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