fs/bcachefs/extents.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  *
   5  * Code for managing the extent btree and dynamically updating the writeback
   6  * dirty sector count.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "bkey_methods.h"
  11 #include "btree_cache.h"
  12 #include "btree_gc.h"
  13 #include "btree_io.h"
  14 #include "btree_iter.h"
  15 #include "buckets.h"
  16 #include "checksum.h"
  17 #include "compress.h"
  18 #include "debug.h"
  19 #include "disk_groups.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "inode.h"
  23 #include "journal.h"
  24 #include "rebalance.h"
  25 #include "replicas.h"
  26 #include "super.h"
  27 #include "super-io.h"
  28 #include "trace.h"
  29 #include "util.h"
  30
  31 static const char * const bch2_extent_flags_strs[] = {
  32 #define x(n, v) [BCH_EXTENT_FLAG_##n] = #n,
  33         BCH_EXTENT_FLAGS()
  34 #undef x
  35         NULL,
  36 };
  37
  38 static unsigned bch2_crc_field_size_max[] = {
  39         [BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
  40         [BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
  41         [BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
  42 };
  43
  44 static void bch2_extent_crc_pack(union bch_extent_crc *,
  45                                  struct bch_extent_crc_unpacked,
  46                                  enum bch_extent_entry_type);
  47
  48 void bch2_io_failures_to_text(struct printbuf *out,
  49                               struct bch_fs *c,
  50                               struct bch_io_failures *failed)
  51 {
  52         static const char * const error_types[] = {
  53                 "btree validate", "io", "checksum", "ec reconstruct", NULL
  54         };
  55
  56         for (struct bch_dev_io_failures *f = failed->devs;
  57              f < failed->devs + failed->nr;
  58              f++) {
  59                 unsigned errflags =
  60                         ((!!f->failed_btree_validate)   << 0) |
  61                         ((!!f->failed_io)               << 1) |
  62                         ((!!f->failed_csum_nr)          << 2) |
  63                         ((!!f->failed_ec)               << 3);
  64
  65                 bch2_printbuf_make_room(out, 1024);
  66                 out->atomic++;
  67                 scoped_guard(rcu) {
  68                         struct bch_dev *ca = bch2_dev_rcu_noerror(c, f->dev);
  69                         if (ca)
  70                                 prt_str(out, ca->name);
  71                         else
  72                                 prt_printf(out, "(invalid device %u)", f->dev);
  73                 }
  74                 --out->atomic;
  75
  76                 prt_char(out, ' ');
  77
  78                 if (!errflags) {
  79                         prt_str(out, "no error - confused");
  80                 } else if (is_power_of_2(errflags)) {
  81                         prt_bitflags(out, error_types, errflags);
  82                         prt_str(out, " error");
  83                 } else {
  84                         prt_str(out, "errors: ");
  85                         prt_bitflags(out, error_types, errflags);
  86                 }
  87                 prt_newline(out);
  88         }
  89 }
  90
  91 struct bch_dev_io_failures *bch2_dev_io_failures(struct bch_io_failures *f,
  92                                                  unsigned dev)
  93 {
  94         struct bch_dev_io_failures *i;
  95
  96         for (i = f->devs; i < f->devs + f->nr; i++)
  97                 if (i->dev == dev)
  98                         return i;
  99
 100         return NULL;
 101 }
 102
 103 void bch2_mark_io_failure(struct bch_io_failures *failed,
 104                           struct extent_ptr_decoded *p,
 105                           bool csum_error)
 106 {
 107         struct bch_dev_io_failures *f = bch2_dev_io_failures(failed, p->ptr.dev);
 108
 109         if (!f) {
 110                 BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
 111
 112                 f = &failed->devs[failed->nr++];
 113                 memset(f, 0, sizeof(*f));
 114                 f->dev = p->ptr.dev;
 115         }
 116
 117         if (p->do_ec_reconstruct)
 118                 f->failed_ec = true;
 119         else if (!csum_error)
 120                 f->failed_io = true;
 121         else
 122                 f->failed_csum_nr++;
 123 }
 124
 125 void bch2_mark_btree_validate_failure(struct bch_io_failures *failed,
 126                                       unsigned dev)
 127 {
 128         struct bch_dev_io_failures *f = bch2_dev_io_failures(failed, dev);
 129
 130         if (!f) {
 131                 BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
 132
 133                 f = &failed->devs[failed->nr++];
 134                 memset(f, 0, sizeof(*f));
 135                 f->dev = dev;
 136         }
 137
 138         f->failed_btree_validate = true;
 139 }
 140
 141 static inline u64 dev_latency(struct bch_dev *ca)
 142 {
 143         return ca ? atomic64_read(&ca->cur_latency[READ]) : S64_MAX;
 144 }
 145
 146 static inline int dev_failed(struct bch_dev *ca)
 147 {
 148         return !ca || ca->mi.state == BCH_MEMBER_STATE_failed;
 149 }
 150
 151 /*
 152  * returns true if p1 is better than p2:
 153  */
 154 static inline bool ptr_better(struct bch_fs *c,
 155                               const struct extent_ptr_decoded p1,
 156                               u64 p1_latency,
 157                               struct bch_dev *ca1,
 158                               const struct extent_ptr_decoded p2,
 159                               u64 p2_latency)
 160 {
 161         struct bch_dev *ca2 = bch2_dev_rcu(c, p2.ptr.dev);
 162
 163         int failed_delta = dev_failed(ca1) - dev_failed(ca2);
 164         if (unlikely(failed_delta))
 165                 return failed_delta < 0;
 166
 167         if (static_branch_unlikely(&bch2_force_reconstruct_read))
 168                 return p1.do_ec_reconstruct > p2.do_ec_reconstruct;
 169
 170         if (unlikely(p1.do_ec_reconstruct || p2.do_ec_reconstruct))
 171                 return p1.do_ec_reconstruct < p2.do_ec_reconstruct;
 172
 173         int crc_retry_delta = (int) p1.crc_retry_nr - (int) p2.crc_retry_nr;
 174         if (unlikely(crc_retry_delta))
 175                 return crc_retry_delta < 0;
 176
 177         /* Pick at random, biased in favor of the faster device: */
 178
 179         return bch2_get_random_u64_below(p1_latency + p2_latency) > p1_latency;
 180 }
 181
 182 /*
 183  * This picks a non-stale pointer, preferably from a device other than @avoid.
 184  * Avoid can be NULL, meaning pick any. If there are no non-stale pointers to
 185  * other devices, it will still pick a pointer from avoid.
 186  */
 187 int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 188                                struct bch_io_failures *failed,
 189                                struct extent_ptr_decoded *pick,
 190                                int dev)
 191 {
 192         bool have_csum_errors = false, have_io_errors = false, have_missing_devs = false;
 193         bool have_dirty_ptrs = false, have_pick = false;
 194
 195         if (k.k->type == KEY_TYPE_error)
 196                 return bch_err_throw(c, key_type_error);
 197
 198         rcu_read_lock();
 199         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 200         const union bch_extent_entry *entry;
 201         struct extent_ptr_decoded p;
 202         u64 pick_latency;
 203
 204         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 205                 have_dirty_ptrs |= !p.ptr.cached;
 206
 207                 /*
 208                  * Unwritten extent: no need to actually read, treat it as a
 209                  * hole and return 0s:
 210                  */
 211                 if (p.ptr.unwritten) {
 212                         rcu_read_unlock();
 213                         return 0;
 214                 }
 215
 216                 /* Are we being asked to read from a specific device? */
 217                 if (dev >= 0 && p.ptr.dev != dev)
 218                         continue;
 219
 220                 struct bch_dev *ca = bch2_dev_rcu_noerror(c, p.ptr.dev);
 221
 222                 if (unlikely(!ca && p.ptr.dev != BCH_SB_MEMBER_INVALID)) {
 223                         rcu_read_unlock();
 224                         int ret = bch2_dev_missing_bkey(c, k, p.ptr.dev);
 225                         if (ret)
 226                                 return ret;
 227                         rcu_read_lock();
 228                 }
 229
 230                 if (p.ptr.cached && (!ca || dev_ptr_stale_rcu(ca, &p.ptr)))
 231                         continue;
 232
 233                 struct bch_dev_io_failures *f =
 234                         unlikely(failed) ? bch2_dev_io_failures(failed, p.ptr.dev) : NULL;
 235                 if (unlikely(f)) {
 236                         p.crc_retry_nr     = f->failed_csum_nr;
 237                         p.has_ec          &= ~f->failed_ec;
 238
 239                         if (ca && ca->mi.state != BCH_MEMBER_STATE_failed) {
 240                                 have_io_errors  |= f->failed_io;
 241                                 have_io_errors  |= f->failed_btree_validate;
 242                                 have_io_errors  |= f->failed_ec;
 243                         }
 244                         have_csum_errors        |= !!f->failed_csum_nr;
 245
 246                         if (p.has_ec && (f->failed_io || f->failed_csum_nr))
 247                                 p.do_ec_reconstruct = true;
 248                         else if (f->failed_io ||
 249                                  f->failed_btree_validate ||
 250                                  f->failed_csum_nr > c->opts.checksum_err_retry_nr)
 251                                 continue;
 252                 }
 253
 254                 have_missing_devs |= ca && !bch2_dev_is_online(ca);
 255
 256                 if (!ca || !bch2_dev_is_online(ca)) {
 257                         if (!p.has_ec)
 258                                 continue;
 259                         p.do_ec_reconstruct = true;
 260                 }
 261
 262                 if (static_branch_unlikely(&bch2_force_reconstruct_read) && p.has_ec)
 263                         p.do_ec_reconstruct = true;
 264
 265                 u64 p_latency = dev_latency(ca);
 266                 /*
 267                  * Square the latencies, to bias more in favor of the faster
 268                  * device - we never want to stop issuing reads to the slower
 269                  * device altogether, so that we can update our latency numbers:
 270                  */
 271                 p_latency *= p_latency;
 272
 273                 if (!have_pick ||
 274                     ptr_better(c,
 275                                p, p_latency, ca,
 276                                *pick, pick_latency)) {
 277                         *pick = p;
 278                         pick_latency = p_latency;
 279                         have_pick = true;
 280                 }
 281         }
 282         rcu_read_unlock();
 283
 284         if (have_pick)
 285                 return 1;
 286         if (!have_dirty_ptrs)
 287                 return 0;
 288         if (have_missing_devs)
 289                 return bch_err_throw(c, no_device_to_read_from);
 290         if (have_csum_errors)
 291                 return bch_err_throw(c, data_read_csum_err);
 292         if (have_io_errors)
 293                 return bch_err_throw(c, data_read_io_err);
 294
 295         /*
 296          * If we get here, we have pointers (bkey_ptrs_validate() ensures that),
 297          * but they don't point to valid devices:
 298          */
 299         return bch_err_throw(c, no_devices_valid);
 300 }
 301
 302 /* KEY_TYPE_btree_ptr: */
 303
 304 int bch2_btree_ptr_validate(struct bch_fs *c, struct bkey_s_c k,
 305                             struct bkey_validate_context from)
 306 {
 307         int ret = 0;
 308
 309         bkey_fsck_err_on(bkey_val_u64s(k.k) > BCH_REPLICAS_MAX,
 310                          c, btree_ptr_val_too_big,
 311                          "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
 312
 313         ret = bch2_bkey_ptrs_validate(c, k, from);
 314 fsck_err:
 315         return ret;
 316 }
 317
 318 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 319                             struct bkey_s_c k)
 320 {
 321         bch2_bkey_ptrs_to_text(out, c, k);
 322 }
 323
 324 int bch2_btree_ptr_v2_validate(struct bch_fs *c, struct bkey_s_c k,
 325                                struct bkey_validate_context from)
 326 {
 327         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 328         int ret = 0;
 329
 330         bkey_fsck_err_on(bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX,
 331                          c, btree_ptr_v2_val_too_big,
 332                          "value too big (%zu > %zu)",
 333                          bkey_val_u64s(k.k), BKEY_BTREE_PTR_VAL_U64s_MAX);
 334
 335         bkey_fsck_err_on(bpos_ge(bp.v->min_key, bp.k->p),
 336                          c, btree_ptr_v2_min_key_bad,
 337                          "min_key > key");
 338
 339         if ((from.flags & BCH_VALIDATE_write) &&
 340             c->sb.version_min >= bcachefs_metadata_version_btree_ptr_sectors_written)
 341                 bkey_fsck_err_on(!bp.v->sectors_written,
 342                                  c, btree_ptr_v2_written_0,
 343                                  "sectors_written == 0");
 344
 345         ret = bch2_bkey_ptrs_validate(c, k, from);
 346 fsck_err:
 347         return ret;
 348 }
 349
 350 void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
 351                                struct bkey_s_c k)
 352 {
 353         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 354
 355         prt_printf(out, "seq %llx written %u min_key %s",
 356                le64_to_cpu(bp.v->seq),
 357                le16_to_cpu(bp.v->sectors_written),
 358                BTREE_PTR_RANGE_UPDATED(bp.v) ? "R " : "");
 359
 360         bch2_bpos_to_text(out, bp.v->min_key);
 361         prt_printf(out, " ");
 362         bch2_bkey_ptrs_to_text(out, c, k);
 363 }
 364
 365 void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
 366                               unsigned big_endian, int write,
 367                               struct bkey_s k)
 368 {
 369         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(k);
 370
 371         compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
 372
 373         if (version < bcachefs_metadata_version_inode_btree_change &&
 374             btree_id_is_extents(btree_id) &&
 375             !bkey_eq(bp.v->min_key, POS_MIN))
 376                 bp.v->min_key = write
 377                         ? bpos_nosnap_predecessor(bp.v->min_key)
 378                         : bpos_nosnap_successor(bp.v->min_key);
 379 }
 380
 381 /* KEY_TYPE_extent: */
 382
 383 bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 384 {
 385         struct bkey_ptrs   l_ptrs = bch2_bkey_ptrs(l);
 386         struct bkey_ptrs_c r_ptrs = bch2_bkey_ptrs_c(r);
 387         union bch_extent_entry *en_l;
 388         const union bch_extent_entry *en_r;
 389         struct extent_ptr_decoded lp, rp;
 390         bool use_right_ptr;
 391
 392         en_l = l_ptrs.start;
 393         en_r = r_ptrs.start;
 394         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 395                 if (extent_entry_type(en_l) != extent_entry_type(en_r))
 396                         return false;
 397
 398                 en_l = extent_entry_next(en_l);
 399                 en_r = extent_entry_next(en_r);
 400         }
 401
 402         if (en_l < l_ptrs.end || en_r < r_ptrs.end)
 403                 return false;
 404
 405         en_l = l_ptrs.start;
 406         en_r = r_ptrs.start;
 407         lp.crc = bch2_extent_crc_unpack(l.k, NULL);
 408         rp.crc = bch2_extent_crc_unpack(r.k, NULL);
 409
 410         guard(rcu)();
 411
 412         while (__bkey_ptr_next_decode(l.k, l_ptrs.end, lp, en_l) &&
 413                __bkey_ptr_next_decode(r.k, r_ptrs.end, rp, en_r)) {
 414                 if (lp.ptr.offset + lp.crc.offset + lp.crc.live_size !=
 415                     rp.ptr.offset + rp.crc.offset ||
 416                     lp.ptr.dev                  != rp.ptr.dev ||
 417                     lp.ptr.gen                  != rp.ptr.gen ||
 418                     lp.ptr.unwritten            != rp.ptr.unwritten ||
 419                     lp.has_ec                   != rp.has_ec)
 420                         return false;
 421
 422                 /* Extents may not straddle buckets: */
 423                 struct bch_dev *ca = bch2_dev_rcu(c, lp.ptr.dev);
 424                 bool same_bucket = ca && PTR_BUCKET_NR(ca, &lp.ptr) == PTR_BUCKET_NR(ca, &rp.ptr);
 425
 426                 if (!same_bucket)
 427                         return false;
 428
 429                 if (lp.has_ec                   != rp.has_ec ||
 430                     (lp.has_ec &&
 431                      (lp.ec.block               != rp.ec.block ||
 432                       lp.ec.redundancy          != rp.ec.redundancy ||
 433                       lp.ec.idx                 != rp.ec.idx)))
 434                         return false;
 435
 436                 if (lp.crc.compression_type     != rp.crc.compression_type ||
 437                     lp.crc.nonce                != rp.crc.nonce)
 438                         return false;
 439
 440                 if (lp.crc.offset + lp.crc.live_size + rp.crc.live_size <=
 441                     lp.crc.uncompressed_size) {
 442                         /* can use left extent's crc entry */
 443                 } else if (lp.crc.live_size <= rp.crc.offset) {
 444                         /* can use right extent's crc entry */
 445                 } else {
 446                         /* check if checksums can be merged: */
 447                         if (lp.crc.csum_type            != rp.crc.csum_type ||
 448                             lp.crc.nonce                != rp.crc.nonce ||
 449                             crc_is_compressed(lp.crc) ||
 450                             !bch2_checksum_mergeable(lp.crc.csum_type))
 451                                 return false;
 452
 453                         if (lp.crc.offset + lp.crc.live_size != lp.crc.compressed_size ||
 454                             rp.crc.offset)
 455                                 return false;
 456
 457                         if (lp.crc.csum_type &&
 458                             lp.crc.uncompressed_size +
 459                             rp.crc.uncompressed_size > (c->opts.encoded_extent_max >> 9))
 460                                 return false;
 461                 }
 462
 463                 en_l = extent_entry_next(en_l);
 464                 en_r = extent_entry_next(en_r);
 465         }
 466
 467         en_l = l_ptrs.start;
 468         en_r = r_ptrs.start;
 469         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 470                 if (extent_entry_is_crc(en_l)) {
 471                         struct bch_extent_crc_unpacked crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 472                         struct bch_extent_crc_unpacked crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 473
 474                         if (crc_l.uncompressed_size + crc_r.uncompressed_size >
 475                             bch2_crc_field_size_max[extent_entry_type(en_l)])
 476                                 return false;
 477                 }
 478
 479                 en_l = extent_entry_next(en_l);
 480                 en_r = extent_entry_next(en_r);
 481         }
 482
 483         use_right_ptr = false;
 484         en_l = l_ptrs.start;
 485         en_r = r_ptrs.start;
 486         while (en_l < l_ptrs.end) {
 487                 if (extent_entry_type(en_l) == BCH_EXTENT_ENTRY_ptr &&
 488                     use_right_ptr)
 489                         en_l->ptr = en_r->ptr;
 490
 491                 if (extent_entry_is_crc(en_l)) {
 492                         struct bch_extent_crc_unpacked crc_l =
 493                                 bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 494                         struct bch_extent_crc_unpacked crc_r =
 495                                 bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 496
 497                         use_right_ptr = false;
 498
 499                         if (crc_l.offset + crc_l.live_size + crc_r.live_size <=
 500                             crc_l.uncompressed_size) {
 501                                 /* can use left extent's crc entry */
 502                         } else if (crc_l.live_size <= crc_r.offset) {
 503                                 /* can use right extent's crc entry */
 504                                 crc_r.offset -= crc_l.live_size;
 505                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_r,
 506                                                      extent_entry_type(en_l));
 507                                 use_right_ptr = true;
 508                         } else {
 509                                 crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
 510                                                                  crc_l.csum,
 511                                                                  crc_r.csum,
 512                                                                  crc_r.uncompressed_size << 9);
 513
 514                                 crc_l.uncompressed_size += crc_r.uncompressed_size;
 515                                 crc_l.compressed_size   += crc_r.compressed_size;
 516                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
 517                                                      extent_entry_type(en_l));
 518                         }
 519                 }
 520
 521                 en_l = extent_entry_next(en_l);
 522                 en_r = extent_entry_next(en_r);
 523         }
 524
 525         bch2_key_resize(l.k, l.k->size + r.k->size);
 526         return true;
 527 }
 528
 529 /* KEY_TYPE_reservation: */
 530
 531 int bch2_reservation_validate(struct bch_fs *c, struct bkey_s_c k,
 532                               struct bkey_validate_context from)
 533 {
 534         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 535         int ret = 0;
 536
 537         bkey_fsck_err_on(!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX,
 538                          c, reservation_key_nr_replicas_invalid,
 539                          "invalid nr_replicas (%u)", r.v->nr_replicas);
 540 fsck_err:
 541         return ret;
 542 }
 543
 544 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
 545                               struct bkey_s_c k)
 546 {
 547         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 548
 549         prt_printf(out, "generation %u replicas %u",
 550                le32_to_cpu(r.v->generation),
 551                r.v->nr_replicas);
 552 }
 553
 554 bool bch2_reservation_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
 555 {
 556         struct bkey_s_reservation l = bkey_s_to_reservation(_l);
 557         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(_r);
 558
 559         if (l.v->generation != r.v->generation ||
 560             l.v->nr_replicas != r.v->nr_replicas)
 561                 return false;
 562
 563         bch2_key_resize(l.k, l.k->size + r.k->size);
 564         return true;
 565 }
 566
 567 /* Extent checksum entries: */
 568
 569 /* returns true if not equal */
 570 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 571                                          struct bch_extent_crc_unpacked r)
 572 {
 573         return (l.csum_type             != r.csum_type ||
 574                 l.compression_type      != r.compression_type ||
 575                 l.compressed_size       != r.compressed_size ||
 576                 l.uncompressed_size     != r.uncompressed_size ||
 577                 l.offset                != r.offset ||
 578                 l.live_size             != r.live_size ||
 579                 l.nonce                 != r.nonce ||
 580                 bch2_crc_cmp(l.csum, r.csum));
 581 }
 582
 583 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 584                                   struct bch_extent_crc_unpacked n)
 585 {
 586         return !crc_is_compressed(u) &&
 587                 u.csum_type &&
 588                 u.uncompressed_size > u.live_size &&
 589                 bch2_csum_type_is_encryption(u.csum_type) ==
 590                 bch2_csum_type_is_encryption(n.csum_type);
 591 }
 592
 593 bool bch2_can_narrow_extent_crcs(struct bkey_s_c k,
 594                                  struct bch_extent_crc_unpacked n)
 595 {
 596         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 597         struct bch_extent_crc_unpacked crc;
 598         const union bch_extent_entry *i;
 599
 600         if (!n.csum_type)
 601                 return false;
 602
 603         bkey_for_each_crc(k.k, ptrs, crc, i)
 604                 if (can_narrow_crc(crc, n))
 605                         return true;
 606
 607         return false;
 608 }
 609
 610 /*
 611  * We're writing another replica for this extent, so while we've got the data in
 612  * memory we'll be computing a new checksum for the currently live data.
 613  *
 614  * If there are other replicas we aren't moving, and they are checksummed but
 615  * not compressed, we can modify them to point to only the data that is
 616  * currently live (so that readers won't have to bounce) while we've got the
 617  * checksum we need:
 618  */
 619 bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 620 {
 621         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 622         struct bch_extent_crc_unpacked u;
 623         struct extent_ptr_decoded p;
 624         union bch_extent_entry *i;
 625         bool ret = false;
 626
 627         /* Find a checksum entry that covers only live data: */
 628         if (!n.csum_type) {
 629                 bkey_for_each_crc(&k->k, ptrs, u, i)
 630                         if (!crc_is_compressed(u) &&
 631                             u.csum_type &&
 632                             u.live_size == u.uncompressed_size) {
 633                                 n = u;
 634                                 goto found;
 635                         }
 636                 return false;
 637         }
 638 found:
 639         BUG_ON(crc_is_compressed(n));
 640         BUG_ON(n.offset);
 641         BUG_ON(n.live_size != k->k.size);
 642
 643 restart_narrow_pointers:
 644         ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 645
 646         bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
 647                 if (can_narrow_crc(p.crc, n)) {
 648                         bch2_bkey_drop_ptr_noerror(bkey_i_to_s(k), &i->ptr);
 649                         p.ptr.offset += p.crc.offset;
 650                         p.crc = n;
 651                         bch2_extent_ptr_decoded_append(k, &p);
 652                         ret = true;
 653                         goto restart_narrow_pointers;
 654                 }
 655
 656         return ret;
 657 }
 658
 659 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
 660                                  struct bch_extent_crc_unpacked src,
 661                                  enum bch_extent_entry_type type)
 662 {
 663 #define common_fields(_src)                                             \
 664                 .type                   = BIT(type),                    \
 665                 .csum_type              = _src.csum_type,               \
 666                 .compression_type       = _src.compression_type,        \
 667                 ._compressed_size       = _src.compressed_size - 1,     \
 668                 ._uncompressed_size     = _src.uncompressed_size - 1,   \
 669                 .offset                 = _src.offset
 670
 671         switch (type) {
 672         case BCH_EXTENT_ENTRY_crc32:
 673                 dst->crc32              = (struct bch_extent_crc32) {
 674                         common_fields(src),
 675                         .csum           = (u32 __force) *((__le32 *) &src.csum.lo),
 676                 };
 677                 break;
 678         case BCH_EXTENT_ENTRY_crc64:
 679                 dst->crc64              = (struct bch_extent_crc64) {
 680                         common_fields(src),
 681                         .nonce          = src.nonce,
 682                         .csum_lo        = (u64 __force) src.csum.lo,
 683                         .csum_hi        = (u64 __force) *((__le16 *) &src.csum.hi),
 684                 };
 685                 break;
 686         case BCH_EXTENT_ENTRY_crc128:
 687                 dst->crc128             = (struct bch_extent_crc128) {
 688                         common_fields(src),
 689                         .nonce          = src.nonce,
 690                         .csum           = src.csum,
 691                 };
 692                 break;
 693         default:
 694                 BUG();
 695         }
 696 #undef set_common_fields
 697 }
 698
 699 void bch2_extent_crc_append(struct bkey_i *k,
 700                             struct bch_extent_crc_unpacked new)
 701 {
 702         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 703         union bch_extent_crc *crc = (void *) ptrs.end;
 704         enum bch_extent_entry_type type;
 705
 706         if (bch_crc_bytes[new.csum_type]        <= 4 &&
 707             new.uncompressed_size               <= CRC32_SIZE_MAX &&
 708             new.nonce                           <= CRC32_NONCE_MAX)
 709                 type = BCH_EXTENT_ENTRY_crc32;
 710         else if (bch_crc_bytes[new.csum_type]   <= 10 &&
 711                    new.uncompressed_size        <= CRC64_SIZE_MAX &&
 712                    new.nonce                    <= CRC64_NONCE_MAX)
 713                 type = BCH_EXTENT_ENTRY_crc64;
 714         else if (bch_crc_bytes[new.csum_type]   <= 16 &&
 715                    new.uncompressed_size        <= CRC128_SIZE_MAX &&
 716                    new.nonce                    <= CRC128_NONCE_MAX)
 717                 type = BCH_EXTENT_ENTRY_crc128;
 718         else
 719                 BUG();
 720
 721         bch2_extent_crc_pack(crc, new, type);
 722
 723         k->k.u64s += extent_entry_u64s(ptrs.end);
 724
 725         EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
 726 }
 727
 728 /* Generic code for keys with pointers: */
 729
 730 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
 731 {
 732         return bch2_bkey_devs(k).nr;
 733 }
 734
 735 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
 736 {
 737         return k.k->type == KEY_TYPE_reservation
 738                 ? bkey_s_c_to_reservation(k).v->nr_replicas
 739                 : bch2_bkey_dirty_devs(k).nr;
 740 }
 741
 742 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
 743 {
 744         unsigned ret = 0;
 745
 746         if (k.k->type == KEY_TYPE_reservation) {
 747                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
 748         } else {
 749                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 750                 const union bch_extent_entry *entry;
 751                 struct extent_ptr_decoded p;
 752
 753                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 754                         ret += !p.ptr.cached && !crc_is_compressed(p.crc);
 755         }
 756
 757         return ret;
 758 }
 759
 760 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
 761 {
 762         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 763         const union bch_extent_entry *entry;
 764         struct extent_ptr_decoded p;
 765         unsigned ret = 0;
 766
 767         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 768                 if (!p.ptr.cached && crc_is_compressed(p.crc))
 769                         ret += p.crc.compressed_size;
 770
 771         return ret;
 772 }
 773
 774 bool bch2_bkey_is_incompressible(struct bkey_s_c k)
 775 {
 776         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 777         const union bch_extent_entry *entry;
 778         struct bch_extent_crc_unpacked crc;
 779
 780         bkey_for_each_crc(k.k, ptrs, crc, entry)
 781                 if (crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 782                         return true;
 783         return false;
 784 }
 785
 786 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 787 {
 788         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 789         const union bch_extent_entry *entry;
 790         struct extent_ptr_decoded p = { 0 };
 791         unsigned replicas = 0;
 792
 793         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 794                 if (p.ptr.cached)
 795                         continue;
 796
 797                 if (p.has_ec)
 798                         replicas += p.ec.redundancy;
 799
 800                 replicas++;
 801
 802         }
 803
 804         return replicas;
 805 }
 806
 807 static inline unsigned __extent_ptr_durability(struct bch_dev *ca, struct extent_ptr_decoded *p)
 808 {
 809         if (p->ptr.cached)
 810                 return 0;
 811
 812         return p->has_ec
 813                 ? p->ec.redundancy + 1
 814                 : ca->mi.durability;
 815 }
 816
 817 unsigned bch2_extent_ptr_desired_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 818 {
 819         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 820
 821         return ca ? __extent_ptr_durability(ca, p) : 0;
 822 }
 823
 824 unsigned bch2_extent_ptr_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 825 {
 826         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 827
 828         if (!ca || ca->mi.state == BCH_MEMBER_STATE_failed)
 829                 return 0;
 830
 831         return __extent_ptr_durability(ca, p);
 832 }
 833
 834 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
 835 {
 836         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 837         const union bch_extent_entry *entry;
 838         struct extent_ptr_decoded p;
 839         unsigned durability = 0;
 840
 841         guard(rcu)();
 842         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 843                 durability += bch2_extent_ptr_durability(c, &p);
 844         return durability;
 845 }
 846
 847 static unsigned bch2_bkey_durability_safe(struct bch_fs *c, struct bkey_s_c k)
 848 {
 849         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 850         const union bch_extent_entry *entry;
 851         struct extent_ptr_decoded p;
 852         unsigned durability = 0;
 853
 854         guard(rcu)();
 855         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 856                 if (p.ptr.dev < c->sb.nr_devices && c->devs[p.ptr.dev])
 857                         durability += bch2_extent_ptr_durability(c, &p);
 858         return durability;
 859 }
 860
 861 void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry)
 862 {
 863         union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
 864         union bch_extent_entry *next = extent_entry_next(entry);
 865
 866         memmove_u64s(entry, next, (u64 *) end - (u64 *) next);
 867         k->k.u64s -= extent_entry_u64s(entry);
 868 }
 869
 870 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
 871                                     struct extent_ptr_decoded *p)
 872 {
 873         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 874         struct bch_extent_crc_unpacked crc =
 875                 bch2_extent_crc_unpack(&k->k, NULL);
 876         union bch_extent_entry *pos;
 877
 878         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 879                 pos = ptrs.start;
 880                 goto found;
 881         }
 882
 883         bkey_for_each_crc(&k->k, ptrs, crc, pos)
 884                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 885                         pos = extent_entry_next(pos);
 886                         goto found;
 887                 }
 888
 889         bch2_extent_crc_append(k, p->crc);
 890         pos = bkey_val_end(bkey_i_to_s(k));
 891 found:
 892         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 893         __extent_entry_insert(k, pos, to_entry(&p->ptr));
 894
 895         if (p->has_ec) {
 896                 p->ec.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
 897                 __extent_entry_insert(k, pos, to_entry(&p->ec));
 898         }
 899 }
 900
 901 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 902                                           union bch_extent_entry *entry)
 903 {
 904         union bch_extent_entry *i = ptrs.start;
 905
 906         if (i == entry)
 907                 return NULL;
 908
 909         while (extent_entry_next(i) != entry)
 910                 i = extent_entry_next(i);
 911         return i;
 912 }
 913
 914 /*
 915  * Returns pointer to the next entry after the one being dropped:
 916  */
 917 void bch2_bkey_drop_ptr_noerror(struct bkey_s k, struct bch_extent_ptr *ptr)
 918 {
 919         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 920         union bch_extent_entry *entry = to_entry(ptr), *next;
 921         bool drop_crc = true;
 922
 923         if (k.k->type == KEY_TYPE_stripe) {
 924                 ptr->dev = BCH_SB_MEMBER_INVALID;
 925                 return;
 926         }
 927
 928         EBUG_ON(ptr < &ptrs.start->ptr ||
 929                 ptr >= &ptrs.end->ptr);
 930         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 931
 932         for (next = extent_entry_next(entry);
 933              next != ptrs.end;
 934              next = extent_entry_next(next)) {
 935                 if (extent_entry_is_crc(next)) {
 936                         break;
 937                 } else if (extent_entry_is_ptr(next)) {
 938                         drop_crc = false;
 939                         break;
 940                 }
 941         }
 942
 943         extent_entry_drop(k, entry);
 944
 945         while ((entry = extent_entry_prev(ptrs, entry))) {
 946                 if (extent_entry_is_ptr(entry))
 947                         break;
 948
 949                 if ((extent_entry_is_crc(entry) && drop_crc) ||
 950                     extent_entry_is_stripe_ptr(entry))
 951                         extent_entry_drop(k, entry);
 952         }
 953 }
 954
 955 void bch2_bkey_drop_ptr(struct bkey_s k, struct bch_extent_ptr *ptr)
 956 {
 957         if (k.k->type != KEY_TYPE_stripe) {
 958                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k.s_c);
 959                 const union bch_extent_entry *entry;
 960                 struct extent_ptr_decoded p;
 961
 962                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 963                         if (p.ptr.dev == ptr->dev && p.has_ec) {
 964                                 ptr->dev = BCH_SB_MEMBER_INVALID;
 965                                 return;
 966                         }
 967         }
 968
 969         bool have_dirty = bch2_bkey_dirty_devs(k.s_c).nr;
 970
 971         bch2_bkey_drop_ptr_noerror(k, ptr);
 972
 973         /*
 974          * If we deleted all the dirty pointers and there's still cached
 975          * pointers, we could set the cached pointers to dirty if they're not
 976          * stale - but to do that correctly we'd need to grab an open_bucket
 977          * reference so that we don't race with bucket reuse:
 978          */
 979         if (have_dirty &&
 980             !bch2_bkey_dirty_devs(k.s_c).nr) {
 981                 k.k->type = KEY_TYPE_error;
 982                 set_bkey_val_u64s(k.k, 0);
 983         } else if (!bch2_bkey_nr_ptrs(k.s_c)) {
 984                 k.k->type = KEY_TYPE_deleted;
 985                 set_bkey_val_u64s(k.k, 0);
 986         }
 987 }
 988
 989 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
 990 {
 991         bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
 992 }
 993
 994 void bch2_bkey_drop_device_noerror(struct bkey_s k, unsigned dev)
 995 {
 996         bch2_bkey_drop_ptrs_noerror(k, ptr, ptr->dev == dev);
 997 }
 998
 999 const struct bch_extent_ptr *bch2_bkey_has_device_c(struct bkey_s_c k, unsigned dev)
1000 {
1001         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1002
1003         bkey_for_each_ptr(ptrs, ptr)
1004                 if (ptr->dev == dev)
1005                         return ptr;
1006
1007         return NULL;
1008 }
1009
1010 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
1011 {
1012         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1013         struct bch_dev *ca;
1014
1015         guard(rcu)();
1016         bkey_for_each_ptr(ptrs, ptr)
1017                 if (bch2_dev_in_target(c, ptr->dev, target) &&
1018                     (ca = bch2_dev_rcu(c, ptr->dev)) &&
1019                     (!ptr->cached ||
1020                      !dev_ptr_stale_rcu(ca, ptr)))
1021                         return true;
1022
1023         return false;
1024 }
1025
1026 bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
1027                            struct bch_extent_ptr m, u64 offset)
1028 {
1029         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1030         const union bch_extent_entry *entry;
1031         struct extent_ptr_decoded p;
1032
1033         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
1034                 if (p.ptr.dev   == m.dev &&
1035                     p.ptr.gen   == m.gen &&
1036                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
1037                     (s64) m.offset  - offset)
1038                         return true;
1039
1040         return false;
1041 }
1042
1043 /*
1044  * Returns true if two extents refer to the same data:
1045  */
1046 bool bch2_extents_match(struct bkey_s_c k1, struct bkey_s_c k2)
1047 {
1048         if (k1.k->type != k2.k->type)
1049                 return false;
1050
1051         if (bkey_extent_is_direct_data(k1.k)) {
1052                 struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(k1);
1053                 struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(k2);
1054                 const union bch_extent_entry *entry1, *entry2;
1055                 struct extent_ptr_decoded p1, p2;
1056
1057                 if (bkey_extent_is_unwritten(k1) != bkey_extent_is_unwritten(k2))
1058                         return false;
1059
1060                 bkey_for_each_ptr_decode(k1.k, ptrs1, p1, entry1)
1061                         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
1062                                 if (p1.ptr.dev          == p2.ptr.dev &&
1063                                     p1.ptr.gen          == p2.ptr.gen &&
1064
1065                                     /*
1066                                      * This checks that the two pointers point
1067                                      * to the same region on disk - adjusting
1068                                      * for the difference in where the extents
1069                                      * start, since one may have been trimmed:
1070                                      */
1071                                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
1072                                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k) &&
1073
1074                                     /*
1075                                      * This additionally checks that the
1076                                      * extents overlap on disk, since the
1077                                      * previous check may trigger spuriously
1078                                      * when one extent is immediately partially
1079                                      * overwritten with another extent (so that
1080                                      * on disk they are adjacent) and
1081                                      * compression is in use:
1082                                      */
1083                                     ((p1.ptr.offset >= p2.ptr.offset &&
1084                                       p1.ptr.offset  < p2.ptr.offset + p2.crc.compressed_size) ||
1085                                      (p2.ptr.offset >= p1.ptr.offset &&
1086                                       p2.ptr.offset  < p1.ptr.offset + p1.crc.compressed_size)))
1087                                         return true;
1088
1089                 return false;
1090         } else {
1091                 /* KEY_TYPE_deleted, etc. */
1092                 return true;
1093         }
1094 }
1095
1096 struct bch_extent_ptr *
1097 bch2_extent_has_ptr(struct bkey_s_c k1, struct extent_ptr_decoded p1, struct bkey_s k2)
1098 {
1099         struct bkey_ptrs ptrs2 = bch2_bkey_ptrs(k2);
1100         union bch_extent_entry *entry2;
1101         struct extent_ptr_decoded p2;
1102
1103         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
1104                 if (p1.ptr.dev          == p2.ptr.dev &&
1105                     p1.ptr.gen          == p2.ptr.gen &&
1106                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
1107                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
1108                         return &entry2->ptr;
1109
1110         return NULL;
1111 }
1112
1113 static bool want_cached_ptr(struct bch_fs *c, struct bch_io_opts *opts,
1114                             struct bch_extent_ptr *ptr)
1115 {
1116         unsigned target = opts->promote_target ?: opts->foreground_target;
1117
1118         if (target && !bch2_dev_in_target(c, ptr->dev, target))
1119                 return false;
1120
1121         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
1122
1123         return ca && bch2_dev_is_healthy(ca) && !dev_ptr_stale_rcu(ca, ptr);
1124 }
1125
1126 void bch2_extent_ptr_set_cached(struct bch_fs *c,
1127                                 struct bch_io_opts *opts,
1128                                 struct bkey_s k,
1129                                 struct bch_extent_ptr *ptr)
1130 {
1131         struct bkey_ptrs ptrs;
1132         union bch_extent_entry *entry;
1133         struct extent_ptr_decoded p;
1134         bool have_cached_ptr;
1135         unsigned drop_dev = ptr->dev;
1136
1137         guard(rcu)();
1138 restart_drop_ptrs:
1139         ptrs = bch2_bkey_ptrs(k);
1140         have_cached_ptr = false;
1141
1142         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1143                 /*
1144                  * Check if it's erasure coded - stripes can't contain cached
1145                  * data. Possibly something we can fix in the future?
1146                  */
1147                 if (&entry->ptr == ptr && p.has_ec)
1148                         goto drop;
1149
1150                 if (p.ptr.cached) {
1151                         if (have_cached_ptr || !want_cached_ptr(c, opts, &p.ptr)) {
1152                                 bch2_bkey_drop_ptr_noerror(k, &entry->ptr);
1153                                 ptr = NULL;
1154                                 goto restart_drop_ptrs;
1155                         }
1156
1157                         have_cached_ptr = true;
1158                 }
1159         }
1160
1161         if (!ptr)
1162                 bkey_for_each_ptr(ptrs, ptr2)
1163                         if (ptr2->dev == drop_dev)
1164                                 ptr = ptr2;
1165
1166         if (have_cached_ptr || !want_cached_ptr(c, opts, ptr))
1167                 goto drop;
1168
1169         ptr->cached = true;
1170         return;
1171 drop:
1172         bch2_bkey_drop_ptr_noerror(k, ptr);
1173 }
1174
1175 /*
1176  * bch2_extent_normalize - clean up an extent, dropping stale pointers etc.
1177  *
1178  * Returns true if @k should be dropped entirely
1179  *
1180  * For existing keys, only called when btree nodes are being rewritten, not when
1181  * they're merely being compacted/resorted in memory.
1182  */
1183 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1184 {
1185         struct bch_dev *ca;
1186
1187         guard(rcu)();
1188         bch2_bkey_drop_ptrs(k, ptr,
1189                 ptr->cached &&
1190                 (!(ca = bch2_dev_rcu(c, ptr->dev)) ||
1191                  dev_ptr_stale_rcu(ca, ptr) > 0));
1192
1193         return bkey_deleted(k.k);
1194 }
1195
1196 /*
1197  * bch2_extent_normalize_by_opts - clean up an extent, dropping stale pointers etc.
1198  *
1199  * Like bch2_extent_normalize(), but also only keeps a single cached pointer on
1200  * the promote target.
1201  */
1202 bool bch2_extent_normalize_by_opts(struct bch_fs *c,
1203                                    struct bch_io_opts *opts,
1204                                    struct bkey_s k)
1205 {
1206         struct bkey_ptrs ptrs;
1207         bool have_cached_ptr;
1208
1209         guard(rcu)();
1210 restart_drop_ptrs:
1211         ptrs = bch2_bkey_ptrs(k);
1212         have_cached_ptr = false;
1213
1214         bkey_for_each_ptr(ptrs, ptr)
1215                 if (ptr->cached) {
1216                         if (have_cached_ptr || !want_cached_ptr(c, opts, ptr)) {
1217                                 bch2_bkey_drop_ptr(k, ptr);
1218                                 goto restart_drop_ptrs;
1219                         }
1220                         have_cached_ptr = true;
1221                 }
1222
1223         return bkey_deleted(k.k);
1224 }
1225
1226 void bch2_extent_ptr_to_text(struct printbuf *out, struct bch_fs *c, const struct bch_extent_ptr *ptr)
1227 {
1228         out->atomic++;
1229         guard(rcu)();
1230         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
1231         if (!ca) {
1232                 prt_printf(out, "ptr: %u:%llu gen %u%s", ptr->dev,
1233                            (u64) ptr->offset, ptr->gen,
1234                            ptr->cached ? " cached" : "");
1235         } else {
1236                 u32 offset;
1237                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
1238
1239                 prt_printf(out, "ptr: %u:%llu:%u gen %u",
1240                            ptr->dev, b, offset, ptr->gen);
1241                 if (ca->mi.durability != 1)
1242                         prt_printf(out, " d=%u", ca->mi.durability);
1243                 if (ptr->cached)
1244                         prt_str(out, " cached");
1245                 if (ptr->unwritten)
1246                         prt_str(out, " unwritten");
1247                 int stale = dev_ptr_stale_rcu(ca, ptr);
1248                 if (stale > 0)
1249                         prt_printf(out, " stale");
1250                 else if (stale)
1251                         prt_printf(out, " invalid");
1252         }
1253         --out->atomic;
1254 }
1255
1256 void bch2_extent_crc_unpacked_to_text(struct printbuf *out, struct bch_extent_crc_unpacked *crc)
1257 {
1258         prt_printf(out, "crc: c_size %u size %u offset %u nonce %u csum ",
1259                    crc->compressed_size,
1260                    crc->uncompressed_size,
1261                    crc->offset, crc->nonce);
1262         bch2_prt_csum_type(out, crc->csum_type);
1263         prt_printf(out, " %0llx:%0llx ", crc->csum.hi, crc->csum.lo);
1264         prt_str(out, " compress ");
1265         bch2_prt_compression_type(out, crc->compression_type);
1266 }
1267
1268 static void bch2_extent_rebalance_to_text(struct printbuf *out, struct bch_fs *c,
1269                                           const struct bch_extent_rebalance *r)
1270 {
1271         prt_str(out, "rebalance:");
1272
1273         prt_printf(out, " replicas=%u", r->data_replicas);
1274         if (r->data_replicas_from_inode)
1275                 prt_str(out, " (inode)");
1276
1277         prt_str(out, " checksum=");
1278         bch2_prt_csum_opt(out, r->data_checksum);
1279         if (r->data_checksum_from_inode)
1280                 prt_str(out, " (inode)");
1281
1282         if (r->background_compression || r->background_compression_from_inode) {
1283                 prt_str(out, " background_compression=");
1284                 bch2_compression_opt_to_text(out, r->background_compression);
1285
1286                 if (r->background_compression_from_inode)
1287                         prt_str(out, " (inode)");
1288         }
1289
1290         if (r->background_target || r->background_target_from_inode) {
1291                 prt_str(out, " background_target=");
1292                 if (c)
1293                         bch2_target_to_text(out, c, r->background_target);
1294                 else
1295                         prt_printf(out, "%u", r->background_target);
1296
1297                 if (r->background_target_from_inode)
1298                         prt_str(out, " (inode)");
1299         }
1300
1301         if (r->promote_target || r->promote_target_from_inode) {
1302                 prt_str(out, " promote_target=");
1303                 if (c)
1304                         bch2_target_to_text(out, c, r->promote_target);
1305                 else
1306                         prt_printf(out, "%u", r->promote_target);
1307
1308                 if (r->promote_target_from_inode)
1309                         prt_str(out, " (inode)");
1310         }
1311
1312         if (r->erasure_code || r->erasure_code_from_inode) {
1313                 prt_printf(out, " ec=%u", r->erasure_code);
1314                 if (r->erasure_code_from_inode)
1315                         prt_str(out, " (inode)");
1316         }
1317 }
1318
1319 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1320                             struct bkey_s_c k)
1321 {
1322         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1323         const union bch_extent_entry *entry;
1324         bool first = true;
1325
1326         if (c)
1327                 prt_printf(out, "durability: %u ", bch2_bkey_durability_safe(c, k));
1328
1329         bkey_extent_entry_for_each(ptrs, entry) {
1330                 if (!first)
1331                         prt_printf(out, " ");
1332
1333                 switch (__extent_entry_type(entry)) {
1334                 case BCH_EXTENT_ENTRY_ptr:
1335                         bch2_extent_ptr_to_text(out, c, entry_to_ptr(entry));
1336                         break;
1337
1338                 case BCH_EXTENT_ENTRY_crc32:
1339                 case BCH_EXTENT_ENTRY_crc64:
1340                 case BCH_EXTENT_ENTRY_crc128: {
1341                         struct bch_extent_crc_unpacked crc =
1342                                 bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1343
1344                         bch2_extent_crc_unpacked_to_text(out, &crc);
1345                         break;
1346                 }
1347                 case BCH_EXTENT_ENTRY_stripe_ptr: {
1348                         const struct bch_extent_stripe_ptr *ec = &entry->stripe_ptr;
1349
1350                         prt_printf(out, "ec: idx %llu block %u",
1351                                (u64) ec->idx, ec->block);
1352                         break;
1353                 }
1354                 case BCH_EXTENT_ENTRY_rebalance:
1355                         bch2_extent_rebalance_to_text(out, c, &entry->rebalance);
1356                         break;
1357
1358                 case BCH_EXTENT_ENTRY_flags:
1359                         prt_bitflags(out, bch2_extent_flags_strs, entry->flags.flags);
1360                         break;
1361
1362                 default:
1363                         prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
1364                         return;
1365                 }
1366
1367                 first = false;
1368         }
1369 }
1370
1371 static int extent_ptr_validate(struct bch_fs *c,
1372                                struct bkey_s_c k,
1373                                struct bkey_validate_context from,
1374                                const struct bch_extent_ptr *ptr,
1375                                unsigned size_ondisk,
1376                                bool metadata)
1377 {
1378         int ret = 0;
1379
1380         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1381         bkey_for_each_ptr(ptrs, ptr2)
1382                 bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
1383                                  c, ptr_to_duplicate_device,
1384                                  "multiple pointers to same device (%u)", ptr->dev);
1385
1386         /* bad pointers are repaired by check_fix_ptrs(): */
1387         rcu_read_lock();
1388         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
1389         if (!ca) {
1390                 rcu_read_unlock();
1391                 return 0;
1392         }
1393         u32 bucket_offset;
1394         u64 bucket = sector_to_bucket_and_offset(ca, ptr->offset, &bucket_offset);
1395         unsigned first_bucket   = ca->mi.first_bucket;
1396         u64 nbuckets            = ca->mi.nbuckets;
1397         unsigned bucket_size    = ca->mi.bucket_size;
1398         rcu_read_unlock();
1399
1400         bkey_fsck_err_on(bucket >= nbuckets,
1401                          c, ptr_after_last_bucket,
1402                          "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
1403         bkey_fsck_err_on(bucket < first_bucket,
1404                          c, ptr_before_first_bucket,
1405                          "pointer before first bucket (%llu < %u)", bucket, first_bucket);
1406         bkey_fsck_err_on(bucket_offset + size_ondisk > bucket_size,
1407                          c, ptr_spans_multiple_buckets,
1408                          "pointer spans multiple buckets (%u + %u > %u)",
1409                        bucket_offset, size_ondisk, bucket_size);
1410 fsck_err:
1411         return ret;
1412 }
1413
1414 int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
1415                             struct bkey_validate_context from)
1416 {
1417         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1418         const union bch_extent_entry *entry;
1419         struct bch_extent_crc_unpacked crc;
1420         unsigned size_ondisk = k.k->size;
1421         unsigned nonce = UINT_MAX;
1422         unsigned nr_ptrs = 0;
1423         bool have_written = false, have_unwritten = false, have_ec = false, crc_since_last_ptr = false;
1424         int ret = 0;
1425
1426         if (bkey_is_btree_ptr(k.k))
1427                 size_ondisk = btree_sectors(c);
1428
1429         bkey_extent_entry_for_each(ptrs, entry) {
1430                 bkey_fsck_err_on(__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX,
1431                                  c, extent_ptrs_invalid_entry,
1432                                  "invalid extent entry type (got %u, max %u)",
1433                                  __extent_entry_type(entry), BCH_EXTENT_ENTRY_MAX);
1434
1435                 bkey_fsck_err_on(bkey_is_btree_ptr(k.k) &&
1436                                  !extent_entry_is_ptr(entry),
1437                                  c, btree_ptr_has_non_ptr,
1438                                  "has non ptr field");
1439
1440                 switch (extent_entry_type(entry)) {
1441                 case BCH_EXTENT_ENTRY_ptr:
1442                         ret = extent_ptr_validate(c, k, from, &entry->ptr, size_ondisk, false);
1443                         if (ret)
1444                                 return ret;
1445
1446                         bkey_fsck_err_on(entry->ptr.cached && have_ec,
1447                                          c, ptr_cached_and_erasure_coded,
1448                                          "cached, erasure coded ptr");
1449
1450                         if (!entry->ptr.unwritten)
1451                                 have_written = true;
1452                         else
1453                                 have_unwritten = true;
1454
1455                         have_ec = false;
1456                         crc_since_last_ptr = false;
1457                         nr_ptrs++;
1458                         break;
1459                 case BCH_EXTENT_ENTRY_crc32:
1460                 case BCH_EXTENT_ENTRY_crc64:
1461                 case BCH_EXTENT_ENTRY_crc128:
1462                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1463
1464                         bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type),
1465                                          c, ptr_crc_csum_type_unknown,
1466                                          "invalid checksum type");
1467                         bkey_fsck_err_on(crc.compression_type >= BCH_COMPRESSION_TYPE_NR,
1468                                          c, ptr_crc_compression_type_unknown,
1469                                          "invalid compression type");
1470
1471                         bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size,
1472                                          c, ptr_crc_uncompressed_size_too_small,
1473                                          "checksum offset + key size > uncompressed size");
1474                         bkey_fsck_err_on(crc_is_encoded(crc) &&
1475                                          (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
1476                                          (from.flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)),
1477                                          c, ptr_crc_uncompressed_size_too_big,
1478                                          "too large encoded extent");
1479                         bkey_fsck_err_on(!crc_is_compressed(crc) &&
1480                                          crc.compressed_size != crc.uncompressed_size,
1481                                          c, ptr_crc_uncompressed_size_mismatch,
1482                                          "not compressed but compressed != uncompressed size");
1483
1484                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1485                                 if (nonce == UINT_MAX)
1486                                         nonce = crc.offset + crc.nonce;
1487                                 else if (nonce != crc.offset + crc.nonce)
1488                                         bkey_fsck_err(c, ptr_crc_nonce_mismatch,
1489                                                       "incorrect nonce");
1490                         }
1491
1492                         bkey_fsck_err_on(crc_since_last_ptr,
1493                                          c, ptr_crc_redundant,
1494                                          "redundant crc entry");
1495                         crc_since_last_ptr = true;
1496
1497                         size_ondisk = crc.compressed_size;
1498                         break;
1499                 case BCH_EXTENT_ENTRY_stripe_ptr:
1500                         bkey_fsck_err_on(have_ec,
1501                                          c, ptr_stripe_redundant,
1502                                          "redundant stripe entry");
1503                         have_ec = true;
1504                         break;
1505                 case BCH_EXTENT_ENTRY_rebalance: {
1506                         /*
1507                          * this shouldn't be a fsck error, for forward
1508                          * compatibility; the rebalance code should just refetch
1509                          * the compression opt if it's unknown
1510                          */
1511 #if 0
1512                         const struct bch_extent_rebalance *r = &entry->rebalance;
1513
1514                         if (!bch2_compression_opt_valid(r->compression)) {
1515                                 struct bch_compression_opt opt = __bch2_compression_decode(r->compression);
1516                                 prt_printf(err, "invalid compression opt %u:%u",
1517                                            opt.type, opt.level);
1518                                 return bch_err_throw(c, invalid_bkey);
1519                         }
1520 #endif
1521                         break;
1522                 }
1523                 case BCH_EXTENT_ENTRY_flags:
1524                         bkey_fsck_err_on(entry != ptrs.start,
1525                                          c, extent_flags_not_at_start,
1526                                          "extent flags entry not at start");
1527                         break;
1528                 }
1529         }
1530
1531         bkey_fsck_err_on(!nr_ptrs,
1532                          c, extent_ptrs_no_ptrs,
1533                          "no ptrs");
1534         bkey_fsck_err_on(nr_ptrs > BCH_BKEY_PTRS_MAX,
1535                          c, extent_ptrs_too_many_ptrs,
1536                          "too many ptrs: %u > %u", nr_ptrs, BCH_BKEY_PTRS_MAX);
1537         bkey_fsck_err_on(have_written && have_unwritten,
1538                          c, extent_ptrs_written_and_unwritten,
1539                          "extent with unwritten and written ptrs");
1540         bkey_fsck_err_on(k.k->type != KEY_TYPE_extent && have_unwritten,
1541                          c, extent_ptrs_unwritten,
1542                          "has unwritten ptrs");
1543         bkey_fsck_err_on(crc_since_last_ptr,
1544                          c, extent_ptrs_redundant_crc,
1545                          "redundant crc entry");
1546         bkey_fsck_err_on(have_ec,
1547                          c, extent_ptrs_redundant_stripe,
1548                          "redundant stripe entry");
1549 fsck_err:
1550         return ret;
1551 }
1552
1553 void bch2_ptr_swab(struct bkey_s k)
1554 {
1555         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1556         union bch_extent_entry *entry;
1557         u64 *d;
1558
1559         for (d =  (u64 *) ptrs.start;
1560              d != (u64 *) ptrs.end;
1561              d++)
1562                 *d = swab64(*d);
1563
1564         for (entry = ptrs.start;
1565              entry < ptrs.end;
1566              entry = extent_entry_next(entry)) {
1567                 switch (__extent_entry_type(entry)) {
1568                 case BCH_EXTENT_ENTRY_ptr:
1569                         break;
1570                 case BCH_EXTENT_ENTRY_crc32:
1571                         entry->crc32.csum = swab32(entry->crc32.csum);
1572                         break;
1573                 case BCH_EXTENT_ENTRY_crc64:
1574                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
1575                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
1576                         break;
1577                 case BCH_EXTENT_ENTRY_crc128:
1578                         entry->crc128.csum.hi = (__force __le64)
1579                                 swab64((__force u64) entry->crc128.csum.hi);
1580                         entry->crc128.csum.lo = (__force __le64)
1581                                 swab64((__force u64) entry->crc128.csum.lo);
1582                         break;
1583                 case BCH_EXTENT_ENTRY_stripe_ptr:
1584                         break;
1585                 case BCH_EXTENT_ENTRY_rebalance:
1586                         break;
1587                 default:
1588                         /* Bad entry type: will be caught by validate() */
1589                         return;
1590                 }
1591         }
1592 }
1593
1594 int bch2_bkey_extent_flags_set(struct bch_fs *c, struct bkey_i *k, u64 flags)
1595 {
1596         int ret = bch2_request_incompat_feature(c, bcachefs_metadata_version_extent_flags);
1597         if (ret)
1598                 return ret;
1599
1600         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
1601
1602         if (ptrs.start != ptrs.end &&
1603             extent_entry_type(ptrs.start) == BCH_EXTENT_ENTRY_flags) {
1604                 ptrs.start->flags.flags = flags;
1605         } else {
1606                 struct bch_extent_flags f = {
1607                         .type   = BIT(BCH_EXTENT_ENTRY_flags),
1608                         .flags  = flags,
1609                 };
1610                 __extent_entry_insert(k, ptrs.start, (union bch_extent_entry *) &f);
1611         }
1612
1613         return 0;
1614 }
1615
1616 /* Generic extent code: */
1617
1618 int bch2_cut_front_s(struct bpos where, struct bkey_s k)
1619 {
1620         unsigned new_val_u64s = bkey_val_u64s(k.k);
1621         int val_u64s_delta;
1622         u64 sub;
1623
1624         if (bkey_le(where, bkey_start_pos(k.k)))
1625                 return 0;
1626
1627         EBUG_ON(bkey_gt(where, k.k->p));
1628
1629         sub = where.offset - bkey_start_offset(k.k);
1630
1631         k.k->size -= sub;
1632
1633         if (!k.k->size) {
1634                 k.k->type = KEY_TYPE_deleted;
1635                 new_val_u64s = 0;
1636         }
1637
1638         switch (k.k->type) {
1639         case KEY_TYPE_extent:
1640         case KEY_TYPE_reflink_v: {
1641                 struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1642                 union bch_extent_entry *entry;
1643                 bool seen_crc = false;
1644
1645                 bkey_extent_entry_for_each(ptrs, entry) {
1646                         switch (extent_entry_type(entry)) {
1647                         case BCH_EXTENT_ENTRY_ptr:
1648                                 if (!seen_crc)
1649                                         entry->ptr.offset += sub;
1650                                 break;
1651                         case BCH_EXTENT_ENTRY_crc32:
1652                                 entry->crc32.offset += sub;
1653                                 break;
1654                         case BCH_EXTENT_ENTRY_crc64:
1655                                 entry->crc64.offset += sub;
1656                                 break;
1657                         case BCH_EXTENT_ENTRY_crc128:
1658                                 entry->crc128.offset += sub;
1659                                 break;
1660                         case BCH_EXTENT_ENTRY_stripe_ptr:
1661                         case BCH_EXTENT_ENTRY_rebalance:
1662                         case BCH_EXTENT_ENTRY_flags:
1663                                 break;
1664                         }
1665
1666                         if (extent_entry_is_crc(entry))
1667                                 seen_crc = true;
1668                 }
1669
1670                 break;
1671         }
1672         case KEY_TYPE_reflink_p: {
1673                 struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
1674
1675                 SET_REFLINK_P_IDX(p.v, REFLINK_P_IDX(p.v) + sub);
1676                 break;
1677         }
1678         case KEY_TYPE_inline_data:
1679         case KEY_TYPE_indirect_inline_data: {
1680                 void *p = bkey_inline_data_p(k);
1681                 unsigned bytes = bkey_inline_data_bytes(k.k);
1682
1683                 sub = min_t(u64, sub << 9, bytes);
1684
1685                 memmove(p, p + sub, bytes - sub);
1686
1687                 new_val_u64s -= sub >> 3;
1688                 break;
1689         }
1690         }
1691
1692         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1693         BUG_ON(val_u64s_delta < 0);
1694
1695         set_bkey_val_u64s(k.k, new_val_u64s);
1696         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1697         return -val_u64s_delta;
1698 }
1699
1700 int bch2_cut_back_s(struct bpos where, struct bkey_s k)
1701 {
1702         unsigned new_val_u64s = bkey_val_u64s(k.k);
1703         int val_u64s_delta;
1704         u64 len = 0;
1705
1706         if (bkey_ge(where, k.k->p))
1707                 return 0;
1708
1709         EBUG_ON(bkey_lt(where, bkey_start_pos(k.k)));
1710
1711         len = where.offset - bkey_start_offset(k.k);
1712
1713         k.k->p.offset = where.offset;
1714         k.k->size = len;
1715
1716         if (!len) {
1717                 k.k->type = KEY_TYPE_deleted;
1718                 new_val_u64s = 0;
1719         }
1720
1721         switch (k.k->type) {
1722         case KEY_TYPE_inline_data:
1723         case KEY_TYPE_indirect_inline_data:
1724                 new_val_u64s = (bkey_inline_data_offset(k.k) +
1725                                 min(bkey_inline_data_bytes(k.k), k.k->size << 9)) >> 3;
1726                 break;
1727         }
1728
1729         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1730         BUG_ON(val_u64s_delta < 0);
1731
1732         set_bkey_val_u64s(k.k, new_val_u64s);
1733         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1734         return -val_u64s_delta;
1735 }