git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/move.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_background.h"
   5 #include "alloc_foreground.h"
   6 #include "backpointers.h"
   7 #include "bkey_buf.h"
   8 #include "btree_gc.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_write_buffer.h"
  12 #include "disk_groups.h"
  13 #include "ec.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "inode.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "move.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22 #include "keylist.h"
  23
  24 #include <linux/ioprio.h>
  25 #include <linux/kthread.h>
  26
  27 #include <trace/events/bcachefs.h>
  28
  29 static void progress_list_add(struct bch_fs *c, struct bch_move_stats *stats)
  30 {
  31         mutex_lock(&c->data_progress_lock);
  32         list_add(&stats->list, &c->data_progress_list);
  33         mutex_unlock(&c->data_progress_lock);
  34 }
  35
  36 static void progress_list_del(struct bch_fs *c, struct bch_move_stats *stats)
  37 {
  38         mutex_lock(&c->data_progress_lock);
  39         list_del(&stats->list);
  40         mutex_unlock(&c->data_progress_lock);
  41 }
  42
  43 struct moving_io {
  44         struct list_head                read_list;
  45         struct list_head                io_list;
  46         struct move_bucket_in_flight    *b;
  47         struct closure                  cl;
  48         bool                            read_completed;
  49
  50         unsigned                        read_sectors;
  51         unsigned                        write_sectors;
  52
  53         struct bch_read_bio             rbio;
  54
  55         struct data_update              write;
  56         /* Must be last since it is variable size */
  57         struct bio_vec                  bi_inline_vecs[0];
  58 };
  59
  60 static void move_free(struct moving_io *io)
  61 {
  62         struct moving_context *ctxt = io->write.ctxt;
  63
  64         if (io->b)
  65                 atomic_dec(&io->b->count);
  66
  67         bch2_data_update_exit(&io->write);
  68
  69         mutex_lock(&ctxt->lock);
  70         list_del(&io->io_list);
  71         wake_up(&ctxt->wait);
  72         mutex_unlock(&ctxt->lock);
  73
  74         kfree(io);
  75 }
  76
  77 static void move_write_done(struct bch_write_op *op)
  78 {
  79         struct moving_io *io = container_of(op, struct moving_io, write.op);
  80         struct moving_context *ctxt = io->write.ctxt;
  81
  82         if (io->write.op.error)
  83                 ctxt->write_error = true;
  84
  85         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
  86         atomic_dec(&io->write.ctxt->write_ios);
  87         move_free(io);
  88         closure_put(&ctxt->cl);
  89 }
  90
  91 static void move_write(struct moving_io *io)
  92 {
  93         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
  94                 move_free(io);
  95                 return;
  96         }
  97
  98         closure_get(&io->write.ctxt->cl);
  99         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 100         atomic_inc(&io->write.ctxt->write_ios);
 101
 102         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 103 }
 104
 105 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 106 {
 107         struct moving_io *io =
 108                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 109
 110         return io && io->read_completed ? io : NULL;
 111 }
 112
 113 static void move_read_endio(struct bio *bio)
 114 {
 115         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 116         struct moving_context *ctxt = io->write.ctxt;
 117
 118         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 119         atomic_dec(&ctxt->read_ios);
 120         io->read_completed = true;
 121
 122         wake_up(&ctxt->wait);
 123         closure_put(&ctxt->cl);
 124 }
 125
 126 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 127                                         struct btree_trans *trans)
 128 {
 129         struct moving_io *io;
 130
 131         if (trans)
 132                 bch2_trans_unlock(trans);
 133
 134         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 135                 list_del(&io->read_list);
 136                 move_write(io);
 137         }
 138 }
 139
 140 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 141                                        struct btree_trans *trans)
 142 {
 143         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 144
 145         move_ctxt_wait_event(ctxt, trans,
 146                 !atomic_read(&ctxt->write_sectors) ||
 147                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 148 }
 149
 150 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 151 {
 152         struct bch_fs *c = ctxt->c;
 153
 154         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 155         closure_sync(&ctxt->cl);
 156
 157         EBUG_ON(atomic_read(&ctxt->write_sectors));
 158         EBUG_ON(atomic_read(&ctxt->write_ios));
 159         EBUG_ON(atomic_read(&ctxt->read_sectors));
 160         EBUG_ON(atomic_read(&ctxt->read_ios));
 161
 162         if (ctxt->stats) {
 163                 progress_list_del(c, ctxt->stats);
 164                 trace_move_data(c,
 165                                 atomic64_read(&ctxt->stats->sectors_moved),
 166                                 atomic64_read(&ctxt->stats->keys_moved));
 167         }
 168
 169         mutex_lock(&c->moving_context_lock);
 170         list_del(&ctxt->list);
 171         mutex_unlock(&c->moving_context_lock);
 172 }
 173
 174 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 175                            struct bch_fs *c,
 176                            struct bch_ratelimit *rate,
 177                            struct bch_move_stats *stats,
 178                            struct write_point_specifier wp,
 179                            bool wait_on_copygc)
 180 {
 181         memset(ctxt, 0, sizeof(*ctxt));
 182
 183         ctxt->c         = c;
 184         ctxt->fn        = (void *) _RET_IP_;
 185         ctxt->rate      = rate;
 186         ctxt->stats     = stats;
 187         ctxt->wp        = wp;
 188         ctxt->wait_on_copygc = wait_on_copygc;
 189
 190         closure_init_stack(&ctxt->cl);
 191
 192         mutex_init(&ctxt->lock);
 193         INIT_LIST_HEAD(&ctxt->reads);
 194         INIT_LIST_HEAD(&ctxt->ios);
 195         init_waitqueue_head(&ctxt->wait);
 196
 197         mutex_lock(&c->moving_context_lock);
 198         list_add(&ctxt->list, &c->moving_context_list);
 199         mutex_unlock(&c->moving_context_lock);
 200
 201         if (stats) {
 202                 progress_list_add(c, stats);
 203                 stats->data_type = BCH_DATA_user;
 204         }
 205 }
 206
 207 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 208 {
 209         memset(stats, 0, sizeof(*stats));
 210         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 211 }
 212
 213 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 214                                  struct btree_iter *iter,
 215                                  struct bkey_s_c k,
 216                                  struct data_update_opts data_opts)
 217 {
 218         struct bch_fs *c = trans->c;
 219         struct bkey_i *n;
 220         int ret;
 221
 222         n = bch2_bkey_make_mut(trans, k);
 223         ret = PTR_ERR_OR_ZERO(n);
 224         if (ret)
 225                 return ret;
 226
 227         while (data_opts.kill_ptrs) {
 228                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 229                 struct bch_extent_ptr *ptr;
 230
 231                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 232                 data_opts.kill_ptrs ^= 1U << drop;
 233         }
 234
 235         /*
 236          * If the new extent no longer has any pointers, bch2_extent_normalize()
 237          * will do the appropriate thing with it (turning it into a
 238          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 239          */
 240         bch2_extent_normalize(c, bkey_i_to_s(n));
 241
 242         /*
 243          * Since we're not inserting through an extent iterator
 244          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 245          * we aren't using the extent overwrite path to delete, we're
 246          * just using the normal key deletion path:
 247          */
 248         if (bkey_deleted(&n->k))
 249                 n->k.size = 0;
 250
 251         return bch2_trans_relock(trans) ?:
 252                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 253                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 254 }
 255
 256 static int bch2_move_extent(struct btree_trans *trans,
 257                             struct btree_iter *iter,
 258                             struct moving_context *ctxt,
 259                             struct move_bucket_in_flight *bucket_in_flight,
 260                             struct bch_io_opts io_opts,
 261                             enum btree_id btree_id,
 262                             struct bkey_s_c k,
 263                             struct data_update_opts data_opts)
 264 {
 265         struct bch_fs *c = trans->c;
 266         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 267         struct moving_io *io;
 268         const union bch_extent_entry *entry;
 269         struct extent_ptr_decoded p;
 270         unsigned sectors = k.k->size, pages;
 271         int ret = -ENOMEM;
 272
 273         bch2_data_update_opts_normalize(k, &data_opts);
 274
 275         if (!data_opts.rewrite_ptrs &&
 276             !data_opts.extra_replicas) {
 277                 if (data_opts.kill_ptrs)
 278                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 279                 return 0;
 280         }
 281
 282         /*
 283          * Before memory allocations & taking nocow locks in
 284          * bch2_data_update_init():
 285          */
 286         bch2_trans_unlock(trans);
 287
 288         /* write path might have to decompress data: */
 289         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 290                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 291
 292         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 293         io = kzalloc(sizeof(struct moving_io) +
 294                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 295         if (!io)
 296                 goto err;
 297
 298         INIT_LIST_HEAD(&io->io_list);
 299         io->write.ctxt          = ctxt;
 300         io->read_sectors        = k.k->size;
 301         io->write_sectors       = k.k->size;
 302
 303         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 304         bio_set_prio(&io->write.op.wbio.bio,
 305                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 306
 307         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 308                                  GFP_KERNEL))
 309                 goto err_free;
 310
 311         io->rbio.c              = c;
 312         io->rbio.opts           = io_opts;
 313         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 314         io->rbio.bio.bi_vcnt = pages;
 315         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 316         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 317
 318         io->rbio.bio.bi_opf             = REQ_OP_READ;
 319         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 320         io->rbio.bio.bi_end_io          = move_read_endio;
 321
 322         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 323                                     io_opts, data_opts, btree_id, k);
 324         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 325                 goto err_free_pages;
 326
 327         if (ret == -BCH_ERR_unwritten_extent_update) {
 328                 bch2_update_unwritten_extent(trans, &io->write);
 329                 move_free(io);
 330                 return 0;
 331         }
 332
 333         BUG_ON(ret);
 334
 335         io->write.ctxt = ctxt;
 336         io->write.op.end_io = move_write_done;
 337
 338         if (ctxt->stats) {
 339                 atomic64_inc(&ctxt->stats->keys_moved);
 340                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 341         }
 342
 343         if (bucket_in_flight) {
 344                 io->b = bucket_in_flight;
 345                 atomic_inc(&io->b->count);
 346         }
 347
 348         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 349         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 350         trace_move_extent_read(k.k);
 351
 352
 353         mutex_lock(&ctxt->lock);
 354         atomic_add(io->read_sectors, &ctxt->read_sectors);
 355         atomic_inc(&ctxt->read_ios);
 356
 357         list_add_tail(&io->read_list, &ctxt->reads);
 358         list_add_tail(&io->io_list, &ctxt->ios);
 359         mutex_unlock(&ctxt->lock);
 360
 361         /*
 362          * dropped by move_read_endio() - guards against use after free of
 363          * ctxt when doing wakeup
 364          */
 365         closure_get(&ctxt->cl);
 366         bch2_read_extent(trans, &io->rbio,
 367                          bkey_start_pos(k.k),
 368                          btree_id, k, 0,
 369                          BCH_READ_NODECODE|
 370                          BCH_READ_LAST_FRAGMENT);
 371         return 0;
 372 err_free_pages:
 373         bio_free_pages(&io->write.op.wbio.bio);
 374 err_free:
 375         kfree(io);
 376 err:
 377         trace_and_count(c, move_extent_alloc_mem_fail, k.k);
 378         return ret;
 379 }
 380
 381 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 382                         struct bch_inode_unpacked *inode)
 383 {
 384         struct btree_iter iter;
 385         struct bkey_s_c k;
 386         int ret;
 387
 388         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 389                              BTREE_ITER_ALL_SNAPSHOTS);
 390         k = bch2_btree_iter_peek(&iter);
 391         ret = bkey_err(k);
 392         if (ret)
 393                 goto err;
 394
 395         if (!k.k || !bkey_eq(k.k->p, pos)) {
 396                 ret = -ENOENT;
 397                 goto err;
 398         }
 399
 400         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 401         if (ret)
 402                 goto err;
 403
 404         ret = bch2_inode_unpack(k, inode);
 405         if (ret)
 406                 goto err;
 407 err:
 408         bch2_trans_iter_exit(trans, &iter);
 409         return ret;
 410 }
 411
 412 static int move_ratelimit(struct btree_trans *trans,
 413                           struct moving_context *ctxt)
 414 {
 415         struct bch_fs *c = trans->c;
 416         u64 delay;
 417
 418         if (ctxt->wait_on_copygc) {
 419                 bch2_trans_unlock(trans);
 420                 wait_event_killable(c->copygc_running_wq,
 421                                     !c->copygc_running ||
 422                                     kthread_should_stop());
 423         }
 424
 425         do {
 426                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 427
 428                 if (delay) {
 429                         bch2_trans_unlock(trans);
 430                         set_current_state(TASK_INTERRUPTIBLE);
 431                 }
 432
 433                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 434                         __set_current_state(TASK_RUNNING);
 435                         return 1;
 436                 }
 437
 438                 if (delay)
 439                         schedule_timeout(delay);
 440
 441                 if (unlikely(freezing(current))) {
 442                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 443                         try_to_freeze();
 444                 }
 445         } while (delay);
 446
 447         /*
 448          * XXX: these limits really ought to be per device, SSDs and hard drives
 449          * will want different limits
 450          */
 451         move_ctxt_wait_event(ctxt, trans,
 452                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 453                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 454                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 455                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 456
 457         return 0;
 458 }
 459
 460 static int move_get_io_opts(struct btree_trans *trans,
 461                             struct bch_io_opts *io_opts,
 462                             struct bkey_s_c k, u64 *cur_inum)
 463 {
 464         struct bch_inode_unpacked inode;
 465         int ret;
 466
 467         if (*cur_inum == k.k->p.inode)
 468                 return 0;
 469
 470         ret = lookup_inode(trans,
 471                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 472                            &inode);
 473         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 474                 return ret;
 475
 476         if (!ret)
 477                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 478         else
 479                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 480         *cur_inum = k.k->p.inode;
 481         return 0;
 482 }
 483
 484 static int __bch2_move_data(struct moving_context *ctxt,
 485                             struct bpos start,
 486                             struct bpos end,
 487                             move_pred_fn pred, void *arg,
 488                             enum btree_id btree_id)
 489 {
 490         struct bch_fs *c = ctxt->c;
 491         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 492         struct bkey_buf sk;
 493         struct btree_trans trans;
 494         struct btree_iter iter;
 495         struct bkey_s_c k;
 496         struct data_update_opts data_opts;
 497         u64 cur_inum = U64_MAX;
 498         int ret = 0, ret2;
 499
 500         bch2_bkey_buf_init(&sk);
 501         bch2_trans_init(&trans, c, 0, 0);
 502
 503         if (ctxt->stats) {
 504                 ctxt->stats->data_type  = BCH_DATA_user;
 505                 ctxt->stats->btree_id   = btree_id;
 506                 ctxt->stats->pos        = start;
 507         }
 508
 509         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 510                              BTREE_ITER_PREFETCH|
 511                              BTREE_ITER_ALL_SNAPSHOTS);
 512
 513         if (ctxt->rate)
 514                 bch2_ratelimit_reset(ctxt->rate);
 515
 516         while (!move_ratelimit(&trans, ctxt)) {
 517                 bch2_trans_begin(&trans);
 518
 519                 k = bch2_btree_iter_peek(&iter);
 520                 if (!k.k)
 521                         break;
 522
 523                 ret = bkey_err(k);
 524                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 525                         continue;
 526                 if (ret)
 527                         break;
 528
 529                 if (bkey_ge(bkey_start_pos(k.k), end))
 530                         break;
 531
 532                 if (ctxt->stats)
 533                         ctxt->stats->pos = iter.pos;
 534
 535                 if (!bkey_extent_is_direct_data(k.k))
 536                         goto next_nondata;
 537
 538                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 539                 if (ret)
 540                         continue;
 541
 542                 memset(&data_opts, 0, sizeof(data_opts));
 543                 if (!pred(c, arg, k, &io_opts, &data_opts))
 544                         goto next;
 545
 546                 /*
 547                  * The iterator gets unlocked by __bch2_read_extent - need to
 548                  * save a copy of @k elsewhere:
 549                  */
 550                 bch2_bkey_buf_reassemble(&sk, c, k);
 551                 k = bkey_i_to_s_c(sk.k);
 552                 bch2_trans_unlock(&trans);
 553
 554                 ret2 = bch2_move_extent(&trans, &iter, ctxt, NULL,
 555                                         io_opts, btree_id, k, data_opts);
 556                 if (ret2) {
 557                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 558                                 continue;
 559
 560                         if (ret2 == -ENOMEM) {
 561                                 /* memory allocation failure, wait for some IO to finish */
 562                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 563                                 continue;
 564                         }
 565
 566                         /* XXX signal failure */
 567                         goto next;
 568                 }
 569
 570                 if (ctxt->rate)
 571                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 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_trans_exit(&trans);
 581         bch2_bkey_buf_exit(&sk, c);
 582
 583         return ret;
 584 }
 585
 586 int bch2_move_data(struct bch_fs *c,
 587                    enum btree_id start_btree_id, struct bpos start_pos,
 588                    enum btree_id end_btree_id,   struct bpos end_pos,
 589                    struct bch_ratelimit *rate,
 590                    struct bch_move_stats *stats,
 591                    struct write_point_specifier wp,
 592                    bool wait_on_copygc,
 593                    move_pred_fn pred, void *arg)
 594 {
 595         struct moving_context ctxt;
 596         enum btree_id id;
 597         int ret;
 598
 599         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 600
 601         for (id = start_btree_id;
 602              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 603              id++) {
 604                 stats->btree_id = id;
 605
 606                 if (id != BTREE_ID_extents &&
 607                     id != BTREE_ID_reflink)
 608                         continue;
 609
 610                 ret = __bch2_move_data(&ctxt,
 611                                        id == start_btree_id ? start_pos : POS_MIN,
 612                                        id == end_btree_id   ? end_pos   : POS_MAX,
 613                                        pred, arg, id);
 614                 if (ret)
 615                         break;
 616         }
 617
 618         bch2_moving_ctxt_exit(&ctxt);
 619
 620         return ret;
 621 }
 622
 623 void bch2_verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
 624 {
 625         struct bch_fs *c = trans->c;
 626         struct btree_iter iter;
 627         struct bkey_s_c k;
 628         struct printbuf buf = PRINTBUF;
 629         struct bch_backpointer bp;
 630         struct bpos bp_pos = POS_MIN;
 631         unsigned nr_bps = 0;
 632         int ret;
 633
 634         bch2_trans_begin(trans);
 635
 636         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 637                              bucket, BTREE_ITER_CACHED);
 638 again:
 639         ret = lockrestart_do(trans,
 640                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 641
 642         if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
 643                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 644
 645                 if (a.v->gen == gen &&
 646                     a.v->dirty_sectors) {
 647                         if (a.v->data_type == BCH_DATA_btree) {
 648                                 bch2_trans_unlock(trans);
 649                                 if (bch2_btree_interior_updates_flush(c))
 650                                         goto again;
 651                                 goto failed_to_evacuate;
 652                         }
 653                 }
 654         }
 655
 656         set_btree_iter_dontneed(&iter);
 657         bch2_trans_iter_exit(trans, &iter);
 658         return;
 659 failed_to_evacuate:
 660         bch2_trans_iter_exit(trans, &iter);
 661
 662         if (test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
 663                 return;
 664
 665         prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
 666         bch2_bkey_val_to_text(&buf, c, k);
 667
 668         while (1) {
 669                 bch2_trans_begin(trans);
 670
 671                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 672                                                 &bp_pos, &bp,
 673                                                 BTREE_ITER_CACHED);
 674                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 675                         continue;
 676                 if (ret)
 677                         break;
 678                 if (bkey_eq(bp_pos, POS_MAX))
 679                         break;
 680
 681                 k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 682                 ret = bkey_err(k);
 683                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 684                         continue;
 685                 if (ret)
 686                         break;
 687                 if (!k.k)
 688                         continue;
 689                 prt_newline(&buf);
 690                 bch2_bkey_val_to_text(&buf, c, k);
 691                 bch2_trans_iter_exit(trans, &iter);
 692
 693                 if (++nr_bps > 10)
 694                         break;
 695                 bp_pos = bpos_nosnap_successor(bp_pos);
 696         }
 697
 698         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 699         printbuf_exit(&buf);
 700 }
 701
 702 int __bch2_evacuate_bucket(struct btree_trans *trans,
 703                            struct moving_context *ctxt,
 704                            struct move_bucket_in_flight *bucket_in_flight,
 705                            struct bpos bucket, int gen,
 706                            struct data_update_opts _data_opts)
 707 {
 708         struct bch_fs *c = ctxt->c;
 709         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 710         struct btree_iter iter;
 711         struct bkey_buf sk;
 712         struct bch_backpointer bp;
 713         struct bch_alloc_v4 a_convert;
 714         const struct bch_alloc_v4 *a;
 715         struct bkey_s_c k;
 716         struct data_update_opts data_opts;
 717         unsigned dirty_sectors, bucket_size;
 718         u64 fragmentation;
 719         u64 cur_inum = U64_MAX;
 720         struct bpos bp_pos = POS_MIN;
 721         int ret = 0;
 722
 723         bch2_bkey_buf_init(&sk);
 724
 725         /*
 726          * We're not run in a context that handles transaction restarts:
 727          */
 728         bch2_trans_begin(trans);
 729
 730         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 731                              bucket, BTREE_ITER_CACHED);
 732         ret = lockrestart_do(trans,
 733                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 734         bch2_trans_iter_exit(trans, &iter);
 735
 736         if (ret) {
 737                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 738                 goto err;
 739         }
 740
 741         a = bch2_alloc_to_v4(k, &a_convert);
 742         dirty_sectors = a->dirty_sectors;
 743         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 744         fragmentation = a->fragmentation_lru;
 745
 746         ret = bch2_btree_write_buffer_flush(trans);
 747         if (ret) {
 748                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 749                 goto err;
 750         }
 751
 752         while (!(ret = move_ratelimit(trans, ctxt))) {
 753                 bch2_trans_begin(trans);
 754
 755                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 756                                                 &bp_pos, &bp,
 757                                                 BTREE_ITER_CACHED);
 758                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 759                         continue;
 760                 if (ret)
 761                         goto err;
 762                 if (bkey_eq(bp_pos, POS_MAX))
 763                         break;
 764
 765                 if (!bp.level) {
 766                         const struct bch_extent_ptr *ptr;
 767                         struct bkey_s_c k;
 768                         unsigned i = 0;
 769
 770                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 771                         ret = bkey_err(k);
 772                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 773                                 continue;
 774                         if (ret)
 775                                 goto err;
 776                         if (!k.k)
 777                                 goto next;
 778
 779                         bch2_bkey_buf_reassemble(&sk, c, k);
 780                         k = bkey_i_to_s_c(sk.k);
 781
 782                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 783                         if (ret) {
 784                                 bch2_trans_iter_exit(trans, &iter);
 785                                 continue;
 786                         }
 787
 788                         data_opts = _data_opts;
 789                         data_opts.target        = io_opts.background_target;
 790                         data_opts.rewrite_ptrs = 0;
 791
 792                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 793                                 if (ptr->dev == bucket.inode) {
 794                                         data_opts.rewrite_ptrs |= 1U << i;
 795                                         if (ptr->cached) {
 796                                                 bch2_trans_iter_exit(trans, &iter);
 797                                                 goto next;
 798                                         }
 799                                 }
 800                                 i++;
 801                         }
 802
 803                         ret = bch2_move_extent(trans, &iter, ctxt,
 804                                         bucket_in_flight,
 805                                         io_opts, bp.btree_id, k, data_opts);
 806                         bch2_trans_iter_exit(trans, &iter);
 807
 808                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 809                                 continue;
 810                         if (ret == -ENOMEM) {
 811                                 /* memory allocation failure, wait for some IO to finish */
 812                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 813                                 continue;
 814                         }
 815                         if (ret)
 816                                 goto err;
 817
 818                         if (ctxt->rate)
 819                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 820                         if (ctxt->stats)
 821                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 822                 } else {
 823                         struct btree *b;
 824
 825                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 826                         ret = PTR_ERR_OR_ZERO(b);
 827                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 828                                 continue;
 829                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 830                                 continue;
 831                         if (ret)
 832                                 goto err;
 833                         if (!b)
 834                                 goto next;
 835
 836                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 837                         bch2_trans_iter_exit(trans, &iter);
 838
 839                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 840                                 continue;
 841                         if (ret)
 842                                 goto err;
 843
 844                         if (ctxt->rate)
 845                                 bch2_ratelimit_increment(ctxt->rate,
 846                                                          c->opts.btree_node_size >> 9);
 847                         if (ctxt->stats) {
 848                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 849                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 850                         }
 851                 }
 852 next:
 853                 bp_pos = bpos_nosnap_successor(bp_pos);
 854         }
 855
 856         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 857 err:
 858         bch2_bkey_buf_exit(&sk, c);
 859         return ret;
 860 }
 861
 862 int bch2_evacuate_bucket(struct bch_fs *c,
 863                          struct bpos bucket, int gen,
 864                          struct data_update_opts data_opts,
 865                          struct bch_ratelimit *rate,
 866                          struct bch_move_stats *stats,
 867                          struct write_point_specifier wp,
 868                          bool wait_on_copygc)
 869 {
 870         struct btree_trans trans;
 871         struct moving_context ctxt;
 872         int ret;
 873
 874         bch2_trans_init(&trans, c, 0, 0);
 875         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 876         ret = __bch2_evacuate_bucket(&trans, &ctxt, NULL, bucket, gen, data_opts);
 877         bch2_moving_ctxt_exit(&ctxt);
 878         bch2_trans_exit(&trans);
 879
 880         return ret;
 881 }
 882
 883 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 884                                 struct btree *, struct bch_io_opts *,
 885                                 struct data_update_opts *);
 886
 887 static int bch2_move_btree(struct bch_fs *c,
 888                            enum btree_id start_btree_id, struct bpos start_pos,
 889                            enum btree_id end_btree_id,   struct bpos end_pos,
 890                            move_btree_pred pred, void *arg,
 891                            struct bch_move_stats *stats)
 892 {
 893         bool kthread = (current->flags & PF_KTHREAD) != 0;
 894         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 895         struct btree_trans trans;
 896         struct btree_iter iter;
 897         struct btree *b;
 898         enum btree_id id;
 899         struct data_update_opts data_opts;
 900         int ret = 0;
 901
 902         bch2_trans_init(&trans, c, 0, 0);
 903         progress_list_add(c, stats);
 904
 905         stats->data_type = BCH_DATA_btree;
 906
 907         for (id = start_btree_id;
 908              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 909              id++) {
 910                 stats->btree_id = id;
 911
 912                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 913                                           BTREE_ITER_PREFETCH);
 914 retry:
 915                 ret = 0;
 916                 while (bch2_trans_begin(&trans),
 917                        (b = bch2_btree_iter_peek_node(&iter)) &&
 918                        !(ret = PTR_ERR_OR_ZERO(b))) {
 919                         if (kthread && kthread_should_stop())
 920                                 break;
 921
 922                         if ((cmp_int(id, end_btree_id) ?:
 923                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 924                                 break;
 925
 926                         stats->pos = iter.pos;
 927
 928                         if (!pred(c, arg, b, &io_opts, &data_opts))
 929                                 goto next;
 930
 931                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 932                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 933                                 continue;
 934                         if (ret)
 935                                 break;
 936 next:
 937                         bch2_btree_iter_next_node(&iter);
 938                 }
 939                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 940                         goto retry;
 941
 942                 bch2_trans_iter_exit(&trans, &iter);
 943
 944                 if (kthread && kthread_should_stop())
 945                         break;
 946         }
 947
 948         bch2_trans_exit(&trans);
 949
 950         if (ret)
 951                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 952
 953         bch2_btree_interior_updates_flush(c);
 954
 955         progress_list_del(c, stats);
 956         return ret;
 957 }
 958
 959 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 960                              struct bkey_s_c k,
 961                              struct bch_io_opts *io_opts,
 962                              struct data_update_opts *data_opts)
 963 {
 964         unsigned nr_good = bch2_bkey_durability(c, k);
 965         unsigned replicas = bkey_is_btree_ptr(k.k)
 966                 ? c->opts.metadata_replicas
 967                 : io_opts->data_replicas;
 968
 969         if (!nr_good || nr_good >= replicas)
 970                 return false;
 971
 972         data_opts->target               = 0;
 973         data_opts->extra_replicas       = replicas - nr_good;
 974         data_opts->btree_insert_flags   = 0;
 975         return true;
 976 }
 977
 978 static bool migrate_pred(struct bch_fs *c, void *arg,
 979                          struct bkey_s_c k,
 980                          struct bch_io_opts *io_opts,
 981                          struct data_update_opts *data_opts)
 982 {
 983         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 984         const struct bch_extent_ptr *ptr;
 985         struct bch_ioctl_data *op = arg;
 986         unsigned i = 0;
 987
 988         data_opts->rewrite_ptrs         = 0;
 989         data_opts->target               = 0;
 990         data_opts->extra_replicas       = 0;
 991         data_opts->btree_insert_flags   = 0;
 992
 993         bkey_for_each_ptr(ptrs, ptr) {
 994                 if (ptr->dev == op->migrate.dev)
 995                         data_opts->rewrite_ptrs |= 1U << i;
 996                 i++;
 997         }
 998
 999         return data_opts->rewrite_ptrs != 0;
1000 }
1001
1002 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
1003                                    struct btree *b,
1004                                    struct bch_io_opts *io_opts,
1005                                    struct data_update_opts *data_opts)
1006 {
1007         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1008 }
1009
1010 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
1011                                struct btree *b,
1012                                struct bch_io_opts *io_opts,
1013                                struct data_update_opts *data_opts)
1014 {
1015         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1016 }
1017
1018 static bool bformat_needs_redo(struct bkey_format *f)
1019 {
1020         unsigned i;
1021
1022         for (i = 0; i < f->nr_fields; i++) {
1023                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
1024                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
1025                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
1026
1027                 if (f->bits_per_field[i] > unpacked_bits)
1028                         return true;
1029
1030                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1031                         return true;
1032
1033                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1034                      unpacked_mask) <
1035                     field_offset)
1036                         return true;
1037         }
1038
1039         return false;
1040 }
1041
1042 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1043                                    struct btree *b,
1044                                    struct bch_io_opts *io_opts,
1045                                    struct data_update_opts *data_opts)
1046 {
1047         if (b->version_ondisk != c->sb.version ||
1048             btree_node_need_rewrite(b) ||
1049             bformat_needs_redo(&b->format)) {
1050                 data_opts->target               = 0;
1051                 data_opts->extra_replicas       = 0;
1052                 data_opts->btree_insert_flags   = 0;
1053                 return true;
1054         }
1055
1056         return false;
1057 }
1058
1059 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1060 {
1061         int ret;
1062
1063         ret = bch2_move_btree(c,
1064                               0,                POS_MIN,
1065                               BTREE_ID_NR,      SPOS_MAX,
1066                               rewrite_old_nodes_pred, c, stats);
1067         if (!ret) {
1068                 mutex_lock(&c->sb_lock);
1069                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1070                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1071                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1072                 bch2_write_super(c);
1073                 mutex_unlock(&c->sb_lock);
1074         }
1075
1076         return ret;
1077 }
1078
1079 int bch2_data_job(struct bch_fs *c,
1080                   struct bch_move_stats *stats,
1081                   struct bch_ioctl_data op)
1082 {
1083         int ret = 0;
1084
1085         switch (op.op) {
1086         case BCH_DATA_OP_REREPLICATE:
1087                 bch2_move_stats_init(stats, "rereplicate");
1088                 stats->data_type = BCH_DATA_journal;
1089                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1090
1091                 ret = bch2_move_btree(c,
1092                                       op.start_btree,   op.start_pos,
1093                                       op.end_btree,     op.end_pos,
1094                                       rereplicate_btree_pred, c, stats) ?: ret;
1095                 ret = bch2_replicas_gc2(c) ?: ret;
1096
1097                 ret = bch2_move_data(c,
1098                                      op.start_btree,    op.start_pos,
1099                                      op.end_btree,      op.end_pos,
1100                                      NULL,
1101                                      stats,
1102                                      writepoint_hashed((unsigned long) current),
1103                                      true,
1104                                      rereplicate_pred, c) ?: ret;
1105                 ret = bch2_replicas_gc2(c) ?: ret;
1106                 break;
1107         case BCH_DATA_OP_MIGRATE:
1108                 if (op.migrate.dev >= c->sb.nr_devices)
1109                         return -EINVAL;
1110
1111                 bch2_move_stats_init(stats, "migrate");
1112                 stats->data_type = BCH_DATA_journal;
1113                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1114
1115                 ret = bch2_move_btree(c,
1116                                       op.start_btree,   op.start_pos,
1117                                       op.end_btree,     op.end_pos,
1118                                       migrate_btree_pred, &op, stats) ?: ret;
1119                 ret = bch2_replicas_gc2(c) ?: ret;
1120
1121                 ret = bch2_move_data(c,
1122                                      op.start_btree,    op.start_pos,
1123                                      op.end_btree,      op.end_pos,
1124                                      NULL,
1125                                      stats,
1126                                      writepoint_hashed((unsigned long) current),
1127                                      true,
1128                                      migrate_pred, &op) ?: ret;
1129                 ret = bch2_replicas_gc2(c) ?: ret;
1130                 break;
1131         case BCH_DATA_OP_REWRITE_OLD_NODES:
1132                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1133                 ret = bch2_scan_old_btree_nodes(c, stats);
1134                 break;
1135         default:
1136                 ret = -EINVAL;
1137         }
1138
1139         return ret;
1140 }
1141
1142 void bch2_data_jobs_to_text(struct printbuf *out, struct bch_fs *c)
1143 {
1144         struct bch_move_stats *stats;
1145
1146         mutex_lock(&c->data_progress_lock);
1147         list_for_each_entry(stats, &c->data_progress_list, list) {
1148                 prt_printf(out, "%s: data type %s btree_id %s position: ",
1149                        stats->name,
1150                        bch2_data_types[stats->data_type],
1151                        bch2_btree_ids[stats->btree_id]);
1152                 bch2_bpos_to_text(out, stats->pos);
1153                 prt_printf(out, "%s", "\n");
1154         }
1155         mutex_unlock(&c->data_progress_lock);
1156 }
1157
1158 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct moving_context *ctxt)
1159 {
1160         struct moving_io *io;
1161
1162         prt_printf(out, "%ps:", ctxt->fn);
1163         prt_newline(out);
1164         printbuf_indent_add(out, 2);
1165
1166         prt_printf(out, "reads: %u sectors %u",
1167                    atomic_read(&ctxt->read_ios),
1168                    atomic_read(&ctxt->read_sectors));
1169         prt_newline(out);
1170
1171         prt_printf(out, "writes: %u sectors %u",
1172                    atomic_read(&ctxt->write_ios),
1173                    atomic_read(&ctxt->write_sectors));
1174         prt_newline(out);
1175
1176         printbuf_indent_add(out, 2);
1177
1178         mutex_lock(&ctxt->lock);
1179         list_for_each_entry(io, &ctxt->ios, io_list) {
1180                 bch2_write_op_to_text(out, &io->write.op);
1181         }
1182         mutex_unlock(&ctxt->lock);
1183
1184         printbuf_indent_sub(out, 4);
1185 }
1186
1187 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1188 {
1189         struct moving_context *ctxt;
1190
1191         mutex_lock(&c->moving_context_lock);
1192         list_for_each_entry(ctxt, &c->moving_context_list, list)
1193                 bch2_moving_ctxt_to_text(out, ctxt);
1194         mutex_unlock(&c->moving_context_lock);
1195 }
1196
1197 void bch2_fs_move_init(struct bch_fs *c)
1198 {
1199         INIT_LIST_HEAD(&c->moving_context_list);
1200         mutex_init(&c->moving_context_lock);
1201
1202         INIT_LIST_HEAD(&c->data_progress_list);
1203         mutex_init(&c->data_progress_lock);
1204 }