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 "buckets.h"
   8 #include "error.h"
   9 #include "extents.h"
  10 #include "extent_update.h"
  11 #include "inode.h"
  12 #include "str_hash.h"
  13 #include "subvolume.h"
  14 #include "varint.h"
  15
  16 #include <linux/random.h>
  17
  18 #include <asm/unaligned.h>
  19
  20 const char * const bch2_inode_opts[] = {
  21 #define x(name, ...)    #name,
  22         BCH_INODE_OPTS()
  23 #undef  x
  24         NULL,
  25 };
  26
  27 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  28
  29 static int inode_decode_field(const u8 *in, const u8 *end,
  30                               u64 out[2], unsigned *out_bits)
  31 {
  32         __be64 be[2] = { 0, 0 };
  33         unsigned bytes, shift;
  34         u8 *p;
  35
  36         if (in >= end)
  37                 return -1;
  38
  39         if (!*in)
  40                 return -1;
  41
  42         /*
  43          * position of highest set bit indicates number of bytes:
  44          * shift = number of bits to remove in high byte:
  45          */
  46         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  47         bytes   = byte_table[shift - 1];
  48
  49         if (in + bytes > end)
  50                 return -1;
  51
  52         p = (u8 *) be + 16 - bytes;
  53         memcpy(p, in, bytes);
  54         *p ^= (1 << 8) >> shift;
  55
  56         out[0] = be64_to_cpu(be[0]);
  57         out[1] = be64_to_cpu(be[1]);
  58         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  59
  60         return bytes;
  61 }
  62
  63 void bch2_inode_pack(struct bch_fs *c,
  64                      struct bkey_inode_buf *packed,
  65                      const struct bch_inode_unpacked *inode)
  66 {
  67         struct bkey_i_inode_v2 *k = &packed->inode;
  68         u8 *out = k->v.fields;
  69         u8 *end = (void *) &packed[1];
  70         u8 *last_nonzero_field = out;
  71         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  72         unsigned bytes;
  73         int ret;
  74
  75         bkey_inode_v2_init(&packed->inode.k_i);
  76         packed->inode.k.p.offset        = inode->bi_inum;
  77         packed->inode.v.bi_journal_seq  = cpu_to_le64(inode->bi_journal_seq);
  78         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  79         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  80         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  81         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
  82
  83 #define x(_name, _bits)                                                 \
  84         nr_fields++;                                                    \
  85                                                                         \
  86         if (inode->_name) {                                             \
  87                 ret = bch2_varint_encode_fast(out, inode->_name);       \
  88                 out += ret;                                             \
  89                                                                         \
  90                 if (_bits > 64)                                         \
  91                         *out++ = 0;                                     \
  92                                                                         \
  93                 last_nonzero_field = out;                               \
  94                 last_nonzero_fieldnr = nr_fields;                       \
  95         } else {                                                        \
  96                 *out++ = 0;                                             \
  97                                                                         \
  98                 if (_bits > 64)                                         \
  99                         *out++ = 0;                                     \
 100         }
 101
 102         BCH_INODE_FIELDS()
 103 #undef  x
 104         BUG_ON(out > end);
 105
 106         out = last_nonzero_field;
 107         nr_fields = last_nonzero_fieldnr;
 108
 109         bytes = out - (u8 *) &packed->inode.v;
 110         set_bkey_val_bytes(&packed->inode.k, bytes);
 111         memset_u64s_tail(&packed->inode.v, 0, bytes);
 112
 113         SET_INODEv2_NR_FIELDS(&k->v, nr_fields);
 114
 115         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 116                 struct bch_inode_unpacked unpacked;
 117
 118                 int ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i),
 119                                            &unpacked);
 120                 BUG_ON(ret);
 121                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 122                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 123                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 124
 125 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 126                         panic("unpacked %llu should be %llu",           \
 127                               (u64) unpacked._name, (u64) inode->_name);
 128                 BCH_INODE_FIELDS()
 129 #undef  x
 130         }
 131 }
 132
 133 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 134                                 struct bch_inode_unpacked *unpacked)
 135 {
 136         const u8 *in = inode.v->fields;
 137         const u8 *end = bkey_val_end(inode);
 138         u64 field[2];
 139         unsigned fieldnr = 0, field_bits;
 140         int ret;
 141
 142 #define x(_name, _bits)                                 \
 143         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 144                 memset(&unpacked->_name, 0,                             \
 145                        sizeof(*unpacked) -                              \
 146                        offsetof(struct bch_inode_unpacked, _name));     \
 147                 return 0;                                               \
 148         }                                                               \
 149                                                                         \
 150         ret = inode_decode_field(in, end, field, &field_bits);          \
 151         if (ret < 0)                                                    \
 152                 return ret;                                             \
 153                                                                         \
 154         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 155                 return -1;                                              \
 156                                                                         \
 157         unpacked->_name = field[1];                                     \
 158         in += ret;
 159
 160         BCH_INODE_FIELDS()
 161 #undef  x
 162
 163         /* XXX: signal if there were more fields than expected? */
 164         return 0;
 165 }
 166
 167 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 168                                 const u8 *in, const u8 *end,
 169                                 unsigned nr_fields)
 170 {
 171         unsigned fieldnr = 0;
 172         int ret;
 173         u64 v[2];
 174
 175 #define x(_name, _bits)                                                 \
 176         if (fieldnr < nr_fields) {                                      \
 177                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 178                 if (ret < 0)                                            \
 179                         return ret;                                     \
 180                 in += ret;                                              \
 181                                                                         \
 182                 if (_bits > 64) {                                       \
 183                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 184                         if (ret < 0)                                    \
 185                                 return ret;                             \
 186                         in += ret;                                      \
 187                 } else {                                                \
 188                         v[1] = 0;                                       \
 189                 }                                                       \
 190         } else {                                                        \
 191                 v[0] = v[1] = 0;                                        \
 192         }                                                               \
 193                                                                         \
 194         unpacked->_name = v[0];                                         \
 195         if (v[1] || v[0] != unpacked->_name)                            \
 196                 return -1;                                              \
 197         fieldnr++;
 198
 199         BCH_INODE_FIELDS()
 200 #undef  x
 201
 202         /* XXX: signal if there were more fields than expected? */
 203         return 0;
 204 }
 205
 206 int bch2_inode_unpack(struct bkey_s_c k,
 207                       struct bch_inode_unpacked *unpacked)
 208 {
 209         switch (k.k->type) {
 210         case KEY_TYPE_inode: {
 211                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 212
 213                 unpacked->bi_inum       = inode.k->p.offset;
 214                 unpacked->bi_journal_seq= 0;
 215                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 216                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 217                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 218
 219                 if (INODE_NEW_VARINT(inode.v)) {
 220                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 221                                                     bkey_val_end(inode),
 222                                                     INODE_NR_FIELDS(inode.v));
 223                 } else {
 224                         return bch2_inode_unpack_v1(inode, unpacked);
 225                 }
 226                 break;
 227         }
 228         case KEY_TYPE_inode_v2: {
 229                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 230
 231                 unpacked->bi_inum       = inode.k->p.offset;
 232                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 233                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 234                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 235                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 236
 237                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 238                                             bkey_val_end(inode),
 239                                             INODEv2_NR_FIELDS(inode.v));
 240         }
 241         default:
 242                 BUG();
 243         }
 244 }
 245
 246 int bch2_inode_peek(struct btree_trans *trans,
 247                     struct btree_iter *iter,
 248                     struct bch_inode_unpacked *inode,
 249                     subvol_inum inum, unsigned flags)
 250 {
 251         struct bkey_s_c k;
 252         u32 snapshot;
 253         int ret;
 254
 255         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 256         if (ret)
 257                 return ret;
 258
 259         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 260                              SPOS(0, inum.inum, snapshot),
 261                              flags|BTREE_ITER_CACHED);
 262         k = bch2_btree_iter_peek_slot(iter);
 263         ret = bkey_err(k);
 264         if (ret)
 265                 goto err;
 266
 267         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 268         if (ret)
 269                 goto err;
 270
 271         ret = bch2_inode_unpack(k, inode);
 272         if (ret)
 273                 goto err;
 274
 275         return 0;
 276 err:
 277         bch2_trans_iter_exit(trans, iter);
 278         return ret;
 279 }
 280
 281 int bch2_inode_write(struct btree_trans *trans,
 282                      struct btree_iter *iter,
 283                      struct bch_inode_unpacked *inode)
 284 {
 285         struct bkey_inode_buf *inode_p;
 286
 287         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 288         if (IS_ERR(inode_p))
 289                 return PTR_ERR(inode_p);
 290
 291         bch2_inode_pack(trans->c, inode_p, inode);
 292         inode_p->inode.k.p.snapshot = iter->snapshot;
 293         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 294 }
 295
 296 static int __bch2_inode_invalid(struct bkey_s_c k, struct printbuf *err)
 297 {
 298         struct bch_inode_unpacked unpacked;
 299
 300         if (k.k->p.inode) {
 301                 pr_buf(err, "nonzero k.p.inode");
 302                 return -EINVAL;
 303         }
 304
 305         if (k.k->p.offset < BLOCKDEV_INODE_MAX) {
 306                 pr_buf(err, "fs inode in blockdev range");
 307                 return -EINVAL;
 308         }
 309
 310         if (bch2_inode_unpack(k, &unpacked)){
 311                 pr_buf(err, "invalid variable length fields");
 312                 return -EINVAL;
 313         }
 314
 315         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1) {
 316                 pr_buf(err, "invalid data checksum type (%u >= %u",
 317                         unpacked.bi_data_checksum, BCH_CSUM_OPT_NR + 1);
 318                 return -EINVAL;
 319         }
 320
 321         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1) {
 322                 pr_buf(err, "invalid data checksum type (%u >= %u)",
 323                        unpacked.bi_compression, BCH_COMPRESSION_OPT_NR + 1);
 324                 return -EINVAL;
 325         }
 326
 327         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 328             unpacked.bi_nlink != 0) {
 329                 pr_buf(err, "flagged as unlinked but bi_nlink != 0");
 330                 return -EINVAL;
 331         }
 332
 333         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode)) {
 334                 pr_buf(err, "subvolume root but not a directory");
 335                 return -EINVAL;
 336         }
 337
 338         return 0;
 339 }
 340
 341 int bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k,
 342                        int rw, struct printbuf *err)
 343 {
 344         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 345
 346         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 347                 pr_buf(err, "incorrect value size (%zu < %zu)",
 348                        bkey_val_bytes(k.k), sizeof(*inode.v));
 349                 return -EINVAL;
 350         }
 351
 352         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 353                 pr_buf(err, "invalid str hash type (%llu >= %u)",
 354                        INODE_STR_HASH(inode.v), BCH_STR_HASH_NR);
 355                 return -EINVAL;
 356         }
 357
 358         return __bch2_inode_invalid(k, err);
 359 }
 360
 361 int bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 362                           int rw, struct printbuf *err)
 363 {
 364         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 365
 366         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 367                 pr_buf(err, "incorrect value size (%zu < %zu)",
 368                        bkey_val_bytes(k.k), sizeof(*inode.v));
 369                 return -EINVAL;
 370         }
 371
 372         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 373                 pr_buf(err, "invalid str hash type (%llu >= %u)",
 374                        INODEv2_STR_HASH(inode.v), BCH_STR_HASH_NR);
 375                 return -EINVAL;
 376         }
 377
 378         return __bch2_inode_invalid(k, err);
 379 }
 380
 381 static void __bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 382 {
 383         pr_buf(out, "mode %o flags %x journal_seq %llu",
 384                inode->bi_mode, inode->bi_flags,
 385                inode->bi_journal_seq);
 386
 387 #define x(_name, _bits)                                         \
 388         pr_buf(out, " "#_name " %llu", (u64) inode->_name);
 389         BCH_INODE_FIELDS()
 390 #undef  x
 391 }
 392
 393 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 394 {
 395         pr_buf(out, "inum: %llu ", inode->bi_inum);
 396         __bch2_inode_unpacked_to_text(out, inode);
 397 }
 398
 399 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 400                        struct bkey_s_c k)
 401 {
 402         struct bch_inode_unpacked inode;
 403
 404         if (bch2_inode_unpack(k, &inode)) {
 405                 pr_buf(out, "(unpack error)");
 406                 return;
 407         }
 408
 409         __bch2_inode_unpacked_to_text(out, &inode);
 410 }
 411
 412 int bch2_inode_generation_invalid(const struct bch_fs *c, struct bkey_s_c k,
 413                                   int rw, struct printbuf *err)
 414 {
 415         if (k.k->p.inode) {
 416                 pr_buf(err, "nonzero k.p.inode");
 417                 return -EINVAL;
 418         }
 419
 420         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation)) {
 421                 pr_buf(err, "incorrect value size (%zu != %zu)",
 422                        bkey_val_bytes(k.k), sizeof(struct bch_inode_generation));
 423                 return -EINVAL;
 424         }
 425
 426         return 0;
 427 }
 428
 429 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 430                                    struct bkey_s_c k)
 431 {
 432         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 433
 434         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 435 }
 436
 437 void bch2_inode_init_early(struct bch_fs *c,
 438                            struct bch_inode_unpacked *inode_u)
 439 {
 440         enum bch_str_hash_type str_hash =
 441                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 442
 443         memset(inode_u, 0, sizeof(*inode_u));
 444
 445         /* ick */
 446         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 447         get_random_bytes(&inode_u->bi_hash_seed,
 448                          sizeof(inode_u->bi_hash_seed));
 449 }
 450
 451 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 452                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 453                           struct bch_inode_unpacked *parent)
 454 {
 455         inode_u->bi_mode        = mode;
 456         inode_u->bi_uid         = uid;
 457         inode_u->bi_gid         = gid;
 458         inode_u->bi_dev         = rdev;
 459         inode_u->bi_atime       = now;
 460         inode_u->bi_mtime       = now;
 461         inode_u->bi_ctime       = now;
 462         inode_u->bi_otime       = now;
 463
 464         if (parent && parent->bi_mode & S_ISGID) {
 465                 inode_u->bi_gid = parent->bi_gid;
 466                 if (S_ISDIR(mode))
 467                         inode_u->bi_mode |= S_ISGID;
 468         }
 469
 470         if (parent) {
 471 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 472                 BCH_INODE_OPTS()
 473 #undef x
 474         }
 475 }
 476
 477 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 478                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 479                      struct bch_inode_unpacked *parent)
 480 {
 481         bch2_inode_init_early(c, inode_u);
 482         bch2_inode_init_late(inode_u, bch2_current_time(c),
 483                              uid, gid, mode, rdev, parent);
 484 }
 485
 486 static inline u32 bkey_generation(struct bkey_s_c k)
 487 {
 488         switch (k.k->type) {
 489         case KEY_TYPE_inode:
 490         case KEY_TYPE_inode_v2:
 491                 BUG();
 492         case KEY_TYPE_inode_generation:
 493                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 494         default:
 495                 return 0;
 496         }
 497 }
 498
 499 /*
 500  * This just finds an empty slot:
 501  */
 502 int bch2_inode_create(struct btree_trans *trans,
 503                       struct btree_iter *iter,
 504                       struct bch_inode_unpacked *inode_u,
 505                       u32 snapshot, u64 cpu)
 506 {
 507         struct bch_fs *c = trans->c;
 508         struct bkey_s_c k;
 509         u64 min, max, start, pos, *hint;
 510         int ret = 0;
 511         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 512
 513         if (c->opts.shard_inode_numbers) {
 514                 bits -= c->inode_shard_bits;
 515
 516                 min = (cpu << bits);
 517                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 518
 519                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 520                 hint = c->unused_inode_hints + cpu;
 521         } else {
 522                 min = BLOCKDEV_INODE_MAX;
 523                 max = ~(ULLONG_MAX << bits);
 524                 hint = c->unused_inode_hints;
 525         }
 526
 527         start = READ_ONCE(*hint);
 528
 529         if (start >= max || start < min)
 530                 start = min;
 531
 532         pos = start;
 533         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 534                              BTREE_ITER_ALL_SNAPSHOTS|
 535                              BTREE_ITER_INTENT);
 536 again:
 537         while ((k = bch2_btree_iter_peek(iter)).k &&
 538                !(ret = bkey_err(k)) &&
 539                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 540                 while (pos < iter->pos.offset) {
 541                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 542                                 goto found_slot;
 543
 544                         pos++;
 545                 }
 546
 547                 if (k.k->p.snapshot == snapshot &&
 548                     !bkey_is_inode(k.k) &&
 549                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 550                         bch2_btree_iter_advance(iter);
 551                         continue;
 552                 }
 553
 554                 /*
 555                  * We don't need to iterate over keys in every snapshot once
 556                  * we've found just one:
 557                  */
 558                 pos = iter->pos.offset + 1;
 559                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 560         }
 561
 562         while (!ret && pos < max) {
 563                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 564                         goto found_slot;
 565
 566                 pos++;
 567         }
 568
 569         if (!ret && start == min)
 570                 ret = -ENOSPC;
 571
 572         if (ret) {
 573                 bch2_trans_iter_exit(trans, iter);
 574                 return ret;
 575         }
 576
 577         /* Retry from start */
 578         pos = start = min;
 579         bch2_btree_iter_set_pos(iter, POS(0, pos));
 580         goto again;
 581 found_slot:
 582         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 583         k = bch2_btree_iter_peek_slot(iter);
 584         ret = bkey_err(k);
 585         if (ret) {
 586                 bch2_trans_iter_exit(trans, iter);
 587                 return ret;
 588         }
 589
 590         /* We may have raced while the iterator wasn't pointing at pos: */
 591         if (bkey_is_inode(k.k) ||
 592             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 593                 goto again;
 594
 595         *hint                   = k.k->p.offset;
 596         inode_u->bi_inum        = k.k->p.offset;
 597         inode_u->bi_generation  = bkey_generation(k);
 598         return 0;
 599 }
 600
 601 static int bch2_inode_delete_keys(struct btree_trans *trans,
 602                                   subvol_inum inum, enum btree_id id)
 603 {
 604         struct btree_iter iter;
 605         struct bkey_s_c k;
 606         struct bkey_i delete;
 607         u32 snapshot;
 608         int ret = 0;
 609
 610         /*
 611          * We're never going to be deleting extents, no need to use an extent
 612          * iterator:
 613          */
 614         bch2_trans_iter_init(trans, &iter, id, POS(inum.inum, 0),
 615                              BTREE_ITER_NOT_EXTENTS|
 616                              BTREE_ITER_INTENT);
 617
 618         while (1) {
 619                 bch2_trans_begin(trans);
 620
 621                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 622                 if (ret)
 623                         goto err;
 624
 625                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 626
 627                 k = bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
 628                 ret = bkey_err(k);
 629                 if (ret)
 630                         goto err;
 631
 632                 if (!k.k)
 633                         break;
 634
 635                 bkey_init(&delete.k);
 636                 delete.k.p = iter.pos;
 637
 638                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 639                       bch2_trans_commit(trans, NULL, NULL,
 640                                         BTREE_INSERT_NOFAIL);
 641 err:
 642                 if (ret && ret != -EINTR)
 643                         break;
 644         }
 645
 646         bch2_trans_iter_exit(trans, &iter);
 647         return ret;
 648 }
 649
 650 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 651 {
 652         struct btree_trans trans;
 653         struct btree_iter iter = { NULL };
 654         struct bkey_i_inode_generation delete;
 655         struct bch_inode_unpacked inode_u;
 656         struct bkey_s_c k;
 657         u32 snapshot;
 658         int ret;
 659
 660         bch2_trans_init(&trans, c, 0, 1024);
 661
 662         /*
 663          * If this was a directory, there shouldn't be any real dirents left -
 664          * but there could be whiteouts (from hash collisions) that we should
 665          * delete:
 666          *
 667          * XXX: the dirent could ideally would delete whiteouts when they're no
 668          * longer needed
 669          */
 670         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 671                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 672                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 673         if (ret)
 674                 goto err;
 675 retry:
 676         bch2_trans_begin(&trans);
 677
 678         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 679         if (ret)
 680                 goto err;
 681
 682         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 683                              SPOS(0, inum.inum, snapshot),
 684                              BTREE_ITER_INTENT|BTREE_ITER_CACHED);
 685         k = bch2_btree_iter_peek_slot(&iter);
 686
 687         ret = bkey_err(k);
 688         if (ret)
 689                 goto err;
 690
 691         if (!bkey_is_inode(k.k)) {
 692                 bch2_fs_inconsistent(trans.c,
 693                                      "inode %llu not found when deleting",
 694                                      inum.inum);
 695                 ret = -EIO;
 696                 goto err;
 697         }
 698
 699         bch2_inode_unpack(k, &inode_u);
 700
 701         /* Subvolume root? */
 702         BUG_ON(inode_u.bi_subvol);
 703
 704         bkey_inode_generation_init(&delete.k_i);
 705         delete.k.p = iter.pos;
 706         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 707
 708         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 709                 bch2_trans_commit(&trans, NULL, NULL,
 710                                 BTREE_INSERT_NOFAIL);
 711 err:
 712         bch2_trans_iter_exit(&trans, &iter);
 713         if (ret == -EINTR)
 714                 goto retry;
 715
 716         bch2_trans_exit(&trans);
 717         return ret;
 718 }
 719
 720 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 721                                   subvol_inum inum,
 722                                   struct bch_inode_unpacked *inode)
 723 {
 724         struct btree_iter iter;
 725         int ret;
 726
 727         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 728         if (!ret)
 729                 bch2_trans_iter_exit(trans, &iter);
 730         return ret;
 731 }
 732
 733 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 734                             struct bch_inode_unpacked *inode)
 735 {
 736         return bch2_trans_do(c, NULL, NULL, 0,
 737                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 738 }