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 trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
  30 {
  31         if (trace_move_extent_enabled()) {
  32                 struct printbuf buf = PRINTBUF;
  33
  34                 bch2_bkey_val_to_text(&buf, c, k);
  35                 trace_move_extent(c, buf.buf);
  36                 printbuf_exit(&buf);
  37         }
  38 }
  39
  40 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
  41 {
  42         if (trace_move_extent_read_enabled()) {
  43                 struct printbuf buf = PRINTBUF;
  44
  45                 bch2_bkey_val_to_text(&buf, c, k);
  46                 trace_move_extent_read(c, buf.buf);
  47                 printbuf_exit(&buf);
  48         }
  49 }
  50
  51 static void trace_move_extent_alloc_mem_fail2(struct bch_fs *c, struct bkey_s_c k)
  52 {
  53         if (trace_move_extent_alloc_mem_fail_enabled()) {
  54                 struct printbuf buf = PRINTBUF;
  55
  56                 bch2_bkey_val_to_text(&buf, c, k);
  57                 trace_move_extent_alloc_mem_fail(c, buf.buf);
  58                 printbuf_exit(&buf);
  59         }
  60 }
  61
  62 static void progress_list_add(struct bch_fs *c, struct bch_move_stats *stats)
  63 {
  64         mutex_lock(&c->data_progress_lock);
  65         list_add(&stats->list, &c->data_progress_list);
  66         mutex_unlock(&c->data_progress_lock);
  67 }
  68
  69 static void progress_list_del(struct bch_fs *c, struct bch_move_stats *stats)
  70 {
  71         mutex_lock(&c->data_progress_lock);
  72         list_del(&stats->list);
  73         mutex_unlock(&c->data_progress_lock);
  74 }
  75
  76 struct moving_io {
  77         struct list_head                read_list;
  78         struct list_head                io_list;
  79         struct move_bucket_in_flight    *b;
  80         struct closure                  cl;
  81         bool                            read_completed;
  82
  83         unsigned                        read_sectors;
  84         unsigned                        write_sectors;
  85
  86         struct bch_read_bio             rbio;
  87
  88         struct data_update              write;
  89         /* Must be last since it is variable size */
  90         struct bio_vec                  bi_inline_vecs[0];
  91 };
  92
  93 static void move_free(struct moving_io *io)
  94 {
  95         struct moving_context *ctxt = io->write.ctxt;
  96
  97         if (io->b)
  98                 atomic_dec(&io->b->count);
  99
 100         bch2_data_update_exit(&io->write);
 101
 102         mutex_lock(&ctxt->lock);
 103         list_del(&io->io_list);
 104         wake_up(&ctxt->wait);
 105         mutex_unlock(&ctxt->lock);
 106
 107         kfree(io);
 108 }
 109
 110 static void move_write_done(struct bch_write_op *op)
 111 {
 112         struct moving_io *io = container_of(op, struct moving_io, write.op);
 113         struct moving_context *ctxt = io->write.ctxt;
 114
 115         if (io->write.op.error)
 116                 ctxt->write_error = true;
 117
 118         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 119         atomic_dec(&io->write.ctxt->write_ios);
 120         move_free(io);
 121         closure_put(&ctxt->cl);
 122 }
 123
 124 static void move_write(struct moving_io *io)
 125 {
 126         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 127                 move_free(io);
 128                 return;
 129         }
 130
 131         closure_get(&io->write.ctxt->cl);
 132         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 133         atomic_inc(&io->write.ctxt->write_ios);
 134
 135         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 136 }
 137
 138 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 139 {
 140         struct moving_io *io =
 141                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 142
 143         return io && io->read_completed ? io : NULL;
 144 }
 145
 146 static void move_read_endio(struct bio *bio)
 147 {
 148         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 149         struct moving_context *ctxt = io->write.ctxt;
 150
 151         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 152         atomic_dec(&ctxt->read_ios);
 153         io->read_completed = true;
 154
 155         wake_up(&ctxt->wait);
 156         closure_put(&ctxt->cl);
 157 }
 158
 159 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
 160                                         struct btree_trans *trans)
 161 {
 162         struct moving_io *io;
 163
 164         if (trans)
 165                 bch2_trans_unlock(trans);
 166
 167         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 168                 list_del(&io->read_list);
 169                 move_write(io);
 170         }
 171 }
 172
 173 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
 174                                        struct btree_trans *trans)
 175 {
 176         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 177
 178         move_ctxt_wait_event(ctxt, trans,
 179                 !atomic_read(&ctxt->write_sectors) ||
 180                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 181 }
 182
 183 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 184 {
 185         struct bch_fs *c = ctxt->c;
 186
 187         move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 188         closure_sync(&ctxt->cl);
 189
 190         EBUG_ON(atomic_read(&ctxt->write_sectors));
 191         EBUG_ON(atomic_read(&ctxt->write_ios));
 192         EBUG_ON(atomic_read(&ctxt->read_sectors));
 193         EBUG_ON(atomic_read(&ctxt->read_ios));
 194
 195         if (ctxt->stats) {
 196                 progress_list_del(c, ctxt->stats);
 197                 trace_move_data(c,
 198                                 atomic64_read(&ctxt->stats->sectors_moved),
 199                                 atomic64_read(&ctxt->stats->keys_moved));
 200         }
 201
 202         mutex_lock(&c->moving_context_lock);
 203         list_del(&ctxt->list);
 204         mutex_unlock(&c->moving_context_lock);
 205 }
 206
 207 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 208                            struct bch_fs *c,
 209                            struct bch_ratelimit *rate,
 210                            struct bch_move_stats *stats,
 211                            struct write_point_specifier wp,
 212                            bool wait_on_copygc)
 213 {
 214         memset(ctxt, 0, sizeof(*ctxt));
 215
 216         ctxt->c         = c;
 217         ctxt->fn        = (void *) _RET_IP_;
 218         ctxt->rate      = rate;
 219         ctxt->stats     = stats;
 220         ctxt->wp        = wp;
 221         ctxt->wait_on_copygc = wait_on_copygc;
 222
 223         closure_init_stack(&ctxt->cl);
 224
 225         mutex_init(&ctxt->lock);
 226         INIT_LIST_HEAD(&ctxt->reads);
 227         INIT_LIST_HEAD(&ctxt->ios);
 228         init_waitqueue_head(&ctxt->wait);
 229
 230         mutex_lock(&c->moving_context_lock);
 231         list_add(&ctxt->list, &c->moving_context_list);
 232         mutex_unlock(&c->moving_context_lock);
 233
 234         if (stats) {
 235                 progress_list_add(c, stats);
 236                 stats->data_type = BCH_DATA_user;
 237         }
 238 }
 239
 240 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 241 {
 242         memset(stats, 0, sizeof(*stats));
 243         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 244 }
 245
 246 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 247                                  struct btree_iter *iter,
 248                                  struct bkey_s_c k,
 249                                  struct data_update_opts data_opts)
 250 {
 251         struct bch_fs *c = trans->c;
 252         struct bkey_i *n;
 253         int ret;
 254
 255         n = bch2_bkey_make_mut(trans, k);
 256         ret = PTR_ERR_OR_ZERO(n);
 257         if (ret)
 258                 return ret;
 259
 260         while (data_opts.kill_ptrs) {
 261                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 262                 struct bch_extent_ptr *ptr;
 263
 264                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 265                 data_opts.kill_ptrs ^= 1U << drop;
 266         }
 267
 268         /*
 269          * If the new extent no longer has any pointers, bch2_extent_normalize()
 270          * will do the appropriate thing with it (turning it into a
 271          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 272          */
 273         bch2_extent_normalize(c, bkey_i_to_s(n));
 274
 275         /*
 276          * Since we're not inserting through an extent iterator
 277          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 278          * we aren't using the extent overwrite path to delete, we're
 279          * just using the normal key deletion path:
 280          */
 281         if (bkey_deleted(&n->k))
 282                 n->k.size = 0;
 283
 284         return bch2_trans_relock(trans) ?:
 285                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 286                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 287 }
 288
 289 static int bch2_move_extent(struct btree_trans *trans,
 290                             struct btree_iter *iter,
 291                             struct moving_context *ctxt,
 292                             struct move_bucket_in_flight *bucket_in_flight,
 293                             struct bch_io_opts io_opts,
 294                             enum btree_id btree_id,
 295                             struct bkey_s_c k,
 296                             struct data_update_opts data_opts)
 297 {
 298         struct bch_fs *c = trans->c;
 299         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 300         struct moving_io *io;
 301         const union bch_extent_entry *entry;
 302         struct extent_ptr_decoded p;
 303         unsigned sectors = k.k->size, pages;
 304         int ret = -ENOMEM;
 305
 306         trace_move_extent2(c, k);
 307
 308         bch2_data_update_opts_normalize(k, &data_opts);
 309
 310         if (!data_opts.rewrite_ptrs &&
 311             !data_opts.extra_replicas) {
 312                 if (data_opts.kill_ptrs)
 313                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 314                 return 0;
 315         }
 316
 317         /*
 318          * Before memory allocations & taking nocow locks in
 319          * bch2_data_update_init():
 320          */
 321         bch2_trans_unlock(trans);
 322
 323         /* write path might have to decompress data: */
 324         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 325                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 326
 327         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 328         io = kzalloc(sizeof(struct moving_io) +
 329                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 330         if (!io)
 331                 goto err;
 332
 333         INIT_LIST_HEAD(&io->io_list);
 334         io->write.ctxt          = ctxt;
 335         io->read_sectors        = k.k->size;
 336         io->write_sectors       = k.k->size;
 337
 338         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 339         bio_set_prio(&io->write.op.wbio.bio,
 340                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 341
 342         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 343                                  GFP_KERNEL))
 344                 goto err_free;
 345
 346         io->rbio.c              = c;
 347         io->rbio.opts           = io_opts;
 348         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 349         io->rbio.bio.bi_vcnt = pages;
 350         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 351         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 352
 353         io->rbio.bio.bi_opf             = REQ_OP_READ;
 354         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 355         io->rbio.bio.bi_end_io          = move_read_endio;
 356
 357         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 358                                     io_opts, data_opts, btree_id, k);
 359         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 360                 goto err_free_pages;
 361
 362         if (ret == -BCH_ERR_unwritten_extent_update) {
 363                 bch2_update_unwritten_extent(trans, &io->write);
 364                 move_free(io);
 365                 return 0;
 366         }
 367
 368         BUG_ON(ret);
 369
 370         io->write.ctxt = ctxt;
 371         io->write.op.end_io = move_write_done;
 372
 373         if (ctxt->stats) {
 374                 atomic64_inc(&ctxt->stats->keys_moved);
 375                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 376         }
 377
 378         if (bucket_in_flight) {
 379                 io->b = bucket_in_flight;
 380                 atomic_inc(&io->b->count);
 381         }
 382
 383         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 384         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 385         trace_move_extent_read2(c, k);
 386
 387         mutex_lock(&ctxt->lock);
 388         atomic_add(io->read_sectors, &ctxt->read_sectors);
 389         atomic_inc(&ctxt->read_ios);
 390
 391         list_add_tail(&io->read_list, &ctxt->reads);
 392         list_add_tail(&io->io_list, &ctxt->ios);
 393         mutex_unlock(&ctxt->lock);
 394
 395         /*
 396          * dropped by move_read_endio() - guards against use after free of
 397          * ctxt when doing wakeup
 398          */
 399         closure_get(&ctxt->cl);
 400         bch2_read_extent(trans, &io->rbio,
 401                          bkey_start_pos(k.k),
 402                          btree_id, k, 0,
 403                          BCH_READ_NODECODE|
 404                          BCH_READ_LAST_FRAGMENT);
 405         return 0;
 406 err_free_pages:
 407         bio_free_pages(&io->write.op.wbio.bio);
 408 err_free:
 409         kfree(io);
 410 err:
 411         this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
 412         trace_move_extent_alloc_mem_fail2(c, k);
 413         return ret;
 414 }
 415
 416 static int lookup_inode(struct btree_trans *trans, struct bpos pos,
 417                         struct bch_inode_unpacked *inode)
 418 {
 419         struct btree_iter iter;
 420         struct bkey_s_c k;
 421         int ret;
 422
 423         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes, pos,
 424                              BTREE_ITER_ALL_SNAPSHOTS);
 425         k = bch2_btree_iter_peek(&iter);
 426         ret = bkey_err(k);
 427         if (ret)
 428                 goto err;
 429
 430         if (!k.k || !bkey_eq(k.k->p, pos)) {
 431                 ret = -ENOENT;
 432                 goto err;
 433         }
 434
 435         ret = bkey_is_inode(k.k) ? 0 : -EIO;
 436         if (ret)
 437                 goto err;
 438
 439         ret = bch2_inode_unpack(k, inode);
 440         if (ret)
 441                 goto err;
 442 err:
 443         bch2_trans_iter_exit(trans, &iter);
 444         return ret;
 445 }
 446
 447 static int move_ratelimit(struct btree_trans *trans,
 448                           struct moving_context *ctxt)
 449 {
 450         struct bch_fs *c = trans->c;
 451         u64 delay;
 452
 453         if (ctxt->wait_on_copygc) {
 454                 bch2_trans_unlock(trans);
 455                 wait_event_killable(c->copygc_running_wq,
 456                                     !c->copygc_running ||
 457                                     kthread_should_stop());
 458         }
 459
 460         do {
 461                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 462
 463                 if (delay) {
 464                         bch2_trans_unlock(trans);
 465                         set_current_state(TASK_INTERRUPTIBLE);
 466                 }
 467
 468                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 469                         __set_current_state(TASK_RUNNING);
 470                         return 1;
 471                 }
 472
 473                 if (delay)
 474                         schedule_timeout(delay);
 475
 476                 if (unlikely(freezing(current))) {
 477                         move_ctxt_wait_event(ctxt, trans, list_empty(&ctxt->reads));
 478                         try_to_freeze();
 479                 }
 480         } while (delay);
 481
 482         /*
 483          * XXX: these limits really ought to be per device, SSDs and hard drives
 484          * will want different limits
 485          */
 486         move_ctxt_wait_event(ctxt, trans,
 487                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 488                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 489                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 490                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 491
 492         return 0;
 493 }
 494
 495 static int move_get_io_opts(struct btree_trans *trans,
 496                             struct bch_io_opts *io_opts,
 497                             struct bkey_s_c k, u64 *cur_inum)
 498 {
 499         struct bch_inode_unpacked inode;
 500         int ret;
 501
 502         if (*cur_inum == k.k->p.inode)
 503                 return 0;
 504
 505         ret = lookup_inode(trans,
 506                            SPOS(0, k.k->p.inode, k.k->p.snapshot),
 507                            &inode);
 508         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 509                 return ret;
 510
 511         if (!ret)
 512                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 513         else
 514                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 515         *cur_inum = k.k->p.inode;
 516         return 0;
 517 }
 518
 519 static int __bch2_move_data(struct moving_context *ctxt,
 520                             struct bpos start,
 521                             struct bpos end,
 522                             move_pred_fn pred, void *arg,
 523                             enum btree_id btree_id)
 524 {
 525         struct bch_fs *c = ctxt->c;
 526         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 527         struct bkey_buf sk;
 528         struct btree_trans trans;
 529         struct btree_iter iter;
 530         struct bkey_s_c k;
 531         struct data_update_opts data_opts;
 532         u64 cur_inum = U64_MAX;
 533         int ret = 0, ret2;
 534
 535         bch2_bkey_buf_init(&sk);
 536         bch2_trans_init(&trans, c, 0, 0);
 537
 538         if (ctxt->stats) {
 539                 ctxt->stats->data_type  = BCH_DATA_user;
 540                 ctxt->stats->btree_id   = btree_id;
 541                 ctxt->stats->pos        = start;
 542         }
 543
 544         bch2_trans_iter_init(&trans, &iter, btree_id, start,
 545                              BTREE_ITER_PREFETCH|
 546                              BTREE_ITER_ALL_SNAPSHOTS);
 547
 548         if (ctxt->rate)
 549                 bch2_ratelimit_reset(ctxt->rate);
 550
 551         while (!move_ratelimit(&trans, ctxt)) {
 552                 bch2_trans_begin(&trans);
 553
 554                 k = bch2_btree_iter_peek(&iter);
 555                 if (!k.k)
 556                         break;
 557
 558                 ret = bkey_err(k);
 559                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 560                         continue;
 561                 if (ret)
 562                         break;
 563
 564                 if (bkey_ge(bkey_start_pos(k.k), end))
 565                         break;
 566
 567                 if (ctxt->stats)
 568                         ctxt->stats->pos = iter.pos;
 569
 570                 if (!bkey_extent_is_direct_data(k.k))
 571                         goto next_nondata;
 572
 573                 ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
 574                 if (ret)
 575                         continue;
 576
 577                 memset(&data_opts, 0, sizeof(data_opts));
 578                 if (!pred(c, arg, k, &io_opts, &data_opts))
 579                         goto next;
 580
 581                 /*
 582                  * The iterator gets unlocked by __bch2_read_extent - need to
 583                  * save a copy of @k elsewhere:
 584                  */
 585                 bch2_bkey_buf_reassemble(&sk, c, k);
 586                 k = bkey_i_to_s_c(sk.k);
 587                 bch2_trans_unlock(&trans);
 588
 589                 ret2 = bch2_move_extent(&trans, &iter, ctxt, NULL,
 590                                         io_opts, btree_id, k, data_opts);
 591                 if (ret2) {
 592                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 593                                 continue;
 594
 595                         if (ret2 == -ENOMEM) {
 596                                 /* memory allocation failure, wait for some IO to finish */
 597                                 bch2_move_ctxt_wait_for_io(ctxt, &trans);
 598                                 continue;
 599                         }
 600
 601                         /* XXX signal failure */
 602                         goto next;
 603                 }
 604
 605                 if (ctxt->rate)
 606                         bch2_ratelimit_increment(ctxt->rate, k.k->size);
 607 next:
 608                 if (ctxt->stats)
 609                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 610 next_nondata:
 611                 bch2_btree_iter_advance(&iter);
 612         }
 613
 614         bch2_trans_iter_exit(&trans, &iter);
 615         bch2_trans_exit(&trans);
 616         bch2_bkey_buf_exit(&sk, c);
 617
 618         return ret;
 619 }
 620
 621 int bch2_move_data(struct bch_fs *c,
 622                    enum btree_id start_btree_id, struct bpos start_pos,
 623                    enum btree_id end_btree_id,   struct bpos end_pos,
 624                    struct bch_ratelimit *rate,
 625                    struct bch_move_stats *stats,
 626                    struct write_point_specifier wp,
 627                    bool wait_on_copygc,
 628                    move_pred_fn pred, void *arg)
 629 {
 630         struct moving_context ctxt;
 631         enum btree_id id;
 632         int ret;
 633
 634         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 635
 636         for (id = start_btree_id;
 637              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 638              id++) {
 639                 stats->btree_id = id;
 640
 641                 if (id != BTREE_ID_extents &&
 642                     id != BTREE_ID_reflink)
 643                         continue;
 644
 645                 ret = __bch2_move_data(&ctxt,
 646                                        id == start_btree_id ? start_pos : POS_MIN,
 647                                        id == end_btree_id   ? end_pos   : POS_MAX,
 648                                        pred, arg, id);
 649                 if (ret)
 650                         break;
 651         }
 652
 653         bch2_moving_ctxt_exit(&ctxt);
 654
 655         return ret;
 656 }
 657
 658 int __bch2_evacuate_bucket(struct btree_trans *trans,
 659                            struct moving_context *ctxt,
 660                            struct move_bucket_in_flight *bucket_in_flight,
 661                            struct bpos bucket, int gen,
 662                            struct data_update_opts _data_opts)
 663 {
 664         struct bch_fs *c = ctxt->c;
 665         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 666         struct btree_iter iter;
 667         struct bkey_buf sk;
 668         struct bch_backpointer bp;
 669         struct bch_alloc_v4 a_convert;
 670         const struct bch_alloc_v4 *a;
 671         struct bkey_s_c k;
 672         struct data_update_opts data_opts;
 673         unsigned dirty_sectors, bucket_size;
 674         u64 fragmentation;
 675         u64 cur_inum = U64_MAX;
 676         struct bpos bp_pos = POS_MIN;
 677         int ret = 0;
 678
 679         trace_bucket_evacuate(c, bucket);
 680
 681         bch2_bkey_buf_init(&sk);
 682
 683         /*
 684          * We're not run in a context that handles transaction restarts:
 685          */
 686         bch2_trans_begin(trans);
 687
 688         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 689                              bucket, BTREE_ITER_CACHED);
 690         ret = lockrestart_do(trans,
 691                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 692         bch2_trans_iter_exit(trans, &iter);
 693
 694         if (ret) {
 695                 bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
 696                 goto err;
 697         }
 698
 699         a = bch2_alloc_to_v4(k, &a_convert);
 700         dirty_sectors = a->dirty_sectors;
 701         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 702         fragmentation = a->fragmentation_lru;
 703
 704         ret = bch2_btree_write_buffer_flush(trans);
 705         if (ret) {
 706                 bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 707                 goto err;
 708         }
 709
 710         while (!(ret = move_ratelimit(trans, ctxt))) {
 711                 bch2_trans_begin(trans);
 712
 713                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 714                                                 &bp_pos, &bp,
 715                                                 BTREE_ITER_CACHED);
 716                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 717                         continue;
 718                 if (ret)
 719                         goto err;
 720                 if (bkey_eq(bp_pos, POS_MAX))
 721                         break;
 722
 723                 if (!bp.level) {
 724                         const struct bch_extent_ptr *ptr;
 725                         struct bkey_s_c k;
 726                         unsigned i = 0;
 727
 728                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 729                         ret = bkey_err(k);
 730                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 731                                 continue;
 732                         if (ret)
 733                                 goto err;
 734                         if (!k.k)
 735                                 goto next;
 736
 737                         bch2_bkey_buf_reassemble(&sk, c, k);
 738                         k = bkey_i_to_s_c(sk.k);
 739
 740                         ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 741                         if (ret) {
 742                                 bch2_trans_iter_exit(trans, &iter);
 743                                 continue;
 744                         }
 745
 746                         data_opts = _data_opts;
 747                         data_opts.target        = io_opts.background_target;
 748                         data_opts.rewrite_ptrs = 0;
 749
 750                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 751                                 if (ptr->dev == bucket.inode) {
 752                                         data_opts.rewrite_ptrs |= 1U << i;
 753                                         if (ptr->cached) {
 754                                                 bch2_trans_iter_exit(trans, &iter);
 755                                                 goto next;
 756                                         }
 757                                 }
 758                                 i++;
 759                         }
 760
 761                         ret = bch2_move_extent(trans, &iter, ctxt,
 762                                         bucket_in_flight,
 763                                         io_opts, bp.btree_id, k, data_opts);
 764                         bch2_trans_iter_exit(trans, &iter);
 765
 766                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 767                                 continue;
 768                         if (ret == -ENOMEM) {
 769                                 /* memory allocation failure, wait for some IO to finish */
 770                                 bch2_move_ctxt_wait_for_io(ctxt, trans);
 771                                 continue;
 772                         }
 773                         if (ret)
 774                                 goto err;
 775
 776                         if (ctxt->rate)
 777                                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 778                         if (ctxt->stats)
 779                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 780                 } else {
 781                         struct btree *b;
 782
 783                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 784                         ret = PTR_ERR_OR_ZERO(b);
 785                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 786                                 continue;
 787                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 788                                 continue;
 789                         if (ret)
 790                                 goto err;
 791                         if (!b)
 792                                 goto next;
 793
 794                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 795                         bch2_trans_iter_exit(trans, &iter);
 796
 797                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 798                                 continue;
 799                         if (ret)
 800                                 goto err;
 801
 802                         if (ctxt->rate)
 803                                 bch2_ratelimit_increment(ctxt->rate,
 804                                                          c->opts.btree_node_size >> 9);
 805                         if (ctxt->stats) {
 806                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 807                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 808                         }
 809                 }
 810 next:
 811                 bp_pos = bpos_nosnap_successor(bp_pos);
 812         }
 813
 814         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 815 err:
 816         bch2_bkey_buf_exit(&sk, c);
 817         return ret;
 818 }
 819
 820 int bch2_evacuate_bucket(struct bch_fs *c,
 821                          struct bpos bucket, int gen,
 822                          struct data_update_opts data_opts,
 823                          struct bch_ratelimit *rate,
 824                          struct bch_move_stats *stats,
 825                          struct write_point_specifier wp,
 826                          bool wait_on_copygc)
 827 {
 828         struct btree_trans trans;
 829         struct moving_context ctxt;
 830         int ret;
 831
 832         bch2_trans_init(&trans, c, 0, 0);
 833         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 834         ret = __bch2_evacuate_bucket(&trans, &ctxt, NULL, bucket, gen, data_opts);
 835         bch2_moving_ctxt_exit(&ctxt);
 836         bch2_trans_exit(&trans);
 837
 838         return ret;
 839 }
 840
 841 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 842                                 struct btree *, struct bch_io_opts *,
 843                                 struct data_update_opts *);
 844
 845 static int bch2_move_btree(struct bch_fs *c,
 846                            enum btree_id start_btree_id, struct bpos start_pos,
 847                            enum btree_id end_btree_id,   struct bpos end_pos,
 848                            move_btree_pred pred, void *arg,
 849                            struct bch_move_stats *stats)
 850 {
 851         bool kthread = (current->flags & PF_KTHREAD) != 0;
 852         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 853         struct btree_trans trans;
 854         struct btree_iter iter;
 855         struct btree *b;
 856         enum btree_id id;
 857         struct data_update_opts data_opts;
 858         int ret = 0;
 859
 860         bch2_trans_init(&trans, c, 0, 0);
 861         progress_list_add(c, stats);
 862
 863         stats->data_type = BCH_DATA_btree;
 864
 865         for (id = start_btree_id;
 866              id <= min_t(unsigned, end_btree_id, BTREE_ID_NR - 1);
 867              id++) {
 868                 stats->btree_id = id;
 869
 870                 bch2_trans_node_iter_init(&trans, &iter, id, POS_MIN, 0, 0,
 871                                           BTREE_ITER_PREFETCH);
 872 retry:
 873                 ret = 0;
 874                 while (bch2_trans_begin(&trans),
 875                        (b = bch2_btree_iter_peek_node(&iter)) &&
 876                        !(ret = PTR_ERR_OR_ZERO(b))) {
 877                         if (kthread && kthread_should_stop())
 878                                 break;
 879
 880                         if ((cmp_int(id, end_btree_id) ?:
 881                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 882                                 break;
 883
 884                         stats->pos = iter.pos;
 885
 886                         if (!pred(c, arg, b, &io_opts, &data_opts))
 887                                 goto next;
 888
 889                         ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
 890                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 891                                 continue;
 892                         if (ret)
 893                                 break;
 894 next:
 895                         bch2_btree_iter_next_node(&iter);
 896                 }
 897                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 898                         goto retry;
 899
 900                 bch2_trans_iter_exit(&trans, &iter);
 901
 902                 if (kthread && kthread_should_stop())
 903                         break;
 904         }
 905
 906         bch2_trans_exit(&trans);
 907
 908         if (ret)
 909                 bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
 910
 911         bch2_btree_interior_updates_flush(c);
 912
 913         progress_list_del(c, stats);
 914         return ret;
 915 }
 916
 917 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 918                              struct bkey_s_c k,
 919                              struct bch_io_opts *io_opts,
 920                              struct data_update_opts *data_opts)
 921 {
 922         unsigned nr_good = bch2_bkey_durability(c, k);
 923         unsigned replicas = bkey_is_btree_ptr(k.k)
 924                 ? c->opts.metadata_replicas
 925                 : io_opts->data_replicas;
 926
 927         if (!nr_good || nr_good >= replicas)
 928                 return false;
 929
 930         data_opts->target               = 0;
 931         data_opts->extra_replicas       = replicas - nr_good;
 932         data_opts->btree_insert_flags   = 0;
 933         return true;
 934 }
 935
 936 static bool migrate_pred(struct bch_fs *c, void *arg,
 937                          struct bkey_s_c k,
 938                          struct bch_io_opts *io_opts,
 939                          struct data_update_opts *data_opts)
 940 {
 941         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 942         const struct bch_extent_ptr *ptr;
 943         struct bch_ioctl_data *op = arg;
 944         unsigned i = 0;
 945
 946         data_opts->rewrite_ptrs         = 0;
 947         data_opts->target               = 0;
 948         data_opts->extra_replicas       = 0;
 949         data_opts->btree_insert_flags   = 0;
 950
 951         bkey_for_each_ptr(ptrs, ptr) {
 952                 if (ptr->dev == op->migrate.dev)
 953                         data_opts->rewrite_ptrs |= 1U << i;
 954                 i++;
 955         }
 956
 957         return data_opts->rewrite_ptrs != 0;
 958 }
 959
 960 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 961                                    struct btree *b,
 962                                    struct bch_io_opts *io_opts,
 963                                    struct data_update_opts *data_opts)
 964 {
 965         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 966 }
 967
 968 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 969                                struct btree *b,
 970                                struct bch_io_opts *io_opts,
 971                                struct data_update_opts *data_opts)
 972 {
 973         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 974 }
 975
 976 static bool bformat_needs_redo(struct bkey_format *f)
 977 {
 978         unsigned i;
 979
 980         for (i = 0; i < f->nr_fields; i++) {
 981                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 982                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 983                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 984
 985                 if (f->bits_per_field[i] > unpacked_bits)
 986                         return true;
 987
 988                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
 989                         return true;
 990
 991                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
 992                      unpacked_mask) <
 993                     field_offset)
 994                         return true;
 995         }
 996
 997         return false;
 998 }
 999
1000 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1001                                    struct btree *b,
1002                                    struct bch_io_opts *io_opts,
1003                                    struct data_update_opts *data_opts)
1004 {
1005         if (b->version_ondisk != c->sb.version ||
1006             btree_node_need_rewrite(b) ||
1007             bformat_needs_redo(&b->format)) {
1008                 data_opts->target               = 0;
1009                 data_opts->extra_replicas       = 0;
1010                 data_opts->btree_insert_flags   = 0;
1011                 return true;
1012         }
1013
1014         return false;
1015 }
1016
1017 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1018 {
1019         int ret;
1020
1021         ret = bch2_move_btree(c,
1022                               0,                POS_MIN,
1023                               BTREE_ID_NR,      SPOS_MAX,
1024                               rewrite_old_nodes_pred, c, stats);
1025         if (!ret) {
1026                 mutex_lock(&c->sb_lock);
1027                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1028                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1029                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1030                 bch2_write_super(c);
1031                 mutex_unlock(&c->sb_lock);
1032         }
1033
1034         return ret;
1035 }
1036
1037 int bch2_data_job(struct bch_fs *c,
1038                   struct bch_move_stats *stats,
1039                   struct bch_ioctl_data op)
1040 {
1041         int ret = 0;
1042
1043         switch (op.op) {
1044         case BCH_DATA_OP_REREPLICATE:
1045                 bch2_move_stats_init(stats, "rereplicate");
1046                 stats->data_type = BCH_DATA_journal;
1047                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1048
1049                 ret = bch2_move_btree(c,
1050                                       op.start_btree,   op.start_pos,
1051                                       op.end_btree,     op.end_pos,
1052                                       rereplicate_btree_pred, c, stats) ?: ret;
1053                 ret = bch2_replicas_gc2(c) ?: ret;
1054
1055                 ret = bch2_move_data(c,
1056                                      op.start_btree,    op.start_pos,
1057                                      op.end_btree,      op.end_pos,
1058                                      NULL,
1059                                      stats,
1060                                      writepoint_hashed((unsigned long) current),
1061                                      true,
1062                                      rereplicate_pred, c) ?: ret;
1063                 ret = bch2_replicas_gc2(c) ?: ret;
1064                 break;
1065         case BCH_DATA_OP_MIGRATE:
1066                 if (op.migrate.dev >= c->sb.nr_devices)
1067                         return -EINVAL;
1068
1069                 bch2_move_stats_init(stats, "migrate");
1070                 stats->data_type = BCH_DATA_journal;
1071                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1072
1073                 ret = bch2_move_btree(c,
1074                                       op.start_btree,   op.start_pos,
1075                                       op.end_btree,     op.end_pos,
1076                                       migrate_btree_pred, &op, stats) ?: ret;
1077                 ret = bch2_replicas_gc2(c) ?: ret;
1078
1079                 ret = bch2_move_data(c,
1080                                      op.start_btree,    op.start_pos,
1081                                      op.end_btree,      op.end_pos,
1082                                      NULL,
1083                                      stats,
1084                                      writepoint_hashed((unsigned long) current),
1085                                      true,
1086                                      migrate_pred, &op) ?: ret;
1087                 ret = bch2_replicas_gc2(c) ?: ret;
1088                 break;
1089         case BCH_DATA_OP_REWRITE_OLD_NODES:
1090                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1091                 ret = bch2_scan_old_btree_nodes(c, stats);
1092                 break;
1093         default:
1094                 ret = -EINVAL;
1095         }
1096
1097         return ret;
1098 }
1099
1100 void bch2_data_jobs_to_text(struct printbuf *out, struct bch_fs *c)
1101 {
1102         struct bch_move_stats *stats;
1103
1104         mutex_lock(&c->data_progress_lock);
1105         list_for_each_entry(stats, &c->data_progress_list, list) {
1106                 prt_printf(out, "%s: data type %s btree_id %s position: ",
1107                        stats->name,
1108                        bch2_data_types[stats->data_type],
1109                        bch2_btree_ids[stats->btree_id]);
1110                 bch2_bpos_to_text(out, stats->pos);
1111                 prt_printf(out, "%s", "\n");
1112         }
1113         mutex_unlock(&c->data_progress_lock);
1114 }
1115
1116 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct moving_context *ctxt)
1117 {
1118         struct moving_io *io;
1119
1120         prt_printf(out, "%ps:", ctxt->fn);
1121         prt_newline(out);
1122         printbuf_indent_add(out, 2);
1123
1124         prt_printf(out, "reads: %u sectors %u",
1125                    atomic_read(&ctxt->read_ios),
1126                    atomic_read(&ctxt->read_sectors));
1127         prt_newline(out);
1128
1129         prt_printf(out, "writes: %u sectors %u",
1130                    atomic_read(&ctxt->write_ios),
1131                    atomic_read(&ctxt->write_sectors));
1132         prt_newline(out);
1133
1134         printbuf_indent_add(out, 2);
1135
1136         mutex_lock(&ctxt->lock);
1137         list_for_each_entry(io, &ctxt->ios, io_list) {
1138                 bch2_write_op_to_text(out, &io->write.op);
1139         }
1140         mutex_unlock(&ctxt->lock);
1141
1142         printbuf_indent_sub(out, 4);
1143 }
1144
1145 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1146 {
1147         struct moving_context *ctxt;
1148
1149         mutex_lock(&c->moving_context_lock);
1150         list_for_each_entry(ctxt, &c->moving_context_list, list)
1151                 bch2_moving_ctxt_to_text(out, ctxt);
1152         mutex_unlock(&c->moving_context_lock);
1153 }
1154
1155 void bch2_fs_move_init(struct bch_fs *c)
1156 {
1157         INIT_LIST_HEAD(&c->moving_context_list);
1158         mutex_init(&c->moving_context_lock);
1159
1160         INIT_LIST_HEAD(&c->data_progress_list);
1161         mutex_init(&c->data_progress_lock);
1162 }