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         bch2_inode_nlink_inc(inode_u);
 208
 209         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_INTENT);
 210         if (ret)
 211                 goto err;
 212
 213         dir_u->bi_mtime = dir_u->bi_ctime = now;
 214
 215         dir_hash = bch2_hash_info_init(c, dir_u);
 216
 217         ret = bch2_dirent_create(trans, dir, &dir_hash,
 218                                  mode_to_type(inode_u->bi_mode),
 219                                  name, inum.inum, &dir_offset,
 220                                  BCH_HASH_SET_MUST_CREATE);
 221         if (ret)
 222                 goto err;
 223
 224         if (c->sb.version >= bcachefs_metadata_version_inode_backpointers) {
 225                 inode_u->bi_dir         = dir.inum;
 226                 inode_u->bi_dir_offset  = dir_offset;
 227         }
 228
 229         ret =   bch2_inode_write(trans, &dir_iter, dir_u) ?:
 230                 bch2_inode_write(trans, &inode_iter, inode_u);
 231 err:
 232         bch2_trans_iter_exit(trans, &dir_iter);
 233         bch2_trans_iter_exit(trans, &inode_iter);
 234         return ret;
 235 }
 236
 237 int bch2_unlink_trans(struct btree_trans *trans,
 238                       subvol_inum dir,
 239                       struct bch_inode_unpacked *dir_u,
 240                       struct bch_inode_unpacked *inode_u,
 241                       const struct qstr *name,
 242                       bool deleting_snapshot)
 243 {
 244         struct bch_fs *c = trans->c;
 245         struct btree_iter dir_iter = { NULL };
 246         struct btree_iter dirent_iter = { NULL };
 247         struct btree_iter inode_iter = { NULL };
 248         struct bch_hash_info dir_hash;
 249         subvol_inum inum;
 250         u64 now = bch2_current_time(c);
 251         struct bkey_s_c k;
 252         int ret;
 253
 254         ret = bch2_inode_peek(trans, &dir_iter, dir_u, dir, BTREE_ITER_INTENT);
 255         if (ret)
 256                 goto err;
 257
 258         dir_hash = bch2_hash_info_init(c, dir_u);
 259
 260         ret = __bch2_dirent_lookup_trans(trans, &dirent_iter, dir, &dir_hash,
 261                                          name, &inum, BTREE_ITER_INTENT);
 262         if (ret)
 263                 goto err;
 264
 265         ret = bch2_inode_peek(trans, &inode_iter, inode_u, inum,
 266                               BTREE_ITER_INTENT);
 267         if (ret)
 268                 goto err;
 269
 270         if (!deleting_snapshot && S_ISDIR(inode_u->bi_mode)) {
 271                 ret = bch2_empty_dir_trans(trans, inum);
 272                 if (ret)
 273                         goto err;
 274         }
 275
 276         if (deleting_snapshot && !inode_u->bi_subvol) {
 277                 ret = -ENOENT;
 278                 goto err;
 279         }
 280
 281         if (deleting_snapshot || inode_u->bi_subvol) {
 282                 ret = bch2_subvolume_unlink(trans, inode_u->bi_subvol);
 283                 if (ret)
 284                         goto err;
 285
 286                 k = bch2_btree_iter_peek_slot(&dirent_iter);
 287                 ret = bkey_err(k);
 288                 if (ret)
 289                         goto err;
 290
 291                 /*
 292                  * If we're deleting a subvolume, we need to really delete the
 293                  * dirent, not just emit a whiteout in the current snapshot:
 294                  */
 295                 bch2_btree_iter_set_snapshot(&dirent_iter, k.k->p.snapshot);
 296                 ret = bch2_btree_iter_traverse(&dirent_iter);
 297                 if (ret)
 298                         goto err;
 299         } else {
 300                 bch2_inode_nlink_dec(inode_u);
 301         }
 302
 303         if (inode_u->bi_dir             == dirent_iter.pos.inode &&
 304             inode_u->bi_dir_offset      == dirent_iter.pos.offset) {
 305                 inode_u->bi_dir         = 0;
 306                 inode_u->bi_dir_offset  = 0;
 307         }
 308
 309         dir_u->bi_mtime = dir_u->bi_ctime = inode_u->bi_ctime = now;
 310         dir_u->bi_nlink -= is_subdir_for_nlink(inode_u);
 311
 312         ret =   bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
 313                                     &dir_hash, &dirent_iter,
 314                                     BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 315                 bch2_inode_write(trans, &dir_iter, dir_u) ?:
 316                 bch2_inode_write(trans, &inode_iter, inode_u);
 317 err:
 318         bch2_trans_iter_exit(trans, &inode_iter);
 319         bch2_trans_iter_exit(trans, &dirent_iter);
 320         bch2_trans_iter_exit(trans, &dir_iter);
 321         return ret;
 322 }
 323
 324 bool bch2_reinherit_attrs(struct bch_inode_unpacked *dst_u,
 325                           struct bch_inode_unpacked *src_u)
 326 {
 327         u64 src, dst;
 328         unsigned id;
 329         bool ret = false;
 330
 331         for (id = 0; id < Inode_opt_nr; id++) {
 332                 /* Skip attributes that were explicitly set on this inode */
 333                 if (dst_u->bi_fields_set & (1 << id))
 334                         continue;
 335
 336                 src = bch2_inode_opt_get(src_u, id);
 337                 dst = bch2_inode_opt_get(dst_u, id);
 338
 339                 if (src == dst)
 340                         continue;
 341
 342                 bch2_inode_opt_set(dst_u, id, src);
 343                 ret = true;
 344         }
 345
 346         return ret;
 347 }
 348
 349 int bch2_rename_trans(struct btree_trans *trans,
 350                       subvol_inum src_dir, struct bch_inode_unpacked *src_dir_u,
 351                       subvol_inum dst_dir, struct bch_inode_unpacked *dst_dir_u,
 352                       struct bch_inode_unpacked *src_inode_u,
 353                       struct bch_inode_unpacked *dst_inode_u,
 354                       const struct qstr *src_name,
 355                       const struct qstr *dst_name,
 356                       enum bch_rename_mode mode)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct btree_iter src_dir_iter = { NULL };
 360         struct btree_iter dst_dir_iter = { NULL };
 361         struct btree_iter src_inode_iter = { NULL };
 362         struct btree_iter dst_inode_iter = { NULL };
 363         struct bch_hash_info src_hash, dst_hash;
 364         subvol_inum src_inum, dst_inum;
 365         u64 src_offset, dst_offset;
 366         u64 now = bch2_current_time(c);
 367         int ret;
 368
 369         ret = bch2_inode_peek(trans, &src_dir_iter, src_dir_u, src_dir,
 370                               BTREE_ITER_INTENT);
 371         if (ret)
 372                 goto err;
 373
 374         src_hash = bch2_hash_info_init(c, src_dir_u);
 375
 376         if (dst_dir.inum        != src_dir.inum ||
 377             dst_dir.subvol      != src_dir.subvol) {
 378                 ret = bch2_inode_peek(trans, &dst_dir_iter, dst_dir_u, dst_dir,
 379                                       BTREE_ITER_INTENT);
 380                 if (ret)
 381                         goto err;
 382
 383                 dst_hash = bch2_hash_info_init(c, dst_dir_u);
 384         } else {
 385                 dst_dir_u = src_dir_u;
 386                 dst_hash = src_hash;
 387         }
 388
 389         ret = bch2_dirent_rename(trans,
 390                                  src_dir, &src_hash,
 391                                  dst_dir, &dst_hash,
 392                                  src_name, &src_inum, &src_offset,
 393                                  dst_name, &dst_inum, &dst_offset,
 394                                  mode);
 395         if (ret)
 396                 goto err;
 397
 398         ret = bch2_inode_peek(trans, &src_inode_iter, src_inode_u, src_inum,
 399                               BTREE_ITER_INTENT);
 400         if (ret)
 401                 goto err;
 402
 403         if (dst_inum.inum) {
 404                 ret = bch2_inode_peek(trans, &dst_inode_iter, dst_inode_u, dst_inum,
 405                                       BTREE_ITER_INTENT);
 406                 if (ret)
 407                         goto err;
 408         }
 409
 410         if (c->sb.version >= bcachefs_metadata_version_inode_backpointers) {
 411                 src_inode_u->bi_dir             = dst_dir_u->bi_inum;
 412                 src_inode_u->bi_dir_offset      = dst_offset;
 413
 414                 if (mode == BCH_RENAME_EXCHANGE) {
 415                         dst_inode_u->bi_dir             = src_dir_u->bi_inum;
 416                         dst_inode_u->bi_dir_offset      = src_offset;
 417                 }
 418
 419                 if (mode == BCH_RENAME_OVERWRITE &&
 420                     dst_inode_u->bi_dir         == dst_dir_u->bi_inum &&
 421                     dst_inode_u->bi_dir_offset  == src_offset) {
 422                         dst_inode_u->bi_dir             = 0;
 423                         dst_inode_u->bi_dir_offset      = 0;
 424                 }
 425         }
 426
 427         if (mode == BCH_RENAME_OVERWRITE) {
 428                 if (S_ISDIR(src_inode_u->bi_mode) !=
 429                     S_ISDIR(dst_inode_u->bi_mode)) {
 430                         ret = -ENOTDIR;
 431                         goto err;
 432                 }
 433
 434                 if (S_ISDIR(dst_inode_u->bi_mode) &&
 435                     bch2_empty_dir_trans(trans, dst_inum)) {
 436                         ret = -ENOTEMPTY;
 437                         goto err;
 438                 }
 439         }
 440
 441         if (bch2_reinherit_attrs(src_inode_u, dst_dir_u) &&
 442             S_ISDIR(src_inode_u->bi_mode)) {
 443                 ret = -EXDEV;
 444                 goto err;
 445         }
 446
 447         if (mode == BCH_RENAME_EXCHANGE &&
 448             bch2_reinherit_attrs(dst_inode_u, src_dir_u) &&
 449             S_ISDIR(dst_inode_u->bi_mode)) {
 450                 ret = -EXDEV;
 451                 goto err;
 452         }
 453
 454         if (is_subdir_for_nlink(src_inode_u)) {
 455                 src_dir_u->bi_nlink--;
 456                 dst_dir_u->bi_nlink++;
 457         }
 458
 459         if (dst_inum.inum && is_subdir_for_nlink(dst_inode_u)) {
 460                 dst_dir_u->bi_nlink--;
 461                 src_dir_u->bi_nlink += mode == BCH_RENAME_EXCHANGE;
 462         }
 463
 464         if (mode == BCH_RENAME_OVERWRITE)
 465                 bch2_inode_nlink_dec(dst_inode_u);
 466
 467         src_dir_u->bi_mtime             = now;
 468         src_dir_u->bi_ctime             = now;
 469
 470         if (src_dir.inum != dst_dir.inum) {
 471                 dst_dir_u->bi_mtime     = now;
 472                 dst_dir_u->bi_ctime     = now;
 473         }
 474
 475         src_inode_u->bi_ctime           = now;
 476
 477         if (dst_inum.inum)
 478                 dst_inode_u->bi_ctime   = now;
 479
 480         ret =   bch2_inode_write(trans, &src_dir_iter, src_dir_u) ?:
 481                 (src_dir.inum != dst_dir.inum
 482                  ? bch2_inode_write(trans, &dst_dir_iter, dst_dir_u)
 483                  : 0 ) ?:
 484                 bch2_inode_write(trans, &src_inode_iter, src_inode_u) ?:
 485                 (dst_inum.inum
 486                  ? bch2_inode_write(trans, &dst_inode_iter, dst_inode_u)
 487                  : 0 );
 488 err:
 489         bch2_trans_iter_exit(trans, &dst_inode_iter);
 490         bch2_trans_iter_exit(trans, &src_inode_iter);
 491         bch2_trans_iter_exit(trans, &dst_dir_iter);
 492         bch2_trans_iter_exit(trans, &src_dir_iter);
 493         return ret;
 494 }