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         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_begin(&trans);
  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                 ret   = bch2_trans_update(&trans, iter, insert, 0) ?:
 167                         bch2_trans_commit(&trans, &op->res,
 168                                 op_journal_seq(op),
 169                                 BTREE_INSERT_NOFAIL|
 170                                 m->data_opts.btree_insert_flags);
 171 err:
 172                 if (!ret)
 173                         atomic_long_inc(&c->extent_migrate_done);
 174                 if (ret == -EINTR)
 175                         ret = 0;
 176                 if (ret)
 177                         break;
 178 next:
 179                 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
 180                         bch2_keylist_pop_front(keys);
 181                         if (bch2_keylist_empty(keys))
 182                                 goto out;
 183                 }
 184                 continue;
 185 nomatch:
 186                 if (m->ctxt) {
 187                         BUG_ON(k.k->p.offset <= iter->pos.offset);
 188                         atomic64_inc(&m->ctxt->stats->keys_raced);
 189                         atomic64_add(k.k->p.offset - iter->pos.offset,
 190                                      &m->ctxt->stats->sectors_raced);
 191                 }
 192                 atomic_long_inc(&c->extent_migrate_raced);
 193                 trace_move_race(&new->k);
 194                 bch2_btree_iter_advance(iter);
 195                 goto next;
 196         }
 197 out:
 198         bch2_trans_iter_put(&trans, iter);
 199         bch2_trans_exit(&trans);
 200         bch2_bkey_buf_exit(&_insert, c);
 201         bch2_bkey_buf_exit(&_new, c);
 202         BUG_ON(ret == -EINTR);
 203         return ret;
 204 }
 205
 206 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 207 {
 208         /* write bio must own pages: */
 209         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 210
 211         m->ptr          = rbio->pick.ptr;
 212         m->offset       = rbio->data_pos.offset - rbio->pick.crc.offset;
 213         m->op.devs_have = rbio->devs_have;
 214         m->op.pos       = rbio->data_pos;
 215         m->op.version   = rbio->version;
 216         m->op.crc       = rbio->pick.crc;
 217         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 218
 219         if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
 220                 m->op.nonce     = m->op.crc.nonce + m->op.crc.offset;
 221                 m->op.csum_type = m->op.crc.csum_type;
 222         }
 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 extent_ptr_decoded p;
 239         int ret;
 240
 241         m->btree_id     = btree_id;
 242         m->data_cmd     = data_cmd;
 243         m->data_opts    = data_opts;
 244         m->nr_ptrs_reserved = 0;
 245
 246         bch2_write_op_init(&m->op, c, io_opts);
 247
 248         if (!bch2_bkey_is_incompressible(k))
 249                 m->op.compression_type =
 250                         bch2_compression_opt_to_type[io_opts.background_compression ?:
 251                                                      io_opts.compression];
 252         else
 253                 m->op.incompressible = true;
 254
 255         m->op.target    = data_opts.target,
 256         m->op.write_point = wp;
 257
 258         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
 259                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 260                 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
 261         } else {
 262                 /* XXX: this should probably be passed in */
 263                 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 264         }
 265
 266         m->op.flags |= BCH_WRITE_PAGES_STABLE|
 267                 BCH_WRITE_PAGES_OWNED|
 268                 BCH_WRITE_DATA_ENCODED|
 269                 BCH_WRITE_FROM_INTERNAL;
 270
 271         m->op.nr_replicas       = data_opts.nr_replicas;
 272         m->op.nr_replicas_required = data_opts.nr_replicas;
 273         m->op.index_update_fn   = bch2_migrate_index_update;
 274
 275         switch (data_cmd) {
 276         case DATA_ADD_REPLICAS: {
 277                 /*
 278                  * DATA_ADD_REPLICAS is used for moving data to a different
 279                  * device in the background, and due to compression the new copy
 280                  * might take up more space than the old copy:
 281                  */
 282 #if 0
 283                 int nr = (int) io_opts.data_replicas -
 284                         bch2_bkey_nr_ptrs_allocated(k);
 285 #endif
 286                 int nr = (int) io_opts.data_replicas;
 287
 288                 if (nr > 0) {
 289                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 290
 291                         ret = bch2_disk_reservation_get(c, &m->op.res,
 292                                         k.k->size, m->op.nr_replicas, 0);
 293                         if (ret)
 294                                 return ret;
 295                 }
 296                 break;
 297         }
 298         case DATA_REWRITE: {
 299                 unsigned compressed_sectors = 0;
 300
 301                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 302                         if (p.ptr.dev == data_opts.rewrite_dev &&
 303                             !p.ptr.cached &&
 304                             crc_is_compressed(p.crc))
 305                                 compressed_sectors += p.crc.compressed_size;
 306
 307                 if (compressed_sectors) {
 308                         ret = bch2_disk_reservation_add(c, &m->op.res,
 309                                         k.k->size * m->op.nr_replicas,
 310                                         BCH_DISK_RESERVATION_NOFAIL);
 311                         if (ret)
 312                                 return ret;
 313                 }
 314                 break;
 315         }
 316         case DATA_PROMOTE:
 317                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 318                 m->op.flags     |= BCH_WRITE_CACHED;
 319                 break;
 320         default:
 321                 BUG();
 322         }
 323
 324         return 0;
 325 }
 326
 327 static void move_free(struct closure *cl)
 328 {
 329         struct moving_io *io = container_of(cl, struct moving_io, cl);
 330         struct moving_context *ctxt = io->write.ctxt;
 331         struct bvec_iter_all iter;
 332         struct bio_vec *bv;
 333
 334         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 335
 336         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 337                 if (bv->bv_page)
 338                         __free_page(bv->bv_page);
 339
 340         wake_up(&ctxt->wait);
 341
 342         kfree(io);
 343 }
 344
 345 static void move_write_done(struct closure *cl)
 346 {
 347         struct moving_io *io = container_of(cl, struct moving_io, cl);
 348
 349         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 350         closure_return_with_destructor(cl, move_free);
 351 }
 352
 353 static void move_write(struct closure *cl)
 354 {
 355         struct moving_io *io = container_of(cl, struct moving_io, cl);
 356
 357         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 358                 closure_return_with_destructor(cl, move_free);
 359                 return;
 360         }
 361
 362         bch2_migrate_read_done(&io->write, &io->rbio);
 363
 364         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 365         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 366         continue_at(cl, move_write_done, NULL);
 367 }
 368
 369 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 370 {
 371         struct moving_io *io =
 372                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 373
 374         return io && io->read_completed ? io : NULL;
 375 }
 376
 377 static void move_read_endio(struct bio *bio)
 378 {
 379         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 380         struct moving_context *ctxt = io->write.ctxt;
 381
 382         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 383         io->read_completed = true;
 384
 385         if (next_pending_write(ctxt))
 386                 wake_up(&ctxt->wait);
 387
 388         closure_put(&ctxt->cl);
 389 }
 390
 391 static void do_pending_writes(struct moving_context *ctxt)
 392 {
 393         struct moving_io *io;
 394
 395         while ((io = next_pending_write(ctxt))) {
 396                 list_del(&io->list);
 397                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 398         }
 399 }
 400
 401 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 402 do {                                                            \
 403         do_pending_writes(_ctxt);                               \
 404                                                                 \
 405         if (_cond)                                              \
 406                 break;                                          \
 407         __wait_event((_ctxt)->wait,                             \
 408                      next_pending_write(_ctxt) || (_cond));     \
 409 } while (1)
 410
 411 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 412 {
 413         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 414
 415         move_ctxt_wait_event(ctxt,
 416                 !atomic_read(&ctxt->write_sectors) ||
 417                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 418 }
 419
 420 static int bch2_move_extent(struct btree_trans *trans,
 421                             struct moving_context *ctxt,
 422                             struct write_point_specifier wp,
 423                             struct bch_io_opts io_opts,
 424                             enum btree_id btree_id,
 425                             struct bkey_s_c k,
 426                             enum data_cmd data_cmd,
 427                             struct data_opts data_opts)
 428 {
 429         struct bch_fs *c = trans->c;
 430         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 431         struct moving_io *io;
 432         const union bch_extent_entry *entry;
 433         struct extent_ptr_decoded p;
 434         unsigned sectors = k.k->size, pages;
 435         int ret = -ENOMEM;
 436
 437         move_ctxt_wait_event(ctxt,
 438                 atomic_read(&ctxt->write_sectors) <
 439                 SECTORS_IN_FLIGHT_PER_DEVICE);
 440
 441         move_ctxt_wait_event(ctxt,
 442                 atomic_read(&ctxt->read_sectors) <
 443                 SECTORS_IN_FLIGHT_PER_DEVICE);
 444
 445         /* write path might have to decompress data: */
 446         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 447                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 448
 449         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 450         io = kzalloc(sizeof(struct moving_io) +
 451                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 452         if (!io)
 453                 goto err;
 454
 455         io->write.ctxt          = ctxt;
 456         io->read_sectors        = k.k->size;
 457         io->write_sectors       = k.k->size;
 458
 459         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 460         bio_set_prio(&io->write.op.wbio.bio,
 461                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 462
 463         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 464                                  GFP_KERNEL))
 465                 goto err_free;
 466
 467         io->rbio.c              = c;
 468         io->rbio.opts           = io_opts;
 469         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 470         io->rbio.bio.bi_vcnt = pages;
 471         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 472         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 473
 474         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 475         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 476         io->rbio.bio.bi_end_io          = move_read_endio;
 477
 478         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 479                                       data_cmd, data_opts, btree_id, k);
 480         if (ret)
 481                 goto err_free_pages;
 482
 483         atomic64_inc(&ctxt->stats->keys_moved);
 484         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 485
 486         trace_move_extent(k.k);
 487
 488         atomic_add(io->read_sectors, &ctxt->read_sectors);
 489         list_add_tail(&io->list, &ctxt->reads);
 490
 491         /*
 492          * dropped by move_read_endio() - guards against use after free of
 493          * ctxt when doing wakeup
 494          */
 495         closure_get(&ctxt->cl);
 496         bch2_read_extent(trans, &io->rbio,
 497                          bkey_start_pos(k.k),
 498                          btree_id, k, 0,
 499                          BCH_READ_NODECODE|
 500                          BCH_READ_LAST_FRAGMENT);
 501         return 0;
 502 err_free_pages:
 503         bio_free_pages(&io->write.op.wbio.bio);
 504 err_free:
 505         kfree(io);
 506 err:
 507         trace_move_alloc_fail(k.k);
 508         return ret;
 509 }
 510
 511 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 512                         struct bch_inode_unpacked *inode)
 513 {
 514         struct btree_iter *iter;
 515         struct bkey_s_c k;
 516         int ret;
 517
 518         iter = bch2_trans_get_iter(trans, BTREE_ID_inodes, pos,
 519                                    BTREE_ITER_ALL_SNAPSHOTS);
 520         k = bch2_btree_iter_peek(iter);
 521         ret = bkey_err(k);
 522         if (ret)
 523                 goto err;
 524
 525         if (!k.k || bkey_cmp(k.k->p, pos)) {
 526                 ret = -ENOENT;
 527                 goto err;
 528         }
 529
 530         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 531         if (ret)
 532                 goto err;
 533
 534         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 535         if (ret)
 536                 goto err;
 537 err:
 538         bch2_trans_iter_put(trans, iter);
 539         return ret;
 540 }
 541
 542 static int __bch2_move_data(struct bch_fs *c,
 543                 struct moving_context *ctxt,
 544                 struct bch_ratelimit *rate,
 545                 struct write_point_specifier wp,
 546                 struct bpos start,
 547                 struct bpos end,
 548                 move_pred_fn pred, void *arg,
 549                 struct bch_move_stats *stats,
 550                 enum btree_id btree_id)
 551 {
 552         bool kthread = (current->flags & PF_KTHREAD) != 0;
 553         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 554         struct bkey_buf sk;
 555         struct btree_trans trans;
 556         struct btree_iter *iter;
 557         struct bkey_s_c k;
 558         struct data_opts data_opts;
 559         enum data_cmd data_cmd;
 560         u64 delay, cur_inum = U64_MAX;
 561         int ret = 0, ret2;
 562
 563         bch2_bkey_buf_init(&sk);
 564         bch2_trans_init(&trans, c, 0, 0);
 565
 566         stats->data_type = BCH_DATA_user;
 567         stats->btree_id = btree_id;
 568         stats->pos      = start;
 569
 570         iter = bch2_trans_get_iter(&trans, btree_id, start,
 571                                    BTREE_ITER_PREFETCH);
 572
 573         if (rate)
 574                 bch2_ratelimit_reset(rate);
 575
 576         while (1) {
 577                 do {
 578                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 579
 580                         if (delay) {
 581                                 bch2_trans_unlock(&trans);
 582                                 set_current_state(TASK_INTERRUPTIBLE);
 583                         }
 584
 585                         if (kthread && (ret = kthread_should_stop())) {
 586                                 __set_current_state(TASK_RUNNING);
 587                                 goto out;
 588                         }
 589
 590                         if (delay)
 591                                 schedule_timeout(delay);
 592
 593                         if (unlikely(freezing(current))) {
 594                                 bch2_trans_unlock(&trans);
 595                                 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 596                                 try_to_freeze();
 597                         }
 598                 } while (delay);
 599
 600                 bch2_trans_begin(&trans);
 601
 602                 k = bch2_btree_iter_peek(iter);
 603
 604                 stats->pos = iter->pos;
 605
 606                 if (!k.k)
 607                         break;
 608                 ret = bkey_err(k);
 609                 if (ret)
 610                         break;
 611                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 612                         break;
 613
 614                 if (!bkey_extent_is_direct_data(k.k))
 615                         goto next_nondata;
 616
 617                 if (btree_id == BTREE_ID_extents &&
 618                     cur_inum != k.k->p.inode) {
 619                         struct bch_inode_unpacked inode;
 620
 621                         io_opts = bch2_opts_to_inode_opts(c->opts);
 622
 623                         ret = lookup_inode(&trans,
 624                                         SPOS(0, k.k->p.inode, k.k->p.snapshot),
 625                                         &inode);
 626                         if (ret == -EINTR)
 627                                 continue;
 628
 629                         if (!ret)
 630                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 631
 632                         cur_inum = k.k->p.inode;
 633                 }
 634
 635                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 636                 case DATA_SKIP:
 637                         goto next;
 638                 case DATA_SCRUB:
 639                         BUG();
 640                 case DATA_ADD_REPLICAS:
 641                 case DATA_REWRITE:
 642                 case DATA_PROMOTE:
 643                         break;
 644                 default:
 645                         BUG();
 646                 }
 647
 648                 /* unlock before doing IO: */
 649                 bch2_bkey_buf_reassemble(&sk, c, k);
 650                 k = bkey_i_to_s_c(sk.k);
 651                 bch2_trans_unlock(&trans);
 652
 653                 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
 654                                         data_cmd, data_opts);
 655                 if (ret2) {
 656                         if (ret2 == -EINTR) {
 657                                 bch2_trans_begin(&trans);
 658                                 continue;
 659                         }
 660
 661                         if (ret2 == -ENOMEM) {
 662                                 /* memory allocation failure, wait for some IO to finish */
 663                                 bch2_move_ctxt_wait_for_io(ctxt);
 664                                 continue;
 665                         }
 666
 667                         /* XXX signal failure */
 668                         goto next;
 669                 }
 670
 671                 if (rate)
 672                         bch2_ratelimit_increment(rate, k.k->size);
 673 next:
 674                 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 675                              &stats->sectors_seen);
 676 next_nondata:
 677                 bch2_btree_iter_advance(iter);
 678                 bch2_trans_cond_resched(&trans);
 679         }
 680 out:
 681
 682         bch2_trans_iter_put(&trans, iter);
 683         ret = bch2_trans_exit(&trans) ?: ret;
 684         bch2_bkey_buf_exit(&sk, c);
 685
 686         return ret;
 687 }
 688
 689 int bch2_move_data(struct bch_fs *c,
 690                    enum btree_id start_btree_id, struct bpos start_pos,
 691                    enum btree_id end_btree_id,   struct bpos end_pos,
 692                    struct bch_ratelimit *rate,
 693                    struct write_point_specifier wp,
 694                    move_pred_fn pred, void *arg,
 695                    struct bch_move_stats *stats)
 696 {
 697         struct moving_context ctxt = { .stats = stats };
 698         enum btree_id id;
 699         int ret;
 700
 701         closure_init_stack(&ctxt.cl);
 702         INIT_LIST_HEAD(&ctxt.reads);
 703         init_waitqueue_head(&ctxt.wait);
 704
 705         stats->data_type = BCH_DATA_user;
 706
 707         for (id = start_btree_id;
 708              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 709              id++) {
 710                 stats->btree_id = id;
 711
 712                 if (id != BTREE_ID_extents &&
 713                     id != BTREE_ID_reflink)
 714                         continue;
 715
 716                 ret = __bch2_move_data(c, &ctxt, rate, wp,
 717                                        id == start_btree_id ? start_pos : POS_MIN,
 718                                        id == end_btree_id   ? end_pos   : POS_MAX,
 719                                        pred, arg, stats, id);
 720                 if (ret)
 721                         break;
 722         }
 723
 724
 725         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 726         closure_sync(&ctxt.cl);
 727
 728         EBUG_ON(atomic_read(&ctxt.write_sectors));
 729
 730         trace_move_data(c,
 731                         atomic64_read(&stats->sectors_moved),
 732                         atomic64_read(&stats->keys_moved));
 733
 734         return ret;
 735 }
 736
 737 typedef enum data_cmd (*move_btree_pred)(struct bch_fs *, void *,
 738                                          struct btree *, struct bch_io_opts *,
 739                                          struct data_opts *);
 740
 741 static int bch2_move_btree(struct bch_fs *c,
 742                            enum btree_id start_btree_id, struct bpos start_pos,
 743                            enum btree_id end_btree_id,   struct bpos end_pos,
 744                            move_btree_pred pred, void *arg,
 745                            struct bch_move_stats *stats)
 746 {
 747         bool kthread = (current->flags & PF_KTHREAD) != 0;
 748         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 749         struct btree_trans trans;
 750         struct btree_iter *iter;
 751         struct btree *b;
 752         enum btree_id id;
 753         struct data_opts data_opts;
 754         enum data_cmd cmd;
 755         int ret = 0;
 756
 757         bch2_trans_init(&trans, c, 0, 0);
 758
 759         stats->data_type = BCH_DATA_btree;
 760
 761         for (id = start_btree_id;
 762              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 763              id++) {
 764                 stats->btree_id = id;
 765
 766                 for_each_btree_node(&trans, iter, id,
 767                                     id == start_btree_id ? start_pos : POS_MIN,
 768                                     BTREE_ITER_PREFETCH, b) {
 769                         if (kthread && kthread_should_stop())
 770                                 break;
 771
 772                         if ((cmp_int(id, end_btree_id) ?:
 773                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 774                                 break;
 775
 776                         stats->pos = iter->pos;
 777
 778                         switch ((cmd = pred(c, arg, b, &io_opts, &data_opts))) {
 779                         case DATA_SKIP:
 780                                 goto next;
 781                         case DATA_SCRUB:
 782                                 BUG();
 783                         case DATA_ADD_REPLICAS:
 784                         case DATA_REWRITE:
 785                                 break;
 786                         default:
 787                                 BUG();
 788                         }
 789
 790                         ret = bch2_btree_node_rewrite(&trans, iter,
 791                                         b->data->keys.seq, 0) ?: ret;
 792 next:
 793                         bch2_trans_cond_resched(&trans);
 794                 }
 795
 796                 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
 797                 if (kthread && kthread_should_stop())
 798                         break;
 799         }
 800
 801         bch2_trans_exit(&trans);
 802
 803         if (ret)
 804                 bch_err(c, "error %i in bch2_move_btree", ret);
 805
 806         return ret;
 807 }
 808
 809 #if 0
 810 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 811                                 struct bkey_s_c k,
 812                                 struct bch_io_opts *io_opts,
 813                                 struct data_opts *data_opts)
 814 {
 815         return DATA_SCRUB;
 816 }
 817 #endif
 818
 819 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 820                                       struct bkey_s_c k,
 821                                       struct bch_io_opts *io_opts,
 822                                       struct data_opts *data_opts)
 823 {
 824         unsigned nr_good = bch2_bkey_durability(c, k);
 825         unsigned replicas = 0;
 826
 827         switch (k.k->type) {
 828         case KEY_TYPE_btree_ptr:
 829                 replicas = c->opts.metadata_replicas;
 830                 break;
 831         case KEY_TYPE_extent:
 832                 replicas = io_opts->data_replicas;
 833                 break;
 834         }
 835
 836         if (!nr_good || nr_good >= replicas)
 837                 return DATA_SKIP;
 838
 839         data_opts->target               = 0;
 840         data_opts->nr_replicas          = 1;
 841         data_opts->btree_insert_flags   = 0;
 842         return DATA_ADD_REPLICAS;
 843 }
 844
 845 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 846                                   struct bkey_s_c k,
 847                                   struct bch_io_opts *io_opts,
 848                                   struct data_opts *data_opts)
 849 {
 850         struct bch_ioctl_data *op = arg;
 851
 852         if (!bch2_bkey_has_device(k, op->migrate.dev))
 853                 return DATA_SKIP;
 854
 855         data_opts->target               = 0;
 856         data_opts->nr_replicas          = 1;
 857         data_opts->btree_insert_flags   = 0;
 858         data_opts->rewrite_dev          = op->migrate.dev;
 859         return DATA_REWRITE;
 860 }
 861
 862 static enum data_cmd rereplicate_btree_pred(struct bch_fs *c, void *arg,
 863                                             struct btree *b,
 864                                             struct bch_io_opts *io_opts,
 865                                             struct data_opts *data_opts)
 866 {
 867         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 868 }
 869
 870 static enum data_cmd migrate_btree_pred(struct bch_fs *c, void *arg,
 871                                         struct btree *b,
 872                                         struct bch_io_opts *io_opts,
 873                                         struct data_opts *data_opts)
 874 {
 875         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 876 }
 877
 878 static bool bformat_needs_redo(struct bkey_format *f)
 879 {
 880         unsigned i;
 881
 882         for (i = 0; i < f->nr_fields; i++) {
 883                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 884                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 885                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 886
 887                 if (f->bits_per_field[i] > unpacked_bits)
 888                         return true;
 889
 890                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 891                         return true;
 892
 893                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 894                      unpacked_mask) <
 895                     field_offset)
 896                         return true;
 897         }
 898
 899         return false;
 900 }
 901
 902 static enum data_cmd rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 903                                             struct btree *b,
 904                                             struct bch_io_opts *io_opts,
 905                                             struct data_opts *data_opts)
 906 {
 907         if (b->version_ondisk != c->sb.version ||
 908             btree_node_need_rewrite(b) ||
 909             bformat_needs_redo(&b->format)) {
 910                 data_opts->target               = 0;
 911                 data_opts->nr_replicas          = 1;
 912                 data_opts->btree_insert_flags   = 0;
 913                 return DATA_REWRITE;
 914         }
 915
 916         return DATA_SKIP;
 917 }
 918
 919 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 920 {
 921         int ret;
 922
 923         ret = bch2_move_btree(c,
 924                               0,                POS_MIN,
 925                               BTREE_ID_NR,      SPOS_MAX,
 926                               rewrite_old_nodes_pred, c, stats);
 927         if (!ret) {
 928                 mutex_lock(&c->sb_lock);
 929                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
 930                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 931                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
 932                 bch2_write_super(c);
 933                 mutex_unlock(&c->sb_lock);
 934         }
 935
 936         return ret;
 937 }
 938
 939 int bch2_data_job(struct bch_fs *c,
 940                   struct bch_move_stats *stats,
 941                   struct bch_ioctl_data op)
 942 {
 943         int ret = 0;
 944
 945         switch (op.op) {
 946         case BCH_DATA_OP_REREPLICATE:
 947                 stats->data_type = BCH_DATA_journal;
 948                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 949
 950                 ret = bch2_move_btree(c,
 951                                       op.start_btree,   op.start_pos,
 952                                       op.end_btree,     op.end_pos,
 953                                       rereplicate_btree_pred, c, stats) ?: ret;
 954
 955                 closure_wait_event(&c->btree_interior_update_wait,
 956                                    !bch2_btree_interior_updates_nr_pending(c));
 957
 958                 ret = bch2_replicas_gc2(c) ?: ret;
 959
 960                 ret = bch2_move_data(c,
 961                                      op.start_btree,    op.start_pos,
 962                                      op.end_btree,      op.end_pos,
 963                                      NULL, writepoint_hashed((unsigned long) current),
 964                                      rereplicate_pred, c, stats) ?: ret;
 965                 ret = bch2_replicas_gc2(c) ?: ret;
 966                 break;
 967         case BCH_DATA_OP_MIGRATE:
 968                 if (op.migrate.dev >= c->sb.nr_devices)
 969                         return -EINVAL;
 970
 971                 stats->data_type = BCH_DATA_journal;
 972                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
 973
 974                 ret = bch2_move_btree(c,
 975                                       op.start_btree,   op.start_pos,
 976                                       op.end_btree,     op.end_pos,
 977                                       migrate_btree_pred, &op, stats) ?: ret;
 978                 ret = bch2_replicas_gc2(c) ?: ret;
 979
 980                 ret = bch2_move_data(c,
 981                                      op.start_btree,    op.start_pos,
 982                                      op.end_btree,      op.end_pos,
 983                                      NULL, writepoint_hashed((unsigned long) current),
 984                                      migrate_pred, &op, stats) ?: ret;
 985                 ret = bch2_replicas_gc2(c) ?: ret;
 986                 break;
 987         case BCH_DATA_OP_REWRITE_OLD_NODES:
 988                 ret = bch2_scan_old_btree_nodes(c, stats);
 989                 break;
 990         default:
 991                 ret = -EINVAL;
 992         }
 993
 994         return ret;
 995 }