp = &cg_clustersfree(cgp)[bpref / NBBY];
1966 bit = 1 << (bpref % NBBY);
1967 for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) {
1968 if ((map & bit) == 0) {
1975 if ((got & (NBBY - 1)) != (NBBY - 1)) {
1982 if (got >= cgp->cg_nclusterblks) {
1988 * Allocate the cluster that we have found.
1990 blksfree = cg_blksfree(cgp);
1991 for (i = 1; i <= len; i++)
1992 if (!ffs_isblock(fs, blksfree, got - run + i))
1993 panic("ffs_clusteralloc: map mismatch");
1994 bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1);
1995 if (dtog(fs, bno) != cg)
1996 panic("ffs_clusteralloc: allocated out of group");
1997 len = blkstofrags(fs, len);
1999 for (i = 0; i < len; i += fs->fs_frag)
2000 if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i)
2001 panic("ffs_clusteralloc: lost block");
2002 ACTIVECLEAR(fs, cg);
2008 static inline struct buf *
2009 getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags)
2014 return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs,
2015 cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
2020 * Synchronous inode initialization is needed only when barrier writes do not
2021 * work as advertised, and will impose a heavy cost on file creation in a newly
2022 * created filesystem.
2024 static int doasyncinodeinit = 1;
2025 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN,
2026 &doasyncinodeinit, 0,
2027 "Perform inode block initialization using asynchronous writes");
2030 * Determine whether an inode can be allocated.
2032 * Check to see if an inode is available, and if it is,
2033 * allocate it using the following policy:
2034 * 1) allocate the requested inode.
2035 * 2) allocate the next available inode after the requested
2036 * inode in the specified cylinder group.
2039 ffs_nodealloccg(ip, cg, ipref, mode, unused)
2048 struct buf *bp, *ibp;
2049 struct ufsmount *ump;
2050 u_int8_t *inosused, *loc;
2051 struct ufs2_dinode *dp2;
2052 int error, start, len, i;
2053 u_int32_t old_initediblk;
2058 if (fs->fs_cs(fs, cg).cs_nifree == 0)
2061 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) {
2066 if (cgp->cg_cs.cs_nifree == 0) {
2071 inosused = cg_inosused(cgp);
2073 ipref %= fs->fs_ipg;
2074 if (isclr(inosused, ipref))
2077 start = cgp->cg_irotor / NBBY;
2078 len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY);
2079 loc = memcchr(&inosused[start], 0xff, len);
2083 loc = memcchr(&inosused[start], 0xff, len);
2085 printf("cg = %d, irotor = %ld, fs = %s\n",
2086 cg, (long)cgp->cg_irotor, fs->fs_fsmnt);
2087 panic("ffs_nodealloccg: map corrupted");
2091 ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
2094 * Check to see if we need to initialize more inodes.
2096 if (fs->fs_magic == FS_UFS2_MAGIC &&
2097 ipref + INOPB(fs) > cgp->cg_initediblk &&
2098 cgp->cg_initediblk < cgp->cg_niblk) {
2099 old_initediblk = cgp->cg_initediblk;
2102 * Free the cylinder group lock before writing the
2103 * initialized inode block. Entering the
2104 * babarrierwrite() with the cylinder group lock
2105 * causes lock order violation between the lock and
2108 * Another thread can decide to initialize the same
2109 * inode block, but whichever thread first gets the
2110 * cylinder group lock after writing the newly
2111 * allocated inode block will update it and the other
2112 * will realize that it has lost and leave the
2113 * cylinder group unchanged.
2115 ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
2119 * The inode block buffer is already owned by
2120 * another thread, which must initialize it.
2121 * Wait on the buffer to allow another thread
2122 * to finish the updates, with dropped cg
2123 * buffer lock, then retry.
2125 ibp = getinobuf(ip, cg, old_initediblk, 0);
2130 bzero(ibp->b_data, (int)fs->fs_bsize);
2131 dp2 = (struct ufs2_dinode *)(ibp->b_data);
2132 for (i = 0; i < INOPB(fs); i++) {
2133 while (dp2->di_gen == 0)
2134 dp2->di_gen = arc4random();
2139 * Rather than adding a soft updates dependency to ensure
2140 * that the new inode block is written before it is claimed
2141 * by the cylinder group map, we just do a barrier write
2142 * here. The barrier write will ensure that the inode block
2143 * gets written before the updated cylinder group map can be
2144 * written. The barrier write should only slow down bulk
2145 * loading of newly created filesystems.
2147 if (doasyncinodeinit)
2148 babarrierwrite(ibp);
2153 * After the inode block is written, try to update the
2154 * cg initediblk pointer. If another thread beat us
2155 * to it, then leave it unchanged as the other thread
2156 * has already set it correctly.
2158 error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp);
2160 ACTIVECLEAR(fs, cg);
2164 if (cgp->cg_initediblk == old_initediblk)
2165 cgp->cg_initediblk += INOPB(fs);
2168 cgp->cg_irotor = ipref;
2170 ACTIVECLEAR(fs, cg);
2171 setbit(inosused, ipref);
2172 cgp->cg_cs.cs_nifree--;
2173 fs->fs_cstotal.cs_nifree--;
2174 fs->fs_cs(fs, cg).cs_nifree--;
2176 if ((mode & IFMT) == IFDIR) {
2177 cgp->cg_cs.cs_ndir++;
2178 fs->fs_cstotal.cs_ndir++;
2179 fs->fs_cs(fs, cg).cs_ndir++;
2182 if (DOINGSOFTDEP(ITOV(ip)))
2183 softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode);
2185 return ((ino_t)(cg * fs->fs_ipg + ipref));
2189 * Free a block or fragment.
2191 * The specified block or fragment is placed back in the
2192 * free map. If a fragment is deallocated, a possible
2193 * block reassembly is checked.
2196 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd)
2197 struct ufsmount *ump;
2199 struct vnode *devvp;
2203 struct workhead *dephd;
2208 ufs1_daddr_t fragno, cgbno;
2209 int i, blk, frags, bbase, error;
2215 if (devvp->v_type == VREG) {
2216 /* devvp is a snapshot */
2217 MPASS(devvp->v_mount->mnt_data == ump);
2218 dev = ump->um_devvp->v_rdev;
2219 } else if (devvp->v_type == VCHR) {
2220 /* devvp is a normal disk device */
2221 dev = devvp->v_rdev;
2222 ASSERT_VOP_LOCKED(devvp, "ffs_blkfree_cg");
2226 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 ||
2227 fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) {
2228 printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n",
2229 devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize,
2230 size, fs->fs_fsmnt);
2231 panic("ffs_blkfree_cg: bad size");
2234 if ((u_int)bno >= fs->fs_size) {
2235 printf("bad block %jd, ino %lu\n", (intmax_t)bno,
2237 ffs_fserr(fs, inum, "bad block");
2240 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2242 cgbno = dtogd(fs, bno);
2243 blksfree = cg_blksfree(cgp);
2245 if (size == fs->fs_bsize) {
2246 fragno = fragstoblks(fs, cgbno);
2247 if (!ffs_isfreeblock(fs, blksfree, fragno)) {
2248 if (devvp->v_type == VREG) {
2250 /* devvp is a snapshot */
2254 printf("dev = %s, block = %jd, fs = %s\n",
2255 devtoname(dev), (intmax_t)bno, fs->fs_fsmnt);
2256 panic("ffs_blkfree_cg: freeing free block");
2258 ffs_setblock(fs, blksfree, fragno);
2259 ffs_clusteracct(fs, cgp, fragno, 1);
2260 cgp->cg_cs.cs_nbfree++;
2261 fs->fs_cstotal.cs_nbfree++;
2262 fs->fs_cs(fs, cg).cs_nbfree++;
2264 bbase = cgbno - fragnum(fs, cgbno);
2266 * decrement the counts associated with the old frags
2268 blk = blkmap(fs, blksfree, bbase);
2269 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
2271 * deallocate the fragment
2273 frags = numfrags(fs, size);
2274 for (i = 0; i < frags; i++) {
2275 if (isset(blksfree, cgbno + i)) {
2276 printf("dev = %s, block = %jd, fs = %s\n",
2277 devtoname(dev), (intmax_t)(bno + i),
2279 panic("ffs_blkfree_cg: freeing free frag");
2281 setbit(blksfree, cgbno + i);
2283 cgp->cg_cs.cs_nffree += i;
2284 fs->fs_cstotal.cs_nffree += i;
2285 fs->fs_cs(fs, cg).cs_nffree += i;
2287 * add back in counts associated with the new frags
2289 blk = blkmap(fs, blksfree, bbase);
2290 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
2292 * if a complete block has been reassembled, account for it
2294 fragno = fragstoblks(fs, bbase);
2295 if (ffs_isblock(fs, blksfree, fragno)) {
2296 cgp->cg_cs.cs_nffree -= fs->fs_frag;
2297 fs->fs_cstotal.cs_nffree -= fs->fs_frag;
2298 fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag;
2299 ffs_clusteracct(fs, cgp, fragno, 1);
2300 cgp->cg_cs.cs_nbfree++;
2301 fs->fs_cstotal.cs_nbfree++;
2302 fs->fs_cs(fs, cg).cs_nbfree++;
2306 ACTIVECLEAR(fs, cg);
2309 if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR)
2310 softdep_setup_blkfree(UFSTOVFS(ump), bp, bno,
2311 numfrags(fs, size), dephd);
2316 * Structures and routines associated with trim management.
2318 * The following requests are passed to trim_lookup to indicate
2319 * the actions that should be taken.
2321 #define NEW 1 /* if found, error else allocate and hash it */
2322 #define OLD 2 /* if not found, error, else return it */
2323 #define REPLACE 3 /* if not found, error else unhash and reallocate it */
2324 #define DONE 4 /* if not found, error else unhash and return it */
2325 #define SINGLE 5 /* don't look up, just allocate it and don't hash it */
2327 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures");
2329 #define TRIMLIST_HASH(ump, key) \
2330 (&(ump)->um_trimhash[(key) & (ump)->um_trimlisthashsize])
2333 * These structures describe each of the block free requests aggregated
2334 * together to make up a trim request.
2336 struct trim_blkreq {
2337 TAILQ_ENTRY(trim_blkreq) blkreqlist;
2340 struct workhead *pdephd;
2341 struct workhead dephd;
2345 * Description of a trim request.
2347 struct ffs_blkfree_trim_params {
2348 TAILQ_HEAD(, trim_blkreq) blklist;
2349 LIST_ENTRY(ffs_blkfree_trim_params) hashlist;
2351 struct ufsmount *ump;
2352 struct vnode *devvp;
2359 static void ffs_blkfree_trim_completed(struct buf *);
2360 static void ffs_blkfree_trim_task(void *ctx, int pending __unused);
2361 static struct ffs_blkfree_trim_params *trim_lookup(struct ufsmount *,
2362 struct vnode *, ufs2_daddr_t, long, ino_t, u_long, int);
2363 static void ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *);
2366 * Called on trim completion to start a task to free the associated block(s).
2369 ffs_blkfree_trim_completed(bp)
2372 struct ffs_blkfree_trim_params *tp;
2374 tp = bp->b_fsprivate1;
2376 TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp);
2377 taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task);
2381 * Trim completion task that free associated block(s).
2384 ffs_blkfree_trim_task(ctx, pending)
2388 struct ffs_blkfree_trim_params *tp;
2389 struct trim_blkreq *blkelm;
2390 struct ufsmount *ump;
2394 while ((blkelm = TAILQ_FIRST(&tp->blklist)) != NULL) {
2395 ffs_blkfree_cg(ump, ump->um_fs, tp->devvp, blkelm->bno,
2396 blkelm->size, tp->inum, blkelm->pdephd);
2397 TAILQ_REMOVE(&tp->blklist, blkelm, blkreqlist);
2398 free(blkelm, M_TRIM);
2400 vn_finished_secondary_write(UFSTOVFS(ump));
2402 ump->um_trim_inflight -= 1;
2403 ump->um_trim_inflight_blks -= numfrags(ump->um_fs, tp->size);
2409 * Lookup a trim request by inode number.
2410 * Allocate if requested (NEW, REPLACE, SINGLE).
2412 static struct ffs_blkfree_trim_params *
2413 trim_lookup(ump, devvp, bno, size, inum, key, alloctype)
2414 struct ufsmount *ump;
2415 struct vnode *devvp;
2422 struct trimlist_hashhead *tphashhead;
2423 struct ffs_blkfree_trim_params *tp, *ntp;
2425 ntp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK);
2426 if (alloctype != SINGLE) {
2427 KASSERT(key >= FIRST_VALID_KEY, ("trim_lookup: invalid key"));
2429 tphashhead = TRIMLIST_HASH(ump, key);
2430 LIST_FOREACH(tp, tphashhead, hashlist)
2434 switch (alloctype) {
2436 KASSERT(tp == NULL, ("trim_lookup: found trim"));
2440 ("trim_lookup: missing call to ffs_blkrelease_start()"));
2445 KASSERT(tp != NULL, ("trim_lookup: missing REPLACE trim"));
2446 LIST_REMOVE(tp, hashlist);
2447 /* tp will be freed by caller */
2450 KASSERT(tp != NULL, ("trim_lookup: missing DONE trim"));
2451 LIST_REMOVE(tp, hashlist);
2456 TAILQ_INIT(&ntp->blklist);
2463 if (alloctype != SINGLE) {
2464 LIST_INSERT_HEAD(tphashhead, ntp, hashlist);
2471 * Dispatch a trim request.
2474 ffs_blkfree_sendtrim(tp)
2475 struct ffs_blkfree_trim_params *tp;
2477 struct ufsmount *ump;
2482 * Postpone the set of the free bit in the cg bitmap until the
2483 * BIO_DELETE is completed. Otherwise, due to disk queue
2484 * reordering, TRIM might be issued after we reuse the block
2485 * and write some new data into it.
2488 bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO);
2489 bp->b_iocmd = BIO_DELETE;
2490 bp->b_iooffset = dbtob(fsbtodb(ump->um_fs, tp->bno));
2491 bp->b_iodone = ffs_blkfree_trim_completed;
2492 bp->b_bcount = tp->size;
2493 bp->b_fsprivate1 = tp;
2495 ump->um_trim_total += 1;
2496 ump->um_trim_inflight += 1;
2497 ump->um_trim_inflight_blks += numfrags(ump->um_fs, tp->size);
2498 ump->um_trim_total_blks += numfrags(ump->um_fs, tp->size);
2502 vn_start_secondary_write(NULL, &mp, 0);
2503 g_vfs_strategy(ump->um_bo, bp);
2507 * Allocate a new key to use to identify a range of blocks.
2510 ffs_blkrelease_start(ump, devvp, inum)
2511 struct ufsmount *ump;
2512 struct vnode *devvp;
2515 static u_long masterkey;
2518 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2519 return (SINGLETON_KEY);
2521 key = atomic_fetchadd_long(&masterkey, 1);
2522 } while (key < FIRST_VALID_KEY);
2523 (void) trim_lookup(ump, devvp, 0, 0, inum, key, NEW);
2528 * Deallocate a key that has been used to identify a range of blocks.
2531 ffs_blkrelease_finish(ump, key)
2532 struct ufsmount *ump;
2535 struct ffs_blkfree_trim_params *tp;
2537 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0)
2540 * If the vfs.ffs.dotrimcons sysctl option is enabled while
2541 * a file deletion is active, specifically after a call
2542 * to ffs_blkrelease_start() but before the call to
2543 * ffs_blkrelease_finish(), ffs_blkrelease_start() will
2544 * have handed out SINGLETON_KEY rather than starting a
2545 * collection sequence. Thus if we get a SINGLETON_KEY
2546 * passed to ffs_blkrelease_finish(), we just return rather
2547 * than trying to finish the nonexistent sequence.
2549 if (key == SINGLETON_KEY) {
2551 printf("%s: vfs.ffs.dotrimcons enabled on active filesystem\n",
2552 ump->um_mountp->mnt_stat.f_mntonname);
2557 * We are done with sending blocks using this key. Look up the key
2558 * using the DONE alloctype (in tp) to request that it be unhashed
2559 * as we will not be adding to it. If the key has never been used,
2560 * tp->size will be zero, so we can just free tp. Otherwise the call
2561 * to ffs_blkfree_sendtrim(tp) causes the block range described by
2562 * tp to be issued (and then tp to be freed).
2564 tp = trim_lookup(ump, NULL, 0, 0, 0, key, DONE);
2568 ffs_blkfree_sendtrim(tp);
2572 * Setup to free a block or fragment.
2574 * Check for snapshots that might want to claim the block.
2575 * If trims are requested, prepare a trim request. Attempt to
2576 * aggregate consecutive blocks into a single trim request.
2579 ffs_blkfree(ump, fs, devvp, bno, size, inum, vtype, dephd, key)
2580 struct ufsmount *ump;
2582 struct vnode *devvp;
2587 struct workhead *dephd;
2590 struct ffs_blkfree_trim_params *tp, *ntp;
2591 struct trim_blkreq *blkelm;
2594 * Check to see if a snapshot wants to claim the block.
2595 * Check that devvp is a normal disk device, not a snapshot,
2596 * it has a snapshot(s) associated with it, and one of the
2597 * snapshots wants to claim the block.
2599 if (devvp->v_type == VCHR &&
2600 (devvp->v_vflag & VV_COPYONWRITE) &&
2601 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) {
2605 * Nothing to delay if TRIM is not required for this block or TRIM
2606 * is disabled or the operation is performed on a snapshot.
2608 if (key == NOTRIM_KEY || ((ump->um_flags & UM_CANDELETE) == 0) ||
2609 devvp->v_type == VREG) {
2610 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd);
2613 blkelm = malloc(sizeof(struct trim_blkreq), M_TRIM, M_WAITOK);
2615 blkelm->size = size;
2616 if (dephd == NULL) {
2617 blkelm->pdephd = NULL;
2619 LIST_INIT(&blkelm->dephd);
2620 LIST_SWAP(dephd, &blkelm->dephd, worklist, wk_list);
2621 blkelm->pdephd = &blkelm->dephd;
2623 if (key == SINGLETON_KEY) {
2625 * Just a single non-contiguous piece. Use the SINGLE
2626 * alloctype to return a trim request that will not be
2627 * hashed for future lookup.
2629 tp = trim_lookup(ump, devvp, bno, size, inum, key, SINGLE);
2630 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2631 ffs_blkfree_sendtrim(tp);
2635 * The callers of this function are not tracking whether or not
2636 * the blocks are contiguous. They are just saying that they
2637 * are freeing a set of blocks. It is this code that determines
2638 * the pieces of that range that are actually contiguous.
2640 * Calling ffs_blkrelease_start() will have created an entry
2643 tp = trim_lookup(ump, devvp, bno, size, inum, key, OLD);
2644 if (tp->size == 0) {
2646 * First block of a potential range, set block and size
2647 * for the trim block.
2651 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2655 * If this block is a continuation of the range (either
2656 * follows at the end or preceeds in the front) then we
2657 * add it to the front or back of the list and return.
2659 * If it is not a continuation of the trim that we were
2660 * building, using the REPLACE alloctype, we request that
2661 * the old trim request (still in tp) be unhashed and a
2662 * new range started (in ntp). The ffs_blkfree_sendtrim(tp)
2663 * call causes the block range described by tp to be issued
2664 * (and then tp to be freed).
2666 if (bno + numfrags(fs, size) == tp->bno) {
2667 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist);
2671 } else if (bno == tp->bno + numfrags(fs, tp->size)) {
2672 TAILQ_INSERT_TAIL(&tp->blklist, blkelm, blkreqlist);
2676 ntp = trim_lookup(ump, devvp, bno, size, inum, key, REPLACE);
2677 TAILQ_INSERT_HEAD(&ntp->blklist, blkelm, blkreqlist);
2678 ffs_blkfree_sendtrim(tp);
2683 * Verify allocation of a block or fragment. Returns true if block or
2684 * fragment is allocated, false if it is free.
2687 ffs_checkblk(ip, bno, size)
2696 int i, error, frags, free;
2700 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) {
2701 printf("bsize = %ld, size = %ld, fs = %s\n",
2702 (long)fs->fs_bsize, size, fs->fs_fsmnt);
2703 panic("ffs_checkblk: bad size");
2705 if ((u_int)bno >= fs->fs_size)
2706 panic("ffs_checkblk: bad block %jd", (intmax_t)bno);
2707 error = ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), &bp, &cgp);
2709 panic("ffs_checkblk: cylinder group read failed");
2710 blksfree = cg_blksfree(cgp);
2711 cgbno = dtogd(fs, bno);
2712 if (size == fs->fs_bsize) {
2713 free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno));
2715 frags = numfrags(fs, size);
2716 for (free = 0, i = 0; i < frags; i++)
2717 if (isset(blksfree, cgbno + i))
2719 if (free != 0 && free != frags)
2720 panic("ffs_checkblk: partially free fragment");
2725 #endif /* INVARIANTS */
2731 ffs_vfree(pvp, ino, mode)
2736 struct ufsmount *ump;
2738 if (DOINGSOFTDEP(pvp)) {
2739 softdep_freefile(pvp, ino, mode);
2742 ump = VFSTOUFS(pvp->v_mount);
2743 return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL));
2747 * Do the actual free operation.
2748 * The specified inode is placed back in the free map.
2751 ffs_freefile(ump, fs, devvp, ino, mode, wkhd)
2752 struct ufsmount *ump;
2754 struct vnode *devvp;
2757 struct workhead *wkhd;
2766 cg = ino_to_cg(fs, ino);
2767 if (devvp->v_type == VREG) {
2768 /* devvp is a snapshot */
2769 MPASS(devvp->v_mount->mnt_data == ump);
2770 dev = ump->um_devvp->v_rdev;
2771 } else if (devvp->v_type == VCHR) {
2772 /* devvp is a normal disk device */
2773 dev = devvp->v_rdev;
2778 if (ino >= fs->fs_ipg * fs->fs_ncg)
2779 panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s",
2780 devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt);
2781 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2783 inosused = cg_inosused(cgp);
2785 if (isclr(inosused, ino)) {
2786 printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev),
2787 (uintmax_t)(ino + cg * fs->fs_ipg), fs->fs_fsmnt);
2788 if (fs->fs_ronly == 0)
2789 panic("ffs_freefile: freeing free inode");
2791 clrbit(inosused, ino);
2792 if (ino < cgp->cg_irotor)
2793 cgp->cg_irotor = ino;
2794 cgp->cg_cs.cs_nifree++;
2796 fs->fs_cstotal.cs_nifree++;
2797 fs->fs_cs(fs, cg).cs_nifree++;
2798 if ((mode & IFMT) == IFDIR) {
2799 cgp->cg_cs.cs_ndir--;
2800 fs->fs_cstotal.cs_ndir--;
2801 fs->fs_cs(fs, cg).cs_ndir--;
2804 ACTIVECLEAR(fs, cg);
2806 if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR)
2807 softdep_setup_inofree(UFSTOVFS(ump), bp,
2808 ino + cg * fs->fs_ipg, wkhd);
2814 * Check to see if a file is free.
2815 * Used to check for allocated files in snapshots.
2818 ffs_checkfreefile(fs, devvp, ino)
2820 struct vnode *devvp;
2829 cg = ino_to_cg(fs, ino);
2830 if ((devvp->v_type != VREG) && (devvp->v_type != VCHR))
2832 if (ino >= fs->fs_ipg * fs->fs_ncg)
2834 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0)
2836 inosused = cg_inosused(cgp);
2838 ret = isclr(inosused, ino);
2844 * Find a block of the specified size in the specified cylinder group.
2846 * It is a panic if a request is made to find a block if none are
2850 ffs_mapsearch(fs, cgp, bpref, allocsiz)
2857 int start, len, loc, i;
2858 int blk, field, subfield, pos;
2862 * find the fragment by searching through the free block
2863 * map for an appropriate bit pattern
2866 start = dtogd(fs, bpref) / NBBY;
2868 start = cgp->cg_frotor / NBBY;
2869 blksfree = cg_blksfree(cgp);
2870 len = howmany(fs->fs_fpg, NBBY) - start;
2871 loc = scanc((u_int)len, (u_char *)&blksfree[start],
2872 fragtbl[fs->fs_frag],
2873 (u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2877 loc = scanc((u_int)len, (u_char *)&blksfree[0],
2878 fragtbl[fs->fs_frag],
2879 (u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY))));
2881 printf("start = %d, len = %d, fs = %s\n",
2882 start, len, fs->fs_fsmnt);
2883 panic("ffs_alloccg: map corrupted");
2887 bno = (start + len - loc) * NBBY;
2888 cgp->cg_frotor = bno;
2890 * found the byte in the map
2891 * sift through the bits to find the selected frag
2893 for (i = bno + NBBY; bno < i; bno += fs->fs_frag) {
2894 blk = blkmap(fs, blksfree, bno);
2896 field = around[allocsiz];
2897 subfield = inside[allocsiz];
2898 for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) {
2899 if ((blk & field) == subfield)
2905 printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt);
2906 panic("ffs_alloccg: block not in map");
2910 static const struct statfs *
2911 ffs_getmntstat(struct vnode *devvp)
2914 if (devvp->v_type == VCHR)
2915 return (&devvp->v_rdev->si_mountpt->mnt_stat);
2916 return (ffs_getmntstat(VFSTOUFS(devvp->v_mount)->um_devvp));
2920 * Fetch and verify a cylinder group.
2923 ffs_getcg(fs, devvp, cg, bpp, cgpp)
2925 struct vnode *devvp;
2932 const struct statfs *sfs;
2938 if ((fs->fs_metackhash & CK_CYLGRP) != 0)
2940 error = breadn_flags(devvp, devvp->v_type == VREG ?
2941 fragstoblks(fs, cgtod(fs, cg)) : fsbtodb(fs, cgtod(fs, cg)),
2942 (int)fs->fs_cgsize, NULL, NULL, 0, NOCRED, flags,
2943 ffs_ckhash_cg, &bp);
2946 cgp = (struct cg *)bp->b_data;
2947 if ((fs->fs_metackhash & CK_CYLGRP) != 0 &&
2948 (bp->b_flags & B_CKHASH) != 0 &&
2949 cgp->cg_ckhash != bp->b_ckhash) {
2950 sfs = ffs_getmntstat(devvp);
2951 printf("UFS %s%s (%s) cylinder checksum failed: cg %u, cgp: "
2952 "0x%x != bp: 0x%jx\n",
2953 devvp->v_type == VCHR ? "" : "snapshot of ",
2954 sfs->f_mntfromname, sfs->f_mntonname,
2955 cg, cgp->cg_ckhash, (uintmax_t)bp->b_ckhash);
2956 bp->b_flags &= ~B_CKHASH;
2957 bp->b_flags |= B_INVAL | B_NOCACHE;
2961 if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) {
2962 sfs = ffs_getmntstat(devvp);
2963 printf("UFS %s%s (%s)",
2964 devvp->v_type == VCHR ? "" : "snapshot of ",
2965 sfs->f_mntfromname, sfs->f_mntonname);
2966 if (!cg_chkmagic(cgp))
2967 printf(" cg %u: bad magic number 0x%x should be 0x%x\n",
2968 cg, cgp->cg_magic, CG_MAGIC);
2970 printf(": wrong cylinder group cg %u != cgx %u\n", cg,
2972 bp->b_flags &= ~B_CKHASH;
2973 bp->b_flags |= B_INVAL | B_NOCACHE;
2977 bp->b_flags &= ~B_CKHASH;
2978 bp->b_xflags |= BX_BKGRDWRITE;
2980 * If we are using check hashes on the cylinder group then we want
2981 * to limit changing the cylinder group time to when we are actually
2982 * going to write it to disk so that its check hash remains correct
2983 * in memory. If the CK_CYLGRP flag is set the time is updated in
2984 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we
2985 * update the time here as we have done historically.
2987 if ((fs->fs_metackhash & CK_CYLGRP) != 0)
2988 bp->b_xflags |= BX_CYLGRP;
2990 cgp->cg_old_time = cgp->cg_time = time_second;
3003 cgp = (struct cg *)bp->b_data;
3004 ckhash = cgp->cg_ckhash;
3006 bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount);
3007 cgp->cg_ckhash = ckhash;
3011 * Fserr prints the name of a filesystem with an error diagnostic.
3013 * The form of the error message is:
3017 ffs_fserr(fs, inum, cp)
3022 struct thread *td = curthread; /* XXX */
3023 struct proc *p = td->td_proc;
3025 log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n",
3026 p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum,
3031 * This function provides the capability for the fsck program to
3032 * update an active filesystem. Fourteen operations are provided:
3034 * adjrefcnt(inode, amt) - adjusts the reference count on the
3035 * specified inode by the specified amount. Under normal
3036 * operation the count should always go down. Decrementing
3037 * the count to zero will cause the inode to be freed.
3038 * adjblkcnt(inode, amt) - adjust the number of blocks used by the
3039 * inode by the specified amount.
3040 * adjsize(inode, size) - set the size of the inode to the
3042 * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) -
3043 * adjust the superblock summary.
3044 * freedirs(inode, count) - directory inodes [inode..inode + count - 1]
3045 * are marked as free. Inodes should never have to be marked
3047 * freefiles(inode, count) - file inodes [inode..inode + count - 1]
3048 * are marked as free. Inodes should never have to be marked
3050 * freeblks(blockno, size) - blocks [blockno..blockno + size - 1]
3051 * are marked as free. Blocks should never have to be marked
3053 * setflags(flags, set/clear) - the fs_flags field has the specified
3054 * flags set (second parameter +1) or cleared (second parameter -1).
3055 * setcwd(dirinode) - set the current directory to dirinode in the
3056 * filesystem associated with the snapshot.
3057 * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".."
3058 * in the current directory is oldvalue then change it to newvalue.
3059 * unlink(nameptr, oldvalue) - Verify that the inode number associated
3060 * with nameptr in the current directory is oldvalue then unlink it.
3062 * The following functions may only be used on a quiescent filesystem
3063 * by the soft updates journal. They are not safe to be run on an active
3066 * setinode(inode, dip) - the specified disk inode is replaced with the
3067 * contents pointed to by dip.
3068 * setbufoutput(fd, flags) - output associated with the specified file
3069 * descriptor (which must reference the character device supporting
3070 * the filesystem) switches from using physio to running through the
3071 * buffer cache when flags is set to 1. The descriptor reverts to
3072 * physio for output when flags is set to zero.
3075 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS);
3077 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt, CTLFLAG_WR|CTLTYPE_STRUCT,
3078 0, 0, sysctl_ffs_fsck, "S,fsck", "Adjust Inode Reference Count");
3080 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt, CTLFLAG_WR,
3081 sysctl_ffs_fsck, "Adjust Inode Used Blocks Count");
3083 static SYSCTL_NODE(_vfs_ffs, FFS_SET_SIZE, setsize, CTLFLAG_WR,
3084 sysctl_ffs_fsck, "Set the inode size");
3086 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir, CTLFLAG_WR,
3087 sysctl_ffs_fsck, "Adjust number of directories");
3089 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree, CTLFLAG_WR,
3090 sysctl_ffs_fsck, "Adjust number of free blocks");
3092 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree, CTLFLAG_WR,
3093 sysctl_ffs_fsck, "Adjust number of free inodes");
3095 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree, CTLFLAG_WR,
3096 sysctl_ffs_fsck, "Adjust number of free frags");
3098 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters, CTLFLAG_WR,
3099 sysctl_ffs_fsck, "Adjust number of free clusters");
3101 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs, CTLFLAG_WR,
3102 sysctl_ffs_fsck, "Free Range of Directory Inodes");
3104 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles, CTLFLAG_WR,
3105 sysctl_ffs_fsck, "Free Range of File Inodes");
3107 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks, CTLFLAG_WR,
3108 sysctl_ffs_fsck, "Free Range of Blocks");
3110 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags, CTLFLAG_WR,
3111 sysctl_ffs_fsck, "Change Filesystem Flags");
3113 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd, CTLFLAG_WR,
3114 sysctl_ffs_fsck, "Set Current Working Directory");
3116 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot, CTLFLAG_WR,
3117 sysctl_ffs_fsck, "Change Value of .. Entry");
3119 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink, CTLFLAG_WR,
3120 sysctl_ffs_fsck, "Unlink a Duplicate Name");
3122 static SYSCTL_NODE(_vfs_ffs, FFS_SET_INODE, setinode, CTLFLAG_WR,
3123 sysctl_ffs_fsck, "Update an On-Disk Inode");
3125 static SYSCTL_NODE(_vfs_ffs, FFS_SET_BUFOUTPUT, setbufoutput, CTLFLAG_WR,
3126 sysctl_ffs_fsck, "Set Buffered Writing for Descriptor");
3130 static int fsckcmds = 0;
3131 SYSCTL_INT(_debug, OID_AUTO, fsckcmds, CTLFLAG_RW, &fsckcmds, 0, "");
3134 static int buffered_write(struct file *, struct uio *, struct ucred *,
3135 int, struct thread *);
3138 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS)
3140 struct thread *td = curthread;
3141 struct fsck_cmd cmd;
3142 struct ufsmount *ump;
3143 struct vnode *vp, *dvp, *fdvp;
3144 struct inode *ip, *dp;
3148 long blkcnt, blksize;
3150 struct file *fp, *vfp;
3151 cap_rights_t rights;
3152 int filetype, error;
3153 static struct fileops *origops, bufferedops;
3155 if (req->newlen > sizeof cmd)
3157 if ((error = SYSCTL_IN(req, &cmd, sizeof cmd)) != 0)
3159 if (cmd.version != FFS_CMD_VERSION)
3160 return (ERPCMISMATCH);
3161 if ((error = getvnode(td, cmd.handle,
3162 cap_rights_init(&rights, CAP_FSCK), &fp)) != 0)
3165 if (vp->v_type != VREG && vp->v_type != VDIR) {
3169 vn_start_write(vp, &mp, V_WAIT);
3171 strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) {
3172 vn_finished_write(mp);
3177 if ((mp->mnt_flag & MNT_RDONLY) &&
3178 ump->um_fsckpid != td->td_proc->p_pid) {
3179 vn_finished_write(mp);
3186 switch (oidp->oid_number) {
3191 printf("%s: %s flags\n", mp->mnt_stat.f_mntonname,
3192 cmd.size > 0 ? "set" : "clear");
3195 fs->fs_flags |= (long)cmd.value;
3197 fs->fs_flags &= ~(long)cmd.value;
3200 case FFS_ADJ_REFCNT:
3203 printf("%s: adjust inode %jd link count by %jd\n",
3204 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3205 (intmax_t)cmd.size);
3208 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3211 ip->i_nlink += cmd.size;
3212 DIP_SET(ip, i_nlink, ip->i_nlink);
3213 ip->i_effnlink += cmd.size;
3214 ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3215 error = ffs_update(vp, 1);
3216 if (DOINGSOFTDEP(vp))
3217 softdep_change_linkcnt(ip);
3221 case FFS_ADJ_BLKCNT:
3224 printf("%s: adjust inode %jd block count by %jd\n",
3225 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3226 (intmax_t)cmd.size);
3229 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3232 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size);
3233 ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3234 error = ffs_update(vp, 1);
3241 printf("%s: set inode %jd size to %jd\n",
3242 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3243 (intmax_t)cmd.size);
3246 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3249 DIP_SET(ip, i_size, cmd.size);
3250 ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3251 error = ffs_update(vp, 1);
3263 printf("%s: free %s inode %ju\n",
3264 mp->mnt_stat.f_mntonname,
3265 filetype == IFDIR ? "directory" : "file",
3266 (uintmax_t)cmd.value);
3268 printf("%s: free %s inodes %ju-%ju\n",
3269 mp->mnt_stat.f_mntonname,
3270 filetype == IFDIR ? "directory" : "file",
3271 (uintmax_t)cmd.value,
3272 (uintmax_t)(cmd.value + cmd.size - 1));
3275 while (cmd.size > 0) {
3276 if ((error = ffs_freefile(ump, fs, ump->um_devvp,
3277 cmd.value, filetype, NULL)))
3288 printf("%s: free block %jd\n",
3289 mp->mnt_stat.f_mntonname,
3290 (intmax_t)cmd.value);
3292 printf("%s: free blocks %jd-%jd\n",
3293 mp->mnt_stat.f_mntonname,
3294 (intmax_t)cmd.value,
3295 (intmax_t)cmd.value + cmd.size - 1);
3300 blksize = fs->fs_frag - (blkno % fs->fs_frag);
3301 key = ffs_blkrelease_start(ump, ump->um_devvp, UFS_ROOTINO);
3302 while (blkcnt > 0) {
3303 if (blkcnt < blksize)
3305 ffs_blkfree(ump, fs, ump->um_devvp, blkno,
3306 blksize * fs->fs_fsize, UFS_ROOTINO,
3310 blksize = fs->fs_frag;
3312 ffs_blkrelease_finish(ump, key);
3316 * Adjust superblock summaries. fsck(8) is expected to
3317 * submit deltas when necessary.
3322 printf("%s: adjust number of directories by %jd\n",
3323 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3326 fs->fs_cstotal.cs_ndir += cmd.value;
3329 case FFS_ADJ_NBFREE:
3332 printf("%s: adjust number of free blocks by %+jd\n",
3333 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3336 fs->fs_cstotal.cs_nbfree += cmd.value;
3339 case FFS_ADJ_NIFREE:
3342 printf("%s: adjust number of free inodes by %+jd\n",
3343 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3346 fs->fs_cstotal.cs_nifree += cmd.value;
3349 case FFS_ADJ_NFFREE:
3352 printf("%s: adjust number of free frags by %+jd\n",
3353 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3356 fs->fs_cstotal.cs_nffree += cmd.value;
3359 case FFS_ADJ_NUMCLUSTERS:
3362 printf("%s: adjust number of free clusters by %+jd\n",
3363 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3366 fs->fs_cstotal.cs_numclusters += cmd.value;
3372 printf("%s: set current directory to inode %jd\n",
3373 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3376 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp)))
3378 AUDIT_ARG_VNODE1(vp);
3379 if ((error = change_dir(vp, td)) != 0) {
3387 case FFS_SET_DOTDOT:
3390 printf("%s: change .. in cwd from %jd to %jd\n",
3391 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value,
3392 (intmax_t)cmd.size);
3396 * First we have to get and lock the parent directory
3397 * to which ".." points.
3399 error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp);
3403 * Now we get and lock the child directory containing "..".
3405 FILEDESC_SLOCK(td->td_proc->p_fd);
3406 dvp = td->td_proc->p_fd->fd_cdir;
3407 FILEDESC_SUNLOCK(td->td_proc->p_fd);
3408 if ((error = vget(dvp, LK_EXCLUSIVE, td)) != 0) {
3413 dp->i_offset = 12; /* XXX mastertemplate.dot_reclen */
3414 error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size,
3427 if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL))
3428 strncpy(buf, "Name_too_long", 32);
3429 printf("%s: unlink %s (inode %jd)\n",
3430 mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size);
3434 * kern_unlinkat will do its own start/finish writes and
3435 * they do not nest, so drop ours here. Setting mp == NULL
3436 * indicates that vn_finished_write is not needed down below.
3438 vn_finished_write(mp);
3440 error = kern_unlinkat(td, AT_FDCWD, (char *)(intptr_t)cmd.value,
3441 UIO_USERSPACE, 0, (ino_t)cmd.size);
3445 if (ump->um_fsckpid != td->td_proc->p_pid) {
3451 printf("%s: update inode %jd\n",
3452 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value);
3455 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp)))
3457 AUDIT_ARG_VNODE1(vp);
3460 error = copyin((void *)(intptr_t)cmd.size, ip->i_din1,
3461 sizeof(struct ufs1_dinode));
3463 error = copyin((void *)(intptr_t)cmd.size, ip->i_din2,
3464 sizeof(struct ufs2_dinode));
3469 ip->i_flag |= IN_CHANGE | IN_MODIFIED;
3470 error = ffs_update(vp, 1);
3474 case FFS_SET_BUFOUTPUT:
3475 if (ump->um_fsckpid != td->td_proc->p_pid) {
3479 if (ITOUMP(VTOI(vp)) != ump) {
3485 printf("%s: %s buffered output for descriptor %jd\n",
3486 mp->mnt_stat.f_mntonname,
3487 cmd.size == 1 ? "enable" : "disable",
3488 (intmax_t)cmd.value);
3491 if ((error = getvnode(td, cmd.value,
3492 cap_rights_init(&rights, CAP_FSCK), &vfp)) != 0)
3494 if (vfp->f_vnode->v_type != VCHR) {
3499 if (origops == NULL) {
3500 origops = vfp->f_ops;
3501 bcopy((void *)origops, (void *)&bufferedops,
3502 sizeof(bufferedops));
3503 bufferedops.fo_write = buffered_write;
3506 atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3507 (uintptr_t)&bufferedops);
3509 atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops,
3510 (uintptr_t)origops);
3517 printf("Invalid request %d from fsck\n",
3526 vn_finished_write(mp);
3531 * Function to switch a descriptor to use the buffer cache to stage
3532 * its I/O. This is needed so that writes to the filesystem device
3533 * will give snapshots a chance to copy modified blocks for which it
3534 * needs to retain copies.
3537 buffered_write(fp, uio, active_cred, flags, td)
3540 struct ucred *active_cred;
3544 struct vnode *devvp, *vp;
3548 struct filedesc *fdp;
3553 * The devvp is associated with the /dev filesystem. To discover
3554 * the filesystem with which the device is associated, we depend
3555 * on the application setting the current directory to a location
3556 * within the filesystem being written. Yes, this is an ugly hack.
3558 devvp = fp->f_vnode;
3559 if (!vn_isdisk(devvp, NULL))
3561 fdp = td->td_proc->p_fd;
3562 FILEDESC_SLOCK(fdp);
3565 FILEDESC_SUNLOCK(fdp);
3566 vn_lock(vp, LK_SHARED | LK_RETRY);
3568 * Check that the current directory vnode indeed belongs to
3569 * UFS before trying to dereference UFS-specific v_data fields.
3571 if (vp->v_op != &ffs_vnodeops1 && vp->v_op != &ffs_vnodeops2) {
3576 if (ITODEVVP(ip) != devvp) {
3582 foffset_lock_uio(fp, uio, flags);
3583 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
3586 printf("%s: buffered write for block %jd\n",
3587 fs->fs_fsmnt, (intmax_t)btodb(uio->uio_offset));
3591 * All I/O must be contained within a filesystem block, start on
3592 * a fragment boundary, and be a multiple of fragments in length.
3594 if (uio->uio_resid > fs->fs_bsize - (uio->uio_offset % fs->fs_bsize) ||
3595 fragoff(fs, uio->uio_offset) != 0 ||
3596 fragoff(fs, uio->uio_resid) != 0) {
3600 lbn = numfrags(fs, uio->uio_offset);
3601 bp = getblk(devvp, lbn, uio->uio_resid, 0, 0, 0);
3602 bp->b_flags |= B_RELBUF;
3603 if ((error = uiomove((char *)bp->b_data, uio->uio_resid, uio)) != 0) {
3609 VOP_UNLOCK(devvp, 0);
3610 foffset_unlock_uio(fp, uio, flags | FOF_NEXTOFF);