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_cache.h"
  12 #include "btree_gc.h"
  13 #include "btree_io.h"
  14 #include "btree_iter.h"
  15 #include "buckets.h"
  16 #include "checksum.h"
  17 #include "compress.h"
  18 #include "debug.h"
  19 #include "disk_groups.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "inode.h"
  23 #include "journal.h"
  24 #include "replicas.h"
  25 #include "super.h"
  26 #include "super-io.h"
  27 #include "trace.h"
  28 #include "util.h"
  29
  30 static unsigned bch2_crc_field_size_max[] = {
  31         [BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
  32         [BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
  33         [BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
  34 };
  35
  36 static void bch2_extent_crc_pack(union bch_extent_crc *,
  37                                  struct bch_extent_crc_unpacked,
  38                                  enum bch_extent_entry_type);
  39
  40 static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f,
  41                                                    unsigned dev)
  42 {
  43         struct bch_dev_io_failures *i;
  44
  45         for (i = f->devs; i < f->devs + f->nr; i++)
  46                 if (i->dev == dev)
  47                         return i;
  48
  49         return NULL;
  50 }
  51
  52 void bch2_mark_io_failure(struct bch_io_failures *failed,
  53                           struct extent_ptr_decoded *p)
  54 {
  55         struct bch_dev_io_failures *f = dev_io_failures(failed, p->ptr.dev);
  56
  57         if (!f) {
  58                 BUG_ON(failed->nr >= ARRAY_SIZE(failed->devs));
  59
  60                 f = &failed->devs[failed->nr++];
  61                 f->dev          = p->ptr.dev;
  62                 f->idx          = p->idx;
  63                 f->nr_failed    = 1;
  64                 f->nr_retries   = 0;
  65         } else if (p->idx != f->idx) {
  66                 f->idx          = p->idx;
  67                 f->nr_failed    = 1;
  68                 f->nr_retries   = 0;
  69         } else {
  70                 f->nr_failed++;
  71         }
  72 }
  73
  74 static inline u64 dev_latency(struct bch_fs *c, unsigned dev)
  75 {
  76         struct bch_dev *ca = bch2_dev_rcu(c, dev);
  77         return ca ? atomic64_read(&ca->cur_latency[READ]) : S64_MAX;
  78 }
  79
  80 /*
  81  * returns true if p1 is better than p2:
  82  */
  83 static inline bool ptr_better(struct bch_fs *c,
  84                               const struct extent_ptr_decoded p1,
  85                               const struct extent_ptr_decoded p2)
  86 {
  87         if (likely(!p1.idx && !p2.idx)) {
  88                 u64 l1 = dev_latency(c, p1.ptr.dev);
  89                 u64 l2 = dev_latency(c, p2.ptr.dev);
  90
  91                 /* Pick at random, biased in favor of the faster device: */
  92
  93                 return bch2_rand_range(l1 + l2) > l1;
  94         }
  95
  96         if (bch2_force_reconstruct_read)
  97                 return p1.idx > p2.idx;
  98
  99         return p1.idx < p2.idx;
 100 }
 101
 102 /*
 103  * This picks a non-stale pointer, preferably from a device other than @avoid.
 104  * Avoid can be NULL, meaning pick any. If there are no non-stale pointers to
 105  * other devices, it will still pick a pointer from avoid.
 106  */
 107 int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 108                                struct bch_io_failures *failed,
 109                                struct extent_ptr_decoded *pick)
 110 {
 111         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 112         const union bch_extent_entry *entry;
 113         struct extent_ptr_decoded p;
 114         struct bch_dev_io_failures *f;
 115         int ret = 0;
 116
 117         if (k.k->type == KEY_TYPE_error)
 118                 return -EIO;
 119
 120         rcu_read_lock();
 121         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 122                 /*
 123                  * Unwritten extent: no need to actually read, treat it as a
 124                  * hole and return 0s:
 125                  */
 126                 if (p.ptr.unwritten) {
 127                         ret = 0;
 128                         break;
 129                 }
 130
 131                 /*
 132                  * If there are any dirty pointers it's an error if we can't
 133                  * read:
 134                  */
 135                 if (!ret && !p.ptr.cached)
 136                         ret = -EIO;
 137
 138                 struct bch_dev *ca = bch2_dev_rcu(c, p.ptr.dev);
 139
 140                 if (p.ptr.cached && (!ca || dev_ptr_stale_rcu(ca, &p.ptr)))
 141                         continue;
 142
 143                 f = failed ? dev_io_failures(failed, p.ptr.dev) : NULL;
 144                 if (f)
 145                         p.idx = f->nr_failed < f->nr_retries
 146                                 ? f->idx
 147                                 : f->idx + 1;
 148
 149                 if (!p.idx && !ca)
 150                         p.idx++;
 151
 152                 if (!p.idx && p.has_ec && bch2_force_reconstruct_read)
 153                         p.idx++;
 154
 155                 if (!p.idx && !bch2_dev_is_readable(ca))
 156                         p.idx++;
 157
 158                 if (p.idx >= (unsigned) p.has_ec + 1)
 159                         continue;
 160
 161                 if (ret > 0 && !ptr_better(c, p, *pick))
 162                         continue;
 163
 164                 *pick = p;
 165                 ret = 1;
 166         }
 167         rcu_read_unlock();
 168
 169         return ret;
 170 }
 171
 172 /* KEY_TYPE_btree_ptr: */
 173
 174 int bch2_btree_ptr_invalid(struct bch_fs *c, struct bkey_s_c k,
 175                            enum bch_validate_flags flags,
 176                            struct printbuf *err)
 177 {
 178         int ret = 0;
 179
 180         bkey_fsck_err_on(bkey_val_u64s(k.k) > BCH_REPLICAS_MAX, c, err,
 181                          btree_ptr_val_too_big,
 182                          "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
 183
 184         ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
 185 fsck_err:
 186         return ret;
 187 }
 188
 189 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 190                             struct bkey_s_c k)
 191 {
 192         bch2_bkey_ptrs_to_text(out, c, k);
 193 }
 194
 195 int bch2_btree_ptr_v2_invalid(struct bch_fs *c, struct bkey_s_c k,
 196                               enum bch_validate_flags flags,
 197                               struct printbuf *err)
 198 {
 199         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 200         int ret = 0;
 201
 202         bkey_fsck_err_on(bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX,
 203                          c, err, btree_ptr_v2_val_too_big,
 204                          "value too big (%zu > %zu)",
 205                          bkey_val_u64s(k.k), BKEY_BTREE_PTR_VAL_U64s_MAX);
 206
 207         bkey_fsck_err_on(bpos_ge(bp.v->min_key, bp.k->p),
 208                          c, err, btree_ptr_v2_min_key_bad,
 209                          "min_key > key");
 210
 211         if (flags & BCH_VALIDATE_write)
 212                 bkey_fsck_err_on(!bp.v->sectors_written,
 213                                  c, err, btree_ptr_v2_written_0,
 214                                  "sectors_written == 0");
 215
 216         ret = bch2_bkey_ptrs_invalid(c, k, flags, err);
 217 fsck_err:
 218         return ret;
 219 }
 220
 221 void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
 222                                struct bkey_s_c k)
 223 {
 224         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 225
 226         prt_printf(out, "seq %llx written %u min_key %s",
 227                le64_to_cpu(bp.v->seq),
 228                le16_to_cpu(bp.v->sectors_written),
 229                BTREE_PTR_RANGE_UPDATED(bp.v) ? "R " : "");
 230
 231         bch2_bpos_to_text(out, bp.v->min_key);
 232         prt_printf(out, " ");
 233         bch2_bkey_ptrs_to_text(out, c, k);
 234 }
 235
 236 void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
 237                               unsigned big_endian, int write,
 238                               struct bkey_s k)
 239 {
 240         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(k);
 241
 242         compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
 243
 244         if (version < bcachefs_metadata_version_inode_btree_change &&
 245             btree_id_is_extents(btree_id) &&
 246             !bkey_eq(bp.v->min_key, POS_MIN))
 247                 bp.v->min_key = write
 248                         ? bpos_nosnap_predecessor(bp.v->min_key)
 249                         : bpos_nosnap_successor(bp.v->min_key);
 250 }
 251
 252 /* KEY_TYPE_extent: */
 253
 254 bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 255 {
 256         struct bkey_ptrs   l_ptrs = bch2_bkey_ptrs(l);
 257         struct bkey_ptrs_c r_ptrs = bch2_bkey_ptrs_c(r);
 258         union bch_extent_entry *en_l;
 259         const union bch_extent_entry *en_r;
 260         struct extent_ptr_decoded lp, rp;
 261         bool use_right_ptr;
 262
 263         en_l = l_ptrs.start;
 264         en_r = r_ptrs.start;
 265         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 266                 if (extent_entry_type(en_l) != extent_entry_type(en_r))
 267                         return false;
 268
 269                 en_l = extent_entry_next(en_l);
 270                 en_r = extent_entry_next(en_r);
 271         }
 272
 273         if (en_l < l_ptrs.end || en_r < r_ptrs.end)
 274                 return false;
 275
 276         en_l = l_ptrs.start;
 277         en_r = r_ptrs.start;
 278         lp.crc = bch2_extent_crc_unpack(l.k, NULL);
 279         rp.crc = bch2_extent_crc_unpack(r.k, NULL);
 280
 281         while (__bkey_ptr_next_decode(l.k, l_ptrs.end, lp, en_l) &&
 282                __bkey_ptr_next_decode(r.k, r_ptrs.end, rp, en_r)) {
 283                 if (lp.ptr.offset + lp.crc.offset + lp.crc.live_size !=
 284                     rp.ptr.offset + rp.crc.offset ||
 285                     lp.ptr.dev                  != rp.ptr.dev ||
 286                     lp.ptr.gen                  != rp.ptr.gen ||
 287                     lp.ptr.unwritten            != rp.ptr.unwritten ||
 288                     lp.has_ec                   != rp.has_ec)
 289                         return false;
 290
 291                 /* Extents may not straddle buckets: */
 292                 rcu_read_lock();
 293                 struct bch_dev *ca = bch2_dev_rcu(c, lp.ptr.dev);
 294                 bool same_bucket = ca && PTR_BUCKET_NR(ca, &lp.ptr) == PTR_BUCKET_NR(ca, &rp.ptr);
 295                 rcu_read_unlock();
 296
 297                 if (!same_bucket)
 298                         return false;
 299
 300                 if (lp.has_ec                   != rp.has_ec ||
 301                     (lp.has_ec &&
 302                      (lp.ec.block               != rp.ec.block ||
 303                       lp.ec.redundancy          != rp.ec.redundancy ||
 304                       lp.ec.idx                 != rp.ec.idx)))
 305                         return false;
 306
 307                 if (lp.crc.compression_type     != rp.crc.compression_type ||
 308                     lp.crc.nonce                != rp.crc.nonce)
 309                         return false;
 310
 311                 if (lp.crc.offset + lp.crc.live_size + rp.crc.live_size <=
 312                     lp.crc.uncompressed_size) {
 313                         /* can use left extent's crc entry */
 314                 } else if (lp.crc.live_size <= rp.crc.offset) {
 315                         /* can use right extent's crc entry */
 316                 } else {
 317                         /* check if checksums can be merged: */
 318                         if (lp.crc.csum_type            != rp.crc.csum_type ||
 319                             lp.crc.nonce                != rp.crc.nonce ||
 320                             crc_is_compressed(lp.crc) ||
 321                             !bch2_checksum_mergeable(lp.crc.csum_type))
 322                                 return false;
 323
 324                         if (lp.crc.offset + lp.crc.live_size != lp.crc.compressed_size ||
 325                             rp.crc.offset)
 326                                 return false;
 327
 328                         if (lp.crc.csum_type &&
 329                             lp.crc.uncompressed_size +
 330                             rp.crc.uncompressed_size > (c->opts.encoded_extent_max >> 9))
 331                                 return false;
 332                 }
 333
 334                 en_l = extent_entry_next(en_l);
 335                 en_r = extent_entry_next(en_r);
 336         }
 337
 338         en_l = l_ptrs.start;
 339         en_r = r_ptrs.start;
 340         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 341                 if (extent_entry_is_crc(en_l)) {
 342                         struct bch_extent_crc_unpacked crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 343                         struct bch_extent_crc_unpacked crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 344
 345                         if (crc_l.uncompressed_size + crc_r.uncompressed_size >
 346                             bch2_crc_field_size_max[extent_entry_type(en_l)])
 347                                 return false;
 348                 }
 349
 350                 en_l = extent_entry_next(en_l);
 351                 en_r = extent_entry_next(en_r);
 352         }
 353
 354         use_right_ptr = false;
 355         en_l = l_ptrs.start;
 356         en_r = r_ptrs.start;
 357         while (en_l < l_ptrs.end) {
 358                 if (extent_entry_type(en_l) == BCH_EXTENT_ENTRY_ptr &&
 359                     use_right_ptr)
 360                         en_l->ptr = en_r->ptr;
 361
 362                 if (extent_entry_is_crc(en_l)) {
 363                         struct bch_extent_crc_unpacked crc_l =
 364                                 bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 365                         struct bch_extent_crc_unpacked crc_r =
 366                                 bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 367
 368                         use_right_ptr = false;
 369
 370                         if (crc_l.offset + crc_l.live_size + crc_r.live_size <=
 371                             crc_l.uncompressed_size) {
 372                                 /* can use left extent's crc entry */
 373                         } else if (crc_l.live_size <= crc_r.offset) {
 374                                 /* can use right extent's crc entry */
 375                                 crc_r.offset -= crc_l.live_size;
 376                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_r,
 377                                                      extent_entry_type(en_l));
 378                                 use_right_ptr = true;
 379                         } else {
 380                                 crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
 381                                                                  crc_l.csum,
 382                                                                  crc_r.csum,
 383                                                                  crc_r.uncompressed_size << 9);
 384
 385                                 crc_l.uncompressed_size += crc_r.uncompressed_size;
 386                                 crc_l.compressed_size   += crc_r.compressed_size;
 387                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
 388                                                      extent_entry_type(en_l));
 389                         }
 390                 }
 391
 392                 en_l = extent_entry_next(en_l);
 393                 en_r = extent_entry_next(en_r);
 394         }
 395
 396         bch2_key_resize(l.k, l.k->size + r.k->size);
 397         return true;
 398 }
 399
 400 /* KEY_TYPE_reservation: */
 401
 402 int bch2_reservation_invalid(struct bch_fs *c, struct bkey_s_c k,
 403                              enum bch_validate_flags flags,
 404                              struct printbuf *err)
 405 {
 406         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 407         int ret = 0;
 408
 409         bkey_fsck_err_on(!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX, c, err,
 410                          reservation_key_nr_replicas_invalid,
 411                          "invalid nr_replicas (%u)", r.v->nr_replicas);
 412 fsck_err:
 413         return ret;
 414 }
 415
 416 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
 417                               struct bkey_s_c k)
 418 {
 419         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 420
 421         prt_printf(out, "generation %u replicas %u",
 422                le32_to_cpu(r.v->generation),
 423                r.v->nr_replicas);
 424 }
 425
 426 bool bch2_reservation_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
 427 {
 428         struct bkey_s_reservation l = bkey_s_to_reservation(_l);
 429         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(_r);
 430
 431         if (l.v->generation != r.v->generation ||
 432             l.v->nr_replicas != r.v->nr_replicas)
 433                 return false;
 434
 435         bch2_key_resize(l.k, l.k->size + r.k->size);
 436         return true;
 437 }
 438
 439 /* Extent checksum entries: */
 440
 441 /* returns true if not equal */
 442 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 443                                          struct bch_extent_crc_unpacked r)
 444 {
 445         return (l.csum_type             != r.csum_type ||
 446                 l.compression_type      != r.compression_type ||
 447                 l.compressed_size       != r.compressed_size ||
 448                 l.uncompressed_size     != r.uncompressed_size ||
 449                 l.offset                != r.offset ||
 450                 l.live_size             != r.live_size ||
 451                 l.nonce                 != r.nonce ||
 452                 bch2_crc_cmp(l.csum, r.csum));
 453 }
 454
 455 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 456                                   struct bch_extent_crc_unpacked n)
 457 {
 458         return !crc_is_compressed(u) &&
 459                 u.csum_type &&
 460                 u.uncompressed_size > u.live_size &&
 461                 bch2_csum_type_is_encryption(u.csum_type) ==
 462                 bch2_csum_type_is_encryption(n.csum_type);
 463 }
 464
 465 bool bch2_can_narrow_extent_crcs(struct bkey_s_c k,
 466                                  struct bch_extent_crc_unpacked n)
 467 {
 468         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 469         struct bch_extent_crc_unpacked crc;
 470         const union bch_extent_entry *i;
 471
 472         if (!n.csum_type)
 473                 return false;
 474
 475         bkey_for_each_crc(k.k, ptrs, crc, i)
 476                 if (can_narrow_crc(crc, n))
 477                         return true;
 478
 479         return false;
 480 }
 481
 482 /*
 483  * We're writing another replica for this extent, so while we've got the data in
 484  * memory we'll be computing a new checksum for the currently live data.
 485  *
 486  * If there are other replicas we aren't moving, and they are checksummed but
 487  * not compressed, we can modify them to point to only the data that is
 488  * currently live (so that readers won't have to bounce) while we've got the
 489  * checksum we need:
 490  */
 491 bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 492 {
 493         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 494         struct bch_extent_crc_unpacked u;
 495         struct extent_ptr_decoded p;
 496         union bch_extent_entry *i;
 497         bool ret = false;
 498
 499         /* Find a checksum entry that covers only live data: */
 500         if (!n.csum_type) {
 501                 bkey_for_each_crc(&k->k, ptrs, u, i)
 502                         if (!crc_is_compressed(u) &&
 503                             u.csum_type &&
 504                             u.live_size == u.uncompressed_size) {
 505                                 n = u;
 506                                 goto found;
 507                         }
 508                 return false;
 509         }
 510 found:
 511         BUG_ON(crc_is_compressed(n));
 512         BUG_ON(n.offset);
 513         BUG_ON(n.live_size != k->k.size);
 514
 515 restart_narrow_pointers:
 516         ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 517
 518         bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
 519                 if (can_narrow_crc(p.crc, n)) {
 520                         bch2_bkey_drop_ptr_noerror(bkey_i_to_s(k), &i->ptr);
 521                         p.ptr.offset += p.crc.offset;
 522                         p.crc = n;
 523                         bch2_extent_ptr_decoded_append(k, &p);
 524                         ret = true;
 525                         goto restart_narrow_pointers;
 526                 }
 527
 528         return ret;
 529 }
 530
 531 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
 532                                  struct bch_extent_crc_unpacked src,
 533                                  enum bch_extent_entry_type type)
 534 {
 535 #define set_common_fields(_dst, _src)                                   \
 536                 _dst.type               = 1 << type;                    \
 537                 _dst.csum_type          = _src.csum_type,               \
 538                 _dst.compression_type   = _src.compression_type,        \
 539                 _dst._compressed_size   = _src.compressed_size - 1,     \
 540                 _dst._uncompressed_size = _src.uncompressed_size - 1,   \
 541                 _dst.offset             = _src.offset
 542
 543         switch (type) {
 544         case BCH_EXTENT_ENTRY_crc32:
 545                 set_common_fields(dst->crc32, src);
 546                 dst->crc32.csum         = (u32 __force) *((__le32 *) &src.csum.lo);
 547                 break;
 548         case BCH_EXTENT_ENTRY_crc64:
 549                 set_common_fields(dst->crc64, src);
 550                 dst->crc64.nonce        = src.nonce;
 551                 dst->crc64.csum_lo      = (u64 __force) src.csum.lo;
 552                 dst->crc64.csum_hi      = (u64 __force) *((__le16 *) &src.csum.hi);
 553                 break;
 554         case BCH_EXTENT_ENTRY_crc128:
 555                 set_common_fields(dst->crc128, src);
 556                 dst->crc128.nonce       = src.nonce;
 557                 dst->crc128.csum        = src.csum;
 558                 break;
 559         default:
 560                 BUG();
 561         }
 562 #undef set_common_fields
 563 }
 564
 565 void bch2_extent_crc_append(struct bkey_i *k,
 566                             struct bch_extent_crc_unpacked new)
 567 {
 568         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 569         union bch_extent_crc *crc = (void *) ptrs.end;
 570         enum bch_extent_entry_type type;
 571
 572         if (bch_crc_bytes[new.csum_type]        <= 4 &&
 573             new.uncompressed_size               <= CRC32_SIZE_MAX &&
 574             new.nonce                           <= CRC32_NONCE_MAX)
 575                 type = BCH_EXTENT_ENTRY_crc32;
 576         else if (bch_crc_bytes[new.csum_type]   <= 10 &&
 577                    new.uncompressed_size        <= CRC64_SIZE_MAX &&
 578                    new.nonce                    <= CRC64_NONCE_MAX)
 579                 type = BCH_EXTENT_ENTRY_crc64;
 580         else if (bch_crc_bytes[new.csum_type]   <= 16 &&
 581                    new.uncompressed_size        <= CRC128_SIZE_MAX &&
 582                    new.nonce                    <= CRC128_NONCE_MAX)
 583                 type = BCH_EXTENT_ENTRY_crc128;
 584         else
 585                 BUG();
 586
 587         bch2_extent_crc_pack(crc, new, type);
 588
 589         k->k.u64s += extent_entry_u64s(ptrs.end);
 590
 591         EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
 592 }
 593
 594 /* Generic code for keys with pointers: */
 595
 596 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
 597 {
 598         return bch2_bkey_devs(k).nr;
 599 }
 600
 601 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
 602 {
 603         return k.k->type == KEY_TYPE_reservation
 604                 ? bkey_s_c_to_reservation(k).v->nr_replicas
 605                 : bch2_bkey_dirty_devs(k).nr;
 606 }
 607
 608 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
 609 {
 610         unsigned ret = 0;
 611
 612         if (k.k->type == KEY_TYPE_reservation) {
 613                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
 614         } else {
 615                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 616                 const union bch_extent_entry *entry;
 617                 struct extent_ptr_decoded p;
 618
 619                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 620                         ret += !p.ptr.cached && !crc_is_compressed(p.crc);
 621         }
 622
 623         return ret;
 624 }
 625
 626 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
 627 {
 628         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 629         const union bch_extent_entry *entry;
 630         struct extent_ptr_decoded p;
 631         unsigned ret = 0;
 632
 633         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 634                 if (!p.ptr.cached && crc_is_compressed(p.crc))
 635                         ret += p.crc.compressed_size;
 636
 637         return ret;
 638 }
 639
 640 bool bch2_bkey_is_incompressible(struct bkey_s_c k)
 641 {
 642         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 643         const union bch_extent_entry *entry;
 644         struct bch_extent_crc_unpacked crc;
 645
 646         bkey_for_each_crc(k.k, ptrs, crc, entry)
 647                 if (crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 648                         return true;
 649         return false;
 650 }
 651
 652 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 653 {
 654         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 655         const union bch_extent_entry *entry;
 656         struct extent_ptr_decoded p = { 0 };
 657         unsigned replicas = 0;
 658
 659         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 660                 if (p.ptr.cached)
 661                         continue;
 662
 663                 if (p.has_ec)
 664                         replicas += p.ec.redundancy;
 665
 666                 replicas++;
 667
 668         }
 669
 670         return replicas;
 671 }
 672
 673 static inline unsigned __extent_ptr_durability(struct bch_dev *ca, struct extent_ptr_decoded *p)
 674 {
 675         if (p->ptr.cached)
 676                 return 0;
 677
 678         return p->has_ec
 679                 ? p->ec.redundancy + 1
 680                 : ca->mi.durability;
 681 }
 682
 683 unsigned bch2_extent_ptr_desired_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 684 {
 685         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 686
 687         return ca ? __extent_ptr_durability(ca, p) : 0;
 688 }
 689
 690 unsigned bch2_extent_ptr_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 691 {
 692         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 693
 694         if (!ca || ca->mi.state == BCH_MEMBER_STATE_failed)
 695                 return 0;
 696
 697         return __extent_ptr_durability(ca, p);
 698 }
 699
 700 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
 701 {
 702         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 703         const union bch_extent_entry *entry;
 704         struct extent_ptr_decoded p;
 705         unsigned durability = 0;
 706
 707         rcu_read_lock();
 708         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 709                 durability += bch2_extent_ptr_durability(c, &p);
 710         rcu_read_unlock();
 711
 712         return durability;
 713 }
 714
 715 static unsigned bch2_bkey_durability_safe(struct bch_fs *c, struct bkey_s_c k)
 716 {
 717         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 718         const union bch_extent_entry *entry;
 719         struct extent_ptr_decoded p;
 720         unsigned durability = 0;
 721
 722         rcu_read_lock();
 723         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 724                 if (p.ptr.dev < c->sb.nr_devices && c->devs[p.ptr.dev])
 725                         durability += bch2_extent_ptr_durability(c, &p);
 726         rcu_read_unlock();
 727
 728         return durability;
 729 }
 730
 731 void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry)
 732 {
 733         union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
 734         union bch_extent_entry *next = extent_entry_next(entry);
 735
 736         memmove_u64s(entry, next, (u64 *) end - (u64 *) next);
 737         k->k.u64s -= extent_entry_u64s(entry);
 738 }
 739
 740 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
 741                                     struct extent_ptr_decoded *p)
 742 {
 743         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 744         struct bch_extent_crc_unpacked crc =
 745                 bch2_extent_crc_unpack(&k->k, NULL);
 746         union bch_extent_entry *pos;
 747
 748         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 749                 pos = ptrs.start;
 750                 goto found;
 751         }
 752
 753         bkey_for_each_crc(&k->k, ptrs, crc, pos)
 754                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 755                         pos = extent_entry_next(pos);
 756                         goto found;
 757                 }
 758
 759         bch2_extent_crc_append(k, p->crc);
 760         pos = bkey_val_end(bkey_i_to_s(k));
 761 found:
 762         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 763         __extent_entry_insert(k, pos, to_entry(&p->ptr));
 764
 765         if (p->has_ec) {
 766                 p->ec.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
 767                 __extent_entry_insert(k, pos, to_entry(&p->ec));
 768         }
 769 }
 770
 771 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 772                                           union bch_extent_entry *entry)
 773 {
 774         union bch_extent_entry *i = ptrs.start;
 775
 776         if (i == entry)
 777                 return NULL;
 778
 779         while (extent_entry_next(i) != entry)
 780                 i = extent_entry_next(i);
 781         return i;
 782 }
 783
 784 /*
 785  * Returns pointer to the next entry after the one being dropped:
 786  */
 787 union bch_extent_entry *bch2_bkey_drop_ptr_noerror(struct bkey_s k,
 788                                                    struct bch_extent_ptr *ptr)
 789 {
 790         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 791         union bch_extent_entry *entry = to_entry(ptr), *next;
 792         union bch_extent_entry *ret = entry;
 793         bool drop_crc = true;
 794
 795         EBUG_ON(ptr < &ptrs.start->ptr ||
 796                 ptr >= &ptrs.end->ptr);
 797         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 798
 799         for (next = extent_entry_next(entry);
 800              next != ptrs.end;
 801              next = extent_entry_next(next)) {
 802                 if (extent_entry_is_crc(next)) {
 803                         break;
 804                 } else if (extent_entry_is_ptr(next)) {
 805                         drop_crc = false;
 806                         break;
 807                 }
 808         }
 809
 810         extent_entry_drop(k, entry);
 811
 812         while ((entry = extent_entry_prev(ptrs, entry))) {
 813                 if (extent_entry_is_ptr(entry))
 814                         break;
 815
 816                 if ((extent_entry_is_crc(entry) && drop_crc) ||
 817                     extent_entry_is_stripe_ptr(entry)) {
 818                         ret = (void *) ret - extent_entry_bytes(entry);
 819                         extent_entry_drop(k, entry);
 820                 }
 821         }
 822
 823         return ret;
 824 }
 825
 826 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
 827                                            struct bch_extent_ptr *ptr)
 828 {
 829         bool have_dirty = bch2_bkey_dirty_devs(k.s_c).nr;
 830         union bch_extent_entry *ret =
 831                 bch2_bkey_drop_ptr_noerror(k, ptr);
 832
 833         /*
 834          * If we deleted all the dirty pointers and there's still cached
 835          * pointers, we could set the cached pointers to dirty if they're not
 836          * stale - but to do that correctly we'd need to grab an open_bucket
 837          * reference so that we don't race with bucket reuse:
 838          */
 839         if (have_dirty &&
 840             !bch2_bkey_dirty_devs(k.s_c).nr) {
 841                 k.k->type = KEY_TYPE_error;
 842                 set_bkey_val_u64s(k.k, 0);
 843                 ret = NULL;
 844         } else if (!bch2_bkey_nr_ptrs(k.s_c)) {
 845                 k.k->type = KEY_TYPE_deleted;
 846                 set_bkey_val_u64s(k.k, 0);
 847                 ret = NULL;
 848         }
 849
 850         return ret;
 851 }
 852
 853 void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
 854 {
 855         bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
 856 }
 857
 858 void bch2_bkey_drop_device_noerror(struct bkey_s k, unsigned dev)
 859 {
 860         struct bch_extent_ptr *ptr = bch2_bkey_has_device(k, dev);
 861
 862         if (ptr)
 863                 bch2_bkey_drop_ptr_noerror(k, ptr);
 864 }
 865
 866 const struct bch_extent_ptr *bch2_bkey_has_device_c(struct bkey_s_c k, unsigned dev)
 867 {
 868         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 869
 870         bkey_for_each_ptr(ptrs, ptr)
 871                 if (ptr->dev == dev)
 872                         return ptr;
 873
 874         return NULL;
 875 }
 876
 877 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
 878 {
 879         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 880         struct bch_dev *ca;
 881         bool ret = false;
 882
 883         rcu_read_lock();
 884         bkey_for_each_ptr(ptrs, ptr)
 885                 if (bch2_dev_in_target(c, ptr->dev, target) &&
 886                     (ca = bch2_dev_rcu(c, ptr->dev)) &&
 887                     (!ptr->cached ||
 888                      !dev_ptr_stale_rcu(ca, ptr))) {
 889                         ret = true;
 890                         break;
 891                 }
 892         rcu_read_unlock();
 893
 894         return ret;
 895 }
 896
 897 bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
 898                            struct bch_extent_ptr m, u64 offset)
 899 {
 900         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 901         const union bch_extent_entry *entry;
 902         struct extent_ptr_decoded p;
 903
 904         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 905                 if (p.ptr.dev   == m.dev &&
 906                     p.ptr.gen   == m.gen &&
 907                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
 908                     (s64) m.offset  - offset)
 909                         return true;
 910
 911         return false;
 912 }
 913
 914 /*
 915  * Returns true if two extents refer to the same data:
 916  */
 917 bool bch2_extents_match(struct bkey_s_c k1, struct bkey_s_c k2)
 918 {
 919         if (k1.k->type != k2.k->type)
 920                 return false;
 921
 922         if (bkey_extent_is_direct_data(k1.k)) {
 923                 struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(k1);
 924                 struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(k2);
 925                 const union bch_extent_entry *entry1, *entry2;
 926                 struct extent_ptr_decoded p1, p2;
 927
 928                 if (bkey_extent_is_unwritten(k1) != bkey_extent_is_unwritten(k2))
 929                         return false;
 930
 931                 bkey_for_each_ptr_decode(k1.k, ptrs1, p1, entry1)
 932                         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
 933                                 if (p1.ptr.dev          == p2.ptr.dev &&
 934                                     p1.ptr.gen          == p2.ptr.gen &&
 935                                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
 936                                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
 937                                         return true;
 938
 939                 return false;
 940         } else {
 941                 /* KEY_TYPE_deleted, etc. */
 942                 return true;
 943         }
 944 }
 945
 946 struct bch_extent_ptr *
 947 bch2_extent_has_ptr(struct bkey_s_c k1, struct extent_ptr_decoded p1, struct bkey_s k2)
 948 {
 949         struct bkey_ptrs ptrs2 = bch2_bkey_ptrs(k2);
 950         union bch_extent_entry *entry2;
 951         struct extent_ptr_decoded p2;
 952
 953         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
 954                 if (p1.ptr.dev          == p2.ptr.dev &&
 955                     p1.ptr.gen          == p2.ptr.gen &&
 956                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
 957                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
 958                         return &entry2->ptr;
 959
 960         return NULL;
 961 }
 962
 963 void bch2_extent_ptr_set_cached(struct bkey_s k, struct bch_extent_ptr *ptr)
 964 {
 965         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 966         union bch_extent_entry *entry;
 967         union bch_extent_entry *ec = NULL;
 968
 969         bkey_extent_entry_for_each(ptrs, entry) {
 970                 if (&entry->ptr == ptr) {
 971                         ptr->cached = true;
 972                         if (ec)
 973                                 extent_entry_drop(k, ec);
 974                         return;
 975                 }
 976
 977                 if (extent_entry_is_stripe_ptr(entry))
 978                         ec = entry;
 979                 else if (extent_entry_is_ptr(entry))
 980                         ec = NULL;
 981         }
 982
 983         BUG();
 984 }
 985
 986 /*
 987  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
 988  *
 989  * Returns true if @k should be dropped entirely
 990  *
 991  * For existing keys, only called when btree nodes are being rewritten, not when
 992  * they're merely being compacted/resorted in memory.
 993  */
 994 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
 995 {
 996         struct bch_dev *ca;
 997
 998         rcu_read_lock();
 999         bch2_bkey_drop_ptrs(k, ptr,
1000                 ptr->cached &&
1001                 (ca = bch2_dev_rcu(c, ptr->dev)) &&
1002                 dev_ptr_stale_rcu(ca, ptr) > 0);
1003         rcu_read_unlock();
1004
1005         return bkey_deleted(k.k);
1006 }
1007
1008 void bch2_extent_ptr_to_text(struct printbuf *out, struct bch_fs *c, const struct bch_extent_ptr *ptr)
1009 {
1010         out->atomic++;
1011         rcu_read_lock();
1012         struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
1013         if (!ca) {
1014                 prt_printf(out, "ptr: %u:%llu gen %u%s", ptr->dev,
1015                            (u64) ptr->offset, ptr->gen,
1016                            ptr->cached ? " cached" : "");
1017         } else {
1018                 u32 offset;
1019                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
1020
1021                 prt_printf(out, "ptr: %u:%llu:%u gen %u",
1022                            ptr->dev, b, offset, ptr->gen);
1023                 if (ptr->cached)
1024                         prt_str(out, " cached");
1025                 if (ptr->unwritten)
1026                         prt_str(out, " unwritten");
1027                 int stale = dev_ptr_stale_rcu(ca, ptr);
1028                 if (stale > 0)
1029                         prt_printf(out, " stale");
1030                 else if (stale)
1031                         prt_printf(out, " invalid");
1032         }
1033         rcu_read_unlock();
1034         --out->atomic;
1035 }
1036
1037 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1038                             struct bkey_s_c k)
1039 {
1040         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1041         const union bch_extent_entry *entry;
1042         bool first = true;
1043
1044         if (c)
1045                 prt_printf(out, "durability: %u ", bch2_bkey_durability_safe(c, k));
1046
1047         bkey_extent_entry_for_each(ptrs, entry) {
1048                 if (!first)
1049                         prt_printf(out, " ");
1050
1051                 switch (__extent_entry_type(entry)) {
1052                 case BCH_EXTENT_ENTRY_ptr:
1053                         bch2_extent_ptr_to_text(out, c, entry_to_ptr(entry));
1054                         break;
1055
1056                 case BCH_EXTENT_ENTRY_crc32:
1057                 case BCH_EXTENT_ENTRY_crc64:
1058                 case BCH_EXTENT_ENTRY_crc128: {
1059                         struct bch_extent_crc_unpacked crc =
1060                                 bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1061
1062                         prt_printf(out, "crc: c_size %u size %u offset %u nonce %u csum ",
1063                                crc.compressed_size,
1064                                crc.uncompressed_size,
1065                                crc.offset, crc.nonce);
1066                         bch2_prt_csum_type(out, crc.csum_type);
1067                         prt_str(out, " compress ");
1068                         bch2_prt_compression_type(out, crc.compression_type);
1069                         break;
1070                 }
1071                 case BCH_EXTENT_ENTRY_stripe_ptr: {
1072                         const struct bch_extent_stripe_ptr *ec = &entry->stripe_ptr;
1073
1074                         prt_printf(out, "ec: idx %llu block %u",
1075                                (u64) ec->idx, ec->block);
1076                         break;
1077                 }
1078                 case BCH_EXTENT_ENTRY_rebalance: {
1079                         const struct bch_extent_rebalance *r = &entry->rebalance;
1080
1081                         prt_str(out, "rebalance: target ");
1082                         if (c)
1083                                 bch2_target_to_text(out, c, r->target);
1084                         else
1085                                 prt_printf(out, "%u", r->target);
1086                         prt_str(out, " compression ");
1087                         bch2_compression_opt_to_text(out, r->compression);
1088                         break;
1089                 }
1090                 default:
1091                         prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
1092                         return;
1093                 }
1094
1095                 first = false;
1096         }
1097 }
1098
1099 static int extent_ptr_invalid(struct bch_fs *c,
1100                               struct bkey_s_c k,
1101                               enum bch_validate_flags flags,
1102                               const struct bch_extent_ptr *ptr,
1103                               unsigned size_ondisk,
1104                               bool metadata,
1105                               struct printbuf *err)
1106 {
1107         int ret = 0;
1108
1109         rcu_read_lock();
1110         struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
1111         if (!ca) {
1112                 rcu_read_unlock();
1113                 return 0;
1114         }
1115         u32 bucket_offset;
1116         u64 bucket = sector_to_bucket_and_offset(ca, ptr->offset, &bucket_offset);
1117         unsigned first_bucket   = ca->mi.first_bucket;
1118         u64 nbuckets            = ca->mi.nbuckets;
1119         unsigned bucket_size    = ca->mi.bucket_size;
1120         rcu_read_unlock();
1121
1122         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1123         bkey_for_each_ptr(ptrs, ptr2)
1124                 bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev, c, err,
1125                                  ptr_to_duplicate_device,
1126                                  "multiple pointers to same device (%u)", ptr->dev);
1127
1128
1129         bkey_fsck_err_on(bucket >= nbuckets, c, err,
1130                          ptr_after_last_bucket,
1131                          "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
1132         bkey_fsck_err_on(bucket < first_bucket, c, err,
1133                          ptr_before_first_bucket,
1134                          "pointer before first bucket (%llu < %u)", bucket, first_bucket);
1135         bkey_fsck_err_on(bucket_offset + size_ondisk > bucket_size, c, err,
1136                          ptr_spans_multiple_buckets,
1137                          "pointer spans multiple buckets (%u + %u > %u)",
1138                        bucket_offset, size_ondisk, bucket_size);
1139 fsck_err:
1140         return ret;
1141 }
1142
1143 int bch2_bkey_ptrs_invalid(struct bch_fs *c, struct bkey_s_c k,
1144                            enum bch_validate_flags flags,
1145                            struct printbuf *err)
1146 {
1147         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1148         const union bch_extent_entry *entry;
1149         struct bch_extent_crc_unpacked crc;
1150         unsigned size_ondisk = k.k->size;
1151         unsigned nonce = UINT_MAX;
1152         unsigned nr_ptrs = 0;
1153         bool have_written = false, have_unwritten = false, have_ec = false, crc_since_last_ptr = false;
1154         int ret = 0;
1155
1156         if (bkey_is_btree_ptr(k.k))
1157                 size_ondisk = btree_sectors(c);
1158
1159         bkey_extent_entry_for_each(ptrs, entry) {
1160                 bkey_fsck_err_on(__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX, c, err,
1161                         extent_ptrs_invalid_entry,
1162                         "invalid extent entry type (got %u, max %u)",
1163                         __extent_entry_type(entry), BCH_EXTENT_ENTRY_MAX);
1164
1165                 bkey_fsck_err_on(bkey_is_btree_ptr(k.k) &&
1166                                  !extent_entry_is_ptr(entry), c, err,
1167                                  btree_ptr_has_non_ptr,
1168                                  "has non ptr field");
1169
1170                 switch (extent_entry_type(entry)) {
1171                 case BCH_EXTENT_ENTRY_ptr:
1172                         ret = extent_ptr_invalid(c, k, flags, &entry->ptr,
1173                                                  size_ondisk, false, err);
1174                         if (ret)
1175                                 return ret;
1176
1177                         bkey_fsck_err_on(entry->ptr.cached && have_ec, c, err,
1178                                          ptr_cached_and_erasure_coded,
1179                                          "cached, erasure coded ptr");
1180
1181                         if (!entry->ptr.unwritten)
1182                                 have_written = true;
1183                         else
1184                                 have_unwritten = true;
1185
1186                         have_ec = false;
1187                         crc_since_last_ptr = false;
1188                         nr_ptrs++;
1189                         break;
1190                 case BCH_EXTENT_ENTRY_crc32:
1191                 case BCH_EXTENT_ENTRY_crc64:
1192                 case BCH_EXTENT_ENTRY_crc128:
1193                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1194
1195                         bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size, c, err,
1196                                          ptr_crc_uncompressed_size_too_small,
1197                                          "checksum offset + key size > uncompressed size");
1198                         bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type), c, err,
1199                                          ptr_crc_csum_type_unknown,
1200                                          "invalid checksum type");
1201                         bkey_fsck_err_on(crc.compression_type >= BCH_COMPRESSION_TYPE_NR, c, err,
1202                                          ptr_crc_compression_type_unknown,
1203                                          "invalid compression type");
1204
1205                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1206                                 if (nonce == UINT_MAX)
1207                                         nonce = crc.offset + crc.nonce;
1208                                 else if (nonce != crc.offset + crc.nonce)
1209                                         bkey_fsck_err(c, err, ptr_crc_nonce_mismatch,
1210                                                       "incorrect nonce");
1211                         }
1212
1213                         bkey_fsck_err_on(crc_since_last_ptr, c, err,
1214                                          ptr_crc_redundant,
1215                                          "redundant crc entry");
1216                         crc_since_last_ptr = true;
1217
1218                         bkey_fsck_err_on(crc_is_encoded(crc) &&
1219                                          (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
1220                                          (flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)), c, err,
1221                                          ptr_crc_uncompressed_size_too_big,
1222                                          "too large encoded extent");
1223
1224                         size_ondisk = crc.compressed_size;
1225                         break;
1226                 case BCH_EXTENT_ENTRY_stripe_ptr:
1227                         bkey_fsck_err_on(have_ec, c, err,
1228                                          ptr_stripe_redundant,
1229                                          "redundant stripe entry");
1230                         have_ec = true;
1231                         break;
1232                 case BCH_EXTENT_ENTRY_rebalance: {
1233                         const struct bch_extent_rebalance *r = &entry->rebalance;
1234
1235                         if (!bch2_compression_opt_valid(r->compression)) {
1236                                 struct bch_compression_opt opt = __bch2_compression_decode(r->compression);
1237                                 prt_printf(err, "invalid compression opt %u:%u",
1238                                            opt.type, opt.level);
1239                                 return -BCH_ERR_invalid_bkey;
1240                         }
1241                         break;
1242                 }
1243                 }
1244         }
1245
1246         bkey_fsck_err_on(!nr_ptrs, c, err,
1247                          extent_ptrs_no_ptrs,
1248                          "no ptrs");
1249         bkey_fsck_err_on(nr_ptrs > BCH_BKEY_PTRS_MAX, c, err,
1250                          extent_ptrs_too_many_ptrs,
1251                          "too many ptrs: %u > %u", nr_ptrs, BCH_BKEY_PTRS_MAX);
1252         bkey_fsck_err_on(have_written && have_unwritten, c, err,
1253                          extent_ptrs_written_and_unwritten,
1254                          "extent with unwritten and written ptrs");
1255         bkey_fsck_err_on(k.k->type != KEY_TYPE_extent && have_unwritten, c, err,
1256                          extent_ptrs_unwritten,
1257                          "has unwritten ptrs");
1258         bkey_fsck_err_on(crc_since_last_ptr, c, err,
1259                          extent_ptrs_redundant_crc,
1260                          "redundant crc entry");
1261         bkey_fsck_err_on(have_ec, c, err,
1262                          extent_ptrs_redundant_stripe,
1263                          "redundant stripe entry");
1264 fsck_err:
1265         return ret;
1266 }
1267
1268 void bch2_ptr_swab(struct bkey_s k)
1269 {
1270         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1271         union bch_extent_entry *entry;
1272         u64 *d;
1273
1274         for (d =  (u64 *) ptrs.start;
1275              d != (u64 *) ptrs.end;
1276              d++)
1277                 *d = swab64(*d);
1278
1279         for (entry = ptrs.start;
1280              entry < ptrs.end;
1281              entry = extent_entry_next(entry)) {
1282                 switch (extent_entry_type(entry)) {
1283                 case BCH_EXTENT_ENTRY_ptr:
1284                         break;
1285                 case BCH_EXTENT_ENTRY_crc32:
1286                         entry->crc32.csum = swab32(entry->crc32.csum);
1287                         break;
1288                 case BCH_EXTENT_ENTRY_crc64:
1289                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
1290                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
1291                         break;
1292                 case BCH_EXTENT_ENTRY_crc128:
1293                         entry->crc128.csum.hi = (__force __le64)
1294                                 swab64((__force u64) entry->crc128.csum.hi);
1295                         entry->crc128.csum.lo = (__force __le64)
1296                                 swab64((__force u64) entry->crc128.csum.lo);
1297                         break;
1298                 case BCH_EXTENT_ENTRY_stripe_ptr:
1299                         break;
1300                 case BCH_EXTENT_ENTRY_rebalance:
1301                         break;
1302                 }
1303         }
1304 }
1305
1306 const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
1307 {
1308         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1309         const union bch_extent_entry *entry;
1310
1311         bkey_extent_entry_for_each(ptrs, entry)
1312                 if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
1313                         return &entry->rebalance;
1314
1315         return NULL;
1316 }
1317
1318 unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
1319                                        unsigned target, unsigned compression)
1320 {
1321         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1322         unsigned rewrite_ptrs = 0;
1323
1324         if (compression) {
1325                 unsigned compression_type = bch2_compression_opt_to_type(compression);
1326                 const union bch_extent_entry *entry;
1327                 struct extent_ptr_decoded p;
1328                 unsigned i = 0;
1329
1330                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1331                         if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
1332                             p.ptr.unwritten) {
1333                                 rewrite_ptrs = 0;
1334                                 goto incompressible;
1335                         }
1336
1337                         if (!p.ptr.cached && p.crc.compression_type != compression_type)
1338                                 rewrite_ptrs |= 1U << i;
1339                         i++;
1340                 }
1341         }
1342 incompressible:
1343         if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
1344                 unsigned i = 0;
1345
1346                 bkey_for_each_ptr(ptrs, ptr) {
1347                         if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, target))
1348                                 rewrite_ptrs |= 1U << i;
1349                         i++;
1350                 }
1351         }
1352
1353         return rewrite_ptrs;
1354 }
1355
1356 bool bch2_bkey_needs_rebalance(struct bch_fs *c, struct bkey_s_c k)
1357 {
1358         const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
1359
1360         /*
1361          * If it's an indirect extent, we don't delete the rebalance entry when
1362          * done so that we know what options were applied - check if it still
1363          * needs work done:
1364          */
1365         if (r &&
1366             k.k->type == KEY_TYPE_reflink_v &&
1367             !bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression))
1368                 r = NULL;
1369
1370         return r != NULL;
1371 }
1372
1373 int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bkey_i *_k,
1374                                   struct bch_io_opts *opts)
1375 {
1376         struct bkey_s k = bkey_i_to_s(_k);
1377         struct bch_extent_rebalance *r;
1378         unsigned target = opts->background_target;
1379         unsigned compression = background_compression(*opts);
1380         bool needs_rebalance;
1381
1382         if (!bkey_extent_is_direct_data(k.k))
1383                 return 0;
1384
1385         /* get existing rebalance entry: */
1386         r = (struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
1387         if (r) {
1388                 if (k.k->type == KEY_TYPE_reflink_v) {
1389                         /*
1390                          * indirect extents: existing options take precedence,
1391                          * so that we don't move extents back and forth if
1392                          * they're referenced by different inodes with different
1393                          * options:
1394                          */
1395                         if (r->target)
1396                                 target = r->target;
1397                         if (r->compression)
1398                                 compression = r->compression;
1399                 }
1400
1401                 r->target       = target;
1402                 r->compression  = compression;
1403         }
1404
1405         needs_rebalance = bch2_bkey_ptrs_need_rebalance(c, k.s_c, target, compression);
1406
1407         if (needs_rebalance && !r) {
1408                 union bch_extent_entry *new = bkey_val_end(k);
1409
1410                 new->rebalance.type             = 1U << BCH_EXTENT_ENTRY_rebalance;
1411                 new->rebalance.compression      = compression;
1412                 new->rebalance.target           = target;
1413                 new->rebalance.unused           = 0;
1414                 k.k->u64s += extent_entry_u64s(new);
1415         } else if (!needs_rebalance && r && k.k->type != KEY_TYPE_reflink_v) {
1416                 /*
1417                  * For indirect extents, don't delete the rebalance entry when
1418                  * we're finished so that we know we specifically moved it or
1419                  * compressed it to its current location/compression type
1420                  */
1421                 extent_entry_drop(k, (union bch_extent_entry *) r);
1422         }
1423
1424         return 0;
1425 }
1426
1427 /* Generic extent code: */
1428
1429 int bch2_cut_front_s(struct bpos where, struct bkey_s k)
1430 {
1431         unsigned new_val_u64s = bkey_val_u64s(k.k);
1432         int val_u64s_delta;
1433         u64 sub;
1434
1435         if (bkey_le(where, bkey_start_pos(k.k)))
1436                 return 0;
1437
1438         EBUG_ON(bkey_gt(where, k.k->p));
1439
1440         sub = where.offset - bkey_start_offset(k.k);
1441
1442         k.k->size -= sub;
1443
1444         if (!k.k->size) {
1445                 k.k->type = KEY_TYPE_deleted;
1446                 new_val_u64s = 0;
1447         }
1448
1449         switch (k.k->type) {
1450         case KEY_TYPE_extent:
1451         case KEY_TYPE_reflink_v: {
1452                 struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1453                 union bch_extent_entry *entry;
1454                 bool seen_crc = false;
1455
1456                 bkey_extent_entry_for_each(ptrs, entry) {
1457                         switch (extent_entry_type(entry)) {
1458                         case BCH_EXTENT_ENTRY_ptr:
1459                                 if (!seen_crc)
1460                                         entry->ptr.offset += sub;
1461                                 break;
1462                         case BCH_EXTENT_ENTRY_crc32:
1463                                 entry->crc32.offset += sub;
1464                                 break;
1465                         case BCH_EXTENT_ENTRY_crc64:
1466                                 entry->crc64.offset += sub;
1467                                 break;
1468                         case BCH_EXTENT_ENTRY_crc128:
1469                                 entry->crc128.offset += sub;
1470                                 break;
1471                         case BCH_EXTENT_ENTRY_stripe_ptr:
1472                                 break;
1473                         case BCH_EXTENT_ENTRY_rebalance:
1474                                 break;
1475                         }
1476
1477                         if (extent_entry_is_crc(entry))
1478                                 seen_crc = true;
1479                 }
1480
1481                 break;
1482         }
1483         case KEY_TYPE_reflink_p: {
1484                 struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
1485
1486                 le64_add_cpu(&p.v->idx, sub);
1487                 break;
1488         }
1489         case KEY_TYPE_inline_data:
1490         case KEY_TYPE_indirect_inline_data: {
1491                 void *p = bkey_inline_data_p(k);
1492                 unsigned bytes = bkey_inline_data_bytes(k.k);
1493
1494                 sub = min_t(u64, sub << 9, bytes);
1495
1496                 memmove(p, p + sub, bytes - sub);
1497
1498                 new_val_u64s -= sub >> 3;
1499                 break;
1500         }
1501         }
1502
1503         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1504         BUG_ON(val_u64s_delta < 0);
1505
1506         set_bkey_val_u64s(k.k, new_val_u64s);
1507         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1508         return -val_u64s_delta;
1509 }
1510
1511 int bch2_cut_back_s(struct bpos where, struct bkey_s k)
1512 {
1513         unsigned new_val_u64s = bkey_val_u64s(k.k);
1514         int val_u64s_delta;
1515         u64 len = 0;
1516
1517         if (bkey_ge(where, k.k->p))
1518                 return 0;
1519
1520         EBUG_ON(bkey_lt(where, bkey_start_pos(k.k)));
1521
1522         len = where.offset - bkey_start_offset(k.k);
1523
1524         k.k->p.offset = where.offset;
1525         k.k->size = len;
1526
1527         if (!len) {
1528                 k.k->type = KEY_TYPE_deleted;
1529                 new_val_u64s = 0;
1530         }
1531
1532         switch (k.k->type) {
1533         case KEY_TYPE_inline_data:
1534         case KEY_TYPE_indirect_inline_data:
1535                 new_val_u64s = (bkey_inline_data_offset(k.k) +
1536                                 min(bkey_inline_data_bytes(k.k), k.k->size << 9)) >> 3;
1537                 break;
1538         }
1539
1540         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1541         BUG_ON(val_u64s_delta < 0);
1542
1543         set_bkey_val_u64s(k.k, new_val_u64s);
1544         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1545         return -val_u64s_delta;
1546 }