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

   1
   2 #include "bcachefs.h"
   3 #include "alloc_foreground.h"
   4 #include "btree_gc.h"
   5 #include "btree_update.h"
   6 #include "btree_update_interior.h"
   7 #include "buckets.h"
   8 #include "disk_groups.h"
   9 #include "inode.h"
  10 #include "io.h"
  11 #include "journal_reclaim.h"
  12 #include "move.h"
  13 #include "replicas.h"
  14 #include "super-io.h"
  15 #include "keylist.h"
  16
  17 #include <linux/ioprio.h>
  18 #include <linux/kthread.h>
  19
  20 #include <trace/events/bcachefs.h>
  21
  22 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  23
  24 struct moving_io {
  25         struct list_head        list;
  26         struct closure          cl;
  27         bool                    read_completed;
  28
  29         unsigned                read_sectors;
  30         unsigned                write_sectors;
  31
  32         struct bch_read_bio     rbio;
  33
  34         struct migrate_write    write;
  35         /* Must be last since it is variable size */
  36         struct bio_vec          bi_inline_vecs[0];
  37 };
  38
  39 struct moving_context {
  40         /* Closure for waiting on all reads and writes to complete */
  41         struct closure          cl;
  42
  43         struct bch_move_stats   *stats;
  44
  45         struct list_head        reads;
  46
  47         /* in flight sectors: */
  48         atomic_t                read_sectors;
  49         atomic_t                write_sectors;
  50
  51         wait_queue_head_t       wait;
  52 };
  53
  54 static int bch2_migrate_index_update(struct bch_write_op *op)
  55 {
  56         struct bch_fs *c = op->c;
  57         struct btree_trans trans;
  58         struct btree_iter *iter;
  59         struct migrate_write *m =
  60                 container_of(op, struct migrate_write, op);
  61         struct keylist *keys = &op->insert_keys;
  62         int ret = 0;
  63
  64         bch2_trans_init(&trans, c);
  65         bch2_trans_preload_iters(&trans);
  66
  67         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
  68                                    bkey_start_pos(&bch2_keylist_front(keys)->k),
  69                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
  70
  71         while (1) {
  72                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
  73                 struct bkey_i_extent *insert, *new =
  74                         bkey_i_to_extent(bch2_keylist_front(keys));
  75                 BKEY_PADDED(k) _new, _insert;
  76                 const union bch_extent_entry *entry;
  77                 struct extent_ptr_decoded p;
  78                 bool did_work = false;
  79                 int nr;
  80
  81                 ret = bkey_err(k);
  82                 if (ret)
  83                         break;
  84
  85                 if (bversion_cmp(k.k->version, new->k.version) ||
  86                     !bkey_extent_is_data(k.k) ||
  87                     !bch2_extent_matches_ptr(c, bkey_s_c_to_extent(k),
  88                                              m->ptr, m->offset))
  89                         goto nomatch;
  90
  91                 if (m->data_cmd == DATA_REWRITE &&
  92                     !bch2_extent_has_device(bkey_s_c_to_extent(k),
  93                                             m->data_opts.rewrite_dev))
  94                         goto nomatch;
  95
  96                 bkey_reassemble(&_insert.k, k);
  97                 insert = bkey_i_to_extent(&_insert.k);
  98
  99                 bkey_copy(&_new.k, bch2_keylist_front(keys));
 100                 new = bkey_i_to_extent(&_new.k);
 101
 102                 bch2_cut_front(iter->pos, &insert->k_i);
 103                 bch2_cut_back(new->k.p, &insert->k);
 104                 bch2_cut_back(insert->k.p, &new->k);
 105
 106                 if (m->data_cmd == DATA_REWRITE)
 107                         bch2_bkey_drop_device(extent_i_to_s(insert).s,
 108                                               m->data_opts.rewrite_dev);
 109
 110                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 111                         if (bch2_extent_has_device(extent_i_to_s_c(insert), p.ptr.dev)) {
 112                                 /*
 113                                  * raced with another move op? extent already
 114                                  * has a pointer to the device we just wrote
 115                                  * data to
 116                                  */
 117                                 continue;
 118                         }
 119
 120                         bch2_extent_ptr_decoded_append(insert, &p);
 121                         did_work = true;
 122                 }
 123
 124                 if (!did_work)
 125                         goto nomatch;
 126
 127                 bch2_extent_narrow_crcs(insert,
 128                                 (struct bch_extent_crc_unpacked) { 0 });
 129                 bch2_extent_normalize(c, extent_i_to_s(insert).s);
 130                 bch2_extent_mark_replicas_cached(c, extent_i_to_s(insert),
 131                                                  op->opts.background_target,
 132                                                  op->opts.data_replicas);
 133
 134                 /*
 135                  * If we're not fully overwriting @k, and it's compressed, we
 136                  * need a reservation for all the pointers in @insert
 137                  */
 138                 nr = bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&insert->k_i)) -
 139                          m->nr_ptrs_reserved;
 140
 141                 if (insert->k.size < k.k->size &&
 142                     bch2_extent_is_compressed(k) &&
 143                     nr > 0) {
 144                         ret = bch2_disk_reservation_add(c, &op->res,
 145                                         keylist_sectors(keys) * nr, 0);
 146                         if (ret)
 147                                 goto out;
 148
 149                         m->nr_ptrs_reserved += nr;
 150                         goto next;
 151                 }
 152
 153                 bch2_trans_update(&trans,
 154                                 BTREE_INSERT_ENTRY(iter, &insert->k_i));
 155
 156                 ret = bch2_trans_commit(&trans, &op->res,
 157                                 op_journal_seq(op),
 158                                 BTREE_INSERT_ATOMIC|
 159                                 BTREE_INSERT_NOFAIL|
 160                                 BTREE_INSERT_USE_RESERVE|
 161                                 m->data_opts.btree_insert_flags);
 162                 if (!ret)
 163                         atomic_long_inc(&c->extent_migrate_done);
 164                 if (ret == -EINTR)
 165                         ret = 0;
 166                 if (ret)
 167                         break;
 168 next:
 169                 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
 170                         bch2_keylist_pop_front(keys);
 171                         if (bch2_keylist_empty(keys))
 172                                 goto out;
 173                 }
 174
 175                 bch2_cut_front(iter->pos, bch2_keylist_front(keys));
 176                 continue;
 177 nomatch:
 178                 if (m->ctxt)
 179                         atomic64_add(k.k->p.offset - iter->pos.offset,
 180                                      &m->ctxt->stats->sectors_raced);
 181                 atomic_long_inc(&c->extent_migrate_raced);
 182                 trace_move_race(&new->k);
 183                 bch2_btree_iter_next_slot(iter);
 184                 goto next;
 185         }
 186 out:
 187         bch2_trans_exit(&trans);
 188         BUG_ON(ret == -EINTR);
 189         return ret;
 190 }
 191
 192 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 193 {
 194         /* write bio must own pages: */
 195         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 196
 197         m->ptr          = rbio->pick.ptr;
 198         m->offset       = rbio->pos.offset - rbio->pick.crc.offset;
 199         m->op.devs_have = rbio->devs_have;
 200         m->op.pos       = rbio->pos;
 201         m->op.version   = rbio->version;
 202         m->op.crc       = rbio->pick.crc;
 203         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 204
 205         if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
 206                 m->op.nonce     = m->op.crc.nonce + m->op.crc.offset;
 207                 m->op.csum_type = m->op.crc.csum_type;
 208         }
 209
 210         if (m->data_cmd == DATA_REWRITE)
 211                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 212 }
 213
 214 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 215                             struct write_point_specifier wp,
 216                             struct bch_io_opts io_opts,
 217                             enum data_cmd data_cmd,
 218                             struct data_opts data_opts,
 219                             struct bkey_s_c k)
 220 {
 221         int ret;
 222
 223         m->data_cmd     = data_cmd;
 224         m->data_opts    = data_opts;
 225         m->nr_ptrs_reserved = 0;
 226
 227         bch2_write_op_init(&m->op, c, io_opts);
 228         m->op.compression_type =
 229                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 230                                              io_opts.compression];
 231         m->op.target    = data_opts.target,
 232         m->op.write_point = wp;
 233
 234         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 235                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 236
 237         m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
 238                 BCH_WRITE_PAGES_STABLE|
 239                 BCH_WRITE_PAGES_OWNED|
 240                 BCH_WRITE_DATA_ENCODED;
 241
 242         m->op.nr_replicas       = 1;
 243         m->op.nr_replicas_required = 1;
 244         m->op.index_update_fn   = bch2_migrate_index_update;
 245
 246         switch (data_cmd) {
 247         case DATA_ADD_REPLICAS: {
 248                 /*
 249                  * DATA_ADD_REPLICAS is used for moving data to a different
 250                  * device in the background, and due to compression the new copy
 251                  * might take up more space than the old copy:
 252                  */
 253 #if 0
 254                 int nr = (int) io_opts.data_replicas -
 255                         bch2_bkey_nr_dirty_ptrs(k);
 256 #endif
 257                 int nr = (int) io_opts.data_replicas;
 258
 259                 if (nr > 0) {
 260                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 261
 262                         ret = bch2_disk_reservation_get(c, &m->op.res,
 263                                         k.k->size, m->op.nr_replicas, 0);
 264                         if (ret)
 265                                 return ret;
 266                 }
 267                 break;
 268         }
 269         case DATA_REWRITE: {
 270                 const union bch_extent_entry *entry;
 271                 struct extent_ptr_decoded p;
 272                 unsigned compressed_sectors = 0;
 273
 274                 extent_for_each_ptr_decode(bkey_s_c_to_extent(k), p, entry)
 275                         if (!p.ptr.cached &&
 276                             p.crc.compression_type != BCH_COMPRESSION_NONE &&
 277                             bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
 278                                 compressed_sectors += p.crc.compressed_size;
 279
 280                 if (compressed_sectors) {
 281                         ret = bch2_disk_reservation_add(c, &m->op.res,
 282                                         compressed_sectors,
 283                                         BCH_DISK_RESERVATION_NOFAIL);
 284                         if (ret)
 285                                 return ret;
 286                 }
 287                 break;
 288         }
 289         case DATA_PROMOTE:
 290                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 291                 m->op.flags     |= BCH_WRITE_CACHED;
 292                 break;
 293         default:
 294                 BUG();
 295         }
 296
 297         return 0;
 298 }
 299
 300 static void move_free(struct closure *cl)
 301 {
 302         struct moving_io *io = container_of(cl, struct moving_io, cl);
 303         struct moving_context *ctxt = io->write.ctxt;
 304         struct bio_vec *bv;
 305         int i;
 306
 307         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 308
 309         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, i)
 310                 if (bv->bv_page)
 311                         __free_page(bv->bv_page);
 312
 313         wake_up(&ctxt->wait);
 314
 315         kfree(io);
 316 }
 317
 318 static void move_write_done(struct closure *cl)
 319 {
 320         struct moving_io *io = container_of(cl, struct moving_io, cl);
 321
 322         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 323         closure_return_with_destructor(cl, move_free);
 324 }
 325
 326 static void move_write(struct closure *cl)
 327 {
 328         struct moving_io *io = container_of(cl, struct moving_io, cl);
 329
 330         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 331                 closure_return_with_destructor(cl, move_free);
 332                 return;
 333         }
 334
 335         bch2_migrate_read_done(&io->write, &io->rbio);
 336
 337         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 338         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 339         continue_at(cl, move_write_done, NULL);
 340 }
 341
 342 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 343 {
 344         struct moving_io *io =
 345                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 346
 347         return io && io->read_completed ? io : NULL;
 348 }
 349
 350 static void move_read_endio(struct bio *bio)
 351 {
 352         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 353         struct moving_context *ctxt = io->write.ctxt;
 354
 355         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 356         io->read_completed = true;
 357
 358         if (next_pending_write(ctxt))
 359                 wake_up(&ctxt->wait);
 360
 361         closure_put(&ctxt->cl);
 362 }
 363
 364 static void do_pending_writes(struct moving_context *ctxt)
 365 {
 366         struct moving_io *io;
 367
 368         while ((io = next_pending_write(ctxt))) {
 369                 list_del(&io->list);
 370                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 371         }
 372 }
 373
 374 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 375 do {                                                            \
 376         do_pending_writes(_ctxt);                               \
 377                                                                 \
 378         if (_cond)                                              \
 379                 break;                                          \
 380         __wait_event((_ctxt)->wait,                             \
 381                      next_pending_write(_ctxt) || (_cond));     \
 382 } while (1)
 383
 384 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 385 {
 386         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 387
 388         move_ctxt_wait_event(ctxt,
 389                 !atomic_read(&ctxt->write_sectors) ||
 390                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 391 }
 392
 393 static int bch2_move_extent(struct bch_fs *c,
 394                             struct moving_context *ctxt,
 395                             struct write_point_specifier wp,
 396                             struct bch_io_opts io_opts,
 397                             struct bkey_s_c_extent e,
 398                             enum data_cmd data_cmd,
 399                             struct data_opts data_opts)
 400 {
 401         struct moving_io *io;
 402         const union bch_extent_entry *entry;
 403         struct extent_ptr_decoded p;
 404         unsigned sectors = e.k->size, pages;
 405         int ret = -ENOMEM;
 406
 407         move_ctxt_wait_event(ctxt,
 408                 atomic_read(&ctxt->write_sectors) <
 409                 SECTORS_IN_FLIGHT_PER_DEVICE);
 410
 411         move_ctxt_wait_event(ctxt,
 412                 atomic_read(&ctxt->read_sectors) <
 413                 SECTORS_IN_FLIGHT_PER_DEVICE);
 414
 415         /* write path might have to decompress data: */
 416         extent_for_each_ptr_decode(e, p, entry)
 417                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 418
 419         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 420         io = kzalloc(sizeof(struct moving_io) +
 421                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 422         if (!io)
 423                 goto err;
 424
 425         io->write.ctxt          = ctxt;
 426         io->read_sectors        = e.k->size;
 427         io->write_sectors       = e.k->size;
 428
 429         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 430         bio_set_prio(&io->write.op.wbio.bio,
 431                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 432         io->write.op.wbio.bio.bi_iter.bi_size = sectors << 9;
 433
 434         bch2_bio_map(&io->write.op.wbio.bio, NULL);
 435         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, GFP_KERNEL))
 436                 goto err_free;
 437
 438         io->rbio.opts = io_opts;
 439         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 440         io->rbio.bio.bi_vcnt = pages;
 441         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 442         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 443
 444         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 445         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(e.k);
 446         io->rbio.bio.bi_end_io          = move_read_endio;
 447
 448         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 449                                       data_cmd, data_opts, e.s_c);
 450         if (ret)
 451                 goto err_free_pages;
 452
 453         atomic64_inc(&ctxt->stats->keys_moved);
 454         atomic64_add(e.k->size, &ctxt->stats->sectors_moved);
 455
 456         trace_move_extent(e.k);
 457
 458         atomic_add(io->read_sectors, &ctxt->read_sectors);
 459         list_add_tail(&io->list, &ctxt->reads);
 460
 461         /*
 462          * dropped by move_read_endio() - guards against use after free of
 463          * ctxt when doing wakeup
 464          */
 465         closure_get(&ctxt->cl);
 466         bch2_read_extent(c, &io->rbio, e.s_c,
 467                          BCH_READ_NODECODE|
 468                          BCH_READ_LAST_FRAGMENT);
 469         return 0;
 470 err_free_pages:
 471         bio_free_pages(&io->write.op.wbio.bio);
 472 err_free:
 473         kfree(io);
 474 err:
 475         trace_move_alloc_fail(e.k);
 476         return ret;
 477 }
 478
 479 int bch2_move_data(struct bch_fs *c,
 480                    struct bch_ratelimit *rate,
 481                    struct write_point_specifier wp,
 482                    struct bpos start,
 483                    struct bpos end,
 484                    move_pred_fn pred, void *arg,
 485                    struct bch_move_stats *stats)
 486 {
 487         bool kthread = (current->flags & PF_KTHREAD) != 0;
 488         struct moving_context ctxt = { .stats = stats };
 489         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 490         BKEY_PADDED(k) tmp;
 491         struct btree_trans trans;
 492         struct btree_iter *iter;
 493         struct bkey_s_c k;
 494         struct data_opts data_opts;
 495         enum data_cmd data_cmd;
 496         u64 delay, cur_inum = U64_MAX;
 497         int ret = 0, ret2;
 498
 499         closure_init_stack(&ctxt.cl);
 500         INIT_LIST_HEAD(&ctxt.reads);
 501         init_waitqueue_head(&ctxt.wait);
 502
 503         bch2_trans_init(&trans, c);
 504
 505         stats->data_type = BCH_DATA_USER;
 506         stats->btree_id = BTREE_ID_EXTENTS;
 507         stats->pos      = POS_MIN;
 508
 509         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, start,
 510                                    BTREE_ITER_PREFETCH);
 511
 512         if (rate)
 513                 bch2_ratelimit_reset(rate);
 514
 515         while (1) {
 516                 do {
 517                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 518
 519                         if (delay) {
 520                                 bch2_trans_unlock(&trans);
 521                                 set_current_state(TASK_INTERRUPTIBLE);
 522                         }
 523
 524                         if (kthread && (ret = kthread_should_stop())) {
 525                                 __set_current_state(TASK_RUNNING);
 526                                 goto out;
 527                         }
 528
 529                         if (delay)
 530                                 schedule_timeout(delay);
 531
 532                         if (unlikely(freezing(current))) {
 533                                 bch2_trans_unlock(&trans);
 534                                 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 535                                 try_to_freeze();
 536                         }
 537                 } while (delay);
 538 peek:
 539                 k = bch2_btree_iter_peek(iter);
 540
 541                 stats->pos = iter->pos;
 542
 543                 if (!k.k)
 544                         break;
 545                 ret = bkey_err(k);
 546                 if (ret)
 547                         break;
 548                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 549                         break;
 550
 551                 if (!bkey_extent_is_data(k.k))
 552                         goto next_nondata;
 553
 554                 if (cur_inum != k.k->p.inode) {
 555                         struct bch_inode_unpacked inode;
 556
 557                         /* don't hold btree locks while looking up inode: */
 558                         bch2_trans_unlock(&trans);
 559
 560                         io_opts = bch2_opts_to_inode_opts(c->opts);
 561                         if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
 562                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 563                         cur_inum = k.k->p.inode;
 564                         goto peek;
 565                 }
 566
 567                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 568                 case DATA_SKIP:
 569                         goto next;
 570                 case DATA_SCRUB:
 571                         BUG();
 572                 case DATA_ADD_REPLICAS:
 573                 case DATA_REWRITE:
 574                 case DATA_PROMOTE:
 575                         break;
 576                 default:
 577                         BUG();
 578                 }
 579
 580                 /* unlock before doing IO: */
 581                 bkey_reassemble(&tmp.k, k);
 582                 k = bkey_i_to_s_c(&tmp.k);
 583                 bch2_trans_unlock(&trans);
 584
 585                 ret2 = bch2_move_extent(c, &ctxt, wp, io_opts,
 586                                         bkey_s_c_to_extent(k),
 587                                         data_cmd, data_opts);
 588                 if (ret2) {
 589                         if (ret2 == -ENOMEM) {
 590                                 /* memory allocation failure, wait for some IO to finish */
 591                                 bch2_move_ctxt_wait_for_io(&ctxt);
 592                                 continue;
 593                         }
 594
 595                         /* XXX signal failure */
 596                         goto next;
 597                 }
 598
 599                 if (rate)
 600                         bch2_ratelimit_increment(rate, k.k->size);
 601 next:
 602                 atomic64_add(k.k->size * bch2_bkey_nr_dirty_ptrs(k),
 603                              &stats->sectors_seen);
 604 next_nondata:
 605                 bch2_btree_iter_next(iter);
 606                 bch2_trans_cond_resched(&trans);
 607         }
 608 out:
 609         bch2_trans_exit(&trans);
 610
 611         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 612         closure_sync(&ctxt.cl);
 613
 614         EBUG_ON(atomic_read(&ctxt.write_sectors));
 615
 616         trace_move_data(c,
 617                         atomic64_read(&stats->sectors_moved),
 618                         atomic64_read(&stats->keys_moved));
 619
 620         return ret;
 621 }
 622
 623 static int bch2_move_btree(struct bch_fs *c,
 624                            move_pred_fn pred,
 625                            void *arg,
 626                            struct bch_move_stats *stats)
 627 {
 628         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 629         struct btree_trans trans;
 630         struct btree_iter *iter;
 631         struct btree *b;
 632         unsigned id;
 633         struct data_opts data_opts;
 634         enum data_cmd cmd;
 635         int ret = 0;
 636
 637         bch2_trans_init(&trans, c);
 638
 639         stats->data_type = BCH_DATA_BTREE;
 640
 641         for (id = 0; id < BTREE_ID_NR; id++) {
 642                 stats->btree_id = id;
 643
 644                 for_each_btree_node(&trans, iter, id, POS_MIN,
 645                                     BTREE_ITER_PREFETCH, b) {
 646                         stats->pos = iter->pos;
 647
 648                         switch ((cmd = pred(c, arg,
 649                                             bkey_i_to_s_c(&b->key),
 650                                             &io_opts, &data_opts))) {
 651                         case DATA_SKIP:
 652                                 goto next;
 653                         case DATA_SCRUB:
 654                                 BUG();
 655                         case DATA_ADD_REPLICAS:
 656                         case DATA_REWRITE:
 657                                 break;
 658                         default:
 659                                 BUG();
 660                         }
 661
 662                         ret = bch2_btree_node_rewrite(c, iter,
 663                                         b->data->keys.seq, 0) ?: ret;
 664 next:
 665                         bch2_trans_cond_resched(&trans);
 666                 }
 667
 668                 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
 669         }
 670
 671         bch2_trans_exit(&trans);
 672
 673         return ret;
 674 }
 675
 676 #if 0
 677 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 678                                 struct bkey_s_c k,
 679                                 struct bch_io_opts *io_opts,
 680                                 struct data_opts *data_opts)
 681 {
 682         return DATA_SCRUB;
 683 }
 684 #endif
 685
 686 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 687                                       struct bkey_s_c k,
 688                                       struct bch_io_opts *io_opts,
 689                                       struct data_opts *data_opts)
 690 {
 691         unsigned nr_good = bch2_bkey_durability(c, k);
 692         unsigned replicas = 0;
 693
 694         switch (k.k->type) {
 695         case KEY_TYPE_btree_ptr:
 696                 replicas = c->opts.metadata_replicas;
 697                 break;
 698         case KEY_TYPE_extent:
 699                 replicas = io_opts->data_replicas;
 700                 break;
 701         }
 702
 703         if (!nr_good || nr_good >= replicas)
 704                 return DATA_SKIP;
 705
 706         data_opts->target               = 0;
 707         data_opts->btree_insert_flags   = 0;
 708         return DATA_ADD_REPLICAS;
 709 }
 710
 711 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 712                                   struct bkey_s_c k,
 713                                   struct bch_io_opts *io_opts,
 714                                   struct data_opts *data_opts)
 715 {
 716         struct bch_ioctl_data *op = arg;
 717
 718         if (!bch2_bkey_has_device(k, op->migrate.dev))
 719                 return DATA_SKIP;
 720
 721         data_opts->target               = 0;
 722         data_opts->btree_insert_flags   = 0;
 723         data_opts->rewrite_dev          = op->migrate.dev;
 724         return DATA_REWRITE;
 725 }
 726
 727 int bch2_data_job(struct bch_fs *c,
 728                   struct bch_move_stats *stats,
 729                   struct bch_ioctl_data op)
 730 {
 731         int ret = 0;
 732
 733         switch (op.op) {
 734         case BCH_DATA_OP_REREPLICATE:
 735                 stats->data_type = BCH_DATA_JOURNAL;
 736                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 737
 738                 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
 739
 740                 while (1) {
 741                         closure_wait_event(&c->btree_interior_update_wait,
 742                                            !bch2_btree_interior_updates_nr_pending(c) ||
 743                                            c->btree_roots_dirty);
 744                         if (!bch2_btree_interior_updates_nr_pending(c))
 745                                 break;
 746                         bch2_journal_meta(&c->journal);
 747                 }
 748
 749                 ret = bch2_replicas_gc2(c) ?: ret;
 750
 751                 ret = bch2_move_data(c, NULL,
 752                                      writepoint_hashed((unsigned long) current),
 753                                      op.start,
 754                                      op.end,
 755                                      rereplicate_pred, c, stats) ?: ret;
 756                 ret = bch2_replicas_gc2(c) ?: ret;
 757                 break;
 758         case BCH_DATA_OP_MIGRATE:
 759                 if (op.migrate.dev >= c->sb.nr_devices)
 760                         return -EINVAL;
 761
 762                 stats->data_type = BCH_DATA_JOURNAL;
 763                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
 764
 765                 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
 766                 ret = bch2_replicas_gc2(c) ?: ret;
 767
 768                 ret = bch2_move_data(c, NULL,
 769                                      writepoint_hashed((unsigned long) current),
 770                                      op.start,
 771                                      op.end,
 772                                      migrate_pred, &op, stats) ?: ret;
 773                 ret = bch2_replicas_gc2(c) ?: ret;
 774                 break;
 775         default:
 776                 ret = -EINVAL;
 777         }
 778
 779         return ret;
 780 }