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_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 231                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 232 }
 233
 234 static int bch2_move_extent(struct btree_trans *trans,
 235                             struct btree_iter *iter,
 236                             struct moving_context *ctxt,
 237                             struct bch_io_opts io_opts,
 238                             enum btree_id btree_id,
 239                             struct bkey_s_c k,
 240                             struct data_update_opts data_opts)
 241 {
 242         struct bch_fs *c = trans->c;
 243         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 244         struct moving_io *io;
 245         const union bch_extent_entry *entry;
 246         struct extent_ptr_decoded p;
 247         unsigned sectors = k.k->size, pages;
 248         int ret = -ENOMEM;
 249
 250         bch2_data_update_opts_normalize(k, &data_opts);
 251
 252         if (!data_opts.rewrite_ptrs &&
 253             !data_opts.extra_replicas) {
 254                 if (data_opts.kill_ptrs)
 255                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 256                 return 0;
 257         }
 258
 259         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_move))
 260                 return -BCH_ERR_erofs_no_writes;
 261
 262         /*
 263          * Before memory allocations & taking nocow locks in
 264          * bch2_data_update_init():
 265          */
 266         bch2_trans_unlock(trans);
 267
 268         /* write path might have to decompress data: */
 269         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 270                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 271
 272         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 273         io = kzalloc(sizeof(struct moving_io) +
 274                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 275         if (!io)
 276                 goto err;
 277
 278         io->write.ctxt          = ctxt;
 279         io->read_sectors        = k.k->size;
 280         io->write_sectors       = k.k->size;
 281
 282         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 283         bio_set_prio(&io->write.op.wbio.bio,
 284                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 285
 286         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 287                                  GFP_KERNEL))
 288                 goto err_free;
 289
 290         io->rbio.c              = c;
 291         io->rbio.opts           = io_opts;
 292         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 293         io->rbio.bio.bi_vcnt = pages;
 294         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 295         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 296
 297         bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
 298         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 299         io->rbio.bio.bi_end_io          = move_read_endio;
 300
 301         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 302                                     io_opts, data_opts, btree_id, k);
 303         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 304                 goto err_free_pages;
 305
 306         io->write.ctxt = ctxt;
 307         io->write.op.end_io = move_write_done;
 308
 309         atomic64_inc(&ctxt->stats->keys_moved);
 310         atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 311
 312         if (ret == -BCH_ERR_unwritten_extent_update) {
 313                 bch2_update_unwritten_extent(trans, &io->write);
 314                 move_free(io);
 315                 return 0;
 316         }
 317
 318         BUG_ON(ret);
 319
 320         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 321         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 322         trace_move_extent_read(k.k);
 323
 324         atomic_add(io->read_sectors, &ctxt->read_sectors);
 325         atomic_inc(&ctxt->read_ios);
 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         bch2_write_ref_put(c, BCH_WRITE_REF_move);
 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         /*
 416          * XXX: these limits really ought to be per device, SSDs and hard drives
 417          * will want different limits
 418          */
 419         move_ctxt_wait_event(ctxt, trans,
 420                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 421                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 422                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 423                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 424
 425         return 0;
 426 }
 427
 428 static int move_get_io_opts(struct btree_trans *trans,
 429                             struct bch_io_opts *io_opts,
 430                             struct bkey_s_c k, u64 *cur_inum)
 431 {
 432         struct bch_inode_unpacked inode;
 433         int ret;
 434
 435         if (*cur_inum == k.k->p.inode)
 436                 return 0;
 437
 438         ret = lookup_inode(trans,
 439                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 440                            &inode);
 441         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 442                 return ret;
 443
 444         if (!ret)
 445                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 446         else
 447                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 448         *cur_inum = k.k->p.inode;
 449         return 0;
 450 }
 451
 452 static int __bch2_move_data(struct moving_context *ctxt,
 453                             struct bpos start,
 454                             struct bpos end,
 455                             move_pred_fn pred, void *arg,
 456                             enum btree_id btree_id)
 457 {
 458         struct bch_fs *c = ctxt->c;
 459         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 460         struct bkey_buf sk;
 461         struct btree_trans trans;
 462         struct btree_iter iter;
 463         struct bkey_s_c k;
 464         struct data_update_opts data_opts;
 465         u64 cur_inum = U64_MAX;
 466         int ret = 0, ret2;
 467
 468         bch2_bkey_buf_init(&sk);
 469         bch2_trans_init(&trans, c, 0, 0);
 470
 471         ctxt->stats->data_type  = BCH_DATA_user;
 472         ctxt->stats->btree_id   = btree_id;
 473         ctxt->stats->pos        = start;
 474
 475         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 476                              BTREE_ITER_PREFETCH|
 477                              BTREE_ITER_ALL_SNAPSHOTS);
 478
 479         if (ctxt->rate)
 480                 bch2_ratelimit_reset(ctxt->rate);
 481
 482         while (!move_ratelimit(&trans, ctxt)) {
 483                 bch2_trans_begin(&trans);
 484
 485                 k = bch2_btree_iter_peek(&iter);
 486                 if (!k.k)
 487                         break;
 488
 489                 ret = bkey_err(k);
 490                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 491                         continue;
 492                 if (ret)
 493                         break;
 494
 495                 if (bkey_ge(bkey_start_pos(k.k), end))
 496                         break;
 497
 498                 ctxt->stats->pos = iter.pos;
 499
 500                 if (!bkey_extent_is_direct_data(k.k))
 501                         goto next_nondata;
 502
 503                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 504                 if (ret)
 505                         continue;
 506
 507                 memset(&data_opts, 0, sizeof(data_opts));
 508                 if (!pred(c, arg, k, &io_opts, &data_opts))
 509                         goto next;
 510
 511                 /*
 512                  * The iterator gets unlocked by __bch2_read_extent - need to
 513                  * save a copy of @k elsewhere:
 514                  */
 515                 bch2_bkey_buf_reassemble(&sk, c, k);
 516                 k = bkey_i_to_s_c(sk.k);
 517                 bch2_trans_unlock(&trans);
 518
 519                 ret2 = bch2_move_extent(&trans, &iter, ctxt, io_opts,
 520                                         btree_id, k, data_opts);
 521                 if (ret2) {
 522                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 523                                 continue;
 524
 525                         if (ret2 == -ENOMEM) {
 526                                 /* memory allocation failure, wait for some IO to finish */
 527                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 528                                 continue;
 529                         }
 530
 531                         /* XXX signal failure */
 532                         goto next;
 533                 }
 534
 535                 if (ctxt->rate)
 536                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 537 next:
 538                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 539 next_nondata:
 540                 bch2_btree_iter_advance(&iter);
 541         }
 542
 543         bch2_trans_iter_exit(&trans, &iter);
 544         bch2_trans_exit(&trans);
 545         bch2_bkey_buf_exit(&sk, c);
 546
 547         return ret;
 548 }
 549
 550 int bch2_move_data(struct bch_fs *c,
 551                    enum btree_id start_btree_id, struct bpos start_pos,
 552                    enum btree_id end_btree_id,   struct bpos end_pos,
 553                    struct bch_ratelimit *rate,
 554                    struct bch_move_stats *stats,
 555                    struct write_point_specifier wp,
 556                    bool wait_on_copygc,
 557                    move_pred_fn pred, void *arg)
 558 {
 559         struct moving_context ctxt;
 560         enum btree_id id;
 561         int ret;
 562
 563         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 564
 565         for (id = start_btree_id;
 566              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 567              id++) {
 568                 stats->btree_id = id;
 569
 570                 if (id != BTREE_ID_extents &&
 571                     id != BTREE_ID_reflink)
 572                         continue;
 573
 574                 ret = __bch2_move_data(&ctxt,
 575                                        id == start_btree_id ? start_pos : POS_MIN,
 576                                        id == end_btree_id   ? end_pos   : POS_MAX,
 577                                        pred, arg, id);
 578                 if (ret)
 579                         break;
 580         }
 581
 582         bch2_moving_ctxt_exit(&ctxt);
 583
 584         return ret;
 585 }
 586
 587 static noinline void verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 588 {
 589         struct bch_fs *c = trans->c;
 590         struct btree_iter iter;
 591         struct bkey_s_c k;
 592         struct printbuf buf = PRINTBUF;
 593         struct bch_backpointer bp;
 594         u64 bp_offset = 0;
 595         int ret;
 596
 597         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 598                              bucket, BTREE_ITER_CACHED);
 599 again:
 600         ret = lockrestart_do(trans,
 601                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 602
 603         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 604                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 605
 606                 if (a.v->gen == gen &&
 607                     a.v->dirty_sectors) {
 608                         if (a.v->data_type == BCH_DATA_btree) {
 609                                 bch2_trans_unlock(trans);
 610                                 if (bch2_btree_interior_updates_flush(c))
 611                                         goto again;
 612                                 goto failed_to_evacuate;
 613                         }
 614                 }
 615         }
 616
 617         bch2_trans_iter_exit(trans, &iter);
 618         return;
 619 failed_to_evacuate:
 620         bch2_trans_iter_exit(trans, &iter);
 621
 622         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 623         bch2_bkey_val_to_text(&buf, c, k);
 624
 625         while (1) {
 626                 bch2_trans_begin(trans);
 627
 628                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 629                                                 &bp_offset, &bp,
 630                                                 BTREE_ITER_CACHED);
 631                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 632                         continue;
 633                 if (ret)
 634                         break;
 635                 if (bp_offset == U64_MAX)
 636                         break;
 637
 638                 k = bch2_backpointer_get_key(trans, &iter,
 639                                              bucket, bp_offset, bp);
 640                 ret = bkey_err(k);
 641                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 642                         continue;
 643                 if (ret)
 644                         break;
 645                 if (!k.k)
 646                         continue;
 647                 prt_newline(&buf);
 648                 bch2_bkey_val_to_text(&buf, c, k);
 649                 bch2_trans_iter_exit(trans, &iter);
 650         }
 651
 652         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 653         printbuf_exit(&buf);
 654 }
 655
 656 int __bch2_evacuate_bucket(struct btree_trans *trans,
 657                            struct moving_context *ctxt,
 658                            struct bpos bucket, int gen,
 659                            struct data_update_opts _data_opts)
 660 {
 661         struct bch_fs *c = ctxt->c;
 662         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 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 fragmentation;
 672         u64 bp_offset = 0, cur_inum = U64_MAX;
 673         int ret = 0;
 674
 675         bch2_bkey_buf_init(&sk);
 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                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 685                 goto err;
 686         }
 687
 688         a = bch2_alloc_to_v4(k, &a_convert);
 689         dirty_sectors = a->dirty_sectors;
 690         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 691         fragmentation = a->fragmentation_lru;
 692
 693         ret = bch2_btree_write_buffer_flush(trans);
 694         if (ret) {
 695                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 696                 goto err;
 697         }
 698
 699         while (!(ret = move_ratelimit(trans, ctxt))) {
 700                 bch2_trans_begin(trans);
 701
 702                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 703                                                 &bp_offset, &bp,
 704                                                 BTREE_ITER_CACHED);
 705                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 706                         continue;
 707                 if (ret)
 708                         goto err;
 709                 if (bp_offset == U64_MAX)
 710                         break;
 711
 712                 if (!bp.level) {
 713                         const struct bch_extent_ptr *ptr;
 714                         struct bkey_s_c k;
 715                         unsigned i = 0;
 716
 717                         k = bch2_backpointer_get_key(trans, &iter,
 718                                                 bucket, bp_offset, bp);
 719                         ret = bkey_err(k);
 720                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 721                                 continue;
 722                         if (ret)
 723                                 goto err;
 724                         if (!k.k)
 725                                 goto next;
 726
 727                         bch2_bkey_buf_reassemble(&sk, c, k);
 728                         k = bkey_i_to_s_c(sk.k);
 729
 730                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 731                         if (ret) {
 732                                 bch2_trans_iter_exit(trans, &iter);
 733                                 continue;
 734                         }
 735
 736                         data_opts = _data_opts;
 737                         data_opts.target        = io_opts.background_target;
 738                         data_opts.rewrite_ptrs = 0;
 739
 740                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 741                                 if (ptr->dev == bucket.inode)
 742                                         data_opts.rewrite_ptrs |= 1U << i;
 743                                 i++;
 744                         }
 745
 746                         ret = bch2_move_extent(trans, &iter, ctxt, io_opts,
 747                                                bp.btree_id, k, data_opts);
 748                         bch2_trans_iter_exit(trans, &iter);
 749
 750                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 751                                 continue;
 752                         if (ret == -ENOMEM) {
 753                                 /* memory allocation failure, wait for some IO to finish */
 754                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 755                                 continue;
 756                         }
 757                         if (ret)
 758                                 goto err;
 759
 760                         if (ctxt->rate)
 761                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 762                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 763                 } else {
 764                         struct btree *b;
 765
 766                         b = bch2_backpointer_get_node(trans, &iter,
 767                                                 bucket, bp_offset, bp);
 768                         ret = PTR_ERR_OR_ZERO(b);
 769                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 770                                 continue;
 771                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 772                                 continue;
 773                         if (ret)
 774                                 goto err;
 775                         if (!b)
 776                                 goto next;
 777
 778                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 779                         bch2_trans_iter_exit(trans, &iter);
 780
 781                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 782                                 continue;
 783                         if (ret)
 784                                 goto err;
 785
 786                         if (ctxt->rate)
 787                                 bch2_ratelimit_increment(ctxt->rate,
 788                                                          c->opts.btree_node_size >> 9);
 789                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 790                         atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 791                 }
 792 next:
 793                 bp_offset++;
 794         }
 795
 796         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 797
 798         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
 799                 bch2_trans_unlock(trans);
 800                 move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 801                 closure_sync(&ctxt->cl);
 802                 if (!ctxt->write_error)
 803                         verify_bucket_evacuated(trans, bucket, gen);
 804         }
 805 err:
 806         bch2_bkey_buf_exit(&sk, c);
 807         return ret;
 808 }
 809
 810 int bch2_evacuate_bucket(struct bch_fs *c,
 811                          struct bpos bucket, int gen,
 812                          struct data_update_opts data_opts,
 813                          struct bch_ratelimit *rate,
 814                          struct bch_move_stats *stats,
 815                          struct write_point_specifier wp,
 816                          bool wait_on_copygc)
 817 {
 818         struct btree_trans trans;
 819         struct moving_context ctxt;
 820         int ret;
 821
 822         bch2_trans_init(&trans, c, 0, 0);
 823         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 824         ret = __bch2_evacuate_bucket(&trans, &ctxt, bucket, gen, data_opts);
 825         bch2_moving_ctxt_exit(&ctxt);
 826         bch2_trans_exit(&trans);
 827
 828         return ret;
 829 }
 830
 831 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 832                                 struct btree *, struct bch_io_opts *,
 833                                 struct data_update_opts *);
 834
 835 static int bch2_move_btree(struct bch_fs *c,
 836                            enum btree_id start_btree_id, struct bpos start_pos,
 837                            enum btree_id end_btree_id,   struct bpos end_pos,
 838                            move_btree_pred pred, void *arg,
 839                            struct bch_move_stats *stats)
 840 {
 841         bool kthread = (current->flags & PF_KTHREAD) != 0;
 842         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 843         struct btree_trans trans;
 844         struct btree_iter iter;
 845         struct btree *b;
 846         enum btree_id id;
 847         struct data_update_opts data_opts;
 848         int ret = 0;
 849
 850         bch2_trans_init(&trans, c, 0, 0);
 851         progress_list_add(c, stats);
 852
 853         stats->data_type = BCH_DATA_btree;
 854
 855         for (id = start_btree_id;
 856              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 857              id++) {
 858                 stats->btree_id = id;
 859
 860                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 861                                           BTREE_ITER_PREFETCH);
 862 retry:
 863                 ret = 0;
 864                 while (bch2_trans_begin(&trans),
 865                        (b = bch2_btree_iter_peek_node(&iter)) &&
 866                        !(ret = PTR_ERR_OR_ZERO(b))) {
 867                         if (kthread && kthread_should_stop())
 868                                 break;
 869
 870                         if ((cmp_int(id, end_btree_id) ?:
 871                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 872                                 break;
 873
 874                         stats->pos = iter.pos;
 875
 876                         if (!pred(c, arg, b, &io_opts, &data_opts))
 877                                 goto next;
 878
 879                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 880                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 881                                 continue;
 882                         if (ret)
 883                                 break;
 884 next:
 885                         bch2_btree_iter_next_node(&iter);
 886                 }
 887                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 888                         goto retry;
 889
 890                 bch2_trans_iter_exit(&trans, &iter);
 891
 892                 if (kthread && kthread_should_stop())
 893                         break;
 894         }
 895
 896         bch2_trans_exit(&trans);
 897
 898         if (ret)
 899                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 900
 901         bch2_btree_interior_updates_flush(c);
 902
 903         progress_list_del(c, stats);
 904         return ret;
 905 }
 906
 907 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 908                              struct bkey_s_c k,
 909                              struct bch_io_opts *io_opts,
 910                              struct data_update_opts *data_opts)
 911 {
 912         unsigned nr_good = bch2_bkey_durability(c, k);
 913         unsigned replicas = bkey_is_btree_ptr(k.k)
 914                 ? c->opts.metadata_replicas
 915                 : io_opts->data_replicas;
 916
 917         if (!nr_good || nr_good >= replicas)
 918                 return false;
 919
 920         data_opts->target               = 0;
 921         data_opts->extra_replicas       = replicas - nr_good;
 922         data_opts->btree_insert_flags   = 0;
 923         return true;
 924 }
 925
 926 static bool migrate_pred(struct bch_fs *c, void *arg,
 927                          struct bkey_s_c k,
 928                          struct bch_io_opts *io_opts,
 929                          struct data_update_opts *data_opts)
 930 {
 931         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 932         const struct bch_extent_ptr *ptr;
 933         struct bch_ioctl_data *op = arg;
 934         unsigned i = 0;
 935
 936         data_opts->rewrite_ptrs         = 0;
 937         data_opts->target               = 0;
 938         data_opts->extra_replicas       = 0;
 939         data_opts->btree_insert_flags   = 0;
 940
 941         bkey_for_each_ptr(ptrs, ptr) {
 942                 if (ptr->dev == op->migrate.dev)
 943                         data_opts->rewrite_ptrs |= 1U << i;
 944                 i++;
 945         }
 946
 947         return data_opts->rewrite_ptrs != 0;
 948 }
 949
 950 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 951                                    struct btree *b,
 952                                    struct bch_io_opts *io_opts,
 953                                    struct data_update_opts *data_opts)
 954 {
 955         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 956 }
 957
 958 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 959                                struct btree *b,
 960                                struct bch_io_opts *io_opts,
 961                                struct data_update_opts *data_opts)
 962 {
 963         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 964 }
 965
 966 static bool bformat_needs_redo(struct bkey_format *f)
 967 {
 968         unsigned i;
 969
 970         for (i = 0; i < f->nr_fields; i++) {
 971                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 972                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 973                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 974
 975                 if (f->bits_per_field[i] > unpacked_bits)
 976                         return true;
 977
 978                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 979                         return true;
 980
 981                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 982                      unpacked_mask) <
 983                     field_offset)
 984                         return true;
 985         }
 986
 987         return false;
 988 }
 989
 990 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 991                                    struct btree *b,
 992                                    struct bch_io_opts *io_opts,
 993                                    struct data_update_opts *data_opts)
 994 {
 995         if (b->version_ondisk != c->sb.version ||
 996             btree_node_need_rewrite(b) ||
 997             bformat_needs_redo(&b->format)) {
 998                 data_opts->target               = 0;
 999                 data_opts->extra_replicas       = 0;
1000                 data_opts->btree_insert_flags   = 0;
1001                 return true;
1002         }
1003
1004         return false;
1005 }
1006
1007 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1008 {
1009         int ret;
1010
1011         ret = bch2_move_btree(c,
1012                               0,                POS_MIN,
1013                               BTREE_ID_NR,      SPOS_MAX,
1014                               rewrite_old_nodes_pred, c, stats);
1015         if (!ret) {
1016                 mutex_lock(&c->sb_lock);
1017                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1018                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1019                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1020                 bch2_write_super(c);
1021                 mutex_unlock(&c->sb_lock);
1022         }
1023
1024         return ret;
1025 }
1026
1027 int bch2_data_job(struct bch_fs *c,
1028                   struct bch_move_stats *stats,
1029                   struct bch_ioctl_data op)
1030 {
1031         int ret = 0;
1032
1033         switch (op.op) {
1034         case BCH_DATA_OP_REREPLICATE:
1035                 bch2_move_stats_init(stats, "rereplicate");
1036                 stats->data_type = BCH_DATA_journal;
1037                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1038
1039                 ret = bch2_move_btree(c,
1040                                       op.start_btree,   op.start_pos,
1041                                       op.end_btree,     op.end_pos,
1042                                       rereplicate_btree_pred, c, stats) ?: ret;
1043                 ret = bch2_replicas_gc2(c) ?: ret;
1044
1045                 ret = bch2_move_data(c,
1046                                      op.start_btree,    op.start_pos,
1047                                      op.end_btree,      op.end_pos,
1048                                      NULL,
1049                                      stats,
1050                                      writepoint_hashed((unsigned long) current),
1051                                      true,
1052                                      rereplicate_pred, c) ?: ret;
1053                 ret = bch2_replicas_gc2(c) ?: ret;
1054                 break;
1055         case BCH_DATA_OP_MIGRATE:
1056                 if (op.migrate.dev >= c->sb.nr_devices)
1057                         return -EINVAL;
1058
1059                 bch2_move_stats_init(stats, "migrate");
1060                 stats->data_type = BCH_DATA_journal;
1061                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1062
1063                 ret = bch2_move_btree(c,
1064                                       op.start_btree,   op.start_pos,
1065                                       op.end_btree,     op.end_pos,
1066                                       migrate_btree_pred, &op, stats) ?: ret;
1067                 ret = bch2_replicas_gc2(c) ?: ret;
1068
1069                 ret = bch2_move_data(c,
1070                                      op.start_btree,    op.start_pos,
1071                                      op.end_btree,      op.end_pos,
1072                                      NULL,
1073                                      stats,
1074                                      writepoint_hashed((unsigned long) current),
1075                                      true,
1076                                      migrate_pred, &op) ?: ret;
1077                 ret = bch2_replicas_gc2(c) ?: ret;
1078                 break;
1079         case BCH_DATA_OP_REWRITE_OLD_NODES:
1080                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1081                 ret = bch2_scan_old_btree_nodes(c, stats);
1082                 break;
1083         default:
1084                 ret = -EINVAL;
1085         }
1086
1087         return ret;
1088 }