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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_background.h"
   5 #include "alloc_foreground.h"
   6 #include "backpointers.h"
   7 #include "bkey_buf.h"
   8 #include "btree_gc.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_write_buffer.h"
  12 #include "disk_groups.h"
  13 #include "ec.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "inode.h"
  17 #include "io_read.h"
  18 #include "io_write.h"
  19 #include "journal_reclaim.h"
  20 #include "keylist.h"
  21 #include "move.h"
  22 #include "replicas.h"
  23 #include "snapshot.h"
  24 #include "super-io.h"
  25 #include "trace.h"
  26
  27 #include <linux/ioprio.h>
  28 #include <linux/kthread.h>
  29
  30 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
  31 {
  32         if (trace_move_extent_enabled()) {
  33                 struct printbuf buf = PRINTBUF;
  34
  35                 bch2_bkey_val_to_text(&buf, c, k);
  36                 trace_move_extent(c, buf.buf);
  37                 printbuf_exit(&buf);
  38         }
  39 }
  40
  41 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
  42 {
  43         if (trace_move_extent_read_enabled()) {
  44                 struct printbuf buf = PRINTBUF;
  45
  46                 bch2_bkey_val_to_text(&buf, c, k);
  47                 trace_move_extent_read(c, buf.buf);
  48                 printbuf_exit(&buf);
  49         }
  50 }
  51
  52 static void trace_move_extent_alloc_mem_fail2(struct bch_fs *c, struct bkey_s_c k)
  53 {
  54         if (trace_move_extent_alloc_mem_fail_enabled()) {
  55                 struct printbuf buf = PRINTBUF;
  56
  57                 bch2_bkey_val_to_text(&buf, c, k);
  58                 trace_move_extent_alloc_mem_fail(c, buf.buf);
  59                 printbuf_exit(&buf);
  60         }
  61 }
  62
  63 struct moving_io {
  64         struct list_head                read_list;
  65         struct list_head                io_list;
  66         struct move_bucket_in_flight    *b;
  67         struct closure                  cl;
  68         bool                            read_completed;
  69
  70         unsigned                        read_sectors;
  71         unsigned                        write_sectors;
  72
  73         struct bch_read_bio             rbio;
  74
  75         struct data_update              write;
  76         /* Must be last since it is variable size */
  77         struct bio_vec                  bi_inline_vecs[0];
  78 };
  79
  80 static void move_free(struct moving_io *io)
  81 {
  82         struct moving_context *ctxt = io->write.ctxt;
  83
  84         if (io->b)
  85                 atomic_dec(&io->b->count);
  86
  87         bch2_data_update_exit(&io->write);
  88
  89         mutex_lock(&ctxt->lock);
  90         list_del(&io->io_list);
  91         wake_up(&ctxt->wait);
  92         mutex_unlock(&ctxt->lock);
  93
  94         kfree(io);
  95 }
  96
  97 static void move_write_done(struct bch_write_op *op)
  98 {
  99         struct moving_io *io = container_of(op, struct moving_io, write.op);
 100         struct moving_context *ctxt = io->write.ctxt;
 101
 102         if (io->write.op.error)
 103                 ctxt->write_error = true;
 104
 105         atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
 106         atomic_dec(&io->write.ctxt->write_ios);
 107         move_free(io);
 108         closure_put(&ctxt->cl);
 109 }
 110
 111 static void move_write(struct moving_io *io)
 112 {
 113         if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
 114                 move_free(io);
 115                 return;
 116         }
 117
 118         closure_get(&io->write.ctxt->cl);
 119         atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
 120         atomic_inc(&io->write.ctxt->write_ios);
 121
 122         bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
 123 }
 124
 125 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
 126 {
 127         struct moving_io *io =
 128                 list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
 129
 130         return io && io->read_completed ? io : NULL;
 131 }
 132
 133 static void move_read_endio(struct bio *bio)
 134 {
 135         struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
 136         struct moving_context *ctxt = io->write.ctxt;
 137
 138         atomic_sub(io->read_sectors, &ctxt->read_sectors);
 139         atomic_dec(&ctxt->read_ios);
 140         io->read_completed = true;
 141
 142         wake_up(&ctxt->wait);
 143         closure_put(&ctxt->cl);
 144 }
 145
 146 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
 147 {
 148         struct moving_io *io;
 149
 150         bch2_trans_unlock(ctxt->trans);
 151
 152         while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
 153                 list_del(&io->read_list);
 154                 move_write(io);
 155         }
 156 }
 157
 158 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
 159 {
 160         unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
 161
 162         move_ctxt_wait_event(ctxt,
 163                 !atomic_read(&ctxt->write_sectors) ||
 164                 atomic_read(&ctxt->write_sectors) != sectors_pending);
 165 }
 166
 167 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 168 {
 169         struct bch_fs *c = ctxt->trans->c;
 170
 171         move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 172         closure_sync(&ctxt->cl);
 173
 174         EBUG_ON(atomic_read(&ctxt->write_sectors));
 175         EBUG_ON(atomic_read(&ctxt->write_ios));
 176         EBUG_ON(atomic_read(&ctxt->read_sectors));
 177         EBUG_ON(atomic_read(&ctxt->read_ios));
 178
 179         mutex_lock(&c->moving_context_lock);
 180         list_del(&ctxt->list);
 181         mutex_unlock(&c->moving_context_lock);
 182
 183         bch2_trans_put(ctxt->trans);
 184         memset(ctxt, 0, sizeof(*ctxt));
 185 }
 186
 187 void bch2_moving_ctxt_init(struct moving_context *ctxt,
 188                            struct bch_fs *c,
 189                            struct bch_ratelimit *rate,
 190                            struct bch_move_stats *stats,
 191                            struct write_point_specifier wp,
 192                            bool wait_on_copygc)
 193 {
 194         memset(ctxt, 0, sizeof(*ctxt));
 195
 196         ctxt->trans     = bch2_trans_get(c);
 197         ctxt->fn        = (void *) _RET_IP_;
 198         ctxt->rate      = rate;
 199         ctxt->stats     = stats;
 200         ctxt->wp        = wp;
 201         ctxt->wait_on_copygc = wait_on_copygc;
 202
 203         closure_init_stack(&ctxt->cl);
 204
 205         mutex_init(&ctxt->lock);
 206         INIT_LIST_HEAD(&ctxt->reads);
 207         INIT_LIST_HEAD(&ctxt->ios);
 208         init_waitqueue_head(&ctxt->wait);
 209
 210         mutex_lock(&c->moving_context_lock);
 211         list_add(&ctxt->list, &c->moving_context_list);
 212         mutex_unlock(&c->moving_context_lock);
 213 }
 214
 215 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
 216 {
 217         trace_move_data(c, stats);
 218 }
 219
 220 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 221 {
 222         memset(stats, 0, sizeof(*stats));
 223         stats->data_type = BCH_DATA_user;
 224         scnprintf(stats->name, sizeof(stats->name), "%s", name);
 225 }
 226
 227 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
 228                                  struct btree_iter *iter,
 229                                  struct bkey_s_c k,
 230                                  struct data_update_opts data_opts)
 231 {
 232         struct bch_fs *c = trans->c;
 233         struct bkey_i *n;
 234         int ret;
 235
 236         n = bch2_bkey_make_mut_noupdate(trans, k);
 237         ret = PTR_ERR_OR_ZERO(n);
 238         if (ret)
 239                 return ret;
 240
 241         while (data_opts.kill_ptrs) {
 242                 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
 243                 struct bch_extent_ptr *ptr;
 244
 245                 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
 246                 data_opts.kill_ptrs ^= 1U << drop;
 247         }
 248
 249         /*
 250          * If the new extent no longer has any pointers, bch2_extent_normalize()
 251          * will do the appropriate thing with it (turning it into a
 252          * KEY_TYPE_error key, or just a discard if it was a cached extent)
 253          */
 254         bch2_extent_normalize(c, bkey_i_to_s(n));
 255
 256         /*
 257          * Since we're not inserting through an extent iterator
 258          * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 259          * we aren't using the extent overwrite path to delete, we're
 260          * just using the normal key deletion path:
 261          */
 262         if (bkey_deleted(&n->k))
 263                 n->k.size = 0;
 264
 265         return bch2_trans_relock(trans) ?:
 266                 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 267                 bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
 268 }
 269
 270 int bch2_move_extent(struct moving_context *ctxt,
 271                      struct move_bucket_in_flight *bucket_in_flight,
 272                      struct btree_iter *iter,
 273                      struct bkey_s_c k,
 274                      struct bch_io_opts io_opts,
 275                      struct data_update_opts data_opts)
 276 {
 277         struct btree_trans *trans = ctxt->trans;
 278         struct bch_fs *c = trans->c;
 279         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 280         struct moving_io *io;
 281         const union bch_extent_entry *entry;
 282         struct extent_ptr_decoded p;
 283         unsigned sectors = k.k->size, pages;
 284         int ret = -ENOMEM;
 285
 286         if (ctxt->stats)
 287                 ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
 288         trace_move_extent2(c, k);
 289
 290         bch2_data_update_opts_normalize(k, &data_opts);
 291
 292         if (!data_opts.rewrite_ptrs &&
 293             !data_opts.extra_replicas) {
 294                 if (data_opts.kill_ptrs)
 295                         return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
 296                 return 0;
 297         }
 298
 299         /*
 300          * Before memory allocations & taking nocow locks in
 301          * bch2_data_update_init():
 302          */
 303         bch2_trans_unlock(trans);
 304
 305         /* write path might have to decompress data: */
 306         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 307                 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
 308
 309         pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 310         io = kzalloc(sizeof(struct moving_io) +
 311                      sizeof(struct bio_vec) * pages, GFP_KERNEL);
 312         if (!io)
 313                 goto err;
 314
 315         INIT_LIST_HEAD(&io->io_list);
 316         io->write.ctxt          = ctxt;
 317         io->read_sectors        = k.k->size;
 318         io->write_sectors       = k.k->size;
 319
 320         bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 321         bio_set_prio(&io->write.op.wbio.bio,
 322                      IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 323
 324         if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
 325                                  GFP_KERNEL))
 326                 goto err_free;
 327
 328         io->rbio.c              = c;
 329         io->rbio.opts           = io_opts;
 330         bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
 331         io->rbio.bio.bi_vcnt = pages;
 332         bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 333         io->rbio.bio.bi_iter.bi_size = sectors << 9;
 334
 335         io->rbio.bio.bi_opf             = REQ_OP_READ;
 336         io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
 337         io->rbio.bio.bi_end_io          = move_read_endio;
 338
 339         ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
 340                                     io_opts, data_opts, iter->btree_id, k);
 341         if (ret && ret != -BCH_ERR_unwritten_extent_update)
 342                 goto err_free_pages;
 343
 344         if (ret == -BCH_ERR_unwritten_extent_update) {
 345                 bch2_update_unwritten_extent(trans, &io->write);
 346                 move_free(io);
 347                 return 0;
 348         }
 349
 350         BUG_ON(ret);
 351
 352         io->write.op.end_io = move_write_done;
 353
 354         if (ctxt->rate)
 355                 bch2_ratelimit_increment(ctxt->rate, k.k->size);
 356
 357         if (ctxt->stats) {
 358                 atomic64_inc(&ctxt->stats->keys_moved);
 359                 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
 360         }
 361
 362         if (bucket_in_flight) {
 363                 io->b = bucket_in_flight;
 364                 atomic_inc(&io->b->count);
 365         }
 366
 367         this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
 368         this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
 369         trace_move_extent_read2(c, k);
 370
 371         mutex_lock(&ctxt->lock);
 372         atomic_add(io->read_sectors, &ctxt->read_sectors);
 373         atomic_inc(&ctxt->read_ios);
 374
 375         list_add_tail(&io->read_list, &ctxt->reads);
 376         list_add_tail(&io->io_list, &ctxt->ios);
 377         mutex_unlock(&ctxt->lock);
 378
 379         /*
 380          * dropped by move_read_endio() - guards against use after free of
 381          * ctxt when doing wakeup
 382          */
 383         closure_get(&ctxt->cl);
 384         bch2_read_extent(trans, &io->rbio,
 385                          bkey_start_pos(k.k),
 386                          iter->btree_id, k, 0,
 387                          BCH_READ_NODECODE|
 388                          BCH_READ_LAST_FRAGMENT);
 389         return 0;
 390 err_free_pages:
 391         bio_free_pages(&io->write.op.wbio.bio);
 392 err_free:
 393         kfree(io);
 394 err:
 395         this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
 396         trace_move_extent_alloc_mem_fail2(c, k);
 397         return ret;
 398 }
 399
 400 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 401                           struct per_snapshot_io_opts *io_opts,
 402                           struct bkey_s_c extent_k)
 403 {
 404         struct bch_fs *c = trans->c;
 405         u32 restart_count = trans->restart_count;
 406         int ret = 0;
 407
 408         if (io_opts->cur_inum != extent_k.k->p.inode) {
 409                 struct btree_iter iter;
 410                 struct bkey_s_c k;
 411
 412                 io_opts->d.nr = 0;
 413
 414                 for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
 415                                    BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
 416                         if (k.k->p.offset != extent_k.k->p.inode)
 417                                 break;
 418
 419                         if (!bkey_is_inode(k.k))
 420                                 continue;
 421
 422                         struct bch_inode_unpacked inode;
 423                         BUG_ON(bch2_inode_unpack(k, &inode));
 424
 425                         struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 426                         bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
 427
 428                         ret = darray_push(&io_opts->d, e);
 429                         if (ret)
 430                                 break;
 431                 }
 432                 bch2_trans_iter_exit(trans, &iter);
 433                 io_opts->cur_inum = extent_k.k->p.inode;
 434         }
 435
 436         ret = ret ?: trans_was_restarted(trans, restart_count);
 437         if (ret)
 438                 return ERR_PTR(ret);
 439
 440         if (extent_k.k->p.snapshot) {
 441                 struct snapshot_io_opts_entry *i;
 442                 darray_for_each(io_opts->d, i)
 443                         if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
 444                                 return &i->io_opts;
 445         }
 446
 447         return &io_opts->fs_io_opts;
 448 }
 449
 450 int bch2_move_get_io_opts_one(struct btree_trans *trans,
 451                               struct bch_io_opts *io_opts,
 452                               struct bkey_s_c extent_k)
 453 {
 454         struct btree_iter iter;
 455         struct bkey_s_c k;
 456         int ret;
 457
 458         /* reflink btree? */
 459         if (!extent_k.k->p.inode) {
 460                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 461                 return 0;
 462         }
 463
 464         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 465                                SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
 466                                BTREE_ITER_CACHED);
 467         ret = bkey_err(k);
 468         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 469                 return ret;
 470
 471         if (!ret && bkey_is_inode(k.k)) {
 472                 struct bch_inode_unpacked inode;
 473                 bch2_inode_unpack(k, &inode);
 474                 bch2_inode_opts_get(io_opts, trans->c, &inode);
 475         } else {
 476                 *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
 477         }
 478
 479         bch2_trans_iter_exit(trans, &iter);
 480         return 0;
 481 }
 482
 483 int bch2_move_ratelimit(struct moving_context *ctxt)
 484 {
 485         struct bch_fs *c = ctxt->trans->c;
 486         u64 delay;
 487
 488         if (ctxt->wait_on_copygc) {
 489                 bch2_trans_unlock(ctxt->trans);
 490                 wait_event_killable(c->copygc_running_wq,
 491                                     !c->copygc_running ||
 492                                     kthread_should_stop());
 493         }
 494
 495         do {
 496                 delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
 497
 498                 if (delay) {
 499                         bch2_trans_unlock(ctxt->trans);
 500                         set_current_state(TASK_INTERRUPTIBLE);
 501                 }
 502
 503                 if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
 504                         __set_current_state(TASK_RUNNING);
 505                         return 1;
 506                 }
 507
 508                 if (delay)
 509                         schedule_timeout(delay);
 510
 511                 if (unlikely(freezing(current))) {
 512                         move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
 513                         try_to_freeze();
 514                 }
 515         } while (delay);
 516
 517         /*
 518          * XXX: these limits really ought to be per device, SSDs and hard drives
 519          * will want different limits
 520          */
 521         move_ctxt_wait_event(ctxt,
 522                 atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 523                 atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
 524                 atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
 525                 atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
 526
 527         return 0;
 528 }
 529
 530 static int bch2_move_data_btree(struct moving_context *ctxt,
 531                                 struct bpos start,
 532                                 struct bpos end,
 533                                 move_pred_fn pred, void *arg,
 534                                 enum btree_id btree_id)
 535 {
 536         struct btree_trans *trans = ctxt->trans;
 537         struct bch_fs *c = trans->c;
 538         struct per_snapshot_io_opts snapshot_io_opts;
 539         struct bch_io_opts *io_opts;
 540         struct bkey_buf sk;
 541         struct btree_iter iter;
 542         struct bkey_s_c k;
 543         struct data_update_opts data_opts;
 544         int ret = 0, ret2;
 545
 546         per_snapshot_io_opts_init(&snapshot_io_opts, c);
 547         bch2_bkey_buf_init(&sk);
 548
 549         if (ctxt->stats) {
 550                 ctxt->stats->data_type  = BCH_DATA_user;
 551                 ctxt->stats->pos        = BBPOS(btree_id, start);
 552         }
 553
 554         bch2_trans_iter_init(trans, &iter, btree_id, start,
 555                              BTREE_ITER_PREFETCH|
 556                              BTREE_ITER_ALL_SNAPSHOTS);
 557
 558         if (ctxt->rate)
 559                 bch2_ratelimit_reset(ctxt->rate);
 560
 561         while (!bch2_move_ratelimit(ctxt)) {
 562                 bch2_trans_begin(trans);
 563
 564                 k = bch2_btree_iter_peek(&iter);
 565                 if (!k.k)
 566                         break;
 567
 568                 ret = bkey_err(k);
 569                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 570                         continue;
 571                 if (ret)
 572                         break;
 573
 574                 if (bkey_ge(bkey_start_pos(k.k), end))
 575                         break;
 576
 577                 if (ctxt->stats)
 578                         ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
 579
 580                 if (!bkey_extent_is_direct_data(k.k))
 581                         goto next_nondata;
 582
 583                 io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
 584                 ret = PTR_ERR_OR_ZERO(io_opts);
 585                 if (ret)
 586                         continue;
 587
 588                 memset(&data_opts, 0, sizeof(data_opts));
 589                 if (!pred(c, arg, k, io_opts, &data_opts))
 590                         goto next;
 591
 592                 /*
 593                  * The iterator gets unlocked by __bch2_read_extent - need to
 594                  * save a copy of @k elsewhere:
 595                  */
 596                 bch2_bkey_buf_reassemble(&sk, c, k);
 597                 k = bkey_i_to_s_c(sk.k);
 598
 599                 ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
 600                 if (ret2) {
 601                         if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 602                                 continue;
 603
 604                         if (ret2 == -ENOMEM) {
 605                                 /* memory allocation failure, wait for some IO to finish */
 606                                 bch2_move_ctxt_wait_for_io(ctxt);
 607                                 continue;
 608                         }
 609
 610                         /* XXX signal failure */
 611                         goto next;
 612                 }
 613 next:
 614                 if (ctxt->stats)
 615                         atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 616 next_nondata:
 617                 bch2_btree_iter_advance(&iter);
 618         }
 619
 620         bch2_trans_iter_exit(trans, &iter);
 621         bch2_bkey_buf_exit(&sk, c);
 622         per_snapshot_io_opts_exit(&snapshot_io_opts);
 623
 624         return ret;
 625 }
 626
 627 int __bch2_move_data(struct moving_context *ctxt,
 628                      struct bbpos start,
 629                      struct bbpos end,
 630                      move_pred_fn pred, void *arg)
 631 {
 632         struct bch_fs *c = ctxt->trans->c;
 633         enum btree_id id;
 634         int ret = 0;
 635
 636         for (id = start.btree;
 637              id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
 638              id++) {
 639                 ctxt->stats->pos = BBPOS(id, POS_MIN);
 640
 641                 if (!btree_type_has_ptrs(id) ||
 642                     !bch2_btree_id_root(c, id)->b)
 643                         continue;
 644
 645                 ret = bch2_move_data_btree(ctxt,
 646                                        id == start.btree ? start.pos : POS_MIN,
 647                                        id == end.btree   ? end.pos   : POS_MAX,
 648                                        pred, arg, id);
 649                 if (ret)
 650                         break;
 651         }
 652
 653         return ret;
 654 }
 655
 656 int bch2_move_data(struct bch_fs *c,
 657                    struct bbpos start,
 658                    struct bbpos end,
 659                    struct bch_ratelimit *rate,
 660                    struct bch_move_stats *stats,
 661                    struct write_point_specifier wp,
 662                    bool wait_on_copygc,
 663                    move_pred_fn pred, void *arg)
 664 {
 665
 666         struct moving_context ctxt;
 667         int ret;
 668
 669         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 670         ret = __bch2_move_data(&ctxt, start, end, pred, arg);
 671         bch2_moving_ctxt_exit(&ctxt);
 672
 673         return ret;
 674 }
 675
 676 int __bch2_evacuate_bucket(struct moving_context *ctxt,
 677                            struct move_bucket_in_flight *bucket_in_flight,
 678                            struct bpos bucket, int gen,
 679                            struct data_update_opts _data_opts)
 680 {
 681         struct btree_trans *trans = ctxt->trans;
 682         struct bch_fs *c = trans->c;
 683         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 684         struct btree_iter iter;
 685         struct bkey_buf sk;
 686         struct bch_backpointer bp;
 687         struct bch_alloc_v4 a_convert;
 688         const struct bch_alloc_v4 *a;
 689         struct bkey_s_c k;
 690         struct data_update_opts data_opts;
 691         unsigned dirty_sectors, bucket_size;
 692         u64 fragmentation;
 693         struct bpos bp_pos = POS_MIN;
 694         int ret = 0;
 695
 696         trace_bucket_evacuate(c, &bucket);
 697
 698         bch2_bkey_buf_init(&sk);
 699
 700         /*
 701          * We're not run in a context that handles transaction restarts:
 702          */
 703         bch2_trans_begin(trans);
 704
 705         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 706                              bucket, BTREE_ITER_CACHED);
 707         ret = lockrestart_do(trans,
 708                         bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 709         bch2_trans_iter_exit(trans, &iter);
 710
 711         if (ret) {
 712                 bch_err_msg(c, ret, "looking up alloc key");
 713                 goto err;
 714         }
 715
 716         a = bch2_alloc_to_v4(k, &a_convert);
 717         dirty_sectors = a->dirty_sectors;
 718         bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
 719         fragmentation = a->fragmentation_lru;
 720
 721         ret = bch2_btree_write_buffer_flush(trans);
 722         if (ret) {
 723                 bch_err_msg(c, ret, "flushing btree write buffer");
 724                 goto err;
 725         }
 726
 727         while (!(ret = bch2_move_ratelimit(ctxt))) {
 728                 bch2_trans_begin(trans);
 729
 730                 ret = bch2_get_next_backpointer(trans, bucket, gen,
 731                                                 &bp_pos, &bp,
 732                                                 BTREE_ITER_CACHED);
 733                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 734                         continue;
 735                 if (ret)
 736                         goto err;
 737                 if (bkey_eq(bp_pos, POS_MAX))
 738                         break;
 739
 740                 if (!bp.level) {
 741                         const struct bch_extent_ptr *ptr;
 742                         unsigned i = 0;
 743
 744                         k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
 745                         ret = bkey_err(k);
 746                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 747                                 continue;
 748                         if (ret)
 749                                 goto err;
 750                         if (!k.k)
 751                                 goto next;
 752
 753                         bch2_bkey_buf_reassemble(&sk, c, k);
 754                         k = bkey_i_to_s_c(sk.k);
 755
 756                         ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
 757                         if (ret) {
 758                                 bch2_trans_iter_exit(trans, &iter);
 759                                 continue;
 760                         }
 761
 762                         data_opts = _data_opts;
 763                         data_opts.target        = io_opts.background_target;
 764                         data_opts.rewrite_ptrs = 0;
 765
 766                         bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
 767                                 if (ptr->dev == bucket.inode) {
 768                                         data_opts.rewrite_ptrs |= 1U << i;
 769                                         if (ptr->cached) {
 770                                                 bch2_trans_iter_exit(trans, &iter);
 771                                                 goto next;
 772                                         }
 773                                 }
 774                                 i++;
 775                         }
 776
 777                         ret = bch2_move_extent(ctxt, bucket_in_flight,
 778                                                &iter, k, io_opts, data_opts);
 779                         bch2_trans_iter_exit(trans, &iter);
 780
 781                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 782                                 continue;
 783                         if (ret == -ENOMEM) {
 784                                 /* memory allocation failure, wait for some IO to finish */
 785                                 bch2_move_ctxt_wait_for_io(ctxt);
 786                                 continue;
 787                         }
 788                         if (ret)
 789                                 goto err;
 790
 791                         if (ctxt->stats)
 792                                 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 793                 } else {
 794                         struct btree *b;
 795
 796                         b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
 797                         ret = PTR_ERR_OR_ZERO(b);
 798                         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 799                                 continue;
 800                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 801                                 continue;
 802                         if (ret)
 803                                 goto err;
 804                         if (!b)
 805                                 goto next;
 806
 807                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
 808                         bch2_trans_iter_exit(trans, &iter);
 809
 810                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 811                                 continue;
 812                         if (ret)
 813                                 goto err;
 814
 815                         if (ctxt->rate)
 816                                 bch2_ratelimit_increment(ctxt->rate,
 817                                                          c->opts.btree_node_size >> 9);
 818                         if (ctxt->stats) {
 819                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
 820                                 atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
 821                         }
 822                 }
 823 next:
 824                 bp_pos = bpos_nosnap_successor(bp_pos);
 825         }
 826
 827         trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 828 err:
 829         bch2_bkey_buf_exit(&sk, c);
 830         return ret;
 831 }
 832
 833 int bch2_evacuate_bucket(struct bch_fs *c,
 834                          struct bpos bucket, int gen,
 835                          struct data_update_opts data_opts,
 836                          struct bch_ratelimit *rate,
 837                          struct bch_move_stats *stats,
 838                          struct write_point_specifier wp,
 839                          bool wait_on_copygc)
 840 {
 841         struct moving_context ctxt;
 842         int ret;
 843
 844         bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
 845         ret = __bch2_evacuate_bucket(&ctxt, NULL, bucket, gen, data_opts);
 846         bch2_moving_ctxt_exit(&ctxt);
 847
 848         return ret;
 849 }
 850
 851 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
 852                                 struct btree *, struct bch_io_opts *,
 853                                 struct data_update_opts *);
 854
 855 static int bch2_move_btree(struct bch_fs *c,
 856                            enum btree_id start_btree_id, struct bpos start_pos,
 857                            enum btree_id end_btree_id,   struct bpos end_pos,
 858                            move_btree_pred pred, void *arg,
 859                            struct bch_move_stats *stats)
 860 {
 861         bool kthread = (current->flags & PF_KTHREAD) != 0;
 862         struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
 863         struct moving_context ctxt;
 864         struct btree_trans *trans;
 865         struct btree_iter iter;
 866         struct btree *b;
 867         enum btree_id id;
 868         struct data_update_opts data_opts;
 869         int ret = 0;
 870
 871         bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
 872                               writepoint_ptr(&c->btree_write_point),
 873                               true);
 874         trans = ctxt.trans;
 875
 876         stats->data_type = BCH_DATA_btree;
 877
 878         for (id = start_btree_id;
 879              id <= min_t(unsigned, end_btree_id, btree_id_nr_alive(c) - 1);
 880              id++) {
 881                 stats->pos = BBPOS(id, POS_MIN);
 882
 883                 if (!bch2_btree_id_root(c, id)->b)
 884                         continue;
 885
 886                 bch2_trans_node_iter_init(trans, &iter, id, POS_MIN, 0, 0,
 887                                           BTREE_ITER_PREFETCH);
 888 retry:
 889                 ret = 0;
 890                 while (bch2_trans_begin(trans),
 891                        (b = bch2_btree_iter_peek_node(&iter)) &&
 892                        !(ret = PTR_ERR_OR_ZERO(b))) {
 893                         if (kthread && kthread_should_stop())
 894                                 break;
 895
 896                         if ((cmp_int(id, end_btree_id) ?:
 897                              bpos_cmp(b->key.k.p, end_pos)) > 0)
 898                                 break;
 899
 900                         stats->pos = BBPOS(iter.btree_id, iter.pos);
 901
 902                         if (!pred(c, arg, b, &io_opts, &data_opts))
 903                                 goto next;
 904
 905                         ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
 906                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 907                                 continue;
 908                         if (ret)
 909                                 break;
 910 next:
 911                         bch2_btree_iter_next_node(&iter);
 912                 }
 913                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 914                         goto retry;
 915
 916                 bch2_trans_iter_exit(trans, &iter);
 917
 918                 if (kthread && kthread_should_stop())
 919                         break;
 920         }
 921
 922         bch_err_fn(c, ret);
 923         bch2_moving_ctxt_exit(&ctxt);
 924         bch2_btree_interior_updates_flush(c);
 925
 926         return ret;
 927 }
 928
 929 static bool rereplicate_pred(struct bch_fs *c, void *arg,
 930                              struct bkey_s_c k,
 931                              struct bch_io_opts *io_opts,
 932                              struct data_update_opts *data_opts)
 933 {
 934         unsigned nr_good = bch2_bkey_durability(c, k);
 935         unsigned replicas = bkey_is_btree_ptr(k.k)
 936                 ? c->opts.metadata_replicas
 937                 : io_opts->data_replicas;
 938
 939         if (!nr_good || nr_good >= replicas)
 940                 return false;
 941
 942         data_opts->target               = 0;
 943         data_opts->extra_replicas       = replicas - nr_good;
 944         data_opts->btree_insert_flags   = 0;
 945         return true;
 946 }
 947
 948 static bool migrate_pred(struct bch_fs *c, void *arg,
 949                          struct bkey_s_c k,
 950                          struct bch_io_opts *io_opts,
 951                          struct data_update_opts *data_opts)
 952 {
 953         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 954         const struct bch_extent_ptr *ptr;
 955         struct bch_ioctl_data *op = arg;
 956         unsigned i = 0;
 957
 958         data_opts->rewrite_ptrs         = 0;
 959         data_opts->target               = 0;
 960         data_opts->extra_replicas       = 0;
 961         data_opts->btree_insert_flags   = 0;
 962
 963         bkey_for_each_ptr(ptrs, ptr) {
 964                 if (ptr->dev == op->migrate.dev)
 965                         data_opts->rewrite_ptrs |= 1U << i;
 966                 i++;
 967         }
 968
 969         return data_opts->rewrite_ptrs != 0;
 970 }
 971
 972 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
 973                                    struct btree *b,
 974                                    struct bch_io_opts *io_opts,
 975                                    struct data_update_opts *data_opts)
 976 {
 977         return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 978 }
 979
 980 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
 981                                struct btree *b,
 982                                struct bch_io_opts *io_opts,
 983                                struct data_update_opts *data_opts)
 984 {
 985         return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
 986 }
 987
 988 static bool bformat_needs_redo(struct bkey_format *f)
 989 {
 990         unsigned i;
 991
 992         for (i = 0; i < f->nr_fields; i++) {
 993                 unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
 994                 u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
 995                 u64 field_offset = le64_to_cpu(f->field_offset[i]);
 996
 997                 if (f->bits_per_field[i] > unpacked_bits)
 998                         return true;
 999
1000                 if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1001                         return true;
1002
1003                 if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1004                      unpacked_mask) <
1005                     field_offset)
1006                         return true;
1007         }
1008
1009         return false;
1010 }
1011
1012 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1013                                    struct btree *b,
1014                                    struct bch_io_opts *io_opts,
1015                                    struct data_update_opts *data_opts)
1016 {
1017         if (b->version_ondisk != c->sb.version ||
1018             btree_node_need_rewrite(b) ||
1019             bformat_needs_redo(&b->format)) {
1020                 data_opts->target               = 0;
1021                 data_opts->extra_replicas       = 0;
1022                 data_opts->btree_insert_flags   = 0;
1023                 return true;
1024         }
1025
1026         return false;
1027 }
1028
1029 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1030 {
1031         int ret;
1032
1033         ret = bch2_move_btree(c,
1034                               0,                POS_MIN,
1035                               BTREE_ID_NR,      SPOS_MAX,
1036                               rewrite_old_nodes_pred, c, stats);
1037         if (!ret) {
1038                 mutex_lock(&c->sb_lock);
1039                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1040                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1041                 c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1042                 bch2_write_super(c);
1043                 mutex_unlock(&c->sb_lock);
1044         }
1045
1046         bch_err_fn(c, ret);
1047         return ret;
1048 }
1049
1050 int bch2_data_job(struct bch_fs *c,
1051                   struct bch_move_stats *stats,
1052                   struct bch_ioctl_data op)
1053 {
1054         int ret = 0;
1055
1056         switch (op.op) {
1057         case BCH_DATA_OP_REREPLICATE:
1058                 bch2_move_stats_init(stats, "rereplicate");
1059                 stats->data_type = BCH_DATA_journal;
1060                 ret = bch2_journal_flush_device_pins(&c->journal, -1);
1061
1062                 ret = bch2_move_btree(c,
1063                                       op.start_btree,   op.start_pos,
1064                                       op.end_btree,     op.end_pos,
1065                                       rereplicate_btree_pred, c, stats) ?: ret;
1066                 ret = bch2_replicas_gc2(c) ?: ret;
1067
1068                 ret = bch2_move_data(c,
1069                                      (struct bbpos) { op.start_btree,   op.start_pos },
1070                                      (struct bbpos) { op.end_btree,     op.end_pos },
1071                                      NULL,
1072                                      stats,
1073                                      writepoint_hashed((unsigned long) current),
1074                                      true,
1075                                      rereplicate_pred, c) ?: ret;
1076                 ret = bch2_replicas_gc2(c) ?: ret;
1077
1078                 bch2_move_stats_exit(stats, c);
1079                 break;
1080         case BCH_DATA_OP_MIGRATE:
1081                 if (op.migrate.dev >= c->sb.nr_devices)
1082                         return -EINVAL;
1083
1084                 bch2_move_stats_init(stats, "migrate");
1085                 stats->data_type = BCH_DATA_journal;
1086                 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1087
1088                 ret = bch2_move_btree(c,
1089                                       op.start_btree,   op.start_pos,
1090                                       op.end_btree,     op.end_pos,
1091                                       migrate_btree_pred, &op, stats) ?: ret;
1092                 ret = bch2_replicas_gc2(c) ?: ret;
1093
1094                 ret = bch2_move_data(c,
1095                                      (struct bbpos) { op.start_btree,   op.start_pos },
1096                                      (struct bbpos) { op.end_btree,     op.end_pos },
1097                                      NULL,
1098                                      stats,
1099                                      writepoint_hashed((unsigned long) current),
1100                                      true,
1101                                      migrate_pred, &op) ?: ret;
1102                 ret = bch2_replicas_gc2(c) ?: ret;
1103
1104                 bch2_move_stats_exit(stats, c);
1105                 break;
1106         case BCH_DATA_OP_REWRITE_OLD_NODES:
1107                 bch2_move_stats_init(stats, "rewrite_old_nodes");
1108                 ret = bch2_scan_old_btree_nodes(c, stats);
1109                 bch2_move_stats_exit(stats, c);
1110                 break;
1111         default:
1112                 ret = -EINVAL;
1113         }
1114
1115         return ret;
1116 }
1117
1118 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1119 {
1120         prt_printf(out, "%s: data type=%s pos=",
1121                    stats->name,
1122                    bch2_data_types[stats->data_type]);
1123         bch2_bbpos_to_text(out, stats->pos);
1124         prt_newline(out);
1125         printbuf_indent_add(out, 2);
1126
1127         prt_str(out, "keys moved:  ");
1128         prt_u64(out, atomic64_read(&stats->keys_moved));
1129         prt_newline(out);
1130
1131         prt_str(out, "keys raced:  ");
1132         prt_u64(out, atomic64_read(&stats->keys_raced));
1133         prt_newline(out);
1134
1135         prt_str(out, "bytes seen:  ");
1136         prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1137         prt_newline(out);
1138
1139         prt_str(out, "bytes moved: ");
1140         prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1141         prt_newline(out);
1142
1143         prt_str(out, "bytes raced: ");
1144         prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1145         prt_newline(out);
1146
1147         printbuf_indent_sub(out, 2);
1148 }
1149
1150 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1151 {
1152         struct moving_io *io;
1153
1154         bch2_move_stats_to_text(out, ctxt->stats);
1155         printbuf_indent_add(out, 2);
1156
1157         prt_printf(out, "reads: ios %u/%u sectors %u/%u",
1158                    atomic_read(&ctxt->read_ios),
1159                    c->opts.move_ios_in_flight,
1160                    atomic_read(&ctxt->read_sectors),
1161                    c->opts.move_bytes_in_flight >> 9);
1162         prt_newline(out);
1163
1164         prt_printf(out, "writes: ios %u/%u sectors %u/%u",
1165                    atomic_read(&ctxt->write_ios),
1166                    c->opts.move_ios_in_flight,
1167                    atomic_read(&ctxt->write_sectors),
1168                    c->opts.move_bytes_in_flight >> 9);
1169         prt_newline(out);
1170
1171         printbuf_indent_add(out, 2);
1172
1173         mutex_lock(&ctxt->lock);
1174         list_for_each_entry(io, &ctxt->ios, io_list)
1175                 bch2_write_op_to_text(out, &io->write.op);
1176         mutex_unlock(&ctxt->lock);
1177
1178         printbuf_indent_sub(out, 4);
1179 }
1180
1181 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1182 {
1183         struct moving_context *ctxt;
1184
1185         mutex_lock(&c->moving_context_lock);
1186         list_for_each_entry(ctxt, &c->moving_context_list, list)
1187                 bch2_moving_ctxt_to_text(out, c, ctxt);
1188         mutex_unlock(&c->moving_context_lock);
1189 }
1190
1191 void bch2_fs_move_init(struct bch_fs *c)
1192 {
1193         INIT_LIST_HEAD(&c->moving_context_list);
1194         mutex_init(&c->moving_context_lock);
1195 }