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_background.h"
   5 #include "alloc_foreground.h"
   6 #include "backpointers.h"
   7 #include "bkey_buf.h"
   8 #include "btree_gc.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "disk_groups.h"
  12 #include "ec.h"
  13 #include "errcode.h"
  14 #include "error.h"
  15 #include "inode.h"
  16 #include "io.h"
  17 #include "journal_reclaim.h"
  18 #include "move.h"
  19 #include "replicas.h"
  20 #include "super-io.h"
  21 #include "keylist.h"
  22
  23 #include <linux/ioprio.h>
  24 #include <linux/kthread.h>
  25
  26 #include <trace/events/bcachefs.h>
  27
  28 static void progress_list_add(struct bch_fs *c, struct bch_move_stats *stats)
  29 {
  30         mutex_lock(&c->data_progress_lock);
  31         list_add(&stats->list, &c->data_progress_list);
  32         mutex_unlock(&c->data_progress_lock);
  33 }
  34
  35 static void progress_list_del(struct bch_fs *c, struct bch_move_stats *stats)
  36 {
  37         mutex_lock(&c->data_progress_lock);
  38         list_del(&stats->list);
  39         mutex_unlock(&c->data_progress_lock);
  40 }
  41
  42 struct moving_io {
  43         struct list_head        list;
  44         struct closure          cl;
  45         bool                    read_completed;
  46
  47         unsigned                read_sectors;
  48         unsigned                write_sectors;
  49
  50         struct bch_read_bio     rbio;
  51
  52         struct data_update      write;
  53         /* Must be last since it is variable size */
  54         struct bio_vec          bi_inline_vecs[0];
  55 };
  56
  57 static void move_free(struct moving_io *io)
  58 {
  59         struct moving_context *ctxt = io->write.ctxt;
  60         struct bch_fs *c = ctxt->c;
  61
  62         bch2_data_update_exit(&io->write);
  63         wake_up(&ctxt->wait);
  64         percpu_ref_put(&c->writes);
  65         kfree(io);
  66 }
  67
  68 static void move_write_done(struct bch_write_op *op)
  69 {
  70         struct moving_io *io = container_of(op, struct moving_io, write.op);
  71         struct moving_context *ctxt = io->write.ctxt;
  72
  73         if (io->write.op.error)
  74                 ctxt->write_error = true;
  75
  76         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
  77         move_free(io);
  78         closure_put(&ctxt->cl);
  79 }
  80
  81 static void move_write(struct moving_io *io)
  82 {
  83         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
  84                 move_free(io);
  85                 return;
  86         }
  87
  88         closure_get(&io->write.ctxt->cl);
  89         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
  90
  91         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
  92 }
  93
  94 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
  95 {
  96         struct moving_io *io =
  97                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
  98
  99         return io && io->read_completed ? io : NULL;
 100 }
 101
 102 static void move_read_endio(struct bio *bio)
 103 {
 104         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 105         struct moving_context *ctxt = io->write.ctxt;
 106
 107         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 108         io->read_completed = true;
 109
 110         wake_up(&ctxt->wait);
 111         closure_put(&ctxt->cl);
 112 }
 113
 114 static void do_pending_writes(struct moving_context *ctxt, struct btree_trans *trans)
 115 {
 116         struct moving_io *io;
 117
 118         if (trans)
 119                 bch2_trans_unlock(trans);
 120
 121         while ((io = next_pending_write(ctxt))) {
 122                 list_del(&io->list);
 123                 move_write(io);
 124         }
 125 }
 126
 127 #define move_ctxt_wait_event(_ctxt, _trans, _cond)              \
 128 do {                                                            \
 129         do_pending_writes(_ctxt, _trans);                       \
 130                                                                 \
 131         if (_cond)                                              \
 132                 break;                                          \
 133         __wait_event((_ctxt)->wait,                             \
 134                      next_pending_write(_ctxt) || (_cond));     \
 135 } while (1)
 136
 137 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 138                                        struct btree_trans *trans)
 139 {
 140         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 141
 142         move_ctxt_wait_event(ctxt, trans,
 143                 !atomic_read(&ctxt->write_sectors) ||
 144                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 145 }
 146
 147 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 148 {
 149         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 150         closure_sync(&ctxt->cl);
 151         EBUG_ON(atomic_read(&ctxt->write_sectors));
 152
 153         if (ctxt->stats) {
 154                 progress_list_del(ctxt->c, ctxt->stats);
 155
 156                 trace_move_data(ctxt->c,
 157                                 atomic64_read(&ctxt->stats->sectors_moved),
 158                                 atomic64_read(&ctxt->stats->keys_moved));
 159         }
 160 }
 161
 162 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 163                            struct bch_fs *c,
 164                            struct bch_ratelimit *rate,
 165                            struct bch_move_stats *stats,
 166                            struct write_point_specifier wp,
 167                            bool wait_on_copygc)
 168 {
 169         memset(ctxt, 0, sizeof(*ctxt));
 170
 171         ctxt->c         = c;
 172         ctxt->rate      = rate;
 173         ctxt->stats     = stats;
 174         ctxt->wp        = wp;
 175         ctxt->wait_on_copygc = wait_on_copygc;
 176
 177         closure_init_stack(&ctxt->cl);
 178         INIT_LIST_HEAD(&ctxt->reads);
 179         init_waitqueue_head(&ctxt->wait);
 180
 181         if (stats) {
 182                 progress_list_add(c, stats);
 183                 stats->data_type = BCH_DATA_user;
 184         }
 185 }
 186
 187 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 188 {
 189         memset(stats, 0, sizeof(*stats));
 190         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 191 }
 192
 193 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 194                                  struct btree_iter *iter,
 195                                  struct bkey_s_c k,
 196                                  struct data_update_opts data_opts)
 197 {
 198         struct bch_fs *c = trans->c;
 199         struct bkey_i *n;
 200         int ret;
 201
 202         n = bch2_bkey_make_mut(trans, k);
 203         ret = PTR_ERR_OR_ZERO(n);
 204         if (ret)
 205                 return ret;
 206
 207         while (data_opts.kill_ptrs) {
 208                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 209                 struct bch_extent_ptr *ptr;
 210
 211                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 212                 data_opts.kill_ptrs ^= 1U << drop;
 213         }
 214
 215         /*
 216          * If the new extent no longer has any pointers, bch2_extent_normalize()
 217          * will do the appropriate thing with it (turning it into a
 218          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 219          */
 220         bch2_extent_normalize(c, bkey_i_to_s(n));
 221
 222         /*
 223          * Since we're not inserting through an extent iterator
 224          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 225          * we aren't using the extent overwrite path to delete, we're
 226          * just using the normal key deletion path:
 227          */
 228         if (bkey_deleted(&n->k))
 229                 n->k.size = 0;
 230
 231         return bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 232                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 233 }
 234
 235 static int bch2_move_extent(struct btree_trans *trans,
 236                             struct btree_iter *iter,
 237                             struct moving_context *ctxt,
 238                             struct bch_io_opts io_opts,
 239                             enum btree_id btree_id,
 240                             struct bkey_s_c k,
 241                             struct data_update_opts data_opts)
 242 {
 243         struct bch_fs *c = trans->c;
 244         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 245         struct moving_io *io;
 246         const union bch_extent_entry *entry;
 247         struct extent_ptr_decoded p;
 248         unsigned sectors = k.k->size, pages;
 249         int ret = -ENOMEM;
 250
 251         bch2_data_update_opts_normalize(k, &data_opts);
 252
 253         if (!data_opts.rewrite_ptrs &&
 254             !data_opts.extra_replicas) {
 255                 if (data_opts.kill_ptrs)
 256                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 257                 return 0;
 258         }
 259
 260         if (!percpu_ref_tryget_live(&c->writes))
 261                 return -BCH_ERR_erofs_no_writes;
 262
 263         /*
 264          * Before memory allocations & taking nocow locks in
 265          * bch2_data_update_init():
 266          */
 267         bch2_trans_unlock(trans);
 268
 269         /* write path might have to decompress data: */
 270         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 271                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 272
 273         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 274         io = kzalloc(sizeof(struct moving_io) +
 275                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 276         if (!io)
 277                 goto err;
 278
 279         io->write.ctxt          = ctxt;
 280         io->read_sectors        = k.k->size;
 281         io->write_sectors       = k.k->size;
 282
 283         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 284         bio_set_prio(&io->write.op.wbio.bio,
 285                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 286
 287         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 288                                  GFP_KERNEL))
 289                 goto err_free;
 290
 291         io->rbio.c              = c;
 292         io->rbio.opts           = io_opts;
 293         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 294         io->rbio.bio.bi_vcnt = pages;
 295         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 296         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 297
 298         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 299         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 300         io->rbio.bio.bi_end_io          = move_read_endio;
 301
 302         ret = bch2_data_update_init(c, &io->write, ctxt->wp, io_opts,
 303                                     data_opts, btree_id, k);
 304         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 305                 goto err_free_pages;
 306
 307         io->write.ctxt = ctxt;
 308         io->write.op.end_io = move_write_done;
 309
 310         atomic64_inc(&ctxt->stats->keys_moved);
 311         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 312
 313         if (ret == -BCH_ERR_unwritten_extent_update) {
 314                 bch2_update_unwritten_extent(trans, &io->write);
 315                 move_free(io);
 316                 return 0;
 317         }
 318
 319         BUG_ON(ret);
 320
 321         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 322         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 323         trace_move_extent_read(k.k);
 324
 325         atomic_add(io->read_sectors, &ctxt->read_sectors);
 326         list_add_tail(&io->list, &ctxt->reads);
 327
 328         /*
 329          * dropped by move_read_endio() - guards against use after free of
 330          * ctxt when doing wakeup
 331          */
 332         closure_get(&ctxt->cl);
 333         bch2_read_extent(trans, &io->rbio,
 334                          bkey_start_pos(k.k),
 335                          btree_id, k, 0,
 336                          BCH_READ_NODECODE|
 337                          BCH_READ_LAST_FRAGMENT);
 338         return 0;
 339 err_free_pages:
 340         bio_free_pages(&io->write.op.wbio.bio);
 341 err_free:
 342         kfree(io);
 343 err:
 344         percpu_ref_put(&c->writes);
 345         trace_and_count(c, move_extent_alloc_mem_fail, k.k);
 346         return ret;
 347 }
 348
 349 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 350                         struct bch_inode_unpacked *inode)
 351 {
 352         struct btree_iter iter;
 353         struct bkey_s_c k;
 354         int ret;
 355
 356         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 357                              BTREE_ITER_ALL_SNAPSHOTS);
 358         k = bch2_btree_iter_peek(&iter);
 359         ret = bkey_err(k);
 360         if (ret)
 361                 goto err;
 362
 363         if (!k.k || !bkey_eq(k.k->p, pos)) {
 364                 ret = -ENOENT;
 365                 goto err;
 366         }
 367
 368         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 369         if (ret)
 370                 goto err;
 371
 372         ret = bch2_inode_unpack(k, inode);
 373         if (ret)
 374                 goto err;
 375 err:
 376         bch2_trans_iter_exit(trans, &iter);
 377         return ret;
 378 }
 379
 380 static int move_ratelimit(struct btree_trans *trans,
 381                           struct moving_context *ctxt)
 382 {
 383         struct bch_fs *c = trans->c;
 384         u64 delay;
 385
 386         if (ctxt->wait_on_copygc) {
 387                 bch2_trans_unlock(trans);
 388                 wait_event_killable(c->copygc_running_wq,
 389                                     !c->copygc_running ||
 390                                     kthread_should_stop());
 391         }
 392
 393         do {
 394                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 395
 396                 if (delay) {
 397                         bch2_trans_unlock(trans);
 398                         set_current_state(TASK_INTERRUPTIBLE);
 399                 }
 400
 401                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 402                         __set_current_state(TASK_RUNNING);
 403                         return 1;
 404                 }
 405
 406                 if (delay)
 407                         schedule_timeout(delay);
 408
 409                 if (unlikely(freezing(current))) {
 410                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 411                         try_to_freeze();
 412                 }
 413         } while (delay);
 414
 415         move_ctxt_wait_event(ctxt, trans,
 416                 atomic_read(&ctxt->write_sectors) <
 417                 c->opts.move_bytes_in_flight >> 9);
 418
 419         move_ctxt_wait_event(ctxt, trans,
 420                 atomic_read(&ctxt->read_sectors) <
 421                 c->opts.move_bytes_in_flight >> 9);
 422
 423         return 0;
 424 }
 425
 426 static int move_get_io_opts(struct btree_trans *trans,
 427                             struct bch_io_opts *io_opts,
 428                             struct bkey_s_c k, u64 *cur_inum)
 429 {
 430         struct bch_inode_unpacked inode;
 431         int ret;
 432
 433         if (*cur_inum == k.k->p.inode)
 434                 return 0;
 435
 436         ret = lookup_inode(trans,
 437                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 438                            &inode);
 439         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 440                 return ret;
 441
 442         if (!ret)
 443                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 444         else
 445                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 446         *cur_inum = k.k->p.inode;
 447         return 0;
 448 }
 449
 450 static int __bch2_move_data(struct moving_context *ctxt,
 451                             struct bpos start,
 452                             struct bpos end,
 453                             move_pred_fn pred, void *arg,
 454                             enum btree_id btree_id)
 455 {
 456         struct bch_fs *c = ctxt->c;
 457         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 458         struct bkey_buf sk;
 459         struct btree_trans trans;
 460         struct btree_iter iter;
 461         struct bkey_s_c k;
 462         struct data_update_opts data_opts;
 463         u64 cur_inum = U64_MAX;
 464         int ret = 0, ret2;
 465
 466         bch2_bkey_buf_init(&sk);
 467         bch2_trans_init(&trans, c, 0, 0);
 468
 469         ctxt->stats->data_type  = BCH_DATA_user;
 470         ctxt->stats->btree_id   = btree_id;
 471         ctxt->stats->pos        = start;
 472
 473         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 474                              BTREE_ITER_PREFETCH|
 475                              BTREE_ITER_ALL_SNAPSHOTS);
 476
 477         if (ctxt->rate)
 478                 bch2_ratelimit_reset(ctxt->rate);
 479
 480         while (!move_ratelimit(&trans, ctxt)) {
 481                 bch2_trans_begin(&trans);
 482
 483                 k = bch2_btree_iter_peek(&iter);
 484                 if (!k.k)
 485                         break;
 486
 487                 ret = bkey_err(k);
 488                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 489                         continue;
 490                 if (ret)
 491                         break;
 492
 493                 if (bkey_ge(bkey_start_pos(k.k), end))
 494                         break;
 495
 496                 ctxt->stats->pos = iter.pos;
 497
 498                 if (!bkey_extent_is_direct_data(k.k))
 499                         goto next_nondata;
 500
 501                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 502                 if (ret)
 503                         continue;
 504
 505                 memset(&data_opts, 0, sizeof(data_opts));
 506                 if (!pred(c, arg, k, &io_opts, &data_opts))
 507                         goto next;
 508
 509                 /*
 510                  * The iterator gets unlocked by __bch2_read_extent - need to
 511                  * save a copy of @k elsewhere:
 512                  */
 513                 bch2_bkey_buf_reassemble(&sk, c, k);
 514                 k = bkey_i_to_s_c(sk.k);
 515                 bch2_trans_unlock(&trans);
 516
 517                 ret2 = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 518                                         btree_id, k, data_opts);
 519                 if (ret2) {
 520                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 521                                 continue;
 522
 523                         if (ret2 == -ENOMEM) {
 524                                 /* memory allocation failure, wait for some IO to finish */
 525                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 526                                 continue;
 527                         }
 528
 529                         /* XXX signal failure */
 530                         goto next;
 531                 }
 532
 533                 if (ctxt->rate)
 534                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 535 next:
 536                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 537 next_nondata:
 538                 bch2_btree_iter_advance(&iter);
 539         }
 540
 541         bch2_trans_iter_exit(&trans, &iter);
 542         bch2_trans_exit(&trans);
 543         bch2_bkey_buf_exit(&sk, c);
 544
 545         return ret;
 546 }
 547
 548 int bch2_move_data(struct bch_fs *c,
 549                    enum btree_id start_btree_id, struct bpos start_pos,
 550                    enum btree_id end_btree_id,   struct bpos end_pos,
 551                    struct bch_ratelimit *rate,
 552                    struct bch_move_stats *stats,
 553                    struct write_point_specifier wp,
 554                    bool wait_on_copygc,
 555                    move_pred_fn pred, void *arg)
 556 {
 557         struct moving_context ctxt;
 558         enum btree_id id;
 559         int ret;
 560
 561         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 562
 563         for (id = start_btree_id;
 564              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 565              id++) {
 566                 stats->btree_id = id;
 567
 568                 if (id != BTREE_ID_extents &&
 569                     id != BTREE_ID_reflink)
 570                         continue;
 571
 572                 ret = __bch2_move_data(&ctxt,
 573                                        id == start_btree_id ? start_pos : POS_MIN,
 574                                        id == end_btree_id   ? end_pos   : POS_MAX,
 575                                        pred, arg, id);
 576                 if (ret)
 577                         break;
 578         }
 579
 580         bch2_moving_ctxt_exit(&ctxt);
 581
 582         return ret;
 583 }
 584
 585 static int verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 586 {
 587         struct bch_fs *c = trans->c;
 588         struct btree_iter iter;
 589         struct bkey_s_c k;
 590         struct printbuf buf = PRINTBUF;
 591         struct bch_backpointer bp;
 592         u64 bp_offset = 0;
 593         int ret;
 594
 595         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 596                              bucket, BTREE_ITER_CACHED);
 597 again:
 598         k = bch2_btree_iter_peek_slot(&iter);
 599         ret = bkey_err(k);
 600
 601         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 602                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 603
 604                 if (a.v->gen == gen &&
 605                     a.v->dirty_sectors) {
 606                         if (a.v->data_type == BCH_DATA_btree) {
 607                                 bch2_trans_unlock(trans);
 608                                 if (bch2_btree_interior_updates_flush(c))
 609                                         goto again;
 610                                 goto failed_to_evacuate;
 611                         }
 612                 }
 613         }
 614
 615         bch2_trans_iter_exit(trans, &iter);
 616         return ret;
 617 failed_to_evacuate:
 618         bch2_trans_iter_exit(trans, &iter);
 619
 620         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 621         bch2_bkey_val_to_text(&buf, c, k);
 622
 623         while (1) {
 624                 bch2_trans_begin(trans);
 625
 626                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 627                                                 &bp_offset, &bp,
 628                                                 BTREE_ITER_CACHED);
 629                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 630                         continue;
 631                 if (ret)
 632                         break;
 633                 if (bp_offset == U64_MAX)
 634                         break;
 635
 636                 k = bch2_backpointer_get_key(trans, &iter,
 637                                              bucket, bp_offset, bp);
 638                 ret = bkey_err(k);
 639                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 640                         continue;
 641                 if (ret)
 642                         break;
 643                 if (!k.k)
 644                         continue;
 645                 prt_newline(&buf);
 646                 bch2_bkey_val_to_text(&buf, c, k);
 647                 bch2_trans_iter_exit(trans, &iter);
 648         }
 649
 650         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 651         printbuf_exit(&buf);
 652         return 0;
 653 }
 654
 655 int __bch2_evacuate_bucket(struct moving_context *ctxt,
 656                            struct bpos bucket, int gen,
 657                            struct data_update_opts _data_opts)
 658 {
 659         struct bch_fs *c = ctxt->c;
 660         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 661         struct btree_trans trans;
 662         struct btree_iter iter;
 663         struct bkey_buf sk;
 664         struct bch_backpointer bp;
 665         struct bch_alloc_v4 a_convert;
 666         const struct bch_alloc_v4 *a;
 667         struct bkey_s_c k;
 668         struct data_update_opts data_opts;
 669         unsigned dirty_sectors, bucket_size;
 670         u64 bp_offset = 0, cur_inum = U64_MAX;
 671         int ret = 0;
 672
 673         bch2_bkey_buf_init(&sk);
 674         bch2_trans_init(&trans, c, 0, 0);
 675
 676         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
 677                              bucket, BTREE_ITER_CACHED);
 678         ret = lockrestart_do(&trans,
 679                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 680         bch2_trans_iter_exit(&trans, &iter);
 681
 682         if (!ret) {
 683                 a = bch2_alloc_to_v4(k, &a_convert);
 684                 dirty_sectors = a->dirty_sectors;
 685                 bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 686         }
 687
 688         while (!(ret = move_ratelimit(&trans, ctxt))) {
 689                 bch2_trans_begin(&trans);
 690
 691                 ret = bch2_get_next_backpointer(&trans, bucket, gen,
 692                                                 &bp_offset, &bp,
 693                                                 BTREE_ITER_CACHED);
 694                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 695                         continue;
 696                 if (ret)
 697                         goto err;
 698                 if (bp_offset == U64_MAX)
 699                         break;
 700
 701                 if (!bp.level) {
 702                         const struct bch_extent_ptr *ptr;
 703                         struct bkey_s_c k;
 704                         unsigned i = 0;
 705
 706                         k = bch2_backpointer_get_key(&trans, &iter,
 707                                                 bucket, bp_offset, bp);
 708                         ret = bkey_err(k);
 709                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 710                                 continue;
 711                         if (ret)
 712                                 goto err;
 713                         if (!k.k)
 714                                 continue;
 715
 716                         bch2_bkey_buf_reassemble(&sk, c, k);
 717                         k = bkey_i_to_s_c(sk.k);
 718
 719                         ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 720                         if (ret) {
 721                                 bch2_trans_iter_exit(&trans, &iter);
 722                                 continue;
 723                         }
 724
 725                         data_opts = _data_opts;
 726                         data_opts.target        = io_opts.background_target;
 727                         data_opts.rewrite_ptrs = 0;
 728
 729                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 730                                 if (ptr->dev == bucket.inode)
 731                                         data_opts.rewrite_ptrs |= 1U << i;
 732                                 i++;
 733                         }
 734
 735                         ret = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 736                                                bp.btree_id, k, data_opts);
 737                         bch2_trans_iter_exit(&trans, &iter);
 738
 739                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 740                                 continue;
 741                         if (ret == -ENOMEM) {
 742                                 /* memory allocation failure, wait for some IO to finish */
 743                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 744                                 continue;
 745                         }
 746                         if (ret)
 747                                 goto err;
 748
 749                         if (ctxt->rate)
 750                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 751                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 752                 } else {
 753                         struct btree *b;
 754
 755                         b = bch2_backpointer_get_node(&trans, &iter,
 756                                                 bucket, bp_offset, bp);
 757                         ret = PTR_ERR_OR_ZERO(b);
 758                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 759                                 continue;
 760                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 761                                 continue;
 762                         if (ret)
 763                                 goto err;
 764                         if (!b)
 765                                 continue;
 766
 767                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
 768                         bch2_trans_iter_exit(&trans, &iter);
 769
 770                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 771                                 continue;
 772                         if (ret)
 773                                 goto err;
 774
 775                         if (ctxt->rate)
 776                                 bch2_ratelimit_increment(ctxt->rate,
 777                                                          c->opts.btree_node_size >> 9);
 778                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 779                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 780                 }
 781
 782                 bp_offset++;
 783         }
 784
 785         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, ret);
 786
 787         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
 788                 bch2_trans_unlock(&trans);
 789                 move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 790                 closure_sync(&ctxt->cl);
 791                 if (!ctxt->write_error)
 792                         lockrestart_do(&trans, verify_bucket_evacuated(&trans, bucket, gen));
 793         }
 794 err:
 795         bch2_trans_exit(&trans);
 796         bch2_bkey_buf_exit(&sk, c);
 797         return ret;
 798 }
 799
 800 int bch2_evacuate_bucket(struct bch_fs *c,
 801                          struct bpos bucket, int gen,
 802                          struct data_update_opts data_opts,
 803                          struct bch_ratelimit *rate,
 804                          struct bch_move_stats *stats,
 805                          struct write_point_specifier wp,
 806                          bool wait_on_copygc)
 807 {
 808         struct moving_context ctxt;
 809         int ret;
 810
 811         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 812         ret = __bch2_evacuate_bucket(&ctxt, bucket, gen, data_opts);
 813         bch2_moving_ctxt_exit(&ctxt);
 814
 815         return ret;
 816 }
 817
 818 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 819                                 struct btree *, struct bch_io_opts *,
 820                                 struct data_update_opts *);
 821
 822 static int bch2_move_btree(struct bch_fs *c,
 823                            enum btree_id start_btree_id, struct bpos start_pos,
 824                            enum btree_id end_btree_id,   struct bpos end_pos,
 825                            move_btree_pred pred, void *arg,
 826                            struct bch_move_stats *stats)
 827 {
 828         bool kthread = (current->flags & PF_KTHREAD) != 0;
 829         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 830         struct btree_trans trans;
 831         struct btree_iter iter;
 832         struct btree *b;
 833         enum btree_id id;
 834         struct data_update_opts data_opts;
 835         int ret = 0;
 836
 837         bch2_trans_init(&trans, c, 0, 0);
 838         progress_list_add(c, stats);
 839
 840         stats->data_type = BCH_DATA_btree;
 841
 842         for (id = start_btree_id;
 843              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 844              id++) {
 845                 stats->btree_id = id;
 846
 847                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 848                                           BTREE_ITER_PREFETCH);
 849 retry:
 850                 ret = 0;
 851                 while (bch2_trans_begin(&trans),
 852                        (b = bch2_btree_iter_peek_node(&iter)) &&
 853                        !(ret = PTR_ERR_OR_ZERO(b))) {
 854                         if (kthread && kthread_should_stop())
 855                                 break;
 856
 857                         if ((cmp_int(id, end_btree_id) ?:
 858                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 859                                 break;
 860
 861                         stats->pos = iter.pos;
 862
 863                         if (!pred(c, arg, b, &io_opts, &data_opts))
 864                                 goto next;
 865
 866                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 867                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 868                                 continue;
 869                         if (ret)
 870                                 break;
 871 next:
 872                         bch2_btree_iter_next_node(&iter);
 873                 }
 874                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 875                         goto retry;
 876
 877                 bch2_trans_iter_exit(&trans, &iter);
 878
 879                 if (kthread && kthread_should_stop())
 880                         break;
 881         }
 882
 883         bch2_trans_exit(&trans);
 884
 885         if (ret)
 886                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 887
 888         bch2_btree_interior_updates_flush(c);
 889
 890         progress_list_del(c, stats);
 891         return ret;
 892 }
 893
 894 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 895                              struct bkey_s_c k,
 896                              struct bch_io_opts *io_opts,
 897                              struct data_update_opts *data_opts)
 898 {
 899         unsigned nr_good = bch2_bkey_durability(c, k);
 900         unsigned replicas = bkey_is_btree_ptr(k.k)
 901                 ? c->opts.metadata_replicas
 902                 : io_opts->data_replicas;
 903
 904         if (!nr_good || nr_good >= replicas)
 905                 return false;
 906
 907         data_opts->target               = 0;
 908         data_opts->extra_replicas       = replicas - nr_good;
 909         data_opts->btree_insert_flags   = 0;
 910         return true;
 911 }
 912
 913 static bool migrate_pred(struct bch_fs *c, void *arg,
 914                          struct bkey_s_c k,
 915                          struct bch_io_opts *io_opts,
 916                          struct data_update_opts *data_opts)
 917 {
 918         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 919         const struct bch_extent_ptr *ptr;
 920         struct bch_ioctl_data *op = arg;
 921         unsigned i = 0;
 922
 923         data_opts->rewrite_ptrs         = 0;
 924         data_opts->target               = 0;
 925         data_opts->extra_replicas       = 0;
 926         data_opts->btree_insert_flags   = 0;
 927
 928         bkey_for_each_ptr(ptrs, ptr) {
 929                 if (ptr->dev == op->migrate.dev)
 930                         data_opts->rewrite_ptrs |= 1U << i;
 931                 i++;
 932         }
 933
 934         return data_opts->rewrite_ptrs != 0;
 935 }
 936
 937 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 938                                    struct btree *b,
 939                                    struct bch_io_opts *io_opts,
 940                                    struct data_update_opts *data_opts)
 941 {
 942         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 943 }
 944
 945 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 946                                struct btree *b,
 947                                struct bch_io_opts *io_opts,
 948                                struct data_update_opts *data_opts)
 949 {
 950         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 951 }
 952
 953 static bool bformat_needs_redo(struct bkey_format *f)
 954 {
 955         unsigned i;
 956
 957         for (i = 0; i < f->nr_fields; i++) {
 958                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 959                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 960                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 961
 962                 if (f->bits_per_field[i] > unpacked_bits)
 963                         return true;
 964
 965                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 966                         return true;
 967
 968                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 969                      unpacked_mask) <
 970                     field_offset)
 971                         return true;
 972         }
 973
 974         return false;
 975 }
 976
 977 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 978                                    struct btree *b,
 979                                    struct bch_io_opts *io_opts,
 980                                    struct data_update_opts *data_opts)
 981 {
 982         if (b->version_ondisk != c->sb.version ||
 983             btree_node_need_rewrite(b) ||
 984             bformat_needs_redo(&b->format)) {
 985                 data_opts->target               = 0;
 986                 data_opts->extra_replicas       = 0;
 987                 data_opts->btree_insert_flags   = 0;
 988                 return true;
 989         }
 990
 991         return false;
 992 }
 993
 994 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 995 {
 996         int ret;
 997
 998         ret = bch2_move_btree(c,
 999                               0,                POS_MIN,
1000                               BTREE_ID_NR,      SPOS_MAX,
1001                               rewrite_old_nodes_pred, c, stats);
1002         if (!ret) {
1003                 mutex_lock(&c->sb_lock);
1004                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1005                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1006                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1007                 bch2_write_super(c);
1008                 mutex_unlock(&c->sb_lock);
1009         }
1010
1011         return ret;
1012 }
1013
1014 int bch2_data_job(struct bch_fs *c,
1015                   struct bch_move_stats *stats,
1016                   struct bch_ioctl_data op)
1017 {
1018         int ret = 0;
1019
1020         switch (op.op) {
1021         case BCH_DATA_OP_REREPLICATE:
1022                 bch2_move_stats_init(stats, "rereplicate");
1023                 stats->data_type = BCH_DATA_journal;
1024                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1025
1026                 ret = bch2_move_btree(c,
1027                                       op.start_btree,   op.start_pos,
1028                                       op.end_btree,     op.end_pos,
1029                                       rereplicate_btree_pred, c, stats) ?: ret;
1030                 ret = bch2_replicas_gc2(c) ?: ret;
1031
1032                 ret = bch2_move_data(c,
1033                                      op.start_btree,    op.start_pos,
1034                                      op.end_btree,      op.end_pos,
1035                                      NULL,
1036                                      stats,
1037                                      writepoint_hashed((unsigned long) current),
1038                                      true,
1039                                      rereplicate_pred, c) ?: ret;
1040                 ret = bch2_replicas_gc2(c) ?: ret;
1041                 break;
1042         case BCH_DATA_OP_MIGRATE:
1043                 if (op.migrate.dev >= c->sb.nr_devices)
1044                         return -EINVAL;
1045
1046                 bch2_move_stats_init(stats, "migrate");
1047                 stats->data_type = BCH_DATA_journal;
1048                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1049
1050                 ret = bch2_move_btree(c,
1051                                       op.start_btree,   op.start_pos,
1052                                       op.end_btree,     op.end_pos,
1053                                       migrate_btree_pred, &op, stats) ?: ret;
1054                 ret = bch2_replicas_gc2(c) ?: ret;
1055
1056                 ret = bch2_move_data(c,
1057                                      op.start_btree,    op.start_pos,
1058                                      op.end_btree,      op.end_pos,
1059                                      NULL,
1060                                      stats,
1061                                      writepoint_hashed((unsigned long) current),
1062                                      true,
1063                                      migrate_pred, &op) ?: ret;
1064                 ret = bch2_replicas_gc2(c) ?: ret;
1065                 break;
1066         case BCH_DATA_OP_REWRITE_OLD_NODES:
1067                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1068                 ret = bch2_scan_old_btree_nodes(c, stats);
1069                 break;
1070         default:
1071                 ret = -EINVAL;
1072         }
1073
1074         return ret;
1075 }