git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/fs-common.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "acl.h"
   5 #include "btree_update.h"
   6 #include "dirent.h"
   7 #include "fs-common.h"
   8 #include "inode.h"
   9 #include "subvolume.h"
  10 #include "xattr.h"
  11
  12 #include <linux/posix_acl.h>
  13
  14 static inline int is_subdir_for_nlink(struct bch_inode_unpacked *inode)
  15 {
  16         return S_ISDIR(inode->bi_mode) && !inode->bi_subvol;
  17 }
  18
  19 int bch2_create_trans(struct btree_trans *trans,
  20                       subvol_inum dir,
  21                       struct bch_inode_unpacked *dir_u,
  22                       struct bch_inode_unpacked *new_inode,
  23                       const struct qstr *name,
  24                       uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
  25                       struct posix_acl *default_acl,
  26                       struct posix_acl *acl,
  27                       subvol_inum snapshot_src,
  28                       unsigned flags)
  29 {
  30         struct bch_fs *c = trans->c;
  31         struct btree_iter dir_iter = { NULL };
  32         struct btree_iter inode_iter = { NULL };
  33         subvol_inum new_inum = dir;
  34         u64 now = bch2_current_time(c);
  35         u64 cpu = raw_smp_processor_id();
  36         u64 dir_target;
  37         u32 snapshot;
  38         unsigned dir_type = mode_to_type(mode);
  39         int ret;
  40
  41         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
  42         if (ret)
  43                 goto err;
  44
  45         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_INTENT);
  46         if (ret)
  47                 goto err;
  48
  49         if (!(flags & BCH_CREATE_SNAPSHOT)) {
  50                 /* Normal create path - allocate a new inode: */
  51                 bch2_inode_init_late(new_inode, now, uid, gid, mode, rdev, dir_u);
  52
  53                 if (flags & BCH_CREATE_TMPFILE)
  54                         new_inode->bi_flags |= BCH_INODE_UNLINKED;
  55
  56                 ret = bch2_inode_create(trans, &inode_iter, new_inode, snapshot, cpu);
  57                 if (ret)
  58                         goto err;
  59
  60                 snapshot_src = (subvol_inum) { 0 };
  61         } else {
  62                 /*
  63                  * Creating a snapshot - we're not allocating a new inode, but
  64                  * we do have to lookup the root inode of the subvolume we're
  65                  * snapshotting and update it (in the new snapshot):
  66                  */
  67
  68                 if (!snapshot_src.inum) {
  69                         /* Inode wasn't specified, just snapshot: */
  70                         struct bch_subvolume s;
  71
  72                         ret = bch2_subvolume_get(trans, snapshot_src.subvol, true,
  73                                                  BTREE_ITER_CACHED, &s);
  74                         if (ret)
  75                                 goto err;
  76
  77                         snapshot_src.inum = le64_to_cpu(s.inode);
  78                 }
  79
  80                 ret = bch2_inode_peek(trans, &inode_iter, new_inode, snapshot_src,
  81                                       BTREE_ITER_INTENT);
  82                 if (ret)
  83                         goto err;
  84
  85                 if (new_inode->bi_subvol != snapshot_src.subvol) {
  86                         /* Not a subvolume root: */
  87                         ret = -EINVAL;
  88                         goto err;
  89                 }
  90
  91                 /*
  92                  * If we're not root, we have to own the subvolume being
  93                  * snapshotted:
  94                  */
  95                 if (uid && new_inode->bi_uid != uid) {
  96                         ret = -EPERM;
  97                         goto err;
  98                 }
  99
 100                 flags |= BCH_CREATE_SUBVOL;
 101         }
 102
 103         new_inum.inum   = new_inode->bi_inum;
 104         dir_target      = new_inode->bi_inum;
 105
 106         if (flags & BCH_CREATE_SUBVOL) {
 107                 u32 new_subvol, dir_snapshot;
 108
 109                 ret = bch2_subvolume_create(trans, new_inode->bi_inum,
 110                                             snapshot_src.subvol,
 111                                             &new_subvol, &snapshot,
 112                                             (flags & BCH_CREATE_SNAPSHOT_RO) != 0);
 113                 if (ret)
 114                         goto err;
 115
 116                 new_inode->bi_parent_subvol     = dir.subvol;
 117                 new_inode->bi_subvol            = new_subvol;
 118                 new_inum.subvol                 = new_subvol;
 119                 dir_target                      = new_subvol;
 120                 dir_type                        = DT_SUBVOL;
 121
 122                 ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &dir_snapshot);
 123                 if (ret)
 124                         goto err;
 125
 126                 bch2_btree_iter_set_snapshot(&dir_iter, dir_snapshot);
 127                 ret = bch2_btree_iter_traverse(&dir_iter);
 128                 if (ret)
 129                         goto err;
 130         }
 131
 132         if (!(flags & BCH_CREATE_SNAPSHOT)) {
 133                 if (default_acl) {
 134                         ret = bch2_set_acl_trans(trans, new_inum, new_inode,
 135                                                  default_acl, ACL_TYPE_DEFAULT);
 136                         if (ret)
 137                                 goto err;
 138                 }
 139
 140                 if (acl) {
 141                         ret = bch2_set_acl_trans(trans, new_inum, new_inode,
 142                                                  acl, ACL_TYPE_ACCESS);
 143                         if (ret)
 144                                 goto err;
 145                 }
 146         }
 147
 148         if (!(flags & BCH_CREATE_TMPFILE)) {
 149                 struct bch_hash_info dir_hash = bch2_hash_info_init(c, dir_u);
 150                 u64 dir_offset;
 151
 152                 if (is_subdir_for_nlink(new_inode))
 153                         dir_u->bi_nlink++;
 154                 dir_u->bi_mtime = dir_u->bi_ctime = now;
 155
 156                 ret = bch2_inode_write(trans, &dir_iter, dir_u);
 157                 if (ret)
 158                         goto err;
 159
 160                 ret = bch2_dirent_create(trans, dir, &dir_hash,
 161                                          dir_type,
 162                                          name,
 163                                          dir_target,
 164                                          &dir_offset,
 165                                          BCH_HASH_SET_MUST_CREATE);
 166                 if (ret)
 167                         goto err;
 168
 169                 if (c->sb.version >= bcachefs_metadata_version_inode_backpointers) {
 170                         new_inode->bi_dir               = dir_u->bi_inum;
 171                         new_inode->bi_dir_offset        = dir_offset;
 172                 }
 173         }
 174
 175         inode_iter.flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 176         bch2_btree_iter_set_snapshot(&inode_iter, snapshot);
 177
 178         ret   = bch2_btree_iter_traverse(&inode_iter) ?:
 179                 bch2_inode_write(trans, &inode_iter, new_inode);
 180 err:
 181         bch2_trans_iter_exit(trans, &inode_iter);
 182         bch2_trans_iter_exit(trans, &dir_iter);
 183         return ret;
 184 }
 185
 186 int bch2_link_trans(struct btree_trans *trans,
 187                     subvol_inum dir,  struct bch_inode_unpacked *dir_u,
 188                     subvol_inum inum, struct bch_inode_unpacked *inode_u,
 189                     const struct qstr *name)
 190 {
 191         struct bch_fs *c = trans->c;
 192         struct btree_iter dir_iter = { NULL };
 193         struct btree_iter inode_iter = { NULL };
 194         struct bch_hash_info dir_hash;
 195         u64 now = bch2_current_time(c);
 196         u64 dir_offset = 0;
 197         int ret;
 198
 199         if (dir.subvol != inum.subvol)
 200                 return -EXDEV;
 201
 202         ret = bch2_inode_peek(trans, &inode_iter, inode_u, inum, BTREE_ITER_INTENT);
 203         if (ret)
 204                 goto err;
 205
 206         inode_u->bi_ctime = now;
 207         ret = bch2_inode_nlink_inc(inode_u);
 208         if (ret)
 209                 return ret;
 210
 211         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_INTENT);
 212         if (ret)
 213                 goto err;
 214
 215         dir_u->bi_mtime = dir_u->bi_ctime = now;
 216
 217         dir_hash = bch2_hash_info_init(c, dir_u);
 218
 219         ret = bch2_dirent_create(trans, dir, &dir_hash,
 220                                  mode_to_type(inode_u->bi_mode),
 221                                  name, inum.inum, &dir_offset,
 222                                  BCH_HASH_SET_MUST_CREATE);
 223         if (ret)
 224                 goto err;
 225
 226         if (c->sb.version >= bcachefs_metadata_version_inode_backpointers) {
 227                 inode_u->bi_dir         = dir.inum;
 228                 inode_u->bi_dir_offset  = dir_offset;
 229         }
 230
 231         ret =   bch2_inode_write(trans, &dir_iter, dir_u) ?:
 232                 bch2_inode_write(trans, &inode_iter, inode_u);
 233 err:
 234         bch2_trans_iter_exit(trans, &dir_iter);
 235         bch2_trans_iter_exit(trans, &inode_iter);
 236         return ret;
 237 }
 238
 239 int bch2_unlink_trans(struct btree_trans *trans,
 240                       subvol_inum dir,
 241                       struct bch_inode_unpacked *dir_u,
 242                       struct bch_inode_unpacked *inode_u,
 243                       const struct qstr *name,
 244                       bool deleting_snapshot)
 245 {
 246         struct bch_fs *c = trans->c;
 247         struct btree_iter dir_iter = { NULL };
 248         struct btree_iter dirent_iter = { NULL };
 249         struct btree_iter inode_iter = { NULL };
 250         struct bch_hash_info dir_hash;
 251         subvol_inum inum;
 252         u64 now = bch2_current_time(c);
 253         struct bkey_s_c k;
 254         int ret;
 255
 256         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_INTENT);
 257         if (ret)
 258                 goto err;
 259
 260         dir_hash = bch2_hash_info_init(c, dir_u);
 261
 262         ret = __bch2_dirent_lookup_trans(trans, &dirent_iter, dir, &dir_hash,
 263                                          name, &inum, BTREE_ITER_INTENT);
 264         if (ret)
 265                 goto err;
 266
 267         ret = bch2_inode_peek(trans, &inode_iter, inode_u, inum,
 268                               BTREE_ITER_INTENT);
 269         if (ret)
 270                 goto err;
 271
 272         if (!deleting_snapshot && S_ISDIR(inode_u->bi_mode)) {
 273                 ret = bch2_empty_dir_trans(trans, inum);
 274                 if (ret)
 275                         goto err;
 276         }
 277
 278         if (deleting_snapshot && !inode_u->bi_subvol) {
 279                 ret = -ENOENT;
 280                 goto err;
 281         }
 282
 283         if (deleting_snapshot || inode_u->bi_subvol) {
 284                 ret = bch2_subvolume_unlink(trans, inode_u->bi_subvol);
 285                 if (ret)
 286                         goto err;
 287
 288                 k = bch2_btree_iter_peek_slot(&dirent_iter);
 289                 ret = bkey_err(k);
 290                 if (ret)
 291                         goto err;
 292
 293                 /*
 294                  * If we're deleting a subvolume, we need to really delete the
 295                  * dirent, not just emit a whiteout in the current snapshot:
 296                  */
 297                 bch2_btree_iter_set_snapshot(&dirent_iter, k.k->p.snapshot);
 298                 ret = bch2_btree_iter_traverse(&dirent_iter);
 299                 if (ret)
 300                         goto err;
 301         } else {
 302                 bch2_inode_nlink_dec(trans, inode_u);
 303         }
 304
 305         if (inode_u->bi_dir             == dirent_iter.pos.inode &&
 306             inode_u->bi_dir_offset      == dirent_iter.pos.offset) {
 307                 inode_u->bi_dir         = 0;
 308                 inode_u->bi_dir_offset  = 0;
 309         }
 310
 311         dir_u->bi_mtime = dir_u->bi_ctime = inode_u->bi_ctime = now;
 312         dir_u->bi_nlink -= is_subdir_for_nlink(inode_u);
 313
 314         ret =   bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
 315                                     &dir_hash, &dirent_iter,
 316                                     BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 317                 bch2_inode_write(trans, &dir_iter, dir_u) ?:
 318                 bch2_inode_write(trans, &inode_iter, inode_u);
 319 err:
 320         bch2_trans_iter_exit(trans, &inode_iter);
 321         bch2_trans_iter_exit(trans, &dirent_iter);
 322         bch2_trans_iter_exit(trans, &dir_iter);
 323         return ret;
 324 }
 325
 326 bool bch2_reinherit_attrs(struct bch_inode_unpacked *dst_u,
 327                           struct bch_inode_unpacked *src_u)
 328 {
 329         u64 src, dst;
 330         unsigned id;
 331         bool ret = false;
 332
 333         for (id = 0; id < Inode_opt_nr; id++) {
 334                 /* Skip attributes that were explicitly set on this inode */
 335                 if (dst_u->bi_fields_set & (1 << id))
 336                         continue;
 337
 338                 src = bch2_inode_opt_get(src_u, id);
 339                 dst = bch2_inode_opt_get(dst_u, id);
 340
 341                 if (src == dst)
 342                         continue;
 343
 344                 bch2_inode_opt_set(dst_u, id, src);
 345                 ret = true;
 346         }
 347
 348         return ret;
 349 }
 350
 351 int bch2_rename_trans(struct btree_trans *trans,
 352                       subvol_inum src_dir, struct bch_inode_unpacked *src_dir_u,
 353                       subvol_inum dst_dir, struct bch_inode_unpacked *dst_dir_u,
 354                       struct bch_inode_unpacked *src_inode_u,
 355                       struct bch_inode_unpacked *dst_inode_u,
 356                       const struct qstr *src_name,
 357                       const struct qstr *dst_name,
 358                       enum bch_rename_mode mode)
 359 {
 360         struct bch_fs *c = trans->c;
 361         struct btree_iter src_dir_iter = { NULL };
 362         struct btree_iter dst_dir_iter = { NULL };
 363         struct btree_iter src_inode_iter = { NULL };
 364         struct btree_iter dst_inode_iter = { NULL };
 365         struct bch_hash_info src_hash, dst_hash;
 366         subvol_inum src_inum, dst_inum;
 367         u64 src_offset, dst_offset;
 368         u64 now = bch2_current_time(c);
 369         int ret;
 370
 371         ret = bch2_inode_peek(trans, &src_dir_iter, src_dir_u, src_dir,
 372                               BTREE_ITER_INTENT);
 373         if (ret)
 374                 goto err;
 375
 376         src_hash = bch2_hash_info_init(c, src_dir_u);
 377
 378         if (dst_dir.inum        != src_dir.inum ||
 379             dst_dir.subvol      != src_dir.subvol) {
 380                 ret = bch2_inode_peek(trans, &dst_dir_iter, dst_dir_u, dst_dir,
 381                                       BTREE_ITER_INTENT);
 382                 if (ret)
 383                         goto err;
 384
 385                 dst_hash = bch2_hash_info_init(c, dst_dir_u);
 386         } else {
 387                 dst_dir_u = src_dir_u;
 388                 dst_hash = src_hash;
 389         }
 390
 391         ret = bch2_dirent_rename(trans,
 392                                  src_dir, &src_hash,
 393                                  dst_dir, &dst_hash,
 394                                  src_name, &src_inum, &src_offset,
 395                                  dst_name, &dst_inum, &dst_offset,
 396                                  mode);
 397         if (ret)
 398                 goto err;
 399
 400         ret = bch2_inode_peek(trans, &src_inode_iter, src_inode_u, src_inum,
 401                               BTREE_ITER_INTENT);
 402         if (ret)
 403                 goto err;
 404
 405         if (dst_inum.inum) {
 406                 ret = bch2_inode_peek(trans, &dst_inode_iter, dst_inode_u, dst_inum,
 407                                       BTREE_ITER_INTENT);
 408                 if (ret)
 409                         goto err;
 410         }
 411
 412         if (c->sb.version >= bcachefs_metadata_version_inode_backpointers) {
 413                 src_inode_u->bi_dir             = dst_dir_u->bi_inum;
 414                 src_inode_u->bi_dir_offset      = dst_offset;
 415
 416                 if (mode == BCH_RENAME_EXCHANGE) {
 417                         dst_inode_u->bi_dir             = src_dir_u->bi_inum;
 418                         dst_inode_u->bi_dir_offset      = src_offset;
 419                 }
 420
 421                 if (mode == BCH_RENAME_OVERWRITE &&
 422                     dst_inode_u->bi_dir         == dst_dir_u->bi_inum &&
 423                     dst_inode_u->bi_dir_offset  == src_offset) {
 424                         dst_inode_u->bi_dir             = 0;
 425                         dst_inode_u->bi_dir_offset      = 0;
 426                 }
 427         }
 428
 429         if (mode == BCH_RENAME_OVERWRITE) {
 430                 if (S_ISDIR(src_inode_u->bi_mode) !=
 431                     S_ISDIR(dst_inode_u->bi_mode)) {
 432                         ret = -ENOTDIR;
 433                         goto err;
 434                 }
 435
 436                 if (S_ISDIR(dst_inode_u->bi_mode) &&
 437                     bch2_empty_dir_trans(trans, dst_inum)) {
 438                         ret = -ENOTEMPTY;
 439                         goto err;
 440                 }
 441         }
 442
 443         if (bch2_reinherit_attrs(src_inode_u, dst_dir_u) &&
 444             S_ISDIR(src_inode_u->bi_mode)) {
 445                 ret = -EXDEV;
 446                 goto err;
 447         }
 448
 449         if (mode == BCH_RENAME_EXCHANGE &&
 450             bch2_reinherit_attrs(dst_inode_u, src_dir_u) &&
 451             S_ISDIR(dst_inode_u->bi_mode)) {
 452                 ret = -EXDEV;
 453                 goto err;
 454         }
 455
 456         if (is_subdir_for_nlink(src_inode_u)) {
 457                 src_dir_u->bi_nlink--;
 458                 dst_dir_u->bi_nlink++;
 459         }
 460
 461         if (dst_inum.inum && is_subdir_for_nlink(dst_inode_u)) {
 462                 dst_dir_u->bi_nlink--;
 463                 src_dir_u->bi_nlink += mode == BCH_RENAME_EXCHANGE;
 464         }
 465
 466         if (mode == BCH_RENAME_OVERWRITE)
 467                 bch2_inode_nlink_dec(trans, dst_inode_u);
 468
 469         src_dir_u->bi_mtime             = now;
 470         src_dir_u->bi_ctime             = now;
 471
 472         if (src_dir.inum != dst_dir.inum) {
 473                 dst_dir_u->bi_mtime     = now;
 474                 dst_dir_u->bi_ctime     = now;
 475         }
 476
 477         src_inode_u->bi_ctime           = now;
 478
 479         if (dst_inum.inum)
 480                 dst_inode_u->bi_ctime   = now;
 481
 482         ret =   bch2_inode_write(trans, &src_dir_iter, src_dir_u) ?:
 483                 (src_dir.inum != dst_dir.inum
 484                  ? bch2_inode_write(trans, &dst_dir_iter, dst_dir_u)
 485                  : 0 ) ?:
 486                 bch2_inode_write(trans, &src_inode_iter, src_inode_u) ?:
 487                 (dst_inum.inum
 488                  ? bch2_inode_write(trans, &dst_inode_iter, dst_inode_u)
 489                  : 0 );
 490 err:
 491         bch2_trans_iter_exit(trans, &dst_inode_iter);
 492         bch2_trans_iter_exit(trans, &src_inode_iter);
 493         bch2_trans_iter_exit(trans, &dst_dir_iter);
 494         bch2_trans_iter_exit(trans, &src_dir_iter);
 495         return ret;
 496 }