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                                 ret += p.crc.compressed_size;
 310         }
 311         }
 312
 313         return ret;
 314 }
 315
 316 bool bch2_extent_matches_ptr(struct bch_fs *c, struct bkey_s_c_extent e,
 317                              struct bch_extent_ptr m, u64 offset)
 318 {
 319         const union bch_extent_entry *entry;
 320         struct extent_ptr_decoded p;
 321
 322         extent_for_each_ptr_decode(e, p, entry)
 323                 if (p.ptr.dev   == m.dev &&
 324                     p.ptr.gen   == m.gen &&
 325                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(e.k) ==
 326                     (s64) m.offset  - offset)
 327                         return true;
 328
 329         return false;
 330 }
 331
 332 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 333                                           union bch_extent_entry *entry)
 334 {
 335         union bch_extent_entry *i = ptrs.start;
 336
 337         if (i == entry)
 338                 return NULL;
 339
 340         while (extent_entry_next(i) != entry)
 341                 i = extent_entry_next(i);
 342         return i;
 343 }
 344
 345 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
 346                                            struct bch_extent_ptr *ptr)
 347 {
 348         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 349         union bch_extent_entry *dst, *src, *prev;
 350         bool drop_crc = true;
 351
 352         EBUG_ON(ptr < &ptrs.start->ptr ||
 353                 ptr >= &ptrs.end->ptr);
 354         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 355
 356         src = extent_entry_next(to_entry(ptr));
 357         if (src != ptrs.end &&
 358             !extent_entry_is_crc(src))
 359                 drop_crc = false;
 360
 361         dst = to_entry(ptr);
 362         while ((prev = extent_entry_prev(ptrs, dst))) {
 363                 if (extent_entry_is_ptr(prev))
 364                         break;
 365
 366                 if (extent_entry_is_crc(prev)) {
 367                         if (drop_crc)
 368                                 dst = prev;
 369                         break;
 370                 }
 371
 372                 dst = prev;
 373         }
 374
 375         memmove_u64s_down(dst, src,
 376                           (u64 *) ptrs.end - (u64 *) src);
 377         k.k->u64s -= (u64 *) src - (u64 *) dst;
 378
 379         return dst;
 380 }
 381
 382 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 383                                   struct bch_extent_crc_unpacked n)
 384 {
 385         return !u.compression_type &&
 386                 u.csum_type &&
 387                 u.uncompressed_size > u.live_size &&
 388                 bch2_csum_type_is_encryption(u.csum_type) ==
 389                 bch2_csum_type_is_encryption(n.csum_type);
 390 }
 391
 392 bool bch2_can_narrow_extent_crcs(struct bkey_s_c_extent e,
 393                                  struct bch_extent_crc_unpacked n)
 394 {
 395         struct bch_extent_crc_unpacked crc;
 396         const union bch_extent_entry *i;
 397
 398         if (!n.csum_type)
 399                 return false;
 400
 401         extent_for_each_crc(e, crc, i)
 402                 if (can_narrow_crc(crc, n))
 403                         return true;
 404
 405         return false;
 406 }
 407
 408 /*
 409  * We're writing another replica for this extent, so while we've got the data in
 410  * memory we'll be computing a new checksum for the currently live data.
 411  *
 412  * If there are other replicas we aren't moving, and they are checksummed but
 413  * not compressed, we can modify them to point to only the data that is
 414  * currently live (so that readers won't have to bounce) while we've got the
 415  * checksum we need:
 416  */
 417 bool bch2_extent_narrow_crcs(struct bkey_i_extent *e,
 418                              struct bch_extent_crc_unpacked n)
 419 {
 420         struct bch_extent_crc_unpacked u;
 421         struct extent_ptr_decoded p;
 422         union bch_extent_entry *i;
 423         bool ret = false;
 424
 425         /* Find a checksum entry that covers only live data: */
 426         if (!n.csum_type) {
 427                 extent_for_each_crc(extent_i_to_s(e), u, i)
 428                         if (!u.compression_type &&
 429                             u.csum_type &&
 430                             u.live_size == u.uncompressed_size) {
 431                                 n = u;
 432                                 goto found;
 433                         }
 434                 return false;
 435         }
 436 found:
 437         BUG_ON(n.compression_type);
 438         BUG_ON(n.offset);
 439         BUG_ON(n.live_size != e->k.size);
 440
 441 restart_narrow_pointers:
 442         extent_for_each_ptr_decode(extent_i_to_s(e), p, i)
 443                 if (can_narrow_crc(p.crc, n)) {
 444                         bch2_bkey_drop_ptr(extent_i_to_s(e).s, &i->ptr);
 445                         p.ptr.offset += p.crc.offset;
 446                         p.crc = n;
 447                         bch2_extent_ptr_decoded_append(e, &p);
 448                         ret = true;
 449                         goto restart_narrow_pointers;
 450                 }
 451
 452         return ret;
 453 }
 454
 455 /* returns true if not equal */
 456 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 457                                          struct bch_extent_crc_unpacked r)
 458 {
 459         return (l.csum_type             != r.csum_type ||
 460                 l.compression_type      != r.compression_type ||
 461                 l.compressed_size       != r.compressed_size ||
 462                 l.uncompressed_size     != r.uncompressed_size ||
 463                 l.offset                != r.offset ||
 464                 l.live_size             != r.live_size ||
 465                 l.nonce                 != r.nonce ||
 466                 bch2_crc_cmp(l.csum, r.csum));
 467 }
 468
 469 void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
 470 {
 471         union bch_extent_entry *entry;
 472         u64 *d = (u64 *) bkeyp_val(f, k);
 473         unsigned i;
 474
 475         for (i = 0; i < bkeyp_val_u64s(f, k); i++)
 476                 d[i] = swab64(d[i]);
 477
 478         for (entry = (union bch_extent_entry *) d;
 479              entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
 480              entry = extent_entry_next(entry)) {
 481                 switch (extent_entry_type(entry)) {
 482                 case BCH_EXTENT_ENTRY_ptr:
 483                         break;
 484                 case BCH_EXTENT_ENTRY_crc32:
 485                         entry->crc32.csum = swab32(entry->crc32.csum);
 486                         break;
 487                 case BCH_EXTENT_ENTRY_crc64:
 488                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
 489                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
 490                         break;
 491                 case BCH_EXTENT_ENTRY_crc128:
 492                         entry->crc128.csum.hi = (__force __le64)
 493                                 swab64((__force u64) entry->crc128.csum.hi);
 494                         entry->crc128.csum.lo = (__force __le64)
 495                                 swab64((__force u64) entry->crc128.csum.lo);
 496                         break;
 497                 case BCH_EXTENT_ENTRY_stripe_ptr:
 498                         break;
 499                 }
 500         }
 501 }
 502
 503 static const char *extent_ptr_invalid(const struct bch_fs *c,
 504                                       struct bkey_s_c k,
 505                                       const struct bch_extent_ptr *ptr,
 506                                       unsigned size_ondisk,
 507                                       bool metadata)
 508 {
 509         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 510         const struct bch_extent_ptr *ptr2;
 511         struct bch_dev *ca;
 512
 513         if (ptr->dev >= c->sb.nr_devices ||
 514             !c->devs[ptr->dev])
 515                 return "pointer to invalid device";
 516
 517         ca = bch_dev_bkey_exists(c, ptr->dev);
 518         if (!ca)
 519                 return "pointer to invalid device";
 520
 521         bkey_for_each_ptr(ptrs, ptr2)
 522                 if (ptr != ptr2 && ptr->dev == ptr2->dev)
 523                         return "multiple pointers to same device";
 524
 525         if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
 526                 return "offset past end of device";
 527
 528         if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
 529                 return "offset before first bucket";
 530
 531         if (bucket_remainder(ca, ptr->offset) +
 532             size_ondisk > ca->mi.bucket_size)
 533                 return "spans multiple buckets";
 534
 535         return NULL;
 536 }
 537
 538 static void bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 539                               struct bkey_s_c k)
 540 {
 541         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 542         const union bch_extent_entry *entry;
 543         struct bch_extent_crc_unpacked crc;
 544         const struct bch_extent_ptr *ptr;
 545         const struct bch_extent_stripe_ptr *ec;
 546         struct bch_dev *ca;
 547         bool first = true;
 548
 549         bkey_extent_entry_for_each(ptrs, entry) {
 550                 if (!first)
 551                         pr_buf(out, " ");
 552
 553                 switch (__extent_entry_type(entry)) {
 554                 case BCH_EXTENT_ENTRY_ptr:
 555                         ptr = entry_to_ptr(entry);
 556                         ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
 557                                 ? bch_dev_bkey_exists(c, ptr->dev)
 558                                 : NULL;
 559
 560                         pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
 561                                (u64) ptr->offset, ptr->gen,
 562                                ptr->cached ? " cached" : "",
 563                                ca && ptr_stale(ca, ptr)
 564                                ? " stale" : "");
 565                         break;
 566                 case BCH_EXTENT_ENTRY_crc32:
 567                 case BCH_EXTENT_ENTRY_crc64:
 568                 case BCH_EXTENT_ENTRY_crc128:
 569                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 570
 571                         pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
 572                                crc.compressed_size,
 573                                crc.uncompressed_size,
 574                                crc.offset, crc.nonce,
 575                                crc.csum_type,
 576                                crc.compression_type);
 577                         break;
 578                 case BCH_EXTENT_ENTRY_stripe_ptr:
 579                         ec = &entry->stripe_ptr;
 580
 581                         pr_buf(out, "ec: idx %llu block %u",
 582                                (u64) ec->idx, ec->block);
 583                         break;
 584                 default:
 585                         pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
 586                         return;
 587                 }
 588
 589                 first = false;
 590         }
 591 }
 592
 593 /* Btree ptrs */
 594
 595 const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
 596 {
 597         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 598         const union bch_extent_entry *entry;
 599         const struct bch_extent_ptr *ptr;
 600         const char *reason;
 601
 602         if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
 603                 return "value too big";
 604
 605         bkey_extent_entry_for_each(ptrs, entry) {
 606                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
 607                         return "invalid extent entry type";
 608
 609                 if (!extent_entry_is_ptr(entry))
 610                         return "has non ptr field";
 611         }
 612
 613         bkey_for_each_ptr(ptrs, ptr) {
 614                 reason = extent_ptr_invalid(c, k, ptr,
 615                                             c->opts.btree_node_size,
 616                                             true);
 617                 if (reason)
 618                         return reason;
 619         }
 620
 621         return NULL;
 622 }
 623
 624 void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct btree *b,
 625                                struct bkey_s_c k)
 626 {
 627         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 628         const struct bch_extent_ptr *ptr;
 629         const char *err;
 630         char buf[160];
 631         struct bucket_mark mark;
 632         struct bch_dev *ca;
 633
 634         bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 635                        !bch2_bkey_replicas_marked(c, k, false), c,
 636                        "btree key bad (replicas not marked in superblock):\n%s",
 637                        (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 638
 639         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 640                 return;
 641
 642         bkey_for_each_ptr(ptrs, ptr) {
 643                 ca = bch_dev_bkey_exists(c, ptr->dev);
 644
 645                 mark = ptr_bucket_mark(ca, ptr);
 646
 647                 err = "stale";
 648                 if (gen_after(mark.gen, ptr->gen))
 649                         goto err;
 650
 651                 err = "inconsistent";
 652                 if (mark.data_type != BCH_DATA_BTREE ||
 653                     mark.dirty_sectors < c->opts.btree_node_size)
 654                         goto err;
 655         }
 656
 657         return;
 658 err:
 659         bch2_bkey_val_to_text(&PBUF(buf), c, k);
 660         bch2_fs_bug(c, "%s btree pointer %s: bucket %zi gen %i mark %08x",
 661                     err, buf, PTR_BUCKET_NR(ca, ptr),
 662                     mark.gen, (unsigned) mark.v.counter);
 663 }
 664
 665 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 666                             struct bkey_s_c k)
 667 {
 668         const char *invalid;
 669
 670         bkey_ptrs_to_text(out, c, k);
 671
 672         invalid = bch2_btree_ptr_invalid(c, k);
 673         if (invalid)
 674                 pr_buf(out, " invalid: %s", invalid);
 675 }
 676
 677 /* Extents */
 678
 679 bool __bch2_cut_front(struct bpos where, struct bkey_s k)
 680 {
 681         u64 len = 0;
 682
 683         if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
 684                 return false;
 685
 686         EBUG_ON(bkey_cmp(where, k.k->p) > 0);
 687
 688         len = k.k->p.offset - where.offset;
 689
 690         BUG_ON(len > k.k->size);
 691
 692         /*
 693          * Don't readjust offset if the key size is now 0, because that could
 694          * cause offset to point to the next bucket:
 695          */
 696         if (!len)
 697                 k.k->type = KEY_TYPE_deleted;
 698         else if (bkey_extent_is_data(k.k)) {
 699                 struct bkey_s_extent e = bkey_s_to_extent(k);
 700                 union bch_extent_entry *entry;
 701                 bool seen_crc = false;
 702
 703                 extent_for_each_entry(e, entry) {
 704                         switch (extent_entry_type(entry)) {
 705                         case BCH_EXTENT_ENTRY_ptr:
 706                                 if (!seen_crc)
 707                                         entry->ptr.offset += e.k->size - len;
 708                                 break;
 709                         case BCH_EXTENT_ENTRY_crc32:
 710                                 entry->crc32.offset += e.k->size - len;
 711                                 break;
 712                         case BCH_EXTENT_ENTRY_crc64:
 713                                 entry->crc64.offset += e.k->size - len;
 714                                 break;
 715                         case BCH_EXTENT_ENTRY_crc128:
 716                                 entry->crc128.offset += e.k->size - len;
 717                                 break;
 718                         case BCH_EXTENT_ENTRY_stripe_ptr:
 719                                 break;
 720                         }
 721
 722                         if (extent_entry_is_crc(entry))
 723                                 seen_crc = true;
 724                 }
 725         }
 726
 727         k.k->size = len;
 728
 729         return true;
 730 }
 731
 732 bool bch2_cut_back(struct bpos where, struct bkey *k)
 733 {
 734         u64 len = 0;
 735
 736         if (bkey_cmp(where, k->p) >= 0)
 737                 return false;
 738
 739         EBUG_ON(bkey_cmp(where, bkey_start_pos(k)) < 0);
 740
 741         len = where.offset - bkey_start_offset(k);
 742
 743         BUG_ON(len > k->size);
 744
 745         k->p = where;
 746         k->size = len;
 747
 748         if (!len)
 749                 k->type = KEY_TYPE_deleted;
 750
 751         return true;
 752 }
 753
 754 /**
 755  * bch_key_resize - adjust size of @k
 756  *
 757  * bkey_start_offset(k) will be preserved, modifies where the extent ends
 758  */
 759 void bch2_key_resize(struct bkey *k,
 760                     unsigned new_size)
 761 {
 762         k->p.offset -= k->size;
 763         k->p.offset += new_size;
 764         k->size = new_size;
 765 }
 766
 767 static bool extent_i_save(struct btree *b, struct bkey_packed *dst,
 768                           struct bkey_i *src)
 769 {
 770         struct bkey_format *f = &b->format;
 771         struct bkey_i *dst_unpacked;
 772         struct bkey_packed tmp;
 773
 774         if ((dst_unpacked = packed_to_bkey(dst)))
 775                 dst_unpacked->k = src->k;
 776         else if (bch2_bkey_pack_key(&tmp, &src->k, f))
 777                 memcpy_u64s(dst, &tmp, f->key_u64s);
 778         else
 779                 return false;
 780
 781         memcpy_u64s(bkeyp_val(f, dst), &src->v, bkey_val_u64s(&src->k));
 782         return true;
 783 }
 784
 785 struct extent_insert_state {
 786         struct btree_insert             *trans;
 787         struct btree_insert_entry       *insert;
 788         struct bpos                     committed;
 789
 790         /* for deleting: */
 791         struct bkey_i                   whiteout;
 792         bool                            update_journal;
 793         bool                            update_btree;
 794         bool                            deleting;
 795 };
 796
 797 static bool bch2_extent_merge_inline(struct bch_fs *,
 798                                      struct btree_iter *,
 799                                      struct bkey_packed *,
 800                                      struct bkey_packed *,
 801                                      bool);
 802
 803 static void verify_extent_nonoverlapping(struct btree *b,
 804                                          struct btree_node_iter *_iter,
 805                                          struct bkey_i *insert)
 806 {
 807 #ifdef CONFIG_BCACHEFS_DEBUG
 808         struct btree_node_iter iter;
 809         struct bkey_packed *k;
 810         struct bkey uk;
 811
 812         iter = *_iter;
 813         k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
 814         BUG_ON(k &&
 815                (uk = bkey_unpack_key(b, k),
 816                 bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
 817
 818         iter = *_iter;
 819         k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
 820 #if 0
 821         BUG_ON(k &&
 822                (uk = bkey_unpack_key(b, k),
 823                 bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
 824 #else
 825         if (k &&
 826             (uk = bkey_unpack_key(b, k),
 827              bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
 828                 char buf1[100];
 829                 char buf2[100];
 830
 831                 bch2_bkey_to_text(&PBUF(buf1), &insert->k);
 832                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 833
 834                 bch2_dump_btree_node(b);
 835                 panic("insert > next :\n"
 836                       "insert %s\n"
 837                       "next   %s\n",
 838                       buf1, buf2);
 839         }
 840 #endif
 841
 842 #endif
 843 }
 844
 845 static void verify_modified_extent(struct btree_iter *iter,
 846                                    struct bkey_packed *k)
 847 {
 848         bch2_btree_iter_verify(iter, iter->l[0].b);
 849         bch2_verify_insert_pos(iter->l[0].b, k, k, k->u64s);
 850 }
 851
 852 static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 853                                struct bkey_i *insert)
 854 {
 855         struct btree_iter_level *l = &iter->l[0];
 856         struct btree_node_iter node_iter;
 857         struct bkey_packed *k;
 858
 859         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
 860
 861         EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
 862         verify_extent_nonoverlapping(l->b, &l->iter, insert);
 863
 864         node_iter = l->iter;
 865         k = bch2_btree_node_iter_prev_filter(&node_iter, l->b, KEY_TYPE_discard);
 866         if (k && !bkey_written(l->b, k) &&
 867             bch2_extent_merge_inline(c, iter, k, bkey_to_packed(insert), true))
 868                 return;
 869
 870         node_iter = l->iter;
 871         k = bch2_btree_node_iter_peek_filter(&node_iter, l->b, KEY_TYPE_discard);
 872         if (k && !bkey_written(l->b, k) &&
 873             bch2_extent_merge_inline(c, iter, bkey_to_packed(insert), k, false))
 874                 return;
 875
 876         k = bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
 877
 878         bch2_bset_insert(l->b, &l->iter, k, insert, 0);
 879         bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
 880         bch2_btree_iter_verify(iter, l->b);
 881 }
 882
 883 static void extent_insert_committed(struct extent_insert_state *s)
 884 {
 885         struct bch_fs *c = s->trans->c;
 886         struct btree_iter *iter = s->insert->iter;
 887         struct bkey_i *insert = s->insert->k;
 888         BKEY_PADDED(k) split;
 889
 890         EBUG_ON(bkey_cmp(insert->k.p, s->committed) < 0);
 891         EBUG_ON(bkey_cmp(s->committed, bkey_start_pos(&insert->k)) < 0);
 892
 893         bkey_copy(&split.k, insert);
 894         if (s->deleting)
 895                 split.k.k.type = KEY_TYPE_discard;
 896
 897         bch2_cut_back(s->committed, &split.k.k);
 898
 899         if (!bkey_cmp(s->committed, iter->pos))
 900                 return;
 901
 902         bch2_btree_iter_set_pos_same_leaf(iter, s->committed);
 903
 904         if (s->update_btree) {
 905                 if (debug_check_bkeys(c))
 906                         bch2_bkey_debugcheck(c, iter->l[0].b,
 907                                              bkey_i_to_s_c(&split.k));
 908
 909                 EBUG_ON(bkey_deleted(&split.k.k) || !split.k.k.size);
 910
 911                 extent_bset_insert(c, iter, &split.k);
 912         }
 913
 914         if (s->update_journal) {
 915                 bkey_copy(&split.k, !s->deleting ? insert : &s->whiteout);
 916                 if (s->deleting)
 917                         split.k.k.type = KEY_TYPE_discard;
 918
 919                 bch2_cut_back(s->committed, &split.k.k);
 920
 921                 EBUG_ON(bkey_deleted(&split.k.k) || !split.k.k.size);
 922
 923                 bch2_btree_journal_key(s->trans, iter, &split.k);
 924         }
 925
 926         bch2_cut_front(s->committed, insert);
 927
 928         insert->k.needs_whiteout        = false;
 929 }
 930
 931 void bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
 932 {
 933         struct btree *b = iter->l[0].b;
 934
 935         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 936
 937         bch2_cut_back(b->key.k.p, &k->k);
 938
 939         BUG_ON(bkey_cmp(bkey_start_pos(&k->k), b->data->min_key) < 0);
 940 }
 941
 942 enum btree_insert_ret
 943 bch2_extent_can_insert(struct btree_insert *trans,
 944                        struct btree_insert_entry *insert,
 945                        unsigned *u64s)
 946 {
 947         struct btree_iter_level *l = &insert->iter->l[0];
 948         struct btree_node_iter node_iter = l->iter;
 949         enum bch_extent_overlap overlap;
 950         struct bkey_packed *_k;
 951         struct bkey unpacked;
 952         struct bkey_s_c k;
 953         int sectors;
 954
 955         BUG_ON(trans->flags & BTREE_INSERT_ATOMIC &&
 956                !bch2_extent_is_atomic(&insert->k->k, insert->iter));
 957
 958         /*
 959          * We avoid creating whiteouts whenever possible when deleting, but
 960          * those optimizations mean we may potentially insert two whiteouts
 961          * instead of one (when we overlap with the front of one extent and the
 962          * back of another):
 963          */
 964         if (bkey_whiteout(&insert->k->k))
 965                 *u64s += BKEY_U64s;
 966
 967         _k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
 968                                               KEY_TYPE_discard);
 969         if (!_k)
 970                 return BTREE_INSERT_OK;
 971
 972         k = bkey_disassemble(l->b, _k, &unpacked);
 973
 974         overlap = bch2_extent_overlap(&insert->k->k, k.k);
 975
 976         /* account for having to split existing extent: */
 977         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
 978                 *u64s += _k->u64s;
 979
 980         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
 981             (sectors = bch2_extent_is_compressed(k))) {
 982                 int flags = BCH_DISK_RESERVATION_BTREE_LOCKS_HELD;
 983
 984                 if (trans->flags & BTREE_INSERT_NOFAIL)
 985                         flags |= BCH_DISK_RESERVATION_NOFAIL;
 986
 987                 switch (bch2_disk_reservation_add(trans->c,
 988                                 trans->disk_res,
 989                                 sectors, flags)) {
 990                 case 0:
 991                         break;
 992                 case -ENOSPC:
 993                         return BTREE_INSERT_ENOSPC;
 994                 case -EINTR:
 995                         return BTREE_INSERT_NEED_GC_LOCK;
 996                 default:
 997                         BUG();
 998                 }
 999         }
1000
1001         return BTREE_INSERT_OK;
1002 }
1003
1004 static void
1005 extent_squash(struct extent_insert_state *s, struct bkey_i *insert,
1006               struct bkey_packed *_k, struct bkey_s k,
1007               enum bch_extent_overlap overlap)
1008 {
1009         struct bch_fs *c = s->trans->c;
1010         struct btree_iter *iter = s->insert->iter;
1011         struct btree_iter_level *l = &iter->l[0];
1012
1013         switch (overlap) {
1014         case BCH_EXTENT_OVERLAP_FRONT:
1015                 /* insert overlaps with start of k: */
1016                 __bch2_cut_front(insert->k.p, k);
1017                 BUG_ON(bkey_deleted(k.k));
1018                 extent_save(l->b, _k, k.k);
1019                 verify_modified_extent(iter, _k);
1020                 break;
1021
1022         case BCH_EXTENT_OVERLAP_BACK:
1023                 /* insert overlaps with end of k: */
1024                 bch2_cut_back(bkey_start_pos(&insert->k), k.k);
1025                 BUG_ON(bkey_deleted(k.k));
1026                 extent_save(l->b, _k, k.k);
1027
1028                 /*
1029                  * As the auxiliary tree is indexed by the end of the
1030                  * key and we've just changed the end, update the
1031                  * auxiliary tree.
1032                  */
1033                 bch2_bset_fix_invalidated_key(l->b, _k);
1034                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1035                                          _k, _k->u64s, _k->u64s);
1036                 verify_modified_extent(iter, _k);
1037                 break;
1038
1039         case BCH_EXTENT_OVERLAP_ALL: {
1040                 /* The insert key completely covers k, invalidate k */
1041                 if (!bkey_whiteout(k.k))
1042                         btree_account_key_drop(l->b, _k);
1043
1044                 k.k->size = 0;
1045                 k.k->type = KEY_TYPE_deleted;
1046
1047                 if (_k >= btree_bset_last(l->b)->start) {
1048                         unsigned u64s = _k->u64s;
1049
1050                         bch2_bset_delete(l->b, _k, _k->u64s);
1051                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1052                                                  _k, u64s, 0);
1053                         bch2_btree_iter_verify(iter, l->b);
1054                 } else {
1055                         extent_save(l->b, _k, k.k);
1056                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
1057                                                  _k, _k->u64s, _k->u64s);
1058                         verify_modified_extent(iter, _k);
1059                 }
1060
1061                 break;
1062         }
1063         case BCH_EXTENT_OVERLAP_MIDDLE: {
1064                 BKEY_PADDED(k) split;
1065                 /*
1066                  * The insert key falls 'in the middle' of k
1067                  * The insert key splits k in 3:
1068                  * - start only in k, preserve
1069                  * - middle common section, invalidate in k
1070                  * - end only in k, preserve
1071                  *
1072                  * We update the old key to preserve the start,
1073                  * insert will be the new common section,
1074                  * we manually insert the end that we are preserving.
1075                  *
1076                  * modify k _before_ doing the insert (which will move
1077                  * what k points to)
1078                  */
1079                 bkey_reassemble(&split.k, k.s_c);
1080                 split.k.k.needs_whiteout |= bkey_written(l->b, _k);
1081
1082                 bch2_cut_back(bkey_start_pos(&insert->k), &split.k.k);
1083                 BUG_ON(bkey_deleted(&split.k.k));
1084
1085                 __bch2_cut_front(insert->k.p, k);
1086                 BUG_ON(bkey_deleted(k.k));
1087                 extent_save(l->b, _k, k.k);
1088                 verify_modified_extent(iter, _k);
1089
1090                 extent_bset_insert(c, iter, &split.k);
1091                 break;
1092         }
1093         }
1094 }
1095
1096 static void __bch2_insert_fixup_extent(struct extent_insert_state *s)
1097 {
1098         struct btree_iter *iter = s->insert->iter;
1099         struct btree_iter_level *l = &iter->l[0];
1100         struct bkey_packed *_k;
1101         struct bkey unpacked;
1102         struct bkey_i *insert = s->insert->k;
1103
1104         while (bkey_cmp(s->committed, insert->k.p) < 0 &&
1105                (_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
1106                                                       KEY_TYPE_discard))) {
1107                 struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
1108                 enum bch_extent_overlap overlap = bch2_extent_overlap(&insert->k, k.k);
1109
1110                 EBUG_ON(bkey_cmp(iter->pos, k.k->p) >= 0);
1111
1112                 if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
1113                         break;
1114
1115                 s->committed = bpos_min(s->insert->k->k.p, k.k->p);
1116
1117                 if (!bkey_whiteout(k.k))
1118                         s->update_journal = true;
1119
1120                 if (!s->update_journal) {
1121                         bch2_cut_front(s->committed, insert);
1122                         bch2_cut_front(s->committed, &s->whiteout);
1123                         bch2_btree_iter_set_pos_same_leaf(iter, s->committed);
1124                         goto next;
1125                 }
1126
1127                 /*
1128                  * When deleting, if possible just do it by switching the type
1129                  * of the key we're deleting, instead of creating and inserting
1130                  * a new whiteout:
1131                  */
1132                 if (s->deleting &&
1133                     !s->update_btree &&
1134                     !bkey_cmp(insert->k.p, k.k->p) &&
1135                     !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
1136                         if (!bkey_whiteout(k.k)) {
1137                                 btree_account_key_drop(l->b, _k);
1138                                 _k->type = KEY_TYPE_discard;
1139                                 reserve_whiteout(l->b, _k);
1140                         }
1141                         break;
1142                 }
1143
1144                 if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
1145                         insert->k.needs_whiteout = true;
1146                         s->update_btree = true;
1147                 }
1148
1149                 if (s->update_btree &&
1150                     overlap == BCH_EXTENT_OVERLAP_ALL &&
1151                     bkey_whiteout(k.k) &&
1152                     k.k->needs_whiteout) {
1153                         unreserve_whiteout(l->b, _k);
1154                         _k->needs_whiteout = false;
1155                 }
1156
1157                 extent_squash(s, insert, _k, k, overlap);
1158
1159                 if (!s->update_btree)
1160                         bch2_cut_front(s->committed, insert);
1161 next:
1162                 if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
1163                     overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1164                         break;
1165         }
1166
1167         if (bkey_cmp(s->committed, insert->k.p) < 0)
1168                 s->committed = bpos_min(s->insert->k->k.p, l->b->key.k.p);
1169
1170         /*
1171          * may have skipped past some deleted extents greater than the insert
1172          * key, before we got to a non deleted extent and knew we could bail out
1173          * rewind the iterator a bit if necessary:
1174          */
1175         {
1176                 struct btree_node_iter node_iter = l->iter;
1177
1178                 while ((_k = bch2_btree_node_iter_prev_all(&node_iter, l->b)) &&
1179                        bkey_cmp_left_packed(l->b, _k, &s->committed) > 0)
1180                         l->iter = node_iter;
1181         }
1182 }
1183
1184 /**
1185  * bch_extent_insert_fixup - insert a new extent and deal with overlaps
1186  *
1187  * this may result in not actually doing the insert, or inserting some subset
1188  * of the insert key. For cmpxchg operations this is where that logic lives.
1189  *
1190  * All subsets of @insert that need to be inserted are inserted using
1191  * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
1192  * returns false, setting @iter->pos for the prefix of @insert that actually got
1193  * inserted.
1194  *
1195  * BSET INVARIANTS: this function is responsible for maintaining all the
1196  * invariants for bsets of extents in memory. things get really hairy with 0
1197  * size extents
1198  *
1199  * within one bset:
1200  *
1201  * bkey_start_pos(bkey_next(k)) >= k
1202  * or bkey_start_offset(bkey_next(k)) >= k->offset
1203  *
1204  * i.e. strict ordering, no overlapping extents.
1205  *
1206  * multiple bsets (i.e. full btree node):
1207  *
1208  * ∀ k, j
1209  *   k.size != 0 ∧ j.size != 0 →
1210  *     ¬ (k > bkey_start_pos(j) ∧ k < j)
1211  *
1212  * i.e. no two overlapping keys _of nonzero size_
1213  *
1214  * We can't realistically maintain this invariant for zero size keys because of
1215  * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
1216  * there may be another 0 size key between them in another bset, and it will
1217  * thus overlap with the merged key.
1218  *
1219  * In addition, the end of iter->pos indicates how much has been processed.
1220  * If the end of iter->pos is not the same as the end of insert, then
1221  * key insertion needs to continue/be retried.
1222  */
1223 enum btree_insert_ret
1224 bch2_insert_fixup_extent(struct btree_insert *trans,
1225                          struct btree_insert_entry *insert)
1226 {
1227         struct btree_iter *iter = insert->iter;
1228         struct btree *b         = iter->l[0].b;
1229         struct extent_insert_state s = {
1230                 .trans          = trans,
1231                 .insert         = insert,
1232                 .committed      = iter->pos,
1233
1234                 .whiteout       = *insert->k,
1235                 .update_journal = !bkey_whiteout(&insert->k->k),
1236                 .update_btree   = !bkey_whiteout(&insert->k->k),
1237                 .deleting       = bkey_whiteout(&insert->k->k),
1238         };
1239
1240         EBUG_ON(iter->level);
1241         EBUG_ON(!insert->k->k.size);
1242
1243         /*
1244          * As we process overlapping extents, we advance @iter->pos both to
1245          * signal to our caller (btree_insert_key()) how much of @insert->k has
1246          * been inserted, and also to keep @iter->pos consistent with
1247          * @insert->k and the node iterator that we're advancing:
1248          */
1249         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1250
1251         __bch2_insert_fixup_extent(&s);
1252
1253         extent_insert_committed(&s);
1254
1255         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1256         EBUG_ON(bkey_cmp(iter->pos, s.committed));
1257
1258         if (insert->k->k.size) {
1259                 /* got to the end of this leaf node */
1260                 BUG_ON(bkey_cmp(iter->pos, b->key.k.p));
1261                 return BTREE_INSERT_NEED_TRAVERSE;
1262         }
1263
1264         return BTREE_INSERT_OK;
1265 }
1266
1267 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
1268 {
1269         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1270         const union bch_extent_entry *entry;
1271         struct bch_extent_crc_unpacked crc;
1272         const struct bch_extent_ptr *ptr;
1273         unsigned size_ondisk = e.k->size;
1274         const char *reason;
1275         unsigned nonce = UINT_MAX;
1276
1277         if (bkey_val_u64s(e.k) > BKEY_EXTENT_VAL_U64s_MAX)
1278                 return "value too big";
1279
1280         extent_for_each_entry(e, entry) {
1281                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
1282                         return "invalid extent entry type";
1283
1284                 switch (extent_entry_type(entry)) {
1285                 case BCH_EXTENT_ENTRY_ptr:
1286                         ptr = entry_to_ptr(entry);
1287
1288                         reason = extent_ptr_invalid(c, e.s_c, &entry->ptr,
1289                                                     size_ondisk, false);
1290                         if (reason)
1291                                 return reason;
1292                         break;
1293                 case BCH_EXTENT_ENTRY_crc32:
1294                 case BCH_EXTENT_ENTRY_crc64:
1295                 case BCH_EXTENT_ENTRY_crc128:
1296                         crc = bch2_extent_crc_unpack(e.k, entry_to_crc(entry));
1297
1298                         if (crc.offset + e.k->size >
1299                             crc.uncompressed_size)
1300                                 return "checksum offset + key size > uncompressed size";
1301
1302                         size_ondisk = crc.compressed_size;
1303
1304                         if (!bch2_checksum_type_valid(c, crc.csum_type))
1305                                 return "invalid checksum type";
1306
1307                         if (crc.compression_type >= BCH_COMPRESSION_NR)
1308                                 return "invalid compression type";
1309
1310                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1311                                 if (nonce == UINT_MAX)
1312                                         nonce = crc.offset + crc.nonce;
1313                                 else if (nonce != crc.offset + crc.nonce)
1314                                         return "incorrect nonce";
1315                         }
1316                         break;
1317                 case BCH_EXTENT_ENTRY_stripe_ptr:
1318                         break;
1319                 }
1320         }
1321
1322         return NULL;
1323 }
1324
1325 void bch2_extent_debugcheck(struct bch_fs *c, struct btree *b,
1326                             struct bkey_s_c k)
1327 {
1328         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1329         const union bch_extent_entry *entry;
1330         struct extent_ptr_decoded p;
1331         char buf[160];
1332
1333         /*
1334          * XXX: we should be doing most/all of these checks at startup time,
1335          * where we check bch2_bkey_invalid() in btree_node_read_done()
1336          *
1337          * But note that we can't check for stale pointers or incorrect gc marks
1338          * until after journal replay is done (it might be an extent that's
1339          * going to get overwritten during replay)
1340          */
1341
1342         bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1343                        !bch2_bkey_replicas_marked(c, e.s_c, false), c,
1344                        "extent key bad (replicas not marked in superblock):\n%s",
1345                        (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
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         const char *invalid;
1384
1385         bkey_ptrs_to_text(out, c, k);
1386
1387         invalid = bch2_extent_invalid(c, k);
1388         if (invalid)
1389                 pr_buf(out, " invalid: %s", invalid);
1390 }
1391
1392 static void bch2_extent_crc_init(union bch_extent_crc *crc,
1393                                  struct bch_extent_crc_unpacked new)
1394 {
1395 #define common_fields(_crc)                                             \
1396                 .csum_type              = _crc.csum_type,               \
1397                 .compression_type       = _crc.compression_type,        \
1398                 ._compressed_size       = _crc.compressed_size - 1,     \
1399                 ._uncompressed_size     = _crc.uncompressed_size - 1,   \
1400                 .offset                 = _crc.offset
1401
1402         if (bch_crc_bytes[new.csum_type]        <= 4 &&
1403             new.uncompressed_size               <= CRC32_SIZE_MAX &&
1404             new.nonce                           <= CRC32_NONCE_MAX) {
1405                 crc->crc32 = (struct bch_extent_crc32) {
1406                         .type = 1 << BCH_EXTENT_ENTRY_crc32,
1407                         common_fields(new),
1408                         .csum                   = *((__le32 *) &new.csum.lo),
1409                 };
1410                 return;
1411         }
1412
1413         if (bch_crc_bytes[new.csum_type]        <= 10 &&
1414             new.uncompressed_size               <= CRC64_SIZE_MAX &&
1415             new.nonce                           <= CRC64_NONCE_MAX) {
1416                 crc->crc64 = (struct bch_extent_crc64) {
1417                         .type = 1 << BCH_EXTENT_ENTRY_crc64,
1418                         common_fields(new),
1419                         .nonce                  = new.nonce,
1420                         .csum_lo                = new.csum.lo,
1421                         .csum_hi                = *((__le16 *) &new.csum.hi),
1422                 };
1423                 return;
1424         }
1425
1426         if (bch_crc_bytes[new.csum_type]        <= 16 &&
1427             new.uncompressed_size               <= CRC128_SIZE_MAX &&
1428             new.nonce                           <= CRC128_NONCE_MAX) {
1429                 crc->crc128 = (struct bch_extent_crc128) {
1430                         .type = 1 << BCH_EXTENT_ENTRY_crc128,
1431                         common_fields(new),
1432                         .nonce                  = new.nonce,
1433                         .csum                   = new.csum,
1434                 };
1435                 return;
1436         }
1437 #undef common_fields
1438         BUG();
1439 }
1440
1441 void bch2_extent_crc_append(struct bkey_i_extent *e,
1442                             struct bch_extent_crc_unpacked new)
1443 {
1444         bch2_extent_crc_init((void *) extent_entry_last(extent_i_to_s(e)), new);
1445         __extent_entry_push(e);
1446 }
1447
1448 static inline void __extent_entry_insert(struct bkey_i_extent *e,
1449                                          union bch_extent_entry *dst,
1450                                          union bch_extent_entry *new)
1451 {
1452         union bch_extent_entry *end = extent_entry_last(extent_i_to_s(e));
1453
1454         memmove_u64s_up((u64 *) dst + extent_entry_u64s(new),
1455                         dst, (u64 *) end - (u64 *) dst);
1456         e->k.u64s += extent_entry_u64s(new);
1457         memcpy(dst, new, extent_entry_bytes(new));
1458 }
1459
1460 void bch2_extent_ptr_decoded_append(struct bkey_i_extent *e,
1461                                     struct extent_ptr_decoded *p)
1462 {
1463         struct bch_extent_crc_unpacked crc;
1464         union bch_extent_entry *pos;
1465         unsigned i;
1466
1467         extent_for_each_crc(extent_i_to_s(e), crc, pos)
1468                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1469                         pos = extent_entry_next(pos);
1470                         goto found;
1471                 }
1472
1473         bch2_extent_crc_append(e, p->crc);
1474         pos = extent_entry_last(extent_i_to_s(e));
1475 found:
1476         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
1477         __extent_entry_insert(e, pos, to_entry(&p->ptr));
1478
1479         for (i = 0; i < p->ec_nr; i++) {
1480                 p->ec[i].type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
1481                 __extent_entry_insert(e, pos, to_entry(&p->ec[i]));
1482         }
1483 }
1484
1485 /*
1486  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
1487  *
1488  * Returns true if @k should be dropped entirely
1489  *
1490  * For existing keys, only called when btree nodes are being rewritten, not when
1491  * they're merely being compacted/resorted in memory.
1492  */
1493 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1494 {
1495         struct bch_extent_ptr *ptr;
1496
1497         bch2_bkey_drop_ptrs(k, ptr,
1498                 ptr->cached &&
1499                 ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
1500
1501         /* will only happen if all pointers were cached: */
1502         if (!bkey_val_u64s(k.k))
1503                 k.k->type = KEY_TYPE_deleted;
1504
1505         return false;
1506 }
1507
1508 void bch2_extent_mark_replicas_cached(struct bch_fs *c,
1509                                       struct bkey_s_extent e,
1510                                       unsigned target,
1511                                       unsigned nr_desired_replicas)
1512 {
1513         union bch_extent_entry *entry;
1514         struct extent_ptr_decoded p;
1515         int extra = bch2_bkey_durability(c, e.s_c) - nr_desired_replicas;
1516
1517         if (target && extra > 0)
1518                 extent_for_each_ptr_decode(e, p, entry) {
1519                         int n = bch2_extent_ptr_durability(c, p);
1520
1521                         if (n && n <= extra &&
1522                             !bch2_dev_in_target(c, p.ptr.dev, target)) {
1523                                 entry->ptr.cached = true;
1524                                 extra -= n;
1525                         }
1526                 }
1527
1528         if (extra > 0)
1529                 extent_for_each_ptr_decode(e, p, entry) {
1530                         int n = bch2_extent_ptr_durability(c, p);
1531
1532                         if (n && n <= extra) {
1533                                 entry->ptr.cached = true;
1534                                 extra -= n;
1535                         }
1536                 }
1537 }
1538
1539 enum merge_result bch2_extent_merge(struct bch_fs *c,
1540                                     struct bkey_i *l, struct bkey_i *r)
1541 {
1542         struct bkey_s_extent el = bkey_i_to_s_extent(l);
1543         struct bkey_s_extent er = bkey_i_to_s_extent(r);
1544         union bch_extent_entry *en_l, *en_r;
1545
1546         if (bkey_val_u64s(&l->k) != bkey_val_u64s(&r->k))
1547                 return BCH_MERGE_NOMERGE;
1548
1549         extent_for_each_entry(el, en_l) {
1550                 struct bch_extent_ptr *lp, *rp;
1551                 struct bch_dev *ca;
1552
1553                 en_r = vstruct_idx(er.v, (u64 *) en_l - el.v->_data);
1554
1555                 if ((extent_entry_type(en_l) !=
1556                      extent_entry_type(en_r)) ||
1557                     !extent_entry_is_ptr(en_l))
1558                         return BCH_MERGE_NOMERGE;
1559
1560                 lp = &en_l->ptr;
1561                 rp = &en_r->ptr;
1562
1563                 if (lp->offset + el.k->size     != rp->offset ||
1564                     lp->dev                     != rp->dev ||
1565                     lp->gen                     != rp->gen)
1566                         return BCH_MERGE_NOMERGE;
1567
1568                 /* We don't allow extents to straddle buckets: */
1569                 ca = bch_dev_bkey_exists(c, lp->dev);
1570
1571                 if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
1572                         return BCH_MERGE_NOMERGE;
1573         }
1574
1575         l->k.needs_whiteout |= r->k.needs_whiteout;
1576
1577         /* Keys with no pointers aren't restricted to one bucket and could
1578          * overflow KEY_SIZE
1579          */
1580         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1581                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1582                 bch2_cut_front(l->k.p, r);
1583                 return BCH_MERGE_PARTIAL;
1584         }
1585
1586         bch2_key_resize(&l->k, l->k.size + r->k.size);
1587
1588         return BCH_MERGE_MERGE;
1589 }
1590
1591 /*
1592  * When merging an extent that we're inserting into a btree node, the new merged
1593  * extent could overlap with an existing 0 size extent - if we don't fix that,
1594  * it'll break the btree node iterator so this code finds those 0 size extents
1595  * and shifts them out of the way.
1596  *
1597  * Also unpacks and repacks.
1598  */
1599 static bool bch2_extent_merge_inline(struct bch_fs *c,
1600                                      struct btree_iter *iter,
1601                                      struct bkey_packed *l,
1602                                      struct bkey_packed *r,
1603                                      bool back_merge)
1604 {
1605         struct btree *b = iter->l[0].b;
1606         struct btree_node_iter *node_iter = &iter->l[0].iter;
1607         BKEY_PADDED(k) li, ri;
1608         struct bkey_packed *m   = back_merge ? l : r;
1609         struct bkey_i *mi       = back_merge ? &li.k : &ri.k;
1610         struct bset_tree *t     = bch2_bkey_to_bset(b, m);
1611         enum merge_result ret;
1612
1613         EBUG_ON(bkey_written(b, m));
1614
1615         /*
1616          * We need to save copies of both l and r, because we might get a
1617          * partial merge (which modifies both) and then fails to repack
1618          */
1619         bch2_bkey_unpack(b, &li.k, l);
1620         bch2_bkey_unpack(b, &ri.k, r);
1621
1622         ret = bch2_bkey_merge(c, &li.k, &ri.k);
1623         if (ret == BCH_MERGE_NOMERGE)
1624                 return false;
1625
1626         /*
1627          * check if we overlap with deleted extents - would break the sort
1628          * order:
1629          */
1630         if (back_merge) {
1631                 struct bkey_packed *n = bkey_next(m);
1632
1633                 if (n != btree_bkey_last(b, t) &&
1634                     bkey_cmp_left_packed(b, n, &li.k.k.p) <= 0 &&
1635                     bkey_deleted(n))
1636                         return false;
1637         } else if (ret == BCH_MERGE_MERGE) {
1638                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1639
1640                 if (prev &&
1641                     bkey_cmp_left_packed_byval(b, prev,
1642                                 bkey_start_pos(&li.k.k)) > 0)
1643                         return false;
1644         }
1645
1646         if (ret == BCH_MERGE_PARTIAL) {
1647                 if (!extent_i_save(b, m, mi))
1648                         return false;
1649
1650                 if (!back_merge)
1651                         bkey_copy(packed_to_bkey(l), &li.k);
1652                 else
1653                         bkey_copy(packed_to_bkey(r), &ri.k);
1654         } else {
1655                 if (!extent_i_save(b, m, &li.k))
1656                         return false;
1657         }
1658
1659         bch2_bset_fix_invalidated_key(b, m);
1660         bch2_btree_node_iter_fix(iter, b, node_iter,
1661                                  m, m->u64s, m->u64s);
1662         verify_modified_extent(iter, m);
1663
1664         return ret == BCH_MERGE_MERGE;
1665 }
1666
1667 int bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size)
1668 {
1669         struct btree_iter iter;
1670         struct bpos end = pos;
1671         struct bkey_s_c k;
1672         int ret = 0;
1673
1674         end.offset += size;
1675
1676         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, pos,
1677                              BTREE_ITER_SLOTS, k) {
1678                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
1679                         break;
1680
1681                 if (!bch2_extent_is_fully_allocated(k)) {
1682                         ret = -ENOSPC;
1683                         break;
1684                 }
1685         }
1686         bch2_btree_iter_unlock(&iter);
1687
1688         return ret;
1689 }
1690
1691 /* KEY_TYPE_reservation: */
1692
1693 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
1694 {
1695         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1696
1697         if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
1698                 return "incorrect value size";
1699
1700         if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
1701                 return "invalid nr_replicas";
1702
1703         return NULL;
1704 }
1705
1706 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
1707                               struct bkey_s_c k)
1708 {
1709         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1710
1711         pr_buf(out, "generation %u replicas %u",
1712                le32_to_cpu(r.v->generation),
1713                r.v->nr_replicas);
1714 }
1715
1716 enum merge_result bch2_reservation_merge(struct bch_fs *c,
1717                                          struct bkey_i *l, struct bkey_i *r)
1718 {
1719         struct bkey_i_reservation *li = bkey_i_to_reservation(l);
1720         struct bkey_i_reservation *ri = bkey_i_to_reservation(r);
1721
1722         if (li->v.generation != ri->v.generation ||
1723             li->v.nr_replicas != ri->v.nr_replicas)
1724                 return BCH_MERGE_NOMERGE;
1725
1726         l->k.needs_whiteout |= r->k.needs_whiteout;
1727
1728         /* Keys with no pointers aren't restricted to one bucket and could
1729          * overflow KEY_SIZE
1730          */
1731         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1732                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1733                 bch2_cut_front(l->k.p, r);
1734                 return BCH_MERGE_PARTIAL;
1735         }
1736
1737         bch2_key_resize(&l->k, l->k.size + r->k.size);
1738
1739         return BCH_MERGE_MERGE;
1740 }