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 const struct bch_hash_desc bch2_dirent_hash_desc = {
  68         .btree_id       = BTREE_ID_dirents,
  69         .key_type       = KEY_TYPE_dirent,
  70         .hash_key       = dirent_hash_key,
  71         .hash_bkey      = dirent_hash_bkey,
  72         .cmp_key        = dirent_cmp_key,
  73         .cmp_bkey       = dirent_cmp_bkey,
  74 };
  75
  76 const char *bch2_dirent_invalid(const struct bch_fs *c, struct bkey_s_c k)
  77 {
  78         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  79         unsigned len;
  80
  81         if (bkey_val_bytes(k.k) < sizeof(struct bch_dirent))
  82                 return "value too small";
  83
  84         len = bch2_dirent_name_bytes(d);
  85         if (!len)
  86                 return "empty name";
  87
  88         if (bkey_val_u64s(k.k) > dirent_val_u64s(len))
  89                 return "value too big";
  90
  91         if (len > BCH_NAME_MAX)
  92                 return "dirent name too big";
  93
  94         if (len == 1 && !memcmp(d.v->d_name, ".", 1))
  95                 return "invalid name";
  96
  97         if (len == 2 && !memcmp(d.v->d_name, "..", 2))
  98                 return "invalid name";
  99
 100         if (memchr(d.v->d_name, '/', len))
 101                 return "invalid name";
 102
 103         if (d.v->d_type != DT_SUBVOL &&
 104             le64_to_cpu(d.v->d_inum) == d.k->p.inode)
 105                 return "dirent points to own directory";
 106
 107         return NULL;
 108 }
 109
 110 void bch2_dirent_to_text(struct printbuf *out, struct bch_fs *c,
 111                          struct bkey_s_c k)
 112 {
 113         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 114
 115         bch_scnmemcpy(out, d.v->d_name,
 116                       bch2_dirent_name_bytes(d));
 117         pr_buf(out, " -> %llu type %s", d.v->d_inum,
 118                d.v->d_type < BCH_DT_MAX
 119                ? bch2_d_types[d.v->d_type]
 120                : "(bad d_type)");
 121 }
 122
 123 static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
 124                                 u8 type, const struct qstr *name, u64 dst)
 125 {
 126         struct bkey_i_dirent *dirent;
 127         unsigned u64s = BKEY_U64s + dirent_val_u64s(name->len);
 128
 129         if (name->len > BCH_NAME_MAX)
 130                 return ERR_PTR(-ENAMETOOLONG);
 131
 132         BUG_ON(u64s > U8_MAX);
 133
 134         dirent = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
 135         if (IS_ERR(dirent))
 136                 return dirent;
 137
 138         bkey_dirent_init(&dirent->k_i);
 139         dirent->k.u64s = u64s;
 140         dirent->v.d_inum = cpu_to_le64(dst);
 141         dirent->v.d_type = type;
 142
 143         memcpy(dirent->v.d_name, name->name, name->len);
 144         memset(dirent->v.d_name + name->len, 0,
 145                bkey_val_bytes(&dirent->k) -
 146                offsetof(struct bch_dirent, d_name) -
 147                name->len);
 148
 149         EBUG_ON(bch2_dirent_name_bytes(dirent_i_to_s_c(dirent)) != name->len);
 150
 151         return dirent;
 152 }
 153
 154 int bch2_dirent_create(struct btree_trans *trans, subvol_inum dir,
 155                        const struct bch_hash_info *hash_info,
 156                        u8 type, const struct qstr *name, u64 dst_inum,
 157                        u64 *dir_offset, int flags)
 158 {
 159         struct bkey_i_dirent *dirent;
 160         int ret;
 161
 162         dirent = dirent_create_key(trans, type, name, dst_inum);
 163         ret = PTR_ERR_OR_ZERO(dirent);
 164         if (ret)
 165                 return ret;
 166
 167         ret = bch2_hash_set(trans, bch2_dirent_hash_desc, hash_info,
 168                             dir, &dirent->k_i, flags);
 169         *dir_offset = dirent->k.p.offset;
 170
 171         return ret;
 172 }
 173
 174 static void dirent_copy_target(struct bkey_i_dirent *dst,
 175                                struct bkey_s_c_dirent src)
 176 {
 177         dst->v.d_inum = src.v->d_inum;
 178         dst->v.d_type = src.v->d_type;
 179 }
 180
 181 int __bch2_dirent_read_target(struct btree_trans *trans,
 182                               struct bkey_s_c_dirent d,
 183                               u32 *subvol, u32 *snapshot, u64 *inum,
 184                               bool is_fsck)
 185 {
 186         struct bch_subvolume s;
 187         int ret = 0;
 188
 189         *subvol         = 0;
 190         *snapshot       = d.k->p.snapshot;
 191
 192         if (likely(d.v->d_type != DT_SUBVOL)) {
 193                 *inum = le64_to_cpu(d.v->d_inum);
 194         } else {
 195                 *subvol = le64_to_cpu(d.v->d_inum);
 196
 197                 ret = bch2_subvolume_get(trans, *subvol, !is_fsck, BTREE_ITER_CACHED, &s);
 198
 199                 *snapshot       = le32_to_cpu(s.snapshot);
 200                 *inum           = le64_to_cpu(s.inode);
 201         }
 202
 203         return ret;
 204 }
 205
 206 static int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
 207                                    struct bkey_s_c_dirent d, subvol_inum *target)
 208 {
 209         u32 snapshot;
 210         int ret = 0;
 211
 212         ret = __bch2_dirent_read_target(trans, d, &target->subvol, &snapshot,
 213                                         &target->inum, false);
 214         if (!target->subvol)
 215                 target->subvol = dir.subvol;
 216
 217         return ret;
 218 }
 219
 220 int bch2_dirent_rename(struct btree_trans *trans,
 221                 subvol_inum src_dir, struct bch_hash_info *src_hash,
 222                 subvol_inum dst_dir, struct bch_hash_info *dst_hash,
 223                 const struct qstr *src_name, subvol_inum *src_inum, u64 *src_offset,
 224                 const struct qstr *dst_name, subvol_inum *dst_inum, u64 *dst_offset,
 225                 enum bch_rename_mode mode)
 226 {
 227         struct btree_iter src_iter = { NULL };
 228         struct btree_iter dst_iter = { NULL };
 229         struct bkey_s_c old_src, old_dst;
 230         struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 231         struct bpos dst_pos =
 232                 POS(dst_dir.inum, bch2_dirent_hash(dst_hash, dst_name));
 233         int ret = 0;
 234
 235         if (src_dir.subvol != dst_dir.subvol)
 236                 return -EXDEV;
 237
 238         memset(src_inum, 0, sizeof(*src_inum));
 239         memset(dst_inum, 0, sizeof(*dst_inum));
 240
 241         /*
 242          * Lookup dst:
 243          *
 244          * Note that in BCH_RENAME mode, we're _not_ checking if
 245          * the target already exists - we're relying on the VFS
 246          * to do that check for us for correctness:
 247          */
 248         ret = mode == BCH_RENAME
 249                 ? bch2_hash_hole(trans, &dst_iter, bch2_dirent_hash_desc,
 250                                  dst_hash, dst_dir, dst_name)
 251                 : bch2_hash_lookup(trans, &dst_iter, bch2_dirent_hash_desc,
 252                                    dst_hash, dst_dir, dst_name,
 253                                    BTREE_ITER_INTENT);
 254         if (ret)
 255                 goto out;
 256
 257         old_dst = bch2_btree_iter_peek_slot(&dst_iter);
 258         ret = bkey_err(old_dst);
 259         if (ret)
 260                 goto out;
 261
 262         if (mode != BCH_RENAME) {
 263                 ret = bch2_dirent_read_target(trans, dst_dir,
 264                                 bkey_s_c_to_dirent(old_dst), dst_inum);
 265                 if (ret)
 266                         goto out;
 267         }
 268         if (mode != BCH_RENAME_EXCHANGE)
 269                 *src_offset = dst_iter.pos.offset;
 270
 271         /* Lookup src: */
 272         ret = bch2_hash_lookup(trans, &src_iter, bch2_dirent_hash_desc,
 273                                src_hash, src_dir, src_name,
 274                                BTREE_ITER_INTENT);
 275         if (ret)
 276                 goto out;
 277
 278         old_src = bch2_btree_iter_peek_slot(&src_iter);
 279         ret = bkey_err(old_src);
 280         if (ret)
 281                 goto out;
 282
 283         ret = bch2_dirent_read_target(trans, src_dir,
 284                         bkey_s_c_to_dirent(old_src), src_inum);
 285         if (ret)
 286                 goto out;
 287
 288         /* Create new dst key: */
 289         new_dst = dirent_create_key(trans, 0, dst_name, 0);
 290         ret = PTR_ERR_OR_ZERO(new_dst);
 291         if (ret)
 292                 goto out;
 293
 294         dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 295         new_dst->k.p = dst_iter.pos;
 296
 297         /* Create new src key: */
 298         if (mode == BCH_RENAME_EXCHANGE) {
 299                 new_src = dirent_create_key(trans, 0, src_name, 0);
 300                 ret = PTR_ERR_OR_ZERO(new_src);
 301                 if (ret)
 302                         goto out;
 303
 304                 dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
 305                 new_src->k.p = src_iter.pos;
 306         } else {
 307                 new_src = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 308                 ret = PTR_ERR_OR_ZERO(new_src);
 309                 if (ret)
 310                         goto out;
 311
 312                 bkey_init(&new_src->k);
 313                 new_src->k.p = src_iter.pos;
 314
 315                 if (bkey_cmp(dst_pos, src_iter.pos) <= 0 &&
 316                     bkey_cmp(src_iter.pos, dst_iter.pos) < 0) {
 317                         /*
 318                          * We have a hash collision for the new dst key,
 319                          * and new_src - the key we're deleting - is between
 320                          * new_dst's hashed slot and the slot we're going to be
 321                          * inserting it into - oops.  This will break the hash
 322                          * table if we don't deal with it:
 323                          */
 324                         if (mode == BCH_RENAME) {
 325                                 /*
 326                                  * If we're not overwriting, we can just insert
 327                                  * new_dst at the src position:
 328                                  */
 329                                 new_dst->k.p = src_iter.pos;
 330                                 bch2_trans_update(trans, &src_iter,
 331                                                   &new_dst->k_i, 0);
 332                                 goto out_set_offset;
 333                         } else {
 334                                 /* If we're overwriting, we can't insert new_dst
 335                                  * at a different slot because it has to
 336                                  * overwrite old_dst - just make sure to use a
 337                                  * whiteout when deleting src:
 338                                  */
 339                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 340                         }
 341                 } else {
 342                         /* Check if we need a whiteout to delete src: */
 343                         ret = bch2_hash_needs_whiteout(trans, bch2_dirent_hash_desc,
 344                                                        src_hash, &src_iter);
 345                         if (ret < 0)
 346                                 goto out;
 347
 348                         if (ret)
 349                                 new_src->k.type = KEY_TYPE_hash_whiteout;
 350                 }
 351         }
 352
 353         bch2_trans_update(trans, &src_iter, &new_src->k_i, 0);
 354         bch2_trans_update(trans, &dst_iter, &new_dst->k_i, 0);
 355 out_set_offset:
 356         if (mode == BCH_RENAME_EXCHANGE)
 357                 *src_offset = new_src->k.p.offset;
 358         *dst_offset = new_dst->k.p.offset;
 359 out:
 360         bch2_trans_iter_exit(trans, &src_iter);
 361         bch2_trans_iter_exit(trans, &dst_iter);
 362         return ret;
 363 }
 364
 365 int __bch2_dirent_lookup_trans(struct btree_trans *trans,
 366                                struct btree_iter *iter,
 367                                subvol_inum dir,
 368                                const struct bch_hash_info *hash_info,
 369                                const struct qstr *name, subvol_inum *inum,
 370                                unsigned flags)
 371 {
 372         struct bkey_s_c k;
 373         struct bkey_s_c_dirent d;
 374         u32 snapshot;
 375         int ret;
 376
 377         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 378         if (ret)
 379                 return ret;
 380
 381         ret = bch2_hash_lookup(trans, iter, bch2_dirent_hash_desc,
 382                                hash_info, dir, name, flags);
 383         if (ret)
 384                 return ret;
 385
 386         k = bch2_btree_iter_peek_slot(iter);
 387         ret = bkey_err(k);
 388         if (ret) {
 389                 bch2_trans_iter_exit(trans, iter);
 390                 return ret;
 391         }
 392
 393         d = bkey_s_c_to_dirent(k);
 394
 395         ret = bch2_dirent_read_target(trans, dir, d, inum);
 396         if (ret)
 397                 bch2_trans_iter_exit(trans, iter);
 398
 399         return ret;
 400 }
 401
 402 u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 403                        const struct bch_hash_info *hash_info,
 404                        const struct qstr *name, subvol_inum *inum)
 405 {
 406         struct btree_trans trans;
 407         struct btree_iter iter;
 408         int ret;
 409
 410         bch2_trans_init(&trans, c, 0, 0);
 411 retry:
 412         bch2_trans_begin(&trans);
 413
 414         ret = __bch2_dirent_lookup_trans(&trans, &iter, dir, hash_info,
 415                                           name, inum, 0);
 416
 417         bch2_trans_iter_exit(&trans, &iter);
 418         if (ret == -EINTR)
 419                 goto retry;
 420         bch2_trans_exit(&trans);
 421         return ret;
 422 }
 423
 424 int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 425 {
 426         struct btree_iter iter;
 427         struct bkey_s_c k;
 428         u32 snapshot;
 429         int ret;
 430
 431         ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
 432         if (ret)
 433                 return ret;
 434
 435         for_each_btree_key(trans, iter, BTREE_ID_dirents,
 436                            SPOS(dir.inum, 0, snapshot), 0, k, ret) {
 437                 if (k.k->p.inode > dir.inum)
 438                         break;
 439
 440                 if (k.k->type == KEY_TYPE_dirent) {
 441                         ret = -ENOTEMPTY;
 442                         break;
 443                 }
 444         }
 445         bch2_trans_iter_exit(trans, &iter);
 446
 447         return ret;
 448 }
 449
 450 int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 451 {
 452         struct btree_trans trans;
 453         struct btree_iter iter;
 454         struct bkey_s_c k;
 455         struct bkey_s_c_dirent dirent;
 456         u32 snapshot;
 457         int ret;
 458
 459         bch2_trans_init(&trans, c, 0, 0);
 460 retry:
 461         bch2_trans_begin(&trans);
 462
 463         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 464         if (ret)
 465                 goto err;
 466
 467         for_each_btree_key(&trans, iter, BTREE_ID_dirents,
 468                            SPOS(inum.inum, ctx->pos, snapshot), 0, k, ret) {
 469                 if (k.k->p.inode > inum.inum)
 470                         break;
 471
 472                 if (k.k->type != KEY_TYPE_dirent)
 473                         continue;
 474
 475                 dirent = bkey_s_c_to_dirent(k);
 476
 477                 /*
 478                  * XXX: dir_emit() can fault and block, while we're holding
 479                  * locks
 480                  */
 481                 ctx->pos = dirent.k->p.offset;
 482                 if (!dir_emit(ctx, dirent.v->d_name,
 483                               bch2_dirent_name_bytes(dirent),
 484                               le64_to_cpu(dirent.v->d_inum),
 485                               vfs_d_type(dirent.v->d_type)))
 486                         break;
 487                 ctx->pos = dirent.k->p.offset + 1;
 488         }
 489         bch2_trans_iter_exit(&trans, &iter);
 490 err:
 491         if (ret == -EINTR)
 492                 goto retry;
 493
 494         ret = bch2_trans_exit(&trans) ?: ret;
 495
 496         return ret;
 497 }