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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "bkey_methods.h"
   6 #include "btree_update.h"
   7 #include "error.h"
   8 #include "extents.h"
   9 #include "extent_update.h"
  10 #include "inode.h"
  11 #include "str_hash.h"
  12 #include "subvolume.h"
  13 #include "varint.h"
  14
  15 #include <linux/random.h>
  16
  17 #include <asm/unaligned.h>
  18
  19 const char * const bch2_inode_opts[] = {
  20 #define x(name, ...)    #name,
  21         BCH_INODE_OPTS()
  22 #undef  x
  23         NULL,
  24 };
  25
  26 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  27 static const u8 bits_table[8] = {
  28         1  * 8 - 1,
  29         2  * 8 - 2,
  30         3  * 8 - 3,
  31         4  * 8 - 4,
  32         6  * 8 - 5,
  33         8  * 8 - 6,
  34         10 * 8 - 7,
  35         13 * 8 - 8,
  36 };
  37
  38 static int inode_decode_field(const u8 *in, const u8 *end,
  39                               u64 out[2], unsigned *out_bits)
  40 {
  41         __be64 be[2] = { 0, 0 };
  42         unsigned bytes, shift;
  43         u8 *p;
  44
  45         if (in >= end)
  46                 return -1;
  47
  48         if (!*in)
  49                 return -1;
  50
  51         /*
  52          * position of highest set bit indicates number of bytes:
  53          * shift = number of bits to remove in high byte:
  54          */
  55         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  56         bytes   = byte_table[shift - 1];
  57
  58         if (in + bytes > end)
  59                 return -1;
  60
  61         p = (u8 *) be + 16 - bytes;
  62         memcpy(p, in, bytes);
  63         *p ^= (1 << 8) >> shift;
  64
  65         out[0] = be64_to_cpu(be[0]);
  66         out[1] = be64_to_cpu(be[1]);
  67         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  68
  69         return bytes;
  70 }
  71
  72 void bch2_inode_pack(struct bch_fs *c,
  73                      struct bkey_inode_buf *packed,
  74                      const struct bch_inode_unpacked *inode)
  75 {
  76         struct bkey_i_inode_v2 *k = &packed->inode;
  77         u8 *out = k->v.fields;
  78         u8 *end = (void *) &packed[1];
  79         u8 *last_nonzero_field = out;
  80         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  81         unsigned bytes;
  82         int ret;
  83
  84         bkey_inode_v2_init(&packed->inode.k_i);
  85         packed->inode.k.p.offset        = inode->bi_inum;
  86         packed->inode.v.bi_journal_seq  = cpu_to_le64(inode->bi_journal_seq);
  87         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  88         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  89         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  90         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
  91
  92 #define x(_name, _bits)                                                 \
  93         nr_fields++;                                                    \
  94                                                                         \
  95         if (inode->_name) {                                             \
  96                 ret = bch2_varint_encode_fast(out, inode->_name);       \
  97                 out += ret;                                             \
  98                                                                         \
  99                 if (_bits > 64)                                         \
 100                         *out++ = 0;                                     \
 101                                                                         \
 102                 last_nonzero_field = out;                               \
 103                 last_nonzero_fieldnr = nr_fields;                       \
 104         } else {                                                        \
 105                 *out++ = 0;                                             \
 106                                                                         \
 107                 if (_bits > 64)                                         \
 108                         *out++ = 0;                                     \
 109         }
 110
 111         BCH_INODE_FIELDS()
 112 #undef  x
 113         BUG_ON(out > end);
 114
 115         out = last_nonzero_field;
 116         nr_fields = last_nonzero_fieldnr;
 117
 118         bytes = out - (u8 *) &packed->inode.v;
 119         set_bkey_val_bytes(&packed->inode.k, bytes);
 120         memset_u64s_tail(&packed->inode.v, 0, bytes);
 121
 122         SET_INODEv2_NR_FIELDS(&k->v, nr_fields);
 123
 124         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 125                 struct bch_inode_unpacked unpacked;
 126
 127                 int ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i),
 128                                            &unpacked);
 129                 BUG_ON(ret);
 130                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 131                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 132                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 133
 134 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 135                         panic("unpacked %llu should be %llu",           \
 136                               (u64) unpacked._name, (u64) inode->_name);
 137                 BCH_INODE_FIELDS()
 138 #undef  x
 139         }
 140 }
 141
 142 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 143                                 struct bch_inode_unpacked *unpacked)
 144 {
 145         const u8 *in = inode.v->fields;
 146         const u8 *end = bkey_val_end(inode);
 147         u64 field[2];
 148         unsigned fieldnr = 0, field_bits;
 149         int ret;
 150
 151 #define x(_name, _bits)                                 \
 152         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 153                 memset(&unpacked->_name, 0,                             \
 154                        sizeof(*unpacked) -                              \
 155                        offsetof(struct bch_inode_unpacked, _name));     \
 156                 return 0;                                               \
 157         }                                                               \
 158                                                                         \
 159         ret = inode_decode_field(in, end, field, &field_bits);          \
 160         if (ret < 0)                                                    \
 161                 return ret;                                             \
 162                                                                         \
 163         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 164                 return -1;                                              \
 165                                                                         \
 166         unpacked->_name = field[1];                                     \
 167         in += ret;
 168
 169         BCH_INODE_FIELDS()
 170 #undef  x
 171
 172         /* XXX: signal if there were more fields than expected? */
 173         return 0;
 174 }
 175
 176 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 177                                 const u8 *in, const u8 *end,
 178                                 unsigned nr_fields)
 179 {
 180         unsigned fieldnr = 0;
 181         int ret;
 182         u64 v[2];
 183
 184 #define x(_name, _bits)                                                 \
 185         if (fieldnr < nr_fields) {                                      \
 186                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 187                 if (ret < 0)                                            \
 188                         return ret;                                     \
 189                 in += ret;                                              \
 190                                                                         \
 191                 if (_bits > 64) {                                       \
 192                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 193                         if (ret < 0)                                    \
 194                                 return ret;                             \
 195                         in += ret;                                      \
 196                 } else {                                                \
 197                         v[1] = 0;                                       \
 198                 }                                                       \
 199         } else {                                                        \
 200                 v[0] = v[1] = 0;                                        \
 201         }                                                               \
 202                                                                         \
 203         unpacked->_name = v[0];                                         \
 204         if (v[1] || v[0] != unpacked->_name)                            \
 205                 return -1;                                              \
 206         fieldnr++;
 207
 208         BCH_INODE_FIELDS()
 209 #undef  x
 210
 211         /* XXX: signal if there were more fields than expected? */
 212         return 0;
 213 }
 214
 215 int bch2_inode_unpack(struct bkey_s_c k,
 216                       struct bch_inode_unpacked *unpacked)
 217 {
 218         switch (k.k->type) {
 219         case KEY_TYPE_inode: {
 220                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 221
 222                 unpacked->bi_inum       = inode.k->p.offset;
 223                 unpacked->bi_journal_seq= 0;
 224                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 225                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 226                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 227
 228                 if (INODE_NEW_VARINT(inode.v)) {
 229                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 230                                                     bkey_val_end(inode),
 231                                                     INODE_NR_FIELDS(inode.v));
 232                 } else {
 233                         return bch2_inode_unpack_v1(inode, unpacked);
 234                 }
 235                 break;
 236         }
 237         case KEY_TYPE_inode_v2: {
 238                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 239
 240                 unpacked->bi_inum       = inode.k->p.offset;
 241                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 242                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 243                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 244                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 245
 246                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 247                                             bkey_val_end(inode),
 248                                             INODEv2_NR_FIELDS(inode.v));
 249         }
 250         default:
 251                 BUG();
 252         }
 253 }
 254
 255 int bch2_inode_peek(struct btree_trans *trans,
 256                     struct btree_iter *iter,
 257                     struct bch_inode_unpacked *inode,
 258                     subvol_inum inum, unsigned flags)
 259 {
 260         struct bkey_s_c k;
 261         u32 snapshot;
 262         int ret;
 263
 264         if (0 && trans->c->opts.inodes_use_key_cache)
 265                 flags |= BTREE_ITER_CACHED;
 266
 267         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 268         if (ret)
 269                 return ret;
 270
 271         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 272                              SPOS(0, inum.inum, snapshot), flags);
 273         k = bch2_btree_iter_peek_slot(iter);
 274         ret = bkey_err(k);
 275         if (ret)
 276                 goto err;
 277
 278         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 279         if (ret)
 280                 goto err;
 281
 282         ret = bch2_inode_unpack(k, inode);
 283         if (ret)
 284                 goto err;
 285
 286         return 0;
 287 err:
 288         bch2_trans_iter_exit(trans, iter);
 289         return ret;
 290 }
 291
 292 int bch2_inode_write(struct btree_trans *trans,
 293                      struct btree_iter *iter,
 294                      struct bch_inode_unpacked *inode)
 295 {
 296         struct bkey_inode_buf *inode_p;
 297
 298         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 299         if (IS_ERR(inode_p))
 300                 return PTR_ERR(inode_p);
 301
 302         bch2_inode_pack(trans->c, inode_p, inode);
 303         inode_p->inode.k.p.snapshot = iter->snapshot;
 304         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 305 }
 306
 307 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 308 {
 309         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 310         struct bch_inode_unpacked unpacked;
 311
 312         if (k.k->p.inode)
 313                 return "nonzero k.p.inode";
 314
 315         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 316                 return "incorrect value size";
 317
 318         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 319                 return "fs inode in blockdev range";
 320
 321         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 322                 return "invalid str hash type";
 323
 324         if (bch2_inode_unpack(k, &unpacked))
 325                 return "invalid variable length fields";
 326
 327         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 328                 return "invalid data checksum type";
 329
 330         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 331                 return "invalid data checksum type";
 332
 333         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 334             unpacked.bi_nlink != 0)
 335                 return "flagged as unlinked but bi_nlink != 0";
 336
 337         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 338                 return "subvolume root but not a directory";
 339
 340         return NULL;
 341 }
 342
 343 const char *bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 344 {
 345         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 346         struct bch_inode_unpacked unpacked;
 347
 348         if (k.k->p.inode)
 349                 return "nonzero k.p.inode";
 350
 351         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 352                 return "incorrect value size";
 353
 354         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 355                 return "fs inode in blockdev range";
 356
 357         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 358                 return "invalid str hash type";
 359
 360         if (bch2_inode_unpack(k, &unpacked))
 361                 return "invalid variable length fields";
 362
 363         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 364                 return "invalid data checksum type";
 365
 366         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 367                 return "invalid data checksum type";
 368
 369         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 370             unpacked.bi_nlink != 0)
 371                 return "flagged as unlinked but bi_nlink != 0";
 372
 373         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 374                 return "subvolume root but not a directory";
 375
 376         return NULL;
 377 }
 378
 379 static void __bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 380 {
 381         pr_buf(out, "mode %o flags %x journal_seq %llu",
 382                inode->bi_mode, inode->bi_flags,
 383                inode->bi_journal_seq);
 384
 385 #define x(_name, _bits)                                         \
 386         pr_buf(out, " "#_name " %llu", (u64) inode->_name);
 387         BCH_INODE_FIELDS()
 388 #undef  x
 389 }
 390
 391 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 392 {
 393         pr_buf(out, "inum: %llu ", inode->bi_inum);
 394         __bch2_inode_unpacked_to_text(out, inode);
 395 }
 396
 397 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 398                        struct bkey_s_c k)
 399 {
 400         struct bch_inode_unpacked inode;
 401
 402         if (bch2_inode_unpack(k, &inode)) {
 403                 pr_buf(out, "(unpack error)");
 404                 return;
 405         }
 406
 407         __bch2_inode_unpacked_to_text(out, &inode);
 408 }
 409
 410 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 411                                           struct bkey_s_c k)
 412 {
 413         if (k.k->p.inode)
 414                 return "nonzero k.p.inode";
 415
 416         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 417                 return "incorrect value size";
 418
 419         return NULL;
 420 }
 421
 422 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 423                                    struct bkey_s_c k)
 424 {
 425         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 426
 427         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 428 }
 429
 430 void bch2_inode_init_early(struct bch_fs *c,
 431                            struct bch_inode_unpacked *inode_u)
 432 {
 433         enum bch_str_hash_type str_hash =
 434                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 435
 436         memset(inode_u, 0, sizeof(*inode_u));
 437
 438         /* ick */
 439         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 440         get_random_bytes(&inode_u->bi_hash_seed,
 441                          sizeof(inode_u->bi_hash_seed));
 442 }
 443
 444 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 445                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 446                           struct bch_inode_unpacked *parent)
 447 {
 448         inode_u->bi_mode        = mode;
 449         inode_u->bi_uid         = uid;
 450         inode_u->bi_gid         = gid;
 451         inode_u->bi_dev         = rdev;
 452         inode_u->bi_atime       = now;
 453         inode_u->bi_mtime       = now;
 454         inode_u->bi_ctime       = now;
 455         inode_u->bi_otime       = now;
 456
 457         if (parent && parent->bi_mode & S_ISGID) {
 458                 inode_u->bi_gid = parent->bi_gid;
 459                 if (S_ISDIR(mode))
 460                         inode_u->bi_mode |= S_ISGID;
 461         }
 462
 463         if (parent) {
 464 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 465                 BCH_INODE_OPTS()
 466 #undef x
 467         }
 468 }
 469
 470 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 471                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 472                      struct bch_inode_unpacked *parent)
 473 {
 474         bch2_inode_init_early(c, inode_u);
 475         bch2_inode_init_late(inode_u, bch2_current_time(c),
 476                              uid, gid, mode, rdev, parent);
 477 }
 478
 479 static inline u32 bkey_generation(struct bkey_s_c k)
 480 {
 481         switch (k.k->type) {
 482         case KEY_TYPE_inode:
 483         case KEY_TYPE_inode_v2:
 484                 BUG();
 485         case KEY_TYPE_inode_generation:
 486                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 487         default:
 488                 return 0;
 489         }
 490 }
 491
 492 /*
 493  * This just finds an empty slot:
 494  */
 495 int bch2_inode_create(struct btree_trans *trans,
 496                       struct btree_iter *iter,
 497                       struct bch_inode_unpacked *inode_u,
 498                       u32 snapshot, u64 cpu)
 499 {
 500         struct bch_fs *c = trans->c;
 501         struct bkey_s_c k;
 502         u64 min, max, start, pos, *hint;
 503         int ret = 0;
 504         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 505
 506         if (c->opts.shard_inode_numbers) {
 507                 bits -= c->inode_shard_bits;
 508
 509                 min = (cpu << bits);
 510                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 511
 512                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 513                 hint = c->unused_inode_hints + cpu;
 514         } else {
 515                 min = BLOCKDEV_INODE_MAX;
 516                 max = ~(ULLONG_MAX << bits);
 517                 hint = c->unused_inode_hints;
 518         }
 519
 520         start = READ_ONCE(*hint);
 521
 522         if (start >= max || start < min)
 523                 start = min;
 524
 525         pos = start;
 526         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 527                              BTREE_ITER_ALL_SNAPSHOTS|
 528                              BTREE_ITER_INTENT);
 529 again:
 530         while ((k = bch2_btree_iter_peek(iter)).k &&
 531                !(ret = bkey_err(k)) &&
 532                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 533                 while (pos < iter->pos.offset) {
 534                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 535                                 goto found_slot;
 536
 537                         pos++;
 538                 }
 539
 540                 if (k.k->p.snapshot == snapshot &&
 541                     !bkey_is_inode(k.k) &&
 542                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 543                         bch2_btree_iter_advance(iter);
 544                         continue;
 545                 }
 546
 547                 /*
 548                  * We don't need to iterate over keys in every snapshot once
 549                  * we've found just one:
 550                  */
 551                 pos = iter->pos.offset + 1;
 552                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 553         }
 554
 555         while (!ret && pos < max) {
 556                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 557                         goto found_slot;
 558
 559                 pos++;
 560         }
 561
 562         if (!ret && start == min)
 563                 ret = -ENOSPC;
 564
 565         if (ret) {
 566                 bch2_trans_iter_exit(trans, iter);
 567                 return ret;
 568         }
 569
 570         /* Retry from start */
 571         pos = start = min;
 572         bch2_btree_iter_set_pos(iter, POS(0, pos));
 573         goto again;
 574 found_slot:
 575         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 576         k = bch2_btree_iter_peek_slot(iter);
 577         ret = bkey_err(k);
 578         if (ret) {
 579                 bch2_trans_iter_exit(trans, iter);
 580                 return ret;
 581         }
 582
 583         /* We may have raced while the iterator wasn't pointing at pos: */
 584         if (bkey_is_inode(k.k) ||
 585             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 586                 goto again;
 587
 588         *hint                   = k.k->p.offset;
 589         inode_u->bi_inum        = k.k->p.offset;
 590         inode_u->bi_generation  = bkey_generation(k);
 591         return 0;
 592 }
 593
 594 static int bch2_inode_delete_keys(struct btree_trans *trans,
 595                                   subvol_inum inum, enum btree_id id)
 596 {
 597         u64 offset = 0;
 598         int ret = 0;
 599
 600         while (!ret || ret == -EINTR) {
 601                 struct btree_iter iter;
 602                 struct bkey_s_c k;
 603                 struct bkey_i delete;
 604                 u32 snapshot;
 605
 606                 bch2_trans_begin(trans);
 607
 608                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 609                 if (ret)
 610                         continue;
 611
 612                 bch2_trans_iter_init(trans, &iter, id,
 613                                      SPOS(inum.inum, offset, snapshot),
 614                                      BTREE_ITER_INTENT);
 615                 k = bch2_btree_iter_peek(&iter);
 616
 617                 if (!k.k || iter.pos.inode != inum.inum) {
 618                         bch2_trans_iter_exit(trans, &iter);
 619                         break;
 620                 }
 621
 622                 ret = bkey_err(k);
 623                 if (ret)
 624                         goto err;
 625
 626                 bkey_init(&delete.k);
 627                 delete.k.p = iter.pos;
 628
 629                 if (btree_node_type_is_extents(iter.btree_id)) {
 630                         unsigned max_sectors =
 631                                 min_t(u64, U64_MAX - iter.pos.offset,
 632                                       KEY_SIZE_MAX & (~0 << trans->c->block_bits));
 633
 634                         /* create the biggest key we can */
 635                         bch2_key_resize(&delete.k, max_sectors);
 636
 637                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 638                         if (ret)
 639                                 goto err;
 640                 }
 641
 642                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 643                       bch2_trans_commit(trans, NULL, NULL,
 644                                         BTREE_INSERT_NOFAIL);
 645 err:
 646                 offset = iter.pos.offset;
 647                 bch2_trans_iter_exit(trans, &iter);
 648         }
 649
 650         return ret;
 651 }
 652
 653 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum, bool cached)
 654 {
 655         struct btree_trans trans;
 656         struct btree_iter iter = { NULL };
 657         struct bkey_i_inode_generation delete;
 658         struct bch_inode_unpacked inode_u;
 659         struct bkey_s_c k;
 660         unsigned iter_flags = BTREE_ITER_INTENT;
 661         u32 snapshot;
 662         int ret;
 663
 664         if (0 && cached && c->opts.inodes_use_key_cache)
 665                 iter_flags |= BTREE_ITER_CACHED;
 666
 667         bch2_trans_init(&trans, c, 0, 1024);
 668
 669         /*
 670          * If this was a directory, there shouldn't be any real dirents left -
 671          * but there could be whiteouts (from hash collisions) that we should
 672          * delete:
 673          *
 674          * XXX: the dirent could ideally would delete whiteouts when they're no
 675          * longer needed
 676          */
 677         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 678                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 679                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 680         if (ret)
 681                 goto err;
 682 retry:
 683         bch2_trans_begin(&trans);
 684
 685         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 686         if (ret)
 687                 goto err;
 688
 689         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 690                              SPOS(0, inum.inum, snapshot), iter_flags);
 691         k = bch2_btree_iter_peek_slot(&iter);
 692
 693         ret = bkey_err(k);
 694         if (ret)
 695                 goto err;
 696
 697         if (!bkey_is_inode(k.k)) {
 698                 bch2_fs_inconsistent(trans.c,
 699                                      "inode %llu not found when deleting",
 700                                      inum.inum);
 701                 ret = -EIO;
 702                 goto err;
 703         }
 704
 705         bch2_inode_unpack(k, &inode_u);
 706
 707         /* Subvolume root? */
 708         BUG_ON(inode_u.bi_subvol);
 709
 710         bkey_inode_generation_init(&delete.k_i);
 711         delete.k.p = iter.pos;
 712         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 713
 714         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 715                 bch2_trans_commit(&trans, NULL, NULL,
 716                                 BTREE_INSERT_NOFAIL);
 717 err:
 718         bch2_trans_iter_exit(&trans, &iter);
 719         if (ret == -EINTR)
 720                 goto retry;
 721
 722         bch2_trans_exit(&trans);
 723         return ret;
 724 }
 725
 726 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 727                                   subvol_inum inum,
 728                                   struct bch_inode_unpacked *inode)
 729 {
 730         struct btree_iter iter;
 731         int ret;
 732
 733         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 734         if (!ret)
 735                 bch2_trans_iter_exit(trans, &iter);
 736         return ret;
 737 }
 738
 739 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 740                             struct bch_inode_unpacked *inode)
 741 {
 742         return bch2_trans_do(c, NULL, NULL, 0,
 743                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 744 }