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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "bkey_methods.h"
   6 #include "btree_update.h"
   7 #include "extents.h"
   8 #include "dirent.h"
   9 #include "fs.h"
  10 #include "keylist.h"
  11 #include "str_hash.h"
  12 #include "subvolume.h"
  13
  14 #include <linux/dcache.h>
  15
  16 unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
  17 {
  18         unsigned len = bkey_val_bytes(d.k) -
  19                 offsetof(struct bch_dirent, d_name);
  20
  21         return strnlen(d.v->d_name, len);
  22 }
  23
  24 static u64 bch2_dirent_hash(const struct bch_hash_info *info,
  25                             const struct qstr *name)
  26 {
  27         struct bch_str_hash_ctx ctx;
  28
  29         bch2_str_hash_init(&ctx, info);
  30         bch2_str_hash_update(&ctx, info, name->name, name->len);
  31
  32         /* [0,2) reserved for dots */
  33         return max_t(u64, bch2_str_hash_end(&ctx, info), 2);
  34 }
  35
  36 static u64 dirent_hash_key(const struct bch_hash_info *info, const void *key)
  37 {
  38         return bch2_dirent_hash(info, key);
  39 }
  40
  41 static u64 dirent_hash_bkey(const struct bch_hash_info *info, struct bkey_s_c k)
  42 {
  43         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  44         struct qstr name = QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
  45
  46         return bch2_dirent_hash(info, &name);
  47 }
  48
  49 static bool dirent_cmp_key(struct bkey_s_c _l, const void *_r)
  50 {
  51         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  52         int len = bch2_dirent_name_bytes(l);
  53         const struct qstr *r = _r;
  54
  55         return len - r->len ?: memcmp(l.v->d_name, r->name, len);
  56 }
  57
  58 static bool dirent_cmp_bkey(struct bkey_s_c _l, struct bkey_s_c _r)
  59 {
  60         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  61         struct bkey_s_c_dirent r = bkey_s_c_to_dirent(_r);
  62         int l_len = bch2_dirent_name_bytes(l);
  63         int r_len = bch2_dirent_name_bytes(r);
  64
  65         return l_len - r_len ?: memcmp(l.v->d_name, r.v->d_name, l_len);
  66 }
  67
  68 static bool dirent_is_visible(subvol_inum inum, struct bkey_s_c k)
  69 {
  70         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  71
  72         if (d.v->d_type == DT_SUBVOL)
  73                 return le32_to_cpu(d.v->d_parent_subvol) == inum.subvol;
  74         return true;
  75 }
  76
  77 const struct bch_hash_desc bch2_dirent_hash_desc = {
  78         .btree_id       = BTREE_ID_dirents,
  79         .key_type       = KEY_TYPE_dirent,
  80         .hash_key       = dirent_hash_key,
  81         .hash_bkey      = dirent_hash_bkey,
  82         .cmp_key        = dirent_cmp_key,
  83         .cmp_bkey       = dirent_cmp_bkey,
  84         .is_visible     = dirent_is_visible,
  85 };
  86
  87 int bch2_dirent_invalid(const struct bch_fs *c, struct bkey_s_c k,
  88                         unsigned flags, struct printbuf *err)
  89 {
  90         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  91         unsigned len;
  92
  93         len = bch2_dirent_name_bytes(d);
  94         if (!len) {
  95                 prt_printf(err, "empty name");
  96                 return -BCH_ERR_invalid_bkey;
  97         }
  98
  99         if (bkey_val_u64s(k.k) > dirent_val_u64s(len)) {
 100                 prt_printf(err, "value too big (%zu > %u)",
 101                        bkey_val_u64s(k.k), dirent_val_u64s(len));
 102                 return -BCH_ERR_invalid_bkey;
 103         }
 104
 105         if (len > BCH_NAME_MAX) {
 106                 prt_printf(err, "dirent name too big (%u > %u)",
 107                        len, BCH_NAME_MAX);
 108                 return -BCH_ERR_invalid_bkey;
 109         }
 110
 111         if (len == 1 && !memcmp(d.v->d_name, ".", 1)) {
 112                 prt_printf(err, "invalid name");
 113                 return -BCH_ERR_invalid_bkey;
 114         }
 115
 116         if (len == 2 && !memcmp(d.v->d_name, "..", 2)) {
 117                 prt_printf(err, "invalid name");
 118                 return -BCH_ERR_invalid_bkey;
 119         }
 120
 121         if (memchr(d.v->d_name, '/', len)) {
 122                 prt_printf(err, "invalid name");
 123                 return -BCH_ERR_invalid_bkey;
 124         }
 125
 126         if (d.v->d_type != DT_SUBVOL &&
 127             le64_to_cpu(d.v->d_inum) == d.k->p.inode) {
 128                 prt_printf(err, "dirent points to own directory");
 129                 return -BCH_ERR_invalid_bkey;
 130         }
 131
 132         return 0;
 133 }
 134
 135 void bch2_dirent_to_text(struct printbuf *out, struct bch_fs *c,
 136                          struct bkey_s_c k)
 137 {
 138         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 139
 140         prt_printf(out, "%.*s -> %llu type %s",
 141                bch2_dirent_name_bytes(d),
 142                d.v->d_name,
 143                d.v->d_type != DT_SUBVOL
 144                ? le64_to_cpu(d.v->d_inum)
 145                : le32_to_cpu(d.v->d_child_subvol),
 146                bch2_d_type_str(d.v->d_type));
 147 }
 148
 149 static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
 150                                 subvol_inum dir, u8 type,
 151                                 const struct qstr *name, u64 dst)
 152 {
 153         struct bkey_i_dirent *dirent;
 154         unsigned u64s = BKEY_U64s + dirent_val_u64s(name->len);
 155
 156         if (name->len > BCH_NAME_MAX)
 157                 return ERR_PTR(-ENAMETOOLONG);
 158
 159         BUG_ON(u64s > U8_MAX);
 160
 161         dirent = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
 162         if (IS_ERR(dirent))
 163                 return dirent;
 164
 165         bkey_dirent_init(&dirent->k_i);
 166         dirent->k.u64s = u64s;
 167
 168         if (type != DT_SUBVOL) {
 169                 dirent->v.d_inum = cpu_to_le64(dst);
 170         } else {
 171                 dirent->v.d_parent_subvol = cpu_to_le32(dir.subvol);
 172                 dirent->v.d_child_subvol = cpu_to_le32(dst);
 173         }
 174
 175         dirent->v.d_type = type;
 176
 177         memcpy(dirent->v.d_name, name->name, name->len);
 178         memset(dirent->v.d_name + name->len, 0,
 179                bkey_val_bytes(&dirent->k) -
 180                offsetof(struct bch_dirent, d_name) -
 181                name->len);
 182
 183         EBUG_ON(bch2_dirent_name_bytes(dirent_i_to_s_c(dirent)) != name->len);
 184
 185         return dirent;
 186 }
 187
 188 int bch2_dirent_create(struct btree_trans *trans, subvol_inum dir,
 189                        const struct bch_hash_info *hash_info,
 190                        u8 type, const struct qstr *name, u64 dst_inum,
 191                        u64 *dir_offset, int flags)
 192 {
 193         struct bkey_i_dirent *dirent;
 194         int ret;
 195
 196         dirent = dirent_create_key(trans, dir, type, name, dst_inum);
 197         ret = PTR_ERR_OR_ZERO(dirent);
 198         if (ret)
 199                 return ret;
 200
 201         ret = bch2_hash_set(trans, bch2_dirent_hash_desc, hash_info,
 202                             dir, &dirent->k_i, flags);
 203         *dir_offset = dirent->k.p.offset;
 204
 205         return ret;
 206 }
 207
 208 static void dirent_copy_target(struct bkey_i_dirent *dst,
 209                                struct bkey_s_c_dirent src)
 210 {
 211         dst->v.d_inum = src.v->d_inum;
 212         dst->v.d_type = src.v->d_type;
 213 }
 214
 215 int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
 216                             struct bkey_s_c_dirent d, subvol_inum *target)
 217 {
 218         struct bch_subvolume s;
 219         int ret = 0;
 220
 221         if (d.v->d_type == DT_SUBVOL &&
 222             d.v->d_parent_subvol != dir.subvol)
 223                 return 1;
 224
 225         if (likely(d.v->d_type != DT_SUBVOL)) {
 226                 target->subvol  = dir.subvol;
 227                 target->inum    = le64_to_cpu(d.v->d_inum);
 228         } else {
 229                 target->subvol  = le32_to_cpu(d.v->d_child_subvol);
 230
 231                 ret = bch2_subvolume_get(trans, target->subvol, true, BTREE_ITER_CACHED, &s);
 232
 233                 target->inum    = le64_to_cpu(s.inode);
 234         }
 235
 236         return ret;
 237 }
 238
 239 int bch2_dirent_rename(struct btree_trans *trans,
 240                 subvol_inum src_dir, struct bch_hash_info *src_hash,
 241                 subvol_inum dst_dir, struct bch_hash_info *dst_hash,
 242                 const struct qstr *src_name, subvol_inum *src_inum, u64 *src_offset,
 243                 const struct qstr *dst_name, subvol_inum *dst_inum, u64 *dst_offset,
 244                 enum bch_rename_mode mode)
 245 {
 246         struct btree_iter src_iter = { NULL };
 247         struct btree_iter dst_iter = { NULL };
 248         struct bkey_s_c old_src, old_dst = bkey_s_c_null;
 249         struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 250         struct bpos dst_pos =
 251                 POS(dst_dir.inum, bch2_dirent_hash(dst_hash, dst_name));
 252         unsigned src_type = 0, dst_type = 0, src_update_flags = 0;
 253         int ret = 0;
 254
 255         if (src_dir.subvol != dst_dir.subvol)
 256                 return -EXDEV;
 257
 258         memset(src_inum, 0, sizeof(*src_inum));
 259         memset(dst_inum, 0, sizeof(*dst_inum));
 260
 261         /* Lookup src: */
 262         ret = bch2_hash_lookup(trans, &src_iter, bch2_dirent_hash_desc,
 263                                src_hash, src_dir, src_name,
 264                                BTREE_ITER_INTENT);
 265         if (ret)
 266                 goto out;
 267
 268         old_src = bch2_btree_iter_peek_slot(&src_iter);
 269         ret = bkey_err(old_src);
 270         if (ret)
 271                 goto out;
 272
 273         ret = bch2_dirent_read_target(trans, src_dir,
 274                         bkey_s_c_to_dirent(old_src), src_inum);
 275         if (ret)
 276                 goto out;
 277
 278         src_type = bkey_s_c_to_dirent(old_src).v->d_type;
 279
 280         if (src_type == DT_SUBVOL && mode == BCH_RENAME_EXCHANGE)
 281                 return -EOPNOTSUPP;
 282
 283
 284         /* Lookup dst: */
 285         if (mode == BCH_RENAME) {
 286                 /*
 287                  * Note that we're _not_ checking if the target already exists -
 288                  * we're relying on the VFS to do that check for us for
 289                  * correctness:
 290                  */
 291                 ret = bch2_hash_hole(trans, &dst_iter, bch2_dirent_hash_desc,
 292                                      dst_hash, dst_dir, dst_name);
 293                 if (ret)
 294                         goto out;
 295         } else {
 296                 ret = bch2_hash_lookup(trans, &dst_iter, bch2_dirent_hash_desc,
 297                                        dst_hash, dst_dir, dst_name,
 298                                        BTREE_ITER_INTENT);
 299                 if (ret)
 300                         goto out;
 301
 302                 old_dst = bch2_btree_iter_peek_slot(&dst_iter);
 303                 ret = bkey_err(old_dst);
 304                 if (ret)
 305                         goto out;
 306
 307                 ret = bch2_dirent_read_target(trans, dst_dir,
 308                                 bkey_s_c_to_dirent(old_dst), dst_inum);
 309                 if (ret)
 310                         goto out;
 311
 312                 dst_type = bkey_s_c_to_dirent(old_dst).v->d_type;
 313
 314                 if (dst_type == DT_SUBVOL)
 315                         return -EOPNOTSUPP;
 316         }
 317
 318         if (mode != BCH_RENAME_EXCHANGE)
 319                 *src_offset = dst_iter.pos.offset;
 320
 321         /* Create new dst key: */
 322         new_dst = dirent_create_key(trans, dst_dir, 0, dst_name, 0);
 323         ret = PTR_ERR_OR_ZERO(new_dst);
 324         if (ret)
 325                 goto out;
 326
 327         dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 328         new_dst->k.p = dst_iter.pos;
 329
 330         /* Create new src key: */
 331         if (mode == BCH_RENAME_EXCHANGE) {
 332                 new_src = dirent_create_key(trans, src_dir, 0, src_name, 0);
 333                 ret = PTR_ERR_OR_ZERO(new_src);
 334                 if (ret)
 335                         goto out;
 336
 337                 dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
 338                 new_src->k.p = src_iter.pos;
 339         } else {
 340                 new_src = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 341                 ret = PTR_ERR_OR_ZERO(new_src);
 342                 if (ret)
 343                         goto out;
 344
 345                 bkey_init(&new_src->k);
 346                 new_src->k.p = src_iter.pos;
 347
 348                 if (bkey_le(dst_pos, src_iter.pos) &&
 349                     bkey_lt(src_iter.pos, dst_iter.pos)) {
 350                         /*
 351                          * We have a hash collision for the new dst key,
 352                          * and new_src - the key we're deleting - is between
 353                          * new_dst's hashed slot and the slot we're going to be
 354                          * inserting it into - oops.  This will break the hash
 355                          * table if we don't deal with it:
 356                          */
 357                         if (mode == BCH_RENAME) {
 358                                 /*
 359                                  * If we're not overwriting, we can just insert
 360                                  * new_dst at the src position:
 361                                  */
 362                                 new_src = new_dst;
 363                                 new_src->k.p = src_iter.pos;
 364                                 goto out_set_src;
 365                         } else {
 366                                 /* If we're overwriting, we can't insert new_dst
 367                                  * at a different slot because it has to
 368                                  * overwrite old_dst - just make sure to use a
 369                                  * whiteout when deleting src:
 370                                  */
 371                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 372                         }
 373                 } else {
 374                         /* Check if we need a whiteout to delete src: */
 375                         ret = bch2_hash_needs_whiteout(trans, bch2_dirent_hash_desc,
 376                                                        src_hash, &src_iter);
 377                         if (ret < 0)
 378                                 goto out;
 379
 380                         if (ret)
 381                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 382                 }
 383         }
 384
 385         ret = bch2_trans_update(trans, &dst_iter, &new_dst->k_i, 0);
 386         if (ret)
 387                 goto out;
 388 out_set_src:
 389
 390         /*
 391          * If we're deleting a subvolume, we need to really delete the dirent,
 392          * not just emit a whiteout in the current snapshot:
 393          */
 394         if (src_type == DT_SUBVOL) {
 395                 bch2_btree_iter_set_snapshot(&src_iter, old_src.k->p.snapshot);
 396                 ret = bch2_btree_iter_traverse(&src_iter);
 397                 if (ret)
 398                         goto out;
 399
 400                 new_src->k.p = src_iter.pos;
 401                 src_update_flags |= BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE;
 402         }
 403
 404         ret = bch2_trans_update(trans, &src_iter, &new_src->k_i, src_update_flags);
 405         if (ret)
 406                 goto out;
 407
 408         if (mode == BCH_RENAME_EXCHANGE)
 409                 *src_offset = new_src->k.p.offset;
 410         *dst_offset = new_dst->k.p.offset;
 411 out:
 412         bch2_trans_iter_exit(trans, &src_iter);
 413         bch2_trans_iter_exit(trans, &dst_iter);
 414         return ret;
 415 }
 416
 417 int __bch2_dirent_lookup_trans(struct btree_trans *trans,
 418                                struct btree_iter *iter,
 419                                subvol_inum dir,
 420                                const struct bch_hash_info *hash_info,
 421                                const struct qstr *name, subvol_inum *inum,
 422                                unsigned flags)
 423 {
 424         struct bkey_s_c k;
 425         struct bkey_s_c_dirent d;
 426         u32 snapshot;
 427         int ret;
 428
 429         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 430         if (ret)
 431                 return ret;
 432
 433         ret = bch2_hash_lookup(trans, iter, bch2_dirent_hash_desc,
 434                                hash_info, dir, name, flags);
 435         if (ret)
 436                 return ret;
 437
 438         k = bch2_btree_iter_peek_slot(iter);
 439         ret = bkey_err(k);
 440         if (ret)
 441                 goto err;
 442
 443         d = bkey_s_c_to_dirent(k);
 444
 445         ret = bch2_dirent_read_target(trans, dir, d, inum);
 446         if (ret > 0)
 447                 ret = -ENOENT;
 448 err:
 449         if (ret)
 450                 bch2_trans_iter_exit(trans, iter);
 451
 452         return ret;
 453 }
 454
 455 u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 456                        const struct bch_hash_info *hash_info,
 457                        const struct qstr *name, subvol_inum *inum)
 458 {
 459         struct btree_trans trans;
 460         struct btree_iter iter;
 461         int ret;
 462
 463         bch2_trans_init(&trans, c, 0, 0);
 464 retry:
 465         bch2_trans_begin(&trans);
 466
 467         ret = __bch2_dirent_lookup_trans(&trans, &iter, dir, hash_info,
 468                                           name, inum, 0);
 469         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 470                 goto retry;
 471         if (!ret)
 472                 bch2_trans_iter_exit(&trans, &iter);
 473         bch2_trans_exit(&trans);
 474         return ret;
 475 }
 476
 477 int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 478 {
 479         struct btree_iter iter;
 480         struct bkey_s_c k;
 481         u32 snapshot;
 482         int ret;
 483
 484         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 485         if (ret)
 486                 return ret;
 487
 488         for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
 489                            SPOS(dir.inum, 0, snapshot),
 490                            POS(dir.inum, U64_MAX), 0, k, ret)
 491                 if (k.k->type == KEY_TYPE_dirent) {
 492                         ret = -ENOTEMPTY;
 493                         break;
 494                 }
 495         bch2_trans_iter_exit(trans, &iter);
 496
 497         return ret;
 498 }
 499
 500 int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 501 {
 502         struct btree_trans trans;
 503         struct btree_iter iter;
 504         struct bkey_s_c k;
 505         struct bkey_s_c_dirent dirent;
 506         subvol_inum target;
 507         u32 snapshot;
 508         struct bkey_buf sk;
 509         int ret;
 510
 511         bch2_bkey_buf_init(&sk);
 512         bch2_trans_init(&trans, c, 0, 0);
 513 retry:
 514         bch2_trans_begin(&trans);
 515
 516         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 517         if (ret)
 518                 goto err;
 519
 520         for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_dirents,
 521                            SPOS(inum.inum, ctx->pos, snapshot),
 522                            POS(inum.inum, U64_MAX), 0, k, ret) {
 523                 if (k.k->type != KEY_TYPE_dirent)
 524                         continue;
 525
 526                 dirent = bkey_s_c_to_dirent(k);
 527
 528                 ret = bch2_dirent_read_target(&trans, inum, dirent, &target);
 529                 if (ret < 0)
 530                         break;
 531                 if (ret)
 532                         continue;
 533
 534                 /* dir_emit() can fault and block: */
 535                 bch2_bkey_buf_reassemble(&sk, c, k);
 536                 dirent = bkey_i_to_s_c_dirent(sk.k);
 537                 bch2_trans_unlock(&trans);
 538
 539                 ctx->pos = dirent.k->p.offset;
 540                 if (!dir_emit(ctx, dirent.v->d_name,
 541                               bch2_dirent_name_bytes(dirent),
 542                               target.inum,
 543                               vfs_d_type(dirent.v->d_type)))
 544                         break;
 545                 ctx->pos = dirent.k->p.offset + 1;
 546
 547                 /*
 548                  * read_target looks up subvolumes, we can overflow paths if the
 549                  * directory has many subvolumes in it
 550                  */
 551                 ret = btree_trans_too_many_iters(&trans);
 552                 if (ret)
 553                         break;
 554         }
 555         bch2_trans_iter_exit(&trans, &iter);
 556 err:
 557         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 558                 goto retry;
 559
 560         bch2_trans_exit(&trans);
 561         bch2_bkey_buf_exit(&sk, c);
 562
 563         return ret;
 564 }