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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_buf.h"
   6 #include "btree_gc.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "buckets.h"
  10 #include "disk_groups.h"
  11 #include "ec.h"
  12 #include "inode.h"
  13 #include "io.h"
  14 #include "journal_reclaim.h"
  15 #include "move.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18 #include "super-io.h"
  19 #include "keylist.h"
  20
  21 #include <linux/ioprio.h>
  22 #include <linux/kthread.h>
  23
  24 #include <trace/events/bcachefs.h>
  25
  26 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  27
  28 struct moving_io {
  29         struct list_head        list;
  30         struct closure          cl;
  31         bool                    read_completed;
  32
  33         unsigned                read_sectors;
  34         unsigned                write_sectors;
  35
  36         struct bch_read_bio     rbio;
  37
  38         struct migrate_write    write;
  39         /* Must be last since it is variable size */
  40         struct bio_vec          bi_inline_vecs[0];
  41 };
  42
  43 struct moving_context {
  44         /* Closure for waiting on all reads and writes to complete */
  45         struct closure          cl;
  46
  47         struct bch_move_stats   *stats;
  48
  49         struct list_head        reads;
  50
  51         /* in flight sectors: */
  52         atomic_t                read_sectors;
  53         atomic_t                write_sectors;
  54
  55         wait_queue_head_t       wait;
  56 };
  57
  58 static int insert_snapshot_whiteouts(struct btree_trans *trans,
  59                                      enum btree_id id,
  60                                      struct bpos old_pos,
  61                                      struct bpos new_pos)
  62 {
  63         struct bch_fs *c = trans->c;
  64         struct btree_iter iter, update_iter;
  65         struct bkey_s_c k;
  66         struct snapshots_seen s;
  67         int ret;
  68
  69         if (!btree_type_has_snapshots(id))
  70                 return 0;
  71
  72         snapshots_seen_init(&s);
  73
  74         if (!bkey_cmp(old_pos, new_pos))
  75                 return 0;
  76
  77         if (!snapshot_t(c, old_pos.snapshot)->children[0])
  78                 return 0;
  79
  80         bch2_trans_iter_init(trans, &iter, id, old_pos,
  81                              BTREE_ITER_NOT_EXTENTS|
  82                              BTREE_ITER_ALL_SNAPSHOTS);
  83         while (1) {
  84 next:
  85                 k = bch2_btree_iter_prev(&iter);
  86                 ret = bkey_err(k);
  87                 if (ret)
  88                         break;
  89
  90                 if (bkey_cmp(old_pos, k.k->p))
  91                         break;
  92
  93                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
  94                         struct bkey_i *update;
  95                         size_t i;
  96
  97                         for (i = 0; i < s.nr; i++)
  98                                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, s.d[i]))
  99                                         goto next;
 100
 101                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 102
 103                         ret = PTR_ERR_OR_ZERO(update);
 104                         if (ret)
 105                                 break;
 106
 107                         bkey_init(&update->k);
 108                         update->k.p = new_pos;
 109                         update->k.p.snapshot = k.k->p.snapshot;
 110
 111                         bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
 112                                              BTREE_ITER_NOT_EXTENTS|
 113                                              BTREE_ITER_ALL_SNAPSHOTS|
 114                                              BTREE_ITER_INTENT);
 115                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
 116                                 bch2_trans_update(trans, &update_iter, update,
 117                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 118                         bch2_trans_iter_exit(trans, &update_iter);
 119                         if (ret)
 120                                 break;
 121
 122                         ret = snapshots_seen_add(c, &s, k.k->p.snapshot);
 123                         if (ret)
 124                                 break;
 125                 }
 126         }
 127         bch2_trans_iter_exit(trans, &iter);
 128         kfree(s.d);
 129
 130         return ret;
 131 }
 132
 133 static int bch2_migrate_index_update(struct bch_write_op *op)
 134 {
 135         struct bch_fs *c = op->c;
 136         struct btree_trans trans;
 137         struct btree_iter iter;
 138         struct migrate_write *m =
 139                 container_of(op, struct migrate_write, op);
 140         struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
 141         struct keylist *keys = &op->insert_keys;
 142         struct bkey_buf _new, _insert;
 143         int ret = 0;
 144
 145         bch2_bkey_buf_init(&_new);
 146         bch2_bkey_buf_init(&_insert);
 147         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 148
 149         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 150
 151         bch2_trans_iter_init(&trans, &iter, m->btree_id,
 152                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 153                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 154
 155         while (1) {
 156                 struct bkey_s_c k;
 157                 struct bkey_i *insert;
 158                 struct bkey_i_extent *new;
 159                 const union bch_extent_entry *entry;
 160                 struct extent_ptr_decoded p;
 161                 struct bpos next_pos;
 162                 bool did_work = false;
 163                 bool extending = false, should_check_enospc;
 164                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 165
 166                 bch2_trans_begin(&trans);
 167
 168                 k = bch2_btree_iter_peek_slot(&iter);
 169                 ret = bkey_err(k);
 170                 if (ret)
 171                         goto err;
 172
 173                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 174
 175                 if (bversion_cmp(k.k->version, new->k.version) ||
 176                     !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
 177                         goto nomatch;
 178
 179                 bkey_reassemble(_insert.k, k);
 180                 insert = _insert.k;
 181
 182                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 183                 new = bkey_i_to_extent(_new.k);
 184                 bch2_cut_front(iter.pos, &new->k_i);
 185
 186                 bch2_cut_front(iter.pos,        insert);
 187                 bch2_cut_back(new->k.p,         insert);
 188                 bch2_cut_back(insert->k.p,      &new->k_i);
 189
 190                 if (m->data_cmd == DATA_REWRITE) {
 191                         struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
 192                                 bch2_bkey_has_device(bkey_i_to_s_c(insert),
 193                                                      m->data_opts.rewrite_dev);
 194                         if (!old_ptr)
 195                                 goto nomatch;
 196
 197                         if (old_ptr->cached)
 198                                 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
 199                                         new_ptr->cached = true;
 200
 201                         __bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
 202                 }
 203
 204                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 205                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 206                                 /*
 207                                  * raced with another move op? extent already
 208                                  * has a pointer to the device we just wrote
 209                                  * data to
 210                                  */
 211                                 continue;
 212                         }
 213
 214                         bch2_extent_ptr_decoded_append(insert, &p);
 215                         did_work = true;
 216                 }
 217
 218                 if (!did_work)
 219                         goto nomatch;
 220
 221                 bch2_bkey_narrow_crcs(insert,
 222                                 (struct bch_extent_crc_unpacked) { 0 });
 223                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 224                 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
 225                                                op->opts.background_target,
 226                                                op->opts.data_replicas);
 227
 228                 ret = bch2_sum_sector_overwrites(&trans, &iter, insert,
 229                                                  &extending,
 230                                                  &should_check_enospc,
 231                                                  &i_sectors_delta,
 232                                                  &disk_sectors_delta);
 233                 if (ret)
 234                         goto err;
 235
 236                 if (disk_sectors_delta > (s64) op->res.sectors) {
 237                         ret = bch2_disk_reservation_add(c, &op->res,
 238                                                 disk_sectors_delta - op->res.sectors,
 239                                                 !should_check_enospc
 240                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 241                         if (ret)
 242                                 goto out;
 243                 }
 244
 245                 next_pos = insert->k.p;
 246
 247                 ret   = insert_snapshot_whiteouts(&trans, m->btree_id,
 248                                                   k.k->p, insert->k.p) ?:
 249                         bch2_trans_update(&trans, &iter, insert,
 250                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 251                         bch2_trans_commit(&trans, &op->res,
 252                                 op_journal_seq(op),
 253                                 BTREE_INSERT_NOFAIL|
 254                                 m->data_opts.btree_insert_flags);
 255                 if (!ret) {
 256                         bch2_btree_iter_set_pos(&iter, next_pos);
 257                         atomic_long_inc(&c->extent_migrate_done);
 258                         if (ec_ob)
 259                                 bch2_ob_add_backpointer(c, ec_ob, &insert->k);
 260                 }
 261 err:
 262                 if (ret == -EINTR)
 263                         ret = 0;
 264                 if (ret)
 265                         break;
 266 next:
 267                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 268                         bch2_keylist_pop_front(keys);
 269                         if (bch2_keylist_empty(keys))
 270                                 goto out;
 271                 }
 272                 continue;
 273 nomatch:
 274                 if (m->ctxt) {
 275                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 276                         atomic64_inc(&m->ctxt->stats->keys_raced);
 277                         atomic64_add(k.k->p.offset - iter.pos.offset,
 278                                      &m->ctxt->stats->sectors_raced);
 279                 }
 280                 atomic_long_inc(&c->extent_migrate_raced);
 281                 trace_move_race(&new->k);
 282                 bch2_btree_iter_advance(&iter);
 283                 goto next;
 284         }
 285 out:
 286         bch2_trans_iter_exit(&trans, &iter);
 287         bch2_trans_exit(&trans);
 288         bch2_bkey_buf_exit(&_insert, c);
 289         bch2_bkey_buf_exit(&_new, c);
 290         BUG_ON(ret == -EINTR);
 291         return ret;
 292 }
 293
 294 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 295 {
 296         /* write bio must own pages: */
 297         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 298
 299         m->ptr          = rbio->pick.ptr;
 300         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 301         m->op.devs_have = rbio->devs_have;
 302         m->op.pos       = rbio->data_pos;
 303         m->op.version   = rbio->version;
 304         m->op.crc       = rbio->pick.crc;
 305         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 306
 307         if (m->data_cmd == DATA_REWRITE)
 308                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 309 }
 310
 311 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 312                             struct write_point_specifier wp,
 313                             struct bch_io_opts io_opts,
 314                             enum data_cmd data_cmd,
 315                             struct data_opts data_opts,
 316                             enum btree_id btree_id,
 317                             struct bkey_s_c k)
 318 {
 319         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 320         const union bch_extent_entry *entry;
 321         struct bch_extent_crc_unpacked crc;
 322         struct extent_ptr_decoded p;
 323         int ret;
 324
 325         m->btree_id     = btree_id;
 326         m->data_cmd     = data_cmd;
 327         m->data_opts    = data_opts;
 328         m->nr_ptrs_reserved = 0;
 329
 330         bch2_write_op_init(&m->op, c, io_opts);
 331
 332         if (!bch2_bkey_is_incompressible(k))
 333                 m->op.compression_type =
 334                         bch2_compression_opt_to_type[io_opts.background_compression ?:
 335                                                      io_opts.compression];
 336         else
 337                 m->op.incompressible = true;
 338
 339         m->op.target    = data_opts.target,
 340         m->op.write_point = wp;
 341
 342         /*
 343          * op->csum_type is normally initialized from the fs/file's current
 344          * options - but if an extent is encrypted, we require that it stays
 345          * encrypted:
 346          */
 347         bkey_for_each_crc(k.k, ptrs, crc, entry)
 348                 if (bch2_csum_type_is_encryption(crc.csum_type)) {
 349                         m->op.nonce     = crc.nonce + crc.offset;
 350                         m->op.csum_type = crc.csum_type;
 351                         break;
 352                 }
 353
 354         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 355                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 356                 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
 357         } else {
 358                 /* XXX: this should probably be passed in */
 359                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 360         }
 361
 362         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 363                 BCH_WRITE_PAGES_OWNED|
 364                 BCH_WRITE_DATA_ENCODED|
 365                 BCH_WRITE_FROM_INTERNAL;
 366
 367         m->op.nr_replicas       = data_opts.nr_replicas;
 368         m->op.nr_replicas_required = data_opts.nr_replicas;
 369         m->op.index_update_fn   = bch2_migrate_index_update;
 370
 371         switch (data_cmd) {
 372         case DATA_ADD_REPLICAS: {
 373                 /*
 374                  * DATA_ADD_REPLICAS is used for moving data to a different
 375                  * device in the background, and due to compression the new copy
 376                  * might take up more space than the old copy:
 377                  */
 378 #if 0
 379                 int nr = (int) io_opts.data_replicas -
 380                         bch2_bkey_nr_ptrs_allocated(k);
 381 #endif
 382                 int nr = (int) io_opts.data_replicas;
 383
 384                 if (nr > 0) {
 385                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 386
 387                         ret = bch2_disk_reservation_get(c, &m->op.res,
 388                                         k.k->size, m->op.nr_replicas, 0);
 389                         if (ret)
 390                                 return ret;
 391                 }
 392                 break;
 393         }
 394         case DATA_REWRITE: {
 395                 unsigned compressed_sectors = 0;
 396
 397                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 398                         if (p.ptr.dev == data_opts.rewrite_dev &&
 399                             !p.ptr.cached &&
 400                             crc_is_compressed(p.crc))
 401                                 compressed_sectors += p.crc.compressed_size;
 402
 403                 if (compressed_sectors) {
 404                         ret = bch2_disk_reservation_add(c, &m->op.res,
 405                                         k.k->size * m->op.nr_replicas,
 406                                         BCH_DISK_RESERVATION_NOFAIL);
 407                         if (ret)
 408                                 return ret;
 409                 }
 410                 break;
 411         }
 412         case DATA_PROMOTE:
 413                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 414                 m->op.flags     |= BCH_WRITE_CACHED;
 415                 break;
 416         default:
 417                 BUG();
 418         }
 419
 420         return 0;
 421 }
 422
 423 static void move_free(struct closure *cl)
 424 {
 425         struct moving_io *io = container_of(cl, struct moving_io, cl);
 426         struct moving_context *ctxt = io->write.ctxt;
 427         struct bvec_iter_all iter;
 428         struct bio_vec *bv;
 429
 430         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 431
 432         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 433                 if (bv->bv_page)
 434                         __free_page(bv->bv_page);
 435
 436         wake_up(&ctxt->wait);
 437
 438         kfree(io);
 439 }
 440
 441 static void move_write_done(struct closure *cl)
 442 {
 443         struct moving_io *io = container_of(cl, struct moving_io, cl);
 444
 445         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 446         closure_return_with_destructor(cl, move_free);
 447 }
 448
 449 static void move_write(struct closure *cl)
 450 {
 451         struct moving_io *io = container_of(cl, struct moving_io, cl);
 452
 453         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 454                 closure_return_with_destructor(cl, move_free);
 455                 return;
 456         }
 457
 458         bch2_migrate_read_done(&io->write, &io->rbio);
 459
 460         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 461         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 462         continue_at(cl, move_write_done, NULL);
 463 }
 464
 465 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 466 {
 467         struct moving_io *io =
 468                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 469
 470         return io && io->read_completed ? io : NULL;
 471 }
 472
 473 static void move_read_endio(struct bio *bio)
 474 {
 475         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 476         struct moving_context *ctxt = io->write.ctxt;
 477
 478         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 479         io->read_completed = true;
 480
 481         if (next_pending_write(ctxt))
 482                 wake_up(&ctxt->wait);
 483
 484         closure_put(&ctxt->cl);
 485 }
 486
 487 static void do_pending_writes(struct moving_context *ctxt)
 488 {
 489         struct moving_io *io;
 490
 491         while ((io = next_pending_write(ctxt))) {
 492                 list_del(&io->list);
 493                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 494         }
 495 }
 496
 497 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 498 do {                                                            \
 499         do_pending_writes(_ctxt);                               \
 500                                                                 \
 501         if (_cond)                                              \
 502                 break;                                          \
 503         __wait_event((_ctxt)->wait,                             \
 504                      next_pending_write(_ctxt) || (_cond));     \
 505 } while (1)
 506
 507 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 508 {
 509         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 510
 511         move_ctxt_wait_event(ctxt,
 512                 !atomic_read(&ctxt->write_sectors) ||
 513                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 514 }
 515
 516 static int bch2_move_extent(struct btree_trans *trans,
 517                             struct moving_context *ctxt,
 518                             struct write_point_specifier wp,
 519                             struct bch_io_opts io_opts,
 520                             enum btree_id btree_id,
 521                             struct bkey_s_c k,
 522                             enum data_cmd data_cmd,
 523                             struct data_opts data_opts)
 524 {
 525         struct bch_fs *c = trans->c;
 526         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 527         struct moving_io *io;
 528         const union bch_extent_entry *entry;
 529         struct extent_ptr_decoded p;
 530         unsigned sectors = k.k->size, pages;
 531         int ret = -ENOMEM;
 532
 533         move_ctxt_wait_event(ctxt,
 534                 atomic_read(&ctxt->write_sectors) <
 535                 SECTORS_IN_FLIGHT_PER_DEVICE);
 536
 537         move_ctxt_wait_event(ctxt,
 538                 atomic_read(&ctxt->read_sectors) <
 539                 SECTORS_IN_FLIGHT_PER_DEVICE);
 540
 541         /* write path might have to decompress data: */
 542         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 543                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 544
 545         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 546         io = kzalloc(sizeof(struct moving_io) +
 547                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 548         if (!io)
 549                 goto err;
 550
 551         io->write.ctxt          = ctxt;
 552         io->read_sectors        = k.k->size;
 553         io->write_sectors       = k.k->size;
 554
 555         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 556         bio_set_prio(&io->write.op.wbio.bio,
 557                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 558
 559         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 560                                  GFP_KERNEL))
 561                 goto err_free;
 562
 563         io->rbio.c              = c;
 564         io->rbio.opts           = io_opts;
 565         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 566         io->rbio.bio.bi_vcnt = pages;
 567         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 568         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 569
 570         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 571         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 572         io->rbio.bio.bi_end_io          = move_read_endio;
 573
 574         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 575                                       data_cmd, data_opts, btree_id, k);
 576         if (ret)
 577                 goto err_free_pages;
 578
 579         atomic64_inc(&ctxt->stats->keys_moved);
 580         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 581
 582         trace_move_extent(k.k);
 583
 584         atomic_add(io->read_sectors, &ctxt->read_sectors);
 585         list_add_tail(&io->list, &ctxt->reads);
 586
 587         /*
 588          * dropped by move_read_endio() - guards against use after free of
 589          * ctxt when doing wakeup
 590          */
 591         closure_get(&ctxt->cl);
 592         bch2_read_extent(trans, &io->rbio,
 593                          bkey_start_pos(k.k),
 594                          btree_id, k, 0,
 595                          BCH_READ_NODECODE|
 596                          BCH_READ_LAST_FRAGMENT);
 597         return 0;
 598 err_free_pages:
 599         bio_free_pages(&io->write.op.wbio.bio);
 600 err_free:
 601         kfree(io);
 602 err:
 603         trace_move_alloc_fail(k.k);
 604         return ret;
 605 }
 606
 607 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 608                         struct bch_inode_unpacked *inode)
 609 {
 610         struct btree_iter iter;
 611         struct bkey_s_c k;
 612         int ret;
 613
 614         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 615                              BTREE_ITER_ALL_SNAPSHOTS);
 616         k = bch2_btree_iter_peek(&iter);
 617         ret = bkey_err(k);
 618         if (ret)
 619                 goto err;
 620
 621         if (!k.k || bkey_cmp(k.k->p, pos)) {
 622                 ret = -ENOENT;
 623                 goto err;
 624         }
 625
 626         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 627         if (ret)
 628                 goto err;
 629
 630         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 631         if (ret)
 632                 goto err;
 633 err:
 634         bch2_trans_iter_exit(trans, &iter);
 635         return ret;
 636 }
 637
 638 static int __bch2_move_data(struct bch_fs *c,
 639                 struct moving_context *ctxt,
 640                 struct bch_ratelimit *rate,
 641                 struct write_point_specifier wp,
 642                 struct bpos start,
 643                 struct bpos end,
 644                 move_pred_fn pred, void *arg,
 645                 struct bch_move_stats *stats,
 646                 enum btree_id btree_id)
 647 {
 648         bool kthread = (current->flags & PF_KTHREAD) != 0;
 649         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 650         struct bkey_buf sk;
 651         struct btree_trans trans;
 652         struct btree_iter iter;
 653         struct bkey_s_c k;
 654         struct data_opts data_opts;
 655         enum data_cmd data_cmd;
 656         u64 delay, cur_inum = U64_MAX;
 657         int ret = 0, ret2;
 658
 659         bch2_bkey_buf_init(&sk);
 660         bch2_trans_init(&trans, c, 0, 0);
 661
 662         stats->data_type = BCH_DATA_user;
 663         stats->btree_id = btree_id;
 664         stats->pos      = start;
 665
 666         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 667                              BTREE_ITER_PREFETCH|
 668                              BTREE_ITER_ALL_SNAPSHOTS);
 669
 670         if (rate)
 671                 bch2_ratelimit_reset(rate);
 672
 673         while (1) {
 674                 do {
 675                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 676
 677                         if (delay) {
 678                                 bch2_trans_unlock(&trans);
 679                                 set_current_state(TASK_INTERRUPTIBLE);
 680                         }
 681
 682                         if (kthread && (ret = kthread_should_stop())) {
 683                                 __set_current_state(TASK_RUNNING);
 684                                 goto out;
 685                         }
 686
 687                         if (delay)
 688                                 schedule_timeout(delay);
 689
 690                         if (unlikely(freezing(current))) {
 691                                 bch2_trans_unlock(&trans);
 692                                 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 693                                 try_to_freeze();
 694                         }
 695                 } while (delay);
 696
 697                 bch2_trans_begin(&trans);
 698
 699                 k = bch2_btree_iter_peek(&iter);
 700
 701                 stats->pos = iter.pos;
 702
 703                 if (!k.k)
 704                         break;
 705                 ret = bkey_err(k);
 706                 if (ret)
 707                         break;
 708                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 709                         break;
 710
 711                 if (!bkey_extent_is_direct_data(k.k))
 712                         goto next_nondata;
 713
 714                 if (btree_id == BTREE_ID_extents &&
 715                     cur_inum != k.k->p.inode) {
 716                         struct bch_inode_unpacked inode;
 717
 718                         io_opts = bch2_opts_to_inode_opts(c->opts);
 719
 720                         ret = lookup_inode(&trans,
 721                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 722                                         &inode);
 723                         if (ret == -EINTR)
 724                                 continue;
 725
 726                         if (!ret)
 727                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 728
 729                         cur_inum = k.k->p.inode;
 730                 }
 731
 732                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 733                 case DATA_SKIP:
 734                         goto next;
 735                 case DATA_SCRUB:
 736                         BUG();
 737                 case DATA_ADD_REPLICAS:
 738                 case DATA_REWRITE:
 739                 case DATA_PROMOTE:
 740                         break;
 741                 default:
 742                         BUG();
 743                 }
 744
 745                 /* unlock before doing IO: */
 746                 bch2_bkey_buf_reassemble(&sk, c, k);
 747                 k = bkey_i_to_s_c(sk.k);
 748                 bch2_trans_unlock(&trans);
 749
 750                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 751                                         data_cmd, data_opts);
 752                 if (ret2) {
 753                         if (ret2 == -EINTR) {
 754                                 bch2_trans_begin(&trans);
 755                                 continue;
 756                         }
 757
 758                         if (ret2 == -ENOMEM) {
 759                                 /* memory allocation failure, wait for some IO to finish */
 760                                 bch2_move_ctxt_wait_for_io(ctxt);
 761                                 continue;
 762                         }
 763
 764                         /* XXX signal failure */
 765                         goto next;
 766                 }
 767
 768                 if (rate)
 769                         bch2_ratelimit_increment(rate, k.k->size);
 770 next:
 771                 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 772                              &stats->sectors_seen);
 773 next_nondata:
 774                 bch2_btree_iter_advance(&iter);
 775         }
 776 out:
 777
 778         bch2_trans_iter_exit(&trans, &iter);
 779         bch2_trans_exit(&trans);
 780         bch2_bkey_buf_exit(&sk, c);
 781
 782         return ret;
 783 }
 784
 785 inline void bch_move_stats_init(struct bch_move_stats *stats, char *name)
 786 {
 787         memset(stats, 0, sizeof(*stats));
 788
 789         scnprintf(stats->name, sizeof(stats->name),
 790                         "%s", name);
 791 }
 792
 793 static inline void progress_list_add(struct bch_fs *c,
 794                                      struct bch_move_stats *stats)
 795 {
 796         mutex_lock(&c->data_progress_lock);
 797         list_add(&stats->list, &c->data_progress_list);
 798         mutex_unlock(&c->data_progress_lock);
 799 }
 800
 801 static inline void progress_list_del(struct bch_fs *c,
 802                                      struct bch_move_stats *stats)
 803 {
 804         mutex_lock(&c->data_progress_lock);
 805         list_del(&stats->list);
 806         mutex_unlock(&c->data_progress_lock);
 807 }
 808
 809 int bch2_move_data(struct bch_fs *c,
 810                    enum btree_id start_btree_id, struct bpos start_pos,
 811                    enum btree_id end_btree_id,   struct bpos end_pos,
 812                    struct bch_ratelimit *rate,
 813                    struct write_point_specifier wp,
 814                    move_pred_fn pred, void *arg,
 815                    struct bch_move_stats *stats)
 816 {
 817         struct moving_context ctxt = { .stats = stats };
 818         enum btree_id id;
 819         int ret;
 820
 821         progress_list_add(c, stats);
 822         closure_init_stack(&ctxt.cl);
 823         INIT_LIST_HEAD(&ctxt.reads);
 824         init_waitqueue_head(&ctxt.wait);
 825
 826         stats->data_type = BCH_DATA_user;
 827
 828         for (id = start_btree_id;
 829              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 830              id++) {
 831                 stats->btree_id = id;
 832
 833                 if (id != BTREE_ID_extents &&
 834                     id != BTREE_ID_reflink)
 835                         continue;
 836
 837                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 838                                        id == start_btree_id ? start_pos : POS_MIN,
 839                                        id == end_btree_id   ? end_pos   : POS_MAX,
 840                                        pred, arg, stats, id);
 841                 if (ret)
 842                         break;
 843         }
 844
 845
 846         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 847         closure_sync(&ctxt.cl);
 848
 849         EBUG_ON(atomic_read(&ctxt.write_sectors));
 850
 851         trace_move_data(c,
 852                         atomic64_read(&stats->sectors_moved),
 853                         atomic64_read(&stats->keys_moved));
 854
 855         progress_list_del(c, stats);
 856         return ret;
 857 }
 858
 859 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 860                                          struct btree *, struct bch_io_opts *,
 861                                          struct data_opts *);
 862
 863 static int bch2_move_btree(struct bch_fs *c,
 864                            enum btree_id start_btree_id, struct bpos start_pos,
 865                            enum btree_id end_btree_id,   struct bpos end_pos,
 866                            move_btree_pred pred, void *arg,
 867                            struct bch_move_stats *stats)
 868 {
 869         bool kthread = (current->flags & PF_KTHREAD) != 0;
 870         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 871         struct btree_trans trans;
 872         struct btree_iter iter;
 873         struct btree *b;
 874         enum btree_id id;
 875         struct data_opts data_opts;
 876         enum data_cmd cmd;
 877         int ret = 0;
 878
 879         bch2_trans_init(&trans, c, 0, 0);
 880         progress_list_add(c, stats);
 881
 882         stats->data_type = BCH_DATA_btree;
 883
 884         for (id = start_btree_id;
 885              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 886              id++) {
 887                 stats->btree_id = id;
 888
 889                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 890                                           BTREE_ITER_PREFETCH);
 891 retry:
 892                 ret = 0;
 893                 while (bch2_trans_begin(&trans),
 894                        (b = bch2_btree_iter_peek_node(&iter)) &&
 895                        !(ret = PTR_ERR_OR_ZERO(b))) {
 896                         if (kthread && kthread_should_stop())
 897                                 break;
 898
 899                         if ((cmp_int(id, end_btree_id) ?:
 900                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 901                                 break;
 902
 903                         stats->pos = iter.pos;
 904
 905                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 906                         case DATA_SKIP:
 907                                 goto next;
 908                         case DATA_SCRUB:
 909                                 BUG();
 910                         case DATA_ADD_REPLICAS:
 911                         case DATA_REWRITE:
 912                                 break;
 913                         default:
 914                                 BUG();
 915                         }
 916
 917                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 918                         if (ret == -EINTR)
 919                                 continue;
 920                         if (ret)
 921                                 break;
 922 next:
 923                         bch2_btree_iter_next_node(&iter);
 924                 }
 925                 if (ret == -EINTR)
 926                         goto retry;
 927
 928                 bch2_trans_iter_exit(&trans, &iter);
 929
 930                 if (kthread && kthread_should_stop())
 931                         break;
 932         }
 933
 934         bch2_trans_exit(&trans);
 935
 936         if (ret)
 937                 bch_err(c, "error %i in bch2_move_btree", ret);
 938
 939         /* flush relevant btree updates */
 940         closure_wait_event(&c->btree_interior_update_wait,
 941                            !bch2_btree_interior_updates_nr_pending(c));
 942
 943         progress_list_del(c, stats);
 944         return ret;
 945 }
 946
 947 #if 0
 948 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 949                                 struct bkey_s_c k,
 950                                 struct bch_io_opts *io_opts,
 951                                 struct data_opts *data_opts)
 952 {
 953         return DATA_SCRUB;
 954 }
 955 #endif
 956
 957 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 958                                       struct bkey_s_c k,
 959                                       struct bch_io_opts *io_opts,
 960                                       struct data_opts *data_opts)
 961 {
 962         unsigned nr_good = bch2_bkey_durability(c, k);
 963         unsigned replicas = bkey_is_btree_ptr(k.k)
 964                 ? c->opts.metadata_replicas
 965                 : io_opts->data_replicas;
 966
 967         if (!nr_good || nr_good >= replicas)
 968                 return DATA_SKIP;
 969
 970         data_opts->target               = 0;
 971         data_opts->nr_replicas          = 1;
 972         data_opts->btree_insert_flags   = 0;
 973         return DATA_ADD_REPLICAS;
 974 }
 975
 976 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 977                                   struct bkey_s_c k,
 978                                   struct bch_io_opts *io_opts,
 979                                   struct data_opts *data_opts)
 980 {
 981         struct bch_ioctl_data *op = arg;
 982
 983         if (!bch2_bkey_has_device(k, op->migrate.dev))
 984                 return DATA_SKIP;
 985
 986         data_opts->target               = 0;
 987         data_opts->nr_replicas          = 1;
 988         data_opts->btree_insert_flags   = 0;
 989         data_opts->rewrite_dev          = op->migrate.dev;
 990         return DATA_REWRITE;
 991 }
 992
 993 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 994                                             struct btree *b,
 995                                             struct bch_io_opts *io_opts,
 996                                             struct data_opts *data_opts)
 997 {
 998         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 999 }
1000
1001 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
1002                                         struct btree *b,
1003                                         struct bch_io_opts *io_opts,
1004                                         struct data_opts *data_opts)
1005 {
1006         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1007 }
1008
1009 static bool bformat_needs_redo(struct bkey_format *f)
1010 {
1011         unsigned i;
1012
1013         for (i = 0; i < f->nr_fields; i++) {
1014                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
1015                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
1016                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
1017
1018                 if (f->bits_per_field[i] > unpacked_bits)
1019                         return true;
1020
1021                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1022                         return true;
1023
1024                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1025                      unpacked_mask) <
1026                     field_offset)
1027                         return true;
1028         }
1029
1030         return false;
1031 }
1032
1033 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1034                                             struct btree *b,
1035                                             struct bch_io_opts *io_opts,
1036                                             struct data_opts *data_opts)
1037 {
1038         if (b->version_ondisk != c->sb.version ||
1039             btree_node_need_rewrite(b) ||
1040             bformat_needs_redo(&b->format)) {
1041                 data_opts->target               = 0;
1042                 data_opts->nr_replicas          = 1;
1043                 data_opts->btree_insert_flags   = 0;
1044                 return DATA_REWRITE;
1045         }
1046
1047         return DATA_SKIP;
1048 }
1049
1050 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1051 {
1052         int ret;
1053
1054         ret = bch2_move_btree(c,
1055                               0,                POS_MIN,
1056                               BTREE_ID_NR,      SPOS_MAX,
1057                               rewrite_old_nodes_pred, c, stats);
1058         if (!ret) {
1059                 mutex_lock(&c->sb_lock);
1060                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1061                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1062                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1063                 bch2_write_super(c);
1064                 mutex_unlock(&c->sb_lock);
1065         }
1066
1067         return ret;
1068 }
1069
1070 int bch2_data_job(struct bch_fs *c,
1071                   struct bch_move_stats *stats,
1072                   struct bch_ioctl_data op)
1073 {
1074         int ret = 0;
1075
1076         switch (op.op) {
1077         case BCH_DATA_OP_REREPLICATE:
1078                 bch_move_stats_init(stats, "rereplicate");
1079                 stats->data_type = BCH_DATA_journal;
1080                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1081
1082                 ret = bch2_move_btree(c,
1083                                       op.start_btree,   op.start_pos,
1084                                       op.end_btree,     op.end_pos,
1085                                       rereplicate_btree_pred, c, stats) ?: ret;
1086                 ret = bch2_replicas_gc2(c) ?: ret;
1087
1088                 ret = bch2_move_data(c,
1089                                      op.start_btree,    op.start_pos,
1090                                      op.end_btree,      op.end_pos,
1091                                      NULL, writepoint_hashed((unsigned long) current),
1092                                      rereplicate_pred, c, stats) ?: ret;
1093                 ret = bch2_replicas_gc2(c) ?: ret;
1094                 break;
1095         case BCH_DATA_OP_MIGRATE:
1096                 if (op.migrate.dev >= c->sb.nr_devices)
1097                         return -EINVAL;
1098
1099                 bch_move_stats_init(stats, "migrate");
1100                 stats->data_type = BCH_DATA_journal;
1101                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1102
1103                 ret = bch2_move_btree(c,
1104                                       op.start_btree,   op.start_pos,
1105                                       op.end_btree,     op.end_pos,
1106                                       migrate_btree_pred, &op, stats) ?: ret;
1107                 ret = bch2_replicas_gc2(c) ?: ret;
1108
1109                 ret = bch2_move_data(c,
1110                                      op.start_btree,    op.start_pos,
1111                                      op.end_btree,      op.end_pos,
1112                                      NULL, writepoint_hashed((unsigned long) current),
1113                                      migrate_pred, &op, stats) ?: ret;
1114                 ret = bch2_replicas_gc2(c) ?: ret;
1115                 break;
1116         case BCH_DATA_OP_REWRITE_OLD_NODES:
1117                 bch_move_stats_init(stats, "rewrite_old_nodes");
1118                 ret = bch2_scan_old_btree_nodes(c, stats);
1119                 break;
1120         default:
1121                 ret = -EINVAL;
1122         }
1123
1124         return ret;
1125 }