fs/bcachefs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "btree_write_buffer.h"
   6 #include "bkey_methods.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "compress.h"
  10 #include "dirent.h"
  11 #include "disk_accounting.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "extent_update.h"
  15 #include "fs.h"
  16 #include "inode.h"
  17 #include "str_hash.h"
  18 #include "snapshot.h"
  19 #include "subvolume.h"
  20 #include "varint.h"
  21
  22 #include <linux/random.h>
  23
  24 #include <linux/unaligned.h>
  25
  26 #define x(name, ...)    #name,
  27 const char * const bch2_inode_opts[] = {
  28         BCH_INODE_OPTS()
  29         NULL,
  30 };
  31
  32 static const char * const bch2_inode_flag_strs[] = {
  33         BCH_INODE_FLAGS()
  34         NULL
  35 };
  36 #undef  x
  37
  38 static int delete_ancestor_snapshot_inodes(struct btree_trans *, struct bpos);
  39
  40 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  41
  42 static int inode_decode_field(const u8 *in, const u8 *end,
  43                               u64 out[2], unsigned *out_bits)
  44 {
  45         __be64 be[2] = { 0, 0 };
  46         unsigned bytes, shift;
  47         u8 *p;
  48
  49         if (in >= end)
  50                 return -1;
  51
  52         if (!*in)
  53                 return -1;
  54
  55         /*
  56          * position of highest set bit indicates number of bytes:
  57          * shift = number of bits to remove in high byte:
  58          */
  59         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  60         bytes   = byte_table[shift - 1];
  61
  62         if (in + bytes > end)
  63                 return -1;
  64
  65         p = (u8 *) be + 16 - bytes;
  66         memcpy(p, in, bytes);
  67         *p ^= (1 << 8) >> shift;
  68
  69         out[0] = be64_to_cpu(be[0]);
  70         out[1] = be64_to_cpu(be[1]);
  71         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  72
  73         return bytes;
  74 }
  75
  76 static inline void bch2_inode_pack_inlined(struct bkey_inode_buf *packed,
  77                                            const struct bch_inode_unpacked *inode)
  78 {
  79         struct bkey_i_inode_v3 *k = &packed->inode;
  80         u8 *out = k->v.fields;
  81         u8 *end = (void *) &packed[1];
  82         u8 *last_nonzero_field = out;
  83         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  84         unsigned bytes;
  85         int ret;
  86
  87         bkey_inode_v3_init(&packed->inode.k_i);
  88         packed->inode.k.p.offset        = inode->bi_inum;
  89         packed->inode.v.bi_journal_seq  = cpu_to_le64(inode->bi_journal_seq);
  90         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  91         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  92         packed->inode.v.bi_sectors      = cpu_to_le64(inode->bi_sectors);
  93         packed->inode.v.bi_size         = cpu_to_le64(inode->bi_size);
  94         packed->inode.v.bi_version      = cpu_to_le64(inode->bi_version);
  95         SET_INODEv3_MODE(&packed->inode.v, inode->bi_mode);
  96         SET_INODEv3_FIELDS_START(&packed->inode.v, INODEv3_FIELDS_START_CUR);
  97
  98
  99 #define x(_name, _bits)                                                 \
 100         nr_fields++;                                                    \
 101                                                                         \
 102         if (inode->_name) {                                             \
 103                 ret = bch2_varint_encode_fast(out, inode->_name);       \
 104                 out += ret;                                             \
 105                                                                         \
 106                 if (_bits > 64)                                         \
 107                         *out++ = 0;                                     \
 108                                                                         \
 109                 last_nonzero_field = out;                               \
 110                 last_nonzero_fieldnr = nr_fields;                       \
 111         } else {                                                        \
 112                 *out++ = 0;                                             \
 113                                                                         \
 114                 if (_bits > 64)                                         \
 115                         *out++ = 0;                                     \
 116         }
 117
 118         BCH_INODE_FIELDS_v3()
 119 #undef  x
 120         BUG_ON(out > end);
 121
 122         out = last_nonzero_field;
 123         nr_fields = last_nonzero_fieldnr;
 124
 125         bytes = out - (u8 *) &packed->inode.v;
 126         set_bkey_val_bytes(&packed->inode.k, bytes);
 127         memset_u64s_tail(&packed->inode.v, 0, bytes);
 128
 129         SET_INODEv3_NR_FIELDS(&k->v, nr_fields);
 130
 131         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 132                 struct bch_inode_unpacked unpacked;
 133
 134                 ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i), &unpacked);
 135                 BUG_ON(ret);
 136                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 137                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 138                 BUG_ON(unpacked.bi_sectors      != inode->bi_sectors);
 139                 BUG_ON(unpacked.bi_size         != inode->bi_size);
 140                 BUG_ON(unpacked.bi_version      != inode->bi_version);
 141                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 142
 143 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 144                         panic("unpacked %llu should be %llu",           \
 145                               (u64) unpacked._name, (u64) inode->_name);
 146                 BCH_INODE_FIELDS_v3()
 147 #undef  x
 148         }
 149 }
 150
 151 void bch2_inode_pack(struct bkey_inode_buf *packed,
 152                      const struct bch_inode_unpacked *inode)
 153 {
 154         bch2_inode_pack_inlined(packed, inode);
 155 }
 156
 157 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 158                                 struct bch_inode_unpacked *unpacked)
 159 {
 160         const u8 *in = inode.v->fields;
 161         const u8 *end = bkey_val_end(inode);
 162         u64 field[2];
 163         unsigned fieldnr = 0, field_bits;
 164         int ret;
 165
 166 #define x(_name, _bits)                                                 \
 167         if (fieldnr++ == INODEv1_NR_FIELDS(inode.v)) {                  \
 168                 unsigned offset = offsetof(struct bch_inode_unpacked, _name);\
 169                 memset((void *) unpacked + offset, 0,                   \
 170                        sizeof(*unpacked) - offset);                     \
 171                 return 0;                                               \
 172         }                                                               \
 173                                                                         \
 174         ret = inode_decode_field(in, end, field, &field_bits);          \
 175         if (ret < 0)                                                    \
 176                 return ret;                                             \
 177                                                                         \
 178         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 179                 return -1;                                              \
 180                                                                         \
 181         unpacked->_name = field[1];                                     \
 182         in += ret;
 183
 184         BCH_INODE_FIELDS_v2()
 185 #undef  x
 186
 187         /* XXX: signal if there were more fields than expected? */
 188         return 0;
 189 }
 190
 191 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 192                                 const u8 *in, const u8 *end,
 193                                 unsigned nr_fields)
 194 {
 195         unsigned fieldnr = 0;
 196         int ret;
 197         u64 v[2];
 198
 199 #define x(_name, _bits)                                                 \
 200         if (fieldnr < nr_fields) {                                      \
 201                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 202                 if (ret < 0)                                            \
 203                         return ret;                                     \
 204                 in += ret;                                              \
 205                                                                         \
 206                 if (_bits > 64) {                                       \
 207                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 208                         if (ret < 0)                                    \
 209                                 return ret;                             \
 210                         in += ret;                                      \
 211                 } else {                                                \
 212                         v[1] = 0;                                       \
 213                 }                                                       \
 214         } else {                                                        \
 215                 v[0] = v[1] = 0;                                        \
 216         }                                                               \
 217                                                                         \
 218         unpacked->_name = v[0];                                         \
 219         if (v[1] || v[0] != unpacked->_name)                            \
 220                 return -1;                                              \
 221         fieldnr++;
 222
 223         BCH_INODE_FIELDS_v2()
 224 #undef  x
 225
 226         /* XXX: signal if there were more fields than expected? */
 227         return 0;
 228 }
 229
 230 static int bch2_inode_unpack_v3(struct bkey_s_c k,
 231                                 struct bch_inode_unpacked *unpacked)
 232 {
 233         struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 234         const u8 *in = inode.v->fields;
 235         const u8 *end = bkey_val_end(inode);
 236         unsigned nr_fields = INODEv3_NR_FIELDS(inode.v);
 237         unsigned fieldnr = 0;
 238         int ret;
 239         u64 v[2];
 240
 241         unpacked->bi_inum       = inode.k->p.offset;
 242         unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 243         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 244         unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 245         unpacked->bi_sectors    = le64_to_cpu(inode.v->bi_sectors);
 246         unpacked->bi_size       = le64_to_cpu(inode.v->bi_size);
 247         unpacked->bi_version    = le64_to_cpu(inode.v->bi_version);
 248         unpacked->bi_mode       = INODEv3_MODE(inode.v);
 249
 250 #define x(_name, _bits)                                                 \
 251         if (fieldnr < nr_fields) {                                      \
 252                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 253                 if (ret < 0)                                            \
 254                         return ret;                                     \
 255                 in += ret;                                              \
 256                                                                         \
 257                 if (_bits > 64) {                                       \
 258                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 259                         if (ret < 0)                                    \
 260                                 return ret;                             \
 261                         in += ret;                                      \
 262                 } else {                                                \
 263                         v[1] = 0;                                       \
 264                 }                                                       \
 265         } else {                                                        \
 266                 v[0] = v[1] = 0;                                        \
 267         }                                                               \
 268                                                                         \
 269         unpacked->_name = v[0];                                         \
 270         if (v[1] || v[0] != unpacked->_name)                            \
 271                 return -1;                                              \
 272         fieldnr++;
 273
 274         BCH_INODE_FIELDS_v3()
 275 #undef  x
 276
 277         /* XXX: signal if there were more fields than expected? */
 278         return 0;
 279 }
 280
 281 static noinline int bch2_inode_unpack_slowpath(struct bkey_s_c k,
 282                                                struct bch_inode_unpacked *unpacked)
 283 {
 284         memset(unpacked, 0, sizeof(*unpacked));
 285
 286         unpacked->bi_snapshot = k.k->p.snapshot;
 287
 288         switch (k.k->type) {
 289         case KEY_TYPE_inode: {
 290                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 291
 292                 unpacked->bi_inum       = inode.k->p.offset;
 293                 unpacked->bi_journal_seq= 0;
 294                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 295                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 296                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 297
 298                 if (INODEv1_NEW_VARINT(inode.v)) {
 299                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 300                                                     bkey_val_end(inode),
 301                                                     INODEv1_NR_FIELDS(inode.v));
 302                 } else {
 303                         return bch2_inode_unpack_v1(inode, unpacked);
 304                 }
 305                 break;
 306         }
 307         case KEY_TYPE_inode_v2: {
 308                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 309
 310                 unpacked->bi_inum       = inode.k->p.offset;
 311                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 312                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 313                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 314                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 315
 316                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 317                                             bkey_val_end(inode),
 318                                             INODEv2_NR_FIELDS(inode.v));
 319         }
 320         default:
 321                 BUG();
 322         }
 323 }
 324
 325 int bch2_inode_unpack(struct bkey_s_c k,
 326                       struct bch_inode_unpacked *unpacked)
 327 {
 328         unpacked->bi_snapshot = k.k->p.snapshot;
 329
 330         return likely(k.k->type == KEY_TYPE_inode_v3)
 331                 ? bch2_inode_unpack_v3(k, unpacked)
 332                 : bch2_inode_unpack_slowpath(k, unpacked);
 333 }
 334
 335 int __bch2_inode_peek(struct btree_trans *trans,
 336                       struct btree_iter *iter,
 337                       struct bch_inode_unpacked *inode,
 338                       subvol_inum inum, unsigned flags,
 339                       bool warn)
 340 {
 341         u32 snapshot;
 342         int ret = __bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot, warn);
 343         if (ret)
 344                 return ret;
 345
 346         struct bkey_s_c k = bch2_bkey_get_iter(trans, iter, BTREE_ID_inodes,
 347                                                SPOS(0, inum.inum, snapshot),
 348                                                flags|BTREE_ITER_cached);
 349         ret = bkey_err(k);
 350         if (ret)
 351                 return ret;
 352
 353         ret = bkey_is_inode(k.k) ? 0 : -BCH_ERR_ENOENT_inode;
 354         if (ret)
 355                 goto err;
 356
 357         ret = bch2_inode_unpack(k, inode);
 358         if (ret)
 359                 goto err;
 360
 361         return 0;
 362 err:
 363         if (warn)
 364                 bch_err_msg(trans->c, ret, "looking up inum %llu:%llu:", inum.subvol, inum.inum);
 365         bch2_trans_iter_exit(trans, iter);
 366         return ret;
 367 }
 368
 369 int bch2_inode_write_flags(struct btree_trans *trans,
 370                      struct btree_iter *iter,
 371                      struct bch_inode_unpacked *inode,
 372                      enum btree_iter_update_trigger_flags flags)
 373 {
 374         struct bkey_inode_buf *inode_p;
 375
 376         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 377         if (IS_ERR(inode_p))
 378                 return PTR_ERR(inode_p);
 379
 380         bch2_inode_pack_inlined(inode_p, inode);
 381         inode_p->inode.k.p.snapshot = iter->snapshot;
 382         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, flags);
 383 }
 384
 385 int __bch2_fsck_write_inode(struct btree_trans *trans, struct bch_inode_unpacked *inode)
 386 {
 387         struct bkey_inode_buf *inode_p =
 388                 bch2_trans_kmalloc(trans, sizeof(*inode_p));
 389
 390         if (IS_ERR(inode_p))
 391                 return PTR_ERR(inode_p);
 392
 393         bch2_inode_pack(inode_p, inode);
 394         inode_p->inode.k.p.snapshot = inode->bi_snapshot;
 395
 396         return bch2_btree_insert_nonextent(trans, BTREE_ID_inodes,
 397                                 &inode_p->inode.k_i,
 398                                 BTREE_UPDATE_internal_snapshot_node);
 399 }
 400
 401 int bch2_fsck_write_inode(struct btree_trans *trans, struct bch_inode_unpacked *inode)
 402 {
 403         int ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 404                             __bch2_fsck_write_inode(trans, inode));
 405         bch_err_fn(trans->c, ret);
 406         return ret;
 407 }
 408
 409 struct bkey_i *bch2_inode_to_v3(struct btree_trans *trans, struct bkey_i *k)
 410 {
 411         struct bch_inode_unpacked u;
 412         struct bkey_inode_buf *inode_p;
 413         int ret;
 414
 415         if (!bkey_is_inode(&k->k))
 416                 return ERR_PTR(-ENOENT);
 417
 418         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 419         if (IS_ERR(inode_p))
 420                 return ERR_CAST(inode_p);
 421
 422         ret = bch2_inode_unpack(bkey_i_to_s_c(k), &u);
 423         if (ret)
 424                 return ERR_PTR(ret);
 425
 426         bch2_inode_pack(inode_p, &u);
 427         return &inode_p->inode.k_i;
 428 }
 429
 430 static int __bch2_inode_validate(struct bch_fs *c, struct bkey_s_c k,
 431                                  enum bch_validate_flags flags)
 432 {
 433         struct bch_inode_unpacked unpacked;
 434         int ret = 0;
 435
 436         bkey_fsck_err_on(k.k->p.inode,
 437                          c, inode_pos_inode_nonzero,
 438                          "nonzero k.p.inode");
 439
 440         bkey_fsck_err_on(k.k->p.offset < BLOCKDEV_INODE_MAX,
 441                          c, inode_pos_blockdev_range,
 442                          "fs inode in blockdev range");
 443
 444         bkey_fsck_err_on(bch2_inode_unpack(k, &unpacked),
 445                          c, inode_unpack_error,
 446                          "invalid variable length fields");
 447
 448         bkey_fsck_err_on(unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1,
 449                          c, inode_checksum_type_invalid,
 450                          "invalid data checksum type (%u >= %u",
 451                          unpacked.bi_data_checksum, BCH_CSUM_OPT_NR + 1);
 452
 453         bkey_fsck_err_on(unpacked.bi_compression &&
 454                          !bch2_compression_opt_valid(unpacked.bi_compression - 1),
 455                          c, inode_compression_type_invalid,
 456                          "invalid compression opt %u", unpacked.bi_compression - 1);
 457
 458         bkey_fsck_err_on((unpacked.bi_flags & BCH_INODE_unlinked) &&
 459                          unpacked.bi_nlink != 0,
 460                          c, inode_unlinked_but_nlink_nonzero,
 461                          "flagged as unlinked but bi_nlink != 0");
 462
 463         bkey_fsck_err_on(unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode),
 464                          c, inode_subvol_root_but_not_dir,
 465                          "subvolume root but not a directory");
 466 fsck_err:
 467         return ret;
 468 }
 469
 470 int bch2_inode_validate(struct bch_fs *c, struct bkey_s_c k,
 471                         enum bch_validate_flags flags)
 472 {
 473         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 474         int ret = 0;
 475
 476         bkey_fsck_err_on(INODEv1_STR_HASH(inode.v) >= BCH_STR_HASH_NR,
 477                          c, inode_str_hash_invalid,
 478                          "invalid str hash type (%llu >= %u)",
 479                          INODEv1_STR_HASH(inode.v), BCH_STR_HASH_NR);
 480
 481         ret = __bch2_inode_validate(c, k, flags);
 482 fsck_err:
 483         return ret;
 484 }
 485
 486 int bch2_inode_v2_validate(struct bch_fs *c, struct bkey_s_c k,
 487                            enum bch_validate_flags flags)
 488 {
 489         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 490         int ret = 0;
 491
 492         bkey_fsck_err_on(INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR,
 493                          c, inode_str_hash_invalid,
 494                          "invalid str hash type (%llu >= %u)",
 495                          INODEv2_STR_HASH(inode.v), BCH_STR_HASH_NR);
 496
 497         ret = __bch2_inode_validate(c, k, flags);
 498 fsck_err:
 499         return ret;
 500 }
 501
 502 int bch2_inode_v3_validate(struct bch_fs *c, struct bkey_s_c k,
 503                            enum bch_validate_flags flags)
 504 {
 505         struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 506         int ret = 0;
 507
 508         bkey_fsck_err_on(INODEv3_FIELDS_START(inode.v) < INODEv3_FIELDS_START_INITIAL ||
 509                          INODEv3_FIELDS_START(inode.v) > bkey_val_u64s(inode.k),
 510                          c, inode_v3_fields_start_bad,
 511                          "invalid fields_start (got %llu, min %u max %zu)",
 512                          INODEv3_FIELDS_START(inode.v),
 513                          INODEv3_FIELDS_START_INITIAL,
 514                          bkey_val_u64s(inode.k));
 515
 516         bkey_fsck_err_on(INODEv3_STR_HASH(inode.v) >= BCH_STR_HASH_NR,
 517                          c, inode_str_hash_invalid,
 518                          "invalid str hash type (%llu >= %u)",
 519                          INODEv3_STR_HASH(inode.v), BCH_STR_HASH_NR);
 520
 521         ret = __bch2_inode_validate(c, k, flags);
 522 fsck_err:
 523         return ret;
 524 }
 525
 526 static void __bch2_inode_unpacked_to_text(struct printbuf *out,
 527                                           struct bch_inode_unpacked *inode)
 528 {
 529         prt_printf(out, "\n");
 530         printbuf_indent_add(out, 2);
 531         prt_printf(out, "mode=%o\n", inode->bi_mode);
 532
 533         prt_str(out, "flags=");
 534         prt_bitflags(out, bch2_inode_flag_strs, inode->bi_flags & ((1U << 20) - 1));
 535         prt_printf(out, "(%x)\n", inode->bi_flags);
 536
 537         prt_printf(out, "journal_seq=%llu\n",   inode->bi_journal_seq);
 538         prt_printf(out, "hash_seed=%llx\n",     inode->bi_hash_seed);
 539         prt_printf(out, "hash_type=");
 540         bch2_prt_str_hash_type(out, INODE_STR_HASH(inode));
 541         prt_newline(out);
 542         prt_printf(out, "bi_size=%llu\n",       inode->bi_size);
 543         prt_printf(out, "bi_sectors=%llu\n",    inode->bi_sectors);
 544         prt_printf(out, "bi_version=%llu\n",    inode->bi_version);
 545
 546 #define x(_name, _bits)                                         \
 547         prt_printf(out, #_name "=%llu\n", (u64) inode->_name);
 548         BCH_INODE_FIELDS_v3()
 549 #undef  x
 550
 551         bch2_printbuf_strip_trailing_newline(out);
 552         printbuf_indent_sub(out, 2);
 553 }
 554
 555 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 556 {
 557         prt_printf(out, "inum: %llu:%u ", inode->bi_inum, inode->bi_snapshot);
 558         __bch2_inode_unpacked_to_text(out, inode);
 559 }
 560
 561 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 562 {
 563         struct bch_inode_unpacked inode;
 564
 565         if (bch2_inode_unpack(k, &inode)) {
 566                 prt_printf(out, "(unpack error)");
 567                 return;
 568         }
 569
 570         __bch2_inode_unpacked_to_text(out, &inode);
 571 }
 572
 573 static inline u64 bkey_inode_flags(struct bkey_s_c k)
 574 {
 575         switch (k.k->type) {
 576         case KEY_TYPE_inode:
 577                 return le32_to_cpu(bkey_s_c_to_inode(k).v->bi_flags);
 578         case KEY_TYPE_inode_v2:
 579                 return le64_to_cpu(bkey_s_c_to_inode_v2(k).v->bi_flags);
 580         case KEY_TYPE_inode_v3:
 581                 return le64_to_cpu(bkey_s_c_to_inode_v3(k).v->bi_flags);
 582         default:
 583                 return 0;
 584         }
 585 }
 586
 587 static inline void bkey_inode_flags_set(struct bkey_s k, u64 f)
 588 {
 589         switch (k.k->type) {
 590         case KEY_TYPE_inode:
 591                 bkey_s_to_inode(k).v->bi_flags = cpu_to_le32(f);
 592                 return;
 593         case KEY_TYPE_inode_v2:
 594                 bkey_s_to_inode_v2(k).v->bi_flags = cpu_to_le64(f);
 595                 return;
 596         case KEY_TYPE_inode_v3:
 597                 bkey_s_to_inode_v3(k).v->bi_flags = cpu_to_le64(f);
 598                 return;
 599         default:
 600                 BUG();
 601         }
 602 }
 603
 604 static inline bool bkey_is_unlinked_inode(struct bkey_s_c k)
 605 {
 606         unsigned f = bkey_inode_flags(k) & BCH_INODE_unlinked;
 607
 608         return (f & BCH_INODE_unlinked) && !(f & BCH_INODE_has_child_snapshot);
 609 }
 610
 611 static struct bkey_s_c
 612 bch2_bkey_get_iter_snapshot_parent(struct btree_trans *trans, struct btree_iter *iter,
 613                                    enum btree_id btree, struct bpos pos,
 614                                    unsigned flags)
 615 {
 616         struct bch_fs *c = trans->c;
 617         struct bkey_s_c k;
 618         int ret = 0;
 619
 620         for_each_btree_key_upto_norestart(trans, *iter, btree,
 621                                           bpos_successor(pos),
 622                                           SPOS(pos.inode, pos.offset, U32_MAX),
 623                                           flags|BTREE_ITER_all_snapshots, k, ret)
 624                 if (bch2_snapshot_is_ancestor(c, pos.snapshot, k.k->p.snapshot))
 625                         return k;
 626
 627         bch2_trans_iter_exit(trans, iter);
 628         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 629 }
 630
 631 static struct bkey_s_c
 632 bch2_inode_get_iter_snapshot_parent(struct btree_trans *trans, struct btree_iter *iter,
 633                                     struct bpos pos, unsigned flags)
 634 {
 635         struct bkey_s_c k;
 636 again:
 637         k = bch2_bkey_get_iter_snapshot_parent(trans, iter, BTREE_ID_inodes, pos, flags);
 638         if (!k.k ||
 639             bkey_err(k) ||
 640             bkey_is_inode(k.k))
 641                 return k;
 642
 643         bch2_trans_iter_exit(trans, iter);
 644         pos = k.k->p;
 645         goto again;
 646 }
 647
 648 int __bch2_inode_has_child_snapshots(struct btree_trans *trans, struct bpos pos)
 649 {
 650         struct bch_fs *c = trans->c;
 651         struct btree_iter iter;
 652         struct bkey_s_c k;
 653         int ret = 0;
 654
 655         for_each_btree_key_upto_norestart(trans, iter,
 656                         BTREE_ID_inodes, POS(0, pos.offset), bpos_predecessor(pos),
 657                         BTREE_ITER_all_snapshots|
 658                         BTREE_ITER_with_updates, k, ret)
 659                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot) &&
 660                     bkey_is_inode(k.k)) {
 661                         ret = 1;
 662                         break;
 663                 }
 664         bch2_trans_iter_exit(trans, &iter);
 665         return ret;
 666 }
 667
 668 static int update_inode_has_children(struct btree_trans *trans,
 669                                      struct bkey_s k,
 670                                      bool have_child)
 671 {
 672         if (!have_child) {
 673                 int ret = bch2_inode_has_child_snapshots(trans, k.k->p);
 674                 if (ret)
 675                         return ret < 0 ? ret : 0;
 676         }
 677
 678         u64 f = bkey_inode_flags(k.s_c);
 679         if (have_child != !!(f & BCH_INODE_has_child_snapshot))
 680                 bkey_inode_flags_set(k, f ^ BCH_INODE_has_child_snapshot);
 681
 682         return 0;
 683 }
 684
 685 static int update_parent_inode_has_children(struct btree_trans *trans, struct bpos pos,
 686                                             bool have_child)
 687 {
 688         struct btree_iter iter;
 689         struct bkey_s_c k = bch2_inode_get_iter_snapshot_parent(trans,
 690                                                 &iter, pos, BTREE_ITER_with_updates);
 691         int ret = bkey_err(k);
 692         if (ret)
 693                 return ret;
 694         if (!k.k)
 695                 return 0;
 696
 697         if (!have_child) {
 698                 ret = bch2_inode_has_child_snapshots(trans, k.k->p);
 699                 if (ret) {
 700                         ret = ret < 0 ? ret : 0;
 701                         goto err;
 702                 }
 703         }
 704
 705         u64 f = bkey_inode_flags(k);
 706         if (have_child != !!(f & BCH_INODE_has_child_snapshot)) {
 707                 struct bkey_i *update = bch2_bkey_make_mut(trans, &iter, &k,
 708                                              BTREE_UPDATE_internal_snapshot_node);
 709                 ret = PTR_ERR_OR_ZERO(update);
 710                 if (ret)
 711                         goto err;
 712
 713                 bkey_inode_flags_set(bkey_i_to_s(update), f ^ BCH_INODE_has_child_snapshot);
 714         }
 715 err:
 716         bch2_trans_iter_exit(trans, &iter);
 717         return ret;
 718 }
 719
 720 int bch2_trigger_inode(struct btree_trans *trans,
 721                        enum btree_id btree_id, unsigned level,
 722                        struct bkey_s_c old,
 723                        struct bkey_s new,
 724                        enum btree_iter_update_trigger_flags flags)
 725 {
 726         struct bch_fs *c = trans->c;
 727
 728         if ((flags & BTREE_TRIGGER_atomic) && (flags & BTREE_TRIGGER_insert)) {
 729                 BUG_ON(!trans->journal_res.seq);
 730                 bkey_s_to_inode_v3(new).v->bi_journal_seq = cpu_to_le64(trans->journal_res.seq);
 731         }
 732
 733         s64 nr = bkey_is_inode(new.k) - bkey_is_inode(old.k);
 734         if ((flags & (BTREE_TRIGGER_transactional|BTREE_TRIGGER_gc)) && nr) {
 735                 struct disk_accounting_pos acc = { .type = BCH_DISK_ACCOUNTING_nr_inodes };
 736                 int ret = bch2_disk_accounting_mod(trans, &acc, &nr, 1, flags & BTREE_TRIGGER_gc);
 737                 if (ret)
 738                         return ret;
 739         }
 740
 741         if (flags & BTREE_TRIGGER_transactional) {
 742                 int unlinked_delta =    (int) bkey_is_unlinked_inode(new.s_c) -
 743                                         (int) bkey_is_unlinked_inode(old);
 744                 if (unlinked_delta) {
 745                         int ret = bch2_btree_bit_mod_buffered(trans, BTREE_ID_deleted_inodes,
 746                                                               new.k->p, unlinked_delta > 0);
 747                         if (ret)
 748                                 return ret;
 749                 }
 750
 751                 /*
 752                  * If we're creating or deleting an inode at this snapshot ID,
 753                  * and there might be an inode in a parent snapshot ID, we might
 754                  * need to set or clear the has_child_snapshot flag on the
 755                  * parent.
 756                  */
 757                 int deleted_delta = (int) bkey_is_inode(new.k) -
 758                                     (int) bkey_is_inode(old.k);
 759                 if (deleted_delta &&
 760                     bch2_snapshot_parent(c, new.k->p.snapshot)) {
 761                         int ret = update_parent_inode_has_children(trans, new.k->p,
 762                                                                    deleted_delta > 0);
 763                         if (ret)
 764                                 return ret;
 765                 }
 766
 767                 /*
 768                  * When an inode is first updated in a new snapshot, we may need
 769                  * to clear has_child_snapshot
 770                  */
 771                 if (deleted_delta > 0) {
 772                         int ret = update_inode_has_children(trans, new, false);
 773                         if (ret)
 774                                 return ret;
 775                 }
 776         }
 777
 778         return 0;
 779 }
 780
 781 int bch2_inode_generation_validate(struct bch_fs *c, struct bkey_s_c k,
 782                                    enum bch_validate_flags flags)
 783 {
 784         int ret = 0;
 785
 786         bkey_fsck_err_on(k.k->p.inode,
 787                          c, inode_pos_inode_nonzero,
 788                          "nonzero k.p.inode");
 789 fsck_err:
 790         return ret;
 791 }
 792
 793 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 794                                    struct bkey_s_c k)
 795 {
 796         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 797
 798         prt_printf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 799 }
 800
 801 void bch2_inode_init_early(struct bch_fs *c,
 802                            struct bch_inode_unpacked *inode_u)
 803 {
 804         enum bch_str_hash_type str_hash =
 805                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 806
 807         memset(inode_u, 0, sizeof(*inode_u));
 808
 809         SET_INODE_STR_HASH(inode_u, str_hash);
 810         get_random_bytes(&inode_u->bi_hash_seed, sizeof(inode_u->bi_hash_seed));
 811 }
 812
 813 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 814                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 815                           struct bch_inode_unpacked *parent)
 816 {
 817         inode_u->bi_mode        = mode;
 818         inode_u->bi_uid         = uid;
 819         inode_u->bi_gid         = gid;
 820         inode_u->bi_dev         = rdev;
 821         inode_u->bi_atime       = now;
 822         inode_u->bi_mtime       = now;
 823         inode_u->bi_ctime       = now;
 824         inode_u->bi_otime       = now;
 825
 826         if (parent && parent->bi_mode & S_ISGID) {
 827                 inode_u->bi_gid = parent->bi_gid;
 828                 if (S_ISDIR(mode))
 829                         inode_u->bi_mode |= S_ISGID;
 830         }
 831
 832         if (parent) {
 833 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 834                 BCH_INODE_OPTS()
 835 #undef x
 836         }
 837 }
 838
 839 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 840                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 841                      struct bch_inode_unpacked *parent)
 842 {
 843         bch2_inode_init_early(c, inode_u);
 844         bch2_inode_init_late(inode_u, bch2_current_time(c),
 845                              uid, gid, mode, rdev, parent);
 846 }
 847
 848 static inline u32 bkey_generation(struct bkey_s_c k)
 849 {
 850         switch (k.k->type) {
 851         case KEY_TYPE_inode:
 852         case KEY_TYPE_inode_v2:
 853                 BUG();
 854         case KEY_TYPE_inode_generation:
 855                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 856         default:
 857                 return 0;
 858         }
 859 }
 860
 861 /*
 862  * This just finds an empty slot:
 863  */
 864 int bch2_inode_create(struct btree_trans *trans,
 865                       struct btree_iter *iter,
 866                       struct bch_inode_unpacked *inode_u,
 867                       u32 snapshot, u64 cpu)
 868 {
 869         struct bch_fs *c = trans->c;
 870         struct bkey_s_c k;
 871         u64 min, max, start, pos, *hint;
 872         int ret = 0;
 873         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 874
 875         if (c->opts.shard_inode_numbers) {
 876                 bits -= c->inode_shard_bits;
 877
 878                 min = (cpu << bits);
 879                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 880
 881                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 882                 hint = c->unused_inode_hints + cpu;
 883         } else {
 884                 min = BLOCKDEV_INODE_MAX;
 885                 max = ~(ULLONG_MAX << bits);
 886                 hint = c->unused_inode_hints;
 887         }
 888
 889         start = READ_ONCE(*hint);
 890
 891         if (start >= max || start < min)
 892                 start = min;
 893
 894         pos = start;
 895         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 896                              BTREE_ITER_all_snapshots|
 897                              BTREE_ITER_intent);
 898 again:
 899         while ((k = bch2_btree_iter_peek(iter)).k &&
 900                !(ret = bkey_err(k)) &&
 901                bkey_lt(k.k->p, POS(0, max))) {
 902                 if (pos < iter->pos.offset)
 903                         goto found_slot;
 904
 905                 /*
 906                  * We don't need to iterate over keys in every snapshot once
 907                  * we've found just one:
 908                  */
 909                 pos = iter->pos.offset + 1;
 910                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 911         }
 912
 913         if (!ret && pos < max)
 914                 goto found_slot;
 915
 916         if (!ret && start == min)
 917                 ret = -BCH_ERR_ENOSPC_inode_create;
 918
 919         if (ret) {
 920                 bch2_trans_iter_exit(trans, iter);
 921                 return ret;
 922         }
 923
 924         /* Retry from start */
 925         pos = start = min;
 926         bch2_btree_iter_set_pos(iter, POS(0, pos));
 927         goto again;
 928 found_slot:
 929         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 930         k = bch2_btree_iter_peek_slot(iter);
 931         ret = bkey_err(k);
 932         if (ret) {
 933                 bch2_trans_iter_exit(trans, iter);
 934                 return ret;
 935         }
 936
 937         *hint                   = k.k->p.offset;
 938         inode_u->bi_inum        = k.k->p.offset;
 939         inode_u->bi_generation  = bkey_generation(k);
 940         return 0;
 941 }
 942
 943 static int bch2_inode_delete_keys(struct btree_trans *trans,
 944                                   subvol_inum inum, enum btree_id id)
 945 {
 946         struct btree_iter iter;
 947         struct bkey_s_c k;
 948         struct bkey_i delete;
 949         struct bpos end = POS(inum.inum, U64_MAX);
 950         u32 snapshot;
 951         int ret = 0;
 952
 953         /*
 954          * We're never going to be deleting partial extents, no need to use an
 955          * extent iterator:
 956          */
 957         bch2_trans_iter_init(trans, &iter, id, POS(inum.inum, 0),
 958                              BTREE_ITER_intent);
 959
 960         while (1) {
 961                 bch2_trans_begin(trans);
 962
 963                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 964                 if (ret)
 965                         goto err;
 966
 967                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 968
 969                 k = bch2_btree_iter_peek_upto(&iter, end);
 970                 ret = bkey_err(k);
 971                 if (ret)
 972                         goto err;
 973
 974                 if (!k.k)
 975                         break;
 976
 977                 bkey_init(&delete.k);
 978                 delete.k.p = iter.pos;
 979
 980                 if (iter.flags & BTREE_ITER_is_extents)
 981                         bch2_key_resize(&delete.k,
 982                                         bpos_min(end, k.k->p).offset -
 983                                         iter.pos.offset);
 984
 985                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 986                       bch2_trans_commit(trans, NULL, NULL,
 987                                         BCH_TRANS_COMMIT_no_enospc);
 988 err:
 989                 if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
 990                         break;
 991         }
 992
 993         bch2_trans_iter_exit(trans, &iter);
 994         return ret;
 995 }
 996
 997 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 998 {
 999         struct btree_trans *trans = bch2_trans_get(c);
1000         struct btree_iter iter = { NULL };
1001         struct bkey_i_inode_generation delete;
1002         struct bch_inode_unpacked inode_u;
1003         struct bkey_s_c k;
1004         u32 snapshot;
1005         int ret;
1006
1007         /*
1008          * If this was a directory, there shouldn't be any real dirents left -
1009          * but there could be whiteouts (from hash collisions) that we should
1010          * delete:
1011          *
1012          * XXX: the dirent could ideally would delete whiteouts when they're no
1013          * longer needed
1014          */
1015         ret   = bch2_inode_delete_keys(trans, inum, BTREE_ID_extents) ?:
1016                 bch2_inode_delete_keys(trans, inum, BTREE_ID_xattrs) ?:
1017                 bch2_inode_delete_keys(trans, inum, BTREE_ID_dirents);
1018         if (ret)
1019                 goto err;
1020 retry:
1021         bch2_trans_begin(trans);
1022
1023         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1024         if (ret)
1025                 goto err;
1026
1027         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
1028                                SPOS(0, inum.inum, snapshot),
1029                                BTREE_ITER_intent|BTREE_ITER_cached);
1030         ret = bkey_err(k);
1031         if (ret)
1032                 goto err;
1033
1034         if (!bkey_is_inode(k.k)) {
1035                 bch2_fs_inconsistent(c,
1036                                      "inode %llu:%u not found when deleting",
1037                                      inum.inum, snapshot);
1038                 ret = -EIO;
1039                 goto err;
1040         }
1041
1042         bch2_inode_unpack(k, &inode_u);
1043
1044         bkey_inode_generation_init(&delete.k_i);
1045         delete.k.p = iter.pos;
1046         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
1047
1048         ret   = bch2_trans_update(trans, &iter, &delete.k_i, 0) ?:
1049                 bch2_trans_commit(trans, NULL, NULL,
1050                                 BCH_TRANS_COMMIT_no_enospc);
1051 err:
1052         bch2_trans_iter_exit(trans, &iter);
1053         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1054                 goto retry;
1055
1056         if (ret)
1057                 goto err2;
1058
1059         ret = delete_ancestor_snapshot_inodes(trans, SPOS(0, inum.inum, snapshot));
1060 err2:
1061         bch2_trans_put(trans);
1062         return ret;
1063 }
1064
1065 int bch2_inode_find_by_inum_nowarn_trans(struct btree_trans *trans,
1066                                   subvol_inum inum,
1067                                   struct bch_inode_unpacked *inode)
1068 {
1069         struct btree_iter iter;
1070         int ret;
1071
1072         ret = bch2_inode_peek_nowarn(trans, &iter, inode, inum, 0);
1073         if (!ret)
1074                 bch2_trans_iter_exit(trans, &iter);
1075         return ret;
1076 }
1077
1078 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
1079                                   subvol_inum inum,
1080                                   struct bch_inode_unpacked *inode)
1081 {
1082         struct btree_iter iter;
1083         int ret;
1084
1085         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
1086         if (!ret)
1087                 bch2_trans_iter_exit(trans, &iter);
1088         return ret;
1089 }
1090
1091 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
1092                             struct bch_inode_unpacked *inode)
1093 {
1094         return bch2_trans_do(c, bch2_inode_find_by_inum_trans(trans, inum, inode));
1095 }
1096
1097 int bch2_inode_nlink_inc(struct bch_inode_unpacked *bi)
1098 {
1099         if (bi->bi_flags & BCH_INODE_unlinked)
1100                 bi->bi_flags &= ~BCH_INODE_unlinked;
1101         else {
1102                 if (bi->bi_nlink == U32_MAX)
1103                         return -EINVAL;
1104
1105                 bi->bi_nlink++;
1106         }
1107
1108         return 0;
1109 }
1110
1111 void bch2_inode_nlink_dec(struct btree_trans *trans, struct bch_inode_unpacked *bi)
1112 {
1113         if (bi->bi_nlink && (bi->bi_flags & BCH_INODE_unlinked)) {
1114                 bch2_trans_inconsistent(trans, "inode %llu unlinked but link count nonzero",
1115                                         bi->bi_inum);
1116                 return;
1117         }
1118
1119         if (bi->bi_flags & BCH_INODE_unlinked) {
1120                 bch2_trans_inconsistent(trans, "inode %llu link count underflow", bi->bi_inum);
1121                 return;
1122         }
1123
1124         if (bi->bi_nlink)
1125                 bi->bi_nlink--;
1126         else
1127                 bi->bi_flags |= BCH_INODE_unlinked;
1128 }
1129
1130 struct bch_opts bch2_inode_opts_to_opts(struct bch_inode_unpacked *inode)
1131 {
1132         struct bch_opts ret = { 0 };
1133 #define x(_name, _bits)                                                 \
1134         if (inode->bi_##_name)                                          \
1135                 opt_set(ret, _name, inode->bi_##_name - 1);
1136         BCH_INODE_OPTS()
1137 #undef x
1138         return ret;
1139 }
1140
1141 void bch2_inode_opts_get(struct bch_io_opts *opts, struct bch_fs *c,
1142                          struct bch_inode_unpacked *inode)
1143 {
1144 #define x(_name, _bits)         opts->_name = inode_opt_get(c, inode, _name);
1145         BCH_INODE_OPTS()
1146 #undef x
1147
1148         if (opts->nocow)
1149                 opts->compression = opts->background_compression = opts->data_checksum = opts->erasure_code = 0;
1150 }
1151
1152 int bch2_inum_opts_get(struct btree_trans *trans, subvol_inum inum, struct bch_io_opts *opts)
1153 {
1154         struct bch_inode_unpacked inode;
1155         int ret = lockrestart_do(trans, bch2_inode_find_by_inum_trans(trans, inum, &inode));
1156
1157         if (ret)
1158                 return ret;
1159
1160         bch2_inode_opts_get(opts, trans->c, &inode);
1161         return 0;
1162 }
1163
1164 static noinline int __bch2_inode_rm_snapshot(struct btree_trans *trans, u64 inum, u32 snapshot)
1165 {
1166         struct bch_fs *c = trans->c;
1167         struct btree_iter iter = { NULL };
1168         struct bkey_i_inode_generation delete;
1169         struct bch_inode_unpacked inode_u;
1170         struct bkey_s_c k;
1171         int ret;
1172
1173         do {
1174                 ret   = bch2_btree_delete_range_trans(trans, BTREE_ID_extents,
1175                                                       SPOS(inum, 0, snapshot),
1176                                                       SPOS(inum, U64_MAX, snapshot),
1177                                                       0, NULL) ?:
1178                         bch2_btree_delete_range_trans(trans, BTREE_ID_dirents,
1179                                                       SPOS(inum, 0, snapshot),
1180                                                       SPOS(inum, U64_MAX, snapshot),
1181                                                       0, NULL) ?:
1182                         bch2_btree_delete_range_trans(trans, BTREE_ID_xattrs,
1183                                                       SPOS(inum, 0, snapshot),
1184                                                       SPOS(inum, U64_MAX, snapshot),
1185                                                       0, NULL);
1186         } while (ret == -BCH_ERR_transaction_restart_nested);
1187         if (ret)
1188                 goto err;
1189 retry:
1190         bch2_trans_begin(trans);
1191
1192         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
1193                                SPOS(0, inum, snapshot), BTREE_ITER_intent);
1194         ret = bkey_err(k);
1195         if (ret)
1196                 goto err;
1197
1198         if (!bkey_is_inode(k.k)) {
1199                 bch2_fs_inconsistent(c,
1200                                      "inode %llu:%u not found when deleting",
1201                                      inum, snapshot);
1202                 ret = -EIO;
1203                 goto err;
1204         }
1205
1206         bch2_inode_unpack(k, &inode_u);
1207
1208         /* Subvolume root? */
1209         if (inode_u.bi_subvol)
1210                 bch_warn(c, "deleting inode %llu marked as unlinked, but also a subvolume root!?", inode_u.bi_inum);
1211
1212         bkey_inode_generation_init(&delete.k_i);
1213         delete.k.p = iter.pos;
1214         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
1215
1216         ret   = bch2_trans_update(trans, &iter, &delete.k_i, 0) ?:
1217                 bch2_trans_commit(trans, NULL, NULL,
1218                                 BCH_TRANS_COMMIT_no_enospc);
1219 err:
1220         bch2_trans_iter_exit(trans, &iter);
1221         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1222                 goto retry;
1223
1224         return ret ?: -BCH_ERR_transaction_restart_nested;
1225 }
1226
1227 /*
1228  * After deleting an inode, there may be versions in older snapshots that should
1229  * also be deleted - if they're not referenced by sibling snapshots and not open
1230  * in other subvolumes:
1231  */
1232 static int delete_ancestor_snapshot_inodes(struct btree_trans *trans, struct bpos pos)
1233 {
1234         struct btree_iter iter;
1235         struct bkey_s_c k;
1236         int ret;
1237 next_parent:
1238         ret = lockrestart_do(trans,
1239                 bkey_err(k = bch2_inode_get_iter_snapshot_parent(trans, &iter, pos, 0)));
1240         if (ret || !k.k)
1241                 return ret;
1242
1243         bool unlinked = bkey_is_unlinked_inode(k);
1244         pos = k.k->p;
1245         bch2_trans_iter_exit(trans, &iter);
1246
1247         if (!unlinked)
1248                 return 0;
1249
1250         ret = lockrestart_do(trans, bch2_inode_or_descendents_is_open(trans, pos));
1251         if (ret)
1252                 return ret < 0 ? ret : 0;
1253
1254         ret = __bch2_inode_rm_snapshot(trans, pos.offset, pos.snapshot);
1255         if (ret)
1256                 return ret;
1257         goto next_parent;
1258 }
1259
1260 int bch2_inode_rm_snapshot(struct btree_trans *trans, u64 inum, u32 snapshot)
1261 {
1262         return __bch2_inode_rm_snapshot(trans, inum, snapshot) ?:
1263                 delete_ancestor_snapshot_inodes(trans, SPOS(0, inum, snapshot));
1264 }
1265
1266 static int may_delete_deleted_inode(struct btree_trans *trans,
1267                                     struct btree_iter *iter,
1268                                     struct bpos pos,
1269                                     bool *need_another_pass)
1270 {
1271         struct bch_fs *c = trans->c;
1272         struct btree_iter inode_iter;
1273         struct bkey_s_c k;
1274         struct bch_inode_unpacked inode;
1275         struct printbuf buf = PRINTBUF;
1276         int ret;
1277
1278         k = bch2_bkey_get_iter(trans, &inode_iter, BTREE_ID_inodes, pos, BTREE_ITER_cached);
1279         ret = bkey_err(k);
1280         if (ret)
1281                 return ret;
1282
1283         ret = bkey_is_inode(k.k) ? 0 : -BCH_ERR_ENOENT_inode;
1284         if (fsck_err_on(!bkey_is_inode(k.k),
1285                         trans, deleted_inode_missing,
1286                         "nonexistent inode %llu:%u in deleted_inodes btree",
1287                         pos.offset, pos.snapshot))
1288                 goto delete;
1289
1290         ret = bch2_inode_unpack(k, &inode);
1291         if (ret)
1292                 goto out;
1293
1294         if (S_ISDIR(inode.bi_mode)) {
1295                 ret = bch2_empty_dir_snapshot(trans, pos.offset, 0, pos.snapshot);
1296                 if (fsck_err_on(bch2_err_matches(ret, ENOTEMPTY),
1297                                 trans, deleted_inode_is_dir,
1298                                 "non empty directory %llu:%u in deleted_inodes btree",
1299                                 pos.offset, pos.snapshot))
1300                         goto delete;
1301                 if (ret)
1302                         goto out;
1303         }
1304
1305         if (fsck_err_on(!(inode.bi_flags & BCH_INODE_unlinked),
1306                         trans, deleted_inode_not_unlinked,
1307                         "non-deleted inode %llu:%u in deleted_inodes btree",
1308                         pos.offset, pos.snapshot))
1309                 goto delete;
1310
1311         if (fsck_err_on(inode.bi_flags & BCH_INODE_has_child_snapshot,
1312                         trans, deleted_inode_has_child_snapshots,
1313                         "inode with child snapshots %llu:%u in deleted_inodes btree",
1314                         pos.offset, pos.snapshot))
1315                 goto delete;
1316
1317         ret = bch2_inode_has_child_snapshots(trans, k.k->p);
1318         if (ret < 0)
1319                 goto out;
1320
1321         if (ret) {
1322                 if (fsck_err(trans, inode_has_child_snapshots_wrong,
1323                              "inode has_child_snapshots flag wrong (should be set)\n%s",
1324                              (printbuf_reset(&buf),
1325                               bch2_inode_unpacked_to_text(&buf, &inode),
1326                               buf.buf))) {
1327                         inode.bi_flags |= BCH_INODE_has_child_snapshot;
1328                         ret = __bch2_fsck_write_inode(trans, &inode);
1329                         if (ret)
1330                                 goto out;
1331                 }
1332                 goto delete;
1333
1334         }
1335
1336         if (test_bit(BCH_FS_clean_recovery, &c->flags) &&
1337             !fsck_err(trans, deleted_inode_but_clean,
1338                       "filesystem marked as clean but have deleted inode %llu:%u",
1339                       pos.offset, pos.snapshot)) {
1340                 ret = 0;
1341                 goto out;
1342         }
1343
1344         ret = 1;
1345 out:
1346 fsck_err:
1347         bch2_trans_iter_exit(trans, &inode_iter);
1348         printbuf_exit(&buf);
1349         return ret;
1350 delete:
1351         ret = bch2_btree_bit_mod_buffered(trans, BTREE_ID_deleted_inodes, pos, false);
1352         goto out;
1353 }
1354
1355 int bch2_delete_dead_inodes(struct bch_fs *c)
1356 {
1357         struct btree_trans *trans = bch2_trans_get(c);
1358         bool need_another_pass;
1359         int ret;
1360 again:
1361         /*
1362          * if we ran check_inodes() unlinked inodes will have already been
1363          * cleaned up but the write buffer will be out of sync; therefore we
1364          * alway need a write buffer flush
1365          */
1366         ret = bch2_btree_write_buffer_flush_sync(trans);
1367         if (ret)
1368                 goto err;
1369
1370         need_another_pass = false;
1371
1372         /*
1373          * Weird transaction restart handling here because on successful delete,
1374          * bch2_inode_rm_snapshot() will return a nested transaction restart,
1375          * but we can't retry because the btree write buffer won't have been
1376          * flushed and we'd spin:
1377          */
1378         ret = for_each_btree_key_commit(trans, iter, BTREE_ID_deleted_inodes, POS_MIN,
1379                                         BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k,
1380                                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1381                 ret = may_delete_deleted_inode(trans, &iter, k.k->p, &need_another_pass);
1382                 if (ret > 0) {
1383                         bch_verbose(c, "deleting unlinked inode %llu:%u", k.k->p.offset, k.k->p.snapshot);
1384
1385                         ret = bch2_inode_rm_snapshot(trans, k.k->p.offset, k.k->p.snapshot);
1386                         /*
1387                          * We don't want to loop here: a transaction restart
1388                          * error here means we handled a transaction restart and
1389                          * we're actually done, but if we loop we'll retry the
1390                          * same key because the write buffer hasn't been flushed
1391                          * yet
1392                          */
1393                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
1394                                 ret = 0;
1395                                 continue;
1396                         }
1397                 }
1398
1399                 ret;
1400         }));
1401
1402         if (!ret && need_another_pass)
1403                 goto again;
1404 err:
1405         bch2_trans_put(trans);
1406         return ret;
1407 }