1 /*- 2 * SPDX-License-Identifier: (BSD-2-Clause-FreeBSD AND BSD-3-Clause) 3 * 4 * Copyright (c) 2002 Networks Associates Technology, Inc. 5 * All rights reserved. 6 * 7 * This software was developed for the FreeBSD Project by Marshall 8 * Kirk McKusick and Network Associates Laboratories, the Security 9 * Research Division of Network Associates, Inc. under DARPA/SPAWAR 10 * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS 11 * research program 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1982, 1986, 1989, 1993 35 * The Regents of the University of California. All rights reserved. 36 * 37 * Redistribution and use in source and binary forms, with or without 38 * modification, are permitted provided that the following conditions 39 * are met: 40 * 1. Redistributions of source code must retain the above copyright 41 * notice, this list of conditions and the following disclaimer. 42 * 2. Redistributions in binary form must reproduce the above copyright 43 * notice, this list of conditions and the following disclaimer in the 44 * documentation and/or other materials provided with the distribution. 45 * 3. Neither the name of the University nor the names of its contributors 46 * may be used to endorse or promote products derived from this software 47 * without specific prior written permission. 48 * 49 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 50 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 51 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 52 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 53 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 54 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 55 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 56 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 57 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 58 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 59 * SUCH DAMAGE. 60 * 61 * @(#)ffs_alloc.c 8.18 (Berkeley) 5/26/95 62 */ 63 64 #include <sys/cdefs.h> 65 __FBSDID("$FreeBSD$"); 66 67 #include "opt_quota.h" 68 69 #include <sys/param.h> 70 #include <sys/capsicum.h> 71 #include <sys/gsb_crc32.h> 72 #include <sys/systm.h> 73 #include <sys/bio.h> 74 #include <sys/buf.h> 75 #include <sys/conf.h> 76 #include <sys/fcntl.h> 77 #include <sys/file.h> 78 #include <sys/filedesc.h> 79 #include <sys/priv.h> 80 #include <sys/proc.h> 81 #include <sys/vnode.h> 82 #include <sys/mount.h> 83 #include <sys/kernel.h> 84 #include <sys/syscallsubr.h> 85 #include <sys/sysctl.h> 86 #include <sys/syslog.h> 87 #include <sys/taskqueue.h> 88 89 #include <security/audit/audit.h> 90 91 #include <geom/geom.h> 92 #include <geom/geom_vfs.h> 93 94 #include <ufs/ufs/dir.h> 95 #include <ufs/ufs/extattr.h> 96 #include <ufs/ufs/quota.h> 97 #include <ufs/ufs/inode.h> 98 #include <ufs/ufs/ufs_extern.h> 99 #include <ufs/ufs/ufsmount.h> 100 101 #include <ufs/ffs/fs.h> 102 #include <ufs/ffs/ffs_extern.h> 103 #include <ufs/ffs/softdep.h> 104 105 typedef ufs2_daddr_t allocfcn_t(struct inode *ip, u_int cg, ufs2_daddr_t bpref, 106 int size, int rsize); 107 108 static ufs2_daddr_t ffs_alloccg(struct inode *, u_int, ufs2_daddr_t, int, int); 109 static ufs2_daddr_t 110 ffs_alloccgblk(struct inode *, struct buf *, ufs2_daddr_t, int); 111 static void ffs_blkfree_cg(struct ufsmount *, struct fs *, 112 struct vnode *, ufs2_daddr_t, long, ino_t, 113 struct workhead *); 114 #ifdef INVARIANTS 115 static int ffs_checkblk(struct inode *, ufs2_daddr_t, long); 116 #endif 117 static ufs2_daddr_t ffs_clusteralloc(struct inode *, u_int, ufs2_daddr_t, int); 118 static ino_t ffs_dirpref(struct inode *); 119 static ufs2_daddr_t ffs_fragextend(struct inode *, u_int, ufs2_daddr_t, 120 int, int); 121 static ufs2_daddr_t ffs_hashalloc 122 (struct inode *, u_int, ufs2_daddr_t, int, int, allocfcn_t *); 123 static ufs2_daddr_t ffs_nodealloccg(struct inode *, u_int, ufs2_daddr_t, int, 124 int); 125 static ufs1_daddr_t ffs_mapsearch(struct fs *, struct cg *, ufs2_daddr_t, int); 126 static int ffs_reallocblks_ufs1(struct vop_reallocblks_args *); 127 static int ffs_reallocblks_ufs2(struct vop_reallocblks_args *); 128 static void ffs_ckhash_cg(struct buf *); 129 130 /* 131 * Allocate a block in the filesystem. 132 * 133 * The size of the requested block is given, which must be some 134 * multiple of fs_fsize and <= fs_bsize. 135 * A preference may be optionally specified. If a preference is given 136 * the following hierarchy is used to allocate a block: 137 * 1) allocate the requested block. 138 * 2) allocate a rotationally optimal block in the same cylinder. 139 * 3) allocate a block in the same cylinder group. 140 * 4) quadradically rehash into other cylinder groups, until an 141 * available block is located. 142 * If no block preference is given the following hierarchy is used 143 * to allocate a block: 144 * 1) allocate a block in the cylinder group that contains the 145 * inode for the file. 146 * 2) quadradically rehash into other cylinder groups, until an 147 * available block is located. 148 */ 149 int 150 ffs_alloc(ip, lbn, bpref, size, flags, cred, bnp) 151 struct inode *ip; 152 ufs2_daddr_t lbn, bpref; 153 int size, flags; 154 struct ucred *cred; 155 ufs2_daddr_t *bnp; 156 { 157 struct fs *fs; 158 struct ufsmount *ump; 159 ufs2_daddr_t bno; 160 u_int cg, reclaimed; 161 int64_t delta; 162 #ifdef QUOTA 163 int error; 164 #endif 165 166 *bnp = 0; 167 ump = ITOUMP(ip); 168 fs = ump->um_fs; 169 mtx_assert(UFS_MTX(ump), MA_OWNED); 170 #ifdef INVARIANTS 171 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) { 172 printf("dev = %s, bsize = %ld, size = %d, fs = %s\n", 173 devtoname(ump->um_dev), (long)fs->fs_bsize, size, 174 fs->fs_fsmnt); 175 panic("ffs_alloc: bad size"); 176 } 177 if (cred == NOCRED) 178 panic("ffs_alloc: missing credential"); 179 #endif /* INVARIANTS */ 180 reclaimed = 0; 181 retry: 182 #ifdef QUOTA 183 UFS_UNLOCK(ump); 184 error = chkdq(ip, btodb(size), cred, 0); 185 if (error) 186 return (error); 187 UFS_LOCK(ump); 188 #endif 189 if (size == fs->fs_bsize && fs->fs_cstotal.cs_nbfree == 0) 190 goto nospace; 191 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) && 192 freespace(fs, fs->fs_minfree) - numfrags(fs, size) < 0) 193 goto nospace; 194 if (bpref >= fs->fs_size) 195 bpref = 0; 196 if (bpref == 0) 197 cg = ino_to_cg(fs, ip->i_number); 198 else 199 cg = dtog(fs, bpref); 200 bno = ffs_hashalloc(ip, cg, bpref, size, size, ffs_alloccg); 201 if (bno > 0) { 202 delta = btodb(size); 203 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 204 if (flags & IO_EXT) 205 ip->i_flag |= IN_CHANGE; 206 else 207 ip->i_flag |= IN_CHANGE | IN_UPDATE; 208 *bnp = bno; 209 return (0); 210 } 211 nospace: 212 #ifdef QUOTA 213 UFS_UNLOCK(ump); 214 /* 215 * Restore user's disk quota because allocation failed. 216 */ 217 (void) chkdq(ip, -btodb(size), cred, FORCE); 218 UFS_LOCK(ump); 219 #endif 220 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) { 221 reclaimed = 1; 222 softdep_request_cleanup(fs, ITOV(ip), cred, FLUSH_BLOCKS_WAIT); 223 goto retry; 224 } 225 if (reclaimed > 0 && 226 ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 227 UFS_UNLOCK(ump); 228 ffs_fserr(fs, ip->i_number, "filesystem full"); 229 uprintf("\n%s: write failed, filesystem is full\n", 230 fs->fs_fsmnt); 231 } else { 232 UFS_UNLOCK(ump); 233 } 234 return (ENOSPC); 235 } 236 237 /* 238 * Reallocate a fragment to a bigger size 239 * 240 * The number and size of the old block is given, and a preference 241 * and new size is also specified. The allocator attempts to extend 242 * the original block. Failing that, the regular block allocator is 243 * invoked to get an appropriate block. 244 */ 245 int 246 ffs_realloccg(ip, lbprev, bprev, bpref, osize, nsize, flags, cred, bpp) 247 struct inode *ip; 248 ufs2_daddr_t lbprev; 249 ufs2_daddr_t bprev; 250 ufs2_daddr_t bpref; 251 int osize, nsize, flags; 252 struct ucred *cred; 253 struct buf **bpp; 254 { 255 struct vnode *vp; 256 struct fs *fs; 257 struct buf *bp; 258 struct ufsmount *ump; 259 u_int cg, request, reclaimed; 260 int error, gbflags; 261 ufs2_daddr_t bno; 262 int64_t delta; 263 264 vp = ITOV(ip); 265 ump = ITOUMP(ip); 266 fs = ump->um_fs; 267 bp = NULL; 268 gbflags = (flags & BA_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 269 270 mtx_assert(UFS_MTX(ump), MA_OWNED); 271 #ifdef INVARIANTS 272 if (vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED) 273 panic("ffs_realloccg: allocation on suspended filesystem"); 274 if ((u_int)osize > fs->fs_bsize || fragoff(fs, osize) != 0 || 275 (u_int)nsize > fs->fs_bsize || fragoff(fs, nsize) != 0) { 276 printf( 277 "dev = %s, bsize = %ld, osize = %d, nsize = %d, fs = %s\n", 278 devtoname(ump->um_dev), (long)fs->fs_bsize, osize, 279 nsize, fs->fs_fsmnt); 280 panic("ffs_realloccg: bad size"); 281 } 282 if (cred == NOCRED) 283 panic("ffs_realloccg: missing credential"); 284 #endif /* INVARIANTS */ 285 reclaimed = 0; 286 retry: 287 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE) && 288 freespace(fs, fs->fs_minfree) - numfrags(fs, nsize - osize) < 0) { 289 goto nospace; 290 } 291 if (bprev == 0) { 292 printf("dev = %s, bsize = %ld, bprev = %jd, fs = %s\n", 293 devtoname(ump->um_dev), (long)fs->fs_bsize, (intmax_t)bprev, 294 fs->fs_fsmnt); 295 panic("ffs_realloccg: bad bprev"); 296 } 297 UFS_UNLOCK(ump); 298 /* 299 * Allocate the extra space in the buffer. 300 */ 301 error = bread_gb(vp, lbprev, osize, NOCRED, gbflags, &bp); 302 if (error) { 303 brelse(bp); 304 return (error); 305 } 306 307 if (bp->b_blkno == bp->b_lblkno) { 308 if (lbprev >= UFS_NDADDR) 309 panic("ffs_realloccg: lbprev out of range"); 310 bp->b_blkno = fsbtodb(fs, bprev); 311 } 312 313 #ifdef QUOTA 314 error = chkdq(ip, btodb(nsize - osize), cred, 0); 315 if (error) { 316 brelse(bp); 317 return (error); 318 } 319 #endif 320 /* 321 * Check for extension in the existing location. 322 */ 323 *bpp = NULL; 324 cg = dtog(fs, bprev); 325 UFS_LOCK(ump); 326 bno = ffs_fragextend(ip, cg, bprev, osize, nsize); 327 if (bno) { 328 if (bp->b_blkno != fsbtodb(fs, bno)) 329 panic("ffs_realloccg: bad blockno"); 330 delta = btodb(nsize - osize); 331 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 332 if (flags & IO_EXT) 333 ip->i_flag |= IN_CHANGE; 334 else 335 ip->i_flag |= IN_CHANGE | IN_UPDATE; 336 allocbuf(bp, nsize); 337 bp->b_flags |= B_DONE; 338 vfs_bio_bzero_buf(bp, osize, nsize - osize); 339 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO) 340 vfs_bio_set_valid(bp, osize, nsize - osize); 341 *bpp = bp; 342 return (0); 343 } 344 /* 345 * Allocate a new disk location. 346 */ 347 if (bpref >= fs->fs_size) 348 bpref = 0; 349 switch ((int)fs->fs_optim) { 350 case FS_OPTSPACE: 351 /* 352 * Allocate an exact sized fragment. Although this makes 353 * best use of space, we will waste time relocating it if 354 * the file continues to grow. If the fragmentation is 355 * less than half of the minimum free reserve, we choose 356 * to begin optimizing for time. 357 */ 358 request = nsize; 359 if (fs->fs_minfree <= 5 || 360 fs->fs_cstotal.cs_nffree > 361 (off_t)fs->fs_dsize * fs->fs_minfree / (2 * 100)) 362 break; 363 log(LOG_NOTICE, "%s: optimization changed from SPACE to TIME\n", 364 fs->fs_fsmnt); 365 fs->fs_optim = FS_OPTTIME; 366 break; 367 case FS_OPTTIME: 368 /* 369 * At this point we have discovered a file that is trying to 370 * grow a small fragment to a larger fragment. To save time, 371 * we allocate a full sized block, then free the unused portion. 372 * If the file continues to grow, the `ffs_fragextend' call 373 * above will be able to grow it in place without further 374 * copying. If aberrant programs cause disk fragmentation to 375 * grow within 2% of the free reserve, we choose to begin 376 * optimizing for space. 377 */ 378 request = fs->fs_bsize; 379 if (fs->fs_cstotal.cs_nffree < 380 (off_t)fs->fs_dsize * (fs->fs_minfree - 2) / 100) 381 break; 382 log(LOG_NOTICE, "%s: optimization changed from TIME to SPACE\n", 383 fs->fs_fsmnt); 384 fs->fs_optim = FS_OPTSPACE; 385 break; 386 default: 387 printf("dev = %s, optim = %ld, fs = %s\n", 388 devtoname(ump->um_dev), (long)fs->fs_optim, fs->fs_fsmnt); 389 panic("ffs_realloccg: bad optim"); 390 /* NOTREACHED */ 391 } 392 bno = ffs_hashalloc(ip, cg, bpref, request, nsize, ffs_alloccg); 393 if (bno > 0) { 394 bp->b_blkno = fsbtodb(fs, bno); 395 if (!DOINGSOFTDEP(vp)) 396 /* 397 * The usual case is that a smaller fragment that 398 * was just allocated has been replaced with a bigger 399 * fragment or a full-size block. If it is marked as 400 * B_DELWRI, the current contents have not been written 401 * to disk. It is possible that the block was written 402 * earlier, but very uncommon. If the block has never 403 * been written, there is no need to send a BIO_DELETE 404 * for it when it is freed. The gain from avoiding the 405 * TRIMs for the common case of unwritten blocks far 406 * exceeds the cost of the write amplification for the 407 * uncommon case of failing to send a TRIM for a block 408 * that had been written. 409 */ 410 ffs_blkfree(ump, fs, ump->um_devvp, bprev, (long)osize, 411 ip->i_number, vp->v_type, NULL, 412 (bp->b_flags & B_DELWRI) != 0 ? 413 NOTRIM_KEY : SINGLETON_KEY); 414 delta = btodb(nsize - osize); 415 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + delta); 416 if (flags & IO_EXT) 417 ip->i_flag |= IN_CHANGE; 418 else 419 ip->i_flag |= IN_CHANGE | IN_UPDATE; 420 allocbuf(bp, nsize); 421 bp->b_flags |= B_DONE; 422 vfs_bio_bzero_buf(bp, osize, nsize - osize); 423 if ((bp->b_flags & (B_MALLOC | B_VMIO)) == B_VMIO) 424 vfs_bio_set_valid(bp, osize, nsize - osize); 425 *bpp = bp; 426 return (0); 427 } 428 #ifdef QUOTA 429 UFS_UNLOCK(ump); 430 /* 431 * Restore user's disk quota because allocation failed. 432 */ 433 (void) chkdq(ip, -btodb(nsize - osize), cred, FORCE); 434 UFS_LOCK(ump); 435 #endif 436 nospace: 437 /* 438 * no space available 439 */ 440 if (reclaimed == 0 && (flags & IO_BUFLOCKED) == 0) { 441 reclaimed = 1; 442 UFS_UNLOCK(ump); 443 if (bp) { 444 brelse(bp); 445 bp = NULL; 446 } 447 UFS_LOCK(ump); 448 softdep_request_cleanup(fs, vp, cred, FLUSH_BLOCKS_WAIT); 449 goto retry; 450 } 451 if (reclaimed > 0 && 452 ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 453 UFS_UNLOCK(ump); 454 ffs_fserr(fs, ip->i_number, "filesystem full"); 455 uprintf("\n%s: write failed, filesystem is full\n", 456 fs->fs_fsmnt); 457 } else { 458 UFS_UNLOCK(ump); 459 } 460 if (bp) 461 brelse(bp); 462 return (ENOSPC); 463 } 464 465 /* 466 * Reallocate a sequence of blocks into a contiguous sequence of blocks. 467 * 468 * The vnode and an array of buffer pointers for a range of sequential 469 * logical blocks to be made contiguous is given. The allocator attempts 470 * to find a range of sequential blocks starting as close as possible 471 * from the end of the allocation for the logical block immediately 472 * preceding the current range. If successful, the physical block numbers 473 * in the buffer pointers and in the inode are changed to reflect the new 474 * allocation. If unsuccessful, the allocation is left unchanged. The 475 * success in doing the reallocation is returned. Note that the error 476 * return is not reflected back to the user. Rather the previous block 477 * allocation will be used. 478 */ 479 480 SYSCTL_NODE(_vfs, OID_AUTO, ffs, CTLFLAG_RW, 0, "FFS filesystem"); 481 482 static int doasyncfree = 1; 483 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncfree, CTLFLAG_RW, &doasyncfree, 0, 484 "do not force synchronous writes when blocks are reallocated"); 485 486 static int doreallocblks = 1; 487 SYSCTL_INT(_vfs_ffs, OID_AUTO, doreallocblks, CTLFLAG_RW, &doreallocblks, 0, 488 "enable block reallocation"); 489 490 static int dotrimcons = 1; 491 SYSCTL_INT(_vfs_ffs, OID_AUTO, dotrimcons, CTLFLAG_RWTUN, &dotrimcons, 0, 492 "enable BIO_DELETE / TRIM consolidation"); 493 494 static int maxclustersearch = 10; 495 SYSCTL_INT(_vfs_ffs, OID_AUTO, maxclustersearch, CTLFLAG_RW, &maxclustersearch, 496 0, "max number of cylinder group to search for contigous blocks"); 497 498 #ifdef DIAGNOSTIC 499 static int prtrealloc = 0; 500 SYSCTL_INT(_debug, OID_AUTO, ffs_prtrealloc, CTLFLAG_RW, &prtrealloc, 0, 501 "print out FFS filesystem block reallocation operations"); 502 #endif 503 504 int 505 ffs_reallocblks(ap) 506 struct vop_reallocblks_args /* { 507 struct vnode *a_vp; 508 struct cluster_save *a_buflist; 509 } */ *ap; 510 { 511 struct ufsmount *ump; 512 513 /* 514 * We used to skip reallocating the blocks of a file into a 515 * contiguous sequence if the underlying flash device requested 516 * BIO_DELETE notifications, because devices that benefit from 517 * BIO_DELETE also benefit from not moving the data. However, 518 * the destination for the data is usually moved before the data 519 * is written to the initially allocated location, so we rarely 520 * suffer the penalty of extra writes. With the addition of the 521 * consolidation of contiguous blocks into single BIO_DELETE 522 * operations, having fewer but larger contiguous blocks reduces 523 * the number of (slow and expensive) BIO_DELETE operations. So 524 * when doing BIO_DELETE consolidation, we do block reallocation. 525 * 526 * Skip if reallocblks has been disabled globally. 527 */ 528 ump = ap->a_vp->v_mount->mnt_data; 529 if ((((ump->um_flags) & UM_CANDELETE) != 0 && dotrimcons == 0) || 530 doreallocblks == 0) 531 return (ENOSPC); 532 533 /* 534 * We can't wait in softdep prealloc as it may fsync and recurse 535 * here. Instead we simply fail to reallocate blocks if this 536 * rare condition arises. 537 */ 538 if (DOINGSOFTDEP(ap->a_vp)) 539 if (softdep_prealloc(ap->a_vp, MNT_NOWAIT) != 0) 540 return (ENOSPC); 541 if (ump->um_fstype == UFS1) 542 return (ffs_reallocblks_ufs1(ap)); 543 return (ffs_reallocblks_ufs2(ap)); 544 } 545 546 static int 547 ffs_reallocblks_ufs1(ap) 548 struct vop_reallocblks_args /* { 549 struct vnode *a_vp; 550 struct cluster_save *a_buflist; 551 } */ *ap; 552 { 553 struct fs *fs; 554 struct inode *ip; 555 struct vnode *vp; 556 struct buf *sbp, *ebp, *bp; 557 ufs1_daddr_t *bap, *sbap, *ebap; 558 struct cluster_save *buflist; 559 struct ufsmount *ump; 560 ufs_lbn_t start_lbn, end_lbn; 561 ufs1_daddr_t soff, newblk, blkno; 562 ufs2_daddr_t pref; 563 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp; 564 int i, cg, len, start_lvl, end_lvl, ssize; 565 566 vp = ap->a_vp; 567 ip = VTOI(vp); 568 ump = ITOUMP(ip); 569 fs = ump->um_fs; 570 /* 571 * If we are not tracking block clusters or if we have less than 4% 572 * free blocks left, then do not attempt to cluster. Running with 573 * less than 5% free block reserve is not recommended and those that 574 * choose to do so do not expect to have good file layout. 575 */ 576 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0) 577 return (ENOSPC); 578 buflist = ap->a_buflist; 579 len = buflist->bs_nchildren; 580 start_lbn = buflist->bs_children[0]->b_lblkno; 581 end_lbn = start_lbn + len - 1; 582 #ifdef INVARIANTS 583 for (i = 0; i < len; i++) 584 if (!ffs_checkblk(ip, 585 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 586 panic("ffs_reallocblks: unallocated block 1"); 587 for (i = 1; i < len; i++) 588 if (buflist->bs_children[i]->b_lblkno != start_lbn + i) 589 panic("ffs_reallocblks: non-logical cluster"); 590 blkno = buflist->bs_children[0]->b_blkno; 591 ssize = fsbtodb(fs, fs->fs_frag); 592 for (i = 1; i < len - 1; i++) 593 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize)) 594 panic("ffs_reallocblks: non-physical cluster %d", i); 595 #endif 596 /* 597 * If the cluster crosses the boundary for the first indirect 598 * block, leave space for the indirect block. Indirect blocks 599 * are initially laid out in a position after the last direct 600 * block. Block reallocation would usually destroy locality by 601 * moving the indirect block out of the way to make room for 602 * data blocks if we didn't compensate here. We should also do 603 * this for other indirect block boundaries, but it is only 604 * important for the first one. 605 */ 606 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR) 607 return (ENOSPC); 608 /* 609 * If the latest allocation is in a new cylinder group, assume that 610 * the filesystem has decided to move and do not force it back to 611 * the previous cylinder group. 612 */ 613 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) != 614 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno))) 615 return (ENOSPC); 616 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) || 617 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl)) 618 return (ENOSPC); 619 /* 620 * Get the starting offset and block map for the first block. 621 */ 622 if (start_lvl == 0) { 623 sbap = &ip->i_din1->di_db[0]; 624 soff = start_lbn; 625 } else { 626 idp = &start_ap[start_lvl - 1]; 627 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) { 628 brelse(sbp); 629 return (ENOSPC); 630 } 631 sbap = (ufs1_daddr_t *)sbp->b_data; 632 soff = idp->in_off; 633 } 634 /* 635 * If the block range spans two block maps, get the second map. 636 */ 637 ebap = NULL; 638 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) { 639 ssize = len; 640 } else { 641 #ifdef INVARIANTS 642 if (start_lvl > 0 && 643 start_ap[start_lvl - 1].in_lbn == idp->in_lbn) 644 panic("ffs_reallocblk: start == end"); 645 #endif 646 ssize = len - (idp->in_off + 1); 647 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp)) 648 goto fail; 649 ebap = (ufs1_daddr_t *)ebp->b_data; 650 } 651 /* 652 * Find the preferred location for the cluster. If we have not 653 * previously failed at this endeavor, then follow our standard 654 * preference calculation. If we have failed at it, then pick up 655 * where we last ended our search. 656 */ 657 UFS_LOCK(ump); 658 if (ip->i_nextclustercg == -1) 659 pref = ffs_blkpref_ufs1(ip, start_lbn, soff, sbap); 660 else 661 pref = cgdata(fs, ip->i_nextclustercg); 662 /* 663 * Search the block map looking for an allocation of the desired size. 664 * To avoid wasting too much time, we limit the number of cylinder 665 * groups that we will search. 666 */ 667 cg = dtog(fs, pref); 668 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) { 669 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0) 670 break; 671 cg += 1; 672 if (cg >= fs->fs_ncg) 673 cg = 0; 674 } 675 /* 676 * If we have failed in our search, record where we gave up for 677 * next time. Otherwise, fall back to our usual search citerion. 678 */ 679 if (newblk == 0) { 680 ip->i_nextclustercg = cg; 681 UFS_UNLOCK(ump); 682 goto fail; 683 } 684 ip->i_nextclustercg = -1; 685 /* 686 * We have found a new contiguous block. 687 * 688 * First we have to replace the old block pointers with the new 689 * block pointers in the inode and indirect blocks associated 690 * with the file. 691 */ 692 #ifdef DIAGNOSTIC 693 if (prtrealloc) 694 printf("realloc: ino %ju, lbns %jd-%jd\n\told:", 695 (uintmax_t)ip->i_number, 696 (intmax_t)start_lbn, (intmax_t)end_lbn); 697 #endif 698 blkno = newblk; 699 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) { 700 if (i == ssize) { 701 bap = ebap; 702 soff = -i; 703 } 704 #ifdef INVARIANTS 705 if (!ffs_checkblk(ip, 706 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 707 panic("ffs_reallocblks: unallocated block 2"); 708 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap) 709 panic("ffs_reallocblks: alloc mismatch"); 710 #endif 711 #ifdef DIAGNOSTIC 712 if (prtrealloc) 713 printf(" %d,", *bap); 714 #endif 715 if (DOINGSOFTDEP(vp)) { 716 if (sbap == &ip->i_din1->di_db[0] && i < ssize) 717 softdep_setup_allocdirect(ip, start_lbn + i, 718 blkno, *bap, fs->fs_bsize, fs->fs_bsize, 719 buflist->bs_children[i]); 720 else 721 softdep_setup_allocindir_page(ip, start_lbn + i, 722 i < ssize ? sbp : ebp, soff + i, blkno, 723 *bap, buflist->bs_children[i]); 724 } 725 *bap++ = blkno; 726 } 727 /* 728 * Next we must write out the modified inode and indirect blocks. 729 * For strict correctness, the writes should be synchronous since 730 * the old block values may have been written to disk. In practise 731 * they are almost never written, but if we are concerned about 732 * strict correctness, the `doasyncfree' flag should be set to zero. 733 * 734 * The test on `doasyncfree' should be changed to test a flag 735 * that shows whether the associated buffers and inodes have 736 * been written. The flag should be set when the cluster is 737 * started and cleared whenever the buffer or inode is flushed. 738 * We can then check below to see if it is set, and do the 739 * synchronous write only when it has been cleared. 740 */ 741 if (sbap != &ip->i_din1->di_db[0]) { 742 if (doasyncfree) 743 bdwrite(sbp); 744 else 745 bwrite(sbp); 746 } else { 747 ip->i_flag |= IN_CHANGE | IN_UPDATE; 748 if (!doasyncfree) 749 ffs_update(vp, 1); 750 } 751 if (ssize < len) { 752 if (doasyncfree) 753 bdwrite(ebp); 754 else 755 bwrite(ebp); 756 } 757 /* 758 * Last, free the old blocks and assign the new blocks to the buffers. 759 */ 760 #ifdef DIAGNOSTIC 761 if (prtrealloc) 762 printf("\n\tnew:"); 763 #endif 764 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) { 765 bp = buflist->bs_children[i]; 766 if (!DOINGSOFTDEP(vp)) 767 /* 768 * The usual case is that a set of N-contiguous blocks 769 * that was just allocated has been replaced with a 770 * set of N+1-contiguous blocks. If they are marked as 771 * B_DELWRI, the current contents have not been written 772 * to disk. It is possible that the blocks were written 773 * earlier, but very uncommon. If the blocks have never 774 * been written, there is no need to send a BIO_DELETE 775 * for them when they are freed. The gain from avoiding 776 * the TRIMs for the common case of unwritten blocks 777 * far exceeds the cost of the write amplification for 778 * the uncommon case of failing to send a TRIM for the 779 * blocks that had been written. 780 */ 781 ffs_blkfree(ump, fs, ump->um_devvp, 782 dbtofsb(fs, bp->b_blkno), 783 fs->fs_bsize, ip->i_number, vp->v_type, NULL, 784 (bp->b_flags & B_DELWRI) != 0 ? 785 NOTRIM_KEY : SINGLETON_KEY); 786 bp->b_blkno = fsbtodb(fs, blkno); 787 #ifdef INVARIANTS 788 if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize)) 789 panic("ffs_reallocblks: unallocated block 3"); 790 #endif 791 #ifdef DIAGNOSTIC 792 if (prtrealloc) 793 printf(" %d,", blkno); 794 #endif 795 } 796 #ifdef DIAGNOSTIC 797 if (prtrealloc) { 798 prtrealloc--; 799 printf("\n"); 800 } 801 #endif 802 return (0); 803 804 fail: 805 if (ssize < len) 806 brelse(ebp); 807 if (sbap != &ip->i_din1->di_db[0]) 808 brelse(sbp); 809 return (ENOSPC); 810 } 811 812 static int 813 ffs_reallocblks_ufs2(ap) 814 struct vop_reallocblks_args /* { 815 struct vnode *a_vp; 816 struct cluster_save *a_buflist; 817 } */ *ap; 818 { 819 struct fs *fs; 820 struct inode *ip; 821 struct vnode *vp; 822 struct buf *sbp, *ebp, *bp; 823 ufs2_daddr_t *bap, *sbap, *ebap; 824 struct cluster_save *buflist; 825 struct ufsmount *ump; 826 ufs_lbn_t start_lbn, end_lbn; 827 ufs2_daddr_t soff, newblk, blkno, pref; 828 struct indir start_ap[UFS_NIADDR + 1], end_ap[UFS_NIADDR + 1], *idp; 829 int i, cg, len, start_lvl, end_lvl, ssize; 830 831 vp = ap->a_vp; 832 ip = VTOI(vp); 833 ump = ITOUMP(ip); 834 fs = ump->um_fs; 835 /* 836 * If we are not tracking block clusters or if we have less than 4% 837 * free blocks left, then do not attempt to cluster. Running with 838 * less than 5% free block reserve is not recommended and those that 839 * choose to do so do not expect to have good file layout. 840 */ 841 if (fs->fs_contigsumsize <= 0 || freespace(fs, 4) < 0) 842 return (ENOSPC); 843 buflist = ap->a_buflist; 844 len = buflist->bs_nchildren; 845 start_lbn = buflist->bs_children[0]->b_lblkno; 846 end_lbn = start_lbn + len - 1; 847 #ifdef INVARIANTS 848 for (i = 0; i < len; i++) 849 if (!ffs_checkblk(ip, 850 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 851 panic("ffs_reallocblks: unallocated block 1"); 852 for (i = 1; i < len; i++) 853 if (buflist->bs_children[i]->b_lblkno != start_lbn + i) 854 panic("ffs_reallocblks: non-logical cluster"); 855 blkno = buflist->bs_children[0]->b_blkno; 856 ssize = fsbtodb(fs, fs->fs_frag); 857 for (i = 1; i < len - 1; i++) 858 if (buflist->bs_children[i]->b_blkno != blkno + (i * ssize)) 859 panic("ffs_reallocblks: non-physical cluster %d", i); 860 #endif 861 /* 862 * If the cluster crosses the boundary for the first indirect 863 * block, do not move anything in it. Indirect blocks are 864 * usually initially laid out in a position between the data 865 * blocks. Block reallocation would usually destroy locality by 866 * moving the indirect block out of the way to make room for 867 * data blocks if we didn't compensate here. We should also do 868 * this for other indirect block boundaries, but it is only 869 * important for the first one. 870 */ 871 if (start_lbn < UFS_NDADDR && end_lbn >= UFS_NDADDR) 872 return (ENOSPC); 873 /* 874 * If the latest allocation is in a new cylinder group, assume that 875 * the filesystem has decided to move and do not force it back to 876 * the previous cylinder group. 877 */ 878 if (dtog(fs, dbtofsb(fs, buflist->bs_children[0]->b_blkno)) != 879 dtog(fs, dbtofsb(fs, buflist->bs_children[len - 1]->b_blkno))) 880 return (ENOSPC); 881 if (ufs_getlbns(vp, start_lbn, start_ap, &start_lvl) || 882 ufs_getlbns(vp, end_lbn, end_ap, &end_lvl)) 883 return (ENOSPC); 884 /* 885 * Get the starting offset and block map for the first block. 886 */ 887 if (start_lvl == 0) { 888 sbap = &ip->i_din2->di_db[0]; 889 soff = start_lbn; 890 } else { 891 idp = &start_ap[start_lvl - 1]; 892 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &sbp)) { 893 brelse(sbp); 894 return (ENOSPC); 895 } 896 sbap = (ufs2_daddr_t *)sbp->b_data; 897 soff = idp->in_off; 898 } 899 /* 900 * If the block range spans two block maps, get the second map. 901 */ 902 ebap = NULL; 903 if (end_lvl == 0 || (idp = &end_ap[end_lvl - 1])->in_off + 1 >= len) { 904 ssize = len; 905 } else { 906 #ifdef INVARIANTS 907 if (start_lvl > 0 && 908 start_ap[start_lvl - 1].in_lbn == idp->in_lbn) 909 panic("ffs_reallocblk: start == end"); 910 #endif 911 ssize = len - (idp->in_off + 1); 912 if (bread(vp, idp->in_lbn, (int)fs->fs_bsize, NOCRED, &ebp)) 913 goto fail; 914 ebap = (ufs2_daddr_t *)ebp->b_data; 915 } 916 /* 917 * Find the preferred location for the cluster. If we have not 918 * previously failed at this endeavor, then follow our standard 919 * preference calculation. If we have failed at it, then pick up 920 * where we last ended our search. 921 */ 922 UFS_LOCK(ump); 923 if (ip->i_nextclustercg == -1) 924 pref = ffs_blkpref_ufs2(ip, start_lbn, soff, sbap); 925 else 926 pref = cgdata(fs, ip->i_nextclustercg); 927 /* 928 * Search the block map looking for an allocation of the desired size. 929 * To avoid wasting too much time, we limit the number of cylinder 930 * groups that we will search. 931 */ 932 cg = dtog(fs, pref); 933 for (i = min(maxclustersearch, fs->fs_ncg); i > 0; i--) { 934 if ((newblk = ffs_clusteralloc(ip, cg, pref, len)) != 0) 935 break; 936 cg += 1; 937 if (cg >= fs->fs_ncg) 938 cg = 0; 939 } 940 /* 941 * If we have failed in our search, record where we gave up for 942 * next time. Otherwise, fall back to our usual search citerion. 943 */ 944 if (newblk == 0) { 945 ip->i_nextclustercg = cg; 946 UFS_UNLOCK(ump); 947 goto fail; 948 } 949 ip->i_nextclustercg = -1; 950 /* 951 * We have found a new contiguous block. 952 * 953 * First we have to replace the old block pointers with the new 954 * block pointers in the inode and indirect blocks associated 955 * with the file. 956 */ 957 #ifdef DIAGNOSTIC 958 if (prtrealloc) 959 printf("realloc: ino %ju, lbns %jd-%jd\n\told:", (uintmax_t)ip->i_number, 960 (intmax_t)start_lbn, (intmax_t)end_lbn); 961 #endif 962 blkno = newblk; 963 for (bap = &sbap[soff], i = 0; i < len; i++, blkno += fs->fs_frag) { 964 if (i == ssize) { 965 bap = ebap; 966 soff = -i; 967 } 968 #ifdef INVARIANTS 969 if (!ffs_checkblk(ip, 970 dbtofsb(fs, buflist->bs_children[i]->b_blkno), fs->fs_bsize)) 971 panic("ffs_reallocblks: unallocated block 2"); 972 if (dbtofsb(fs, buflist->bs_children[i]->b_blkno) != *bap) 973 panic("ffs_reallocblks: alloc mismatch"); 974 #endif 975 #ifdef DIAGNOSTIC 976 if (prtrealloc) 977 printf(" %jd,", (intmax_t)*bap); 978 #endif 979 if (DOINGSOFTDEP(vp)) { 980 if (sbap == &ip->i_din2->di_db[0] && i < ssize) 981 softdep_setup_allocdirect(ip, start_lbn + i, 982 blkno, *bap, fs->fs_bsize, fs->fs_bsize, 983 buflist->bs_children[i]); 984 else 985 softdep_setup_allocindir_page(ip, start_lbn + i, 986 i < ssize ? sbp : ebp, soff + i, blkno, 987 *bap, buflist->bs_children[i]); 988 } 989 *bap++ = blkno; 990 } 991 /* 992 * Next we must write out the modified inode and indirect blocks. 993 * For strict correctness, the writes should be synchronous since 994 * the old block values may have been written to disk. In practise 995 * they are almost never written, but if we are concerned about 996 * strict correctness, the `doasyncfree' flag should be set to zero. 997 * 998 * The test on `doasyncfree' should be changed to test a flag 999 * that shows whether the associated buffers and inodes have 1000 * been written. The flag should be set when the cluster is 1001 * started and cleared whenever the buffer or inode is flushed. 1002 * We can then check below to see if it is set, and do the 1003 * synchronous write only when it has been cleared. 1004 */ 1005 if (sbap != &ip->i_din2->di_db[0]) { 1006 if (doasyncfree) 1007 bdwrite(sbp); 1008 else 1009 bwrite(sbp); 1010 } else { 1011 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1012 if (!doasyncfree) 1013 ffs_update(vp, 1); 1014 } 1015 if (ssize < len) { 1016 if (doasyncfree) 1017 bdwrite(ebp); 1018 else 1019 bwrite(ebp); 1020 } 1021 /* 1022 * Last, free the old blocks and assign the new blocks to the buffers. 1023 */ 1024 #ifdef DIAGNOSTIC 1025 if (prtrealloc) 1026 printf("\n\tnew:"); 1027 #endif 1028 for (blkno = newblk, i = 0; i < len; i++, blkno += fs->fs_frag) { 1029 bp = buflist->bs_children[i]; 1030 if (!DOINGSOFTDEP(vp)) 1031 /* 1032 * The usual case is that a set of N-contiguous blocks 1033 * that was just allocated has been replaced with a 1034 * set of N+1-contiguous blocks. If they are marked as 1035 * B_DELWRI, the current contents have not been written 1036 * to disk. It is possible that the blocks were written 1037 * earlier, but very uncommon. If the blocks have never 1038 * been written, there is no need to send a BIO_DELETE 1039 * for them when they are freed. The gain from avoiding 1040 * the TRIMs for the common case of unwritten blocks 1041 * far exceeds the cost of the write amplification for 1042 * the uncommon case of failing to send a TRIM for the 1043 * blocks that had been written. 1044 */ 1045 ffs_blkfree(ump, fs, ump->um_devvp, 1046 dbtofsb(fs, bp->b_blkno), 1047 fs->fs_bsize, ip->i_number, vp->v_type, NULL, 1048 (bp->b_flags & B_DELWRI) != 0 ? 1049 NOTRIM_KEY : SINGLETON_KEY); 1050 bp->b_blkno = fsbtodb(fs, blkno); 1051 #ifdef INVARIANTS 1052 if (!ffs_checkblk(ip, dbtofsb(fs, bp->b_blkno), fs->fs_bsize)) 1053 panic("ffs_reallocblks: unallocated block 3"); 1054 #endif 1055 #ifdef DIAGNOSTIC 1056 if (prtrealloc) 1057 printf(" %jd,", (intmax_t)blkno); 1058 #endif 1059 } 1060 #ifdef DIAGNOSTIC 1061 if (prtrealloc) { 1062 prtrealloc--; 1063 printf("\n"); 1064 } 1065 #endif 1066 return (0); 1067 1068 fail: 1069 if (ssize < len) 1070 brelse(ebp); 1071 if (sbap != &ip->i_din2->di_db[0]) 1072 brelse(sbp); 1073 return (ENOSPC); 1074 } 1075 1076 /* 1077 * Allocate an inode in the filesystem. 1078 * 1079 * If allocating a directory, use ffs_dirpref to select the inode. 1080 * If allocating in a directory, the following hierarchy is followed: 1081 * 1) allocate the preferred inode. 1082 * 2) allocate an inode in the same cylinder group. 1083 * 3) quadradically rehash into other cylinder groups, until an 1084 * available inode is located. 1085 * If no inode preference is given the following hierarchy is used 1086 * to allocate an inode: 1087 * 1) allocate an inode in cylinder group 0. 1088 * 2) quadradically rehash into other cylinder groups, until an 1089 * available inode is located. 1090 */ 1091 int 1092 ffs_valloc(pvp, mode, cred, vpp) 1093 struct vnode *pvp; 1094 int mode; 1095 struct ucred *cred; 1096 struct vnode **vpp; 1097 { 1098 struct inode *pip; 1099 struct fs *fs; 1100 struct inode *ip; 1101 struct timespec ts; 1102 struct ufsmount *ump; 1103 ino_t ino, ipref; 1104 u_int cg; 1105 int error, error1, reclaimed; 1106 1107 *vpp = NULL; 1108 pip = VTOI(pvp); 1109 ump = ITOUMP(pip); 1110 fs = ump->um_fs; 1111 1112 UFS_LOCK(ump); 1113 reclaimed = 0; 1114 retry: 1115 if (fs->fs_cstotal.cs_nifree == 0) 1116 goto noinodes; 1117 1118 if ((mode & IFMT) == IFDIR) 1119 ipref = ffs_dirpref(pip); 1120 else 1121 ipref = pip->i_number; 1122 if (ipref >= fs->fs_ncg * fs->fs_ipg) 1123 ipref = 0; 1124 cg = ino_to_cg(fs, ipref); 1125 /* 1126 * Track number of dirs created one after another 1127 * in a same cg without intervening by files. 1128 */ 1129 if ((mode & IFMT) == IFDIR) { 1130 if (fs->fs_contigdirs[cg] < 255) 1131 fs->fs_contigdirs[cg]++; 1132 } else { 1133 if (fs->fs_contigdirs[cg] > 0) 1134 fs->fs_contigdirs[cg]--; 1135 } 1136 ino = (ino_t)ffs_hashalloc(pip, cg, ipref, mode, 0, 1137 (allocfcn_t *)ffs_nodealloccg); 1138 if (ino == 0) 1139 goto noinodes; 1140 1141 /* 1142 * Get rid of the cached old vnode, force allocation of a new vnode 1143 * for this inode. 1144 */ 1145 error = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp, FFSV_REPLACE); 1146 if (error) { 1147 error1 = ffs_vgetf(pvp->v_mount, ino, LK_EXCLUSIVE, vpp, 1148 FFSV_FORCEINSMQ | FFSV_REPLACE); 1149 ffs_vfree(pvp, ino, mode); 1150 if (error1 == 0) { 1151 ip = VTOI(*vpp); 1152 if (ip->i_mode) 1153 goto dup_alloc; 1154 ip->i_flag |= IN_MODIFIED; 1155 vput(*vpp); 1156 } 1157 return (error); 1158 } 1159 ip = VTOI(*vpp); 1160 if (ip->i_mode) { 1161 dup_alloc: 1162 printf("mode = 0%o, inum = %ju, fs = %s\n", 1163 ip->i_mode, (uintmax_t)ip->i_number, fs->fs_fsmnt); 1164 panic("ffs_valloc: dup alloc"); 1165 } 1166 if (DIP(ip, i_blocks) && (fs->fs_flags & FS_UNCLEAN) == 0) { /* XXX */ 1167 printf("free inode %s/%lu had %ld blocks\n", 1168 fs->fs_fsmnt, (u_long)ino, (long)DIP(ip, i_blocks)); 1169 DIP_SET(ip, i_blocks, 0); 1170 } 1171 ip->i_flags = 0; 1172 DIP_SET(ip, i_flags, 0); 1173 /* 1174 * Set up a new generation number for this inode. 1175 */ 1176 while (ip->i_gen == 0 || ++ip->i_gen == 0) 1177 ip->i_gen = arc4random(); 1178 DIP_SET(ip, i_gen, ip->i_gen); 1179 if (fs->fs_magic == FS_UFS2_MAGIC) { 1180 vfs_timestamp(&ts); 1181 ip->i_din2->di_birthtime = ts.tv_sec; 1182 ip->i_din2->di_birthnsec = ts.tv_nsec; 1183 } 1184 ip->i_flag = 0; 1185 (*vpp)->v_vflag = 0; 1186 (*vpp)->v_type = VNON; 1187 if (fs->fs_magic == FS_UFS2_MAGIC) { 1188 (*vpp)->v_op = &ffs_vnodeops2; 1189 ip->i_flag |= IN_UFS2; 1190 } else { 1191 (*vpp)->v_op = &ffs_vnodeops1; 1192 } 1193 return (0); 1194 noinodes: 1195 if (reclaimed == 0) { 1196 reclaimed = 1; 1197 softdep_request_cleanup(fs, pvp, cred, FLUSH_INODES_WAIT); 1198 goto retry; 1199 } 1200 if (ppsratecheck(&ump->um_last_fullmsg, &ump->um_secs_fullmsg, 1)) { 1201 UFS_UNLOCK(ump); 1202 ffs_fserr(fs, pip->i_number, "out of inodes"); 1203 uprintf("\n%s: create/symlink failed, no inodes free\n", 1204 fs->fs_fsmnt); 1205 } else { 1206 UFS_UNLOCK(ump); 1207 } 1208 return (ENOSPC); 1209 } 1210 1211 /* 1212 * Find a cylinder group to place a directory. 1213 * 1214 * The policy implemented by this algorithm is to allocate a 1215 * directory inode in the same cylinder group as its parent 1216 * directory, but also to reserve space for its files inodes 1217 * and data. Restrict the number of directories which may be 1218 * allocated one after another in the same cylinder group 1219 * without intervening allocation of files. 1220 * 1221 * If we allocate a first level directory then force allocation 1222 * in another cylinder group. 1223 */ 1224 static ino_t 1225 ffs_dirpref(pip) 1226 struct inode *pip; 1227 { 1228 struct fs *fs; 1229 int cg, prefcg, dirsize, cgsize; 1230 u_int avgifree, avgbfree, avgndir, curdirsize; 1231 u_int minifree, minbfree, maxndir; 1232 u_int mincg, minndir; 1233 u_int maxcontigdirs; 1234 1235 mtx_assert(UFS_MTX(ITOUMP(pip)), MA_OWNED); 1236 fs = ITOFS(pip); 1237 1238 avgifree = fs->fs_cstotal.cs_nifree / fs->fs_ncg; 1239 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1240 avgndir = fs->fs_cstotal.cs_ndir / fs->fs_ncg; 1241 1242 /* 1243 * Force allocation in another cg if creating a first level dir. 1244 */ 1245 ASSERT_VOP_LOCKED(ITOV(pip), "ffs_dirpref"); 1246 if (ITOV(pip)->v_vflag & VV_ROOT) { 1247 prefcg = arc4random() % fs->fs_ncg; 1248 mincg = prefcg; 1249 minndir = fs->fs_ipg; 1250 for (cg = prefcg; cg < fs->fs_ncg; cg++) 1251 if (fs->fs_cs(fs, cg).cs_ndir < minndir && 1252 fs->fs_cs(fs, cg).cs_nifree >= avgifree && 1253 fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1254 mincg = cg; 1255 minndir = fs->fs_cs(fs, cg).cs_ndir; 1256 } 1257 for (cg = 0; cg < prefcg; cg++) 1258 if (fs->fs_cs(fs, cg).cs_ndir < minndir && 1259 fs->fs_cs(fs, cg).cs_nifree >= avgifree && 1260 fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1261 mincg = cg; 1262 minndir = fs->fs_cs(fs, cg).cs_ndir; 1263 } 1264 return ((ino_t)(fs->fs_ipg * mincg)); 1265 } 1266 1267 /* 1268 * Count various limits which used for 1269 * optimal allocation of a directory inode. 1270 */ 1271 maxndir = min(avgndir + fs->fs_ipg / 16, fs->fs_ipg); 1272 minifree = avgifree - avgifree / 4; 1273 if (minifree < 1) 1274 minifree = 1; 1275 minbfree = avgbfree - avgbfree / 4; 1276 if (minbfree < 1) 1277 minbfree = 1; 1278 cgsize = fs->fs_fsize * fs->fs_fpg; 1279 dirsize = fs->fs_avgfilesize * fs->fs_avgfpdir; 1280 curdirsize = avgndir ? (cgsize - avgbfree * fs->fs_bsize) / avgndir : 0; 1281 if (dirsize < curdirsize) 1282 dirsize = curdirsize; 1283 if (dirsize <= 0) 1284 maxcontigdirs = 0; /* dirsize overflowed */ 1285 else 1286 maxcontigdirs = min((avgbfree * fs->fs_bsize) / dirsize, 255); 1287 if (fs->fs_avgfpdir > 0) 1288 maxcontigdirs = min(maxcontigdirs, 1289 fs->fs_ipg / fs->fs_avgfpdir); 1290 if (maxcontigdirs == 0) 1291 maxcontigdirs = 1; 1292 1293 /* 1294 * Limit number of dirs in one cg and reserve space for 1295 * regular files, but only if we have no deficit in 1296 * inodes or space. 1297 * 1298 * We are trying to find a suitable cylinder group nearby 1299 * our preferred cylinder group to place a new directory. 1300 * We scan from our preferred cylinder group forward looking 1301 * for a cylinder group that meets our criterion. If we get 1302 * to the final cylinder group and do not find anything, 1303 * we start scanning forwards from the beginning of the 1304 * filesystem. While it might seem sensible to start scanning 1305 * backwards or even to alternate looking forward and backward, 1306 * this approach fails badly when the filesystem is nearly full. 1307 * Specifically, we first search all the areas that have no space 1308 * and finally try the one preceding that. We repeat this on 1309 * every request and in the case of the final block end up 1310 * searching the entire filesystem. By jumping to the front 1311 * of the filesystem, our future forward searches always look 1312 * in new cylinder groups so finds every possible block after 1313 * one pass over the filesystem. 1314 */ 1315 prefcg = ino_to_cg(fs, pip->i_number); 1316 for (cg = prefcg; cg < fs->fs_ncg; cg++) 1317 if (fs->fs_cs(fs, cg).cs_ndir < maxndir && 1318 fs->fs_cs(fs, cg).cs_nifree >= minifree && 1319 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) { 1320 if (fs->fs_contigdirs[cg] < maxcontigdirs) 1321 return ((ino_t)(fs->fs_ipg * cg)); 1322 } 1323 for (cg = 0; cg < prefcg; cg++) 1324 if (fs->fs_cs(fs, cg).cs_ndir < maxndir && 1325 fs->fs_cs(fs, cg).cs_nifree >= minifree && 1326 fs->fs_cs(fs, cg).cs_nbfree >= minbfree) { 1327 if (fs->fs_contigdirs[cg] < maxcontigdirs) 1328 return ((ino_t)(fs->fs_ipg * cg)); 1329 } 1330 /* 1331 * This is a backstop when we have deficit in space. 1332 */ 1333 for (cg = prefcg; cg < fs->fs_ncg; cg++) 1334 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree) 1335 return ((ino_t)(fs->fs_ipg * cg)); 1336 for (cg = 0; cg < prefcg; cg++) 1337 if (fs->fs_cs(fs, cg).cs_nifree >= avgifree) 1338 break; 1339 return ((ino_t)(fs->fs_ipg * cg)); 1340 } 1341 1342 /* 1343 * Select the desired position for the next block in a file. The file is 1344 * logically divided into sections. The first section is composed of the 1345 * direct blocks and the next fs_maxbpg blocks. Each additional section 1346 * contains fs_maxbpg blocks. 1347 * 1348 * If no blocks have been allocated in the first section, the policy is to 1349 * request a block in the same cylinder group as the inode that describes 1350 * the file. The first indirect is allocated immediately following the last 1351 * direct block and the data blocks for the first indirect immediately 1352 * follow it. 1353 * 1354 * If no blocks have been allocated in any other section, the indirect 1355 * block(s) are allocated in the same cylinder group as its inode in an 1356 * area reserved immediately following the inode blocks. The policy for 1357 * the data blocks is to place them in a cylinder group with a greater than 1358 * average number of free blocks. An appropriate cylinder group is found 1359 * by using a rotor that sweeps the cylinder groups. When a new group of 1360 * blocks is needed, the sweep begins in the cylinder group following the 1361 * cylinder group from which the previous allocation was made. The sweep 1362 * continues until a cylinder group with greater than the average number 1363 * of free blocks is found. If the allocation is for the first block in an 1364 * indirect block or the previous block is a hole, then the information on 1365 * the previous allocation is unavailable; here a best guess is made based 1366 * on the logical block number being allocated. 1367 * 1368 * If a section is already partially allocated, the policy is to 1369 * allocate blocks contiguously within the section if possible. 1370 */ 1371 ufs2_daddr_t 1372 ffs_blkpref_ufs1(ip, lbn, indx, bap) 1373 struct inode *ip; 1374 ufs_lbn_t lbn; 1375 int indx; 1376 ufs1_daddr_t *bap; 1377 { 1378 struct fs *fs; 1379 u_int cg, inocg; 1380 u_int avgbfree, startcg; 1381 ufs2_daddr_t pref, prevbn; 1382 1383 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap")); 1384 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1385 fs = ITOFS(ip); 1386 /* 1387 * Allocation of indirect blocks is indicated by passing negative 1388 * values in indx: -1 for single indirect, -2 for double indirect, 1389 * -3 for triple indirect. As noted below, we attempt to allocate 1390 * the first indirect inline with the file data. For all later 1391 * indirect blocks, the data is often allocated in other cylinder 1392 * groups. However to speed random file access and to speed up 1393 * fsck, the filesystem reserves the first fs_metaspace blocks 1394 * (typically half of fs_minfree) of the data area of each cylinder 1395 * group to hold these later indirect blocks. 1396 */ 1397 inocg = ino_to_cg(fs, ip->i_number); 1398 if (indx < 0) { 1399 /* 1400 * Our preference for indirect blocks is the zone at the 1401 * beginning of the inode's cylinder group data area that 1402 * we try to reserve for indirect blocks. 1403 */ 1404 pref = cgmeta(fs, inocg); 1405 /* 1406 * If we are allocating the first indirect block, try to 1407 * place it immediately following the last direct block. 1408 */ 1409 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) && 1410 ip->i_din1->di_db[UFS_NDADDR - 1] != 0) 1411 pref = ip->i_din1->di_db[UFS_NDADDR - 1] + fs->fs_frag; 1412 return (pref); 1413 } 1414 /* 1415 * If we are allocating the first data block in the first indirect 1416 * block and the indirect has been allocated in the data block area, 1417 * try to place it immediately following the indirect block. 1418 */ 1419 if (lbn == UFS_NDADDR) { 1420 pref = ip->i_din1->di_ib[0]; 1421 if (pref != 0 && pref >= cgdata(fs, inocg) && 1422 pref < cgbase(fs, inocg + 1)) 1423 return (pref + fs->fs_frag); 1424 } 1425 /* 1426 * If we are at the beginning of a file, or we have already allocated 1427 * the maximum number of blocks per cylinder group, or we do not 1428 * have a block allocated immediately preceding us, then we need 1429 * to decide where to start allocating new blocks. 1430 */ 1431 if (indx == 0) { 1432 prevbn = 0; 1433 } else { 1434 prevbn = bap[indx - 1]; 1435 if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn, 1436 fs->fs_bsize) != 0) 1437 prevbn = 0; 1438 } 1439 if (indx % fs->fs_maxbpg == 0 || prevbn == 0) { 1440 /* 1441 * If we are allocating a directory data block, we want 1442 * to place it in the metadata area. 1443 */ 1444 if ((ip->i_mode & IFMT) == IFDIR) 1445 return (cgmeta(fs, inocg)); 1446 /* 1447 * Until we fill all the direct and all the first indirect's 1448 * blocks, we try to allocate in the data area of the inode's 1449 * cylinder group. 1450 */ 1451 if (lbn < UFS_NDADDR + NINDIR(fs)) 1452 return (cgdata(fs, inocg)); 1453 /* 1454 * Find a cylinder with greater than average number of 1455 * unused data blocks. 1456 */ 1457 if (indx == 0 || prevbn == 0) 1458 startcg = inocg + lbn / fs->fs_maxbpg; 1459 else 1460 startcg = dtog(fs, prevbn) + 1; 1461 startcg %= fs->fs_ncg; 1462 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1463 for (cg = startcg; cg < fs->fs_ncg; cg++) 1464 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1465 fs->fs_cgrotor = cg; 1466 return (cgdata(fs, cg)); 1467 } 1468 for (cg = 0; cg <= startcg; cg++) 1469 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1470 fs->fs_cgrotor = cg; 1471 return (cgdata(fs, cg)); 1472 } 1473 return (0); 1474 } 1475 /* 1476 * Otherwise, we just always try to lay things out contiguously. 1477 */ 1478 return (prevbn + fs->fs_frag); 1479 } 1480 1481 /* 1482 * Same as above, but for UFS2 1483 */ 1484 ufs2_daddr_t 1485 ffs_blkpref_ufs2(ip, lbn, indx, bap) 1486 struct inode *ip; 1487 ufs_lbn_t lbn; 1488 int indx; 1489 ufs2_daddr_t *bap; 1490 { 1491 struct fs *fs; 1492 u_int cg, inocg; 1493 u_int avgbfree, startcg; 1494 ufs2_daddr_t pref, prevbn; 1495 1496 KASSERT(indx <= 0 || bap != NULL, ("need non-NULL bap")); 1497 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1498 fs = ITOFS(ip); 1499 /* 1500 * Allocation of indirect blocks is indicated by passing negative 1501 * values in indx: -1 for single indirect, -2 for double indirect, 1502 * -3 for triple indirect. As noted below, we attempt to allocate 1503 * the first indirect inline with the file data. For all later 1504 * indirect blocks, the data is often allocated in other cylinder 1505 * groups. However to speed random file access and to speed up 1506 * fsck, the filesystem reserves the first fs_metaspace blocks 1507 * (typically half of fs_minfree) of the data area of each cylinder 1508 * group to hold these later indirect blocks. 1509 */ 1510 inocg = ino_to_cg(fs, ip->i_number); 1511 if (indx < 0) { 1512 /* 1513 * Our preference for indirect blocks is the zone at the 1514 * beginning of the inode's cylinder group data area that 1515 * we try to reserve for indirect blocks. 1516 */ 1517 pref = cgmeta(fs, inocg); 1518 /* 1519 * If we are allocating the first indirect block, try to 1520 * place it immediately following the last direct block. 1521 */ 1522 if (indx == -1 && lbn < UFS_NDADDR + NINDIR(fs) && 1523 ip->i_din2->di_db[UFS_NDADDR - 1] != 0) 1524 pref = ip->i_din2->di_db[UFS_NDADDR - 1] + fs->fs_frag; 1525 return (pref); 1526 } 1527 /* 1528 * If we are allocating the first data block in the first indirect 1529 * block and the indirect has been allocated in the data block area, 1530 * try to place it immediately following the indirect block. 1531 */ 1532 if (lbn == UFS_NDADDR) { 1533 pref = ip->i_din2->di_ib[0]; 1534 if (pref != 0 && pref >= cgdata(fs, inocg) && 1535 pref < cgbase(fs, inocg + 1)) 1536 return (pref + fs->fs_frag); 1537 } 1538 /* 1539 * If we are at the beginning of a file, or we have already allocated 1540 * the maximum number of blocks per cylinder group, or we do not 1541 * have a block allocated immediately preceding us, then we need 1542 * to decide where to start allocating new blocks. 1543 */ 1544 if (indx == 0) { 1545 prevbn = 0; 1546 } else { 1547 prevbn = bap[indx - 1]; 1548 if (UFS_CHECK_BLKNO(ITOVFS(ip), ip->i_number, prevbn, 1549 fs->fs_bsize) != 0) 1550 prevbn = 0; 1551 } 1552 if (indx % fs->fs_maxbpg == 0 || prevbn == 0) { 1553 /* 1554 * If we are allocating a directory data block, we want 1555 * to place it in the metadata area. 1556 */ 1557 if ((ip->i_mode & IFMT) == IFDIR) 1558 return (cgmeta(fs, inocg)); 1559 /* 1560 * Until we fill all the direct and all the first indirect's 1561 * blocks, we try to allocate in the data area of the inode's 1562 * cylinder group. 1563 */ 1564 if (lbn < UFS_NDADDR + NINDIR(fs)) 1565 return (cgdata(fs, inocg)); 1566 /* 1567 * Find a cylinder with greater than average number of 1568 * unused data blocks. 1569 */ 1570 if (indx == 0 || prevbn == 0) 1571 startcg = inocg + lbn / fs->fs_maxbpg; 1572 else 1573 startcg = dtog(fs, prevbn) + 1; 1574 startcg %= fs->fs_ncg; 1575 avgbfree = fs->fs_cstotal.cs_nbfree / fs->fs_ncg; 1576 for (cg = startcg; cg < fs->fs_ncg; cg++) 1577 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1578 fs->fs_cgrotor = cg; 1579 return (cgdata(fs, cg)); 1580 } 1581 for (cg = 0; cg <= startcg; cg++) 1582 if (fs->fs_cs(fs, cg).cs_nbfree >= avgbfree) { 1583 fs->fs_cgrotor = cg; 1584 return (cgdata(fs, cg)); 1585 } 1586 return (0); 1587 } 1588 /* 1589 * Otherwise, we just always try to lay things out contiguously. 1590 */ 1591 return (prevbn + fs->fs_frag); 1592 } 1593 1594 /* 1595 * Implement the cylinder overflow algorithm. 1596 * 1597 * The policy implemented by this algorithm is: 1598 * 1) allocate the block in its requested cylinder group. 1599 * 2) quadradically rehash on the cylinder group number. 1600 * 3) brute force search for a free block. 1601 * 1602 * Must be called with the UFS lock held. Will release the lock on success 1603 * and return with it held on failure. 1604 */ 1605 /*VARARGS5*/ 1606 static ufs2_daddr_t 1607 ffs_hashalloc(ip, cg, pref, size, rsize, allocator) 1608 struct inode *ip; 1609 u_int cg; 1610 ufs2_daddr_t pref; 1611 int size; /* Search size for data blocks, mode for inodes */ 1612 int rsize; /* Real allocated size. */ 1613 allocfcn_t *allocator; 1614 { 1615 struct fs *fs; 1616 ufs2_daddr_t result; 1617 u_int i, icg = cg; 1618 1619 mtx_assert(UFS_MTX(ITOUMP(ip)), MA_OWNED); 1620 #ifdef INVARIANTS 1621 if (ITOV(ip)->v_mount->mnt_kern_flag & MNTK_SUSPENDED) 1622 panic("ffs_hashalloc: allocation on suspended filesystem"); 1623 #endif 1624 fs = ITOFS(ip); 1625 /* 1626 * 1: preferred cylinder group 1627 */ 1628 result = (*allocator)(ip, cg, pref, size, rsize); 1629 if (result) 1630 return (result); 1631 /* 1632 * 2: quadratic rehash 1633 */ 1634 for (i = 1; i < fs->fs_ncg; i *= 2) { 1635 cg += i; 1636 if (cg >= fs->fs_ncg) 1637 cg -= fs->fs_ncg; 1638 result = (*allocator)(ip, cg, 0, size, rsize); 1639 if (result) 1640 return (result); 1641 } 1642 /* 1643 * 3: brute force search 1644 * Note that we start at i == 2, since 0 was checked initially, 1645 * and 1 is always checked in the quadratic rehash. 1646 */ 1647 cg = (icg + 2) % fs->fs_ncg; 1648 for (i = 2; i < fs->fs_ncg; i++) { 1649 result = (*allocator)(ip, cg, 0, size, rsize); 1650 if (result) 1651 return (result); 1652 cg++; 1653 if (cg == fs->fs_ncg) 1654 cg = 0; 1655 } 1656 return (0); 1657 } 1658 1659 /* 1660 * Determine whether a fragment can be extended. 1661 * 1662 * Check to see if the necessary fragments are available, and 1663 * if they are, allocate them. 1664 */ 1665 static ufs2_daddr_t 1666 ffs_fragextend(ip, cg, bprev, osize, nsize) 1667 struct inode *ip; 1668 u_int cg; 1669 ufs2_daddr_t bprev; 1670 int osize, nsize; 1671 { 1672 struct fs *fs; 1673 struct cg *cgp; 1674 struct buf *bp; 1675 struct ufsmount *ump; 1676 int nffree; 1677 long bno; 1678 int frags, bbase; 1679 int i, error; 1680 u_int8_t *blksfree; 1681 1682 ump = ITOUMP(ip); 1683 fs = ump->um_fs; 1684 if (fs->fs_cs(fs, cg).cs_nffree < numfrags(fs, nsize - osize)) 1685 return (0); 1686 frags = numfrags(fs, nsize); 1687 bbase = fragnum(fs, bprev); 1688 if (bbase > fragnum(fs, (bprev + frags - 1))) { 1689 /* cannot extend across a block boundary */ 1690 return (0); 1691 } 1692 UFS_UNLOCK(ump); 1693 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) 1694 goto fail; 1695 bno = dtogd(fs, bprev); 1696 blksfree = cg_blksfree(cgp); 1697 for (i = numfrags(fs, osize); i < frags; i++) 1698 if (isclr(blksfree, bno + i)) 1699 goto fail; 1700 /* 1701 * the current fragment can be extended 1702 * deduct the count on fragment being extended into 1703 * increase the count on the remaining fragment (if any) 1704 * allocate the extended piece 1705 */ 1706 for (i = frags; i < fs->fs_frag - bbase; i++) 1707 if (isclr(blksfree, bno + i)) 1708 break; 1709 cgp->cg_frsum[i - numfrags(fs, osize)]--; 1710 if (i != frags) 1711 cgp->cg_frsum[i - frags]++; 1712 for (i = numfrags(fs, osize), nffree = 0; i < frags; i++) { 1713 clrbit(blksfree, bno + i); 1714 cgp->cg_cs.cs_nffree--; 1715 nffree++; 1716 } 1717 UFS_LOCK(ump); 1718 fs->fs_cstotal.cs_nffree -= nffree; 1719 fs->fs_cs(fs, cg).cs_nffree -= nffree; 1720 fs->fs_fmod = 1; 1721 ACTIVECLEAR(fs, cg); 1722 UFS_UNLOCK(ump); 1723 if (DOINGSOFTDEP(ITOV(ip))) 1724 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), bprev, 1725 frags, numfrags(fs, osize)); 1726 bdwrite(bp); 1727 return (bprev); 1728 1729 fail: 1730 brelse(bp); 1731 UFS_LOCK(ump); 1732 return (0); 1733 1734 } 1735 1736 /* 1737 * Determine whether a block can be allocated. 1738 * 1739 * Check to see if a block of the appropriate size is available, 1740 * and if it is, allocate it. 1741 */ 1742 static ufs2_daddr_t 1743 ffs_alloccg(ip, cg, bpref, size, rsize) 1744 struct inode *ip; 1745 u_int cg; 1746 ufs2_daddr_t bpref; 1747 int size; 1748 int rsize; 1749 { 1750 struct fs *fs; 1751 struct cg *cgp; 1752 struct buf *bp; 1753 struct ufsmount *ump; 1754 ufs1_daddr_t bno; 1755 ufs2_daddr_t blkno; 1756 int i, allocsiz, error, frags; 1757 u_int8_t *blksfree; 1758 1759 ump = ITOUMP(ip); 1760 fs = ump->um_fs; 1761 if (fs->fs_cs(fs, cg).cs_nbfree == 0 && size == fs->fs_bsize) 1762 return (0); 1763 UFS_UNLOCK(ump); 1764 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0 || 1765 (cgp->cg_cs.cs_nbfree == 0 && size == fs->fs_bsize)) 1766 goto fail; 1767 if (size == fs->fs_bsize) { 1768 UFS_LOCK(ump); 1769 blkno = ffs_alloccgblk(ip, bp, bpref, rsize); 1770 ACTIVECLEAR(fs, cg); 1771 UFS_UNLOCK(ump); 1772 bdwrite(bp); 1773 return (blkno); 1774 } 1775 /* 1776 * check to see if any fragments are already available 1777 * allocsiz is the size which will be allocated, hacking 1778 * it down to a smaller size if necessary 1779 */ 1780 blksfree = cg_blksfree(cgp); 1781 frags = numfrags(fs, size); 1782 for (allocsiz = frags; allocsiz < fs->fs_frag; allocsiz++) 1783 if (cgp->cg_frsum[allocsiz] != 0) 1784 break; 1785 if (allocsiz == fs->fs_frag) { 1786 /* 1787 * no fragments were available, so a block will be 1788 * allocated, and hacked up 1789 */ 1790 if (cgp->cg_cs.cs_nbfree == 0) 1791 goto fail; 1792 UFS_LOCK(ump); 1793 blkno = ffs_alloccgblk(ip, bp, bpref, rsize); 1794 ACTIVECLEAR(fs, cg); 1795 UFS_UNLOCK(ump); 1796 bdwrite(bp); 1797 return (blkno); 1798 } 1799 KASSERT(size == rsize, 1800 ("ffs_alloccg: size(%d) != rsize(%d)", size, rsize)); 1801 bno = ffs_mapsearch(fs, cgp, bpref, allocsiz); 1802 if (bno < 0) 1803 goto fail; 1804 for (i = 0; i < frags; i++) 1805 clrbit(blksfree, bno + i); 1806 cgp->cg_cs.cs_nffree -= frags; 1807 cgp->cg_frsum[allocsiz]--; 1808 if (frags != allocsiz) 1809 cgp->cg_frsum[allocsiz - frags]++; 1810 UFS_LOCK(ump); 1811 fs->fs_cstotal.cs_nffree -= frags; 1812 fs->fs_cs(fs, cg).cs_nffree -= frags; 1813 fs->fs_fmod = 1; 1814 blkno = cgbase(fs, cg) + bno; 1815 ACTIVECLEAR(fs, cg); 1816 UFS_UNLOCK(ump); 1817 if (DOINGSOFTDEP(ITOV(ip))) 1818 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, frags, 0); 1819 bdwrite(bp); 1820 return (blkno); 1821 1822 fail: 1823 brelse(bp); 1824 UFS_LOCK(ump); 1825 return (0); 1826 } 1827 1828 /* 1829 * Allocate a block in a cylinder group. 1830 * 1831 * This algorithm implements the following policy: 1832 * 1) allocate the requested block. 1833 * 2) allocate a rotationally optimal block in the same cylinder. 1834 * 3) allocate the next available block on the block rotor for the 1835 * specified cylinder group. 1836 * Note that this routine only allocates fs_bsize blocks; these 1837 * blocks may be fragmented by the routine that allocates them. 1838 */ 1839 static ufs2_daddr_t 1840 ffs_alloccgblk(ip, bp, bpref, size) 1841 struct inode *ip; 1842 struct buf *bp; 1843 ufs2_daddr_t bpref; 1844 int size; 1845 { 1846 struct fs *fs; 1847 struct cg *cgp; 1848 struct ufsmount *ump; 1849 ufs1_daddr_t bno; 1850 ufs2_daddr_t blkno; 1851 u_int8_t *blksfree; 1852 int i, cgbpref; 1853 1854 ump = ITOUMP(ip); 1855 fs = ump->um_fs; 1856 mtx_assert(UFS_MTX(ump), MA_OWNED); 1857 cgp = (struct cg *)bp->b_data; 1858 blksfree = cg_blksfree(cgp); 1859 if (bpref == 0) { 1860 bpref = cgbase(fs, cgp->cg_cgx) + cgp->cg_rotor + fs->fs_frag; 1861 } else if ((cgbpref = dtog(fs, bpref)) != cgp->cg_cgx) { 1862 /* map bpref to correct zone in this cg */ 1863 if (bpref < cgdata(fs, cgbpref)) 1864 bpref = cgmeta(fs, cgp->cg_cgx); 1865 else 1866 bpref = cgdata(fs, cgp->cg_cgx); 1867 } 1868 /* 1869 * if the requested block is available, use it 1870 */ 1871 bno = dtogd(fs, blknum(fs, bpref)); 1872 if (ffs_isblock(fs, blksfree, fragstoblks(fs, bno))) 1873 goto gotit; 1874 /* 1875 * Take the next available block in this cylinder group. 1876 */ 1877 bno = ffs_mapsearch(fs, cgp, bpref, (int)fs->fs_frag); 1878 if (bno < 0) 1879 return (0); 1880 /* Update cg_rotor only if allocated from the data zone */ 1881 if (bno >= dtogd(fs, cgdata(fs, cgp->cg_cgx))) 1882 cgp->cg_rotor = bno; 1883 gotit: 1884 blkno = fragstoblks(fs, bno); 1885 ffs_clrblock(fs, blksfree, (long)blkno); 1886 ffs_clusteracct(fs, cgp, blkno, -1); 1887 cgp->cg_cs.cs_nbfree--; 1888 fs->fs_cstotal.cs_nbfree--; 1889 fs->fs_cs(fs, cgp->cg_cgx).cs_nbfree--; 1890 fs->fs_fmod = 1; 1891 blkno = cgbase(fs, cgp->cg_cgx) + bno; 1892 /* 1893 * If the caller didn't want the whole block free the frags here. 1894 */ 1895 size = numfrags(fs, size); 1896 if (size != fs->fs_frag) { 1897 bno = dtogd(fs, blkno); 1898 for (i = size; i < fs->fs_frag; i++) 1899 setbit(blksfree, bno + i); 1900 i = fs->fs_frag - size; 1901 cgp->cg_cs.cs_nffree += i; 1902 fs->fs_cstotal.cs_nffree += i; 1903 fs->fs_cs(fs, cgp->cg_cgx).cs_nffree += i; 1904 fs->fs_fmod = 1; 1905 cgp->cg_frsum[i]++; 1906 } 1907 /* XXX Fixme. */ 1908 UFS_UNLOCK(ump); 1909 if (DOINGSOFTDEP(ITOV(ip))) 1910 softdep_setup_blkmapdep(bp, UFSTOVFS(ump), blkno, size, 0); 1911 UFS_LOCK(ump); 1912 return (blkno); 1913 } 1914 1915 /* 1916 * Determine whether a cluster can be allocated. 1917 * 1918 * We do not currently check for optimal rotational layout if there 1919 * are multiple choices in the same cylinder group. Instead we just 1920 * take the first one that we find following bpref. 1921 */ 1922 static ufs2_daddr_t 1923 ffs_clusteralloc(ip, cg, bpref, len) 1924 struct inode *ip; 1925 u_int cg; 1926 ufs2_daddr_t bpref; 1927 int len; 1928 { 1929 struct fs *fs; 1930 struct cg *cgp; 1931 struct buf *bp; 1932 struct ufsmount *ump; 1933 int i, run, bit, map, got, error; 1934 ufs2_daddr_t bno; 1935 u_char *mapp; 1936 int32_t *lp; 1937 u_int8_t *blksfree; 1938 1939 ump = ITOUMP(ip); 1940 fs = ump->um_fs; 1941 if (fs->fs_maxcluster[cg] < len) 1942 return (0); 1943 UFS_UNLOCK(ump); 1944 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) { 1945 UFS_LOCK(ump); 1946 return (0); 1947 } 1948 /* 1949 * Check to see if a cluster of the needed size (or bigger) is 1950 * available in this cylinder group. 1951 */ 1952 lp = &cg_clustersum(cgp)[len]; 1953 for (i = len; i <= fs->fs_contigsumsize; i++) 1954 if (*lp++ > 0) 1955 break; 1956 if (i > fs->fs_contigsumsize) { 1957 /* 1958 * This is the first time looking for a cluster in this 1959 * cylinder group. Update the cluster summary information 1960 * to reflect the true maximum sized cluster so that 1961 * future cluster allocation requests can avoid reading 1962 * the cylinder group map only to find no clusters. 1963 */ 1964 lp = &cg_clustersum(cgp)[len - 1]; 1965 for (i = len - 1; i > 0; i--) 1966 if (*lp-- > 0) 1967 break; 1968 UFS_LOCK(ump); 1969 fs->fs_maxcluster[cg] = i; 1970 brelse(bp); 1971 return (0); 1972 } 1973 /* 1974 * Search the cluster map to find a big enough cluster. 1975 * We take the first one that we find, even if it is larger 1976 * than we need as we prefer to get one close to the previous 1977 * block allocation. We do not search before the current 1978 * preference point as we do not want to allocate a block 1979 * that is allocated before the previous one (as we will 1980 * then have to wait for another pass of the elevator 1981 * algorithm before it will be read). We prefer to fail and 1982 * be recalled to try an allocation in the next cylinder group. 1983 */ 1984 if (dtog(fs, bpref) != cg) 1985 bpref = cgdata(fs, cg); 1986 else 1987 bpref = blknum(fs, bpref); 1988 bpref = fragstoblks(fs, dtogd(fs, bpref)); 1989 mapp = &cg_clustersfree(cgp)[bpref / NBBY]; 1990 map = *mapp++; 1991 bit = 1 << (bpref % NBBY); 1992 for (run = 0, got = bpref; got < cgp->cg_nclusterblks; got++) { 1993 if ((map & bit) == 0) { 1994 run = 0; 1995 } else { 1996 run++; 1997 if (run == len) 1998 break; 1999 } 2000 if ((got & (NBBY - 1)) != (NBBY - 1)) { 2001 bit <<= 1; 2002 } else { 2003 map = *mapp++; 2004 bit = 1; 2005 } 2006 } 2007 if (got >= cgp->cg_nclusterblks) { 2008 UFS_LOCK(ump); 2009 brelse(bp); 2010 return (0); 2011 } 2012 /* 2013 * Allocate the cluster that we have found. 2014 */ 2015 blksfree = cg_blksfree(cgp); 2016 for (i = 1; i <= len; i++) 2017 if (!ffs_isblock(fs, blksfree, got - run + i)) 2018 panic("ffs_clusteralloc: map mismatch"); 2019 bno = cgbase(fs, cg) + blkstofrags(fs, got - run + 1); 2020 if (dtog(fs, bno) != cg) 2021 panic("ffs_clusteralloc: allocated out of group"); 2022 len = blkstofrags(fs, len); 2023 UFS_LOCK(ump); 2024 for (i = 0; i < len; i += fs->fs_frag) 2025 if (ffs_alloccgblk(ip, bp, bno + i, fs->fs_bsize) != bno + i) 2026 panic("ffs_clusteralloc: lost block"); 2027 ACTIVECLEAR(fs, cg); 2028 UFS_UNLOCK(ump); 2029 bdwrite(bp); 2030 return (bno); 2031 } 2032 2033 static inline struct buf * 2034 getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags) 2035 { 2036 struct fs *fs; 2037 2038 fs = ITOFS(ip); 2039 return (getblk(ITODEVVP(ip), fsbtodb(fs, ino_to_fsba(fs, 2040 cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0, 2041 gbflags)); 2042 } 2043 2044 /* 2045 * Synchronous inode initialization is needed only when barrier writes do not 2046 * work as advertised, and will impose a heavy cost on file creation in a newly 2047 * created filesystem. 2048 */ 2049 static int doasyncinodeinit = 1; 2050 SYSCTL_INT(_vfs_ffs, OID_AUTO, doasyncinodeinit, CTLFLAG_RWTUN, 2051 &doasyncinodeinit, 0, 2052 "Perform inode block initialization using asynchronous writes"); 2053 2054 /* 2055 * Determine whether an inode can be allocated. 2056 * 2057 * Check to see if an inode is available, and if it is, 2058 * allocate it using the following policy: 2059 * 1) allocate the requested inode. 2060 * 2) allocate the next available inode after the requested 2061 * inode in the specified cylinder group. 2062 */ 2063 static ufs2_daddr_t 2064 ffs_nodealloccg(ip, cg, ipref, mode, unused) 2065 struct inode *ip; 2066 u_int cg; 2067 ufs2_daddr_t ipref; 2068 int mode; 2069 int unused; 2070 { 2071 struct fs *fs; 2072 struct cg *cgp; 2073 struct buf *bp, *ibp; 2074 struct ufsmount *ump; 2075 u_int8_t *inosused, *loc; 2076 struct ufs2_dinode *dp2; 2077 int error, start, len, i; 2078 u_int32_t old_initediblk; 2079 2080 ump = ITOUMP(ip); 2081 fs = ump->um_fs; 2082 check_nifree: 2083 if (fs->fs_cs(fs, cg).cs_nifree == 0) 2084 return (0); 2085 UFS_UNLOCK(ump); 2086 if ((error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp)) != 0) { 2087 UFS_LOCK(ump); 2088 return (0); 2089 } 2090 restart: 2091 if (cgp->cg_cs.cs_nifree == 0) { 2092 brelse(bp); 2093 UFS_LOCK(ump); 2094 return (0); 2095 } 2096 inosused = cg_inosused(cgp); 2097 if (ipref) { 2098 ipref %= fs->fs_ipg; 2099 if (isclr(inosused, ipref)) 2100 goto gotit; 2101 } 2102 start = cgp->cg_irotor / NBBY; 2103 len = howmany(fs->fs_ipg - cgp->cg_irotor, NBBY); 2104 loc = memcchr(&inosused[start], 0xff, len); 2105 if (loc == NULL) { 2106 len = start + 1; 2107 start = 0; 2108 loc = memcchr(&inosused[start], 0xff, len); 2109 if (loc == NULL) { 2110 printf("cg = %d, irotor = %ld, fs = %s\n", 2111 cg, (long)cgp->cg_irotor, fs->fs_fsmnt); 2112 panic("ffs_nodealloccg: map corrupted"); 2113 /* NOTREACHED */ 2114 } 2115 } 2116 ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1; 2117 gotit: 2118 /* 2119 * Check to see if we need to initialize more inodes. 2120 */ 2121 if (fs->fs_magic == FS_UFS2_MAGIC && 2122 ipref + INOPB(fs) > cgp->cg_initediblk && 2123 cgp->cg_initediblk < cgp->cg_niblk) { 2124 old_initediblk = cgp->cg_initediblk; 2125 2126 /* 2127 * Free the cylinder group lock before writing the 2128 * initialized inode block. Entering the 2129 * babarrierwrite() with the cylinder group lock 2130 * causes lock order violation between the lock and 2131 * snaplk. 2132 * 2133 * Another thread can decide to initialize the same 2134 * inode block, but whichever thread first gets the 2135 * cylinder group lock after writing the newly 2136 * allocated inode block will update it and the other 2137 * will realize that it has lost and leave the 2138 * cylinder group unchanged. 2139 */ 2140 ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT); 2141 brelse(bp); 2142 if (ibp == NULL) { 2143 /* 2144 * The inode block buffer is already owned by 2145 * another thread, which must initialize it. 2146 * Wait on the buffer to allow another thread 2147 * to finish the updates, with dropped cg 2148 * buffer lock, then retry. 2149 */ 2150 ibp = getinobuf(ip, cg, old_initediblk, 0); 2151 brelse(ibp); 2152 UFS_LOCK(ump); 2153 goto check_nifree; 2154 } 2155 bzero(ibp->b_data, (int)fs->fs_bsize); 2156 dp2 = (struct ufs2_dinode *)(ibp->b_data); 2157 for (i = 0; i < INOPB(fs); i++) { 2158 while (dp2->di_gen == 0) 2159 dp2->di_gen = arc4random(); 2160 dp2++; 2161 } 2162 2163 /* 2164 * Rather than adding a soft updates dependency to ensure 2165 * that the new inode block is written before it is claimed 2166 * by the cylinder group map, we just do a barrier write 2167 * here. The barrier write will ensure that the inode block 2168 * gets written before the updated cylinder group map can be 2169 * written. The barrier write should only slow down bulk 2170 * loading of newly created filesystems. 2171 */ 2172 if (doasyncinodeinit) 2173 babarrierwrite(ibp); 2174 else 2175 bwrite(ibp); 2176 2177 /* 2178 * After the inode block is written, try to update the 2179 * cg initediblk pointer. If another thread beat us 2180 * to it, then leave it unchanged as the other thread 2181 * has already set it correctly. 2182 */ 2183 error = ffs_getcg(fs, ump->um_devvp, cg, &bp, &cgp); 2184 UFS_LOCK(ump); 2185 ACTIVECLEAR(fs, cg); 2186 UFS_UNLOCK(ump); 2187 if (error != 0) 2188 return (error); 2189 if (cgp->cg_initediblk == old_initediblk) 2190 cgp->cg_initediblk += INOPB(fs); 2191 goto restart; 2192 } 2193 cgp->cg_irotor = ipref; 2194 UFS_LOCK(ump); 2195 ACTIVECLEAR(fs, cg); 2196 setbit(inosused, ipref); 2197 cgp->cg_cs.cs_nifree--; 2198 fs->fs_cstotal.cs_nifree--; 2199 fs->fs_cs(fs, cg).cs_nifree--; 2200 fs->fs_fmod = 1; 2201 if ((mode & IFMT) == IFDIR) { 2202 cgp->cg_cs.cs_ndir++; 2203 fs->fs_cstotal.cs_ndir++; 2204 fs->fs_cs(fs, cg).cs_ndir++; 2205 } 2206 UFS_UNLOCK(ump); 2207 if (DOINGSOFTDEP(ITOV(ip))) 2208 softdep_setup_inomapdep(bp, ip, cg * fs->fs_ipg + ipref, mode); 2209 bdwrite(bp); 2210 return ((ino_t)(cg * fs->fs_ipg + ipref)); 2211 } 2212 2213 /* 2214 * Free a block or fragment. 2215 * 2216 * The specified block or fragment is placed back in the 2217 * free map. If a fragment is deallocated, a possible 2218 * block reassembly is checked. 2219 */ 2220 static void 2221 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd) 2222 struct ufsmount *ump; 2223 struct fs *fs; 2224 struct vnode *devvp; 2225 ufs2_daddr_t bno; 2226 long size; 2227 ino_t inum; 2228 struct workhead *dephd; 2229 { 2230 struct mount *mp; 2231 struct cg *cgp; 2232 struct buf *bp; 2233 ufs1_daddr_t fragno, cgbno; 2234 int i, blk, frags, bbase, error; 2235 u_int cg; 2236 u_int8_t *blksfree; 2237 struct cdev *dev; 2238 2239 cg = dtog(fs, bno); 2240 if (devvp->v_type == VREG) { 2241 /* devvp is a snapshot */ 2242 MPASS(devvp->v_mount->mnt_data == ump); 2243 dev = ump->um_devvp->v_rdev; 2244 } else if (devvp->v_type == VCHR) { 2245 /* devvp is a normal disk device */ 2246 dev = devvp->v_rdev; 2247 ASSERT_VOP_LOCKED(devvp, "ffs_blkfree_cg"); 2248 } else 2249 return; 2250 #ifdef INVARIANTS 2251 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0 || 2252 fragnum(fs, bno) + numfrags(fs, size) > fs->fs_frag) { 2253 printf("dev=%s, bno = %jd, bsize = %ld, size = %ld, fs = %s\n", 2254 devtoname(dev), (intmax_t)bno, (long)fs->fs_bsize, 2255 size, fs->fs_fsmnt); 2256 panic("ffs_blkfree_cg: bad size"); 2257 } 2258 #endif 2259 if ((u_int)bno >= fs->fs_size) { 2260 printf("bad block %jd, ino %lu\n", (intmax_t)bno, 2261 (u_long)inum); 2262 ffs_fserr(fs, inum, "bad block"); 2263 return; 2264 } 2265 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0) 2266 return; 2267 cgbno = dtogd(fs, bno); 2268 blksfree = cg_blksfree(cgp); 2269 UFS_LOCK(ump); 2270 if (size == fs->fs_bsize) { 2271 fragno = fragstoblks(fs, cgbno); 2272 if (!ffs_isfreeblock(fs, blksfree, fragno)) { 2273 if (devvp->v_type == VREG) { 2274 UFS_UNLOCK(ump); 2275 /* devvp is a snapshot */ 2276 brelse(bp); 2277 return; 2278 } 2279 printf("dev = %s, block = %jd, fs = %s\n", 2280 devtoname(dev), (intmax_t)bno, fs->fs_fsmnt); 2281 panic("ffs_blkfree_cg: freeing free block"); 2282 } 2283 ffs_setblock(fs, blksfree, fragno); 2284 ffs_clusteracct(fs, cgp, fragno, 1); 2285 cgp->cg_cs.cs_nbfree++; 2286 fs->fs_cstotal.cs_nbfree++; 2287 fs->fs_cs(fs, cg).cs_nbfree++; 2288 } else { 2289 bbase = cgbno - fragnum(fs, cgbno); 2290 /* 2291 * decrement the counts associated with the old frags 2292 */ 2293 blk = blkmap(fs, blksfree, bbase); 2294 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 2295 /* 2296 * deallocate the fragment 2297 */ 2298 frags = numfrags(fs, size); 2299 for (i = 0; i < frags; i++) { 2300 if (isset(blksfree, cgbno + i)) { 2301 printf("dev = %s, block = %jd, fs = %s\n", 2302 devtoname(dev), (intmax_t)(bno + i), 2303 fs->fs_fsmnt); 2304 panic("ffs_blkfree_cg: freeing free frag"); 2305 } 2306 setbit(blksfree, cgbno + i); 2307 } 2308 cgp->cg_cs.cs_nffree += i; 2309 fs->fs_cstotal.cs_nffree += i; 2310 fs->fs_cs(fs, cg).cs_nffree += i; 2311 /* 2312 * add back in counts associated with the new frags 2313 */ 2314 blk = blkmap(fs, blksfree, bbase); 2315 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 2316 /* 2317 * if a complete block has been reassembled, account for it 2318 */ 2319 fragno = fragstoblks(fs, bbase); 2320 if (ffs_isblock(fs, blksfree, fragno)) { 2321 cgp->cg_cs.cs_nffree -= fs->fs_frag; 2322 fs->fs_cstotal.cs_nffree -= fs->fs_frag; 2323 fs->fs_cs(fs, cg).cs_nffree -= fs->fs_frag; 2324 ffs_clusteracct(fs, cgp, fragno, 1); 2325 cgp->cg_cs.cs_nbfree++; 2326 fs->fs_cstotal.cs_nbfree++; 2327 fs->fs_cs(fs, cg).cs_nbfree++; 2328 } 2329 } 2330 fs->fs_fmod = 1; 2331 ACTIVECLEAR(fs, cg); 2332 UFS_UNLOCK(ump); 2333 mp = UFSTOVFS(ump); 2334 if (MOUNTEDSOFTDEP(mp) && devvp->v_type == VCHR) 2335 softdep_setup_blkfree(UFSTOVFS(ump), bp, bno, 2336 numfrags(fs, size), dephd); 2337 bdwrite(bp); 2338 } 2339 2340 /* 2341 * Structures and routines associated with trim management. 2342 * 2343 * The following requests are passed to trim_lookup to indicate 2344 * the actions that should be taken. 2345 */ 2346 #define NEW 1 /* if found, error else allocate and hash it */ 2347 #define OLD 2 /* if not found, error, else return it */ 2348 #define REPLACE 3 /* if not found, error else unhash and reallocate it */ 2349 #define DONE 4 /* if not found, error else unhash and return it */ 2350 #define SINGLE 5 /* don't look up, just allocate it and don't hash it */ 2351 2352 MALLOC_DEFINE(M_TRIM, "ufs_trim", "UFS trim structures"); 2353 2354 #define TRIMLIST_HASH(ump, key) \ 2355 (&(ump)->um_trimhash[(key) & (ump)->um_trimlisthashsize]) 2356 2357 /* 2358 * These structures describe each of the block free requests aggregated 2359 * together to make up a trim request. 2360 */ 2361 struct trim_blkreq { 2362 TAILQ_ENTRY(trim_blkreq) blkreqlist; 2363 ufs2_daddr_t bno; 2364 long size; 2365 struct workhead *pdephd; 2366 struct workhead dephd; 2367 }; 2368 2369 /* 2370 * Description of a trim request. 2371 */ 2372 struct ffs_blkfree_trim_params { 2373 TAILQ_HEAD(, trim_blkreq) blklist; 2374 LIST_ENTRY(ffs_blkfree_trim_params) hashlist; 2375 struct task task; 2376 struct ufsmount *ump; 2377 struct vnode *devvp; 2378 ino_t inum; 2379 ufs2_daddr_t bno; 2380 long size; 2381 long key; 2382 }; 2383 2384 static void ffs_blkfree_trim_completed(struct buf *); 2385 static void ffs_blkfree_trim_task(void *ctx, int pending __unused); 2386 static struct ffs_blkfree_trim_params *trim_lookup(struct ufsmount *, 2387 struct vnode *, ufs2_daddr_t, long, ino_t, u_long, int); 2388 static void ffs_blkfree_sendtrim(struct ffs_blkfree_trim_params *); 2389 2390 /* 2391 * Called on trim completion to start a task to free the associated block(s). 2392 */ 2393 static void 2394 ffs_blkfree_trim_completed(bp) 2395 struct buf *bp; 2396 { 2397 struct ffs_blkfree_trim_params *tp; 2398 2399 tp = bp->b_fsprivate1; 2400 free(bp, M_TRIM); 2401 TASK_INIT(&tp->task, 0, ffs_blkfree_trim_task, tp); 2402 taskqueue_enqueue(tp->ump->um_trim_tq, &tp->task); 2403 } 2404 2405 /* 2406 * Trim completion task that free associated block(s). 2407 */ 2408 static void 2409 ffs_blkfree_trim_task(ctx, pending) 2410 void *ctx; 2411 int pending; 2412 { 2413 struct ffs_blkfree_trim_params *tp; 2414 struct trim_blkreq *blkelm; 2415 struct ufsmount *ump; 2416 2417 tp = ctx; 2418 ump = tp->ump; 2419 while ((blkelm = TAILQ_FIRST(&tp->blklist)) != NULL) { 2420 ffs_blkfree_cg(ump, ump->um_fs, tp->devvp, blkelm->bno, 2421 blkelm->size, tp->inum, blkelm->pdephd); 2422 TAILQ_REMOVE(&tp->blklist, blkelm, blkreqlist); 2423 free(blkelm, M_TRIM); 2424 } 2425 vn_finished_secondary_write(UFSTOVFS(ump)); 2426 UFS_LOCK(ump); 2427 ump->um_trim_inflight -= 1; 2428 ump->um_trim_inflight_blks -= numfrags(ump->um_fs, tp->size); 2429 UFS_UNLOCK(ump); 2430 free(tp, M_TRIM); 2431 } 2432 2433 /* 2434 * Lookup a trim request by inode number. 2435 * Allocate if requested (NEW, REPLACE, SINGLE). 2436 */ 2437 static struct ffs_blkfree_trim_params * 2438 trim_lookup(ump, devvp, bno, size, inum, key, alloctype) 2439 struct ufsmount *ump; 2440 struct vnode *devvp; 2441 ufs2_daddr_t bno; 2442 long size; 2443 ino_t inum; 2444 u_long key; 2445 int alloctype; 2446 { 2447 struct trimlist_hashhead *tphashhead; 2448 struct ffs_blkfree_trim_params *tp, *ntp; 2449 2450 ntp = malloc(sizeof(struct ffs_blkfree_trim_params), M_TRIM, M_WAITOK); 2451 if (alloctype != SINGLE) { 2452 KASSERT(key >= FIRST_VALID_KEY, ("trim_lookup: invalid key")); 2453 UFS_LOCK(ump); 2454 tphashhead = TRIMLIST_HASH(ump, key); 2455 LIST_FOREACH(tp, tphashhead, hashlist) 2456 if (key == tp->key) 2457 break; 2458 } 2459 switch (alloctype) { 2460 case NEW: 2461 KASSERT(tp == NULL, ("trim_lookup: found trim")); 2462 break; 2463 case OLD: 2464 KASSERT(tp != NULL, 2465 ("trim_lookup: missing call to ffs_blkrelease_start()")); 2466 UFS_UNLOCK(ump); 2467 free(ntp, M_TRIM); 2468 return (tp); 2469 case REPLACE: 2470 KASSERT(tp != NULL, ("trim_lookup: missing REPLACE trim")); 2471 LIST_REMOVE(tp, hashlist); 2472 /* tp will be freed by caller */ 2473 break; 2474 case DONE: 2475 KASSERT(tp != NULL, ("trim_lookup: missing DONE trim")); 2476 LIST_REMOVE(tp, hashlist); 2477 UFS_UNLOCK(ump); 2478 free(ntp, M_TRIM); 2479 return (tp); 2480 } 2481 TAILQ_INIT(&ntp->blklist); 2482 ntp->ump = ump; 2483 ntp->devvp = devvp; 2484 ntp->bno = bno; 2485 ntp->size = size; 2486 ntp->inum = inum; 2487 ntp->key = key; 2488 if (alloctype != SINGLE) { 2489 LIST_INSERT_HEAD(tphashhead, ntp, hashlist); 2490 UFS_UNLOCK(ump); 2491 } 2492 return (ntp); 2493 } 2494 2495 /* 2496 * Dispatch a trim request. 2497 */ 2498 static void 2499 ffs_blkfree_sendtrim(tp) 2500 struct ffs_blkfree_trim_params *tp; 2501 { 2502 struct ufsmount *ump; 2503 struct mount *mp; 2504 struct buf *bp; 2505 2506 /* 2507 * Postpone the set of the free bit in the cg bitmap until the 2508 * BIO_DELETE is completed. Otherwise, due to disk queue 2509 * reordering, TRIM might be issued after we reuse the block 2510 * and write some new data into it. 2511 */ 2512 ump = tp->ump; 2513 bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO); 2514 bp->b_iocmd = BIO_DELETE; 2515 bp->b_iooffset = dbtob(fsbtodb(ump->um_fs, tp->bno)); 2516 bp->b_iodone = ffs_blkfree_trim_completed; 2517 bp->b_bcount = tp->size; 2518 bp->b_fsprivate1 = tp; 2519 UFS_LOCK(ump); 2520 ump->um_trim_total += 1; 2521 ump->um_trim_inflight += 1; 2522 ump->um_trim_inflight_blks += numfrags(ump->um_fs, tp->size); 2523 ump->um_trim_total_blks += numfrags(ump->um_fs, tp->size); 2524 UFS_UNLOCK(ump); 2525 2526 mp = UFSTOVFS(ump); 2527 vn_start_secondary_write(NULL, &mp, 0); 2528 g_vfs_strategy(ump->um_bo, bp); 2529 } 2530 2531 /* 2532 * Allocate a new key to use to identify a range of blocks. 2533 */ 2534 u_long 2535 ffs_blkrelease_start(ump, devvp, inum) 2536 struct ufsmount *ump; 2537 struct vnode *devvp; 2538 ino_t inum; 2539 { 2540 static u_long masterkey; 2541 u_long key; 2542 2543 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0) 2544 return (SINGLETON_KEY); 2545 do { 2546 key = atomic_fetchadd_long(&masterkey, 1); 2547 } while (key < FIRST_VALID_KEY); 2548 (void) trim_lookup(ump, devvp, 0, 0, inum, key, NEW); 2549 return (key); 2550 } 2551 2552 /* 2553 * Deallocate a key that has been used to identify a range of blocks. 2554 */ 2555 void 2556 ffs_blkrelease_finish(ump, key) 2557 struct ufsmount *ump; 2558 u_long key; 2559 { 2560 struct ffs_blkfree_trim_params *tp; 2561 2562 if (((ump->um_flags & UM_CANDELETE) == 0) || dotrimcons == 0) 2563 return; 2564 /* 2565 * If the vfs.ffs.dotrimcons sysctl option is enabled while 2566 * a file deletion is active, specifically after a call 2567 * to ffs_blkrelease_start() but before the call to 2568 * ffs_blkrelease_finish(), ffs_blkrelease_start() will 2569 * have handed out SINGLETON_KEY rather than starting a 2570 * collection sequence. Thus if we get a SINGLETON_KEY 2571 * passed to ffs_blkrelease_finish(), we just return rather 2572 * than trying to finish the nonexistent sequence. 2573 */ 2574 if (key == SINGLETON_KEY) { 2575 #ifdef INVARIANTS 2576 printf("%s: vfs.ffs.dotrimcons enabled on active filesystem\n", 2577 ump->um_mountp->mnt_stat.f_mntonname); 2578 #endif 2579 return; 2580 } 2581 /* 2582 * We are done with sending blocks using this key. Look up the key 2583 * using the DONE alloctype (in tp) to request that it be unhashed 2584 * as we will not be adding to it. If the key has never been used, 2585 * tp->size will be zero, so we can just free tp. Otherwise the call 2586 * to ffs_blkfree_sendtrim(tp) causes the block range described by 2587 * tp to be issued (and then tp to be freed). 2588 */ 2589 tp = trim_lookup(ump, NULL, 0, 0, 0, key, DONE); 2590 if (tp->size == 0) 2591 free(tp, M_TRIM); 2592 else 2593 ffs_blkfree_sendtrim(tp); 2594 } 2595 2596 /* 2597 * Setup to free a block or fragment. 2598 * 2599 * Check for snapshots that might want to claim the block. 2600 * If trims are requested, prepare a trim request. Attempt to 2601 * aggregate consecutive blocks into a single trim request. 2602 */ 2603 void 2604 ffs_blkfree(ump, fs, devvp, bno, size, inum, vtype, dephd, key) 2605 struct ufsmount *ump; 2606 struct fs *fs; 2607 struct vnode *devvp; 2608 ufs2_daddr_t bno; 2609 long size; 2610 ino_t inum; 2611 enum vtype vtype; 2612 struct workhead *dephd; 2613 u_long key; 2614 { 2615 struct ffs_blkfree_trim_params *tp, *ntp; 2616 struct trim_blkreq *blkelm; 2617 2618 /* 2619 * Check to see if a snapshot wants to claim the block. 2620 * Check that devvp is a normal disk device, not a snapshot, 2621 * it has a snapshot(s) associated with it, and one of the 2622 * snapshots wants to claim the block. 2623 */ 2624 if (devvp->v_type == VCHR && 2625 (devvp->v_vflag & VV_COPYONWRITE) && 2626 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, dephd)) { 2627 return; 2628 } 2629 /* 2630 * Nothing to delay if TRIM is not required for this block or TRIM 2631 * is disabled or the operation is performed on a snapshot. 2632 */ 2633 if (key == NOTRIM_KEY || ((ump->um_flags & UM_CANDELETE) == 0) || 2634 devvp->v_type == VREG) { 2635 ffs_blkfree_cg(ump, fs, devvp, bno, size, inum, dephd); 2636 return; 2637 } 2638 blkelm = malloc(sizeof(struct trim_blkreq), M_TRIM, M_WAITOK); 2639 blkelm->bno = bno; 2640 blkelm->size = size; 2641 if (dephd == NULL) { 2642 blkelm->pdephd = NULL; 2643 } else { 2644 LIST_INIT(&blkelm->dephd); 2645 LIST_SWAP(dephd, &blkelm->dephd, worklist, wk_list); 2646 blkelm->pdephd = &blkelm->dephd; 2647 } 2648 if (key == SINGLETON_KEY) { 2649 /* 2650 * Just a single non-contiguous piece. Use the SINGLE 2651 * alloctype to return a trim request that will not be 2652 * hashed for future lookup. 2653 */ 2654 tp = trim_lookup(ump, devvp, bno, size, inum, key, SINGLE); 2655 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2656 ffs_blkfree_sendtrim(tp); 2657 return; 2658 } 2659 /* 2660 * The callers of this function are not tracking whether or not 2661 * the blocks are contiguous. They are just saying that they 2662 * are freeing a set of blocks. It is this code that determines 2663 * the pieces of that range that are actually contiguous. 2664 * 2665 * Calling ffs_blkrelease_start() will have created an entry 2666 * that we will use. 2667 */ 2668 tp = trim_lookup(ump, devvp, bno, size, inum, key, OLD); 2669 if (tp->size == 0) { 2670 /* 2671 * First block of a potential range, set block and size 2672 * for the trim block. 2673 */ 2674 tp->bno = bno; 2675 tp->size = size; 2676 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2677 return; 2678 } 2679 /* 2680 * If this block is a continuation of the range (either 2681 * follows at the end or preceeds in the front) then we 2682 * add it to the front or back of the list and return. 2683 * 2684 * If it is not a continuation of the trim that we were 2685 * building, using the REPLACE alloctype, we request that 2686 * the old trim request (still in tp) be unhashed and a 2687 * new range started (in ntp). The ffs_blkfree_sendtrim(tp) 2688 * call causes the block range described by tp to be issued 2689 * (and then tp to be freed). 2690 */ 2691 if (bno + numfrags(fs, size) == tp->bno) { 2692 TAILQ_INSERT_HEAD(&tp->blklist, blkelm, blkreqlist); 2693 tp->bno = bno; 2694 tp->size += size; 2695 return; 2696 } else if (bno == tp->bno + numfrags(fs, tp->size)) { 2697 TAILQ_INSERT_TAIL(&tp->blklist, blkelm, blkreqlist); 2698 tp->size += size; 2699 return; 2700 } 2701 ntp = trim_lookup(ump, devvp, bno, size, inum, key, REPLACE); 2702 TAILQ_INSERT_HEAD(&ntp->blklist, blkelm, blkreqlist); 2703 ffs_blkfree_sendtrim(tp); 2704 } 2705 2706 #ifdef INVARIANTS 2707 /* 2708 * Verify allocation of a block or fragment. Returns true if block or 2709 * fragment is allocated, false if it is free. 2710 */ 2711 static int 2712 ffs_checkblk(ip, bno, size) 2713 struct inode *ip; 2714 ufs2_daddr_t bno; 2715 long size; 2716 { 2717 struct fs *fs; 2718 struct cg *cgp; 2719 struct buf *bp; 2720 ufs1_daddr_t cgbno; 2721 int i, error, frags, free; 2722 u_int8_t *blksfree; 2723 2724 fs = ITOFS(ip); 2725 if ((u_int)size > fs->fs_bsize || fragoff(fs, size) != 0) { 2726 printf("bsize = %ld, size = %ld, fs = %s\n", 2727 (long)fs->fs_bsize, size, fs->fs_fsmnt); 2728 panic("ffs_checkblk: bad size"); 2729 } 2730 if ((u_int)bno >= fs->fs_size) 2731 panic("ffs_checkblk: bad block %jd", (intmax_t)bno); 2732 error = ffs_getcg(fs, ITODEVVP(ip), dtog(fs, bno), &bp, &cgp); 2733 if (error) 2734 panic("ffs_checkblk: cylinder group read failed"); 2735 blksfree = cg_blksfree(cgp); 2736 cgbno = dtogd(fs, bno); 2737 if (size == fs->fs_bsize) { 2738 free = ffs_isblock(fs, blksfree, fragstoblks(fs, cgbno)); 2739 } else { 2740 frags = numfrags(fs, size); 2741 for (free = 0, i = 0; i < frags; i++) 2742 if (isset(blksfree, cgbno + i)) 2743 free++; 2744 if (free != 0 && free != frags) 2745 panic("ffs_checkblk: partially free fragment"); 2746 } 2747 brelse(bp); 2748 return (!free); 2749 } 2750 #endif /* INVARIANTS */ 2751 2752 /* 2753 * Free an inode. 2754 */ 2755 int 2756 ffs_vfree(pvp, ino, mode) 2757 struct vnode *pvp; 2758 ino_t ino; 2759 int mode; 2760 { 2761 struct ufsmount *ump; 2762 2763 if (DOINGSOFTDEP(pvp)) { 2764 softdep_freefile(pvp, ino, mode); 2765 return (0); 2766 } 2767 ump = VFSTOUFS(pvp->v_mount); 2768 return (ffs_freefile(ump, ump->um_fs, ump->um_devvp, ino, mode, NULL)); 2769 } 2770 2771 /* 2772 * Do the actual free operation. 2773 * The specified inode is placed back in the free map. 2774 */ 2775 int 2776 ffs_freefile(ump, fs, devvp, ino, mode, wkhd) 2777 struct ufsmount *ump; 2778 struct fs *fs; 2779 struct vnode *devvp; 2780 ino_t ino; 2781 int mode; 2782 struct workhead *wkhd; 2783 { 2784 struct cg *cgp; 2785 struct buf *bp; 2786 int error; 2787 u_int cg; 2788 u_int8_t *inosused; 2789 struct cdev *dev; 2790 2791 cg = ino_to_cg(fs, ino); 2792 if (devvp->v_type == VREG) { 2793 /* devvp is a snapshot */ 2794 MPASS(devvp->v_mount->mnt_data == ump); 2795 dev = ump->um_devvp->v_rdev; 2796 } else if (devvp->v_type == VCHR) { 2797 /* devvp is a normal disk device */ 2798 dev = devvp->v_rdev; 2799 } else { 2800 bp = NULL; 2801 return (0); 2802 } 2803 if (ino >= fs->fs_ipg * fs->fs_ncg) 2804 panic("ffs_freefile: range: dev = %s, ino = %ju, fs = %s", 2805 devtoname(dev), (uintmax_t)ino, fs->fs_fsmnt); 2806 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0) 2807 return (error); 2808 inosused = cg_inosused(cgp); 2809 ino %= fs->fs_ipg; 2810 if (isclr(inosused, ino)) { 2811 printf("dev = %s, ino = %ju, fs = %s\n", devtoname(dev), 2812 (uintmax_t)(ino + cg * fs->fs_ipg), fs->fs_fsmnt); 2813 if (fs->fs_ronly == 0) 2814 panic("ffs_freefile: freeing free inode"); 2815 } 2816 clrbit(inosused, ino); 2817 if (ino < cgp->cg_irotor) 2818 cgp->cg_irotor = ino; 2819 cgp->cg_cs.cs_nifree++; 2820 UFS_LOCK(ump); 2821 fs->fs_cstotal.cs_nifree++; 2822 fs->fs_cs(fs, cg).cs_nifree++; 2823 if ((mode & IFMT) == IFDIR) { 2824 cgp->cg_cs.cs_ndir--; 2825 fs->fs_cstotal.cs_ndir--; 2826 fs->fs_cs(fs, cg).cs_ndir--; 2827 } 2828 fs->fs_fmod = 1; 2829 ACTIVECLEAR(fs, cg); 2830 UFS_UNLOCK(ump); 2831 if (MOUNTEDSOFTDEP(UFSTOVFS(ump)) && devvp->v_type == VCHR) 2832 softdep_setup_inofree(UFSTOVFS(ump), bp, 2833 ino + cg * fs->fs_ipg, wkhd); 2834 bdwrite(bp); 2835 return (0); 2836 } 2837 2838 /* 2839 * Check to see if a file is free. 2840 * Used to check for allocated files in snapshots. 2841 */ 2842 int 2843 ffs_checkfreefile(fs, devvp, ino) 2844 struct fs *fs; 2845 struct vnode *devvp; 2846 ino_t ino; 2847 { 2848 struct cg *cgp; 2849 struct buf *bp; 2850 int ret, error; 2851 u_int cg; 2852 u_int8_t *inosused; 2853 2854 cg = ino_to_cg(fs, ino); 2855 if ((devvp->v_type != VREG) && (devvp->v_type != VCHR)) 2856 return (1); 2857 if (ino >= fs->fs_ipg * fs->fs_ncg) 2858 return (1); 2859 if ((error = ffs_getcg(fs, devvp, cg, &bp, &cgp)) != 0) 2860 return (1); 2861 inosused = cg_inosused(cgp); 2862 ino %= fs->fs_ipg; 2863 ret = isclr(inosused, ino); 2864 brelse(bp); 2865 return (ret); 2866 } 2867 2868 /* 2869 * Find a block of the specified size in the specified cylinder group. 2870 * 2871 * It is a panic if a request is made to find a block if none are 2872 * available. 2873 */ 2874 static ufs1_daddr_t 2875 ffs_mapsearch(fs, cgp, bpref, allocsiz) 2876 struct fs *fs; 2877 struct cg *cgp; 2878 ufs2_daddr_t bpref; 2879 int allocsiz; 2880 { 2881 ufs1_daddr_t bno; 2882 int start, len, loc, i; 2883 int blk, field, subfield, pos; 2884 u_int8_t *blksfree; 2885 2886 /* 2887 * find the fragment by searching through the free block 2888 * map for an appropriate bit pattern 2889 */ 2890 if (bpref) 2891 start = dtogd(fs, bpref) / NBBY; 2892 else 2893 start = cgp->cg_frotor / NBBY; 2894 blksfree = cg_blksfree(cgp); 2895 len = howmany(fs->fs_fpg, NBBY) - start; 2896 loc = scanc((u_int)len, (u_char *)&blksfree[start], 2897 fragtbl[fs->fs_frag], 2898 (u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY)))); 2899 if (loc == 0) { 2900 len = start + 1; 2901 start = 0; 2902 loc = scanc((u_int)len, (u_char *)&blksfree[0], 2903 fragtbl[fs->fs_frag], 2904 (u_char)(1 << (allocsiz - 1 + (fs->fs_frag % NBBY)))); 2905 if (loc == 0) { 2906 printf("start = %d, len = %d, fs = %s\n", 2907 start, len, fs->fs_fsmnt); 2908 panic("ffs_alloccg: map corrupted"); 2909 /* NOTREACHED */ 2910 } 2911 } 2912 bno = (start + len - loc) * NBBY; 2913 cgp->cg_frotor = bno; 2914 /* 2915 * found the byte in the map 2916 * sift through the bits to find the selected frag 2917 */ 2918 for (i = bno + NBBY; bno < i; bno += fs->fs_frag) { 2919 blk = blkmap(fs, blksfree, bno); 2920 blk <<= 1; 2921 field = around[allocsiz]; 2922 subfield = inside[allocsiz]; 2923 for (pos = 0; pos <= fs->fs_frag - allocsiz; pos++) { 2924 if ((blk & field) == subfield) 2925 return (bno + pos); 2926 field <<= 1; 2927 subfield <<= 1; 2928 } 2929 } 2930 printf("bno = %lu, fs = %s\n", (u_long)bno, fs->fs_fsmnt); 2931 panic("ffs_alloccg: block not in map"); 2932 return (-1); 2933 } 2934 2935 static const struct statfs * 2936 ffs_getmntstat(struct vnode *devvp) 2937 { 2938 2939 if (devvp->v_type == VCHR) 2940 return (&devvp->v_rdev->si_mountpt->mnt_stat); 2941 return (ffs_getmntstat(VFSTOUFS(devvp->v_mount)->um_devvp)); 2942 } 2943 2944 /* 2945 * Fetch and verify a cylinder group. 2946 */ 2947 int 2948 ffs_getcg(fs, devvp, cg, bpp, cgpp) 2949 struct fs *fs; 2950 struct vnode *devvp; 2951 u_int cg; 2952 struct buf **bpp; 2953 struct cg **cgpp; 2954 { 2955 struct buf *bp; 2956 struct cg *cgp; 2957 const struct statfs *sfs; 2958 int flags, error; 2959 2960 *bpp = NULL; 2961 *cgpp = NULL; 2962 flags = 0; 2963 if ((fs->fs_metackhash & CK_CYLGRP) != 0) 2964 flags |= GB_CKHASH; 2965 error = breadn_flags(devvp, devvp->v_type == VREG ? 2966 fragstoblks(fs, cgtod(fs, cg)) : fsbtodb(fs, cgtod(fs, cg)), 2967 (int)fs->fs_cgsize, NULL, NULL, 0, NOCRED, flags, 2968 ffs_ckhash_cg, &bp); 2969 if (error != 0) 2970 return (error); 2971 cgp = (struct cg *)bp->b_data; 2972 if ((fs->fs_metackhash & CK_CYLGRP) != 0 && 2973 (bp->b_flags & B_CKHASH) != 0 && 2974 cgp->cg_ckhash != bp->b_ckhash) { 2975 sfs = ffs_getmntstat(devvp); 2976 printf("UFS %s%s (%s) cylinder checksum failed: cg %u, cgp: " 2977 "0x%x != bp: 0x%jx\n", 2978 devvp->v_type == VCHR ? "" : "snapshot of ", 2979 sfs->f_mntfromname, sfs->f_mntonname, 2980 cg, cgp->cg_ckhash, (uintmax_t)bp->b_ckhash); 2981 bp->b_flags &= ~B_CKHASH; 2982 bp->b_flags |= B_INVAL | B_NOCACHE; 2983 brelse(bp); 2984 return (EIO); 2985 } 2986 if (!cg_chkmagic(cgp) || cgp->cg_cgx != cg) { 2987 sfs = ffs_getmntstat(devvp); 2988 printf("UFS %s%s (%s)", 2989 devvp->v_type == VCHR ? "" : "snapshot of ", 2990 sfs->f_mntfromname, sfs->f_mntonname); 2991 if (!cg_chkmagic(cgp)) 2992 printf(" cg %u: bad magic number 0x%x should be 0x%x\n", 2993 cg, cgp->cg_magic, CG_MAGIC); 2994 else 2995 printf(": wrong cylinder group cg %u != cgx %u\n", cg, 2996 cgp->cg_cgx); 2997 bp->b_flags &= ~B_CKHASH; 2998 bp->b_flags |= B_INVAL | B_NOCACHE; 2999 brelse(bp); 3000 return (EIO); 3001 } 3002 bp->b_flags &= ~B_CKHASH; 3003 bp->b_xflags |= BX_BKGRDWRITE; 3004 /* 3005 * If we are using check hashes on the cylinder group then we want 3006 * to limit changing the cylinder group time to when we are actually 3007 * going to write it to disk so that its check hash remains correct 3008 * in memory. If the CK_CYLGRP flag is set the time is updated in 3009 * ffs_bufwrite() as the buffer is queued for writing. Otherwise we 3010 * update the time here as we have done historically. 3011 */ 3012 if ((fs->fs_metackhash & CK_CYLGRP) != 0) 3013 bp->b_xflags |= BX_CYLGRP; 3014 else 3015 cgp->cg_old_time = cgp->cg_time = time_second; 3016 *bpp = bp; 3017 *cgpp = cgp; 3018 return (0); 3019 } 3020 3021 static void 3022 ffs_ckhash_cg(bp) 3023 struct buf *bp; 3024 { 3025 uint32_t ckhash; 3026 struct cg *cgp; 3027 3028 cgp = (struct cg *)bp->b_data; 3029 ckhash = cgp->cg_ckhash; 3030 cgp->cg_ckhash = 0; 3031 bp->b_ckhash = calculate_crc32c(~0L, bp->b_data, bp->b_bcount); 3032 cgp->cg_ckhash = ckhash; 3033 } 3034 3035 /* 3036 * Fserr prints the name of a filesystem with an error diagnostic. 3037 * 3038 * The form of the error message is: 3039 * fs: error message 3040 */ 3041 void 3042 ffs_fserr(fs, inum, cp) 3043 struct fs *fs; 3044 ino_t inum; 3045 char *cp; 3046 { 3047 struct thread *td = curthread; /* XXX */ 3048 struct proc *p = td->td_proc; 3049 3050 log(LOG_ERR, "pid %d (%s), uid %d inumber %ju on %s: %s\n", 3051 p->p_pid, p->p_comm, td->td_ucred->cr_uid, (uintmax_t)inum, 3052 fs->fs_fsmnt, cp); 3053 } 3054 3055 /* 3056 * This function provides the capability for the fsck program to 3057 * update an active filesystem. Fourteen operations are provided: 3058 * 3059 * adjrefcnt(inode, amt) - adjusts the reference count on the 3060 * specified inode by the specified amount. Under normal 3061 * operation the count should always go down. Decrementing 3062 * the count to zero will cause the inode to be freed. 3063 * adjblkcnt(inode, amt) - adjust the number of blocks used by the 3064 * inode by the specified amount. 3065 * adjsize(inode, size) - set the size of the inode to the 3066 * specified size. 3067 * adjndir, adjbfree, adjifree, adjffree, adjnumclusters(amt) - 3068 * adjust the superblock summary. 3069 * freedirs(inode, count) - directory inodes [inode..inode + count - 1] 3070 * are marked as free. Inodes should never have to be marked 3071 * as in use. 3072 * freefiles(inode, count) - file inodes [inode..inode + count - 1] 3073 * are marked as free. Inodes should never have to be marked 3074 * as in use. 3075 * freeblks(blockno, size) - blocks [blockno..blockno + size - 1] 3076 * are marked as free. Blocks should never have to be marked 3077 * as in use. 3078 * setflags(flags, set/clear) - the fs_flags field has the specified 3079 * flags set (second parameter +1) or cleared (second parameter -1). 3080 * setcwd(dirinode) - set the current directory to dirinode in the 3081 * filesystem associated with the snapshot. 3082 * setdotdot(oldvalue, newvalue) - Verify that the inode number for ".." 3083 * in the current directory is oldvalue then change it to newvalue. 3084 * unlink(nameptr, oldvalue) - Verify that the inode number associated 3085 * with nameptr in the current directory is oldvalue then unlink it. 3086 * 3087 * The following functions may only be used on a quiescent filesystem 3088 * by the soft updates journal. They are not safe to be run on an active 3089 * filesystem. 3090 * 3091 * setinode(inode, dip) - the specified disk inode is replaced with the 3092 * contents pointed to by dip. 3093 * setbufoutput(fd, flags) - output associated with the specified file 3094 * descriptor (which must reference the character device supporting 3095 * the filesystem) switches from using physio to running through the 3096 * buffer cache when flags is set to 1. The descriptor reverts to 3097 * physio for output when flags is set to zero. 3098 */ 3099 3100 static int sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS); 3101 3102 SYSCTL_PROC(_vfs_ffs, FFS_ADJ_REFCNT, adjrefcnt, CTLFLAG_WR|CTLTYPE_STRUCT, 3103 0, 0, sysctl_ffs_fsck, "S,fsck", "Adjust Inode Reference Count"); 3104 3105 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_BLKCNT, adjblkcnt, CTLFLAG_WR, 3106 sysctl_ffs_fsck, "Adjust Inode Used Blocks Count"); 3107 3108 static SYSCTL_NODE(_vfs_ffs, FFS_SET_SIZE, setsize, CTLFLAG_WR, 3109 sysctl_ffs_fsck, "Set the inode size"); 3110 3111 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NDIR, adjndir, CTLFLAG_WR, 3112 sysctl_ffs_fsck, "Adjust number of directories"); 3113 3114 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NBFREE, adjnbfree, CTLFLAG_WR, 3115 sysctl_ffs_fsck, "Adjust number of free blocks"); 3116 3117 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NIFREE, adjnifree, CTLFLAG_WR, 3118 sysctl_ffs_fsck, "Adjust number of free inodes"); 3119 3120 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NFFREE, adjnffree, CTLFLAG_WR, 3121 sysctl_ffs_fsck, "Adjust number of free frags"); 3122 3123 static SYSCTL_NODE(_vfs_ffs, FFS_ADJ_NUMCLUSTERS, adjnumclusters, CTLFLAG_WR, 3124 sysctl_ffs_fsck, "Adjust number of free clusters"); 3125 3126 static SYSCTL_NODE(_vfs_ffs, FFS_DIR_FREE, freedirs, CTLFLAG_WR, 3127 sysctl_ffs_fsck, "Free Range of Directory Inodes"); 3128 3129 static SYSCTL_NODE(_vfs_ffs, FFS_FILE_FREE, freefiles, CTLFLAG_WR, 3130 sysctl_ffs_fsck, "Free Range of File Inodes"); 3131 3132 static SYSCTL_NODE(_vfs_ffs, FFS_BLK_FREE, freeblks, CTLFLAG_WR, 3133 sysctl_ffs_fsck, "Free Range of Blocks"); 3134 3135 static SYSCTL_NODE(_vfs_ffs, FFS_SET_FLAGS, setflags, CTLFLAG_WR, 3136 sysctl_ffs_fsck, "Change Filesystem Flags"); 3137 3138 static SYSCTL_NODE(_vfs_ffs, FFS_SET_CWD, setcwd, CTLFLAG_WR, 3139 sysctl_ffs_fsck, "Set Current Working Directory"); 3140 3141 static SYSCTL_NODE(_vfs_ffs, FFS_SET_DOTDOT, setdotdot, CTLFLAG_WR, 3142 sysctl_ffs_fsck, "Change Value of .. Entry"); 3143 3144 static SYSCTL_NODE(_vfs_ffs, FFS_UNLINK, unlink, CTLFLAG_WR, 3145 sysctl_ffs_fsck, "Unlink a Duplicate Name"); 3146 3147 static SYSCTL_NODE(_vfs_ffs, FFS_SET_INODE, setinode, CTLFLAG_WR, 3148 sysctl_ffs_fsck, "Update an On-Disk Inode"); 3149 3150 static SYSCTL_NODE(_vfs_ffs, FFS_SET_BUFOUTPUT, setbufoutput, CTLFLAG_WR, 3151 sysctl_ffs_fsck, "Set Buffered Writing for Descriptor"); 3152 3153 #ifdef DIAGNOSTIC 3154 static int fsckcmds = 0; 3155 SYSCTL_INT(_debug, OID_AUTO, ffs_fsckcmds, CTLFLAG_RW, &fsckcmds, 0, 3156 "print out fsck_ffs-based filesystem update commands"); 3157 #endif /* DIAGNOSTIC */ 3158 3159 static int buffered_write(struct file *, struct uio *, struct ucred *, 3160 int, struct thread *); 3161 3162 static int 3163 sysctl_ffs_fsck(SYSCTL_HANDLER_ARGS) 3164 { 3165 struct thread *td = curthread; 3166 struct fsck_cmd cmd; 3167 struct ufsmount *ump; 3168 struct vnode *vp, *dvp, *fdvp; 3169 struct inode *ip, *dp; 3170 struct mount *mp; 3171 struct fs *fs; 3172 ufs2_daddr_t blkno; 3173 long blkcnt, blksize; 3174 u_long key; 3175 struct file *fp, *vfp; 3176 cap_rights_t rights; 3177 int filetype, error; 3178 static struct fileops *origops, bufferedops; 3179 3180 if (req->newlen > sizeof cmd) 3181 return (EBADRPC); 3182 if ((error = SYSCTL_IN(req, &cmd, sizeof cmd)) != 0) 3183 return (error); 3184 if (cmd.version != FFS_CMD_VERSION) 3185 return (ERPCMISMATCH); 3186 if ((error = getvnode(td, cmd.handle, 3187 cap_rights_init(&rights, CAP_FSCK), &fp)) != 0) 3188 return (error); 3189 vp = fp->f_data; 3190 if (vp->v_type != VREG && vp->v_type != VDIR) { 3191 fdrop(fp, td); 3192 return (EINVAL); 3193 } 3194 vn_start_write(vp, &mp, V_WAIT); 3195 if (mp == NULL || 3196 strncmp(mp->mnt_stat.f_fstypename, "ufs", MFSNAMELEN)) { 3197 vn_finished_write(mp); 3198 fdrop(fp, td); 3199 return (EINVAL); 3200 } 3201 ump = VFSTOUFS(mp); 3202 if ((mp->mnt_flag & MNT_RDONLY) && 3203 ump->um_fsckpid != td->td_proc->p_pid) { 3204 vn_finished_write(mp); 3205 fdrop(fp, td); 3206 return (EROFS); 3207 } 3208 fs = ump->um_fs; 3209 filetype = IFREG; 3210 3211 switch (oidp->oid_number) { 3212 3213 case FFS_SET_FLAGS: 3214 #ifdef DIAGNOSTIC 3215 if (fsckcmds) 3216 printf("%s: %s flags\n", mp->mnt_stat.f_mntonname, 3217 cmd.size > 0 ? "set" : "clear"); 3218 #endif /* DIAGNOSTIC */ 3219 if (cmd.size > 0) 3220 fs->fs_flags |= (long)cmd.value; 3221 else 3222 fs->fs_flags &= ~(long)cmd.value; 3223 break; 3224 3225 case FFS_ADJ_REFCNT: 3226 #ifdef DIAGNOSTIC 3227 if (fsckcmds) { 3228 printf("%s: adjust inode %jd link count by %jd\n", 3229 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3230 (intmax_t)cmd.size); 3231 } 3232 #endif /* DIAGNOSTIC */ 3233 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3234 break; 3235 ip = VTOI(vp); 3236 ip->i_nlink += cmd.size; 3237 DIP_SET(ip, i_nlink, ip->i_nlink); 3238 ip->i_effnlink += cmd.size; 3239 ip->i_flag |= IN_CHANGE | IN_MODIFIED; 3240 error = ffs_update(vp, 1); 3241 if (DOINGSOFTDEP(vp)) 3242 softdep_change_linkcnt(ip); 3243 vput(vp); 3244 break; 3245 3246 case FFS_ADJ_BLKCNT: 3247 #ifdef DIAGNOSTIC 3248 if (fsckcmds) { 3249 printf("%s: adjust inode %jd block count by %jd\n", 3250 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3251 (intmax_t)cmd.size); 3252 } 3253 #endif /* DIAGNOSTIC */ 3254 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3255 break; 3256 ip = VTOI(vp); 3257 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + cmd.size); 3258 ip->i_flag |= IN_CHANGE | IN_MODIFIED; 3259 error = ffs_update(vp, 1); 3260 vput(vp); 3261 break; 3262 3263 case FFS_SET_SIZE: 3264 #ifdef DIAGNOSTIC 3265 if (fsckcmds) { 3266 printf("%s: set inode %jd size to %jd\n", 3267 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3268 (intmax_t)cmd.size); 3269 } 3270 #endif /* DIAGNOSTIC */ 3271 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3272 break; 3273 ip = VTOI(vp); 3274 DIP_SET(ip, i_size, cmd.size); 3275 ip->i_flag |= IN_CHANGE | IN_MODIFIED; 3276 error = ffs_update(vp, 1); 3277 vput(vp); 3278 break; 3279 3280 case FFS_DIR_FREE: 3281 filetype = IFDIR; 3282 /* fall through */ 3283 3284 case FFS_FILE_FREE: 3285 #ifdef DIAGNOSTIC 3286 if (fsckcmds) { 3287 if (cmd.size == 1) 3288 printf("%s: free %s inode %ju\n", 3289 mp->mnt_stat.f_mntonname, 3290 filetype == IFDIR ? "directory" : "file", 3291 (uintmax_t)cmd.value); 3292 else 3293 printf("%s: free %s inodes %ju-%ju\n", 3294 mp->mnt_stat.f_mntonname, 3295 filetype == IFDIR ? "directory" : "file", 3296 (uintmax_t)cmd.value, 3297 (uintmax_t)(cmd.value + cmd.size - 1)); 3298 } 3299 #endif /* DIAGNOSTIC */ 3300 while (cmd.size > 0) { 3301 if ((error = ffs_freefile(ump, fs, ump->um_devvp, 3302 cmd.value, filetype, NULL))) 3303 break; 3304 cmd.size -= 1; 3305 cmd.value += 1; 3306 } 3307 break; 3308 3309 case FFS_BLK_FREE: 3310 #ifdef DIAGNOSTIC 3311 if (fsckcmds) { 3312 if (cmd.size == 1) 3313 printf("%s: free block %jd\n", 3314 mp->mnt_stat.f_mntonname, 3315 (intmax_t)cmd.value); 3316 else 3317 printf("%s: free blocks %jd-%jd\n", 3318 mp->mnt_stat.f_mntonname, 3319 (intmax_t)cmd.value, 3320 (intmax_t)cmd.value + cmd.size - 1); 3321 } 3322 #endif /* DIAGNOSTIC */ 3323 blkno = cmd.value; 3324 blkcnt = cmd.size; 3325 blksize = fs->fs_frag - (blkno % fs->fs_frag); 3326 key = ffs_blkrelease_start(ump, ump->um_devvp, UFS_ROOTINO); 3327 while (blkcnt > 0) { 3328 if (blkcnt < blksize) 3329 blksize = blkcnt; 3330 ffs_blkfree(ump, fs, ump->um_devvp, blkno, 3331 blksize * fs->fs_fsize, UFS_ROOTINO, 3332 VDIR, NULL, key); 3333 blkno += blksize; 3334 blkcnt -= blksize; 3335 blksize = fs->fs_frag; 3336 } 3337 ffs_blkrelease_finish(ump, key); 3338 break; 3339 3340 /* 3341 * Adjust superblock summaries. fsck(8) is expected to 3342 * submit deltas when necessary. 3343 */ 3344 case FFS_ADJ_NDIR: 3345 #ifdef DIAGNOSTIC 3346 if (fsckcmds) { 3347 printf("%s: adjust number of directories by %jd\n", 3348 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3349 } 3350 #endif /* DIAGNOSTIC */ 3351 fs->fs_cstotal.cs_ndir += cmd.value; 3352 break; 3353 3354 case FFS_ADJ_NBFREE: 3355 #ifdef DIAGNOSTIC 3356 if (fsckcmds) { 3357 printf("%s: adjust number of free blocks by %+jd\n", 3358 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3359 } 3360 #endif /* DIAGNOSTIC */ 3361 fs->fs_cstotal.cs_nbfree += cmd.value; 3362 break; 3363 3364 case FFS_ADJ_NIFREE: 3365 #ifdef DIAGNOSTIC 3366 if (fsckcmds) { 3367 printf("%s: adjust number of free inodes by %+jd\n", 3368 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3369 } 3370 #endif /* DIAGNOSTIC */ 3371 fs->fs_cstotal.cs_nifree += cmd.value; 3372 break; 3373 3374 case FFS_ADJ_NFFREE: 3375 #ifdef DIAGNOSTIC 3376 if (fsckcmds) { 3377 printf("%s: adjust number of free frags by %+jd\n", 3378 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3379 } 3380 #endif /* DIAGNOSTIC */ 3381 fs->fs_cstotal.cs_nffree += cmd.value; 3382 break; 3383 3384 case FFS_ADJ_NUMCLUSTERS: 3385 #ifdef DIAGNOSTIC 3386 if (fsckcmds) { 3387 printf("%s: adjust number of free clusters by %+jd\n", 3388 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3389 } 3390 #endif /* DIAGNOSTIC */ 3391 fs->fs_cstotal.cs_numclusters += cmd.value; 3392 break; 3393 3394 case FFS_SET_CWD: 3395 #ifdef DIAGNOSTIC 3396 if (fsckcmds) { 3397 printf("%s: set current directory to inode %jd\n", 3398 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3399 } 3400 #endif /* DIAGNOSTIC */ 3401 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_SHARED, &vp))) 3402 break; 3403 AUDIT_ARG_VNODE1(vp); 3404 if ((error = change_dir(vp, td)) != 0) { 3405 vput(vp); 3406 break; 3407 } 3408 VOP_UNLOCK(vp, 0); 3409 pwd_chdir(td, vp); 3410 break; 3411 3412 case FFS_SET_DOTDOT: 3413 #ifdef DIAGNOSTIC 3414 if (fsckcmds) { 3415 printf("%s: change .. in cwd from %jd to %jd\n", 3416 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value, 3417 (intmax_t)cmd.size); 3418 } 3419 #endif /* DIAGNOSTIC */ 3420 /* 3421 * First we have to get and lock the parent directory 3422 * to which ".." points. 3423 */ 3424 error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &fdvp); 3425 if (error) 3426 break; 3427 /* 3428 * Now we get and lock the child directory containing "..". 3429 */ 3430 FILEDESC_SLOCK(td->td_proc->p_fd); 3431 dvp = td->td_proc->p_fd->fd_cdir; 3432 FILEDESC_SUNLOCK(td->td_proc->p_fd); 3433 if ((error = vget(dvp, LK_EXCLUSIVE, td)) != 0) { 3434 vput(fdvp); 3435 break; 3436 } 3437 dp = VTOI(dvp); 3438 dp->i_offset = 12; /* XXX mastertemplate.dot_reclen */ 3439 error = ufs_dirrewrite(dp, VTOI(fdvp), (ino_t)cmd.size, 3440 DT_DIR, 0); 3441 cache_purge(fdvp); 3442 cache_purge(dvp); 3443 vput(dvp); 3444 vput(fdvp); 3445 break; 3446 3447 case FFS_UNLINK: 3448 #ifdef DIAGNOSTIC 3449 if (fsckcmds) { 3450 char buf[32]; 3451 3452 if (copyinstr((char *)(intptr_t)cmd.value, buf,32,NULL)) 3453 strncpy(buf, "Name_too_long", 32); 3454 printf("%s: unlink %s (inode %jd)\n", 3455 mp->mnt_stat.f_mntonname, buf, (intmax_t)cmd.size); 3456 } 3457 #endif /* DIAGNOSTIC */ 3458 /* 3459 * kern_funlinkat will do its own start/finish writes and 3460 * they do not nest, so drop ours here. Setting mp == NULL 3461 * indicates that vn_finished_write is not needed down below. 3462 */ 3463 vn_finished_write(mp); 3464 mp = NULL; 3465 error = kern_funlinkat(td, AT_FDCWD, 3466 (char *)(intptr_t)cmd.value, FD_NONE, UIO_USERSPACE, 3467 0, (ino_t)cmd.size); 3468 break; 3469 3470 case FFS_SET_INODE: 3471 if (ump->um_fsckpid != td->td_proc->p_pid) { 3472 error = EPERM; 3473 break; 3474 } 3475 #ifdef DIAGNOSTIC 3476 if (fsckcmds) { 3477 printf("%s: update inode %jd\n", 3478 mp->mnt_stat.f_mntonname, (intmax_t)cmd.value); 3479 } 3480 #endif /* DIAGNOSTIC */ 3481 if ((error = ffs_vget(mp, (ino_t)cmd.value, LK_EXCLUSIVE, &vp))) 3482 break; 3483 AUDIT_ARG_VNODE1(vp); 3484 ip = VTOI(vp); 3485 if (I_IS_UFS1(ip)) 3486 error = copyin((void *)(intptr_t)cmd.size, ip->i_din1, 3487 sizeof(struct ufs1_dinode)); 3488 else 3489 error = copyin((void *)(intptr_t)cmd.size, ip->i_din2, 3490 sizeof(struct ufs2_dinode)); 3491 if (error) { 3492 vput(vp); 3493 break; 3494 } 3495 ip->i_flag |= IN_CHANGE | IN_MODIFIED; 3496 error = ffs_update(vp, 1); 3497 vput(vp); 3498 break; 3499 3500 case FFS_SET_BUFOUTPUT: 3501 if (ump->um_fsckpid != td->td_proc->p_pid) { 3502 error = EPERM; 3503 break; 3504 } 3505 if (ITOUMP(VTOI(vp)) != ump) { 3506 error = EINVAL; 3507 break; 3508 } 3509 #ifdef DIAGNOSTIC 3510 if (fsckcmds) { 3511 printf("%s: %s buffered output for descriptor %jd\n", 3512 mp->mnt_stat.f_mntonname, 3513 cmd.size == 1 ? "enable" : "disable", 3514 (intmax_t)cmd.value); 3515 } 3516 #endif /* DIAGNOSTIC */ 3517 if ((error = getvnode(td, cmd.value, 3518 cap_rights_init(&rights, CAP_FSCK), &vfp)) != 0) 3519 break; 3520 if (vfp->f_vnode->v_type != VCHR) { 3521 fdrop(vfp, td); 3522 error = EINVAL; 3523 break; 3524 } 3525 if (origops == NULL) { 3526 origops = vfp->f_ops; 3527 bcopy((void *)origops, (void *)&bufferedops, 3528 sizeof(bufferedops)); 3529 bufferedops.fo_write = buffered_write; 3530 } 3531 if (cmd.size == 1) 3532 atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops, 3533 (uintptr_t)&bufferedops); 3534 else 3535 atomic_store_rel_ptr((volatile uintptr_t *)&vfp->f_ops, 3536 (uintptr_t)origops); 3537 fdrop(vfp, td); 3538 break; 3539 3540 default: 3541 #ifdef DIAGNOSTIC 3542 if (fsckcmds) { 3543 printf("Invalid request %d from fsck\n", 3544 oidp->oid_number); 3545 } 3546 #endif /* DIAGNOSTIC */ 3547 error = EINVAL; 3548 break; 3549 3550 } 3551 fdrop(fp, td); 3552 vn_finished_write(mp); 3553 return (error); 3554 } 3555 3556 /* 3557 * Function to switch a descriptor to use the buffer cache to stage 3558 * its I/O. This is needed so that writes to the filesystem device 3559 * will give snapshots a chance to copy modified blocks for which it 3560 * needs to retain copies. 3561 */ 3562 static int 3563 buffered_write(fp, uio, active_cred, flags, td) 3564 struct file *fp; 3565 struct uio *uio; 3566 struct ucred *active_cred; 3567 int flags; 3568 struct thread *td; 3569 { 3570 struct vnode *devvp, *vp; 3571 struct inode *ip; 3572 struct buf *bp; 3573 struct fs *fs; 3574 struct filedesc *fdp; 3575 int error; 3576 daddr_t lbn; 3577 3578 /* 3579 * The devvp is associated with the /dev filesystem. To discover 3580 * the filesystem with which the device is associated, we depend 3581 * on the application setting the current directory to a location 3582 * within the filesystem being written. Yes, this is an ugly hack. 3583 */ 3584 devvp = fp->f_vnode; 3585 if (!vn_isdisk(devvp, NULL)) 3586 return (EINVAL); 3587 fdp = td->td_proc->p_fd; 3588 FILEDESC_SLOCK(fdp); 3589 vp = fdp->fd_cdir; 3590 vref(vp); 3591 FILEDESC_SUNLOCK(fdp); 3592 vn_lock(vp, LK_SHARED | LK_RETRY); 3593 /* 3594 * Check that the current directory vnode indeed belongs to 3595 * UFS before trying to dereference UFS-specific v_data fields. 3596 */ 3597 if (vp->v_op != &ffs_vnodeops1 && vp->v_op != &ffs_vnodeops2) { 3598 vput(vp); 3599 return (EINVAL); 3600 } 3601 ip = VTOI(vp); 3602 if (ITODEVVP(ip) != devvp) { 3603 vput(vp); 3604 return (EINVAL); 3605 } 3606 fs = ITOFS(ip); 3607 vput(vp); 3608 foffset_lock_uio(fp, uio, flags); 3609 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 3610 #ifdef DIAGNOSTIC 3611 if (fsckcmds) { 3612 printf("%s: buffered write for block %jd\n", 3613 fs->fs_fsmnt, (intmax_t)btodb(uio->uio_offset)); 3614 } 3615 #endif /* DIAGNOSTIC */ 3616 /* 3617 * All I/O must be contained within a filesystem block, start on 3618 * a fragment boundary, and be a multiple of fragments in length. 3619 */ 3620 if (uio->uio_resid > fs->fs_bsize - (uio->uio_offset % fs->fs_bsize) || 3621 fragoff(fs, uio->uio_offset) != 0 || 3622 fragoff(fs, uio->uio_resid) != 0) { 3623 error = EINVAL; 3624 goto out; 3625 } 3626 lbn = numfrags(fs, uio->uio_offset); 3627 bp = getblk(devvp, lbn, uio->uio_resid, 0, 0, 0); 3628 bp->b_flags |= B_RELBUF; 3629 if ((error = uiomove((char *)bp->b_data, uio->uio_resid, uio)) != 0) { 3630 brelse(bp); 3631 goto out; 3632 } 3633 error = bwrite(bp); 3634 out: 3635 VOP_UNLOCK(devvp, 0); 3636 foffset_unlock_uio(fp, uio, flags | FOF_NEXTOFF); 3637 return (error); 3638 } 3639