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 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         if (p.has_ec) {
  82                 struct stripe *s =
  83                         genradix_ptr(&c->stripes[0], p.ec.idx);
  84
  85                 if (WARN_ON(!s))
  86                         goto out;
  87
  88                 durability = max_t(unsigned, durability, s->nr_redundant);
  89         }
  90 out:
  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.has_ec)
 210                         p.idx++;
 211
 212                 if (p.idx >= (unsigned) p.has_ec + 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 unsigned bch2_bkey_nr_alloc_ptrs(struct bkey_s_c k)
 810 {
 811         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 812         const union bch_extent_entry *entry;
 813         unsigned ret = 0;
 814
 815         bkey_extent_entry_for_each(ptrs, entry) {
 816                 switch (__extent_entry_type(entry)) {
 817                 case BCH_EXTENT_ENTRY_ptr:
 818                 case BCH_EXTENT_ENTRY_stripe_ptr:
 819                         ret++;
 820                 }
 821         }
 822
 823         return ret;
 824 }
 825
 826 static int count_iters_for_insert(struct btree_trans *trans,
 827                                   struct bkey_s_c k,
 828                                   unsigned offset,
 829                                   struct bpos *end,
 830                                   unsigned *nr_iters,
 831                                   unsigned max_iters,
 832                                   bool overwrite)
 833 {
 834         int ret = 0;
 835
 836         switch (k.k->type) {
 837         case KEY_TYPE_extent:
 838         case KEY_TYPE_reflink_v:
 839                 *nr_iters += bch2_bkey_nr_alloc_ptrs(k);
 840
 841                 if (*nr_iters >= max_iters) {
 842                         *end = bpos_min(*end, k.k->p);
 843                         ret = 1;
 844                 }
 845
 846                 break;
 847         case KEY_TYPE_reflink_p: {
 848                 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
 849                 u64 idx = le64_to_cpu(p.v->idx);
 850                 unsigned sectors = bpos_min(*end, p.k->p).offset -
 851                         bkey_start_offset(p.k);
 852                 struct btree_iter *iter;
 853                 struct bkey_s_c r_k;
 854
 855                 for_each_btree_key(trans, iter,
 856                                    BTREE_ID_REFLINK, POS(0, idx + offset),
 857                                    BTREE_ITER_SLOTS, r_k, ret) {
 858                         if (bkey_cmp(bkey_start_pos(r_k.k),
 859                                      POS(0, idx + sectors)) >= 0)
 860                                 break;
 861
 862                         *nr_iters += 1 + bch2_bkey_nr_alloc_ptrs(r_k);
 863
 864                         if (*nr_iters >= max_iters) {
 865                                 struct bpos pos = bkey_start_pos(k.k);
 866                                 pos.offset += r_k.k->p.offset - idx;
 867
 868                                 *end = bpos_min(*end, pos);
 869                                 ret = 1;
 870                                 break;
 871                         }
 872                 }
 873
 874                 bch2_trans_iter_put(trans, iter);
 875                 break;
 876         }
 877         }
 878
 879         return ret;
 880 }
 881
 882 #define EXTENT_ITERS_MAX        (BTREE_ITER_MAX / 3)
 883
 884 int bch2_extent_atomic_end(struct btree_iter *iter,
 885                            struct bkey_i *insert,
 886                            struct bpos *end)
 887 {
 888         struct btree_trans *trans = iter->trans;
 889         struct btree *b;
 890         struct btree_node_iter  node_iter;
 891         struct bkey_packed      *_k;
 892         unsigned                nr_iters = 0;
 893         int ret;
 894
 895         ret = bch2_btree_iter_traverse(iter);
 896         if (ret)
 897                 return ret;
 898
 899         b = iter->l[0].b;
 900         node_iter = iter->l[0].iter;
 901
 902         BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0);
 903
 904         *end = bpos_min(insert->k.p, b->key.k.p);
 905
 906         ret = count_iters_for_insert(trans, bkey_i_to_s_c(insert), 0, end,
 907                                      &nr_iters, EXTENT_ITERS_MAX / 2, false);
 908         if (ret < 0)
 909                 return ret;
 910
 911         while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
 912                                                       KEY_TYPE_discard))) {
 913                 struct bkey     unpacked;
 914                 struct bkey_s_c k = bkey_disassemble(b, _k, &unpacked);
 915                 unsigned offset = 0;
 916
 917                 if (bkey_cmp(bkey_start_pos(k.k), *end) >= 0)
 918                         break;
 919
 920                 if (bkey_cmp(bkey_start_pos(&insert->k),
 921                              bkey_start_pos(k.k)) > 0)
 922                         offset = bkey_start_offset(&insert->k) -
 923                                 bkey_start_offset(k.k);
 924
 925                 ret = count_iters_for_insert(trans, k, offset, end,
 926                                         &nr_iters, EXTENT_ITERS_MAX, true);
 927                 if (ret)
 928                         break;
 929
 930                 bch2_btree_node_iter_advance(&node_iter, b);
 931         }
 932
 933         return ret < 0 ? ret : 0;
 934 }
 935
 936 int bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
 937 {
 938         struct bpos end;
 939         int ret;
 940
 941         ret = bch2_extent_atomic_end(iter, k, &end);
 942         if (ret)
 943                 return ret;
 944
 945         bch2_cut_back(end, &k->k);
 946         return 0;
 947 }
 948
 949 int bch2_extent_is_atomic(struct bkey_i *k, struct btree_iter *iter)
 950 {
 951         struct bpos end;
 952         int ret;
 953
 954         ret = bch2_extent_atomic_end(iter, k, &end);
 955         if (ret)
 956                 return ret;
 957
 958         return !bkey_cmp(end, k->k.p);
 959 }
 960
 961 enum btree_insert_ret
 962 bch2_extent_can_insert(struct btree_trans *trans,
 963                        struct btree_insert_entry *insert,
 964                        unsigned *u64s)
 965 {
 966         struct btree_iter_level *l = &insert->iter->l[0];
 967         struct btree_node_iter node_iter = l->iter;
 968         enum bch_extent_overlap overlap;
 969         struct bkey_packed *_k;
 970         struct bkey unpacked;
 971         struct bkey_s_c k;
 972         int sectors;
 973
 974         /*
 975          * We avoid creating whiteouts whenever possible when deleting, but
 976          * those optimizations mean we may potentially insert two whiteouts
 977          * instead of one (when we overlap with the front of one extent and the
 978          * back of another):
 979          */
 980         if (bkey_whiteout(&insert->k->k))
 981                 *u64s += BKEY_U64s;
 982
 983         _k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
 984                                               KEY_TYPE_discard);
 985         if (!_k)
 986                 return BTREE_INSERT_OK;
 987
 988         k = bkey_disassemble(l->b, _k, &unpacked);
 989
 990         overlap = bch2_extent_overlap(&insert->k->k, k.k);
 991
 992         /* account for having to split existing extent: */
 993         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
 994                 *u64s += _k->u64s;
 995
 996         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
 997             (sectors = bch2_extent_is_compressed(k))) {
 998                 int flags = trans->flags & BTREE_INSERT_NOFAIL
 999                         ? BCH_DISK_RESERVATION_NOFAIL : 0;
1000
1001                 switch (bch2_disk_reservation_add(trans->c,
1002                                 trans->disk_res,
1003                                 sectors, flags)) {
1004                 case 0:
1005                         break;
1006                 case -ENOSPC:
1007                         return BTREE_INSERT_ENOSPC;
1008                 default:
1009                         BUG();
1010                 }
1011         }
1012
1013         return BTREE_INSERT_OK;
1014 }
1015
1016 static void verify_extent_nonoverlapping(struct bch_fs *c,
1017                                          struct btree *b,
1018                                          struct btree_node_iter *_iter,
1019                                          struct bkey_i *insert)
1020 {
1021 #ifdef CONFIG_BCACHEFS_DEBUG
1022         struct btree_node_iter iter;
1023         struct bkey_packed *k;
1024         struct bkey uk;
1025
1026         if (!expensive_debug_checks(c))
1027                 return;
1028
1029         iter = *_iter;
1030         k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
1031         BUG_ON(k &&
1032                (uk = bkey_unpack_key(b, k),
1033                 bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
1034
1035         iter = *_iter;
1036         k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
1037 #if 0
1038         BUG_ON(k &&
1039                (uk = bkey_unpack_key(b, k),
1040                 bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
1041 #else
1042         if (k &&
1043             (uk = bkey_unpack_key(b, k),
1044              bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
1045                 char buf1[100];
1046                 char buf2[100];
1047
1048                 bch2_bkey_to_text(&PBUF(buf1), &insert->k);
1049                 bch2_bkey_to_text(&PBUF(buf2), &uk);
1050
1051                 bch2_dump_btree_node(b);
1052                 panic("insert > next :\n"
1053                       "insert %s\n"
1054                       "next   %s\n",
1055                       buf1, buf2);
1056         }
1057 #endif
1058
1059 #endif
1060 }
1061
1062 static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
1063                                struct bkey_i *insert)
1064 {
1065         struct btree_iter_level *l = &iter->l[0];
1066         struct bkey_packed *k =
1067                 bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
1068
1069         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
1070
1071         EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
1072         verify_extent_nonoverlapping(c, l->b, &l->iter, insert);
1073
1074         if (debug_check_bkeys(c))
1075                 bch2_bkey_debugcheck(c, l->b, bkey_i_to_s_c(insert));
1076
1077         bch2_bset_insert(l->b, &l->iter, k, insert, 0);
1078         bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
1079 }
1080
1081 static void
1082 extent_squash(struct bch_fs *c, struct btree_iter *iter,
1083               struct bkey_i *insert,
1084               struct bkey_packed *_k, struct bkey_s k,
1085               enum bch_extent_overlap overlap)
1086 {
1087         struct btree_iter_level *l = &iter->l[0];
1088
1089         switch (overlap) {
1090         case BCH_EXTENT_OVERLAP_FRONT:
1091                 /* insert overlaps with start of k: */
1092                 __bch2_cut_front(insert->k.p, k);
1093                 EBUG_ON(bkey_deleted(k.k));
1094                 extent_save(l->b, _k, k.k);
1095                 bch2_btree_iter_fix_key_modified(iter, l->b, _k);
1096                 break;
1097
1098         case BCH_EXTENT_OVERLAP_BACK:
1099                 /* insert overlaps with end of k: */
1100                 bch2_cut_back(bkey_start_pos(&insert->k), k.k);
1101                 EBUG_ON(bkey_deleted(k.k));
1102                 extent_save(l->b, _k, k.k);
1103
1104                 /*
1105                  * As the auxiliary tree is indexed by the end of the
1106                  * key and we've just changed the end, update the
1107                  * auxiliary tree.
1108                  */
1109                 bch2_bset_fix_invalidated_key(l->b, _k);
1110                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1111                                          _k, _k->u64s, _k->u64s);
1112                 break;
1113
1114         case BCH_EXTENT_OVERLAP_ALL: {
1115                 /* The insert key completely covers k, invalidate k */
1116                 if (!bkey_whiteout(k.k))
1117                         btree_account_key_drop(l->b, _k);
1118
1119                 k.k->size = 0;
1120                 k.k->type = KEY_TYPE_deleted;
1121
1122                 if (_k >= btree_bset_last(l->b)->start) {
1123                         unsigned u64s = _k->u64s;
1124
1125                         bch2_bset_delete(l->b, _k, _k->u64s);
1126                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1127                                                  _k, u64s, 0);
1128                 } else {
1129                         extent_save(l->b, _k, k.k);
1130                         bch2_btree_iter_fix_key_modified(iter, l->b, _k);
1131                 }
1132
1133                 break;
1134         }
1135         case BCH_EXTENT_OVERLAP_MIDDLE: {
1136                 BKEY_PADDED(k) split;
1137                 /*
1138                  * The insert key falls 'in the middle' of k
1139                  * The insert key splits k in 3:
1140                  * - start only in k, preserve
1141                  * - middle common section, invalidate in k
1142                  * - end only in k, preserve
1143                  *
1144                  * We update the old key to preserve the start,
1145                  * insert will be the new common section,
1146                  * we manually insert the end that we are preserving.
1147                  *
1148                  * modify k _before_ doing the insert (which will move
1149                  * what k points to)
1150                  */
1151                 bkey_reassemble(&split.k, k.s_c);
1152                 split.k.k.needs_whiteout |= bkey_written(l->b, _k);
1153
1154                 bch2_cut_back(bkey_start_pos(&insert->k), &split.k.k);
1155                 BUG_ON(bkey_deleted(&split.k.k));
1156
1157                 __bch2_cut_front(insert->k.p, k);
1158                 BUG_ON(bkey_deleted(k.k));
1159                 extent_save(l->b, _k, k.k);
1160                 bch2_btree_iter_fix_key_modified(iter, l->b, _k);
1161
1162                 extent_bset_insert(c, iter, &split.k);
1163                 break;
1164         }
1165         }
1166 }
1167
1168 /**
1169  * bch_extent_insert_fixup - insert a new extent and deal with overlaps
1170  *
1171  * this may result in not actually doing the insert, or inserting some subset
1172  * of the insert key. For cmpxchg operations this is where that logic lives.
1173  *
1174  * All subsets of @insert that need to be inserted are inserted using
1175  * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
1176  * returns false, setting @iter->pos for the prefix of @insert that actually got
1177  * inserted.
1178  *
1179  * BSET INVARIANTS: this function is responsible for maintaining all the
1180  * invariants for bsets of extents in memory. things get really hairy with 0
1181  * size extents
1182  *
1183  * within one bset:
1184  *
1185  * bkey_start_pos(bkey_next(k)) >= k
1186  * or bkey_start_offset(bkey_next(k)) >= k->offset
1187  *
1188  * i.e. strict ordering, no overlapping extents.
1189  *
1190  * multiple bsets (i.e. full btree node):
1191  *
1192  * ∀ k, j
1193  *   k.size != 0 ∧ j.size != 0 →
1194  *     ¬ (k > bkey_start_pos(j) ∧ k < j)
1195  *
1196  * i.e. no two overlapping keys _of nonzero size_
1197  *
1198  * We can't realistically maintain this invariant for zero size keys because of
1199  * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
1200  * there may be another 0 size key between them in another bset, and it will
1201  * thus overlap with the merged key.
1202  *
1203  * In addition, the end of iter->pos indicates how much has been processed.
1204  * If the end of iter->pos is not the same as the end of insert, then
1205  * key insertion needs to continue/be retried.
1206  */
1207 void bch2_insert_fixup_extent(struct btree_trans *trans,
1208                               struct btree_insert_entry *insert_entry)
1209 {
1210         struct bch_fs *c = trans->c;
1211         struct btree_iter *iter = insert_entry->iter;
1212         struct bkey_i *insert   = insert_entry->k;
1213         struct btree_iter_level *l = &iter->l[0];
1214         struct btree_node_iter node_iter = l->iter;
1215         bool deleting           = bkey_whiteout(&insert->k);
1216         bool update_journal     = !deleting;
1217         bool update_btree       = !deleting;
1218         struct bkey_i whiteout  = *insert;
1219         struct bkey_packed *_k;
1220         struct bkey unpacked;
1221         BKEY_PADDED(k) tmp;
1222
1223         EBUG_ON(iter->level);
1224         EBUG_ON(!insert->k.size);
1225         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k)));
1226
1227         while ((_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
1228                                                       KEY_TYPE_discard))) {
1229                 struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
1230                 struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
1231                 enum bch_extent_overlap overlap =
1232                         bch2_extent_overlap(&insert->k, k.k);
1233
1234                 if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
1235                         break;
1236
1237                 if (!bkey_whiteout(k.k))
1238                         update_journal = true;
1239
1240                 if (!update_journal) {
1241                         bch2_cut_front(cur_end, insert);
1242                         bch2_cut_front(cur_end, &whiteout);
1243                         bch2_btree_iter_set_pos_same_leaf(iter, cur_end);
1244                         goto next;
1245                 }
1246
1247                 /*
1248                  * When deleting, if possible just do it by switching the type
1249                  * of the key we're deleting, instead of creating and inserting
1250                  * a new whiteout:
1251                  */
1252                 if (deleting &&
1253                     !update_btree &&
1254                     !bkey_cmp(insert->k.p, k.k->p) &&
1255                     !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
1256                         if (!bkey_whiteout(k.k)) {
1257                                 btree_account_key_drop(l->b, _k);
1258                                 _k->type = KEY_TYPE_discard;
1259                                 reserve_whiteout(l->b, _k);
1260                                 bch2_btree_iter_fix_key_modified(iter,
1261                                                                  l->b, _k);
1262                         }
1263                         break;
1264                 }
1265
1266                 if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
1267                         insert->k.needs_whiteout = true;
1268                         update_btree = true;
1269                 }
1270
1271                 if (update_btree &&
1272                     overlap == BCH_EXTENT_OVERLAP_ALL &&
1273                     bkey_whiteout(k.k) &&
1274                     k.k->needs_whiteout) {
1275                         unreserve_whiteout(l->b, _k);
1276                         _k->needs_whiteout = false;
1277                 }
1278
1279                 extent_squash(c, iter, insert, _k, k, overlap);
1280
1281                 if (!update_btree)
1282                         bch2_cut_front(cur_end, insert);
1283 next:
1284                 node_iter = l->iter;
1285
1286                 if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
1287                     overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1288                         break;
1289         }
1290
1291         l->iter = node_iter;
1292         bch2_btree_iter_set_pos_same_leaf(iter, insert->k.p);
1293
1294         if (update_btree) {
1295                 bkey_copy(&tmp.k, insert);
1296
1297                 if (deleting)
1298                         tmp.k.k.type = KEY_TYPE_discard;
1299
1300                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1301
1302                 extent_bset_insert(c, iter, &tmp.k);
1303         }
1304
1305         if (update_journal) {
1306                 bkey_copy(&tmp.k, !deleting ? insert : &whiteout);
1307
1308                 if (deleting)
1309                         tmp.k.k.type = KEY_TYPE_discard;
1310
1311                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1312
1313                 bch2_btree_journal_key(trans, iter, &tmp.k);
1314         }
1315
1316         bch2_cut_front(insert->k.p, insert);
1317 }
1318
1319 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
1320 {
1321         return bch2_bkey_ptrs_invalid(c, k);
1322 }
1323
1324 void bch2_extent_debugcheck(struct bch_fs *c, struct bkey_s_c k)
1325 {
1326         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1327         const union bch_extent_entry *entry;
1328         struct extent_ptr_decoded p;
1329         char buf[160];
1330
1331         /*
1332          * XXX: we should be doing most/all of these checks at startup time,
1333          * where we check bch2_bkey_invalid() in btree_node_read_done()
1334          *
1335          * But note that we can't check for stale pointers or incorrect gc marks
1336          * until after journal replay is done (it might be an extent that's
1337          * going to get overwritten during replay)
1338          */
1339
1340         if (percpu_down_read_trylock(&c->mark_lock)) {
1341                 bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1342                                !bch2_bkey_replicas_marked_locked(c, e.s_c, false), c,
1343                                "extent key bad (replicas not marked in superblock):\n%s",
1344                                (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
1345                 percpu_up_read(&c->mark_lock);
1346         }
1347         /*
1348          * If journal replay hasn't finished, we might be seeing keys
1349          * that will be overwritten by the time journal replay is done:
1350          */
1351         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1352                 return;
1353
1354         extent_for_each_ptr_decode(e, p, entry) {
1355                 struct bch_dev *ca      = bch_dev_bkey_exists(c, p.ptr.dev);
1356                 struct bucket_mark mark = ptr_bucket_mark(ca, &p.ptr);
1357                 unsigned stale          = gen_after(mark.gen, p.ptr.gen);
1358                 unsigned disk_sectors   = ptr_disk_sectors(p);
1359                 unsigned mark_sectors   = p.ptr.cached
1360                         ? mark.cached_sectors
1361                         : mark.dirty_sectors;
1362
1363                 bch2_fs_bug_on(stale && !p.ptr.cached, c,
1364                                "stale dirty pointer (ptr gen %u bucket %u",
1365                                p.ptr.gen, mark.gen);
1366
1367                 bch2_fs_bug_on(stale > 96, c, "key too stale: %i", stale);
1368
1369                 bch2_fs_bug_on(!stale &&
1370                                (mark.data_type != BCH_DATA_USER ||
1371                                 mark_sectors < disk_sectors), c,
1372                                "extent pointer not marked: %s:\n"
1373                                "type %u sectors %u < %u",
1374                                (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf),
1375                                mark.data_type,
1376                                mark_sectors, disk_sectors);
1377         }
1378 }
1379
1380 void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
1381                          struct bkey_s_c k)
1382 {
1383         bch2_bkey_ptrs_to_text(out, c, k);
1384 }
1385
1386 static unsigned bch2_crc_field_size_max[] = {
1387         [BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
1388         [BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
1389         [BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
1390 };
1391
1392 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
1393                                  struct bch_extent_crc_unpacked src)
1394 {
1395 #define set_common_fields(_dst, _src)                                   \
1396                 _dst.csum_type          = _src.csum_type,               \
1397                 _dst.compression_type   = _src.compression_type,        \
1398                 _dst._compressed_size   = _src.compressed_size - 1,     \
1399                 _dst._uncompressed_size = _src.uncompressed_size - 1,   \
1400                 _dst.offset             = _src.offset
1401
1402         switch (extent_entry_type(to_entry(dst))) {
1403         case BCH_EXTENT_ENTRY_crc32:
1404                 set_common_fields(dst->crc32, src);
1405                 dst->crc32.csum  = *((__le32 *) &src.csum.lo);
1406                 break;
1407         case BCH_EXTENT_ENTRY_crc64:
1408                 set_common_fields(dst->crc64, src);
1409                 dst->crc64.nonce        = src.nonce;
1410                 dst->crc64.csum_lo      = src.csum.lo;
1411                 dst->crc64.csum_hi      = *((__le16 *) &src.csum.hi);
1412                 break;
1413         case BCH_EXTENT_ENTRY_crc128:
1414                 set_common_fields(dst->crc128, src);
1415                 dst->crc128.nonce       = src.nonce;
1416                 dst->crc128.csum        = src.csum;
1417                 break;
1418         default:
1419                 BUG();
1420         }
1421 #undef set_common_fields
1422 }
1423
1424 void bch2_extent_crc_append(struct bkey_i *k,
1425                             struct bch_extent_crc_unpacked new)
1426 {
1427         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
1428         union bch_extent_crc *crc = (void *) ptrs.end;
1429
1430         if (bch_crc_bytes[new.csum_type]        <= 4 &&
1431             new.uncompressed_size - 1           <= CRC32_SIZE_MAX &&
1432             new.nonce                           <= CRC32_NONCE_MAX)
1433                 crc->type = 1 << BCH_EXTENT_ENTRY_crc32;
1434         else if (bch_crc_bytes[new.csum_type]   <= 10 &&
1435                    new.uncompressed_size - 1    <= CRC64_SIZE_MAX &&
1436                    new.nonce                    <= CRC64_NONCE_MAX)
1437                 crc->type = 1 << BCH_EXTENT_ENTRY_crc64;
1438         else if (bch_crc_bytes[new.csum_type]   <= 16 &&
1439                    new.uncompressed_size - 1    <= CRC128_SIZE_MAX &&
1440                    new.nonce                    <= CRC128_NONCE_MAX)
1441                 crc->type = 1 << BCH_EXTENT_ENTRY_crc128;
1442         else
1443                 BUG();
1444
1445         bch2_extent_crc_pack(crc, new);
1446
1447         k->k.u64s += extent_entry_u64s(ptrs.end);
1448
1449         EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
1450 }
1451
1452 static inline void __extent_entry_insert(struct bkey_i *k,
1453                                          union bch_extent_entry *dst,
1454                                          union bch_extent_entry *new)
1455 {
1456         union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
1457
1458         memmove_u64s_up_small((u64 *) dst + extent_entry_u64s(new),
1459                               dst, (u64 *) end - (u64 *) dst);
1460         k->k.u64s += extent_entry_u64s(new);
1461         memcpy(dst, new, extent_entry_bytes(new));
1462 }
1463
1464 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
1465                                     struct extent_ptr_decoded *p)
1466 {
1467         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
1468         struct bch_extent_crc_unpacked crc =
1469                 bch2_extent_crc_unpack(&k->k, NULL);
1470         union bch_extent_entry *pos;
1471
1472         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1473                 pos = ptrs.start;
1474                 goto found;
1475         }
1476
1477         bkey_for_each_crc(&k->k, ptrs, crc, pos)
1478                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1479                         pos = extent_entry_next(pos);
1480                         goto found;
1481                 }
1482
1483         bch2_extent_crc_append(k, p->crc);
1484         pos = bkey_val_end(bkey_i_to_s(k));
1485 found:
1486         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
1487         __extent_entry_insert(k, pos, to_entry(&p->ptr));
1488
1489         if (p->has_ec) {
1490                 p->ec.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
1491                 __extent_entry_insert(k, pos, to_entry(&p->ec));
1492         }
1493 }
1494
1495 /*
1496  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
1497  *
1498  * Returns true if @k should be dropped entirely
1499  *
1500  * For existing keys, only called when btree nodes are being rewritten, not when
1501  * they're merely being compacted/resorted in memory.
1502  */
1503 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1504 {
1505         struct bch_extent_ptr *ptr;
1506
1507         bch2_bkey_drop_ptrs(k, ptr,
1508                 ptr->cached &&
1509                 ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
1510
1511         /* will only happen if all pointers were cached: */
1512         if (!bkey_val_u64s(k.k))
1513                 k.k->type = KEY_TYPE_discard;
1514
1515         return bkey_whiteout(k.k);
1516 }
1517
1518 void bch2_bkey_mark_replicas_cached(struct bch_fs *c, struct bkey_s k,
1519                                     unsigned target,
1520                                     unsigned nr_desired_replicas)
1521 {
1522         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1523         union bch_extent_entry *entry;
1524         struct extent_ptr_decoded p;
1525         int extra = bch2_bkey_durability(c, k.s_c) - nr_desired_replicas;
1526
1527         if (target && extra > 0)
1528                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1529                         int n = bch2_extent_ptr_durability(c, p);
1530
1531                         if (n && n <= extra &&
1532                             !bch2_dev_in_target(c, p.ptr.dev, target)) {
1533                                 entry->ptr.cached = true;
1534                                 extra -= n;
1535                         }
1536                 }
1537
1538         if (extra > 0)
1539                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1540                         int n = bch2_extent_ptr_durability(c, p);
1541
1542                         if (n && n <= extra) {
1543                                 entry->ptr.cached = true;
1544                                 extra -= n;
1545                         }
1546                 }
1547 }
1548
1549 enum merge_result bch2_extent_merge(struct bch_fs *c,
1550                                     struct bkey_s _l, struct bkey_s _r)
1551 {
1552         struct bkey_s_extent l = bkey_s_to_extent(_l);
1553         struct bkey_s_extent r = bkey_s_to_extent(_r);
1554         union bch_extent_entry *en_l = l.v->start;
1555         union bch_extent_entry *en_r = r.v->start;
1556         struct bch_extent_crc_unpacked crc_l, crc_r;
1557
1558         if (bkey_val_u64s(l.k) != bkey_val_u64s(r.k))
1559                 return BCH_MERGE_NOMERGE;
1560
1561         crc_l = bch2_extent_crc_unpack(l.k, NULL);
1562
1563         extent_for_each_entry(l, en_l) {
1564                 en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
1565
1566                 if (extent_entry_type(en_l) != extent_entry_type(en_r))
1567                         return BCH_MERGE_NOMERGE;
1568
1569                 switch (extent_entry_type(en_l)) {
1570                 case BCH_EXTENT_ENTRY_ptr: {
1571                         const struct bch_extent_ptr *lp = &en_l->ptr;
1572                         const struct bch_extent_ptr *rp = &en_r->ptr;
1573                         struct bch_dev *ca;
1574
1575                         if (lp->offset + crc_l.compressed_size != rp->offset ||
1576                             lp->dev                     != rp->dev ||
1577                             lp->gen                     != rp->gen)
1578                                 return BCH_MERGE_NOMERGE;
1579
1580                         /* We don't allow extents to straddle buckets: */
1581                         ca = bch_dev_bkey_exists(c, lp->dev);
1582
1583                         if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
1584                                 return BCH_MERGE_NOMERGE;
1585
1586                         break;
1587                 }
1588                 case BCH_EXTENT_ENTRY_stripe_ptr:
1589                         if (en_l->stripe_ptr.block      != en_r->stripe_ptr.block ||
1590                             en_l->stripe_ptr.idx        != en_r->stripe_ptr.idx)
1591                                 return BCH_MERGE_NOMERGE;
1592                         break;
1593                 case BCH_EXTENT_ENTRY_crc32:
1594                 case BCH_EXTENT_ENTRY_crc64:
1595                 case BCH_EXTENT_ENTRY_crc128:
1596                         crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
1597                         crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
1598
1599                         if (crc_l.csum_type             != crc_r.csum_type ||
1600                             crc_l.compression_type      != crc_r.compression_type ||
1601                             crc_l.nonce                 != crc_r.nonce)
1602                                 return BCH_MERGE_NOMERGE;
1603
1604                         if (crc_l.offset + crc_l.live_size != crc_l.compressed_size ||
1605                             crc_r.offset)
1606                                 return BCH_MERGE_NOMERGE;
1607
1608                         if (!bch2_checksum_mergeable(crc_l.csum_type))
1609                                 return BCH_MERGE_NOMERGE;
1610
1611                         if (crc_l.compression_type)
1612                                 return BCH_MERGE_NOMERGE;
1613
1614                         if (crc_l.csum_type &&
1615                             crc_l.uncompressed_size +
1616                             crc_r.uncompressed_size > c->sb.encoded_extent_max)
1617                                 return BCH_MERGE_NOMERGE;
1618
1619                         if (crc_l.uncompressed_size + crc_r.uncompressed_size - 1 >
1620                             bch2_crc_field_size_max[extent_entry_type(en_l)])
1621                                 return BCH_MERGE_NOMERGE;
1622
1623                         break;
1624                 default:
1625                         return BCH_MERGE_NOMERGE;
1626                 }
1627         }
1628
1629         extent_for_each_entry(l, en_l) {
1630                 struct bch_extent_crc_unpacked crc_l, crc_r;
1631
1632                 en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
1633
1634                 if (!extent_entry_is_crc(en_l))
1635                         continue;
1636
1637                 crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
1638                 crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
1639
1640                 crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
1641                                                  crc_l.csum,
1642                                                  crc_r.csum,
1643                                                  crc_r.uncompressed_size << 9);
1644
1645                 crc_l.uncompressed_size += crc_r.uncompressed_size;
1646                 crc_l.compressed_size   += crc_r.compressed_size;
1647
1648                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_l);
1649         }
1650
1651         bch2_key_resize(l.k, l.k->size + r.k->size);
1652
1653         return BCH_MERGE_MERGE;
1654 }
1655
1656 bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
1657                                unsigned nr_replicas)
1658 {
1659         struct btree_trans trans;
1660         struct btree_iter *iter;
1661         struct bpos end = pos;
1662         struct bkey_s_c k;
1663         bool ret = true;
1664         int err;
1665
1666         end.offset += size;
1667
1668         bch2_trans_init(&trans, c, 0, 0);
1669
1670         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, pos,
1671                            BTREE_ITER_SLOTS, k, err) {
1672                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
1673                         break;
1674
1675                 if (nr_replicas > bch2_bkey_nr_ptrs_allocated(k)) {
1676                         ret = false;
1677                         break;
1678                 }
1679         }
1680         bch2_trans_exit(&trans);
1681
1682         return ret;
1683 }
1684
1685 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
1686 {
1687         unsigned ret = 0;
1688
1689         switch (k.k->type) {
1690         case KEY_TYPE_extent: {
1691                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1692                 const union bch_extent_entry *entry;
1693                 struct extent_ptr_decoded p;
1694
1695                 extent_for_each_ptr_decode(e, p, entry)
1696                         ret += !p.ptr.cached &&
1697                                 p.crc.compression_type == BCH_COMPRESSION_NONE;
1698                 break;
1699         }
1700         case KEY_TYPE_reservation:
1701                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
1702                 break;
1703         }
1704
1705         return ret;
1706 }
1707
1708 /* KEY_TYPE_reservation: */
1709
1710 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
1711 {
1712         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1713
1714         if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
1715                 return "incorrect value size";
1716
1717         if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
1718                 return "invalid nr_replicas";
1719
1720         return NULL;
1721 }
1722
1723 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
1724                               struct bkey_s_c k)
1725 {
1726         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1727
1728         pr_buf(out, "generation %u replicas %u",
1729                le32_to_cpu(r.v->generation),
1730                r.v->nr_replicas);
1731 }
1732
1733 enum merge_result bch2_reservation_merge(struct bch_fs *c,
1734                                          struct bkey_s _l, struct bkey_s _r)
1735 {
1736         struct bkey_s_reservation l = bkey_s_to_reservation(_l);
1737         struct bkey_s_reservation r = bkey_s_to_reservation(_r);
1738
1739         if (l.v->generation != r.v->generation ||
1740             l.v->nr_replicas != r.v->nr_replicas)
1741                 return BCH_MERGE_NOMERGE;
1742
1743         if ((u64) l.k->size + r.k->size > KEY_SIZE_MAX) {
1744                 bch2_key_resize(l.k, KEY_SIZE_MAX);
1745                 __bch2_cut_front(l.k->p, r.s);
1746                 return BCH_MERGE_PARTIAL;
1747         }
1748
1749         bch2_key_resize(l.k, l.k->size + r.k->size);
1750
1751         return BCH_MERGE_MERGE;
1752 }