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

   1
   2 #include "bcachefs.h"
   3 #include "alloc_foreground.h"
   4 #include "btree_gc.h"
   5 #include "btree_update.h"
   6 #include "buckets.h"
   7 #include "disk_groups.h"
   8 #include "inode.h"
   9 #include "io.h"
  10 #include "journal_reclaim.h"
  11 #include "move.h"
  12 #include "replicas.h"
  13 #include "super-io.h"
  14 #include "keylist.h"
  15
  16 #include <linux/ioprio.h>
  17 #include <linux/kthread.h>
  18
  19 #include <trace/events/bcachefs.h>
  20
  21 #define SECTORS_IN_FLIGHT_PER_DEVICE    2048
  22
  23 struct moving_io {
  24         struct list_head        list;
  25         struct closure          cl;
  26         bool                    read_completed;
  27
  28         unsigned                read_sectors;
  29         unsigned                write_sectors;
  30
  31         struct bch_read_bio     rbio;
  32
  33         struct migrate_write    write;
  34         /* Must be last since it is variable size */
  35         struct bio_vec          bi_inline_vecs[0];
  36 };
  37
  38 struct moving_context {
  39         /* Closure for waiting on all reads and writes to complete */
  40         struct closure          cl;
  41
  42         struct bch_move_stats   *stats;
  43
  44         struct list_head        reads;
  45
  46         /* in flight sectors: */
  47         atomic_t                read_sectors;
  48         atomic_t                write_sectors;
  49
  50         wait_queue_head_t       wait;
  51 };
  52
  53 static int bch2_migrate_index_update(struct bch_write_op *op)
  54 {
  55         struct bch_fs *c = op->c;
  56         struct migrate_write *m =
  57                 container_of(op, struct migrate_write, op);
  58         struct keylist *keys = &op->insert_keys;
  59         struct btree_iter iter;
  60         int ret = 0;
  61
  62         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS,
  63                              bkey_start_pos(&bch2_keylist_front(keys)->k),
  64                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
  65
  66         while (1) {
  67                 struct bkey_s_c k = bch2_btree_iter_peek_slot(&iter);
  68                 struct bkey_i_extent *insert, *new =
  69                         bkey_i_to_extent(bch2_keylist_front(keys));
  70                 BKEY_PADDED(k) _new, _insert;
  71                 const union bch_extent_entry *entry;
  72                 struct extent_ptr_decoded p;
  73                 bool did_work = false;
  74                 int nr;
  75
  76                 if (btree_iter_err(k)) {
  77                         ret = bch2_btree_iter_unlock(&iter);
  78                         break;
  79                 }
  80
  81                 if (bversion_cmp(k.k->version, new->k.version) ||
  82                     !bkey_extent_is_data(k.k) ||
  83                     !bch2_extent_matches_ptr(c, bkey_s_c_to_extent(k),
  84                                              m->ptr, m->offset))
  85                         goto nomatch;
  86
  87                 if (m->data_cmd == DATA_REWRITE &&
  88                     !bch2_extent_has_device(bkey_s_c_to_extent(k),
  89                                             m->data_opts.rewrite_dev))
  90                         goto nomatch;
  91
  92                 bkey_reassemble(&_insert.k, k);
  93                 insert = bkey_i_to_extent(&_insert.k);
  94
  95                 bkey_copy(&_new.k, bch2_keylist_front(keys));
  96                 new = bkey_i_to_extent(&_new.k);
  97
  98                 bch2_cut_front(iter.pos, &insert->k_i);
  99                 bch2_cut_back(new->k.p, &insert->k);
 100                 bch2_cut_back(insert->k.p, &new->k);
 101
 102                 if (m->data_cmd == DATA_REWRITE)
 103                         bch2_bkey_drop_device(extent_i_to_s(insert).s,
 104                                               m->data_opts.rewrite_dev);
 105
 106                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 107                         if (bch2_extent_has_device(extent_i_to_s_c(insert), p.ptr.dev)) {
 108                                 /*
 109                                  * raced with another move op? extent already
 110                                  * has a pointer to the device we just wrote
 111                                  * data to
 112                                  */
 113                                 continue;
 114                         }
 115
 116                         bch2_extent_ptr_decoded_append(insert, &p);
 117                         did_work = true;
 118                 }
 119
 120                 if (!did_work)
 121                         goto nomatch;
 122
 123                 bch2_extent_narrow_crcs(insert,
 124                                 (struct bch_extent_crc_unpacked) { 0 });
 125                 bch2_extent_normalize(c, extent_i_to_s(insert).s);
 126                 bch2_extent_mark_replicas_cached(c, extent_i_to_s(insert),
 127                                                  op->opts.background_target,
 128                                                  op->opts.data_replicas);
 129
 130                 /*
 131                  * It's possible we race, and for whatever reason the extent now
 132                  * has fewer replicas than when we last looked at it - meaning
 133                  * we need to get a disk reservation here:
 134                  */
 135                 nr = bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&insert->k_i)) -
 136                         (bch2_bkey_nr_dirty_ptrs(k) + m->nr_ptrs_reserved);
 137                 if (nr > 0) {
 138                         /*
 139                          * can't call bch2_disk_reservation_add() with btree
 140                          * locks held, at least not without a song and dance
 141                          */
 142                         bch2_btree_iter_unlock(&iter);
 143
 144                         ret = bch2_disk_reservation_add(c, &op->res,
 145                                         keylist_sectors(keys) * nr, 0);
 146                         if (ret)
 147                                 goto out;
 148
 149                         m->nr_ptrs_reserved += nr;
 150                         goto next;
 151                 }
 152
 153                 ret = bch2_btree_insert_at(c, &op->res,
 154                                 op_journal_seq(op),
 155                                 BTREE_INSERT_ATOMIC|
 156                                 BTREE_INSERT_NOFAIL|
 157                                 BTREE_INSERT_USE_RESERVE|
 158                                 m->data_opts.btree_insert_flags,
 159                                 BTREE_INSERT_ENTRY(&iter, &insert->k_i));
 160                 if (!ret)
 161                         atomic_long_inc(&c->extent_migrate_done);
 162                 if (ret == -EINTR)
 163                         ret = 0;
 164                 if (ret)
 165                         break;
 166 next:
 167                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 168                         bch2_keylist_pop_front(keys);
 169                         if (bch2_keylist_empty(keys))
 170                                 goto out;
 171                 }
 172
 173                 bch2_cut_front(iter.pos, bch2_keylist_front(keys));
 174                 continue;
 175 nomatch:
 176                 if (m->ctxt)
 177                         atomic64_add(k.k->p.offset - iter.pos.offset,
 178                                      &m->ctxt->stats->sectors_raced);
 179                 atomic_long_inc(&c->extent_migrate_raced);
 180                 trace_move_race(&new->k);
 181                 bch2_btree_iter_next_slot(&iter);
 182                 goto next;
 183         }
 184 out:
 185         bch2_btree_iter_unlock(&iter);
 186         return ret;
 187 }
 188
 189 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
 190 {
 191         /* write bio must own pages: */
 192         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 193
 194         m->ptr          = rbio->pick.ptr;
 195         m->offset       = rbio->pos.offset - rbio->pick.crc.offset;
 196         m->op.devs_have = rbio->devs_have;
 197         m->op.pos       = rbio->pos;
 198         m->op.version   = rbio->version;
 199         m->op.crc       = rbio->pick.crc;
 200         m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
 201
 202         if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
 203                 m->op.nonce     = m->op.crc.nonce + m->op.crc.offset;
 204                 m->op.csum_type = m->op.crc.csum_type;
 205         }
 206
 207         if (m->data_cmd == DATA_REWRITE)
 208                 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
 209 }
 210
 211 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 212                             struct write_point_specifier wp,
 213                             struct bch_io_opts io_opts,
 214                             enum data_cmd data_cmd,
 215                             struct data_opts data_opts,
 216                             struct bkey_s_c k)
 217 {
 218         int ret;
 219
 220         m->data_cmd     = data_cmd;
 221         m->data_opts    = data_opts;
 222         m->nr_ptrs_reserved = 0;
 223
 224         bch2_write_op_init(&m->op, c, io_opts);
 225         m->op.compression_type =
 226                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 227                                              io_opts.compression];
 228         m->op.target    = data_opts.target,
 229         m->op.write_point = wp;
 230
 231         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 232                 m->op.alloc_reserve = RESERVE_MOVINGGC;
 233
 234         m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
 235                 BCH_WRITE_PAGES_STABLE|
 236                 BCH_WRITE_PAGES_OWNED|
 237                 BCH_WRITE_DATA_ENCODED;
 238
 239         m->op.nr_replicas       = 1;
 240         m->op.nr_replicas_required = 1;
 241         m->op.index_update_fn   = bch2_migrate_index_update;
 242
 243         switch (data_cmd) {
 244         case DATA_ADD_REPLICAS: {
 245                 int nr = (int) io_opts.data_replicas -
 246                         bch2_bkey_nr_dirty_ptrs(k);
 247
 248                 if (nr > 0) {
 249                         m->op.nr_replicas = m->nr_ptrs_reserved = nr;
 250
 251                         ret = bch2_disk_reservation_get(c, &m->op.res,
 252                                         k.k->size, m->op.nr_replicas, 0);
 253                         if (ret)
 254                                 return ret;
 255                 }
 256                 break;
 257         }
 258         case DATA_REWRITE: {
 259                 const union bch_extent_entry *entry;
 260                 struct extent_ptr_decoded p;
 261                 unsigned compressed_sectors = 0;
 262
 263                 extent_for_each_ptr_decode(bkey_s_c_to_extent(k), p, entry)
 264                         if (!p.ptr.cached &&
 265                             p.crc.compression_type != BCH_COMPRESSION_NONE &&
 266                             bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
 267                                 compressed_sectors += p.crc.compressed_size;
 268
 269                 if (compressed_sectors) {
 270                         ret = bch2_disk_reservation_add(c, &m->op.res,
 271                                         compressed_sectors,
 272                                         BCH_DISK_RESERVATION_NOFAIL);
 273                         if (ret)
 274                                 return ret;
 275                 }
 276                 break;
 277         }
 278         case DATA_PROMOTE:
 279                 m->op.flags     |= BCH_WRITE_ALLOC_NOWAIT;
 280                 m->op.flags     |= BCH_WRITE_CACHED;
 281                 break;
 282         default:
 283                 BUG();
 284         }
 285
 286         return 0;
 287 }
 288
 289 static void move_free(struct closure *cl)
 290 {
 291         struct moving_io *io = container_of(cl, struct moving_io, cl);
 292         struct moving_context *ctxt = io->write.ctxt;
 293         struct bio_vec *bv;
 294         int i;
 295
 296         bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 297
 298         bio_for_each_segment_all(bv, &io->write.op.wbio.bio, i)
 299                 if (bv->bv_page)
 300                         __free_page(bv->bv_page);
 301
 302         wake_up(&ctxt->wait);
 303
 304         kfree(io);
 305 }
 306
 307 static void move_write_done(struct closure *cl)
 308 {
 309         struct moving_io *io = container_of(cl, struct moving_io, cl);
 310
 311         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 312         closure_return_with_destructor(cl, move_free);
 313 }
 314
 315 static void move_write(struct closure *cl)
 316 {
 317         struct moving_io *io = container_of(cl, struct moving_io, cl);
 318
 319         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 320                 closure_return_with_destructor(cl, move_free);
 321                 return;
 322         }
 323
 324         bch2_migrate_read_done(&io->write, &io->rbio);
 325
 326         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 327         closure_call(&io->write.op.cl, bch2_write, NULL, cl);
 328         continue_at(cl, move_write_done, NULL);
 329 }
 330
 331 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
 332 {
 333         struct moving_io *io =
 334                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 335
 336         return io && io->read_completed ? io : NULL;
 337 }
 338
 339 static void move_read_endio(struct bio *bio)
 340 {
 341         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 342         struct moving_context *ctxt = io->write.ctxt;
 343
 344         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 345         io->read_completed = true;
 346
 347         if (next_pending_write(ctxt))
 348                 wake_up(&ctxt->wait);
 349
 350         closure_put(&ctxt->cl);
 351 }
 352
 353 static void do_pending_writes(struct moving_context *ctxt)
 354 {
 355         struct moving_io *io;
 356
 357         while ((io = next_pending_write(ctxt))) {
 358                 list_del(&io->list);
 359                 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
 360         }
 361 }
 362
 363 #define move_ctxt_wait_event(_ctxt, _cond)                      \
 364 do {                                                            \
 365         do_pending_writes(_ctxt);                               \
 366                                                                 \
 367         if (_cond)                                              \
 368                 break;                                          \
 369         __wait_event((_ctxt)->wait,                             \
 370                      next_pending_write(_ctxt) || (_cond));     \
 371 } while (1)
 372
 373 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 374 {
 375         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 376
 377         move_ctxt_wait_event(ctxt,
 378                 !atomic_read(&ctxt->write_sectors) ||
 379                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 380 }
 381
 382 static int bch2_move_extent(struct bch_fs *c,
 383                             struct moving_context *ctxt,
 384                             struct write_point_specifier wp,
 385                             struct bch_io_opts io_opts,
 386                             struct bkey_s_c_extent e,
 387                             enum data_cmd data_cmd,
 388                             struct data_opts data_opts)
 389 {
 390         struct moving_io *io;
 391         const union bch_extent_entry *entry;
 392         struct extent_ptr_decoded p;
 393         unsigned sectors = e.k->size, pages;
 394         int ret = -ENOMEM;
 395
 396         move_ctxt_wait_event(ctxt,
 397                 atomic_read(&ctxt->write_sectors) <
 398                 SECTORS_IN_FLIGHT_PER_DEVICE);
 399
 400         move_ctxt_wait_event(ctxt,
 401                 atomic_read(&ctxt->read_sectors) <
 402                 SECTORS_IN_FLIGHT_PER_DEVICE);
 403
 404         /* write path might have to decompress data: */
 405         extent_for_each_ptr_decode(e, p, entry)
 406                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 407
 408         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 409         io = kzalloc(sizeof(struct moving_io) +
 410                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 411         if (!io)
 412                 goto err;
 413
 414         io->write.ctxt          = ctxt;
 415         io->read_sectors        = e.k->size;
 416         io->write_sectors       = e.k->size;
 417
 418         bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
 419         bio_set_prio(&io->write.op.wbio.bio,
 420                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 421         io->write.op.wbio.bio.bi_iter.bi_size = sectors << 9;
 422
 423         bch2_bio_map(&io->write.op.wbio.bio, NULL);
 424         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, GFP_KERNEL))
 425                 goto err_free;
 426
 427         io->rbio.opts = io_opts;
 428         bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
 429         io->rbio.bio.bi_vcnt = pages;
 430         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 431         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 432
 433         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 434         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(e.k);
 435         io->rbio.bio.bi_end_io          = move_read_endio;
 436
 437         ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
 438                                       data_cmd, data_opts, e.s_c);
 439         if (ret)
 440                 goto err_free_pages;
 441
 442         atomic64_inc(&ctxt->stats->keys_moved);
 443         atomic64_add(e.k->size, &ctxt->stats->sectors_moved);
 444
 445         trace_move_extent(e.k);
 446
 447         atomic_add(io->read_sectors, &ctxt->read_sectors);
 448         list_add_tail(&io->list, &ctxt->reads);
 449
 450         /*
 451          * dropped by move_read_endio() - guards against use after free of
 452          * ctxt when doing wakeup
 453          */
 454         closure_get(&ctxt->cl);
 455         bch2_read_extent(c, &io->rbio, e.s_c,
 456                          BCH_READ_NODECODE|
 457                          BCH_READ_LAST_FRAGMENT);
 458         return 0;
 459 err_free_pages:
 460         bio_free_pages(&io->write.op.wbio.bio);
 461 err_free:
 462         kfree(io);
 463 err:
 464         trace_move_alloc_fail(e.k);
 465         return ret;
 466 }
 467
 468 int bch2_move_data(struct bch_fs *c,
 469                    struct bch_ratelimit *rate,
 470                    struct write_point_specifier wp,
 471                    struct bpos start,
 472                    struct bpos end,
 473                    move_pred_fn pred, void *arg,
 474                    struct bch_move_stats *stats)
 475 {
 476         bool kthread = (current->flags & PF_KTHREAD) != 0;
 477         struct moving_context ctxt = { .stats = stats };
 478         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 479         BKEY_PADDED(k) tmp;
 480         struct bkey_s_c k;
 481         struct data_opts data_opts;
 482         enum data_cmd data_cmd;
 483         u64 delay, cur_inum = U64_MAX;
 484         int ret = 0, ret2;
 485
 486         closure_init_stack(&ctxt.cl);
 487         INIT_LIST_HEAD(&ctxt.reads);
 488         init_waitqueue_head(&ctxt.wait);
 489
 490         stats->data_type = BCH_DATA_USER;
 491         bch2_btree_iter_init(&stats->iter, c, BTREE_ID_EXTENTS, start,
 492                              BTREE_ITER_PREFETCH);
 493
 494         if (rate)
 495                 bch2_ratelimit_reset(rate);
 496
 497         while (1) {
 498                 do {
 499                         delay = rate ? bch2_ratelimit_delay(rate) : 0;
 500
 501                         if (delay) {
 502                                 bch2_btree_iter_unlock(&stats->iter);
 503                                 set_current_state(TASK_INTERRUPTIBLE);
 504                         }
 505
 506                         if (kthread && (ret = kthread_should_stop())) {
 507                                 __set_current_state(TASK_RUNNING);
 508                                 goto out;
 509                         }
 510
 511                         if (delay)
 512                                 schedule_timeout(delay);
 513
 514                         if (unlikely(freezing(current))) {
 515                                 bch2_btree_iter_unlock(&stats->iter);
 516                                 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 517                                 try_to_freeze();
 518                         }
 519                 } while (delay);
 520 peek:
 521                 k = bch2_btree_iter_peek(&stats->iter);
 522                 if (!k.k)
 523                         break;
 524                 ret = btree_iter_err(k);
 525                 if (ret)
 526                         break;
 527                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 528                         break;
 529
 530                 if (!bkey_extent_is_data(k.k))
 531                         goto next_nondata;
 532
 533                 if (cur_inum != k.k->p.inode) {
 534                         struct bch_inode_unpacked inode;
 535
 536                         /* don't hold btree locks while looking up inode: */
 537                         bch2_btree_iter_unlock(&stats->iter);
 538
 539                         io_opts = bch2_opts_to_inode_opts(c->opts);
 540                         if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
 541                                 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
 542                         cur_inum = k.k->p.inode;
 543                         goto peek;
 544                 }
 545
 546                 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
 547                 case DATA_SKIP:
 548                         goto next;
 549                 case DATA_SCRUB:
 550                         BUG();
 551                 case DATA_ADD_REPLICAS:
 552                 case DATA_REWRITE:
 553                 case DATA_PROMOTE:
 554                         break;
 555                 default:
 556                         BUG();
 557                 }
 558
 559                 /* unlock before doing IO: */
 560                 bkey_reassemble(&tmp.k, k);
 561                 k = bkey_i_to_s_c(&tmp.k);
 562                 bch2_btree_iter_unlock(&stats->iter);
 563
 564                 ret2 = bch2_move_extent(c, &ctxt, wp, io_opts,
 565                                         bkey_s_c_to_extent(k),
 566                                         data_cmd, data_opts);
 567                 if (ret2) {
 568                         if (ret2 == -ENOMEM) {
 569                                 /* memory allocation failure, wait for some IO to finish */
 570                                 bch2_move_ctxt_wait_for_io(&ctxt);
 571                                 continue;
 572                         }
 573
 574                         /* XXX signal failure */
 575                         goto next;
 576                 }
 577
 578                 if (rate)
 579                         bch2_ratelimit_increment(rate, k.k->size);
 580 next:
 581                 atomic64_add(k.k->size * bch2_bkey_nr_dirty_ptrs(k),
 582                              &stats->sectors_seen);
 583 next_nondata:
 584                 bch2_btree_iter_next(&stats->iter);
 585                 bch2_btree_iter_cond_resched(&stats->iter);
 586         }
 587 out:
 588         bch2_btree_iter_unlock(&stats->iter);
 589
 590         move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
 591         closure_sync(&ctxt.cl);
 592
 593         EBUG_ON(atomic_read(&ctxt.write_sectors));
 594
 595         trace_move_data(c,
 596                         atomic64_read(&stats->sectors_moved),
 597                         atomic64_read(&stats->keys_moved));
 598
 599         return ret;
 600 }
 601
 602 static int bch2_gc_data_replicas(struct bch_fs *c)
 603 {
 604         struct btree_iter iter;
 605         struct bkey_s_c k;
 606         int ret;
 607
 608         mutex_lock(&c->replicas_gc_lock);
 609         bch2_replicas_gc_start(c, (1 << BCH_DATA_USER)|(1 << BCH_DATA_CACHED));
 610
 611         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
 612                            BTREE_ITER_PREFETCH, k) {
 613                 ret = bch2_mark_bkey_replicas(c, k);
 614                 if (ret)
 615                         break;
 616         }
 617         ret = bch2_btree_iter_unlock(&iter) ?: ret;
 618
 619         bch2_replicas_gc_end(c, ret);
 620         mutex_unlock(&c->replicas_gc_lock);
 621
 622         return ret;
 623 }
 624
 625 static int bch2_gc_btree_replicas(struct bch_fs *c)
 626 {
 627         struct btree_iter iter;
 628         struct btree *b;
 629         unsigned id;
 630         int ret = 0;
 631
 632         mutex_lock(&c->replicas_gc_lock);
 633         bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
 634
 635         for (id = 0; id < BTREE_ID_NR; id++) {
 636                 for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 637                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 638
 639                         bch2_btree_iter_cond_resched(&iter);
 640                 }
 641
 642                 ret = bch2_btree_iter_unlock(&iter) ?: ret;
 643         }
 644
 645         bch2_replicas_gc_end(c, ret);
 646         mutex_unlock(&c->replicas_gc_lock);
 647
 648         return ret;
 649 }
 650
 651 static int bch2_move_btree(struct bch_fs *c,
 652                            move_pred_fn pred,
 653                            void *arg,
 654                            struct bch_move_stats *stats)
 655 {
 656         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 657         struct btree *b;
 658         unsigned id;
 659         struct data_opts data_opts;
 660         enum data_cmd cmd;
 661         int ret = 0;
 662
 663         stats->data_type = BCH_DATA_BTREE;
 664
 665         for (id = 0; id < BTREE_ID_NR; id++) {
 666                 for_each_btree_node(&stats->iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 667                         switch ((cmd = pred(c, arg,
 668                                             bkey_i_to_s_c(&b->key),
 669                                             &io_opts, &data_opts))) {
 670                         case DATA_SKIP:
 671                                 goto next;
 672                         case DATA_SCRUB:
 673                                 BUG();
 674                         case DATA_ADD_REPLICAS:
 675                         case DATA_REWRITE:
 676                                 break;
 677                         default:
 678                                 BUG();
 679                         }
 680
 681                         ret = bch2_btree_node_rewrite(c, &stats->iter,
 682                                         b->data->keys.seq, 0) ?: ret;
 683 next:
 684                         bch2_btree_iter_cond_resched(&stats->iter);
 685                 }
 686
 687                 ret = bch2_btree_iter_unlock(&stats->iter) ?: ret;
 688         }
 689
 690         return ret;
 691 }
 692
 693 #if 0
 694 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
 695                                 struct bkey_s_c k,
 696                                 struct bch_io_opts *io_opts,
 697                                 struct data_opts *data_opts)
 698 {
 699         return DATA_SCRUB;
 700 }
 701 #endif
 702
 703 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
 704                                       struct bkey_s_c k,
 705                                       struct bch_io_opts *io_opts,
 706                                       struct data_opts *data_opts)
 707 {
 708         unsigned nr_good = bch2_bkey_durability(c, k);
 709         unsigned replicas = 0;
 710
 711         switch (k.k->type) {
 712         case KEY_TYPE_btree_ptr:
 713                 replicas = c->opts.metadata_replicas;
 714                 break;
 715         case KEY_TYPE_extent:
 716                 replicas = io_opts->data_replicas;
 717                 break;
 718         }
 719
 720         if (!nr_good || nr_good >= replicas)
 721                 return DATA_SKIP;
 722
 723         data_opts->target               = 0;
 724         data_opts->btree_insert_flags   = 0;
 725         return DATA_ADD_REPLICAS;
 726 }
 727
 728 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
 729                                   struct bkey_s_c k,
 730                                   struct bch_io_opts *io_opts,
 731                                   struct data_opts *data_opts)
 732 {
 733         struct bch_ioctl_data *op = arg;
 734
 735         if (!bch2_bkey_has_device(k, op->migrate.dev))
 736                 return DATA_SKIP;
 737
 738         data_opts->target               = 0;
 739         data_opts->btree_insert_flags   = 0;
 740         data_opts->rewrite_dev          = op->migrate.dev;
 741         return DATA_REWRITE;
 742 }
 743
 744 int bch2_data_job(struct bch_fs *c,
 745                   struct bch_move_stats *stats,
 746                   struct bch_ioctl_data op)
 747 {
 748         int ret = 0;
 749
 750         switch (op.op) {
 751         case BCH_DATA_OP_REREPLICATE:
 752                 stats->data_type = BCH_DATA_JOURNAL;
 753                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
 754
 755                 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
 756                 ret = bch2_gc_btree_replicas(c) ?: ret;
 757
 758                 ret = bch2_move_data(c, NULL,
 759                                      writepoint_hashed((unsigned long) current),
 760                                      op.start,
 761                                      op.end,
 762                                      rereplicate_pred, c, stats) ?: ret;
 763                 ret = bch2_gc_data_replicas(c) ?: ret;
 764                 break;
 765         case BCH_DATA_OP_MIGRATE:
 766                 if (op.migrate.dev >= c->sb.nr_devices)
 767                         return -EINVAL;
 768
 769                 stats->data_type = BCH_DATA_JOURNAL;
 770                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
 771
 772                 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
 773                 ret = bch2_gc_btree_replicas(c) ?: ret;
 774
 775                 ret = bch2_move_data(c, NULL,
 776                                      writepoint_hashed((unsigned long) current),
 777                                      op.start,
 778                                      op.end,
 779                                      migrate_pred, &op, stats) ?: ret;
 780                 ret = bch2_gc_data_replicas(c) ?: ret;
 781                 break;
 782         default:
 783                 ret = -EINVAL;
 784         }
 785
 786         return ret;
 787 }