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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_buf.h"
   6 #include "btree_gc.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "buckets.h"
  10 #include "disk_groups.h"
  11 #include "ec.h"
  12 #include "inode.h"
  13 #include "io.h"
  14 #include "journal_reclaim.h"
  15 #include "move.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18 #include "super-io.h"
  19 #include "keylist.h"
  20
  21 #include <linux/ioprio.h>
  22 #include <linux/kthread.h>
  23
  24 #include <trace/events/bcachefs.h>
  25
  26 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  27
  28 struct moving_io {
  29         struct list_head        list;
  30         struct closure          cl;
  31         bool                    read_completed;
  32
  33         unsigned                read_sectors;
  34         unsigned                write_sectors;
  35
  36         struct bch_read_bio     rbio;
  37
  38         struct migrate_write    write;
  39         /* Must be last since it is variable size */
  40         struct bio_vec          bi_inline_vecs[0];
  41 };
  42
  43 struct moving_context {
  44         /* Closure for waiting on all reads and writes to complete */
  45         struct closure          cl;
  46
  47         struct bch_move_stats   *stats;
  48
  49         struct list_head        reads;
  50
  51         /* in flight sectors: */
  52         atomic_t                read_sectors;
  53         atomic_t                write_sectors;
  54
  55         wait_queue_head_t       wait;
  56 };
  57
  58 static int insert_snapshot_whiteouts(struct btree_trans *trans,
  59                                      enum btree_id id,
  60                                      struct bpos old_pos,
  61                                      struct bpos new_pos)
  62 {
  63         struct bch_fs *c = trans->c;
  64         struct btree_iter iter, update_iter;
  65         struct bkey_s_c k;
  66         struct snapshots_seen s;
  67         int ret;
  68
  69         if (!btree_type_has_snapshots(id))
  70                 return 0;
  71
  72         snapshots_seen_init(&s);
  73
  74         if (!bkey_cmp(old_pos, new_pos))
  75                 return 0;
  76
  77         if (!snapshot_t(c, old_pos.snapshot)->children[0])
  78                 return 0;
  79
  80         bch2_trans_iter_init(trans, &iter, id, old_pos,
  81                              BTREE_ITER_NOT_EXTENTS|
  82                              BTREE_ITER_ALL_SNAPSHOTS);
  83         while (1) {
  84 next:
  85                 k = bch2_btree_iter_prev(&iter);
  86                 ret = bkey_err(k);
  87                 if (ret)
  88                         break;
  89
  90                 if (bkey_cmp(old_pos, k.k->p))
  91                         break;
  92
  93                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
  94                         struct bkey_i *update;
  95                         u32 *i;
  96
  97                         darray_for_each(s.ids, i)
  98                                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, *i))
  99                                         goto next;
 100
 101                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 102
 103                         ret = PTR_ERR_OR_ZERO(update);
 104                         if (ret)
 105                                 break;
 106
 107                         bkey_init(&update->k);
 108                         update->k.p = new_pos;
 109                         update->k.p.snapshot = k.k->p.snapshot;
 110
 111                         bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
 112                                              BTREE_ITER_NOT_EXTENTS|
 113                                              BTREE_ITER_ALL_SNAPSHOTS|
 114                                              BTREE_ITER_INTENT);
 115                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
 116                                 bch2_trans_update(trans, &update_iter, update,
 117                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 118                         bch2_trans_iter_exit(trans, &update_iter);
 119                         if (ret)
 120                                 break;
 121
 122                         ret = snapshots_seen_add(c, &s, k.k->p.snapshot);
 123                         if (ret)
 124                                 break;
 125                 }
 126         }
 127         bch2_trans_iter_exit(trans, &iter);
 128         darray_exit(&s.ids);
 129
 130         return ret;
 131 }
 132
 133 static int bch2_migrate_index_update(struct bch_write_op *op)
 134 {
 135         struct bch_fs *c = op->c;
 136         struct btree_trans trans;
 137         struct btree_iter iter;
 138         struct migrate_write *m =
 139                 container_of(op, struct migrate_write, op);
 140         struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
 141         struct keylist *keys = &op->insert_keys;
 142         struct bkey_buf _new, _insert;
 143         int ret = 0;
 144
 145         bch2_bkey_buf_init(&_new);
 146         bch2_bkey_buf_init(&_insert);
 147         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 148
 149         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 150
 151         bch2_trans_iter_init(&trans, &iter, m->btree_id,
 152                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 153                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 154
 155         while (1) {
 156                 struct bkey_s_c k;
 157                 struct bkey_i *insert;
 158                 struct bkey_i_extent *new;
 159                 const union bch_extent_entry *entry;
 160                 struct extent_ptr_decoded p;
 161                 struct bpos next_pos;
 162                 bool did_work = false;
 163                 bool should_check_enospc;
 164                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 165
 166                 bch2_trans_begin(&trans);
 167
 168                 k = bch2_btree_iter_peek_slot(&iter);
 169                 ret = bkey_err(k);
 170                 if (ret)
 171                         goto err;
 172
 173                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 174
 175                 if (bversion_cmp(k.k->version, new->k.version) ||
 176                     !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
 177                         goto nomatch;
 178
 179                 bkey_reassemble(_insert.k, k);
 180                 insert = _insert.k;
 181
 182                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 183                 new = bkey_i_to_extent(_new.k);
 184                 bch2_cut_front(iter.pos, &new->k_i);
 185
 186                 bch2_cut_front(iter.pos,        insert);
 187                 bch2_cut_back(new->k.p,         insert);
 188                 bch2_cut_back(insert->k.p,      &new->k_i);
 189
 190                 if (m->data_cmd == DATA_REWRITE) {
 191                         struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
 192                                 bch2_bkey_has_device(bkey_i_to_s_c(insert),
 193                                                      m->data_opts.rewrite_dev);
 194                         if (!old_ptr)
 195                                 goto nomatch;
 196
 197                         if (old_ptr->cached)
 198                                 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
 199                                         new_ptr->cached = true;
 200
 201                         __bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
 202                 }
 203
 204                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 205                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 206                                 /*
 207                                  * raced with another move op? extent already
 208                                  * has a pointer to the device we just wrote
 209                                  * data to
 210                                  */
 211                                 continue;
 212                         }
 213
 214                         bch2_extent_ptr_decoded_append(insert, &p);
 215                         did_work = true;
 216                 }
 217
 218                 if (!did_work)
 219                         goto nomatch;
 220
 221                 bch2_bkey_narrow_crcs(insert,
 222                                 (struct bch_extent_crc_unpacked) { 0 });
 223                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 224                 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
 225                                                op->opts.background_target,
 226                                                op->opts.data_replicas);
 227
 228                 ret = bch2_sum_sector_overwrites(&trans, &iter, insert,
 229                                                  &should_check_enospc,
 230                                                  &i_sectors_delta,
 231                                                  &disk_sectors_delta);
 232                 if (ret)
 233                         goto err;
 234
 235                 if (disk_sectors_delta > (s64) op->res.sectors) {
 236                         ret = bch2_disk_reservation_add(c, &op->res,
 237                                                 disk_sectors_delta - op->res.sectors,
 238                                                 !should_check_enospc
 239                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 240                         if (ret)
 241                                 goto out;
 242                 }
 243
 244                 next_pos = insert->k.p;
 245
 246                 ret   = insert_snapshot_whiteouts(&trans, m->btree_id,
 247                                                   k.k->p, insert->k.p) ?:
 248                         bch2_trans_update(&trans, &iter, insert,
 249                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 250                         bch2_trans_commit(&trans, &op->res,
 251                                 op_journal_seq(op),
 252                                 BTREE_INSERT_NOFAIL|
 253                                 m->data_opts.btree_insert_flags);
 254                 if (!ret) {
 255                         bch2_btree_iter_set_pos(&iter, next_pos);
 256                         atomic_long_inc(&c->extent_migrate_done);
 257                         if (ec_ob)
 258                                 bch2_ob_add_backpointer(c, ec_ob, &insert->k);
 259                 }
 260 err:
 261                 if (ret == -EINTR)
 262                         ret = 0;
 263                 if (ret)
 264                         break;
 265 next:
 266                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 267                         bch2_keylist_pop_front(keys);
 268                         if (bch2_keylist_empty(keys))
 269                                 goto out;
 270                 }
 271                 continue;
 272 nomatch:
 273                 if (m->ctxt) {
 274                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 275                         atomic64_inc(&m->ctxt->stats->keys_raced);
 276                         atomic64_add(k.k->p.offset - iter.pos.offset,
 277                                      &m->ctxt->stats->sectors_raced);
 278                 }
 279                 atomic_long_inc(&c->extent_migrate_raced);
 280                 trace_move_race(&new->k);
 281                 bch2_btree_iter_advance(&iter);
 282                 goto next;
 283         }
 284 out:
 285         bch2_trans_iter_exit(&trans, &iter);
 286         bch2_trans_exit(&trans);
 287         bch2_bkey_buf_exit(&_insert, c);
 288         bch2_bkey_buf_exit(&_new, c);
 289         BUG_ON(ret == -EINTR);
 290         return ret;
 291 }
 292
 293 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 294 {
 295         /* write bio must own pages: */
 296         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 297
 298         m->ptr          = rbio->pick.ptr;
 299         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 300         m->op.devs_have = rbio->devs_have;
 301         m->op.pos       = rbio->data_pos;
 302         m->op.version   = rbio->version;
 303         m->op.crc       = rbio->pick.crc;
 304         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 305
 306         if (m->data_cmd == DATA_REWRITE)
 307                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 308 }
 309
 310 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 311                             struct write_point_specifier wp,
 312                             struct bch_io_opts io_opts,
 313                             enum data_cmd data_cmd,
 314                             struct data_opts data_opts,
 315                             enum btree_id btree_id,
 316                             struct bkey_s_c k)
 317 {
 318         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 319         const union bch_extent_entry *entry;
 320         struct bch_extent_crc_unpacked crc;
 321         struct extent_ptr_decoded p;
 322         int ret;
 323
 324         m->btree_id     = btree_id;
 325         m->data_cmd     = data_cmd;
 326         m->data_opts    = data_opts;
 327         m->nr_ptrs_reserved = 0;
 328
 329         bch2_write_op_init(&m->op, c, io_opts);
 330
 331         if (!bch2_bkey_is_incompressible(k))
 332                 m->op.compression_type =
 333                         bch2_compression_opt_to_type[io_opts.background_compression ?:
 334                                                      io_opts.compression];
 335         else
 336                 m->op.incompressible = true;
 337
 338         m->op.target    = data_opts.target,
 339         m->op.write_point = wp;
 340
 341         /*
 342          * op->csum_type is normally initialized from the fs/file's current
 343          * options - but if an extent is encrypted, we require that it stays
 344          * encrypted:
 345          */
 346         bkey_for_each_crc(k.k, ptrs, crc, entry)
 347                 if (bch2_csum_type_is_encryption(crc.csum_type)) {
 348                         m->op.nonce     = crc.nonce + crc.offset;
 349                         m->op.csum_type = crc.csum_type;
 350                         break;
 351                 }
 352
 353         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 354                 m->op.alloc_reserve = RESERVE_movinggc;
 355         } else {
 356                 /* XXX: this should probably be passed in */
 357                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 358         }
 359
 360         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 361                 BCH_WRITE_PAGES_OWNED|
 362                 BCH_WRITE_DATA_ENCODED|
 363                 BCH_WRITE_FROM_INTERNAL;
 364
 365         m->op.nr_replicas       = data_opts.nr_replicas;
 366         m->op.nr_replicas_required = data_opts.nr_replicas;
 367         m->op.index_update_fn   = bch2_migrate_index_update;
 368
 369         switch (data_cmd) {
 370         case DATA_ADD_REPLICAS: {
 371                 /*
 372                  * DATA_ADD_REPLICAS is used for moving data to a different
 373                  * device in the background, and due to compression the new copy
 374                  * might take up more space than the old copy:
 375                  */
 376 #if 0
 377                 int nr = (int) io_opts.data_replicas -
 378                         bch2_bkey_nr_ptrs_allocated(k);
 379 #endif
 380                 int nr = (int) io_opts.data_replicas;
 381
 382                 if (nr > 0) {
 383                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 384
 385                         ret = bch2_disk_reservation_get(c, &m->op.res,
 386                                         k.k->size, m->op.nr_replicas, 0);
 387                         if (ret)
 388                                 return ret;
 389                 }
 390                 break;
 391         }
 392         case DATA_REWRITE: {
 393                 unsigned compressed_sectors = 0;
 394
 395                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 396                         if (p.ptr.dev == data_opts.rewrite_dev) {
 397                                 if (p.ptr.cached)
 398                                         m->op.flags |= BCH_WRITE_CACHED;
 399
 400                                 if (!p.ptr.cached &&
 401                                     crc_is_compressed(p.crc))
 402                                         compressed_sectors += p.crc.compressed_size;
 403                         }
 404
 405                 if (compressed_sectors) {
 406                         ret = bch2_disk_reservation_add(c, &m->op.res,
 407                                         k.k->size * m->op.nr_replicas,
 408                                         BCH_DISK_RESERVATION_NOFAIL);
 409                         if (ret)
 410                                 return ret;
 411                 }
 412                 break;
 413         }
 414         case DATA_PROMOTE:
 415                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 416                 m->op.flags     |= BCH_WRITE_CACHED;
 417                 break;
 418         default:
 419                 BUG();
 420         }
 421
 422         return 0;
 423 }
 424
 425 static void move_free(struct closure *cl)
 426 {
 427         struct moving_io *io = container_of(cl, struct moving_io, cl);
 428         struct moving_context *ctxt = io->write.ctxt;
 429         struct bvec_iter_all iter;
 430         struct bio_vec *bv;
 431
 432         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 433
 434         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 435                 if (bv->bv_page)
 436                         __free_page(bv->bv_page);
 437
 438         wake_up(&ctxt->wait);
 439
 440         kfree(io);
 441 }
 442
 443 static void move_write_done(struct closure *cl)
 444 {
 445         struct moving_io *io = container_of(cl, struct moving_io, cl);
 446
 447         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 448         closure_return_with_destructor(cl, move_free);
 449 }
 450
 451 static void move_write(struct closure *cl)
 452 {
 453         struct moving_io *io = container_of(cl, struct moving_io, cl);
 454
 455         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 456                 closure_return_with_destructor(cl, move_free);
 457                 return;
 458         }
 459
 460         bch2_migrate_read_done(&io->write, &io->rbio);
 461
 462         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 463         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 464         continue_at(cl, move_write_done, NULL);
 465 }
 466
 467 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 468 {
 469         struct moving_io *io =
 470                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 471
 472         return io && io->read_completed ? io : NULL;
 473 }
 474
 475 static void move_read_endio(struct bio *bio)
 476 {
 477         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 478         struct moving_context *ctxt = io->write.ctxt;
 479
 480         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 481         io->read_completed = true;
 482
 483         wake_up(&ctxt->wait);
 484         closure_put(&ctxt->cl);
 485 }
 486
 487 static void do_pending_writes(struct moving_context *ctxt, struct btree_trans *trans)
 488 {
 489         struct moving_io *io;
 490
 491         if (trans)
 492                 bch2_trans_unlock(trans);
 493
 494         while ((io = next_pending_write(ctxt))) {
 495                 list_del(&io->list);
 496                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 497         }
 498 }
 499
 500 #define move_ctxt_wait_event(_ctxt, _trans, _cond)              \
 501 do {                                                            \
 502         do_pending_writes(_ctxt, _trans);                       \
 503                                                                 \
 504         if (_cond)                                              \
 505                 break;                                          \
 506         __wait_event((_ctxt)->wait,                             \
 507                      next_pending_write(_ctxt) || (_cond));     \
 508 } while (1)
 509
 510 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 511                                        struct btree_trans *trans)
 512 {
 513         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 514
 515         move_ctxt_wait_event(ctxt, trans,
 516                 !atomic_read(&ctxt->write_sectors) ||
 517                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 518 }
 519
 520 static int bch2_move_extent(struct btree_trans *trans,
 521                             struct moving_context *ctxt,
 522                             struct write_point_specifier wp,
 523                             struct bch_io_opts io_opts,
 524                             enum btree_id btree_id,
 525                             struct bkey_s_c k,
 526                             enum data_cmd data_cmd,
 527                             struct data_opts data_opts)
 528 {
 529         struct bch_fs *c = trans->c;
 530         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 531         struct moving_io *io;
 532         const union bch_extent_entry *entry;
 533         struct extent_ptr_decoded p;
 534         unsigned sectors = k.k->size, pages;
 535         int ret = -ENOMEM;
 536
 537         /* write path might have to decompress data: */
 538         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 539                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 540
 541         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 542         io = kzalloc(sizeof(struct moving_io) +
 543                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 544         if (!io)
 545                 goto err;
 546
 547         io->write.ctxt          = ctxt;
 548         io->read_sectors        = k.k->size;
 549         io->write_sectors       = k.k->size;
 550
 551         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 552         bio_set_prio(&io->write.op.wbio.bio,
 553                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 554
 555         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 556                                  GFP_KERNEL))
 557                 goto err_free;
 558
 559         io->rbio.c              = c;
 560         io->rbio.opts           = io_opts;
 561         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 562         io->rbio.bio.bi_vcnt = pages;
 563         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 564         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 565
 566         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 567         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 568         io->rbio.bio.bi_end_io          = move_read_endio;
 569
 570         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 571                                       data_cmd, data_opts, btree_id, k);
 572         if (ret)
 573                 goto err_free_pages;
 574
 575         atomic64_inc(&ctxt->stats->keys_moved);
 576         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 577         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 578
 579         trace_move_extent(k.k);
 580
 581         atomic_add(io->read_sectors, &ctxt->read_sectors);
 582         list_add_tail(&io->list, &ctxt->reads);
 583
 584         /*
 585          * dropped by move_read_endio() - guards against use after free of
 586          * ctxt when doing wakeup
 587          */
 588         closure_get(&ctxt->cl);
 589         bch2_read_extent(trans, &io->rbio,
 590                          bkey_start_pos(k.k),
 591                          btree_id, k, 0,
 592                          BCH_READ_NODECODE|
 593                          BCH_READ_LAST_FRAGMENT);
 594         return 0;
 595 err_free_pages:
 596         bio_free_pages(&io->write.op.wbio.bio);
 597 err_free:
 598         kfree(io);
 599 err:
 600         trace_move_alloc_mem_fail(k.k);
 601         return ret;
 602 }
 603
 604 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 605                         struct bch_inode_unpacked *inode)
 606 {
 607         struct btree_iter iter;
 608         struct bkey_s_c k;
 609         int ret;
 610
 611         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 612                              BTREE_ITER_ALL_SNAPSHOTS);
 613         k = bch2_btree_iter_peek(&iter);
 614         ret = bkey_err(k);
 615         if (ret)
 616                 goto err;
 617
 618         if (!k.k || bkey_cmp(k.k->p, pos)) {
 619                 ret = -ENOENT;
 620                 goto err;
 621         }
 622
 623         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 624         if (ret)
 625                 goto err;
 626
 627         ret = bch2_inode_unpack(k, inode);
 628         if (ret)
 629                 goto err;
 630 err:
 631         bch2_trans_iter_exit(trans, &iter);
 632         return ret;
 633 }
 634
 635 static int __bch2_move_data(struct bch_fs *c,
 636                 struct moving_context *ctxt,
 637                 struct bch_ratelimit *rate,
 638                 struct write_point_specifier wp,
 639                 struct bpos start,
 640                 struct bpos end,
 641                 move_pred_fn pred, void *arg,
 642                 struct bch_move_stats *stats,
 643                 enum btree_id btree_id)
 644 {
 645         bool kthread = (current->flags & PF_KTHREAD) != 0;
 646         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 647         struct bkey_buf sk;
 648         struct btree_trans trans;
 649         struct btree_iter iter;
 650         struct bkey_s_c k;
 651         struct data_opts data_opts;
 652         enum data_cmd data_cmd;
 653         u64 delay, cur_inum = U64_MAX;
 654         int ret = 0, ret2;
 655
 656         bch2_bkey_buf_init(&sk);
 657         bch2_trans_init(&trans, c, 0, 0);
 658
 659         stats->data_type = BCH_DATA_user;
 660         stats->btree_id = btree_id;
 661         stats->pos      = start;
 662
 663         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 664                              BTREE_ITER_PREFETCH|
 665                              BTREE_ITER_ALL_SNAPSHOTS);
 666
 667         if (rate)
 668                 bch2_ratelimit_reset(rate);
 669
 670         while (1) {
 671                 do {
 672                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 673
 674                         if (delay) {
 675                                 bch2_trans_unlock(&trans);
 676                                 set_current_state(TASK_INTERRUPTIBLE);
 677                         }
 678
 679                         if (kthread && (ret = kthread_should_stop())) {
 680                                 __set_current_state(TASK_RUNNING);
 681                                 goto out;
 682                         }
 683
 684                         if (delay)
 685                                 schedule_timeout(delay);
 686
 687                         if (unlikely(freezing(current))) {
 688                                 move_ctxt_wait_event(ctxt, &trans, list_empty(&ctxt->reads));
 689                                 try_to_freeze();
 690                         }
 691                 } while (delay);
 692
 693                 move_ctxt_wait_event(ctxt, &trans,
 694                         atomic_read(&ctxt->write_sectors) <
 695                         SECTORS_IN_FLIGHT_PER_DEVICE);
 696
 697                 move_ctxt_wait_event(ctxt, &trans,
 698                         atomic_read(&ctxt->read_sectors) <
 699                         SECTORS_IN_FLIGHT_PER_DEVICE);
 700
 701                 bch2_trans_begin(&trans);
 702
 703                 k = bch2_btree_iter_peek(&iter);
 704                 if (!k.k)
 705                         break;
 706
 707                 ret = bkey_err(k);
 708                 if (ret == -EINTR)
 709                         continue;
 710                 if (ret)
 711                         break;
 712
 713                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 714                         break;
 715
 716                 stats->pos = iter.pos;
 717
 718                 if (!bkey_extent_is_direct_data(k.k))
 719                         goto next_nondata;
 720
 721                 if (btree_id == BTREE_ID_extents &&
 722                     cur_inum != k.k->p.inode) {
 723                         struct bch_inode_unpacked inode;
 724
 725                         io_opts = bch2_opts_to_inode_opts(c->opts);
 726
 727                         ret = lookup_inode(&trans,
 728                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 729                                         &inode);
 730                         if (ret == -EINTR)
 731                                 continue;
 732
 733                         if (!ret)
 734                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 735
 736                         cur_inum = k.k->p.inode;
 737                 }
 738
 739                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 740                 case DATA_SKIP:
 741                         goto next;
 742                 case DATA_SCRUB:
 743                         BUG();
 744                 case DATA_ADD_REPLICAS:
 745                 case DATA_REWRITE:
 746                 case DATA_PROMOTE:
 747                         break;
 748                 default:
 749                         BUG();
 750                 }
 751
 752                 /*
 753                  * The iterator gets unlocked by __bch2_read_extent - need to
 754                  * save a copy of @k elsewhere:
 755                   */
 756                 bch2_bkey_buf_reassemble(&sk, c, k);
 757                 k = bkey_i_to_s_c(sk.k);
 758
 759                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 760                                         data_cmd, data_opts);
 761                 if (ret2) {
 762                         if (ret2 == -EINTR)
 763                                 continue;
 764
 765                         if (ret2 == -ENOMEM) {
 766                                 /* memory allocation failure, wait for some IO to finish */
 767                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 768                                 continue;
 769                         }
 770
 771                         /* XXX signal failure */
 772                         goto next;
 773                 }
 774
 775                 if (rate)
 776                         bch2_ratelimit_increment(rate, k.k->size);
 777 next:
 778                 atomic64_add(k.k->size, &stats->sectors_seen);
 779 next_nondata:
 780                 bch2_btree_iter_advance(&iter);
 781         }
 782 out:
 783
 784         bch2_trans_iter_exit(&trans, &iter);
 785         bch2_trans_exit(&trans);
 786         bch2_bkey_buf_exit(&sk, c);
 787
 788         return ret;
 789 }
 790
 791 inline void bch_move_stats_init(struct bch_move_stats *stats, char *name)
 792 {
 793         memset(stats, 0, sizeof(*stats));
 794
 795         scnprintf(stats->name, sizeof(stats->name),
 796                         "%s", name);
 797 }
 798
 799 static inline void progress_list_add(struct bch_fs *c,
 800                                      struct bch_move_stats *stats)
 801 {
 802         mutex_lock(&c->data_progress_lock);
 803         list_add(&stats->list, &c->data_progress_list);
 804         mutex_unlock(&c->data_progress_lock);
 805 }
 806
 807 static inline void progress_list_del(struct bch_fs *c,
 808                                      struct bch_move_stats *stats)
 809 {
 810         mutex_lock(&c->data_progress_lock);
 811         list_del(&stats->list);
 812         mutex_unlock(&c->data_progress_lock);
 813 }
 814
 815 int bch2_move_data(struct bch_fs *c,
 816                    enum btree_id start_btree_id, struct bpos start_pos,
 817                    enum btree_id end_btree_id,   struct bpos end_pos,
 818                    struct bch_ratelimit *rate,
 819                    struct write_point_specifier wp,
 820                    move_pred_fn pred, void *arg,
 821                    struct bch_move_stats *stats)
 822 {
 823         struct moving_context ctxt = { .stats = stats };
 824         enum btree_id id;
 825         int ret;
 826
 827         progress_list_add(c, stats);
 828         closure_init_stack(&ctxt.cl);
 829         INIT_LIST_HEAD(&ctxt.reads);
 830         init_waitqueue_head(&ctxt.wait);
 831
 832         stats->data_type = BCH_DATA_user;
 833
 834         for (id = start_btree_id;
 835              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 836              id++) {
 837                 stats->btree_id = id;
 838
 839                 if (id != BTREE_ID_extents &&
 840                     id != BTREE_ID_reflink)
 841                         continue;
 842
 843                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 844                                        id == start_btree_id ? start_pos : POS_MIN,
 845                                        id == end_btree_id   ? end_pos   : POS_MAX,
 846                                        pred, arg, stats, id);
 847                 if (ret)
 848                         break;
 849         }
 850
 851
 852         move_ctxt_wait_event(&ctxt, NULL, list_empty(&ctxt.reads));
 853         closure_sync(&ctxt.cl);
 854
 855         EBUG_ON(atomic_read(&ctxt.write_sectors));
 856
 857         trace_move_data(c,
 858                         atomic64_read(&stats->sectors_moved),
 859                         atomic64_read(&stats->keys_moved));
 860
 861         progress_list_del(c, stats);
 862         return ret;
 863 }
 864
 865 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 866                                          struct btree *, struct bch_io_opts *,
 867                                          struct data_opts *);
 868
 869 static int bch2_move_btree(struct bch_fs *c,
 870                            enum btree_id start_btree_id, struct bpos start_pos,
 871                            enum btree_id end_btree_id,   struct bpos end_pos,
 872                            move_btree_pred pred, void *arg,
 873                            struct bch_move_stats *stats)
 874 {
 875         bool kthread = (current->flags & PF_KTHREAD) != 0;
 876         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 877         struct btree_trans trans;
 878         struct btree_iter iter;
 879         struct btree *b;
 880         enum btree_id id;
 881         struct data_opts data_opts;
 882         enum data_cmd cmd;
 883         int ret = 0;
 884
 885         bch2_trans_init(&trans, c, 0, 0);
 886         progress_list_add(c, stats);
 887
 888         stats->data_type = BCH_DATA_btree;
 889
 890         for (id = start_btree_id;
 891              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 892              id++) {
 893                 stats->btree_id = id;
 894
 895                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 896                                           BTREE_ITER_PREFETCH);
 897 retry:
 898                 ret = 0;
 899                 while (bch2_trans_begin(&trans),
 900                        (b = bch2_btree_iter_peek_node(&iter)) &&
 901                        !(ret = PTR_ERR_OR_ZERO(b))) {
 902                         if (kthread && kthread_should_stop())
 903                                 break;
 904
 905                         if ((cmp_int(id, end_btree_id) ?:
 906                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 907                                 break;
 908
 909                         stats->pos = iter.pos;
 910
 911                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 912                         case DATA_SKIP:
 913                                 goto next;
 914                         case DATA_SCRUB:
 915                                 BUG();
 916                         case DATA_ADD_REPLICAS:
 917                         case DATA_REWRITE:
 918                                 break;
 919                         default:
 920                                 BUG();
 921                         }
 922
 923                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 924                         if (ret == -EINTR)
 925                                 continue;
 926                         if (ret)
 927                                 break;
 928 next:
 929                         bch2_btree_iter_next_node(&iter);
 930                 }
 931                 if (ret == -EINTR)
 932                         goto retry;
 933
 934                 bch2_trans_iter_exit(&trans, &iter);
 935
 936                 if (kthread && kthread_should_stop())
 937                         break;
 938         }
 939
 940         bch2_trans_exit(&trans);
 941
 942         if (ret)
 943                 bch_err(c, "error %i in bch2_move_btree", ret);
 944
 945         bch2_btree_interior_updates_flush(c);
 946
 947         progress_list_del(c, stats);
 948         return ret;
 949 }
 950
 951 #if 0
 952 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 953                                 struct bkey_s_c k,
 954                                 struct bch_io_opts *io_opts,
 955                                 struct data_opts *data_opts)
 956 {
 957         return DATA_SCRUB;
 958 }
 959 #endif
 960
 961 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 962                                       struct bkey_s_c k,
 963                                       struct bch_io_opts *io_opts,
 964                                       struct data_opts *data_opts)
 965 {
 966         unsigned nr_good = bch2_bkey_durability(c, k);
 967         unsigned replicas = bkey_is_btree_ptr(k.k)
 968                 ? c->opts.metadata_replicas
 969                 : io_opts->data_replicas;
 970
 971         if (!nr_good || nr_good >= replicas)
 972                 return DATA_SKIP;
 973
 974         data_opts->target               = 0;
 975         data_opts->nr_replicas          = 1;
 976         data_opts->btree_insert_flags   = 0;
 977         return DATA_ADD_REPLICAS;
 978 }
 979
 980 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 981                                   struct bkey_s_c k,
 982                                   struct bch_io_opts *io_opts,
 983                                   struct data_opts *data_opts)
 984 {
 985         struct bch_ioctl_data *op = arg;
 986
 987         if (!bch2_bkey_has_device(k, op->migrate.dev))
 988                 return DATA_SKIP;
 989
 990         data_opts->target               = 0;
 991         data_opts->nr_replicas          = 1;
 992         data_opts->btree_insert_flags   = 0;
 993         data_opts->rewrite_dev          = op->migrate.dev;
 994         return DATA_REWRITE;
 995 }
 996
 997 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 998                                             struct btree *b,
 999                                             struct bch_io_opts *io_opts,
1000                                             struct data_opts *data_opts)
1001 {
1002         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1003 }
1004
1005 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
1006                                         struct btree *b,
1007                                         struct bch_io_opts *io_opts,
1008                                         struct data_opts *data_opts)
1009 {
1010         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1011 }
1012
1013 static bool bformat_needs_redo(struct bkey_format *f)
1014 {
1015         unsigned i;
1016
1017         for (i = 0; i < f->nr_fields; i++) {
1018                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
1019                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
1020                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
1021
1022                 if (f->bits_per_field[i] > unpacked_bits)
1023                         return true;
1024
1025                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1026                         return true;
1027
1028                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1029                      unpacked_mask) <
1030                     field_offset)
1031                         return true;
1032         }
1033
1034         return false;
1035 }
1036
1037 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1038                                             struct btree *b,
1039                                             struct bch_io_opts *io_opts,
1040                                             struct data_opts *data_opts)
1041 {
1042         if (b->version_ondisk != c->sb.version ||
1043             btree_node_need_rewrite(b) ||
1044             bformat_needs_redo(&b->format)) {
1045                 data_opts->target               = 0;
1046                 data_opts->nr_replicas          = 1;
1047                 data_opts->btree_insert_flags   = 0;
1048                 return DATA_REWRITE;
1049         }
1050
1051         return DATA_SKIP;
1052 }
1053
1054 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1055 {
1056         int ret;
1057
1058         ret = bch2_move_btree(c,
1059                               0,                POS_MIN,
1060                               BTREE_ID_NR,      SPOS_MAX,
1061                               rewrite_old_nodes_pred, c, stats);
1062         if (!ret) {
1063                 mutex_lock(&c->sb_lock);
1064                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1065                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1066                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1067                 bch2_write_super(c);
1068                 mutex_unlock(&c->sb_lock);
1069         }
1070
1071         return ret;
1072 }
1073
1074 int bch2_data_job(struct bch_fs *c,
1075                   struct bch_move_stats *stats,
1076                   struct bch_ioctl_data op)
1077 {
1078         int ret = 0;
1079
1080         switch (op.op) {
1081         case BCH_DATA_OP_REREPLICATE:
1082                 bch_move_stats_init(stats, "rereplicate");
1083                 stats->data_type = BCH_DATA_journal;
1084                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1085
1086                 ret = bch2_move_btree(c,
1087                                       op.start_btree,   op.start_pos,
1088                                       op.end_btree,     op.end_pos,
1089                                       rereplicate_btree_pred, c, stats) ?: ret;
1090                 ret = bch2_replicas_gc2(c) ?: ret;
1091
1092                 ret = bch2_move_data(c,
1093                                      op.start_btree,    op.start_pos,
1094                                      op.end_btree,      op.end_pos,
1095                                      NULL, writepoint_hashed((unsigned long) current),
1096                                      rereplicate_pred, c, stats) ?: ret;
1097                 ret = bch2_replicas_gc2(c) ?: ret;
1098                 break;
1099         case BCH_DATA_OP_MIGRATE:
1100                 if (op.migrate.dev >= c->sb.nr_devices)
1101                         return -EINVAL;
1102
1103                 bch_move_stats_init(stats, "migrate");
1104                 stats->data_type = BCH_DATA_journal;
1105                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1106
1107                 ret = bch2_move_btree(c,
1108                                       op.start_btree,   op.start_pos,
1109                                       op.end_btree,     op.end_pos,
1110                                       migrate_btree_pred, &op, stats) ?: ret;
1111                 ret = bch2_replicas_gc2(c) ?: ret;
1112
1113                 ret = bch2_move_data(c,
1114                                      op.start_btree,    op.start_pos,
1115                                      op.end_btree,      op.end_pos,
1116                                      NULL, writepoint_hashed((unsigned long) current),
1117                                      migrate_pred, &op, stats) ?: ret;
1118                 ret = bch2_replicas_gc2(c) ?: ret;
1119                 break;
1120         case BCH_DATA_OP_REWRITE_OLD_NODES:
1121                 bch_move_stats_init(stats, "rewrite_old_nodes");
1122                 ret = bch2_scan_old_btree_nodes(c, stats);
1123                 break;
1124         default:
1125                 ret = -EINVAL;
1126         }
1127
1128         return ret;
1129 }