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         bch2_write_ref_put(c, BCH_WRITE_REF_move);
  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         atomic_dec(&io->write.ctxt->write_ios);
  78         move_free(io);
  79         closure_put(&ctxt->cl);
  80 }
  81
  82 static void move_write(struct moving_io *io)
  83 {
  84         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
  85                 move_free(io);
  86                 return;
  87         }
  88
  89         closure_get(&io->write.ctxt->cl);
  90         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
  91         atomic_inc(&io->write.ctxt->write_ios);
  92
  93         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
  94 }
  95
  96 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
  97 {
  98         struct moving_io *io =
  99                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 100
 101         return io && io->read_completed ? io : NULL;
 102 }
 103
 104 static void move_read_endio(struct bio *bio)
 105 {
 106         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 107         struct moving_context *ctxt = io->write.ctxt;
 108
 109         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 110         atomic_dec(&ctxt->read_ios);
 111         io->read_completed = true;
 112
 113         wake_up(&ctxt->wait);
 114         closure_put(&ctxt->cl);
 115 }
 116
 117 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 118                                         struct btree_trans *trans)
 119 {
 120         struct moving_io *io;
 121
 122         if (trans)
 123                 bch2_trans_unlock(trans);
 124
 125         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 126                 list_del(&io->list);
 127                 move_write(io);
 128         }
 129 }
 130
 131 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 132                                        struct btree_trans *trans)
 133 {
 134         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 135
 136         move_ctxt_wait_event(ctxt, trans,
 137                 !atomic_read(&ctxt->write_sectors) ||
 138                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 139 }
 140
 141 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 142 {
 143         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 144         closure_sync(&ctxt->cl);
 145
 146         EBUG_ON(atomic_read(&ctxt->write_sectors));
 147         EBUG_ON(atomic_read(&ctxt->write_ios));
 148         EBUG_ON(atomic_read(&ctxt->read_sectors));
 149         EBUG_ON(atomic_read(&ctxt->read_ios));
 150
 151         if (ctxt->stats) {
 152                 progress_list_del(ctxt->c, ctxt->stats);
 153
 154                 trace_move_data(ctxt->c,
 155                                 atomic64_read(&ctxt->stats->sectors_moved),
 156                                 atomic64_read(&ctxt->stats->keys_moved));
 157         }
 158 }
 159
 160 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 161                            struct bch_fs *c,
 162                            struct bch_ratelimit *rate,
 163                            struct bch_move_stats *stats,
 164                            struct write_point_specifier wp,
 165                            bool wait_on_copygc)
 166 {
 167         memset(ctxt, 0, sizeof(*ctxt));
 168
 169         ctxt->c         = c;
 170         ctxt->rate      = rate;
 171         ctxt->stats     = stats;
 172         ctxt->wp        = wp;
 173         ctxt->wait_on_copygc = wait_on_copygc;
 174
 175         closure_init_stack(&ctxt->cl);
 176         INIT_LIST_HEAD(&ctxt->reads);
 177         init_waitqueue_head(&ctxt->wait);
 178
 179         if (stats) {
 180                 progress_list_add(c, stats);
 181                 stats->data_type = BCH_DATA_user;
 182         }
 183 }
 184
 185 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 186 {
 187         memset(stats, 0, sizeof(*stats));
 188         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 189 }
 190
 191 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 192                                  struct btree_iter *iter,
 193                                  struct bkey_s_c k,
 194                                  struct data_update_opts data_opts)
 195 {
 196         struct bch_fs *c = trans->c;
 197         struct bkey_i *n;
 198         int ret;
 199
 200         n = bch2_bkey_make_mut(trans, k);
 201         ret = PTR_ERR_OR_ZERO(n);
 202         if (ret)
 203                 return ret;
 204
 205         while (data_opts.kill_ptrs) {
 206                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 207                 struct bch_extent_ptr *ptr;
 208
 209                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 210                 data_opts.kill_ptrs ^= 1U << drop;
 211         }
 212
 213         /*
 214          * If the new extent no longer has any pointers, bch2_extent_normalize()
 215          * will do the appropriate thing with it (turning it into a
 216          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 217          */
 218         bch2_extent_normalize(c, bkey_i_to_s(n));
 219
 220         /*
 221          * Since we're not inserting through an extent iterator
 222          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 223          * we aren't using the extent overwrite path to delete, we're
 224          * just using the normal key deletion path:
 225          */
 226         if (bkey_deleted(&n->k))
 227                 n->k.size = 0;
 228
 229         return bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 230                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 231 }
 232
 233 static int bch2_move_extent(struct btree_trans *trans,
 234                             struct btree_iter *iter,
 235                             struct moving_context *ctxt,
 236                             struct bch_io_opts io_opts,
 237                             enum btree_id btree_id,
 238                             struct bkey_s_c k,
 239                             struct data_update_opts data_opts)
 240 {
 241         struct bch_fs *c = trans->c;
 242         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 243         struct moving_io *io;
 244         const union bch_extent_entry *entry;
 245         struct extent_ptr_decoded p;
 246         unsigned sectors = k.k->size, pages;
 247         int ret = -ENOMEM;
 248
 249         bch2_data_update_opts_normalize(k, &data_opts);
 250
 251         if (!data_opts.rewrite_ptrs &&
 252             !data_opts.extra_replicas) {
 253                 if (data_opts.kill_ptrs)
 254                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 255                 return 0;
 256         }
 257
 258         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_move))
 259                 return -BCH_ERR_erofs_no_writes;
 260
 261         /*
 262          * Before memory allocations & taking nocow locks in
 263          * bch2_data_update_init():
 264          */
 265         bch2_trans_unlock(trans);
 266
 267         /* write path might have to decompress data: */
 268         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 269                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 270
 271         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 272         io = kzalloc(sizeof(struct moving_io) +
 273                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 274         if (!io)
 275                 goto err;
 276
 277         io->write.ctxt          = ctxt;
 278         io->read_sectors        = k.k->size;
 279         io->write_sectors       = k.k->size;
 280
 281         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 282         bio_set_prio(&io->write.op.wbio.bio,
 283                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 284
 285         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 286                                  GFP_KERNEL))
 287                 goto err_free;
 288
 289         io->rbio.c              = c;
 290         io->rbio.opts           = io_opts;
 291         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 292         io->rbio.bio.bi_vcnt = pages;
 293         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 294         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 295
 296         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 297         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 298         io->rbio.bio.bi_end_io          = move_read_endio;
 299
 300         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 301                                     io_opts, data_opts, btree_id, k);
 302         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 303                 goto err_free_pages;
 304
 305         io->write.ctxt = ctxt;
 306         io->write.op.end_io = move_write_done;
 307
 308         atomic64_inc(&ctxt->stats->keys_moved);
 309         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 310
 311         if (ret == -BCH_ERR_unwritten_extent_update) {
 312                 bch2_update_unwritten_extent(trans, &io->write);
 313                 move_free(io);
 314                 return 0;
 315         }
 316
 317         BUG_ON(ret);
 318
 319         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 320         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 321         trace_move_extent_read(k.k);
 322
 323         atomic_add(io->read_sectors, &ctxt->read_sectors);
 324         atomic_inc(&ctxt->read_ios);
 325         list_add_tail(&io->list, &ctxt->reads);
 326
 327         /*
 328          * dropped by move_read_endio() - guards against use after free of
 329          * ctxt when doing wakeup
 330          */
 331         closure_get(&ctxt->cl);
 332         bch2_read_extent(trans, &io->rbio,
 333                          bkey_start_pos(k.k),
 334                          btree_id, k, 0,
 335                          BCH_READ_NODECODE|
 336                          BCH_READ_LAST_FRAGMENT);
 337         return 0;
 338 err_free_pages:
 339         bio_free_pages(&io->write.op.wbio.bio);
 340 err_free:
 341         kfree(io);
 342 err:
 343         bch2_write_ref_put(c, BCH_WRITE_REF_move);
 344         trace_and_count(c, move_extent_alloc_mem_fail, k.k);
 345         return ret;
 346 }
 347
 348 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 349                         struct bch_inode_unpacked *inode)
 350 {
 351         struct btree_iter iter;
 352         struct bkey_s_c k;
 353         int ret;
 354
 355         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 356                              BTREE_ITER_ALL_SNAPSHOTS);
 357         k = bch2_btree_iter_peek(&iter);
 358         ret = bkey_err(k);
 359         if (ret)
 360                 goto err;
 361
 362         if (!k.k || !bkey_eq(k.k->p, pos)) {
 363                 ret = -ENOENT;
 364                 goto err;
 365         }
 366
 367         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 368         if (ret)
 369                 goto err;
 370
 371         ret = bch2_inode_unpack(k, inode);
 372         if (ret)
 373                 goto err;
 374 err:
 375         bch2_trans_iter_exit(trans, &iter);
 376         return ret;
 377 }
 378
 379 static int move_ratelimit(struct btree_trans *trans,
 380                           struct moving_context *ctxt)
 381 {
 382         struct bch_fs *c = trans->c;
 383         u64 delay;
 384
 385         if (ctxt->wait_on_copygc) {
 386                 bch2_trans_unlock(trans);
 387                 wait_event_killable(c->copygc_running_wq,
 388                                     !c->copygc_running ||
 389                                     kthread_should_stop());
 390         }
 391
 392         do {
 393                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 394
 395                 if (delay) {
 396                         bch2_trans_unlock(trans);
 397                         set_current_state(TASK_INTERRUPTIBLE);
 398                 }
 399
 400                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 401                         __set_current_state(TASK_RUNNING);
 402                         return 1;
 403                 }
 404
 405                 if (delay)
 406                         schedule_timeout(delay);
 407
 408                 if (unlikely(freezing(current))) {
 409                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 410                         try_to_freeze();
 411                 }
 412         } while (delay);
 413
 414         /*
 415          * XXX: these limits really ought to be per device, SSDs and hard drives
 416          * will want different limits
 417          */
 418         move_ctxt_wait_event(ctxt, trans,
 419                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 420                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 421                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 422                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 423
 424         return 0;
 425 }
 426
 427 static int move_get_io_opts(struct btree_trans *trans,
 428                             struct bch_io_opts *io_opts,
 429                             struct bkey_s_c k, u64 *cur_inum)
 430 {
 431         struct bch_inode_unpacked inode;
 432         int ret;
 433
 434         if (*cur_inum == k.k->p.inode)
 435                 return 0;
 436
 437         ret = lookup_inode(trans,
 438                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 439                            &inode);
 440         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 441                 return ret;
 442
 443         if (!ret)
 444                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 445         else
 446                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 447         *cur_inum = k.k->p.inode;
 448         return 0;
 449 }
 450
 451 static int __bch2_move_data(struct moving_context *ctxt,
 452                             struct bpos start,
 453                             struct bpos end,
 454                             move_pred_fn pred, void *arg,
 455                             enum btree_id btree_id)
 456 {
 457         struct bch_fs *c = ctxt->c;
 458         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 459         struct bkey_buf sk;
 460         struct btree_trans trans;
 461         struct btree_iter iter;
 462         struct bkey_s_c k;
 463         struct data_update_opts data_opts;
 464         u64 cur_inum = U64_MAX;
 465         int ret = 0, ret2;
 466
 467         bch2_bkey_buf_init(&sk);
 468         bch2_trans_init(&trans, c, 0, 0);
 469
 470         ctxt->stats->data_type  = BCH_DATA_user;
 471         ctxt->stats->btree_id   = btree_id;
 472         ctxt->stats->pos        = start;
 473
 474         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 475                              BTREE_ITER_PREFETCH|
 476                              BTREE_ITER_ALL_SNAPSHOTS);
 477
 478         if (ctxt->rate)
 479                 bch2_ratelimit_reset(ctxt->rate);
 480
 481         while (!move_ratelimit(&trans, ctxt)) {
 482                 bch2_trans_begin(&trans);
 483
 484                 k = bch2_btree_iter_peek(&iter);
 485                 if (!k.k)
 486                         break;
 487
 488                 ret = bkey_err(k);
 489                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 490                         continue;
 491                 if (ret)
 492                         break;
 493
 494                 if (bkey_ge(bkey_start_pos(k.k), end))
 495                         break;
 496
 497                 ctxt->stats->pos = iter.pos;
 498
 499                 if (!bkey_extent_is_direct_data(k.k))
 500                         goto next_nondata;
 501
 502                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 503                 if (ret)
 504                         continue;
 505
 506                 memset(&data_opts, 0, sizeof(data_opts));
 507                 if (!pred(c, arg, k, &io_opts, &data_opts))
 508                         goto next;
 509
 510                 /*
 511                  * The iterator gets unlocked by __bch2_read_extent - need to
 512                  * save a copy of @k elsewhere:
 513                  */
 514                 bch2_bkey_buf_reassemble(&sk, c, k);
 515                 k = bkey_i_to_s_c(sk.k);
 516                 bch2_trans_unlock(&trans);
 517
 518                 ret2 = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 519                                         btree_id, k, data_opts);
 520                 if (ret2) {
 521                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 522                                 continue;
 523
 524                         if (ret2 == -ENOMEM) {
 525                                 /* memory allocation failure, wait for some IO to finish */
 526                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 527                                 continue;
 528                         }
 529
 530                         /* XXX signal failure */
 531                         goto next;
 532                 }
 533
 534                 if (ctxt->rate)
 535                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 536 next:
 537                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 538 next_nondata:
 539                 bch2_btree_iter_advance(&iter);
 540         }
 541
 542         bch2_trans_iter_exit(&trans, &iter);
 543         bch2_trans_exit(&trans);
 544         bch2_bkey_buf_exit(&sk, c);
 545
 546         return ret;
 547 }
 548
 549 int bch2_move_data(struct bch_fs *c,
 550                    enum btree_id start_btree_id, struct bpos start_pos,
 551                    enum btree_id end_btree_id,   struct bpos end_pos,
 552                    struct bch_ratelimit *rate,
 553                    struct bch_move_stats *stats,
 554                    struct write_point_specifier wp,
 555                    bool wait_on_copygc,
 556                    move_pred_fn pred, void *arg)
 557 {
 558         struct moving_context ctxt;
 559         enum btree_id id;
 560         int ret;
 561
 562         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 563
 564         for (id = start_btree_id;
 565              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 566              id++) {
 567                 stats->btree_id = id;
 568
 569                 if (id != BTREE_ID_extents &&
 570                     id != BTREE_ID_reflink)
 571                         continue;
 572
 573                 ret = __bch2_move_data(&ctxt,
 574                                        id == start_btree_id ? start_pos : POS_MIN,
 575                                        id == end_btree_id   ? end_pos   : POS_MAX,
 576                                        pred, arg, id);
 577                 if (ret)
 578                         break;
 579         }
 580
 581         bch2_moving_ctxt_exit(&ctxt);
 582
 583         return ret;
 584 }
 585
 586 static int verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 587 {
 588         struct bch_fs *c = trans->c;
 589         struct btree_iter iter;
 590         struct bkey_s_c k;
 591         struct printbuf buf = PRINTBUF;
 592         struct bch_backpointer bp;
 593         u64 bp_offset = 0;
 594         int ret;
 595
 596         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 597                              bucket, BTREE_ITER_CACHED);
 598 again:
 599         k = bch2_btree_iter_peek_slot(&iter);
 600         ret = bkey_err(k);
 601
 602         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 603                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 604
 605                 if (a.v->gen == gen &&
 606                     a.v->dirty_sectors) {
 607                         if (a.v->data_type == BCH_DATA_btree) {
 608                                 bch2_trans_unlock(trans);
 609                                 if (bch2_btree_interior_updates_flush(c))
 610                                         goto again;
 611                                 goto failed_to_evacuate;
 612                         }
 613                 }
 614         }
 615
 616         bch2_trans_iter_exit(trans, &iter);
 617         return ret;
 618 failed_to_evacuate:
 619         bch2_trans_iter_exit(trans, &iter);
 620
 621         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 622         bch2_bkey_val_to_text(&buf, c, k);
 623
 624         while (1) {
 625                 bch2_trans_begin(trans);
 626
 627                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 628                                                 &bp_offset, &bp,
 629                                                 BTREE_ITER_CACHED);
 630                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 631                         continue;
 632                 if (ret)
 633                         break;
 634                 if (bp_offset == U64_MAX)
 635                         break;
 636
 637                 k = bch2_backpointer_get_key(trans, &iter,
 638                                              bucket, bp_offset, bp);
 639                 ret = bkey_err(k);
 640                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 641                         continue;
 642                 if (ret)
 643                         break;
 644                 if (!k.k)
 645                         continue;
 646                 prt_newline(&buf);
 647                 bch2_bkey_val_to_text(&buf, c, k);
 648                 bch2_trans_iter_exit(trans, &iter);
 649         }
 650
 651         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 652         printbuf_exit(&buf);
 653         return 0;
 654 }
 655
 656 int __bch2_evacuate_bucket(struct moving_context *ctxt,
 657                            struct bpos bucket, int gen,
 658                            struct data_update_opts _data_opts)
 659 {
 660         struct bch_fs *c = ctxt->c;
 661         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 662         struct btree_trans trans;
 663         struct btree_iter iter;
 664         struct bkey_buf sk;
 665         struct bch_backpointer bp;
 666         struct bch_alloc_v4 a_convert;
 667         const struct bch_alloc_v4 *a;
 668         struct bkey_s_c k;
 669         struct data_update_opts data_opts;
 670         unsigned dirty_sectors, bucket_size;
 671         u64 bp_offset = 0, cur_inum = U64_MAX;
 672         int ret = 0;
 673
 674         bch2_bkey_buf_init(&sk);
 675         bch2_trans_init(&trans, c, 0, 0);
 676
 677         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
 678                              bucket, BTREE_ITER_CACHED);
 679         ret = lockrestart_do(&trans,
 680                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 681         bch2_trans_iter_exit(&trans, &iter);
 682
 683         if (!ret) {
 684                 a = bch2_alloc_to_v4(k, &a_convert);
 685                 dirty_sectors = a->dirty_sectors;
 686                 bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 687         }
 688
 689         while (!(ret = move_ratelimit(&trans, ctxt))) {
 690                 bch2_trans_begin(&trans);
 691
 692                 ret = bch2_get_next_backpointer(&trans, bucket, gen,
 693                                                 &bp_offset, &bp,
 694                                                 BTREE_ITER_CACHED);
 695                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 696                         continue;
 697                 if (ret)
 698                         goto err;
 699                 if (bp_offset == U64_MAX)
 700                         break;
 701
 702                 if (!bp.level) {
 703                         const struct bch_extent_ptr *ptr;
 704                         struct bkey_s_c k;
 705                         unsigned i = 0;
 706
 707                         k = bch2_backpointer_get_key(&trans, &iter,
 708                                                 bucket, bp_offset, bp);
 709                         ret = bkey_err(k);
 710                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 711                                 continue;
 712                         if (ret)
 713                                 goto err;
 714                         if (!k.k)
 715                                 continue;
 716
 717                         bch2_bkey_buf_reassemble(&sk, c, k);
 718                         k = bkey_i_to_s_c(sk.k);
 719
 720                         ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 721                         if (ret) {
 722                                 bch2_trans_iter_exit(&trans, &iter);
 723                                 continue;
 724                         }
 725
 726                         data_opts = _data_opts;
 727                         data_opts.target        = io_opts.background_target;
 728                         data_opts.rewrite_ptrs = 0;
 729
 730                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 731                                 if (ptr->dev == bucket.inode)
 732                                         data_opts.rewrite_ptrs |= 1U << i;
 733                                 i++;
 734                         }
 735
 736                         ret = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 737                                                bp.btree_id, k, data_opts);
 738                         bch2_trans_iter_exit(&trans, &iter);
 739
 740                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 741                                 continue;
 742                         if (ret == -ENOMEM) {
 743                                 /* memory allocation failure, wait for some IO to finish */
 744                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 745                                 continue;
 746                         }
 747                         if (ret)
 748                                 goto err;
 749
 750                         if (ctxt->rate)
 751                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 752                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 753                 } else {
 754                         struct btree *b;
 755
 756                         b = bch2_backpointer_get_node(&trans, &iter,
 757                                                 bucket, bp_offset, bp);
 758                         ret = PTR_ERR_OR_ZERO(b);
 759                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 760                                 continue;
 761                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 762                                 continue;
 763                         if (ret)
 764                                 goto err;
 765                         if (!b)
 766                                 continue;
 767
 768                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
 769                         bch2_trans_iter_exit(&trans, &iter);
 770
 771                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 772                                 continue;
 773                         if (ret)
 774                                 goto err;
 775
 776                         if (ctxt->rate)
 777                                 bch2_ratelimit_increment(ctxt->rate,
 778                                                          c->opts.btree_node_size >> 9);
 779                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 780                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 781                 }
 782
 783                 bp_offset++;
 784         }
 785
 786         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, ret);
 787
 788         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
 789                 bch2_trans_unlock(&trans);
 790                 move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 791                 closure_sync(&ctxt->cl);
 792                 if (!ctxt->write_error)
 793                         lockrestart_do(&trans, verify_bucket_evacuated(&trans, bucket, gen));
 794         }
 795 err:
 796         bch2_trans_exit(&trans);
 797         bch2_bkey_buf_exit(&sk, c);
 798         return ret;
 799 }
 800
 801 int bch2_evacuate_bucket(struct bch_fs *c,
 802                          struct bpos bucket, int gen,
 803                          struct data_update_opts data_opts,
 804                          struct bch_ratelimit *rate,
 805                          struct bch_move_stats *stats,
 806                          struct write_point_specifier wp,
 807                          bool wait_on_copygc)
 808 {
 809         struct moving_context ctxt;
 810         int ret;
 811
 812         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 813         ret = __bch2_evacuate_bucket(&ctxt, bucket, gen, data_opts);
 814         bch2_moving_ctxt_exit(&ctxt);
 815
 816         return ret;
 817 }
 818
 819 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 820                                 struct btree *, struct bch_io_opts *,
 821                                 struct data_update_opts *);
 822
 823 static int bch2_move_btree(struct bch_fs *c,
 824                            enum btree_id start_btree_id, struct bpos start_pos,
 825                            enum btree_id end_btree_id,   struct bpos end_pos,
 826                            move_btree_pred pred, void *arg,
 827                            struct bch_move_stats *stats)
 828 {
 829         bool kthread = (current->flags & PF_KTHREAD) != 0;
 830         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 831         struct btree_trans trans;
 832         struct btree_iter iter;
 833         struct btree *b;
 834         enum btree_id id;
 835         struct data_update_opts data_opts;
 836         int ret = 0;
 837
 838         bch2_trans_init(&trans, c, 0, 0);
 839         progress_list_add(c, stats);
 840
 841         stats->data_type = BCH_DATA_btree;
 842
 843         for (id = start_btree_id;
 844              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 845              id++) {
 846                 stats->btree_id = id;
 847
 848                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 849                                           BTREE_ITER_PREFETCH);
 850 retry:
 851                 ret = 0;
 852                 while (bch2_trans_begin(&trans),
 853                        (b = bch2_btree_iter_peek_node(&iter)) &&
 854                        !(ret = PTR_ERR_OR_ZERO(b))) {
 855                         if (kthread && kthread_should_stop())
 856                                 break;
 857
 858                         if ((cmp_int(id, end_btree_id) ?:
 859                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 860                                 break;
 861
 862                         stats->pos = iter.pos;
 863
 864                         if (!pred(c, arg, b, &io_opts, &data_opts))
 865                                 goto next;
 866
 867                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 868                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 869                                 continue;
 870                         if (ret)
 871                                 break;
 872 next:
 873                         bch2_btree_iter_next_node(&iter);
 874                 }
 875                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 876                         goto retry;
 877
 878                 bch2_trans_iter_exit(&trans, &iter);
 879
 880                 if (kthread && kthread_should_stop())
 881                         break;
 882         }
 883
 884         bch2_trans_exit(&trans);
 885
 886         if (ret)
 887                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 888
 889         bch2_btree_interior_updates_flush(c);
 890
 891         progress_list_del(c, stats);
 892         return ret;
 893 }
 894
 895 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 896                              struct bkey_s_c k,
 897                              struct bch_io_opts *io_opts,
 898                              struct data_update_opts *data_opts)
 899 {
 900         unsigned nr_good = bch2_bkey_durability(c, k);
 901         unsigned replicas = bkey_is_btree_ptr(k.k)
 902                 ? c->opts.metadata_replicas
 903                 : io_opts->data_replicas;
 904
 905         if (!nr_good || nr_good >= replicas)
 906                 return false;
 907
 908         data_opts->target               = 0;
 909         data_opts->extra_replicas       = replicas - nr_good;
 910         data_opts->btree_insert_flags   = 0;
 911         return true;
 912 }
 913
 914 static bool migrate_pred(struct bch_fs *c, void *arg,
 915                          struct bkey_s_c k,
 916                          struct bch_io_opts *io_opts,
 917                          struct data_update_opts *data_opts)
 918 {
 919         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 920         const struct bch_extent_ptr *ptr;
 921         struct bch_ioctl_data *op = arg;
 922         unsigned i = 0;
 923
 924         data_opts->rewrite_ptrs         = 0;
 925         data_opts->target               = 0;
 926         data_opts->extra_replicas       = 0;
 927         data_opts->btree_insert_flags   = 0;
 928
 929         bkey_for_each_ptr(ptrs, ptr) {
 930                 if (ptr->dev == op->migrate.dev)
 931                         data_opts->rewrite_ptrs |= 1U << i;
 932                 i++;
 933         }
 934
 935         return data_opts->rewrite_ptrs != 0;
 936 }
 937
 938 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 939                                    struct btree *b,
 940                                    struct bch_io_opts *io_opts,
 941                                    struct data_update_opts *data_opts)
 942 {
 943         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 944 }
 945
 946 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 947                                struct btree *b,
 948                                struct bch_io_opts *io_opts,
 949                                struct data_update_opts *data_opts)
 950 {
 951         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 952 }
 953
 954 static bool bformat_needs_redo(struct bkey_format *f)
 955 {
 956         unsigned i;
 957
 958         for (i = 0; i < f->nr_fields; i++) {
 959                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 960                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 961                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 962
 963                 if (f->bits_per_field[i] > unpacked_bits)
 964                         return true;
 965
 966                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 967                         return true;
 968
 969                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 970                      unpacked_mask) <
 971                     field_offset)
 972                         return true;
 973         }
 974
 975         return false;
 976 }
 977
 978 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 979                                    struct btree *b,
 980                                    struct bch_io_opts *io_opts,
 981                                    struct data_update_opts *data_opts)
 982 {
 983         if (b->version_ondisk != c->sb.version ||
 984             btree_node_need_rewrite(b) ||
 985             bformat_needs_redo(&b->format)) {
 986                 data_opts->target               = 0;
 987                 data_opts->extra_replicas       = 0;
 988                 data_opts->btree_insert_flags   = 0;
 989                 return true;
 990         }
 991
 992         return false;
 993 }
 994
 995 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 996 {
 997         int ret;
 998
 999         ret = bch2_move_btree(c,
1000                               0,                POS_MIN,
1001                               BTREE_ID_NR,      SPOS_MAX,
1002                               rewrite_old_nodes_pred, c, stats);
1003         if (!ret) {
1004                 mutex_lock(&c->sb_lock);
1005                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1006                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1007                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1008                 bch2_write_super(c);
1009                 mutex_unlock(&c->sb_lock);
1010         }
1011
1012         return ret;
1013 }
1014
1015 int bch2_data_job(struct bch_fs *c,
1016                   struct bch_move_stats *stats,
1017                   struct bch_ioctl_data op)
1018 {
1019         int ret = 0;
1020
1021         switch (op.op) {
1022         case BCH_DATA_OP_REREPLICATE:
1023                 bch2_move_stats_init(stats, "rereplicate");
1024                 stats->data_type = BCH_DATA_journal;
1025                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1026
1027                 ret = bch2_move_btree(c,
1028                                       op.start_btree,   op.start_pos,
1029                                       op.end_btree,     op.end_pos,
1030                                       rereplicate_btree_pred, c, stats) ?: ret;
1031                 ret = bch2_replicas_gc2(c) ?: ret;
1032
1033                 ret = bch2_move_data(c,
1034                                      op.start_btree,    op.start_pos,
1035                                      op.end_btree,      op.end_pos,
1036                                      NULL,
1037                                      stats,
1038                                      writepoint_hashed((unsigned long) current),
1039                                      true,
1040                                      rereplicate_pred, c) ?: ret;
1041                 ret = bch2_replicas_gc2(c) ?: ret;
1042                 break;
1043         case BCH_DATA_OP_MIGRATE:
1044                 if (op.migrate.dev >= c->sb.nr_devices)
1045                         return -EINVAL;
1046
1047                 bch2_move_stats_init(stats, "migrate");
1048                 stats->data_type = BCH_DATA_journal;
1049                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1050
1051                 ret = bch2_move_btree(c,
1052                                       op.start_btree,   op.start_pos,
1053                                       op.end_btree,     op.end_pos,
1054                                       migrate_btree_pred, &op, stats) ?: ret;
1055                 ret = bch2_replicas_gc2(c) ?: ret;
1056
1057                 ret = bch2_move_data(c,
1058                                      op.start_btree,    op.start_pos,
1059                                      op.end_btree,      op.end_pos,
1060                                      NULL,
1061                                      stats,
1062                                      writepoint_hashed((unsigned long) current),
1063                                      true,
1064                                      migrate_pred, &op) ?: ret;
1065                 ret = bch2_replicas_gc2(c) ?: ret;
1066                 break;
1067         case BCH_DATA_OP_REWRITE_OLD_NODES:
1068                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1069                 ret = bch2_scan_old_btree_nodes(c, stats);
1070                 break;
1071         default:
1072                 ret = -EINVAL;
1073         }
1074
1075         return ret;
1076 }