git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/extents.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  *
   4  * Code for managing the extent btree and dynamically updating the writeback
   5  * dirty sector count.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "bkey_methods.h"
  10 #include "btree_gc.h"
  11 #include "btree_update.h"
  12 #include "btree_update_interior.h"
  13 #include "buckets.h"
  14 #include "checksum.h"
  15 #include "debug.h"
  16 #include "dirent.h"
  17 #include "disk_groups.h"
  18 #include "error.h"
  19 #include "extents.h"
  20 #include "inode.h"
  21 #include "journal.h"
  22 #include "replicas.h"
  23 #include "super.h"
  24 #include "super-io.h"
  25 #include "util.h"
  26 #include "xattr.h"
  27
  28 #include <trace/events/bcachefs.h>
  29
  30 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
  31 {
  32         struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
  33         const struct bch_extent_ptr *ptr;
  34         unsigned nr_ptrs = 0;
  35
  36         bkey_for_each_ptr(p, ptr)
  37                 nr_ptrs++;
  38
  39         return nr_ptrs;
  40 }
  41
  42 unsigned bch2_bkey_nr_dirty_ptrs(struct bkey_s_c k)
  43 {
  44         unsigned nr_ptrs = 0;
  45
  46         switch (k.k->type) {
  47         case KEY_TYPE_btree_ptr:
  48         case KEY_TYPE_extent: {
  49                 struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
  50                 const struct bch_extent_ptr *ptr;
  51
  52                 bkey_for_each_ptr(p, ptr)
  53                         nr_ptrs += !ptr->cached;
  54                 BUG_ON(!nr_ptrs);
  55                 break;
  56         }
  57         case KEY_TYPE_reservation:
  58                 nr_ptrs = bkey_s_c_to_reservation(k).v->nr_replicas;
  59                 break;
  60         }
  61
  62         return nr_ptrs;
  63 }
  64
  65 static unsigned bch2_extent_ptr_durability(struct bch_fs *c,
  66                                            struct extent_ptr_decoded p)
  67 {
  68         unsigned i, durability = 0;
  69         struct bch_dev *ca;
  70
  71         if (p.ptr.cached)
  72                 return 0;
  73
  74         ca = bch_dev_bkey_exists(c, p.ptr.dev);
  75
  76         if (ca->mi.state != BCH_MEMBER_STATE_FAILED)
  77                 durability = max_t(unsigned, durability, ca->mi.durability);
  78
  79         for (i = 0; i < p.ec_nr; i++) {
  80                 struct stripe *s =
  81                         genradix_ptr(&c->stripes[0], p.idx);
  82
  83                 if (WARN_ON(!s))
  84                         continue;
  85
  86                 durability = max_t(unsigned, durability, s->nr_redundant);
  87         }
  88
  89         return durability;
  90 }
  91
  92 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
  93 {
  94         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  95         const union bch_extent_entry *entry;
  96         struct extent_ptr_decoded p;
  97         unsigned durability = 0;
  98
  99         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 100                 durability += bch2_extent_ptr_durability(c, p);
 101
 102         return durability;
 103 }
 104
 105 static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f,
 106                                                    unsigned dev)
 107 {
 108         struct bch_dev_io_failures *i;
 109
 110         for (i = f->devs; i < f->devs + f->nr; i++)
 111                 if (i->dev == dev)
 112                         return i;
 113
 114         return NULL;
 115 }
 116
 117 void bch2_mark_io_failure(struct bch_io_failures *failed,
 118                           struct extent_ptr_decoded *p)
 119 {
 120         struct bch_dev_io_failures *f = dev_io_failures(failed, p->ptr.dev);
 121
 122         if (!f) {
 123                 BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
 124
 125                 f = &failed->devs[failed->nr++];
 126                 f->dev          = p->ptr.dev;
 127                 f->idx          = p->idx;
 128                 f->nr_failed    = 1;
 129                 f->nr_retries   = 0;
 130         } else if (p->idx != f->idx) {
 131                 f->idx          = p->idx;
 132                 f->nr_failed    = 1;
 133                 f->nr_retries   = 0;
 134         } else {
 135                 f->nr_failed++;
 136         }
 137 }
 138
 139 /*
 140  * returns true if p1 is better than p2:
 141  */
 142 static inline bool ptr_better(struct bch_fs *c,
 143                               const struct extent_ptr_decoded p1,
 144                               const struct extent_ptr_decoded p2)
 145 {
 146         if (likely(!p1.idx && !p2.idx)) {
 147                 struct bch_dev *dev1 = bch_dev_bkey_exists(c, p1.ptr.dev);
 148                 struct bch_dev *dev2 = bch_dev_bkey_exists(c, p2.ptr.dev);
 149
 150                 u64 l1 = atomic64_read(&dev1->cur_latency[READ]);
 151                 u64 l2 = atomic64_read(&dev2->cur_latency[READ]);
 152
 153                 /* Pick at random, biased in favor of the faster device: */
 154
 155                 return bch2_rand_range(l1 + l2) > l1;
 156         }
 157
 158         if (force_reconstruct_read(c))
 159                 return p1.idx > p2.idx;
 160
 161         return p1.idx < p2.idx;
 162 }
 163
 164 /*
 165  * This picks a non-stale pointer, preferably from a device other than @avoid.
 166  * Avoid can be NULL, meaning pick any. If there are no non-stale pointers to
 167  * other devices, it will still pick a pointer from avoid.
 168  */
 169 int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 170                                struct bch_io_failures *failed,
 171                                struct extent_ptr_decoded *pick)
 172 {
 173         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 174         const union bch_extent_entry *entry;
 175         struct extent_ptr_decoded p;
 176         struct bch_dev_io_failures *f;
 177         struct bch_dev *ca;
 178         int ret = 0;
 179
 180         if (k.k->type == KEY_TYPE_error)
 181                 return -EIO;
 182
 183         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 184                 ca = bch_dev_bkey_exists(c, p.ptr.dev);
 185
 186                 /*
 187                  * If there are any dirty pointers it's an error if we can't
 188                  * read:
 189                  */
 190                 if (!ret && !p.ptr.cached)
 191                         ret = -EIO;
 192
 193                 if (p.ptr.cached && ptr_stale(ca, &p.ptr))
 194                         continue;
 195
 196                 f = failed ? dev_io_failures(failed, p.ptr.dev) : NULL;
 197                 if (f)
 198                         p.idx = f->nr_failed < f->nr_retries
 199                                 ? f->idx
 200                                 : f->idx + 1;
 201
 202                 if (!p.idx &&
 203                     !bch2_dev_is_readable(ca))
 204                         p.idx++;
 205
 206                 if (force_reconstruct_read(c) &&
 207                     !p.idx && p.ec_nr)
 208                         p.idx++;
 209
 210                 if (p.idx >= p.ec_nr + 1)
 211                         continue;
 212
 213                 if (ret > 0 && !ptr_better(c, p, *pick))
 214                         continue;
 215
 216                 *pick = p;
 217                 ret = 1;
 218         }
 219
 220         return ret;
 221 }
 222
 223 void bch2_bkey_append_ptr(struct bkey_i *k,
 224                           struct bch_extent_ptr ptr)
 225 {
 226         EBUG_ON(bch2_bkey_has_device(bkey_i_to_s_c(k), ptr.dev));
 227
 228         switch (k->k.type) {
 229         case KEY_TYPE_btree_ptr:
 230         case KEY_TYPE_extent:
 231                 EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
 232
 233                 ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 234
 235                 memcpy((void *) &k->v + bkey_val_bytes(&k->k),
 236                        &ptr,
 237                        sizeof(ptr));
 238                 k->u64s++;
 239                 break;
 240         default:
 241                 BUG();
 242         }
 243 }
 244
 245 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
 246 {
 247         struct bch_extent_ptr *ptr;
 248
 249         bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
 250 }
 251
 252 /* extent specific utility code */
 253
 254 const struct bch_extent_ptr *
 255 bch2_extent_has_device(struct bkey_s_c_extent e, unsigned dev)
 256 {
 257         const struct bch_extent_ptr *ptr;
 258
 259         extent_for_each_ptr(e, ptr)
 260                 if (ptr->dev == dev)
 261                         return ptr;
 262
 263         return NULL;
 264 }
 265
 266 const struct bch_extent_ptr *
 267 bch2_extent_has_group(struct bch_fs *c, struct bkey_s_c_extent e, unsigned group)
 268 {
 269         const struct bch_extent_ptr *ptr;
 270
 271         extent_for_each_ptr(e, ptr) {
 272                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 273
 274                 if (ca->mi.group &&
 275                     ca->mi.group - 1 == group)
 276                         return ptr;
 277         }
 278
 279         return NULL;
 280 }
 281
 282 const struct bch_extent_ptr *
 283 bch2_extent_has_target(struct bch_fs *c, struct bkey_s_c_extent e, unsigned target)
 284 {
 285         const struct bch_extent_ptr *ptr;
 286
 287         extent_for_each_ptr(e, ptr)
 288                 if (bch2_dev_in_target(c, ptr->dev, target) &&
 289                     (!ptr->cached ||
 290                      !ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr)))
 291                         return ptr;
 292
 293         return NULL;
 294 }
 295
 296 unsigned bch2_extent_is_compressed(struct bkey_s_c k)
 297 {
 298         unsigned ret = 0;
 299
 300         switch (k.k->type) {
 301         case KEY_TYPE_extent: {
 302                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 303                 const union bch_extent_entry *entry;
 304                 struct extent_ptr_decoded p;
 305
 306                 extent_for_each_ptr_decode(e, p, entry)
 307                         if (!p.ptr.cached &&
 308                             p.crc.compression_type != BCH_COMPRESSION_NONE &&
 309                             p.crc.compressed_size < p.crc.live_size)
 310                                 ret += p.crc.compressed_size;
 311         }
 312         }
 313
 314         return ret;
 315 }
 316
 317 bool bch2_extent_matches_ptr(struct bch_fs *c, struct bkey_s_c_extent e,
 318                              struct bch_extent_ptr m, u64 offset)
 319 {
 320         const union bch_extent_entry *entry;
 321         struct extent_ptr_decoded p;
 322
 323         extent_for_each_ptr_decode(e, p, entry)
 324                 if (p.ptr.dev   == m.dev &&
 325                     p.ptr.gen   == m.gen &&
 326                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(e.k) ==
 327                     (s64) m.offset  - offset)
 328                         return true;
 329
 330         return false;
 331 }
 332
 333 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 334                                           union bch_extent_entry *entry)
 335 {
 336         union bch_extent_entry *i = ptrs.start;
 337
 338         if (i == entry)
 339                 return NULL;
 340
 341         while (extent_entry_next(i) != entry)
 342                 i = extent_entry_next(i);
 343         return i;
 344 }
 345
 346 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
 347                                            struct bch_extent_ptr *ptr)
 348 {
 349         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 350         union bch_extent_entry *dst, *src, *prev;
 351         bool drop_crc = true;
 352
 353         EBUG_ON(ptr < &ptrs.start->ptr ||
 354                 ptr >= &ptrs.end->ptr);
 355         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 356
 357         src = extent_entry_next(to_entry(ptr));
 358         if (src != ptrs.end &&
 359             !extent_entry_is_crc(src))
 360                 drop_crc = false;
 361
 362         dst = to_entry(ptr);
 363         while ((prev = extent_entry_prev(ptrs, dst))) {
 364                 if (extent_entry_is_ptr(prev))
 365                         break;
 366
 367                 if (extent_entry_is_crc(prev)) {
 368                         if (drop_crc)
 369                                 dst = prev;
 370                         break;
 371                 }
 372
 373                 dst = prev;
 374         }
 375
 376         memmove_u64s_down(dst, src,
 377                           (u64 *) ptrs.end - (u64 *) src);
 378         k.k->u64s -= (u64 *) src - (u64 *) dst;
 379
 380         return dst;
 381 }
 382
 383 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 384                                   struct bch_extent_crc_unpacked n)
 385 {
 386         return !u.compression_type &&
 387                 u.csum_type &&
 388                 u.uncompressed_size > u.live_size &&
 389                 bch2_csum_type_is_encryption(u.csum_type) ==
 390                 bch2_csum_type_is_encryption(n.csum_type);
 391 }
 392
 393 bool bch2_can_narrow_extent_crcs(struct bkey_s_c_extent e,
 394                                  struct bch_extent_crc_unpacked n)
 395 {
 396         struct bch_extent_crc_unpacked crc;
 397         const union bch_extent_entry *i;
 398
 399         if (!n.csum_type)
 400                 return false;
 401
 402         extent_for_each_crc(e, crc, i)
 403                 if (can_narrow_crc(crc, n))
 404                         return true;
 405
 406         return false;
 407 }
 408
 409 /*
 410  * We're writing another replica for this extent, so while we've got the data in
 411  * memory we'll be computing a new checksum for the currently live data.
 412  *
 413  * If there are other replicas we aren't moving, and they are checksummed but
 414  * not compressed, we can modify them to point to only the data that is
 415  * currently live (so that readers won't have to bounce) while we've got the
 416  * checksum we need:
 417  */
 418 bool bch2_extent_narrow_crcs(struct bkey_i_extent *e,
 419                              struct bch_extent_crc_unpacked n)
 420 {
 421         struct bch_extent_crc_unpacked u;
 422         struct extent_ptr_decoded p;
 423         union bch_extent_entry *i;
 424         bool ret = false;
 425
 426         /* Find a checksum entry that covers only live data: */
 427         if (!n.csum_type) {
 428                 extent_for_each_crc(extent_i_to_s(e), u, i)
 429                         if (!u.compression_type &&
 430                             u.csum_type &&
 431                             u.live_size == u.uncompressed_size) {
 432                                 n = u;
 433                                 goto found;
 434                         }
 435                 return false;
 436         }
 437 found:
 438         BUG_ON(n.compression_type);
 439         BUG_ON(n.offset);
 440         BUG_ON(n.live_size != e->k.size);
 441
 442 restart_narrow_pointers:
 443         extent_for_each_ptr_decode(extent_i_to_s(e), p, i)
 444                 if (can_narrow_crc(p.crc, n)) {
 445                         bch2_bkey_drop_ptr(extent_i_to_s(e).s, &i->ptr);
 446                         p.ptr.offset += p.crc.offset;
 447                         p.crc = n;
 448                         bch2_extent_ptr_decoded_append(e, &p);
 449                         ret = true;
 450                         goto restart_narrow_pointers;
 451                 }
 452
 453         return ret;
 454 }
 455
 456 /* returns true if not equal */
 457 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 458                                          struct bch_extent_crc_unpacked r)
 459 {
 460         return (l.csum_type             != r.csum_type ||
 461                 l.compression_type      != r.compression_type ||
 462                 l.compressed_size       != r.compressed_size ||
 463                 l.uncompressed_size     != r.uncompressed_size ||
 464                 l.offset                != r.offset ||
 465                 l.live_size             != r.live_size ||
 466                 l.nonce                 != r.nonce ||
 467                 bch2_crc_cmp(l.csum, r.csum));
 468 }
 469
 470 void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
 471 {
 472         union bch_extent_entry *entry;
 473         u64 *d = (u64 *) bkeyp_val(f, k);
 474         unsigned i;
 475
 476         for (i = 0; i < bkeyp_val_u64s(f, k); i++)
 477                 d[i] = swab64(d[i]);
 478
 479         for (entry = (union bch_extent_entry *) d;
 480              entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
 481              entry = extent_entry_next(entry)) {
 482                 switch (extent_entry_type(entry)) {
 483                 case BCH_EXTENT_ENTRY_ptr:
 484                         break;
 485                 case BCH_EXTENT_ENTRY_crc32:
 486                         entry->crc32.csum = swab32(entry->crc32.csum);
 487                         break;
 488                 case BCH_EXTENT_ENTRY_crc64:
 489                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
 490                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
 491                         break;
 492                 case BCH_EXTENT_ENTRY_crc128:
 493                         entry->crc128.csum.hi = (__force __le64)
 494                                 swab64((__force u64) entry->crc128.csum.hi);
 495                         entry->crc128.csum.lo = (__force __le64)
 496                                 swab64((__force u64) entry->crc128.csum.lo);
 497                         break;
 498                 case BCH_EXTENT_ENTRY_stripe_ptr:
 499                         break;
 500                 }
 501         }
 502 }
 503
 504 static const char *extent_ptr_invalid(const struct bch_fs *c,
 505                                       struct bkey_s_c k,
 506                                       const struct bch_extent_ptr *ptr,
 507                                       unsigned size_ondisk,
 508                                       bool metadata)
 509 {
 510         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 511         const struct bch_extent_ptr *ptr2;
 512         struct bch_dev *ca;
 513
 514         if (ptr->dev >= c->sb.nr_devices ||
 515             !c->devs[ptr->dev])
 516                 return "pointer to invalid device";
 517
 518         ca = bch_dev_bkey_exists(c, ptr->dev);
 519         if (!ca)
 520                 return "pointer to invalid device";
 521
 522         bkey_for_each_ptr(ptrs, ptr2)
 523                 if (ptr != ptr2 && ptr->dev == ptr2->dev)
 524                         return "multiple pointers to same device";
 525
 526         if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
 527                 return "offset past end of device";
 528
 529         if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
 530                 return "offset before first bucket";
 531
 532         if (bucket_remainder(ca, ptr->offset) +
 533             size_ondisk > ca->mi.bucket_size)
 534                 return "spans multiple buckets";
 535
 536         return NULL;
 537 }
 538
 539 static void bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 540                               struct bkey_s_c k)
 541 {
 542         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 543         const union bch_extent_entry *entry;
 544         struct bch_extent_crc_unpacked crc;
 545         const struct bch_extent_ptr *ptr;
 546         const struct bch_extent_stripe_ptr *ec;
 547         struct bch_dev *ca;
 548         bool first = true;
 549
 550         bkey_extent_entry_for_each(ptrs, entry) {
 551                 if (!first)
 552                         pr_buf(out, " ");
 553
 554                 switch (__extent_entry_type(entry)) {
 555                 case BCH_EXTENT_ENTRY_ptr:
 556                         ptr = entry_to_ptr(entry);
 557                         ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
 558                                 ? bch_dev_bkey_exists(c, ptr->dev)
 559                                 : NULL;
 560
 561                         pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
 562                                (u64) ptr->offset, ptr->gen,
 563                                ptr->cached ? " cached" : "",
 564                                ca && ptr_stale(ca, ptr)
 565                                ? " stale" : "");
 566                         break;
 567                 case BCH_EXTENT_ENTRY_crc32:
 568                 case BCH_EXTENT_ENTRY_crc64:
 569                 case BCH_EXTENT_ENTRY_crc128:
 570                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 571
 572                         pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
 573                                crc.compressed_size,
 574                                crc.uncompressed_size,
 575                                crc.offset, crc.nonce,
 576                                crc.csum_type,
 577                                crc.compression_type);
 578                         break;
 579                 case BCH_EXTENT_ENTRY_stripe_ptr:
 580                         ec = &entry->stripe_ptr;
 581
 582                         pr_buf(out, "ec: idx %llu block %u",
 583                                (u64) ec->idx, ec->block);
 584                         break;
 585                 default:
 586                         pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
 587                         return;
 588                 }
 589
 590                 first = false;
 591         }
 592 }
 593
 594 /* Btree ptrs */
 595
 596 const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
 597 {
 598         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 599         const union bch_extent_entry *entry;
 600         const struct bch_extent_ptr *ptr;
 601         const char *reason;
 602
 603         if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
 604                 return "value too big";
 605
 606         bkey_extent_entry_for_each(ptrs, entry) {
 607                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
 608                         return "invalid extent entry type";
 609
 610                 if (!extent_entry_is_ptr(entry))
 611                         return "has non ptr field";
 612         }
 613
 614         bkey_for_each_ptr(ptrs, ptr) {
 615                 reason = extent_ptr_invalid(c, k, ptr,
 616                                             c->opts.btree_node_size,
 617                                             true);
 618                 if (reason)
 619                         return reason;
 620         }
 621
 622         return NULL;
 623 }
 624
 625 void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct btree *b,
 626                                struct bkey_s_c k)
 627 {
 628         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 629         const struct bch_extent_ptr *ptr;
 630         unsigned seq;
 631         const char *err;
 632         char buf[160];
 633         struct bucket_mark mark;
 634         struct bch_dev *ca;
 635         unsigned replicas = 0;
 636         bool bad;
 637
 638         bkey_for_each_ptr(ptrs, ptr) {
 639                 ca = bch_dev_bkey_exists(c, ptr->dev);
 640                 replicas++;
 641
 642                 if (!test_bit(BCH_FS_ALLOC_READ_DONE, &c->flags))
 643                         continue;
 644
 645                 err = "stale";
 646                 if (ptr_stale(ca, ptr))
 647                         goto err;
 648
 649                 do {
 650                         seq = read_seqcount_begin(&c->gc_pos_lock);
 651                         mark = ptr_bucket_mark(ca, ptr);
 652
 653                         bad = gc_pos_cmp(c->gc_pos, gc_pos_btree_node(b)) > 0 &&
 654                                 (mark.data_type != BCH_DATA_BTREE ||
 655                                  mark.dirty_sectors < c->opts.btree_node_size);
 656                 } while (read_seqcount_retry(&c->gc_pos_lock, seq));
 657
 658                 err = "inconsistent";
 659                 if (bad)
 660                         goto err;
 661         }
 662
 663         if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 664             !bch2_bkey_replicas_marked(c, k, false)) {
 665                 bch2_bkey_val_to_text(&PBUF(buf), c, k);
 666                 bch2_fs_bug(c,
 667                         "btree key bad (replicas not marked in superblock):\n%s",
 668                         buf);
 669                 return;
 670         }
 671
 672         return;
 673 err:
 674         bch2_bkey_val_to_text(&PBUF(buf), c, k);
 675         bch2_fs_bug(c, "%s btree pointer %s: bucket %zi gen %i mark %08x",
 676                     err, buf, PTR_BUCKET_NR(ca, ptr),
 677                     mark.gen, (unsigned) mark.v.counter);
 678 }
 679
 680 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 681                             struct bkey_s_c k)
 682 {
 683         const char *invalid;
 684
 685         bkey_ptrs_to_text(out, c, k);
 686
 687         invalid = bch2_btree_ptr_invalid(c, k);
 688         if (invalid)
 689                 pr_buf(out, " invalid: %s", invalid);
 690 }
 691
 692 /* Extents */
 693
 694 bool __bch2_cut_front(struct bpos where, struct bkey_s k)
 695 {
 696         u64 len = 0;
 697
 698         if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
 699                 return false;
 700
 701         EBUG_ON(bkey_cmp(where, k.k->p) > 0);
 702
 703         len = k.k->p.offset - where.offset;
 704
 705         BUG_ON(len > k.k->size);
 706
 707         /*
 708          * Don't readjust offset if the key size is now 0, because that could
 709          * cause offset to point to the next bucket:
 710          */
 711         if (!len)
 712                 k.k->type = KEY_TYPE_deleted;
 713         else if (bkey_extent_is_data(k.k)) {
 714                 struct bkey_s_extent e = bkey_s_to_extent(k);
 715                 union bch_extent_entry *entry;
 716                 bool seen_crc = false;
 717
 718                 extent_for_each_entry(e, entry) {
 719                         switch (extent_entry_type(entry)) {
 720                         case BCH_EXTENT_ENTRY_ptr:
 721                                 if (!seen_crc)
 722                                         entry->ptr.offset += e.k->size - len;
 723                                 break;
 724                         case BCH_EXTENT_ENTRY_crc32:
 725                                 entry->crc32.offset += e.k->size - len;
 726                                 break;
 727                         case BCH_EXTENT_ENTRY_crc64:
 728                                 entry->crc64.offset += e.k->size - len;
 729                                 break;
 730                         case BCH_EXTENT_ENTRY_crc128:
 731                                 entry->crc128.offset += e.k->size - len;
 732                                 break;
 733                         case BCH_EXTENT_ENTRY_stripe_ptr:
 734                                 break;
 735                         }
 736
 737                         if (extent_entry_is_crc(entry))
 738                                 seen_crc = true;
 739                 }
 740         }
 741
 742         k.k->size = len;
 743
 744         return true;
 745 }
 746
 747 bool bch2_cut_back(struct bpos where, struct bkey *k)
 748 {
 749         u64 len = 0;
 750
 751         if (bkey_cmp(where, k->p) >= 0)
 752                 return false;
 753
 754         EBUG_ON(bkey_cmp(where, bkey_start_pos(k)) < 0);
 755
 756         len = where.offset - bkey_start_offset(k);
 757
 758         BUG_ON(len > k->size);
 759
 760         k->p = where;
 761         k->size = len;
 762
 763         if (!len)
 764                 k->type = KEY_TYPE_deleted;
 765
 766         return true;
 767 }
 768
 769 /**
 770  * bch_key_resize - adjust size of @k
 771  *
 772  * bkey_start_offset(k) will be preserved, modifies where the extent ends
 773  */
 774 void bch2_key_resize(struct bkey *k,
 775                     unsigned new_size)
 776 {
 777         k->p.offset -= k->size;
 778         k->p.offset += new_size;
 779         k->size = new_size;
 780 }
 781
 782 static bool extent_i_save(struct btree *b, struct bkey_packed *dst,
 783                           struct bkey_i *src)
 784 {
 785         struct bkey_format *f = &b->format;
 786         struct bkey_i *dst_unpacked;
 787         struct bkey_packed tmp;
 788
 789         if ((dst_unpacked = packed_to_bkey(dst)))
 790                 dst_unpacked->k = src->k;
 791         else if (bch2_bkey_pack_key(&tmp, &src->k, f))
 792                 memcpy_u64s(dst, &tmp, f->key_u64s);
 793         else
 794                 return false;
 795
 796         memcpy_u64s(bkeyp_val(f, dst), &src->v, bkey_val_u64s(&src->k));
 797         return true;
 798 }
 799
 800 struct extent_insert_state {
 801         struct btree_insert             *trans;
 802         struct btree_insert_entry       *insert;
 803         struct bpos                     committed;
 804
 805         /* for deleting: */
 806         struct bkey_i                   whiteout;
 807         bool                            update_journal;
 808         bool                            update_btree;
 809         bool                            deleting;
 810 };
 811
 812 static bool bch2_extent_merge_inline(struct bch_fs *,
 813                                      struct btree_iter *,
 814                                      struct bkey_packed *,
 815                                      struct bkey_packed *,
 816                                      bool);
 817
 818 static void verify_extent_nonoverlapping(struct btree *b,
 819                                          struct btree_node_iter *_iter,
 820                                          struct bkey_i *insert)
 821 {
 822 #ifdef CONFIG_BCACHEFS_DEBUG
 823         struct btree_node_iter iter;
 824         struct bkey_packed *k;
 825         struct bkey uk;
 826
 827         iter = *_iter;
 828         k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
 829         BUG_ON(k &&
 830                (uk = bkey_unpack_key(b, k),
 831                 bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
 832
 833         iter = *_iter;
 834         k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
 835 #if 0
 836         BUG_ON(k &&
 837                (uk = bkey_unpack_key(b, k),
 838                 bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
 839 #else
 840         if (k &&
 841             (uk = bkey_unpack_key(b, k),
 842              bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
 843                 char buf1[100];
 844                 char buf2[100];
 845
 846                 bch2_bkey_to_text(&PBUF(buf1), &insert->k);
 847                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 848
 849                 bch2_dump_btree_node(b);
 850                 panic("insert > next :\n"
 851                       "insert %s\n"
 852                       "next   %s\n",
 853                       buf1, buf2);
 854         }
 855 #endif
 856
 857 #endif
 858 }
 859
 860 static void verify_modified_extent(struct btree_iter *iter,
 861                                    struct bkey_packed *k)
 862 {
 863         bch2_btree_iter_verify(iter, iter->l[0].b);
 864         bch2_verify_insert_pos(iter->l[0].b, k, k, k->u64s);
 865 }
 866
 867 static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 868                                struct bkey_i *insert)
 869 {
 870         struct btree_iter_level *l = &iter->l[0];
 871         struct btree_node_iter node_iter;
 872         struct bkey_packed *k;
 873
 874         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
 875
 876         EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
 877         verify_extent_nonoverlapping(l->b, &l->iter, insert);
 878
 879         node_iter = l->iter;
 880         k = bch2_btree_node_iter_prev_filter(&node_iter, l->b, KEY_TYPE_discard);
 881         if (k && !bkey_written(l->b, k) &&
 882             bch2_extent_merge_inline(c, iter, k, bkey_to_packed(insert), true))
 883                 return;
 884
 885         node_iter = l->iter;
 886         k = bch2_btree_node_iter_peek_filter(&node_iter, l->b, KEY_TYPE_discard);
 887         if (k && !bkey_written(l->b, k) &&
 888             bch2_extent_merge_inline(c, iter, bkey_to_packed(insert), k, false))
 889                 return;
 890
 891         k = bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
 892
 893         bch2_bset_insert(l->b, &l->iter, k, insert, 0);
 894         bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
 895         bch2_btree_iter_verify(iter, l->b);
 896 }
 897
 898 static void extent_insert_committed(struct extent_insert_state *s)
 899 {
 900         struct bch_fs *c = s->trans->c;
 901         struct btree_iter *iter = s->insert->iter;
 902         struct bkey_i *insert = s->insert->k;
 903         BKEY_PADDED(k) split;
 904
 905         EBUG_ON(bkey_cmp(insert->k.p, s->committed) < 0);
 906         EBUG_ON(bkey_cmp(s->committed, bkey_start_pos(&insert->k)) < 0);
 907
 908         bkey_copy(&split.k, insert);
 909         if (s->deleting)
 910                 split.k.k.type = KEY_TYPE_discard;
 911
 912         bch2_cut_back(s->committed, &split.k.k);
 913
 914         if (!bkey_cmp(s->committed, iter->pos))
 915                 return;
 916
 917         bch2_btree_iter_set_pos_same_leaf(iter, s->committed);
 918
 919         if (s->update_btree) {
 920                 if (debug_check_bkeys(c))
 921                         bch2_bkey_debugcheck(c, iter->l[0].b,
 922                                              bkey_i_to_s_c(&split.k));
 923
 924                 EBUG_ON(bkey_deleted(&split.k.k) || !split.k.k.size);
 925
 926                 extent_bset_insert(c, iter, &split.k);
 927         }
 928
 929         if (s->update_journal) {
 930                 bkey_copy(&split.k, !s->deleting ? insert : &s->whiteout);
 931                 if (s->deleting)
 932                         split.k.k.type = KEY_TYPE_discard;
 933
 934                 bch2_cut_back(s->committed, &split.k.k);
 935
 936                 EBUG_ON(bkey_deleted(&split.k.k) || !split.k.k.size);
 937
 938                 bch2_btree_journal_key(s->trans, iter, &split.k);
 939         }
 940
 941         bch2_cut_front(s->committed, insert);
 942
 943         insert->k.needs_whiteout        = false;
 944 }
 945
 946 void bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
 947 {
 948         struct btree *b = iter->l[0].b;
 949
 950         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 951
 952         bch2_cut_back(b->key.k.p, &k->k);
 953
 954         BUG_ON(bkey_cmp(bkey_start_pos(&k->k), b->data->min_key) < 0);
 955 }
 956
 957 enum btree_insert_ret
 958 bch2_extent_can_insert(struct btree_insert *trans,
 959                        struct btree_insert_entry *insert,
 960                        unsigned *u64s)
 961 {
 962         struct btree_iter_level *l = &insert->iter->l[0];
 963         struct btree_node_iter node_iter = l->iter;
 964         enum bch_extent_overlap overlap;
 965         struct bkey_packed *_k;
 966         struct bkey unpacked;
 967         struct bkey_s_c k;
 968         int sectors;
 969
 970         BUG_ON(trans->flags & BTREE_INSERT_ATOMIC &&
 971                !bch2_extent_is_atomic(&insert->k->k, insert->iter));
 972
 973         /*
 974          * We avoid creating whiteouts whenever possible when deleting, but
 975          * those optimizations mean we may potentially insert two whiteouts
 976          * instead of one (when we overlap with the front of one extent and the
 977          * back of another):
 978          */
 979         if (bkey_whiteout(&insert->k->k))
 980                 *u64s += BKEY_U64s;
 981
 982         _k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
 983                                               KEY_TYPE_discard);
 984         if (!_k)
 985                 return BTREE_INSERT_OK;
 986
 987         k = bkey_disassemble(l->b, _k, &unpacked);
 988
 989         overlap = bch2_extent_overlap(&insert->k->k, k.k);
 990
 991         /* account for having to split existing extent: */
 992         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
 993                 *u64s += _k->u64s;
 994
 995         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
 996             (sectors = bch2_extent_is_compressed(k))) {
 997                 int flags = BCH_DISK_RESERVATION_BTREE_LOCKS_HELD;
 998
 999                 if (trans->flags & BTREE_INSERT_NOFAIL)
1000                         flags |= BCH_DISK_RESERVATION_NOFAIL;
1001
1002                 switch (bch2_disk_reservation_add(trans->c,
1003                                 trans->disk_res,
1004                                 sectors, flags)) {
1005                 case 0:
1006                         break;
1007                 case -ENOSPC:
1008                         return BTREE_INSERT_ENOSPC;
1009                 case -EINTR:
1010                         return BTREE_INSERT_NEED_GC_LOCK;
1011                 default:
1012                         BUG();
1013                 }
1014         }
1015
1016         return BTREE_INSERT_OK;
1017 }
1018
1019 static void
1020 extent_squash(struct extent_insert_state *s, struct bkey_i *insert,
1021               struct bkey_packed *_k, struct bkey_s k,
1022               enum bch_extent_overlap overlap)
1023 {
1024         struct bch_fs *c = s->trans->c;
1025         struct btree_iter *iter = s->insert->iter;
1026         struct btree_iter_level *l = &iter->l[0];
1027
1028         switch (overlap) {
1029         case BCH_EXTENT_OVERLAP_FRONT:
1030                 /* insert overlaps with start of k: */
1031                 __bch2_cut_front(insert->k.p, k);
1032                 BUG_ON(bkey_deleted(k.k));
1033                 extent_save(l->b, _k, k.k);
1034                 verify_modified_extent(iter, _k);
1035                 break;
1036
1037         case BCH_EXTENT_OVERLAP_BACK:
1038                 /* insert overlaps with end of k: */
1039                 bch2_cut_back(bkey_start_pos(&insert->k), k.k);
1040                 BUG_ON(bkey_deleted(k.k));
1041                 extent_save(l->b, _k, k.k);
1042
1043                 /*
1044                  * As the auxiliary tree is indexed by the end of the
1045                  * key and we've just changed the end, update the
1046                  * auxiliary tree.
1047                  */
1048                 bch2_bset_fix_invalidated_key(l->b, _k);
1049                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1050                                          _k, _k->u64s, _k->u64s);
1051                 verify_modified_extent(iter, _k);
1052                 break;
1053
1054         case BCH_EXTENT_OVERLAP_ALL: {
1055                 /* The insert key completely covers k, invalidate k */
1056                 if (!bkey_whiteout(k.k))
1057                         btree_account_key_drop(l->b, _k);
1058
1059                 k.k->size = 0;
1060                 k.k->type = KEY_TYPE_deleted;
1061
1062                 if (_k >= btree_bset_last(l->b)->start) {
1063                         unsigned u64s = _k->u64s;
1064
1065                         bch2_bset_delete(l->b, _k, _k->u64s);
1066                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1067                                                  _k, u64s, 0);
1068                         bch2_btree_iter_verify(iter, l->b);
1069                 } else {
1070                         extent_save(l->b, _k, k.k);
1071                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1072                                                  _k, _k->u64s, _k->u64s);
1073                         verify_modified_extent(iter, _k);
1074                 }
1075
1076                 break;
1077         }
1078         case BCH_EXTENT_OVERLAP_MIDDLE: {
1079                 BKEY_PADDED(k) split;
1080                 /*
1081                  * The insert key falls 'in the middle' of k
1082                  * The insert key splits k in 3:
1083                  * - start only in k, preserve
1084                  * - middle common section, invalidate in k
1085                  * - end only in k, preserve
1086                  *
1087                  * We update the old key to preserve the start,
1088                  * insert will be the new common section,
1089                  * we manually insert the end that we are preserving.
1090                  *
1091                  * modify k _before_ doing the insert (which will move
1092                  * what k points to)
1093                  */
1094                 bkey_reassemble(&split.k, k.s_c);
1095                 split.k.k.needs_whiteout |= bkey_written(l->b, _k);
1096
1097                 bch2_cut_back(bkey_start_pos(&insert->k), &split.k.k);
1098                 BUG_ON(bkey_deleted(&split.k.k));
1099
1100                 __bch2_cut_front(insert->k.p, k);
1101                 BUG_ON(bkey_deleted(k.k));
1102                 extent_save(l->b, _k, k.k);
1103                 verify_modified_extent(iter, _k);
1104
1105                 extent_bset_insert(c, iter, &split.k);
1106                 break;
1107         }
1108         }
1109 }
1110
1111 static void __bch2_insert_fixup_extent(struct extent_insert_state *s)
1112 {
1113         struct btree_iter *iter = s->insert->iter;
1114         struct btree_iter_level *l = &iter->l[0];
1115         struct bkey_packed *_k;
1116         struct bkey unpacked;
1117         struct bkey_i *insert = s->insert->k;
1118
1119         while (bkey_cmp(s->committed, insert->k.p) < 0 &&
1120                (_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
1121                                                       KEY_TYPE_discard))) {
1122                 struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
1123                 enum bch_extent_overlap overlap = bch2_extent_overlap(&insert->k, k.k);
1124
1125                 EBUG_ON(bkey_cmp(iter->pos, k.k->p) >= 0);
1126
1127                 if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
1128                         break;
1129
1130                 s->committed = bpos_min(s->insert->k->k.p, k.k->p);
1131
1132                 if (!bkey_whiteout(k.k))
1133                         s->update_journal = true;
1134
1135                 if (!s->update_journal) {
1136                         bch2_cut_front(s->committed, insert);
1137                         bch2_cut_front(s->committed, &s->whiteout);
1138                         bch2_btree_iter_set_pos_same_leaf(iter, s->committed);
1139                         goto next;
1140                 }
1141
1142                 /*
1143                  * When deleting, if possible just do it by switching the type
1144                  * of the key we're deleting, instead of creating and inserting
1145                  * a new whiteout:
1146                  */
1147                 if (s->deleting &&
1148                     !s->update_btree &&
1149                     !bkey_cmp(insert->k.p, k.k->p) &&
1150                     !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
1151                         if (!bkey_whiteout(k.k)) {
1152                                 btree_account_key_drop(l->b, _k);
1153                                 _k->type = KEY_TYPE_discard;
1154                                 reserve_whiteout(l->b, _k);
1155                         }
1156                         break;
1157                 }
1158
1159                 if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
1160                         insert->k.needs_whiteout = true;
1161                         s->update_btree = true;
1162                 }
1163
1164                 if (s->update_btree &&
1165                     overlap == BCH_EXTENT_OVERLAP_ALL &&
1166                     bkey_whiteout(k.k) &&
1167                     k.k->needs_whiteout) {
1168                         unreserve_whiteout(l->b, _k);
1169                         _k->needs_whiteout = false;
1170                 }
1171
1172                 extent_squash(s, insert, _k, k, overlap);
1173
1174                 if (!s->update_btree)
1175                         bch2_cut_front(s->committed, insert);
1176 next:
1177                 if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
1178                     overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1179                         break;
1180         }
1181
1182         if (bkey_cmp(s->committed, insert->k.p) < 0)
1183                 s->committed = bpos_min(s->insert->k->k.p, l->b->key.k.p);
1184
1185         /*
1186          * may have skipped past some deleted extents greater than the insert
1187          * key, before we got to a non deleted extent and knew we could bail out
1188          * rewind the iterator a bit if necessary:
1189          */
1190         {
1191                 struct btree_node_iter node_iter = l->iter;
1192
1193                 while ((_k = bch2_btree_node_iter_prev_all(&node_iter, l->b)) &&
1194                        bkey_cmp_left_packed(l->b, _k, &s->committed) > 0)
1195                         l->iter = node_iter;
1196         }
1197 }
1198
1199 /**
1200  * bch_extent_insert_fixup - insert a new extent and deal with overlaps
1201  *
1202  * this may result in not actually doing the insert, or inserting some subset
1203  * of the insert key. For cmpxchg operations this is where that logic lives.
1204  *
1205  * All subsets of @insert that need to be inserted are inserted using
1206  * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
1207  * returns false, setting @iter->pos for the prefix of @insert that actually got
1208  * inserted.
1209  *
1210  * BSET INVARIANTS: this function is responsible for maintaining all the
1211  * invariants for bsets of extents in memory. things get really hairy with 0
1212  * size extents
1213  *
1214  * within one bset:
1215  *
1216  * bkey_start_pos(bkey_next(k)) >= k
1217  * or bkey_start_offset(bkey_next(k)) >= k->offset
1218  *
1219  * i.e. strict ordering, no overlapping extents.
1220  *
1221  * multiple bsets (i.e. full btree node):
1222  *
1223  * ∀ k, j
1224  *   k.size != 0 ∧ j.size != 0 →
1225  *     ¬ (k > bkey_start_pos(j) ∧ k < j)
1226  *
1227  * i.e. no two overlapping keys _of nonzero size_
1228  *
1229  * We can't realistically maintain this invariant for zero size keys because of
1230  * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
1231  * there may be another 0 size key between them in another bset, and it will
1232  * thus overlap with the merged key.
1233  *
1234  * In addition, the end of iter->pos indicates how much has been processed.
1235  * If the end of iter->pos is not the same as the end of insert, then
1236  * key insertion needs to continue/be retried.
1237  */
1238 enum btree_insert_ret
1239 bch2_insert_fixup_extent(struct btree_insert *trans,
1240                          struct btree_insert_entry *insert)
1241 {
1242         struct btree_iter *iter = insert->iter;
1243         struct btree *b         = iter->l[0].b;
1244         struct extent_insert_state s = {
1245                 .trans          = trans,
1246                 .insert         = insert,
1247                 .committed      = iter->pos,
1248
1249                 .whiteout       = *insert->k,
1250                 .update_journal = !bkey_whiteout(&insert->k->k),
1251                 .update_btree   = !bkey_whiteout(&insert->k->k),
1252                 .deleting       = bkey_whiteout(&insert->k->k),
1253         };
1254
1255         EBUG_ON(iter->level);
1256         EBUG_ON(!insert->k->k.size);
1257
1258         /*
1259          * As we process overlapping extents, we advance @iter->pos both to
1260          * signal to our caller (btree_insert_key()) how much of @insert->k has
1261          * been inserted, and also to keep @iter->pos consistent with
1262          * @insert->k and the node iterator that we're advancing:
1263          */
1264         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1265
1266         __bch2_insert_fixup_extent(&s);
1267
1268         extent_insert_committed(&s);
1269
1270         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1271         EBUG_ON(bkey_cmp(iter->pos, s.committed));
1272
1273         if (insert->k->k.size) {
1274                 /* got to the end of this leaf node */
1275                 BUG_ON(bkey_cmp(iter->pos, b->key.k.p));
1276                 return BTREE_INSERT_NEED_TRAVERSE;
1277         }
1278
1279         return BTREE_INSERT_OK;
1280 }
1281
1282 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
1283 {
1284         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1285         const union bch_extent_entry *entry;
1286         struct bch_extent_crc_unpacked crc;
1287         const struct bch_extent_ptr *ptr;
1288         unsigned size_ondisk = e.k->size;
1289         const char *reason;
1290         unsigned nonce = UINT_MAX;
1291
1292         if (bkey_val_u64s(e.k) > BKEY_EXTENT_VAL_U64s_MAX)
1293                 return "value too big";
1294
1295         extent_for_each_entry(e, entry) {
1296                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
1297                         return "invalid extent entry type";
1298
1299                 switch (extent_entry_type(entry)) {
1300                 case BCH_EXTENT_ENTRY_ptr:
1301                         ptr = entry_to_ptr(entry);
1302
1303                         reason = extent_ptr_invalid(c, e.s_c, &entry->ptr,
1304                                                     size_ondisk, false);
1305                         if (reason)
1306                                 return reason;
1307                         break;
1308                 case BCH_EXTENT_ENTRY_crc32:
1309                 case BCH_EXTENT_ENTRY_crc64:
1310                 case BCH_EXTENT_ENTRY_crc128:
1311                         crc = bch2_extent_crc_unpack(e.k, entry_to_crc(entry));
1312
1313                         if (crc.offset + e.k->size >
1314                             crc.uncompressed_size)
1315                                 return "checksum offset + key size > uncompressed size";
1316
1317                         size_ondisk = crc.compressed_size;
1318
1319                         if (!bch2_checksum_type_valid(c, crc.csum_type))
1320                                 return "invalid checksum type";
1321
1322                         if (crc.compression_type >= BCH_COMPRESSION_NR)
1323                                 return "invalid compression type";
1324
1325                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1326                                 if (nonce == UINT_MAX)
1327                                         nonce = crc.offset + crc.nonce;
1328                                 else if (nonce != crc.offset + crc.nonce)
1329                                         return "incorrect nonce";
1330                         }
1331                         break;
1332                 case BCH_EXTENT_ENTRY_stripe_ptr:
1333                         break;
1334                 }
1335         }
1336
1337         return NULL;
1338 }
1339
1340 void bch2_extent_debugcheck(struct bch_fs *c, struct btree *b,
1341                             struct bkey_s_c k)
1342 {
1343         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1344         const struct bch_extent_ptr *ptr;
1345         struct bch_dev *ca;
1346         struct bucket_mark mark;
1347         unsigned seq, stale;
1348         char buf[160];
1349         bool bad;
1350         unsigned replicas = 0;
1351
1352         /*
1353          * XXX: we should be doing most/all of these checks at startup time,
1354          * where we check bch2_bkey_invalid() in btree_node_read_done()
1355          *
1356          * But note that we can't check for stale pointers or incorrect gc marks
1357          * until after journal replay is done (it might be an extent that's
1358          * going to get overwritten during replay)
1359          */
1360
1361         extent_for_each_ptr(e, ptr) {
1362                 ca = bch_dev_bkey_exists(c, ptr->dev);
1363                 replicas++;
1364
1365                 /*
1366                  * If journal replay hasn't finished, we might be seeing keys
1367                  * that will be overwritten by the time journal replay is done:
1368                  */
1369                 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1370                         continue;
1371
1372                 stale = 0;
1373
1374                 do {
1375                         seq = read_seqcount_begin(&c->gc_pos_lock);
1376                         mark = ptr_bucket_mark(ca, ptr);
1377
1378                         /* between mark and bucket gen */
1379                         smp_rmb();
1380
1381                         stale = ptr_stale(ca, ptr);
1382
1383                         bch2_fs_bug_on(stale && !ptr->cached, c,
1384                                          "stale dirty pointer");
1385
1386                         bch2_fs_bug_on(stale > 96, c,
1387                                          "key too stale: %i",
1388                                          stale);
1389
1390                         if (stale)
1391                                 break;
1392
1393                         bad = gc_pos_cmp(c->gc_pos, gc_pos_btree_node(b)) > 0 &&
1394                                 (mark.data_type != BCH_DATA_USER ||
1395                                  !(ptr->cached
1396                                    ? mark.cached_sectors
1397                                    : mark.dirty_sectors));
1398                 } while (read_seqcount_retry(&c->gc_pos_lock, seq));
1399
1400                 if (bad)
1401                         goto bad_ptr;
1402         }
1403
1404         if (replicas > BCH_REPLICAS_MAX) {
1405                 bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c);
1406                 bch2_fs_bug(c,
1407                         "extent key bad (too many replicas: %u): %s",
1408                         replicas, buf);
1409                 return;
1410         }
1411
1412         if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1413             !bch2_bkey_replicas_marked(c, e.s_c, false)) {
1414                 bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c);
1415                 bch2_fs_bug(c,
1416                         "extent key bad (replicas not marked in superblock):\n%s",
1417                         buf);
1418                 return;
1419         }
1420
1421         return;
1422
1423 bad_ptr:
1424         bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c);
1425         bch2_fs_bug(c, "extent pointer bad gc mark: %s:\nbucket %zu "
1426                    "gen %i type %u", buf,
1427                    PTR_BUCKET_NR(ca, ptr), mark.gen, mark.data_type);
1428 }
1429
1430 void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
1431                          struct bkey_s_c k)
1432 {
1433         const char *invalid;
1434
1435         bkey_ptrs_to_text(out, c, k);
1436
1437         invalid = bch2_extent_invalid(c, k);
1438         if (invalid)
1439                 pr_buf(out, " invalid: %s", invalid);
1440 }
1441
1442 static void bch2_extent_crc_init(union bch_extent_crc *crc,
1443                                  struct bch_extent_crc_unpacked new)
1444 {
1445 #define common_fields(_crc)                                             \
1446                 .csum_type              = _crc.csum_type,               \
1447                 .compression_type       = _crc.compression_type,        \
1448                 ._compressed_size       = _crc.compressed_size - 1,     \
1449                 ._uncompressed_size     = _crc.uncompressed_size - 1,   \
1450                 .offset                 = _crc.offset
1451
1452         if (bch_crc_bytes[new.csum_type]        <= 4 &&
1453             new.uncompressed_size               <= CRC32_SIZE_MAX &&
1454             new.nonce                           <= CRC32_NONCE_MAX) {
1455                 crc->crc32 = (struct bch_extent_crc32) {
1456                         .type = 1 << BCH_EXTENT_ENTRY_crc32,
1457                         common_fields(new),
1458                         .csum                   = *((__le32 *) &new.csum.lo),
1459                 };
1460                 return;
1461         }
1462
1463         if (bch_crc_bytes[new.csum_type]        <= 10 &&
1464             new.uncompressed_size               <= CRC64_SIZE_MAX &&
1465             new.nonce                           <= CRC64_NONCE_MAX) {
1466                 crc->crc64 = (struct bch_extent_crc64) {
1467                         .type = 1 << BCH_EXTENT_ENTRY_crc64,
1468                         common_fields(new),
1469                         .nonce                  = new.nonce,
1470                         .csum_lo                = new.csum.lo,
1471                         .csum_hi                = *((__le16 *) &new.csum.hi),
1472                 };
1473                 return;
1474         }
1475
1476         if (bch_crc_bytes[new.csum_type]        <= 16 &&
1477             new.uncompressed_size               <= CRC128_SIZE_MAX &&
1478             new.nonce                           <= CRC128_NONCE_MAX) {
1479                 crc->crc128 = (struct bch_extent_crc128) {
1480                         .type = 1 << BCH_EXTENT_ENTRY_crc128,
1481                         common_fields(new),
1482                         .nonce                  = new.nonce,
1483                         .csum                   = new.csum,
1484                 };
1485                 return;
1486         }
1487 #undef common_fields
1488         BUG();
1489 }
1490
1491 void bch2_extent_crc_append(struct bkey_i_extent *e,
1492                             struct bch_extent_crc_unpacked new)
1493 {
1494         bch2_extent_crc_init((void *) extent_entry_last(extent_i_to_s(e)), new);
1495         __extent_entry_push(e);
1496 }
1497
1498 void bch2_extent_ptr_decoded_append(struct bkey_i_extent *e,
1499                                     struct extent_ptr_decoded *p)
1500 {
1501         struct bch_extent_crc_unpacked crc;
1502         union bch_extent_entry *pos;
1503         unsigned i;
1504
1505         extent_for_each_crc(extent_i_to_s(e), crc, pos)
1506                 if (!bch2_crc_unpacked_cmp(crc, p->crc))
1507                         goto found;
1508
1509         bch2_extent_crc_append(e, p->crc);
1510         pos = extent_entry_last(extent_i_to_s(e));
1511 found:
1512         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
1513         __extent_entry_insert(e, pos, to_entry(&p->ptr));
1514
1515         for (i = 0; i < p->ec_nr; i++) {
1516                 p->ec[i].type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
1517                 __extent_entry_insert(e, pos, to_entry(&p->ec[i]));
1518         }
1519 }
1520
1521 /*
1522  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
1523  *
1524  * Returns true if @k should be dropped entirely
1525  *
1526  * For existing keys, only called when btree nodes are being rewritten, not when
1527  * they're merely being compacted/resorted in memory.
1528  */
1529 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1530 {
1531         struct bch_extent_ptr *ptr;
1532
1533         bch2_bkey_drop_ptrs(k, ptr,
1534                 ptr->cached &&
1535                 ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
1536
1537         /* will only happen if all pointers were cached: */
1538         if (!bkey_val_u64s(k.k))
1539                 k.k->type = KEY_TYPE_deleted;
1540
1541         return false;
1542 }
1543
1544 void bch2_extent_mark_replicas_cached(struct bch_fs *c,
1545                                       struct bkey_s_extent e,
1546                                       unsigned target,
1547                                       unsigned nr_desired_replicas)
1548 {
1549         union bch_extent_entry *entry;
1550         struct extent_ptr_decoded p;
1551         int extra = bch2_bkey_durability(c, e.s_c) - nr_desired_replicas;
1552
1553         if (target && extra > 0)
1554                 extent_for_each_ptr_decode(e, p, entry) {
1555                         int n = bch2_extent_ptr_durability(c, p);
1556
1557                         if (n && n <= extra &&
1558                             !bch2_dev_in_target(c, p.ptr.dev, target)) {
1559                                 entry->ptr.cached = true;
1560                                 extra -= n;
1561                         }
1562                 }
1563
1564         if (extra > 0)
1565                 extent_for_each_ptr_decode(e, p, entry) {
1566                         int n = bch2_extent_ptr_durability(c, p);
1567
1568                         if (n && n <= extra) {
1569                                 entry->ptr.cached = true;
1570                                 extra -= n;
1571                         }
1572                 }
1573 }
1574
1575 enum merge_result bch2_extent_merge(struct bch_fs *c,
1576                                     struct bkey_i *l, struct bkey_i *r)
1577 {
1578         struct bkey_s_extent el = bkey_i_to_s_extent(l);
1579         struct bkey_s_extent er = bkey_i_to_s_extent(r);
1580         union bch_extent_entry *en_l, *en_r;
1581
1582         if (bkey_val_u64s(&l->k) != bkey_val_u64s(&r->k))
1583                 return BCH_MERGE_NOMERGE;
1584
1585         extent_for_each_entry(el, en_l) {
1586                 struct bch_extent_ptr *lp, *rp;
1587                 struct bch_dev *ca;
1588
1589                 en_r = vstruct_idx(er.v, (u64 *) en_l - el.v->_data);
1590
1591                 if ((extent_entry_type(en_l) !=
1592                      extent_entry_type(en_r)) ||
1593                     !extent_entry_is_ptr(en_l))
1594                         return BCH_MERGE_NOMERGE;
1595
1596                 lp = &en_l->ptr;
1597                 rp = &en_r->ptr;
1598
1599                 if (lp->offset + el.k->size     != rp->offset ||
1600                     lp->dev                     != rp->dev ||
1601                     lp->gen                     != rp->gen)
1602                         return BCH_MERGE_NOMERGE;
1603
1604                 /* We don't allow extents to straddle buckets: */
1605                 ca = bch_dev_bkey_exists(c, lp->dev);
1606
1607                 if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
1608                         return BCH_MERGE_NOMERGE;
1609         }
1610
1611         l->k.needs_whiteout |= r->k.needs_whiteout;
1612
1613         /* Keys with no pointers aren't restricted to one bucket and could
1614          * overflow KEY_SIZE
1615          */
1616         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1617                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1618                 bch2_cut_front(l->k.p, r);
1619                 return BCH_MERGE_PARTIAL;
1620         }
1621
1622         bch2_key_resize(&l->k, l->k.size + r->k.size);
1623
1624         return BCH_MERGE_MERGE;
1625 }
1626
1627 /*
1628  * When merging an extent that we're inserting into a btree node, the new merged
1629  * extent could overlap with an existing 0 size extent - if we don't fix that,
1630  * it'll break the btree node iterator so this code finds those 0 size extents
1631  * and shifts them out of the way.
1632  *
1633  * Also unpacks and repacks.
1634  */
1635 static bool bch2_extent_merge_inline(struct bch_fs *c,
1636                                      struct btree_iter *iter,
1637                                      struct bkey_packed *l,
1638                                      struct bkey_packed *r,
1639                                      bool back_merge)
1640 {
1641         struct btree *b = iter->l[0].b;
1642         struct btree_node_iter *node_iter = &iter->l[0].iter;
1643         BKEY_PADDED(k) li, ri;
1644         struct bkey_packed *m   = back_merge ? l : r;
1645         struct bkey_i *mi       = back_merge ? &li.k : &ri.k;
1646         struct bset_tree *t     = bch2_bkey_to_bset(b, m);
1647         enum merge_result ret;
1648
1649         EBUG_ON(bkey_written(b, m));
1650
1651         /*
1652          * We need to save copies of both l and r, because we might get a
1653          * partial merge (which modifies both) and then fails to repack
1654          */
1655         bch2_bkey_unpack(b, &li.k, l);
1656         bch2_bkey_unpack(b, &ri.k, r);
1657
1658         ret = bch2_bkey_merge(c, &li.k, &ri.k);
1659         if (ret == BCH_MERGE_NOMERGE)
1660                 return false;
1661
1662         /*
1663          * check if we overlap with deleted extents - would break the sort
1664          * order:
1665          */
1666         if (back_merge) {
1667                 struct bkey_packed *n = bkey_next(m);
1668
1669                 if (n != btree_bkey_last(b, t) &&
1670                     bkey_cmp_left_packed(b, n, &li.k.k.p) <= 0 &&
1671                     bkey_deleted(n))
1672                         return false;
1673         } else if (ret == BCH_MERGE_MERGE) {
1674                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1675
1676                 if (prev &&
1677                     bkey_cmp_left_packed_byval(b, prev,
1678                                 bkey_start_pos(&li.k.k)) > 0)
1679                         return false;
1680         }
1681
1682         if (ret == BCH_MERGE_PARTIAL) {
1683                 if (!extent_i_save(b, m, mi))
1684                         return false;
1685
1686                 if (!back_merge)
1687                         bkey_copy(packed_to_bkey(l), &li.k);
1688                 else
1689                         bkey_copy(packed_to_bkey(r), &ri.k);
1690         } else {
1691                 if (!extent_i_save(b, m, &li.k))
1692                         return false;
1693         }
1694
1695         bch2_bset_fix_invalidated_key(b, m);
1696         bch2_btree_node_iter_fix(iter, b, node_iter,
1697                                  m, m->u64s, m->u64s);
1698         verify_modified_extent(iter, m);
1699
1700         return ret == BCH_MERGE_MERGE;
1701 }
1702
1703 int bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size)
1704 {
1705         struct btree_iter iter;
1706         struct bpos end = pos;
1707         struct bkey_s_c k;
1708         int ret = 0;
1709
1710         end.offset += size;
1711
1712         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, pos,
1713                              BTREE_ITER_SLOTS, k) {
1714                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
1715                         break;
1716
1717                 if (!bch2_extent_is_fully_allocated(k)) {
1718                         ret = -ENOSPC;
1719                         break;
1720                 }
1721         }
1722         bch2_btree_iter_unlock(&iter);
1723
1724         return ret;
1725 }
1726
1727 /* KEY_TYPE_reservation: */
1728
1729 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
1730 {
1731         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1732
1733         if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
1734                 return "incorrect value size";
1735
1736         if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
1737                 return "invalid nr_replicas";
1738
1739         return NULL;
1740 }
1741
1742 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
1743                               struct bkey_s_c k)
1744 {
1745         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1746
1747         pr_buf(out, "generation %u replicas %u",
1748                le32_to_cpu(r.v->generation),
1749                r.v->nr_replicas);
1750 }
1751
1752 enum merge_result bch2_reservation_merge(struct bch_fs *c,
1753                                          struct bkey_i *l, struct bkey_i *r)
1754 {
1755         struct bkey_i_reservation *li = bkey_i_to_reservation(l);
1756         struct bkey_i_reservation *ri = bkey_i_to_reservation(r);
1757
1758         if (li->v.generation != ri->v.generation ||
1759             li->v.nr_replicas != ri->v.nr_replicas)
1760                 return BCH_MERGE_NOMERGE;
1761
1762         l->k.needs_whiteout |= r->k.needs_whiteout;
1763
1764         /* Keys with no pointers aren't restricted to one bucket and could
1765          * overflow KEY_SIZE
1766          */
1767         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1768                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1769                 bch2_cut_front(l->k.p, r);
1770                 return BCH_MERGE_PARTIAL;
1771         }
1772
1773         bch2_key_resize(&l->k, l->k.size + r->k.size);
1774
1775         return BCH_MERGE_MERGE;
1776 }