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 "btree_write_buffer.h"
  12 #include "disk_groups.h"
  13 #include "ec.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "inode.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "move.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22 #include "keylist.h"
  23
  24 #include <linux/ioprio.h>
  25 #include <linux/kthread.h>
  26
  27 #include <trace/events/bcachefs.h>
  28
  29 static void progress_list_add(struct bch_fs *c, struct bch_move_stats *stats)
  30 {
  31         mutex_lock(&c->data_progress_lock);
  32         list_add(&stats->list, &c->data_progress_list);
  33         mutex_unlock(&c->data_progress_lock);
  34 }
  35
  36 static void progress_list_del(struct bch_fs *c, struct bch_move_stats *stats)
  37 {
  38         mutex_lock(&c->data_progress_lock);
  39         list_del(&stats->list);
  40         mutex_unlock(&c->data_progress_lock);
  41 }
  42
  43 struct moving_io {
  44         struct list_head        list;
  45         struct closure          cl;
  46         bool                    read_completed;
  47
  48         unsigned                read_sectors;
  49         unsigned                write_sectors;
  50
  51         struct bch_read_bio     rbio;
  52
  53         struct data_update      write;
  54         /* Must be last since it is variable size */
  55         struct bio_vec          bi_inline_vecs[0];
  56 };
  57
  58 static void move_free(struct moving_io *io)
  59 {
  60         struct moving_context *ctxt = io->write.ctxt;
  61         struct bch_fs *c = ctxt->c;
  62
  63         bch2_data_update_exit(&io->write);
  64         wake_up(&ctxt->wait);
  65         bch2_write_ref_put(c, BCH_WRITE_REF_move);
  66         kfree(io);
  67 }
  68
  69 static void move_write_done(struct bch_write_op *op)
  70 {
  71         struct moving_io *io = container_of(op, struct moving_io, write.op);
  72         struct moving_context *ctxt = io->write.ctxt;
  73
  74         if (io->write.op.error)
  75                 ctxt->write_error = true;
  76
  77         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
  78         atomic_dec(&io->write.ctxt->write_ios);
  79         move_free(io);
  80         closure_put(&ctxt->cl);
  81 }
  82
  83 static void move_write(struct moving_io *io)
  84 {
  85         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
  86                 move_free(io);
  87                 return;
  88         }
  89
  90         closure_get(&io->write.ctxt->cl);
  91         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
  92         atomic_inc(&io->write.ctxt->write_ios);
  93
  94         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
  95 }
  96
  97 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
  98 {
  99         struct moving_io *io =
 100                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 101
 102         return io && io->read_completed ? io : NULL;
 103 }
 104
 105 static void move_read_endio(struct bio *bio)
 106 {
 107         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 108         struct moving_context *ctxt = io->write.ctxt;
 109
 110         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 111         atomic_dec(&ctxt->read_ios);
 112         io->read_completed = true;
 113
 114         wake_up(&ctxt->wait);
 115         closure_put(&ctxt->cl);
 116 }
 117
 118 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 119                                         struct btree_trans *trans)
 120 {
 121         struct moving_io *io;
 122
 123         if (trans)
 124                 bch2_trans_unlock(trans);
 125
 126         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 127                 list_del(&io->list);
 128                 move_write(io);
 129         }
 130 }
 131
 132 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 133                                        struct btree_trans *trans)
 134 {
 135         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 136
 137         move_ctxt_wait_event(ctxt, trans,
 138                 !atomic_read(&ctxt->write_sectors) ||
 139                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 140 }
 141
 142 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 143 {
 144         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 145         closure_sync(&ctxt->cl);
 146
 147         EBUG_ON(atomic_read(&ctxt->write_sectors));
 148         EBUG_ON(atomic_read(&ctxt->write_ios));
 149         EBUG_ON(atomic_read(&ctxt->read_sectors));
 150         EBUG_ON(atomic_read(&ctxt->read_ios));
 151
 152         if (ctxt->stats) {
 153                 progress_list_del(ctxt->c, ctxt->stats);
 154
 155                 trace_move_data(ctxt->c,
 156                                 atomic64_read(&ctxt->stats->sectors_moved),
 157                                 atomic64_read(&ctxt->stats->keys_moved));
 158         }
 159 }
 160
 161 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 162                            struct bch_fs *c,
 163                            struct bch_ratelimit *rate,
 164                            struct bch_move_stats *stats,
 165                            struct write_point_specifier wp,
 166                            bool wait_on_copygc)
 167 {
 168         memset(ctxt, 0, sizeof(*ctxt));
 169
 170         ctxt->c         = c;
 171         ctxt->rate      = rate;
 172         ctxt->stats     = stats;
 173         ctxt->wp        = wp;
 174         ctxt->wait_on_copygc = wait_on_copygc;
 175
 176         closure_init_stack(&ctxt->cl);
 177         INIT_LIST_HEAD(&ctxt->reads);
 178         init_waitqueue_head(&ctxt->wait);
 179
 180         if (stats) {
 181                 progress_list_add(c, stats);
 182                 stats->data_type = BCH_DATA_user;
 183         }
 184 }
 185
 186 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 187 {
 188         memset(stats, 0, sizeof(*stats));
 189         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 190 }
 191
 192 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 193                                  struct btree_iter *iter,
 194                                  struct bkey_s_c k,
 195                                  struct data_update_opts data_opts)
 196 {
 197         struct bch_fs *c = trans->c;
 198         struct bkey_i *n;
 199         int ret;
 200
 201         n = bch2_bkey_make_mut(trans, k);
 202         ret = PTR_ERR_OR_ZERO(n);
 203         if (ret)
 204                 return ret;
 205
 206         while (data_opts.kill_ptrs) {
 207                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 208                 struct bch_extent_ptr *ptr;
 209
 210                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 211                 data_opts.kill_ptrs ^= 1U << drop;
 212         }
 213
 214         /*
 215          * If the new extent no longer has any pointers, bch2_extent_normalize()
 216          * will do the appropriate thing with it (turning it into a
 217          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 218          */
 219         bch2_extent_normalize(c, bkey_i_to_s(n));
 220
 221         /*
 222          * Since we're not inserting through an extent iterator
 223          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 224          * we aren't using the extent overwrite path to delete, we're
 225          * just using the normal key deletion path:
 226          */
 227         if (bkey_deleted(&n->k))
 228                 n->k.size = 0;
 229
 230         return bch2_trans_relock(trans) ?:
 231                 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 (!bch2_write_ref_tryget(c, BCH_WRITE_REF_move))
 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(trans, ctxt, &io->write, ctxt->wp,
 303                                     io_opts, data_opts, btree_id, k);
 304         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 305                 goto err_free_pages;
 306
 307         if (ret == -BCH_ERR_unwritten_extent_update) {
 308                 bch2_update_unwritten_extent(trans, &io->write);
 309                 move_free(io);
 310                 return 0;
 311         }
 312
 313         BUG_ON(ret);
 314
 315         io->write.ctxt = ctxt;
 316         io->write.op.end_io = move_write_done;
 317
 318         if (ctxt->stats) {
 319                 atomic64_inc(&ctxt->stats->keys_moved);
 320                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 321         }
 322
 323         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 324         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 325         trace_move_extent_read(k.k);
 326
 327         atomic_add(io->read_sectors, &ctxt->read_sectors);
 328         atomic_inc(&ctxt->read_ios);
 329         list_add_tail(&io->list, &ctxt->reads);
 330
 331         /*
 332          * dropped by move_read_endio() - guards against use after free of
 333          * ctxt when doing wakeup
 334          */
 335         closure_get(&ctxt->cl);
 336         bch2_read_extent(trans, &io->rbio,
 337                          bkey_start_pos(k.k),
 338                          btree_id, k, 0,
 339                          BCH_READ_NODECODE|
 340                          BCH_READ_LAST_FRAGMENT);
 341         return 0;
 342 err_free_pages:
 343         bio_free_pages(&io->write.op.wbio.bio);
 344 err_free:
 345         kfree(io);
 346 err:
 347         bch2_write_ref_put(c, BCH_WRITE_REF_move);
 348         trace_and_count(c, move_extent_alloc_mem_fail, k.k);
 349         return ret;
 350 }
 351
 352 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 353                         struct bch_inode_unpacked *inode)
 354 {
 355         struct btree_iter iter;
 356         struct bkey_s_c k;
 357         int ret;
 358
 359         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 360                              BTREE_ITER_ALL_SNAPSHOTS);
 361         k = bch2_btree_iter_peek(&iter);
 362         ret = bkey_err(k);
 363         if (ret)
 364                 goto err;
 365
 366         if (!k.k || !bkey_eq(k.k->p, pos)) {
 367                 ret = -ENOENT;
 368                 goto err;
 369         }
 370
 371         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 372         if (ret)
 373                 goto err;
 374
 375         ret = bch2_inode_unpack(k, inode);
 376         if (ret)
 377                 goto err;
 378 err:
 379         bch2_trans_iter_exit(trans, &iter);
 380         return ret;
 381 }
 382
 383 static int move_ratelimit(struct btree_trans *trans,
 384                           struct moving_context *ctxt)
 385 {
 386         struct bch_fs *c = trans->c;
 387         u64 delay;
 388
 389         if (ctxt->wait_on_copygc) {
 390                 bch2_trans_unlock(trans);
 391                 wait_event_killable(c->copygc_running_wq,
 392                                     !c->copygc_running ||
 393                                     kthread_should_stop());
 394         }
 395
 396         do {
 397                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 398
 399                 if (delay) {
 400                         bch2_trans_unlock(trans);
 401                         set_current_state(TASK_INTERRUPTIBLE);
 402                 }
 403
 404                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 405                         __set_current_state(TASK_RUNNING);
 406                         return 1;
 407                 }
 408
 409                 if (delay)
 410                         schedule_timeout(delay);
 411
 412                 if (unlikely(freezing(current))) {
 413                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 414                         try_to_freeze();
 415                 }
 416         } while (delay);
 417
 418         /*
 419          * XXX: these limits really ought to be per device, SSDs and hard drives
 420          * will want different limits
 421          */
 422         move_ctxt_wait_event(ctxt, trans,
 423                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 424                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 425                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 426                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 427
 428         return 0;
 429 }
 430
 431 static int move_get_io_opts(struct btree_trans *trans,
 432                             struct bch_io_opts *io_opts,
 433                             struct bkey_s_c k, u64 *cur_inum)
 434 {
 435         struct bch_inode_unpacked inode;
 436         int ret;
 437
 438         if (*cur_inum == k.k->p.inode)
 439                 return 0;
 440
 441         ret = lookup_inode(trans,
 442                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 443                            &inode);
 444         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 445                 return ret;
 446
 447         if (!ret)
 448                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 449         else
 450                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 451         *cur_inum = k.k->p.inode;
 452         return 0;
 453 }
 454
 455 static int __bch2_move_data(struct moving_context *ctxt,
 456                             struct bpos start,
 457                             struct bpos end,
 458                             move_pred_fn pred, void *arg,
 459                             enum btree_id btree_id)
 460 {
 461         struct bch_fs *c = ctxt->c;
 462         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 463         struct bkey_buf sk;
 464         struct btree_trans trans;
 465         struct btree_iter iter;
 466         struct bkey_s_c k;
 467         struct data_update_opts data_opts;
 468         u64 cur_inum = U64_MAX;
 469         int ret = 0, ret2;
 470
 471         bch2_bkey_buf_init(&sk);
 472         bch2_trans_init(&trans, c, 0, 0);
 473
 474         if (ctxt->stats) {
 475                 ctxt->stats->data_type  = BCH_DATA_user;
 476                 ctxt->stats->btree_id   = btree_id;
 477                 ctxt->stats->pos        = start;
 478         }
 479
 480         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 481                              BTREE_ITER_PREFETCH|
 482                              BTREE_ITER_ALL_SNAPSHOTS);
 483
 484         if (ctxt->rate)
 485                 bch2_ratelimit_reset(ctxt->rate);
 486
 487         while (!move_ratelimit(&trans, ctxt)) {
 488                 bch2_trans_begin(&trans);
 489
 490                 k = bch2_btree_iter_peek(&iter);
 491                 if (!k.k)
 492                         break;
 493
 494                 ret = bkey_err(k);
 495                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 496                         continue;
 497                 if (ret)
 498                         break;
 499
 500                 if (bkey_ge(bkey_start_pos(k.k), end))
 501                         break;
 502
 503                 if (ctxt->stats)
 504                         ctxt->stats->pos = iter.pos;
 505
 506                 if (!bkey_extent_is_direct_data(k.k))
 507                         goto next_nondata;
 508
 509                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 510                 if (ret)
 511                         continue;
 512
 513                 memset(&data_opts, 0, sizeof(data_opts));
 514                 if (!pred(c, arg, k, &io_opts, &data_opts))
 515                         goto next;
 516
 517                 /*
 518                  * The iterator gets unlocked by __bch2_read_extent - need to
 519                  * save a copy of @k elsewhere:
 520                  */
 521                 bch2_bkey_buf_reassemble(&sk, c, k);
 522                 k = bkey_i_to_s_c(sk.k);
 523                 bch2_trans_unlock(&trans);
 524
 525                 ret2 = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 526                                         btree_id, k, data_opts);
 527                 if (ret2) {
 528                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 529                                 continue;
 530
 531                         if (ret2 == -ENOMEM) {
 532                                 /* memory allocation failure, wait for some IO to finish */
 533                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 534                                 continue;
 535                         }
 536
 537                         /* XXX signal failure */
 538                         goto next;
 539                 }
 540
 541                 if (ctxt->rate)
 542                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 543 next:
 544                 if (ctxt->stats)
 545                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 546 next_nondata:
 547                 bch2_btree_iter_advance(&iter);
 548         }
 549
 550         bch2_trans_iter_exit(&trans, &iter);
 551         bch2_trans_exit(&trans);
 552         bch2_bkey_buf_exit(&sk, c);
 553
 554         return ret;
 555 }
 556
 557 int bch2_move_data(struct bch_fs *c,
 558                    enum btree_id start_btree_id, struct bpos start_pos,
 559                    enum btree_id end_btree_id,   struct bpos end_pos,
 560                    struct bch_ratelimit *rate,
 561                    struct bch_move_stats *stats,
 562                    struct write_point_specifier wp,
 563                    bool wait_on_copygc,
 564                    move_pred_fn pred, void *arg)
 565 {
 566         struct moving_context ctxt;
 567         enum btree_id id;
 568         int ret;
 569
 570         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 571
 572         for (id = start_btree_id;
 573              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 574              id++) {
 575                 stats->btree_id = id;
 576
 577                 if (id != BTREE_ID_extents &&
 578                     id != BTREE_ID_reflink)
 579                         continue;
 580
 581                 ret = __bch2_move_data(&ctxt,
 582                                        id == start_btree_id ? start_pos : POS_MIN,
 583                                        id == end_btree_id   ? end_pos   : POS_MAX,
 584                                        pred, arg, id);
 585                 if (ret)
 586                         break;
 587         }
 588
 589         bch2_moving_ctxt_exit(&ctxt);
 590
 591         return ret;
 592 }
 593
 594 void bch2_verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 595 {
 596         struct bch_fs *c = trans->c;
 597         struct btree_iter iter;
 598         struct bkey_s_c k;
 599         struct printbuf buf = PRINTBUF;
 600         struct bch_backpointer bp;
 601         u64 bp_offset = 0;
 602         int ret;
 603
 604         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 605                              bucket, BTREE_ITER_CACHED);
 606 again:
 607         ret = lockrestart_do(trans,
 608                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 609
 610         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 611                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 612
 613                 if (a.v->gen == gen &&
 614                     a.v->dirty_sectors) {
 615                         if (a.v->data_type == BCH_DATA_btree) {
 616                                 bch2_trans_unlock(trans);
 617                                 if (bch2_btree_interior_updates_flush(c))
 618                                         goto again;
 619                                 goto failed_to_evacuate;
 620                         }
 621                 }
 622         }
 623
 624         bch2_trans_iter_exit(trans, &iter);
 625         return;
 626 failed_to_evacuate:
 627         bch2_trans_iter_exit(trans, &iter);
 628
 629         if (test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
 630                 return;
 631
 632         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 633         bch2_bkey_val_to_text(&buf, c, k);
 634
 635         while (1) {
 636                 bch2_trans_begin(trans);
 637
 638                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 639                                                 &bp_offset, &bp,
 640                                                 BTREE_ITER_CACHED);
 641                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 642                         continue;
 643                 if (ret)
 644                         break;
 645                 if (bp_offset == U64_MAX)
 646                         break;
 647
 648                 k = bch2_backpointer_get_key(trans, &iter,
 649                                              bucket, bp_offset, bp);
 650                 ret = bkey_err(k);
 651                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 652                         continue;
 653                 if (ret)
 654                         break;
 655                 if (!k.k)
 656                         continue;
 657                 prt_newline(&buf);
 658                 bch2_bkey_val_to_text(&buf, c, k);
 659                 bch2_trans_iter_exit(trans, &iter);
 660         }
 661
 662         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 663         printbuf_exit(&buf);
 664 }
 665
 666 int __bch2_evacuate_bucket(struct btree_trans *trans,
 667                            struct moving_context *ctxt,
 668                            struct bpos bucket, int gen,
 669                            struct data_update_opts _data_opts)
 670 {
 671         struct bch_fs *c = ctxt->c;
 672         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 673         struct btree_iter iter;
 674         struct bkey_buf sk;
 675         struct bch_backpointer bp;
 676         struct bch_alloc_v4 a_convert;
 677         const struct bch_alloc_v4 *a;
 678         struct bkey_s_c k;
 679         struct data_update_opts data_opts;
 680         unsigned dirty_sectors, bucket_size;
 681         u64 fragmentation;
 682         u64 bp_offset = 0, cur_inum = U64_MAX;
 683         int ret = 0;
 684
 685         bch2_bkey_buf_init(&sk);
 686
 687         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 688                              bucket, BTREE_ITER_CACHED);
 689         ret = lockrestart_do(trans,
 690                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 691         bch2_trans_iter_exit(trans, &iter);
 692
 693         if (ret) {
 694                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 695                 goto err;
 696         }
 697
 698         a = bch2_alloc_to_v4(k, &a_convert);
 699         dirty_sectors = a->dirty_sectors;
 700         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 701         fragmentation = a->fragmentation_lru;
 702
 703         ret = bch2_btree_write_buffer_flush(trans);
 704         if (ret) {
 705                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 706                 goto err;
 707         }
 708
 709         while (!(ret = move_ratelimit(trans, ctxt))) {
 710                 bch2_trans_begin(trans);
 711
 712                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 713                                                 &bp_offset, &bp,
 714                                                 BTREE_ITER_CACHED);
 715                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 716                         continue;
 717                 if (ret)
 718                         goto err;
 719                 if (bp_offset == U64_MAX)
 720                         break;
 721
 722                 if (!bp.level) {
 723                         const struct bch_extent_ptr *ptr;
 724                         struct bkey_s_c k;
 725                         unsigned i = 0;
 726
 727                         k = bch2_backpointer_get_key(trans, &iter,
 728                                                 bucket, bp_offset, bp);
 729                         ret = bkey_err(k);
 730                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 731                                 continue;
 732                         if (ret)
 733                                 goto err;
 734                         if (!k.k)
 735                                 goto next;
 736
 737                         bch2_bkey_buf_reassemble(&sk, c, k);
 738                         k = bkey_i_to_s_c(sk.k);
 739
 740                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 741                         if (ret) {
 742                                 bch2_trans_iter_exit(trans, &iter);
 743                                 continue;
 744                         }
 745
 746                         data_opts = _data_opts;
 747                         data_opts.target        = io_opts.background_target;
 748                         data_opts.rewrite_ptrs = 0;
 749
 750                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 751                                 if (ptr->dev == bucket.inode)
 752                                         data_opts.rewrite_ptrs |= 1U << i;
 753                                 i++;
 754                         }
 755
 756                         ret = bch2_move_extent(trans, &iter, ctxt, io_opts,
 757                                                bp.btree_id, k, data_opts);
 758                         bch2_trans_iter_exit(trans, &iter);
 759
 760                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 761                                 continue;
 762                         if (ret == -ENOMEM) {
 763                                 /* memory allocation failure, wait for some IO to finish */
 764                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 765                                 continue;
 766                         }
 767                         if (ret)
 768                                 goto err;
 769
 770                         if (ctxt->rate)
 771                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 772                         if (ctxt->stats)
 773                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 774                 } else {
 775                         struct btree *b;
 776
 777                         b = bch2_backpointer_get_node(trans, &iter,
 778                                                 bucket, bp_offset, bp);
 779                         ret = PTR_ERR_OR_ZERO(b);
 780                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 781                                 continue;
 782                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 783                                 continue;
 784                         if (ret)
 785                                 goto err;
 786                         if (!b)
 787                                 goto next;
 788
 789                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 790                         bch2_trans_iter_exit(trans, &iter);
 791
 792                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 793                                 continue;
 794                         if (ret)
 795                                 goto err;
 796
 797                         if (ctxt->rate)
 798                                 bch2_ratelimit_increment(ctxt->rate,
 799                                                          c->opts.btree_node_size >> 9);
 800                         if (ctxt->stats) {
 801                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 802                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 803                         }
 804                 }
 805 next:
 806                 bp_offset++;
 807         }
 808
 809         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 810
 811         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
 812                 bch2_trans_unlock(trans);
 813                 move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 814                 closure_sync(&ctxt->cl);
 815                 if (!ctxt->write_error)
 816                         bch2_verify_bucket_evacuated(trans, bucket, gen);
 817         }
 818 err:
 819         bch2_bkey_buf_exit(&sk, c);
 820         return ret;
 821 }
 822
 823 int bch2_evacuate_bucket(struct bch_fs *c,
 824                          struct bpos bucket, int gen,
 825                          struct data_update_opts data_opts,
 826                          struct bch_ratelimit *rate,
 827                          struct bch_move_stats *stats,
 828                          struct write_point_specifier wp,
 829                          bool wait_on_copygc)
 830 {
 831         struct btree_trans trans;
 832         struct moving_context ctxt;
 833         int ret;
 834
 835         bch2_trans_init(&trans, c, 0, 0);
 836         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 837         ret = __bch2_evacuate_bucket(&trans, &ctxt, bucket, gen, data_opts);
 838         bch2_moving_ctxt_exit(&ctxt);
 839         bch2_trans_exit(&trans);
 840
 841         return ret;
 842 }
 843
 844 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 845                                 struct btree *, struct bch_io_opts *,
 846                                 struct data_update_opts *);
 847
 848 static int bch2_move_btree(struct bch_fs *c,
 849                            enum btree_id start_btree_id, struct bpos start_pos,
 850                            enum btree_id end_btree_id,   struct bpos end_pos,
 851                            move_btree_pred pred, void *arg,
 852                            struct bch_move_stats *stats)
 853 {
 854         bool kthread = (current->flags & PF_KTHREAD) != 0;
 855         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 856         struct btree_trans trans;
 857         struct btree_iter iter;
 858         struct btree *b;
 859         enum btree_id id;
 860         struct data_update_opts data_opts;
 861         int ret = 0;
 862
 863         bch2_trans_init(&trans, c, 0, 0);
 864         progress_list_add(c, stats);
 865
 866         stats->data_type = BCH_DATA_btree;
 867
 868         for (id = start_btree_id;
 869              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 870              id++) {
 871                 stats->btree_id = id;
 872
 873                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 874                                           BTREE_ITER_PREFETCH);
 875 retry:
 876                 ret = 0;
 877                 while (bch2_trans_begin(&trans),
 878                        (b = bch2_btree_iter_peek_node(&iter)) &&
 879                        !(ret = PTR_ERR_OR_ZERO(b))) {
 880                         if (kthread && kthread_should_stop())
 881                                 break;
 882
 883                         if ((cmp_int(id, end_btree_id) ?:
 884                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 885                                 break;
 886
 887                         stats->pos = iter.pos;
 888
 889                         if (!pred(c, arg, b, &io_opts, &data_opts))
 890                                 goto next;
 891
 892                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 893                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 894                                 continue;
 895                         if (ret)
 896                                 break;
 897 next:
 898                         bch2_btree_iter_next_node(&iter);
 899                 }
 900                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 901                         goto retry;
 902
 903                 bch2_trans_iter_exit(&trans, &iter);
 904
 905                 if (kthread && kthread_should_stop())
 906                         break;
 907         }
 908
 909         bch2_trans_exit(&trans);
 910
 911         if (ret)
 912                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 913
 914         bch2_btree_interior_updates_flush(c);
 915
 916         progress_list_del(c, stats);
 917         return ret;
 918 }
 919
 920 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 921                              struct bkey_s_c k,
 922                              struct bch_io_opts *io_opts,
 923                              struct data_update_opts *data_opts)
 924 {
 925         unsigned nr_good = bch2_bkey_durability(c, k);
 926         unsigned replicas = bkey_is_btree_ptr(k.k)
 927                 ? c->opts.metadata_replicas
 928                 : io_opts->data_replicas;
 929
 930         if (!nr_good || nr_good >= replicas)
 931                 return false;
 932
 933         data_opts->target               = 0;
 934         data_opts->extra_replicas       = replicas - nr_good;
 935         data_opts->btree_insert_flags   = 0;
 936         return true;
 937 }
 938
 939 static bool migrate_pred(struct bch_fs *c, void *arg,
 940                          struct bkey_s_c k,
 941                          struct bch_io_opts *io_opts,
 942                          struct data_update_opts *data_opts)
 943 {
 944         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 945         const struct bch_extent_ptr *ptr;
 946         struct bch_ioctl_data *op = arg;
 947         unsigned i = 0;
 948
 949         data_opts->rewrite_ptrs         = 0;
 950         data_opts->target               = 0;
 951         data_opts->extra_replicas       = 0;
 952         data_opts->btree_insert_flags   = 0;
 953
 954         bkey_for_each_ptr(ptrs, ptr) {
 955                 if (ptr->dev == op->migrate.dev)
 956                         data_opts->rewrite_ptrs |= 1U << i;
 957                 i++;
 958         }
 959
 960         return data_opts->rewrite_ptrs != 0;
 961 }
 962
 963 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 964                                    struct btree *b,
 965                                    struct bch_io_opts *io_opts,
 966                                    struct data_update_opts *data_opts)
 967 {
 968         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 969 }
 970
 971 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 972                                struct btree *b,
 973                                struct bch_io_opts *io_opts,
 974                                struct data_update_opts *data_opts)
 975 {
 976         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 977 }
 978
 979 static bool bformat_needs_redo(struct bkey_format *f)
 980 {
 981         unsigned i;
 982
 983         for (i = 0; i < f->nr_fields; i++) {
 984                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 985                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 986                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 987
 988                 if (f->bits_per_field[i] > unpacked_bits)
 989                         return true;
 990
 991                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 992                         return true;
 993
 994                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 995                      unpacked_mask) <
 996                     field_offset)
 997                         return true;
 998         }
 999
1000         return false;
1001 }
1002
1003 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1004                                    struct btree *b,
1005                                    struct bch_io_opts *io_opts,
1006                                    struct data_update_opts *data_opts)
1007 {
1008         if (b->version_ondisk != c->sb.version ||
1009             btree_node_need_rewrite(b) ||
1010             bformat_needs_redo(&b->format)) {
1011                 data_opts->target               = 0;
1012                 data_opts->extra_replicas       = 0;
1013                 data_opts->btree_insert_flags   = 0;
1014                 return true;
1015         }
1016
1017         return false;
1018 }
1019
1020 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1021 {
1022         int ret;
1023
1024         ret = bch2_move_btree(c,
1025                               0,                POS_MIN,
1026                               BTREE_ID_NR,      SPOS_MAX,
1027                               rewrite_old_nodes_pred, c, stats);
1028         if (!ret) {
1029                 mutex_lock(&c->sb_lock);
1030                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1031                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1032                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1033                 bch2_write_super(c);
1034                 mutex_unlock(&c->sb_lock);
1035         }
1036
1037         return ret;
1038 }
1039
1040 int bch2_data_job(struct bch_fs *c,
1041                   struct bch_move_stats *stats,
1042                   struct bch_ioctl_data op)
1043 {
1044         int ret = 0;
1045
1046         switch (op.op) {
1047         case BCH_DATA_OP_REREPLICATE:
1048                 bch2_move_stats_init(stats, "rereplicate");
1049                 stats->data_type = BCH_DATA_journal;
1050                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1051
1052                 ret = bch2_move_btree(c,
1053                                       op.start_btree,   op.start_pos,
1054                                       op.end_btree,     op.end_pos,
1055                                       rereplicate_btree_pred, c, stats) ?: ret;
1056                 ret = bch2_replicas_gc2(c) ?: ret;
1057
1058                 ret = bch2_move_data(c,
1059                                      op.start_btree,    op.start_pos,
1060                                      op.end_btree,      op.end_pos,
1061                                      NULL,
1062                                      stats,
1063                                      writepoint_hashed((unsigned long) current),
1064                                      true,
1065                                      rereplicate_pred, c) ?: ret;
1066                 ret = bch2_replicas_gc2(c) ?: ret;
1067                 break;
1068         case BCH_DATA_OP_MIGRATE:
1069                 if (op.migrate.dev >= c->sb.nr_devices)
1070                         return -EINVAL;
1071
1072                 bch2_move_stats_init(stats, "migrate");
1073                 stats->data_type = BCH_DATA_journal;
1074                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1075
1076                 ret = bch2_move_btree(c,
1077                                       op.start_btree,   op.start_pos,
1078                                       op.end_btree,     op.end_pos,
1079                                       migrate_btree_pred, &op, stats) ?: ret;
1080                 ret = bch2_replicas_gc2(c) ?: ret;
1081
1082                 ret = bch2_move_data(c,
1083                                      op.start_btree,    op.start_pos,
1084                                      op.end_btree,      op.end_pos,
1085                                      NULL,
1086                                      stats,
1087                                      writepoint_hashed((unsigned long) current),
1088                                      true,
1089                                      migrate_pred, &op) ?: ret;
1090                 ret = bch2_replicas_gc2(c) ?: ret;
1091                 break;
1092         case BCH_DATA_OP_REWRITE_OLD_NODES:
1093                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1094                 ret = bch2_scan_old_btree_nodes(c, stats);
1095                 break;
1096         default:
1097                 ret = -EINVAL;
1098         }
1099
1100         return ret;
1101 }