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_read.h"
  18 #include "io_write.h"
  19 #include "journal_reclaim.h"
  20 #include "keylist.h"
  21 #include "move.h"
  22 #include "replicas.h"
  23 #include "snapshot.h"
  24 #include "super-io.h"
  25 #include "trace.h"
  26
  27 #include <linux/ioprio.h>
  28 #include <linux/kthread.h>
  29
  30 const char * const bch2_data_ops_strs[] = {
  31 #define x(t, n, ...) [n] = #t,
  32         BCH_DATA_OPS()
  33 #undef x
  34         NULL
  35 };
  36
  37 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
  38 {
  39         if (trace_move_extent_enabled()) {
  40                 struct printbuf buf = PRINTBUF;
  41
  42                 bch2_bkey_val_to_text(&buf, c, k);
  43                 trace_move_extent(c, buf.buf);
  44                 printbuf_exit(&buf);
  45         }
  46 }
  47
  48 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
  49 {
  50         if (trace_move_extent_read_enabled()) {
  51                 struct printbuf buf = PRINTBUF;
  52
  53                 bch2_bkey_val_to_text(&buf, c, k);
  54                 trace_move_extent_read(c, buf.buf);
  55                 printbuf_exit(&buf);
  56         }
  57 }
  58
  59 static void trace_move_extent_alloc_mem_fail2(struct bch_fs *c, struct bkey_s_c k)
  60 {
  61         if (trace_move_extent_alloc_mem_fail_enabled()) {
  62                 struct printbuf buf = PRINTBUF;
  63
  64                 bch2_bkey_val_to_text(&buf, c, k);
  65                 trace_move_extent_alloc_mem_fail(c, buf.buf);
  66                 printbuf_exit(&buf);
  67         }
  68 }
  69
  70 struct moving_io {
  71         struct list_head                read_list;
  72         struct list_head                io_list;
  73         struct move_bucket_in_flight    *b;
  74         struct closure                  cl;
  75         bool                            read_completed;
  76
  77         unsigned                        read_sectors;
  78         unsigned                        write_sectors;
  79
  80         struct bch_read_bio             rbio;
  81
  82         struct data_update              write;
  83         /* Must be last since it is variable size */
  84         struct bio_vec                  bi_inline_vecs[0];
  85 };
  86
  87 static void move_free(struct moving_io *io)
  88 {
  89         struct moving_context *ctxt = io->write.ctxt;
  90
  91         if (io->b)
  92                 atomic_dec(&io->b->count);
  93
  94         bch2_data_update_exit(&io->write);
  95
  96         mutex_lock(&ctxt->lock);
  97         list_del(&io->io_list);
  98         wake_up(&ctxt->wait);
  99         mutex_unlock(&ctxt->lock);
 100
 101         kfree(io);
 102 }
 103
 104 static void move_write_done(struct bch_write_op *op)
 105 {
 106         struct moving_io *io = container_of(op, struct moving_io, write.op);
 107         struct moving_context *ctxt = io->write.ctxt;
 108
 109         if (io->write.op.error)
 110                 ctxt->write_error = true;
 111
 112         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 113         atomic_dec(&io->write.ctxt->write_ios);
 114         move_free(io);
 115         closure_put(&ctxt->cl);
 116 }
 117
 118 static void move_write(struct moving_io *io)
 119 {
 120         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 121                 move_free(io);
 122                 return;
 123         }
 124
 125         closure_get(&io->write.ctxt->cl);
 126         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 127         atomic_inc(&io->write.ctxt->write_ios);
 128
 129         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 130 }
 131
 132 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 133 {
 134         struct moving_io *io =
 135                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 136
 137         return io && io->read_completed ? io : NULL;
 138 }
 139
 140 static void move_read_endio(struct bio *bio)
 141 {
 142         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 143         struct moving_context *ctxt = io->write.ctxt;
 144
 145         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 146         atomic_dec(&ctxt->read_ios);
 147         io->read_completed = true;
 148
 149         wake_up(&ctxt->wait);
 150         closure_put(&ctxt->cl);
 151 }
 152
 153 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
 154 {
 155         struct moving_io *io;
 156
 157         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 158                 bch2_trans_unlock_long(ctxt->trans);
 159                 list_del(&io->read_list);
 160                 move_write(io);
 161         }
 162 }
 163
 164 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 165 {
 166         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 167
 168         move_ctxt_wait_event(ctxt,
 169                 !atomic_read(&ctxt->write_sectors) ||
 170                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 171 }
 172
 173 static void bch2_moving_ctxt_flush_all(struct moving_context *ctxt)
 174 {
 175         move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 176         bch2_trans_unlock_long(ctxt->trans);
 177         closure_sync(&ctxt->cl);
 178 }
 179
 180 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 181 {
 182         struct bch_fs *c = ctxt->trans->c;
 183
 184         bch2_moving_ctxt_flush_all(ctxt);
 185
 186         EBUG_ON(atomic_read(&ctxt->write_sectors));
 187         EBUG_ON(atomic_read(&ctxt->write_ios));
 188         EBUG_ON(atomic_read(&ctxt->read_sectors));
 189         EBUG_ON(atomic_read(&ctxt->read_ios));
 190
 191         mutex_lock(&c->moving_context_lock);
 192         list_del(&ctxt->list);
 193         mutex_unlock(&c->moving_context_lock);
 194
 195         bch2_trans_put(ctxt->trans);
 196         memset(ctxt, 0, sizeof(*ctxt));
 197 }
 198
 199 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 200                            struct bch_fs *c,
 201                            struct bch_ratelimit *rate,
 202                            struct bch_move_stats *stats,
 203                            struct write_point_specifier wp,
 204                            bool wait_on_copygc)
 205 {
 206         memset(ctxt, 0, sizeof(*ctxt));
 207
 208         ctxt->trans     = bch2_trans_get(c);
 209         ctxt->fn        = (void *) _RET_IP_;
 210         ctxt->rate      = rate;
 211         ctxt->stats     = stats;
 212         ctxt->wp        = wp;
 213         ctxt->wait_on_copygc = wait_on_copygc;
 214
 215         closure_init_stack(&ctxt->cl);
 216
 217         mutex_init(&ctxt->lock);
 218         INIT_LIST_HEAD(&ctxt->reads);
 219         INIT_LIST_HEAD(&ctxt->ios);
 220         init_waitqueue_head(&ctxt->wait);
 221
 222         mutex_lock(&c->moving_context_lock);
 223         list_add(&ctxt->list, &c->moving_context_list);
 224         mutex_unlock(&c->moving_context_lock);
 225 }
 226
 227 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
 228 {
 229         trace_move_data(c, stats);
 230 }
 231
 232 void bch2_move_stats_init(struct bch_move_stats *stats, const char *name)
 233 {
 234         memset(stats, 0, sizeof(*stats));
 235         stats->data_type = BCH_DATA_user;
 236         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 237 }
 238
 239 int bch2_move_extent(struct moving_context *ctxt,
 240                      struct move_bucket_in_flight *bucket_in_flight,
 241                      struct btree_iter *iter,
 242                      struct bkey_s_c k,
 243                      struct bch_io_opts io_opts,
 244                      struct data_update_opts data_opts)
 245 {
 246         struct btree_trans *trans = ctxt->trans;
 247         struct bch_fs *c = trans->c;
 248         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 249         struct moving_io *io;
 250         const union bch_extent_entry *entry;
 251         struct extent_ptr_decoded p;
 252         unsigned sectors = k.k->size, pages;
 253         int ret = -ENOMEM;
 254
 255         if (ctxt->stats)
 256                 ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
 257         trace_move_extent2(c, k);
 258
 259         bch2_data_update_opts_normalize(k, &data_opts);
 260
 261         if (!data_opts.rewrite_ptrs &&
 262             !data_opts.extra_replicas) {
 263                 if (data_opts.kill_ptrs)
 264                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 265                 return 0;
 266         }
 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         INIT_LIST_HEAD(&io->io_list);
 285         io->write.ctxt          = ctxt;
 286         io->read_sectors        = k.k->size;
 287         io->write_sectors       = k.k->size;
 288
 289         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 290         bio_set_prio(&io->write.op.wbio.bio,
 291                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 292
 293         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 294                                  GFP_KERNEL))
 295                 goto err_free;
 296
 297         io->rbio.c              = c;
 298         io->rbio.opts           = io_opts;
 299         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 300         io->rbio.bio.bi_vcnt = pages;
 301         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 302         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 303
 304         io->rbio.bio.bi_opf             = REQ_OP_READ;
 305         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 306         io->rbio.bio.bi_end_io          = move_read_endio;
 307
 308         ret = bch2_data_update_init(trans, iter, ctxt, &io->write, ctxt->wp,
 309                                     io_opts, data_opts, iter->btree_id, k);
 310         if (ret)
 311                 goto err_free_pages;
 312
 313         io->write.op.end_io = move_write_done;
 314
 315         if (ctxt->rate)
 316                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 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         if (bucket_in_flight) {
 324                 io->b = bucket_in_flight;
 325                 atomic_inc(&io->b->count);
 326         }
 327
 328         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 329         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 330         trace_move_extent_read2(c, k);
 331
 332         mutex_lock(&ctxt->lock);
 333         atomic_add(io->read_sectors, &ctxt->read_sectors);
 334         atomic_inc(&ctxt->read_ios);
 335
 336         list_add_tail(&io->read_list, &ctxt->reads);
 337         list_add_tail(&io->io_list, &ctxt->ios);
 338         mutex_unlock(&ctxt->lock);
 339
 340         /*
 341          * dropped by move_read_endio() - guards against use after free of
 342          * ctxt when doing wakeup
 343          */
 344         closure_get(&ctxt->cl);
 345         bch2_read_extent(trans, &io->rbio,
 346                          bkey_start_pos(k.k),
 347                          iter->btree_id, k, 0,
 348                          BCH_READ_NODECODE|
 349                          BCH_READ_LAST_FRAGMENT);
 350         return 0;
 351 err_free_pages:
 352         bio_free_pages(&io->write.op.wbio.bio);
 353 err_free:
 354         kfree(io);
 355 err:
 356         if (ret == -BCH_ERR_data_update_done)
 357                 return 0;
 358
 359         this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
 360         trace_move_extent_alloc_mem_fail2(c, k);
 361         return ret;
 362 }
 363
 364 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 365                           struct per_snapshot_io_opts *io_opts,
 366                           struct bkey_s_c extent_k)
 367 {
 368         struct bch_fs *c = trans->c;
 369         u32 restart_count = trans->restart_count;
 370         int ret = 0;
 371
 372         if (io_opts->cur_inum != extent_k.k->p.inode) {
 373                 struct btree_iter iter;
 374                 struct bkey_s_c k;
 375
 376                 io_opts->d.nr = 0;
 377
 378                 for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
 379                                    BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
 380                         if (k.k->p.offset != extent_k.k->p.inode)
 381                                 break;
 382
 383                         if (!bkey_is_inode(k.k))
 384                                 continue;
 385
 386                         struct bch_inode_unpacked inode;
 387                         BUG_ON(bch2_inode_unpack(k, &inode));
 388
 389                         struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 390                         bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
 391
 392                         ret = darray_push(&io_opts->d, e);
 393                         if (ret)
 394                                 break;
 395                 }
 396                 bch2_trans_iter_exit(trans, &iter);
 397                 io_opts->cur_inum = extent_k.k->p.inode;
 398         }
 399
 400         ret = ret ?: trans_was_restarted(trans, restart_count);
 401         if (ret)
 402                 return ERR_PTR(ret);
 403
 404         if (extent_k.k->p.snapshot) {
 405                 struct snapshot_io_opts_entry *i;
 406                 darray_for_each(io_opts->d, i)
 407                         if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
 408                                 return &i->io_opts;
 409         }
 410
 411         return &io_opts->fs_io_opts;
 412 }
 413
 414 int bch2_move_get_io_opts_one(struct btree_trans *trans,
 415                               struct bch_io_opts *io_opts,
 416                               struct bkey_s_c extent_k)
 417 {
 418         struct btree_iter iter;
 419         struct bkey_s_c k;
 420         int ret;
 421
 422         /* reflink btree? */
 423         if (!extent_k.k->p.inode) {
 424                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 425                 return 0;
 426         }
 427
 428         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 429                                SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
 430                                BTREE_ITER_CACHED);
 431         ret = bkey_err(k);
 432         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 433                 return ret;
 434
 435         if (!ret && bkey_is_inode(k.k)) {
 436                 struct bch_inode_unpacked inode;
 437                 bch2_inode_unpack(k, &inode);
 438                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 439         } else {
 440                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 441         }
 442
 443         bch2_trans_iter_exit(trans, &iter);
 444         return 0;
 445 }
 446
 447 int bch2_move_ratelimit(struct moving_context *ctxt)
 448 {
 449         struct bch_fs *c = ctxt->trans->c;
 450         u64 delay;
 451
 452         if (ctxt->wait_on_copygc && c->copygc_running) {
 453                 bch2_moving_ctxt_flush_all(ctxt);
 454                 wait_event_killable(c->copygc_running_wq,
 455                                     !c->copygc_running ||
 456                                     kthread_should_stop());
 457         }
 458
 459         do {
 460                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 461
 462                 if (kthread_should_stop())
 463                         return 1;
 464
 465                 if (delay)
 466                         move_ctxt_wait_event_timeout(ctxt,
 467                                         freezing(current) || kthread_should_stop(),
 468                                         delay);
 469
 470                 if (unlikely(freezing(current))) {
 471                         bch2_moving_ctxt_flush_all(ctxt);
 472                         try_to_freeze();
 473                 }
 474         } while (delay);
 475
 476         /*
 477          * XXX: these limits really ought to be per device, SSDs and hard drives
 478          * will want different limits
 479          */
 480         move_ctxt_wait_event(ctxt,
 481                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 482                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 483                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 484                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 485
 486         return 0;
 487 }
 488
 489 static int bch2_move_data_btree(struct moving_context *ctxt,
 490                                 struct bpos start,
 491                                 struct bpos end,
 492                                 move_pred_fn pred, void *arg,
 493                                 enum btree_id btree_id)
 494 {
 495         struct btree_trans *trans = ctxt->trans;
 496         struct bch_fs *c = trans->c;
 497         struct per_snapshot_io_opts snapshot_io_opts;
 498         struct bch_io_opts *io_opts;
 499         struct bkey_buf sk;
 500         struct btree_iter iter;
 501         struct bkey_s_c k;
 502         struct data_update_opts data_opts;
 503         int ret = 0, ret2;
 504
 505         per_snapshot_io_opts_init(&snapshot_io_opts, c);
 506         bch2_bkey_buf_init(&sk);
 507
 508         if (ctxt->stats) {
 509                 ctxt->stats->data_type  = BCH_DATA_user;
 510                 ctxt->stats->pos        = BBPOS(btree_id, start);
 511         }
 512
 513         bch2_trans_iter_init(trans, &iter, btree_id, start,
 514                              BTREE_ITER_PREFETCH|
 515                              BTREE_ITER_ALL_SNAPSHOTS);
 516
 517         if (ctxt->rate)
 518                 bch2_ratelimit_reset(ctxt->rate);
 519
 520         while (!bch2_move_ratelimit(ctxt)) {
 521                 bch2_trans_begin(trans);
 522
 523                 k = bch2_btree_iter_peek(&iter);
 524                 if (!k.k)
 525                         break;
 526
 527                 ret = bkey_err(k);
 528                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 529                         continue;
 530                 if (ret)
 531                         break;
 532
 533                 if (bkey_ge(bkey_start_pos(k.k), end))
 534                         break;
 535
 536                 if (ctxt->stats)
 537                         ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
 538
 539                 if (!bkey_extent_is_direct_data(k.k))
 540                         goto next_nondata;
 541
 542                 io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
 543                 ret = PTR_ERR_OR_ZERO(io_opts);
 544                 if (ret)
 545                         continue;
 546
 547                 memset(&data_opts, 0, sizeof(data_opts));
 548                 if (!pred(c, arg, k, io_opts, &data_opts))
 549                         goto next;
 550
 551                 /*
 552                  * The iterator gets unlocked by __bch2_read_extent - need to
 553                  * save a copy of @k elsewhere:
 554                  */
 555                 bch2_bkey_buf_reassemble(&sk, c, k);
 556                 k = bkey_i_to_s_c(sk.k);
 557
 558                 ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
 559                 if (ret2) {
 560                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 561                                 continue;
 562
 563                         if (ret2 == -ENOMEM) {
 564                                 /* memory allocation failure, wait for some IO to finish */
 565                                 bch2_move_ctxt_wait_for_io(ctxt);
 566                                 continue;
 567                         }
 568
 569                         /* XXX signal failure */
 570                         goto next;
 571                 }
 572 next:
 573                 if (ctxt->stats)
 574                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 575 next_nondata:
 576                 bch2_btree_iter_advance(&iter);
 577         }
 578
 579         bch2_trans_iter_exit(trans, &iter);
 580         bch2_bkey_buf_exit(&sk, c);
 581         per_snapshot_io_opts_exit(&snapshot_io_opts);
 582
 583         return ret;
 584 }
 585
 586 int __bch2_move_data(struct moving_context *ctxt,
 587                      struct bbpos start,
 588                      struct bbpos end,
 589                      move_pred_fn pred, void *arg)
 590 {
 591         struct bch_fs *c = ctxt->trans->c;
 592         enum btree_id id;
 593         int ret = 0;
 594
 595         for (id = start.btree;
 596              id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 597              id++) {
 598                 ctxt->stats->pos = BBPOS(id, POS_MIN);
 599
 600                 if (!btree_type_has_ptrs(id) ||
 601                     !bch2_btree_id_root(c, id)->b)
 602                         continue;
 603
 604                 ret = bch2_move_data_btree(ctxt,
 605                                        id == start.btree ? start.pos : POS_MIN,
 606                                        id == end.btree   ? end.pos   : POS_MAX,
 607                                        pred, arg, id);
 608                 if (ret)
 609                         break;
 610         }
 611
 612         return ret;
 613 }
 614
 615 int bch2_move_data(struct bch_fs *c,
 616                    struct bbpos start,
 617                    struct bbpos end,
 618                    struct bch_ratelimit *rate,
 619                    struct bch_move_stats *stats,
 620                    struct write_point_specifier wp,
 621                    bool wait_on_copygc,
 622                    move_pred_fn pred, void *arg)
 623 {
 624
 625         struct moving_context ctxt;
 626         int ret;
 627
 628         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 629         ret = __bch2_move_data(&ctxt, start, end, pred, arg);
 630         bch2_moving_ctxt_exit(&ctxt);
 631
 632         return ret;
 633 }
 634
 635 int __bch2_evacuate_bucket(struct moving_context *ctxt,
 636                            struct move_bucket_in_flight *bucket_in_flight,
 637                            struct bpos bucket, int gen,
 638                            struct data_update_opts _data_opts)
 639 {
 640         struct btree_trans *trans = ctxt->trans;
 641         struct bch_fs *c = trans->c;
 642         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 643         struct btree_iter iter;
 644         struct bkey_buf sk;
 645         struct bch_backpointer bp;
 646         struct bch_alloc_v4 a_convert;
 647         const struct bch_alloc_v4 *a;
 648         struct bkey_s_c k;
 649         struct data_update_opts data_opts;
 650         unsigned dirty_sectors, bucket_size;
 651         u64 fragmentation;
 652         struct bpos bp_pos = POS_MIN;
 653         int ret = 0;
 654
 655         trace_bucket_evacuate(c, &bucket);
 656
 657         bch2_bkey_buf_init(&sk);
 658
 659         /*
 660          * We're not run in a context that handles transaction restarts:
 661          */
 662         bch2_trans_begin(trans);
 663
 664         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 665                              bucket, BTREE_ITER_CACHED);
 666         ret = lockrestart_do(trans,
 667                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 668         bch2_trans_iter_exit(trans, &iter);
 669
 670         if (ret) {
 671                 bch_err_msg(c, ret, "looking up alloc key");
 672                 goto err;
 673         }
 674
 675         a = bch2_alloc_to_v4(k, &a_convert);
 676         dirty_sectors = bch2_bucket_sectors_dirty(*a);
 677         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 678         fragmentation = a->fragmentation_lru;
 679
 680         ret = bch2_btree_write_buffer_tryflush(trans);
 681         bch_err_msg(c, ret, "flushing btree write buffer");
 682         if (ret)
 683                 goto err;
 684
 685         while (!(ret = bch2_move_ratelimit(ctxt))) {
 686                 bch2_trans_begin(trans);
 687
 688                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 689                                                 &bp_pos, &bp,
 690                                                 BTREE_ITER_CACHED);
 691                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 692                         continue;
 693                 if (ret)
 694                         goto err;
 695                 if (bkey_eq(bp_pos, POS_MAX))
 696                         break;
 697
 698                 if (!bp.level) {
 699                         const struct bch_extent_ptr *ptr;
 700                         unsigned i = 0;
 701
 702                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 703                         ret = bkey_err(k);
 704                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 705                                 continue;
 706                         if (ret)
 707                                 goto err;
 708                         if (!k.k)
 709                                 goto next;
 710
 711                         bch2_bkey_buf_reassemble(&sk, c, k);
 712                         k = bkey_i_to_s_c(sk.k);
 713
 714                         ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
 715                         if (ret) {
 716                                 bch2_trans_iter_exit(trans, &iter);
 717                                 continue;
 718                         }
 719
 720                         data_opts = _data_opts;
 721                         data_opts.target        = io_opts.background_target;
 722                         data_opts.rewrite_ptrs = 0;
 723
 724                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 725                                 if (ptr->dev == bucket.inode) {
 726                                         data_opts.rewrite_ptrs |= 1U << i;
 727                                         if (ptr->cached) {
 728                                                 bch2_trans_iter_exit(trans, &iter);
 729                                                 goto next;
 730                                         }
 731                                 }
 732                                 i++;
 733                         }
 734
 735                         ret = bch2_move_extent(ctxt, bucket_in_flight,
 736                                                &iter, k, io_opts, data_opts);
 737                         bch2_trans_iter_exit(trans, &iter);
 738
 739                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 740                                 continue;
 741                         if (ret == -ENOMEM) {
 742                                 /* memory allocation failure, wait for some IO to finish */
 743                                 bch2_move_ctxt_wait_for_io(ctxt);
 744                                 continue;
 745                         }
 746                         if (ret)
 747                                 goto err;
 748
 749                         if (ctxt->stats)
 750                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 751                 } else {
 752                         struct btree *b;
 753
 754                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 755                         ret = PTR_ERR_OR_ZERO(b);
 756                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 757                                 continue;
 758                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 759                                 continue;
 760                         if (ret)
 761                                 goto err;
 762                         if (!b)
 763                                 goto next;
 764
 765                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 766                         bch2_trans_iter_exit(trans, &iter);
 767
 768                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 769                                 continue;
 770                         if (ret)
 771                                 goto err;
 772
 773                         if (ctxt->rate)
 774                                 bch2_ratelimit_increment(ctxt->rate,
 775                                                          c->opts.btree_node_size >> 9);
 776                         if (ctxt->stats) {
 777                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 778                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 779                         }
 780                 }
 781 next:
 782                 bp_pos = bpos_nosnap_successor(bp_pos);
 783         }
 784
 785         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 786 err:
 787         bch2_bkey_buf_exit(&sk, c);
 788         return ret;
 789 }
 790
 791 int bch2_evacuate_bucket(struct bch_fs *c,
 792                          struct bpos bucket, int gen,
 793                          struct data_update_opts data_opts,
 794                          struct bch_ratelimit *rate,
 795                          struct bch_move_stats *stats,
 796                          struct write_point_specifier wp,
 797                          bool wait_on_copygc)
 798 {
 799         struct moving_context ctxt;
 800         int ret;
 801
 802         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 803         ret = __bch2_evacuate_bucket(&ctxt, NULL, bucket, gen, data_opts);
 804         bch2_moving_ctxt_exit(&ctxt);
 805
 806         return ret;
 807 }
 808
 809 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 810                                 struct btree *, struct bch_io_opts *,
 811                                 struct data_update_opts *);
 812
 813 static int bch2_move_btree(struct bch_fs *c,
 814                            struct bbpos start,
 815                            struct bbpos end,
 816                            move_btree_pred pred, void *arg,
 817                            struct bch_move_stats *stats)
 818 {
 819         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 820         struct moving_context ctxt;
 821         struct btree_trans *trans;
 822         struct btree_iter iter;
 823         struct btree *b;
 824         enum btree_id btree;
 825         struct data_update_opts data_opts;
 826         int ret = 0;
 827
 828         bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
 829                               writepoint_ptr(&c->btree_write_point),
 830                               true);
 831         trans = ctxt.trans;
 832
 833         stats->data_type = BCH_DATA_btree;
 834
 835         for (btree = start.btree;
 836              btree <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 837              btree ++) {
 838                 stats->pos = BBPOS(btree, POS_MIN);
 839
 840                 if (!bch2_btree_id_root(c, btree)->b)
 841                         continue;
 842
 843                 bch2_trans_node_iter_init(trans, &iter, btree, POS_MIN, 0, 0,
 844                                           BTREE_ITER_PREFETCH);
 845 retry:
 846                 ret = 0;
 847                 while (bch2_trans_begin(trans),
 848                        (b = bch2_btree_iter_peek_node(&iter)) &&
 849                        !(ret = PTR_ERR_OR_ZERO(b))) {
 850                         if (kthread_should_stop())
 851                                 break;
 852
 853                         if ((cmp_int(btree, end.btree) ?:
 854                              bpos_cmp(b->key.k.p, end.pos)) > 0)
 855                                 break;
 856
 857                         stats->pos = BBPOS(iter.btree_id, iter.pos);
 858
 859                         if (!pred(c, arg, b, &io_opts, &data_opts))
 860                                 goto next;
 861
 862                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
 863                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 864                                 continue;
 865                         if (ret)
 866                                 break;
 867 next:
 868                         bch2_btree_iter_next_node(&iter);
 869                 }
 870                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 871                         goto retry;
 872
 873                 bch2_trans_iter_exit(trans, &iter);
 874
 875                 if (kthread_should_stop())
 876                         break;
 877         }
 878
 879         bch_err_fn(c, ret);
 880         bch2_moving_ctxt_exit(&ctxt);
 881         bch2_btree_interior_updates_flush(c);
 882
 883         return ret;
 884 }
 885
 886 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 887                              struct bkey_s_c k,
 888                              struct bch_io_opts *io_opts,
 889                              struct data_update_opts *data_opts)
 890 {
 891         unsigned nr_good = bch2_bkey_durability(c, k);
 892         unsigned replicas = bkey_is_btree_ptr(k.k)
 893                 ? c->opts.metadata_replicas
 894                 : io_opts->data_replicas;
 895
 896         if (!nr_good || nr_good >= replicas)
 897                 return false;
 898
 899         data_opts->target               = 0;
 900         data_opts->extra_replicas       = replicas - nr_good;
 901         data_opts->btree_insert_flags   = 0;
 902         return true;
 903 }
 904
 905 static bool migrate_pred(struct bch_fs *c, void *arg,
 906                          struct bkey_s_c k,
 907                          struct bch_io_opts *io_opts,
 908                          struct data_update_opts *data_opts)
 909 {
 910         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 911         const struct bch_extent_ptr *ptr;
 912         struct bch_ioctl_data *op = arg;
 913         unsigned i = 0;
 914
 915         data_opts->rewrite_ptrs         = 0;
 916         data_opts->target               = 0;
 917         data_opts->extra_replicas       = 0;
 918         data_opts->btree_insert_flags   = 0;
 919
 920         bkey_for_each_ptr(ptrs, ptr) {
 921                 if (ptr->dev == op->migrate.dev)
 922                         data_opts->rewrite_ptrs |= 1U << i;
 923                 i++;
 924         }
 925
 926         return data_opts->rewrite_ptrs != 0;
 927 }
 928
 929 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 930                                    struct btree *b,
 931                                    struct bch_io_opts *io_opts,
 932                                    struct data_update_opts *data_opts)
 933 {
 934         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 935 }
 936
 937 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 938                                struct btree *b,
 939                                struct bch_io_opts *io_opts,
 940                                struct data_update_opts *data_opts)
 941 {
 942         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 943 }
 944
 945 static bool bformat_needs_redo(struct bkey_format *f)
 946 {
 947         unsigned i;
 948
 949         for (i = 0; i < f->nr_fields; i++) {
 950                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 951                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 952                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 953
 954                 if (f->bits_per_field[i] > unpacked_bits)
 955                         return true;
 956
 957                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 958                         return true;
 959
 960                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 961                      unpacked_mask) <
 962                     field_offset)
 963                         return true;
 964         }
 965
 966         return false;
 967 }
 968
 969 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 970                                    struct btree *b,
 971                                    struct bch_io_opts *io_opts,
 972                                    struct data_update_opts *data_opts)
 973 {
 974         if (b->version_ondisk != c->sb.version ||
 975             btree_node_need_rewrite(b) ||
 976             bformat_needs_redo(&b->format)) {
 977                 data_opts->target               = 0;
 978                 data_opts->extra_replicas       = 0;
 979                 data_opts->btree_insert_flags   = 0;
 980                 return true;
 981         }
 982
 983         return false;
 984 }
 985
 986 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 987 {
 988         int ret;
 989
 990         ret = bch2_move_btree(c,
 991                               BBPOS_MIN,
 992                               BBPOS_MAX,
 993                               rewrite_old_nodes_pred, c, stats);
 994         if (!ret) {
 995                 mutex_lock(&c->sb_lock);
 996                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
 997                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 998                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
 999                 bch2_write_super(c);
1000                 mutex_unlock(&c->sb_lock);
1001         }
1002
1003         bch_err_fn(c, ret);
1004         return ret;
1005 }
1006
1007 static bool drop_extra_replicas_pred(struct bch_fs *c, void *arg,
1008                              struct bkey_s_c k,
1009                              struct bch_io_opts *io_opts,
1010                              struct data_update_opts *data_opts)
1011 {
1012         unsigned durability = bch2_bkey_durability(c, k);
1013         unsigned replicas = bkey_is_btree_ptr(k.k)
1014                 ? c->opts.metadata_replicas
1015                 : io_opts->data_replicas;
1016         const union bch_extent_entry *entry;
1017         struct extent_ptr_decoded p;
1018         unsigned i = 0;
1019
1020         bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
1021                 unsigned d = bch2_extent_ptr_durability(c, &p);
1022
1023                 if (d && durability - d >= replicas) {
1024                         data_opts->kill_ptrs |= BIT(i);
1025                         durability -= d;
1026                 }
1027
1028                 i++;
1029         }
1030
1031         return data_opts->kill_ptrs != 0;
1032 }
1033
1034 static bool drop_extra_replicas_btree_pred(struct bch_fs *c, void *arg,
1035                                    struct btree *b,
1036                                    struct bch_io_opts *io_opts,
1037                                    struct data_update_opts *data_opts)
1038 {
1039         return drop_extra_replicas_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1040 }
1041
1042 int bch2_data_job(struct bch_fs *c,
1043                   struct bch_move_stats *stats,
1044                   struct bch_ioctl_data op)
1045 {
1046         struct bbpos start      = BBPOS(op.start_btree, op.start_pos);
1047         struct bbpos end        = BBPOS(op.end_btree, op.end_pos);
1048         int ret = 0;
1049
1050         if (op.op >= BCH_DATA_OP_NR)
1051                 return -EINVAL;
1052
1053         bch2_move_stats_init(stats, bch2_data_ops_strs[op.op]);
1054
1055         switch (op.op) {
1056         case BCH_DATA_OP_rereplicate:
1057                 stats->data_type = BCH_DATA_journal;
1058                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1059                 ret = bch2_move_btree(c, start, end,
1060                                       rereplicate_btree_pred, c, stats) ?: ret;
1061                 ret = bch2_move_data(c, start, end,
1062                                      NULL,
1063                                      stats,
1064                                      writepoint_hashed((unsigned long) current),
1065                                      true,
1066                                      rereplicate_pred, c) ?: ret;
1067                 ret = bch2_replicas_gc2(c) ?: ret;
1068                 break;
1069         case BCH_DATA_OP_migrate:
1070                 if (op.migrate.dev >= c->sb.nr_devices)
1071                         return -EINVAL;
1072
1073                 stats->data_type = BCH_DATA_journal;
1074                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1075                 ret = bch2_move_btree(c, start, end,
1076                                       migrate_btree_pred, &op, stats) ?: ret;
1077                 ret = bch2_move_data(c, start, end,
1078                                      NULL,
1079                                      stats,
1080                                      writepoint_hashed((unsigned long) current),
1081                                      true,
1082                                      migrate_pred, &op) ?: ret;
1083                 ret = bch2_replicas_gc2(c) ?: ret;
1084                 break;
1085         case BCH_DATA_OP_rewrite_old_nodes:
1086                 ret = bch2_scan_old_btree_nodes(c, stats);
1087                 break;
1088         case BCH_DATA_OP_drop_extra_replicas:
1089                 ret = bch2_move_btree(c, start, end,
1090                                 drop_extra_replicas_btree_pred, c, stats) ?: ret;
1091                 ret = bch2_move_data(c, start, end, NULL, stats,
1092                                 writepoint_hashed((unsigned long) current),
1093                                 true,
1094                                 drop_extra_replicas_pred, c) ?: ret;
1095                 ret = bch2_replicas_gc2(c) ?: ret;
1096                 break;
1097         default:
1098                 ret = -EINVAL;
1099         }
1100
1101         bch2_move_stats_exit(stats, c);
1102         return ret;
1103 }
1104
1105 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1106 {
1107         prt_printf(out, "%s: data type=%s pos=",
1108                    stats->name,
1109                    bch2_data_types[stats->data_type]);
1110         bch2_bbpos_to_text(out, stats->pos);
1111         prt_newline(out);
1112         printbuf_indent_add(out, 2);
1113
1114         prt_str(out, "keys moved:  ");
1115         prt_u64(out, atomic64_read(&stats->keys_moved));
1116         prt_newline(out);
1117
1118         prt_str(out, "keys raced:  ");
1119         prt_u64(out, atomic64_read(&stats->keys_raced));
1120         prt_newline(out);
1121
1122         prt_str(out, "bytes seen:  ");
1123         prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1124         prt_newline(out);
1125
1126         prt_str(out, "bytes moved: ");
1127         prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1128         prt_newline(out);
1129
1130         prt_str(out, "bytes raced: ");
1131         prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1132         prt_newline(out);
1133
1134         printbuf_indent_sub(out, 2);
1135 }
1136
1137 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1138 {
1139         struct moving_io *io;
1140
1141         bch2_move_stats_to_text(out, ctxt->stats);
1142         printbuf_indent_add(out, 2);
1143
1144         prt_printf(out, "reads: ios %u/%u sectors %u/%u",
1145                    atomic_read(&ctxt->read_ios),
1146                    c->opts.move_ios_in_flight,
1147                    atomic_read(&ctxt->read_sectors),
1148                    c->opts.move_bytes_in_flight >> 9);
1149         prt_newline(out);
1150
1151         prt_printf(out, "writes: ios %u/%u sectors %u/%u",
1152                    atomic_read(&ctxt->write_ios),
1153                    c->opts.move_ios_in_flight,
1154                    atomic_read(&ctxt->write_sectors),
1155                    c->opts.move_bytes_in_flight >> 9);
1156         prt_newline(out);
1157
1158         printbuf_indent_add(out, 2);
1159
1160         mutex_lock(&ctxt->lock);
1161         list_for_each_entry(io, &ctxt->ios, io_list)
1162                 bch2_write_op_to_text(out, &io->write.op);
1163         mutex_unlock(&ctxt->lock);
1164
1165         printbuf_indent_sub(out, 4);
1166 }
1167
1168 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1169 {
1170         struct moving_context *ctxt;
1171
1172         mutex_lock(&c->moving_context_lock);
1173         list_for_each_entry(ctxt, &c->moving_context_list, list)
1174                 bch2_moving_ctxt_to_text(out, c, ctxt);
1175         mutex_unlock(&c->moving_context_lock);
1176 }
1177
1178 void bch2_fs_move_init(struct bch_fs *c)
1179 {
1180         INIT_LIST_HEAD(&c->moving_context_list);
1181         mutex_init(&c->moving_context_lock);
1182 }