fs/bcachefs/extents.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  *
   5  * Code for managing the extent btree and dynamically updating the writeback
   6  * dirty sector count.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "bkey_methods.h"
  11 #include "btree_gc.h"
  12 #include "btree_update.h"
  13 #include "btree_update_interior.h"
  14 #include "buckets.h"
  15 #include "checksum.h"
  16 #include "debug.h"
  17 #include "dirent.h"
  18 #include "disk_groups.h"
  19 #include "error.h"
  20 #include "extents.h"
  21 #include "inode.h"
  22 #include "journal.h"
  23 #include "replicas.h"
  24 #include "super.h"
  25 #include "super-io.h"
  26 #include "trace.h"
  27 #include "util.h"
  28 #include "xattr.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 static bool bch2_extent_merge_inline(struct bch_fs *,
 786                                      struct btree_iter *,
 787                                      struct bkey_packed *,
 788                                      struct bkey_packed *,
 789                                      bool);
 790
 791 static void verify_extent_nonoverlapping(struct btree *b,
 792                                          struct btree_node_iter *_iter,
 793                                          struct bkey_i *insert)
 794 {
 795 #ifdef CONFIG_BCACHEFS_DEBUG
 796         struct btree_node_iter iter;
 797         struct bkey_packed *k;
 798         struct bkey uk;
 799
 800         iter = *_iter;
 801         k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
 802         BUG_ON(k &&
 803                (uk = bkey_unpack_key(b, k),
 804                 bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
 805
 806         iter = *_iter;
 807         k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
 808 #if 0
 809         BUG_ON(k &&
 810                (uk = bkey_unpack_key(b, k),
 811                 bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
 812 #else
 813         if (k &&
 814             (uk = bkey_unpack_key(b, k),
 815              bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
 816                 char buf1[100];
 817                 char buf2[100];
 818
 819                 bch2_bkey_to_text(&PBUF(buf1), &insert->k);
 820                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 821
 822                 bch2_dump_btree_node(b);
 823                 panic("insert > next :\n"
 824                       "insert %s\n"
 825                       "next   %s\n",
 826                       buf1, buf2);
 827         }
 828 #endif
 829
 830 #endif
 831 }
 832
 833 static void verify_modified_extent(struct btree_iter *iter,
 834                                    struct bkey_packed *k)
 835 {
 836         bch2_btree_iter_verify(iter, iter->l[0].b);
 837         bch2_verify_insert_pos(iter->l[0].b, k, k, k->u64s);
 838 }
 839
 840 static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 841                                struct bkey_i *insert)
 842 {
 843         struct btree_iter_level *l = &iter->l[0];
 844         struct btree_node_iter node_iter;
 845         struct bkey_packed *k;
 846
 847         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
 848
 849         EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
 850         verify_extent_nonoverlapping(l->b, &l->iter, insert);
 851
 852         node_iter = l->iter;
 853         k = bch2_btree_node_iter_prev_filter(&node_iter, l->b, KEY_TYPE_discard);
 854         if (k && !bkey_written(l->b, k) &&
 855             bch2_extent_merge_inline(c, iter, k, bkey_to_packed(insert), true))
 856                 return;
 857
 858         node_iter = l->iter;
 859         k = bch2_btree_node_iter_peek_filter(&node_iter, l->b, KEY_TYPE_discard);
 860         if (k && !bkey_written(l->b, k) &&
 861             bch2_extent_merge_inline(c, iter, bkey_to_packed(insert), k, false))
 862                 return;
 863
 864         k = bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
 865
 866         bch2_bset_insert(l->b, &l->iter, k, insert, 0);
 867         bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
 868         bch2_btree_iter_verify(iter, l->b);
 869 }
 870
 871 static inline struct bpos
 872 bch2_extent_atomic_end(struct bkey_i *k, struct btree_iter *iter)
 873 {
 874         struct btree *b = iter->l[0].b;
 875
 876         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 877         BUG_ON(bkey_cmp(bkey_start_pos(&k->k), b->data->min_key) < 0);
 878
 879         return bpos_min(k->k.p, b->key.k.p);
 880 }
 881
 882 void bch2_extent_trim_atomic(struct bkey_i *k, struct btree_iter *iter)
 883 {
 884         bch2_cut_back(bch2_extent_atomic_end(k, iter), &k->k);
 885 }
 886
 887 bool bch2_extent_is_atomic(struct bkey_i *k, struct btree_iter *iter)
 888 {
 889         return !bkey_cmp(bch2_extent_atomic_end(k, iter), k->k.p);
 890 }
 891
 892 enum btree_insert_ret
 893 bch2_extent_can_insert(struct btree_insert *trans,
 894                        struct btree_insert_entry *insert,
 895                        unsigned *u64s)
 896 {
 897         struct btree_iter_level *l = &insert->iter->l[0];
 898         struct btree_node_iter node_iter = l->iter;
 899         enum bch_extent_overlap overlap;
 900         struct bkey_packed *_k;
 901         struct bkey unpacked;
 902         struct bkey_s_c k;
 903         int sectors;
 904
 905         /*
 906          * We avoid creating whiteouts whenever possible when deleting, but
 907          * those optimizations mean we may potentially insert two whiteouts
 908          * instead of one (when we overlap with the front of one extent and the
 909          * back of another):
 910          */
 911         if (bkey_whiteout(&insert->k->k))
 912                 *u64s += BKEY_U64s;
 913
 914         _k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
 915                                               KEY_TYPE_discard);
 916         if (!_k)
 917                 return BTREE_INSERT_OK;
 918
 919         k = bkey_disassemble(l->b, _k, &unpacked);
 920
 921         overlap = bch2_extent_overlap(&insert->k->k, k.k);
 922
 923         /* account for having to split existing extent: */
 924         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
 925                 *u64s += _k->u64s;
 926
 927         if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
 928             (sectors = bch2_extent_is_compressed(k))) {
 929                 int flags = trans->flags & BTREE_INSERT_NOFAIL
 930                         ? BCH_DISK_RESERVATION_NOFAIL : 0;
 931
 932                 switch (bch2_disk_reservation_add(trans->c,
 933                                 trans->disk_res,
 934                                 sectors, flags)) {
 935                 case 0:
 936                         break;
 937                 case -ENOSPC:
 938                         return BTREE_INSERT_ENOSPC;
 939                 default:
 940                         BUG();
 941                 }
 942         }
 943
 944         return BTREE_INSERT_OK;
 945 }
 946
 947 static void
 948 extent_squash(struct bch_fs *c, struct btree_iter *iter,
 949               struct bkey_i *insert,
 950               struct bkey_packed *_k, struct bkey_s k,
 951               enum bch_extent_overlap overlap)
 952 {
 953         struct btree_iter_level *l = &iter->l[0];
 954
 955         switch (overlap) {
 956         case BCH_EXTENT_OVERLAP_FRONT:
 957                 /* insert overlaps with start of k: */
 958                 __bch2_cut_front(insert->k.p, k);
 959                 BUG_ON(bkey_deleted(k.k));
 960                 extent_save(l->b, _k, k.k);
 961                 verify_modified_extent(iter, _k);
 962                 break;
 963
 964         case BCH_EXTENT_OVERLAP_BACK:
 965                 /* insert overlaps with end of k: */
 966                 bch2_cut_back(bkey_start_pos(&insert->k), k.k);
 967                 BUG_ON(bkey_deleted(k.k));
 968                 extent_save(l->b, _k, k.k);
 969
 970                 /*
 971                  * As the auxiliary tree is indexed by the end of the
 972                  * key and we've just changed the end, update the
 973                  * auxiliary tree.
 974                  */
 975                 bch2_bset_fix_invalidated_key(l->b, _k);
 976                 bch2_btree_node_iter_fix(iter, l->b, &l->iter,
 977                                          _k, _k->u64s, _k->u64s);
 978                 verify_modified_extent(iter, _k);
 979                 break;
 980
 981         case BCH_EXTENT_OVERLAP_ALL: {
 982                 /* The insert key completely covers k, invalidate k */
 983                 if (!bkey_whiteout(k.k))
 984                         btree_account_key_drop(l->b, _k);
 985
 986                 k.k->size = 0;
 987                 k.k->type = KEY_TYPE_deleted;
 988
 989                 if (_k >= btree_bset_last(l->b)->start) {
 990                         unsigned u64s = _k->u64s;
 991
 992                         bch2_bset_delete(l->b, _k, _k->u64s);
 993                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
 994                                                  _k, u64s, 0);
 995                         bch2_btree_iter_verify(iter, l->b);
 996                 } else {
 997                         extent_save(l->b, _k, k.k);
 998                         bch2_btree_node_iter_fix(iter, l->b, &l->iter,
 999                                                  _k, _k->u64s, _k->u64s);
1000                         verify_modified_extent(iter, _k);
1001                 }
1002
1003                 break;
1004         }
1005         case BCH_EXTENT_OVERLAP_MIDDLE: {
1006                 BKEY_PADDED(k) split;
1007                 /*
1008                  * The insert key falls 'in the middle' of k
1009                  * The insert key splits k in 3:
1010                  * - start only in k, preserve
1011                  * - middle common section, invalidate in k
1012                  * - end only in k, preserve
1013                  *
1014                  * We update the old key to preserve the start,
1015                  * insert will be the new common section,
1016                  * we manually insert the end that we are preserving.
1017                  *
1018                  * modify k _before_ doing the insert (which will move
1019                  * what k points to)
1020                  */
1021                 bkey_reassemble(&split.k, k.s_c);
1022                 split.k.k.needs_whiteout |= bkey_written(l->b, _k);
1023
1024                 bch2_cut_back(bkey_start_pos(&insert->k), &split.k.k);
1025                 BUG_ON(bkey_deleted(&split.k.k));
1026
1027                 __bch2_cut_front(insert->k.p, k);
1028                 BUG_ON(bkey_deleted(k.k));
1029                 extent_save(l->b, _k, k.k);
1030                 verify_modified_extent(iter, _k);
1031
1032                 extent_bset_insert(c, iter, &split.k);
1033                 break;
1034         }
1035         }
1036 }
1037
1038 struct extent_insert_state {
1039         struct bkey_i                   whiteout;
1040         bool                            update_journal;
1041         bool                            update_btree;
1042         bool                            deleting;
1043 };
1044
1045 static void __bch2_insert_fixup_extent(struct bch_fs *c,
1046                                        struct btree_iter *iter,
1047                                        struct bkey_i *insert,
1048                                        struct extent_insert_state *s)
1049 {
1050         struct btree_iter_level *l = &iter->l[0];
1051         struct bkey_packed *_k;
1052         struct bkey unpacked;
1053
1054         while ((_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
1055                                                       KEY_TYPE_discard))) {
1056                 struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
1057                 struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
1058                 enum bch_extent_overlap overlap =
1059                         bch2_extent_overlap(&insert->k, k.k);
1060
1061                 if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
1062                         break;
1063
1064                 if (!bkey_whiteout(k.k))
1065                         s->update_journal = true;
1066
1067                 if (!s->update_journal) {
1068                         bch2_cut_front(cur_end, insert);
1069                         bch2_cut_front(cur_end, &s->whiteout);
1070                         bch2_btree_iter_set_pos_same_leaf(iter, cur_end);
1071                         goto next;
1072                 }
1073
1074                 /*
1075                  * When deleting, if possible just do it by switching the type
1076                  * of the key we're deleting, instead of creating and inserting
1077                  * a new whiteout:
1078                  */
1079                 if (s->deleting &&
1080                     !s->update_btree &&
1081                     !bkey_cmp(insert->k.p, k.k->p) &&
1082                     !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
1083                         if (!bkey_whiteout(k.k)) {
1084                                 btree_account_key_drop(l->b, _k);
1085                                 _k->type = KEY_TYPE_discard;
1086                                 reserve_whiteout(l->b, _k);
1087                         }
1088                         break;
1089                 }
1090
1091                 if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
1092                         insert->k.needs_whiteout = true;
1093                         s->update_btree = true;
1094                 }
1095
1096                 if (s->update_btree &&
1097                     overlap == BCH_EXTENT_OVERLAP_ALL &&
1098                     bkey_whiteout(k.k) &&
1099                     k.k->needs_whiteout) {
1100                         unreserve_whiteout(l->b, _k);
1101                         _k->needs_whiteout = false;
1102                 }
1103
1104                 extent_squash(c, iter, insert, _k, k, overlap);
1105
1106                 if (!s->update_btree)
1107                         bch2_cut_front(cur_end, insert);
1108 next:
1109                 if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
1110                     overlap == BCH_EXTENT_OVERLAP_MIDDLE)
1111                         break;
1112         }
1113
1114         /*
1115          * may have skipped past some deleted extents greater than the insert
1116          * key, before we got to a non deleted extent and knew we could bail out
1117          * rewind the iterator a bit if necessary:
1118          */
1119         {
1120                 struct btree_node_iter node_iter = l->iter;
1121
1122                 while ((_k = bch2_btree_node_iter_prev_all(&node_iter, l->b)) &&
1123                        bkey_cmp_left_packed(l->b, _k, &insert->k.p) > 0)
1124                         l->iter = node_iter;
1125         }
1126 }
1127
1128 /**
1129  * bch_extent_insert_fixup - insert a new extent and deal with overlaps
1130  *
1131  * this may result in not actually doing the insert, or inserting some subset
1132  * of the insert key. For cmpxchg operations this is where that logic lives.
1133  *
1134  * All subsets of @insert that need to be inserted are inserted using
1135  * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
1136  * returns false, setting @iter->pos for the prefix of @insert that actually got
1137  * inserted.
1138  *
1139  * BSET INVARIANTS: this function is responsible for maintaining all the
1140  * invariants for bsets of extents in memory. things get really hairy with 0
1141  * size extents
1142  *
1143  * within one bset:
1144  *
1145  * bkey_start_pos(bkey_next(k)) >= k
1146  * or bkey_start_offset(bkey_next(k)) >= k->offset
1147  *
1148  * i.e. strict ordering, no overlapping extents.
1149  *
1150  * multiple bsets (i.e. full btree node):
1151  *
1152  * ∀ k, j
1153  *   k.size != 0 ∧ j.size != 0 →
1154  *     ¬ (k > bkey_start_pos(j) ∧ k < j)
1155  *
1156  * i.e. no two overlapping keys _of nonzero size_
1157  *
1158  * We can't realistically maintain this invariant for zero size keys because of
1159  * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
1160  * there may be another 0 size key between them in another bset, and it will
1161  * thus overlap with the merged key.
1162  *
1163  * In addition, the end of iter->pos indicates how much has been processed.
1164  * If the end of iter->pos is not the same as the end of insert, then
1165  * key insertion needs to continue/be retried.
1166  */
1167 void bch2_insert_fixup_extent(struct btree_insert *trans,
1168                               struct btree_insert_entry *insert)
1169 {
1170         struct bch_fs *c = trans->c;
1171         struct btree_iter *iter = insert->iter;
1172         struct extent_insert_state s = {
1173                 .whiteout       = *insert->k,
1174                 .update_journal = !bkey_whiteout(&insert->k->k),
1175                 .update_btree   = !bkey_whiteout(&insert->k->k),
1176                 .deleting       = bkey_whiteout(&insert->k->k),
1177         };
1178         BKEY_PADDED(k) tmp;
1179
1180         EBUG_ON(iter->level);
1181         EBUG_ON(!insert->k->k.size);
1182         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
1183
1184         __bch2_insert_fixup_extent(c, iter, insert->k, &s);
1185
1186         bch2_btree_iter_set_pos_same_leaf(iter, insert->k->k.p);
1187
1188         if (s.update_btree) {
1189                 bkey_copy(&tmp.k, insert->k);
1190
1191                 if (s.deleting)
1192                         tmp.k.k.type = KEY_TYPE_discard;
1193
1194                 if (debug_check_bkeys(c))
1195                         bch2_bkey_debugcheck(c, iter->l[0].b,
1196                                              bkey_i_to_s_c(&tmp.k));
1197
1198                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1199
1200                 extent_bset_insert(c, iter, &tmp.k);
1201         }
1202
1203         if (s.update_journal) {
1204                 bkey_copy(&tmp.k, !s.deleting ? insert->k : &s.whiteout);
1205
1206                 if (s.deleting)
1207                         tmp.k.k.type = KEY_TYPE_discard;
1208
1209                 EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
1210
1211                 bch2_btree_journal_key(trans, iter, &tmp.k);
1212         }
1213
1214         bch2_cut_front(insert->k->k.p, insert->k);
1215 }
1216
1217 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
1218 {
1219         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1220         const union bch_extent_entry *entry;
1221         struct bch_extent_crc_unpacked crc;
1222         const struct bch_extent_ptr *ptr;
1223         unsigned size_ondisk = e.k->size;
1224         const char *reason;
1225         unsigned nonce = UINT_MAX;
1226
1227         if (bkey_val_u64s(e.k) > BKEY_EXTENT_VAL_U64s_MAX)
1228                 return "value too big";
1229
1230         extent_for_each_entry(e, entry) {
1231                 if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
1232                         return "invalid extent entry type";
1233
1234                 switch (extent_entry_type(entry)) {
1235                 case BCH_EXTENT_ENTRY_ptr:
1236                         ptr = entry_to_ptr(entry);
1237
1238                         reason = extent_ptr_invalid(c, e.s_c, &entry->ptr,
1239                                                     size_ondisk, false);
1240                         if (reason)
1241                                 return reason;
1242                         break;
1243                 case BCH_EXTENT_ENTRY_crc32:
1244                 case BCH_EXTENT_ENTRY_crc64:
1245                 case BCH_EXTENT_ENTRY_crc128:
1246                         crc = bch2_extent_crc_unpack(e.k, entry_to_crc(entry));
1247
1248                         if (crc.offset + e.k->size >
1249                             crc.uncompressed_size)
1250                                 return "checksum offset + key size > uncompressed size";
1251
1252                         size_ondisk = crc.compressed_size;
1253
1254                         if (!bch2_checksum_type_valid(c, crc.csum_type))
1255                                 return "invalid checksum type";
1256
1257                         if (crc.compression_type >= BCH_COMPRESSION_NR)
1258                                 return "invalid compression type";
1259
1260                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1261                                 if (nonce == UINT_MAX)
1262                                         nonce = crc.offset + crc.nonce;
1263                                 else if (nonce != crc.offset + crc.nonce)
1264                                         return "incorrect nonce";
1265                         }
1266                         break;
1267                 case BCH_EXTENT_ENTRY_stripe_ptr:
1268                         break;
1269                 }
1270         }
1271
1272         return NULL;
1273 }
1274
1275 void bch2_extent_debugcheck(struct bch_fs *c, struct btree *b,
1276                             struct bkey_s_c k)
1277 {
1278         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1279         const union bch_extent_entry *entry;
1280         struct extent_ptr_decoded p;
1281         char buf[160];
1282
1283         /*
1284          * XXX: we should be doing most/all of these checks at startup time,
1285          * where we check bch2_bkey_invalid() in btree_node_read_done()
1286          *
1287          * But note that we can't check for stale pointers or incorrect gc marks
1288          * until after journal replay is done (it might be an extent that's
1289          * going to get overwritten during replay)
1290          */
1291
1292         bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1293                        !bch2_bkey_replicas_marked(c, e.s_c, false), c,
1294                        "extent key bad (replicas not marked in superblock):\n%s",
1295                        (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
1296
1297         /*
1298          * If journal replay hasn't finished, we might be seeing keys
1299          * that will be overwritten by the time journal replay is done:
1300          */
1301         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1302                 return;
1303
1304         extent_for_each_ptr_decode(e, p, entry) {
1305                 struct bch_dev *ca      = bch_dev_bkey_exists(c, p.ptr.dev);
1306                 struct bucket_mark mark = ptr_bucket_mark(ca, &p.ptr);
1307                 unsigned stale          = gen_after(mark.gen, p.ptr.gen);
1308                 unsigned disk_sectors   = ptr_disk_sectors(p);
1309                 unsigned mark_sectors   = p.ptr.cached
1310                         ? mark.cached_sectors
1311                         : mark.dirty_sectors;
1312
1313                 bch2_fs_bug_on(stale && !p.ptr.cached, c,
1314                                "stale dirty pointer (ptr gen %u bucket %u",
1315                                p.ptr.gen, mark.gen);
1316
1317                 bch2_fs_bug_on(stale > 96, c, "key too stale: %i", stale);
1318
1319                 bch2_fs_bug_on(!stale &&
1320                                (mark.data_type != BCH_DATA_USER ||
1321                                 mark_sectors < disk_sectors), c,
1322                                "extent pointer not marked: %s:\n"
1323                                "type %u sectors %u < %u",
1324                                (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf),
1325                                mark.data_type,
1326                                mark_sectors, disk_sectors);
1327         }
1328 }
1329
1330 void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
1331                          struct bkey_s_c k)
1332 {
1333         const char *invalid;
1334
1335         bkey_ptrs_to_text(out, c, k);
1336
1337         invalid = bch2_extent_invalid(c, k);
1338         if (invalid)
1339                 pr_buf(out, " invalid: %s", invalid);
1340 }
1341
1342 static void bch2_extent_crc_init(union bch_extent_crc *crc,
1343                                  struct bch_extent_crc_unpacked new)
1344 {
1345 #define common_fields(_crc)                                             \
1346                 .csum_type              = _crc.csum_type,               \
1347                 .compression_type       = _crc.compression_type,        \
1348                 ._compressed_size       = _crc.compressed_size - 1,     \
1349                 ._uncompressed_size     = _crc.uncompressed_size - 1,   \
1350                 .offset                 = _crc.offset
1351
1352         if (bch_crc_bytes[new.csum_type]        <= 4 &&
1353             new.uncompressed_size               <= CRC32_SIZE_MAX &&
1354             new.nonce                           <= CRC32_NONCE_MAX) {
1355                 crc->crc32 = (struct bch_extent_crc32) {
1356                         .type = 1 << BCH_EXTENT_ENTRY_crc32,
1357                         common_fields(new),
1358                         .csum                   = *((__le32 *) &new.csum.lo),
1359                 };
1360                 return;
1361         }
1362
1363         if (bch_crc_bytes[new.csum_type]        <= 10 &&
1364             new.uncompressed_size               <= CRC64_SIZE_MAX &&
1365             new.nonce                           <= CRC64_NONCE_MAX) {
1366                 crc->crc64 = (struct bch_extent_crc64) {
1367                         .type = 1 << BCH_EXTENT_ENTRY_crc64,
1368                         common_fields(new),
1369                         .nonce                  = new.nonce,
1370                         .csum_lo                = new.csum.lo,
1371                         .csum_hi                = *((__le16 *) &new.csum.hi),
1372                 };
1373                 return;
1374         }
1375
1376         if (bch_crc_bytes[new.csum_type]        <= 16 &&
1377             new.uncompressed_size               <= CRC128_SIZE_MAX &&
1378             new.nonce                           <= CRC128_NONCE_MAX) {
1379                 crc->crc128 = (struct bch_extent_crc128) {
1380                         .type = 1 << BCH_EXTENT_ENTRY_crc128,
1381                         common_fields(new),
1382                         .nonce                  = new.nonce,
1383                         .csum                   = new.csum,
1384                 };
1385                 return;
1386         }
1387 #undef common_fields
1388         BUG();
1389 }
1390
1391 void bch2_extent_crc_append(struct bkey_i_extent *e,
1392                             struct bch_extent_crc_unpacked new)
1393 {
1394         bch2_extent_crc_init((void *) extent_entry_last(extent_i_to_s(e)), new);
1395         __extent_entry_push(e);
1396 }
1397
1398 static inline void __extent_entry_insert(struct bkey_i_extent *e,
1399                                          union bch_extent_entry *dst,
1400                                          union bch_extent_entry *new)
1401 {
1402         union bch_extent_entry *end = extent_entry_last(extent_i_to_s(e));
1403
1404         memmove_u64s_up((u64 *) dst + extent_entry_u64s(new),
1405                         dst, (u64 *) end - (u64 *) dst);
1406         e->k.u64s += extent_entry_u64s(new);
1407         memcpy_u64s_small(dst, new, extent_entry_u64s(new));
1408 }
1409
1410 void bch2_extent_ptr_decoded_append(struct bkey_i_extent *e,
1411                                     struct extent_ptr_decoded *p)
1412 {
1413         struct bch_extent_crc_unpacked crc = bch2_extent_crc_unpack(&e->k, NULL);
1414         union bch_extent_entry *pos;
1415         unsigned i;
1416
1417         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1418                 pos = e->v.start;
1419                 goto found;
1420         }
1421
1422         extent_for_each_crc(extent_i_to_s(e), crc, pos)
1423                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
1424                         pos = extent_entry_next(pos);
1425                         goto found;
1426                 }
1427
1428         bch2_extent_crc_append(e, p->crc);
1429         pos = extent_entry_last(extent_i_to_s(e));
1430 found:
1431         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
1432         __extent_entry_insert(e, pos, to_entry(&p->ptr));
1433
1434         for (i = 0; i < p->ec_nr; i++) {
1435                 p->ec[i].type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
1436                 __extent_entry_insert(e, pos, to_entry(&p->ec[i]));
1437         }
1438 }
1439
1440 /*
1441  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
1442  *
1443  * Returns true if @k should be dropped entirely
1444  *
1445  * For existing keys, only called when btree nodes are being rewritten, not when
1446  * they're merely being compacted/resorted in memory.
1447  */
1448 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1449 {
1450         struct bch_extent_ptr *ptr;
1451
1452         bch2_bkey_drop_ptrs(k, ptr,
1453                 ptr->cached &&
1454                 ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
1455
1456         /* will only happen if all pointers were cached: */
1457         if (!bkey_val_u64s(k.k))
1458                 k.k->type = KEY_TYPE_deleted;
1459
1460         return false;
1461 }
1462
1463 void bch2_extent_mark_replicas_cached(struct bch_fs *c,
1464                                       struct bkey_s_extent e,
1465                                       unsigned target,
1466                                       unsigned nr_desired_replicas)
1467 {
1468         union bch_extent_entry *entry;
1469         struct extent_ptr_decoded p;
1470         int extra = bch2_bkey_durability(c, e.s_c) - nr_desired_replicas;
1471
1472         if (target && extra > 0)
1473                 extent_for_each_ptr_decode(e, p, entry) {
1474                         int n = bch2_extent_ptr_durability(c, p);
1475
1476                         if (n && n <= extra &&
1477                             !bch2_dev_in_target(c, p.ptr.dev, target)) {
1478                                 entry->ptr.cached = true;
1479                                 extra -= n;
1480                         }
1481                 }
1482
1483         if (extra > 0)
1484                 extent_for_each_ptr_decode(e, p, entry) {
1485                         int n = bch2_extent_ptr_durability(c, p);
1486
1487                         if (n && n <= extra) {
1488                                 entry->ptr.cached = true;
1489                                 extra -= n;
1490                         }
1491                 }
1492 }
1493
1494 enum merge_result bch2_extent_merge(struct bch_fs *c,
1495                                     struct bkey_i *l, struct bkey_i *r)
1496 {
1497         struct bkey_s_extent el = bkey_i_to_s_extent(l);
1498         struct bkey_s_extent er = bkey_i_to_s_extent(r);
1499         union bch_extent_entry *en_l, *en_r;
1500
1501         if (bkey_val_u64s(&l->k) != bkey_val_u64s(&r->k))
1502                 return BCH_MERGE_NOMERGE;
1503
1504         extent_for_each_entry(el, en_l) {
1505                 struct bch_extent_ptr *lp, *rp;
1506                 struct bch_dev *ca;
1507
1508                 en_r = vstruct_idx(er.v, (u64 *) en_l - el.v->_data);
1509
1510                 if ((extent_entry_type(en_l) !=
1511                      extent_entry_type(en_r)) ||
1512                     !extent_entry_is_ptr(en_l))
1513                         return BCH_MERGE_NOMERGE;
1514
1515                 lp = &en_l->ptr;
1516                 rp = &en_r->ptr;
1517
1518                 if (lp->offset + el.k->size     != rp->offset ||
1519                     lp->dev                     != rp->dev ||
1520                     lp->gen                     != rp->gen)
1521                         return BCH_MERGE_NOMERGE;
1522
1523                 /* We don't allow extents to straddle buckets: */
1524                 ca = bch_dev_bkey_exists(c, lp->dev);
1525
1526                 if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
1527                         return BCH_MERGE_NOMERGE;
1528         }
1529
1530         l->k.needs_whiteout |= r->k.needs_whiteout;
1531
1532         /* Keys with no pointers aren't restricted to one bucket and could
1533          * overflow KEY_SIZE
1534          */
1535         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1536                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1537                 bch2_cut_front(l->k.p, r);
1538                 return BCH_MERGE_PARTIAL;
1539         }
1540
1541         bch2_key_resize(&l->k, l->k.size + r->k.size);
1542
1543         return BCH_MERGE_MERGE;
1544 }
1545
1546 /*
1547  * When merging an extent that we're inserting into a btree node, the new merged
1548  * extent could overlap with an existing 0 size extent - if we don't fix that,
1549  * it'll break the btree node iterator so this code finds those 0 size extents
1550  * and shifts them out of the way.
1551  *
1552  * Also unpacks and repacks.
1553  */
1554 static bool bch2_extent_merge_inline(struct bch_fs *c,
1555                                      struct btree_iter *iter,
1556                                      struct bkey_packed *l,
1557                                      struct bkey_packed *r,
1558                                      bool back_merge)
1559 {
1560         struct btree *b = iter->l[0].b;
1561         struct btree_node_iter *node_iter = &iter->l[0].iter;
1562         BKEY_PADDED(k) li, ri;
1563         struct bkey_packed *m   = back_merge ? l : r;
1564         struct bkey_i *mi       = back_merge ? &li.k : &ri.k;
1565         struct bset_tree *t     = bch2_bkey_to_bset(b, m);
1566         enum merge_result ret;
1567
1568         EBUG_ON(bkey_written(b, m));
1569
1570         /*
1571          * We need to save copies of both l and r, because we might get a
1572          * partial merge (which modifies both) and then fails to repack
1573          */
1574         bch2_bkey_unpack(b, &li.k, l);
1575         bch2_bkey_unpack(b, &ri.k, r);
1576
1577         ret = bch2_bkey_merge(c, &li.k, &ri.k);
1578         if (ret == BCH_MERGE_NOMERGE)
1579                 return false;
1580
1581         /*
1582          * check if we overlap with deleted extents - would break the sort
1583          * order:
1584          */
1585         if (back_merge) {
1586                 struct bkey_packed *n = bkey_next(m);
1587
1588                 if (n != btree_bkey_last(b, t) &&
1589                     bkey_cmp_left_packed(b, n, &li.k.k.p) <= 0 &&
1590                     bkey_deleted(n))
1591                         return false;
1592         } else if (ret == BCH_MERGE_MERGE) {
1593                 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1594
1595                 if (prev &&
1596                     bkey_cmp_left_packed_byval(b, prev,
1597                                 bkey_start_pos(&li.k.k)) > 0)
1598                         return false;
1599         }
1600
1601         if (ret == BCH_MERGE_PARTIAL) {
1602                 if (!extent_i_save(b, m, mi))
1603                         return false;
1604
1605                 if (!back_merge)
1606                         bkey_copy(packed_to_bkey(l), &li.k);
1607                 else
1608                         bkey_copy(packed_to_bkey(r), &ri.k);
1609         } else {
1610                 if (!extent_i_save(b, m, &li.k))
1611                         return false;
1612         }
1613
1614         bch2_bset_fix_invalidated_key(b, m);
1615         bch2_btree_node_iter_fix(iter, b, node_iter,
1616                                  m, m->u64s, m->u64s);
1617         verify_modified_extent(iter, m);
1618
1619         return ret == BCH_MERGE_MERGE;
1620 }
1621
1622 bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
1623                                unsigned nr_replicas)
1624 {
1625         struct btree_iter iter;
1626         struct bpos end = pos;
1627         struct bkey_s_c k;
1628         bool ret = true;
1629
1630         end.offset += size;
1631
1632         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, pos,
1633                              BTREE_ITER_SLOTS, k) {
1634                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
1635                         break;
1636
1637                 if (nr_replicas > bch2_bkey_nr_ptrs_allocated(k)) {
1638                         ret = false;
1639                         break;
1640                 }
1641         }
1642         bch2_btree_iter_unlock(&iter);
1643
1644         return ret;
1645 }
1646
1647 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
1648 {
1649         unsigned ret = 0;
1650
1651         switch (k.k->type) {
1652         case KEY_TYPE_extent: {
1653                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
1654                 const union bch_extent_entry *entry;
1655                 struct extent_ptr_decoded p;
1656
1657                 extent_for_each_ptr_decode(e, p, entry)
1658                         ret += !p.ptr.cached &&
1659                                 p.crc.compression_type == BCH_COMPRESSION_NONE;
1660                 break;
1661         }
1662         case KEY_TYPE_reservation:
1663                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
1664                 break;
1665         }
1666
1667         return ret;
1668 }
1669
1670 /* KEY_TYPE_reservation: */
1671
1672 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
1673 {
1674         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1675
1676         if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
1677                 return "incorrect value size";
1678
1679         if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
1680                 return "invalid nr_replicas";
1681
1682         return NULL;
1683 }
1684
1685 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
1686                               struct bkey_s_c k)
1687 {
1688         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
1689
1690         pr_buf(out, "generation %u replicas %u",
1691                le32_to_cpu(r.v->generation),
1692                r.v->nr_replicas);
1693 }
1694
1695 enum merge_result bch2_reservation_merge(struct bch_fs *c,
1696                                          struct bkey_i *l, struct bkey_i *r)
1697 {
1698         struct bkey_i_reservation *li = bkey_i_to_reservation(l);
1699         struct bkey_i_reservation *ri = bkey_i_to_reservation(r);
1700
1701         if (li->v.generation != ri->v.generation ||
1702             li->v.nr_replicas != ri->v.nr_replicas)
1703                 return BCH_MERGE_NOMERGE;
1704
1705         l->k.needs_whiteout |= r->k.needs_whiteout;
1706
1707         /* Keys with no pointers aren't restricted to one bucket and could
1708          * overflow KEY_SIZE
1709          */
1710         if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
1711                 bch2_key_resize(&l->k, KEY_SIZE_MAX);
1712                 bch2_cut_front(l->k.p, r);
1713                 return BCH_MERGE_PARTIAL;
1714         }
1715
1716         bch2_key_resize(&l->k, l->k.size + r->k.size);
1717
1718         return BCH_MERGE_MERGE;
1719 }