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_methods.h"
   5 #include "btree_update.h"
   6 #include "extents.h"
   7 #include "dirent.h"
   8 #include "fs.h"
   9 #include "keylist.h"
  10 #include "str_hash.h"
  11 #include "subvolume.h"
  12
  13 #include <linux/dcache.h>
  14
  15 unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
  16 {
  17         unsigned len = bkey_val_bytes(d.k) -
  18                 offsetof(struct bch_dirent, d_name);
  19
  20         return strnlen(d.v->d_name, len);
  21 }
  22
  23 static u64 bch2_dirent_hash(const struct bch_hash_info *info,
  24                             const struct qstr *name)
  25 {
  26         struct bch_str_hash_ctx ctx;
  27
  28         bch2_str_hash_init(&ctx, info);
  29         bch2_str_hash_update(&ctx, info, name->name, name->len);
  30
  31         /* [0,2) reserved for dots */
  32         return max_t(u64, bch2_str_hash_end(&ctx, info), 2);
  33 }
  34
  35 static u64 dirent_hash_key(const struct bch_hash_info *info, const void *key)
  36 {
  37         return bch2_dirent_hash(info, key);
  38 }
  39
  40 static u64 dirent_hash_bkey(const struct bch_hash_info *info, struct bkey_s_c k)
  41 {
  42         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  43         struct qstr name = QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
  44
  45         return bch2_dirent_hash(info, &name);
  46 }
  47
  48 static bool dirent_cmp_key(struct bkey_s_c _l, const void *_r)
  49 {
  50         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  51         int len = bch2_dirent_name_bytes(l);
  52         const struct qstr *r = _r;
  53
  54         return len - r->len ?: memcmp(l.v->d_name, r->name, len);
  55 }
  56
  57 static bool dirent_cmp_bkey(struct bkey_s_c _l, struct bkey_s_c _r)
  58 {
  59         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  60         struct bkey_s_c_dirent r = bkey_s_c_to_dirent(_r);
  61         int l_len = bch2_dirent_name_bytes(l);
  62         int r_len = bch2_dirent_name_bytes(r);
  63
  64         return l_len - r_len ?: memcmp(l.v->d_name, r.v->d_name, l_len);
  65 }
  66
  67 static bool dirent_is_visible(subvol_inum inum, struct bkey_s_c k)
  68 {
  69         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  70
  71         if (d.v->d_type == DT_SUBVOL)
  72                 return le32_to_cpu(d.v->d_parent_subvol) == inum.subvol;
  73         return true;
  74 }
  75
  76 const struct bch_hash_desc bch2_dirent_hash_desc = {
  77         .btree_id       = BTREE_ID_dirents,
  78         .key_type       = KEY_TYPE_dirent,
  79         .hash_key       = dirent_hash_key,
  80         .hash_bkey      = dirent_hash_bkey,
  81         .cmp_key        = dirent_cmp_key,
  82         .cmp_bkey       = dirent_cmp_bkey,
  83         .is_visible     = dirent_is_visible,
  84 };
  85
  86 int bch2_dirent_invalid(const struct bch_fs *c, struct bkey_s_c k,
  87                         unsigned flags, struct printbuf *err)
  88 {
  89         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  90         unsigned len;
  91
  92         if (bkey_val_bytes(k.k) < sizeof(struct bch_dirent)) {
  93                 prt_printf(err, "incorrect value size (%zu < %zu)",
  94                        bkey_val_bytes(k.k), sizeof(*d.v));
  95                 return -BCH_ERR_invalid_bkey;
  96         }
  97
  98         len = bch2_dirent_name_bytes(d);
  99         if (!len) {
 100                 prt_printf(err, "empty name");
 101                 return -BCH_ERR_invalid_bkey;
 102         }
 103
 104         if (bkey_val_u64s(k.k) > dirent_val_u64s(len)) {
 105                 prt_printf(err, "value too big (%zu > %u)",
 106                        bkey_val_u64s(k.k), dirent_val_u64s(len));
 107                 return -BCH_ERR_invalid_bkey;
 108         }
 109
 110         if (len > BCH_NAME_MAX) {
 111                 prt_printf(err, "dirent name too big (%u > %u)",
 112                        len, BCH_NAME_MAX);
 113                 return -BCH_ERR_invalid_bkey;
 114         }
 115
 116         if (len == 1 && !memcmp(d.v->d_name, ".", 1)) {
 117                 prt_printf(err, "invalid name");
 118                 return -BCH_ERR_invalid_bkey;
 119         }
 120
 121         if (len == 2 && !memcmp(d.v->d_name, "..", 2)) {
 122                 prt_printf(err, "invalid name");
 123                 return -BCH_ERR_invalid_bkey;
 124         }
 125
 126         if (memchr(d.v->d_name, '/', len)) {
 127                 prt_printf(err, "invalid name");
 128                 return -BCH_ERR_invalid_bkey;
 129         }
 130
 131         if (d.v->d_type != DT_SUBVOL &&
 132             le64_to_cpu(d.v->d_inum) == d.k->p.inode) {
 133                 prt_printf(err, "dirent points to own directory");
 134                 return -BCH_ERR_invalid_bkey;
 135         }
 136
 137         return 0;
 138 }
 139
 140 void bch2_dirent_to_text(struct printbuf *out, struct bch_fs *c,
 141                          struct bkey_s_c k)
 142 {
 143         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 144
 145         prt_printf(out, "%.*s -> %llu type %s",
 146                bch2_dirent_name_bytes(d),
 147                d.v->d_name,
 148                d.v->d_type != DT_SUBVOL
 149                ? le64_to_cpu(d.v->d_inum)
 150                : le32_to_cpu(d.v->d_child_subvol),
 151                bch2_d_type_str(d.v->d_type));
 152 }
 153
 154 static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
 155                                 subvol_inum dir, u8 type,
 156                                 const struct qstr *name, u64 dst)
 157 {
 158         struct bkey_i_dirent *dirent;
 159         unsigned u64s = BKEY_U64s + dirent_val_u64s(name->len);
 160
 161         if (name->len > BCH_NAME_MAX)
 162                 return ERR_PTR(-ENAMETOOLONG);
 163
 164         BUG_ON(u64s > U8_MAX);
 165
 166         dirent = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
 167         if (IS_ERR(dirent))
 168                 return dirent;
 169
 170         bkey_dirent_init(&dirent->k_i);
 171         dirent->k.u64s = u64s;
 172
 173         if (type != DT_SUBVOL) {
 174                 dirent->v.d_inum = cpu_to_le64(dst);
 175         } else {
 176                 dirent->v.d_parent_subvol = cpu_to_le32(dir.subvol);
 177                 dirent->v.d_child_subvol = cpu_to_le32(dst);
 178         }
 179
 180         dirent->v.d_type = type;
 181
 182         memcpy(dirent->v.d_name, name->name, name->len);
 183         memset(dirent->v.d_name + name->len, 0,
 184                bkey_val_bytes(&dirent->k) -
 185                offsetof(struct bch_dirent, d_name) -
 186                name->len);
 187
 188         EBUG_ON(bch2_dirent_name_bytes(dirent_i_to_s_c(dirent)) != name->len);
 189
 190         return dirent;
 191 }
 192
 193 int bch2_dirent_create(struct btree_trans *trans, subvol_inum dir,
 194                        const struct bch_hash_info *hash_info,
 195                        u8 type, const struct qstr *name, u64 dst_inum,
 196                        u64 *dir_offset, int flags)
 197 {
 198         struct bkey_i_dirent *dirent;
 199         int ret;
 200
 201         dirent = dirent_create_key(trans, dir, type, name, dst_inum);
 202         ret = PTR_ERR_OR_ZERO(dirent);
 203         if (ret)
 204                 return ret;
 205
 206         ret = bch2_hash_set(trans, bch2_dirent_hash_desc, hash_info,
 207                             dir, &dirent->k_i, flags);
 208         *dir_offset = dirent->k.p.offset;
 209
 210         return ret;
 211 }
 212
 213 static void dirent_copy_target(struct bkey_i_dirent *dst,
 214                                struct bkey_s_c_dirent src)
 215 {
 216         dst->v.d_inum = src.v->d_inum;
 217         dst->v.d_type = src.v->d_type;
 218 }
 219
 220 int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
 221                             struct bkey_s_c_dirent d, subvol_inum *target)
 222 {
 223         struct bch_subvolume s;
 224         int ret = 0;
 225
 226         if (d.v->d_type == DT_SUBVOL &&
 227             d.v->d_parent_subvol != dir.subvol)
 228                 return 1;
 229
 230         if (likely(d.v->d_type != DT_SUBVOL)) {
 231                 target->subvol  = dir.subvol;
 232                 target->inum    = le64_to_cpu(d.v->d_inum);
 233         } else {
 234                 target->subvol  = le32_to_cpu(d.v->d_child_subvol);
 235
 236                 ret = bch2_subvolume_get(trans, target->subvol, true, BTREE_ITER_CACHED, &s);
 237
 238                 target->inum    = le64_to_cpu(s.inode);
 239         }
 240
 241         return ret;
 242 }
 243
 244 int bch2_dirent_rename(struct btree_trans *trans,
 245                 subvol_inum src_dir, struct bch_hash_info *src_hash,
 246                 subvol_inum dst_dir, struct bch_hash_info *dst_hash,
 247                 const struct qstr *src_name, subvol_inum *src_inum, u64 *src_offset,
 248                 const struct qstr *dst_name, subvol_inum *dst_inum, u64 *dst_offset,
 249                 enum bch_rename_mode mode)
 250 {
 251         struct btree_iter src_iter = { NULL };
 252         struct btree_iter dst_iter = { NULL };
 253         struct bkey_s_c old_src, old_dst = bkey_s_c_null;
 254         struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 255         struct bpos dst_pos =
 256                 POS(dst_dir.inum, bch2_dirent_hash(dst_hash, dst_name));
 257         unsigned src_type = 0, dst_type = 0, src_update_flags = 0;
 258         int ret = 0;
 259
 260         if (src_dir.subvol != dst_dir.subvol)
 261                 return -EXDEV;
 262
 263         memset(src_inum, 0, sizeof(*src_inum));
 264         memset(dst_inum, 0, sizeof(*dst_inum));
 265
 266         /* Lookup src: */
 267         ret = bch2_hash_lookup(trans, &src_iter, bch2_dirent_hash_desc,
 268                                src_hash, src_dir, src_name,
 269                                BTREE_ITER_INTENT);
 270         if (ret)
 271                 goto out;
 272
 273         old_src = bch2_btree_iter_peek_slot(&src_iter);
 274         ret = bkey_err(old_src);
 275         if (ret)
 276                 goto out;
 277
 278         ret = bch2_dirent_read_target(trans, src_dir,
 279                         bkey_s_c_to_dirent(old_src), src_inum);
 280         if (ret)
 281                 goto out;
 282
 283         src_type = bkey_s_c_to_dirent(old_src).v->d_type;
 284
 285         if (src_type == DT_SUBVOL && mode == BCH_RENAME_EXCHANGE)
 286                 return -EOPNOTSUPP;
 287
 288
 289         /* Lookup dst: */
 290         if (mode == BCH_RENAME) {
 291                 /*
 292                  * Note that we're _not_ checking if the target already exists -
 293                  * we're relying on the VFS to do that check for us for
 294                  * correctness:
 295                  */
 296                 ret = bch2_hash_hole(trans, &dst_iter, bch2_dirent_hash_desc,
 297                                      dst_hash, dst_dir, dst_name);
 298                 if (ret)
 299                         goto out;
 300         } else {
 301                 ret = bch2_hash_lookup(trans, &dst_iter, bch2_dirent_hash_desc,
 302                                        dst_hash, dst_dir, dst_name,
 303                                        BTREE_ITER_INTENT);
 304                 if (ret)
 305                         goto out;
 306
 307                 old_dst = bch2_btree_iter_peek_slot(&dst_iter);
 308                 ret = bkey_err(old_dst);
 309                 if (ret)
 310                         goto out;
 311
 312                 ret = bch2_dirent_read_target(trans, dst_dir,
 313                                 bkey_s_c_to_dirent(old_dst), dst_inum);
 314                 if (ret)
 315                         goto out;
 316
 317                 dst_type = bkey_s_c_to_dirent(old_dst).v->d_type;
 318
 319                 if (dst_type == DT_SUBVOL)
 320                         return -EOPNOTSUPP;
 321         }
 322
 323         if (mode != BCH_RENAME_EXCHANGE)
 324                 *src_offset = dst_iter.pos.offset;
 325
 326         /* Create new dst key: */
 327         new_dst = dirent_create_key(trans, dst_dir, 0, dst_name, 0);
 328         ret = PTR_ERR_OR_ZERO(new_dst);
 329         if (ret)
 330                 goto out;
 331
 332         dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 333         new_dst->k.p = dst_iter.pos;
 334
 335         /* Create new src key: */
 336         if (mode == BCH_RENAME_EXCHANGE) {
 337                 new_src = dirent_create_key(trans, src_dir, 0, src_name, 0);
 338                 ret = PTR_ERR_OR_ZERO(new_src);
 339                 if (ret)
 340                         goto out;
 341
 342                 dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
 343                 new_src->k.p = src_iter.pos;
 344         } else {
 345                 new_src = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 346                 ret = PTR_ERR_OR_ZERO(new_src);
 347                 if (ret)
 348                         goto out;
 349
 350                 bkey_init(&new_src->k);
 351                 new_src->k.p = src_iter.pos;
 352
 353                 if (bkey_le(dst_pos, src_iter.pos) &&
 354                     bkey_lt(src_iter.pos, dst_iter.pos)) {
 355                         /*
 356                          * We have a hash collision for the new dst key,
 357                          * and new_src - the key we're deleting - is between
 358                          * new_dst's hashed slot and the slot we're going to be
 359                          * inserting it into - oops.  This will break the hash
 360                          * table if we don't deal with it:
 361                          */
 362                         if (mode == BCH_RENAME) {
 363                                 /*
 364                                  * If we're not overwriting, we can just insert
 365                                  * new_dst at the src position:
 366                                  */
 367                                 new_src = new_dst;
 368                                 new_src->k.p = src_iter.pos;
 369                                 goto out_set_src;
 370                         } else {
 371                                 /* If we're overwriting, we can't insert new_dst
 372                                  * at a different slot because it has to
 373                                  * overwrite old_dst - just make sure to use a
 374                                  * whiteout when deleting src:
 375                                  */
 376                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 377                         }
 378                 } else {
 379                         /* Check if we need a whiteout to delete src: */
 380                         ret = bch2_hash_needs_whiteout(trans, bch2_dirent_hash_desc,
 381                                                        src_hash, &src_iter);
 382                         if (ret < 0)
 383                                 goto out;
 384
 385                         if (ret)
 386                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 387                 }
 388         }
 389
 390         ret = bch2_trans_update(trans, &dst_iter, &new_dst->k_i, 0);
 391         if (ret)
 392                 goto out;
 393 out_set_src:
 394
 395         /*
 396          * If we're deleting a subvolume, we need to really delete the dirent,
 397          * not just emit a whiteout in the current snapshot:
 398          */
 399         if (src_type == DT_SUBVOL) {
 400                 bch2_btree_iter_set_snapshot(&src_iter, old_src.k->p.snapshot);
 401                 ret = bch2_btree_iter_traverse(&src_iter);
 402                 if (ret)
 403                         goto out;
 404
 405                 new_src->k.p = src_iter.pos;
 406                 src_update_flags |= BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE;
 407         }
 408
 409         ret = bch2_trans_update(trans, &src_iter, &new_src->k_i, src_update_flags);
 410         if (ret)
 411                 goto out;
 412
 413         if (mode == BCH_RENAME_EXCHANGE)
 414                 *src_offset = new_src->k.p.offset;
 415         *dst_offset = new_dst->k.p.offset;
 416 out:
 417         bch2_trans_iter_exit(trans, &src_iter);
 418         bch2_trans_iter_exit(trans, &dst_iter);
 419         return ret;
 420 }
 421
 422 int __bch2_dirent_lookup_trans(struct btree_trans *trans,
 423                                struct btree_iter *iter,
 424                                subvol_inum dir,
 425                                const struct bch_hash_info *hash_info,
 426                                const struct qstr *name, subvol_inum *inum,
 427                                unsigned flags)
 428 {
 429         struct bkey_s_c k;
 430         struct bkey_s_c_dirent d;
 431         u32 snapshot;
 432         int ret;
 433
 434         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 435         if (ret)
 436                 return ret;
 437
 438         ret = bch2_hash_lookup(trans, iter, bch2_dirent_hash_desc,
 439                                hash_info, dir, name, flags);
 440         if (ret)
 441                 return ret;
 442
 443         k = bch2_btree_iter_peek_slot(iter);
 444         ret = bkey_err(k);
 445         if (ret)
 446                 goto err;
 447
 448         d = bkey_s_c_to_dirent(k);
 449
 450         ret = bch2_dirent_read_target(trans, dir, d, inum);
 451         if (ret > 0)
 452                 ret = -ENOENT;
 453 err:
 454         if (ret)
 455                 bch2_trans_iter_exit(trans, iter);
 456
 457         return ret;
 458 }
 459
 460 u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 461                        const struct bch_hash_info *hash_info,
 462                        const struct qstr *name, subvol_inum *inum)
 463 {
 464         struct btree_trans trans;
 465         struct btree_iter iter;
 466         int ret;
 467
 468         bch2_trans_init(&trans, c, 0, 0);
 469 retry:
 470         bch2_trans_begin(&trans);
 471
 472         ret = __bch2_dirent_lookup_trans(&trans, &iter, dir, hash_info,
 473                                           name, inum, 0);
 474         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 475                 goto retry;
 476         if (!ret)
 477                 bch2_trans_iter_exit(&trans, &iter);
 478         bch2_trans_exit(&trans);
 479         return ret;
 480 }
 481
 482 int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 483 {
 484         struct btree_iter iter;
 485         struct bkey_s_c k;
 486         u32 snapshot;
 487         int ret;
 488
 489         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 490         if (ret)
 491                 return ret;
 492
 493         for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
 494                            SPOS(dir.inum, 0, snapshot),
 495                            POS(dir.inum, U64_MAX), 0, k, ret)
 496                 if (k.k->type == KEY_TYPE_dirent) {
 497                         ret = -ENOTEMPTY;
 498                         break;
 499                 }
 500         bch2_trans_iter_exit(trans, &iter);
 501
 502         return ret;
 503 }
 504
 505 int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 506 {
 507         struct btree_trans trans;
 508         struct btree_iter iter;
 509         struct bkey_s_c k;
 510         struct bkey_s_c_dirent dirent;
 511         subvol_inum target;
 512         u32 snapshot;
 513         int ret;
 514
 515         bch2_trans_init(&trans, c, 0, 0);
 516 retry:
 517         bch2_trans_begin(&trans);
 518
 519         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 520         if (ret)
 521                 goto err;
 522
 523         for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_dirents,
 524                            SPOS(inum.inum, ctx->pos, snapshot),
 525                            POS(inum.inum, U64_MAX), 0, k, ret) {
 526                 if (k.k->type != KEY_TYPE_dirent)
 527                         continue;
 528
 529                 dirent = bkey_s_c_to_dirent(k);
 530
 531                 ret = bch2_dirent_read_target(&trans, inum, dirent, &target);
 532                 if (ret < 0)
 533                         break;
 534                 if (ret)
 535                         continue;
 536
 537                 /*
 538                  * XXX: dir_emit() can fault and block, while we're holding
 539                  * locks
 540                  */
 541                 ctx->pos = dirent.k->p.offset;
 542                 if (!dir_emit(ctx, dirent.v->d_name,
 543                               bch2_dirent_name_bytes(dirent),
 544                               target.inum,
 545                               vfs_d_type(dirent.v->d_type)))
 546                         break;
 547                 ctx->pos = dirent.k->p.offset + 1;
 548
 549                 /*
 550                  * read_target looks up subvolumes, we can overflow paths if the
 551                  * directory has many subvolumes in it
 552                  */
 553                 ret = btree_trans_too_many_iters(&trans);
 554                 if (ret)
 555                         break;
 556         }
 557         bch2_trans_iter_exit(&trans, &iter);
 558 err:
 559         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 560                 goto retry;
 561
 562         bch2_trans_exit(&trans);
 563
 564         return ret;
 565 }