1 /*- 2 * SPDX-License-Identifier: (BSD-2-Clause AND BSD-3-Clause) 3 * 4 * Copyright (c) 2002 Networks Associates Technology, Inc. 5 * All rights reserved. 6 * 7 * This software was developed for the FreeBSD Project by Marshall 8 * Kirk McKusick and Network Associates Laboratories, the Security 9 * Research Division of Network Associates, Inc. under DARPA/SPAWAR 10 * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS 11 * research program 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1982, 1986, 1989, 1993 35 * The Regents of the University of California. All rights reserved. 36 * 37 * Redistribution and use in source and binary forms, with or without 38 * modification, are permitted provided that the following conditions 39 * are met: 40 * 1. Redistributions of source code must retain the above copyright 41 * notice, this list of conditions and the following disclaimer. 42 * 2. Redistributions in binary form must reproduce the above copyright 43 * notice, this list of conditions and the following disclaimer in the 44 * documentation and/or other materials provided with the distribution. 45 * 3. Neither the name of the University nor the names of its contributors 46 * may be used to endorse or promote products derived from this software 47 * without specific prior written permission. 48 * 49 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 50 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 51 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 52 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 53 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 54 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 55 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 56 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 57 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 58 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 59 * SUCH DAMAGE. 60 */ 61 62 #include <sys/cdefs.h> 63 #include "opt_quota.h" 64 65 #include <sys/param.h> 66 #include <sys/systm.h> 67 #include <sys/bio.h> 68 #include <sys/buf.h> 69 #include <sys/capsicum.h> 70 #include <sys/conf.h> 71 #include <sys/fcntl.h> 72 #include <sys/file.h> 73 #include <sys/filedesc.h> 74 #include <sys/gsb_crc32.h> 75 #include <sys/kernel.h> 76 #include <sys/mount.h> 77 #include <sys/priv.h> 78 #include <sys/proc.h> 79 #include <sys/stat.h> 80 #include <sys/syscallsubr.h> 81 #include <sys/sysctl.h> 82 #include <sys/syslog.h> 83 #include <sys/taskqueue.h> 84 #include <sys/vnode.h> 85 86 #include <security/audit/audit.h> 87 88 #include <geom/geom.h> 89 #include <geom/geom_vfs.h> 90 91 #include <ufs/ufs/dir.h> 92 #include <ufs/ufs/extattr.h> 93 #include <ufs/ufs/quota.h> 94 #include <ufs/ufs/inode.h> 95 #include <ufs/ufs/ufs_extern.h> 96 #include <ufs/ufs/ufsmount.h> 97 98 #include <ufs/ffs/fs.h> 99 #include <ufs/ffs/ffs_extern.h> 100 #include <ufs/ffs/softdep.h> 101 102 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, uint64_t cg, 103 ufs2_daddr_t bpref, int size, int rsize); 104 105 static ufs2_daddr_t ffs_alloccg(struct inode *, uint64_t, ufs2_daddr_t, int, 106 int); 107 static ufs2_daddr_t 108 ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int); 109 static void ffs_blkfree_cg(struct ufsmount *, struct fs *, 110 struct vnode *, ufs2_daddr_t, long, ino_t, 111 struct workhead *); 112 #ifdef INVARIANTS 113 static int ffs_checkfreeblk(struct inode *, ufs2_daddr_t, long); 114 #endif 115 static void ffs_checkcgintegrity(struct fs *, uint64_t, int); 116 static ufs2_daddr_t ffs_clusteralloc(struct inode *, uint64_t, ufs2_daddr_t, 117 int); 118 static ino_t ffs_dirpref(struct inode *); 119 static ufs2_daddr_t ffs_fragextend(struct inode *, uint64_t, ufs2_daddr_t, 120 int, int); 121 static ufs2_daddr_t ffs_hashalloc(struct inode *, uint64_t, ufs2_daddr_t, 122 int, int, allocfcn_t *); 123 static ufs2_daddr_t ffs_nodealloccg(struct inode *, uint64_t, 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) quadratically 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) quadratically rehash into other cylinder groups, until an 147 * available block is located. 148 */ 149 int 150 ffs_alloc(struct inode *ip, 151 ufs2_daddr_t lbn, 152 ufs2_daddr_t bpref, 153 int size, 154 int flags, 155 struct ucred *cred, 156 ufs2_daddr_t *bnp) 157 { 158 struct fs *fs; 159 struct ufsmount *ump; 160 ufs2_daddr_t bno; 161 uint64_t cg, reclaimed; 162 int64_t delta; 163 #ifdef QUOTA 164 int error; 165 #endif 166 167 *bnp = 0; 168 ump = ITOUMP(ip); 169 fs = ump->um_fs; 170 mtx_assert(UFS_MTX(ump), MA_OWNED); 171 #ifdef INVARIANTS 172 if ((uint64_t)size > fs->fs_bsize || fragoff(fs, size) != 0) { 173 printf("dev = %s, bsize = %ld, size = %d, fs = %s\n", 174 devtoname(ump->um_dev), (long)fs->fs_bsize, size, 175 fs->fs_fsmnt); 176 panic("ffs_alloc: bad size"); 177 } 178 if (cred == NOCRED) 179 panic("ffs_alloc: missing credential"); 180 #endif /* INVARIANTS */ 181 reclaimed = 0; 182 retry: 183 #ifdef QUOTA 184 UFS_UNLOCK(ump); 185 error = chkdq(ip, btodb(size), cred, 0); 186 if (error) 187 return (error); 188 UFS_LOCK(ump); 189 #endif 190 if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0) 191 goto nospace; 192 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) && 193 freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0) 194 goto nospace; 195 if (bpref >= fs->fs_size) 196 bpref = 0; 197 if (bpref == 0) 198 cg = ino_to_cg(fs, ip->i_number); 199 else 200 cg = dtog(fs, bpref); 201 bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg); 202 if (bno > 0) { 203 delta = btodb(size); 204 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 205 if (flags & IO_EXT) 206 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 207 else 208 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); 209 *bnp = bno; 210 return (0); 211 } 212 nospace: 213 #ifdef QUOTA 214 UFS_UNLOCK(ump); 215 /* 216 * Restore user's disk quota because allocation failed. 217 */ 218 (void) chkdq(ip, -btodb(size), cred, FORCE); 219 UFS_LOCK(ump); 220 #endif 221 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) { 222 reclaimed = 1; 223 softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT); 224 goto retry; 225 } 226 if (ffs_fsfail_cleanup_locked(ump, 0)) { 227 UFS_UNLOCK(ump); 228 return (ENXIO); 229 } 230 if (reclaimed > 0 && 231 ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 232 UFS_UNLOCK(ump); 233 ffs_fserr(fs, ip->i_number, "filesystem full"); 234 uprintf("\n%s: write failed, filesystem is full\n", 235 fs->fs_fsmnt); 236 } else { 237 UFS_UNLOCK(ump); 238 } 239 return (ENOSPC); 240 } 241 242 /* 243 * Reallocate a fragment to a bigger size 244 * 245 * The number and size of the old block is given, and a preference 246 * and new size is also specified. The allocator attempts to extend 247 * the original block. Failing that, the regular block allocator is 248 * invoked to get an appropriate block. 249 */ 250 int 251 ffs_realloccg(struct inode *ip, 252 ufs2_daddr_t lbprev, 253 ufs2_daddr_t bprev, 254 ufs2_daddr_t bpref, 255 int osize, 256 int nsize, 257 int flags, 258 struct ucred *cred, 259 struct buf **bpp) 260 { 261 struct vnode *vp; 262 struct fs *fs; 263 struct buf *bp; 264 struct ufsmount *ump; 265 uint64_t cg, request, reclaimed; 266 int error, gbflags; 267 ufs2_daddr_t bno; 268 int64_t delta; 269 270 vp = ITOV(ip); 271 ump = ITOUMP(ip); 272 fs = ump->um_fs; 273 bp = NULL; 274 gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 275 #ifdef WITNESS 276 gbflags |= IS_SNAPSHOT(ip) ? GB_NOWITNESS : 0; 277 #endif 278 279 mtx_assert(UFS_MTX(ump), MA_OWNED); 280 #ifdef INVARIANTS 281 if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED) 282 panic("ffs_realloccg: allocation on suspended filesystem"); 283 if ((uint64_t)osize > fs->fs_bsize || fragoff(fs, osize) != 0 || 284 (uint64_t)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) { 285 printf( 286 "dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n", 287 devtoname(ump->um_dev), (long)fs->fs_bsize, osize, 288 nsize, fs->fs_fsmnt); 289 panic("ffs_realloccg: bad size"); 290 } 291 if (cred == NOCRED) 292 panic("ffs_realloccg: missing credential"); 293 #endif /* INVARIANTS */ 294 reclaimed = 0; 295 retry: 296 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) && 297 freespace(fs, fs->fs_minfree) - numfrags(fs, nsize - osize) < 0) { 298 goto nospace; 299 } 300 if (bprev == 0) { 301 printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n", 302 devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev, 303 fs->fs_fsmnt); 304 panic("ffs_realloccg: bad bprev"); 305 } 306 UFS_UNLOCK(ump); 307 /* 308 * Allocate the extra space in the buffer. 309 */ 310 error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp); 311 if (error) { 312 return (error); 313 } 314 315 if (bp->b_blkno == bp->b_lblkno) { 316 if (lbprev >= UFS_NDADDR) 317 panic("ffs_realloccg: lbprev out of range"); 318 bp->b_blkno = fsbtodb(fs, bprev); 319 } 320 321 #ifdef QUOTA 322 error = chkdq(ip, btodb(nsize - osize), cred, 0); 323 if (error) { 324 brelse(bp); 325 return (error); 326 } 327 #endif 328 /* 329 * Check for extension in the existing location. 330 */ 331 *bpp = NULL; 332 cg = dtog(fs, bprev); 333 UFS_LOCK(ump); 334 bno = ffs_fragextend(ip, cg, bprev, osize, nsize); 335 if (bno) { 336 if (bp->b_blkno != fsbtodb(fs, bno)) 337 panic("ffs_realloccg: bad blockno"); 338 delta = btodb(nsize - osize); 339 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 340 if (flags & IO_EXT) 341 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 342 else 343 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); 344 allocbuf(bp, nsize); 345 bp->b_flags |= B_DONE; 346 vfs_bio_bzero_buf(bp, osize, nsize - osize); 347 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO) 348 vfs_bio_set_valid(bp, osize, nsize - osize); 349 *bpp = bp; 350 return (0); 351 } 352 /* 353 * Allocate a new disk location. 354 */ 355 if (bpref >= fs->fs_size) 356 bpref = 0; 357 switch ((int)fs->fs_optim) { 358 case FS_OPTSPACE: 359 /* 360 * Allocate an exact sized fragment. Although this makes 361 * best use of space, we will waste time relocating it if 362 * the file continues to grow. If the fragmentation is 363 * less than half of the minimum free reserve, we choose 364 * to begin optimizing for time. 365 */ 366 request = nsize; 367 if (fs->fs_minfree <= 5 || 368 fs->fs_cstotal.cs_nffree > 369 (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100)) 370 break; 371 log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n", 372 fs->fs_fsmnt); 373 fs->fs_optim = FS_OPTTIME; 374 break; 375 case FS_OPTTIME: 376 /* 377 * At this point we have discovered a file that is trying to 378 * grow a small fragment to a larger fragment. To save time, 379 * we allocate a full sized block, then free the unused portion. 380 * If the file continues to grow, the `ffs_fragextend' call 381 * above will be able to grow it in place without further 382 * copying. If aberrant programs cause disk fragmentation to 383 * grow within 2% of the free reserve, we choose to begin 384 * optimizing for space. 385 */ 386 request = fs->fs_bsize; 387 if (fs->fs_cstotal.cs_nffree < 388 (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100) 389 break; 390 log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n", 391 fs->fs_fsmnt); 392 fs->fs_optim = FS_OPTSPACE; 393 break; 394 default: 395 printf("dev = %s, optim = %ld, fs = %s\n", 396 devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt); 397 panic("ffs_realloccg: bad optim"); 398 /* NOTREACHED */ 399 } 400 bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg); 401 if (bno > 0) { 402 bp->b_blkno = fsbtodb(fs, bno); 403 if (!DOINGSOFTDEP(vp)) 404 /* 405 * The usual case is that a smaller fragment that 406 * was just allocated has been replaced with a bigger 407 * fragment or a full-size block. If it is marked as 408 * B_DELWRI, the current contents have not been written 409 * to disk. It is possible that the block was written 410 * earlier, but very uncommon. If the block has never 411 * been written, there is no need to send a BIO_DELETE 412 * for it when it is freed. The gain from avoiding the 413 * TRIMs for the common case of unwritten blocks far 414 * exceeds the cost of the write amplification for the 415 * uncommon case of failing to send a TRIM for a block 416 * that had been written. 417 */ 418 ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize, 419 ip->i_number, vp->v_type, NULL, 420 (bp->b_flags & B_DELWRI) != 0 ? 421 NOTRIM_KEY : SINGLETON_KEY); 422 delta = btodb(nsize - osize); 423 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 424 if (flags & IO_EXT) 425 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 426 else 427 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); 428 allocbuf(bp, nsize); 429 bp->b_flags |= B_DONE; 430 vfs_bio_bzero_buf(bp, osize, nsize - osize); 431 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO) 432 vfs_bio_set_valid(bp, osize, nsize - osize); 433 *bpp = bp; 434 return (0); 435 } 436 #ifdef QUOTA 437 UFS_UNLOCK(ump); 438 /* 439 * Restore user's disk quota because allocation failed. 440 */ 441 (void) chkdq(ip, -btodb(nsize - osize), cred, FORCE); 442 UFS_LOCK(ump); 443 #endif 444 nospace: 445 /* 446 * no space available 447 */ 448 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) { 449 reclaimed = 1; 450 UFS_UNLOCK(ump); 451 if (bp) { 452 brelse(bp); 453 bp = NULL; 454 } 455 UFS_LOCK(ump); 456 softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT); 457 goto retry; 458 } 459 if (bp) 460 brelse(bp); 461 if (ffs_fsfail_cleanup_locked(ump, 0)) { 462 UFS_UNLOCK(ump); 463 return (ENXIO); 464 } 465 if (reclaimed > 0 && 466 ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 467 UFS_UNLOCK(ump); 468 ffs_fserr(fs, ip->i_number, "filesystem full"); 469 uprintf("\n%s: write failed, filesystem is full\n", 470 fs->fs_fsmnt); 471 } else { 472 UFS_UNLOCK(ump); 473 } 474 return (ENOSPC); 475 } 476 477 /* 478 * Reallocate a sequence of blocks into a contiguous sequence of blocks. 479 * 480 * The vnode and an array of buffer pointers for a range of sequential 481 * logical blocks to be made contiguous is given. The allocator attempts 482 * to find a range of sequential blocks starting as close as possible 483 * from the end of the allocation for the logical block immediately 484 * preceding the current range. If successful, the physical block numbers 485 * in the buffer pointers and in the inode are changed to reflect the new 486 * allocation. If unsuccessful, the allocation is left unchanged. The 487 * success in doing the reallocation is returned. Note that the error 488 * return is not reflected back to the user. Rather the previous block 489 * allocation will be used. 490 */ 491 492 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 493 "FFS filesystem"); 494 495 static int doasyncfree = 1; 496 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0, 497 "do not force synchronous writes when blocks are reallocated"); 498 499 static int doreallocblks = 1; 500 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0, 501 "enable block reallocation"); 502 503 static int dotrimcons = 1; 504 SYSCTL_INT(_vfs_ffs, OID_AUTO, dotrimcons, CTLFLAG_RWTUN, &dotrimcons, 0, 505 "enable BIO_DELETE / TRIM consolidation"); 506 507 static int maxclustersearch = 10; 508 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch, 509 0, "max number of cylinder group to search for contigous blocks"); 510 511 #ifdef DIAGNOSTIC 512 static int prtrealloc = 0; 513 SYSCTL_INT(_debug, OID_AUTO, ffs_prtrealloc, CTLFLAG_RW, &prtrealloc, 0, 514 "print out FFS filesystem block reallocation operations"); 515 #endif 516 517 int 518 ffs_reallocblks( 519 struct vop_reallocblks_args /* { 520 struct vnode *a_vp; 521 struct cluster_save *a_buflist; 522 } */ *ap) 523 { 524 struct ufsmount *ump; 525 int error; 526 527 /* 528 * We used to skip reallocating the blocks of a file into a 529 * contiguous sequence if the underlying flash device requested 530 * BIO_DELETE notifications, because devices that benefit from 531 * BIO_DELETE also benefit from not moving the data. However, 532 * the destination for the data is usually moved before the data 533 * is written to the initially allocated location, so we rarely 534 * suffer the penalty of extra writes. With the addition of the 535 * consolidation of contiguous blocks into single BIO_DELETE 536 * operations, having fewer but larger contiguous blocks reduces 537 * the number of (slow and expensive) BIO_DELETE operations. So 538 * when doing BIO_DELETE consolidation, we do block reallocation. 539 * 540 * Skip if reallocblks has been disabled globally. 541 */ 542 ump = ap->a_vp->v_mount->mnt_data; 543 if ((((ump->um_flags) & UM_CANDELETE) != 0 && dotrimcons == 0) || 544 doreallocblks == 0) 545 return (ENOSPC); 546 547 /* 548 * We can't wait in softdep prealloc as it may fsync and recurse 549 * here. Instead we simply fail to reallocate blocks if this 550 * rare condition arises. 551 */ 552 if (DOINGSUJ(ap->a_vp)) 553 if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0) 554 return (ENOSPC); 555 vn_seqc_write_begin(ap->a_vp); 556 error = ump->um_fstype == UFS1 ? ffs_reallocblks_ufs1(ap) : 557 ffs_reallocblks_ufs2(ap); 558 vn_seqc_write_end(ap->a_vp); 559 return (error); 560 } 561 562 static int 563 ffs_reallocblks_ufs1( 564 struct vop_reallocblks_args /* { 565 struct vnode *a_vp; 566 struct cluster_save *a_buflist; 567 } */ *ap) 568 { 569 struct fs *fs; 570 struct inode *ip; 571 struct vnode *vp; 572 struct buf *sbp, *ebp, *bp; 573 ufs1_daddr_t *bap, *sbap, *ebap; 574 struct cluster_save *buflist; 575 struct ufsmount *ump; 576 ufs_lbn_t start_lbn, end_lbn; 577 ufs1_daddr_t soff, newblk, blkno; 578 ufs2_daddr_t pref; 579 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp; 580 int i, cg, len, start_lvl, end_lvl, ssize; 581 582 vp = ap->a_vp; 583 ip = VTOI(vp); 584 ump = ITOUMP(ip); 585 fs = ump->um_fs; 586 /* 587 * If we are not tracking block clusters or if we have less than 4% 588 * free blocks left, then do not attempt to cluster. Running with 589 * less than 5% free block reserve is not recommended and those that 590 * choose to do so do not expect to have good file layout. 591 */ 592 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0) 593 return (ENOSPC); 594 buflist = ap->a_buflist; 595 len = buflist->bs_nchildren; 596 start_lbn = buflist->bs_children[0]->b_lblkno; 597 end_lbn = start_lbn + len - 1; 598 #ifdef INVARIANTS 599 for (i = 0; i < len; i++) 600 if (!ffs_checkfreeblk(ip, 601 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 602 panic("ffs_reallocblks: unallocated block 1"); 603 for (i = 1; i < len; i++) 604 if (buflist->bs_children[i]->b_lblkno != start_lbn + i) 605 panic("ffs_reallocblks: non-logical cluster"); 606 blkno = buflist->bs_children[0]->b_blkno; 607 ssize = fsbtodb(fs, fs->fs_frag); 608 for (i = 1; i < len - 1; i++) 609 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize)) 610 panic("ffs_reallocblks: non-physical cluster %d", i); 611 #endif 612 /* 613 * If the cluster crosses the boundary for the first indirect 614 * block, leave space for the indirect block. Indirect blocks 615 * are initially laid out in a position after the last direct 616 * block. Block reallocation would usually destroy locality by 617 * moving the indirect block out of the way to make room for 618 * data blocks if we didn't compensate here. We should also do 619 * this for other indirect block boundaries, but it is only 620 * important for the first one. 621 */ 622 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR) 623 return (ENOSPC); 624 /* 625 * If the latest allocation is in a new cylinder group, assume that 626 * the filesystem has decided to move and do not force it back to 627 * the previous cylinder group. 628 */ 629 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) != 630 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno))) 631 return (ENOSPC); 632 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) || 633 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl)) 634 return (ENOSPC); 635 /* 636 * Get the starting offset and block map for the first block. 637 */ 638 if (start_lvl == 0) { 639 sbap = &ip->i_din1->di_db[0]; 640 soff = start_lbn; 641 } else { 642 idp = &start_ap[start_lvl - 1]; 643 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) { 644 brelse(sbp); 645 return (ENOSPC); 646 } 647 sbap = (ufs1_daddr_t *)sbp->b_data; 648 soff = idp->in_off; 649 } 650 /* 651 * If the block range spans two block maps, get the second map. 652 */ 653 ebap = NULL; 654 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) { 655 ssize = len; 656 } else { 657 #ifdef INVARIANTS 658 if (start_lvl > 0 && 659 start_ap[start_lvl - 1].in_lbn == idp->in_lbn) 660 panic("ffs_reallocblk: start == end"); 661 #endif 662 ssize = len - (idp->in_off + 1); 663 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp)) 664 goto fail; 665 ebap = (ufs1_daddr_t *)ebp->b_data; 666 } 667 /* 668 * Find the preferred location for the cluster. If we have not 669 * previously failed at this endeavor, then follow our standard 670 * preference calculation. If we have failed at it, then pick up 671 * where we last ended our search. 672 */ 673 UFS_LOCK(ump); 674 if (ip->i_nextclustercg == -1) 675 pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap); 676 else 677 pref = cgdata(fs, ip->i_nextclustercg); 678 /* 679 * Search the block map looking for an allocation of the desired size. 680 * To avoid wasting too much time, we limit the number of cylinder 681 * groups that we will search. 682 */ 683 cg = dtog(fs, pref); 684 MPASS(cg < fs->fs_ncg); 685 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) { 686 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0) 687 break; 688 cg += 1; 689 if (cg >= fs->fs_ncg) 690 cg = 0; 691 } 692 /* 693 * If we have failed in our search, record where we gave up for 694 * next time. Otherwise, fall back to our usual search citerion. 695 */ 696 if (newblk == 0) { 697 ip->i_nextclustercg = cg; 698 UFS_UNLOCK(ump); 699 goto fail; 700 } 701 ip->i_nextclustercg = -1; 702 /* 703 * We have found a new contiguous block. 704 * 705 * First we have to replace the old block pointers with the new 706 * block pointers in the inode and indirect blocks associated 707 * with the file. 708 */ 709 #ifdef DIAGNOSTIC 710 if (prtrealloc) 711 printf("realloc: ino %ju, lbns %jd-%jd\n\told:", 712 (uintmax_t)ip->i_number, 713 (intmax_t)start_lbn, (intmax_t)end_lbn); 714 #endif 715 blkno = newblk; 716 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) { 717 if (i == ssize) { 718 bap = ebap; 719 soff = -i; 720 } 721 #ifdef INVARIANTS 722 if (!ffs_checkfreeblk(ip, 723 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 724 panic("ffs_reallocblks: unallocated block 2"); 725 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap) 726 panic("ffs_reallocblks: alloc mismatch"); 727 #endif 728 #ifdef DIAGNOSTIC 729 if (prtrealloc) 730 printf(" %d,", *bap); 731 #endif 732 if (DOINGSOFTDEP(vp)) { 733 if (sbap == &ip->i_din1->di_db[0] && i < ssize) 734 softdep_setup_allocdirect(ip, start_lbn + i, 735 blkno, *bap, fs->fs_bsize, fs->fs_bsize, 736 buflist->bs_children[i]); 737 else 738 softdep_setup_allocindir_page(ip, start_lbn + i, 739 i < ssize ? sbp : ebp, soff + i, blkno, 740 *bap, buflist->bs_children[i]); 741 } 742 *bap++ = blkno; 743 } 744 /* 745 * Next we must write out the modified inode and indirect blocks. 746 * For strict correctness, the writes should be synchronous since 747 * the old block values may have been written to disk. In practise 748 * they are almost never written, but if we are concerned about 749 * strict correctness, the `doasyncfree' flag should be set to zero. 750 * 751 * The test on `doasyncfree' should be changed to test a flag 752 * that shows whether the associated buffers and inodes have 753 * been written. The flag should be set when the cluster is 754 * started and cleared whenever the buffer or inode is flushed. 755 * We can then check below to see if it is set, and do the 756 * synchronous write only when it has been cleared. 757 */ 758 if (sbap != &ip->i_din1->di_db[0]) { 759 if (doasyncfree) 760 bdwrite(sbp); 761 else 762 bwrite(sbp); 763 } else { 764 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); 765 if (!doasyncfree) 766 ffs_update(vp, 1); 767 } 768 if (ssize < len) { 769 if (doasyncfree) 770 bdwrite(ebp); 771 else 772 bwrite(ebp); 773 } 774 /* 775 * Last, free the old blocks and assign the new blocks to the buffers. 776 */ 777 #ifdef DIAGNOSTIC 778 if (prtrealloc) 779 printf("\n\tnew:"); 780 #endif 781 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) { 782 bp = buflist->bs_children[i]; 783 if (!DOINGSOFTDEP(vp)) 784 /* 785 * The usual case is that a set of N-contiguous blocks 786 * that was just allocated has been replaced with a 787 * set of N+1-contiguous blocks. If they are marked as 788 * B_DELWRI, the current contents have not been written 789 * to disk. It is possible that the blocks were written 790 * earlier, but very uncommon. If the blocks have never 791 * been written, there is no need to send a BIO_DELETE 792 * for them when they are freed. The gain from avoiding 793 * the TRIMs for the common case of unwritten blocks 794 * far exceeds the cost of the write amplification for 795 * the uncommon case of failing to send a TRIM for the 796 * blocks that had been written. 797 */ 798 ffs_blkfree(ump, fs, ump->um_devvp, 799 dbtofsb(fs, bp->b_blkno), 800 fs->fs_bsize, ip->i_number, vp->v_type, NULL, 801 (bp->b_flags & B_DELWRI) != 0 ? 802 NOTRIM_KEY : SINGLETON_KEY); 803 bp->b_blkno = fsbtodb(fs, blkno); 804 #ifdef INVARIANTS 805 if (!ffs_checkfreeblk(ip, dbtofsb(fs, bp->b_blkno), 806 fs->fs_bsize)) 807 panic("ffs_reallocblks: unallocated block 3"); 808 #endif 809 #ifdef DIAGNOSTIC 810 if (prtrealloc) 811 printf(" %d,", blkno); 812 #endif 813 } 814 #ifdef DIAGNOSTIC 815 if (prtrealloc) { 816 prtrealloc--; 817 printf("\n"); 818 } 819 #endif 820 return (0); 821 822 fail: 823 if (ssize < len) 824 brelse(ebp); 825 if (sbap != &ip->i_din1->di_db[0]) 826 brelse(sbp); 827 return (ENOSPC); 828 } 829 830 static int 831 ffs_reallocblks_ufs2( 832 struct vop_reallocblks_args /* { 833 struct vnode *a_vp; 834 struct cluster_save *a_buflist; 835 } */ *ap) 836 { 837 struct fs *fs; 838 struct inode *ip; 839 struct vnode *vp; 840 struct buf *sbp, *ebp, *bp; 841 ufs2_daddr_t *bap, *sbap, *ebap; 842 struct cluster_save *buflist; 843 struct ufsmount *ump; 844 ufs_lbn_t start_lbn, end_lbn; 845 ufs2_daddr_t soff, newblk, blkno, pref; 846 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp; 847 int i, cg, len, start_lvl, end_lvl, ssize; 848 849 vp = ap->a_vp; 850 ip = VTOI(vp); 851 ump = ITOUMP(ip); 852 fs = ump->um_fs; 853 /* 854 * If we are not tracking block clusters or if we have less than 4% 855 * free blocks left, then do not attempt to cluster. Running with 856 * less than 5% free block reserve is not recommended and those that 857 * choose to do so do not expect to have good file layout. 858 */ 859 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0) 860 return (ENOSPC); 861 buflist = ap->a_buflist; 862 len = buflist->bs_nchildren; 863 start_lbn = buflist->bs_children[0]->b_lblkno; 864 end_lbn = start_lbn + len - 1; 865 #ifdef INVARIANTS 866 for (i = 0; i < len; i++) 867 if (!ffs_checkfreeblk(ip, 868 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 869 panic("ffs_reallocblks: unallocated block 1"); 870 for (i = 1; i < len; i++) 871 if (buflist->bs_children[i]->b_lblkno != start_lbn + i) 872 panic("ffs_reallocblks: non-logical cluster"); 873 blkno = buflist->bs_children[0]->b_blkno; 874 ssize = fsbtodb(fs, fs->fs_frag); 875 for (i = 1; i < len - 1; i++) 876 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize)) 877 panic("ffs_reallocblks: non-physical cluster %d", i); 878 #endif 879 /* 880 * If the cluster crosses the boundary for the first indirect 881 * block, do not move anything in it. Indirect blocks are 882 * usually initially laid out in a position between the data 883 * blocks. Block reallocation would usually destroy locality by 884 * moving the indirect block out of the way to make room for 885 * data blocks if we didn't compensate here. We should also do 886 * this for other indirect block boundaries, but it is only 887 * important for the first one. 888 */ 889 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR) 890 return (ENOSPC); 891 /* 892 * If the latest allocation is in a new cylinder group, assume that 893 * the filesystem has decided to move and do not force it back to 894 * the previous cylinder group. 895 */ 896 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) != 897 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno))) 898 return (ENOSPC); 899 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) || 900 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl)) 901 return (ENOSPC); 902 /* 903 * Get the starting offset and block map for the first block. 904 */ 905 if (start_lvl == 0) { 906 sbap = &ip->i_din2->di_db[0]; 907 soff = start_lbn; 908 } else { 909 idp = &start_ap[start_lvl - 1]; 910 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) { 911 brelse(sbp); 912 return (ENOSPC); 913 } 914 sbap = (ufs2_daddr_t *)sbp->b_data; 915 soff = idp->in_off; 916 } 917 /* 918 * If the block range spans two block maps, get the second map. 919 */ 920 ebap = NULL; 921 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) { 922 ssize = len; 923 } else { 924 #ifdef INVARIANTS 925 if (start_lvl > 0 && 926 start_ap[start_lvl - 1].in_lbn == idp->in_lbn) 927 panic("ffs_reallocblk: start == end"); 928 #endif 929 ssize = len - (idp->in_off + 1); 930 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp)) 931 goto fail; 932 ebap = (ufs2_daddr_t *)ebp->b_data; 933 } 934 /* 935 * Find the preferred location for the cluster. If we have not 936 * previously failed at this endeavor, then follow our standard 937 * preference calculation. If we have failed at it, then pick up 938 * where we last ended our search. 939 */ 940 UFS_LOCK(ump); 941 if (ip->i_nextclustercg == -1) 942 pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap); 943 else 944 pref = cgdata(fs, ip->i_nextclustercg); 945 /* 946 * Search the block map looking for an allocation of the desired size. 947 * To avoid wasting too much time, we limit the number of cylinder 948 * groups that we will search. 949 */ 950 cg = dtog(fs, pref); 951 MPASS(cg < fs->fs_ncg); 952 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) { 953 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0) 954 break; 955 cg += 1; 956 if (cg >= fs->fs_ncg) 957 cg = 0; 958 } 959 /* 960 * If we have failed in our search, record where we gave up for 961 * next time. Otherwise, fall back to our usual search citerion. 962 */ 963 if (newblk == 0) { 964 ip->i_nextclustercg = cg; 965 UFS_UNLOCK(ump); 966 goto fail; 967 } 968 ip->i_nextclustercg = -1; 969 /* 970 * We have found a new contiguous block. 971 * 972 * First we have to replace the old block pointers with the new 973 * block pointers in the inode and indirect blocks associated 974 * with the file. 975 */ 976 #ifdef DIAGNOSTIC 977 if (prtrealloc) 978 printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number, 979 (intmax_t)start_lbn, (intmax_t)end_lbn); 980 #endif 981 blkno = newblk; 982 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) { 983 if (i == ssize) { 984 bap = ebap; 985 soff = -i; 986 } 987 #ifdef INVARIANTS 988 if (!ffs_checkfreeblk(ip, 989 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 990 panic("ffs_reallocblks: unallocated block 2"); 991 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap) 992 panic("ffs_reallocblks: alloc mismatch"); 993 #endif 994 #ifdef DIAGNOSTIC 995 if (prtrealloc) 996 printf(" %jd,", (intmax_t)*bap); 997 #endif 998 if (DOINGSOFTDEP(vp)) { 999 if (sbap == &ip->i_din2->di_db[0] && i < ssize) 1000 softdep_setup_allocdirect(ip, start_lbn + i, 1001 blkno, *bap, fs->fs_bsize, fs->fs_bsize, 1002 buflist->bs_children[i]); 1003 else 1004 softdep_setup_allocindir_page(ip, start_lbn + i, 1005 i < ssize ? sbp : ebp, soff + i, blkno, 1006 *bap, buflist->bs_children[i]); 1007 } 1008 *bap++ = blkno; 1009 } 1010 /* 1011 * Next we must write out the modified inode and indirect blocks. 1012 * For strict correctness, the writes should be synchronous since 1013 * the old block values may have been written to disk. In practise 1014 * they are almost never written, but if we are concerned about 1015 * strict correctness, the `doasyncfree' flag should be set to zero. 1016 * 1017 * The test on `doasyncfree' should be changed to test a flag 1018 * that shows whether the associated buffers and inodes have 1019 * been written. The flag should be set when the cluster is 1020 * started and cleared whenever the buffer or inode is flushed. 1021 * We can then check below to see if it is set, and do the 1022 * synchronous write only when it has been cleared. 1023 */ 1024 if (sbap != &ip->i_din2->di_db[0]) { 1025 if (doasyncfree) 1026 bdwrite(sbp); 1027 else 1028 bwrite(sbp); 1029 } else { 1030 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); 1031 if (!doasyncfree) 1032 ffs_update(vp, 1); 1033 } 1034 if (ssize < len) { 1035 if (doasyncfree) 1036 bdwrite(ebp); 1037 else 1038 bwrite(ebp); 1039 } 1040 /* 1041 * Last, free the old blocks and assign the new blocks to the buffers. 1042 */ 1043 #ifdef DIAGNOSTIC 1044 if (prtrealloc) 1045 printf("\n\tnew:"); 1046 #endif 1047 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) { 1048 bp = buflist->bs_children[i]; 1049 if (!DOINGSOFTDEP(vp)) 1050 /* 1051 * The usual case is that a set of N-contiguous blocks 1052 * that was just allocated has been replaced with a 1053 * set of N+1-contiguous blocks. If they are marked as 1054 * B_DELWRI, the current contents have not been written 1055 * to disk. It is possible that the blocks were written 1056 * earlier, but very uncommon. If the blocks have never 1057 * been written, there is no need to send a BIO_DELETE 1058 * for them when they are freed. The gain from avoiding 1059 * the TRIMs for the common case of unwritten blocks 1060 * far exceeds the cost of the write amplification for 1061 * the uncommon case of failing to send a TRIM for the 1062 * blocks that had been written. 1063 */ 1064 ffs_blkfree(ump, fs, ump->um_devvp, 1065 dbtofsb(fs, bp->b_blkno), 1066 fs->fs_bsize, ip->i_number, vp->v_type, NULL, 1067 (bp->b_flags & B_DELWRI) != 0 ? 1068 NOTRIM_KEY : SINGLETON_KEY); 1069 bp->b_blkno = fsbtodb(fs, blkno); 1070 #ifdef INVARIANTS 1071 if (!ffs_checkfreeblk(ip, dbtofsb(fs, bp->b_blkno), 1072 fs->fs_bsize)) 1073 panic("ffs_reallocblks: unallocated block 3"); 1074 #endif 1075 #ifdef DIAGNOSTIC 1076 if (prtrealloc) 1077 printf(" %jd,", (intmax_t)blkno); 1078 #endif 1079 } 1080 #ifdef DIAGNOSTIC 1081 if (prtrealloc) { 1082 prtrealloc--; 1083 printf("\n"); 1084 } 1085 #endif 1086 return (0); 1087 1088 fail: 1089 if (ssize < len) 1090 brelse(ebp); 1091 if (sbap != &ip->i_din2->di_db[0]) 1092 brelse(sbp); 1093 return (ENOSPC); 1094 } 1095 1096 /* 1097 * Allocate an inode in the filesystem. 1098 * 1099 * If allocating a directory, use ffs_dirpref to select the inode. 1100 * If allocating in a directory, the following hierarchy is followed: 1101 * 1) allocate the preferred inode. 1102 * 2) allocate an inode in the same cylinder group. 1103 * 3) quadratically rehash into other cylinder groups, until an 1104 * available inode is located. 1105 * If no inode preference is given the following hierarchy is used 1106 * to allocate an inode: 1107 * 1) allocate an inode in cylinder group 0. 1108 * 2) quadratically rehash into other cylinder groups, until an 1109 * available inode is located. 1110 */ 1111 int 1112 ffs_valloc(struct vnode *pvp, 1113 int mode, 1114 struct ucred *cred, 1115 struct vnode **vpp) 1116 { 1117 struct inode *pip; 1118 struct fs *fs; 1119 struct inode *ip; 1120 struct timespec ts; 1121 struct ufsmount *ump; 1122 ino_t ino, ipref; 1123 uint64_t cg; 1124 int error, reclaimed; 1125 1126 *vpp = NULL; 1127 pip = VTOI(pvp); 1128 ump = ITOUMP(pip); 1129 fs = ump->um_fs; 1130 1131 UFS_LOCK(ump); 1132 reclaimed = 0; 1133 retry: 1134 if (fs->fs_cstotal.cs_nifree == 0) 1135 goto noinodes; 1136 1137 if ((mode & IFMT) == IFDIR) 1138 ipref = ffs_dirpref(pip); 1139 else 1140 ipref = pip->i_number; 1141 if (ipref >= fs->fs_ncg * fs->fs_ipg) 1142 ipref = 0; 1143 cg = ino_to_cg(fs, ipref); 1144 /* 1145 * Track number of dirs created one after another 1146 * in a same cg without intervening by files. 1147 */ 1148 if ((mode & IFMT) == IFDIR) { 1149 if (fs->fs_contigdirs[cg] < 255) 1150 fs->fs_contigdirs[cg]++; 1151 } else { 1152 if (fs->fs_contigdirs[cg] > 0) 1153 fs->fs_contigdirs[cg]--; 1154 } 1155 ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0, 1156 (allocfcn_t *)ffs_nodealloccg); 1157 if (ino == 0) 1158 goto noinodes; 1159 /* 1160 * Get rid of the cached old vnode, force allocation of a new vnode 1161 * for this inode. If this fails, release the allocated ino and 1162 * return the error. 1163 */ 1164 if ((error = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp, 1165 FFSV_FORCEINSMQ | FFSV_REPLACE | FFSV_NEWINODE)) != 0) { 1166 ffs_vfree(pvp, ino, mode); 1167 return (error); 1168 } 1169 /* 1170 * We got an inode, so check mode and panic if it is already allocated. 1171 */ 1172 ip = VTOI(*vpp); 1173 if (ip->i_mode) { 1174 printf("mode = 0%o, inum = %ju, fs = %s\n", 1175 ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt); 1176 panic("ffs_valloc: dup alloc"); 1177 } 1178 if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) { /* XXX */ 1179 printf("free inode %s/%ju had %ld blocks\n", 1180 fs->fs_fsmnt, (intmax_t)ino, (long)DIP(ip, i_blocks)); 1181 DIP_SET(ip, i_blocks, 0); 1182 } 1183 ip->i_flags = 0; 1184 DIP_SET(ip, i_flags, 0); 1185 if ((mode & IFMT) == IFDIR) 1186 DIP_SET(ip, i_dirdepth, DIP(pip, i_dirdepth) + 1); 1187 /* 1188 * Set up a new generation number for this inode. 1189 */ 1190 while (ip->i_gen == 0 || ++ip->i_gen == 0) 1191 ip->i_gen = arc4random(); 1192 DIP_SET(ip, i_gen, ip->i_gen); 1193 if (fs->fs_magic == FS_UFS2_MAGIC) { 1194 vfs_timestamp(&ts); 1195 ip->i_din2->di_birthtime = ts.tv_sec; 1196 ip->i_din2->di_birthnsec = ts.tv_nsec; 1197 } 1198 ip->i_flag = 0; 1199 (*vpp)->v_vflag = 0; 1200 (*vpp)->v_type = VNON; 1201 if (fs->fs_magic == FS_UFS2_MAGIC) { 1202 (*vpp)->v_op = &ffs_vnodeops2; 1203 UFS_INODE_SET_FLAG(ip, IN_UFS2); 1204 } else { 1205 (*vpp)->v_op = &ffs_vnodeops1; 1206 } 1207 return (0); 1208 noinodes: 1209 if (reclaimed == 0) { 1210 reclaimed = 1; 1211 softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT); 1212 goto retry; 1213 } 1214 if (ffs_fsfail_cleanup_locked(ump, 0)) { 1215 UFS_UNLOCK(ump); 1216 return (ENXIO); 1217 } 1218 if (ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 1219 UFS_UNLOCK(ump); 1220 ffs_fserr(fs, pip->i_number, "out of inodes"); 1221 uprintf("\n%s: create/symlink failed, no inodes free\n", 1222 fs->fs_fsmnt); 1223 } else { 1224 UFS_UNLOCK(ump); 1225 } 1226 return (ENOSPC); 1227 } 1228 1229 /* 1230 * Find a cylinder group to place a directory. 1231 * 1232 * The policy implemented by this algorithm is to allocate a 1233 * directory inode in the same cylinder group as its parent 1234 * directory, but also to reserve space for its files inodes 1235 * and data. Restrict the number of directories which may be 1236 * allocated one after another in the same cylinder group 1237 * without intervening allocation of files. 1238 * 1239 * If we allocate a first level directory then force allocation 1240 * in another cylinder group. 1241 */ 1242 static ino_t 1243 ffs_dirpref(struct inode *pip) 1244 { 1245 struct fs *fs; 1246 int cg, prefcg, curcg, dirsize, cgsize; 1247 int depth, range, start, end, numdirs, power, numerator, denominator; 1248 uint64_t avgifree, avgbfree, avgndir, curdirsize; 1249 uint64_t minifree, minbfree, maxndir; 1250 uint64_t maxcontigdirs; 1251 1252 mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED); 1253 fs = ITOFS(pip); 1254 1255 avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg; 1256 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1257 avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg; 1258 1259 /* 1260 * Select a preferred cylinder group to place a new directory. 1261 * If we are near the root of the filesystem we aim to spread 1262 * them out as much as possible. As we descend deeper from the 1263 * root we cluster them closer together around their parent as 1264 * we expect them to be more closely interactive. Higher-level 1265 * directories like usr/src/sys and usr/src/bin should be 1266 * separated while the directories in these areas are more 1267 * likely to be accessed together so should be closer. 1268 * 1269 * We pick a range of cylinder groups around the cylinder group 1270 * of the directory in which we are being created. The size of 1271 * the range for our search is based on our depth from the root 1272 * of our filesystem. We then probe that range based on how many 1273 * directories are already present. The first new directory is at 1274 * 1/2 (middle) of the range; the second is in the first 1/4 of the 1275 * range, then at 3/4, 1/8, 3/8, 5/8, 7/8, 1/16, 3/16, 5/16, etc. 1276 */ 1277 depth = DIP(pip, i_dirdepth); 1278 range = fs->fs_ncg / (1 << depth); 1279 curcg = ino_to_cg(fs, pip->i_number); 1280 start = curcg - (range / 2); 1281 if (start < 0) 1282 start += fs->fs_ncg; 1283 end = curcg + (range / 2); 1284 if (end >= fs->fs_ncg) 1285 end -= fs->fs_ncg; 1286 numdirs = pip->i_effnlink - 1; 1287 power = fls(numdirs); 1288 numerator = (numdirs & ~(1 << (power - 1))) * 2 + 1; 1289 denominator = 1 << power; 1290 prefcg = (curcg - (range / 2) + (range * numerator / denominator)); 1291 if (prefcg < 0) 1292 prefcg += fs->fs_ncg; 1293 if (prefcg >= fs->fs_ncg) 1294 prefcg -= fs->fs_ncg; 1295 /* 1296 * If this filesystem is not tracking directory depths, 1297 * revert to the old algorithm. 1298 */ 1299 if (depth == 0 && pip->i_number != UFS_ROOTINO) 1300 prefcg = curcg; 1301 1302 /* 1303 * Count various limits which used for 1304 * optimal allocation of a directory inode. 1305 */ 1306 maxndir = min(avgndir + (1 << depth), fs->fs_ipg); 1307 minifree = avgifree - avgifree / 4; 1308 if (minifree < 1) 1309 minifree = 1; 1310 minbfree = avgbfree - avgbfree / 4; 1311 if (minbfree < 1) 1312 minbfree = 1; 1313 cgsize = fs->fs_fsize * fs->fs_fpg; 1314 dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir; 1315 curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0; 1316 if (dirsize < curdirsize) 1317 dirsize = curdirsize; 1318 if (dirsize <= 0) 1319 maxcontigdirs = 0; /* dirsize overflowed */ 1320 else 1321 maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255); 1322 if (fs->fs_avgfpdir > 0) 1323 maxcontigdirs = min(maxcontigdirs, 1324 fs->fs_ipg / fs->fs_avgfpdir); 1325 if (maxcontigdirs == 0) 1326 maxcontigdirs = 1; 1327 1328 /* 1329 * Limit number of dirs in one cg and reserve space for 1330 * regular files, but only if we have no deficit in 1331 * inodes or space. 1332 * 1333 * We are trying to find a suitable cylinder group nearby 1334 * our preferred cylinder group to place a new directory. 1335 * We scan from our preferred cylinder group forward looking 1336 * for a cylinder group that meets our criterion. If we get 1337 * to the final cylinder group and do not find anything, 1338 * we start scanning forwards from the beginning of the 1339 * filesystem. While it might seem sensible to start scanning 1340 * backwards or even to alternate looking forward and backward, 1341 * this approach fails badly when the filesystem is nearly full. 1342 * Specifically, we first search all the areas that have no space 1343 * and finally try the one preceding that. We repeat this on 1344 * every request and in the case of the final block end up 1345 * searching the entire filesystem. By jumping to the front 1346 * of the filesystem, our future forward searches always look 1347 * in new cylinder groups so finds every possible block after 1348 * one pass over the filesystem. 1349 */ 1350 for (cg = prefcg; cg < fs->fs_ncg; cg++) 1351 if (fs->fs_cs(fs, cg).cs_ndir < maxndir && 1352 fs->fs_cs(fs, cg).cs_nifree >= minifree && 1353 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) { 1354 if (fs->fs_contigdirs[cg] < maxcontigdirs) 1355 return ((ino_t)(fs->fs_ipg * cg)); 1356 } 1357 for (cg = 0; cg < prefcg; cg++) 1358 if (fs->fs_cs(fs, cg).cs_ndir < maxndir && 1359 fs->fs_cs(fs, cg).cs_nifree >= minifree && 1360 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) { 1361 if (fs->fs_contigdirs[cg] < maxcontigdirs) 1362 return ((ino_t)(fs->fs_ipg * cg)); 1363 } 1364 /* 1365 * This is a backstop when we have deficit in space. 1366 */ 1367 for (cg = prefcg; cg < fs->fs_ncg; cg++) 1368 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree) 1369 return ((ino_t)(fs->fs_ipg * cg)); 1370 for (cg = 0; cg < prefcg; cg++) 1371 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree) 1372 break; 1373 return ((ino_t)(fs->fs_ipg * cg)); 1374 } 1375 1376 /* 1377 * Select the desired position for the next block in a file. The file is 1378 * logically divided into sections. The first section is composed of the 1379 * direct blocks and the next fs_maxbpg blocks. Each additional section 1380 * contains fs_maxbpg blocks. 1381 * 1382 * If no blocks have been allocated in the first section, the policy is to 1383 * request a block in the same cylinder group as the inode that describes 1384 * the file. The first indirect is allocated immediately following the last 1385 * direct block and the data blocks for the first indirect immediately 1386 * follow it. 1387 * 1388 * If no blocks have been allocated in any other section, the indirect 1389 * block(s) are allocated in the same cylinder group as its inode in an 1390 * area reserved immediately following the inode blocks. The policy for 1391 * the data blocks is to place them in a cylinder group with a greater than 1392 * average number of free blocks. An appropriate cylinder group is found 1393 * by using a rotor that sweeps the cylinder groups. When a new group of 1394 * blocks is needed, the sweep begins in the cylinder group following the 1395 * cylinder group from which the previous allocation was made. The sweep 1396 * continues until a cylinder group with greater than the average number 1397 * of free blocks is found. If the allocation is for the first block in an 1398 * indirect block or the previous block is a hole, then the information on 1399 * the previous allocation is unavailable; here a best guess is made based 1400 * on the logical block number being allocated. 1401 * 1402 * If a section is already partially allocated, the policy is to 1403 * allocate blocks contiguously within the section if possible. 1404 */ 1405 ufs2_daddr_t 1406 ffs_blkpref_ufs1(struct inode *ip, 1407 ufs_lbn_t lbn, 1408 int indx, 1409 ufs1_daddr_t *bap) 1410 { 1411 struct fs *fs; 1412 uint64_t cg, inocg; 1413 uint64_t avgbfree, startcg; 1414 ufs2_daddr_t pref, prevbn; 1415 1416 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap")); 1417 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1418 fs = ITOFS(ip); 1419 /* 1420 * Allocation of indirect blocks is indicated by passing negative 1421 * values in indx: -1 for single indirect, -2 for double indirect, 1422 * -3 for triple indirect. As noted below, we attempt to allocate 1423 * the first indirect inline with the file data. For all later 1424 * indirect blocks, the data is often allocated in other cylinder 1425 * groups. However to speed random file access and to speed up 1426 * fsck, the filesystem reserves the first fs_metaspace blocks 1427 * (typically half of fs_minfree) of the data area of each cylinder 1428 * group to hold these later indirect blocks. 1429 */ 1430 inocg = ino_to_cg(fs, ip->i_number); 1431 if (indx < 0) { 1432 /* 1433 * Our preference for indirect blocks is the zone at the 1434 * beginning of the inode's cylinder group data area that 1435 * we try to reserve for indirect blocks. 1436 */ 1437 pref = cgmeta(fs, inocg); 1438 /* 1439 * If we are allocating the first indirect block, try to 1440 * place it immediately following the last direct block. 1441 */ 1442 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) && 1443 ip->i_din1->di_db[UFS_NDADDR - 1] != 0) { 1444 pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag; 1445 if (dtog(fs, pref) >= fs->fs_ncg) 1446 pref = 0; 1447 } 1448 return (pref); 1449 } 1450 /* 1451 * If we are allocating the first data block in the first indirect 1452 * block and the indirect has been allocated in the data block area, 1453 * try to place it immediately following the indirect block. 1454 */ 1455 if (lbn == UFS_NDADDR) { 1456 pref = ip->i_din1->di_ib[0]; 1457 if (pref != 0 && pref >= cgdata(fs, inocg) && 1458 pref < cgbase(fs, inocg + 1)) { 1459 if (dtog(fs, pref + fs->fs_frag) >= fs->fs_ncg) 1460 return (0); 1461 return (pref + fs->fs_frag); 1462 } 1463 } 1464 /* 1465 * If we are at the beginning of a file, or we have already allocated 1466 * the maximum number of blocks per cylinder group, or we do not 1467 * have a block allocated immediately preceding us, then we need 1468 * to decide where to start allocating new blocks. 1469 */ 1470 if (indx == 0) { 1471 prevbn = 0; 1472 } else { 1473 prevbn = bap[indx - 1]; 1474 if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn, 1475 fs->fs_bsize) != 0) 1476 prevbn = 0; 1477 } 1478 if (indx % fs->fs_maxbpg == 0 || prevbn == 0) { 1479 /* 1480 * If we are allocating a directory data block, we want 1481 * to place it in the metadata area. 1482 */ 1483 if ((ip->i_mode & IFMT) == IFDIR) 1484 return (cgmeta(fs, inocg)); 1485 /* 1486 * Until we fill all the direct and all the first indirect's 1487 * blocks, we try to allocate in the data area of the inode's 1488 * cylinder group. 1489 */ 1490 if (lbn < UFS_NDADDR + NINDIR(fs)) 1491 return (cgdata(fs, inocg)); 1492 /* 1493 * Find a cylinder with greater than average number of 1494 * unused data blocks. 1495 */ 1496 if (indx == 0 || prevbn == 0) 1497 startcg = inocg + lbn / fs->fs_maxbpg; 1498 else 1499 startcg = dtog(fs, prevbn) + 1; 1500 startcg %= fs->fs_ncg; 1501 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1502 for (cg = startcg; cg < fs->fs_ncg; cg++) 1503 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1504 fs->fs_cgrotor = cg; 1505 return (cgdata(fs, cg)); 1506 } 1507 for (cg = 0; cg < startcg; cg++) 1508 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1509 fs->fs_cgrotor = cg; 1510 return (cgdata(fs, cg)); 1511 } 1512 return (0); 1513 } 1514 /* 1515 * Otherwise, we just always try to lay things out contiguously. 1516 */ 1517 if (dtog(fs, prevbn + fs->fs_frag) >= fs->fs_ncg) 1518 return (0); 1519 return (prevbn + fs->fs_frag); 1520 } 1521 1522 /* 1523 * Same as above, but for UFS2 1524 */ 1525 ufs2_daddr_t 1526 ffs_blkpref_ufs2(struct inode *ip, 1527 ufs_lbn_t lbn, 1528 int indx, 1529 ufs2_daddr_t *bap) 1530 { 1531 struct fs *fs; 1532 uint64_t cg, inocg; 1533 uint64_t avgbfree, startcg; 1534 ufs2_daddr_t pref, prevbn; 1535 1536 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap")); 1537 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1538 fs = ITOFS(ip); 1539 /* 1540 * Allocation of indirect blocks is indicated by passing negative 1541 * values in indx: -1 for single indirect, -2 for double indirect, 1542 * -3 for triple indirect. As noted below, we attempt to allocate 1543 * the first indirect inline with the file data. For all later 1544 * indirect blocks, the data is often allocated in other cylinder 1545 * groups. However to speed random file access and to speed up 1546 * fsck, the filesystem reserves the first fs_metaspace blocks 1547 * (typically half of fs_minfree) of the data area of each cylinder 1548 * group to hold these later indirect blocks. 1549 */ 1550 inocg = ino_to_cg(fs, ip->i_number); 1551 if (indx < 0) { 1552 /* 1553 * Our preference for indirect blocks is the zone at the 1554 * beginning of the inode's cylinder group data area that 1555 * we try to reserve for indirect blocks. 1556 */ 1557 pref = cgmeta(fs, inocg); 1558 /* 1559 * If we are allocating the first indirect block, try to 1560 * place it immediately following the last direct block. 1561 */ 1562 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) && 1563 ip->i_din2->di_db[UFS_NDADDR - 1] != 0) { 1564 pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag; 1565 if (dtog(fs, pref) >= fs->fs_ncg) 1566 pref = 0; 1567 } 1568 return (pref); 1569 } 1570 /* 1571 * If we are allocating the first data block in the first indirect 1572 * block and the indirect has been allocated in the data block area, 1573 * try to place it immediately following the indirect block. 1574 */ 1575 if (lbn == UFS_NDADDR) { 1576 pref = ip->i_din2->di_ib[0]; 1577 if (pref != 0 && pref >= cgdata(fs, inocg) && 1578 pref < cgbase(fs, inocg + 1)) { 1579 if (dtog(fs, pref + fs->fs_frag) >= fs->fs_ncg) 1580 return (0); 1581 return (pref + fs->fs_frag); 1582 } 1583 } 1584 /* 1585 * If we are at the beginning of a file, or we have already allocated 1586 * the maximum number of blocks per cylinder group, or we do not 1587 * have a block allocated immediately preceding us, then we need 1588 * to decide where to start allocating new blocks. 1589 */ 1590 if (indx == 0) { 1591 prevbn = 0; 1592 } else { 1593 prevbn = bap[indx - 1]; 1594 if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn, 1595 fs->fs_bsize) != 0) 1596 prevbn = 0; 1597 } 1598 if (indx % fs->fs_maxbpg == 0 || prevbn == 0) { 1599 /* 1600 * If we are allocating a directory data block, we want 1601 * to place it in the metadata area. 1602 */ 1603 if ((ip->i_mode & IFMT) == IFDIR) 1604 return (cgmeta(fs, inocg)); 1605 /* 1606 * Until we fill all the direct and all the first indirect's 1607 * blocks, we try to allocate in the data area of the inode's 1608 * cylinder group. 1609 */ 1610 if (lbn < UFS_NDADDR + NINDIR(fs)) 1611 return (cgdata(fs, inocg)); 1612 /* 1613 * Find a cylinder with greater than average number of 1614 * unused data blocks. 1615 */ 1616 if (indx == 0 || prevbn == 0) 1617 startcg = inocg + lbn / fs->fs_maxbpg; 1618 else 1619 startcg = dtog(fs, prevbn) + 1; 1620 startcg %= fs->fs_ncg; 1621 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1622 for (cg = startcg; cg < fs->fs_ncg; cg++) 1623 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1624 fs->fs_cgrotor = cg; 1625 return (cgdata(fs, cg)); 1626 } 1627 for (cg = 0; cg < startcg; cg++) 1628 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1629 fs->fs_cgrotor = cg; 1630 return (cgdata(fs, cg)); 1631 } 1632 return (0); 1633 } 1634 /* 1635 * Otherwise, we just always try to lay things out contiguously. 1636 */ 1637 if (dtog(fs, prevbn + fs->fs_frag) >= fs->fs_ncg) 1638 return (0); 1639 return (prevbn + fs->fs_frag); 1640 } 1641 1642 /* 1643 * Implement the cylinder overflow algorithm. 1644 * 1645 * The policy implemented by this algorithm is: 1646 * 1) allocate the block in its requested cylinder group. 1647 * 2) quadratically rehash on the cylinder group number. 1648 * 3) brute force search for a free block. 1649 * 1650 * Must be called with the UFS lock held. Will release the lock on success 1651 * and return with it held on failure. 1652 */ 1653 /*VARARGS5*/ 1654 static ufs2_daddr_t 1655 ffs_hashalloc(struct inode *ip, 1656 uint64_t cg, 1657 ufs2_daddr_t pref, 1658 int size, /* Search size for data blocks, mode for inodes */ 1659 int rsize, /* Real allocated size. */ 1660 allocfcn_t *allocator) 1661 { 1662 struct fs *fs; 1663 ufs2_daddr_t result; 1664 uint64_t i, icg = cg; 1665 1666 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1667 #ifdef INVARIANTS 1668 if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED) 1669 panic("ffs_hashalloc: allocation on suspended filesystem"); 1670 #endif 1671 fs = ITOFS(ip); 1672 /* 1673 * 1: preferred cylinder group 1674 */ 1675 result = (*allocator)(ip, cg, pref, size, rsize); 1676 if (result) 1677 return (result); 1678 /* 1679 * 2: quadratic rehash 1680 */ 1681 for (i = 1; i < fs->fs_ncg; i *= 2) { 1682 cg += i; 1683 if (cg >= fs->fs_ncg) 1684 cg -= fs->fs_ncg; 1685 result = (*allocator)(ip, cg, 0, size, rsize); 1686 if (result) 1687 return (result); 1688 } 1689 /* 1690 * 3: brute force search 1691 * Note that we start at i == 2, since 0 was checked initially, 1692 * and 1 is always checked in the quadratic rehash. 1693 */ 1694 cg = (icg + 2) % fs->fs_ncg; 1695 for (i = 2; i < fs->fs_ncg; i++) { 1696 result = (*allocator)(ip, cg, 0, size, rsize); 1697 if (result) 1698 return (result); 1699 cg++; 1700 if (cg == fs->fs_ncg) 1701 cg = 0; 1702 } 1703 return (0); 1704 } 1705 1706 /* 1707 * Determine whether a fragment can be extended. 1708 * 1709 * Check to see if the necessary fragments are available, and 1710 * if they are, allocate them. 1711 */ 1712 static ufs2_daddr_t 1713 ffs_fragextend(struct inode *ip, 1714 uint64_t cg, 1715 ufs2_daddr_t bprev, 1716 int osize, 1717 int nsize) 1718 { 1719 struct fs *fs; 1720 struct cg *cgp; 1721 struct buf *bp; 1722 struct ufsmount *ump; 1723 int nffree; 1724 long bno; 1725 int frags, bbase; 1726 int i, error; 1727 uint8_t *blksfree; 1728 1729 ump = ITOUMP(ip); 1730 fs = ump->um_fs; 1731 if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize)) 1732 return (0); 1733 frags = numfrags(fs, nsize); 1734 bbase = fragnum(fs, bprev); 1735 if (bbase > fragnum(fs, (bprev + frags - 1))) { 1736 /* cannot extend across a block boundary */ 1737 return (0); 1738 } 1739 UFS_UNLOCK(ump); 1740 if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) { 1741 ffs_checkcgintegrity(fs, cg, error); 1742 goto fail; 1743 } 1744 bno = dtogd(fs, bprev); 1745 blksfree = cg_blksfree(cgp); 1746 for (i = numfrags(fs, osize); i < frags; i++) 1747 if (isclr(blksfree, bno + i)) 1748 goto fail; 1749 /* 1750 * the current fragment can be extended 1751 * deduct the count on fragment being extended into 1752 * increase the count on the remaining fragment (if any) 1753 * allocate the extended piece 1754 */ 1755 for (i = frags; i < fs->fs_frag - bbase; i++) 1756 if (isclr(blksfree, bno + i)) 1757 break; 1758 cgp->cg_frsum[i - numfrags(fs, osize)]--; 1759 if (i != frags) 1760 cgp->cg_frsum[i - frags]++; 1761 for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) { 1762 clrbit(blksfree, bno + i); 1763 cgp->cg_cs.cs_nffree--; 1764 nffree++; 1765 } 1766 UFS_LOCK(ump); 1767 fs->fs_cstotal.cs_nffree -= nffree; 1768 fs->fs_cs(fs, cg).cs_nffree -= nffree; 1769 fs->fs_fmod = 1; 1770 ACTIVECLEAR(fs, cg); 1771 UFS_UNLOCK(ump); 1772 if (DOINGSOFTDEP(ITOV(ip))) 1773 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev, 1774 frags, numfrags(fs, osize)); 1775 bdwrite(bp); 1776 return (bprev); 1777 1778 fail: 1779 brelse(bp); 1780 UFS_LOCK(ump); 1781 return (0); 1782 1783 } 1784 1785 /* 1786 * Determine whether a block can be allocated. 1787 * 1788 * Check to see if a block of the appropriate size is available, 1789 * and if it is, allocate it. 1790 */ 1791 static ufs2_daddr_t 1792 ffs_alloccg(struct inode *ip, 1793 uint64_t cg, 1794 ufs2_daddr_t bpref, 1795 int size, 1796 int rsize) 1797 { 1798 struct fs *fs; 1799 struct cg *cgp; 1800 struct buf *bp; 1801 struct ufsmount *ump; 1802 ufs1_daddr_t bno; 1803 ufs2_daddr_t blkno; 1804 int i, allocsiz, error, frags; 1805 uint8_t *blksfree; 1806 1807 ump = ITOUMP(ip); 1808 fs = ump->um_fs; 1809 if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize) 1810 return (0); 1811 UFS_UNLOCK(ump); 1812 if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0 || 1813 (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize)) { 1814 ffs_checkcgintegrity(fs, cg, error); 1815 goto fail; 1816 } 1817 if (size == fs->fs_bsize) { 1818 UFS_LOCK(ump); 1819 blkno = ffs_alloccgblk(ip, bp, bpref, rsize); 1820 ACTIVECLEAR(fs, cg); 1821 UFS_UNLOCK(ump); 1822 bdwrite(bp); 1823 return (blkno); 1824 } 1825 /* 1826 * check to see if any fragments are already available 1827 * allocsiz is the size which will be allocated, hacking 1828 * it down to a smaller size if necessary 1829 */ 1830 blksfree = cg_blksfree(cgp); 1831 frags = numfrags(fs, size); 1832 for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++) 1833 if (cgp->cg_frsum[allocsiz] != 0) 1834 break; 1835 if (allocsiz == fs->fs_frag) { 1836 /* 1837 * no fragments were available, so a block will be 1838 * allocated, and hacked up 1839 */ 1840 if (cgp->cg_cs.cs_nbfree == 0) 1841 goto fail; 1842 UFS_LOCK(ump); 1843 blkno = ffs_alloccgblk(ip, bp, bpref, rsize); 1844 ACTIVECLEAR(fs, cg); 1845 UFS_UNLOCK(ump); 1846 bdwrite(bp); 1847 return (blkno); 1848 } 1849 KASSERT(size == rsize, 1850 ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize)); 1851 bno = ffs_mapsearch(fs, cgp, bpref, allocsiz); 1852 if (bno < 0) 1853 goto fail; 1854 for (i = 0; i < frags; i++) 1855 clrbit(blksfree, bno + i); 1856 cgp->cg_cs.cs_nffree -= frags; 1857 cgp->cg_frsum[allocsiz]--; 1858 if (frags != allocsiz) 1859 cgp->cg_frsum[allocsiz - frags]++; 1860 UFS_LOCK(ump); 1861 fs->fs_cstotal.cs_nffree -= frags; 1862 fs->fs_cs(fs, cg).cs_nffree -= frags; 1863 fs->fs_fmod = 1; 1864 blkno = cgbase(fs, cg) + bno; 1865 ACTIVECLEAR(fs, cg); 1866 UFS_UNLOCK(ump); 1867 if (DOINGSOFTDEP(ITOV(ip))) 1868 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0); 1869 bdwrite(bp); 1870 return (blkno); 1871 1872 fail: 1873 brelse(bp); 1874 UFS_LOCK(ump); 1875 return (0); 1876 } 1877 1878 /* 1879 * Allocate a block in a cylinder group. 1880 * 1881 * This algorithm implements the following policy: 1882 * 1) allocate the requested block. 1883 * 2) allocate a rotationally optimal block in the same cylinder. 1884 * 3) allocate the next available block on the block rotor for the 1885 * specified cylinder group. 1886 * Note that this routine only allocates fs_bsize blocks; these 1887 * blocks may be fragmented by the routine that allocates them. 1888 */ 1889 static ufs2_daddr_t 1890 ffs_alloccgblk(struct inode *ip, 1891 struct buf *bp, 1892 ufs2_daddr_t bpref, 1893 int size) 1894 { 1895 struct fs *fs; 1896 struct cg *cgp; 1897 struct ufsmount *ump; 1898 ufs1_daddr_t bno; 1899 ufs2_daddr_t blkno; 1900 uint8_t *blksfree; 1901 int i, cgbpref; 1902 1903 ump = ITOUMP(ip); 1904 fs = ump->um_fs; 1905 mtx_assert(UFS_MTX(ump), MA_OWNED); 1906 cgp = (struct cg *)bp->b_data; 1907 blksfree = cg_blksfree(cgp); 1908 if (bpref == 0) { 1909 bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag; 1910 } else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) { 1911 /* map bpref to correct zone in this cg */ 1912 if (bpref < cgdata(fs, cgbpref)) 1913 bpref = cgmeta(fs, cgp->cg_cgx); 1914 else 1915 bpref = cgdata(fs, cgp->cg_cgx); 1916 } 1917 /* 1918 * if the requested block is available, use it 1919 */ 1920 bno = dtogd(fs, blknum(fs, bpref)); 1921 if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno))) 1922 goto gotit; 1923 /* 1924 * Take the next available block in this cylinder group. 1925 */ 1926 bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag); 1927 if (bno < 0) 1928 return (0); 1929 /* Update cg_rotor only if allocated from the data zone */ 1930 if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx))) 1931 cgp->cg_rotor = bno; 1932 gotit: 1933 blkno = fragstoblks(fs, bno); 1934 ffs_clrblock(fs, blksfree, (long)blkno); 1935 ffs_clusteracct(fs, cgp, blkno, -1); 1936 cgp->cg_cs.cs_nbfree--; 1937 fs->fs_cstotal.cs_nbfree--; 1938 fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--; 1939 fs->fs_fmod = 1; 1940 blkno = cgbase(fs, cgp->cg_cgx) + bno; 1941 /* 1942 * If the caller didn't want the whole block free the frags here. 1943 */ 1944 size = numfrags(fs, size); 1945 if (size != fs->fs_frag) { 1946 bno = dtogd(fs, blkno); 1947 for (i = size; i < fs->fs_frag; i++) 1948 setbit(blksfree, bno + i); 1949 i = fs->fs_frag - size; 1950 cgp->cg_cs.cs_nffree += i; 1951 fs->fs_cstotal.cs_nffree += i; 1952 fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i; 1953 fs->fs_fmod = 1; 1954 cgp->cg_frsum[i]++; 1955 } 1956 /* XXX Fixme. */ 1957 UFS_UNLOCK(ump); 1958 if (DOINGSOFTDEP(ITOV(ip))) 1959 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0); 1960 UFS_LOCK(ump); 1961 return (blkno); 1962 } 1963 1964 /* 1965 * Determine whether a cluster can be allocated. 1966 * 1967 * We do not currently check for optimal rotational layout if there 1968 * are multiple choices in the same cylinder group. Instead we just 1969 * take the first one that we find following bpref. 1970 */ 1971 static ufs2_daddr_t 1972 ffs_clusteralloc(struct inode *ip, 1973 uint64_t cg, 1974 ufs2_daddr_t bpref, 1975 int len) 1976 { 1977 struct fs *fs; 1978 struct cg *cgp; 1979 struct buf *bp; 1980 struct ufsmount *ump; 1981 int i, run, bit, map, got, error; 1982 ufs2_daddr_t bno; 1983 uint8_t *mapp; 1984 int32_t *lp; 1985 uint8_t *blksfree; 1986 1987 ump = ITOUMP(ip); 1988 fs = ump->um_fs; 1989 MPASS(cg < fs->fs_ncg); 1990 if (fs->fs_maxcluster[cg] < len) 1991 return (0); 1992 UFS_UNLOCK(ump); 1993 if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) { 1994 ffs_checkcgintegrity(fs, cg, error); 1995 UFS_LOCK(ump); 1996 return (0); 1997 } 1998 /* 1999 * Check to see if a cluster of the needed size (or bigger) is 2000 * available in this cylinder group. 2001 */ 2002 lp = &cg_clustersum(cgp)[len]; 2003 for (i = len; i <= fs->fs_contigsumsize; i++) 2004 if (*lp++ > 0) 2005 break; 2006 if (i > fs->fs_contigsumsize) { 2007 /* 2008 * This is the first time looking for a cluster in this 2009 * cylinder group. Update the cluster summary information 2010 * to reflect the true maximum sized cluster so that 2011 * future cluster allocation requests can avoid reading 2012 * the cylinder group map only to find no clusters. 2013 */ 2014 lp = &cg_clustersum(cgp)[len - 1]; 2015 for (i = len - 1; i > 0; i--) 2016 if (*lp-- > 0) 2017 break; 2018 UFS_LOCK(ump); 2019 fs->fs_maxcluster[cg] = i; 2020 brelse(bp); 2021 return (0); 2022 } 2023 /* 2024 * Search the cluster map to find a big enough cluster. 2025 * We take the first one that we find, even if it is larger 2026 * than we need as we prefer to get one close to the previous 2027 * block allocation. We do not search before the current 2028 * preference point as we do not want to allocate a block 2029 * that is allocated before the previous one (as we will 2030 * then have to wait for another pass of the elevator 2031 * algorithm before it will be read). We prefer to fail and 2032 * be recalled to try an allocation in the next cylinder group. 2033 */ 2034 if (dtog(fs, bpref) != cg) 2035 bpref = cgdata(fs, cg); 2036 else 2037 bpref = blknum(fs, bpref); 2038 bpref = fragstoblks(fs, dtogd(fs, bpref)); 2039 mapp = &cg_clustersfree(cgp)[bpref / NBBY]; 2040 map = *mapp++; 2041 bit = 1 << (bpref % NBBY); 2042 for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) { 2043 if ((map & bit) == 0) { 2044 run = 0; 2045 } else { 2046 run++; 2047 if (run == len) 2048 break; 2049 } 2050 if ((got & (NBBY - 1)) != (NBBY - 1)) { 2051 bit <<= 1; 2052 } else { 2053 map = *mapp++; 2054 bit = 1; 2055 } 2056 } 2057 if (got >= cgp->cg_nclusterblks) { 2058 UFS_LOCK(ump); 2059 brelse(bp); 2060 return (0); 2061 } 2062 /* 2063 * Allocate the cluster that we have found. 2064 */ 2065 blksfree = cg_blksfree(cgp); 2066 for (i = 1; i <= len; i++) 2067 if (!ffs_isblock(fs, blksfree, got - run + i)) 2068 panic("ffs_clusteralloc: map mismatch"); 2069 bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1); 2070 if (dtog(fs, bno) != cg) 2071 panic("ffs_clusteralloc: allocated out of group"); 2072 len = blkstofrags(fs, len); 2073 UFS_LOCK(ump); 2074 for (i = 0; i < len; i += fs->fs_frag) 2075 if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i) 2076 panic("ffs_clusteralloc: lost block"); 2077 ACTIVECLEAR(fs, cg); 2078 UFS_UNLOCK(ump); 2079 bdwrite(bp); 2080 return (bno); 2081 } 2082 2083 static inline struct buf * 2084 getinobuf(struct inode *ip, 2085 uint64_t cg, 2086 uint32_t cginoblk, 2087 int gbflags) 2088 { 2089 struct fs *fs; 2090 2091 fs = ITOFS(ip); 2092 return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs, 2093 cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0, 2094 gbflags)); 2095 } 2096 2097 /* 2098 * Synchronous inode initialization is needed only when barrier writes do not 2099 * work as advertised, and will impose a heavy cost on file creation in a newly 2100 * created filesystem. 2101 */ 2102 static int doasyncinodeinit = 1; 2103 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN, 2104 &doasyncinodeinit, 0, 2105 "Perform inode block initialization using asynchronous writes"); 2106 2107 /* 2108 * Determine whether an inode can be allocated. 2109 * 2110 * Check to see if an inode is available, and if it is, 2111 * allocate it using the following policy: 2112 * 1) allocate the requested inode. 2113 * 2) allocate the next available inode after the requested 2114 * inode in the specified cylinder group. 2115 */ 2116 static ufs2_daddr_t 2117 ffs_nodealloccg(struct inode *ip, 2118 uint64_t cg, 2119 ufs2_daddr_t ipref, 2120 int mode, 2121 int unused) 2122 { 2123 struct fs *fs; 2124 struct cg *cgp; 2125 struct buf *bp, *ibp; 2126 struct ufsmount *ump; 2127 uint8_t *inosused, *loc; 2128 struct ufs2_dinode *dp2; 2129 int error, start, len, i; 2130 uint32_t old_initediblk; 2131 2132 ump = ITOUMP(ip); 2133 fs = ump->um_fs; 2134 check_nifree: 2135 if (fs->fs_cs(fs, cg).cs_nifree == 0) 2136 return (0); 2137 UFS_UNLOCK(ump); 2138 if ((error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp)) != 0) { 2139 ffs_checkcgintegrity(fs, cg, error); 2140 UFS_LOCK(ump); 2141 return (0); 2142 } 2143 restart: 2144 if (cgp->cg_cs.cs_nifree == 0) { 2145 brelse(bp); 2146 UFS_LOCK(ump); 2147 return (0); 2148 } 2149 inosused = cg_inosused(cgp); 2150 if (ipref) { 2151 ipref %= fs->fs_ipg; 2152 if (isclr(inosused, ipref)) 2153 goto gotit; 2154 } 2155 start = cgp->cg_irotor / NBBY; 2156 len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY); 2157 loc = memcchr(&inosused[start], 0xff, len); 2158 if (loc == NULL) { 2159 len = start + 1; 2160 start = 0; 2161 loc = memcchr(&inosused[start], 0xff, len); 2162 if (loc == NULL) { 2163 printf("cg = %ju, irotor = %ld, fs = %s\n", 2164 (intmax_t)cg, (long)cgp->cg_irotor, fs->fs_fsmnt); 2165 panic("ffs_nodealloccg: map corrupted"); 2166 /* NOTREACHED */ 2167 } 2168 } 2169 ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1; 2170 gotit: 2171 /* 2172 * Check to see if we need to initialize more inodes. 2173 */ 2174 if (fs->fs_magic == FS_UFS2_MAGIC && 2175 ipref + INOPB(fs) > cgp->cg_initediblk && 2176 cgp->cg_initediblk < cgp->cg_niblk) { 2177 old_initediblk = cgp->cg_initediblk; 2178 2179 /* 2180 * Free the cylinder group lock before writing the 2181 * initialized inode block. Entering the 2182 * babarrierwrite() with the cylinder group lock 2183 * causes lock order violation between the lock and 2184 * snaplk. 2185 * 2186 * Another thread can decide to initialize the same 2187 * inode block, but whichever thread first gets the 2188 * cylinder group lock after writing the newly 2189 * allocated inode block will update it and the other 2190 * will realize that it has lost and leave the 2191 * cylinder group unchanged. 2192 */ 2193 ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT); 2194 brelse(bp); 2195 if (ibp == NULL) { 2196 /* 2197 * The inode block buffer is already owned by 2198 * another thread, which must initialize it. 2199 * Wait on the buffer to allow another thread 2200 * to finish the updates, with dropped cg 2201 * buffer lock, then retry. 2202 */ 2203 ibp = getinobuf(ip, cg, old_initediblk, 0); 2204 brelse(ibp); 2205 UFS_LOCK(ump); 2206 goto check_nifree; 2207 } 2208 bzero(ibp->b_data, (int)fs->fs_bsize); 2209 dp2 = (struct ufs2_dinode *)(ibp->b_data); 2210 for (i = 0; i < INOPB(fs); i++) { 2211 while (dp2->di_gen == 0) 2212 dp2->di_gen = arc4random(); 2213 dp2++; 2214 } 2215 2216 /* 2217 * Rather than adding a soft updates dependency to ensure 2218 * that the new inode block is written before it is claimed 2219 * by the cylinder group map, we just do a barrier write 2220 * here. The barrier write will ensure that the inode block 2221 * gets written before the updated cylinder group map can be 2222 * written. The barrier write should only slow down bulk 2223 * loading of newly created filesystems. 2224 */ 2225 if (doasyncinodeinit) 2226 babarrierwrite(ibp); 2227 else 2228 bwrite(ibp); 2229 2230 /* 2231 * After the inode block is written, try to update the 2232 * cg initediblk pointer. If another thread beat us 2233 * to it, then leave it unchanged as the other thread 2234 * has already set it correctly. 2235 */ 2236 error = ffs_getcg(fs, ump->um_devvp, cg, 0, &bp, &cgp); 2237 UFS_LOCK(ump); 2238 ACTIVECLEAR(fs, cg); 2239 UFS_UNLOCK(ump); 2240 if (error != 0) 2241 return (error); 2242 if (cgp->cg_initediblk == old_initediblk) 2243 cgp->cg_initediblk += INOPB(fs); 2244 goto restart; 2245 } 2246 cgp->cg_irotor = ipref; 2247 UFS_LOCK(ump); 2248 ACTIVECLEAR(fs, cg); 2249 setbit(inosused, ipref); 2250 cgp->cg_cs.cs_nifree--; 2251 fs->fs_cstotal.cs_nifree--; 2252 fs->fs_cs(fs, cg).cs_nifree--; 2253 fs->fs_fmod = 1; 2254 if ((mode & IFMT) == IFDIR) { 2255 cgp->cg_cs.cs_ndir++; 2256 fs->fs_cstotal.cs_ndir++; 2257 fs->fs_cs(fs, cg).cs_ndir++; 2258 } 2259 UFS_UNLOCK(ump); 2260 if (DOINGSOFTDEP(ITOV(ip))) 2261 softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode); 2262 bdwrite(bp); 2263 return ((ino_t)(cg * fs->fs_ipg + ipref)); 2264 } 2265 2266 /* 2267 * Free a block or fragment. 2268 * 2269 * The specified block or fragment is placed back in the 2270 * free map. If a fragment is deallocated, a possible 2271 * block reassembly is checked. 2272 */ 2273 static void 2274 ffs_blkfree_cg(struct ufsmount *ump, 2275 struct fs *fs, 2276 struct vnode *devvp, 2277 ufs2_daddr_t bno, 2278 long size, 2279 ino_t inum, 2280 struct workhead *dephd) 2281 { 2282 struct mount *mp; 2283 struct cg *cgp; 2284 struct buf *bp; 2285 daddr_t dbn; 2286 ufs1_daddr_t fragno, cgbno; 2287 int i, blk, frags, bbase, error; 2288 uint64_t cg; 2289 uint8_t *blksfree; 2290 struct cdev *dev; 2291 2292 cg = dtog(fs, bno); 2293 if (devvp->v_type == VREG) { 2294 /* devvp is a snapshot */ 2295 MPASS(devvp->v_mount->mnt_data == ump); 2296 dev = ump->um_devvp->v_rdev; 2297 } else if (devvp->v_type == VCHR) { 2298 /* 2299 * devvp is a normal disk device 2300 * XXXKIB: devvp is not locked there, v_rdev access depends on 2301 * busy mount, which prevents mntfs devvp from reclamation. 2302 */ 2303 dev = devvp->v_rdev; 2304 } else 2305 return; 2306 #ifdef INVARIANTS 2307 if ((uint64_t)size > fs->fs_bsize || fragoff(fs, size) != 0 || 2308 fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) { 2309 printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n", 2310 devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize, 2311 size, fs->fs_fsmnt); 2312 panic("ffs_blkfree_cg: invalid size"); 2313 } 2314 #endif 2315 if ((uint64_t)bno >= fs->fs_size) { 2316 printf("bad block %jd, ino %ju\n", (intmax_t)bno, 2317 (intmax_t)inum); 2318 ffs_fserr(fs, inum, "bad block"); 2319 return; 2320 } 2321 if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) { 2322 if (!MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR) 2323 return; 2324 /* 2325 * Would like to just downgrade to read-only. Until that 2326 * capability is available, just toss the cylinder group 2327 * update and mark the filesystem as needing to run fsck. 2328 */ 2329 fs->fs_flags |= FS_NEEDSFSCK; 2330 if (devvp->v_type == VREG) 2331 dbn = fragstoblks(fs, cgtod(fs, cg)); 2332 else 2333 dbn = fsbtodb(fs, cgtod(fs, cg)); 2334 error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp); 2335 KASSERT(error == 0, ("getblkx failed")); 2336 softdep_setup_blkfree(UFSTOVFS(ump), bp, bno, 2337 numfrags(fs, size), dephd, true); 2338 bp->b_flags |= B_RELBUF | B_NOCACHE; 2339 bp->b_flags &= ~B_CACHE; 2340 bawrite(bp); 2341 return; 2342 } 2343 cgbno = dtogd(fs, bno); 2344 blksfree = cg_blksfree(cgp); 2345 UFS_LOCK(ump); 2346 if (size == fs->fs_bsize) { 2347 fragno = fragstoblks(fs, cgbno); 2348 if (!ffs_isfreeblock(fs, blksfree, fragno)) { 2349 if (devvp->v_type == VREG) { 2350 UFS_UNLOCK(ump); 2351 /* devvp is a snapshot */ 2352 brelse(bp); 2353 return; 2354 } 2355 printf("dev = %s, block = %jd, fs = %s\n", 2356 devtoname(dev), (intmax_t)bno, fs->fs_fsmnt); 2357 panic("ffs_blkfree_cg: freeing free block"); 2358 } 2359 ffs_setblock(fs, blksfree, fragno); 2360 ffs_clusteracct(fs, cgp, fragno, 1); 2361 cgp->cg_cs.cs_nbfree++; 2362 fs->fs_cstotal.cs_nbfree++; 2363 fs->fs_cs(fs, cg).cs_nbfree++; 2364 } else { 2365 bbase = cgbno - fragnum(fs, cgbno); 2366 /* 2367 * decrement the counts associated with the old frags 2368 */ 2369 blk = blkmap(fs, blksfree, bbase); 2370 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 2371 /* 2372 * deallocate the fragment 2373 */ 2374 frags = numfrags(fs, size); 2375 for (i = 0; i < frags; i++) { 2376 if (isset(blksfree, cgbno + i)) { 2377 printf("dev = %s, block = %jd, fs = %s\n", 2378 devtoname(dev), (intmax_t)(bno + i), 2379 fs->fs_fsmnt); 2380 panic("ffs_blkfree_cg: freeing free frag"); 2381 } 2382 setbit(blksfree, cgbno + i); 2383 } 2384 cgp->cg_cs.cs_nffree += i; 2385 fs->fs_cstotal.cs_nffree += i; 2386 fs->fs_cs(fs, cg).cs_nffree += i; 2387 /* 2388 * add back in counts associated with the new frags 2389 */ 2390 blk = blkmap(fs, blksfree, bbase); 2391 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 2392 /* 2393 * if a complete block has been reassembled, account for it 2394 */ 2395 fragno = fragstoblks(fs, bbase); 2396 if (ffs_isblock(fs, blksfree, fragno)) { 2397 cgp->cg_cs.cs_nffree -= fs->fs_frag; 2398 fs->fs_cstotal.cs_nffree -= fs->fs_frag; 2399 fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag; 2400 ffs_clusteracct(fs, cgp, fragno, 1); 2401 cgp->cg_cs.cs_nbfree++; 2402 fs->fs_cstotal.cs_nbfree++; 2403 fs->fs_cs(fs, cg).cs_nbfree++; 2404 } 2405 } 2406 fs->fs_fmod = 1; 2407 ACTIVECLEAR(fs, cg); 2408 UFS_UNLOCK(ump); 2409 mp = UFSTOVFS(ump); 2410 if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR) 2411 softdep_setup_blkfree(UFSTOVFS(ump), bp, bno, 2412 numfrags(fs, size), dephd, false); 2413 bdwrite(bp); 2414 } 2415 2416 /* 2417 * Structures and routines associated with trim management. 2418 * 2419 * The following requests are passed to trim_lookup to indicate 2420 * the actions that should be taken. 2421 */ 2422 #define NEW 1 /* if found, error else allocate and hash it */ 2423 #define OLD 2 /* if not found, error, else return it */ 2424 #define REPLACE 3 /* if not found, error else unhash and reallocate it */ 2425 #define DONE 4 /* if not found, error else unhash and return it */ 2426 #define SINGLE 5 /* don't look up, just allocate it and don't hash it */ 2427 2428 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures"); 2429 2430 #define TRIMLIST_HASH(ump, key) \ 2431 (&(ump)->um_trimhash[(key) & (ump)->um_trimlisthashsize]) 2432 2433 /* 2434 * These structures describe each of the block free requests aggregated 2435 * together to make up a trim request. 2436 */ 2437 struct trim_blkreq { 2438 TAILQ_ENTRY(trim_blkreq) blkreqlist; 2439 ufs2_daddr_t bno; 2440 long size; 2441 struct workhead *pdephd; 2442 struct workhead dephd; 2443 }; 2444 2445 /* 2446 * Description of a trim request. 2447 */ 2448 struct ffs_blkfree_trim_params { 2449 TAILQ_HEAD(, trim_blkreq) blklist; 2450 LIST_ENTRY(ffs_blkfree_trim_params) hashlist; 2451 struct task task; 2452 struct ufsmount *ump; 2453 struct vnode *devvp; 2454 ino_t inum; 2455 ufs2_daddr_t bno; 2456 long size; 2457 long key; 2458 }; 2459 2460 static void ffs_blkfree_trim_completed(struct buf *); 2461 static void ffs_blkfree_trim_task(void *ctx, int pending __unused); 2462 static struct ffs_blkfree_trim_params *trim_lookup(struct ufsmount *, 2463 struct vnode *, ufs2_daddr_t, long, ino_t, uint64_t, int); 2464 static void ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *); 2465 2466 /* 2467 * Called on trim completion to start a task to free the associated block(s). 2468 */ 2469 static void 2470 ffs_blkfree_trim_completed(struct buf *bp) 2471 { 2472 struct ffs_blkfree_trim_params *tp; 2473 2474 tp = bp->b_fsprivate1; 2475 free(bp, M_TRIM); 2476 TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp); 2477 taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task); 2478 } 2479 2480 /* 2481 * Trim completion task that free associated block(s). 2482 */ 2483 static void 2484 ffs_blkfree_trim_task(void *ctx, int pending) 2485 { 2486 struct ffs_blkfree_trim_params *tp; 2487 struct trim_blkreq *blkelm; 2488 struct ufsmount *ump; 2489 2490 tp = ctx; 2491 ump = tp->ump; 2492 while ((blkelm = TAILQ_FIRST(&tp->blklist)) != NULL) { 2493 ffs_blkfree_cg(ump, ump->um_fs, tp->devvp, blkelm->bno, 2494 blkelm->size, tp->inum, blkelm->pdephd); 2495 TAILQ_REMOVE(&tp->blklist, blkelm, blkreqlist); 2496 free(blkelm, M_TRIM); 2497 } 2498 vn_finished_secondary_write(UFSTOVFS(ump)); 2499 UFS_LOCK(ump); 2500 ump->um_trim_inflight -= 1; 2501 ump->um_trim_inflight_blks -= numfrags(ump->um_fs, tp->size); 2502 UFS_UNLOCK(ump); 2503 free(tp, M_TRIM); 2504 } 2505 2506 /* 2507 * Lookup a trim request by inode number. 2508 * Allocate if requested (NEW, REPLACE, SINGLE). 2509 */ 2510 static struct ffs_blkfree_trim_params * 2511 trim_lookup(struct ufsmount *ump, 2512 struct vnode *devvp, 2513 ufs2_daddr_t bno, 2514 long size, 2515 ino_t inum, 2516 uint64_t key, 2517 int alloctype) 2518 { 2519 struct trimlist_hashhead *tphashhead; 2520 struct ffs_blkfree_trim_params *tp, *ntp; 2521 2522 ntp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK); 2523 if (alloctype != SINGLE) { 2524 KASSERT(key >= FIRST_VALID_KEY, ("trim_lookup: invalid key")); 2525 UFS_LOCK(ump); 2526 tphashhead = TRIMLIST_HASH(ump, key); 2527 LIST_FOREACH(tp, tphashhead, hashlist) 2528 if (key == tp->key) 2529 break; 2530 } 2531 switch (alloctype) { 2532 case NEW: 2533 KASSERT(tp == NULL, ("trim_lookup: found trim")); 2534 break; 2535 case OLD: 2536 KASSERT(tp != NULL, 2537 ("trim_lookup: missing call to ffs_blkrelease_start()")); 2538 UFS_UNLOCK(ump); 2539 free(ntp, M_TRIM); 2540 return (tp); 2541 case REPLACE: 2542 KASSERT(tp != NULL, ("trim_lookup: missing REPLACE trim")); 2543 LIST_REMOVE(tp, hashlist); 2544 /* tp will be freed by caller */ 2545 break; 2546 case DONE: 2547 KASSERT(tp != NULL, ("trim_lookup: missing DONE trim")); 2548 LIST_REMOVE(tp, hashlist); 2549 UFS_UNLOCK(ump); 2550 free(ntp, M_TRIM); 2551 return (tp); 2552 } 2553 TAILQ_INIT(&ntp->blklist); 2554 ntp->ump = ump; 2555 ntp->devvp = devvp; 2556 ntp->bno = bno; 2557 ntp->size = size; 2558 ntp->inum = inum; 2559 ntp->key = key; 2560 if (alloctype != SINGLE) { 2561 LIST_INSERT_HEAD(tphashhead, ntp, hashlist); 2562 UFS_UNLOCK(ump); 2563 } 2564 return (ntp); 2565 } 2566 2567 /* 2568 * Dispatch a trim request. 2569 */ 2570 static void 2571 ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *tp) 2572 { 2573 struct ufsmount *ump; 2574 struct mount *mp; 2575 struct buf *bp; 2576 2577 /* 2578 * Postpone the set of the free bit in the cg bitmap until the 2579 * BIO_DELETE is completed. Otherwise, due to disk queue 2580 * reordering, TRIM might be issued after we reuse the block 2581 * and write some new data into it. 2582 */ 2583 ump = tp->ump; 2584 bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO); 2585 bp->b_iocmd = BIO_DELETE; 2586 bp->b_iooffset = dbtob(fsbtodb(ump->um_fs, tp->bno)); 2587 bp->b_iodone = ffs_blkfree_trim_completed; 2588 bp->b_bcount = tp->size; 2589 bp->b_fsprivate1 = tp; 2590 UFS_LOCK(ump); 2591 ump->um_trim_total += 1; 2592 ump->um_trim_inflight += 1; 2593 ump->um_trim_inflight_blks += numfrags(ump->um_fs, tp->size); 2594 ump->um_trim_total_blks += numfrags(ump->um_fs, tp->size); 2595 UFS_UNLOCK(ump); 2596 2597 mp = UFSTOVFS(ump); 2598 vn_start_secondary_write(NULL, &mp, 0); 2599 g_vfs_strategy(ump->um_bo, bp); 2600 } 2601 2602 /* 2603 * Allocate a new key to use to identify a range of blocks. 2604 */ 2605 uint64_t 2606 ffs_blkrelease_start(struct ufsmount *ump, 2607 struct vnode *devvp, 2608 ino_t inum) 2609 { 2610 static u_long masterkey; 2611 uint64_t key; 2612 2613 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0) 2614 return (SINGLETON_KEY); 2615 do { 2616 key = atomic_fetchadd_long(&masterkey, 1); 2617 } while (key < FIRST_VALID_KEY); 2618 (void) trim_lookup(ump, devvp, 0, 0, inum, key, NEW); 2619 return (key); 2620 } 2621 2622 /* 2623 * Deallocate a key that has been used to identify a range of blocks. 2624 */ 2625 void 2626 ffs_blkrelease_finish(struct ufsmount *ump, uint64_t key) 2627 { 2628 struct ffs_blkfree_trim_params *tp; 2629 2630 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0) 2631 return; 2632 /* 2633 * If the vfs.ffs.dotrimcons sysctl option is enabled while 2634 * a file deletion is active, specifically after a call 2635 * to ffs_blkrelease_start() but before the call to 2636 * ffs_blkrelease_finish(), ffs_blkrelease_start() will 2637 * have handed out SINGLETON_KEY rather than starting a 2638 * collection sequence. Thus if we get a SINGLETON_KEY 2639 * passed to ffs_blkrelease_finish(), we just return rather 2640 * than trying to finish the nonexistent sequence. 2641 */ 2642 if (key == SINGLETON_KEY) { 2643 #ifdef INVARIANTS 2644 printf("%s: vfs.ffs.dotrimcons enabled on active filesystem\n", 2645 ump->um_mountp->mnt_stat.f_mntonname); 2646 #endif 2647 return; 2648 } 2649 /* 2650 * We are done with sending blocks using this key. Look up the key 2651 * using the DONE alloctype (in tp) to request that it be unhashed 2652 * as we will not be adding to it. If the key has never been used, 2653 * tp->size will be zero, so we can just free tp. Otherwise the call 2654 * to ffs_blkfree_sendtrim(tp) causes the block range described by 2655 * tp to be issued (and then tp to be freed). 2656 */ 2657 tp = trim_lookup(ump, NULL, 0, 0, 0, key, DONE); 2658 if (tp->size == 0) 2659 free(tp, M_TRIM); 2660 else 2661 ffs_blkfree_sendtrim(tp); 2662 } 2663 2664 /* 2665 * Setup to free a block or fragment. 2666 * 2667 * Check for snapshots that might want to claim the block. 2668 * If trims are requested, prepare a trim request. Attempt to 2669 * aggregate consecutive blocks into a single trim request. 2670 */ 2671 void 2672 ffs_blkfree(struct ufsmount *ump, 2673 struct fs *fs, 2674 struct vnode *devvp, 2675 ufs2_daddr_t bno, 2676 long size, 2677 ino_t inum, 2678 __enum_uint8(vtype) vtype, 2679 struct workhead *dephd, 2680 uint64_t key) 2681 { 2682 struct ffs_blkfree_trim_params *tp, *ntp; 2683 struct trim_blkreq *blkelm; 2684 2685 /* 2686 * Check to see if a snapshot wants to claim the block. 2687 * Check that devvp is a normal disk device, not a snapshot, 2688 * it has a snapshot(s) associated with it, and one of the 2689 * snapshots wants to claim the block. 2690 */ 2691 if (devvp->v_type == VCHR && 2692 (devvp->v_vflag & VV_COPYONWRITE) && 2693 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) { 2694 return; 2695 } 2696 /* 2697 * Nothing to delay if TRIM is not required for this block or TRIM 2698 * is disabled or the operation is performed on a snapshot. 2699 */ 2700 if (key == NOTRIM_KEY || ((ump->um_flags & UM_CANDELETE) == 0) || 2701 devvp->v_type == VREG) { 2702 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd); 2703 return; 2704 } 2705 blkelm = malloc(sizeof(struct trim_blkreq), M_TRIM, M_WAITOK); 2706 blkelm->bno = bno; 2707 blkelm->size = size; 2708 if (dephd == NULL) { 2709 blkelm->pdephd = NULL; 2710 } else { 2711 LIST_INIT(&blkelm->dephd); 2712 LIST_SWAP(dephd, &blkelm->dephd, worklist, wk_list); 2713 blkelm->pdephd = &blkelm->dephd; 2714 } 2715 if (key == SINGLETON_KEY) { 2716 /* 2717 * Just a single non-contiguous piece. Use the SINGLE 2718 * alloctype to return a trim request that will not be 2719 * hashed for future lookup. 2720 */ 2721 tp = trim_lookup(ump, devvp, bno, size, inum, key, SINGLE); 2722 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2723 ffs_blkfree_sendtrim(tp); 2724 return; 2725 } 2726 /* 2727 * The callers of this function are not tracking whether or not 2728 * the blocks are contiguous. They are just saying that they 2729 * are freeing a set of blocks. It is this code that determines 2730 * the pieces of that range that are actually contiguous. 2731 * 2732 * Calling ffs_blkrelease_start() will have created an entry 2733 * that we will use. 2734 */ 2735 tp = trim_lookup(ump, devvp, bno, size, inum, key, OLD); 2736 if (tp->size == 0) { 2737 /* 2738 * First block of a potential range, set block and size 2739 * for the trim block. 2740 */ 2741 tp->bno = bno; 2742 tp->size = size; 2743 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2744 return; 2745 } 2746 /* 2747 * If this block is a continuation of the range (either 2748 * follows at the end or preceeds in the front) then we 2749 * add it to the front or back of the list and return. 2750 * 2751 * If it is not a continuation of the trim that we were 2752 * building, using the REPLACE alloctype, we request that 2753 * the old trim request (still in tp) be unhashed and a 2754 * new range started (in ntp). The ffs_blkfree_sendtrim(tp) 2755 * call causes the block range described by tp to be issued 2756 * (and then tp to be freed). 2757 */ 2758 if (bno + numfrags(fs, size) == tp->bno) { 2759 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2760 tp->bno = bno; 2761 tp->size += size; 2762 return; 2763 } else if (bno == tp->bno + numfrags(fs, tp->size)) { 2764 TAILQ_INSERT_TAIL(&tp->blklist, blkelm, blkreqlist); 2765 tp->size += size; 2766 return; 2767 } 2768 ntp = trim_lookup(ump, devvp, bno, size, inum, key, REPLACE); 2769 TAILQ_INSERT_HEAD(&ntp->blklist, blkelm, blkreqlist); 2770 ffs_blkfree_sendtrim(tp); 2771 } 2772 2773 #ifdef INVARIANTS 2774 /* 2775 * Verify allocation of a block or fragment. 2776 * Return 1 if block or fragment is free. 2777 */ 2778 static int 2779 ffs_checkfreeblk(struct inode *ip, 2780 ufs2_daddr_t bno, 2781 long size) 2782 { 2783 struct fs *fs; 2784 struct cg *cgp; 2785 struct buf *bp; 2786 ufs1_daddr_t cgbno; 2787 int i, frags, blkalloced; 2788 uint8_t *blksfree; 2789 2790 fs = ITOFS(ip); 2791 if ((uint64_t)size > fs->fs_bsize || fragoff(fs, size) != 0) { 2792 printf("bsize = %ld, size = %ld, fs = %s\n", 2793 (long)fs->fs_bsize, size, fs->fs_fsmnt); 2794 panic("ffs_checkfreeblk: bad size"); 2795 } 2796 if ((uint64_t)bno >= fs->fs_size) 2797 panic("ffs_checkfreeblk: too big block %jd", (intmax_t)bno); 2798 if (ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), 0, &bp, &cgp) != 0) 2799 return (0); 2800 blksfree = cg_blksfree(cgp); 2801 cgbno = dtogd(fs, bno); 2802 if (size == fs->fs_bsize) { 2803 blkalloced = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno)); 2804 } else { 2805 frags = numfrags(fs, size); 2806 for (blkalloced = 0, i = 0; i < frags; i++) 2807 if (isset(blksfree, cgbno + i)) 2808 blkalloced++; 2809 if (blkalloced != 0 && blkalloced != frags) 2810 panic("ffs_checkfreeblk: partially free fragment"); 2811 } 2812 brelse(bp); 2813 return (blkalloced == 0); 2814 } 2815 #endif /* INVARIANTS */ 2816 2817 /* 2818 * Free an inode. 2819 */ 2820 int 2821 ffs_vfree(struct vnode *pvp, 2822 ino_t ino, 2823 int mode) 2824 { 2825 struct ufsmount *ump; 2826 2827 if (DOINGSOFTDEP(pvp)) { 2828 softdep_freefile(pvp, ino, mode); 2829 return (0); 2830 } 2831 ump = VFSTOUFS(pvp->v_mount); 2832 return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL)); 2833 } 2834 2835 /* 2836 * Do the actual free operation. 2837 * The specified inode is placed back in the free map. 2838 */ 2839 int 2840 ffs_freefile(struct ufsmount *ump, 2841 struct fs *fs, 2842 struct vnode *devvp, 2843 ino_t ino, 2844 int mode, 2845 struct workhead *wkhd) 2846 { 2847 struct cg *cgp; 2848 struct buf *bp; 2849 daddr_t dbn; 2850 int error; 2851 uint64_t cg; 2852 uint8_t *inosused; 2853 struct cdev *dev; 2854 ino_t cgino; 2855 2856 cg = ino_to_cg(fs, ino); 2857 if (devvp->v_type == VREG) { 2858 /* devvp is a snapshot */ 2859 MPASS(devvp->v_mount->mnt_data == ump); 2860 dev = ump->um_devvp->v_rdev; 2861 } else if (devvp->v_type == VCHR) { 2862 /* devvp is a normal disk device */ 2863 dev = devvp->v_rdev; 2864 } else { 2865 bp = NULL; 2866 return (0); 2867 } 2868 if (ino >= fs->fs_ipg * fs->fs_ncg) 2869 panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s", 2870 devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt); 2871 if ((error = ffs_getcg(fs, devvp, cg, GB_CVTENXIO, &bp, &cgp)) != 0) { 2872 if (!MOUNTEDSOFTDEP(UFSTOVFS(ump)) || devvp->v_type != VCHR) 2873 return (error); 2874 /* 2875 * Would like to just downgrade to read-only. Until that 2876 * capability is available, just toss the cylinder group 2877 * update and mark the filesystem as needing to run fsck. 2878 */ 2879 fs->fs_flags |= FS_NEEDSFSCK; 2880 if (devvp->v_type == VREG) 2881 dbn = fragstoblks(fs, cgtod(fs, cg)); 2882 else 2883 dbn = fsbtodb(fs, cgtod(fs, cg)); 2884 error = getblkx(devvp, dbn, dbn, fs->fs_cgsize, 0, 0, 0, &bp); 2885 KASSERT(error == 0, ("getblkx failed")); 2886 softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd, true); 2887 bp->b_flags |= B_RELBUF | B_NOCACHE; 2888 bp->b_flags &= ~B_CACHE; 2889 bawrite(bp); 2890 return (error); 2891 } 2892 inosused = cg_inosused(cgp); 2893 cgino = ino % fs->fs_ipg; 2894 if (isclr(inosused, cgino)) { 2895 printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev), 2896 (uintmax_t)ino, fs->fs_fsmnt); 2897 if (fs->fs_ronly == 0) 2898 panic("ffs_freefile: freeing free inode"); 2899 } 2900 clrbit(inosused, cgino); 2901 if (cgino < cgp->cg_irotor) 2902 cgp->cg_irotor = cgino; 2903 cgp->cg_cs.cs_nifree++; 2904 UFS_LOCK(ump); 2905 fs->fs_cstotal.cs_nifree++; 2906 fs->fs_cs(fs, cg).cs_nifree++; 2907 if ((mode & IFMT) == IFDIR) { 2908 cgp->cg_cs.cs_ndir--; 2909 fs->fs_cstotal.cs_ndir--; 2910 fs->fs_cs(fs, cg).cs_ndir--; 2911 } 2912 fs->fs_fmod = 1; 2913 ACTIVECLEAR(fs, cg); 2914 UFS_UNLOCK(ump); 2915 if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR) 2916 softdep_setup_inofree(UFSTOVFS(ump), bp, ino, wkhd, false); 2917 bdwrite(bp); 2918 return (0); 2919 } 2920 2921 /* 2922 * Check to see if a file is free. 2923 * Used to check for allocated files in snapshots. 2924 * Return 1 if file is free. 2925 */ 2926 int 2927 ffs_checkfreefile(struct fs *fs, 2928 struct vnode *devvp, 2929 ino_t ino) 2930 { 2931 struct cg *cgp; 2932 struct buf *bp; 2933 int ret, error; 2934 uint64_t cg; 2935 uint8_t *inosused; 2936 2937 cg = ino_to_cg(fs, ino); 2938 if ((devvp->v_type != VREG) && (devvp->v_type != VCHR)) 2939 return (1); 2940 if (ino >= fs->fs_ipg * fs->fs_ncg) 2941 return (1); 2942 if ((error = ffs_getcg(fs, devvp, cg, 0, &bp, &cgp)) != 0) 2943 return (1); 2944 inosused = cg_inosused(cgp); 2945 ino %= fs->fs_ipg; 2946 ret = isclr(inosused, ino); 2947 brelse(bp); 2948 return (ret); 2949 } 2950 2951 /* 2952 * Find a block of the specified size in the specified cylinder group. 2953 * 2954 * It is a panic if a request is made to find a block if none are 2955 * available. 2956 */ 2957 static ufs1_daddr_t 2958 ffs_mapsearch(struct fs *fs, 2959 struct cg *cgp, 2960 ufs2_daddr_t bpref, 2961 int allocsiz) 2962 { 2963 ufs1_daddr_t bno; 2964 int start, len, loc, i; 2965 int blk, field, subfield, pos; 2966 uint8_t *blksfree; 2967 2968 /* 2969 * find the fragment by searching through the free block 2970 * map for an appropriate bit pattern 2971 */ 2972 if (bpref) 2973 start = dtogd(fs, bpref) / NBBY; 2974 else 2975 start = cgp->cg_frotor / NBBY; 2976 blksfree = cg_blksfree(cgp); 2977 len = howmany(fs->fs_fpg, NBBY) - start; 2978 loc = scanc((uint64_t)len, (uint8_t *)&blksfree[start], 2979 fragtbl[fs->fs_frag], 2980 (uint8_t)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY)))); 2981 if (loc == 0) { 2982 len = start + 1; 2983 start = 0; 2984 loc = scanc((uint64_t)len, (uint8_t *)&blksfree[0], 2985 fragtbl[fs->fs_frag], 2986 (uint8_t)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY)))); 2987 if (loc == 0) { 2988 printf("start = %d, len = %d, fs = %s\n", 2989 start, len, fs->fs_fsmnt); 2990 panic("ffs_alloccg: map corrupted"); 2991 /* NOTREACHED */ 2992 } 2993 } 2994 bno = (start + len - loc) * NBBY; 2995 cgp->cg_frotor = bno; 2996 /* 2997 * found the byte in the map 2998 * sift through the bits to find the selected frag 2999 */ 3000 for (i = bno + NBBY; bno < i; bno += fs->fs_frag) { 3001 blk = blkmap(fs, blksfree, bno); 3002 blk <<= 1; 3003 field = around[allocsiz]; 3004 subfield = inside[allocsiz]; 3005 for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) { 3006 if ((blk & field) == subfield) 3007 return (bno + pos); 3008 field <<= 1; 3009 subfield <<= 1; 3010 } 3011 } 3012 printf("bno = %ju, fs = %s\n", (intmax_t)bno, fs->fs_fsmnt); 3013 panic("ffs_alloccg: block not in map"); 3014 return (-1); 3015 } 3016 3017 /* 3018 * Fetch and verify a cylinder group. 3019 */ 3020 int 3021 ffs_getcg(struct fs *fs, 3022 struct vnode *devvp, 3023 uint64_t cg, 3024 int flags, 3025 struct buf **bpp, 3026 struct cg **cgpp) 3027 { 3028 struct buf *bp; 3029 struct cg *cgp; 3030 struct mount *mp; 3031 const struct statfs *sfs; 3032 daddr_t blkno; 3033 int error; 3034 3035 *bpp = NULL; 3036 *cgpp = NULL; 3037 if ((fs->fs_metackhash & CK_CYLGRP) != 0) 3038 flags |= GB_CKHASH; 3039 if (devvp->v_type == VCHR) { 3040 blkno = fsbtodb(fs, cgtod(fs, cg)); 3041 mp = devvp->v_rdev->si_mountpt; 3042 } else { 3043 blkno = fragstoblks(fs, cgtod(fs, cg)); 3044 mp = devvp->v_mount; 3045 } 3046 error = breadn_flags(devvp, blkno, blkno, (int)fs->fs_cgsize, NULL, 3047 NULL, 0, NOCRED, flags, ffs_ckhash_cg, &bp); 3048 if (error != 0) 3049 return (error); 3050 cgp = (struct cg *)bp->b_data; 3051 if ((fs->fs_metackhash & CK_CYLGRP) != 0 && 3052 (bp->b_flags & B_CKHASH) != 0 && 3053 cgp->cg_ckhash != bp->b_ckhash) { 3054 if (ppsratecheck(&VFSTOUFS(mp)->um_last_integritymsg, 3055 &VFSTOUFS(mp)->um_secs_integritymsg, 1)) { 3056 sfs = &mp->mnt_stat; 3057 printf("UFS %s%s (%s) cylinder checkhash failed: " 3058 "cg %ju, cgp: 0x%x != bp: 0x%jx\n", 3059 devvp->v_type == VCHR ? "" : "snapshot of ", 3060 sfs->f_mntfromname, sfs->f_mntonname, (intmax_t)cg, 3061 cgp->cg_ckhash, (uintmax_t)bp->b_ckhash); 3062 } 3063 bp->b_flags &= ~B_CKHASH; 3064 bp->b_flags |= B_INVAL | B_NOCACHE; 3065 brelse(bp); 3066 return (EINTEGRITY); 3067 } 3068 if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) { 3069 if (ppsratecheck(&VFSTOUFS(mp)->um_last_integritymsg, 3070 &VFSTOUFS(mp)->um_secs_integritymsg, 1)) { 3071 sfs = &mp->mnt_stat; 3072 printf("UFS %s%s (%s)", 3073 devvp->v_type == VCHR ? "" : "snapshot of ", 3074 sfs->f_mntfromname, sfs->f_mntonname); 3075 if (!cg_chkmagic(cgp)) 3076 printf(" cg %ju: bad magic number 0x%x should " 3077 "be 0x%x\n", (intmax_t)cg, cgp->cg_magic, 3078 CG_MAGIC); 3079 else 3080 printf(": wrong cylinder group cg %ju != " 3081 "cgx %u\n", (intmax_t)cg, cgp->cg_cgx); 3082 } 3083 bp->b_flags &= ~B_CKHASH; 3084 bp->b_flags |= B_INVAL | B_NOCACHE; 3085 brelse(bp); 3086 return (EINTEGRITY); 3087 } 3088 bp->b_flags &= ~B_CKHASH; 3089 bp->b_xflags |= BX_BKGRDWRITE; 3090 /* 3091 * If we are using check hashes on the cylinder group then we want 3092 * to limit changing the cylinder group time to when we are actually 3093 * going to write it to disk so that its check hash remains correct 3094 * in memory. If the CK_CYLGRP flag is set the time is updated in 3095 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we 3096 * update the time here as we have done historically. 3097 */ 3098 if ((fs->fs_metackhash & CK_CYLGRP) != 0) 3099 bp->b_xflags |= BX_CYLGRP; 3100 else 3101 cgp->cg_old_time = cgp->cg_time = time_second; 3102 *bpp = bp; 3103 *cgpp = cgp; 3104 return (0); 3105 } 3106 3107 static void 3108 ffs_ckhash_cg(struct buf *bp) 3109 { 3110 uint32_t ckhash; 3111 struct cg *cgp; 3112 3113 cgp = (struct cg *)bp->b_data; 3114 ckhash = cgp->cg_ckhash; 3115 cgp->cg_ckhash = 0; 3116 bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount); 3117 cgp->cg_ckhash = ckhash; 3118 } 3119 3120 /* 3121 * Called when a cylinder group read has failed. If an integrity check 3122 * is the cause of failure then the cylinder group will not be usable 3123 * until the filesystem has been unmounted and fsck has been run to 3124 * repair it. To avoid future attempts to allocate resources from the 3125 * cylinder group, its available resources are set to zero in the 3126 * superblock summary information. Since it will appear to have no 3127 * resources available, no further calls will be made to allocate 3128 * resources from it. When resources are freed to the cylinder group 3129 * the resource free routines will find the cylinder group unusable so 3130 * the resource will simply be discarded and thus will not show up in 3131 * the superblock summary information until they are recovered by fsck. 3132 */ 3133 static void 3134 ffs_checkcgintegrity(struct fs *fs, 3135 uint64_t cg, 3136 int error) 3137 { 3138 3139 if (error != EINTEGRITY) 3140 return; 3141 fs->fs_cstotal.cs_nffree -= fs->fs_cs(fs, cg).cs_nffree; 3142 fs->fs_cs(fs, cg).cs_nffree = 0; 3143 fs->fs_cstotal.cs_nbfree -= fs->fs_cs(fs, cg).cs_nbfree; 3144 fs->fs_cs(fs, cg).cs_nbfree = 0; 3145 fs->fs_cstotal.cs_nifree -= fs->fs_cs(fs, cg).cs_nifree; 3146 fs->fs_cs(fs, cg).cs_nifree = 0; 3147 fs->fs_maxcluster[cg] = 0; 3148 fs->fs_flags |= FS_NEEDSFSCK; 3149 fs->fs_fmod = 1; 3150 } 3151 3152 /* 3153 * Fserr prints the name of a filesystem with an error diagnostic. 3154 * 3155 * The form of the error message is: 3156 * fs: error message 3157 */ 3158 void 3159 ffs_fserr(struct fs *fs, 3160 ino_t inum, 3161 char *cp) 3162 { 3163 struct thread *td = curthread; /* XXX */ 3164 struct proc *p = td->td_proc; 3165 3166 log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n", 3167 p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum, 3168 fs->fs_fsmnt, cp); 3169 } 3170 3171 /* 3172 * This function provides the capability for the fsck program to 3173 * update an active filesystem. Sixteen operations are provided: 3174 * 3175 * adjrefcnt(inode, amt) - adjusts the reference count on the 3176 * specified inode by the specified amount. Under normal 3177 * operation the count should always go down. Decrementing 3178 * the count to zero will cause the inode to be freed. 3179 * adjblkcnt(inode, amt) - adjust the number of blocks used by the 3180 * inode by the specified amount. 3181 * adjdepth(inode, amt) - adjust the depth of the specified directory 3182 * inode by the specified amount. 3183 * setsize(inode, size) - set the size of the inode to the 3184 * specified size. 3185 * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) - 3186 * adjust the superblock summary. 3187 * freedirs(inode, count) - directory inodes [inode..inode + count - 1] 3188 * are marked as free. Inodes should never have to be marked 3189 * as in use. 3190 * freefiles(inode, count) - file inodes [inode..inode + count - 1] 3191 * are marked as free. Inodes should never have to be marked 3192 * as in use. 3193 * freeblks(blockno, size) - blocks [blockno..blockno + size - 1] 3194 * are marked as free. Blocks should never have to be marked 3195 * as in use. 3196 * setflags(flags, set/clear) - the fs_flags field has the specified 3197 * flags set (second parameter +1) or cleared (second parameter -1). 3198 * setcwd(dirinode) - set the current directory to dirinode in the 3199 * filesystem associated with the snapshot. 3200 * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".." 3201 * in the current directory is oldvalue then change it to newvalue. 3202 * unlink(nameptr, oldvalue) - Verify that the inode number associated 3203 * with nameptr in the current directory is oldvalue then unlink it. 3204 */ 3205 3206 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS); 3207 3208 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt, 3209 CTLFLAG_WR | CTLTYPE_STRUCT | CTLFLAG_NEEDGIANT, 3210 0, 0, sysctl_ffs_fsck, "S,fsck", 3211 "Adjust Inode Reference Count"); 3212 3213 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt, 3214 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3215 "Adjust Inode Used Blocks Count"); 3216 3217 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_DEPTH, adjdepth, 3218 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3219 "Adjust Directory Inode Depth"); 3220 3221 static SYSCTL_NODE(_vfs_ffs, FFS_SET_SIZE, setsize, 3222 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3223 "Set the inode size"); 3224 3225 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir, 3226 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3227 "Adjust number of directories"); 3228 3229 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree, 3230 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3231 "Adjust number of free blocks"); 3232 3233 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree, 3234 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3235 "Adjust number of free inodes"); 3236 3237 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree, 3238 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3239 "Adjust number of free frags"); 3240 3241 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters, 3242 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3243 "Adjust number of free clusters"); 3244 3245 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs, 3246 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3247 "Free Range of Directory Inodes"); 3248 3249 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles, 3250 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3251 "Free Range of File Inodes"); 3252 3253 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks, 3254 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3255 "Free Range of Blocks"); 3256 3257 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags, 3258 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3259 "Change Filesystem Flags"); 3260 3261 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd, 3262 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3263 "Set Current Working Directory"); 3264 3265 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot, 3266 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3267 "Change Value of .. Entry"); 3268 3269 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink, 3270 CTLFLAG_WR | CTLFLAG_NEEDGIANT, sysctl_ffs_fsck, 3271 "Unlink a Duplicate Name"); 3272 3273 #ifdef DIAGNOSTIC 3274 static int fsckcmds = 0; 3275 SYSCTL_INT(_debug, OID_AUTO, ffs_fsckcmds, CTLFLAG_RW, &fsckcmds, 0, 3276 "print out fsck_ffs-based filesystem update commands"); 3277 #endif /* DIAGNOSTIC */ 3278 3279 static int 3280 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS) 3281 { 3282 struct thread *td = curthread; 3283 struct fsck_cmd cmd; 3284 struct ufsmount *ump; 3285 struct vnode *vp, *dvp, *fdvp; 3286 struct inode *ip, *dp; 3287 struct mount *mp; 3288 struct fs *fs; 3289 struct pwd *pwd; 3290 ufs2_daddr_t blkno; 3291 long blkcnt, blksize; 3292 uint64_t key; 3293 struct file *fp; 3294 cap_rights_t rights; 3295 int filetype, error; 3296 3297 if (req->newptr == NULL || req->newlen > sizeof(cmd)) 3298 return (EBADRPC); 3299 if ((error = SYSCTL_IN(req, &cmd, sizeof(cmd))) != 0) 3300 return (error); 3301 if (cmd.version != FFS_CMD_VERSION) 3302 return (ERPCMISMATCH); 3303 if ((error = getvnode(td, cmd.handle, 3304 cap_rights_init_one(&rights, CAP_FSCK), &fp)) != 0) 3305 return (error); 3306 vp = fp->f_vnode; 3307 if (vp->v_type != VREG && vp->v_type != VDIR) { 3308 fdrop(fp, td); 3309 return (EINVAL); 3310 } 3311 vn_start_write(vp, &mp, V_WAIT); 3312 if (mp == NULL || 3313 strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) { 3314 vn_finished_write(mp); 3315 fdrop(fp, td); 3316 return (EINVAL); 3317 } 3318 ump = VFSTOUFS(mp); 3319 if (mp->mnt_flag & MNT_RDONLY) { 3320 vn_finished_write(mp); 3321 fdrop(fp, td); 3322 return (EROFS); 3323 } 3324 fs = ump->um_fs; 3325 filetype = IFREG; 3326 3327 switch (oidp->oid_number) { 3328 case FFS_SET_FLAGS: 3329 #ifdef DIAGNOSTIC 3330 if (fsckcmds) 3331 printf("%s: %s flags\n", mp->mnt_stat.f_mntonname, 3332 cmd.size > 0 ? "set" : "clear"); 3333 #endif /* DIAGNOSTIC */ 3334 if (cmd.size > 0) 3335 fs->fs_flags |= (long)cmd.value; 3336 else 3337 fs->fs_flags &= ~(long)cmd.value; 3338 break; 3339 3340 case FFS_ADJ_REFCNT: 3341 #ifdef DIAGNOSTIC 3342 if (fsckcmds) { 3343 printf("%s: adjust inode %jd link count by %jd\n", 3344 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3345 (intmax_t)cmd.size); 3346 } 3347 #endif /* DIAGNOSTIC */ 3348 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3349 break; 3350 ip = VTOI(vp); 3351 ip->i_nlink += cmd.size; 3352 DIP_SET_NLINK(ip, ip->i_nlink); 3353 ip->i_effnlink += cmd.size; 3354 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED); 3355 error = ffs_update(vp, 1); 3356 if (DOINGSOFTDEP(vp)) 3357 softdep_change_linkcnt(ip); 3358 vput(vp); 3359 break; 3360 3361 case FFS_ADJ_BLKCNT: 3362 #ifdef DIAGNOSTIC 3363 if (fsckcmds) { 3364 printf("%s: adjust inode %jd block count by %jd\n", 3365 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3366 (intmax_t)cmd.size); 3367 } 3368 #endif /* DIAGNOSTIC */ 3369 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3370 break; 3371 ip = VTOI(vp); 3372 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size); 3373 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED); 3374 error = ffs_update(vp, 1); 3375 vput(vp); 3376 break; 3377 3378 case FFS_ADJ_DEPTH: 3379 #ifdef DIAGNOSTIC 3380 if (fsckcmds) { 3381 printf("%s: adjust directory inode %jd depth by %jd\n", 3382 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3383 (intmax_t)cmd.size); 3384 } 3385 #endif /* DIAGNOSTIC */ 3386 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3387 break; 3388 if (vp->v_type != VDIR) { 3389 vput(vp); 3390 error = ENOTDIR; 3391 break; 3392 } 3393 ip = VTOI(vp); 3394 DIP_SET(ip, i_dirdepth, DIP(ip, i_dirdepth) + cmd.size); 3395 UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_MODIFIED); 3396 error = ffs_update(vp, 1); 3397 vput(vp); 3398 break; 3399 3400 case FFS_SET_SIZE: 3401 #ifdef DIAGNOSTIC 3402 if (fsckcmds) { 3403 printf("%s: set inode %jd size to %jd\n", 3404 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3405 (intmax_t)cmd.size); 3406 } 3407 #endif /* DIAGNOSTIC */ 3408 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3409 break; 3410 ip = VTOI(vp); 3411 DIP_SET(ip, i_size, cmd.size); 3412 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE | IN_MODIFIED); 3413 error = ffs_update(vp, 1); 3414 vput(vp); 3415 break; 3416 3417 case FFS_DIR_FREE: 3418 filetype = IFDIR; 3419 /* fall through */ 3420 3421 case FFS_FILE_FREE: 3422 #ifdef DIAGNOSTIC 3423 if (fsckcmds) { 3424 if (cmd.size == 1) 3425 printf("%s: free %s inode %ju\n", 3426 mp->mnt_stat.f_mntonname, 3427 filetype == IFDIR ? "directory" : "file", 3428 (uintmax_t)cmd.value); 3429 else 3430 printf("%s: free %s inodes %ju-%ju\n", 3431 mp->mnt_stat.f_mntonname, 3432 filetype == IFDIR ? "directory" : "file", 3433 (uintmax_t)cmd.value, 3434 (uintmax_t)(cmd.value + cmd.size - 1)); 3435 } 3436 #endif /* DIAGNOSTIC */ 3437 while (cmd.size > 0) { 3438 if ((error = ffs_freefile(ump, fs, ump->um_devvp, 3439 cmd.value, filetype, NULL))) 3440 break; 3441 cmd.size -= 1; 3442 cmd.value += 1; 3443 } 3444 break; 3445 3446 case FFS_BLK_FREE: 3447 #ifdef DIAGNOSTIC 3448 if (fsckcmds) { 3449 if (cmd.size == 1) 3450 printf("%s: free block %jd\n", 3451 mp->mnt_stat.f_mntonname, 3452 (intmax_t)cmd.value); 3453 else 3454 printf("%s: free blocks %jd-%jd\n", 3455 mp->mnt_stat.f_mntonname, 3456 (intmax_t)cmd.value, 3457 (intmax_t)cmd.value + cmd.size - 1); 3458 } 3459 #endif /* DIAGNOSTIC */ 3460 blkno = cmd.value; 3461 blkcnt = cmd.size; 3462 blksize = fs->fs_frag - (blkno % fs->fs_frag); 3463 key = ffs_blkrelease_start(ump, ump->um_devvp, UFS_ROOTINO); 3464 while (blkcnt > 0) { 3465 if (blkcnt < blksize) 3466 blksize = blkcnt; 3467 ffs_blkfree(ump, fs, ump->um_devvp, blkno, 3468 blksize * fs->fs_fsize, UFS_ROOTINO, 3469 VDIR, NULL, key); 3470 blkno += blksize; 3471 blkcnt -= blksize; 3472 blksize = fs->fs_frag; 3473 } 3474 ffs_blkrelease_finish(ump, key); 3475 break; 3476 3477 /* 3478 * Adjust superblock summaries. fsck(8) is expected to 3479 * submit deltas when necessary. 3480 */ 3481 case FFS_ADJ_NDIR: 3482 #ifdef DIAGNOSTIC 3483 if (fsckcmds) { 3484 printf("%s: adjust number of directories by %jd\n", 3485 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3486 } 3487 #endif /* DIAGNOSTIC */ 3488 fs->fs_cstotal.cs_ndir += cmd.value; 3489 break; 3490 3491 case FFS_ADJ_NBFREE: 3492 #ifdef DIAGNOSTIC 3493 if (fsckcmds) { 3494 printf("%s: adjust number of free blocks by %+jd\n", 3495 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3496 } 3497 #endif /* DIAGNOSTIC */ 3498 fs->fs_cstotal.cs_nbfree += cmd.value; 3499 break; 3500 3501 case FFS_ADJ_NIFREE: 3502 #ifdef DIAGNOSTIC 3503 if (fsckcmds) { 3504 printf("%s: adjust number of free inodes by %+jd\n", 3505 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3506 } 3507 #endif /* DIAGNOSTIC */ 3508 fs->fs_cstotal.cs_nifree += cmd.value; 3509 break; 3510 3511 case FFS_ADJ_NFFREE: 3512 #ifdef DIAGNOSTIC 3513 if (fsckcmds) { 3514 printf("%s: adjust number of free frags by %+jd\n", 3515 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3516 } 3517 #endif /* DIAGNOSTIC */ 3518 fs->fs_cstotal.cs_nffree += cmd.value; 3519 break; 3520 3521 case FFS_ADJ_NUMCLUSTERS: 3522 #ifdef DIAGNOSTIC 3523 if (fsckcmds) { 3524 printf("%s: adjust number of free clusters by %+jd\n", 3525 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3526 } 3527 #endif /* DIAGNOSTIC */ 3528 fs->fs_cstotal.cs_numclusters += cmd.value; 3529 break; 3530 3531 case FFS_SET_CWD: 3532 #ifdef DIAGNOSTIC 3533 if (fsckcmds) { 3534 printf("%s: set current directory to inode %jd\n", 3535 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3536 } 3537 #endif /* DIAGNOSTIC */ 3538 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp))) 3539 break; 3540 AUDIT_ARG_VNODE1(vp); 3541 if ((error = change_dir(vp, td)) != 0) { 3542 vput(vp); 3543 break; 3544 } 3545 VOP_UNLOCK(vp); 3546 pwd_chdir(td, vp); 3547 break; 3548 3549 case FFS_SET_DOTDOT: 3550 #ifdef DIAGNOSTIC 3551 if (fsckcmds) { 3552 printf("%s: change .. in cwd from %jd to %jd\n", 3553 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3554 (intmax_t)cmd.size); 3555 } 3556 #endif /* DIAGNOSTIC */ 3557 /* 3558 * First we have to get and lock the parent directory 3559 * to which ".." points. 3560 */ 3561 error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp); 3562 if (error) 3563 break; 3564 /* 3565 * Now we get and lock the child directory containing "..". 3566 */ 3567 pwd = pwd_hold(td); 3568 dvp = pwd->pwd_cdir; 3569 if ((error = vget(dvp, LK_EXCLUSIVE)) != 0) { 3570 vput(fdvp); 3571 pwd_drop(pwd); 3572 break; 3573 } 3574 dp = VTOI(dvp); 3575 SET_I_OFFSET(dp, 12); /* XXX mastertemplate.dot_reclen */ 3576 error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size, 3577 DT_DIR, 0); 3578 cache_purge(fdvp); 3579 cache_purge(dvp); 3580 vput(dvp); 3581 vput(fdvp); 3582 pwd_drop(pwd); 3583 break; 3584 3585 case FFS_UNLINK: 3586 #ifdef DIAGNOSTIC 3587 if (fsckcmds) { 3588 char buf[32]; 3589 3590 if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL)) 3591 strncpy(buf, "Name_too_long", 32); 3592 printf("%s: unlink %s (inode %jd)\n", 3593 mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size); 3594 } 3595 #endif /* DIAGNOSTIC */ 3596 /* 3597 * kern_funlinkat will do its own start/finish writes and 3598 * they do not nest, so drop ours here. Setting mp == NULL 3599 * indicates that vn_finished_write is not needed down below. 3600 */ 3601 vn_finished_write(mp); 3602 mp = NULL; 3603 error = kern_funlinkat(td, AT_FDCWD, 3604 (char *)(intptr_t)cmd.value, FD_NONE, UIO_USERSPACE, 3605 0, (ino_t)cmd.size); 3606 break; 3607 3608 default: 3609 #ifdef DIAGNOSTIC 3610 if (fsckcmds) { 3611 printf("Invalid request %d from fsck\n", 3612 oidp->oid_number); 3613 } 3614 #endif /* DIAGNOSTIC */ 3615 error = EINVAL; 3616 break; 3617 } 3618 fdrop(fp, td); 3619 vn_finished_write(mp); 3620 return (error); 3621 } 3622