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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_buf.h"
   6 #include "btree_gc.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "buckets.h"
  10 #include "disk_groups.h"
  11 #include "inode.h"
  12 #include "io.h"
  13 #include "journal_reclaim.h"
  14 #include "move.h"
  15 #include "replicas.h"
  16 #include "super-io.h"
  17 #include "keylist.h"
  18
  19 #include <linux/ioprio.h>
  20 #include <linux/kthread.h>
  21
  22 #include <trace/events/bcachefs.h>
  23
  24 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  25
  26 struct moving_io {
  27         struct list_head        list;
  28         struct closure          cl;
  29         bool                    read_completed;
  30
  31         unsigned                read_sectors;
  32         unsigned                write_sectors;
  33
  34         struct bch_read_bio     rbio;
  35
  36         struct migrate_write    write;
  37         /* Must be last since it is variable size */
  38         struct bio_vec          bi_inline_vecs[0];
  39 };
  40
  41 struct moving_context {
  42         /* Closure for waiting on all reads and writes to complete */
  43         struct closure          cl;
  44
  45         struct bch_move_stats   *stats;
  46
  47         struct list_head        reads;
  48
  49         /* in flight sectors: */
  50         atomic_t                read_sectors;
  51         atomic_t                write_sectors;
  52
  53         wait_queue_head_t       wait;
  54 };
  55
  56 static int bch2_migrate_index_update(struct bch_write_op *op)
  57 {
  58         struct bch_fs *c = op->c;
  59         struct btree_trans trans;
  60         struct btree_iter *iter;
  61         struct migrate_write *m =
  62                 container_of(op, struct migrate_write, op);
  63         struct keylist *keys = &op->insert_keys;
  64         struct bkey_buf _new, _insert;
  65         int ret = 0;
  66
  67         bch2_bkey_buf_init(&_new);
  68         bch2_bkey_buf_init(&_insert);
  69         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
  70
  71         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
  72
  73         iter = bch2_trans_get_iter(&trans, m->btree_id,
  74                                    bkey_start_pos(&bch2_keylist_front(keys)->k),
  75                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
  76
  77         while (1) {
  78                 struct bkey_s_c k;
  79                 struct bkey_i *insert;
  80                 struct bkey_i_extent *new;
  81                 const union bch_extent_entry *entry;
  82                 struct extent_ptr_decoded p;
  83                 bool did_work = false;
  84                 bool extending = false, should_check_enospc;
  85                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
  86
  87                 bch2_trans_reset(&trans, 0);
  88
  89                 k = bch2_btree_iter_peek_slot(iter);
  90                 ret = bkey_err(k);
  91                 if (ret)
  92                         goto err;
  93
  94                 new = bkey_i_to_extent(bch2_keylist_front(keys));
  95
  96                 if (bversion_cmp(k.k->version, new->k.version) ||
  97                     !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
  98                         goto nomatch;
  99
 100                 bkey_reassemble(_insert.k, k);
 101                 insert = _insert.k;
 102
 103                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 104                 new = bkey_i_to_extent(_new.k);
 105                 bch2_cut_front(iter->pos, &new->k_i);
 106
 107                 bch2_cut_front(iter->pos,       insert);
 108                 bch2_cut_back(new->k.p,         insert);
 109                 bch2_cut_back(insert->k.p,      &new->k_i);
 110
 111                 if (m->data_cmd == DATA_REWRITE) {
 112                         struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
 113                                 bch2_bkey_has_device(bkey_i_to_s_c(insert),
 114                                                      m->data_opts.rewrite_dev);
 115                         if (!old_ptr)
 116                                 goto nomatch;
 117
 118                         if (old_ptr->cached)
 119                                 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
 120                                         new_ptr->cached = true;
 121
 122                         bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
 123                 }
 124
 125                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 126                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 127                                 /*
 128                                  * raced with another move op? extent already
 129                                  * has a pointer to the device we just wrote
 130                                  * data to
 131                                  */
 132                                 continue;
 133                         }
 134
 135                         bch2_extent_ptr_decoded_append(insert, &p);
 136                         did_work = true;
 137                 }
 138
 139                 if (!did_work)
 140                         goto nomatch;
 141
 142                 bch2_bkey_narrow_crcs(insert,
 143                                 (struct bch_extent_crc_unpacked) { 0 });
 144                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 145                 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
 146                                                op->opts.background_target,
 147                                                op->opts.data_replicas);
 148
 149                 ret = bch2_sum_sector_overwrites(&trans, iter, insert,
 150                                                  &extending,
 151                                                  &should_check_enospc,
 152                                                  &i_sectors_delta,
 153                                                  &disk_sectors_delta);
 154                 if (ret)
 155                         goto err;
 156
 157                 if (disk_sectors_delta > (s64) op->res.sectors) {
 158                         ret = bch2_disk_reservation_add(c, &op->res,
 159                                                 disk_sectors_delta - op->res.sectors,
 160                                                 !should_check_enospc
 161                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 162                         if (ret)
 163                                 goto out;
 164                 }
 165
 166                 bch2_trans_update(&trans, iter, insert, 0);
 167
 168                 ret = bch2_trans_commit(&trans, &op->res,
 169                                 op_journal_seq(op),
 170                                 BTREE_INSERT_NOFAIL|
 171                                 m->data_opts.btree_insert_flags);
 172 err:
 173                 if (!ret)
 174                         atomic_long_inc(&c->extent_migrate_done);
 175                 if (ret == -EINTR)
 176                         ret = 0;
 177                 if (ret)
 178                         break;
 179 next:
 180                 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
 181                         bch2_keylist_pop_front(keys);
 182                         if (bch2_keylist_empty(keys))
 183                                 goto out;
 184                 }
 185                 continue;
 186 nomatch:
 187                 if (m->ctxt) {
 188                         BUG_ON(k.k->p.offset <= iter->pos.offset);
 189                         atomic64_inc(&m->ctxt->stats->keys_raced);
 190                         atomic64_add(k.k->p.offset - iter->pos.offset,
 191                                      &m->ctxt->stats->sectors_raced);
 192                 }
 193                 atomic_long_inc(&c->extent_migrate_raced);
 194                 trace_move_race(&new->k);
 195                 bch2_btree_iter_next_slot(iter);
 196                 goto next;
 197         }
 198 out:
 199         bch2_trans_iter_put(&trans, iter);
 200         bch2_trans_exit(&trans);
 201         bch2_bkey_buf_exit(&_insert, c);
 202         bch2_bkey_buf_exit(&_new, c);
 203         BUG_ON(ret == -EINTR);
 204         return ret;
 205 }
 206
 207 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 208 {
 209         /* write bio must own pages: */
 210         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 211
 212         m->ptr          = rbio->pick.ptr;
 213         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 214         m->op.devs_have = rbio->devs_have;
 215         m->op.pos       = rbio->data_pos;
 216         m->op.version   = rbio->version;
 217         m->op.crc       = rbio->pick.crc;
 218         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 219
 220         if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
 221                 m->op.nonce     = m->op.crc.nonce + m->op.crc.offset;
 222                 m->op.csum_type = m->op.crc.csum_type;
 223         }
 224
 225         if (m->data_cmd == DATA_REWRITE)
 226                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 227 }
 228
 229 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 230                             struct write_point_specifier wp,
 231                             struct bch_io_opts io_opts,
 232                             enum data_cmd data_cmd,
 233                             struct data_opts data_opts,
 234                             enum btree_id btree_id,
 235                             struct bkey_s_c k)
 236 {
 237         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 238         const union bch_extent_entry *entry;
 239         struct extent_ptr_decoded p;
 240         int ret;
 241
 242         m->btree_id     = btree_id;
 243         m->data_cmd     = data_cmd;
 244         m->data_opts    = data_opts;
 245         m->nr_ptrs_reserved = 0;
 246
 247         bch2_write_op_init(&m->op, c, io_opts);
 248
 249         if (!bch2_bkey_is_incompressible(k))
 250                 m->op.compression_type =
 251                         bch2_compression_opt_to_type[io_opts.background_compression ?:
 252                                                      io_opts.compression];
 253         else
 254                 m->op.incompressible = true;
 255
 256         m->op.target    = data_opts.target,
 257         m->op.write_point = wp;
 258
 259         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 260                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 261                 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
 262         } else {
 263                 /* XXX: this should probably be passed in */
 264                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 265         }
 266
 267         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 268                 BCH_WRITE_PAGES_OWNED|
 269                 BCH_WRITE_DATA_ENCODED|
 270                 BCH_WRITE_FROM_INTERNAL;
 271
 272         m->op.nr_replicas       = data_opts.nr_replicas;
 273         m->op.nr_replicas_required = data_opts.nr_replicas;
 274         m->op.index_update_fn   = bch2_migrate_index_update;
 275
 276         switch (data_cmd) {
 277         case DATA_ADD_REPLICAS: {
 278                 /*
 279                  * DATA_ADD_REPLICAS is used for moving data to a different
 280                  * device in the background, and due to compression the new copy
 281                  * might take up more space than the old copy:
 282                  */
 283 #if 0
 284                 int nr = (int) io_opts.data_replicas -
 285                         bch2_bkey_nr_ptrs_allocated(k);
 286 #endif
 287                 int nr = (int) io_opts.data_replicas;
 288
 289                 if (nr > 0) {
 290                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 291
 292                         ret = bch2_disk_reservation_get(c, &m->op.res,
 293                                         k.k->size, m->op.nr_replicas, 0);
 294                         if (ret)
 295                                 return ret;
 296                 }
 297                 break;
 298         }
 299         case DATA_REWRITE: {
 300                 unsigned compressed_sectors = 0;
 301
 302                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 303                         if (p.ptr.dev == data_opts.rewrite_dev &&
 304                             !p.ptr.cached &&
 305                             crc_is_compressed(p.crc))
 306                                 compressed_sectors += p.crc.compressed_size;
 307
 308                 if (compressed_sectors) {
 309                         ret = bch2_disk_reservation_add(c, &m->op.res,
 310                                         k.k->size * m->op.nr_replicas,
 311                                         BCH_DISK_RESERVATION_NOFAIL);
 312                         if (ret)
 313                                 return ret;
 314                 }
 315                 break;
 316         }
 317         case DATA_PROMOTE:
 318                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 319                 m->op.flags     |= BCH_WRITE_CACHED;
 320                 break;
 321         default:
 322                 BUG();
 323         }
 324
 325         return 0;
 326 }
 327
 328 static void move_free(struct closure *cl)
 329 {
 330         struct moving_io *io = container_of(cl, struct moving_io, cl);
 331         struct moving_context *ctxt = io->write.ctxt;
 332         struct bvec_iter_all iter;
 333         struct bio_vec *bv;
 334
 335         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 336
 337         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 338                 if (bv->bv_page)
 339                         __free_page(bv->bv_page);
 340
 341         wake_up(&ctxt->wait);
 342
 343         kfree(io);
 344 }
 345
 346 static void move_write_done(struct closure *cl)
 347 {
 348         struct moving_io *io = container_of(cl, struct moving_io, cl);
 349
 350         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 351         closure_return_with_destructor(cl, move_free);
 352 }
 353
 354 static void move_write(struct closure *cl)
 355 {
 356         struct moving_io *io = container_of(cl, struct moving_io, cl);
 357
 358         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 359                 closure_return_with_destructor(cl, move_free);
 360                 return;
 361         }
 362
 363         bch2_migrate_read_done(&io->write, &io->rbio);
 364
 365         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 366         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 367         continue_at(cl, move_write_done, NULL);
 368 }
 369
 370 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 371 {
 372         struct moving_io *io =
 373                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 374
 375         return io && io->read_completed ? io : NULL;
 376 }
 377
 378 static void move_read_endio(struct bio *bio)
 379 {
 380         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 381         struct moving_context *ctxt = io->write.ctxt;
 382
 383         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 384         io->read_completed = true;
 385
 386         if (next_pending_write(ctxt))
 387                 wake_up(&ctxt->wait);
 388
 389         closure_put(&ctxt->cl);
 390 }
 391
 392 static void do_pending_writes(struct moving_context *ctxt)
 393 {
 394         struct moving_io *io;
 395
 396         while ((io = next_pending_write(ctxt))) {
 397                 list_del(&io->list);
 398                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 399         }
 400 }
 401
 402 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 403 do {                                                            \
 404         do_pending_writes(_ctxt);                               \
 405                                                                 \
 406         if (_cond)                                              \
 407                 break;                                          \
 408         __wait_event((_ctxt)->wait,                             \
 409                      next_pending_write(_ctxt) || (_cond));     \
 410 } while (1)
 411
 412 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 413 {
 414         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 415
 416         move_ctxt_wait_event(ctxt,
 417                 !atomic_read(&ctxt->write_sectors) ||
 418                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 419 }
 420
 421 static int bch2_move_extent(struct btree_trans *trans,
 422                             struct moving_context *ctxt,
 423                             struct write_point_specifier wp,
 424                             struct bch_io_opts io_opts,
 425                             enum btree_id btree_id,
 426                             struct bkey_s_c k,
 427                             enum data_cmd data_cmd,
 428                             struct data_opts data_opts)
 429 {
 430         struct bch_fs *c = trans->c;
 431         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 432         struct moving_io *io;
 433         const union bch_extent_entry *entry;
 434         struct extent_ptr_decoded p;
 435         unsigned sectors = k.k->size, pages;
 436         int ret = -ENOMEM;
 437
 438         move_ctxt_wait_event(ctxt,
 439                 atomic_read(&ctxt->write_sectors) <
 440                 SECTORS_IN_FLIGHT_PER_DEVICE);
 441
 442         move_ctxt_wait_event(ctxt,
 443                 atomic_read(&ctxt->read_sectors) <
 444                 SECTORS_IN_FLIGHT_PER_DEVICE);
 445
 446         /* write path might have to decompress data: */
 447         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 448                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 449
 450         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 451         io = kzalloc(sizeof(struct moving_io) +
 452                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 453         if (!io)
 454                 goto err;
 455
 456         io->write.ctxt          = ctxt;
 457         io->read_sectors        = k.k->size;
 458         io->write_sectors       = k.k->size;
 459
 460         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 461         bio_set_prio(&io->write.op.wbio.bio,
 462                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 463
 464         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 465                                  GFP_KERNEL))
 466                 goto err_free;
 467
 468         io->rbio.c              = c;
 469         io->rbio.opts           = io_opts;
 470         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 471         io->rbio.bio.bi_vcnt = pages;
 472         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 473         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 474
 475         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 476         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 477         io->rbio.bio.bi_end_io          = move_read_endio;
 478
 479         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 480                                       data_cmd, data_opts, btree_id, k);
 481         if (ret)
 482                 goto err_free_pages;
 483
 484         atomic64_inc(&ctxt->stats->keys_moved);
 485         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 486
 487         trace_move_extent(k.k);
 488
 489         atomic_add(io->read_sectors, &ctxt->read_sectors);
 490         list_add_tail(&io->list, &ctxt->reads);
 491
 492         /*
 493          * dropped by move_read_endio() - guards against use after free of
 494          * ctxt when doing wakeup
 495          */
 496         closure_get(&ctxt->cl);
 497         bch2_read_extent(trans, &io->rbio,
 498                          bkey_start_pos(k.k),
 499                          btree_id, k, 0,
 500                          BCH_READ_NODECODE|
 501                          BCH_READ_LAST_FRAGMENT);
 502         return 0;
 503 err_free_pages:
 504         bio_free_pages(&io->write.op.wbio.bio);
 505 err_free:
 506         kfree(io);
 507 err:
 508         trace_move_alloc_fail(k.k);
 509         return ret;
 510 }
 511
 512 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 513                         struct bch_inode_unpacked *inode)
 514 {
 515         struct btree_iter *iter;
 516         struct bkey_s_c k;
 517         int ret;
 518
 519         iter = bch2_trans_get_iter(trans, BTREE_ID_inodes, pos,
 520                                    BTREE_ITER_ALL_SNAPSHOTS);
 521         k = bch2_btree_iter_peek(iter);
 522         ret = bkey_err(k);
 523         if (ret)
 524                 goto err;
 525
 526         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 527         if (ret)
 528                 goto err;
 529
 530         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 531         if (ret)
 532                 goto err;
 533 err:
 534         bch2_trans_iter_put(trans, iter);
 535         return ret;
 536 }
 537
 538 static int __bch2_move_data(struct bch_fs *c,
 539                 struct moving_context *ctxt,
 540                 struct bch_ratelimit *rate,
 541                 struct write_point_specifier wp,
 542                 struct bpos start,
 543                 struct bpos end,
 544                 move_pred_fn pred, void *arg,
 545                 struct bch_move_stats *stats,
 546                 enum btree_id btree_id)
 547 {
 548         bool kthread = (current->flags & PF_KTHREAD) != 0;
 549         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 550         struct bkey_buf sk;
 551         struct btree_trans trans;
 552         struct btree_iter *iter;
 553         struct bkey_s_c k;
 554         struct data_opts data_opts;
 555         enum data_cmd data_cmd;
 556         u64 delay, cur_inum = U64_MAX;
 557         int ret = 0, ret2;
 558
 559         bch2_bkey_buf_init(&sk);
 560         bch2_trans_init(&trans, c, 0, 0);
 561
 562         stats->data_type = BCH_DATA_user;
 563         stats->btree_id = btree_id;
 564         stats->pos      = start;
 565
 566         iter = bch2_trans_get_iter(&trans, btree_id, start,
 567                                    BTREE_ITER_PREFETCH);
 568
 569         if (rate)
 570                 bch2_ratelimit_reset(rate);
 571
 572         while (1) {
 573                 do {
 574                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 575
 576                         if (delay) {
 577                                 bch2_trans_unlock(&trans);
 578                                 set_current_state(TASK_INTERRUPTIBLE);
 579                         }
 580
 581                         if (kthread && (ret = kthread_should_stop())) {
 582                                 __set_current_state(TASK_RUNNING);
 583                                 goto out;
 584                         }
 585
 586                         if (delay)
 587                                 schedule_timeout(delay);
 588
 589                         if (unlikely(freezing(current))) {
 590                                 bch2_trans_unlock(&trans);
 591                                 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 592                                 try_to_freeze();
 593                         }
 594                 } while (delay);
 595
 596                 k = bch2_btree_iter_peek(iter);
 597
 598                 stats->pos = iter->pos;
 599
 600                 if (!k.k)
 601                         break;
 602                 ret = bkey_err(k);
 603                 if (ret)
 604                         break;
 605                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 606                         break;
 607
 608                 if (!bkey_extent_is_direct_data(k.k))
 609                         goto next_nondata;
 610
 611                 if (btree_id == BTREE_ID_extents &&
 612                     cur_inum != k.k->p.inode) {
 613                         struct bch_inode_unpacked inode;
 614
 615                         io_opts = bch2_opts_to_inode_opts(c->opts);
 616
 617                         ret = lookup_inode(&trans,
 618                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 619                                         &inode);
 620                         if (ret == -EINTR)
 621                                 continue;
 622
 623                         if (!ret)
 624                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 625
 626                         cur_inum = k.k->p.inode;
 627                 }
 628
 629                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 630                 case DATA_SKIP:
 631                         goto next;
 632                 case DATA_SCRUB:
 633                         BUG();
 634                 case DATA_ADD_REPLICAS:
 635                 case DATA_REWRITE:
 636                 case DATA_PROMOTE:
 637                         break;
 638                 default:
 639                         BUG();
 640                 }
 641
 642                 /* unlock before doing IO: */
 643                 bch2_bkey_buf_reassemble(&sk, c, k);
 644                 k = bkey_i_to_s_c(sk.k);
 645                 bch2_trans_unlock(&trans);
 646
 647                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 648                                         data_cmd, data_opts);
 649                 if (ret2) {
 650                         if (ret2 == -EINTR) {
 651                                 bch2_trans_reset(&trans, 0);
 652                                 bch2_trans_cond_resched(&trans);
 653                                 continue;
 654                         }
 655
 656                         if (ret2 == -ENOMEM) {
 657                                 /* memory allocation failure, wait for some IO to finish */
 658                                 bch2_move_ctxt_wait_for_io(ctxt);
 659                                 continue;
 660                         }
 661
 662                         /* XXX signal failure */
 663                         goto next;
 664                 }
 665
 666                 if (rate)
 667                         bch2_ratelimit_increment(rate, k.k->size);
 668 next:
 669                 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 670                              &stats->sectors_seen);
 671 next_nondata:
 672                 bch2_btree_iter_advance(iter);
 673                 bch2_trans_cond_resched(&trans);
 674         }
 675 out:
 676
 677         bch2_trans_iter_put(&trans, iter);
 678         ret = bch2_trans_exit(&trans) ?: ret;
 679         bch2_bkey_buf_exit(&sk, c);
 680
 681         return ret;
 682 }
 683
 684 int bch2_move_data(struct bch_fs *c,
 685                    enum btree_id start_btree_id, struct bpos start_pos,
 686                    enum btree_id end_btree_id,   struct bpos end_pos,
 687                    struct bch_ratelimit *rate,
 688                    struct write_point_specifier wp,
 689                    move_pred_fn pred, void *arg,
 690                    struct bch_move_stats *stats)
 691 {
 692         struct moving_context ctxt = { .stats = stats };
 693         enum btree_id id;
 694         int ret;
 695
 696         closure_init_stack(&ctxt.cl);
 697         INIT_LIST_HEAD(&ctxt.reads);
 698         init_waitqueue_head(&ctxt.wait);
 699
 700         stats->data_type = BCH_DATA_user;
 701
 702         for (id = start_btree_id;
 703              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 704              id++) {
 705                 stats->btree_id = id;
 706
 707                 if (id != BTREE_ID_extents &&
 708                     id != BTREE_ID_reflink)
 709                         continue;
 710
 711                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 712                                        id == start_btree_id ? start_pos : POS_MIN,
 713                                        id == end_btree_id   ? end_pos   : POS_MAX,
 714                                        pred, arg, stats, id);
 715                 if (ret)
 716                         break;
 717         }
 718
 719
 720         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 721         closure_sync(&ctxt.cl);
 722
 723         EBUG_ON(atomic_read(&ctxt.write_sectors));
 724
 725         trace_move_data(c,
 726                         atomic64_read(&stats->sectors_moved),
 727                         atomic64_read(&stats->keys_moved));
 728
 729         return ret;
 730 }
 731
 732 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 733                                          struct btree *, struct bch_io_opts *,
 734                                          struct data_opts *);
 735
 736 static int bch2_move_btree(struct bch_fs *c,
 737                            enum btree_id start_btree_id, struct bpos start_pos,
 738                            enum btree_id end_btree_id,   struct bpos end_pos,
 739                            move_btree_pred pred, void *arg,
 740                            struct bch_move_stats *stats)
 741 {
 742         bool kthread = (current->flags & PF_KTHREAD) != 0;
 743         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 744         struct btree_trans trans;
 745         struct btree_iter *iter;
 746         struct btree *b;
 747         enum btree_id id;
 748         struct data_opts data_opts;
 749         enum data_cmd cmd;
 750         int ret = 0;
 751
 752         bch2_trans_init(&trans, c, 0, 0);
 753
 754         stats->data_type = BCH_DATA_btree;
 755
 756         for (id = start_btree_id;
 757              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 758              id++) {
 759                 stats->btree_id = id;
 760
 761                 for_each_btree_node(&trans, iter, id,
 762                                     id == start_btree_id ? start_pos : POS_MIN,
 763                                     BTREE_ITER_PREFETCH, b) {
 764                         if (kthread && kthread_should_stop())
 765                                 goto out;
 766
 767                         if ((cmp_int(id, end_btree_id) ?:
 768                              bkey_cmp(b->key.k.p, end_pos)) > 0)
 769                                 break;
 770
 771                         stats->pos = iter->pos;
 772
 773                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 774                         case DATA_SKIP:
 775                                 goto next;
 776                         case DATA_SCRUB:
 777                                 BUG();
 778                         case DATA_ADD_REPLICAS:
 779                         case DATA_REWRITE:
 780                                 break;
 781                         default:
 782                                 BUG();
 783                         }
 784
 785                         ret = bch2_btree_node_rewrite(c, iter,
 786                                         b->data->keys.seq, 0) ?: ret;
 787 next:
 788                         bch2_trans_cond_resched(&trans);
 789                 }
 790
 791                 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
 792         }
 793 out:
 794         bch2_trans_exit(&trans);
 795
 796         if (ret)
 797                 bch_err(c, "error %i in bch2_move_btree", ret);
 798
 799         return ret;
 800 }
 801
 802 #if 0
 803 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 804                                 struct bkey_s_c k,
 805                                 struct bch_io_opts *io_opts,
 806                                 struct data_opts *data_opts)
 807 {
 808         return DATA_SCRUB;
 809 }
 810 #endif
 811
 812 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 813                                       struct bkey_s_c k,
 814                                       struct bch_io_opts *io_opts,
 815                                       struct data_opts *data_opts)
 816 {
 817         unsigned nr_good = bch2_bkey_durability(c, k);
 818         unsigned replicas = 0;
 819
 820         switch (k.k->type) {
 821         case KEY_TYPE_btree_ptr:
 822                 replicas = c->opts.metadata_replicas;
 823                 break;
 824         case KEY_TYPE_extent:
 825                 replicas = io_opts->data_replicas;
 826                 break;
 827         }
 828
 829         if (!nr_good || nr_good >= replicas)
 830                 return DATA_SKIP;
 831
 832         data_opts->target               = 0;
 833         data_opts->nr_replicas          = 1;
 834         data_opts->btree_insert_flags   = 0;
 835         return DATA_ADD_REPLICAS;
 836 }
 837
 838 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 839                                   struct bkey_s_c k,
 840                                   struct bch_io_opts *io_opts,
 841                                   struct data_opts *data_opts)
 842 {
 843         struct bch_ioctl_data *op = arg;
 844
 845         if (!bch2_bkey_has_device(k, op->migrate.dev))
 846                 return DATA_SKIP;
 847
 848         data_opts->target               = 0;
 849         data_opts->nr_replicas          = 1;
 850         data_opts->btree_insert_flags   = 0;
 851         data_opts->rewrite_dev          = op->migrate.dev;
 852         return DATA_REWRITE;
 853 }
 854
 855 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 856                                             struct btree *b,
 857                                             struct bch_io_opts *io_opts,
 858                                             struct data_opts *data_opts)
 859 {
 860         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 861 }
 862
 863 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
 864                                         struct btree *b,
 865                                         struct bch_io_opts *io_opts,
 866                                         struct data_opts *data_opts)
 867 {
 868         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 869 }
 870
 871 static bool bformat_needs_redo(struct bkey_format *f)
 872 {
 873         unsigned i;
 874
 875         for (i = 0; i < f->nr_fields; i++) {
 876                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 877                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 878                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 879
 880                 if (f->bits_per_field[i] > unpacked_bits)
 881                         return true;
 882
 883                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 884                         return true;
 885
 886                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 887                      unpacked_mask) <
 888                     field_offset)
 889                         return true;
 890         }
 891
 892         return false;
 893 }
 894
 895 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 896                                             struct btree *b,
 897                                             struct bch_io_opts *io_opts,
 898                                             struct data_opts *data_opts)
 899 {
 900         if (b->version_ondisk != c->sb.version ||
 901             btree_node_need_rewrite(b) ||
 902             bformat_needs_redo(&b->format)) {
 903                 data_opts->target               = 0;
 904                 data_opts->nr_replicas          = 1;
 905                 data_opts->btree_insert_flags   = 0;
 906                 return DATA_REWRITE;
 907         }
 908
 909         return DATA_SKIP;
 910 }
 911
 912 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 913 {
 914         int ret;
 915
 916         ret = bch2_move_btree(c,
 917                               0,                POS_MIN,
 918                               BTREE_ID_NR,      POS_MAX,
 919                               rewrite_old_nodes_pred, c, stats);
 920         if (!ret) {
 921                 mutex_lock(&c->sb_lock);
 922                 c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_extents_above_btree_updates_done;
 923                 c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_bformat_overflow_done;
 924                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
 925                 bch2_write_super(c);
 926                 mutex_unlock(&c->sb_lock);
 927         }
 928
 929         return ret;
 930 }
 931
 932 int bch2_data_job(struct bch_fs *c,
 933                   struct bch_move_stats *stats,
 934                   struct bch_ioctl_data op)
 935 {
 936         int ret = 0;
 937
 938         switch (op.op) {
 939         case BCH_DATA_OP_REREPLICATE:
 940                 stats->data_type = BCH_DATA_journal;
 941                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 942
 943                 ret = bch2_move_btree(c,
 944                                       op.start_btree,   op.start_pos,
 945                                       op.end_btree,     op.end_pos,
 946                                       rereplicate_btree_pred, c, stats) ?: ret;
 947
 948                 closure_wait_event(&c->btree_interior_update_wait,
 949                                    !bch2_btree_interior_updates_nr_pending(c));
 950
 951                 ret = bch2_replicas_gc2(c) ?: ret;
 952
 953                 ret = bch2_move_data(c,
 954                                      op.start_btree,    op.start_pos,
 955                                      op.end_btree,      op.end_pos,
 956                                      NULL, writepoint_hashed((unsigned long) current),
 957                                      rereplicate_pred, c, stats) ?: ret;
 958                 ret = bch2_replicas_gc2(c) ?: ret;
 959                 break;
 960         case BCH_DATA_OP_MIGRATE:
 961                 if (op.migrate.dev >= c->sb.nr_devices)
 962                         return -EINVAL;
 963
 964                 stats->data_type = BCH_DATA_journal;
 965                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
 966
 967                 ret = bch2_move_btree(c,
 968                                       op.start_btree,   op.start_pos,
 969                                       op.end_btree,     op.end_pos,
 970                                       migrate_btree_pred, &op, stats) ?: ret;
 971                 ret = bch2_replicas_gc2(c) ?: ret;
 972
 973                 ret = bch2_move_data(c,
 974                                      op.start_btree,    op.start_pos,
 975                                      op.end_btree,      op.end_pos,
 976                                      NULL, writepoint_hashed((unsigned long) current),
 977                                      migrate_pred, &op, stats) ?: ret;
 978                 ret = bch2_replicas_gc2(c) ?: ret;
 979                 break;
 980         case BCH_DATA_OP_REWRITE_OLD_NODES:
 981                 ret = bch2_scan_old_btree_nodes(c, stats);
 982                 break;
 983         default:
 984                 ret = -EINVAL;
 985         }
 986
 987         return ret;
 988 }