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 "inode.h"
  12 #include "io.h"
  13 #include "journal_reclaim.h"
  14 #include "move.h"
  15 #include "replicas.h"
  16 #include "subvolume.h"
  17 #include "super-io.h"
  18 #include "keylist.h"
  19
  20 #include <linux/ioprio.h>
  21 #include <linux/kthread.h>
  22
  23 #include <trace/events/bcachefs.h>
  24
  25 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  26
  27 struct moving_io {
  28         struct list_head        list;
  29         struct closure          cl;
  30         bool                    read_completed;
  31
  32         unsigned                read_sectors;
  33         unsigned                write_sectors;
  34
  35         struct bch_read_bio     rbio;
  36
  37         struct migrate_write    write;
  38         /* Must be last since it is variable size */
  39         struct bio_vec          bi_inline_vecs[0];
  40 };
  41
  42 struct moving_context {
  43         /* Closure for waiting on all reads and writes to complete */
  44         struct closure          cl;
  45
  46         struct bch_move_stats   *stats;
  47
  48         struct list_head        reads;
  49
  50         /* in flight sectors: */
  51         atomic_t                read_sectors;
  52         atomic_t                write_sectors;
  53
  54         wait_queue_head_t       wait;
  55 };
  56
  57 static int insert_snapshot_whiteouts(struct btree_trans *trans,
  58                                      enum btree_id id,
  59                                      struct bpos old_pos,
  60                                      struct bpos new_pos)
  61 {
  62         struct bch_fs *c = trans->c;
  63         struct btree_iter iter, update_iter;
  64         struct bkey_s_c k;
  65         struct snapshots_seen s;
  66         int ret;
  67
  68         if (!btree_type_has_snapshots(id))
  69                 return 0;
  70
  71         snapshots_seen_init(&s);
  72
  73         if (!bkey_cmp(old_pos, new_pos))
  74                 return 0;
  75
  76         if (!snapshot_t(c, old_pos.snapshot)->children[0])
  77                 return 0;
  78
  79         bch2_trans_iter_init(trans, &iter, id, old_pos,
  80                              BTREE_ITER_NOT_EXTENTS|
  81                              BTREE_ITER_ALL_SNAPSHOTS);
  82         while (1) {
  83 next:
  84                 k = bch2_btree_iter_prev(&iter);
  85                 ret = bkey_err(k);
  86                 if (ret)
  87                         break;
  88
  89                 if (bkey_cmp(old_pos, k.k->p))
  90                         break;
  91
  92                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
  93                         struct bkey_i *update;
  94                         size_t i;
  95
  96                         for (i = 0; i < s.nr; i++)
  97                                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, s.d[i]))
  98                                         goto next;
  99
 100                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 101
 102                         ret = PTR_ERR_OR_ZERO(update);
 103                         if (ret)
 104                                 break;
 105
 106                         bkey_init(&update->k);
 107                         update->k.p = new_pos;
 108                         update->k.p.snapshot = k.k->p.snapshot;
 109
 110                         bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
 111                                              BTREE_ITER_NOT_EXTENTS|
 112                                              BTREE_ITER_ALL_SNAPSHOTS|
 113                                              BTREE_ITER_INTENT);
 114                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
 115                                 bch2_trans_update(trans, &update_iter, update,
 116                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 117                         bch2_trans_iter_exit(trans, &update_iter);
 118                         if (ret)
 119                                 break;
 120
 121                         ret = snapshots_seen_add(c, &s, k.k->p.snapshot);
 122                         if (ret)
 123                                 break;
 124                 }
 125         }
 126         bch2_trans_iter_exit(trans, &iter);
 127         kfree(s.d);
 128
 129         return ret;
 130 }
 131
 132 static int bch2_migrate_index_update(struct bch_write_op *op)
 133 {
 134         struct bch_fs *c = op->c;
 135         struct btree_trans trans;
 136         struct btree_iter iter;
 137         struct migrate_write *m =
 138                 container_of(op, struct migrate_write, op);
 139         struct keylist *keys = &op->insert_keys;
 140         struct bkey_buf _new, _insert;
 141         int ret = 0;
 142
 143         bch2_bkey_buf_init(&_new);
 144         bch2_bkey_buf_init(&_insert);
 145         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 146
 147         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 148
 149         bch2_trans_iter_init(&trans, &iter, m->btree_id,
 150                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 151                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 152
 153         while (1) {
 154                 struct bkey_s_c k;
 155                 struct bkey_i *insert;
 156                 struct bkey_i_extent *new;
 157                 const union bch_extent_entry *entry;
 158                 struct extent_ptr_decoded p;
 159                 struct bpos next_pos;
 160                 bool did_work = false;
 161                 bool extending = false, should_check_enospc;
 162                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 163
 164                 bch2_trans_begin(&trans);
 165
 166                 k = bch2_btree_iter_peek_slot(&iter);
 167                 ret = bkey_err(k);
 168                 if (ret)
 169                         goto err;
 170
 171                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 172
 173                 if (bversion_cmp(k.k->version, new->k.version) ||
 174                     !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
 175                         goto nomatch;
 176
 177                 bkey_reassemble(_insert.k, k);
 178                 insert = _insert.k;
 179
 180                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 181                 new = bkey_i_to_extent(_new.k);
 182                 bch2_cut_front(iter.pos, &new->k_i);
 183
 184                 bch2_cut_front(iter.pos,        insert);
 185                 bch2_cut_back(new->k.p,         insert);
 186                 bch2_cut_back(insert->k.p,      &new->k_i);
 187
 188                 if (m->data_cmd == DATA_REWRITE) {
 189                         struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
 190                                 bch2_bkey_has_device(bkey_i_to_s_c(insert),
 191                                                      m->data_opts.rewrite_dev);
 192                         if (!old_ptr)
 193                                 goto nomatch;
 194
 195                         if (old_ptr->cached)
 196                                 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
 197                                         new_ptr->cached = true;
 198
 199                         bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
 200                 }
 201
 202                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 203                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 204                                 /*
 205                                  * raced with another move op? extent already
 206                                  * has a pointer to the device we just wrote
 207                                  * data to
 208                                  */
 209                                 continue;
 210                         }
 211
 212                         bch2_extent_ptr_decoded_append(insert, &p);
 213                         did_work = true;
 214                 }
 215
 216                 if (!did_work)
 217                         goto nomatch;
 218
 219                 bch2_bkey_narrow_crcs(insert,
 220                                 (struct bch_extent_crc_unpacked) { 0 });
 221                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 222                 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
 223                                                op->opts.background_target,
 224                                                op->opts.data_replicas);
 225
 226                 ret = bch2_sum_sector_overwrites(&trans, &iter, insert,
 227                                                  &extending,
 228                                                  &should_check_enospc,
 229                                                  &i_sectors_delta,
 230                                                  &disk_sectors_delta);
 231                 if (ret)
 232                         goto err;
 233
 234                 if (disk_sectors_delta > (s64) op->res.sectors) {
 235                         ret = bch2_disk_reservation_add(c, &op->res,
 236                                                 disk_sectors_delta - op->res.sectors,
 237                                                 !should_check_enospc
 238                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 239                         if (ret)
 240                                 goto out;
 241                 }
 242
 243                 next_pos = insert->k.p;
 244
 245                 ret   = insert_snapshot_whiteouts(&trans, m->btree_id,
 246                                                   k.k->p, insert->k.p) ?:
 247                         bch2_trans_update(&trans, &iter, insert,
 248                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 249                         bch2_trans_commit(&trans, &op->res,
 250                                 op_journal_seq(op),
 251                                 BTREE_INSERT_NOFAIL|
 252                                 m->data_opts.btree_insert_flags);
 253                 if (!ret) {
 254                         bch2_btree_iter_set_pos(&iter, next_pos);
 255                         atomic_long_inc(&c->extent_migrate_done);
 256                 }
 257 err:
 258                 if (ret == -EINTR)
 259                         ret = 0;
 260                 if (ret)
 261                         break;
 262 next:
 263                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 264                         bch2_keylist_pop_front(keys);
 265                         if (bch2_keylist_empty(keys))
 266                                 goto out;
 267                 }
 268                 continue;
 269 nomatch:
 270                 if (m->ctxt) {
 271                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 272                         atomic64_inc(&m->ctxt->stats->keys_raced);
 273                         atomic64_add(k.k->p.offset - iter.pos.offset,
 274                                      &m->ctxt->stats->sectors_raced);
 275                 }
 276                 atomic_long_inc(&c->extent_migrate_raced);
 277                 trace_move_race(&new->k);
 278                 bch2_btree_iter_advance(&iter);
 279                 goto next;
 280         }
 281 out:
 282         bch2_trans_iter_exit(&trans, &iter);
 283         bch2_trans_exit(&trans);
 284         bch2_bkey_buf_exit(&_insert, c);
 285         bch2_bkey_buf_exit(&_new, c);
 286         BUG_ON(ret == -EINTR);
 287         return ret;
 288 }
 289
 290 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 291 {
 292         /* write bio must own pages: */
 293         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 294
 295         m->ptr          = rbio->pick.ptr;
 296         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 297         m->op.devs_have = rbio->devs_have;
 298         m->op.pos       = rbio->data_pos;
 299         m->op.version   = rbio->version;
 300         m->op.crc       = rbio->pick.crc;
 301         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 302
 303         if (m->data_cmd == DATA_REWRITE)
 304                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 305 }
 306
 307 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 308                             struct write_point_specifier wp,
 309                             struct bch_io_opts io_opts,
 310                             enum data_cmd data_cmd,
 311                             struct data_opts data_opts,
 312                             enum btree_id btree_id,
 313                             struct bkey_s_c k)
 314 {
 315         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 316         const union bch_extent_entry *entry;
 317         struct bch_extent_crc_unpacked crc;
 318         struct extent_ptr_decoded p;
 319         int ret;
 320
 321         m->btree_id     = btree_id;
 322         m->data_cmd     = data_cmd;
 323         m->data_opts    = data_opts;
 324         m->nr_ptrs_reserved = 0;
 325
 326         bch2_write_op_init(&m->op, c, io_opts);
 327
 328         if (!bch2_bkey_is_incompressible(k))
 329                 m->op.compression_type =
 330                         bch2_compression_opt_to_type[io_opts.background_compression ?:
 331                                                      io_opts.compression];
 332         else
 333                 m->op.incompressible = true;
 334
 335         m->op.target    = data_opts.target,
 336         m->op.write_point = wp;
 337
 338         /*
 339          * op->csum_type is normally initialized from the fs/file's current
 340          * options - but if an extent is encrypted, we require that it stays
 341          * encrypted:
 342          */
 343         bkey_for_each_crc(k.k, ptrs, crc, entry)
 344                 if (bch2_csum_type_is_encryption(crc.csum_type)) {
 345                         m->op.nonce     = crc.nonce + crc.offset;
 346                         m->op.csum_type = crc.csum_type;
 347                         break;
 348                 }
 349
 350         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 351                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 352                 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
 353         } else {
 354                 /* XXX: this should probably be passed in */
 355                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 356         }
 357
 358         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 359                 BCH_WRITE_PAGES_OWNED|
 360                 BCH_WRITE_DATA_ENCODED|
 361                 BCH_WRITE_FROM_INTERNAL;
 362
 363         m->op.nr_replicas       = data_opts.nr_replicas;
 364         m->op.nr_replicas_required = data_opts.nr_replicas;
 365         m->op.index_update_fn   = bch2_migrate_index_update;
 366
 367         switch (data_cmd) {
 368         case DATA_ADD_REPLICAS: {
 369                 /*
 370                  * DATA_ADD_REPLICAS is used for moving data to a different
 371                  * device in the background, and due to compression the new copy
 372                  * might take up more space than the old copy:
 373                  */
 374 #if 0
 375                 int nr = (int) io_opts.data_replicas -
 376                         bch2_bkey_nr_ptrs_allocated(k);
 377 #endif
 378                 int nr = (int) io_opts.data_replicas;
 379
 380                 if (nr > 0) {
 381                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 382
 383                         ret = bch2_disk_reservation_get(c, &m->op.res,
 384                                         k.k->size, m->op.nr_replicas, 0);
 385                         if (ret)
 386                                 return ret;
 387                 }
 388                 break;
 389         }
 390         case DATA_REWRITE: {
 391                 unsigned compressed_sectors = 0;
 392
 393                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 394                         if (p.ptr.dev == data_opts.rewrite_dev &&
 395                             !p.ptr.cached &&
 396                             crc_is_compressed(p.crc))
 397                                 compressed_sectors += p.crc.compressed_size;
 398
 399                 if (compressed_sectors) {
 400                         ret = bch2_disk_reservation_add(c, &m->op.res,
 401                                         k.k->size * m->op.nr_replicas,
 402                                         BCH_DISK_RESERVATION_NOFAIL);
 403                         if (ret)
 404                                 return ret;
 405                 }
 406                 break;
 407         }
 408         case DATA_PROMOTE:
 409                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 410                 m->op.flags     |= BCH_WRITE_CACHED;
 411                 break;
 412         default:
 413                 BUG();
 414         }
 415
 416         return 0;
 417 }
 418
 419 static void move_free(struct closure *cl)
 420 {
 421         struct moving_io *io = container_of(cl, struct moving_io, cl);
 422         struct moving_context *ctxt = io->write.ctxt;
 423         struct bvec_iter_all iter;
 424         struct bio_vec *bv;
 425
 426         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 427
 428         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 429                 if (bv->bv_page)
 430                         __free_page(bv->bv_page);
 431
 432         wake_up(&ctxt->wait);
 433
 434         kfree(io);
 435 }
 436
 437 static void move_write_done(struct closure *cl)
 438 {
 439         struct moving_io *io = container_of(cl, struct moving_io, cl);
 440
 441         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 442         closure_return_with_destructor(cl, move_free);
 443 }
 444
 445 static void move_write(struct closure *cl)
 446 {
 447         struct moving_io *io = container_of(cl, struct moving_io, cl);
 448
 449         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 450                 closure_return_with_destructor(cl, move_free);
 451                 return;
 452         }
 453
 454         bch2_migrate_read_done(&io->write, &io->rbio);
 455
 456         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 457         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 458         continue_at(cl, move_write_done, NULL);
 459 }
 460
 461 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 462 {
 463         struct moving_io *io =
 464                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 465
 466         return io && io->read_completed ? io : NULL;
 467 }
 468
 469 static void move_read_endio(struct bio *bio)
 470 {
 471         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 472         struct moving_context *ctxt = io->write.ctxt;
 473
 474         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 475         io->read_completed = true;
 476
 477         if (next_pending_write(ctxt))
 478                 wake_up(&ctxt->wait);
 479
 480         closure_put(&ctxt->cl);
 481 }
 482
 483 static void do_pending_writes(struct moving_context *ctxt)
 484 {
 485         struct moving_io *io;
 486
 487         while ((io = next_pending_write(ctxt))) {
 488                 list_del(&io->list);
 489                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 490         }
 491 }
 492
 493 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 494 do {                                                            \
 495         do_pending_writes(_ctxt);                               \
 496                                                                 \
 497         if (_cond)                                              \
 498                 break;                                          \
 499         __wait_event((_ctxt)->wait,                             \
 500                      next_pending_write(_ctxt) || (_cond));     \
 501 } while (1)
 502
 503 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 504 {
 505         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 506
 507         move_ctxt_wait_event(ctxt,
 508                 !atomic_read(&ctxt->write_sectors) ||
 509                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 510 }
 511
 512 static int bch2_move_extent(struct btree_trans *trans,
 513                             struct moving_context *ctxt,
 514                             struct write_point_specifier wp,
 515                             struct bch_io_opts io_opts,
 516                             enum btree_id btree_id,
 517                             struct bkey_s_c k,
 518                             enum data_cmd data_cmd,
 519                             struct data_opts data_opts)
 520 {
 521         struct bch_fs *c = trans->c;
 522         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 523         struct moving_io *io;
 524         const union bch_extent_entry *entry;
 525         struct extent_ptr_decoded p;
 526         unsigned sectors = k.k->size, pages;
 527         int ret = -ENOMEM;
 528
 529         move_ctxt_wait_event(ctxt,
 530                 atomic_read(&ctxt->write_sectors) <
 531                 SECTORS_IN_FLIGHT_PER_DEVICE);
 532
 533         move_ctxt_wait_event(ctxt,
 534                 atomic_read(&ctxt->read_sectors) <
 535                 SECTORS_IN_FLIGHT_PER_DEVICE);
 536
 537         /* write path might have to decompress data: */
 538         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 539                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 540
 541         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 542         io = kzalloc(sizeof(struct moving_io) +
 543                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 544         if (!io)
 545                 goto err;
 546
 547         io->write.ctxt          = ctxt;
 548         io->read_sectors        = k.k->size;
 549         io->write_sectors       = k.k->size;
 550
 551         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 552         bio_set_prio(&io->write.op.wbio.bio,
 553                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 554
 555         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 556                                  GFP_KERNEL))
 557                 goto err_free;
 558
 559         io->rbio.c              = c;
 560         io->rbio.opts           = io_opts;
 561         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 562         io->rbio.bio.bi_vcnt = pages;
 563         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 564         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 565
 566         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 567         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 568         io->rbio.bio.bi_end_io          = move_read_endio;
 569
 570         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 571                                       data_cmd, data_opts, btree_id, k);
 572         if (ret)
 573                 goto err_free_pages;
 574
 575         atomic64_inc(&ctxt->stats->keys_moved);
 576         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 577
 578         trace_move_extent(k.k);
 579
 580         atomic_add(io->read_sectors, &ctxt->read_sectors);
 581         list_add_tail(&io->list, &ctxt->reads);
 582
 583         /*
 584          * dropped by move_read_endio() - guards against use after free of
 585          * ctxt when doing wakeup
 586          */
 587         closure_get(&ctxt->cl);
 588         bch2_read_extent(trans, &io->rbio,
 589                          bkey_start_pos(k.k),
 590                          btree_id, k, 0,
 591                          BCH_READ_NODECODE|
 592                          BCH_READ_LAST_FRAGMENT);
 593         return 0;
 594 err_free_pages:
 595         bio_free_pages(&io->write.op.wbio.bio);
 596 err_free:
 597         kfree(io);
 598 err:
 599         trace_move_alloc_fail(k.k);
 600         return ret;
 601 }
 602
 603 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 604                         struct bch_inode_unpacked *inode)
 605 {
 606         struct btree_iter iter;
 607         struct bkey_s_c k;
 608         int ret;
 609
 610         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 611                              BTREE_ITER_ALL_SNAPSHOTS);
 612         k = bch2_btree_iter_peek(&iter);
 613         ret = bkey_err(k);
 614         if (ret)
 615                 goto err;
 616
 617         if (!k.k || bkey_cmp(k.k->p, pos)) {
 618                 ret = -ENOENT;
 619                 goto err;
 620         }
 621
 622         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 623         if (ret)
 624                 goto err;
 625
 626         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 627         if (ret)
 628                 goto err;
 629 err:
 630         bch2_trans_iter_exit(trans, &iter);
 631         return ret;
 632 }
 633
 634 static int __bch2_move_data(struct bch_fs *c,
 635                 struct moving_context *ctxt,
 636                 struct bch_ratelimit *rate,
 637                 struct write_point_specifier wp,
 638                 struct bpos start,
 639                 struct bpos end,
 640                 move_pred_fn pred, void *arg,
 641                 struct bch_move_stats *stats,
 642                 enum btree_id btree_id)
 643 {
 644         bool kthread = (current->flags & PF_KTHREAD) != 0;
 645         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 646         struct bkey_buf sk;
 647         struct btree_trans trans;
 648         struct btree_iter iter;
 649         struct bkey_s_c k;
 650         struct data_opts data_opts;
 651         enum data_cmd data_cmd;
 652         u64 delay, cur_inum = U64_MAX;
 653         int ret = 0, ret2;
 654
 655         bch2_bkey_buf_init(&sk);
 656         bch2_trans_init(&trans, c, 0, 0);
 657
 658         stats->data_type = BCH_DATA_user;
 659         stats->btree_id = btree_id;
 660         stats->pos      = start;
 661
 662         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 663                              BTREE_ITER_PREFETCH|
 664                              BTREE_ITER_ALL_SNAPSHOTS);
 665
 666         if (rate)
 667                 bch2_ratelimit_reset(rate);
 668
 669         while (1) {
 670                 do {
 671                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 672
 673                         if (delay) {
 674                                 bch2_trans_unlock(&trans);
 675                                 set_current_state(TASK_INTERRUPTIBLE);
 676                         }
 677
 678                         if (kthread && (ret = kthread_should_stop())) {
 679                                 __set_current_state(TASK_RUNNING);
 680                                 goto out;
 681                         }
 682
 683                         if (delay)
 684                                 schedule_timeout(delay);
 685
 686                         if (unlikely(freezing(current))) {
 687                                 bch2_trans_unlock(&trans);
 688                                 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 689                                 try_to_freeze();
 690                         }
 691                 } while (delay);
 692
 693                 bch2_trans_begin(&trans);
 694
 695                 k = bch2_btree_iter_peek(&iter);
 696
 697                 stats->pos = iter.pos;
 698
 699                 if (!k.k)
 700                         break;
 701                 ret = bkey_err(k);
 702                 if (ret)
 703                         break;
 704                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 705                         break;
 706
 707                 if (!bkey_extent_is_direct_data(k.k))
 708                         goto next_nondata;
 709
 710                 if (btree_id == BTREE_ID_extents &&
 711                     cur_inum != k.k->p.inode) {
 712                         struct bch_inode_unpacked inode;
 713
 714                         io_opts = bch2_opts_to_inode_opts(c->opts);
 715
 716                         ret = lookup_inode(&trans,
 717                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 718                                         &inode);
 719                         if (ret == -EINTR)
 720                                 continue;
 721
 722                         if (!ret)
 723                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 724
 725                         cur_inum = k.k->p.inode;
 726                 }
 727
 728                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 729                 case DATA_SKIP:
 730                         goto next;
 731                 case DATA_SCRUB:
 732                         BUG();
 733                 case DATA_ADD_REPLICAS:
 734                 case DATA_REWRITE:
 735                 case DATA_PROMOTE:
 736                         break;
 737                 default:
 738                         BUG();
 739                 }
 740
 741                 /* unlock before doing IO: */
 742                 bch2_bkey_buf_reassemble(&sk, c, k);
 743                 k = bkey_i_to_s_c(sk.k);
 744                 bch2_trans_unlock(&trans);
 745
 746                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 747                                         data_cmd, data_opts);
 748                 if (ret2) {
 749                         if (ret2 == -EINTR) {
 750                                 bch2_trans_begin(&trans);
 751                                 continue;
 752                         }
 753
 754                         if (ret2 == -ENOMEM) {
 755                                 /* memory allocation failure, wait for some IO to finish */
 756                                 bch2_move_ctxt_wait_for_io(ctxt);
 757                                 continue;
 758                         }
 759
 760                         /* XXX signal failure */
 761                         goto next;
 762                 }
 763
 764                 if (rate)
 765                         bch2_ratelimit_increment(rate, k.k->size);
 766 next:
 767                 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 768                              &stats->sectors_seen);
 769 next_nondata:
 770                 bch2_btree_iter_advance(&iter);
 771                 bch2_trans_cond_resched(&trans);
 772         }
 773 out:
 774
 775         bch2_trans_iter_exit(&trans, &iter);
 776         ret = bch2_trans_exit(&trans) ?: ret;
 777         bch2_bkey_buf_exit(&sk, c);
 778
 779         return ret;
 780 }
 781
 782 inline void bch_move_stats_init(struct bch_move_stats *stats, char *name)
 783 {
 784         memset(stats, 0, sizeof(*stats));
 785
 786         scnprintf(stats->name, sizeof(stats->name),
 787                         "%s", name);
 788 }
 789
 790 static inline void progress_list_add(struct bch_fs *c,
 791                                      struct bch_move_stats *stats)
 792 {
 793         mutex_lock(&c->data_progress_lock);
 794         list_add(&stats->list, &c->data_progress_list);
 795         mutex_unlock(&c->data_progress_lock);
 796 }
 797
 798 static inline void progress_list_del(struct bch_fs *c,
 799                                      struct bch_move_stats *stats)
 800 {
 801         mutex_lock(&c->data_progress_lock);
 802         list_del(&stats->list);
 803         mutex_unlock(&c->data_progress_lock);
 804 }
 805
 806 int bch2_move_data(struct bch_fs *c,
 807                    enum btree_id start_btree_id, struct bpos start_pos,
 808                    enum btree_id end_btree_id,   struct bpos end_pos,
 809                    struct bch_ratelimit *rate,
 810                    struct write_point_specifier wp,
 811                    move_pred_fn pred, void *arg,
 812                    struct bch_move_stats *stats)
 813 {
 814         struct moving_context ctxt = { .stats = stats };
 815         enum btree_id id;
 816         int ret;
 817
 818         progress_list_add(c, stats);
 819         closure_init_stack(&ctxt.cl);
 820         INIT_LIST_HEAD(&ctxt.reads);
 821         init_waitqueue_head(&ctxt.wait);
 822
 823         stats->data_type = BCH_DATA_user;
 824
 825         for (id = start_btree_id;
 826              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 827              id++) {
 828                 stats->btree_id = id;
 829
 830                 if (id != BTREE_ID_extents &&
 831                     id != BTREE_ID_reflink)
 832                         continue;
 833
 834                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 835                                        id == start_btree_id ? start_pos : POS_MIN,
 836                                        id == end_btree_id   ? end_pos   : POS_MAX,
 837                                        pred, arg, stats, id);
 838                 if (ret)
 839                         break;
 840         }
 841
 842
 843         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 844         closure_sync(&ctxt.cl);
 845
 846         EBUG_ON(atomic_read(&ctxt.write_sectors));
 847
 848         trace_move_data(c,
 849                         atomic64_read(&stats->sectors_moved),
 850                         atomic64_read(&stats->keys_moved));
 851
 852         progress_list_del(c, stats);
 853         return ret;
 854 }
 855
 856 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 857                                          struct btree *, struct bch_io_opts *,
 858                                          struct data_opts *);
 859
 860 static int bch2_move_btree(struct bch_fs *c,
 861                            enum btree_id start_btree_id, struct bpos start_pos,
 862                            enum btree_id end_btree_id,   struct bpos end_pos,
 863                            move_btree_pred pred, void *arg,
 864                            struct bch_move_stats *stats)
 865 {
 866         bool kthread = (current->flags & PF_KTHREAD) != 0;
 867         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 868         struct btree_trans trans;
 869         struct btree_iter iter;
 870         struct btree *b;
 871         enum btree_id id;
 872         struct data_opts data_opts;
 873         enum data_cmd cmd;
 874         int ret = 0;
 875
 876         bch2_trans_init(&trans, c, 0, 0);
 877         progress_list_add(c, stats);
 878
 879         stats->data_type = BCH_DATA_btree;
 880
 881         for (id = start_btree_id;
 882              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 883              id++) {
 884                 stats->btree_id = id;
 885
 886                 for_each_btree_node(&trans, iter, id,
 887                                     id == start_btree_id ? start_pos : POS_MIN,
 888                                     BTREE_ITER_PREFETCH, b) {
 889                         if (kthread && kthread_should_stop())
 890                                 break;
 891
 892                         if ((cmp_int(id, end_btree_id) ?:
 893                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 894                                 break;
 895
 896                         stats->pos = iter.pos;
 897
 898                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 899                         case DATA_SKIP:
 900                                 goto next;
 901                         case DATA_SCRUB:
 902                                 BUG();
 903                         case DATA_ADD_REPLICAS:
 904                         case DATA_REWRITE:
 905                                 break;
 906                         default:
 907                                 BUG();
 908                         }
 909
 910                         ret = bch2_btree_node_rewrite(&trans, &iter,
 911                                         b->data->keys.seq, 0) ?: ret;
 912 next:
 913                         bch2_trans_cond_resched(&trans);
 914                 }
 915                 bch2_trans_iter_exit(&trans, &iter);
 916
 917                 if (kthread && kthread_should_stop())
 918                         break;
 919         }
 920
 921         bch2_trans_exit(&trans);
 922
 923         if (ret)
 924                 bch_err(c, "error %i in bch2_move_btree", ret);
 925
 926         progress_list_del(c, stats);
 927         return ret;
 928 }
 929
 930 #if 0
 931 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 932                                 struct bkey_s_c k,
 933                                 struct bch_io_opts *io_opts,
 934                                 struct data_opts *data_opts)
 935 {
 936         return DATA_SCRUB;
 937 }
 938 #endif
 939
 940 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 941                                       struct bkey_s_c k,
 942                                       struct bch_io_opts *io_opts,
 943                                       struct data_opts *data_opts)
 944 {
 945         unsigned nr_good = bch2_bkey_durability(c, k);
 946         unsigned replicas = 0;
 947
 948         switch (k.k->type) {
 949         case KEY_TYPE_btree_ptr:
 950                 replicas = c->opts.metadata_replicas;
 951                 break;
 952         case KEY_TYPE_extent:
 953                 replicas = io_opts->data_replicas;
 954                 break;
 955         }
 956
 957         if (!nr_good || nr_good >= replicas)
 958                 return DATA_SKIP;
 959
 960         data_opts->target               = 0;
 961         data_opts->nr_replicas          = 1;
 962         data_opts->btree_insert_flags   = 0;
 963         return DATA_ADD_REPLICAS;
 964 }
 965
 966 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 967                                   struct bkey_s_c k,
 968                                   struct bch_io_opts *io_opts,
 969                                   struct data_opts *data_opts)
 970 {
 971         struct bch_ioctl_data *op = arg;
 972
 973         if (!bch2_bkey_has_device(k, op->migrate.dev))
 974                 return DATA_SKIP;
 975
 976         data_opts->target               = 0;
 977         data_opts->nr_replicas          = 1;
 978         data_opts->btree_insert_flags   = 0;
 979         data_opts->rewrite_dev          = op->migrate.dev;
 980         return DATA_REWRITE;
 981 }
 982
 983 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 984                                             struct btree *b,
 985                                             struct bch_io_opts *io_opts,
 986                                             struct data_opts *data_opts)
 987 {
 988         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 989 }
 990
 991 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
 992                                         struct btree *b,
 993                                         struct bch_io_opts *io_opts,
 994                                         struct data_opts *data_opts)
 995 {
 996         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 997 }
 998
 999 static bool bformat_needs_redo(struct bkey_format *f)
1000 {
1001         unsigned i;
1002
1003         for (i = 0; i < f->nr_fields; i++) {
1004                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
1005                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
1006                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
1007
1008                 if (f->bits_per_field[i] > unpacked_bits)
1009                         return true;
1010
1011                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1012                         return true;
1013
1014                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1015                      unpacked_mask) <
1016                     field_offset)
1017                         return true;
1018         }
1019
1020         return false;
1021 }
1022
1023 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1024                                             struct btree *b,
1025                                             struct bch_io_opts *io_opts,
1026                                             struct data_opts *data_opts)
1027 {
1028         if (b->version_ondisk != c->sb.version ||
1029             btree_node_need_rewrite(b) ||
1030             bformat_needs_redo(&b->format)) {
1031                 data_opts->target               = 0;
1032                 data_opts->nr_replicas          = 1;
1033                 data_opts->btree_insert_flags   = 0;
1034                 return DATA_REWRITE;
1035         }
1036
1037         return DATA_SKIP;
1038 }
1039
1040 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1041 {
1042         int ret;
1043
1044         ret = bch2_move_btree(c,
1045                               0,                POS_MIN,
1046                               BTREE_ID_NR,      SPOS_MAX,
1047                               rewrite_old_nodes_pred, c, stats);
1048         if (!ret) {
1049                 mutex_lock(&c->sb_lock);
1050                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1051                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1052                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1053                 bch2_write_super(c);
1054                 mutex_unlock(&c->sb_lock);
1055         }
1056
1057         return ret;
1058 }
1059
1060 int bch2_data_job(struct bch_fs *c,
1061                   struct bch_move_stats *stats,
1062                   struct bch_ioctl_data op)
1063 {
1064         int ret = 0;
1065
1066         switch (op.op) {
1067         case BCH_DATA_OP_REREPLICATE:
1068                 bch_move_stats_init(stats, "rereplicate");
1069                 stats->data_type = BCH_DATA_journal;
1070                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1071
1072                 ret = bch2_move_btree(c,
1073                                       op.start_btree,   op.start_pos,
1074                                       op.end_btree,     op.end_pos,
1075                                       rereplicate_btree_pred, c, stats) ?: ret;
1076
1077                 closure_wait_event(&c->btree_interior_update_wait,
1078                                    !bch2_btree_interior_updates_nr_pending(c));
1079
1080                 ret = bch2_replicas_gc2(c) ?: ret;
1081
1082                 ret = bch2_move_data(c,
1083                                      op.start_btree,    op.start_pos,
1084                                      op.end_btree,      op.end_pos,
1085                                      NULL, writepoint_hashed((unsigned long) current),
1086                                      rereplicate_pred, c, stats) ?: ret;
1087                 ret = bch2_replicas_gc2(c) ?: ret;
1088                 break;
1089         case BCH_DATA_OP_MIGRATE:
1090                 if (op.migrate.dev >= c->sb.nr_devices)
1091                         return -EINVAL;
1092
1093                 bch_move_stats_init(stats, "migrate");
1094                 stats->data_type = BCH_DATA_journal;
1095                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1096
1097                 ret = bch2_move_btree(c,
1098                                       op.start_btree,   op.start_pos,
1099                                       op.end_btree,     op.end_pos,
1100                                       migrate_btree_pred, &op, stats) ?: ret;
1101                 ret = bch2_replicas_gc2(c) ?: ret;
1102
1103                 ret = bch2_move_data(c,
1104                                      op.start_btree,    op.start_pos,
1105                                      op.end_btree,      op.end_pos,
1106                                      NULL, writepoint_hashed((unsigned long) current),
1107                                      migrate_pred, &op, stats) ?: ret;
1108                 ret = bch2_replicas_gc2(c) ?: ret;
1109                 break;
1110         case BCH_DATA_OP_REWRITE_OLD_NODES:
1111                 bch_move_stats_init(stats, "rewrite_old_nodes");
1112                 ret = bch2_scan_old_btree_nodes(c, stats);
1113                 break;
1114         default:
1115                 ret = -EINVAL;
1116         }
1117
1118         return ret;
1119 }