git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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_gc.h"
  12 #include "btree_update.h"
  13 #include "btree_update_interior.h"
  14 #include "buckets.h"
  15 #include "checksum.h"
  16 #include "debug.h"
  17 #include "dirent.h"
  18 #include "disk_groups.h"
  19 #include "error.h"
  20 #include "extents.h"
  21 #include "inode.h"
  22 #include "journal.h"
  23 #include "replicas.h"
  24 #include "super.h"
  25 #include "super-io.h"
  26 #include "util.h"
  27 #include "xattr.h"
  28
  29 #include <trace/events/bcachefs.h>
  30
  31 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
  32 {
  33         struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
  34         const struct bch_extent_ptr *ptr;
  35         unsigned nr_ptrs = 0;
  36
  37         bkey_for_each_ptr(p, ptr)
  38                 nr_ptrs++;
  39
  40         return nr_ptrs;
  41 }
  42
  43 unsigned bch2_bkey_nr_dirty_ptrs(struct bkey_s_c k)
  44 {
  45         unsigned nr_ptrs = 0;
  46
  47         switch (k.k->type) {
  48         case KEY_TYPE_btree_ptr:
  49         case KEY_TYPE_extent:
  50         case KEY_TYPE_reflink_v: {
  51                 struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
  52                 const struct bch_extent_ptr *ptr;
  53
  54                 bkey_for_each_ptr(p, ptr)
  55                         nr_ptrs += !ptr->cached;
  56                 BUG_ON(!nr_ptrs);
  57                 break;
  58         }
  59         case KEY_TYPE_reservation:
  60                 nr_ptrs = bkey_s_c_to_reservation(k).v->nr_replicas;
  61                 break;
  62         }
  63
  64         return nr_ptrs;
  65 }
  66
  67 static unsigned bch2_extent_ptr_durability(struct bch_fs *c,
  68                                            struct extent_ptr_decoded p)
  69 {
  70         unsigned i, durability = 0;
  71         struct bch_dev *ca;
  72
  73         if (p.ptr.cached)
  74                 return 0;
  75
  76         ca = bch_dev_bkey_exists(c, p.ptr.dev);
  77
  78         if (ca->mi.state != BCH_MEMBER_STATE_FAILED)
  79                 durability = max_t(unsigned, durability, ca->mi.durability);
  80
  81         for (i = 0; i < p.ec_nr; i++) {
  82                 struct stripe *s =
  83                         genradix_ptr(&c->stripes[0], p.idx);
  84
  85                 if (WARN_ON(!s))
  86                         continue;
  87
  88                 durability = max_t(unsigned, durability, s->nr_redundant);
  89         }
  90
  91         return durability;
  92 }
  93
  94 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
  95 {
  96         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  97         const union bch_extent_entry *entry;
  98         struct extent_ptr_decoded p;
  99         unsigned durability = 0;
 100
 101         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 102                 durability += bch2_extent_ptr_durability(c, p);
 103
 104         return durability;
 105 }
 106
 107 static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f,
 108                                                    unsigned dev)
 109 {
 110         struct bch_dev_io_failures *i;
 111
 112         for (i = f->devs; i < f->devs + f->nr; i++)
 113                 if (i->dev == dev)
 114                         return i;
 115
 116         return NULL;
 117 }
 118
 119 void bch2_mark_io_failure(struct bch_io_failures *failed,
 120                           struct extent_ptr_decoded *p)
 121 {
 122         struct bch_dev_io_failures *f = dev_io_failures(failed, p->ptr.dev);
 123
 124         if (!f) {
 125                 BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
 126
 127                 f = &failed->devs[failed->nr++];
 128                 f->dev          = p->ptr.dev;
 129                 f->idx          = p->idx;
 130                 f->nr_failed    = 1;
 131                 f->nr_retries   = 0;
 132         } else if (p->idx != f->idx) {
 133                 f->idx          = p->idx;
 134                 f->nr_failed    = 1;
 135                 f->nr_retries   = 0;
 136         } else {
 137                 f->nr_failed++;
 138         }
 139 }
 140
 141 /*
 142  * returns true if p1 is better than p2:
 143  */
 144 static inline bool ptr_better(struct bch_fs *c,
 145                               const struct extent_ptr_decoded p1,
 146                               const struct extent_ptr_decoded p2)
 147 {
 148         if (likely(!p1.idx && !p2.idx)) {
 149                 struct bch_dev *dev1 = bch_dev_bkey_exists(c, p1.ptr.dev);
 150                 struct bch_dev *dev2 = bch_dev_bkey_exists(c, p2.ptr.dev);
 151
 152                 u64 l1 = atomic64_read(&dev1->cur_latency[READ]);
 153                 u64 l2 = atomic64_read(&dev2->cur_latency[READ]);
 154
 155                 /* Pick at random, biased in favor of the faster device: */
 156
 157                 return bch2_rand_range(l1 + l2) > l1;
 158         }
 159
 160         if (force_reconstruct_read(c))
 161                 return p1.idx > p2.idx;
 162
 163         return p1.idx < p2.idx;
 164 }
 165
 166 /*
 167  * This picks a non-stale pointer, preferably from a device other than @avoid.
 168  * Avoid can be NULL, meaning pick any. If there are no non-stale pointers to
 169  * other devices, it will still pick a pointer from avoid.
 170  */
 171 int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 172                                struct bch_io_failures *failed,
 173                                struct extent_ptr_decoded *pick)
 174 {
 175         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 176         const union bch_extent_entry *entry;
 177         struct extent_ptr_decoded p;
 178         struct bch_dev_io_failures *f;
 179         struct bch_dev *ca;
 180         int ret = 0;
 181
 182         if (k.k->type == KEY_TYPE_error)
 183                 return -EIO;
 184
 185         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 186                 ca = bch_dev_bkey_exists(c, p.ptr.dev);
 187
 188                 /*
 189                  * If there are any dirty pointers it's an error if we can't
 190                  * read:
 191                  */
 192                 if (!ret && !p.ptr.cached)
 193                         ret = -EIO;
 194
 195                 if (p.ptr.cached && ptr_stale(ca, &p.ptr))
 196                         continue;
 197
 198                 f = failed ? dev_io_failures(failed, p.ptr.dev) : NULL;
 199                 if (f)
 200                         p.idx = f->nr_failed < f->nr_retries
 201                                 ? f->idx
 202                                 : f->idx + 1;
 203
 204                 if (!p.idx &&
 205                     !bch2_dev_is_readable(ca))
 206                         p.idx++;
 207
 208                 if (force_reconstruct_read(c) &&
 209                     !p.idx && p.ec_nr)
 210                         p.idx++;
 211
 212                 if (p.idx >= p.ec_nr + 1)
 213                         continue;
 214
 215                 if (ret > 0 && !ptr_better(c, p, *pick))
 216                         continue;
 217
 218                 *pick = p;
 219                 ret = 1;
 220         }
 221
 222         return ret;
 223 }
 224
 225 void bch2_bkey_append_ptr(struct bkey_i *k,
 226                           struct bch_extent_ptr ptr)
 227 {
 228         EBUG_ON(bch2_bkey_has_device(bkey_i_to_s_c(k), ptr.dev));
 229
 230         switch (k->k.type) {
 231         case KEY_TYPE_btree_ptr:
 232         case KEY_TYPE_extent:
 233                 EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
 234
 235                 ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 236
 237                 memcpy((void *) &k->v + bkey_val_bytes(&k->k),
 238                        &ptr,
 239                        sizeof(ptr));
 240                 k->u64s++;
 241                 break;
 242         default:
 243                 BUG();
 244         }
 245 }
 246
 247 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
 248 {
 249         struct bch_extent_ptr *ptr;
 250
 251         bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
 252 }
 253
 254 const struct bch_extent_ptr *
 255 bch2_bkey_has_device(struct bkey_s_c k, unsigned dev)
 256 {
 257         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 258         const struct bch_extent_ptr *ptr;
 259
 260         bkey_for_each_ptr(ptrs, ptr)
 261                 if (ptr->dev == dev)
 262                         return ptr;
 263
 264         return NULL;
 265 }
 266
 267 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
 268 {
 269         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 270         const struct bch_extent_ptr *ptr;
 271
 272         bkey_for_each_ptr(ptrs, ptr)
 273                 if (bch2_dev_in_target(c, ptr->dev, target) &&
 274                     (!ptr->cached ||
 275                      !ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr)))
 276                         return true;
 277
 278         return false;
 279 }
 280
 281 /* extent specific utility code */
 282
 283 const struct bch_extent_ptr *
 284 bch2_extent_has_device(struct bkey_s_c_extent e, unsigned dev)
 285 {
 286         const struct bch_extent_ptr *ptr;
 287
 288         extent_for_each_ptr(e, ptr)
 289                 if (ptr->dev == dev)
 290                         return ptr;
 291
 292         return NULL;
 293 }
 294
 295 const struct bch_extent_ptr *
 296 bch2_extent_has_group(struct bch_fs *c, struct bkey_s_c_extent e, unsigned group)
 297 {
 298         const struct bch_extent_ptr *ptr;
 299
 300         extent_for_each_ptr(e, ptr) {
 301                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 302
 303                 if (ca->mi.group &&
 304                     ca->mi.group - 1 == group)
 305                         return ptr;
 306         }
 307
 308         return NULL;
 309 }
 310
 311 unsigned bch2_extent_is_compressed(struct bkey_s_c k)
 312 {
 313         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 314         const union bch_extent_entry *entry;
 315         struct extent_ptr_decoded p;
 316         unsigned ret = 0;
 317
 318         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 319                 if (!p.ptr.cached &&
 320                     p.crc.compression_type != BCH_COMPRESSION_NONE)
 321                         ret += p.crc.compressed_size;
 322
 323         return ret;
 324 }
 325
 326 bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
 327                            struct bch_extent_ptr m, u64 offset)
 328 {
 329         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 330         const union bch_extent_entry *entry;
 331         struct extent_ptr_decoded p;
 332
 333         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 334                 if (p.ptr.dev   == m.dev &&
 335                     p.ptr.gen   == m.gen &&
 336                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
 337                     (s64) m.offset  - offset)
 338                         return true;
 339
 340         return false;
 341 }
 342
 343 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 344                                           union bch_extent_entry *entry)
 345 {
 346         union bch_extent_entry *i = ptrs.start;
 347
 348         if (i == entry)
 349                 return NULL;
 350
 351         while (extent_entry_next(i) != entry)
 352                 i = extent_entry_next(i);
 353         return i;
 354 }
 355
 356 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
 357                                            struct bch_extent_ptr *ptr)
 358 {
 359         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 360         union bch_extent_entry *dst, *src, *prev;
 361         bool drop_crc = true;
 362
 363         EBUG_ON(ptr < &ptrs.start->ptr ||
 364                 ptr >= &ptrs.end->ptr);
 365         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 366
 367         src = extent_entry_next(to_entry(ptr));
 368         if (src != ptrs.end &&
 369             !extent_entry_is_crc(src))
 370                 drop_crc = false;
 371
 372         dst = to_entry(ptr);
 373         while ((prev = extent_entry_prev(ptrs, dst))) {
 374                 if (extent_entry_is_ptr(prev))
 375                         break;
 376
 377                 if (extent_entry_is_crc(prev)) {
 378                         if (drop_crc)
 379                                 dst = prev;
 380                         break;
 381                 }
 382
 383                 dst = prev;
 384         }
 385
 386         memmove_u64s_down(dst, src,
 387                           (u64 *) ptrs.end - (u64 *) src);
 388         k.k->u64s -= (u64 *) src - (u64 *) dst;
 389
 390         return dst;
 391 }
 392
 393 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 394                                   struct bch_extent_crc_unpacked n)
 395 {
 396         return !u.compression_type &&
 397                 u.csum_type &&
 398                 u.uncompressed_size > u.live_size &&
 399                 bch2_csum_type_is_encryption(u.csum_type) ==
 400                 bch2_csum_type_is_encryption(n.csum_type);
 401 }
 402
 403 bool bch2_can_narrow_extent_crcs(struct bkey_s_c k,
 404                                  struct bch_extent_crc_unpacked n)
 405 {
 406         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 407         struct bch_extent_crc_unpacked crc;
 408         const union bch_extent_entry *i;
 409
 410         if (!n.csum_type)
 411                 return false;
 412
 413         bkey_for_each_crc(k.k, ptrs, crc, i)
 414                 if (can_narrow_crc(crc, n))
 415                         return true;
 416
 417         return false;
 418 }
 419
 420 /*
 421  * We're writing another replica for this extent, so while we've got the data in
 422  * memory we'll be computing a new checksum for the currently live data.
 423  *
 424  * If there are other replicas we aren't moving, and they are checksummed but
 425  * not compressed, we can modify them to point to only the data that is
 426  * currently live (so that readers won't have to bounce) while we've got the
 427  * checksum we need:
 428  */
 429 bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 430 {
 431         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 432         struct bch_extent_crc_unpacked u;
 433         struct extent_ptr_decoded p;
 434         union bch_extent_entry *i;
 435         bool ret = false;
 436
 437         /* Find a checksum entry that covers only live data: */
 438         if (!n.csum_type) {
 439                 bkey_for_each_crc(&k->k, ptrs, u, i)
 440                         if (!u.compression_type &&
 441                             u.csum_type &&
 442                             u.live_size == u.uncompressed_size) {
 443                                 n = u;
 444                                 goto found;
 445                         }
 446                 return false;
 447         }
 448 found:
 449         BUG_ON(n.compression_type);
 450         BUG_ON(n.offset);
 451         BUG_ON(n.live_size != k->k.size);
 452
 453 restart_narrow_pointers:
 454         ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 455
 456         bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
 457                 if (can_narrow_crc(p.crc, n)) {
 458                         bch2_bkey_drop_ptr(bkey_i_to_s(k), &i->ptr);
 459                         p.ptr.offset += p.crc.offset;
 460                         p.crc = n;
 461                         bch2_extent_ptr_decoded_append(k, &p);
 462                         ret = true;
 463                         goto restart_narrow_pointers;
 464                 }
 465
 466         return ret;
 467 }
 468
 469 /* returns true if not equal */
 470 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 471                                          struct bch_extent_crc_unpacked r)
 472 {
 473         return (l.csum_type             != r.csum_type ||
 474                 l.compression_type      != r.compression_type ||
 475                 l.compressed_size       != r.compressed_size ||
 476                 l.uncompressed_size     != r.uncompressed_size ||
 477                 l.offset                != r.offset ||
 478                 l.live_size             != r.live_size ||
 479                 l.nonce                 != r.nonce ||
 480                 bch2_crc_cmp(l.csum, r.csum));
 481 }
 482
 483 void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
 484 {
 485         union bch_extent_entry *entry;
 486         u64 *d = (u64 *) bkeyp_val(f, k);
 487         unsigned i;
 488
 489         for (i = 0; i < bkeyp_val_u64s(f, k); i++)
 490                 d[i] = swab64(d[i]);
 491
 492         for (entry = (union bch_extent_entry *) d;
 493              entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
 494              entry = extent_entry_next(entry)) {
 495                 switch (extent_entry_type(entry)) {
 496                 case BCH_EXTENT_ENTRY_ptr:
 497                         break;
 498                 case BCH_EXTENT_ENTRY_crc32:
 499                         entry->crc32.csum = swab32(entry->crc32.csum);
 500                         break;
 501                 case BCH_EXTENT_ENTRY_crc64:
 502                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
 503                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
 504                         break;
 505                 case BCH_EXTENT_ENTRY_crc128:
 506                         entry->crc128.csum.hi = (__force __le64)
 507                                 swab64((__force u64) entry->crc128.csum.hi);
 508                         entry->crc128.csum.lo = (__force __le64)
 509                                 swab64((__force u64) entry->crc128.csum.lo);
 510                         break;
 511                 case BCH_EXTENT_ENTRY_stripe_ptr:
 512                         break;
 513                 }
 514         }
 515 }
 516
 517 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 518                             struct bkey_s_c k)
 519 {
 520         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 521         const union bch_extent_entry *entry;
 522         struct bch_extent_crc_unpacked crc;
 523         const struct bch_extent_ptr *ptr;
 524         const struct bch_extent_stripe_ptr *ec;
 525         struct bch_dev *ca;
 526         bool first = true;
 527
 528         bkey_extent_entry_for_each(ptrs, entry) {
 529                 if (!first)
 530                         pr_buf(out, " ");
 531
 532                 switch (__extent_entry_type(entry)) {
 533                 case BCH_EXTENT_ENTRY_ptr:
 534                         ptr = entry_to_ptr(entry);
 535                         ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
 536                                 ? bch_dev_bkey_exists(c, ptr->dev)
 537                                 : NULL;
 538
 539                         pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
 540                                (u64) ptr->offset, ptr->gen,
 541                                ptr->cached ? " cached" : "",
 542                                ca && ptr_stale(ca, ptr)
 543                                ? " stale" : "");
 544                         break;
 545                 case BCH_EXTENT_ENTRY_crc32:
 546                 case BCH_EXTENT_ENTRY_crc64:
 547                 case BCH_EXTENT_ENTRY_crc128:
 548                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 549
 550                         pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
 551                                crc.compressed_size,
 552                                crc.uncompressed_size,
 553                                crc.offset, crc.nonce,
 554                                crc.csum_type,
 555                                crc.compression_type);
 556                         break;
 557                 case BCH_EXTENT_ENTRY_stripe_ptr:
 558                         ec = &entry->stripe_ptr;
 559
 560                         pr_buf(out, "ec: idx %llu block %u",
 561                                (u64) ec->idx, ec->block);
 562                         break;
 563                 default:
 564                         pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
 565                         return;
 566                 }
 567
 568                 first = false;
 569         }
 570 }
 571
 572 static const char *extent_ptr_invalid(const struct bch_fs *c,
 573                                       struct bkey_s_c k,
 574                                       const struct bch_extent_ptr *ptr,
 575                                       unsigned size_ondisk,
 576                                       bool metadata)
 577 {
 578         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 579         const struct bch_extent_ptr *ptr2;
 580         struct bch_dev *ca;
 581
 582         if (!bch2_dev_exists2(c, ptr->dev))
 583                 return "pointer to invalid device";
 584
 585         ca = bch_dev_bkey_exists(c, ptr->dev);
 586         if (!ca)
 587                 return "pointer to invalid device";
 588
 589         bkey_for_each_ptr(ptrs, ptr2)
 590                 if (ptr != ptr2 && ptr->dev == ptr2->dev)
 591                         return "multiple pointers to same device";
 592
 593         if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
 594                 return "offset past end of device";
 595
 596         if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
 597                 return "offset before first bucket";
 598
 599         if (bucket_remainder(ca, ptr->offset) +
 600             size_ondisk > ca->mi.bucket_size)
 601                 return "spans multiple buckets";
 602
 603         return NULL;
 604 }
 605
 606 const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
 607 {
 608         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 609         const union bch_extent_entry *entry;
 610         struct bch_extent_crc_unpacked crc;
 611         unsigned size_ondisk = k.k->size;
 612         const char *reason;
 613         unsigned nonce = UINT_MAX;
 614
 615         if (k.k->type == KEY_TYPE_btree_ptr)
 616                 size_ondisk = c->opts.btree_node_size;
 617
 618         bkey_extent_entry_for_each(ptrs, entry) {
 619                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
 620                         return "invalid extent entry type";
 621
 622                 if (k.k->type == KEY_TYPE_btree_ptr &&
 623                     !extent_entry_is_ptr(entry))
 624                         return "has non ptr field";
 625
 626                 switch (extent_entry_type(entry)) {
 627                 case BCH_EXTENT_ENTRY_ptr:
 628                         reason = extent_ptr_invalid(c, k, &entry->ptr,
 629                                                     size_ondisk, false);
 630                         if (reason)
 631                                 return reason;
 632                         break;
 633                 case BCH_EXTENT_ENTRY_crc32:
 634                 case BCH_EXTENT_ENTRY_crc64:
 635                 case BCH_EXTENT_ENTRY_crc128:
 636                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 637
 638                         if (crc.offset + crc.live_size >
 639                             crc.uncompressed_size)
 640                                 return "checksum offset + key size > uncompressed size";
 641
 642                         size_ondisk = crc.compressed_size;
 643
 644                         if (!bch2_checksum_type_valid(c, crc.csum_type))
 645                                 return "invalid checksum type";
 646
 647                         if (crc.compression_type >= BCH_COMPRESSION_NR)
 648                                 return "invalid compression type";
 649
 650                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
 651                                 if (nonce == UINT_MAX)
 652                                         nonce = crc.offset + crc.nonce;
 653                                 else if (nonce != crc.offset + crc.nonce)
 654                                         return "incorrect nonce";
 655                         }
 656                         break;
 657                 case BCH_EXTENT_ENTRY_stripe_ptr:
 658                         break;
 659                 }
 660         }
 661
 662         return NULL;
 663 }
 664
 665 /* Btree ptrs */
 666
 667 const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
 668 {
 669         if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
 670                 return "value too big";
 671
 672         return bch2_bkey_ptrs_invalid(c, k);
 673 }
 674
 675 void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct bkey_s_c k)
 676 {
 677         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 678         const struct bch_extent_ptr *ptr;
 679         const char *err;
 680         char buf[160];
 681         struct bucket_mark mark;
 682         struct bch_dev *ca;
 683
 684         bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 685                        !bch2_bkey_replicas_marked(c, k, false), c,
 686                        "btree key bad (replicas not marked in superblock):\n%s",
 687                        (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 688
 689         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 690                 return;
 691
 692         bkey_for_each_ptr(ptrs, ptr) {
 693                 ca = bch_dev_bkey_exists(c, ptr->dev);
 694
 695                 mark = ptr_bucket_mark(ca, ptr);
 696
 697                 err = "stale";
 698                 if (gen_after(mark.gen, ptr->gen))
 699                         goto err;
 700
 701                 err = "inconsistent";
 702                 if (mark.data_type != BCH_DATA_BTREE ||
 703                     mark.dirty_sectors < c->opts.btree_node_size)
 704                         goto err;
 705         }
 706
 707         return;
 708 err:
 709         bch2_bkey_val_to_text(&PBUF(buf), c, k);
 710         bch2_fs_bug(c, "%s btree pointer %s: bucket %zi gen %i mark %08x",
 711                     err, buf, PTR_BUCKET_NR(ca, ptr),
 712                     mark.gen, (unsigned) mark.v.counter);
 713 }
 714
 715 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 716                             struct bkey_s_c k)
 717 {
 718         bch2_bkey_ptrs_to_text(out, c, k);
 719 }
 720
 721 /* Extents */
 722
 723 void __bch2_cut_front(struct bpos where, struct bkey_s k)
 724 {
 725         u64 sub;
 726
 727         if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
 728                 return;
 729
 730         EBUG_ON(bkey_cmp(where, k.k->p) > 0);
 731
 732         sub = where.offset - bkey_start_offset(k.k);
 733
 734         k.k->size -= sub;
 735
 736         if (!k.k->size)
 737                 k.k->type = KEY_TYPE_deleted;
 738
 739         switch (k.k->type) {
 740         case KEY_TYPE_deleted:
 741         case KEY_TYPE_discard:
 742         case KEY_TYPE_error:
 743         case KEY_TYPE_cookie:
 744                 break;
 745         case KEY_TYPE_extent:
 746         case KEY_TYPE_reflink_v: {
 747                 struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 748                 union bch_extent_entry *entry;
 749                 bool seen_crc = false;
 750
 751                 bkey_extent_entry_for_each(ptrs, entry) {
 752                         switch (extent_entry_type(entry)) {
 753                         case BCH_EXTENT_ENTRY_ptr:
 754                                 if (!seen_crc)
 755                                         entry->ptr.offset += sub;
 756                                 break;
 757                         case BCH_EXTENT_ENTRY_crc32:
 758                                 entry->crc32.offset += sub;
 759                                 break;
 760                         case BCH_EXTENT_ENTRY_crc64:
 761                                 entry->crc64.offset += sub;
 762                                 break;
 763                         case BCH_EXTENT_ENTRY_crc128:
 764                                 entry->crc128.offset += sub;
 765                                 break;
 766                         case BCH_EXTENT_ENTRY_stripe_ptr:
 767                                 break;
 768                         }
 769
 770                         if (extent_entry_is_crc(entry))
 771                                 seen_crc = true;
 772                 }
 773
 774                 break;
 775         }
 776         case KEY_TYPE_reflink_p: {
 777                 struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
 778
 779                 le64_add_cpu(&p.v->idx, sub);
 780                 break;
 781         }
 782         case KEY_TYPE_reservation:
 783                 break;
 784         default:
 785                 BUG();
 786         }
 787 }
 788
 789 bool bch2_cut_back(struct bpos where, struct bkey *k)
 790 {
 791         u64 len = 0;
 792
 793         if (bkey_cmp(where, k->p) >= 0)
 794                 return false;
 795
 796         EBUG_ON(bkey_cmp(where, bkey_start_pos(k)) < 0);
 797
 798         len = where.offset - bkey_start_offset(k);
 799
 800         k->p = where;
 801         k->size = len;
 802
 803         if (!len)
 804                 k->type = KEY_TYPE_deleted;
 805
 806         return true;
 807 }
 808
 809 static bool extent_i_save(struct btree *b, struct bkey_packed *dst,
 810                           struct bkey_i *src)
 811 {
 812         struct bkey_format *f = &b->format;
 813         struct bkey_i *dst_unpacked;
 814         struct bkey_packed tmp;
 815
 816         if ((dst_unpacked = packed_to_bkey(dst)))
 817                 dst_unpacked->k = src->k;
 818         else if (bch2_bkey_pack_key(&tmp, &src->k, f))
 819                 memcpy_u64s(dst, &tmp, f->key_u64s);
 820         else
 821                 return false;
 822
 823         memcpy_u64s(bkeyp_val(f, dst), &src->v, bkey_val_u64s(&src->k));
 824         return true;
 825 }
 826
 827 static bool bch2_extent_merge_inline(struct bch_fs *,
 828                                      struct btree_iter *,
 829                                      struct bkey_packed *,
 830                                      struct bkey_packed *,
 831                                      bool);
 832
 833 static void verify_extent_nonoverlapping(struct bch_fs *c,
 834                                          struct btree *b,
 835                                          struct btree_node_iter *_iter,
 836                                          struct bkey_i *insert)
 837 {
 838 #ifdef CONFIG_BCACHEFS_DEBUG
 839         struct btree_node_iter iter;
 840         struct bkey_packed *k;
 841         struct bkey uk;
 842
 843         if (!expensive_debug_checks(c))
 844                 return;
 845
 846         iter = *_iter;
 847         k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
 848         BUG_ON(k &&
 849                (uk = bkey_unpack_key(b, k),
 850                 bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
 851
 852         iter = *_iter;
 853         k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
 854 #if 0
 855         BUG_ON(k &&
 856                (uk = bkey_unpack_key(b, k),
 857                 bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
 858 #else
 859         if (k &&
 860             (uk = bkey_unpack_key(b, k),
 861              bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
 862                 char buf1[100];
 863                 char buf2[100];
 864
 865                 bch2_bkey_to_text(&PBUF(buf1), &insert->k);
 866                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 867
 868                 bch2_dump_btree_node(b);
 869                 panic("insert > next :\n"
 870                       "insert %s\n"
 871                       "next   %s\n",
 872                       buf1, buf2);
 873         }
 874 #endif
 875
 876 #endif
 877 }
 878
 879 static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 880                                struct bkey_i *insert)
 881 {
 882         struct btree_iter_level *l = &iter->l[0];
 883         struct btree_node_iter node_iter;
 884         struct bkey_packed *k;
 885
 886         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
 887
 888         EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
 889         verify_extent_nonoverlapping(c, l->b, &l->iter, insert);
 890
 891         if (debug_check_bkeys(c))
 892                 bch2_bkey_debugcheck(c, l->b, bkey_i_to_s_c(insert));
 893
 894         node_iter = l->iter;
 895         k = bch2_btree_node_iter_prev_filter(&node_iter, l->b, KEY_TYPE_discard);
 896         if (k && !bkey_written(l->b, k) &&
 897             bch2_extent_merge_inline(c, iter, k, bkey_to_packed(insert), true))
 898                 return;
 899
 900         node_iter = l->iter;
 901         k = bch2_btree_node_iter_peek_filter(&node_iter, l->b, KEY_TYPE_discard);
 902         if (k && !bkey_written(l->b, k) &&
 903             bch2_extent_merge_inline(c, iter, bkey_to_packed(insert), k, false))
 904                 return;
 905
 906         /*
 907          * may have skipped past some deleted extents greater than the insert
 908          * key, before we got to a non deleted extent and knew we could bail out
 909          * rewind the iterator a bit if necessary:
 910          */
 911         node_iter = l->iter;
 912         while ((k = bch2_btree_node_iter_prev_all(&node_iter, l->b)) &&
 913                bkey_cmp_left_packed(l->b, k, &insert->k.p) > 0)
 914                 l->iter = node_iter;
 915
 916         k = bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
 917
 918         bch2_bset_insert(l->b, &l->iter, k, insert, 0);
 919         bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
 920 }
 921
 922 static unsigned bch2_bkey_nr_alloc_ptrs(struct bkey_s_c k)
 923 {
 924         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 925         const union bch_extent_entry *entry;
 926         unsigned ret = 0;
 927
 928         bkey_extent_entry_for_each(ptrs, entry) {
 929                 switch (__extent_entry_type(entry)) {
 930                 case BCH_EXTENT_ENTRY_ptr:
 931                 case BCH_EXTENT_ENTRY_stripe_ptr:
 932                         ret++;
 933                 }
 934         }
 935
 936         return ret;
 937 }
 938
 939 static int count_iters_for_insert(struct btree_trans *trans,
 940                                   struct bkey_s_c k,
 941                                   unsigned offset,
 942                                   struct bpos *end,
 943                                   unsigned *nr_iters,
 944                                   unsigned max_iters,
 945                                   bool overwrite)
 946 {
 947         int ret = 0;
 948
 949         switch (k.k->type) {
 950         case KEY_TYPE_extent:
 951         case KEY_TYPE_reflink_v:
 952                 *nr_iters += bch2_bkey_nr_alloc_ptrs(k);
 953
 954                 if (*nr_iters >= max_iters) {
 955                         *end = bpos_min(*end, k.k->p);
 956                         return 0;
 957                 }
 958
 959                 break;
 960         case KEY_TYPE_reflink_p: {
 961                 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
 962                 u64 idx = le64_to_cpu(p.v->idx);
 963                 unsigned sectors = end->offset - bkey_start_offset(p.k);
 964                 struct btree_iter *iter;
 965                 struct bkey_s_c r_k;
 966
 967                 for_each_btree_key(trans, iter,
 968                                    BTREE_ID_REFLINK, POS(0, idx + offset),
 969                                    BTREE_ITER_SLOTS, r_k, ret) {
 970                         if (bkey_cmp(bkey_start_pos(r_k.k),
 971                                      POS(0, idx + sectors)) >= 0)
 972                                 break;
 973
 974                         *nr_iters += 1;
 975
 976                         if (overwrite &&
 977                             k.k->type == KEY_TYPE_reflink_v) {
 978                                 struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
 979
 980                                 if (le64_to_cpu(r.v->refcount) == 1)
 981                                         *nr_iters += bch2_bkey_nr_alloc_ptrs(k);
 982                         }
 983
 984                         /*
 985                          * if we're going to be deleting an entry from
 986                          * the reflink btree, need more iters...
 987                          */
 988
 989                         if (*nr_iters >= max_iters) {
 990                                 struct bpos pos = bkey_start_pos(k.k);
 991                                 pos.offset += r_k.k->p.offset - idx;
 992
 993                                 *end = bpos_min(*end, pos);
 994                                 break;
 995                         }
 996                 }
 997
 998                 bch2_trans_iter_put(trans, iter);
 999                 break;
1000         }
1001         }
1002
1003         return ret;
1004 }
1005
1006 int bch2_extent_atomic_end(struct btree_iter *iter,
1007                            struct bkey_i *insert,
1008                            struct bpos *end)
1009 {
1010         struct btree_trans *trans = iter->trans;
1011         struct btree *b = iter->l[0].b;
1012         struct btree_node_iter  node_iter = iter->l[0].iter;
1013         struct bkey_packed      *_k;
1014         unsigned                nr_iters =
1015                 bch2_bkey_nr_alloc_ptrs(bkey_i_to_s_c(insert));
1016         int ret = 0;
1017
1018         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
1019         BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0);
1020
1021         *end = bpos_min(insert->k.p, b->key.k.p);
1022
1023         ret = count_iters_for_insert(trans, bkey_i_to_s_c(insert),
1024                                      0, end, &nr_iters, 10, false);
1025         if (ret)
1026                 return ret;
1027
1028         while (nr_iters < 20 &&
1029                (_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1030                                                       KEY_TYPE_discard))) {
1031                 struct bkey     unpacked;
1032                 struct bkey_s_c k = bkey_disassemble(b, _k, &unpacked);
1033                 unsigned offset = 0;
1034
1035                 if (bkey_cmp(bkey_start_pos(k.k), *end) >= 0)
1036                         break;
1037
1038                 if (bkey_cmp(bkey_start_pos(&insert->k),
1039                              bkey_start_pos(k.k)) > 0)
1040                         offset = bkey_start_offset(&insert->k) -
1041                                 bkey_start_offset(k.k);
1042
1043                 ret = count_iters_for_insert(trans, k, offset,
1044                                              end, &nr_iters, 20, true);
1045                 if (ret)
1046                         return ret;
1047
1048                 if (nr_iters >= 20)
1049                         break;
1050
1051                 bch2_btree_node_iter_advance(&node_iter, b);
1052         }
1053
1054         return 0;
1055 }
1056
1057 int bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
1058 {
1059         struct bpos end;
1060         int ret;
1061
1062         ret = bch2_extent_atomic_end(iter, k, &end);
1063         if (ret)
1064                 return ret;
1065
1066         bch2_cut_back(end, &k->k);
1067         return 0;
1068 }
1069
1070 int bch2_extent_is_atomic(struct bkey_i *k, struct btree_iter *iter)
1071 {
1072         struct bpos end;
1073         int ret;
1074
1075         ret = bch2_extent_atomic_end(iter, k, &end);
1076         if (ret)
1077                 return ret;
1078
1079         return !bkey_cmp(end, k->k.p);
1080 }
1081
1082 enum btree_insert_ret
1083 bch2_extent_can_insert(struct btree_trans *trans,
1084                        struct btree_insert_entry *insert,
1085                        unsigned *u64s)
1086 {
1087         struct btree_iter_level *l = &insert->iter->l[0];
1088         struct btree_node_iter node_iter = l->iter;
1089         enum bch_extent_overlap overlap;
1090         struct bkey_packed *_k;
1091         struct bkey unpacked;
1092         struct bkey_s_c k;
1093         int sectors;
1094
1095         /*
1096          * We avoid creating whiteouts whenever possible when deleting, but
1097          * those optimizations mean we may potentially insert two whiteouts
1098          * instead of one (when we overlap with the front of one extent and the
1099          * back of another):
1100          */
1101         if (bkey_whiteout(&insert->k->k))
1102                 *u64s += BKEY_U64s;
1103
1104         _k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
1105                                               KEY_TYPE_discard);
1106         if (!_k)
1107                 return BTREE_INSERT_OK;
1108
1109         k = bkey_disassemble(l->b, _k, &unpacked);
1110
1111         overlap = bch2_extent_overlap(&insert->k->k, k.k);
1112
1113         /* account for having to split existing extent: */
1114         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1115                 *u64s += _k->u64s;
1116
1117         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
1118             (sectors = bch2_extent_is_compressed(k))) {
1119                 int flags = trans->flags & BTREE_INSERT_NOFAIL
1120                         ? BCH_DISK_RESERVATION_NOFAIL : 0;
1121
1122                 switch (bch2_disk_reservation_add(trans->c,
1123                                 trans->disk_res,
1124                                 sectors, flags)) {
1125                 case 0:
1126                         break;
1127                 case -ENOSPC:
1128                         return BTREE_INSERT_ENOSPC;
1129                 default:
1130                         BUG();
1131                 }
1132         }
1133
1134         return BTREE_INSERT_OK;
1135 }
1136
1137 static void
1138 extent_squash(struct bch_fs *c, struct btree_iter *iter,
1139               struct bkey_i *insert,
1140               struct bkey_packed *_k, struct bkey_s k,
1141               enum bch_extent_overlap overlap)
1142 {
1143         struct btree_iter_level *l = &iter->l[0];
1144
1145         switch (overlap) {
1146         case BCH_EXTENT_OVERLAP_FRONT:
1147                 /* insert overlaps with start of k: */
1148                 __bch2_cut_front(insert->k.p, k);
1149                 EBUG_ON(bkey_deleted(k.k));
1150                 extent_save(l->b, _k, k.k);
1151                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1152                                          _k, _k->u64s, _k->u64s);
1153                 break;
1154
1155         case BCH_EXTENT_OVERLAP_BACK:
1156                 /* insert overlaps with end of k: */
1157                 bch2_cut_back(bkey_start_pos(&insert->k), k.k);
1158                 EBUG_ON(bkey_deleted(k.k));
1159                 extent_save(l->b, _k, k.k);
1160
1161                 /*
1162                  * As the auxiliary tree is indexed by the end of the
1163                  * key and we've just changed the end, update the
1164                  * auxiliary tree.
1165                  */
1166                 bch2_bset_fix_invalidated_key(l->b, _k);
1167                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1168                                          _k, _k->u64s, _k->u64s);
1169                 break;
1170
1171         case BCH_EXTENT_OVERLAP_ALL: {
1172                 /* The insert key completely covers k, invalidate k */
1173                 if (!bkey_whiteout(k.k))
1174                         btree_account_key_drop(l->b, _k);
1175
1176                 k.k->size = 0;
1177                 k.k->type = KEY_TYPE_deleted;
1178
1179                 if (_k >= btree_bset_last(l->b)->start) {
1180                         unsigned u64s = _k->u64s;
1181
1182                         bch2_bset_delete(l->b, _k, _k->u64s);
1183                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1184                                                  _k, u64s, 0);
1185                 } else {
1186                         extent_save(l->b, _k, k.k);
1187                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1188                                                  _k, _k->u64s, _k->u64s);
1189                 }
1190
1191                 break;
1192         }
1193         case BCH_EXTENT_OVERLAP_MIDDLE: {
1194                 BKEY_PADDED(k) split;
1195                 /*
1196                  * The insert key falls 'in the middle' of k
1197                  * The insert key splits k in 3:
1198                  * - start only in k, preserve
1199                  * - middle common section, invalidate in k
1200                  * - end only in k, preserve
1201                  *
1202                  * We update the old key to preserve the start,
1203                  * insert will be the new common section,
1204                  * we manually insert the end that we are preserving.
1205                  *
1206                  * modify k _before_ doing the insert (which will move
1207                  * what k points to)
1208                  */
1209                 bkey_reassemble(&split.k, k.s_c);
1210                 split.k.k.needs_whiteout |= bkey_written(l->b, _k);
1211
1212                 bch2_cut_back(bkey_start_pos(&insert->k), &split.k.k);
1213                 BUG_ON(bkey_deleted(&split.k.k));
1214
1215                 __bch2_cut_front(insert->k.p, k);
1216                 BUG_ON(bkey_deleted(k.k));
1217                 extent_save(l->b, _k, k.k);
1218                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1219                                          _k, _k->u64s, _k->u64s);
1220
1221                 extent_bset_insert(c, iter, &split.k);
1222                 break;
1223         }
1224         }
1225 }
1226
1227 struct extent_insert_state {
1228         struct bkey_i                   whiteout;
1229         bool                            update_journal;
1230         bool                            update_btree;
1231         bool                            deleting;
1232 };
1233
1234 static void __bch2_insert_fixup_extent(struct bch_fs *c,
1235                                        struct btree_iter *iter,
1236                                        struct bkey_i *insert,
1237                                        struct extent_insert_state *s)
1238 {
1239         struct btree_iter_level *l = &iter->l[0];
1240         struct bkey_packed *_k;
1241         struct bkey unpacked;
1242
1243         while ((_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
1244                                                       KEY_TYPE_discard))) {
1245                 struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
1246                 struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
1247                 enum bch_extent_overlap overlap =
1248                         bch2_extent_overlap(&insert->k, k.k);
1249
1250                 if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
1251                         break;
1252
1253                 if (!bkey_whiteout(k.k))
1254                         s->update_journal = true;
1255
1256                 if (!s->update_journal) {
1257                         bch2_cut_front(cur_end, insert);
1258                         bch2_cut_front(cur_end, &s->whiteout);
1259                         bch2_btree_iter_set_pos_same_leaf(iter, cur_end);
1260                         goto next;
1261                 }
1262
1263                 /*
1264                  * When deleting, if possible just do it by switching the type
1265                  * of the key we're deleting, instead of creating and inserting
1266                  * a new whiteout:
1267                  */
1268                 if (s->deleting &&
1269                     !s->update_btree &&
1270                     !bkey_cmp(insert->k.p, k.k->p) &&
1271                     !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
1272                         if (!bkey_whiteout(k.k)) {
1273                                 btree_account_key_drop(l->b, _k);
1274                                 _k->type = KEY_TYPE_discard;
1275                                 reserve_whiteout(l->b, _k);
1276                                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1277                                                         _k, _k->u64s, _k->u64s);
1278                         }
1279                         break;
1280                 }
1281
1282                 if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
1283                         insert->k.needs_whiteout = true;
1284                         s->update_btree = true;
1285                 }
1286
1287                 if (s->update_btree &&
1288                     overlap == BCH_EXTENT_OVERLAP_ALL &&
1289                     bkey_whiteout(k.k) &&
1290                     k.k->needs_whiteout) {
1291                         unreserve_whiteout(l->b, _k);
1292                         _k->needs_whiteout = false;
1293                 }
1294
1295                 extent_squash(c, iter, insert, _k, k, overlap);
1296
1297                 if (!s->update_btree)
1298                         bch2_cut_front(cur_end, insert);
1299 next:
1300                 if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
1301                     overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1302                         break;
1303         }
1304 }
1305
1306 /**
1307  * bch_extent_insert_fixup - insert a new extent and deal with overlaps
1308  *
1309  * this may result in not actually doing the insert, or inserting some subset
1310  * of the insert key. For cmpxchg operations this is where that logic lives.
1311  *
1312  * All subsets of @insert that need to be inserted are inserted using
1313  * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
1314  * returns false, setting @iter->pos for the prefix of @insert that actually got
1315  * inserted.
1316  *
1317  * BSET INVARIANTS: this function is responsible for maintaining all the
1318  * invariants for bsets of extents in memory. things get really hairy with 0
1319  * size extents
1320  *
1321  * within one bset:
1322  *
1323  * bkey_start_pos(bkey_next(k)) >= k
1324  * or bkey_start_offset(bkey_next(k)) >= k->offset
1325  *
1326  * i.e. strict ordering, no overlapping extents.
1327  *
1328  * multiple bsets (i.e. full btree node):
1329  *
1330  * ∀ k, j
1331  *   k.size != 0 ∧ j.size != 0 →
1332  *     ¬ (k > bkey_start_pos(j) ∧ k < j)
1333  *
1334  * i.e. no two overlapping keys _of nonzero size_
1335  *
1336  * We can't realistically maintain this invariant for zero size keys because of
1337  * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
1338  * there may be another 0 size key between them in another bset, and it will
1339  * thus overlap with the merged key.
1340  *
1341  * In addition, the end of iter->pos indicates how much has been processed.
1342  * If the end of iter->pos is not the same as the end of insert, then
1343  * key insertion needs to continue/be retried.
1344  */
1345 void bch2_insert_fixup_extent(struct btree_trans *trans,
1346                               struct btree_insert_entry *insert)
1347 {
1348         struct bch_fs *c = trans->c;
1349         struct btree_iter *iter = insert->iter;
1350         struct extent_insert_state s = {
1351                 .whiteout       = *insert->k,
1352                 .update_journal = !bkey_whiteout(&insert->k->k),
1353                 .update_btree   = !bkey_whiteout(&insert->k->k),
1354                 .deleting       = bkey_whiteout(&insert->k->k),
1355         };
1356         BKEY_PADDED(k) tmp;
1357
1358         EBUG_ON(iter->level);
1359         EBUG_ON(!insert->k->k.size);
1360         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1361
1362         __bch2_insert_fixup_extent(c, iter, insert->k, &s);
1363
1364         bch2_btree_iter_set_pos_same_leaf(iter, insert->k->k.p);
1365
1366         if (s.update_btree) {
1367                 bkey_copy(&tmp.k, insert->k);
1368
1369                 if (s.deleting)
1370                         tmp.k.k.type = KEY_TYPE_discard;
1371
1372                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1373
1374                 extent_bset_insert(c, iter, &tmp.k);
1375         }
1376
1377         if (s.update_journal) {
1378                 bkey_copy(&tmp.k, !s.deleting ? insert->k : &s.whiteout);
1379
1380                 if (s.deleting)
1381                         tmp.k.k.type = KEY_TYPE_discard;
1382
1383                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1384
1385                 bch2_btree_journal_key(trans, iter, &tmp.k);
1386         }
1387
1388         bch2_cut_front(insert->k->k.p, insert->k);
1389 }
1390
1391 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
1392 {
1393         return bch2_bkey_ptrs_invalid(c, k);
1394 }
1395
1396 void bch2_extent_debugcheck(struct bch_fs *c, struct bkey_s_c k)
1397 {
1398         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1399         const union bch_extent_entry *entry;
1400         struct extent_ptr_decoded p;
1401         char buf[160];
1402
1403         /*
1404          * XXX: we should be doing most/all of these checks at startup time,
1405          * where we check bch2_bkey_invalid() in btree_node_read_done()
1406          *
1407          * But note that we can't check for stale pointers or incorrect gc marks
1408          * until after journal replay is done (it might be an extent that's
1409          * going to get overwritten during replay)
1410          */
1411
1412         if (percpu_down_read_trylock(&c->mark_lock)) {
1413                 bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1414                                !bch2_bkey_replicas_marked_locked(c, e.s_c, false), c,
1415                                "extent key bad (replicas not marked in superblock):\n%s",
1416                                (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
1417                 percpu_up_read(&c->mark_lock);
1418         }
1419         /*
1420          * If journal replay hasn't finished, we might be seeing keys
1421          * that will be overwritten by the time journal replay is done:
1422          */
1423         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1424                 return;
1425
1426         extent_for_each_ptr_decode(e, p, entry) {
1427                 struct bch_dev *ca      = bch_dev_bkey_exists(c, p.ptr.dev);
1428                 struct bucket_mark mark = ptr_bucket_mark(ca, &p.ptr);
1429                 unsigned stale          = gen_after(mark.gen, p.ptr.gen);
1430                 unsigned disk_sectors   = ptr_disk_sectors(p);
1431                 unsigned mark_sectors   = p.ptr.cached
1432                         ? mark.cached_sectors
1433                         : mark.dirty_sectors;
1434
1435                 bch2_fs_bug_on(stale && !p.ptr.cached, c,
1436                                "stale dirty pointer (ptr gen %u bucket %u",
1437                                p.ptr.gen, mark.gen);
1438
1439                 bch2_fs_bug_on(stale > 96, c, "key too stale: %i", stale);
1440
1441                 bch2_fs_bug_on(!stale &&
1442                                (mark.data_type != BCH_DATA_USER ||
1443                                 mark_sectors < disk_sectors), c,
1444                                "extent pointer not marked: %s:\n"
1445                                "type %u sectors %u < %u",
1446                                (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf),
1447                                mark.data_type,
1448                                mark_sectors, disk_sectors);
1449         }
1450 }
1451
1452 void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
1453                          struct bkey_s_c k)
1454 {
1455         bch2_bkey_ptrs_to_text(out, c, k);
1456 }
1457
1458 static unsigned bch2_crc_field_size_max[] = {
1459         [BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
1460         [BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
1461         [BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
1462 };
1463
1464 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
1465                                  struct bch_extent_crc_unpacked src)
1466 {
1467 #define set_common_fields(_dst, _src)                                   \
1468                 _dst.csum_type          = _src.csum_type,               \
1469                 _dst.compression_type   = _src.compression_type,        \
1470                 _dst._compressed_size   = _src.compressed_size - 1,     \
1471                 _dst._uncompressed_size = _src.uncompressed_size - 1,   \
1472                 _dst.offset             = _src.offset
1473
1474         switch (extent_entry_type(to_entry(dst))) {
1475         case BCH_EXTENT_ENTRY_crc32:
1476                 set_common_fields(dst->crc32, src);
1477                 dst->crc32.csum  = *((__le32 *) &src.csum.lo);
1478                 break;
1479         case BCH_EXTENT_ENTRY_crc64:
1480                 set_common_fields(dst->crc64, src);
1481                 dst->crc64.nonce        = src.nonce;
1482                 dst->crc64.csum_lo      = src.csum.lo;
1483                 dst->crc64.csum_hi      = *((__le16 *) &src.csum.hi);
1484                 break;
1485         case BCH_EXTENT_ENTRY_crc128:
1486                 set_common_fields(dst->crc128, src);
1487                 dst->crc128.nonce       = src.nonce;
1488                 dst->crc128.csum        = src.csum;
1489                 break;
1490         default:
1491                 BUG();
1492         }
1493 #undef set_common_fields
1494 }
1495
1496 static void bch2_extent_crc_append(struct bkey_i *k,
1497                                    struct bch_extent_crc_unpacked new)
1498 {
1499         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
1500         union bch_extent_crc *crc = (void *) ptrs.end;
1501
1502         if (bch_crc_bytes[new.csum_type]        <= 4 &&
1503             new.uncompressed_size - 1           <= CRC32_SIZE_MAX &&
1504             new.nonce                           <= CRC32_NONCE_MAX)
1505                 crc->type = 1 << BCH_EXTENT_ENTRY_crc32;
1506         else if (bch_crc_bytes[new.csum_type]   <= 10 &&
1507                    new.uncompressed_size - 1    <= CRC64_SIZE_MAX &&
1508                    new.nonce                    <= CRC64_NONCE_MAX)
1509                 crc->type = 1 << BCH_EXTENT_ENTRY_crc64;
1510         else if (bch_crc_bytes[new.csum_type]   <= 16 &&
1511                    new.uncompressed_size - 1    <= CRC128_SIZE_MAX &&
1512                    new.nonce                    <= CRC128_NONCE_MAX)
1513                 crc->type = 1 << BCH_EXTENT_ENTRY_crc128;
1514         else
1515                 BUG();
1516
1517         bch2_extent_crc_pack(crc, new);
1518
1519         k->k.u64s += extent_entry_u64s(ptrs.end);
1520
1521         EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
1522 }
1523
1524 static inline void __extent_entry_insert(struct bkey_i *k,
1525                                          union bch_extent_entry *dst,
1526                                          union bch_extent_entry *new)
1527 {
1528         union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
1529
1530         memmove_u64s_up((u64 *) dst + extent_entry_u64s(new),
1531                         dst, (u64 *) end - (u64 *) dst);
1532         k->k.u64s += extent_entry_u64s(new);
1533         memcpy(dst, new, extent_entry_bytes(new));
1534 }
1535
1536 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
1537                                     struct extent_ptr_decoded *p)
1538 {
1539         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
1540         struct bch_extent_crc_unpacked crc =
1541                 bch2_extent_crc_unpack(&k->k, NULL);
1542         union bch_extent_entry *pos;
1543         unsigned i;
1544
1545         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1546                 pos = ptrs.start;
1547                 goto found;
1548         }
1549
1550         bkey_for_each_crc(&k->k, ptrs, crc, pos)
1551                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1552                         pos = extent_entry_next(pos);
1553                         goto found;
1554                 }
1555
1556         bch2_extent_crc_append(k, p->crc);
1557         pos = bkey_val_end(bkey_i_to_s(k));
1558 found:
1559         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
1560         __extent_entry_insert(k, pos, to_entry(&p->ptr));
1561
1562         for (i = 0; i < p->ec_nr; i++) {
1563                 p->ec[i].type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
1564                 __extent_entry_insert(k, pos, to_entry(&p->ec[i]));
1565         }
1566 }
1567
1568 /*
1569  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
1570  *
1571  * Returns true if @k should be dropped entirely
1572  *
1573  * For existing keys, only called when btree nodes are being rewritten, not when
1574  * they're merely being compacted/resorted in memory.
1575  */
1576 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1577 {
1578         struct bch_extent_ptr *ptr;
1579
1580         bch2_bkey_drop_ptrs(k, ptr,
1581                 ptr->cached &&
1582                 ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
1583
1584         /* will only happen if all pointers were cached: */
1585         if (!bkey_val_u64s(k.k))
1586                 k.k->type = KEY_TYPE_discard;
1587
1588         return bkey_whiteout(k.k);
1589 }
1590
1591 void bch2_bkey_mark_replicas_cached(struct bch_fs *c, struct bkey_s k,
1592                                     unsigned target,
1593                                     unsigned nr_desired_replicas)
1594 {
1595         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1596         union bch_extent_entry *entry;
1597         struct extent_ptr_decoded p;
1598         int extra = bch2_bkey_durability(c, k.s_c) - nr_desired_replicas;
1599
1600         if (target && extra > 0)
1601                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1602                         int n = bch2_extent_ptr_durability(c, p);
1603
1604                         if (n && n <= extra &&
1605                             !bch2_dev_in_target(c, p.ptr.dev, target)) {
1606                                 entry->ptr.cached = true;
1607                                 extra -= n;
1608                         }
1609                 }
1610
1611         if (extra > 0)
1612                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1613                         int n = bch2_extent_ptr_durability(c, p);
1614
1615                         if (n && n <= extra) {
1616                                 entry->ptr.cached = true;
1617                                 extra -= n;
1618                         }
1619                 }
1620 }
1621
1622 enum merge_result bch2_extent_merge(struct bch_fs *c,
1623                                     struct bkey_s _l, struct bkey_s _r)
1624 {
1625         struct bkey_s_extent l = bkey_s_to_extent(_l);
1626         struct bkey_s_extent r = bkey_s_to_extent(_r);
1627         union bch_extent_entry *en_l = l.v->start;
1628         union bch_extent_entry *en_r = r.v->start;
1629         struct bch_extent_crc_unpacked crc_l, crc_r;
1630
1631         if (bkey_val_u64s(l.k) != bkey_val_u64s(r.k))
1632                 return BCH_MERGE_NOMERGE;
1633
1634         crc_l = bch2_extent_crc_unpack(l.k, NULL);
1635
1636         extent_for_each_entry(l, en_l) {
1637                 en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
1638
1639                 if (extent_entry_type(en_l) != extent_entry_type(en_r))
1640                         return BCH_MERGE_NOMERGE;
1641
1642                 switch (extent_entry_type(en_l)) {
1643                 case BCH_EXTENT_ENTRY_ptr: {
1644                         const struct bch_extent_ptr *lp = &en_l->ptr;
1645                         const struct bch_extent_ptr *rp = &en_r->ptr;
1646                         struct bch_dev *ca;
1647
1648                         if (lp->offset + crc_l.compressed_size != rp->offset ||
1649                             lp->dev                     != rp->dev ||
1650                             lp->gen                     != rp->gen)
1651                                 return BCH_MERGE_NOMERGE;
1652
1653                         /* We don't allow extents to straddle buckets: */
1654                         ca = bch_dev_bkey_exists(c, lp->dev);
1655
1656                         if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
1657                                 return BCH_MERGE_NOMERGE;
1658
1659                         break;
1660                 }
1661                 case BCH_EXTENT_ENTRY_stripe_ptr:
1662                         if (en_l->stripe_ptr.block      != en_r->stripe_ptr.block ||
1663                             en_l->stripe_ptr.idx        != en_r->stripe_ptr.idx)
1664                                 return BCH_MERGE_NOMERGE;
1665                         break;
1666                 case BCH_EXTENT_ENTRY_crc32:
1667                 case BCH_EXTENT_ENTRY_crc64:
1668                 case BCH_EXTENT_ENTRY_crc128:
1669                         crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
1670                         crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
1671
1672                         if (crc_l.csum_type             != crc_r.csum_type ||
1673                             crc_l.compression_type      != crc_r.compression_type ||
1674                             crc_l.nonce                 != crc_r.nonce)
1675                                 return BCH_MERGE_NOMERGE;
1676
1677                         if (crc_l.offset + crc_l.live_size != crc_l.compressed_size ||
1678                             crc_r.offset)
1679                                 return BCH_MERGE_NOMERGE;
1680
1681                         if (!bch2_checksum_mergeable(crc_l.csum_type))
1682                                 return BCH_MERGE_NOMERGE;
1683
1684                         if (crc_l.compression_type)
1685                                 return BCH_MERGE_NOMERGE;
1686
1687                         if (crc_l.csum_type &&
1688                             crc_l.uncompressed_size +
1689                             crc_r.uncompressed_size > c->sb.encoded_extent_max)
1690                                 return BCH_MERGE_NOMERGE;
1691
1692                         if (crc_l.uncompressed_size + crc_r.uncompressed_size - 1 >
1693                             bch2_crc_field_size_max[extent_entry_type(en_l)])
1694                                 return BCH_MERGE_NOMERGE;
1695
1696                         break;
1697                 default:
1698                         return BCH_MERGE_NOMERGE;
1699                 }
1700         }
1701
1702         extent_for_each_entry(l, en_l) {
1703                 struct bch_extent_crc_unpacked crc_l, crc_r;
1704
1705                 en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
1706
1707                 if (!extent_entry_is_crc(en_l))
1708                         continue;
1709
1710                 crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
1711                 crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
1712
1713                 crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
1714                                                  crc_l.csum,
1715                                                  crc_r.csum,
1716                                                  crc_r.uncompressed_size << 9);
1717
1718                 crc_l.uncompressed_size += crc_r.uncompressed_size;
1719                 crc_l.compressed_size   += crc_r.compressed_size;
1720
1721                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_l);
1722         }
1723
1724         bch2_key_resize(l.k, l.k->size + r.k->size);
1725
1726         return BCH_MERGE_MERGE;
1727 }
1728
1729 /*
1730  * When merging an extent that we're inserting into a btree node, the new merged
1731  * extent could overlap with an existing 0 size extent - if we don't fix that,
1732  * it'll break the btree node iterator so this code finds those 0 size extents
1733  * and shifts them out of the way.
1734  *
1735  * Also unpacks and repacks.
1736  */
1737 static bool bch2_extent_merge_inline(struct bch_fs *c,
1738                                      struct btree_iter *iter,
1739                                      struct bkey_packed *l,
1740                                      struct bkey_packed *r,
1741                                      bool back_merge)
1742 {
1743         struct btree *b = iter->l[0].b;
1744         struct btree_node_iter *node_iter = &iter->l[0].iter;
1745         BKEY_PADDED(k) li, ri;
1746         struct bkey_packed *m   = back_merge ? l : r;
1747         struct bkey_i *mi       = back_merge ? &li.k : &ri.k;
1748         struct bset_tree *t     = bch2_bkey_to_bset(b, m);
1749         enum merge_result ret;
1750
1751         EBUG_ON(bkey_written(b, m));
1752
1753         if (bkey_val_u64s(l) > BKEY_EXTENT_VAL_U64s_MAX ||
1754             bkey_val_u64s(r) > BKEY_EXTENT_VAL_U64s_MAX)
1755                 return BCH_MERGE_NOMERGE;
1756
1757         /*
1758          * We need to save copies of both l and r, because we might get a
1759          * partial merge (which modifies both) and then fails to repack
1760          */
1761         bch2_bkey_unpack(b, &li.k, l);
1762         bch2_bkey_unpack(b, &ri.k, r);
1763
1764         ret = bch2_bkey_merge(c,
1765                               bkey_i_to_s(&li.k),
1766                               bkey_i_to_s(&ri.k));
1767         if (ret == BCH_MERGE_NOMERGE)
1768                 return false;
1769
1770         if (debug_check_bkeys(c))
1771                 bch2_bkey_debugcheck(c, b, bkey_i_to_s_c(&li.k));
1772         if (debug_check_bkeys(c) &&
1773             ret == BCH_MERGE_PARTIAL)
1774                 bch2_bkey_debugcheck(c, b, bkey_i_to_s_c(&ri.k));
1775
1776         /*
1777          * check if we overlap with deleted extents - would break the sort
1778          * order:
1779          */
1780         if (back_merge) {
1781                 struct bkey_packed *n = bkey_next(m);
1782
1783                 if (n != btree_bkey_last(b, t) &&
1784                     bkey_cmp_left_packed(b, n, &li.k.k.p) <= 0 &&
1785                     bkey_deleted(n))
1786                         return false;
1787         } else if (ret == BCH_MERGE_MERGE) {
1788                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1789
1790                 if (prev &&
1791                     bkey_cmp_left_packed_byval(b, prev,
1792                                 bkey_start_pos(&li.k.k)) > 0)
1793                         return false;
1794         }
1795
1796         if (ret == BCH_MERGE_PARTIAL) {
1797                 if (!extent_i_save(b, m, mi))
1798                         return false;
1799
1800                 if (!back_merge)
1801                         bkey_copy(packed_to_bkey(l), &li.k);
1802                 else
1803                         bkey_copy(packed_to_bkey(r), &ri.k);
1804         } else {
1805                 if (!extent_i_save(b, m, &li.k))
1806                         return false;
1807         }
1808
1809         bch2_bset_fix_invalidated_key(b, m);
1810         bch2_btree_node_iter_fix(iter, b, node_iter,
1811                                  m, m->u64s, m->u64s);
1812
1813         return ret == BCH_MERGE_MERGE;
1814 }
1815
1816 bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
1817                                unsigned nr_replicas)
1818 {
1819         struct btree_trans trans;
1820         struct btree_iter *iter;
1821         struct bpos end = pos;
1822         struct bkey_s_c k;
1823         bool ret = true;
1824         int err;
1825
1826         end.offset += size;
1827
1828         bch2_trans_init(&trans, c, 0, 0);
1829
1830         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, pos,
1831                            BTREE_ITER_SLOTS, k, err) {
1832                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
1833                         break;
1834
1835                 if (nr_replicas > bch2_bkey_nr_ptrs_allocated(k)) {
1836                         ret = false;
1837                         break;
1838                 }
1839         }
1840         bch2_trans_exit(&trans);
1841
1842         return ret;
1843 }
1844
1845 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
1846 {
1847         unsigned ret = 0;
1848
1849         switch (k.k->type) {
1850         case KEY_TYPE_extent: {
1851                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1852                 const union bch_extent_entry *entry;
1853                 struct extent_ptr_decoded p;
1854
1855                 extent_for_each_ptr_decode(e, p, entry)
1856                         ret += !p.ptr.cached &&
1857                                 p.crc.compression_type == BCH_COMPRESSION_NONE;
1858                 break;
1859         }
1860         case KEY_TYPE_reservation:
1861                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
1862                 break;
1863         }
1864
1865         return ret;
1866 }
1867
1868 /* KEY_TYPE_reservation: */
1869
1870 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
1871 {
1872         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1873
1874         if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
1875                 return "incorrect value size";
1876
1877         if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
1878                 return "invalid nr_replicas";
1879
1880         return NULL;
1881 }
1882
1883 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
1884                               struct bkey_s_c k)
1885 {
1886         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1887
1888         pr_buf(out, "generation %u replicas %u",
1889                le32_to_cpu(r.v->generation),
1890                r.v->nr_replicas);
1891 }
1892
1893 enum merge_result bch2_reservation_merge(struct bch_fs *c,
1894                                          struct bkey_s _l, struct bkey_s _r)
1895 {
1896         struct bkey_s_reservation l = bkey_s_to_reservation(_l);
1897         struct bkey_s_reservation r = bkey_s_to_reservation(_r);
1898
1899         if (l.v->generation != r.v->generation ||
1900             l.v->nr_replicas != r.v->nr_replicas)
1901                 return BCH_MERGE_NOMERGE;
1902
1903         if ((u64) l.k->size + r.k->size > KEY_SIZE_MAX) {
1904                 bch2_key_resize(l.k, KEY_SIZE_MAX);
1905                 __bch2_cut_front(l.k->p, r.s);
1906                 return BCH_MERGE_PARTIAL;
1907         }
1908
1909         bch2_key_resize(l.k, l.k->size + r.k->size);
1910
1911         return BCH_MERGE_MERGE;
1912 }