1 /*- 2 * Copyright 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * Further information about snapshots can be obtained from: 5 * 6 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 7 * 1614 Oxford Street mckusick@mckusick.com 8 * Berkeley, CA 94709-1608 +1-510-843-9542 9 * USA 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 22 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 23 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 24 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 25 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * @(#)ffs_snapshot.c 8.11 (McKusick) 7/23/00 34 */ 35 36 #include <sys/cdefs.h> 37 __FBSDID("$FreeBSD$"); 38 39 #include "opt_quota.h" 40 41 #include <sys/param.h> 42 #include <sys/kernel.h> 43 #include <sys/systm.h> 44 #include <sys/conf.h> 45 #include <sys/bio.h> 46 #include <sys/buf.h> 47 #include <sys/fcntl.h> 48 #include <sys/proc.h> 49 #include <sys/namei.h> 50 #include <sys/sched.h> 51 #include <sys/stat.h> 52 #include <sys/malloc.h> 53 #include <sys/mount.h> 54 #include <sys/resource.h> 55 #include <sys/resourcevar.h> 56 #include <sys/rwlock.h> 57 #include <sys/vnode.h> 58 59 #include <geom/geom.h> 60 61 #include <ufs/ufs/extattr.h> 62 #include <ufs/ufs/quota.h> 63 #include <ufs/ufs/ufsmount.h> 64 #include <ufs/ufs/inode.h> 65 #include <ufs/ufs/ufs_extern.h> 66 67 #include <ufs/ffs/fs.h> 68 #include <ufs/ffs/ffs_extern.h> 69 70 #define KERNCRED thread0.td_ucred 71 #define DEBUG 1 72 73 #include "opt_ffs.h" 74 75 #ifdef NO_FFS_SNAPSHOT 76 int 77 ffs_snapshot(mp, snapfile) 78 struct mount *mp; 79 char *snapfile; 80 { 81 return (EINVAL); 82 } 83 84 int 85 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, wkhd) 86 struct fs *fs; 87 struct vnode *devvp; 88 ufs2_daddr_t bno; 89 long size; 90 ino_t inum; 91 enum vtype vtype; 92 struct workhead *wkhd; 93 { 94 return (EINVAL); 95 } 96 97 void 98 ffs_snapremove(vp) 99 struct vnode *vp; 100 { 101 } 102 103 void 104 ffs_snapshot_mount(mp) 105 struct mount *mp; 106 { 107 } 108 109 void 110 ffs_snapshot_unmount(mp) 111 struct mount *mp; 112 { 113 } 114 115 void 116 ffs_snapgone(ip) 117 struct inode *ip; 118 { 119 } 120 121 int 122 ffs_copyonwrite(devvp, bp) 123 struct vnode *devvp; 124 struct buf *bp; 125 { 126 return (EINVAL); 127 } 128 129 void 130 ffs_sync_snap(mp, waitfor) 131 struct mount *mp; 132 int waitfor; 133 { 134 } 135 136 #else 137 FEATURE(ffs_snapshot, "FFS snapshot support"); 138 139 LIST_HEAD(, snapdata) snapfree; 140 static struct mtx snapfree_lock; 141 MTX_SYSINIT(ffs_snapfree, &snapfree_lock, "snapdata free list", MTX_DEF); 142 143 static int cgaccount(int, struct vnode *, struct buf *, int); 144 static int expunge_ufs1(struct vnode *, struct inode *, struct fs *, 145 int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *, 146 ufs_lbn_t, int), int, int); 147 static int indiracct_ufs1(struct vnode *, struct vnode *, int, 148 ufs1_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *, 149 int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *, 150 ufs_lbn_t, int), int); 151 static int fullacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, 152 struct fs *, ufs_lbn_t, int); 153 static int snapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, 154 struct fs *, ufs_lbn_t, int); 155 static int mapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, 156 struct fs *, ufs_lbn_t, int); 157 static int expunge_ufs2(struct vnode *, struct inode *, struct fs *, 158 int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *, 159 ufs_lbn_t, int), int, int); 160 static int indiracct_ufs2(struct vnode *, struct vnode *, int, 161 ufs2_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *, 162 int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *, 163 ufs_lbn_t, int), int); 164 static int fullacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, 165 struct fs *, ufs_lbn_t, int); 166 static int snapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, 167 struct fs *, ufs_lbn_t, int); 168 static int mapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, 169 struct fs *, ufs_lbn_t, int); 170 static int readblock(struct vnode *vp, struct buf *, ufs2_daddr_t); 171 static void try_free_snapdata(struct vnode *devvp); 172 static struct snapdata *ffs_snapdata_acquire(struct vnode *devvp); 173 static int ffs_bp_snapblk(struct vnode *, struct buf *); 174 175 /* 176 * To ensure the consistency of snapshots across crashes, we must 177 * synchronously write out copied blocks before allowing the 178 * originals to be modified. Because of the rather severe speed 179 * penalty that this imposes, the code normally only ensures 180 * persistence for the filesystem metadata contained within a 181 * snapshot. Setting the following flag allows this crash 182 * persistence to be enabled for file contents. 183 */ 184 int dopersistence = 0; 185 186 #ifdef DEBUG 187 #include <sys/sysctl.h> 188 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, ""); 189 static int snapdebug = 0; 190 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, ""); 191 int collectsnapstats = 0; 192 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats, 193 0, ""); 194 #endif /* DEBUG */ 195 196 /* 197 * Create a snapshot file and initialize it for the filesystem. 198 */ 199 int 200 ffs_snapshot(mp, snapfile) 201 struct mount *mp; 202 char *snapfile; 203 { 204 ufs2_daddr_t numblks, blkno, *blkp, *snapblklist; 205 int error, cg, snaploc; 206 int i, size, len, loc; 207 ufs2_daddr_t blockno; 208 uint64_t flag; 209 struct timespec starttime = {0, 0}, endtime; 210 char saved_nice = 0; 211 long redo = 0, snaplistsize = 0; 212 int32_t *lp; 213 void *space; 214 struct fs *copy_fs = NULL, *fs; 215 struct thread *td = curthread; 216 struct inode *ip, *xp; 217 struct buf *bp, *nbp, *ibp; 218 struct nameidata nd; 219 struct mount *wrtmp; 220 struct vattr vat; 221 struct vnode *vp, *xvp, *mvp, *devvp; 222 struct uio auio; 223 struct iovec aiov; 224 struct snapdata *sn; 225 struct ufsmount *ump; 226 227 ump = VFSTOUFS(mp); 228 fs = ump->um_fs; 229 sn = NULL; 230 /* 231 * At the moment, journaled soft updates cannot support 232 * taking snapshots. 233 */ 234 if (MOUNTEDSUJ(mp)) { 235 vfs_mount_error(mp, "%s: Snapshots are not yet supported when " 236 "running with journaled soft updates", fs->fs_fsmnt); 237 return (EOPNOTSUPP); 238 } 239 MNT_ILOCK(mp); 240 flag = mp->mnt_flag; 241 MNT_IUNLOCK(mp); 242 /* 243 * Need to serialize access to snapshot code per filesystem. 244 */ 245 /* 246 * Assign a snapshot slot in the superblock. 247 */ 248 UFS_LOCK(ump); 249 for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) 250 if (fs->fs_snapinum[snaploc] == 0) 251 break; 252 UFS_UNLOCK(ump); 253 if (snaploc == FSMAXSNAP) 254 return (ENOSPC); 255 /* 256 * Create the snapshot file. 257 */ 258 restart: 259 NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF | NOCACHE, UIO_SYSSPACE, 260 snapfile, td); 261 if ((error = namei(&nd)) != 0) 262 return (error); 263 if (nd.ni_vp != NULL) { 264 vput(nd.ni_vp); 265 error = EEXIST; 266 } 267 if (nd.ni_dvp->v_mount != mp) 268 error = EXDEV; 269 if (error) { 270 NDFREE(&nd, NDF_ONLY_PNBUF); 271 if (nd.ni_dvp == nd.ni_vp) 272 vrele(nd.ni_dvp); 273 else 274 vput(nd.ni_dvp); 275 return (error); 276 } 277 VATTR_NULL(&vat); 278 vat.va_type = VREG; 279 vat.va_mode = S_IRUSR; 280 vat.va_vaflags |= VA_EXCLUSIVE; 281 if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp)) 282 wrtmp = NULL; 283 if (wrtmp != mp) 284 panic("ffs_snapshot: mount mismatch"); 285 vfs_rel(wrtmp); 286 if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) { 287 NDFREE(&nd, NDF_ONLY_PNBUF); 288 vput(nd.ni_dvp); 289 if ((error = vn_start_write(NULL, &wrtmp, 290 V_XSLEEP | PCATCH)) != 0) 291 return (error); 292 goto restart; 293 } 294 error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat); 295 VOP_UNLOCK(nd.ni_dvp, 0); 296 if (error) { 297 NDFREE(&nd, NDF_ONLY_PNBUF); 298 vn_finished_write(wrtmp); 299 vrele(nd.ni_dvp); 300 return (error); 301 } 302 vp = nd.ni_vp; 303 vp->v_vflag |= VV_SYSTEM; 304 ip = VTOI(vp); 305 devvp = ITODEVVP(ip); 306 /* 307 * Allocate and copy the last block contents so as to be able 308 * to set size to that of the filesystem. 309 */ 310 numblks = howmany(fs->fs_size, fs->fs_frag); 311 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)), 312 fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp); 313 if (error) 314 goto out; 315 ip->i_size = lblktosize(fs, (off_t)numblks); 316 DIP_SET(ip, i_size, ip->i_size); 317 ip->i_flag |= IN_CHANGE | IN_UPDATE; 318 error = readblock(vp, bp, numblks - 1); 319 bawrite(bp); 320 if (error != 0) 321 goto out; 322 /* 323 * Preallocate critical data structures so that we can copy 324 * them in without further allocation after we suspend all 325 * operations on the filesystem. We would like to just release 326 * the allocated buffers without writing them since they will 327 * be filled in below once we are ready to go, but this upsets 328 * the soft update code, so we go ahead and write the new buffers. 329 * 330 * Allocate all indirect blocks and mark all of them as not 331 * needing to be copied. 332 */ 333 for (blkno = UFS_NDADDR; blkno < numblks; blkno += NINDIR(fs)) { 334 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno), 335 fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp); 336 if (error) 337 goto out; 338 bawrite(ibp); 339 } 340 /* 341 * Allocate copies for the superblock and its summary information. 342 */ 343 error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED, 344 0, &nbp); 345 if (error) 346 goto out; 347 bawrite(nbp); 348 blkno = fragstoblks(fs, fs->fs_csaddr); 349 len = howmany(fs->fs_cssize, fs->fs_bsize); 350 for (loc = 0; loc < len; loc++) { 351 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)), 352 fs->fs_bsize, KERNCRED, 0, &nbp); 353 if (error) 354 goto out; 355 bawrite(nbp); 356 } 357 /* 358 * Allocate all cylinder group blocks. 359 */ 360 for (cg = 0; cg < fs->fs_ncg; cg++) { 361 error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)), 362 fs->fs_bsize, KERNCRED, 0, &nbp); 363 if (error) 364 goto out; 365 bawrite(nbp); 366 if (cg % 10 == 0) 367 ffs_syncvnode(vp, MNT_WAIT, 0); 368 } 369 /* 370 * Copy all the cylinder group maps. Although the 371 * filesystem is still active, we hope that only a few 372 * cylinder groups will change between now and when we 373 * suspend operations. Thus, we will be able to quickly 374 * touch up the few cylinder groups that changed during 375 * the suspension period. 376 */ 377 len = howmany(fs->fs_ncg, NBBY); 378 space = malloc(len, M_DEVBUF, M_WAITOK|M_ZERO); 379 UFS_LOCK(ump); 380 fs->fs_active = space; 381 UFS_UNLOCK(ump); 382 for (cg = 0; cg < fs->fs_ncg; cg++) { 383 error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)), 384 fs->fs_bsize, KERNCRED, 0, &nbp); 385 if (error) 386 goto out; 387 error = cgaccount(cg, vp, nbp, 1); 388 bawrite(nbp); 389 if (cg % 10 == 0) 390 ffs_syncvnode(vp, MNT_WAIT, 0); 391 if (error) 392 goto out; 393 } 394 /* 395 * Change inode to snapshot type file. 396 */ 397 ip->i_flags |= SF_SNAPSHOT; 398 DIP_SET(ip, i_flags, ip->i_flags); 399 ip->i_flag |= IN_CHANGE | IN_UPDATE; 400 /* 401 * Ensure that the snapshot is completely on disk. 402 * Since we have marked it as a snapshot it is safe to 403 * unlock it as no process will be allowed to write to it. 404 */ 405 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)) != 0) 406 goto out; 407 VOP_UNLOCK(vp, 0); 408 /* 409 * All allocations are done, so we can now snapshot the system. 410 * 411 * Recind nice scheduling while running with the filesystem suspended. 412 */ 413 if (td->td_proc->p_nice > 0) { 414 struct proc *p; 415 416 p = td->td_proc; 417 PROC_LOCK(p); 418 saved_nice = p->p_nice; 419 sched_nice(p, 0); 420 PROC_UNLOCK(p); 421 } 422 /* 423 * Suspend operation on filesystem. 424 */ 425 for (;;) { 426 vn_finished_write(wrtmp); 427 if ((error = vfs_write_suspend(vp->v_mount, 0)) != 0) { 428 vn_start_write(NULL, &wrtmp, V_WAIT); 429 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 430 goto out; 431 } 432 if (mp->mnt_kern_flag & MNTK_SUSPENDED) 433 break; 434 vn_start_write(NULL, &wrtmp, V_WAIT); 435 } 436 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 437 if (ip->i_effnlink == 0) { 438 error = ENOENT; /* Snapshot file unlinked */ 439 goto out1; 440 } 441 if (collectsnapstats) 442 nanotime(&starttime); 443 444 /* The last block might have changed. Copy it again to be sure. */ 445 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)), 446 fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp); 447 if (error != 0) 448 goto out1; 449 error = readblock(vp, bp, numblks - 1); 450 bp->b_flags |= B_VALIDSUSPWRT; 451 bawrite(bp); 452 if (error != 0) 453 goto out1; 454 /* 455 * First, copy all the cylinder group maps that have changed. 456 */ 457 for (cg = 0; cg < fs->fs_ncg; cg++) { 458 if ((ACTIVECGNUM(fs, cg) & ACTIVECGOFF(cg)) != 0) 459 continue; 460 redo++; 461 error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)), 462 fs->fs_bsize, KERNCRED, 0, &nbp); 463 if (error) 464 goto out1; 465 error = cgaccount(cg, vp, nbp, 2); 466 bawrite(nbp); 467 if (error) 468 goto out1; 469 } 470 /* 471 * Grab a copy of the superblock and its summary information. 472 * We delay writing it until the suspension is released below. 473 */ 474 copy_fs = malloc((u_long)fs->fs_bsize, M_UFSMNT, M_WAITOK); 475 bcopy(fs, copy_fs, fs->fs_sbsize); 476 if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0) 477 copy_fs->fs_clean = 1; 478 size = fs->fs_bsize < SBLOCKSIZE ? fs->fs_bsize : SBLOCKSIZE; 479 if (fs->fs_sbsize < size) 480 bzero(&((char *)copy_fs)[fs->fs_sbsize], 481 size - fs->fs_sbsize); 482 size = blkroundup(fs, fs->fs_cssize); 483 if (fs->fs_contigsumsize > 0) 484 size += fs->fs_ncg * sizeof(int32_t); 485 space = malloc((u_long)size, M_UFSMNT, M_WAITOK); 486 copy_fs->fs_csp = space; 487 bcopy(fs->fs_csp, copy_fs->fs_csp, fs->fs_cssize); 488 space = (char *)space + fs->fs_cssize; 489 loc = howmany(fs->fs_cssize, fs->fs_fsize); 490 i = fs->fs_frag - loc % fs->fs_frag; 491 len = (i == fs->fs_frag) ? 0 : i * fs->fs_fsize; 492 if (len > 0) { 493 if ((error = bread(devvp, fsbtodb(fs, fs->fs_csaddr + loc), 494 len, KERNCRED, &bp)) != 0) { 495 brelse(bp); 496 free(copy_fs->fs_csp, M_UFSMNT); 497 free(copy_fs, M_UFSMNT); 498 copy_fs = NULL; 499 goto out1; 500 } 501 bcopy(bp->b_data, space, (u_int)len); 502 space = (char *)space + len; 503 bp->b_flags |= B_INVAL | B_NOCACHE; 504 brelse(bp); 505 } 506 if (fs->fs_contigsumsize > 0) { 507 copy_fs->fs_maxcluster = lp = space; 508 for (i = 0; i < fs->fs_ncg; i++) 509 *lp++ = fs->fs_contigsumsize; 510 } 511 /* 512 * We must check for active files that have been unlinked 513 * (e.g., with a zero link count). We have to expunge all 514 * trace of these files from the snapshot so that they are 515 * not reclaimed prematurely by fsck or unnecessarily dumped. 516 * We turn off the MNTK_SUSPENDED flag to avoid a panic from 517 * spec_strategy about writing on a suspended filesystem. 518 * Note that we skip unlinked snapshot files as they will 519 * be handled separately below. 520 * 521 * We also calculate the needed size for the snapshot list. 522 */ 523 snaplistsize = fs->fs_ncg + howmany(fs->fs_cssize, fs->fs_bsize) + 524 FSMAXSNAP + 1 /* superblock */ + 1 /* last block */ + 1 /* size */; 525 MNT_ILOCK(mp); 526 mp->mnt_kern_flag &= ~MNTK_SUSPENDED; 527 MNT_IUNLOCK(mp); 528 loop: 529 MNT_VNODE_FOREACH_ALL(xvp, mp, mvp) { 530 if ((xvp->v_usecount == 0 && 531 (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) || 532 xvp->v_type == VNON || 533 IS_SNAPSHOT(VTOI(xvp))) { 534 VI_UNLOCK(xvp); 535 continue; 536 } 537 /* 538 * We can skip parent directory vnode because it must have 539 * this snapshot file in it. 540 */ 541 if (xvp == nd.ni_dvp) { 542 VI_UNLOCK(xvp); 543 continue; 544 } 545 vholdl(xvp); 546 if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK) != 0) { 547 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 548 vdrop(xvp); 549 goto loop; 550 } 551 VI_LOCK(xvp); 552 if (xvp->v_usecount == 0 && 553 (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) { 554 VI_UNLOCK(xvp); 555 VOP_UNLOCK(xvp, 0); 556 vdrop(xvp); 557 continue; 558 } 559 VI_UNLOCK(xvp); 560 if (snapdebug) 561 vn_printf(xvp, "ffs_snapshot: busy vnode "); 562 if (VOP_GETATTR(xvp, &vat, td->td_ucred) == 0 && 563 vat.va_nlink > 0) { 564 VOP_UNLOCK(xvp, 0); 565 vdrop(xvp); 566 continue; 567 } 568 xp = VTOI(xvp); 569 if (ffs_checkfreefile(copy_fs, vp, xp->i_number)) { 570 VOP_UNLOCK(xvp, 0); 571 vdrop(xvp); 572 continue; 573 } 574 /* 575 * If there is a fragment, clear it here. 576 */ 577 blkno = 0; 578 loc = howmany(xp->i_size, fs->fs_bsize) - 1; 579 if (loc < UFS_NDADDR) { 580 len = fragroundup(fs, blkoff(fs, xp->i_size)); 581 if (len != 0 && len < fs->fs_bsize) { 582 ffs_blkfree(ump, copy_fs, vp, 583 DIP(xp, i_db[loc]), len, xp->i_number, 584 xvp->v_type, NULL); 585 blkno = DIP(xp, i_db[loc]); 586 DIP_SET(xp, i_db[loc], 0); 587 } 588 } 589 snaplistsize += 1; 590 if (I_IS_UFS1(xp)) 591 error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1, 592 BLK_NOCOPY, 1); 593 else 594 error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2, 595 BLK_NOCOPY, 1); 596 if (blkno) 597 DIP_SET(xp, i_db[loc], blkno); 598 if (!error) 599 error = ffs_freefile(ump, copy_fs, vp, xp->i_number, 600 xp->i_mode, NULL); 601 VOP_UNLOCK(xvp, 0); 602 vdrop(xvp); 603 if (error) { 604 free(copy_fs->fs_csp, M_UFSMNT); 605 free(copy_fs, M_UFSMNT); 606 copy_fs = NULL; 607 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 608 goto out1; 609 } 610 } 611 /* 612 * Erase the journal file from the snapshot. 613 */ 614 if (fs->fs_flags & FS_SUJ) { 615 error = softdep_journal_lookup(mp, &xvp); 616 if (error) { 617 free(copy_fs->fs_csp, M_UFSMNT); 618 free(copy_fs, M_UFSMNT); 619 copy_fs = NULL; 620 goto out1; 621 } 622 xp = VTOI(xvp); 623 if (I_IS_UFS1(xp)) 624 error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1, 625 BLK_NOCOPY, 0); 626 else 627 error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2, 628 BLK_NOCOPY, 0); 629 vput(xvp); 630 } 631 /* 632 * Acquire a lock on the snapdata structure, creating it if necessary. 633 */ 634 sn = ffs_snapdata_acquire(devvp); 635 /* 636 * Change vnode to use shared snapshot lock instead of the original 637 * private lock. 638 */ 639 vp->v_vnlock = &sn->sn_lock; 640 lockmgr(&vp->v_lock, LK_RELEASE, NULL); 641 xp = TAILQ_FIRST(&sn->sn_head); 642 /* 643 * If this is the first snapshot on this filesystem, then we need 644 * to allocate the space for the list of preallocated snapshot blocks. 645 * This list will be refined below, but this preliminary one will 646 * keep us out of deadlock until the full one is ready. 647 */ 648 if (xp == NULL) { 649 snapblklist = malloc(snaplistsize * sizeof(daddr_t), 650 M_UFSMNT, M_WAITOK); 651 blkp = &snapblklist[1]; 652 *blkp++ = lblkno(fs, fs->fs_sblockloc); 653 blkno = fragstoblks(fs, fs->fs_csaddr); 654 for (cg = 0; cg < fs->fs_ncg; cg++) { 655 if (fragstoblks(fs, cgtod(fs, cg) > blkno)) 656 break; 657 *blkp++ = fragstoblks(fs, cgtod(fs, cg)); 658 } 659 len = howmany(fs->fs_cssize, fs->fs_bsize); 660 for (loc = 0; loc < len; loc++) 661 *blkp++ = blkno + loc; 662 for (; cg < fs->fs_ncg; cg++) 663 *blkp++ = fragstoblks(fs, cgtod(fs, cg)); 664 snapblklist[0] = blkp - snapblklist; 665 VI_LOCK(devvp); 666 if (sn->sn_blklist != NULL) 667 panic("ffs_snapshot: non-empty list"); 668 sn->sn_blklist = snapblklist; 669 sn->sn_listsize = blkp - snapblklist; 670 VI_UNLOCK(devvp); 671 } 672 /* 673 * Record snapshot inode. Since this is the newest snapshot, 674 * it must be placed at the end of the list. 675 */ 676 VI_LOCK(devvp); 677 fs->fs_snapinum[snaploc] = ip->i_number; 678 if (ip->i_nextsnap.tqe_prev != 0) 679 panic("ffs_snapshot: %ju already on list", 680 (uintmax_t)ip->i_number); 681 TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap); 682 devvp->v_vflag |= VV_COPYONWRITE; 683 VI_UNLOCK(devvp); 684 ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp"); 685 out1: 686 KASSERT((sn != NULL && copy_fs != NULL && error == 0) || 687 (sn == NULL && copy_fs == NULL && error != 0), 688 ("email phk@ and mckusick@")); 689 /* 690 * Resume operation on filesystem. 691 */ 692 vfs_write_resume(vp->v_mount, VR_START_WRITE | VR_NO_SUSPCLR); 693 if (collectsnapstats && starttime.tv_sec > 0) { 694 nanotime(&endtime); 695 timespecsub(&endtime, &starttime); 696 printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n", 697 vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec, 698 endtime.tv_nsec / 1000000, redo, fs->fs_ncg); 699 } 700 if (copy_fs == NULL) 701 goto out; 702 /* 703 * Copy allocation information from all the snapshots in 704 * this snapshot and then expunge them from its view. 705 */ 706 TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) { 707 if (xp == ip) 708 break; 709 if (I_IS_UFS1(xp)) 710 error = expunge_ufs1(vp, xp, fs, snapacct_ufs1, 711 BLK_SNAP, 0); 712 else 713 error = expunge_ufs2(vp, xp, fs, snapacct_ufs2, 714 BLK_SNAP, 0); 715 if (error == 0 && xp->i_effnlink == 0) { 716 error = ffs_freefile(ump, 717 copy_fs, 718 vp, 719 xp->i_number, 720 xp->i_mode, NULL); 721 } 722 if (error) { 723 fs->fs_snapinum[snaploc] = 0; 724 goto done; 725 } 726 } 727 /* 728 * Allocate space for the full list of preallocated snapshot blocks. 729 */ 730 snapblklist = malloc(snaplistsize * sizeof(daddr_t), 731 M_UFSMNT, M_WAITOK); 732 ip->i_snapblklist = &snapblklist[1]; 733 /* 734 * Expunge the blocks used by the snapshots from the set of 735 * blocks marked as used in the snapshot bitmaps. Also, collect 736 * the list of allocated blocks in i_snapblklist. 737 */ 738 if (I_IS_UFS1(ip)) 739 error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1, 740 BLK_SNAP, 0); 741 else 742 error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2, 743 BLK_SNAP, 0); 744 if (error) { 745 fs->fs_snapinum[snaploc] = 0; 746 free(snapblklist, M_UFSMNT); 747 goto done; 748 } 749 if (snaplistsize < ip->i_snapblklist - snapblklist) 750 panic("ffs_snapshot: list too small"); 751 snaplistsize = ip->i_snapblklist - snapblklist; 752 snapblklist[0] = snaplistsize; 753 ip->i_snapblklist = 0; 754 /* 755 * Write out the list of allocated blocks to the end of the snapshot. 756 */ 757 auio.uio_iov = &aiov; 758 auio.uio_iovcnt = 1; 759 aiov.iov_base = (void *)snapblklist; 760 aiov.iov_len = snaplistsize * sizeof(daddr_t); 761 auio.uio_resid = aiov.iov_len; 762 auio.uio_offset = ip->i_size; 763 auio.uio_segflg = UIO_SYSSPACE; 764 auio.uio_rw = UIO_WRITE; 765 auio.uio_td = td; 766 if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) { 767 fs->fs_snapinum[snaploc] = 0; 768 free(snapblklist, M_UFSMNT); 769 goto done; 770 } 771 /* 772 * Write the superblock and its summary information 773 * to the snapshot. 774 */ 775 blkno = fragstoblks(fs, fs->fs_csaddr); 776 len = howmany(fs->fs_cssize, fs->fs_bsize); 777 space = copy_fs->fs_csp; 778 for (loc = 0; loc < len; loc++) { 779 error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp); 780 if (error) { 781 brelse(nbp); 782 fs->fs_snapinum[snaploc] = 0; 783 free(snapblklist, M_UFSMNT); 784 goto done; 785 } 786 bcopy(space, nbp->b_data, fs->fs_bsize); 787 space = (char *)space + fs->fs_bsize; 788 bawrite(nbp); 789 } 790 error = bread(vp, lblkno(fs, fs->fs_sblockloc), fs->fs_bsize, 791 KERNCRED, &nbp); 792 if (error) { 793 brelse(nbp); 794 } else { 795 loc = blkoff(fs, fs->fs_sblockloc); 796 bcopy((char *)copy_fs, &nbp->b_data[loc], (u_int)fs->fs_sbsize); 797 bawrite(nbp); 798 } 799 /* 800 * As this is the newest list, it is the most inclusive, so 801 * should replace the previous list. 802 */ 803 VI_LOCK(devvp); 804 space = sn->sn_blklist; 805 sn->sn_blklist = snapblklist; 806 sn->sn_listsize = snaplistsize; 807 VI_UNLOCK(devvp); 808 if (space != NULL) 809 free(space, M_UFSMNT); 810 /* 811 * Preallocate all the direct blocks in the snapshot inode so 812 * that we never have to write the inode itself to commit an 813 * update to the contents of the snapshot. Note that once 814 * created, the size of the snapshot will never change, so 815 * there will never be a need to write the inode except to 816 * update the non-integrity-critical time fields and 817 * allocated-block count. 818 */ 819 for (blockno = 0; blockno < UFS_NDADDR; blockno++) { 820 if (DIP(ip, i_db[blockno]) != 0) 821 continue; 822 error = UFS_BALLOC(vp, lblktosize(fs, blockno), 823 fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp); 824 if (error) 825 break; 826 error = readblock(vp, bp, blockno); 827 bawrite(bp); 828 if (error != 0) 829 break; 830 } 831 done: 832 free(copy_fs->fs_csp, M_UFSMNT); 833 free(copy_fs, M_UFSMNT); 834 copy_fs = NULL; 835 out: 836 NDFREE(&nd, NDF_ONLY_PNBUF); 837 if (saved_nice > 0) { 838 struct proc *p; 839 840 p = td->td_proc; 841 PROC_LOCK(p); 842 sched_nice(td->td_proc, saved_nice); 843 PROC_UNLOCK(td->td_proc); 844 } 845 UFS_LOCK(ump); 846 if (fs->fs_active != 0) { 847 free(fs->fs_active, M_DEVBUF); 848 fs->fs_active = 0; 849 } 850 UFS_UNLOCK(ump); 851 MNT_ILOCK(mp); 852 mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA); 853 MNT_IUNLOCK(mp); 854 if (error) 855 (void) ffs_truncate(vp, (off_t)0, 0, NOCRED); 856 (void) ffs_syncvnode(vp, MNT_WAIT, 0); 857 if (error) 858 vput(vp); 859 else 860 VOP_UNLOCK(vp, 0); 861 vrele(nd.ni_dvp); 862 vn_finished_write(wrtmp); 863 process_deferred_inactive(mp); 864 return (error); 865 } 866 867 /* 868 * Copy a cylinder group map. All the unallocated blocks are marked 869 * BLK_NOCOPY so that the snapshot knows that it need not copy them 870 * if they are later written. If passno is one, then this is a first 871 * pass, so only setting needs to be done. If passno is 2, then this 872 * is a revision to a previous pass which must be undone as the 873 * replacement pass is done. 874 */ 875 static int 876 cgaccount(cg, vp, nbp, passno) 877 int cg; 878 struct vnode *vp; 879 struct buf *nbp; 880 int passno; 881 { 882 struct buf *bp, *ibp; 883 struct inode *ip; 884 struct cg *cgp; 885 struct fs *fs; 886 ufs2_daddr_t base, numblks; 887 int error, len, loc, indiroff; 888 889 ip = VTOI(vp); 890 fs = ITOFS(ip); 891 error = bread(ITODEVVP(ip), fsbtodb(fs, cgtod(fs, cg)), 892 (int)fs->fs_cgsize, KERNCRED, &bp); 893 if (error) { 894 brelse(bp); 895 return (error); 896 } 897 cgp = (struct cg *)bp->b_data; 898 if (!cg_chkmagic(cgp)) { 899 brelse(bp); 900 return (EIO); 901 } 902 UFS_LOCK(ITOUMP(ip)); 903 ACTIVESET(fs, cg); 904 /* 905 * Recomputation of summary information might not have been performed 906 * at mount time. Sync up summary information for current cylinder 907 * group while data is in memory to ensure that result of background 908 * fsck is slightly more consistent. 909 */ 910 fs->fs_cs(fs, cg) = cgp->cg_cs; 911 UFS_UNLOCK(ITOUMP(ip)); 912 bcopy(bp->b_data, nbp->b_data, fs->fs_cgsize); 913 if (fs->fs_cgsize < fs->fs_bsize) 914 bzero(&nbp->b_data[fs->fs_cgsize], 915 fs->fs_bsize - fs->fs_cgsize); 916 cgp = (struct cg *)nbp->b_data; 917 bqrelse(bp); 918 if (passno == 2) 919 nbp->b_flags |= B_VALIDSUSPWRT; 920 numblks = howmany(fs->fs_size, fs->fs_frag); 921 len = howmany(fs->fs_fpg, fs->fs_frag); 922 base = cgbase(fs, cg) / fs->fs_frag; 923 if (base + len >= numblks) 924 len = numblks - base - 1; 925 loc = 0; 926 if (base < UFS_NDADDR) { 927 for ( ; loc < UFS_NDADDR; loc++) { 928 if (ffs_isblock(fs, cg_blksfree(cgp), loc)) 929 DIP_SET(ip, i_db[loc], BLK_NOCOPY); 930 else if (passno == 2 && DIP(ip, i_db[loc])== BLK_NOCOPY) 931 DIP_SET(ip, i_db[loc], 0); 932 else if (passno == 1 && DIP(ip, i_db[loc])== BLK_NOCOPY) 933 panic("ffs_snapshot: lost direct block"); 934 } 935 } 936 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(base + loc)), 937 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 938 if (error) { 939 return (error); 940 } 941 indiroff = (base + loc - UFS_NDADDR) % NINDIR(fs); 942 for ( ; loc < len; loc++, indiroff++) { 943 if (indiroff >= NINDIR(fs)) { 944 if (passno == 2) 945 ibp->b_flags |= B_VALIDSUSPWRT; 946 bawrite(ibp); 947 error = UFS_BALLOC(vp, 948 lblktosize(fs, (off_t)(base + loc)), 949 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 950 if (error) { 951 return (error); 952 } 953 indiroff = 0; 954 } 955 if (I_IS_UFS1(ip)) { 956 if (ffs_isblock(fs, cg_blksfree(cgp), loc)) 957 ((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 958 BLK_NOCOPY; 959 else if (passno == 2 && ((ufs1_daddr_t *)(ibp->b_data)) 960 [indiroff] == BLK_NOCOPY) 961 ((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 0; 962 else if (passno == 1 && ((ufs1_daddr_t *)(ibp->b_data)) 963 [indiroff] == BLK_NOCOPY) 964 panic("ffs_snapshot: lost indirect block"); 965 continue; 966 } 967 if (ffs_isblock(fs, cg_blksfree(cgp), loc)) 968 ((ufs2_daddr_t *)(ibp->b_data))[indiroff] = BLK_NOCOPY; 969 else if (passno == 2 && 970 ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY) 971 ((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 0; 972 else if (passno == 1 && 973 ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY) 974 panic("ffs_snapshot: lost indirect block"); 975 } 976 if (passno == 2) 977 ibp->b_flags |= B_VALIDSUSPWRT; 978 bdwrite(ibp); 979 return (0); 980 } 981 982 /* 983 * Before expunging a snapshot inode, note all the 984 * blocks that it claims with BLK_SNAP so that fsck will 985 * be able to account for those blocks properly and so 986 * that this snapshot knows that it need not copy them 987 * if the other snapshot holding them is freed. This code 988 * is reproduced once each for UFS1 and UFS2. 989 */ 990 static int 991 expunge_ufs1(snapvp, cancelip, fs, acctfunc, expungetype, clearmode) 992 struct vnode *snapvp; 993 struct inode *cancelip; 994 struct fs *fs; 995 int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, 996 struct fs *, ufs_lbn_t, int); 997 int expungetype; 998 int clearmode; 999 { 1000 int i, error, indiroff; 1001 ufs_lbn_t lbn, rlbn; 1002 ufs2_daddr_t len, blkno, numblks, blksperindir; 1003 struct ufs1_dinode *dip; 1004 struct thread *td = curthread; 1005 struct buf *bp; 1006 1007 /* 1008 * Prepare to expunge the inode. If its inode block has not 1009 * yet been copied, then allocate and fill the copy. 1010 */ 1011 lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number)); 1012 blkno = 0; 1013 if (lbn < UFS_NDADDR) { 1014 blkno = VTOI(snapvp)->i_din1->di_db[lbn]; 1015 } else { 1016 if (DOINGSOFTDEP(snapvp)) 1017 softdep_prealloc(snapvp, MNT_WAIT); 1018 td->td_pflags |= TDP_COWINPROGRESS; 1019 error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn), 1020 fs->fs_bsize, KERNCRED, BA_METAONLY, &bp); 1021 td->td_pflags &= ~TDP_COWINPROGRESS; 1022 if (error) 1023 return (error); 1024 indiroff = (lbn - UFS_NDADDR) % NINDIR(fs); 1025 blkno = ((ufs1_daddr_t *)(bp->b_data))[indiroff]; 1026 bqrelse(bp); 1027 } 1028 if (blkno != 0) { 1029 if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp))) 1030 return (error); 1031 } else { 1032 error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn), 1033 fs->fs_bsize, KERNCRED, 0, &bp); 1034 if (error) 1035 return (error); 1036 if ((error = readblock(snapvp, bp, lbn)) != 0) 1037 return (error); 1038 } 1039 /* 1040 * Set a snapshot inode to be a zero length file, regular files 1041 * or unlinked snapshots to be completely unallocated. 1042 */ 1043 dip = (struct ufs1_dinode *)bp->b_data + 1044 ino_to_fsbo(fs, cancelip->i_number); 1045 if (clearmode || cancelip->i_effnlink == 0) 1046 dip->di_mode = 0; 1047 dip->di_size = 0; 1048 dip->di_blocks = 0; 1049 dip->di_flags &= ~SF_SNAPSHOT; 1050 bzero(&dip->di_db[0], (UFS_NDADDR + UFS_NIADDR) * sizeof(ufs1_daddr_t)); 1051 bdwrite(bp); 1052 /* 1053 * Now go through and expunge all the blocks in the file 1054 * using the function requested. 1055 */ 1056 numblks = howmany(cancelip->i_size, fs->fs_bsize); 1057 if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_db[0], 1058 &cancelip->i_din1->di_db[UFS_NDADDR], fs, 0, expungetype))) 1059 return (error); 1060 if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_ib[0], 1061 &cancelip->i_din1->di_ib[UFS_NIADDR], fs, -1, expungetype))) 1062 return (error); 1063 blksperindir = 1; 1064 lbn = -UFS_NDADDR; 1065 len = numblks - UFS_NDADDR; 1066 rlbn = UFS_NDADDR; 1067 for (i = 0; len > 0 && i < UFS_NIADDR; i++) { 1068 error = indiracct_ufs1(snapvp, ITOV(cancelip), i, 1069 cancelip->i_din1->di_ib[i], lbn, rlbn, len, 1070 blksperindir, fs, acctfunc, expungetype); 1071 if (error) 1072 return (error); 1073 blksperindir *= NINDIR(fs); 1074 lbn -= blksperindir + 1; 1075 len -= blksperindir; 1076 rlbn += blksperindir; 1077 } 1078 return (0); 1079 } 1080 1081 /* 1082 * Descend an indirect block chain for vnode cancelvp accounting for all 1083 * its indirect blocks in snapvp. 1084 */ 1085 static int 1086 indiracct_ufs1(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks, 1087 blksperindir, fs, acctfunc, expungetype) 1088 struct vnode *snapvp; 1089 struct vnode *cancelvp; 1090 int level; 1091 ufs1_daddr_t blkno; 1092 ufs_lbn_t lbn; 1093 ufs_lbn_t rlbn; 1094 ufs_lbn_t remblks; 1095 ufs_lbn_t blksperindir; 1096 struct fs *fs; 1097 int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, 1098 struct fs *, ufs_lbn_t, int); 1099 int expungetype; 1100 { 1101 int error, num, i; 1102 ufs_lbn_t subblksperindir; 1103 struct indir indirs[UFS_NIADDR + 2]; 1104 ufs1_daddr_t last, *bap; 1105 struct buf *bp; 1106 1107 if (blkno == 0) { 1108 if (expungetype == BLK_NOCOPY) 1109 return (0); 1110 panic("indiracct_ufs1: missing indir"); 1111 } 1112 if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0) 1113 return (error); 1114 if (lbn != indirs[num - 1 - level].in_lbn || num < 2) 1115 panic("indiracct_ufs1: botched params"); 1116 /* 1117 * We have to expand bread here since it will deadlock looking 1118 * up the block number for any blocks that are not in the cache. 1119 */ 1120 bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0); 1121 bp->b_blkno = fsbtodb(fs, blkno); 1122 if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 && 1123 (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) { 1124 brelse(bp); 1125 return (error); 1126 } 1127 /* 1128 * Account for the block pointers in this indirect block. 1129 */ 1130 last = howmany(remblks, blksperindir); 1131 if (last > NINDIR(fs)) 1132 last = NINDIR(fs); 1133 bap = malloc(fs->fs_bsize, M_DEVBUF, M_WAITOK); 1134 bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize); 1135 bqrelse(bp); 1136 error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs, 1137 level == 0 ? rlbn : -1, expungetype); 1138 if (error || level == 0) 1139 goto out; 1140 /* 1141 * Account for the block pointers in each of the indirect blocks 1142 * in the levels below us. 1143 */ 1144 subblksperindir = blksperindir / NINDIR(fs); 1145 for (lbn++, level--, i = 0; i < last; i++) { 1146 error = indiracct_ufs1(snapvp, cancelvp, level, bap[i], lbn, 1147 rlbn, remblks, subblksperindir, fs, acctfunc, expungetype); 1148 if (error) 1149 goto out; 1150 rlbn += blksperindir; 1151 lbn -= blksperindir; 1152 remblks -= blksperindir; 1153 } 1154 out: 1155 free(bap, M_DEVBUF); 1156 return (error); 1157 } 1158 1159 /* 1160 * Do both snap accounting and map accounting. 1161 */ 1162 static int 1163 fullacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype) 1164 struct vnode *vp; 1165 ufs1_daddr_t *oldblkp, *lastblkp; 1166 struct fs *fs; 1167 ufs_lbn_t lblkno; 1168 int exptype; /* BLK_SNAP or BLK_NOCOPY */ 1169 { 1170 int error; 1171 1172 if ((error = snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype))) 1173 return (error); 1174 return (mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)); 1175 } 1176 1177 /* 1178 * Identify a set of blocks allocated in a snapshot inode. 1179 */ 1180 static int 1181 snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype) 1182 struct vnode *vp; 1183 ufs1_daddr_t *oldblkp, *lastblkp; 1184 struct fs *fs; 1185 ufs_lbn_t lblkno; 1186 int expungetype; /* BLK_SNAP or BLK_NOCOPY */ 1187 { 1188 struct inode *ip = VTOI(vp); 1189 ufs1_daddr_t blkno, *blkp; 1190 ufs_lbn_t lbn; 1191 struct buf *ibp; 1192 int error; 1193 1194 for ( ; oldblkp < lastblkp; oldblkp++) { 1195 blkno = *oldblkp; 1196 if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP) 1197 continue; 1198 lbn = fragstoblks(fs, blkno); 1199 if (lbn < UFS_NDADDR) { 1200 blkp = &ip->i_din1->di_db[lbn]; 1201 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1202 } else { 1203 error = ffs_balloc_ufs1(vp, lblktosize(fs, (off_t)lbn), 1204 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 1205 if (error) 1206 return (error); 1207 blkp = &((ufs1_daddr_t *)(ibp->b_data)) 1208 [(lbn - UFS_NDADDR) % NINDIR(fs)]; 1209 } 1210 /* 1211 * If we are expunging a snapshot vnode and we 1212 * find a block marked BLK_NOCOPY, then it is 1213 * one that has been allocated to this snapshot after 1214 * we took our current snapshot and can be ignored. 1215 */ 1216 if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) { 1217 if (lbn >= UFS_NDADDR) 1218 brelse(ibp); 1219 } else { 1220 if (*blkp != 0) 1221 panic("snapacct_ufs1: bad block"); 1222 *blkp = expungetype; 1223 if (lbn >= UFS_NDADDR) 1224 bdwrite(ibp); 1225 } 1226 } 1227 return (0); 1228 } 1229 1230 /* 1231 * Account for a set of blocks allocated in a snapshot inode. 1232 */ 1233 static int 1234 mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype) 1235 struct vnode *vp; 1236 ufs1_daddr_t *oldblkp, *lastblkp; 1237 struct fs *fs; 1238 ufs_lbn_t lblkno; 1239 int expungetype; 1240 { 1241 ufs1_daddr_t blkno; 1242 struct inode *ip; 1243 ino_t inum; 1244 int acctit; 1245 1246 ip = VTOI(vp); 1247 inum = ip->i_number; 1248 if (lblkno == -1) 1249 acctit = 0; 1250 else 1251 acctit = 1; 1252 for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) { 1253 blkno = *oldblkp; 1254 if (blkno == 0 || blkno == BLK_NOCOPY) 1255 continue; 1256 if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP) 1257 *ip->i_snapblklist++ = lblkno; 1258 if (blkno == BLK_SNAP) 1259 blkno = blkstofrags(fs, lblkno); 1260 ffs_blkfree(ITOUMP(ip), fs, vp, blkno, fs->fs_bsize, inum, 1261 vp->v_type, NULL); 1262 } 1263 return (0); 1264 } 1265 1266 /* 1267 * Before expunging a snapshot inode, note all the 1268 * blocks that it claims with BLK_SNAP so that fsck will 1269 * be able to account for those blocks properly and so 1270 * that this snapshot knows that it need not copy them 1271 * if the other snapshot holding them is freed. This code 1272 * is reproduced once each for UFS1 and UFS2. 1273 */ 1274 static int 1275 expunge_ufs2(snapvp, cancelip, fs, acctfunc, expungetype, clearmode) 1276 struct vnode *snapvp; 1277 struct inode *cancelip; 1278 struct fs *fs; 1279 int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, 1280 struct fs *, ufs_lbn_t, int); 1281 int expungetype; 1282 int clearmode; 1283 { 1284 int i, error, indiroff; 1285 ufs_lbn_t lbn, rlbn; 1286 ufs2_daddr_t len, blkno, numblks, blksperindir; 1287 struct ufs2_dinode *dip; 1288 struct thread *td = curthread; 1289 struct buf *bp; 1290 1291 /* 1292 * Prepare to expunge the inode. If its inode block has not 1293 * yet been copied, then allocate and fill the copy. 1294 */ 1295 lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number)); 1296 blkno = 0; 1297 if (lbn < UFS_NDADDR) { 1298 blkno = VTOI(snapvp)->i_din2->di_db[lbn]; 1299 } else { 1300 if (DOINGSOFTDEP(snapvp)) 1301 softdep_prealloc(snapvp, MNT_WAIT); 1302 td->td_pflags |= TDP_COWINPROGRESS; 1303 error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn), 1304 fs->fs_bsize, KERNCRED, BA_METAONLY, &bp); 1305 td->td_pflags &= ~TDP_COWINPROGRESS; 1306 if (error) 1307 return (error); 1308 indiroff = (lbn - UFS_NDADDR) % NINDIR(fs); 1309 blkno = ((ufs2_daddr_t *)(bp->b_data))[indiroff]; 1310 bqrelse(bp); 1311 } 1312 if (blkno != 0) { 1313 if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp))) 1314 return (error); 1315 } else { 1316 error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn), 1317 fs->fs_bsize, KERNCRED, 0, &bp); 1318 if (error) 1319 return (error); 1320 if ((error = readblock(snapvp, bp, lbn)) != 0) 1321 return (error); 1322 } 1323 /* 1324 * Set a snapshot inode to be a zero length file, regular files 1325 * to be completely unallocated. 1326 */ 1327 dip = (struct ufs2_dinode *)bp->b_data + 1328 ino_to_fsbo(fs, cancelip->i_number); 1329 if (clearmode || cancelip->i_effnlink == 0) 1330 dip->di_mode = 0; 1331 dip->di_size = 0; 1332 dip->di_blocks = 0; 1333 dip->di_flags &= ~SF_SNAPSHOT; 1334 bzero(&dip->di_db[0], (UFS_NDADDR + UFS_NIADDR) * sizeof(ufs2_daddr_t)); 1335 bdwrite(bp); 1336 /* 1337 * Now go through and expunge all the blocks in the file 1338 * using the function requested. 1339 */ 1340 numblks = howmany(cancelip->i_size, fs->fs_bsize); 1341 if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_db[0], 1342 &cancelip->i_din2->di_db[UFS_NDADDR], fs, 0, expungetype))) 1343 return (error); 1344 if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_ib[0], 1345 &cancelip->i_din2->di_ib[UFS_NIADDR], fs, -1, expungetype))) 1346 return (error); 1347 blksperindir = 1; 1348 lbn = -UFS_NDADDR; 1349 len = numblks - UFS_NDADDR; 1350 rlbn = UFS_NDADDR; 1351 for (i = 0; len > 0 && i < UFS_NIADDR; i++) { 1352 error = indiracct_ufs2(snapvp, ITOV(cancelip), i, 1353 cancelip->i_din2->di_ib[i], lbn, rlbn, len, 1354 blksperindir, fs, acctfunc, expungetype); 1355 if (error) 1356 return (error); 1357 blksperindir *= NINDIR(fs); 1358 lbn -= blksperindir + 1; 1359 len -= blksperindir; 1360 rlbn += blksperindir; 1361 } 1362 return (0); 1363 } 1364 1365 /* 1366 * Descend an indirect block chain for vnode cancelvp accounting for all 1367 * its indirect blocks in snapvp. 1368 */ 1369 static int 1370 indiracct_ufs2(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks, 1371 blksperindir, fs, acctfunc, expungetype) 1372 struct vnode *snapvp; 1373 struct vnode *cancelvp; 1374 int level; 1375 ufs2_daddr_t blkno; 1376 ufs_lbn_t lbn; 1377 ufs_lbn_t rlbn; 1378 ufs_lbn_t remblks; 1379 ufs_lbn_t blksperindir; 1380 struct fs *fs; 1381 int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, 1382 struct fs *, ufs_lbn_t, int); 1383 int expungetype; 1384 { 1385 int error, num, i; 1386 ufs_lbn_t subblksperindir; 1387 struct indir indirs[UFS_NIADDR + 2]; 1388 ufs2_daddr_t last, *bap; 1389 struct buf *bp; 1390 1391 if (blkno == 0) { 1392 if (expungetype == BLK_NOCOPY) 1393 return (0); 1394 panic("indiracct_ufs2: missing indir"); 1395 } 1396 if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0) 1397 return (error); 1398 if (lbn != indirs[num - 1 - level].in_lbn || num < 2) 1399 panic("indiracct_ufs2: botched params"); 1400 /* 1401 * We have to expand bread here since it will deadlock looking 1402 * up the block number for any blocks that are not in the cache. 1403 */ 1404 bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0); 1405 bp->b_blkno = fsbtodb(fs, blkno); 1406 if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 && 1407 (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) { 1408 brelse(bp); 1409 return (error); 1410 } 1411 /* 1412 * Account for the block pointers in this indirect block. 1413 */ 1414 last = howmany(remblks, blksperindir); 1415 if (last > NINDIR(fs)) 1416 last = NINDIR(fs); 1417 bap = malloc(fs->fs_bsize, M_DEVBUF, M_WAITOK); 1418 bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize); 1419 bqrelse(bp); 1420 error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs, 1421 level == 0 ? rlbn : -1, expungetype); 1422 if (error || level == 0) 1423 goto out; 1424 /* 1425 * Account for the block pointers in each of the indirect blocks 1426 * in the levels below us. 1427 */ 1428 subblksperindir = blksperindir / NINDIR(fs); 1429 for (lbn++, level--, i = 0; i < last; i++) { 1430 error = indiracct_ufs2(snapvp, cancelvp, level, bap[i], lbn, 1431 rlbn, remblks, subblksperindir, fs, acctfunc, expungetype); 1432 if (error) 1433 goto out; 1434 rlbn += blksperindir; 1435 lbn -= blksperindir; 1436 remblks -= blksperindir; 1437 } 1438 out: 1439 free(bap, M_DEVBUF); 1440 return (error); 1441 } 1442 1443 /* 1444 * Do both snap accounting and map accounting. 1445 */ 1446 static int 1447 fullacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype) 1448 struct vnode *vp; 1449 ufs2_daddr_t *oldblkp, *lastblkp; 1450 struct fs *fs; 1451 ufs_lbn_t lblkno; 1452 int exptype; /* BLK_SNAP or BLK_NOCOPY */ 1453 { 1454 int error; 1455 1456 if ((error = snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype))) 1457 return (error); 1458 return (mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)); 1459 } 1460 1461 /* 1462 * Identify a set of blocks allocated in a snapshot inode. 1463 */ 1464 static int 1465 snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype) 1466 struct vnode *vp; 1467 ufs2_daddr_t *oldblkp, *lastblkp; 1468 struct fs *fs; 1469 ufs_lbn_t lblkno; 1470 int expungetype; /* BLK_SNAP or BLK_NOCOPY */ 1471 { 1472 struct inode *ip = VTOI(vp); 1473 ufs2_daddr_t blkno, *blkp; 1474 ufs_lbn_t lbn; 1475 struct buf *ibp; 1476 int error; 1477 1478 for ( ; oldblkp < lastblkp; oldblkp++) { 1479 blkno = *oldblkp; 1480 if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP) 1481 continue; 1482 lbn = fragstoblks(fs, blkno); 1483 if (lbn < UFS_NDADDR) { 1484 blkp = &ip->i_din2->di_db[lbn]; 1485 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1486 } else { 1487 error = ffs_balloc_ufs2(vp, lblktosize(fs, (off_t)lbn), 1488 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 1489 if (error) 1490 return (error); 1491 blkp = &((ufs2_daddr_t *)(ibp->b_data)) 1492 [(lbn - UFS_NDADDR) % NINDIR(fs)]; 1493 } 1494 /* 1495 * If we are expunging a snapshot vnode and we 1496 * find a block marked BLK_NOCOPY, then it is 1497 * one that has been allocated to this snapshot after 1498 * we took our current snapshot and can be ignored. 1499 */ 1500 if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) { 1501 if (lbn >= UFS_NDADDR) 1502 brelse(ibp); 1503 } else { 1504 if (*blkp != 0) 1505 panic("snapacct_ufs2: bad block"); 1506 *blkp = expungetype; 1507 if (lbn >= UFS_NDADDR) 1508 bdwrite(ibp); 1509 } 1510 } 1511 return (0); 1512 } 1513 1514 /* 1515 * Account for a set of blocks allocated in a snapshot inode. 1516 */ 1517 static int 1518 mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype) 1519 struct vnode *vp; 1520 ufs2_daddr_t *oldblkp, *lastblkp; 1521 struct fs *fs; 1522 ufs_lbn_t lblkno; 1523 int expungetype; 1524 { 1525 ufs2_daddr_t blkno; 1526 struct inode *ip; 1527 ino_t inum; 1528 int acctit; 1529 1530 ip = VTOI(vp); 1531 inum = ip->i_number; 1532 if (lblkno == -1) 1533 acctit = 0; 1534 else 1535 acctit = 1; 1536 for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) { 1537 blkno = *oldblkp; 1538 if (blkno == 0 || blkno == BLK_NOCOPY) 1539 continue; 1540 if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP) 1541 *ip->i_snapblklist++ = lblkno; 1542 if (blkno == BLK_SNAP) 1543 blkno = blkstofrags(fs, lblkno); 1544 ffs_blkfree(ITOUMP(ip), fs, vp, blkno, fs->fs_bsize, inum, 1545 vp->v_type, NULL); 1546 } 1547 return (0); 1548 } 1549 1550 /* 1551 * Decrement extra reference on snapshot when last name is removed. 1552 * It will not be freed until the last open reference goes away. 1553 */ 1554 void 1555 ffs_snapgone(ip) 1556 struct inode *ip; 1557 { 1558 struct inode *xp; 1559 struct fs *fs; 1560 int snaploc; 1561 struct snapdata *sn; 1562 struct ufsmount *ump; 1563 1564 /* 1565 * Find snapshot in incore list. 1566 */ 1567 xp = NULL; 1568 sn = ITODEVVP(ip)->v_rdev->si_snapdata; 1569 if (sn != NULL) 1570 TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) 1571 if (xp == ip) 1572 break; 1573 if (xp != NULL) 1574 vrele(ITOV(ip)); 1575 else if (snapdebug) 1576 printf("ffs_snapgone: lost snapshot vnode %ju\n", 1577 (uintmax_t)ip->i_number); 1578 /* 1579 * Delete snapshot inode from superblock. Keep list dense. 1580 */ 1581 ump = ITOUMP(ip); 1582 fs = ump->um_fs; 1583 UFS_LOCK(ump); 1584 for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) 1585 if (fs->fs_snapinum[snaploc] == ip->i_number) 1586 break; 1587 if (snaploc < FSMAXSNAP) { 1588 for (snaploc++; snaploc < FSMAXSNAP; snaploc++) { 1589 if (fs->fs_snapinum[snaploc] == 0) 1590 break; 1591 fs->fs_snapinum[snaploc - 1] = fs->fs_snapinum[snaploc]; 1592 } 1593 fs->fs_snapinum[snaploc - 1] = 0; 1594 } 1595 UFS_UNLOCK(ump); 1596 } 1597 1598 /* 1599 * Prepare a snapshot file for being removed. 1600 */ 1601 void 1602 ffs_snapremove(vp) 1603 struct vnode *vp; 1604 { 1605 struct inode *ip; 1606 struct vnode *devvp; 1607 struct buf *ibp; 1608 struct fs *fs; 1609 ufs2_daddr_t numblks, blkno, dblk; 1610 int error, loc, last; 1611 struct snapdata *sn; 1612 1613 ip = VTOI(vp); 1614 fs = ITOFS(ip); 1615 devvp = ITODEVVP(ip); 1616 /* 1617 * If active, delete from incore list (this snapshot may 1618 * already have been in the process of being deleted, so 1619 * would not have been active). 1620 * 1621 * Clear copy-on-write flag if last snapshot. 1622 */ 1623 VI_LOCK(devvp); 1624 if (ip->i_nextsnap.tqe_prev != 0) { 1625 sn = devvp->v_rdev->si_snapdata; 1626 TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap); 1627 ip->i_nextsnap.tqe_prev = 0; 1628 VI_UNLOCK(devvp); 1629 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL); 1630 KASSERT(vp->v_vnlock == &sn->sn_lock, 1631 ("ffs_snapremove: lost lock mutation")); 1632 vp->v_vnlock = &vp->v_lock; 1633 VI_LOCK(devvp); 1634 lockmgr(&sn->sn_lock, LK_RELEASE, NULL); 1635 try_free_snapdata(devvp); 1636 } else 1637 VI_UNLOCK(devvp); 1638 /* 1639 * Clear all BLK_NOCOPY fields. Pass any block claims to other 1640 * snapshots that want them (see ffs_snapblkfree below). 1641 */ 1642 for (blkno = 1; blkno < UFS_NDADDR; blkno++) { 1643 dblk = DIP(ip, i_db[blkno]); 1644 if (dblk == 0) 1645 continue; 1646 if (dblk == BLK_NOCOPY || dblk == BLK_SNAP) 1647 DIP_SET(ip, i_db[blkno], 0); 1648 else if ((dblk == blkstofrags(fs, blkno) && 1649 ffs_snapblkfree(fs, ITODEVVP(ip), dblk, fs->fs_bsize, 1650 ip->i_number, vp->v_type, NULL))) { 1651 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - 1652 btodb(fs->fs_bsize)); 1653 DIP_SET(ip, i_db[blkno], 0); 1654 } 1655 } 1656 numblks = howmany(ip->i_size, fs->fs_bsize); 1657 for (blkno = UFS_NDADDR; blkno < numblks; blkno += NINDIR(fs)) { 1658 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno), 1659 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 1660 if (error) 1661 continue; 1662 if (fs->fs_size - blkno > NINDIR(fs)) 1663 last = NINDIR(fs); 1664 else 1665 last = fs->fs_size - blkno; 1666 for (loc = 0; loc < last; loc++) { 1667 if (I_IS_UFS1(ip)) { 1668 dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc]; 1669 if (dblk == 0) 1670 continue; 1671 if (dblk == BLK_NOCOPY || dblk == BLK_SNAP) 1672 ((ufs1_daddr_t *)(ibp->b_data))[loc]= 0; 1673 else if ((dblk == blkstofrags(fs, blkno) && 1674 ffs_snapblkfree(fs, ITODEVVP(ip), dblk, 1675 fs->fs_bsize, ip->i_number, vp->v_type, 1676 NULL))) { 1677 ip->i_din1->di_blocks -= 1678 btodb(fs->fs_bsize); 1679 ((ufs1_daddr_t *)(ibp->b_data))[loc]= 0; 1680 } 1681 continue; 1682 } 1683 dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc]; 1684 if (dblk == 0) 1685 continue; 1686 if (dblk == BLK_NOCOPY || dblk == BLK_SNAP) 1687 ((ufs2_daddr_t *)(ibp->b_data))[loc] = 0; 1688 else if ((dblk == blkstofrags(fs, blkno) && 1689 ffs_snapblkfree(fs, ITODEVVP(ip), dblk, 1690 fs->fs_bsize, ip->i_number, vp->v_type, NULL))) { 1691 ip->i_din2->di_blocks -= btodb(fs->fs_bsize); 1692 ((ufs2_daddr_t *)(ibp->b_data))[loc] = 0; 1693 } 1694 } 1695 bawrite(ibp); 1696 } 1697 /* 1698 * Clear snapshot flag and drop reference. 1699 */ 1700 ip->i_flags &= ~SF_SNAPSHOT; 1701 DIP_SET(ip, i_flags, ip->i_flags); 1702 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1703 /* 1704 * The dirtied indirects must be written out before 1705 * softdep_setup_freeblocks() is called. Otherwise indir_trunc() 1706 * may find indirect pointers using the magic BLK_* values. 1707 */ 1708 if (DOINGSOFTDEP(vp)) 1709 ffs_syncvnode(vp, MNT_WAIT, 0); 1710 #ifdef QUOTA 1711 /* 1712 * Reenable disk quotas for ex-snapshot file. 1713 */ 1714 if (!getinoquota(ip)) 1715 (void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE); 1716 #endif 1717 } 1718 1719 /* 1720 * Notification that a block is being freed. Return zero if the free 1721 * should be allowed to proceed. Return non-zero if the snapshot file 1722 * wants to claim the block. The block will be claimed if it is an 1723 * uncopied part of one of the snapshots. It will be freed if it is 1724 * either a BLK_NOCOPY or has already been copied in all of the snapshots. 1725 * If a fragment is being freed, then all snapshots that care about 1726 * it must make a copy since a snapshot file can only claim full sized 1727 * blocks. Note that if more than one snapshot file maps the block, 1728 * we can pick one at random to claim it. Since none of the snapshots 1729 * can change, we are assurred that they will all see the same unmodified 1730 * image. When deleting a snapshot file (see ffs_snapremove above), we 1731 * must push any of these claimed blocks to one of the other snapshots 1732 * that maps it. These claimed blocks are easily identified as they will 1733 * have a block number equal to their logical block number within the 1734 * snapshot. A copied block can never have this property because they 1735 * must always have been allocated from a BLK_NOCOPY location. 1736 */ 1737 int 1738 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, wkhd) 1739 struct fs *fs; 1740 struct vnode *devvp; 1741 ufs2_daddr_t bno; 1742 long size; 1743 ino_t inum; 1744 enum vtype vtype; 1745 struct workhead *wkhd; 1746 { 1747 struct buf *ibp, *cbp, *savedcbp = NULL; 1748 struct thread *td = curthread; 1749 struct inode *ip; 1750 struct vnode *vp = NULL; 1751 ufs_lbn_t lbn; 1752 ufs2_daddr_t blkno; 1753 int indiroff = 0, error = 0, claimedblk = 0; 1754 struct snapdata *sn; 1755 1756 lbn = fragstoblks(fs, bno); 1757 retry: 1758 VI_LOCK(devvp); 1759 sn = devvp->v_rdev->si_snapdata; 1760 if (sn == NULL) { 1761 VI_UNLOCK(devvp); 1762 return (0); 1763 } 1764 if (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL, 1765 VI_MTX(devvp)) != 0) 1766 goto retry; 1767 TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) { 1768 vp = ITOV(ip); 1769 if (DOINGSOFTDEP(vp)) 1770 softdep_prealloc(vp, MNT_WAIT); 1771 /* 1772 * Lookup block being written. 1773 */ 1774 if (lbn < UFS_NDADDR) { 1775 blkno = DIP(ip, i_db[lbn]); 1776 } else { 1777 td->td_pflags |= TDP_COWINPROGRESS; 1778 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn), 1779 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 1780 td->td_pflags &= ~TDP_COWINPROGRESS; 1781 if (error) 1782 break; 1783 indiroff = (lbn - UFS_NDADDR) % NINDIR(fs); 1784 if (I_IS_UFS1(ip)) 1785 blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff]; 1786 else 1787 blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff]; 1788 } 1789 /* 1790 * Check to see if block needs to be copied. 1791 */ 1792 if (blkno == 0) { 1793 /* 1794 * A block that we map is being freed. If it has not 1795 * been claimed yet, we will claim or copy it (below). 1796 */ 1797 claimedblk = 1; 1798 } else if (blkno == BLK_SNAP) { 1799 /* 1800 * No previous snapshot claimed the block, 1801 * so it will be freed and become a BLK_NOCOPY 1802 * (don't care) for us. 1803 */ 1804 if (claimedblk) 1805 panic("snapblkfree: inconsistent block type"); 1806 if (lbn < UFS_NDADDR) { 1807 DIP_SET(ip, i_db[lbn], BLK_NOCOPY); 1808 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1809 } else if (I_IS_UFS1(ip)) { 1810 ((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 1811 BLK_NOCOPY; 1812 bdwrite(ibp); 1813 } else { 1814 ((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 1815 BLK_NOCOPY; 1816 bdwrite(ibp); 1817 } 1818 continue; 1819 } else /* BLK_NOCOPY or default */ { 1820 /* 1821 * If the snapshot has already copied the block 1822 * (default), or does not care about the block, 1823 * it is not needed. 1824 */ 1825 if (lbn >= UFS_NDADDR) 1826 bqrelse(ibp); 1827 continue; 1828 } 1829 /* 1830 * If this is a full size block, we will just grab it 1831 * and assign it to the snapshot inode. Otherwise we 1832 * will proceed to copy it. See explanation for this 1833 * routine as to why only a single snapshot needs to 1834 * claim this block. 1835 */ 1836 if (size == fs->fs_bsize) { 1837 #ifdef DEBUG 1838 if (snapdebug) 1839 printf("%s %ju lbn %jd from inum %ju\n", 1840 "Grabonremove: snapino", 1841 (uintmax_t)ip->i_number, 1842 (intmax_t)lbn, (uintmax_t)inum); 1843 #endif 1844 /* 1845 * If journaling is tracking this write we must add 1846 * the work to the inode or indirect being written. 1847 */ 1848 if (wkhd != NULL) { 1849 if (lbn < UFS_NDADDR) 1850 softdep_inode_append(ip, 1851 curthread->td_ucred, wkhd); 1852 else 1853 softdep_buf_append(ibp, wkhd); 1854 } 1855 if (lbn < UFS_NDADDR) { 1856 DIP_SET(ip, i_db[lbn], bno); 1857 } else if (I_IS_UFS1(ip)) { 1858 ((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno; 1859 bdwrite(ibp); 1860 } else { 1861 ((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno; 1862 bdwrite(ibp); 1863 } 1864 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size)); 1865 ip->i_flag |= IN_CHANGE | IN_UPDATE; 1866 lockmgr(vp->v_vnlock, LK_RELEASE, NULL); 1867 return (1); 1868 } 1869 if (lbn >= UFS_NDADDR) 1870 bqrelse(ibp); 1871 /* 1872 * Allocate the block into which to do the copy. Note that this 1873 * allocation will never require any additional allocations for 1874 * the snapshot inode. 1875 */ 1876 td->td_pflags |= TDP_COWINPROGRESS; 1877 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn), 1878 fs->fs_bsize, KERNCRED, 0, &cbp); 1879 td->td_pflags &= ~TDP_COWINPROGRESS; 1880 if (error) 1881 break; 1882 #ifdef DEBUG 1883 if (snapdebug) 1884 printf("%s%ju lbn %jd %s %ju size %ld to blkno %jd\n", 1885 "Copyonremove: snapino ", (uintmax_t)ip->i_number, 1886 (intmax_t)lbn, "for inum", (uintmax_t)inum, size, 1887 (intmax_t)cbp->b_blkno); 1888 #endif 1889 /* 1890 * If we have already read the old block contents, then 1891 * simply copy them to the new block. Note that we need 1892 * to synchronously write snapshots that have not been 1893 * unlinked, and hence will be visible after a crash, 1894 * to ensure their integrity. At a minimum we ensure the 1895 * integrity of the filesystem metadata, but use the 1896 * dopersistence sysctl-setable flag to decide on the 1897 * persistence needed for file content data. 1898 */ 1899 if (savedcbp != NULL) { 1900 bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize); 1901 bawrite(cbp); 1902 if ((vtype == VDIR || dopersistence) && 1903 ip->i_effnlink > 0) 1904 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 1905 continue; 1906 } 1907 /* 1908 * Otherwise, read the old block contents into the buffer. 1909 */ 1910 if ((error = readblock(vp, cbp, lbn)) != 0) { 1911 bzero(cbp->b_data, fs->fs_bsize); 1912 bawrite(cbp); 1913 if ((vtype == VDIR || dopersistence) && 1914 ip->i_effnlink > 0) 1915 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 1916 break; 1917 } 1918 savedcbp = cbp; 1919 } 1920 /* 1921 * Note that we need to synchronously write snapshots that 1922 * have not been unlinked, and hence will be visible after 1923 * a crash, to ensure their integrity. At a minimum we 1924 * ensure the integrity of the filesystem metadata, but 1925 * use the dopersistence sysctl-setable flag to decide on 1926 * the persistence needed for file content data. 1927 */ 1928 if (savedcbp) { 1929 vp = savedcbp->b_vp; 1930 bawrite(savedcbp); 1931 if ((vtype == VDIR || dopersistence) && 1932 VTOI(vp)->i_effnlink > 0) 1933 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 1934 } 1935 /* 1936 * If we have been unable to allocate a block in which to do 1937 * the copy, then return non-zero so that the fragment will 1938 * not be freed. Although space will be lost, the snapshot 1939 * will stay consistent. 1940 */ 1941 if (error != 0 && wkhd != NULL) 1942 softdep_freework(wkhd); 1943 lockmgr(vp->v_vnlock, LK_RELEASE, NULL); 1944 return (error); 1945 } 1946 1947 /* 1948 * Associate snapshot files when mounting. 1949 */ 1950 void 1951 ffs_snapshot_mount(mp) 1952 struct mount *mp; 1953 { 1954 struct ufsmount *ump = VFSTOUFS(mp); 1955 struct vnode *devvp = ump->um_devvp; 1956 struct fs *fs = ump->um_fs; 1957 struct thread *td = curthread; 1958 struct snapdata *sn; 1959 struct vnode *vp; 1960 struct vnode *lastvp; 1961 struct inode *ip; 1962 struct uio auio; 1963 struct iovec aiov; 1964 void *snapblklist; 1965 char *reason; 1966 daddr_t snaplistsize; 1967 int error, snaploc, loc; 1968 1969 /* 1970 * XXX The following needs to be set before ffs_truncate or 1971 * VOP_READ can be called. 1972 */ 1973 mp->mnt_stat.f_iosize = fs->fs_bsize; 1974 /* 1975 * Process each snapshot listed in the superblock. 1976 */ 1977 vp = NULL; 1978 lastvp = NULL; 1979 sn = NULL; 1980 for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) { 1981 if (fs->fs_snapinum[snaploc] == 0) 1982 break; 1983 if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc], 1984 LK_EXCLUSIVE, &vp)) != 0){ 1985 printf("ffs_snapshot_mount: vget failed %d\n", error); 1986 continue; 1987 } 1988 ip = VTOI(vp); 1989 if (!IS_SNAPSHOT(ip) || ip->i_size == 1990 lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) { 1991 if (!IS_SNAPSHOT(ip)) { 1992 reason = "non-snapshot"; 1993 } else { 1994 reason = "old format snapshot"; 1995 (void)ffs_truncate(vp, (off_t)0, 0, NOCRED); 1996 (void)ffs_syncvnode(vp, MNT_WAIT, 0); 1997 } 1998 printf("ffs_snapshot_mount: %s inode %d\n", 1999 reason, fs->fs_snapinum[snaploc]); 2000 vput(vp); 2001 vp = NULL; 2002 for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) { 2003 if (fs->fs_snapinum[loc] == 0) 2004 break; 2005 fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc]; 2006 } 2007 fs->fs_snapinum[loc - 1] = 0; 2008 snaploc--; 2009 continue; 2010 } 2011 /* 2012 * Acquire a lock on the snapdata structure, creating it if 2013 * necessary. 2014 */ 2015 sn = ffs_snapdata_acquire(devvp); 2016 /* 2017 * Change vnode to use shared snapshot lock instead of the 2018 * original private lock. 2019 */ 2020 vp->v_vnlock = &sn->sn_lock; 2021 lockmgr(&vp->v_lock, LK_RELEASE, NULL); 2022 /* 2023 * Link it onto the active snapshot list. 2024 */ 2025 VI_LOCK(devvp); 2026 if (ip->i_nextsnap.tqe_prev != 0) 2027 panic("ffs_snapshot_mount: %ju already on list", 2028 (uintmax_t)ip->i_number); 2029 else 2030 TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap); 2031 vp->v_vflag |= VV_SYSTEM; 2032 VI_UNLOCK(devvp); 2033 VOP_UNLOCK(vp, 0); 2034 lastvp = vp; 2035 } 2036 vp = lastvp; 2037 /* 2038 * No usable snapshots found. 2039 */ 2040 if (sn == NULL || vp == NULL) 2041 return; 2042 /* 2043 * Allocate the space for the block hints list. We always want to 2044 * use the list from the newest snapshot. 2045 */ 2046 auio.uio_iov = &aiov; 2047 auio.uio_iovcnt = 1; 2048 aiov.iov_base = (void *)&snaplistsize; 2049 aiov.iov_len = sizeof(snaplistsize); 2050 auio.uio_resid = aiov.iov_len; 2051 auio.uio_offset = 2052 lblktosize(fs, howmany(fs->fs_size, fs->fs_frag)); 2053 auio.uio_segflg = UIO_SYSSPACE; 2054 auio.uio_rw = UIO_READ; 2055 auio.uio_td = td; 2056 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2057 if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) { 2058 printf("ffs_snapshot_mount: read_1 failed %d\n", error); 2059 VOP_UNLOCK(vp, 0); 2060 return; 2061 } 2062 snapblklist = malloc(snaplistsize * sizeof(daddr_t), 2063 M_UFSMNT, M_WAITOK); 2064 auio.uio_iovcnt = 1; 2065 aiov.iov_base = snapblklist; 2066 aiov.iov_len = snaplistsize * sizeof (daddr_t); 2067 auio.uio_resid = aiov.iov_len; 2068 auio.uio_offset -= sizeof(snaplistsize); 2069 if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) { 2070 printf("ffs_snapshot_mount: read_2 failed %d\n", error); 2071 VOP_UNLOCK(vp, 0); 2072 free(snapblklist, M_UFSMNT); 2073 return; 2074 } 2075 VOP_UNLOCK(vp, 0); 2076 VI_LOCK(devvp); 2077 ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount"); 2078 sn->sn_listsize = snaplistsize; 2079 sn->sn_blklist = (daddr_t *)snapblklist; 2080 devvp->v_vflag |= VV_COPYONWRITE; 2081 VI_UNLOCK(devvp); 2082 } 2083 2084 /* 2085 * Disassociate snapshot files when unmounting. 2086 */ 2087 void 2088 ffs_snapshot_unmount(mp) 2089 struct mount *mp; 2090 { 2091 struct vnode *devvp = VFSTOUFS(mp)->um_devvp; 2092 struct snapdata *sn; 2093 struct inode *xp; 2094 struct vnode *vp; 2095 2096 VI_LOCK(devvp); 2097 sn = devvp->v_rdev->si_snapdata; 2098 while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) { 2099 vp = ITOV(xp); 2100 TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap); 2101 xp->i_nextsnap.tqe_prev = 0; 2102 lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE, 2103 VI_MTX(devvp)); 2104 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL); 2105 KASSERT(vp->v_vnlock == &sn->sn_lock, 2106 ("ffs_snapshot_unmount: lost lock mutation")); 2107 vp->v_vnlock = &vp->v_lock; 2108 lockmgr(&vp->v_lock, LK_RELEASE, NULL); 2109 lockmgr(&sn->sn_lock, LK_RELEASE, NULL); 2110 if (xp->i_effnlink > 0) 2111 vrele(vp); 2112 VI_LOCK(devvp); 2113 sn = devvp->v_rdev->si_snapdata; 2114 } 2115 try_free_snapdata(devvp); 2116 ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount"); 2117 } 2118 2119 /* 2120 * Check the buffer block to be belong to device buffer that shall be 2121 * locked after snaplk. devvp shall be locked on entry, and will be 2122 * leaved locked upon exit. 2123 */ 2124 static int 2125 ffs_bp_snapblk(devvp, bp) 2126 struct vnode *devvp; 2127 struct buf *bp; 2128 { 2129 struct snapdata *sn; 2130 struct fs *fs; 2131 ufs2_daddr_t lbn, *snapblklist; 2132 int lower, upper, mid; 2133 2134 ASSERT_VI_LOCKED(devvp, "ffs_bp_snapblk"); 2135 KASSERT(devvp->v_type == VCHR, ("Not a device %p", devvp)); 2136 sn = devvp->v_rdev->si_snapdata; 2137 if (sn == NULL || TAILQ_FIRST(&sn->sn_head) == NULL) 2138 return (0); 2139 fs = ITOFS(TAILQ_FIRST(&sn->sn_head)); 2140 lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno)); 2141 snapblklist = sn->sn_blklist; 2142 upper = sn->sn_listsize - 1; 2143 lower = 1; 2144 while (lower <= upper) { 2145 mid = (lower + upper) / 2; 2146 if (snapblklist[mid] == lbn) 2147 break; 2148 if (snapblklist[mid] < lbn) 2149 lower = mid + 1; 2150 else 2151 upper = mid - 1; 2152 } 2153 if (lower <= upper) 2154 return (1); 2155 return (0); 2156 } 2157 2158 void 2159 ffs_bdflush(bo, bp) 2160 struct bufobj *bo; 2161 struct buf *bp; 2162 { 2163 struct thread *td; 2164 struct vnode *vp, *devvp; 2165 struct buf *nbp; 2166 int bp_bdskip; 2167 2168 if (bo->bo_dirty.bv_cnt <= dirtybufthresh) 2169 return; 2170 2171 td = curthread; 2172 vp = bp->b_vp; 2173 devvp = bo2vnode(bo); 2174 KASSERT(vp == devvp, ("devvp != vp %p %p", bo, bp)); 2175 2176 VI_LOCK(devvp); 2177 bp_bdskip = ffs_bp_snapblk(devvp, bp); 2178 if (bp_bdskip) 2179 bdwriteskip++; 2180 VI_UNLOCK(devvp); 2181 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10 && !bp_bdskip) { 2182 (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 2183 altbufferflushes++; 2184 } else { 2185 BO_LOCK(bo); 2186 /* 2187 * Try to find a buffer to flush. 2188 */ 2189 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2190 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2191 BUF_LOCK(nbp, 2192 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2193 continue; 2194 if (bp == nbp) 2195 panic("bdwrite: found ourselves"); 2196 BO_UNLOCK(bo); 2197 /* 2198 * Don't countdeps with the bo lock 2199 * held. 2200 */ 2201 if (buf_countdeps(nbp, 0)) { 2202 BO_LOCK(bo); 2203 BUF_UNLOCK(nbp); 2204 continue; 2205 } 2206 if (bp_bdskip) { 2207 VI_LOCK(devvp); 2208 if (!ffs_bp_snapblk(vp, nbp)) { 2209 VI_UNLOCK(devvp); 2210 BO_LOCK(bo); 2211 BUF_UNLOCK(nbp); 2212 continue; 2213 } 2214 VI_UNLOCK(devvp); 2215 } 2216 if (nbp->b_flags & B_CLUSTEROK) { 2217 vfs_bio_awrite(nbp); 2218 } else { 2219 bremfree(nbp); 2220 bawrite(nbp); 2221 } 2222 dirtybufferflushes++; 2223 break; 2224 } 2225 if (nbp == NULL) 2226 BO_UNLOCK(bo); 2227 } 2228 } 2229 2230 /* 2231 * Check for need to copy block that is about to be written, 2232 * copying the block if necessary. 2233 */ 2234 int 2235 ffs_copyonwrite(devvp, bp) 2236 struct vnode *devvp; 2237 struct buf *bp; 2238 { 2239 struct snapdata *sn; 2240 struct buf *ibp, *cbp, *savedcbp = NULL; 2241 struct thread *td = curthread; 2242 struct fs *fs; 2243 struct inode *ip; 2244 struct vnode *vp = NULL; 2245 ufs2_daddr_t lbn, blkno, *snapblklist; 2246 int lower, upper, mid, indiroff, error = 0; 2247 int launched_async_io, prev_norunningbuf; 2248 long saved_runningbufspace; 2249 2250 if (devvp != bp->b_vp && IS_SNAPSHOT(VTOI(bp->b_vp))) 2251 return (0); /* Update on a snapshot file */ 2252 if (td->td_pflags & TDP_COWINPROGRESS) 2253 panic("ffs_copyonwrite: recursive call"); 2254 /* 2255 * First check to see if it is in the preallocated list. 2256 * By doing this check we avoid several potential deadlocks. 2257 */ 2258 VI_LOCK(devvp); 2259 sn = devvp->v_rdev->si_snapdata; 2260 if (sn == NULL || 2261 TAILQ_EMPTY(&sn->sn_head)) { 2262 VI_UNLOCK(devvp); 2263 return (0); /* No snapshot */ 2264 } 2265 ip = TAILQ_FIRST(&sn->sn_head); 2266 fs = ITOFS(ip); 2267 lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno)); 2268 snapblklist = sn->sn_blklist; 2269 upper = sn->sn_listsize - 1; 2270 lower = 1; 2271 while (lower <= upper) { 2272 mid = (lower + upper) / 2; 2273 if (snapblklist[mid] == lbn) 2274 break; 2275 if (snapblklist[mid] < lbn) 2276 lower = mid + 1; 2277 else 2278 upper = mid - 1; 2279 } 2280 if (lower <= upper) { 2281 VI_UNLOCK(devvp); 2282 return (0); 2283 } 2284 launched_async_io = 0; 2285 prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF; 2286 /* 2287 * Since I/O on bp isn't yet in progress and it may be blocked 2288 * for a long time waiting on snaplk, back it out of 2289 * runningbufspace, possibly waking other threads waiting for space. 2290 */ 2291 saved_runningbufspace = bp->b_runningbufspace; 2292 if (saved_runningbufspace != 0) 2293 runningbufwakeup(bp); 2294 /* 2295 * Not in the precomputed list, so check the snapshots. 2296 */ 2297 while (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL, 2298 VI_MTX(devvp)) != 0) { 2299 VI_LOCK(devvp); 2300 sn = devvp->v_rdev->si_snapdata; 2301 if (sn == NULL || 2302 TAILQ_EMPTY(&sn->sn_head)) { 2303 VI_UNLOCK(devvp); 2304 if (saved_runningbufspace != 0) { 2305 bp->b_runningbufspace = saved_runningbufspace; 2306 atomic_add_long(&runningbufspace, 2307 bp->b_runningbufspace); 2308 } 2309 return (0); /* Snapshot gone */ 2310 } 2311 } 2312 TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) { 2313 vp = ITOV(ip); 2314 if (DOINGSOFTDEP(vp)) 2315 softdep_prealloc(vp, MNT_WAIT); 2316 /* 2317 * We ensure that everything of our own that needs to be 2318 * copied will be done at the time that ffs_snapshot is 2319 * called. Thus we can skip the check here which can 2320 * deadlock in doing the lookup in UFS_BALLOC. 2321 */ 2322 if (bp->b_vp == vp) 2323 continue; 2324 /* 2325 * Check to see if block needs to be copied. We do not have 2326 * to hold the snapshot lock while doing this lookup as it 2327 * will never require any additional allocations for the 2328 * snapshot inode. 2329 */ 2330 if (lbn < UFS_NDADDR) { 2331 blkno = DIP(ip, i_db[lbn]); 2332 } else { 2333 td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF; 2334 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn), 2335 fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp); 2336 td->td_pflags &= ~TDP_COWINPROGRESS; 2337 if (error) 2338 break; 2339 indiroff = (lbn - UFS_NDADDR) % NINDIR(fs); 2340 if (I_IS_UFS1(ip)) 2341 blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff]; 2342 else 2343 blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff]; 2344 bqrelse(ibp); 2345 } 2346 #ifdef INVARIANTS 2347 if (blkno == BLK_SNAP && bp->b_lblkno >= 0) 2348 panic("ffs_copyonwrite: bad copy block"); 2349 #endif 2350 if (blkno != 0) 2351 continue; 2352 /* 2353 * Allocate the block into which to do the copy. Since 2354 * multiple processes may all try to copy the same block, 2355 * we have to recheck our need to do a copy if we sleep 2356 * waiting for the lock. 2357 * 2358 * Because all snapshots on a filesystem share a single 2359 * lock, we ensure that we will never be in competition 2360 * with another process to allocate a block. 2361 */ 2362 td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF; 2363 error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn), 2364 fs->fs_bsize, KERNCRED, 0, &cbp); 2365 td->td_pflags &= ~TDP_COWINPROGRESS; 2366 if (error) 2367 break; 2368 #ifdef DEBUG 2369 if (snapdebug) { 2370 printf("Copyonwrite: snapino %ju lbn %jd for ", 2371 (uintmax_t)ip->i_number, (intmax_t)lbn); 2372 if (bp->b_vp == devvp) 2373 printf("fs metadata"); 2374 else 2375 printf("inum %ju", 2376 (uintmax_t)VTOI(bp->b_vp)->i_number); 2377 printf(" lblkno %jd to blkno %jd\n", 2378 (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno); 2379 } 2380 #endif 2381 /* 2382 * If we have already read the old block contents, then 2383 * simply copy them to the new block. Note that we need 2384 * to synchronously write snapshots that have not been 2385 * unlinked, and hence will be visible after a crash, 2386 * to ensure their integrity. At a minimum we ensure the 2387 * integrity of the filesystem metadata, but use the 2388 * dopersistence sysctl-setable flag to decide on the 2389 * persistence needed for file content data. 2390 */ 2391 if (savedcbp != NULL) { 2392 bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize); 2393 bawrite(cbp); 2394 if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR || 2395 dopersistence) && ip->i_effnlink > 0) 2396 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 2397 else 2398 launched_async_io = 1; 2399 continue; 2400 } 2401 /* 2402 * Otherwise, read the old block contents into the buffer. 2403 */ 2404 if ((error = readblock(vp, cbp, lbn)) != 0) { 2405 bzero(cbp->b_data, fs->fs_bsize); 2406 bawrite(cbp); 2407 if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR || 2408 dopersistence) && ip->i_effnlink > 0) 2409 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 2410 else 2411 launched_async_io = 1; 2412 break; 2413 } 2414 savedcbp = cbp; 2415 } 2416 /* 2417 * Note that we need to synchronously write snapshots that 2418 * have not been unlinked, and hence will be visible after 2419 * a crash, to ensure their integrity. At a minimum we 2420 * ensure the integrity of the filesystem metadata, but 2421 * use the dopersistence sysctl-setable flag to decide on 2422 * the persistence needed for file content data. 2423 */ 2424 if (savedcbp) { 2425 vp = savedcbp->b_vp; 2426 bawrite(savedcbp); 2427 if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR || 2428 dopersistence) && VTOI(vp)->i_effnlink > 0) 2429 (void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT); 2430 else 2431 launched_async_io = 1; 2432 } 2433 lockmgr(vp->v_vnlock, LK_RELEASE, NULL); 2434 td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) | 2435 prev_norunningbuf; 2436 if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0) 2437 waitrunningbufspace(); 2438 /* 2439 * I/O on bp will now be started, so count it in runningbufspace. 2440 */ 2441 if (saved_runningbufspace != 0) { 2442 bp->b_runningbufspace = saved_runningbufspace; 2443 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 2444 } 2445 return (error); 2446 } 2447 2448 /* 2449 * sync snapshots to force freework records waiting on snapshots to claim 2450 * blocks to free. 2451 */ 2452 void 2453 ffs_sync_snap(mp, waitfor) 2454 struct mount *mp; 2455 int waitfor; 2456 { 2457 struct snapdata *sn; 2458 struct vnode *devvp; 2459 struct vnode *vp; 2460 struct inode *ip; 2461 2462 devvp = VFSTOUFS(mp)->um_devvp; 2463 if ((devvp->v_vflag & VV_COPYONWRITE) == 0) 2464 return; 2465 for (;;) { 2466 VI_LOCK(devvp); 2467 sn = devvp->v_rdev->si_snapdata; 2468 if (sn == NULL) { 2469 VI_UNLOCK(devvp); 2470 return; 2471 } 2472 if (lockmgr(&sn->sn_lock, 2473 LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL, 2474 VI_MTX(devvp)) == 0) 2475 break; 2476 } 2477 TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) { 2478 vp = ITOV(ip); 2479 ffs_syncvnode(vp, waitfor, NO_INO_UPDT); 2480 } 2481 lockmgr(&sn->sn_lock, LK_RELEASE, NULL); 2482 } 2483 2484 /* 2485 * Read the specified block into the given buffer. 2486 * Much of this boiler-plate comes from bwrite(). 2487 */ 2488 static int 2489 readblock(vp, bp, lbn) 2490 struct vnode *vp; 2491 struct buf *bp; 2492 ufs2_daddr_t lbn; 2493 { 2494 struct inode *ip = VTOI(vp); 2495 struct bio *bip; 2496 struct fs *fs; 2497 2498 ip = VTOI(vp); 2499 fs = ITOFS(ip); 2500 2501 bip = g_alloc_bio(); 2502 bip->bio_cmd = BIO_READ; 2503 bip->bio_offset = dbtob(fsbtodb(fs, blkstofrags(fs, lbn))); 2504 bip->bio_data = bp->b_data; 2505 bip->bio_length = bp->b_bcount; 2506 bip->bio_done = NULL; 2507 2508 g_io_request(bip, ITODEVVP(ip)->v_bufobj.bo_private); 2509 bp->b_error = biowait(bip, "snaprdb"); 2510 g_destroy_bio(bip); 2511 return (bp->b_error); 2512 } 2513 2514 #endif 2515 2516 /* 2517 * Process file deletes that were deferred by ufs_inactive() due to 2518 * the file system being suspended. Transfer IN_LAZYACCESS into 2519 * IN_MODIFIED for vnodes that were accessed during suspension. 2520 */ 2521 void 2522 process_deferred_inactive(struct mount *mp) 2523 { 2524 struct vnode *vp, *mvp; 2525 struct inode *ip; 2526 struct thread *td; 2527 int error; 2528 2529 td = curthread; 2530 (void) vn_start_secondary_write(NULL, &mp, V_WAIT); 2531 loop: 2532 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 2533 /* 2534 * IN_LAZYACCESS is checked here without holding any 2535 * vnode lock, but this flag is set only while holding 2536 * vnode interlock. 2537 */ 2538 if (vp->v_type == VNON || 2539 ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 && 2540 ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0))) { 2541 VI_UNLOCK(vp); 2542 continue; 2543 } 2544 vholdl(vp); 2545 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); 2546 if (error != 0) { 2547 vdrop(vp); 2548 if (error == ENOENT) 2549 continue; /* vnode recycled */ 2550 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); 2551 goto loop; 2552 } 2553 ip = VTOI(vp); 2554 if ((ip->i_flag & IN_LAZYACCESS) != 0) { 2555 ip->i_flag &= ~IN_LAZYACCESS; 2556 ip->i_flag |= IN_MODIFIED; 2557 } 2558 VI_LOCK(vp); 2559 if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) { 2560 VI_UNLOCK(vp); 2561 VOP_UNLOCK(vp, 0); 2562 vdrop(vp); 2563 continue; 2564 } 2565 vinactive(vp, td); 2566 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, 2567 ("process_deferred_inactive: got VI_OWEINACT")); 2568 VI_UNLOCK(vp); 2569 VOP_UNLOCK(vp, 0); 2570 vdrop(vp); 2571 } 2572 vn_finished_secondary_write(mp); 2573 } 2574 2575 #ifndef NO_FFS_SNAPSHOT 2576 2577 static struct snapdata * 2578 ffs_snapdata_alloc(void) 2579 { 2580 struct snapdata *sn; 2581 2582 /* 2583 * Fetch a snapdata from the free list if there is one available. 2584 */ 2585 mtx_lock(&snapfree_lock); 2586 sn = LIST_FIRST(&snapfree); 2587 if (sn != NULL) 2588 LIST_REMOVE(sn, sn_link); 2589 mtx_unlock(&snapfree_lock); 2590 if (sn != NULL) 2591 return (sn); 2592 /* 2593 * If there were no free snapdatas allocate one. 2594 */ 2595 sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO); 2596 TAILQ_INIT(&sn->sn_head); 2597 lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT, 2598 LK_CANRECURSE | LK_NOSHARE); 2599 return (sn); 2600 } 2601 2602 /* 2603 * The snapdata is never freed because we can not be certain that 2604 * there are no threads sleeping on the snap lock. Persisting 2605 * them permanently avoids costly synchronization in ffs_lock(). 2606 */ 2607 static void 2608 ffs_snapdata_free(struct snapdata *sn) 2609 { 2610 mtx_lock(&snapfree_lock); 2611 LIST_INSERT_HEAD(&snapfree, sn, sn_link); 2612 mtx_unlock(&snapfree_lock); 2613 } 2614 2615 /* Try to free snapdata associated with devvp */ 2616 static void 2617 try_free_snapdata(struct vnode *devvp) 2618 { 2619 struct snapdata *sn; 2620 ufs2_daddr_t *snapblklist; 2621 2622 ASSERT_VI_LOCKED(devvp, "try_free_snapdata"); 2623 sn = devvp->v_rdev->si_snapdata; 2624 2625 if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL || 2626 (devvp->v_vflag & VV_COPYONWRITE) == 0) { 2627 VI_UNLOCK(devvp); 2628 return; 2629 } 2630 2631 devvp->v_rdev->si_snapdata = NULL; 2632 devvp->v_vflag &= ~VV_COPYONWRITE; 2633 lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp)); 2634 snapblklist = sn->sn_blklist; 2635 sn->sn_blklist = NULL; 2636 sn->sn_listsize = 0; 2637 lockmgr(&sn->sn_lock, LK_RELEASE, NULL); 2638 if (snapblklist != NULL) 2639 free(snapblklist, M_UFSMNT); 2640 ffs_snapdata_free(sn); 2641 } 2642 2643 static struct snapdata * 2644 ffs_snapdata_acquire(struct vnode *devvp) 2645 { 2646 struct snapdata *nsn; 2647 struct snapdata *sn; 2648 2649 /* 2650 * Allocate a free snapdata. This is done before acquiring the 2651 * devvp lock to avoid allocation while the devvp interlock is 2652 * held. 2653 */ 2654 nsn = ffs_snapdata_alloc(); 2655 /* 2656 * If there snapshots already exist on this filesystem grab a 2657 * reference to the shared lock. Otherwise this is the first 2658 * snapshot on this filesystem and we need to use our 2659 * pre-allocated snapdata. 2660 */ 2661 VI_LOCK(devvp); 2662 if (devvp->v_rdev->si_snapdata == NULL) { 2663 devvp->v_rdev->si_snapdata = nsn; 2664 nsn = NULL; 2665 } 2666 sn = devvp->v_rdev->si_snapdata; 2667 /* 2668 * Acquire the snapshot lock. 2669 */ 2670 lockmgr(&sn->sn_lock, 2671 LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, VI_MTX(devvp)); 2672 /* 2673 * Free any unused snapdata. 2674 */ 2675 if (nsn != NULL) 2676 ffs_snapdata_free(nsn); 2677 2678 return (sn); 2679 } 2680 2681 #endif 2682