fs/bcachefs/extents.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <[email protected]>
   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 struct bch_dev_io_failures *bch2_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 = bch2_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 -BCH_ERR_key_type_error;
 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 = -BCH_ERR_no_device_to_read_from;
 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 ? bch2_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 || !bch2_dev_is_readable(ca)))
 150                         p.idx++;
 151
 152                 if (!p.idx && p.has_ec && bch2_force_reconstruct_read)
 153                         p.idx++;
 154
 155                 if (p.idx > (unsigned) p.has_ec)
 156                         continue;
 157
 158                 if (ret > 0 && !ptr_better(c, p, *pick))
 159                         continue;
 160
 161                 *pick = p;
 162                 ret = 1;
 163         }
 164         rcu_read_unlock();
 165
 166         return ret;
 167 }
 168
 169 /* KEY_TYPE_btree_ptr: */
 170
 171 int bch2_btree_ptr_validate(struct bch_fs *c, struct bkey_s_c k,
 172                             enum bch_validate_flags flags)
 173 {
 174         int ret = 0;
 175
 176         bkey_fsck_err_on(bkey_val_u64s(k.k) > BCH_REPLICAS_MAX,
 177                          c, btree_ptr_val_too_big,
 178                          "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
 179
 180         ret = bch2_bkey_ptrs_validate(c, k, flags);
 181 fsck_err:
 182         return ret;
 183 }
 184
 185 void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 186                             struct bkey_s_c k)
 187 {
 188         bch2_bkey_ptrs_to_text(out, c, k);
 189 }
 190
 191 int bch2_btree_ptr_v2_validate(struct bch_fs *c, struct bkey_s_c k,
 192                                enum bch_validate_flags flags)
 193 {
 194         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 195         int ret = 0;
 196
 197         bkey_fsck_err_on(bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX,
 198                          c, btree_ptr_v2_val_too_big,
 199                          "value too big (%zu > %zu)",
 200                          bkey_val_u64s(k.k), BKEY_BTREE_PTR_VAL_U64s_MAX);
 201
 202         bkey_fsck_err_on(bpos_ge(bp.v->min_key, bp.k->p),
 203                          c, btree_ptr_v2_min_key_bad,
 204                          "min_key > key");
 205
 206         if (flags & BCH_VALIDATE_write)
 207                 bkey_fsck_err_on(!bp.v->sectors_written,
 208                                  c, btree_ptr_v2_written_0,
 209                                  "sectors_written == 0");
 210
 211         ret = bch2_bkey_ptrs_validate(c, k, flags);
 212 fsck_err:
 213         return ret;
 214 }
 215
 216 void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
 217                                struct bkey_s_c k)
 218 {
 219         struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 220
 221         prt_printf(out, "seq %llx written %u min_key %s",
 222                le64_to_cpu(bp.v->seq),
 223                le16_to_cpu(bp.v->sectors_written),
 224                BTREE_PTR_RANGE_UPDATED(bp.v) ? "R " : "");
 225
 226         bch2_bpos_to_text(out, bp.v->min_key);
 227         prt_printf(out, " ");
 228         bch2_bkey_ptrs_to_text(out, c, k);
 229 }
 230
 231 void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
 232                               unsigned big_endian, int write,
 233                               struct bkey_s k)
 234 {
 235         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(k);
 236
 237         compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
 238
 239         if (version < bcachefs_metadata_version_inode_btree_change &&
 240             btree_id_is_extents(btree_id) &&
 241             !bkey_eq(bp.v->min_key, POS_MIN))
 242                 bp.v->min_key = write
 243                         ? bpos_nosnap_predecessor(bp.v->min_key)
 244                         : bpos_nosnap_successor(bp.v->min_key);
 245 }
 246
 247 /* KEY_TYPE_extent: */
 248
 249 bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 250 {
 251         struct bkey_ptrs   l_ptrs = bch2_bkey_ptrs(l);
 252         struct bkey_ptrs_c r_ptrs = bch2_bkey_ptrs_c(r);
 253         union bch_extent_entry *en_l;
 254         const union bch_extent_entry *en_r;
 255         struct extent_ptr_decoded lp, rp;
 256         bool use_right_ptr;
 257
 258         en_l = l_ptrs.start;
 259         en_r = r_ptrs.start;
 260         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 261                 if (extent_entry_type(en_l) != extent_entry_type(en_r))
 262                         return false;
 263
 264                 en_l = extent_entry_next(en_l);
 265                 en_r = extent_entry_next(en_r);
 266         }
 267
 268         if (en_l < l_ptrs.end || en_r < r_ptrs.end)
 269                 return false;
 270
 271         en_l = l_ptrs.start;
 272         en_r = r_ptrs.start;
 273         lp.crc = bch2_extent_crc_unpack(l.k, NULL);
 274         rp.crc = bch2_extent_crc_unpack(r.k, NULL);
 275
 276         while (__bkey_ptr_next_decode(l.k, l_ptrs.end, lp, en_l) &&
 277                __bkey_ptr_next_decode(r.k, r_ptrs.end, rp, en_r)) {
 278                 if (lp.ptr.offset + lp.crc.offset + lp.crc.live_size !=
 279                     rp.ptr.offset + rp.crc.offset ||
 280                     lp.ptr.dev                  != rp.ptr.dev ||
 281                     lp.ptr.gen                  != rp.ptr.gen ||
 282                     lp.ptr.unwritten            != rp.ptr.unwritten ||
 283                     lp.has_ec                   != rp.has_ec)
 284                         return false;
 285
 286                 /* Extents may not straddle buckets: */
 287                 rcu_read_lock();
 288                 struct bch_dev *ca = bch2_dev_rcu(c, lp.ptr.dev);
 289                 bool same_bucket = ca && PTR_BUCKET_NR(ca, &lp.ptr) == PTR_BUCKET_NR(ca, &rp.ptr);
 290                 rcu_read_unlock();
 291
 292                 if (!same_bucket)
 293                         return false;
 294
 295                 if (lp.has_ec                   != rp.has_ec ||
 296                     (lp.has_ec &&
 297                      (lp.ec.block               != rp.ec.block ||
 298                       lp.ec.redundancy          != rp.ec.redundancy ||
 299                       lp.ec.idx                 != rp.ec.idx)))
 300                         return false;
 301
 302                 if (lp.crc.compression_type     != rp.crc.compression_type ||
 303                     lp.crc.nonce                != rp.crc.nonce)
 304                         return false;
 305
 306                 if (lp.crc.offset + lp.crc.live_size + rp.crc.live_size <=
 307                     lp.crc.uncompressed_size) {
 308                         /* can use left extent's crc entry */
 309                 } else if (lp.crc.live_size <= rp.crc.offset) {
 310                         /* can use right extent's crc entry */
 311                 } else {
 312                         /* check if checksums can be merged: */
 313                         if (lp.crc.csum_type            != rp.crc.csum_type ||
 314                             lp.crc.nonce                != rp.crc.nonce ||
 315                             crc_is_compressed(lp.crc) ||
 316                             !bch2_checksum_mergeable(lp.crc.csum_type))
 317                                 return false;
 318
 319                         if (lp.crc.offset + lp.crc.live_size != lp.crc.compressed_size ||
 320                             rp.crc.offset)
 321                                 return false;
 322
 323                         if (lp.crc.csum_type &&
 324                             lp.crc.uncompressed_size +
 325                             rp.crc.uncompressed_size > (c->opts.encoded_extent_max >> 9))
 326                                 return false;
 327                 }
 328
 329                 en_l = extent_entry_next(en_l);
 330                 en_r = extent_entry_next(en_r);
 331         }
 332
 333         en_l = l_ptrs.start;
 334         en_r = r_ptrs.start;
 335         while (en_l < l_ptrs.end && en_r < r_ptrs.end) {
 336                 if (extent_entry_is_crc(en_l)) {
 337                         struct bch_extent_crc_unpacked crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 338                         struct bch_extent_crc_unpacked crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 339
 340                         if (crc_l.uncompressed_size + crc_r.uncompressed_size >
 341                             bch2_crc_field_size_max[extent_entry_type(en_l)])
 342                                 return false;
 343                 }
 344
 345                 en_l = extent_entry_next(en_l);
 346                 en_r = extent_entry_next(en_r);
 347         }
 348
 349         use_right_ptr = false;
 350         en_l = l_ptrs.start;
 351         en_r = r_ptrs.start;
 352         while (en_l < l_ptrs.end) {
 353                 if (extent_entry_type(en_l) == BCH_EXTENT_ENTRY_ptr &&
 354                     use_right_ptr)
 355                         en_l->ptr = en_r->ptr;
 356
 357                 if (extent_entry_is_crc(en_l)) {
 358                         struct bch_extent_crc_unpacked crc_l =
 359                                 bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
 360                         struct bch_extent_crc_unpacked crc_r =
 361                                 bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
 362
 363                         use_right_ptr = false;
 364
 365                         if (crc_l.offset + crc_l.live_size + crc_r.live_size <=
 366                             crc_l.uncompressed_size) {
 367                                 /* can use left extent's crc entry */
 368                         } else if (crc_l.live_size <= crc_r.offset) {
 369                                 /* can use right extent's crc entry */
 370                                 crc_r.offset -= crc_l.live_size;
 371                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_r,
 372                                                      extent_entry_type(en_l));
 373                                 use_right_ptr = true;
 374                         } else {
 375                                 crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
 376                                                                  crc_l.csum,
 377                                                                  crc_r.csum,
 378                                                                  crc_r.uncompressed_size << 9);
 379
 380                                 crc_l.uncompressed_size += crc_r.uncompressed_size;
 381                                 crc_l.compressed_size   += crc_r.compressed_size;
 382                                 bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
 383                                                      extent_entry_type(en_l));
 384                         }
 385                 }
 386
 387                 en_l = extent_entry_next(en_l);
 388                 en_r = extent_entry_next(en_r);
 389         }
 390
 391         bch2_key_resize(l.k, l.k->size + r.k->size);
 392         return true;
 393 }
 394
 395 /* KEY_TYPE_reservation: */
 396
 397 int bch2_reservation_validate(struct bch_fs *c, struct bkey_s_c k,
 398                               enum bch_validate_flags flags)
 399 {
 400         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 401         int ret = 0;
 402
 403         bkey_fsck_err_on(!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX,
 404                          c, reservation_key_nr_replicas_invalid,
 405                          "invalid nr_replicas (%u)", r.v->nr_replicas);
 406 fsck_err:
 407         return ret;
 408 }
 409
 410 void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
 411                               struct bkey_s_c k)
 412 {
 413         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 414
 415         prt_printf(out, "generation %u replicas %u",
 416                le32_to_cpu(r.v->generation),
 417                r.v->nr_replicas);
 418 }
 419
 420 bool bch2_reservation_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
 421 {
 422         struct bkey_s_reservation l = bkey_s_to_reservation(_l);
 423         struct bkey_s_c_reservation r = bkey_s_c_to_reservation(_r);
 424
 425         if (l.v->generation != r.v->generation ||
 426             l.v->nr_replicas != r.v->nr_replicas)
 427                 return false;
 428
 429         bch2_key_resize(l.k, l.k->size + r.k->size);
 430         return true;
 431 }
 432
 433 /* Extent checksum entries: */
 434
 435 /* returns true if not equal */
 436 static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 437                                          struct bch_extent_crc_unpacked r)
 438 {
 439         return (l.csum_type             != r.csum_type ||
 440                 l.compression_type      != r.compression_type ||
 441                 l.compressed_size       != r.compressed_size ||
 442                 l.uncompressed_size     != r.uncompressed_size ||
 443                 l.offset                != r.offset ||
 444                 l.live_size             != r.live_size ||
 445                 l.nonce                 != r.nonce ||
 446                 bch2_crc_cmp(l.csum, r.csum));
 447 }
 448
 449 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 450                                   struct bch_extent_crc_unpacked n)
 451 {
 452         return !crc_is_compressed(u) &&
 453                 u.csum_type &&
 454                 u.uncompressed_size > u.live_size &&
 455                 bch2_csum_type_is_encryption(u.csum_type) ==
 456                 bch2_csum_type_is_encryption(n.csum_type);
 457 }
 458
 459 bool bch2_can_narrow_extent_crcs(struct bkey_s_c k,
 460                                  struct bch_extent_crc_unpacked n)
 461 {
 462         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 463         struct bch_extent_crc_unpacked crc;
 464         const union bch_extent_entry *i;
 465
 466         if (!n.csum_type)
 467                 return false;
 468
 469         bkey_for_each_crc(k.k, ptrs, crc, i)
 470                 if (can_narrow_crc(crc, n))
 471                         return true;
 472
 473         return false;
 474 }
 475
 476 /*
 477  * We're writing another replica for this extent, so while we've got the data in
 478  * memory we'll be computing a new checksum for the currently live data.
 479  *
 480  * If there are other replicas we aren't moving, and they are checksummed but
 481  * not compressed, we can modify them to point to only the data that is
 482  * currently live (so that readers won't have to bounce) while we've got the
 483  * checksum we need:
 484  */
 485 bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 486 {
 487         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 488         struct bch_extent_crc_unpacked u;
 489         struct extent_ptr_decoded p;
 490         union bch_extent_entry *i;
 491         bool ret = false;
 492
 493         /* Find a checksum entry that covers only live data: */
 494         if (!n.csum_type) {
 495                 bkey_for_each_crc(&k->k, ptrs, u, i)
 496                         if (!crc_is_compressed(u) &&
 497                             u.csum_type &&
 498                             u.live_size == u.uncompressed_size) {
 499                                 n = u;
 500                                 goto found;
 501                         }
 502                 return false;
 503         }
 504 found:
 505         BUG_ON(crc_is_compressed(n));
 506         BUG_ON(n.offset);
 507         BUG_ON(n.live_size != k->k.size);
 508
 509 restart_narrow_pointers:
 510         ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 511
 512         bkey_for_each_ptr_decode(&k->k, ptrs, p, i)
 513                 if (can_narrow_crc(p.crc, n)) {
 514                         bch2_bkey_drop_ptr_noerror(bkey_i_to_s(k), &i->ptr);
 515                         p.ptr.offset += p.crc.offset;
 516                         p.crc = n;
 517                         bch2_extent_ptr_decoded_append(k, &p);
 518                         ret = true;
 519                         goto restart_narrow_pointers;
 520                 }
 521
 522         return ret;
 523 }
 524
 525 static void bch2_extent_crc_pack(union bch_extent_crc *dst,
 526                                  struct bch_extent_crc_unpacked src,
 527                                  enum bch_extent_entry_type type)
 528 {
 529 #define set_common_fields(_dst, _src)                                   \
 530                 _dst.type               = 1 << type;                    \
 531                 _dst.csum_type          = _src.csum_type,               \
 532                 _dst.compression_type   = _src.compression_type,        \
 533                 _dst._compressed_size   = _src.compressed_size - 1,     \
 534                 _dst._uncompressed_size = _src.uncompressed_size - 1,   \
 535                 _dst.offset             = _src.offset
 536
 537         switch (type) {
 538         case BCH_EXTENT_ENTRY_crc32:
 539                 set_common_fields(dst->crc32, src);
 540                 dst->crc32.csum         = (u32 __force) *((__le32 *) &src.csum.lo);
 541                 break;
 542         case BCH_EXTENT_ENTRY_crc64:
 543                 set_common_fields(dst->crc64, src);
 544                 dst->crc64.nonce        = src.nonce;
 545                 dst->crc64.csum_lo      = (u64 __force) src.csum.lo;
 546                 dst->crc64.csum_hi      = (u64 __force) *((__le16 *) &src.csum.hi);
 547                 break;
 548         case BCH_EXTENT_ENTRY_crc128:
 549                 set_common_fields(dst->crc128, src);
 550                 dst->crc128.nonce       = src.nonce;
 551                 dst->crc128.csum        = src.csum;
 552                 break;
 553         default:
 554                 BUG();
 555         }
 556 #undef set_common_fields
 557 }
 558
 559 void bch2_extent_crc_append(struct bkey_i *k,
 560                             struct bch_extent_crc_unpacked new)
 561 {
 562         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 563         union bch_extent_crc *crc = (void *) ptrs.end;
 564         enum bch_extent_entry_type type;
 565
 566         if (bch_crc_bytes[new.csum_type]        <= 4 &&
 567             new.uncompressed_size               <= CRC32_SIZE_MAX &&
 568             new.nonce                           <= CRC32_NONCE_MAX)
 569                 type = BCH_EXTENT_ENTRY_crc32;
 570         else if (bch_crc_bytes[new.csum_type]   <= 10 &&
 571                    new.uncompressed_size        <= CRC64_SIZE_MAX &&
 572                    new.nonce                    <= CRC64_NONCE_MAX)
 573                 type = BCH_EXTENT_ENTRY_crc64;
 574         else if (bch_crc_bytes[new.csum_type]   <= 16 &&
 575                    new.uncompressed_size        <= CRC128_SIZE_MAX &&
 576                    new.nonce                    <= CRC128_NONCE_MAX)
 577                 type = BCH_EXTENT_ENTRY_crc128;
 578         else
 579                 BUG();
 580
 581         bch2_extent_crc_pack(crc, new, type);
 582
 583         k->k.u64s += extent_entry_u64s(ptrs.end);
 584
 585         EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
 586 }
 587
 588 /* Generic code for keys with pointers: */
 589
 590 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
 591 {
 592         return bch2_bkey_devs(k).nr;
 593 }
 594
 595 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
 596 {
 597         return k.k->type == KEY_TYPE_reservation
 598                 ? bkey_s_c_to_reservation(k).v->nr_replicas
 599                 : bch2_bkey_dirty_devs(k).nr;
 600 }
 601
 602 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
 603 {
 604         unsigned ret = 0;
 605
 606         if (k.k->type == KEY_TYPE_reservation) {
 607                 ret = bkey_s_c_to_reservation(k).v->nr_replicas;
 608         } else {
 609                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 610                 const union bch_extent_entry *entry;
 611                 struct extent_ptr_decoded p;
 612
 613                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 614                         ret += !p.ptr.cached && !crc_is_compressed(p.crc);
 615         }
 616
 617         return ret;
 618 }
 619
 620 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
 621 {
 622         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 623         const union bch_extent_entry *entry;
 624         struct extent_ptr_decoded p;
 625         unsigned ret = 0;
 626
 627         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 628                 if (!p.ptr.cached && crc_is_compressed(p.crc))
 629                         ret += p.crc.compressed_size;
 630
 631         return ret;
 632 }
 633
 634 bool bch2_bkey_is_incompressible(struct bkey_s_c k)
 635 {
 636         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 637         const union bch_extent_entry *entry;
 638         struct bch_extent_crc_unpacked crc;
 639
 640         bkey_for_each_crc(k.k, ptrs, crc, entry)
 641                 if (crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 642                         return true;
 643         return false;
 644 }
 645
 646 unsigned bch2_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 647 {
 648         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 649         const union bch_extent_entry *entry;
 650         struct extent_ptr_decoded p = { 0 };
 651         unsigned replicas = 0;
 652
 653         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 654                 if (p.ptr.cached)
 655                         continue;
 656
 657                 if (p.has_ec)
 658                         replicas += p.ec.redundancy;
 659
 660                 replicas++;
 661
 662         }
 663
 664         return replicas;
 665 }
 666
 667 static inline unsigned __extent_ptr_durability(struct bch_dev *ca, struct extent_ptr_decoded *p)
 668 {
 669         if (p->ptr.cached)
 670                 return 0;
 671
 672         return p->has_ec
 673                 ? p->ec.redundancy + 1
 674                 : ca->mi.durability;
 675 }
 676
 677 unsigned bch2_extent_ptr_desired_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 678 {
 679         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 680
 681         return ca ? __extent_ptr_durability(ca, p) : 0;
 682 }
 683
 684 unsigned bch2_extent_ptr_durability(struct bch_fs *c, struct extent_ptr_decoded *p)
 685 {
 686         struct bch_dev *ca = bch2_dev_rcu(c, p->ptr.dev);
 687
 688         if (!ca || ca->mi.state == BCH_MEMBER_STATE_failed)
 689                 return 0;
 690
 691         return __extent_ptr_durability(ca, p);
 692 }
 693
 694 unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
 695 {
 696         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 697         const union bch_extent_entry *entry;
 698         struct extent_ptr_decoded p;
 699         unsigned durability = 0;
 700
 701         rcu_read_lock();
 702         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 703                 durability += bch2_extent_ptr_durability(c, &p);
 704         rcu_read_unlock();
 705
 706         return durability;
 707 }
 708
 709 static unsigned bch2_bkey_durability_safe(struct bch_fs *c, struct bkey_s_c k)
 710 {
 711         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 712         const union bch_extent_entry *entry;
 713         struct extent_ptr_decoded p;
 714         unsigned durability = 0;
 715
 716         rcu_read_lock();
 717         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 718                 if (p.ptr.dev < c->sb.nr_devices && c->devs[p.ptr.dev])
 719                         durability += bch2_extent_ptr_durability(c, &p);
 720         rcu_read_unlock();
 721
 722         return durability;
 723 }
 724
 725 void bch2_bkey_extent_entry_drop(struct bkey_i *k, union bch_extent_entry *entry)
 726 {
 727         union bch_extent_entry *end = bkey_val_end(bkey_i_to_s(k));
 728         union bch_extent_entry *next = extent_entry_next(entry);
 729
 730         memmove_u64s(entry, next, (u64 *) end - (u64 *) next);
 731         k->k.u64s -= extent_entry_u64s(entry);
 732 }
 733
 734 void bch2_extent_ptr_decoded_append(struct bkey_i *k,
 735                                     struct extent_ptr_decoded *p)
 736 {
 737         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
 738         struct bch_extent_crc_unpacked crc =
 739                 bch2_extent_crc_unpack(&k->k, NULL);
 740         union bch_extent_entry *pos;
 741
 742         if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 743                 pos = ptrs.start;
 744                 goto found;
 745         }
 746
 747         bkey_for_each_crc(&k->k, ptrs, crc, pos)
 748                 if (!bch2_crc_unpacked_cmp(crc, p->crc)) {
 749                         pos = extent_entry_next(pos);
 750                         goto found;
 751                 }
 752
 753         bch2_extent_crc_append(k, p->crc);
 754         pos = bkey_val_end(bkey_i_to_s(k));
 755 found:
 756         p->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 757         __extent_entry_insert(k, pos, to_entry(&p->ptr));
 758
 759         if (p->has_ec) {
 760                 p->ec.type = 1 << BCH_EXTENT_ENTRY_stripe_ptr;
 761                 __extent_entry_insert(k, pos, to_entry(&p->ec));
 762         }
 763 }
 764
 765 static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
 766                                           union bch_extent_entry *entry)
 767 {
 768         union bch_extent_entry *i = ptrs.start;
 769
 770         if (i == entry)
 771                 return NULL;
 772
 773         while (extent_entry_next(i) != entry)
 774                 i = extent_entry_next(i);
 775         return i;
 776 }
 777
 778 /*
 779  * Returns pointer to the next entry after the one being dropped:
 780  */
 781 void bch2_bkey_drop_ptr_noerror(struct bkey_s k, struct bch_extent_ptr *ptr)
 782 {
 783         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 784         union bch_extent_entry *entry = to_entry(ptr), *next;
 785         bool drop_crc = true;
 786
 787         if (k.k->type == KEY_TYPE_stripe) {
 788                 ptr->dev = BCH_SB_MEMBER_INVALID;
 789                 return;
 790         }
 791
 792         EBUG_ON(ptr < &ptrs.start->ptr ||
 793                 ptr >= &ptrs.end->ptr);
 794         EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
 795
 796         for (next = extent_entry_next(entry);
 797              next != ptrs.end;
 798              next = extent_entry_next(next)) {
 799                 if (extent_entry_is_crc(next)) {
 800                         break;
 801                 } else if (extent_entry_is_ptr(next)) {
 802                         drop_crc = false;
 803                         break;
 804                 }
 805         }
 806
 807         extent_entry_drop(k, entry);
 808
 809         while ((entry = extent_entry_prev(ptrs, entry))) {
 810                 if (extent_entry_is_ptr(entry))
 811                         break;
 812
 813                 if ((extent_entry_is_crc(entry) && drop_crc) ||
 814                     extent_entry_is_stripe_ptr(entry))
 815                         extent_entry_drop(k, entry);
 816         }
 817 }
 818
 819 void bch2_bkey_drop_ptr(struct bkey_s k, struct bch_extent_ptr *ptr)
 820 {
 821         if (k.k->type != KEY_TYPE_stripe) {
 822                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k.s_c);
 823                 const union bch_extent_entry *entry;
 824                 struct extent_ptr_decoded p;
 825
 826                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 827                         if (p.ptr.dev == ptr->dev && p.has_ec) {
 828                                 ptr->dev = BCH_SB_MEMBER_INVALID;
 829                                 return;
 830                         }
 831         }
 832
 833         bool have_dirty = bch2_bkey_dirty_devs(k.s_c).nr;
 834
 835         bch2_bkey_drop_ptr_noerror(k, ptr);
 836
 837         /*
 838          * If we deleted all the dirty pointers and there's still cached
 839          * pointers, we could set the cached pointers to dirty if they're not
 840          * stale - but to do that correctly we'd need to grab an open_bucket
 841          * reference so that we don't race with bucket reuse:
 842          */
 843         if (have_dirty &&
 844             !bch2_bkey_dirty_devs(k.s_c).nr) {
 845                 k.k->type = KEY_TYPE_error;
 846                 set_bkey_val_u64s(k.k, 0);
 847         } else if (!bch2_bkey_nr_ptrs(k.s_c)) {
 848                 k.k->type = KEY_TYPE_deleted;
 849                 set_bkey_val_u64s(k.k, 0);
 850         }
 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         bch2_bkey_drop_ptrs_noerror(k, ptr, ptr->dev == dev);
 861 }
 862
 863 const struct bch_extent_ptr *bch2_bkey_has_device_c(struct bkey_s_c k, unsigned dev)
 864 {
 865         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 866
 867         bkey_for_each_ptr(ptrs, ptr)
 868                 if (ptr->dev == dev)
 869                         return ptr;
 870
 871         return NULL;
 872 }
 873
 874 bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
 875 {
 876         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 877         struct bch_dev *ca;
 878         bool ret = false;
 879
 880         rcu_read_lock();
 881         bkey_for_each_ptr(ptrs, ptr)
 882                 if (bch2_dev_in_target(c, ptr->dev, target) &&
 883                     (ca = bch2_dev_rcu(c, ptr->dev)) &&
 884                     (!ptr->cached ||
 885                      !dev_ptr_stale_rcu(ca, ptr))) {
 886                         ret = true;
 887                         break;
 888                 }
 889         rcu_read_unlock();
 890
 891         return ret;
 892 }
 893
 894 bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
 895                            struct bch_extent_ptr m, u64 offset)
 896 {
 897         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 898         const union bch_extent_entry *entry;
 899         struct extent_ptr_decoded p;
 900
 901         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 902                 if (p.ptr.dev   == m.dev &&
 903                     p.ptr.gen   == m.gen &&
 904                     (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
 905                     (s64) m.offset  - offset)
 906                         return true;
 907
 908         return false;
 909 }
 910
 911 /*
 912  * Returns true if two extents refer to the same data:
 913  */
 914 bool bch2_extents_match(struct bkey_s_c k1, struct bkey_s_c k2)
 915 {
 916         if (k1.k->type != k2.k->type)
 917                 return false;
 918
 919         if (bkey_extent_is_direct_data(k1.k)) {
 920                 struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(k1);
 921                 struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(k2);
 922                 const union bch_extent_entry *entry1, *entry2;
 923                 struct extent_ptr_decoded p1, p2;
 924
 925                 if (bkey_extent_is_unwritten(k1) != bkey_extent_is_unwritten(k2))
 926                         return false;
 927
 928                 bkey_for_each_ptr_decode(k1.k, ptrs1, p1, entry1)
 929                         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
 930                                 if (p1.ptr.dev          == p2.ptr.dev &&
 931                                     p1.ptr.gen          == p2.ptr.gen &&
 932
 933                                     /*
 934                                      * This checks that the two pointers point
 935                                      * to the same region on disk - adjusting
 936                                      * for the difference in where the extents
 937                                      * start, since one may have been trimmed:
 938                                      */
 939                                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
 940                                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k) &&
 941
 942                                     /*
 943                                      * This additionally checks that the
 944                                      * extents overlap on disk, since the
 945                                      * previous check may trigger spuriously
 946                                      * when one extent is immediately partially
 947                                      * overwritten with another extent (so that
 948                                      * on disk they are adjacent) and
 949                                      * compression is in use:
 950                                      */
 951                                     ((p1.ptr.offset >= p2.ptr.offset &&
 952                                       p1.ptr.offset  < p2.ptr.offset + p2.crc.compressed_size) ||
 953                                      (p2.ptr.offset >= p1.ptr.offset &&
 954                                       p2.ptr.offset  < p1.ptr.offset + p1.crc.compressed_size)))
 955                                         return true;
 956
 957                 return false;
 958         } else {
 959                 /* KEY_TYPE_deleted, etc. */
 960                 return true;
 961         }
 962 }
 963
 964 struct bch_extent_ptr *
 965 bch2_extent_has_ptr(struct bkey_s_c k1, struct extent_ptr_decoded p1, struct bkey_s k2)
 966 {
 967         struct bkey_ptrs ptrs2 = bch2_bkey_ptrs(k2);
 968         union bch_extent_entry *entry2;
 969         struct extent_ptr_decoded p2;
 970
 971         bkey_for_each_ptr_decode(k2.k, ptrs2, p2, entry2)
 972                 if (p1.ptr.dev          == p2.ptr.dev &&
 973                     p1.ptr.gen          == p2.ptr.gen &&
 974                     (s64) p1.ptr.offset + p1.crc.offset - bkey_start_offset(k1.k) ==
 975                     (s64) p2.ptr.offset + p2.crc.offset - bkey_start_offset(k2.k))
 976                         return &entry2->ptr;
 977
 978         return NULL;
 979 }
 980
 981 static bool want_cached_ptr(struct bch_fs *c, struct bch_io_opts *opts,
 982                             struct bch_extent_ptr *ptr)
 983 {
 984         if (!opts->promote_target ||
 985             !bch2_dev_in_target(c, ptr->dev, opts->promote_target))
 986                 return false;
 987
 988         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
 989
 990         return ca && bch2_dev_is_readable(ca) && !dev_ptr_stale_rcu(ca, ptr);
 991 }
 992
 993 void bch2_extent_ptr_set_cached(struct bch_fs *c,
 994                                 struct bch_io_opts *opts,
 995                                 struct bkey_s k,
 996                                 struct bch_extent_ptr *ptr)
 997 {
 998         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 999         union bch_extent_entry *entry;
1000         struct extent_ptr_decoded p;
1001
1002         rcu_read_lock();
1003         if (!want_cached_ptr(c, opts, ptr)) {
1004                 bch2_bkey_drop_ptr_noerror(k, ptr);
1005                 goto out;
1006         }
1007
1008         /*
1009          * Stripes can't contain cached data, for - reasons.
1010          *
1011          * Possibly something we can fix in the future?
1012          */
1013         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
1014                 if (&entry->ptr == ptr) {
1015                         if (p.has_ec)
1016                                 bch2_bkey_drop_ptr_noerror(k, ptr);
1017                         else
1018                                 ptr->cached = true;
1019                         goto out;
1020                 }
1021
1022         BUG();
1023 out:
1024         rcu_read_unlock();
1025 }
1026
1027 /*
1028  * bch2_extent_normalize - clean up an extent, dropping stale pointers etc.
1029  *
1030  * Returns true if @k should be dropped entirely
1031  *
1032  * For existing keys, only called when btree nodes are being rewritten, not when
1033  * they're merely being compacted/resorted in memory.
1034  */
1035 bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
1036 {
1037         struct bch_dev *ca;
1038
1039         rcu_read_lock();
1040         bch2_bkey_drop_ptrs(k, ptr,
1041                 ptr->cached &&
1042                 (!(ca = bch2_dev_rcu(c, ptr->dev)) ||
1043                  dev_ptr_stale_rcu(ca, ptr) > 0));
1044         rcu_read_unlock();
1045
1046         return bkey_deleted(k.k);
1047 }
1048
1049 /*
1050  * bch2_extent_normalize_by_opts - clean up an extent, dropping stale pointers etc.
1051  *
1052  * Like bch2_extent_normalize(), but also only keeps a single cached pointer on
1053  * the promote target.
1054  */
1055 bool bch2_extent_normalize_by_opts(struct bch_fs *c,
1056                                    struct bch_io_opts *opts,
1057                                    struct bkey_s k)
1058 {
1059         struct bkey_ptrs ptrs;
1060         bool have_cached_ptr;
1061
1062         rcu_read_lock();
1063 restart_drop_ptrs:
1064         ptrs = bch2_bkey_ptrs(k);
1065         have_cached_ptr = false;
1066
1067         bkey_for_each_ptr(ptrs, ptr)
1068                 if (ptr->cached) {
1069                         if (have_cached_ptr || !want_cached_ptr(c, opts, ptr)) {
1070                                 bch2_bkey_drop_ptr(k, ptr);
1071                                 goto restart_drop_ptrs;
1072                         }
1073                         have_cached_ptr = true;
1074                 }
1075         rcu_read_unlock();
1076
1077         return bkey_deleted(k.k);
1078 }
1079
1080 void bch2_extent_ptr_to_text(struct printbuf *out, struct bch_fs *c, const struct bch_extent_ptr *ptr)
1081 {
1082         out->atomic++;
1083         rcu_read_lock();
1084         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
1085         if (!ca) {
1086                 prt_printf(out, "ptr: %u:%llu gen %u%s", ptr->dev,
1087                            (u64) ptr->offset, ptr->gen,
1088                            ptr->cached ? " cached" : "");
1089         } else {
1090                 u32 offset;
1091                 u64 b = sector_to_bucket_and_offset(ca, ptr->offset, &offset);
1092
1093                 prt_printf(out, "ptr: %u:%llu:%u gen %u",
1094                            ptr->dev, b, offset, ptr->gen);
1095                 if (ca->mi.durability != 1)
1096                         prt_printf(out, " d=%u", ca->mi.durability);
1097                 if (ptr->cached)
1098                         prt_str(out, " cached");
1099                 if (ptr->unwritten)
1100                         prt_str(out, " unwritten");
1101                 int stale = dev_ptr_stale_rcu(ca, ptr);
1102                 if (stale > 0)
1103                         prt_printf(out, " stale");
1104                 else if (stale)
1105                         prt_printf(out, " invalid");
1106         }
1107         rcu_read_unlock();
1108         --out->atomic;
1109 }
1110
1111 void bch2_extent_crc_unpacked_to_text(struct printbuf *out, struct bch_extent_crc_unpacked *crc)
1112 {
1113         prt_printf(out, "crc: c_size %u size %u offset %u nonce %u csum ",
1114                    crc->compressed_size,
1115                    crc->uncompressed_size,
1116                    crc->offset, crc->nonce);
1117         bch2_prt_csum_type(out, crc->csum_type);
1118         prt_printf(out, " %0llx:%0llx ", crc->csum.hi, crc->csum.lo);
1119         prt_str(out, " compress ");
1120         bch2_prt_compression_type(out, crc->compression_type);
1121 }
1122
1123 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1124                             struct bkey_s_c k)
1125 {
1126         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1127         const union bch_extent_entry *entry;
1128         bool first = true;
1129
1130         if (c)
1131                 prt_printf(out, "durability: %u ", bch2_bkey_durability_safe(c, k));
1132
1133         bkey_extent_entry_for_each(ptrs, entry) {
1134                 if (!first)
1135                         prt_printf(out, " ");
1136
1137                 switch (__extent_entry_type(entry)) {
1138                 case BCH_EXTENT_ENTRY_ptr:
1139                         bch2_extent_ptr_to_text(out, c, entry_to_ptr(entry));
1140                         break;
1141
1142                 case BCH_EXTENT_ENTRY_crc32:
1143                 case BCH_EXTENT_ENTRY_crc64:
1144                 case BCH_EXTENT_ENTRY_crc128: {
1145                         struct bch_extent_crc_unpacked crc =
1146                                 bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1147
1148                         bch2_extent_crc_unpacked_to_text(out, &crc);
1149                         break;
1150                 }
1151                 case BCH_EXTENT_ENTRY_stripe_ptr: {
1152                         const struct bch_extent_stripe_ptr *ec = &entry->stripe_ptr;
1153
1154                         prt_printf(out, "ec: idx %llu block %u",
1155                                (u64) ec->idx, ec->block);
1156                         break;
1157                 }
1158                 case BCH_EXTENT_ENTRY_rebalance: {
1159                         const struct bch_extent_rebalance *r = &entry->rebalance;
1160
1161                         prt_str(out, "rebalance: target ");
1162                         if (c)
1163                                 bch2_target_to_text(out, c, r->target);
1164                         else
1165                                 prt_printf(out, "%u", r->target);
1166                         prt_str(out, " compression ");
1167                         bch2_compression_opt_to_text(out, r->compression);
1168                         break;
1169                 }
1170                 default:
1171                         prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
1172                         return;
1173                 }
1174
1175                 first = false;
1176         }
1177 }
1178
1179 static int extent_ptr_validate(struct bch_fs *c,
1180                                struct bkey_s_c k,
1181                                enum bch_validate_flags flags,
1182                                const struct bch_extent_ptr *ptr,
1183                                unsigned size_ondisk,
1184                                bool metadata)
1185 {
1186         int ret = 0;
1187
1188         /* bad pointers are repaired by check_fix_ptrs(): */
1189         rcu_read_lock();
1190         struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
1191         if (!ca) {
1192                 rcu_read_unlock();
1193                 return 0;
1194         }
1195         u32 bucket_offset;
1196         u64 bucket = sector_to_bucket_and_offset(ca, ptr->offset, &bucket_offset);
1197         unsigned first_bucket   = ca->mi.first_bucket;
1198         u64 nbuckets            = ca->mi.nbuckets;
1199         unsigned bucket_size    = ca->mi.bucket_size;
1200         rcu_read_unlock();
1201
1202         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1203         bkey_for_each_ptr(ptrs, ptr2)
1204                 bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
1205                                  c, ptr_to_duplicate_device,
1206                                  "multiple pointers to same device (%u)", ptr->dev);
1207
1208
1209         bkey_fsck_err_on(bucket >= nbuckets,
1210                          c, ptr_after_last_bucket,
1211                          "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
1212         bkey_fsck_err_on(bucket < first_bucket,
1213                          c, ptr_before_first_bucket,
1214                          "pointer before first bucket (%llu < %u)", bucket, first_bucket);
1215         bkey_fsck_err_on(bucket_offset + size_ondisk > bucket_size,
1216                          c, ptr_spans_multiple_buckets,
1217                          "pointer spans multiple buckets (%u + %u > %u)",
1218                        bucket_offset, size_ondisk, bucket_size);
1219 fsck_err:
1220         return ret;
1221 }
1222
1223 int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
1224                             enum bch_validate_flags flags)
1225 {
1226         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1227         const union bch_extent_entry *entry;
1228         struct bch_extent_crc_unpacked crc;
1229         unsigned size_ondisk = k.k->size;
1230         unsigned nonce = UINT_MAX;
1231         unsigned nr_ptrs = 0;
1232         bool have_written = false, have_unwritten = false, have_ec = false, crc_since_last_ptr = false;
1233         int ret = 0;
1234
1235         if (bkey_is_btree_ptr(k.k))
1236                 size_ondisk = btree_sectors(c);
1237
1238         bkey_extent_entry_for_each(ptrs, entry) {
1239                 bkey_fsck_err_on(__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX,
1240                                  c, extent_ptrs_invalid_entry,
1241                                  "invalid extent entry type (got %u, max %u)",
1242                                  __extent_entry_type(entry), BCH_EXTENT_ENTRY_MAX);
1243
1244                 bkey_fsck_err_on(bkey_is_btree_ptr(k.k) &&
1245                                  !extent_entry_is_ptr(entry),
1246                                  c, btree_ptr_has_non_ptr,
1247                                  "has non ptr field");
1248
1249                 switch (extent_entry_type(entry)) {
1250                 case BCH_EXTENT_ENTRY_ptr:
1251                         ret = extent_ptr_validate(c, k, flags, &entry->ptr, size_ondisk, false);
1252                         if (ret)
1253                                 return ret;
1254
1255                         bkey_fsck_err_on(entry->ptr.cached && have_ec,
1256                                          c, ptr_cached_and_erasure_coded,
1257                                          "cached, erasure coded ptr");
1258
1259                         if (!entry->ptr.unwritten)
1260                                 have_written = true;
1261                         else
1262                                 have_unwritten = true;
1263
1264                         have_ec = false;
1265                         crc_since_last_ptr = false;
1266                         nr_ptrs++;
1267                         break;
1268                 case BCH_EXTENT_ENTRY_crc32:
1269                 case BCH_EXTENT_ENTRY_crc64:
1270                 case BCH_EXTENT_ENTRY_crc128:
1271                         crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
1272
1273                         bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size,
1274                                          c, ptr_crc_uncompressed_size_too_small,
1275                                          "checksum offset + key size > uncompressed size");
1276                         bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type),
1277                                          c, ptr_crc_csum_type_unknown,
1278                                          "invalid checksum type");
1279                         bkey_fsck_err_on(crc.compression_type >= BCH_COMPRESSION_TYPE_NR,
1280                                          c, ptr_crc_compression_type_unknown,
1281                                          "invalid compression type");
1282
1283                         if (bch2_csum_type_is_encryption(crc.csum_type)) {
1284                                 if (nonce == UINT_MAX)
1285                                         nonce = crc.offset + crc.nonce;
1286                                 else if (nonce != crc.offset + crc.nonce)
1287                                         bkey_fsck_err(c, ptr_crc_nonce_mismatch,
1288                                                       "incorrect nonce");
1289                         }
1290
1291                         bkey_fsck_err_on(crc_since_last_ptr,
1292                                          c, ptr_crc_redundant,
1293                                          "redundant crc entry");
1294                         crc_since_last_ptr = true;
1295
1296                         bkey_fsck_err_on(crc_is_encoded(crc) &&
1297                                          (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
1298                                          (flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)),
1299                                          c, ptr_crc_uncompressed_size_too_big,
1300                                          "too large encoded extent");
1301
1302                         size_ondisk = crc.compressed_size;
1303                         break;
1304                 case BCH_EXTENT_ENTRY_stripe_ptr:
1305                         bkey_fsck_err_on(have_ec,
1306                                          c, ptr_stripe_redundant,
1307                                          "redundant stripe entry");
1308                         have_ec = true;
1309                         break;
1310                 case BCH_EXTENT_ENTRY_rebalance: {
1311                         /*
1312                          * this shouldn't be a fsck error, for forward
1313                          * compatibility; the rebalance code should just refetch
1314                          * the compression opt if it's unknown
1315                          */
1316 #if 0
1317                         const struct bch_extent_rebalance *r = &entry->rebalance;
1318
1319                         if (!bch2_compression_opt_valid(r->compression)) {
1320                                 struct bch_compression_opt opt = __bch2_compression_decode(r->compression);
1321                                 prt_printf(err, "invalid compression opt %u:%u",
1322                                            opt.type, opt.level);
1323                                 return -BCH_ERR_invalid_bkey;
1324                         }
1325 #endif
1326                         break;
1327                 }
1328                 }
1329         }
1330
1331         bkey_fsck_err_on(!nr_ptrs,
1332                          c, extent_ptrs_no_ptrs,
1333                          "no ptrs");
1334         bkey_fsck_err_on(nr_ptrs > BCH_BKEY_PTRS_MAX,
1335                          c, extent_ptrs_too_many_ptrs,
1336                          "too many ptrs: %u > %u", nr_ptrs, BCH_BKEY_PTRS_MAX);
1337         bkey_fsck_err_on(have_written && have_unwritten,
1338                          c, extent_ptrs_written_and_unwritten,
1339                          "extent with unwritten and written ptrs");
1340         bkey_fsck_err_on(k.k->type != KEY_TYPE_extent && have_unwritten,
1341                          c, extent_ptrs_unwritten,
1342                          "has unwritten ptrs");
1343         bkey_fsck_err_on(crc_since_last_ptr,
1344                          c, extent_ptrs_redundant_crc,
1345                          "redundant crc entry");
1346         bkey_fsck_err_on(have_ec,
1347                          c, extent_ptrs_redundant_stripe,
1348                          "redundant stripe entry");
1349 fsck_err:
1350         return ret;
1351 }
1352
1353 void bch2_ptr_swab(struct bkey_s k)
1354 {
1355         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1356         union bch_extent_entry *entry;
1357         u64 *d;
1358
1359         for (d =  (u64 *) ptrs.start;
1360              d != (u64 *) ptrs.end;
1361              d++)
1362                 *d = swab64(*d);
1363
1364         for (entry = ptrs.start;
1365              entry < ptrs.end;
1366              entry = extent_entry_next(entry)) {
1367                 switch (__extent_entry_type(entry)) {
1368                 case BCH_EXTENT_ENTRY_ptr:
1369                         break;
1370                 case BCH_EXTENT_ENTRY_crc32:
1371                         entry->crc32.csum = swab32(entry->crc32.csum);
1372                         break;
1373                 case BCH_EXTENT_ENTRY_crc64:
1374                         entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
1375                         entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
1376                         break;
1377                 case BCH_EXTENT_ENTRY_crc128:
1378                         entry->crc128.csum.hi = (__force __le64)
1379                                 swab64((__force u64) entry->crc128.csum.hi);
1380                         entry->crc128.csum.lo = (__force __le64)
1381                                 swab64((__force u64) entry->crc128.csum.lo);
1382                         break;
1383                 case BCH_EXTENT_ENTRY_stripe_ptr:
1384                         break;
1385                 case BCH_EXTENT_ENTRY_rebalance:
1386                         break;
1387                 default:
1388                         /* Bad entry type: will be caught by validate() */
1389                         return;
1390                 }
1391         }
1392 }
1393
1394 const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
1395 {
1396         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1397         const union bch_extent_entry *entry;
1398
1399         bkey_extent_entry_for_each(ptrs, entry)
1400                 if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
1401                         return &entry->rebalance;
1402
1403         return NULL;
1404 }
1405
1406 unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
1407                                        unsigned target, unsigned compression)
1408 {
1409         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1410         unsigned rewrite_ptrs = 0;
1411
1412         if (compression) {
1413                 unsigned compression_type = bch2_compression_opt_to_type(compression);
1414                 const union bch_extent_entry *entry;
1415                 struct extent_ptr_decoded p;
1416                 unsigned i = 0;
1417
1418                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1419                         if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
1420                             p.ptr.unwritten) {
1421                                 rewrite_ptrs = 0;
1422                                 goto incompressible;
1423                         }
1424
1425                         if (!p.ptr.cached && p.crc.compression_type != compression_type)
1426                                 rewrite_ptrs |= 1U << i;
1427                         i++;
1428                 }
1429         }
1430 incompressible:
1431         if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
1432                 unsigned i = 0;
1433
1434                 bkey_for_each_ptr(ptrs, ptr) {
1435                         if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, target))
1436                                 rewrite_ptrs |= 1U << i;
1437                         i++;
1438                 }
1439         }
1440
1441         return rewrite_ptrs;
1442 }
1443
1444 bool bch2_bkey_needs_rebalance(struct bch_fs *c, struct bkey_s_c k)
1445 {
1446         const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
1447
1448         /*
1449          * If it's an indirect extent, we don't delete the rebalance entry when
1450          * done so that we know what options were applied - check if it still
1451          * needs work done:
1452          */
1453         if (r &&
1454             k.k->type == KEY_TYPE_reflink_v &&
1455             !bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression))
1456                 r = NULL;
1457
1458         return r != NULL;
1459 }
1460
1461 static u64 __bch2_bkey_sectors_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
1462                                        unsigned target, unsigned compression)
1463 {
1464         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1465         const union bch_extent_entry *entry;
1466         struct extent_ptr_decoded p;
1467         u64 sectors = 0;
1468
1469         if (compression) {
1470                 unsigned compression_type = bch2_compression_opt_to_type(compression);
1471
1472                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1473                         if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
1474                             p.ptr.unwritten) {
1475                                 sectors = 0;
1476                                 goto incompressible;
1477                         }
1478
1479                         if (!p.ptr.cached && p.crc.compression_type != compression_type)
1480                                 sectors += p.crc.compressed_size;
1481                 }
1482         }
1483 incompressible:
1484         if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
1485                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
1486                         if (!p.ptr.cached && !bch2_dev_in_target(c, p.ptr.dev, target))
1487                                 sectors += p.crc.compressed_size;
1488         }
1489
1490         return sectors;
1491 }
1492
1493 u64 bch2_bkey_sectors_need_rebalance(struct bch_fs *c, struct bkey_s_c k)
1494 {
1495         const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
1496
1497         return r ? __bch2_bkey_sectors_need_rebalance(c, k, r->target, r->compression) : 0;
1498 }
1499
1500 int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bkey_i *_k,
1501                                   struct bch_io_opts *opts)
1502 {
1503         struct bkey_s k = bkey_i_to_s(_k);
1504         struct bch_extent_rebalance *r;
1505         unsigned target = opts->background_target;
1506         unsigned compression = background_compression(*opts);
1507         bool needs_rebalance;
1508
1509         if (!bkey_extent_is_direct_data(k.k))
1510                 return 0;
1511
1512         /* get existing rebalance entry: */
1513         r = (struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
1514         if (r) {
1515                 if (k.k->type == KEY_TYPE_reflink_v) {
1516                         /*
1517                          * indirect extents: existing options take precedence,
1518                          * so that we don't move extents back and forth if
1519                          * they're referenced by different inodes with different
1520                          * options:
1521                          */
1522                         if (r->target)
1523                                 target = r->target;
1524                         if (r->compression)
1525                                 compression = r->compression;
1526                 }
1527
1528                 r->target       = target;
1529                 r->compression  = compression;
1530         }
1531
1532         needs_rebalance = bch2_bkey_ptrs_need_rebalance(c, k.s_c, target, compression);
1533
1534         if (needs_rebalance && !r) {
1535                 union bch_extent_entry *new = bkey_val_end(k);
1536
1537                 new->rebalance.type             = 1U << BCH_EXTENT_ENTRY_rebalance;
1538                 new->rebalance.compression      = compression;
1539                 new->rebalance.target           = target;
1540                 new->rebalance.unused           = 0;
1541                 k.k->u64s += extent_entry_u64s(new);
1542         } else if (!needs_rebalance && r && k.k->type != KEY_TYPE_reflink_v) {
1543                 /*
1544                  * For indirect extents, don't delete the rebalance entry when
1545                  * we're finished so that we know we specifically moved it or
1546                  * compressed it to its current location/compression type
1547                  */
1548                 extent_entry_drop(k, (union bch_extent_entry *) r);
1549         }
1550
1551         return 0;
1552 }
1553
1554 /* Generic extent code: */
1555
1556 int bch2_cut_front_s(struct bpos where, struct bkey_s k)
1557 {
1558         unsigned new_val_u64s = bkey_val_u64s(k.k);
1559         int val_u64s_delta;
1560         u64 sub;
1561
1562         if (bkey_le(where, bkey_start_pos(k.k)))
1563                 return 0;
1564
1565         EBUG_ON(bkey_gt(where, k.k->p));
1566
1567         sub = where.offset - bkey_start_offset(k.k);
1568
1569         k.k->size -= sub;
1570
1571         if (!k.k->size) {
1572                 k.k->type = KEY_TYPE_deleted;
1573                 new_val_u64s = 0;
1574         }
1575
1576         switch (k.k->type) {
1577         case KEY_TYPE_extent:
1578         case KEY_TYPE_reflink_v: {
1579                 struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
1580                 union bch_extent_entry *entry;
1581                 bool seen_crc = false;
1582
1583                 bkey_extent_entry_for_each(ptrs, entry) {
1584                         switch (extent_entry_type(entry)) {
1585                         case BCH_EXTENT_ENTRY_ptr:
1586                                 if (!seen_crc)
1587                                         entry->ptr.offset += sub;
1588                                 break;
1589                         case BCH_EXTENT_ENTRY_crc32:
1590                                 entry->crc32.offset += sub;
1591                                 break;
1592                         case BCH_EXTENT_ENTRY_crc64:
1593                                 entry->crc64.offset += sub;
1594                                 break;
1595                         case BCH_EXTENT_ENTRY_crc128:
1596                                 entry->crc128.offset += sub;
1597                                 break;
1598                         case BCH_EXTENT_ENTRY_stripe_ptr:
1599                                 break;
1600                         case BCH_EXTENT_ENTRY_rebalance:
1601                                 break;
1602                         }
1603
1604                         if (extent_entry_is_crc(entry))
1605                                 seen_crc = true;
1606                 }
1607
1608                 break;
1609         }
1610         case KEY_TYPE_reflink_p: {
1611                 struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
1612
1613                 le64_add_cpu(&p.v->idx, sub);
1614                 break;
1615         }
1616         case KEY_TYPE_inline_data:
1617         case KEY_TYPE_indirect_inline_data: {
1618                 void *p = bkey_inline_data_p(k);
1619                 unsigned bytes = bkey_inline_data_bytes(k.k);
1620
1621                 sub = min_t(u64, sub << 9, bytes);
1622
1623                 memmove(p, p + sub, bytes - sub);
1624
1625                 new_val_u64s -= sub >> 3;
1626                 break;
1627         }
1628         }
1629
1630         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1631         BUG_ON(val_u64s_delta < 0);
1632
1633         set_bkey_val_u64s(k.k, new_val_u64s);
1634         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1635         return -val_u64s_delta;
1636 }
1637
1638 int bch2_cut_back_s(struct bpos where, struct bkey_s k)
1639 {
1640         unsigned new_val_u64s = bkey_val_u64s(k.k);
1641         int val_u64s_delta;
1642         u64 len = 0;
1643
1644         if (bkey_ge(where, k.k->p))
1645                 return 0;
1646
1647         EBUG_ON(bkey_lt(where, bkey_start_pos(k.k)));
1648
1649         len = where.offset - bkey_start_offset(k.k);
1650
1651         k.k->p.offset = where.offset;
1652         k.k->size = len;
1653
1654         if (!len) {
1655                 k.k->type = KEY_TYPE_deleted;
1656                 new_val_u64s = 0;
1657         }
1658
1659         switch (k.k->type) {
1660         case KEY_TYPE_inline_data:
1661         case KEY_TYPE_indirect_inline_data:
1662                 new_val_u64s = (bkey_inline_data_offset(k.k) +
1663                                 min(bkey_inline_data_bytes(k.k), k.k->size << 9)) >> 3;
1664                 break;
1665         }
1666
1667         val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
1668         BUG_ON(val_u64s_delta < 0);
1669
1670         set_bkey_val_u64s(k.k, new_val_u64s);
1671         memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
1672         return -val_u64s_delta;
1673 }