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, 1024);
  72
  73         bch2_trans_iter_init(&trans, &iter, 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                 struct bpos next_pos;
  84                 bool did_work = false;
  85                 bool extending = false, should_check_enospc;
  86                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
  87
  88                 bch2_trans_begin(&trans);
  89
  90                 k = bch2_btree_iter_peek_slot(&iter);
  91                 ret = bkey_err(k);
  92                 if (ret)
  93                         goto err;
  94
  95                 new = bkey_i_to_extent(bch2_keylist_front(keys));
  96
  97                 if (bversion_cmp(k.k->version, new->k.version) ||
  98                     !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
  99                         goto nomatch;
 100
 101                 bkey_reassemble(_insert.k, k);
 102                 insert = _insert.k;
 103
 104                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 105                 new = bkey_i_to_extent(_new.k);
 106                 bch2_cut_front(iter.pos, &new->k_i);
 107
 108                 bch2_cut_front(iter.pos,        insert);
 109                 bch2_cut_back(new->k.p,         insert);
 110                 bch2_cut_back(insert->k.p,      &new->k_i);
 111
 112                 if (m->data_cmd == DATA_REWRITE) {
 113                         struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
 114                                 bch2_bkey_has_device(bkey_i_to_s_c(insert),
 115                                                      m->data_opts.rewrite_dev);
 116                         if (!old_ptr)
 117                                 goto nomatch;
 118
 119                         if (old_ptr->cached)
 120                                 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
 121                                         new_ptr->cached = true;
 122
 123                         bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
 124                 }
 125
 126                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 127                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 128                                 /*
 129                                  * raced with another move op? extent already
 130                                  * has a pointer to the device we just wrote
 131                                  * data to
 132                                  */
 133                                 continue;
 134                         }
 135
 136                         bch2_extent_ptr_decoded_append(insert, &p);
 137                         did_work = true;
 138                 }
 139
 140                 if (!did_work)
 141                         goto nomatch;
 142
 143                 bch2_bkey_narrow_crcs(insert,
 144                                 (struct bch_extent_crc_unpacked) { 0 });
 145                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 146                 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
 147                                                op->opts.background_target,
 148                                                op->opts.data_replicas);
 149
 150                 ret = bch2_sum_sector_overwrites(&trans, &iter, insert,
 151                                                  &extending,
 152                                                  &should_check_enospc,
 153                                                  &i_sectors_delta,
 154                                                  &disk_sectors_delta);
 155                 if (ret)
 156                         goto err;
 157
 158                 if (disk_sectors_delta > (s64) op->res.sectors) {
 159                         ret = bch2_disk_reservation_add(c, &op->res,
 160                                                 disk_sectors_delta - op->res.sectors,
 161                                                 !should_check_enospc
 162                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 163                         if (ret)
 164                                 goto out;
 165                 }
 166
 167                 next_pos = insert->k.p;
 168
 169                 ret   = bch2_trans_update(&trans, &iter, insert, 0) ?:
 170                         bch2_trans_commit(&trans, &op->res,
 171                                 op_journal_seq(op),
 172                                 BTREE_INSERT_NOFAIL|
 173                                 m->data_opts.btree_insert_flags);
 174                 if (!ret) {
 175                         bch2_btree_iter_set_pos(&iter, next_pos);
 176                         atomic_long_inc(&c->extent_migrate_done);
 177                 }
 178 err:
 179                 if (ret == -EINTR)
 180                         ret = 0;
 181                 if (ret)
 182                         break;
 183 next:
 184                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 185                         bch2_keylist_pop_front(keys);
 186                         if (bch2_keylist_empty(keys))
 187                                 goto out;
 188                 }
 189                 continue;
 190 nomatch:
 191                 if (m->ctxt) {
 192                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 193                         atomic64_inc(&m->ctxt->stats->keys_raced);
 194                         atomic64_add(k.k->p.offset - iter.pos.offset,
 195                                      &m->ctxt->stats->sectors_raced);
 196                 }
 197                 atomic_long_inc(&c->extent_migrate_raced);
 198                 trace_move_race(&new->k);
 199                 bch2_btree_iter_advance(&iter);
 200                 goto next;
 201         }
 202 out:
 203         bch2_trans_iter_exit(&trans, &iter);
 204         bch2_trans_exit(&trans);
 205         bch2_bkey_buf_exit(&_insert, c);
 206         bch2_bkey_buf_exit(&_new, c);
 207         BUG_ON(ret == -EINTR);
 208         return ret;
 209 }
 210
 211 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 212 {
 213         /* write bio must own pages: */
 214         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 215
 216         m->ptr          = rbio->pick.ptr;
 217         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 218         m->op.devs_have = rbio->devs_have;
 219         m->op.pos       = rbio->data_pos;
 220         m->op.version   = rbio->version;
 221         m->op.crc       = rbio->pick.crc;
 222         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 223
 224         if (m->data_cmd == DATA_REWRITE)
 225                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 226 }
 227
 228 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 229                             struct write_point_specifier wp,
 230                             struct bch_io_opts io_opts,
 231                             enum data_cmd data_cmd,
 232                             struct data_opts data_opts,
 233                             enum btree_id btree_id,
 234                             struct bkey_s_c k)
 235 {
 236         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 237         const union bch_extent_entry *entry;
 238         struct bch_extent_crc_unpacked crc;
 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         /*
 260          * op->csum_type is normally initialized from the fs/file's current
 261          * options - but if an extent is encrypted, we require that it stays
 262          * encrypted:
 263          */
 264         bkey_for_each_crc(k.k, ptrs, crc, entry)
 265                 if (bch2_csum_type_is_encryption(crc.csum_type)) {
 266                         m->op.nonce     = crc.nonce + m->op.crc.offset;
 267                         m->op.csum_type = crc.csum_type;
 268                         break;
 269                 }
 270
 271         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 272                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 273                 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
 274         } else {
 275                 /* XXX: this should probably be passed in */
 276                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 277         }
 278
 279         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 280                 BCH_WRITE_PAGES_OWNED|
 281                 BCH_WRITE_DATA_ENCODED|
 282                 BCH_WRITE_FROM_INTERNAL;
 283
 284         m->op.nr_replicas       = data_opts.nr_replicas;
 285         m->op.nr_replicas_required = data_opts.nr_replicas;
 286         m->op.index_update_fn   = bch2_migrate_index_update;
 287
 288         switch (data_cmd) {
 289         case DATA_ADD_REPLICAS: {
 290                 /*
 291                  * DATA_ADD_REPLICAS is used for moving data to a different
 292                  * device in the background, and due to compression the new copy
 293                  * might take up more space than the old copy:
 294                  */
 295 #if 0
 296                 int nr = (int) io_opts.data_replicas -
 297                         bch2_bkey_nr_ptrs_allocated(k);
 298 #endif
 299                 int nr = (int) io_opts.data_replicas;
 300
 301                 if (nr > 0) {
 302                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 303
 304                         ret = bch2_disk_reservation_get(c, &m->op.res,
 305                                         k.k->size, m->op.nr_replicas, 0);
 306                         if (ret)
 307                                 return ret;
 308                 }
 309                 break;
 310         }
 311         case DATA_REWRITE: {
 312                 unsigned compressed_sectors = 0;
 313
 314                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 315                         if (p.ptr.dev == data_opts.rewrite_dev &&
 316                             !p.ptr.cached &&
 317                             crc_is_compressed(p.crc))
 318                                 compressed_sectors += p.crc.compressed_size;
 319
 320                 if (compressed_sectors) {
 321                         ret = bch2_disk_reservation_add(c, &m->op.res,
 322                                         k.k->size * m->op.nr_replicas,
 323                                         BCH_DISK_RESERVATION_NOFAIL);
 324                         if (ret)
 325                                 return ret;
 326                 }
 327                 break;
 328         }
 329         case DATA_PROMOTE:
 330                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 331                 m->op.flags     |= BCH_WRITE_CACHED;
 332                 break;
 333         default:
 334                 BUG();
 335         }
 336
 337         return 0;
 338 }
 339
 340 static void move_free(struct closure *cl)
 341 {
 342         struct moving_io *io = container_of(cl, struct moving_io, cl);
 343         struct moving_context *ctxt = io->write.ctxt;
 344         struct bvec_iter_all iter;
 345         struct bio_vec *bv;
 346
 347         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 348
 349         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 350                 if (bv->bv_page)
 351                         __free_page(bv->bv_page);
 352
 353         wake_up(&ctxt->wait);
 354
 355         kfree(io);
 356 }
 357
 358 static void move_write_done(struct closure *cl)
 359 {
 360         struct moving_io *io = container_of(cl, struct moving_io, cl);
 361
 362         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 363         closure_return_with_destructor(cl, move_free);
 364 }
 365
 366 static void move_write(struct closure *cl)
 367 {
 368         struct moving_io *io = container_of(cl, struct moving_io, cl);
 369
 370         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 371                 closure_return_with_destructor(cl, move_free);
 372                 return;
 373         }
 374
 375         bch2_migrate_read_done(&io->write, &io->rbio);
 376
 377         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 378         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 379         continue_at(cl, move_write_done, NULL);
 380 }
 381
 382 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 383 {
 384         struct moving_io *io =
 385                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 386
 387         return io && io->read_completed ? io : NULL;
 388 }
 389
 390 static void move_read_endio(struct bio *bio)
 391 {
 392         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 393         struct moving_context *ctxt = io->write.ctxt;
 394
 395         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 396         io->read_completed = true;
 397
 398         if (next_pending_write(ctxt))
 399                 wake_up(&ctxt->wait);
 400
 401         closure_put(&ctxt->cl);
 402 }
 403
 404 static void do_pending_writes(struct moving_context *ctxt)
 405 {
 406         struct moving_io *io;
 407
 408         while ((io = next_pending_write(ctxt))) {
 409                 list_del(&io->list);
 410                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 411         }
 412 }
 413
 414 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 415 do {                                                            \
 416         do_pending_writes(_ctxt);                               \
 417                                                                 \
 418         if (_cond)                                              \
 419                 break;                                          \
 420         __wait_event((_ctxt)->wait,                             \
 421                      next_pending_write(_ctxt) || (_cond));     \
 422 } while (1)
 423
 424 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 425 {
 426         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 427
 428         move_ctxt_wait_event(ctxt,
 429                 !atomic_read(&ctxt->write_sectors) ||
 430                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 431 }
 432
 433 static int bch2_move_extent(struct btree_trans *trans,
 434                             struct moving_context *ctxt,
 435                             struct write_point_specifier wp,
 436                             struct bch_io_opts io_opts,
 437                             enum btree_id btree_id,
 438                             struct bkey_s_c k,
 439                             enum data_cmd data_cmd,
 440                             struct data_opts data_opts)
 441 {
 442         struct bch_fs *c = trans->c;
 443         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 444         struct moving_io *io;
 445         const union bch_extent_entry *entry;
 446         struct extent_ptr_decoded p;
 447         unsigned sectors = k.k->size, pages;
 448         int ret = -ENOMEM;
 449
 450         move_ctxt_wait_event(ctxt,
 451                 atomic_read(&ctxt->write_sectors) <
 452                 SECTORS_IN_FLIGHT_PER_DEVICE);
 453
 454         move_ctxt_wait_event(ctxt,
 455                 atomic_read(&ctxt->read_sectors) <
 456                 SECTORS_IN_FLIGHT_PER_DEVICE);
 457
 458         /* write path might have to decompress data: */
 459         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 460                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 461
 462         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 463         io = kzalloc(sizeof(struct moving_io) +
 464                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 465         if (!io)
 466                 goto err;
 467
 468         io->write.ctxt          = ctxt;
 469         io->read_sectors        = k.k->size;
 470         io->write_sectors       = k.k->size;
 471
 472         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 473         bio_set_prio(&io->write.op.wbio.bio,
 474                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 475
 476         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 477                                  GFP_KERNEL))
 478                 goto err_free;
 479
 480         io->rbio.c              = c;
 481         io->rbio.opts           = io_opts;
 482         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 483         io->rbio.bio.bi_vcnt = pages;
 484         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 485         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 486
 487         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 488         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 489         io->rbio.bio.bi_end_io          = move_read_endio;
 490
 491         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 492                                       data_cmd, data_opts, btree_id, k);
 493         if (ret)
 494                 goto err_free_pages;
 495
 496         atomic64_inc(&ctxt->stats->keys_moved);
 497         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 498
 499         trace_move_extent(k.k);
 500
 501         atomic_add(io->read_sectors, &ctxt->read_sectors);
 502         list_add_tail(&io->list, &ctxt->reads);
 503
 504         /*
 505          * dropped by move_read_endio() - guards against use after free of
 506          * ctxt when doing wakeup
 507          */
 508         closure_get(&ctxt->cl);
 509         bch2_read_extent(trans, &io->rbio,
 510                          bkey_start_pos(k.k),
 511                          btree_id, k, 0,
 512                          BCH_READ_NODECODE|
 513                          BCH_READ_LAST_FRAGMENT);
 514         return 0;
 515 err_free_pages:
 516         bio_free_pages(&io->write.op.wbio.bio);
 517 err_free:
 518         kfree(io);
 519 err:
 520         trace_move_alloc_fail(k.k);
 521         return ret;
 522 }
 523
 524 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 525                         struct bch_inode_unpacked *inode)
 526 {
 527         struct btree_iter iter;
 528         struct bkey_s_c k;
 529         int ret;
 530
 531         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 532                              BTREE_ITER_ALL_SNAPSHOTS);
 533         k = bch2_btree_iter_peek(&iter);
 534         ret = bkey_err(k);
 535         if (ret)
 536                 goto err;
 537
 538         if (!k.k || bkey_cmp(k.k->p, pos)) {
 539                 ret = -ENOENT;
 540                 goto err;
 541         }
 542
 543         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 544         if (ret)
 545                 goto err;
 546
 547         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 548         if (ret)
 549                 goto err;
 550 err:
 551         bch2_trans_iter_exit(trans, &iter);
 552         return ret;
 553 }
 554
 555 static int __bch2_move_data(struct bch_fs *c,
 556                 struct moving_context *ctxt,
 557                 struct bch_ratelimit *rate,
 558                 struct write_point_specifier wp,
 559                 struct bpos start,
 560                 struct bpos end,
 561                 move_pred_fn pred, void *arg,
 562                 struct bch_move_stats *stats,
 563                 enum btree_id btree_id)
 564 {
 565         bool kthread = (current->flags & PF_KTHREAD) != 0;
 566         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 567         struct bkey_buf sk;
 568         struct btree_trans trans;
 569         struct btree_iter iter;
 570         struct bkey_s_c k;
 571         struct data_opts data_opts;
 572         enum data_cmd data_cmd;
 573         u64 delay, cur_inum = U64_MAX;
 574         int ret = 0, ret2;
 575
 576         bch2_bkey_buf_init(&sk);
 577         bch2_trans_init(&trans, c, 0, 0);
 578
 579         stats->data_type = BCH_DATA_user;
 580         stats->btree_id = btree_id;
 581         stats->pos      = start;
 582
 583         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 584                              BTREE_ITER_PREFETCH);
 585
 586         if (rate)
 587                 bch2_ratelimit_reset(rate);
 588
 589         while (1) {
 590                 do {
 591                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 592
 593                         if (delay) {
 594                                 bch2_trans_unlock(&trans);
 595                                 set_current_state(TASK_INTERRUPTIBLE);
 596                         }
 597
 598                         if (kthread && (ret = kthread_should_stop())) {
 599                                 __set_current_state(TASK_RUNNING);
 600                                 goto out;
 601                         }
 602
 603                         if (delay)
 604                                 schedule_timeout(delay);
 605
 606                         if (unlikely(freezing(current))) {
 607                                 bch2_trans_unlock(&trans);
 608                                 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 609                                 try_to_freeze();
 610                         }
 611                 } while (delay);
 612
 613                 bch2_trans_begin(&trans);
 614
 615                 k = bch2_btree_iter_peek(&iter);
 616
 617                 stats->pos = iter.pos;
 618
 619                 if (!k.k)
 620                         break;
 621                 ret = bkey_err(k);
 622                 if (ret)
 623                         break;
 624                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 625                         break;
 626
 627                 if (!bkey_extent_is_direct_data(k.k))
 628                         goto next_nondata;
 629
 630                 if (btree_id == BTREE_ID_extents &&
 631                     cur_inum != k.k->p.inode) {
 632                         struct bch_inode_unpacked inode;
 633
 634                         io_opts = bch2_opts_to_inode_opts(c->opts);
 635
 636                         ret = lookup_inode(&trans,
 637                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 638                                         &inode);
 639                         if (ret == -EINTR)
 640                                 continue;
 641
 642                         if (!ret)
 643                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 644
 645                         cur_inum = k.k->p.inode;
 646                 }
 647
 648                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 649                 case DATA_SKIP:
 650                         goto next;
 651                 case DATA_SCRUB:
 652                         BUG();
 653                 case DATA_ADD_REPLICAS:
 654                 case DATA_REWRITE:
 655                 case DATA_PROMOTE:
 656                         break;
 657                 default:
 658                         BUG();
 659                 }
 660
 661                 /* unlock before doing IO: */
 662                 bch2_bkey_buf_reassemble(&sk, c, k);
 663                 k = bkey_i_to_s_c(sk.k);
 664                 bch2_trans_unlock(&trans);
 665
 666                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 667                                         data_cmd, data_opts);
 668                 if (ret2) {
 669                         if (ret2 == -EINTR) {
 670                                 bch2_trans_begin(&trans);
 671                                 continue;
 672                         }
 673
 674                         if (ret2 == -ENOMEM) {
 675                                 /* memory allocation failure, wait for some IO to finish */
 676                                 bch2_move_ctxt_wait_for_io(ctxt);
 677                                 continue;
 678                         }
 679
 680                         /* XXX signal failure */
 681                         goto next;
 682                 }
 683
 684                 if (rate)
 685                         bch2_ratelimit_increment(rate, k.k->size);
 686 next:
 687                 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 688                              &stats->sectors_seen);
 689 next_nondata:
 690                 bch2_btree_iter_advance(&iter);
 691                 bch2_trans_cond_resched(&trans);
 692         }
 693 out:
 694
 695         bch2_trans_iter_exit(&trans, &iter);
 696         ret = bch2_trans_exit(&trans) ?: ret;
 697         bch2_bkey_buf_exit(&sk, c);
 698
 699         return ret;
 700 }
 701
 702 inline void bch_move_stats_init(struct bch_move_stats *stats, char *name)
 703 {
 704         memset(stats, 0, sizeof(*stats));
 705
 706         scnprintf(stats->name, sizeof(stats->name),
 707                         "%s", name);
 708 }
 709
 710 static inline void progress_list_add(struct bch_fs *c,
 711                                      struct bch_move_stats *stats)
 712 {
 713         mutex_lock(&c->data_progress_lock);
 714         list_add(&stats->list, &c->data_progress_list);
 715         mutex_unlock(&c->data_progress_lock);
 716 }
 717
 718 static inline void progress_list_del(struct bch_fs *c,
 719                                      struct bch_move_stats *stats)
 720 {
 721         mutex_lock(&c->data_progress_lock);
 722         list_del(&stats->list);
 723         mutex_unlock(&c->data_progress_lock);
 724 }
 725
 726 int bch2_move_data(struct bch_fs *c,
 727                    enum btree_id start_btree_id, struct bpos start_pos,
 728                    enum btree_id end_btree_id,   struct bpos end_pos,
 729                    struct bch_ratelimit *rate,
 730                    struct write_point_specifier wp,
 731                    move_pred_fn pred, void *arg,
 732                    struct bch_move_stats *stats)
 733 {
 734         struct moving_context ctxt = { .stats = stats };
 735         enum btree_id id;
 736         int ret;
 737
 738         progress_list_add(c, stats);
 739         closure_init_stack(&ctxt.cl);
 740         INIT_LIST_HEAD(&ctxt.reads);
 741         init_waitqueue_head(&ctxt.wait);
 742
 743         stats->data_type = BCH_DATA_user;
 744
 745         for (id = start_btree_id;
 746              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 747              id++) {
 748                 stats->btree_id = id;
 749
 750                 if (id != BTREE_ID_extents &&
 751                     id != BTREE_ID_reflink)
 752                         continue;
 753
 754                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 755                                        id == start_btree_id ? start_pos : POS_MIN,
 756                                        id == end_btree_id   ? end_pos   : POS_MAX,
 757                                        pred, arg, stats, id);
 758                 if (ret)
 759                         break;
 760         }
 761
 762
 763         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 764         closure_sync(&ctxt.cl);
 765
 766         EBUG_ON(atomic_read(&ctxt.write_sectors));
 767
 768         trace_move_data(c,
 769                         atomic64_read(&stats->sectors_moved),
 770                         atomic64_read(&stats->keys_moved));
 771
 772         progress_list_del(c, stats);
 773         return ret;
 774 }
 775
 776 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 777                                          struct btree *, struct bch_io_opts *,
 778                                          struct data_opts *);
 779
 780 static int bch2_move_btree(struct bch_fs *c,
 781                            enum btree_id start_btree_id, struct bpos start_pos,
 782                            enum btree_id end_btree_id,   struct bpos end_pos,
 783                            move_btree_pred pred, void *arg,
 784                            struct bch_move_stats *stats)
 785 {
 786         bool kthread = (current->flags & PF_KTHREAD) != 0;
 787         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 788         struct btree_trans trans;
 789         struct btree_iter iter;
 790         struct btree *b;
 791         enum btree_id id;
 792         struct data_opts data_opts;
 793         enum data_cmd cmd;
 794         int ret = 0;
 795
 796         bch2_trans_init(&trans, c, 0, 0);
 797         progress_list_add(c, stats);
 798
 799         stats->data_type = BCH_DATA_btree;
 800
 801         for (id = start_btree_id;
 802              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 803              id++) {
 804                 stats->btree_id = id;
 805
 806                 for_each_btree_node(&trans, iter, id,
 807                                     id == start_btree_id ? start_pos : POS_MIN,
 808                                     BTREE_ITER_PREFETCH, b) {
 809                         if (kthread && kthread_should_stop())
 810                                 break;
 811
 812                         if ((cmp_int(id, end_btree_id) ?:
 813                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 814                                 break;
 815
 816                         stats->pos = iter.pos;
 817
 818                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 819                         case DATA_SKIP:
 820                                 goto next;
 821                         case DATA_SCRUB:
 822                                 BUG();
 823                         case DATA_ADD_REPLICAS:
 824                         case DATA_REWRITE:
 825                                 break;
 826                         default:
 827                                 BUG();
 828                         }
 829
 830                         ret = bch2_btree_node_rewrite(&trans, &iter,
 831                                         b->data->keys.seq, 0) ?: ret;
 832 next:
 833                         bch2_trans_cond_resched(&trans);
 834                 }
 835                 bch2_trans_iter_exit(&trans, &iter);
 836
 837                 if (kthread && kthread_should_stop())
 838                         break;
 839         }
 840
 841         bch2_trans_exit(&trans);
 842
 843         if (ret)
 844                 bch_err(c, "error %i in bch2_move_btree", ret);
 845
 846         progress_list_del(c, stats);
 847         return ret;
 848 }
 849
 850 #if 0
 851 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 852                                 struct bkey_s_c k,
 853                                 struct bch_io_opts *io_opts,
 854                                 struct data_opts *data_opts)
 855 {
 856         return DATA_SCRUB;
 857 }
 858 #endif
 859
 860 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 861                                       struct bkey_s_c k,
 862                                       struct bch_io_opts *io_opts,
 863                                       struct data_opts *data_opts)
 864 {
 865         unsigned nr_good = bch2_bkey_durability(c, k);
 866         unsigned replicas = 0;
 867
 868         switch (k.k->type) {
 869         case KEY_TYPE_btree_ptr:
 870                 replicas = c->opts.metadata_replicas;
 871                 break;
 872         case KEY_TYPE_extent:
 873                 replicas = io_opts->data_replicas;
 874                 break;
 875         }
 876
 877         if (!nr_good || nr_good >= replicas)
 878                 return DATA_SKIP;
 879
 880         data_opts->target               = 0;
 881         data_opts->nr_replicas          = 1;
 882         data_opts->btree_insert_flags   = 0;
 883         return DATA_ADD_REPLICAS;
 884 }
 885
 886 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 887                                   struct bkey_s_c k,
 888                                   struct bch_io_opts *io_opts,
 889                                   struct data_opts *data_opts)
 890 {
 891         struct bch_ioctl_data *op = arg;
 892
 893         if (!bch2_bkey_has_device(k, op->migrate.dev))
 894                 return DATA_SKIP;
 895
 896         data_opts->target               = 0;
 897         data_opts->nr_replicas          = 1;
 898         data_opts->btree_insert_flags   = 0;
 899         data_opts->rewrite_dev          = op->migrate.dev;
 900         return DATA_REWRITE;
 901 }
 902
 903 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 904                                             struct btree *b,
 905                                             struct bch_io_opts *io_opts,
 906                                             struct data_opts *data_opts)
 907 {
 908         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 909 }
 910
 911 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
 912                                         struct btree *b,
 913                                         struct bch_io_opts *io_opts,
 914                                         struct data_opts *data_opts)
 915 {
 916         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 917 }
 918
 919 static bool bformat_needs_redo(struct bkey_format *f)
 920 {
 921         unsigned i;
 922
 923         for (i = 0; i < f->nr_fields; i++) {
 924                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 925                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 926                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 927
 928                 if (f->bits_per_field[i] > unpacked_bits)
 929                         return true;
 930
 931                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 932                         return true;
 933
 934                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 935                      unpacked_mask) <
 936                     field_offset)
 937                         return true;
 938         }
 939
 940         return false;
 941 }
 942
 943 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 944                                             struct btree *b,
 945                                             struct bch_io_opts *io_opts,
 946                                             struct data_opts *data_opts)
 947 {
 948         if (b->version_ondisk != c->sb.version ||
 949             btree_node_need_rewrite(b) ||
 950             bformat_needs_redo(&b->format)) {
 951                 data_opts->target               = 0;
 952                 data_opts->nr_replicas          = 1;
 953                 data_opts->btree_insert_flags   = 0;
 954                 return DATA_REWRITE;
 955         }
 956
 957         return DATA_SKIP;
 958 }
 959
 960 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 961 {
 962         int ret;
 963
 964         ret = bch2_move_btree(c,
 965                               0,                POS_MIN,
 966                               BTREE_ID_NR,      SPOS_MAX,
 967                               rewrite_old_nodes_pred, c, stats);
 968         if (!ret) {
 969                 mutex_lock(&c->sb_lock);
 970                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
 971                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 972                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
 973                 bch2_write_super(c);
 974                 mutex_unlock(&c->sb_lock);
 975         }
 976
 977         return ret;
 978 }
 979
 980 int bch2_data_job(struct bch_fs *c,
 981                   struct bch_move_stats *stats,
 982                   struct bch_ioctl_data op)
 983 {
 984         int ret = 0;
 985
 986         switch (op.op) {
 987         case BCH_DATA_OP_REREPLICATE:
 988                 bch_move_stats_init(stats, "rereplicate");
 989                 stats->data_type = BCH_DATA_journal;
 990                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 991
 992                 ret = bch2_move_btree(c,
 993                                       op.start_btree,   op.start_pos,
 994                                       op.end_btree,     op.end_pos,
 995                                       rereplicate_btree_pred, c, stats) ?: ret;
 996
 997                 closure_wait_event(&c->btree_interior_update_wait,
 998                                    !bch2_btree_interior_updates_nr_pending(c));
 999
1000                 ret = bch2_replicas_gc2(c) ?: ret;
1001
1002                 ret = bch2_move_data(c,
1003                                      op.start_btree,    op.start_pos,
1004                                      op.end_btree,      op.end_pos,
1005                                      NULL, writepoint_hashed((unsigned long) current),
1006                                      rereplicate_pred, c, stats) ?: ret;
1007                 ret = bch2_replicas_gc2(c) ?: ret;
1008                 break;
1009         case BCH_DATA_OP_MIGRATE:
1010                 if (op.migrate.dev >= c->sb.nr_devices)
1011                         return -EINVAL;
1012
1013                 bch_move_stats_init(stats, "migrate");
1014                 stats->data_type = BCH_DATA_journal;
1015                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1016
1017                 ret = bch2_move_btree(c,
1018                                       op.start_btree,   op.start_pos,
1019                                       op.end_btree,     op.end_pos,
1020                                       migrate_btree_pred, &op, stats) ?: ret;
1021                 ret = bch2_replicas_gc2(c) ?: ret;
1022
1023                 ret = bch2_move_data(c,
1024                                      op.start_btree,    op.start_pos,
1025                                      op.end_btree,      op.end_pos,
1026                                      NULL, writepoint_hashed((unsigned long) current),
1027                                      migrate_pred, &op, stats) ?: ret;
1028                 ret = bch2_replicas_gc2(c) ?: ret;
1029                 break;
1030         case BCH_DATA_OP_REWRITE_OLD_NODES:
1031                 bch_move_stats_init(stats, "rewrite_old_nodes");
1032                 ret = bch2_scan_old_btree_nodes(c, stats);
1033                 break;
1034         default:
1035                 ret = -EINVAL;
1036         }
1037
1038         return ret;
1039 }