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 move_bucket_in_flight    *b;
  46         struct closure                  cl;
  47         bool                            read_completed;
  48
  49         unsigned                        read_sectors;
  50         unsigned                        write_sectors;
  51
  52         struct bch_read_bio             rbio;
  53
  54         struct data_update              write;
  55         /* Must be last since it is variable size */
  56         struct bio_vec                  bi_inline_vecs[0];
  57 };
  58
  59 static void move_free(struct moving_io *io)
  60 {
  61         struct moving_context *ctxt = io->write.ctxt;
  62         struct bch_fs *c = ctxt->c;
  63
  64         if (io->b)
  65                 atomic_dec(&io->b->count);
  66
  67         bch2_data_update_exit(&io->write);
  68         wake_up(&ctxt->wait);
  69         bch2_write_ref_put(c, BCH_WRITE_REF_move);
  70         kfree(io);
  71 }
  72
  73 static void move_write_done(struct bch_write_op *op)
  74 {
  75         struct moving_io *io = container_of(op, struct moving_io, write.op);
  76         struct moving_context *ctxt = io->write.ctxt;
  77
  78         if (io->write.op.error)
  79                 ctxt->write_error = true;
  80
  81         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
  82         atomic_dec(&io->write.ctxt->write_ios);
  83         move_free(io);
  84         closure_put(&ctxt->cl);
  85 }
  86
  87 static void move_write(struct moving_io *io)
  88 {
  89         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
  90                 move_free(io);
  91                 return;
  92         }
  93
  94         closure_get(&io->write.ctxt->cl);
  95         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
  96         atomic_inc(&io->write.ctxt->write_ios);
  97
  98         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
  99 }
 100
 101 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 102 {
 103         struct moving_io *io =
 104                 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
 105
 106         return io && io->read_completed ? io : NULL;
 107 }
 108
 109 static void move_read_endio(struct bio *bio)
 110 {
 111         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 112         struct moving_context *ctxt = io->write.ctxt;
 113
 114         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 115         atomic_dec(&ctxt->read_ios);
 116         io->read_completed = true;
 117
 118         wake_up(&ctxt->wait);
 119         closure_put(&ctxt->cl);
 120 }
 121
 122 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 123                                         struct btree_trans *trans)
 124 {
 125         struct moving_io *io;
 126
 127         if (trans)
 128                 bch2_trans_unlock(trans);
 129
 130         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 131                 list_del(&io->list);
 132                 move_write(io);
 133         }
 134 }
 135
 136 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 137                                        struct btree_trans *trans)
 138 {
 139         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 140
 141         move_ctxt_wait_event(ctxt, trans,
 142                 !atomic_read(&ctxt->write_sectors) ||
 143                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 144 }
 145
 146 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 147 {
 148         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 149         closure_sync(&ctxt->cl);
 150
 151         EBUG_ON(atomic_read(&ctxt->write_sectors));
 152         EBUG_ON(atomic_read(&ctxt->write_ios));
 153         EBUG_ON(atomic_read(&ctxt->read_sectors));
 154         EBUG_ON(atomic_read(&ctxt->read_ios));
 155
 156         if (ctxt->stats) {
 157                 progress_list_del(ctxt->c, ctxt->stats);
 158
 159                 trace_move_data(ctxt->c,
 160                                 atomic64_read(&ctxt->stats->sectors_moved),
 161                                 atomic64_read(&ctxt->stats->keys_moved));
 162         }
 163 }
 164
 165 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 166                            struct bch_fs *c,
 167                            struct bch_ratelimit *rate,
 168                            struct bch_move_stats *stats,
 169                            struct write_point_specifier wp,
 170                            bool wait_on_copygc)
 171 {
 172         memset(ctxt, 0, sizeof(*ctxt));
 173
 174         ctxt->c         = c;
 175         ctxt->rate      = rate;
 176         ctxt->stats     = stats;
 177         ctxt->wp        = wp;
 178         ctxt->wait_on_copygc = wait_on_copygc;
 179
 180         closure_init_stack(&ctxt->cl);
 181         INIT_LIST_HEAD(&ctxt->reads);
 182         init_waitqueue_head(&ctxt->wait);
 183
 184         if (stats) {
 185                 progress_list_add(c, stats);
 186                 stats->data_type = BCH_DATA_user;
 187         }
 188 }
 189
 190 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 191 {
 192         memset(stats, 0, sizeof(*stats));
 193         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 194 }
 195
 196 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 197                                  struct btree_iter *iter,
 198                                  struct bkey_s_c k,
 199                                  struct data_update_opts data_opts)
 200 {
 201         struct bch_fs *c = trans->c;
 202         struct bkey_i *n;
 203         int ret;
 204
 205         n = bch2_bkey_make_mut(trans, k);
 206         ret = PTR_ERR_OR_ZERO(n);
 207         if (ret)
 208                 return ret;
 209
 210         while (data_opts.kill_ptrs) {
 211                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 212                 struct bch_extent_ptr *ptr;
 213
 214                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 215                 data_opts.kill_ptrs ^= 1U << drop;
 216         }
 217
 218         /*
 219          * If the new extent no longer has any pointers, bch2_extent_normalize()
 220          * will do the appropriate thing with it (turning it into a
 221          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 222          */
 223         bch2_extent_normalize(c, bkey_i_to_s(n));
 224
 225         /*
 226          * Since we're not inserting through an extent iterator
 227          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 228          * we aren't using the extent overwrite path to delete, we're
 229          * just using the normal key deletion path:
 230          */
 231         if (bkey_deleted(&n->k))
 232                 n->k.size = 0;
 233
 234         return bch2_trans_relock(trans) ?:
 235                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 236                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 237 }
 238
 239 static int bch2_move_extent(struct btree_trans *trans,
 240                             struct btree_iter *iter,
 241                             struct moving_context *ctxt,
 242                             struct move_bucket_in_flight *bucket_in_flight,
 243                             struct bch_io_opts io_opts,
 244                             enum btree_id btree_id,
 245                             struct bkey_s_c k,
 246                             struct data_update_opts data_opts)
 247 {
 248         struct bch_fs *c = trans->c;
 249         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 250         struct moving_io *io;
 251         const union bch_extent_entry *entry;
 252         struct extent_ptr_decoded p;
 253         unsigned sectors = k.k->size, pages;
 254         int ret = -ENOMEM;
 255
 256         bch2_data_update_opts_normalize(k, &data_opts);
 257
 258         if (!data_opts.rewrite_ptrs &&
 259             !data_opts.extra_replicas) {
 260                 if (data_opts.kill_ptrs)
 261                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 262                 return 0;
 263         }
 264
 265         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_move))
 266                 return -BCH_ERR_erofs_no_writes;
 267
 268         /*
 269          * Before memory allocations & taking nocow locks in
 270          * bch2_data_update_init():
 271          */
 272         bch2_trans_unlock(trans);
 273
 274         /* write path might have to decompress data: */
 275         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 276                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 277
 278         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 279         io = kzalloc(sizeof(struct moving_io) +
 280                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 281         if (!io)
 282                 goto err;
 283
 284         io->write.ctxt          = ctxt;
 285         io->read_sectors        = k.k->size;
 286         io->write_sectors       = k.k->size;
 287
 288         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 289         bio_set_prio(&io->write.op.wbio.bio,
 290                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 291
 292         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 293                                  GFP_KERNEL))
 294                 goto err_free;
 295
 296         io->rbio.c              = c;
 297         io->rbio.opts           = io_opts;
 298         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 299         io->rbio.bio.bi_vcnt = pages;
 300         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 301         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 302
 303         io->rbio.bio.bi_opf             = REQ_OP_READ;
 304         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 305         io->rbio.bio.bi_end_io          = move_read_endio;
 306
 307         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 308                                     io_opts, data_opts, btree_id, k);
 309         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 310                 goto err_free_pages;
 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         io->write.ctxt = ctxt;
 321         io->write.op.end_io = move_write_done;
 322
 323         if (ctxt->stats) {
 324                 atomic64_inc(&ctxt->stats->keys_moved);
 325                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 326         }
 327
 328         if (bucket_in_flight) {
 329                 io->b = bucket_in_flight;
 330                 atomic_inc(&io->b->count);
 331         }
 332
 333         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 334         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 335         trace_move_extent_read(k.k);
 336
 337         atomic_add(io->read_sectors, &ctxt->read_sectors);
 338         atomic_inc(&ctxt->read_ios);
 339         list_add_tail(&io->list, &ctxt->reads);
 340
 341         /*
 342          * dropped by move_read_endio() - guards against use after free of
 343          * ctxt when doing wakeup
 344          */
 345         closure_get(&ctxt->cl);
 346         bch2_read_extent(trans, &io->rbio,
 347                          bkey_start_pos(k.k),
 348                          btree_id, k, 0,
 349                          BCH_READ_NODECODE|
 350                          BCH_READ_LAST_FRAGMENT);
 351         return 0;
 352 err_free_pages:
 353         bio_free_pages(&io->write.op.wbio.bio);
 354 err_free:
 355         kfree(io);
 356 err:
 357         bch2_write_ref_put(c, BCH_WRITE_REF_move);
 358         trace_and_count(c, move_extent_alloc_mem_fail, k.k);
 359         return ret;
 360 }
 361
 362 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 363                         struct bch_inode_unpacked *inode)
 364 {
 365         struct btree_iter iter;
 366         struct bkey_s_c k;
 367         int ret;
 368
 369         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 370                              BTREE_ITER_ALL_SNAPSHOTS);
 371         k = bch2_btree_iter_peek(&iter);
 372         ret = bkey_err(k);
 373         if (ret)
 374                 goto err;
 375
 376         if (!k.k || !bkey_eq(k.k->p, pos)) {
 377                 ret = -ENOENT;
 378                 goto err;
 379         }
 380
 381         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 382         if (ret)
 383                 goto err;
 384
 385         ret = bch2_inode_unpack(k, inode);
 386         if (ret)
 387                 goto err;
 388 err:
 389         bch2_trans_iter_exit(trans, &iter);
 390         return ret;
 391 }
 392
 393 static int move_ratelimit(struct btree_trans *trans,
 394                           struct moving_context *ctxt)
 395 {
 396         struct bch_fs *c = trans->c;
 397         u64 delay;
 398
 399         if (ctxt->wait_on_copygc) {
 400                 bch2_trans_unlock(trans);
 401                 wait_event_killable(c->copygc_running_wq,
 402                                     !c->copygc_running ||
 403                                     kthread_should_stop());
 404         }
 405
 406         do {
 407                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 408
 409                 if (delay) {
 410                         bch2_trans_unlock(trans);
 411                         set_current_state(TASK_INTERRUPTIBLE);
 412                 }
 413
 414                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 415                         __set_current_state(TASK_RUNNING);
 416                         return 1;
 417                 }
 418
 419                 if (delay)
 420                         schedule_timeout(delay);
 421
 422                 if (unlikely(freezing(current))) {
 423                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 424                         try_to_freeze();
 425                 }
 426         } while (delay);
 427
 428         /*
 429          * XXX: these limits really ought to be per device, SSDs and hard drives
 430          * will want different limits
 431          */
 432         move_ctxt_wait_event(ctxt, trans,
 433                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 434                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 435                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 436                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 437
 438         return 0;
 439 }
 440
 441 static int move_get_io_opts(struct btree_trans *trans,
 442                             struct bch_io_opts *io_opts,
 443                             struct bkey_s_c k, u64 *cur_inum)
 444 {
 445         struct bch_inode_unpacked inode;
 446         int ret;
 447
 448         if (*cur_inum == k.k->p.inode)
 449                 return 0;
 450
 451         ret = lookup_inode(trans,
 452                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 453                            &inode);
 454         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 455                 return ret;
 456
 457         if (!ret)
 458                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 459         else
 460                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 461         *cur_inum = k.k->p.inode;
 462         return 0;
 463 }
 464
 465 static int __bch2_move_data(struct moving_context *ctxt,
 466                             struct bpos start,
 467                             struct bpos end,
 468                             move_pred_fn pred, void *arg,
 469                             enum btree_id btree_id)
 470 {
 471         struct bch_fs *c = ctxt->c;
 472         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 473         struct bkey_buf sk;
 474         struct btree_trans trans;
 475         struct btree_iter iter;
 476         struct bkey_s_c k;
 477         struct data_update_opts data_opts;
 478         u64 cur_inum = U64_MAX;
 479         int ret = 0, ret2;
 480
 481         bch2_bkey_buf_init(&sk);
 482         bch2_trans_init(&trans, c, 0, 0);
 483
 484         if (ctxt->stats) {
 485                 ctxt->stats->data_type  = BCH_DATA_user;
 486                 ctxt->stats->btree_id   = btree_id;
 487                 ctxt->stats->pos        = start;
 488         }
 489
 490         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 491                              BTREE_ITER_PREFETCH|
 492                              BTREE_ITER_ALL_SNAPSHOTS);
 493
 494         if (ctxt->rate)
 495                 bch2_ratelimit_reset(ctxt->rate);
 496
 497         while (!move_ratelimit(&trans, ctxt)) {
 498                 bch2_trans_begin(&trans);
 499
 500                 k = bch2_btree_iter_peek(&iter);
 501                 if (!k.k)
 502                         break;
 503
 504                 ret = bkey_err(k);
 505                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 506                         continue;
 507                 if (ret)
 508                         break;
 509
 510                 if (bkey_ge(bkey_start_pos(k.k), end))
 511                         break;
 512
 513                 if (ctxt->stats)
 514                         ctxt->stats->pos = iter.pos;
 515
 516                 if (!bkey_extent_is_direct_data(k.k))
 517                         goto next_nondata;
 518
 519                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 520                 if (ret)
 521                         continue;
 522
 523                 memset(&data_opts, 0, sizeof(data_opts));
 524                 if (!pred(c, arg, k, &io_opts, &data_opts))
 525                         goto next;
 526
 527                 /*
 528                  * The iterator gets unlocked by __bch2_read_extent - need to
 529                  * save a copy of @k elsewhere:
 530                  */
 531                 bch2_bkey_buf_reassemble(&sk, c, k);
 532                 k = bkey_i_to_s_c(sk.k);
 533                 bch2_trans_unlock(&trans);
 534
 535                 ret2 = bch2_move_extent(&trans, &iter, ctxt, NULL,
 536                                         io_opts, btree_id, k, data_opts);
 537                 if (ret2) {
 538                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 539                                 continue;
 540
 541                         if (ret2 == -ENOMEM) {
 542                                 /* memory allocation failure, wait for some IO to finish */
 543                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 544                                 continue;
 545                         }
 546
 547                         /* XXX signal failure */
 548                         goto next;
 549                 }
 550
 551                 if (ctxt->rate)
 552                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 553 next:
 554                 if (ctxt->stats)
 555                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 556 next_nondata:
 557                 bch2_btree_iter_advance(&iter);
 558         }
 559
 560         bch2_trans_iter_exit(&trans, &iter);
 561         bch2_trans_exit(&trans);
 562         bch2_bkey_buf_exit(&sk, c);
 563
 564         return ret;
 565 }
 566
 567 int bch2_move_data(struct bch_fs *c,
 568                    enum btree_id start_btree_id, struct bpos start_pos,
 569                    enum btree_id end_btree_id,   struct bpos end_pos,
 570                    struct bch_ratelimit *rate,
 571                    struct bch_move_stats *stats,
 572                    struct write_point_specifier wp,
 573                    bool wait_on_copygc,
 574                    move_pred_fn pred, void *arg)
 575 {
 576         struct moving_context ctxt;
 577         enum btree_id id;
 578         int ret;
 579
 580         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 581
 582         for (id = start_btree_id;
 583              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 584              id++) {
 585                 stats->btree_id = id;
 586
 587                 if (id != BTREE_ID_extents &&
 588                     id != BTREE_ID_reflink)
 589                         continue;
 590
 591                 ret = __bch2_move_data(&ctxt,
 592                                        id == start_btree_id ? start_pos : POS_MIN,
 593                                        id == end_btree_id   ? end_pos   : POS_MAX,
 594                                        pred, arg, id);
 595                 if (ret)
 596                         break;
 597         }
 598
 599         bch2_moving_ctxt_exit(&ctxt);
 600
 601         return ret;
 602 }
 603
 604 void bch2_verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 605 {
 606         struct bch_fs *c = trans->c;
 607         struct btree_iter iter;
 608         struct bkey_s_c k;
 609         struct printbuf buf = PRINTBUF;
 610         struct bch_backpointer bp;
 611         u64 bp_offset = 0;
 612         int ret;
 613
 614         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 615                              bucket, BTREE_ITER_CACHED);
 616 again:
 617         ret = lockrestart_do(trans,
 618                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 619
 620         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 621                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 622
 623                 if (a.v->gen == gen &&
 624                     a.v->dirty_sectors) {
 625                         if (a.v->data_type == BCH_DATA_btree) {
 626                                 bch2_trans_unlock(trans);
 627                                 if (bch2_btree_interior_updates_flush(c))
 628                                         goto again;
 629                                 goto failed_to_evacuate;
 630                         }
 631                 }
 632         }
 633
 634         bch2_trans_iter_exit(trans, &iter);
 635         return;
 636 failed_to_evacuate:
 637         bch2_trans_iter_exit(trans, &iter);
 638
 639         if (test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
 640                 return;
 641
 642         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 643         bch2_bkey_val_to_text(&buf, c, k);
 644
 645         while (1) {
 646                 bch2_trans_begin(trans);
 647
 648                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 649                                                 &bp_offset, &bp,
 650                                                 BTREE_ITER_CACHED);
 651                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 652                         continue;
 653                 if (ret)
 654                         break;
 655                 if (bp_offset == U64_MAX)
 656                         break;
 657
 658                 k = bch2_backpointer_get_key(trans, &iter,
 659                                              bucket, bp_offset, bp);
 660                 ret = bkey_err(k);
 661                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 662                         continue;
 663                 if (ret)
 664                         break;
 665                 if (!k.k)
 666                         continue;
 667                 prt_newline(&buf);
 668                 bch2_bkey_val_to_text(&buf, c, k);
 669                 bch2_trans_iter_exit(trans, &iter);
 670         }
 671
 672         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 673         printbuf_exit(&buf);
 674 }
 675
 676 int __bch2_evacuate_bucket(struct btree_trans *trans,
 677                            struct moving_context *ctxt,
 678                            struct move_bucket_in_flight *bucket_in_flight,
 679                            struct bpos bucket, int gen,
 680                            struct data_update_opts _data_opts)
 681 {
 682         struct bch_fs *c = ctxt->c;
 683         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 684         struct btree_iter iter;
 685         struct bkey_buf sk;
 686         struct bch_backpointer bp;
 687         struct bch_alloc_v4 a_convert;
 688         const struct bch_alloc_v4 *a;
 689         struct bkey_s_c k;
 690         struct data_update_opts data_opts;
 691         unsigned dirty_sectors, bucket_size;
 692         u64 fragmentation;
 693         u64 bp_offset = 0, cur_inum = U64_MAX;
 694         int ret = 0;
 695
 696         bch2_bkey_buf_init(&sk);
 697
 698         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 699                              bucket, BTREE_ITER_CACHED);
 700         ret = lockrestart_do(trans,
 701                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 702         bch2_trans_iter_exit(trans, &iter);
 703
 704         if (ret) {
 705                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 706                 goto err;
 707         }
 708
 709         a = bch2_alloc_to_v4(k, &a_convert);
 710         dirty_sectors = a->dirty_sectors;
 711         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 712         fragmentation = a->fragmentation_lru;
 713
 714         ret = bch2_btree_write_buffer_flush(trans);
 715         if (ret) {
 716                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 717                 goto err;
 718         }
 719
 720         while (!(ret = move_ratelimit(trans, ctxt))) {
 721                 bch2_trans_begin(trans);
 722
 723                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 724                                                 &bp_offset, &bp,
 725                                                 BTREE_ITER_CACHED);
 726                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 727                         continue;
 728                 if (ret)
 729                         goto err;
 730                 if (bp_offset == U64_MAX)
 731                         break;
 732
 733                 if (!bp.level) {
 734                         const struct bch_extent_ptr *ptr;
 735                         struct bkey_s_c k;
 736                         unsigned i = 0;
 737
 738                         k = bch2_backpointer_get_key(trans, &iter,
 739                                                 bucket, bp_offset, bp);
 740                         ret = bkey_err(k);
 741                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 742                                 continue;
 743                         if (ret)
 744                                 goto err;
 745                         if (!k.k)
 746                                 goto next;
 747
 748                         bch2_bkey_buf_reassemble(&sk, c, k);
 749                         k = bkey_i_to_s_c(sk.k);
 750
 751                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 752                         if (ret) {
 753                                 bch2_trans_iter_exit(trans, &iter);
 754                                 continue;
 755                         }
 756
 757                         data_opts = _data_opts;
 758                         data_opts.target        = io_opts.background_target;
 759                         data_opts.rewrite_ptrs = 0;
 760
 761                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 762                                 if (ptr->dev == bucket.inode)
 763                                         data_opts.rewrite_ptrs |= 1U << i;
 764                                 i++;
 765                         }
 766
 767                         ret = bch2_move_extent(trans, &iter, ctxt,
 768                                         bucket_in_flight,
 769                                         io_opts, bp.btree_id, k, data_opts);
 770                         bch2_trans_iter_exit(trans, &iter);
 771
 772                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 773                                 continue;
 774                         if (ret == -ENOMEM) {
 775                                 /* memory allocation failure, wait for some IO to finish */
 776                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 777                                 continue;
 778                         }
 779                         if (ret)
 780                                 goto err;
 781
 782                         if (ctxt->rate)
 783                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 784                         if (ctxt->stats)
 785                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 786                 } else {
 787                         struct btree *b;
 788
 789                         b = bch2_backpointer_get_node(trans, &iter,
 790                                                 bucket, bp_offset, bp);
 791                         ret = PTR_ERR_OR_ZERO(b);
 792                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 793                                 continue;
 794                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 795                                 continue;
 796                         if (ret)
 797                                 goto err;
 798                         if (!b)
 799                                 goto next;
 800
 801                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 802                         bch2_trans_iter_exit(trans, &iter);
 803
 804                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 805                                 continue;
 806                         if (ret)
 807                                 goto err;
 808
 809                         if (ctxt->rate)
 810                                 bch2_ratelimit_increment(ctxt->rate,
 811                                                          c->opts.btree_node_size >> 9);
 812                         if (ctxt->stats) {
 813                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 814                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 815                         }
 816                 }
 817 next:
 818                 bp_offset++;
 819         }
 820
 821         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 822
 823         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
 824                 bch2_trans_unlock(trans);
 825                 move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 826                 closure_sync(&ctxt->cl);
 827                 if (!ctxt->write_error)
 828                         bch2_verify_bucket_evacuated(trans, bucket, gen);
 829         }
 830 err:
 831         bch2_bkey_buf_exit(&sk, c);
 832         return ret;
 833 }
 834
 835 int bch2_evacuate_bucket(struct bch_fs *c,
 836                          struct bpos bucket, int gen,
 837                          struct data_update_opts data_opts,
 838                          struct bch_ratelimit *rate,
 839                          struct bch_move_stats *stats,
 840                          struct write_point_specifier wp,
 841                          bool wait_on_copygc)
 842 {
 843         struct btree_trans trans;
 844         struct moving_context ctxt;
 845         int ret;
 846
 847         bch2_trans_init(&trans, c, 0, 0);
 848         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 849         ret = __bch2_evacuate_bucket(&trans, &ctxt, NULL, bucket, gen, data_opts);
 850         bch2_moving_ctxt_exit(&ctxt);
 851         bch2_trans_exit(&trans);
 852
 853         return ret;
 854 }
 855
 856 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 857                                 struct btree *, struct bch_io_opts *,
 858                                 struct data_update_opts *);
 859
 860 static int bch2_move_btree(struct bch_fs *c,
 861                            enum btree_id start_btree_id, struct bpos start_pos,
 862                            enum btree_id end_btree_id,   struct bpos end_pos,
 863                            move_btree_pred pred, void *arg,
 864                            struct bch_move_stats *stats)
 865 {
 866         bool kthread = (current->flags & PF_KTHREAD) != 0;
 867         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 868         struct btree_trans trans;
 869         struct btree_iter iter;
 870         struct btree *b;
 871         enum btree_id id;
 872         struct data_update_opts data_opts;
 873         int ret = 0;
 874
 875         bch2_trans_init(&trans, c, 0, 0);
 876         progress_list_add(c, stats);
 877
 878         stats->data_type = BCH_DATA_btree;
 879
 880         for (id = start_btree_id;
 881              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 882              id++) {
 883                 stats->btree_id = id;
 884
 885                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 886                                           BTREE_ITER_PREFETCH);
 887 retry:
 888                 ret = 0;
 889                 while (bch2_trans_begin(&trans),
 890                        (b = bch2_btree_iter_peek_node(&iter)) &&
 891                        !(ret = PTR_ERR_OR_ZERO(b))) {
 892                         if (kthread && kthread_should_stop())
 893                                 break;
 894
 895                         if ((cmp_int(id, end_btree_id) ?:
 896                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 897                                 break;
 898
 899                         stats->pos = iter.pos;
 900
 901                         if (!pred(c, arg, b, &io_opts, &data_opts))
 902                                 goto next;
 903
 904                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 905                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 906                                 continue;
 907                         if (ret)
 908                                 break;
 909 next:
 910                         bch2_btree_iter_next_node(&iter);
 911                 }
 912                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 913                         goto retry;
 914
 915                 bch2_trans_iter_exit(&trans, &iter);
 916
 917                 if (kthread && kthread_should_stop())
 918                         break;
 919         }
 920
 921         bch2_trans_exit(&trans);
 922
 923         if (ret)
 924                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 925
 926         bch2_btree_interior_updates_flush(c);
 927
 928         progress_list_del(c, stats);
 929         return ret;
 930 }
 931
 932 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 933                              struct bkey_s_c k,
 934                              struct bch_io_opts *io_opts,
 935                              struct data_update_opts *data_opts)
 936 {
 937         unsigned nr_good = bch2_bkey_durability(c, k);
 938         unsigned replicas = bkey_is_btree_ptr(k.k)
 939                 ? c->opts.metadata_replicas
 940                 : io_opts->data_replicas;
 941
 942         if (!nr_good || nr_good >= replicas)
 943                 return false;
 944
 945         data_opts->target               = 0;
 946         data_opts->extra_replicas       = replicas - nr_good;
 947         data_opts->btree_insert_flags   = 0;
 948         return true;
 949 }
 950
 951 static bool migrate_pred(struct bch_fs *c, void *arg,
 952                          struct bkey_s_c k,
 953                          struct bch_io_opts *io_opts,
 954                          struct data_update_opts *data_opts)
 955 {
 956         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 957         const struct bch_extent_ptr *ptr;
 958         struct bch_ioctl_data *op = arg;
 959         unsigned i = 0;
 960
 961         data_opts->rewrite_ptrs         = 0;
 962         data_opts->target               = 0;
 963         data_opts->extra_replicas       = 0;
 964         data_opts->btree_insert_flags   = 0;
 965
 966         bkey_for_each_ptr(ptrs, ptr) {
 967                 if (ptr->dev == op->migrate.dev)
 968                         data_opts->rewrite_ptrs |= 1U << i;
 969                 i++;
 970         }
 971
 972         return data_opts->rewrite_ptrs != 0;
 973 }
 974
 975 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 976                                    struct btree *b,
 977                                    struct bch_io_opts *io_opts,
 978                                    struct data_update_opts *data_opts)
 979 {
 980         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 981 }
 982
 983 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 984                                struct btree *b,
 985                                struct bch_io_opts *io_opts,
 986                                struct data_update_opts *data_opts)
 987 {
 988         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 989 }
 990
 991 static bool bformat_needs_redo(struct bkey_format *f)
 992 {
 993         unsigned i;
 994
 995         for (i = 0; i < f->nr_fields; i++) {
 996                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 997                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 998                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 999
1000                 if (f->bits_per_field[i] > unpacked_bits)
1001                         return true;
1002
1003                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1004                         return true;
1005
1006                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1007                      unpacked_mask) <
1008                     field_offset)
1009                         return true;
1010         }
1011
1012         return false;
1013 }
1014
1015 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1016                                    struct btree *b,
1017                                    struct bch_io_opts *io_opts,
1018                                    struct data_update_opts *data_opts)
1019 {
1020         if (b->version_ondisk != c->sb.version ||
1021             btree_node_need_rewrite(b) ||
1022             bformat_needs_redo(&b->format)) {
1023                 data_opts->target               = 0;
1024                 data_opts->extra_replicas       = 0;
1025                 data_opts->btree_insert_flags   = 0;
1026                 return true;
1027         }
1028
1029         return false;
1030 }
1031
1032 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1033 {
1034         int ret;
1035
1036         ret = bch2_move_btree(c,
1037                               0,                POS_MIN,
1038                               BTREE_ID_NR,      SPOS_MAX,
1039                               rewrite_old_nodes_pred, c, stats);
1040         if (!ret) {
1041                 mutex_lock(&c->sb_lock);
1042                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1043                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1044                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1045                 bch2_write_super(c);
1046                 mutex_unlock(&c->sb_lock);
1047         }
1048
1049         return ret;
1050 }
1051
1052 int bch2_data_job(struct bch_fs *c,
1053                   struct bch_move_stats *stats,
1054                   struct bch_ioctl_data op)
1055 {
1056         int ret = 0;
1057
1058         switch (op.op) {
1059         case BCH_DATA_OP_REREPLICATE:
1060                 bch2_move_stats_init(stats, "rereplicate");
1061                 stats->data_type = BCH_DATA_journal;
1062                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1063
1064                 ret = bch2_move_btree(c,
1065                                       op.start_btree,   op.start_pos,
1066                                       op.end_btree,     op.end_pos,
1067                                       rereplicate_btree_pred, c, stats) ?: ret;
1068                 ret = bch2_replicas_gc2(c) ?: ret;
1069
1070                 ret = bch2_move_data(c,
1071                                      op.start_btree,    op.start_pos,
1072                                      op.end_btree,      op.end_pos,
1073                                      NULL,
1074                                      stats,
1075                                      writepoint_hashed((unsigned long) current),
1076                                      true,
1077                                      rereplicate_pred, c) ?: ret;
1078                 ret = bch2_replicas_gc2(c) ?: ret;
1079                 break;
1080         case BCH_DATA_OP_MIGRATE:
1081                 if (op.migrate.dev >= c->sb.nr_devices)
1082                         return -EINVAL;
1083
1084                 bch2_move_stats_init(stats, "migrate");
1085                 stats->data_type = BCH_DATA_journal;
1086                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1087
1088                 ret = bch2_move_btree(c,
1089                                       op.start_btree,   op.start_pos,
1090                                       op.end_btree,     op.end_pos,
1091                                       migrate_btree_pred, &op, stats) ?: ret;
1092                 ret = bch2_replicas_gc2(c) ?: ret;
1093
1094                 ret = bch2_move_data(c,
1095                                      op.start_btree,    op.start_pos,
1096                                      op.end_btree,      op.end_pos,
1097                                      NULL,
1098                                      stats,
1099                                      writepoint_hashed((unsigned long) current),
1100                                      true,
1101                                      migrate_pred, &op) ?: ret;
1102                 ret = bch2_replicas_gc2(c) ?: ret;
1103                 break;
1104         case BCH_DATA_OP_REWRITE_OLD_NODES:
1105                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1106                 ret = bch2_scan_old_btree_nodes(c, stats);
1107                 break;
1108         default:
1109                 ret = -EINVAL;
1110         }
1111
1112         return ret;
1113 }