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 struct moving_io {
  60         struct list_head                read_list;
  61         struct list_head                io_list;
  62         struct move_bucket_in_flight    *b;
  63         struct closure                  cl;
  64         bool                            read_completed;
  65
  66         unsigned                        read_sectors;
  67         unsigned                        write_sectors;
  68
  69         struct bch_read_bio             rbio;
  70
  71         struct data_update              write;
  72         /* Must be last since it is variable size */
  73         struct bio_vec                  bi_inline_vecs[];
  74 };
  75
  76 static void move_free(struct moving_io *io)
  77 {
  78         struct moving_context *ctxt = io->write.ctxt;
  79
  80         if (io->b)
  81                 atomic_dec(&io->b->count);
  82
  83         bch2_data_update_exit(&io->write);
  84
  85         mutex_lock(&ctxt->lock);
  86         list_del(&io->io_list);
  87         wake_up(&ctxt->wait);
  88         mutex_unlock(&ctxt->lock);
  89
  90         kfree(io);
  91 }
  92
  93 static void move_write_done(struct bch_write_op *op)
  94 {
  95         struct moving_io *io = container_of(op, struct moving_io, write.op);
  96         struct moving_context *ctxt = io->write.ctxt;
  97
  98         if (io->write.op.error)
  99                 ctxt->write_error = true;
 100
 101         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 102         atomic_dec(&io->write.ctxt->write_ios);
 103         move_free(io);
 104         closure_put(&ctxt->cl);
 105 }
 106
 107 static void move_write(struct moving_io *io)
 108 {
 109         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 110                 move_free(io);
 111                 return;
 112         }
 113
 114         closure_get(&io->write.ctxt->cl);
 115         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 116         atomic_inc(&io->write.ctxt->write_ios);
 117
 118         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 119 }
 120
 121 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 122 {
 123         struct moving_io *io =
 124                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 125
 126         return io && io->read_completed ? io : NULL;
 127 }
 128
 129 static void move_read_endio(struct bio *bio)
 130 {
 131         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 132         struct moving_context *ctxt = io->write.ctxt;
 133
 134         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 135         atomic_dec(&ctxt->read_ios);
 136         io->read_completed = true;
 137
 138         wake_up(&ctxt->wait);
 139         closure_put(&ctxt->cl);
 140 }
 141
 142 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
 143 {
 144         struct moving_io *io;
 145
 146         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 147                 bch2_trans_unlock_long(ctxt->trans);
 148                 list_del(&io->read_list);
 149                 move_write(io);
 150         }
 151 }
 152
 153 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 154 {
 155         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 156
 157         move_ctxt_wait_event(ctxt,
 158                 !atomic_read(&ctxt->write_sectors) ||
 159                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 160 }
 161
 162 void bch2_moving_ctxt_flush_all(struct moving_context *ctxt)
 163 {
 164         move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 165         bch2_trans_unlock_long(ctxt->trans);
 166         closure_sync(&ctxt->cl);
 167 }
 168
 169 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 170 {
 171         struct bch_fs *c = ctxt->trans->c;
 172
 173         bch2_moving_ctxt_flush_all(ctxt);
 174
 175         EBUG_ON(atomic_read(&ctxt->write_sectors));
 176         EBUG_ON(atomic_read(&ctxt->write_ios));
 177         EBUG_ON(atomic_read(&ctxt->read_sectors));
 178         EBUG_ON(atomic_read(&ctxt->read_ios));
 179
 180         mutex_lock(&c->moving_context_lock);
 181         list_del(&ctxt->list);
 182         mutex_unlock(&c->moving_context_lock);
 183
 184         bch2_trans_put(ctxt->trans);
 185         memset(ctxt, 0, sizeof(*ctxt));
 186 }
 187
 188 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 189                            struct bch_fs *c,
 190                            struct bch_ratelimit *rate,
 191                            struct bch_move_stats *stats,
 192                            struct write_point_specifier wp,
 193                            bool wait_on_copygc)
 194 {
 195         memset(ctxt, 0, sizeof(*ctxt));
 196
 197         ctxt->trans     = bch2_trans_get(c);
 198         ctxt->fn        = (void *) _RET_IP_;
 199         ctxt->rate      = rate;
 200         ctxt->stats     = stats;
 201         ctxt->wp        = wp;
 202         ctxt->wait_on_copygc = wait_on_copygc;
 203
 204         closure_init_stack(&ctxt->cl);
 205
 206         mutex_init(&ctxt->lock);
 207         INIT_LIST_HEAD(&ctxt->reads);
 208         INIT_LIST_HEAD(&ctxt->ios);
 209         init_waitqueue_head(&ctxt->wait);
 210
 211         mutex_lock(&c->moving_context_lock);
 212         list_add(&ctxt->list, &c->moving_context_list);
 213         mutex_unlock(&c->moving_context_lock);
 214 }
 215
 216 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
 217 {
 218         trace_move_data(c, stats);
 219 }
 220
 221 void bch2_move_stats_init(struct bch_move_stats *stats, const char *name)
 222 {
 223         memset(stats, 0, sizeof(*stats));
 224         stats->data_type = BCH_DATA_user;
 225         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 226 }
 227
 228 int bch2_move_extent(struct moving_context *ctxt,
 229                      struct move_bucket_in_flight *bucket_in_flight,
 230                      struct btree_iter *iter,
 231                      struct bkey_s_c k,
 232                      struct bch_io_opts io_opts,
 233                      struct data_update_opts data_opts)
 234 {
 235         struct btree_trans *trans = ctxt->trans;
 236         struct bch_fs *c = trans->c;
 237         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 238         struct moving_io *io;
 239         const union bch_extent_entry *entry;
 240         struct extent_ptr_decoded p;
 241         unsigned sectors = k.k->size, pages;
 242         int ret = -ENOMEM;
 243
 244         if (ctxt->stats)
 245                 ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
 246         trace_move_extent2(c, k);
 247
 248         bch2_data_update_opts_normalize(k, &data_opts);
 249
 250         if (!data_opts.rewrite_ptrs &&
 251             !data_opts.extra_replicas) {
 252                 if (data_opts.kill_ptrs)
 253                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 254                 return 0;
 255         }
 256
 257         /*
 258          * Before memory allocations & taking nocow locks in
 259          * bch2_data_update_init():
 260          */
 261         bch2_trans_unlock(trans);
 262
 263         /* write path might have to decompress data: */
 264         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 265                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 266
 267         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 268         io = kzalloc(sizeof(struct moving_io) +
 269                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 270         if (!io)
 271                 goto err;
 272
 273         INIT_LIST_HEAD(&io->io_list);
 274         io->write.ctxt          = ctxt;
 275         io->read_sectors        = k.k->size;
 276         io->write_sectors       = k.k->size;
 277
 278         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 279         bio_set_prio(&io->write.op.wbio.bio,
 280                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 281
 282         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 283                                  GFP_KERNEL))
 284                 goto err_free;
 285
 286         io->rbio.c              = c;
 287         io->rbio.opts           = io_opts;
 288         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 289         io->rbio.bio.bi_vcnt = pages;
 290         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 291         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 292
 293         io->rbio.bio.bi_opf             = REQ_OP_READ;
 294         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 295         io->rbio.bio.bi_end_io          = move_read_endio;
 296
 297         ret = bch2_data_update_init(trans, iter, ctxt, &io->write, ctxt->wp,
 298                                     io_opts, data_opts, iter->btree_id, k);
 299         if (ret)
 300                 goto err_free_pages;
 301
 302         io->write.op.end_io = move_write_done;
 303
 304         if (ctxt->rate)
 305                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 306
 307         if (ctxt->stats) {
 308                 atomic64_inc(&ctxt->stats->keys_moved);
 309                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 310         }
 311
 312         if (bucket_in_flight) {
 313                 io->b = bucket_in_flight;
 314                 atomic_inc(&io->b->count);
 315         }
 316
 317         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 318         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 319         trace_move_extent_read2(c, k);
 320
 321         mutex_lock(&ctxt->lock);
 322         atomic_add(io->read_sectors, &ctxt->read_sectors);
 323         atomic_inc(&ctxt->read_ios);
 324
 325         list_add_tail(&io->read_list, &ctxt->reads);
 326         list_add_tail(&io->io_list, &ctxt->ios);
 327         mutex_unlock(&ctxt->lock);
 328
 329         /*
 330          * dropped by move_read_endio() - guards against use after free of
 331          * ctxt when doing wakeup
 332          */
 333         closure_get(&ctxt->cl);
 334         bch2_read_extent(trans, &io->rbio,
 335                          bkey_start_pos(k.k),
 336                          iter->btree_id, k, 0,
 337                          BCH_READ_NODECODE|
 338                          BCH_READ_LAST_FRAGMENT);
 339         return 0;
 340 err_free_pages:
 341         bio_free_pages(&io->write.op.wbio.bio);
 342 err_free:
 343         kfree(io);
 344 err:
 345         if (ret == -BCH_ERR_data_update_done)
 346                 return 0;
 347
 348         if (bch2_err_matches(ret, EROFS) ||
 349             bch2_err_matches(ret, BCH_ERR_transaction_restart))
 350                 return ret;
 351
 352         count_event(c, move_extent_start_fail);
 353
 354         if (trace_move_extent_start_fail_enabled()) {
 355                 struct printbuf buf = PRINTBUF;
 356
 357                 bch2_bkey_val_to_text(&buf, c, k);
 358                 prt_str(&buf, ": ");
 359                 prt_str(&buf, bch2_err_str(ret));
 360                 trace_move_extent_start_fail(c, buf.buf);
 361                 printbuf_exit(&buf);
 362         }
 363         return ret;
 364 }
 365
 366 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 367                           struct per_snapshot_io_opts *io_opts,
 368                           struct bkey_s_c extent_k)
 369 {
 370         struct bch_fs *c = trans->c;
 371         u32 restart_count = trans->restart_count;
 372         int ret = 0;
 373
 374         if (io_opts->cur_inum != extent_k.k->p.inode) {
 375                 struct btree_iter iter;
 376                 struct bkey_s_c k;
 377
 378                 io_opts->d.nr = 0;
 379
 380                 for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
 381                                    BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
 382                         if (k.k->p.offset != extent_k.k->p.inode)
 383                                 break;
 384
 385                         if (!bkey_is_inode(k.k))
 386                                 continue;
 387
 388                         struct bch_inode_unpacked inode;
 389                         BUG_ON(bch2_inode_unpack(k, &inode));
 390
 391                         struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 392                         bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
 393
 394                         ret = darray_push(&io_opts->d, e);
 395                         if (ret)
 396                                 break;
 397                 }
 398                 bch2_trans_iter_exit(trans, &iter);
 399                 io_opts->cur_inum = extent_k.k->p.inode;
 400         }
 401
 402         ret = ret ?: trans_was_restarted(trans, restart_count);
 403         if (ret)
 404                 return ERR_PTR(ret);
 405
 406         if (extent_k.k->p.snapshot) {
 407                 struct snapshot_io_opts_entry *i;
 408                 darray_for_each(io_opts->d, i)
 409                         if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
 410                                 return &i->io_opts;
 411         }
 412
 413         return &io_opts->fs_io_opts;
 414 }
 415
 416 int bch2_move_get_io_opts_one(struct btree_trans *trans,
 417                               struct bch_io_opts *io_opts,
 418                               struct bkey_s_c extent_k)
 419 {
 420         struct btree_iter iter;
 421         struct bkey_s_c k;
 422         int ret;
 423
 424         /* reflink btree? */
 425         if (!extent_k.k->p.inode) {
 426                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 427                 return 0;
 428         }
 429
 430         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 431                                SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
 432                                BTREE_ITER_CACHED);
 433         ret = bkey_err(k);
 434         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 435                 return ret;
 436
 437         if (!ret && bkey_is_inode(k.k)) {
 438                 struct bch_inode_unpacked inode;
 439                 bch2_inode_unpack(k, &inode);
 440                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 441         } else {
 442                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 443         }
 444
 445         bch2_trans_iter_exit(trans, &iter);
 446         return 0;
 447 }
 448
 449 int bch2_move_ratelimit(struct moving_context *ctxt)
 450 {
 451         struct bch_fs *c = ctxt->trans->c;
 452         u64 delay;
 453
 454         if (ctxt->wait_on_copygc && c->copygc_running) {
 455                 bch2_moving_ctxt_flush_all(ctxt);
 456                 wait_event_killable(c->copygc_running_wq,
 457                                     !c->copygc_running ||
 458                                     kthread_should_stop());
 459         }
 460
 461         do {
 462                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 463
 464                 if (kthread_should_stop())
 465                         return 1;
 466
 467                 if (delay)
 468                         move_ctxt_wait_event_timeout(ctxt,
 469                                         freezing(current) ||
 470                                         kthread_should_stop(),
 471                                         delay);
 472
 473                 if (unlikely(freezing(current))) {
 474                         bch2_moving_ctxt_flush_all(ctxt);
 475                         try_to_freeze();
 476                 }
 477         } while (delay);
 478
 479         /*
 480          * XXX: these limits really ought to be per device, SSDs and hard drives
 481          * will want different limits
 482          */
 483         move_ctxt_wait_event(ctxt,
 484                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 485                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 486                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 487                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 488
 489         return 0;
 490 }
 491
 492 static int bch2_move_data_btree(struct moving_context *ctxt,
 493                                 struct bpos start,
 494                                 struct bpos end,
 495                                 move_pred_fn pred, void *arg,
 496                                 enum btree_id btree_id)
 497 {
 498         struct btree_trans *trans = ctxt->trans;
 499         struct bch_fs *c = trans->c;
 500         struct per_snapshot_io_opts snapshot_io_opts;
 501         struct bch_io_opts *io_opts;
 502         struct bkey_buf sk;
 503         struct btree_iter iter;
 504         struct bkey_s_c k;
 505         struct data_update_opts data_opts;
 506         int ret = 0, ret2;
 507
 508         per_snapshot_io_opts_init(&snapshot_io_opts, c);
 509         bch2_bkey_buf_init(&sk);
 510
 511         if (ctxt->stats) {
 512                 ctxt->stats->data_type  = BCH_DATA_user;
 513                 ctxt->stats->pos        = BBPOS(btree_id, start);
 514         }
 515
 516         bch2_trans_iter_init(trans, &iter, btree_id, start,
 517                              BTREE_ITER_PREFETCH|
 518                              BTREE_ITER_ALL_SNAPSHOTS);
 519
 520         if (ctxt->rate)
 521                 bch2_ratelimit_reset(ctxt->rate);
 522
 523         while (!bch2_move_ratelimit(ctxt)) {
 524                 bch2_trans_begin(trans);
 525
 526                 k = bch2_btree_iter_peek(&iter);
 527                 if (!k.k)
 528                         break;
 529
 530                 ret = bkey_err(k);
 531                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 532                         continue;
 533                 if (ret)
 534                         break;
 535
 536                 if (bkey_ge(bkey_start_pos(k.k), end))
 537                         break;
 538
 539                 if (ctxt->stats)
 540                         ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
 541
 542                 if (!bkey_extent_is_direct_data(k.k))
 543                         goto next_nondata;
 544
 545                 io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
 546                 ret = PTR_ERR_OR_ZERO(io_opts);
 547                 if (ret)
 548                         continue;
 549
 550                 memset(&data_opts, 0, sizeof(data_opts));
 551                 if (!pred(c, arg, k, io_opts, &data_opts))
 552                         goto next;
 553
 554                 /*
 555                  * The iterator gets unlocked by __bch2_read_extent - need to
 556                  * save a copy of @k elsewhere:
 557                  */
 558                 bch2_bkey_buf_reassemble(&sk, c, k);
 559                 k = bkey_i_to_s_c(sk.k);
 560
 561                 ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
 562                 if (ret2) {
 563                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 564                                 continue;
 565
 566                         if (ret2 == -ENOMEM) {
 567                                 /* memory allocation failure, wait for some IO to finish */
 568                                 bch2_move_ctxt_wait_for_io(ctxt);
 569                                 continue;
 570                         }
 571
 572                         /* XXX signal failure */
 573                         goto next;
 574                 }
 575 next:
 576                 if (ctxt->stats)
 577                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 578 next_nondata:
 579                 bch2_btree_iter_advance(&iter);
 580         }
 581
 582         bch2_trans_iter_exit(trans, &iter);
 583         bch2_bkey_buf_exit(&sk, c);
 584         per_snapshot_io_opts_exit(&snapshot_io_opts);
 585
 586         return ret;
 587 }
 588
 589 int __bch2_move_data(struct moving_context *ctxt,
 590                      struct bbpos start,
 591                      struct bbpos end,
 592                      move_pred_fn pred, void *arg)
 593 {
 594         struct bch_fs *c = ctxt->trans->c;
 595         enum btree_id id;
 596         int ret = 0;
 597
 598         for (id = start.btree;
 599              id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 600              id++) {
 601                 ctxt->stats->pos = BBPOS(id, POS_MIN);
 602
 603                 if (!btree_type_has_ptrs(id) ||
 604                     !bch2_btree_id_root(c, id)->b)
 605                         continue;
 606
 607                 ret = bch2_move_data_btree(ctxt,
 608                                        id == start.btree ? start.pos : POS_MIN,
 609                                        id == end.btree   ? end.pos   : POS_MAX,
 610                                        pred, arg, id);
 611                 if (ret)
 612                         break;
 613         }
 614
 615         return ret;
 616 }
 617
 618 int bch2_move_data(struct bch_fs *c,
 619                    struct bbpos start,
 620                    struct bbpos end,
 621                    struct bch_ratelimit *rate,
 622                    struct bch_move_stats *stats,
 623                    struct write_point_specifier wp,
 624                    bool wait_on_copygc,
 625                    move_pred_fn pred, void *arg)
 626 {
 627
 628         struct moving_context ctxt;
 629         int ret;
 630
 631         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 632         ret = __bch2_move_data(&ctxt, start, end, pred, arg);
 633         bch2_moving_ctxt_exit(&ctxt);
 634
 635         return ret;
 636 }
 637
 638 int bch2_evacuate_bucket(struct moving_context *ctxt,
 639                            struct move_bucket_in_flight *bucket_in_flight,
 640                            struct bpos bucket, int gen,
 641                            struct data_update_opts _data_opts)
 642 {
 643         struct btree_trans *trans = ctxt->trans;
 644         struct bch_fs *c = trans->c;
 645         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 646         struct btree_iter iter;
 647         struct bkey_buf sk;
 648         struct bch_backpointer bp;
 649         struct bch_alloc_v4 a_convert;
 650         const struct bch_alloc_v4 *a;
 651         struct bkey_s_c k;
 652         struct data_update_opts data_opts;
 653         unsigned dirty_sectors, bucket_size;
 654         u64 fragmentation;
 655         struct bpos bp_pos = POS_MIN;
 656         int ret = 0;
 657
 658         trace_bucket_evacuate(c, &bucket);
 659
 660         bch2_bkey_buf_init(&sk);
 661
 662         /*
 663          * We're not run in a context that handles transaction restarts:
 664          */
 665         bch2_trans_begin(trans);
 666
 667         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 668                              bucket, BTREE_ITER_CACHED);
 669         ret = lockrestart_do(trans,
 670                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 671         bch2_trans_iter_exit(trans, &iter);
 672
 673         if (ret) {
 674                 bch_err_msg(c, ret, "looking up alloc key");
 675                 goto err;
 676         }
 677
 678         a = bch2_alloc_to_v4(k, &a_convert);
 679         dirty_sectors = bch2_bucket_sectors_dirty(*a);
 680         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 681         fragmentation = a->fragmentation_lru;
 682
 683         ret = bch2_btree_write_buffer_tryflush(trans);
 684         bch_err_msg(c, ret, "flushing btree write buffer");
 685         if (ret)
 686                 goto err;
 687
 688         while (!(ret = bch2_move_ratelimit(ctxt))) {
 689                 if (kthread_should_stop())
 690                         break;
 691
 692                 bch2_trans_begin(trans);
 693
 694                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 695                                                 &bp_pos, &bp,
 696                                                 BTREE_ITER_CACHED);
 697                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 698                         continue;
 699                 if (ret)
 700                         goto err;
 701                 if (bkey_eq(bp_pos, POS_MAX))
 702                         break;
 703
 704                 if (!bp.level) {
 705                         const struct bch_extent_ptr *ptr;
 706                         unsigned i = 0;
 707
 708                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 709                         ret = bkey_err(k);
 710                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 711                                 continue;
 712                         if (ret)
 713                                 goto err;
 714                         if (!k.k)
 715                                 goto next;
 716
 717                         bch2_bkey_buf_reassemble(&sk, c, k);
 718                         k = bkey_i_to_s_c(sk.k);
 719
 720                         ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
 721                         if (ret) {
 722                                 bch2_trans_iter_exit(trans, &iter);
 723                                 continue;
 724                         }
 725
 726                         data_opts = _data_opts;
 727                         data_opts.target        = io_opts.background_target;
 728                         data_opts.rewrite_ptrs = 0;
 729
 730                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 731                                 if (ptr->dev == bucket.inode) {
 732                                         data_opts.rewrite_ptrs |= 1U << i;
 733                                         if (ptr->cached) {
 734                                                 bch2_trans_iter_exit(trans, &iter);
 735                                                 goto next;
 736                                         }
 737                                 }
 738                                 i++;
 739                         }
 740
 741                         ret = bch2_move_extent(ctxt, bucket_in_flight,
 742                                                &iter, k, io_opts, data_opts);
 743                         bch2_trans_iter_exit(trans, &iter);
 744
 745                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 746                                 continue;
 747                         if (ret == -ENOMEM) {
 748                                 /* memory allocation failure, wait for some IO to finish */
 749                                 bch2_move_ctxt_wait_for_io(ctxt);
 750                                 continue;
 751                         }
 752                         if (ret)
 753                                 goto err;
 754
 755                         if (ctxt->stats)
 756                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 757                 } else {
 758                         struct btree *b;
 759
 760                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 761                         ret = PTR_ERR_OR_ZERO(b);
 762                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 763                                 continue;
 764                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 765                                 continue;
 766                         if (ret)
 767                                 goto err;
 768                         if (!b)
 769                                 goto next;
 770
 771                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 772                         bch2_trans_iter_exit(trans, &iter);
 773
 774                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 775                                 continue;
 776                         if (ret)
 777                                 goto err;
 778
 779                         if (ctxt->rate)
 780                                 bch2_ratelimit_increment(ctxt->rate,
 781                                                          c->opts.btree_node_size >> 9);
 782                         if (ctxt->stats) {
 783                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 784                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 785                         }
 786                 }
 787 next:
 788                 bp_pos = bpos_nosnap_successor(bp_pos);
 789         }
 790
 791         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 792 err:
 793         bch2_bkey_buf_exit(&sk, c);
 794         return ret;
 795 }
 796
 797 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 798                                 struct btree *, struct bch_io_opts *,
 799                                 struct data_update_opts *);
 800
 801 static int bch2_move_btree(struct bch_fs *c,
 802                            struct bbpos start,
 803                            struct bbpos end,
 804                            move_btree_pred pred, void *arg,
 805                            struct bch_move_stats *stats)
 806 {
 807         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 808         struct moving_context ctxt;
 809         struct btree_trans *trans;
 810         struct btree_iter iter;
 811         struct btree *b;
 812         enum btree_id btree;
 813         struct data_update_opts data_opts;
 814         int ret = 0;
 815
 816         bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
 817                               writepoint_ptr(&c->btree_write_point),
 818                               true);
 819         trans = ctxt.trans;
 820
 821         stats->data_type = BCH_DATA_btree;
 822
 823         for (btree = start.btree;
 824              btree <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 825              btree ++) {
 826                 stats->pos = BBPOS(btree, POS_MIN);
 827
 828                 if (!bch2_btree_id_root(c, btree)->b)
 829                         continue;
 830
 831                 bch2_trans_node_iter_init(trans, &iter, btree, POS_MIN, 0, 0,
 832                                           BTREE_ITER_PREFETCH);
 833 retry:
 834                 ret = 0;
 835                 while (bch2_trans_begin(trans),
 836                        (b = bch2_btree_iter_peek_node(&iter)) &&
 837                        !(ret = PTR_ERR_OR_ZERO(b))) {
 838                         if (kthread_should_stop())
 839                                 break;
 840
 841                         if ((cmp_int(btree, end.btree) ?:
 842                              bpos_cmp(b->key.k.p, end.pos)) > 0)
 843                                 break;
 844
 845                         stats->pos = BBPOS(iter.btree_id, iter.pos);
 846
 847                         if (!pred(c, arg, b, &io_opts, &data_opts))
 848                                 goto next;
 849
 850                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
 851                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 852                                 continue;
 853                         if (ret)
 854                                 break;
 855 next:
 856                         bch2_btree_iter_next_node(&iter);
 857                 }
 858                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 859                         goto retry;
 860
 861                 bch2_trans_iter_exit(trans, &iter);
 862
 863                 if (kthread_should_stop())
 864                         break;
 865         }
 866
 867         bch_err_fn(c, ret);
 868         bch2_moving_ctxt_exit(&ctxt);
 869         bch2_btree_interior_updates_flush(c);
 870
 871         return ret;
 872 }
 873
 874 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 875                              struct bkey_s_c k,
 876                              struct bch_io_opts *io_opts,
 877                              struct data_update_opts *data_opts)
 878 {
 879         unsigned nr_good = bch2_bkey_durability(c, k);
 880         unsigned replicas = bkey_is_btree_ptr(k.k)
 881                 ? c->opts.metadata_replicas
 882                 : io_opts->data_replicas;
 883
 884         if (!nr_good || nr_good >= replicas)
 885                 return false;
 886
 887         data_opts->target               = 0;
 888         data_opts->extra_replicas       = replicas - nr_good;
 889         data_opts->btree_insert_flags   = 0;
 890         return true;
 891 }
 892
 893 static bool migrate_pred(struct bch_fs *c, void *arg,
 894                          struct bkey_s_c k,
 895                          struct bch_io_opts *io_opts,
 896                          struct data_update_opts *data_opts)
 897 {
 898         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 899         const struct bch_extent_ptr *ptr;
 900         struct bch_ioctl_data *op = arg;
 901         unsigned i = 0;
 902
 903         data_opts->rewrite_ptrs         = 0;
 904         data_opts->target               = 0;
 905         data_opts->extra_replicas       = 0;
 906         data_opts->btree_insert_flags   = 0;
 907
 908         bkey_for_each_ptr(ptrs, ptr) {
 909                 if (ptr->dev == op->migrate.dev)
 910                         data_opts->rewrite_ptrs |= 1U << i;
 911                 i++;
 912         }
 913
 914         return data_opts->rewrite_ptrs != 0;
 915 }
 916
 917 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 918                                    struct btree *b,
 919                                    struct bch_io_opts *io_opts,
 920                                    struct data_update_opts *data_opts)
 921 {
 922         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 923 }
 924
 925 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 926                                struct btree *b,
 927                                struct bch_io_opts *io_opts,
 928                                struct data_update_opts *data_opts)
 929 {
 930         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 931 }
 932
 933 static bool bformat_needs_redo(struct bkey_format *f)
 934 {
 935         unsigned i;
 936
 937         for (i = 0; i < f->nr_fields; i++) {
 938                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 939                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 940                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 941
 942                 if (f->bits_per_field[i] > unpacked_bits)
 943                         return true;
 944
 945                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 946                         return true;
 947
 948                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 949                      unpacked_mask) <
 950                     field_offset)
 951                         return true;
 952         }
 953
 954         return false;
 955 }
 956
 957 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
 958                                    struct btree *b,
 959                                    struct bch_io_opts *io_opts,
 960                                    struct data_update_opts *data_opts)
 961 {
 962         if (b->version_ondisk != c->sb.version ||
 963             btree_node_need_rewrite(b) ||
 964             bformat_needs_redo(&b->format)) {
 965                 data_opts->target               = 0;
 966                 data_opts->extra_replicas       = 0;
 967                 data_opts->btree_insert_flags   = 0;
 968                 return true;
 969         }
 970
 971         return false;
 972 }
 973
 974 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
 975 {
 976         int ret;
 977
 978         ret = bch2_move_btree(c,
 979                               BBPOS_MIN,
 980                               BBPOS_MAX,
 981                               rewrite_old_nodes_pred, c, stats);
 982         if (!ret) {
 983                 mutex_lock(&c->sb_lock);
 984                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
 985                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
 986                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
 987                 bch2_write_super(c);
 988                 mutex_unlock(&c->sb_lock);
 989         }
 990
 991         bch_err_fn(c, ret);
 992         return ret;
 993 }
 994
 995 static bool drop_extra_replicas_pred(struct bch_fs *c, void *arg,
 996                              struct bkey_s_c k,
 997                              struct bch_io_opts *io_opts,
 998                              struct data_update_opts *data_opts)
 999 {
1000         unsigned durability = bch2_bkey_durability(c, k);
1001         unsigned replicas = bkey_is_btree_ptr(k.k)
1002                 ? c->opts.metadata_replicas
1003                 : io_opts->data_replicas;
1004         const union bch_extent_entry *entry;
1005         struct extent_ptr_decoded p;
1006         unsigned i = 0;
1007
1008         bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
1009                 unsigned d = bch2_extent_ptr_durability(c, &p);
1010
1011                 if (d && durability - d >= replicas) {
1012                         data_opts->kill_ptrs |= BIT(i);
1013                         durability -= d;
1014                 }
1015
1016                 i++;
1017         }
1018
1019         return data_opts->kill_ptrs != 0;
1020 }
1021
1022 static bool drop_extra_replicas_btree_pred(struct bch_fs *c, void *arg,
1023                                    struct btree *b,
1024                                    struct bch_io_opts *io_opts,
1025                                    struct data_update_opts *data_opts)
1026 {
1027         return drop_extra_replicas_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1028 }
1029
1030 int bch2_data_job(struct bch_fs *c,
1031                   struct bch_move_stats *stats,
1032                   struct bch_ioctl_data op)
1033 {
1034         struct bbpos start      = BBPOS(op.start_btree, op.start_pos);
1035         struct bbpos end        = BBPOS(op.end_btree, op.end_pos);
1036         int ret = 0;
1037
1038         if (op.op >= BCH_DATA_OP_NR)
1039                 return -EINVAL;
1040
1041         bch2_move_stats_init(stats, bch2_data_ops_strs[op.op]);
1042
1043         switch (op.op) {
1044         case BCH_DATA_OP_rereplicate:
1045                 stats->data_type = BCH_DATA_journal;
1046                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1047                 ret = bch2_move_btree(c, start, end,
1048                                       rereplicate_btree_pred, c, stats) ?: ret;
1049                 ret = bch2_move_data(c, start, end,
1050                                      NULL,
1051                                      stats,
1052                                      writepoint_hashed((unsigned long) current),
1053                                      true,
1054                                      rereplicate_pred, c) ?: ret;
1055                 ret = bch2_replicas_gc2(c) ?: ret;
1056                 break;
1057         case BCH_DATA_OP_migrate:
1058                 if (op.migrate.dev >= c->sb.nr_devices)
1059                         return -EINVAL;
1060
1061                 stats->data_type = BCH_DATA_journal;
1062                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1063                 ret = bch2_move_btree(c, start, end,
1064                                       migrate_btree_pred, &op, stats) ?: ret;
1065                 ret = bch2_move_data(c, start, end,
1066                                      NULL,
1067                                      stats,
1068                                      writepoint_hashed((unsigned long) current),
1069                                      true,
1070                                      migrate_pred, &op) ?: ret;
1071                 ret = bch2_replicas_gc2(c) ?: ret;
1072                 break;
1073         case BCH_DATA_OP_rewrite_old_nodes:
1074                 ret = bch2_scan_old_btree_nodes(c, stats);
1075                 break;
1076         case BCH_DATA_OP_drop_extra_replicas:
1077                 ret = bch2_move_btree(c, start, end,
1078                                 drop_extra_replicas_btree_pred, c, stats) ?: ret;
1079                 ret = bch2_move_data(c, start, end, NULL, stats,
1080                                 writepoint_hashed((unsigned long) current),
1081                                 true,
1082                                 drop_extra_replicas_pred, c) ?: ret;
1083                 ret = bch2_replicas_gc2(c) ?: ret;
1084                 break;
1085         default:
1086                 ret = -EINVAL;
1087         }
1088
1089         bch2_move_stats_exit(stats, c);
1090         return ret;
1091 }
1092
1093 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1094 {
1095         prt_printf(out, "%s: data type=%s pos=",
1096                    stats->name,
1097                    bch2_data_types[stats->data_type]);
1098         bch2_bbpos_to_text(out, stats->pos);
1099         prt_newline(out);
1100         printbuf_indent_add(out, 2);
1101
1102         prt_str(out, "keys moved:  ");
1103         prt_u64(out, atomic64_read(&stats->keys_moved));
1104         prt_newline(out);
1105
1106         prt_str(out, "keys raced:  ");
1107         prt_u64(out, atomic64_read(&stats->keys_raced));
1108         prt_newline(out);
1109
1110         prt_str(out, "bytes seen:  ");
1111         prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1112         prt_newline(out);
1113
1114         prt_str(out, "bytes moved: ");
1115         prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1116         prt_newline(out);
1117
1118         prt_str(out, "bytes raced: ");
1119         prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1120         prt_newline(out);
1121
1122         printbuf_indent_sub(out, 2);
1123 }
1124
1125 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1126 {
1127         struct moving_io *io;
1128
1129         bch2_move_stats_to_text(out, ctxt->stats);
1130         printbuf_indent_add(out, 2);
1131
1132         prt_printf(out, "reads: ios %u/%u sectors %u/%u",
1133                    atomic_read(&ctxt->read_ios),
1134                    c->opts.move_ios_in_flight,
1135                    atomic_read(&ctxt->read_sectors),
1136                    c->opts.move_bytes_in_flight >> 9);
1137         prt_newline(out);
1138
1139         prt_printf(out, "writes: ios %u/%u sectors %u/%u",
1140                    atomic_read(&ctxt->write_ios),
1141                    c->opts.move_ios_in_flight,
1142                    atomic_read(&ctxt->write_sectors),
1143                    c->opts.move_bytes_in_flight >> 9);
1144         prt_newline(out);
1145
1146         printbuf_indent_add(out, 2);
1147
1148         mutex_lock(&ctxt->lock);
1149         list_for_each_entry(io, &ctxt->ios, io_list)
1150                 bch2_write_op_to_text(out, &io->write.op);
1151         mutex_unlock(&ctxt->lock);
1152
1153         printbuf_indent_sub(out, 4);
1154 }
1155
1156 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1157 {
1158         struct moving_context *ctxt;
1159
1160         mutex_lock(&c->moving_context_lock);
1161         list_for_each_entry(ctxt, &c->moving_context_list, list)
1162                 bch2_moving_ctxt_to_text(out, c, ctxt);
1163         mutex_unlock(&c->moving_context_lock);
1164 }
1165
1166 void bch2_fs_move_init(struct bch_fs *c)
1167 {
1168         INIT_LIST_HEAD(&c->moving_context_list);
1169         mutex_init(&c->moving_context_lock);
1170 }