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
  66         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
  67                                    bkey_start_pos(&bch2_keylist_front(keys)->k),
  68                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
  69
  70         while (1) {
  71                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
  72                 struct bkey_i_extent *insert, *new =
  73                         bkey_i_to_extent(bch2_keylist_front(keys));
  74                 BKEY_PADDED(k) _new, _insert;
  75                 const union bch_extent_entry *entry;
  76                 struct extent_ptr_decoded p;
  77                 bool did_work = false;
  78                 int nr;
  79
  80                 ret = btree_iter_err(k);
  81                 if (ret)
  82                         break;
  83
  84                 if (bversion_cmp(k.k->version, new->k.version) ||
  85                     !bkey_extent_is_data(k.k) ||
  86                     !bch2_extent_matches_ptr(c, bkey_s_c_to_extent(k),
  87                                              m->ptr, m->offset))
  88                         goto nomatch;
  89
  90                 if (m->data_cmd == DATA_REWRITE &&
  91                     !bch2_extent_has_device(bkey_s_c_to_extent(k),
  92                                             m->data_opts.rewrite_dev))
  93                         goto nomatch;
  94
  95                 bkey_reassemble(&_insert.k, k);
  96                 insert = bkey_i_to_extent(&_insert.k);
  97
  98                 bkey_copy(&_new.k, bch2_keylist_front(keys));
  99                 new = bkey_i_to_extent(&_new.k);
 100
 101                 bch2_cut_front(iter->pos, &insert->k_i);
 102                 bch2_cut_back(new->k.p, &insert->k);
 103                 bch2_cut_back(insert->k.p, &new->k);
 104
 105                 if (m->data_cmd == DATA_REWRITE)
 106                         bch2_bkey_drop_device(extent_i_to_s(insert).s,
 107                                               m->data_opts.rewrite_dev);
 108
 109                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 110                         if (bch2_extent_has_device(extent_i_to_s_c(insert), p.ptr.dev)) {
 111                                 /*
 112                                  * raced with another move op? extent already
 113                                  * has a pointer to the device we just wrote
 114                                  * data to
 115                                  */
 116                                 continue;
 117                         }
 118
 119                         bch2_extent_ptr_decoded_append(insert, &p);
 120                         did_work = true;
 121                 }
 122
 123                 if (!did_work)
 124                         goto nomatch;
 125
 126                 bch2_extent_narrow_crcs(insert,
 127                                 (struct bch_extent_crc_unpacked) { 0 });
 128                 bch2_extent_normalize(c, extent_i_to_s(insert).s);
 129                 bch2_extent_mark_replicas_cached(c, extent_i_to_s(insert),
 130                                                  op->opts.background_target,
 131                                                  op->opts.data_replicas);
 132
 133                 /*
 134                  * If we're not fully overwriting @k, and it's compressed, we
 135                  * need a reservation for all the pointers in @insert
 136                  */
 137                 nr = bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&insert->k_i)) -
 138                          m->nr_ptrs_reserved;
 139
 140                 if (insert->k.size < k.k->size &&
 141                     bch2_extent_is_compressed(k) &&
 142                     nr > 0) {
 143                         ret = bch2_disk_reservation_add(c, &op->res,
 144                                         keylist_sectors(keys) * nr, 0);
 145                         if (ret)
 146                                 goto out;
 147
 148                         m->nr_ptrs_reserved += nr;
 149                         goto next;
 150                 }
 151
 152                 bch2_trans_update(&trans,
 153                                 BTREE_INSERT_ENTRY(iter, &insert->k_i));
 154
 155                 ret = bch2_trans_commit(&trans, &op->res,
 156                                 op_journal_seq(op),
 157                                 BTREE_INSERT_ATOMIC|
 158                                 BTREE_INSERT_NOFAIL|
 159                                 BTREE_INSERT_USE_RESERVE|
 160                                 m->data_opts.btree_insert_flags);
 161                 if (!ret)
 162                         atomic_long_inc(&c->extent_migrate_done);
 163                 if (ret == -EINTR)
 164                         ret = 0;
 165                 if (ret)
 166                         break;
 167 next:
 168                 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
 169                         bch2_keylist_pop_front(keys);
 170                         if (bch2_keylist_empty(keys))
 171                                 goto out;
 172                 }
 173
 174                 bch2_cut_front(iter->pos, bch2_keylist_front(keys));
 175                 continue;
 176 nomatch:
 177                 if (m->ctxt)
 178                         atomic64_add(k.k->p.offset - iter->pos.offset,
 179                                      &m->ctxt->stats->sectors_raced);
 180                 atomic_long_inc(&c->extent_migrate_raced);
 181                 trace_move_race(&new->k);
 182                 bch2_btree_iter_next_slot(iter);
 183                 goto next;
 184         }
 185 out:
 186         bch2_trans_exit(&trans);
 187         return ret;
 188 }
 189
 190 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 191 {
 192         /* write bio must own pages: */
 193         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 194
 195         m->ptr          = rbio->pick.ptr;
 196         m->offset       = rbio->pos.offset - rbio->pick.crc.offset;
 197         m->op.devs_have = rbio->devs_have;
 198         m->op.pos       = rbio->pos;
 199         m->op.version   = rbio->version;
 200         m->op.crc       = rbio->pick.crc;
 201         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 202
 203         if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
 204                 m->op.nonce     = m->op.crc.nonce + m->op.crc.offset;
 205                 m->op.csum_type = m->op.crc.csum_type;
 206         }
 207
 208         if (m->data_cmd == DATA_REWRITE)
 209                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 210 }
 211
 212 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 213                             struct write_point_specifier wp,
 214                             struct bch_io_opts io_opts,
 215                             enum data_cmd data_cmd,
 216                             struct data_opts data_opts,
 217                             struct bkey_s_c k)
 218 {
 219         int ret;
 220
 221         m->data_cmd     = data_cmd;
 222         m->data_opts    = data_opts;
 223         m->nr_ptrs_reserved = 0;
 224
 225         bch2_write_op_init(&m->op, c, io_opts);
 226         m->op.compression_type =
 227                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 228                                              io_opts.compression];
 229         m->op.target    = data_opts.target,
 230         m->op.write_point = wp;
 231
 232         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 233                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 234
 235         m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
 236                 BCH_WRITE_PAGES_STABLE|
 237                 BCH_WRITE_PAGES_OWNED|
 238                 BCH_WRITE_DATA_ENCODED;
 239
 240         m->op.nr_replicas       = 1;
 241         m->op.nr_replicas_required = 1;
 242         m->op.index_update_fn   = bch2_migrate_index_update;
 243
 244         switch (data_cmd) {
 245         case DATA_ADD_REPLICAS: {
 246                 /*
 247                  * DATA_ADD_REPLICAS is used for moving data to a different
 248                  * device in the background, and due to compression the new copy
 249                  * might take up more space than the old copy:
 250                  */
 251 #if 0
 252                 int nr = (int) io_opts.data_replicas -
 253                         bch2_bkey_nr_dirty_ptrs(k);
 254 #endif
 255                 int nr = (int) io_opts.data_replicas;
 256
 257                 if (nr > 0) {
 258                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 259
 260                         ret = bch2_disk_reservation_get(c, &m->op.res,
 261                                         k.k->size, m->op.nr_replicas, 0);
 262                         if (ret)
 263                                 return ret;
 264                 }
 265                 break;
 266         }
 267         case DATA_REWRITE: {
 268                 const union bch_extent_entry *entry;
 269                 struct extent_ptr_decoded p;
 270                 unsigned compressed_sectors = 0;
 271
 272                 extent_for_each_ptr_decode(bkey_s_c_to_extent(k), p, entry)
 273                         if (!p.ptr.cached &&
 274                             p.crc.compression_type != BCH_COMPRESSION_NONE &&
 275                             bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
 276                                 compressed_sectors += p.crc.compressed_size;
 277
 278                 if (compressed_sectors) {
 279                         ret = bch2_disk_reservation_add(c, &m->op.res,
 280                                         compressed_sectors,
 281                                         BCH_DISK_RESERVATION_NOFAIL);
 282                         if (ret)
 283                                 return ret;
 284                 }
 285                 break;
 286         }
 287         case DATA_PROMOTE:
 288                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 289                 m->op.flags     |= BCH_WRITE_CACHED;
 290                 break;
 291         default:
 292                 BUG();
 293         }
 294
 295         return 0;
 296 }
 297
 298 static void move_free(struct closure *cl)
 299 {
 300         struct moving_io *io = container_of(cl, struct moving_io, cl);
 301         struct moving_context *ctxt = io->write.ctxt;
 302         struct bio_vec *bv;
 303         int i;
 304
 305         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 306
 307         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, i)
 308                 if (bv->bv_page)
 309                         __free_page(bv->bv_page);
 310
 311         wake_up(&ctxt->wait);
 312
 313         kfree(io);
 314 }
 315
 316 static void move_write_done(struct closure *cl)
 317 {
 318         struct moving_io *io = container_of(cl, struct moving_io, cl);
 319
 320         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 321         closure_return_with_destructor(cl, move_free);
 322 }
 323
 324 static void move_write(struct closure *cl)
 325 {
 326         struct moving_io *io = container_of(cl, struct moving_io, cl);
 327
 328         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 329                 closure_return_with_destructor(cl, move_free);
 330                 return;
 331         }
 332
 333         bch2_migrate_read_done(&io->write, &io->rbio);
 334
 335         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 336         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 337         continue_at(cl, move_write_done, NULL);
 338 }
 339
 340 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 341 {
 342         struct moving_io *io =
 343                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 344
 345         return io && io->read_completed ? io : NULL;
 346 }
 347
 348 static void move_read_endio(struct bio *bio)
 349 {
 350         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 351         struct moving_context *ctxt = io->write.ctxt;
 352
 353         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 354         io->read_completed = true;
 355
 356         if (next_pending_write(ctxt))
 357                 wake_up(&ctxt->wait);
 358
 359         closure_put(&ctxt->cl);
 360 }
 361
 362 static void do_pending_writes(struct moving_context *ctxt)
 363 {
 364         struct moving_io *io;
 365
 366         while ((io = next_pending_write(ctxt))) {
 367                 list_del(&io->list);
 368                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 369         }
 370 }
 371
 372 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 373 do {                                                            \
 374         do_pending_writes(_ctxt);                               \
 375                                                                 \
 376         if (_cond)                                              \
 377                 break;                                          \
 378         __wait_event((_ctxt)->wait,                             \
 379                      next_pending_write(_ctxt) || (_cond));     \
 380 } while (1)
 381
 382 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 383 {
 384         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 385
 386         move_ctxt_wait_event(ctxt,
 387                 !atomic_read(&ctxt->write_sectors) ||
 388                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 389 }
 390
 391 static int bch2_move_extent(struct bch_fs *c,
 392                             struct moving_context *ctxt,
 393                             struct write_point_specifier wp,
 394                             struct bch_io_opts io_opts,
 395                             struct bkey_s_c_extent e,
 396                             enum data_cmd data_cmd,
 397                             struct data_opts data_opts)
 398 {
 399         struct moving_io *io;
 400         const union bch_extent_entry *entry;
 401         struct extent_ptr_decoded p;
 402         unsigned sectors = e.k->size, pages;
 403         int ret = -ENOMEM;
 404
 405         move_ctxt_wait_event(ctxt,
 406                 atomic_read(&ctxt->write_sectors) <
 407                 SECTORS_IN_FLIGHT_PER_DEVICE);
 408
 409         move_ctxt_wait_event(ctxt,
 410                 atomic_read(&ctxt->read_sectors) <
 411                 SECTORS_IN_FLIGHT_PER_DEVICE);
 412
 413         /* write path might have to decompress data: */
 414         extent_for_each_ptr_decode(e, p, entry)
 415                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 416
 417         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 418         io = kzalloc(sizeof(struct moving_io) +
 419                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 420         if (!io)
 421                 goto err;
 422
 423         io->write.ctxt          = ctxt;
 424         io->read_sectors        = e.k->size;
 425         io->write_sectors       = e.k->size;
 426
 427         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 428         bio_set_prio(&io->write.op.wbio.bio,
 429                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 430         io->write.op.wbio.bio.bi_iter.bi_size = sectors << 9;
 431
 432         bch2_bio_map(&io->write.op.wbio.bio, NULL);
 433         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, GFP_KERNEL))
 434                 goto err_free;
 435
 436         io->rbio.opts = io_opts;
 437         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 438         io->rbio.bio.bi_vcnt = pages;
 439         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 440         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 441
 442         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 443         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(e.k);
 444         io->rbio.bio.bi_end_io          = move_read_endio;
 445
 446         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 447                                       data_cmd, data_opts, e.s_c);
 448         if (ret)
 449                 goto err_free_pages;
 450
 451         atomic64_inc(&ctxt->stats->keys_moved);
 452         atomic64_add(e.k->size, &ctxt->stats->sectors_moved);
 453
 454         trace_move_extent(e.k);
 455
 456         atomic_add(io->read_sectors, &ctxt->read_sectors);
 457         list_add_tail(&io->list, &ctxt->reads);
 458
 459         /*
 460          * dropped by move_read_endio() - guards against use after free of
 461          * ctxt when doing wakeup
 462          */
 463         closure_get(&ctxt->cl);
 464         bch2_read_extent(c, &io->rbio, e.s_c,
 465                          BCH_READ_NODECODE|
 466                          BCH_READ_LAST_FRAGMENT);
 467         return 0;
 468 err_free_pages:
 469         bio_free_pages(&io->write.op.wbio.bio);
 470 err_free:
 471         kfree(io);
 472 err:
 473         trace_move_alloc_fail(e.k);
 474         return ret;
 475 }
 476
 477 int bch2_move_data(struct bch_fs *c,
 478                    struct bch_ratelimit *rate,
 479                    struct write_point_specifier wp,
 480                    struct bpos start,
 481                    struct bpos end,
 482                    move_pred_fn pred, void *arg,
 483                    struct bch_move_stats *stats)
 484 {
 485         bool kthread = (current->flags & PF_KTHREAD) != 0;
 486         struct moving_context ctxt = { .stats = stats };
 487         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 488         BKEY_PADDED(k) tmp;
 489         struct bkey_s_c k;
 490         struct data_opts data_opts;
 491         enum data_cmd data_cmd;
 492         u64 delay, cur_inum = U64_MAX;
 493         int ret = 0, ret2;
 494
 495         closure_init_stack(&ctxt.cl);
 496         INIT_LIST_HEAD(&ctxt.reads);
 497         init_waitqueue_head(&ctxt.wait);
 498
 499         stats->data_type = BCH_DATA_USER;
 500         bch2_btree_iter_init(&stats->iter, c, BTREE_ID_EXTENTS, start,
 501                              BTREE_ITER_PREFETCH);
 502
 503         if (rate)
 504                 bch2_ratelimit_reset(rate);
 505
 506         while (1) {
 507                 do {
 508                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 509
 510                         if (delay) {
 511                                 bch2_btree_iter_unlock(&stats->iter);
 512                                 set_current_state(TASK_INTERRUPTIBLE);
 513                         }
 514
 515                         if (kthread && (ret = kthread_should_stop())) {
 516                                 __set_current_state(TASK_RUNNING);
 517                                 goto out;
 518                         }
 519
 520                         if (delay)
 521                                 schedule_timeout(delay);
 522
 523                         if (unlikely(freezing(current))) {
 524                                 bch2_btree_iter_unlock(&stats->iter);
 525                                 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 526                                 try_to_freeze();
 527                         }
 528                 } while (delay);
 529 peek:
 530                 k = bch2_btree_iter_peek(&stats->iter);
 531                 if (!k.k)
 532                         break;
 533                 ret = btree_iter_err(k);
 534                 if (ret)
 535                         break;
 536                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 537                         break;
 538
 539                 if (!bkey_extent_is_data(k.k))
 540                         goto next_nondata;
 541
 542                 if (cur_inum != k.k->p.inode) {
 543                         struct bch_inode_unpacked inode;
 544
 545                         /* don't hold btree locks while looking up inode: */
 546                         bch2_btree_iter_unlock(&stats->iter);
 547
 548                         io_opts = bch2_opts_to_inode_opts(c->opts);
 549                         if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
 550                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 551                         cur_inum = k.k->p.inode;
 552                         goto peek;
 553                 }
 554
 555                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 556                 case DATA_SKIP:
 557                         goto next;
 558                 case DATA_SCRUB:
 559                         BUG();
 560                 case DATA_ADD_REPLICAS:
 561                 case DATA_REWRITE:
 562                 case DATA_PROMOTE:
 563                         break;
 564                 default:
 565                         BUG();
 566                 }
 567
 568                 /* unlock before doing IO: */
 569                 bkey_reassemble(&tmp.k, k);
 570                 k = bkey_i_to_s_c(&tmp.k);
 571                 bch2_btree_iter_unlock(&stats->iter);
 572
 573                 ret2 = bch2_move_extent(c, &ctxt, wp, io_opts,
 574                                         bkey_s_c_to_extent(k),
 575                                         data_cmd, data_opts);
 576                 if (ret2) {
 577                         if (ret2 == -ENOMEM) {
 578                                 /* memory allocation failure, wait for some IO to finish */
 579                                 bch2_move_ctxt_wait_for_io(&ctxt);
 580                                 continue;
 581                         }
 582
 583                         /* XXX signal failure */
 584                         goto next;
 585                 }
 586
 587                 if (rate)
 588                         bch2_ratelimit_increment(rate, k.k->size);
 589 next:
 590                 atomic64_add(k.k->size * bch2_bkey_nr_dirty_ptrs(k),
 591                              &stats->sectors_seen);
 592 next_nondata:
 593                 bch2_btree_iter_next(&stats->iter);
 594                 bch2_btree_iter_cond_resched(&stats->iter);
 595         }
 596 out:
 597         bch2_btree_iter_unlock(&stats->iter);
 598
 599         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 600         closure_sync(&ctxt.cl);
 601
 602         EBUG_ON(atomic_read(&ctxt.write_sectors));
 603
 604         trace_move_data(c,
 605                         atomic64_read(&stats->sectors_moved),
 606                         atomic64_read(&stats->keys_moved));
 607
 608         return ret;
 609 }
 610
 611 static int bch2_gc_data_replicas(struct bch_fs *c)
 612 {
 613         struct btree_iter iter;
 614         struct bkey_s_c k;
 615         int ret;
 616
 617         mutex_lock(&c->replicas_gc_lock);
 618         bch2_replicas_gc_start(c, (1 << BCH_DATA_USER)|(1 << BCH_DATA_CACHED));
 619
 620         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
 621                            BTREE_ITER_PREFETCH, k) {
 622                 ret = bch2_mark_bkey_replicas(c, k);
 623                 if (ret)
 624                         break;
 625         }
 626         ret = bch2_btree_iter_unlock(&iter) ?: ret;
 627
 628         bch2_replicas_gc_end(c, ret);
 629         mutex_unlock(&c->replicas_gc_lock);
 630
 631         return ret;
 632 }
 633
 634 static int bch2_gc_btree_replicas(struct bch_fs *c)
 635 {
 636         struct btree_iter iter;
 637         struct btree *b;
 638         unsigned id;
 639         int ret = 0;
 640
 641         mutex_lock(&c->replicas_gc_lock);
 642         bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
 643
 644         for (id = 0; id < BTREE_ID_NR; id++) {
 645                 for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 646                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 647
 648                         bch2_btree_iter_cond_resched(&iter);
 649                 }
 650
 651                 ret = bch2_btree_iter_unlock(&iter) ?: ret;
 652         }
 653
 654         bch2_replicas_gc_end(c, ret);
 655         mutex_unlock(&c->replicas_gc_lock);
 656
 657         return ret;
 658 }
 659
 660 static int bch2_move_btree(struct bch_fs *c,
 661                            move_pred_fn pred,
 662                            void *arg,
 663                            struct bch_move_stats *stats)
 664 {
 665         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 666         struct btree *b;
 667         unsigned id;
 668         struct data_opts data_opts;
 669         enum data_cmd cmd;
 670         int ret = 0;
 671
 672         stats->data_type = BCH_DATA_BTREE;
 673
 674         for (id = 0; id < BTREE_ID_NR; id++) {
 675                 for_each_btree_node(&stats->iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 676                         switch ((cmd = pred(c, arg,
 677                                             bkey_i_to_s_c(&b->key),
 678                                             &io_opts, &data_opts))) {
 679                         case DATA_SKIP:
 680                                 goto next;
 681                         case DATA_SCRUB:
 682                                 BUG();
 683                         case DATA_ADD_REPLICAS:
 684                         case DATA_REWRITE:
 685                                 break;
 686                         default:
 687                                 BUG();
 688                         }
 689
 690                         ret = bch2_btree_node_rewrite(c, &stats->iter,
 691                                         b->data->keys.seq, 0) ?: ret;
 692 next:
 693                         bch2_btree_iter_cond_resched(&stats->iter);
 694                 }
 695
 696                 ret = bch2_btree_iter_unlock(&stats->iter) ?: ret;
 697         }
 698
 699         return ret;
 700 }
 701
 702 #if 0
 703 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 704                                 struct bkey_s_c k,
 705                                 struct bch_io_opts *io_opts,
 706                                 struct data_opts *data_opts)
 707 {
 708         return DATA_SCRUB;
 709 }
 710 #endif
 711
 712 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 713                                       struct bkey_s_c k,
 714                                       struct bch_io_opts *io_opts,
 715                                       struct data_opts *data_opts)
 716 {
 717         unsigned nr_good = bch2_bkey_durability(c, k);
 718         unsigned replicas = 0;
 719
 720         switch (k.k->type) {
 721         case KEY_TYPE_btree_ptr:
 722                 replicas = c->opts.metadata_replicas;
 723                 break;
 724         case KEY_TYPE_extent:
 725                 replicas = io_opts->data_replicas;
 726                 break;
 727         }
 728
 729         if (!nr_good || nr_good >= replicas)
 730                 return DATA_SKIP;
 731
 732         data_opts->target               = 0;
 733         data_opts->btree_insert_flags   = 0;
 734         return DATA_ADD_REPLICAS;
 735 }
 736
 737 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 738                                   struct bkey_s_c k,
 739                                   struct bch_io_opts *io_opts,
 740                                   struct data_opts *data_opts)
 741 {
 742         struct bch_ioctl_data *op = arg;
 743
 744         if (!bch2_bkey_has_device(k, op->migrate.dev))
 745                 return DATA_SKIP;
 746
 747         data_opts->target               = 0;
 748         data_opts->btree_insert_flags   = 0;
 749         data_opts->rewrite_dev          = op->migrate.dev;
 750         return DATA_REWRITE;
 751 }
 752
 753 int bch2_data_job(struct bch_fs *c,
 754                   struct bch_move_stats *stats,
 755                   struct bch_ioctl_data op)
 756 {
 757         int ret = 0;
 758
 759         switch (op.op) {
 760         case BCH_DATA_OP_REREPLICATE:
 761                 stats->data_type = BCH_DATA_JOURNAL;
 762                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 763
 764                 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
 765
 766                 while (1) {
 767                         closure_wait_event(&c->btree_interior_update_wait,
 768                                            !bch2_btree_interior_updates_nr_pending(c) ||
 769                                            c->btree_roots_dirty);
 770                         if (!bch2_btree_interior_updates_nr_pending(c))
 771                                 break;
 772                         bch2_journal_meta(&c->journal);
 773                 }
 774
 775                 ret = bch2_gc_btree_replicas(c) ?: ret;
 776
 777                 ret = bch2_move_data(c, NULL,
 778                                      writepoint_hashed((unsigned long) current),
 779                                      op.start,
 780                                      op.end,
 781                                      rereplicate_pred, c, stats) ?: ret;
 782                 ret = bch2_gc_data_replicas(c) ?: ret;
 783                 break;
 784         case BCH_DATA_OP_MIGRATE:
 785                 if (op.migrate.dev >= c->sb.nr_devices)
 786                         return -EINVAL;
 787
 788                 stats->data_type = BCH_DATA_JOURNAL;
 789                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
 790
 791                 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
 792                 ret = bch2_gc_btree_replicas(c) ?: ret;
 793
 794                 ret = bch2_move_data(c, NULL,
 795                                      writepoint_hashed((unsigned long) current),
 796                                      op.start,
 797                                      op.end,
 798                                      migrate_pred, &op, stats) ?: ret;
 799                 ret = bch2_gc_data_replicas(c) ?: ret;
 800                 break;
 801         default:
 802                 ret = -EINVAL;
 803         }
 804
 805         return ret;
 806 }